Randall Allemang

Peter Avitabile
Editors

Handbook of
Experimental
Structural
Dynamics
Handbook of Experimental Structural
Dynamics
Randall Allemang • Peter Avitabile
Editors

Handbook of
Experimental Structural
Dynamics

With 667 Figures and 70 Tables

123
Editors
Randall Allemang Peter Avitabile
Mechanical and Materials Engineering Structural Dynamics and
University of Cincinnati Acoustic Systems
Cincinnati, OH, USA Laboratory, University
of Massachusetts
Lowell, Lowell, MA, USA

ISBN 978-1-4614-4546-3 ISBN 978-1-4614-4547-0 (eBook)

https://doi.org/10.1007/978-1-4614-4547-0

© The Society for Experimental Mechanics 2022

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Science+Business Media, LLC
part of Springer Nature.
The registered company address is: 1 New York Plaza, New York, NY 10004, U.S.A.
Preface and Introduction

The purpose of the SEM Handbook of Experimental Structural Dynamics is to

provide a summary/overview of information and reference material in the areas of
analytical and experimental structural dynamics that are important to the Society
of Experimental Mechanics. The topics covered in this Handbook all address the
essential material that is germane to the International Modal Analysis Conference
that has been held yearly for close to 40 years beginning in 1982.
This Handbook can be viewed as covering the macro topics of experimental
mechanics while the traditional SEM Handbook of Experimental Solid Mechanics
covers the micro topics of experimental mechanics. In past revisions of the SEM
Handbook of Experimental Mechanics, all subject areas were contained in one
single volume. Due to the expansions of topics, the Society decided to divide the
material into two Handbooks with this Handbook consisting of two volumes.
The content of the SEM Handbook of Experimental Structural Dynamics
primarily involves the areas of structural mechanics (statics and dynamics) that are
served by the SEM International Modal Analysis Conference (IMAC) and covered
by several Technical Divisions (TDs) or Focus Groups (FGs) of the Society. A few
of the topics that will be included in this Handbook will come from TDs and FGs
that are active in the SEM Annual Conference but were not included in the recently
revised SEM Handbook of Experimental Solid Mechanics.
While the general topic of each chapter is defined by the chapter title, in all cases
the content of the chapter reflects the experimental nature of the Handbook and
the focus/content of the almost 40 years of science and technology representative
of the IMAC Conference. Therefore, even a chapter like Finite Element Modeling
reflects the experimental concerns/issues rather than a purely theoretical explanation
of the methodology. Terminology and nomenclature are suggested (using the IMAC
suggested nomenclature list) but not required due to overlapping standards of
nomenclature used in different technical areas. The focus of the material is on
proven methods and not simply a literature review. Some overlap between various
chapters cannot be avoided for several chapters due to the inter-relationship of
certain topics.
Each chapter generally involves a review of the important theory involved in the
specific analytical and experimental methods as well as the specific science and
technology that is involved. Additionally, though, each chapter covers the relevant
practical needs of scientists and engineers who are new to the field involved. In most
v
vi Preface and Introduction

cases both the pertinent theory and particularly the practical issues have not yet been
presented formally in an academic textbook. Each chapter should be a ’must read’
for someone new to the field or for someone returning to the field after an absence.
The reference list in each chapter may not be all inclusive but does consist of the
seminal papers and references for the area of the chapter.
While much of the material in this Handbook may be found scattered in a variety
of different publications, this Handbook attempts to pull together all that material
into one location to facilitate current and future research as this field moves forward.

June 2022 Randall Allemang

Peter Avitabile
Editors
Contents

Volume 1

Part I Sensors and Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1 Recent History of Experimental Structural Dynamics . . . . . . . . . . . 3

R. J. Allemang
2 Sensors and Their Signal Conditioning for Dynamic
Acceleration, Force, Pressure, and Sound Applications . . . . . . . . . . 45
Gary Foss, Jessica Meloy, Mark Valentino, and Patrick Walter
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 103
Paolo Chiariotti, Christian Rembe, Paolo Castellini, and Matt Allen
4 Applied Digital Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Robert B. Randall, Jerome Antoni, and Pietro Borghesani
5 Introduction to Spectral and Correlation Analysis: Basic
Measurements and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Anders Brandt and Stefano Manzoni
6 Frequency Response Function Estimation . . . . . . . . . . . . . . . . . . . . . . 259
A. W. Phillips and R. J. Allemang
7 Random Vibration and Mechanical Shock . . . . . . . . . . . . . . . . . . . . . 349
Thomas L. Paez, Norman F. Hunter, and David O. Smallwood
8 DIC and Photogrammetry for Structural Dynamic Analysis
and High-Speed Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Christopher Niezrecki, Phillip L. Reu, Javad Baqersad,
and Daniel P. Rohe

Part II Modal Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

9 Design of Modal Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481

Thomas Carne, Ralph Brillhart, Daniel Kammer,
and Kevin Napolitano
vii
viii Contents

10 Experimental Modal Analysis Methods . . . . . . . . . . . . . . . . . . . . . . . . 533

R. J. Allemang and D. L. Brown
11 Experimental Modal Parameter Evaluation Methods . . . . . . . . . . . . 615
R. J. Allemang and A. W. Phillips
12 Damping of Materials and Structures . . . . . . . . . . . . . . . . . . . . . . . . . 681
Lothar Gaul and André Schmidt
13 Modal Analysis of Nonlinear Mechanical Systems . . . . . . . . . . . . . . . 799
G. Kerschen and A. F. Vakakis

Volume 2

Part III Analytical/Experimental Modeling Applications . . . . . . . . . . . 831

14 Substructuring Concepts and Component Mode Synthesis . . . . . . . 833

Daniel Rixen
15 Finite Element Model Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
Peter Avitabile and Michael Mains
16 Model Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 897
John E. Mottershead, Michael Link, Michael I. Friswell,
and Carsten Schedlinski
17 Nonlinear System Analysis Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 951
Janette J. Meyer, Raymond M. Bond, and Douglas E. Adams
18 Structural Health Monitoring and Damage Identification . . . . . . . . 989
R. Fuentes, E. J. Cross, P. A. Gardner, L. A. Bull, T. J. Rogers,
R. J. Barthorpe, H. Shi, N. Dervilis, C. R. Farrar, and K. Worden
19 Experimental Dynamic Substructures . . . . . . . . . . . . . . . . . . . . . . . . . 1063
Randall L. Mayes and Matthew S. Allen
20 Structural Dynamics Modification and Modal Modeling . . . . . . . . . 1093
Mark Richardson and David Formenti
21 Toward Robust Response Models: Theoretical and
Experimental Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1153
Nuno Maia, António Urgueira, Raquel Almeida, and Tiago Silva
22 Linear Modal Substructuring with Nonlinear Connections . . . . . . . 1211
Peter Avitabile

Part IV Applications and Miscellaneous Topics . . . . . . . . . . . . . . . . . . . 1245

23 Civil Structural Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1247

Ruben Boroschek and Joao Pedro Santos
Contents ix

24 Aerospace Perspective for Modeling and Validation . . . . . . . . . . . . . 1343

Robert Coppolino
25 Applied Math for Experimental Structural Dynamics . . . . . . . . . . . 1379
Chuck Van Karsen and Andrew Barnard

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1403
About the Editors

Dr. Allemang is a Professor Emeritus and member of

the faculty of the Mechanical Engineering Program in
Mechanical and Materials Engineering, University of
Cincinnati, where he currently also serves as Direc-
tor of the Structural Dynamics Research Laboratory
(UC-SDRL). Dr. Allemang has been actively involved
in the area of experimental structural dynamics and
modal analysis for over forty years, pioneering the
use of multiple input, multiple output estimation of
frequency response functions, developing the concept
of cyclic averaging, formulating the modal assurance
criterion (MAC) and the enhanced frequency response
function and reformulating modal parameter estimation
algorithms into the unified matrix (coefficient) poly-
nomial approach (UMPA). Dr. Allemang has served
on the Advisory Board for the International Modal
Analysis Conference (1981-2015, Chairman, 1986-
1995), on the Editorial Board of Sound and Vibration
Magazine (1983-2017) and on the Scientific Commit-
tee of the ISMA Noise and Vibration Conference at
the Katholieke University of Leuven, Belgium (1984-
present). Dr. Allemang has also served on the Executive
Board of the Society for Experimental Mechanics
(SEM) from 1998-2006, including President for the
Society during 2003-2004. Dr. Allemang is currently
involved as Senior Collaborator for the AFRL Struc-
tural Sciences Center at WPAFB, working on the
verification and validation (experimental quantification
of margin and uncertainty) of experimental and ana-
lytical dynamic models and has been recognized by
three professional societies as a Fellow (SEM, SAE
and ASME). Dr. Allemang has also mentored two PhD
graduates who went on to win Presidential Early Career
Award for Scientists and Engineers (PECASE).
xi
xii About the Editors

Dr. Peter Avitabile is a Professor Emeritus and mem-

ber of the faculty at the University of Massachusetts
Lowell (UML) in the Mechanical Engineering Depart-
ment. He serves a Co-Director of the Structural
Dynamics and Acoustic Systems Laboratory (SDASL)
at UML.
Dr. Avitabile has over 4 decades of experience in
design and analysis using FEM and experimental tech-
niques. His main area of research is structural dynamics
specializing in the areas of modeling, testing and cor-
relation of analytical and experimental models along
with advanced applications for developing structural
dynamic models.
He is a Fellow of the Society for Experimental
Mechanics (SEM), has received the DeMichele award
(2004), has held numerous board and executive posi-
tions in the SEM Organization and the International
Modal Analysis Community and served as President of
SEM during 2016-2017.
Peter has contributed well over 300 technical papers
in the area as well as his 17 year “Modal Space”
article series in the Experimental Techniques magazine
published by the Society for Experimental Mechanics.
Pete has also published a book entitled “Modal Testing:
A Practitioner’s Guide” (Wiley 2017).
Contributors

Douglas E. Adams Vanderbilt University, Nashville, TN, USA

R. J. Allemang Structural Dynamics Research Laboratory, Mechanical and Mate-
rials Engineering, University of Cincinnati, Cincinnati, OH, USA
Matt Allen University of Wisconsin-Madison, Madison, WI, USA
Matthew S. Allen Department of Engineering Physics, University of Wisconsin,
Madison, WI, USA
Raquel Almeida UNIDEMI, Department of Mechanical and Industrial Engineer-
ing, FCT NOVA – Faculty of Science and Technology, Universidade NOVA de
Lisboa, Caparica, Portugal
Jerome Antoni Univ Lyon, INSA-Lyon, Villeurbanne, France
Peter Avitabile Structural Dynamics and Acoustic Systems Laboratory, University
of Massachusetts Lowell, Lowell, MA, USA
Javad Baqersad Department of Mechanical Engineering, Kettering University,
Flint, MI, USA
Andrew Barnard Mechanical Engineering-Engineering Mechanics Department,
Michigan Technological University, Houghton, MI, USA
R. J. Barthorpe Dynamics Research Group, Department of Mechanical Engineer-
ing, University of Sheffield, Sheffield, UK
Raymond M. Bond Vanderbilt University, Nashville, TN, USA
Pietro Borghesani University of New South Wales, Sydney, NSW, Australia
Queensland University of Technology, Brisbane, QLD, Australia
Ruben Boroschek University of Chile, Santiago, Chile
Anders Brandt Department of Mechanical and Electrical Engineering, University
of Southern Denmark, Odense, Denmark
Ralph Brillhart ATA Engineering, Inc., San Diego, CA, USA

xiii
xiv Contributors

D. L. Brown Structural Dynamics Research Laboratory, Mechanical and Materials

Engineering, University of Cincinnati, Cincinnati, OH, USA
L. A. Bull Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
Thomas Carne Sandia National Laboratories, Retired, Albuquerque, NM, USA
Paolo Castellini Università Politecnica delle Marche, Ancona, Italy
Paolo Chiariotti Università Politecnica delle Marche, Ancona, Italy
Robert Coppolino Measurement Analysis Corporation, Torrance, CA, USA
E. J. Cross Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
N. Dervilis Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
C. R. Farrar Engineering Institute, Los Alamos National Laboratory, Los Alamos,
NM, USA
David Formenti BlackHawk Technology, Santa Cruz, CA, USA
Gary Foss The Boeing Company, Chicago, IL, USA
Michael I. Friswell College of Engineering, Swansea University, Swansea, UK
R. Fuentes Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
P. A. Gardner Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
Lothar Gaul Institut für Nichtlineare Mechanik, Universität Stuttgart, Stuttgart,
Germany
Norman F. Hunter Thomas Paez Consulting, Albuquerque, NM, USA
Daniel Kammer University of Wisconsin – Madison, Madison, WI, USA
G. Kerschen Space Structures and Systems Laboratory, Department of Aerospace
and Mechanical Engineering, University of Liège, Liege, Belgium
Michael Link Department of Civil and Environmental Engineering, University of
Kassel, Kassel, Germany
Nuno Maia Department of Mechanical Engineering, LAETA/IDMEC/IST (Insti-
tuto Superior Tecnico), University of Lisbon, Lisbon, Portugal
Michael Mains University of Cincinnati, Cincinnati, OH, USA

Lothar Gaul: deceased.

Contributors xv

Stefano Manzoni Department of Mechanical Engineering, Politecnico di Milano,

Milan, Italy
Randall L. Mayes Sandia National Laboratories, Albuquerque, NM, USA
Jessica Meloy Pensar Development, Seattle, WA, USA
Janette J. Meyer School of Engineering, Vanderbilt University, Nashville, TN,
USA
John E. Mottershead School of Engineering, University of Liverpool, Liverpool,
UK
Kevin Napolitano ATA Engineering, Inc., San Diego, CA, USA
Christopher Niezrecki Department of Mechanical Engineering, University of
Massachusetts Lowell, Lowell, MA, USA
Thomas L. Paez Thomas Paez Consulting, Albuquerque, NM, USA
A. W. Phillips Structural Dynamics Research Laboratory, Mechanical and Materi-
als Engineering, University of Cincinnati, Cincinnati, OH, USA
Robert B. Randall University of New South Wales, Sydney, NSW, Australia
Christian Rembe TU Clausthal, Clausthal-Zellerfeld, Germany
Phillip L. Reu Sandia National Laboratories, Albuquerque, NM, USA
Mark Richardson Vibrant Technology, Inc., Centennial, CA, USA
Daniel Rixen Technische Universität München, Garching, Germany
T. J. Rogers Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
Daniel P. Rohe Sandia National Laboratories, Albuquerque, NM, USA
Joao Pedro Santos Structural Monitoring Unit, National Laboratory for Civil
Engineering, Lisboa, Portugal
Carsten Schedlinski ICS Engineering GmbH, Dreieich, Germany
André Schmidt Institut für Nichtlineare Mechanik, Universität Stuttgart, Stuttgart,
Germany
H. Shi Dynamics Research Group, Department of Mechanical Engineering, Uni-
versity of Sheffield, Sheffield, UK
Tiago Silva UNIDEMI, Department of Mechanical and Industrial Engineering,
FCT NOVA – Faculty of Science and Technology, Universidade NOVA de Lisboa,
Caparica, Portugal
David O. Smallwood Thomas Paez Consulting, Albuquerque, NM, USA
xvi Contributors

António Urgueira UNIDEMI, Department of Mechanical and Industrial Engineer-

ing, FCT NOVA – Faculty of Science and Technology, Universidade NOVA de
Lisboa, Caparica, Portugal
A. F. Vakakis Department of Mechanical Science and Engineering, University of
Illinois at Urbana-Champaign, Champaign, IL, USA
Mark Valentino PCB Piezotronics, Inc., Depew, NY, USA
Chuck Van Karsen Mechanical Engineering-Engineering Mechanics Department,
Michigan Technological University, Houghton, MI, USA
Patrick Walter Texas Christian University & PCB Piezotronics, Inc., Fort Worth,
TX, USA
K. Worden Dynamics Research Group, Department of Mechanical Engineering,
University of Sheffield, Sheffield, UK
 Part I
Sensors and Measurements
Recent History of Experimental Structural
Dynamics 1
R. J. Allemang

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Timeline History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3 Technology Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4 Experimental Structural Dynamics Methods 1965–1985 . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.1 Classification Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Data Acquisition Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Conferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1 ISMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 IMAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.3 Other Conferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6 Publications and Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7 Pioneers/Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.1 Pioneers: Experimental Structural Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.2 Contributors: Handbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7.3 Contributors: SEM IMAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
8 Summary/Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

Abstract

This chapter on the recent history of experimental structural dynamics puts much
of the Handbook in a historical perspective that begins with the development
of digital data methodology and computerized data processing that began in

R. J. Allemang ()
Structural Dynamics Research Laboratory, Mechanical and Materials Engineering, University of
Cincinnati, Cincinnati, OH, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 3

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_1
4 R. J. Allemang

the mid-1960s. Experimental structural dynamics began much earlier with

analog, single frequency data acquisition and mostly visual data processing
that began in the 1800s with the rail and marine industries, particularly when
the steam engine impacted those technologies. The analog, single frequency
data acquisition methodology, continued in the automotive and the aircraft
industries in the first half of the 1900s. This Handbook mostly chronicles data
acquisition and processing methods that began more recently, in the mid-1960s,
with the advent of the Fourier transform, analog to digital data conversion,
and digital minicomputers to the present time period. The Handbook also
discusses many methods and techniques in use during the 1960s and 1970s that
utilize experimentally derived models, both linear and nonlinear, to calibrate and
validate corresponding analytical models. Part of this discussion includes the
issue of the varying dimensionality of the number of degrees of freedom (DOF)
between experimental and analytical models. This chapter also discusses the
researchers and educators that were part of the development of the experimental
structural dynamics methodology in the 1960s to 1980s that led to the current
technical state of the art. This discussion includes the identification of researchers
and educators that were instrumental to the Society for Experimental Mechanics
(SEM) in the development of this area of interest within the Society over the last
50 years.

Keywords

Experimental structural dynamics · Experimental modal analysis · Modal

parameter estimation · Forced normal mode · Phase resonance · Phase
separation · Sinusoidal input-output model · Frequency response function
(FRF) Model · Damped complex exponential response model · General
input-output model · Frequency response function (FRF) · Impulse response
function (IRF) · IMAC Advisory Board · Handbook chapters and authors ·
Pioneers and contributors

Nomenclature

Ni Number of inputs
No Number of outputs
NS Short Dimension (min(Ni , No ))
NL Long Dimension (max(Ni , No ))
Nf Number of spectral lines (frequencies)
Ne Number of effective modal frequencies
N Number of modal frequencies
Fmax Maximum frequency (Hz)
ωk Frequency (rad/sec)
ωmax Maximum frequency (rad/sec)
sk Generalized Frequency - Complex Valued (rad/sec)
λr Complex modal frequency
T Observation period (sec)
1 Recent History of Experimental Structural Dynamics 5

Δf Frequency resolution (Hz)

si Generalized frequency variable
m Model order for denominator polynomial
n Model order for numerator polynomial
[α] Denominator polynomial matrix coefficient
[β] Numerator polynomial matrix coefficient
[I ] Identity matrix
[M] Mass matrix
[C] Damping matrix
[C] Companion matrix (alternate usage)
[K] Stiffness matrix
[A] State space model matrix
[B] State space model matrix
[A] Controls model matrix
[B] Controls model matrix
[C] Controls model matrix
[D] Controls model matrix
[H (s)] Transfer function matrix. (No × Ni )
[H (ω)] Frequency response function matrix. (No × Ni )
Apqr Residue, output DOF p, input DOF q, mode r
[T ] Transformation matrix
[U ] Left singular vector matrix
[Σ] Singular value matrix (diagonal)
[Λ] Eigenvalue matrix (diagonal)
[V ] Right singular vector, or eigenvector, matrix

1 Introduction

The purpose of the SEM Handbook of Experimental Structural Dynamics is to

provide a summary/overview of information and reference material in the areas of
analytical and experimental structural dynamics that are important to the Society
for Experimental Mechanics. The topics covered in this Handbook all address the
essential material that is germane to the International Modal Analysis Conference
that has been held annually for close to 40 years beginning in 1982.
This Handbook can be viewed as covering the macro topics of experimental
mechanics, while the traditional SEM Handbook of Experimental Solid Mechanics
covers the micro topics of experimental mechanics. In past revisions of the SEM
Handbook of Experimental Mechanics, all subject areas were contained in one
single volume. Due to the expansion of topics, the Society decided to divide the
material into two volumes. The content of this SEM Handbook of Experimental
Structural Dynamics primarily involves the areas of structural mechanics (statics
and dynamics) that are served by the SEM International Modal Analysis Conference
(IMAC) and are covered by several technical divisions (TDs) or focus groups (FGs)
of the Society. A few of the topics that will be included in this Handbook will come
from TDs and FGs that are active in the SEM Annual Conference but were not
included in previous SEM handbooks and were not covered in the 2008 revision of
the SEM Handbook of Experimental Solid Mechanics.
6 R. J. Allemang

While the general topic of each chapter is defined by the chapter title, in all cases
the content of the chapter reflects the experimental nature of the Handbook and the
focus/content of the almost 40 years of science and technology representative of the
IMAC Conference. Therefore, even a chapter like finite element modeling (FEM)
reflects the experimental concerns/issues rather than a purely theoretical explanation
of the FEM methodology. Terminology and nomenclature are suggested (using the
IMAC suggested nomenclature list) but not required due to overlapping standards
of nomenclature used in different technical areas. The focus of the material is on
proven methods and not simply a literature review. Some overlap between various
chapters cannot be avoided for several chapters due to the interrelationship of certain
topics. Each chapter generally involves a review of the important theory involved in
the specific analytical and experimental methods as well as the specific science and
technology that is involved. Additionally, though, each chapter covers the relevant
practical needs of scientists and engineers who are new to the field involved. In most
cases both the pertinent theory and particularly the practical issues have not yet been
presented formally in an academic textbook. Each chapter should be a “must read”
for someone new to the field or for someone returning to the field after an absence.
The reference list in each chapter may not be all inclusive but does consist of the
seminal papers and references for the area of the chapter. While much of the material
in this Handbook may be found scattered in a variety of different publications, this
Handbook attempts to pull together all that material into one location to facilitate
current and future research as this field moves forward.

2 Timeline History

Experimental structural dynamics is defined as estimating the dynamic properties

of structures via estimation methods often involving the determination of modal
properties of the structures (modal frequencies and damping factors, modal vectors,
and modal scaling) from experimentally measured data in the time domain. The
history of these methods goes back to the early 1900s or before in terms of visual
or analog methods, mostly conducted with single frequency experiments. Most
experimental methods prior to 1965 were limited by the sensors available to the time
period in addition to the limits of analog measurement methods that used analog
bandpass filters and single frequency methods.
By 1965, sensors to measure analog force and response were available, and
computer-based data acquisition was becoming available to generate digital time
data via analog to digital conversion (ADC) of analog voltage measurements. The
fast Fourier transform (FFT) algorithm made processing of digital data in the
frequency domain possible.
By 1985, several conferences with archived proceedings began to become
available. The current experimental structural dynamics methods are covered in
other chapters in this SEM Handbook. The discussion in this chapter will primarily
focus on the period from 1965 to 1985 where publications and references are harder
1 Recent History of Experimental Structural Dynamics 7

to find. A relatively complete list of references that were available prior to 1982 is
included in the first (1982) IMAC Conference proceedings [1].

3 Technology Developments

Prior to 1965 and the development of the fast Fourier transform (FFT) [2],
only analog methods were available to measure experimental structural dynamics.
Some sensors were developed using variable resistance potentiometers beginning
in the 1850s. These sensors were often directly attached to moving structures
with mechanical connections that involved friction and limitations of very low
frequency. Sensors involving bonded strain gages (1930s) and piezoelectric sensors
(1950s) were developed which made observations other than visual, mechanical, or
resistance measurements possible. By the late 1960s, minicomputer data acquisition
using ADCs made digital time domain measurements along with the associated
digital frequency domain measurement possible.
Beginning in the early 1980s, computer workstation and personal computer
(PC)-based data acquisition began to be available with increased processing speed,
increased memory, and more flexible operating system environments. This time
period also brought synchronous multiple channel data acquisition systems which
made true multiple input, multiple output (MIMO) data acquisition a commonly
used methodology. MIMO data acquisition also made the estimation of close or
repeated modal frequencies possible for first time, and many multiple reference
parameter estimation algorithms were developed over the next 20–30 years.

3.1 Sensors

For the purpose of discussion in this text, only sensors and sensor technologies that
are capable of giving a measurable voltage signal proportional to load or motion will
be discussed. Most of these sensor technologies are still in use today for specialized
applications.

3.1.1 Resistance Technology

Linear and rotational resistance potentiometers originated somewhere around the
mid-1850s. The rotational version was referred to as a rheostat by Sir Charles
Wheatstone about that time. Displacement motion is measured by connecting the
moving structure to the linear or rotational potentiometer so that resistance varies
with the displacement motion. The resistance is used in a voltage divider electrical
circuit to generate a voltage signal proportional to displacement. This technology
is limited to very low frequency and relatively large displacements. This resistance
technology is still used at the present time in automotive applications in the form of
string potentiometers where a multi-turn rotational potentiometer has a string wound
around the rotational axis of the potentiometer.
8 R. J. Allemang

3.1.2 Bonded Strain Gage Technology

Bonded strain gage technology provided a revolution in the development of both
load and response (mostly accelerometers) sensors. The bonded strain gage has a
unique connection to the early creation of the Society for Experimental Mechanics.
The Society for Experimental Mechanics (SEM) began as the New England
Photoelasticity Conference in 1935. This led to the Society for Experimental Stress
Analysis (SESA) founded in 1943. SESA had two foci at that time: photoelasticity
and strain. Strain gages were developed somewhat independently by Edward
Simmons, a graduate student working at the California Institute of Technology
(CalTech), and by J. Hans Meier, a graduate student working at the Massachusetts
Institute of Technology (MIT) for Professor Ruge. Both groups were awarded
patents for their overlapping developments in 1941 which created a conflict. About
the time SESA was founded, the bonded strain gage was of great commercial
interest, and the company Baldwin-Southwark was able to get an agreement from
both inventors in order to produce commercially available strain gages. J. Hans
Meier stayed very active in the area of strain gage technology and SESA becoming
president of the Society in 1951–1952. The SEM name change officially came about
in 1985 [3].
The 1990 article by Peter Stein [4] gives a very complete history of the
development of bonded stain gages and the early developments of sensor technology
based upon bonded strain gages.

3.1.3 Piezoelectric Crystal Technology

In order to overcome some of the limitations of bonded strain gage technology in
1943, a number of companies and research groups began working on piezoelectric
crystal technology. Early technology used ferroelectric ceramics (barium titanate)
and quartz crystals, and these early designs required the development of charge
amplifiers. These charge amplifier designs were used in both load and acceler-
ation sensor designs yielding relatively flat frequency response characteristics to
10,000 Hz as opposed to flat frequency response characteristics to 200 Hz for
bonded strain gage devices. While extremely useful in sensor designs, since the
sensor signals are charge based, the capacitance of the cables connected to the
sensors created some problems when different cable lengths were utilized (different
capacitance). Also note that the piezoelectric crystal technology is limited at the
low-frequency end of the usable sensor range due to the bleed off of the steady-state
charge signal associated with DC and low-frequency signals.
A significant improvement in sensor designs for both load and response sensors
came about as integrated circuit technology developed. A two wire field-effect
transistor (FET) design allowed for the FET to be embedded in the sensor resulting
in low noise, low impedance, voltage level designs. This technology is referred to as
integrated circuit piezotronic (ICP® ) or integrated electronics piezoelectric (IEPE).
Except for shock and very high frequency applications, the (ICP® ) or (IEPE)
designs are the most widely used technology for measuring both force and
acceleration response at the present time. The 2007 article by Pat Walter [5] gives
1 Recent History of Experimental Structural Dynamics 9

a very complete history of the development of piezoelectric technology and its

use in accelerometers. Also, please refer to the chapter in this Handbook entitled
 Chap. 2, “Sensors and Their Signal Conditioning for Dynamic Acceleration,
Force, Pressure, and Sound Applications”.

3.1.4 Inductance and Capacitance Technology

A number of displacement sensors, primarily used in the experimental rotational
analysis area, are based upon inductance or capacitance technology. This sensor
technology yields an output voltage proportional to the gap between the sensor
and an object (normally ferrous material) and is used to measure very small
displacements. The inductance technology essentially involves a very small eddy
current device embedded in nonconductive resin that is placed in close proximity to
a ferrous material. The output of the eddy current circuit is proportional to the gap
between the probe and the ferrous material. This gives a response sensor that has a
relatively flat frequency response beginning at 0 Hz. This eddy current, proximity
probe (now referred to generically as an inductance sensor), began as a sensor
developed in 1961, primarily for use in rotational systems by the Bentley Nevada
Corporation. This sensor is often referred to as a Bentley probe based upon the
original development.

3.1.5 Optical Technology

Two distinct optical sensor technologies have evolved over more recent years. The
first optical technology is the use of laser Doppler velocimetry to measure velocity
normal to the direction of the laser. While this technology was originally used for
alignment of large equipment dating to the 1970s, measurement of velocity based
upon a single beam laser dates to at least the early 1990s. Today this technology
has been extended to large grids of velocity data using scanning lasers and to three
dimension (3D) using multiple, aligned laser heads. Some laser systems use multiple
beam lasers so that a grid of laser beams can give motion over a 4 × 4, 2 × 8,
or 1 × 16 configuration. Laser-based optical technologies are obviously limited
to measurements along the axis of the laser(s) and can be difficult maintaining
resolution if large distances or if there is large displacement motion such as in free-
free testing. Please refer to the chapter in this Handbook entitled  Chap. 3, “Laser
Doppler Vibrometry Measurements in Structural Dynamics” for more complete
information.
The second optical technology is the more recent development of digital image
correlation (DIC). This technology utilizes high-speed video cameras with extensive
post processing of the frame by frame photogrammetry data to yield displacement
data. Each frame of the high-speed video is correlated with successive frames,
and feature extraction methodology is used to reduce the high resolution (1024 ×
1024 pixels or greater) to a series of 8 × 8 pixels in a time stream. The centroids
of these 8 × 8 pixels essentially serve as virtual displacement sensors. Much of this
technology was developed for traditional static displacement and strain applications
within SEM by Dr. Michael Sutton (another interesting aside is that Dr. Sutton has
also recently served as president of SEM). Resolution of DIC technology is limited
10 R. J. Allemang

to a number of pixels in each frame of the video (e.g., 1024 × 1024), the size of
the pixels used to create a virtual displacement sensor (e.g., 8 × 8), and the optical
frame physical size of the camera (e.g., 100 feet × 100 feet versus 1 inch × 1 inch).
DIC can be very effective for relatively small objects but will become limited as
the field of view is large. Note that the sampling frequency of the high-speed video
cameras is a form of digital data acquisition, and anti-aliasing filtering can be an
issue. For this reason, the maximum frequency of interest is often limited to one half
of the video sampling frequency (framing rate). For complete details of applying
DIC optical technology to the measurement of structural systems, please refer to
the chapter in this Handbook entitled  Chap. 8, “DIC and Photogrammetry for
Structural Dynamic Analysis and High-Speed Testing”.
Both methods of optical technology require placement of optical devices (lasers
or high-speed cameras) in the region of the test article, with a primary orientation
normal to the surface being measured. Objects that have significant 3D character-
istics present more difficulty and require multiple optical systems or a number of
different optical devices synchronized together. Both methods may require surface
treatment (speckling or photoluminescence dots) to obtain optimal measurement
conditions. Both methods are most easily applied to 1D or 2D measurement
situations although 3D measurements are now gaining applications.

3.2 Data Acquisition

Analog data acquisition began with visual observations and gradually moved to
various forms of strip chart recording. This was followed by oscilloscope time
traces, which included two channel Lissajous patterns and analog voltage meters.
All of these devices were made more useful via the use of analog signal filtering,
especially narrow bandpass filtering. Included along this path was the development
of hybrid real-time analyzers that implemented a discrete frequency methodology
via time compression (or sometimes referred to as frequency shifting or heterodyne
technology).
Digital data acquisition became the default methodology beginning in the
late 1960s with the development of the FFT algorithm [2], the development of
minicomputers that implemented the FFT algorithm in relatively real time and the
development of analog to digital conversion (ADC) of continuous time data.

3.2.1 Analog Technology

Most analog methods before the 1960s used single-frequency excitation and the
associated response with analog, bandpass filtering. If input and response channels
were measured simultaneously, the two time domain signals could be presented
on a two-channel oscilloscope in an X-Y format to generate a Lissajous pattern.
This single pattern gave the input-response ratio between the two sinusoidal signals,
and the major axis of the Lissajous ellipse gave the phase relationship. This was
the beginning of the concept of the transfer function analyzer (TFA) developed by
Spectral Dynamics [6].
1 Recent History of Experimental Structural Dynamics 11

The TFA used very slowly swept sine methodology with tracking narrow
bandpass analog filters on each of two channels. The sine frequency was used
to adjust the center frequency of narrow bandpass analog filters applied to both
channels being measured. These two channels were compared using a Co-Quad
meter to determine the coincident information in the two signals separately from the
quadrature information between the two signals. Today, this information is referred
to as the real and imaginary parts of the frequency response function (FRF) at one
single frequency. The analog output of the Co and Quad meter was plotted as the
frequency was slowly swept along to generate an analog version of the FRF.
A structural system with light to moderate damping required a very slowly swept
sine to allow for structural system response to stabilize, frequency by frequency.
The stabilization time required by the narrow bandpass analog filters also limit the
speed at which the sine wave can be swept [6].
A short time later in the 1960s and early 1970s, a number of companies developed
real-time analyzers (RTA), most notably Spectral Dynamics and Federal Scientific
Corp [7]. The real-time analyzer had many names at that time that included
time compression methodology that was based upon tape recorder playback at
altered playback tape speeds. This was also referred to as frequency shifting or
heterodyne technology. Most of these technologies involved only single channel
analysis of measurement data but gave real-time analysis of measured data in terms
of the frequency content. As with the TFA, the RTA generally had no data storage
capability, and data was plotted to paper or captured with a photo [8].

3.2.2 Digital Technology

While the TFA and the RTA methodologies gave apparent frequency by frequency
information, these technologies were at best considered hybrid technologies. Very
useful at the time they were developed but soon to be mostly eclipsed by the true
digital technologies. The digital technology depended on two major innovations:
(1) development of ADCs and (2) implementation of the FFT algorithm. From a
practical point of view, these two innovations arrived right when minicomputers
were first becoming available. This made everything possible.

Analog to Digital Conversion Development of analog to digital conversion (ADC)

hardware began in the 1960s once the need was identified by the potential
use of the FFT algorithm. Most of the initial designs were based upon circuit
board implementations of 8 bit ADCs. These circuit board designs gave way to
potted modules and with the development of integrated circuits, true chip-based
technology. Today, ADCs are rarely developed for measurement applications since
audiovisual hardware uses six or more channels of 24 bit or higher ADCs on a
single chip. Initial ADC designs were 8 bit and 10 bit ADCs giving 256–1024 levels
of resolution over a voltage range (typically +V to −V). By 1980, most ADCs were
12–14 bit giving 4096–16384 levels of resolution over a voltage range. Today most
ADCs are quite inexpensive and are based upon delta-sigma ADC designs that are
popularized by current audio-video applications. These delta-sigma ADC designs
are 24–40 bit equivalent designs, and some implementations use multiple ADCs to
get even higher bit equivalent designs. Note that these higher bit designs allow a
12 R. J. Allemang

single large voltage range (+V to −V) without the need for multiple voltage ranges
and/or auto-ranging of the ADC.

Minicomputer FFT systems Many of the first implementations of the FFT

algorithm occurred in physically large mainframe computers that were the basis for
engineering and science computations in the 1960s. A few of these large mainframe
computers were hybrid computers with an analog computer directly connected to
the digital computer. For experimental studies that could be placed in this computing
environment, these hybrid computers are where the potential of the FFT algorithm
was unlocked. Some large mainframe computers used measured digital time data on
magnetic tape as an operational methodology.
The implementation of the fast Fourier transform (FFT) in a minicomputer
environment allowed for users in a wide variety of industries to begin to use the
technology. The Hewlett Packard Company and Time Data Corporation (eventually
purchased by General Radio) both developed minicomputer FFT systems at the end
of the 1960s. These two implementations were similar: Hewlett Packard utilizing
the HP minicomputer and Time Data utilizing the PDP minicomputer. Both of these
initial implementations utilized from two to four channels of data acquisition with
sampling frequencies of 50 KHz or above make them suitable for structural and most
acoustic measurements. By the mid-1970s, a second wave of FFT analyzers was
being developed by other manufacturers, and smaller, more portable, FFT analyzers
were available.
Beginning in the late 1970s, multiplexed data acquisition was being implemented
to allow for more channels of acquisition. However, the multiplexing sample rate
induced a time delay, channel to channel, that limited multiple channel use to
relatively low frequencies, acceptable for most structural dynamics applications.
True multichannel, multiple input, multiple output (MIMO) capabilities became
available around 1990.
Further perspectives concerning two of the companies (Brüel & Kjær [9, 10]
and Spectral Dynamics [11]) involved in many of these developments have been
recently published. Unfortunately no such historical look back at the early years at
the Hewlett Packard Company and Time Data Corporation could be found.
A good perspective of the state of vibration equipment (shakers, test equipment,
etc.) in the late 1960s can be found in articles by Tustin [12] and Bickel [7].
Welaratna [13] gave a summary of the development of hybrid and digital analyzers
in his 1997 paper. A recent look back by Lang [14] also adds some detail to last
50 years of experimental equipment development.

4 Experimental Structural Dynamics Methods 1965–1985

In reviewing the literature in the area of experimental structural dynamics (experi-

mental modal analysis), some sort of outline of the various techniques is helpful in
1 Recent History of Experimental Structural Dynamics 13

categorizing the different methods that have been developed over the last 55 years.
Most of these methods began during the years 1965–1985.

4.1 Classification Methods

One approach is to group the methods according to whether one mode or multiple
modes will be excited at one time. The terminology that is used for this is:

• Phase resonance (one mode excited, all other modes suppressed)

• Phase separation (all modes excited simultaneously)
At the current time, almost all experimental modal analyses would fall into the
phase separation category. Phase resonance methods are used by an increasingly
smaller group of aerospace testing activities.
A slightly more detailed approach, and the one that is used in the following text,
is to group the methods according to the type of measured data that is acquired.
When this approach is utilized, the relevant terminology is:

• Sinusoidal input-output model

• Frequency response function model
• Damped complex exponential response model
• General input-output model

A very common concept in comparing and contrasting experimental modal

analysis methodologies that is often used in the literature is based upon the type
of model that will be used in the modal parameter estimation stage that follows the
acquisition of the data. The relevant nomenclature for this approach is:

• Parametric model (Unknowns have physical significance or meaning)

– Modal model
– [M], [C], [K] Model
• Nonparametric model (Unknowns are mathematical conveniences)
– Polynomial model
– Autoregressive moving-average (ARMA) model
– [A], [B] Model
– [A], [B], [C], [D] Model

Finally, the different experimental modal analysis approaches may be grouped

according to the domain that the modal parameter estimation model will be
formulated. The relevant nomenclature for this approach is:

• Time domain
• Frequency domain
• Spatial domain
14 R. J. Allemang

Regardless of the approach used to organize or classify the different approaches

to generating modal parameters from experimental data, the fundamental underlying
theory is the same. The differences largely are a matter of logistics, user experience
requirements, and numerical or compute limitations, rather than a fundamentally
superior or inferior method.

4.2 Data Acquisition Classification

Over the past 50 years, at least four general categories of experimental structural
dynamics methods, which are based upon the type of data that is acquired, can be
identified as follows:

• Sinusoidal input-output model

• Frequency response function
• Damped complex exponential response (free decay and impulse response func-
tions)
• General input-output model

Historically, the modal characteristics of mechanical systems have been esti-

mated by techniques that fall into either the first or second category. The experi-
mental modal analysis methods that fall into the last two categories are composite
approaches that utilize elaborate parameter estimation algorithms based upon
structural models. This section reviews and provides references for the initial work
in each of these areas. Over the years from 1965 until the late 1980s. Methods
developed from the late 1980s to present are covered in later chapters in this
Handbook.
In order to evaluate and improve any approach to experimental modal analysis,
the relative merits of all viable techniques must be well understood. To that goal,
many articles have been written to try to compare and contrast the value of one
method over another. Unfortunately, most of these comparisons have been heavily
concerned with differences that are a function of specific implementations of the
various techniques. These comparisons were also potentially biased by the expertise
of the test engineers being restricted to only one of the areas of testing. Since each
method involves very special testing awareness, this sort of analysis has limited
value.
In the evaluation of experimental modal analysis methods, the differences in
the theoretical approach are obviously of prime concern. Since most experimental
modal analysis methods involve similar theoretical basis, the only significant areas
of difference concern the concept of real versus complex modal vectors, the explicit
measurement of the input, and the different numerical approaches used. The debate
over the need to describe complex valued modes of vibration may never end.
Certainly, the concept of a complex mode, since it contains a real mode as a special
case, appears to be the most general case. Likewise, some experimental modal
analysis methods do not require the measurement of the input. While this can be
1 Recent History of Experimental Structural Dynamics 15

advantageous at times where the implicit nature of the input is known or assumed,
it seems prudent, where the input can be measured, to do so.
Beyond the direct theoretical differences, though, there are several key evaluation
considerations which may or may not be a direct function of the theory. The
availability of confidence factors, the potential for implementation, stability and
precision of the solution algorithm, sensitivity to random and/or bias errors in
the measured data, and the need for operator expertise may control the ability
to estimate valid modal parameters. Specifically, through the knowledge of these
aspects with respect to other experimental modal analysis methods, the use of each
of the approaches may be enhanced due to this transfer of technology between the
methods.
Generally, the methods that utilize frequency response function data, damped
complex exponential response data, and/or general input-output data can all be
explained using a unified matrix polynomial approach (UMPA). Most modal param-
eter estimation algorithms can be reformulated into this consistent mathematical
formulation with a corresponding set of definitions and unifying concepts [15].
Particularly, this matrix polynomial approach is used to unify the presentation
with respect to current algorithms such as the least-squares complex exponential
(LSCE), the polyreference time domain (PTD), Ibrahim time domain (ITD),
eigensystem realization algorithm (ERA), rational fraction polynomial (RFP),
polyreference frequency domain (PFD), and the complex mode indicator function
(CMIF) methods. The unified matrix polynomial approach provides a common
formulation that encourages a discussion of the similarities and differences of the
commonly used methods as well as a discussion of the numerical characteristics.
Complete details concerning all experimental structural dynamics methods in use
today can be found in another Handbook chapter titled  Chap. 10, “Experimental
Modal Analysis Methods.”

4.2.1 Sinusoidal Input-Output Method

Methods covered by this category involve excitation that consists of only one fre-
quency during the observation period. While excitation involves only one frequency,
the response will initially involve many frequencies due to the initiation of the
excitation (transient). Even after this initial transient, the response may contain
energy at more than one frequency due to the harmonic distortion of the excitation
caused by system nonlinearities. This harmonic distortion is normally removed by
filtering the response before the data is processed, leaving a single frequency of
information in both the input and output signals.
The sinusoidal input/output methods require minimal data acquisition capa-
bilities but generally involve more sophistication in the test setup or in the
post-processing of the acquired data. Since only a single frequency is present in the
input and output signals, analog time domain methods or, at most, small block size
fast Fourier transform (FFT) methods can be used to determine signal amplitudes.
Depending upon the approach used, this data will yield the modal parameters
somewhat directly or will require considerable post-processing. In the forced normal
mode approach, the configuration of the test setup (location and phasing of multiple
16 R. J. Allemang

exciters) yields the modal parameters somewhat directly. In the forced response
decomposition approach, the test setup is very general (and similar), but the modal
parameters are found using an elaborate post-processing procedure.

Forced Normal Mode Excitation Method The forced normal mode excitation
method of experimental modal analysis is the oldest approach to the estimation
of dynamic structural parameters. This approach is the first method to use the
application of multiple inputs in the estimation of modal parameters. Currently, this
method is still used in the aerospace industry for ground vibration testing of aircraft
structures. This method was originally outlined in an article by Lewis and Wrisley
in 1950 [16] and begins with the matrix form of the differential equation for the
system being tested, Equation 1.

[M] {ẍ(t)} + [C] {ẋ(t)} + [K] {x(t)} = {f (t)} (1)

Very simply stated, Lewis and Wrisley found that a number of exciters, utilizing
a common frequency and monophase amplitudes, could be tuned to exactly balance
the dissipative forces in a structure. This is represented by Equations 2 and 3 and
occurs when the phase lag angle φ lags the input force by 90◦ at every response
location.

{f (t)} = {F } sin(ωt) {x(t)} = {X} sin(ωt + φ) (2)

{f (t)} = [C] {ẋ(t)} (3)

When the force balance is achieved, the differential equations of motion describ-
ing the structure can be reduced to the undamped homogeneous differential
equations of motion at that particular frequency. This is represented by Equation 4

[M] {ẍ(t)} + [K] {x(t)} = {0} (4)

In order to more thoroughly explain this phenomena, De Veubeke published

an article [17] which explains the theoretical basis for this testing in terms of
characteristic phase lag theory. Further development of the practical application of
this theory was enhanced by the concept of effective number of degrees of freedom
by Trail-Nash [18].
This concept explains that the required number of exciters is a function of the
effective number of degrees of freedom, not the total number of degrees of freedom.
The effective number of degrees of freedom (Ne ) is a function of modal density and
damping. Finally, Asher utilized the determinant of the real part of the frequency
response matrix to locate damped natural frequencies and determine the effective
number of degrees of freedom [19]. Most of the work done since 1958 has been
concerned with improvements in the implementation of the method, primarily the
force appropriation [20, 21] using digital data and FFT data processing.
1 Recent History of Experimental Structural Dynamics 17

One of the most advanced implementations of this method involves approxi-

mately 500 channels of data acquisition and co/quad analysis equipment controlled
from a minicomputer. Extensive tuning criteria are utilized as well as real-time
animated displays of the modal vector as well as of the out of phase response.
This can be particularly useful for optimum exciter location as well as force
appropriation. Once a modal vector is tuned using a 90◦ phase lag angle criteria, the
excitation frequency can be varied with no change in modal vector. Theoretically,
this can be used as a check to determine whether the modal vector has been
adequately tuned.
In addition to this potential confidence check, the excitation can be removed from
the system once a modal vector is tuned. If the modal vector contains only responses
due to a single mode of vibration, the exponential decay at all response positions
should contain only the excitation frequency, and the envelope of exponential decay
should give an accurate estimate of the system damping.
The forced normal mode excitation method works well in the presence of
proportional damping but theoretically does not include the concept of complex
modes of vibration nor the concept of repeated and multiple roots. (Proportional
damping is a mathematical approach to the description of damping that states that
the damping matrix resulting from whatever damping mechanism that is present
is either proportional to the mass matrix, the stiffness matrix, or to some linear
combination of the two.) Due to this theoretical limitation, the practical application
of the 90◦ lag criteria is normally applied only to within plus or minus 10◦ .
Likewise, added difficulty is encountered in evaluating the exponential decay purity
as well as force appropriation. Much work has been done on automated tuning
algorithms to alleviate this. These algorithms alter excitation magnitude and phase
to try to achieve 90◦ phase lag criteria under severe impedance matching situations.
Unfortunately, the location of the excitation cannot be evaluated automatically in
this process.

Forced Response Decomposition Method The forced response decomposition

method uses an array of exciters (multiple input) to excite the system into a
forced response at a single frequency. While the magnitude of each input and the
phasing between inputs may be chosen randomly or according to some particular
regime, the inputs are held constant during the observation period. Therefore, after
the initial transient decays, the response is a steady-state forced response of this
system. A forced response vector is created by using all output points of interest
simultaneously. This is represented by Equations 5 through 7.

[M] {ẍ(t)} + [C] {ẋ(t)} + [K] {x(t)} = {f (t)} (5)

{f (t)} = {F } sin(ωk t) (6)

{x(t)} = {X} sin(ωk t − φk ) (7)

18 R. J. Allemang

The forced response vector that is generated, by definition based upon linear
vibration theory, must be a linear superposition of the modal vectors of the system
(expansion theorem). Since in a given frequency range of interest, N modal vectors
will contribute to the response, the individual modal vectors cannot be determined
from one forced response vector. If N or more independent forced response vectors
can be generated, the N modal vectors can be determined. A large number of
independent forced response vectors can be generated by two approaches. First
of all, at a single frequency, many different input vectors can be generated with
randomly chosen magnitudes and relative phasing. Each of these choices will
yield a potentially independent forced response vector. Secondly, this process can
be repeated for different frequencies which will, once again, yield potentially
independent forced response vectors. Since the degree of independence among the
forced response vectors is unknown and since the best estimate of the modal vectors
is desired, many more forced response vectors, compared to the number of expected
modal vectors (N ), are acquired.
The post-processing of the forced response vectors involves using a singular
value decomposition of the data spanned by the forced response vectors. If NV
forced response vectors are acquired (where NV  N ), the number of significant
singular values in the NV × NV data space is an indication of the number of
contributing modal vectors in the data. The singular vectors, associated with the
significant singular values, provide a transformation matrix to transform the forced
response vectors to modal vectors. Frequency and damping values are found in a
second stage solution process. If frequency response functions or the input vectors
required to force a specific normal mode are desired, these characteristics can be
found at this time as well.
The forced response decomposition method, in one form or another, is currently
receiving much research attention [22, 23].
While the methods have not been used commercially to this point, the methods
are very attractive due to minimal data acquisition requirements (when the data
acquisition has been specifically designed for this method) and due to sophisticated
post-processing techniques which require minimal computational power (micro-
computers or minicomputers).

4.2.2 Frequency Response Function Method

The frequency response function method of experimental modal analysis is the
most commonly used approach to the estimation of modal parameters. This method
originated as a testing technique as a result of the use of frequency response
functions in the forced normal mode excitation method to determine natural
frequencies and effective number of degrees of freedom. With the advent of the
computer and minicomputer, the frequency response function method became a
separate, viable technique [24, 25, 26].
In this method, frequency response functions are measured using excitation at
single, or multiple, points. The relationships between the input (F (ω)) and the
response (X(ω)) for both single and multiple inputs are shown in Equations 8
through 12.
1 Recent History of Experimental Structural Dynamics 19

Single Input Relationship

Xp = Hpq Fq (8)

While Equation 8 could be used directly to estimate the FRF, averaged power
spectra methods are normally used to reduce noise and get a best least-squares
estimate if the FRF. This formulation is shown in Equation 9.

GXF pq
Hpq = (9)
GF F qq

Multiple Input Relationship

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X1 H11 · · · · · · · H1q F1
⎢X ⎥ ⎢H · ⎥ ⎢F ⎥
⎢ 2⎥ ⎢ 21 ⎥ ⎢ 2⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ · ⎥ =⎢ · · ⎥ ⎢ · ⎥ (10)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ · ⎦ ⎣ · · ⎦ ⎣ · ⎦
Xp N Hp1 · · · · · · · Hpq N Fq N ×1
o ×1 o ×Ni i

While Equation 10 represents the multiple input, multiple output (MIMO)

relationship, averaged power spectra are used to estimate the FRF matrix, typically
one row at a time. The MIMO estimation procedure utilizes an inverse of the [GF F ]
matrix. This places a numerical constraint on the [GF F ] matrix. Details concerning
the estimation of FRFs can be found in the Handbook chapter titled  Chap. 6,
“Frequency Response Function Estimation.”

[H ] [GF F ] = [GXF ] (11)

[H ] = [GXF ] [GF F ]−1 (12)

If all or part of the elements of the frequency response matrix can be measured,
each column will contain information which can be used to estimate modal vectors.
Since the frequency response matrix is considered to be symmetric due to the
Maxwell-Betti relations, each row will also contain the information needed to
estimate modal vectors.
Often Equations 13 through 16 and Equation 17 are altered by an assumption
of real modes, a specific damping mechanism, or known system poles. Under such
assumptions, the estimation of the modal parameters may become simpler. Most
current research and development in the area of the frequency response function
method involves the modal parameter estimation algorithms that are related to the
20 R. J. Allemang

time or frequency domain models equivalent to Equations 13 through 17. Much

of this work involves algorithms that utilize as much of the redundant information
within multiple rows and columns of the frequency response function matrix as
possible.
Modal parameters are estimated from FRFs using a variety of modal parameter
estimations today, but during the time period from 1965 to 1985, modal parameter
estimation was limited to mostly single degree of freedom (SDOF), single reference
methods. The time period from 1978 to 1985 was the beginning of multi-reference
modal parameter estimation algorithms that are now widely used to process MIMO
FRFs into modal parameters.
The frequency response functions are used as input data to modal parameter
estimation algorithms that estimate modal parameters using a frequency domain
model. Through the use of the fast Fourier transform, the Fourier transform of the
frequency response function, the impulse response function, can be calculated for
use in modal parameter estimation algorithms involving time domain models:


N
[Ar ]NL ×NS [A∗r ]NL ×NS 
2N
[Ar ]NL ×NS
[H (ωi )]NL ×NS = + = (13)
j ωi − λr j ωi − λ∗r j ωi − λr
r=1 r=1


2N Apqr NL ×1
Hpq (ω) NL ×1
= (14)
j ω − λr
r=1


2N
[ Ar ]NS ×NL
[ H (ω) ]NS ×NL = (15)
j ω − λr
r=1

Noting that

Apqr = Lpr ψqr (16)

When Equations 13 through 16 are applied to the estimation of modal vectors,

the relationships between the references (known as the modal participation vectors
([L])), and the complex valued modal frequencies, λr , are already known. This
allows the equation to be rearranged to take this a priori information into account.
Most of the work over the last 25 years has focused on the measurement
of multiple columns of the frequency response function matrix simultaneously.
This work has involved establishing the numerical and excitation requirements for
solving the relationship identified in Equation 10, developing alternate estimation
algorithms and developing modal parameter estimation algorithms that are matched
to this new data acquisition/estimation procedure. This work has revolutionized
experimental modal analysis testing for several reasons. First of all, in order to be
sure that all modal vectors have been found experimentally, a number of excitation
(reference) points must be utilized, either one at a time or simultaneously. This
1 Recent History of Experimental Structural Dynamics 21

minimizes the possibility of exciting the system at or near a node of one of the modal
vectors which would provide inaccurate estimates of that modal vector. Secondly,
multiple columns (or rows) of the frequency response function matrix are necessary
for the detection of repeated or pseudo-repeated (close) modal frequencies. The
ability to measure, detect, and identify the presence of repeated or close modes
was not generally possible prior to this research. Finally, the increased number of
measurements per measurement cycle obtained using multiple inputs does not affect
the time required to acquire frequency response function data adversely. Therefore,
a more complete set of data, allowing for a more complete dynamic model of the
system to be validated, is possible in the same measurement time.

4.2.3 Damped Complex Exponential Methods

The damped complex exponential response methods of experimental modal analysis
are approaches that have received considerable attention in the 1970s. These
methods are normally formulated to utilize data corresponding to the free decay
of a system generated by the release of an initial condition but apply quite generally
to impulse response function data as well. Since impulse response function data
is scaled to include the forcing condition, use of this method on impulse response
function data yields properly scaled modal parameters that can be used to calculate
proper modal scaling (generalized mass and stiffness). This is not possible if free
decay responses are used. Even so, the formulation of the impulse response function
generally involves the computation of the frequency response function (FRF) via
a fast Fourier transform (FFT), potentially introducing digital signal processing
(DSP) bias errors such as leakage, which may degrade the estimation of the modal
parameters.
In order to obtain estimates of the modal vectors, the damped complex exponen-
tial response functions, normally the impulse response functions, are used as input
in a parameter estimation scheme based on one of Equation 17. Note that if the
true damped complex exponential response function can be measured in the time
domain, such as a free decay response, bias errors such as leakage will not be a
problem since the fast Fourier transform is not used.


N
∗ 
2N
[h(t)]NL ×NS = [Ar ]NL ×NS eλr t + [A∗r ]NL ×NS eλr t = [Ar ]NL ×NS eλr t
r=1 r=1
(17)

While the current implementation of methods based upon damped complex

exponentials is relatively recent, the basis of much of the work was formulated in the
eighteenth century by Prony [27]. Currently, the three approaches, that are widely
used, are the Ibrahim time domain (ITD) approach, the polyreference time domain
(PTD) approach, and the eigensystem realization algorithm (ERA) approach. These
methods are discussed briefly in the following sections as well as in detail in the
Handbook chapter on  Chap. 10, “Experimental Modal Analysis Methods”.
22 R. J. Allemang

Ibrahim Time Domain (ITD) Approach One practical implementation of the

damped complex exponential method is the Ibrahim time domain (ITD) method
[28, 29] developed to extract the modal parameters from damped complex expo-
nential response information. Digital free decay response data are measured at
various points on the structure. If response data from all the selected measurement
positions cannot be obtained simultaneously because of equipment restrictions, a
common position is retained between measurement groups. A recurrence matrix is
created from the free decay data, and the eigenvalues of this matrix are exponential
functions of the poles of the system, from which the poles are easily computed.
The eigenvectors of the recurrence matrix are response residues, from which the
mode shapes are determined. The Ibrahim time domain method generates a matrix
polynomial characteristic equation with matrix dimension equal to the number of
response sensors (No ) and low model order.
The damped complex exponential response method is rather straightforward
in application if the necessary data acquisition hardware, computer facilities, and
software are available. This approach computes the poles and residues based upon
a specific initial vibration condition of the structure. A number of different initial
conditions can be established, analogous to the practice of using several exciter
positions in ordinary single input modal surveys, until all the important modes have
been excited. All the modes cannot be established from one exciter position, and
likewise all the modes cannot be determined from one initial condition.
Although this technique is based upon free decay data, the ITD method can also
be used with operating inputs if the free decay is computed from the operating
inputs by using random-decrement averaging or from measured auto- and cross-
correlation functions. Again, it should be emphasized that this can only be done if
there are no poles or zeros in the input spectrum in the frequency range of interest.
An additional development by Ibrahim with respect to this technique is the concept
of modal confidence factor [29].
The modal confidence factor is a complex number calculated for each identified
mode of the structure, while undergoing an exponential decay form of vibration
test. The modal confidence factor is based upon the modal deflection at a particular
measurement point being related to the modal deflection at that same measurement
point at any time earlier or later in the free decay response. Therefore, if modal
vectors are estimated from exponential decay data and two separate estimates are
calculated from sets of data taken some fixed time Δt apart, the relationship between
the measured second estimate of the modal vector and the calculated estimate of the
modal vector based upon the measured first estimate of the modal vector is defined
as the modal confidence factor. The purpose of the modal confidence factor is to
provide an indicator for determining whether an estimated modal vector is real or
computational.

Polyreference Time Domain (PTD) Approach In 1979, a least-squares complex

exponential (LSCE) algorithm was developed to utilize all measured impulse
response function (IRF) data (or free decay data) to estimate a single set of modal
frequencies [30]. Note that IRF data is commonly estimated from the inverse FFT
1 Recent History of Experimental Structural Dynamics 23

of the measured FRFs. While the LSCE method could utilize data from multiple
references (inputs), the algorithm did not formulate the estimation problem in
a matrix sense and only used a scalar coefficient polynomial solution. Most of
the comments relative to the ITD approach can also be repeated with respect to
the LSCE approach. Of particular importance, once again, is the desirability of
acquiring all response data simultaneously to reduce time invariance problems.
A more complete MIMO implementation of the damped complex exponential
approach to experimental modal analysis is the polyreference time domain (PTD)
approach developed by Vold [31, 32, 33]. In contrast to the LSCE approach,
the polyreference time domain approach utilizes all measured damped complex
exponential information, from all references or initial conditions, simultaneously in
the estimation of modal frequencies. Since multiple initial conditions or reference
data is accounted for in the algorithm, the polyreference time domain method is the
first estimation algorithm to be able to solve for closely spaced or repeated roots
(matrix coefficient polynomial solution). Additionally, the PTD approach broke
new ground in the estimation of a single modal coefficient for each measurement
degree of freedom in the presence of multiple initial conditions or references.
This characteristic is shared by the frequency response function method when the
polyreference frequency domain (PFD) approach is used as the parameter estimation
algorithm and with some of the approaches within the mathematical input-output
model methods. The formulation of the algorithm such that constraints are included
to account for redundant information is an advantage but requires that the total
dataset be acquired so as to match this assumption. The data acquisition best
matches the analysis procedure when all of the data can be acquired simultaneously.
The polyreference time domain method generates a matrix polynomial characteristic
equation with matrix dimension equal to the number of reference sensors (Ni for a
multi-shaker test and No for a roving impact hammer test) and high model order.
While much of the work utilizing the polyreference approach is quite recent,
the evaluation of the method based upon comparisons between experimentally
measured and synthesized frequency response functions is quite impressive when
compared to other modal parameter estimation approaches.

Eigensystem Realization Algorithm Approach The eigensystem realization

algorithm (ERA) approach is another technique that is basically an extended version
of the Ho-Kalman system realization algorithm [34, 35, 36]. The ERA algorithm
was developed at NASA-Langley Research Center under an interdisciplinary effort
involving structural dynamics and controls. This method is similar to the other
damped complex exponential methods in that all involve solutions of a matrix
eigenvalue problem. Since the ERA approach utilizes multiple reference data,
the ERA approach is similar to the polyreference time domain (PTD) approach.
This means that repeated roots can be identified with this approach as well as the
polyreference time domain approach. Other significant attributes of this approach
include the extensive use of accuracy indicators to assess effects of noise and
nonlinearities as well as rank information provided by singular value decomposition
techniques.
24 R. J. Allemang

The ERA approach is based upon well-established realization (state-space) the-

ory using the concepts of controllability and observability. The approach determines
a complete state-space model based upon the important principles of minimal
realization theory attributed to Ho and Kalman. The Ho-Kalman procedure uses
a sequence of real matrices known as Markov parameters (impulse response
functions) to construct a state-space representation of a linear system. The ERA
approach begins with a block data matrix formulated from damped complex
exponential functions, such as free decay responses. This block data matrix is
similar to a general Hankel matrix and includes information from several initial
conditions and a weighted set of damped complex exponential functions. The
weighted set of functions means that points of interest or points with large response
can be emphasized without loss of capability of the method. The state-space
matrices are found from the block data matrix by factorization of the block data
matrix using singular value decomposition. Based upon the rank evaluation of the
block data matrix in this factorization procedure, a state-space set of matrices
can be formulated based on the reduced order. In eigenvalues and eigenvectors
of this reduced order, state-space model are then found. Accuracy indicators
such as the rank of the block data matrix, modal amplitude coherence, modal
phase collinearity, and data reconstruction are used to identify the final set of
modal parameters. The eigensystem realization algorithm (ERA) generates a matrix
polynomial characteristic equation with matrix dimension equal to the number of
response sensors (No ) and low model order.
The ERA approach is a recent method that demonstrates extensive use of accu-
racy indicators. Several studies comparing multiple reference algorithms indicate
good agreement between all methods, but the identification of nonrealistic modal
parameters is still a significant problem. The ERA approach, through the extensive
development and use of accuracy indicators, attempts to deal with this part of the
identification problem more completely than most other approaches. The accuracy
indicators utilized in the ERA approach are already being applied to several other
approaches with similar success. The primary limitation of this and other low- order
methods is the amount of computer memory required to solve the problem for cases
with a large number of response sensors.

4.2.4 Mathematical Input-Output Model Method

The experimental modal analysis methods that are included within the category
of mathematical input-output model methods are those approaches that generally
involve input and response data independently without the need for creating auto
and cross moment functions. There is no other restriction with regard to time
or frequency domain models, effective number of degrees of freedom, etc. On
this basis, two approaches are currently in use that can be described in this
fashion. First of all, the autoregressive moving average approach is a time domain
formulation that utilizes a pole-zero model as the basis for the description of the
system characteristics. While this model is appropriate, the model cannot be easily
constrained to account for known system information. Additionally, although the
current application of this technique does not involve multiple inputs, the theoretical
background for the multiple input case is well developed.
1 Recent History of Experimental Structural Dynamics 25

The other approach currently in use involves a reduced structural matrix model
for the basis of the description of the system characteristics. This model involves
the reduced mass, stiffness, and damping matrices with regard to the measured
degrees of freedom. This model easily accounts for constraints such as known
elements in the mass, stiffness, and damping matrices or known characteristics of the
distribution within the matrices such as symmetric or banded characteristics. This
method also incorporates the multiple input case routinely as known terms in the
forcing vector of the matrix differential equation that serves as the mathematical
model. These two approaches to experimental modal analysis will be briefly
described in the following paragraphs.

Autoregressive Moving Average Approach One approach to estimating the

modal characteristics from time domain input-output data is the autoregressive
moving-average (ARMA) procedure. This method has been applied to the determi-
nation of structural parameters by Gersch [37, 38, 39, 40, 41] and Pandit [42, 43].
With this technique the response data is assumed to be caused by a white random
noise input to the structure. The technique computes the best statistical model of
the system in terms of its poles (from the autoregressive part) and zeros (from the
moving-average part), as well as statistical confidence factors on the parameters. It
has been primarily used to estimate the characteristics of buildings being excited
by wind forces. The data used in the computational process are the autocorrelation
functions of the responses measured at various points on the structure. Since in
the general case the inputs are not measured, the modal vectors are determined
by referencing each response function to a single response to provide relative
magnitude and phase information.
Paralleling the development of the FRF equations (Equations 8 and 10), a time
domain model representing the relationship between a single response degree of
freedom and a single input degree of freedom can be stated as follows:


m 
n
αk x (ti+k ) = βk f (ti+k ) (18)
k=0 k=0

For the general multiple input, multiple output case:


m 
n
[αk ] {x (ti+k )} = [βk ] {f (ti+k )} (19)
k=0 k=0

The above model, in the time domain, is also known as an autoregressive moving-
average (ARMA(m,n)) model when developed from a set of discrete time equations
in the time domain.
If the discussion is limited to the use of free decay or impulse response function
data, the previous time domain equations can be simplified by noting that the forcing
function can be assumed to be zero for all time greater than zero. If this is the case,
the [βk ] coefficients can be eliminated from the equations.
26 R. J. Allemang


m
[αk ] [h (ti+k )] = 0 (20)
k=0

Additional equations can be developed by repeating Equation 20 at different time

shifts (initial times ti ) into the data until all data or a sufficient overdetermination
factor is achieved. Note that at least one time shift is required in order to accurately
estimate conjugate modal frequencies. More properly, this model is known as the
autoregressive with eXogenous inputs (ARX(m,n)) model.
The solution for the autoregressive moving average coefficients proceeds in
a two stage least-squares fashion in the Gersch solution. In the first stage a
long autoregressive model is solved linearly by using the Yule-Walker equations.
This process uses output covariance functions to determine the autoregressive
coefficients based upon a determination of the order of the autoregressive model.
The second stage involves setting up an equivalent moving average model for
the output involving convolution of the impulse response function and the input
function. This procedure also involves computations using covariance functions
and results in the least- squares computation of the moving average coefficients.
Since the solution for the autoregressive moving average coefficients, and thus the
structural parameter estimates, are statistically based, statistical confidence factors,
called coefficients of variation, for the natural frequencies and damping can be
easily calculated. These coefficients represent the ratio of standard deviation of each
parameter with respect to the actual parameter.
Note that if the response function x(t) is replaced by the impulse response
function h(t) and the forcing function f (t) is replaced by an impulsive force (unity
at time 0, zero after time 0), the ARMA approach is essentially the same approach
as the damped complex exponential response methods (PTD, ITD, ERA).

Reduced Structural Matrix Approach Over the period of time from 1965
to 1985, there has been increasing interest in being able to estimate reduced
structural (mass, stiffness, and damping) matrices from experimental data. Most of
these methods are based upon an indirect approach utilizing the estimated modal
parameters to synthesize the reduced matrices.
An algorithm has been developed in Germany by Link and Vollan [44] which
attempts to use frequency domain input and response data to directly estimate the
reduced matrices. This method has been designated Identification of Structural
System Parameters (ISSPA) [45]. Leuridan used this formulation as a starting
point and has published results [46, 47, 48] using the same general approach for
the estimation of modal parameters referred to as the direct system parameter
identification (DSPI) method. Since modal parameters are found as a result of the
solution of the eigenvalue problem using the reduced matrix estimations, the process
of estimating the reduced matrices may represent the ultimate goal in experimental
modal analysis. If this process could be correlated with a purely theoretical finite
element approach, the engineering design cycle would be complete.
1 Recent History of Experimental Structural Dynamics 27

Since there are many more known pieces of information (input and output
information at different times) than unknowns that must be estimated, the solution,
then, is a function of the pseudo-inverse procedure chosen. Leuridan and Vold
have evaluated pseudo-inverse numerical procedures utilizing the normal equations,
least-squares approach, and the Householder reflections approach. Since the system
matrix that results is often ill-conditioned, the Householder reflections approach
yields more numerical precision for a given computational word size but at a
sacrifice in speed.
Link and Vollan formulate the pseudo-inverse based upon a singular value
decomposition procedure under the restriction that the rank of the data-dependent
matrices is equal to the effective number of degrees of freedom. This effective
number of degrees of freedom is dependent upon the number of theoretical system
poles in the frequency range of interest, the accuracy of the measured data, and
the computational precision of the computer with respect to the solution algorithm
utilized.
Once the unknown elements of the mass, stiffness, and damping matrices are
found, the modal parameters are estimated from the [M], [C], and [K] matrices
by way of a complex eigenvalue-eigenvector solution algorithm such as the QR
algorithm.
A confidence or validity check of the frequencies, damping factors, and modal
vectors can be performed using a back substitution procedure. The dynamic
response is calculated and compared to the original measured response. The
agreement between these responses is regarded as a measure of the accuracy of
the estimated modal parameters.
This approach has not generated the anticipated results due to a number of
reasons. First of all, regardless of the approach used, the solution for the reduced
matrices is not unique. There are many combinations of matrix relationships that can
be generated from the given set of estimated modal parameters. Second, the reduced
frequency range of the modal parameter estimates means that the matrices will be
weighted to represent an incomplete model. Third, the limitation of the precision
of the modal parameter estimates as a result of commonly accepted experimental
error tends to desensitize the process of estimating the reduced matrices. Finally,
the problem of invalid modal parameter estimates will obviously result in invalid
estimates of reduced matrices.

4.3 Summary

The modal parameter estimation classifications that are included in Sect. 4.2 are
often presented in the literature as one of many modal parameter estimation methods
summarized by the acronyms, titles, and references cited in Table 1. These modal
parameter estimation methods are explained in more detail in  Chap. 10, “Experi-
mental Modal Analysis Methods.”
28 R. J. Allemang

Table 1 Acronyms – Experimental modal parameter estimation algorithms

Modal parameter estimation algorithms
CEA Complex exponential algorithm [49, 30]
LSCE Least-squares complex exponential [30]
PTD Polyreference time domain [31, 32]
ITD Ibrahim time domain [28, 50]
MRITD Multiple reference Ibrahim time domain [51]
ERA Eigensystem realization algorithm [34, 35, 52]
PFD Polyreference frequency domain [53, 54, 55, 56]
FDPI Frequency domain direct parameter identification [55, 56]
SFD Simultaneous frequency domain [57]
MRFD Multi-reference frequency domain [58]
RFP Rational fraction polynomial [47, 59, 60]
OP Orthogonal polynomial [60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71]
PLSCF Polyreference least-squares complex frequency [72, 73, 74, 75, 76, 77]
CMIF Complex mode indication function [78]

5 Conferences

A number of conferences began during the time period shortly after experimental
data acquisition and analysis hardware became available to serve both researchers
and practitioners in the area of experimental structural dynamics. The notable
conferences that began due to the experimental hardware becoming available were
the International Seminar on Modal Analysis (ISMA) at the Katholieke University
of Leuven (KUL) in Belgium, the International Modal Analysis Conference (IMAC)
sponsored originally by Union College in Schenectady, New York, the SAE Noise
and Vibration Conference (SAE-NVC), and the Shock and Vibration Symposium
(SVS). These conferences continue to this day. Several other professional societies
(ASME, AIAA, and the American Helicopter Society (AHS)) also began to
have focused sessions on experimental structural dynamics topics after these four
conferences led the way. All of the following conferences have archives of the
conference papers available from the sponsors.

5.1 ISMA

The International Seminar on Modal Analysis (ISMA) began in 1975 at the KUL
in Leuven, Belgium. Professor Raymond Snoeys of KUL together with Professor
David Brown of the University of Cincinnati (UC) and a number of KUL graduate
students developed a seminar for industry on the topic of modal analysis. Hewlett
Packard assisted in inviting researchers from European companies and universities
that were starting to use the data acquisition hardware who were interested in
learning about the topic. Beginning in 1980, an ISMA conference was added, and
1 Recent History of Experimental Structural Dynamics 29

researchers were invited to submit papers on their research and topics of interest.
This conference has been held biennially since that time. The 30th edition of the
Biennial ISMA Conference on Noise and Vibration Engineering will be held in
2022 at KUL at the same time as the 2022 ISMA Seminar. Further details can be
found at https://www.isma-isaac.be/.

5.2 IMAC

The International Modal Analysis Conference (IMAC) began with its first con-
ference in 1982, organized by Dick DeMichele and Peter Juhl at Union College
in Schenectady, New York. Dick and Peter solicited ideas and potential attendee
lists from many researchers in the area of experimental modal analysis as well
as from many of the companies involved in hardware development and/or use
of the technology. During the years 1982–1986, the conference was sponsored
by Union College followed by cosponsorship by Union College and the Society
for Experimental Mechanics (SEM) during the years 1987–1991. Beginning in
1991, SEM assumed responsibility for the organization and management of IMAC,
while Union College continued in the role of consulting partner until 1996. Dick
DeMichele, a well-known researcher already a member of SEM, continued as IMAC
technical director until 1995 as SEM began to take the lead on most management
and organizational roles [3]. Today IMAC is one of two annual conferences
sponsored by SEM where the organization and management of both conferences is
under the technical divisions (TDs) and focus groups (FGs) of the Society. Further
details about the next IMAC can be found at https://sem.org/imac.

5.3 Other Conferences

SAE-NVC In the 1980s, the Society of Automotive Engineering (SAE) began to

notice an increase in publications in the noise and vibration areas. In response
to this, the SAE developed the Noise and Vibration Conference (NVC) first held
in Traverse City, Michigan. Due to the strong interest in automotive applications,
the SAE-NVC held a biennial conference that advanced many applications in the
area of experimental structural dynamics. Generally, this conference has been held
biennially, in May or June, at locations in the midwestern region of the USA, close
to the automotive industry (Traverse City, Grand Rapids, etc.), but the schedule
has changed due to recent events. The next scheduled SAE-NVC is in September
2023. Further details about the next SAE-NVC can be found at https://www.sae.org/
attend/nvh.

SVS The annual Shock and Vibration Symposium (SVS) is another forum for
the structural dynamics and vibration community to present and discuss new
developments and ongoing research. The Symposium was established in 1947
and typically includes both unclassified (unlimited and limited distribution) and
30 R. J. Allemang

classified sessions. Topics covered at the symposium include shock-ship testing,

water shock, weapons effects (air blast, ground shock, cratering, penetration) shock
physics, earthquake engineering, structural dynamics, and shock and vibration
instrumentation and experimental techniques. The next scheduled conference is the
91st Symposium, and it will be held in September 2023. Further details about the
next SVS can be found at http://savecenter.org/symposium.html.

6 Publications and Books

There are a number of publications and books that have been heavily referenced
by conference proceedings and journal publications over the past years. Some of
these references are well-known, and some other important references are a bit
obscure. The following are some of interest to those in the experimental structural
dynamics area.

Sound and Vibration Magazine The Sound and Vibration Magazine began
publication in 1967 specifically to serve practitioners and researchers in the sound
and vibration industry. Jackson “Jack” Mowry served as the editor for this entire
period. This magazine is notable in that (1) topical research articles, many in the
area of experimental structural dynamics, were published throughout the more than
50 years of the magazine’s existence, and (2) the magazine was made available to
readers at no charge. The importance of the Sound and Vibration Magazine can not
be overstated with respect to the dissemination of experimental structural dynamics
information over the last 50 years.
The 50th anniversary issue was celebrated in January 2017. After 2017, the
publication moved to bimonthly, and then, beginning in 2018, the magazine became
part of Tech Science Press as a bimonthly and now as a quarterly, online journal.
Past issues of Sound and Vibration Magazine can be found at SandV.com at no
charge. The online version of the current Sound and Vibration can be found at
TechSciencePress.com

Handbooks A number of handbooks have been published since the early 1990s
that have attempted to collect and organize material pertinent to the area of
experimental structural dynamics, much in the same way as this Handbook of
Experimental Structural Dynamics is attempting. The following list of handbooks
contains the most well-known, but the authors apologize for any that have been
left out.
The oldest handbook that included a number of areas of interest to those in
the experimental structural dynamics area is the Shock and Vibration Handbook,
published in 1961, edited by C.M. Harris of Columbia University and C.E. Crede of
the California Institute of Technology (CalTech). The chapter topic of Experimental
Modal Analysis first appears in the 1988 Third Edition [79] and subsequent editions.
The First Edition of the SEM Handbook on Experimental Mechanics edited by
A.S. Kobayashi and published in 1987 focused on experimental mechanics topics of
1 Recent History of Experimental Structural Dynamics 31

interest to the Society. This first edition of the Handbook and the subsequent second
edition in 1994 also included a chapter topic of Experimental Modal Analysis [80].
The first handbook dedicated to the area of structural testing is the SEM
Handbook on Structural Testing published in 1993 just as modern, digital methods
were being applied to many areas of structural testing. The editors for this handbook,
Reese and Kawahara, and many of the chapter authors were SEM members with
strong affiliations with Sandia National Labs [81].

Government Reports A number of technical reports have provided initial research

and information that is still useful background and reference material. It is notable
that several of these reports eventually became textbooks that are still widely cited
and used today.
A number of technical reports generated by USAF contracts beginning in 1965
provided the basis of much of the early digital signal processing methodology. Two
of these early contracts were with Measurement Analysis Corporation, providing
the numerical and statistical background of frequency response, coherence, and
multiple coherence [82, 83]. Note that many of these concepts did not find their
way into experimental hardware and software until the early to late 1970s. These
concepts were in practice in the late 1970s and early 1980s and were subsequently
documented in several later USAF contract reports [84, 85, 86, 87, 88, 89, 90].
A number of these same authors developed monographs for the Naval Research
Laboratory as part of the US Department of Defense (DOD), Shock and Vibration
Information Center’s (SAVIAC) Shock and Vibration Monograph Series (SVM-1
through SVM-12) during the time period 1967–1986. Probably the most notable
monograph in the series is SVM-3 by L.D. Enochson and R.K. Otnes [91].
A more recent contribution (1996) from a government agency is the Dynamic
Test Agency Handbook from the Dynamic Testing Agency (DTA) in the UK [92].
The original DTA was set up in the late 1980s through a collaboration between
industry and the UK’s Department of Trade and Industry to disseminate best practice
to the structural dynamic community. Its output was primarily biased toward the
testing community not only to provide fundamental advice on how to implement
effective testing but also to inform users on the analysis and interpretation of results.
One of the objectives of the original group was to bring together test and analysis to
exploit the complementary aspects of each approach. The Handbook was originally
available only to the DTA members but now can be accessed via NAFEMS
at https://www.nafems.org/community/working-groups/dynamics-working-groups/
structural-dynamics/dta_handbook/.

Textbooks There are really only two or three textbooks that have been written
that focus on experimental modal analysis. The first text was by Dr. David Ewins
from Imperial College in 1984 [93]. This text is comprehensive for the year that
it was written, but many contributions have been made from 1980 until now that
were not included. A more recent text is a contribution by Dr. Peter Avitabile
[94] that includes many of the advances particularly in MIMO methodology and
experimentally based modeling of the last 25 years. Another recent text is from
32 R. J. Allemang

Dr. Robert Coppolino from Measurement Analysis Corporation [95] that is a little
more narrow in scope but adds much to the textbook publication arena to fill in
the state of the art. Hopefully, this Handbook will be even more comprehensive
and expansive and will provide the documentation of all important activities in
experimental structural dynamics over the last 50–60 years.
Some other texts are notable starting with texts in the applied mathematics
area. The oldest text concerns elementary matrix methods by Frazer, Duncan, and
Collar from 1938 [96]. This text introduced the matrix operation methodology that
moves higher-order polynomials to first order and allowed second-order differential
equations to be written as first-order, eigenvalue form. Today, this is known as state-
space methodology, and it is used in all higher-order, matrix coefficient polynomials
used in MIMO modal parameter estimation methods.
Many methods in experimental modal analysis and in experimental modeling
utilize eigenvalue-eigenvector methods that were first comprehensively discussed
by Wilkinson in 1965 [97]. Other texts in the area of applied linear algebra were
introduced about that time in the first edition textbook by Strang [98] somewhere
around 1965.
One very notable series of textbooks in the applied digital signal processing area
were authored by Julius Bendat and Allan Piersol beginning in 1966 [99]. This book
was followed by several editions of a more detailed book by Bendat and Piersol in
1971 and 1986 [100,24]. It should be noted that the 1971 first edition is widely cited
in the experimental structural dynamics area, but the second edition made important
changes to concepts like partial coherence. A similar book to the 1971 first edition
of Bendat and Piersol is the book by Otnes and Enochson. This book first appeared
in 1972 [101] after appearing in print as part as the Naval Research Lab Monograph
SVM-3 in 1968 [91]. All of these texts were extensively cited in early work in the
experimental structural dynamics area.
Several current mechanical vibration textbooks have some documentation of
vibration testing and the relationship to experimental structural dynamics. These
are too numerous to reference here. However, the mechanical vibrations text
by J.P. Den Hartog, first edition, published in 1934 was unique in that the
text extensively documented experimental problems and applications [102] of
that era. While obviously quite dated, the Den Hartog book is still enjoyable
reading.
Dr. Roy Craig, Jr. authored a textbook [103] in 1982 that is one of the few
textbooks to address many structural dynamic related topics. The text contains basic
SDOF/MDOF vibration theory found in many vibration texts but also continues
on with finite element modeling basics and expands into system modeling with
component mode synthesis. In addition, the text includes very detailed material
on all the generic numerical processing from static decomposition techniques to
eigensolution techniques followed by propagation solution techniques rounding
out all the numerical processing typically required. The text also contains basic
information on experimental structural dynamic methods. Dr. Roy Craig, Jr. was
an early IMAC participant, serving on the IMAC Advisory Board for many
years.
1 Recent History of Experimental Structural Dynamics 33

7 Pioneers/Contributors

There are a large number of contributors to the area of experimental structural

dynamics particularly from the mid-1960s to the mid-1980s when the ISMA and
IMAC conferences began to flourish. Most of these contributors can be identified
from the author and attendee list from IMAC and ISMA covering the time period
from 1982 until now. Prior to 1982, a bibliography list of papers and references can
be found as one of the papers in the first (1982) IMAC Conference proceedings
[1]. This list of papers and authors includes most references prior to 1982 and
is relatively comprehensive. Two other publications include discussion of early
contributors Pioneers of Shock and Vibration [104] and The Old and the New
. . . A narrative on the history of the Society for Experimental Mechanics [3].
These texts provide additional information for early researchers that is not limited
to those involved in experimental structural dynamics. Certainly the authors of
textbooks mentioned in the previous section would also be important contributors
to the topic.

7.1 Pioneers: Experimental Structural Dynamics

There are two significant contributors to the field of experimental structural

dynamics that would be on the list of pioneers for this field regardless of who is
asked. This very short list would include Professor David L. Brown of the University
of Cincinnati and Professor David J. Ewins of Imperial College. Both of these
researchers started in the mid-1960s at their respective universities and took leading
roles in the development of many aspects of the field of experimental structural
dynamics as is known today. Simply doing a literature search during the time period
of 1960 to the present day for these two authors would allow the reader to get a
comprehensive idea of the field.
Dr. David L. Brown initially focused on the dynamics of machine tool cutting
at the University of Cincinnati at a time when analytical methods (finite element
methods) were just being developed. The only practical approach to understanding
the dynamics of machine tools and the effect of machine tool dynamics on cutting
was the emerging analog experimental modal analysis method based upon a slow
swept sine estimation of frequency response functions (FRFs). This initial research
work led Dr. Brown to the fast Fourier transform and developing general broadband
approach to the estimation of frequency response functions (FRFs) utilizing digital
minicomputers. That starting point led to improved FRF estimation methods and the
development of single and multiple degree of freedom modal parameter estimation
methods during the 1970s along with multiple input estimation of FRFs around
1980.
Dr. David J. Ewins followed a similar path at Imperial College where he ended
up focusing on the experimental dynamic analysis of rotating turbo-machinery. This
focus required specialized methods to understand and experimentally measure the
34 R. J. Allemang

dynamics of many similar circular objects such as brakes, wheels, and the dynamics
of circular complex structures involved in jet engines and turbines. Dr. Ewins first
textbook [93] was the first to focus on experimental structural dynamics and the
first text to document the measurement and parameter estimation methods that
were becoming common practice and, to this point in time, were only presented
in conference proceedings.
Both Dr. Brown and Dr. Ewins have distinguished themselves as first rate
researchers, educators, and technologists. Both have presented the theoretical side of
their research as well as the practical and applied experimental side of the research.
Both Dr. Brown and Dr. Ewins became heavily involved internationally in giving
applied short courses to industry on the experimental structural dynamic methods
that they were pioneering. Once the IMAC Conference began, Dr. Brown and Dr.
Ewins worked together on the modal analysis course given each year at IMAC in
addition to their own experimental modal analysis courses. Both Dr. Brown and Dr.
Ewins brought graduate student researchers with them to IMAC, and these graduate
students have gone on to numerous professor positions and now bring a second and
third generation of graduate student researchers to IMAC.
Both Dr. Brown and Dr. Ewins have received many awards from the technical
community across many professional societies. Limiting the discussion to the
Society for Experimental Mechanics (SEM), both Dr. Brown and Dr. Ewins served
on the IMAC Advisory Board for the first 25 years as the IMAC Conference evolved.
Dr. Ewins served as the chairman of the IMAC Advisory Board for the first 5 years.
Both Dr. Brown and Dr. Ewins have received many awards from the Society for
Experimental Mechanics including the SEM Fellow Award (2014), the M. M. Frocht
Award (2002), and the D. J. DeMichele Award (1993) for Dr. Ewins and the SEM
Fellow Award (2013), the William M. Murray Lecture (2006), the B.J. Lazan (1987),
and the D.J. DeMichele Award (1992) for Dr. Brown. Both Dr. Brown and Dr. Ewins
have been honored by giving two IMAC keynote speeches. Dr. Brown in the 1st
(1982) and the 25th (2007) and Dr. Ewins (1984) in the 2nd and the 25th (2007).

7.2 Contributors: Handbook

Certainly the authors of chapters for this Handbook are important contributors to
the field of experimental structural dynamics. The list is not limited to the authors
of the chapters of this Handbook, but this is a good starting point (Table 2).

7.3 Contributors: SEM IMAC

The list of the SEM IMAC contributors is very long with 200–300 papers published
in each IMAC since 1982. The best way to get a view of these contributors,
particularly the first 20 years, is to acquire a DVD copy of all of the papers
along with the Table of Conference for each IMAC Proceedings. This DVD is still
available through the Society for Experimental Mechanics.
1 Recent History of Experimental Structural Dynamics 35

Table 2 Contributors to the Handbook of Experimental Structural Dynamics

Author Affiliation Topic
Randall J University of History of Experimental Structural Mechanics
Allemang Cincinnati-SDRL
Peter Avitabile University of History of Experimental Structural Mechanics
Massachusetts Lowell
Christopher University of DIC and Photogrammetry Measurements
Niezrecki Massachusetts Lowell
Phillip L. Reu Sandia National DIC and Photogrammetry Measurements
Laboratories
Javad Baqersad Kettering University DIC and Photogrammetry Measurements
Daniel P. Rohe Sandia National DIC and Photogrammetry Measurements
Laboratories
Paolo Chiariotti Università Politecnica Laser Doppler Vibrometry Measurements
delle Marche
Christian Rembe Technical University Laser Doppler Vibrometry Measurements
Clausthal
Paolo Castellini Università Politecnica Laser Doppler Vibrometry Measurements
delle Marche
Matthew S. Allen University of Laser Doppler Vibrometry Measurementss
Wisconsin-Madison
Robert B. Randall University of New South Applied Digital Signal Processing
Wales
Jerome Antoni “University of Lyon, Applied Digital Signal Processing
INSA-Lyon”
Pietro Borghesani University of New South Applied Digital Signal Processing
Wales
Anders Brandt University of Southern Spectral and Correlation Analysis Methods
Denmark
Stefano Manzoni Politecnico di Milano Spectral and Correlation Analysis Methods
Allyn W. Phillips University of Frequency Response Function Estimation
Cincinnati-SDRL
Randall J University of Frequency Response Function Estimation
Allemang Cincinnati-SDRL
Thomas L. Paez Sandia National Random Vibration and Mechanical Shock
Laboratories
Norm F. Hunter Los Alamos National Random Vibration and Mechanical Shock
Laboratory
David Smallwood Sandia National Random Vibration and Mechanical Shock
Laboratories
Mark Valentino PCB Piezotronics, Inc.” Sensors and their Signal Conditioning for
Dynamics
Patrick L. Walter Texas Christian Sensors and their Signal Conditioning for
University Dynamics
Gary Foss The Boeing Company Sensors and their Signal Conditioning
for Dynamics
Jessica Meloy The Boeing Company Sensors and their Signal Conditioning
for Dynamics
(continued)
36 R. J. Allemang

Table 2 (continued)
Author Affiliation Topic
Thomas Carne Sandia National Design of Modal Tests
Laboratories
Ralph Brillhart ATA Engineering, Inc. Design of Modal Tests
Daniel Kammer University of Wisconsin Design of Modal Tests
– Madison
Mr. Kevin ATA Engineering, Inc. Design of Modal Tests
Napolitano
Randall J University of Experimental Modal Parameter Evaluation
Allemang Cincinnati-SDRL Methods
Allyn W. Phillips University of Experimental Modal Parameter Evaluation
Cincinnati-SDRL Methods
Carlos Ventura University of British Operating Modal Analysis Methods
Columbia
Randall J University of Experimental Modal Analysis Methods
Allemang Cincinnati-SDRL
David L. Brown University of Experimental Modal Analysis Methods
Cincinnati-SDRL
Daniel Rixen Technische Universität Substructuring and Component Mode
München Synthesis
Peter Avitabile University of Finite Element Model Correlation
Massachusetts Lowell
Mike Mains University of Finite Element Model Correlation
Cincinnati-SDRL
John E. University of Liverpool Model Updating
Mottershead
Michael Link University of Kassel Model Updating
Michael I. Swansea University Model Updating
Friswell
Carsten ICS Engineering GmbH Model Updating
Schedlinski
Lothar Gaul University of Stuttgart Damping of Materials and Structures
Andre Schmidt University of Stuttgart Damping of Materials and Structures
Francois Hemez Lawrence Livermore Uncertainty Quantification: UQ & QMU
National Lab
Kendra Lu Van Lawrence Livermore Uncertainty Quantification: UQ & QMU
Buren National Lab
Janette J. Meyer Vanderbilt University Nonlinear System Analysis Methods
Raymond M. Vanderbilt University Nonlinear System Analysis Methods
Bond
Douglas E. Vanderbilt University Nonlinear System Analysis Methods
Adams
Keith Worden University of Sheffield Structural Health Monitoring, Damage
Identification
Ramon Fuentes Los Alamos National Structural Health Monitoring, Damage
Laboratory Identification
Charles R. Farrar Los Alamos National Structural Health Monitoring, Damage
Laboratory Identification
(continued)
1 Recent History of Experimental Structural Dynamics 37

Table 2 (continued)
Author Affiliation Topic
Randall L. Mayes Sandia National Experimental Dynamic Substructures
Laboratories
Matthew S. Allen University of Wisconsin Experimental Dynamic Substructures
Mark Richardson Vibrant Technology, Inc. Structural Dynamics Modification, Modal
Modeling
David Formenti Blackhawk Technology Structural Dynamics Modification, Modal
Modeling
Nuno Maia University of Lisbon Robust Response Models: Issues
António Urgueira Universidade NOVA de Robust Response Models: Issues
Lisboa
Raquel Almeida Universidade NOVA de Robust Response Models: Issues
Lisboa
Tiago Silva Universidade NOVA de Robust Response Models: Issues
Lisboa
Gaetan Kerschen University of Liège Modal Analysis of Nonlinear Mechanical
Systems
Alexander F. University of Illinois Modal Analysis of Nonlinear Mechanical
Vakakis (UIUC) Systems
Peter Avitabile University of Linear Modal Substructuring, Nonlinear
Massachusetts Lowell Connections
Jim De Clerck Michigan Technological Automotive Structural Testing
University
Ruben Boroschek University of Chile Civil Structural Testing
Joao Pedro Santos National Laboratory for Civil Structural Testing
Civil
Engineering, Lisboa,
Portugal
Robert Coppolino Measurement Analysis Aerospace Perspective for Modeling and
Corporation Validation
Chuck Van Michigan Technological Applied Math for Experimental Structural
Karsen University Dynamics
Andrew Barnard Michigan Technological Applied Math for Experimental Structural
University Dynamics

Contributions to the Society and to IMAC take on many forms. One notable
contribution to the Society for Experimental Mechanics is to serve on the executive
board and/or as an officer of the Society. Three of the active researchers in the area of
experimental structural dynamics have gone on to serve as president for the Society:
Dr. Randall Allemang (2003–2004), Dr. Carlos Ventura (2012–2013), and Dr. Peter
Avitabile (2016–2017).
In recognition of their contributions to the Society, a number of researchers in
the area of experimental structural dynamics have been honored with the award of
SEM Fellow. Those researchers are noted in the following table (Table 3).
38 R. J. Allemang

Table 3 Recipients of the SEM Fellows Award

Recipient (Year) Recipient (Year) Recipient (Year)
Peter Avitabile (2022) Randall Mayes (2022) Douglas Adams (2021)
Charles Ferrar (2018) Carlos Ventura (2018) Dan Inman (2017)
David J. Ewins (2014) David L. Brown (2013) Randall J. Allemang (2010)

Table 4 Recipients of the D.J. DeMichele Award

Recipient (Year) Recipient (Year) Recipient (Year)
Jason Blough (2021)
Scott Cogan (2020) Ralph D. Brillhart (2019) Charles D. Van Karsen (2018)
Mark Schiefer (2017) Michael D. Todd (2016) Randy Mayes (2015)
Keith Worden (2014) Charles R. Farrar (2013) Raj K. Singhal (2012)
Mark Richardson (2011) François M. Hemez (2010) Douglas E. Adams (2009)
Álvaro Cunha (2008) Daniel J. Inman (2007) William R. Shapton (2006)
James Lally (2005) Peter Avitabile (2004) Carlos E. Ventura (2003)
Roy R. Craig Jr. (2002) Sam Ibrahim (2001) Thomas G. Carne (2000)
Alfred L. Wicks (1999) Randall Allemang (1998) Kenneth G. McConnell (1997)
Nobuyuki Okubo (1996) Larry D. Mitchell (1995) Tzu Chuen Huang (1994)
David J. Ewins (1993) David L. Brown (1992) Bruno Piombo (1991)
Dominick J. DeMichele (1990)

In addition, there is a specific Society award to recognize researchers and

educators active in the structural dynamics area. This award is named after D.J. Dick
DeMichele who led the effort to get IMAC started back in 1982. Those recipients
are listed in the above table (Table 4).
Most of the researchers in the experimental structural dynamics have also been
recognized by other organizations and/or professional societies too numerous to be
included here.

7.3.1 IMAC Advisory Board

The original IMAC Advisory Board was established by Peter Juhl and Dr. Dick
DeMichele as they interviewed many active researchers trying to determine the
interest level for a new organization and conference centered on experimental
structural dynamics. Both Mr. Juhl and Dr. DeMichele obtained input from many
of the active researchers about the format and content of the potential conference.
Many of the researchers listed as part of the original IMAC Advisory Board
participated in this process in the 3–4 years before the first conference in 1982.
Both Mr. Juhl and Dr. DeMichele were lecturers at Union College, and they involved
Union College in the organization of the first few conferences as the primary sponsor
(1982–1986). Gradually, there was a desire to have some sort of professional society
affiliation, and the Society for Experimental Mechanics became the home of the
conference after 1991 as the Society assumed the responsibility for organization and
1 Recent History of Experimental Structural Dynamics 39

Table 5 Members of the Original IMAC Advisory Board

Member Affiliation
David J. Ewins Imperial College of Science and Technology, England
Dominique Bonnecase AS&I Rhone-Alpes, France
David L. Brown University of Cincinnati-SDRL, USA
Dominick J. DeMichele Union College/IMTS, USA
Keneth A. Galione Society for Experimental Mechanics, Inc., USA
Arnold E.S. Gussin, Union College, USA
Larry E.S. Mitchell Virginia Polytechnic Institute, USA
Nobuyuki Okubo Chuo University, Japan
Michael P. Pakstys NKF Engineering Associates, Inc., USA
Anders Reveman Tre Konsultar AB, Sweden
Mark H. Richardson Structural Measurement Systems, Inc., USA
Paul Sas Katholieke Universiteit Leuven, Blegium
Havard Vold, Structural Dynamics Research Corporation, USA
Lingmi Zhang Nanjing Aeronautical Institute, China
Cheng Yaodong Zhejiang University, China
Randall J Allemang University of Cincinnati-SDRL, USA

management of IMAC. Union College continued as a cosponsor with the Society

from 1987 to 1991 and ceased any further involvement in 1996 (Table 5).
Dr. David Ewins was elected as the first chair of the IMAC Advisory Board and
served from 1982 to 1987. Dr. Ewins was followed by Dr. Randall Allemang from
1987 to 1995 and by Dr. Mark Richardson from 1995 to 2015. Much of the growth
of IMAC and the integration of the IMAC activity into the Society for Experimental
Mechanics was the result of hard work by the original Advisory Board and the
chairman during this time period along with the willingness of Union College
to discontinue involvement after 1996. In recent years, the Advisory Board and
chairman are now part of the Society administrative structure with regular changes
in Advisory Board members and chairman.

8 Summary/Conclusions

This introductory chapter for the Handbook reviews the recent history of modern
experimental structural dynamics methods and highlights the active researchers and
educators involved in the developments. Modern experimental structural dynamics
refers to the time period since digital computing, the fast Fourier transform, and
analog to digital conversion of measured data became readily available, somewhere
around the mid-1960s. The experimental methods introduced in this chapter are
covered in great detail in other chapters of this Handbook along with advances that
took place after 1975. This chapter simply represents the starting point that became
the focus of activities within the Society for Experimental Mechanics since the mid-
1980s.
40 R. J. Allemang

References
1. Allemang RJ (1982) Experimental modal analysis bibliography. In: Proceedings, international
modal analysis conference (IMAC), pp 714–726
2. Cooley JW, Tukey J (1965) An algorithm for the machine calculation of complex Fourier
series. Math. Comput. 19(90):297
3. Sciammarella CA, Zimmerman KB (2018) The old and new . . . a narrative on the history of
the society for experimental mechanics. Morgan and Claypool Publishers, San Rafael
4. Stein PK (1990) A brief history of bonded resistance strain gages from conception to
commercialization. Exp. Tech. 14(5):13–20
5. Walter PL (2007) The history of the acccelerometer, 1920s–1996 – prologue and epilogue,
2006. Sound Vib Mag (SVM) 41(1):84–92
6. Keller AC (1969) Vector component techniques: a modern way to measure modes. Sound Vib
Mag (SVM) 3(3):18–26
7. Bickel HJ (1971) Real time spectrum analysis. Sound Vib Mag (SVM) 5(3):14–20
8. Keller AC (1975) Real-time spectrum analysis of machinery dynamics. Sound Vib Mag
(SVM) 9(4):40–48
9. Gade S, Herlufsen H (2017) 35 years of structural measurements at Brüel & Kjær. Sound Vib
Mag (SVM) 51(1):26–30
10. Mowry J, Borring G (2012) Journey to greatness. The story of Brüel & Kjær. Acoustical
Publications, Inc. Bay Village, OH, www.SandV.com
11. Keller T (2017) The making of a great company: how spectral dynamics came to be. Sound
Vib Mag (SVM) 51(1):64–84
12. Tustin W (1969) Vibration test equipment. Sound Vib Equip 3(3):27–32
13. Welaratna S (1997) Thirty years of FFT analyzers. Sound Vib Mag (SVM) 31(1):5
14. Lang GF (2017) Historical snippets from our half century. Sound Vib Mag (SVM) 51(1):2–4
15. Allemang R, Brown D, Fladung W (1994) Modal parameter estimation: a unified matrix
polynomial approach. In: Proceedings, international modal analysis conference (IMAC),
pp 501–514
16. Lewis RC, Wrisley DL (1950) A system for the excitation of pure natural modes of complex
structures. J Aeronaut Sci 17(11):705–722
17. De Veubeke BF (1956) A variational approach to pure mode excitation based on characteristic
phase lag theory. AGARD, Report, vol 39, p 35
18. Trail-Nash RW (1958) On the excitation of pure natural modes in aircraft resonance testing.
J Aeronaut Sci 25(12):775–778
19. Asher GW (1958) A method of normal mode excitation utilizing admittance measurements.
In: Dynamics of aeroelasticity, Proceedings. Institute of the Aeronautical Sciences, pp 69–76
20. Stahle CV, Forlifer WR (1958) Ground vibration testing of complex structures. In: Flight
flutter testing symposium, NASA-SP-, vol 385, pp 83–90
21. Stahle CV Jr (1962) Phase separation technique for ground vibration testing. Aerosp Eng
1962:8
22. DeBlauwe F (1991) Investigation of a parameter estimation algorithm for spatial sine testing.
Doctoral Dissertation, University of Cincinnati, vol 117
23. Deblauwe F, Shih CY, Rost R, Brown DL (1987) Survey of parameter estimation algorithms
applicable to spatial domain sine testing. In: Proceedings, international seminar on modal
analysis (ISMA), vol 15
24. Bendat JS, Piersol AG (1986) Random data: analysis and measurement procedures, 2nd edn.
Wiley-Interscience, New York
25. Napolitano KL (2016) Using singular value decomposition to estimate frequency response
functions. Top Modal Anal Test 10:27–44
26. Allemang R, Patwardhan R, Kolluri M, Phillips A (2022) Frequency response function
estimation techniques and the corresponding coherence functions: a review and update. Mech
Syst Sig Process (MSSP) 162(1):108100
1 Recent History of Experimental Structural Dynamics 41

27. Prony R (1795) Essai Experimental et Analytique sur les Lois de la Dilatabilite des Fluides
Elastiques et sur Celles de la Force Expansive de la Vapeur de l‘eau et de la Vapeur de
l‘Alkool, a Differentes Temperatures. Journal de l‘ Ecole Polytechnique (Paris) 1(2):24–76
28. Ibrahim S, Mikulcik E (1977) A method for the direct identification of vibration parameters
from the free response. Shock Vib Bull 47:183–198
29. Ibrahim S (1978) Modal confidence factor in vibration testing. Shock Vib Bull 48:65–75
30. Brown D, Allemang R, Zimmerman R, Mergeay M (1979) Parameter estimation techniques
for modal analysis. SAE Paper Number 790221, SAE Trans 88:828–846
31. Vold H, Kundrat J, Rocklin T, Russell R (1982) A multi-input modal estimation algorithm for
mini-computers. SAE Trans 91(1):815–821
32. Vold H and Rocklin T (1982) The numerical implementation of a multi-input modal estima-
tion algorithm for mini-computers. In: Proceedings, international modal analysis conference
(IMAC), pp 542–548
33. Vold H, Leuridan J (1982) A generalized frequency domain matrix estimation method for
structural parameter identification. In: Proceedings, international seminar on modal analysis
(ISMA)
34. Juang J-N, Pappa RS (1985) An eigensystem realization algorithm for modal parameter
identification and model reduction. AIAA J Guid Control Dyn 8(4):620–627
35. Juang J (1987) Mathematical correlation of modal parameter identification methods via
system realization theory. J Anal Exp Modal Anal 2(1):1–18
36. Juang JN, Pappa RS (1985) Effects of noise on ERA-identified modal parameters. AAS Paper,
vol 23, No AAS-85-422
37. Gersch W (1970) Estimation of the auto regressive parameters of a mixed autoregressive
moving-average time series. IEEE Trans Autom Control AC-15:583–588
38. Gersch W, Luo S (1972) Discrete time series synthesis of randomly excited structural system
responses. J Acoust Soc Am 51(1):402–408
39. Gersch W, Sharpe DR (1973) Estimation of power spectra with finite-order autoregressive
models. IEEE Trans Autom Control AC-18:367–369
40. Gersch W, Nielsen NN, Akaike H (1973) Maximum likelihood estimation of structural
parameters from random vibration data. J Sound Vib (JSV) 31(3):295–308
41. Gersch W (1974) On the achievable accuracy of structural system parameter estimates.
J Sound Vib (JSV) 34(1):63–79
42. Pandit SM (1977) Analysis of vibration records by data dependent systems. Shock Vib Bull
(47), 161–174
43. Pandit SM, Suzuki H (1979) Application of data dependent systems to diagnostic vibration
analysis. ASME Paper, No. 79-DET-7, p 9
44. Link M,and Vollan A (1978) Identification of structural system parameters from dynamic
response data. Zeitschrift Fur Flugwissenschaften 2(3):165–174
45. Leuridan J (1981) Direct system parameter identification of mechanical structures with appli-
cation to modal analysis. Master of Science Thesis, Department of Mechanical Engineering,
University of Cincinnati, p 200
46. Leuridan J, Mergeay M, Vandeurzen U, Desanghere G (1984) Multiple input estimation of
frequency response functions for experimental modal analysis: currently used method and
some new developments. In: International seminar on modal analysis, p 34
47. Leuridan J (1984) Some direct parameter model identification methods applicable for multiple
input modal analysis. Doctoral Dissertation, p 384+
48. Leuridan J, Brown D, Allemang R (1985) Time domain parameter identification methods for
linear modal analysis: a unifying approach. ASME Paper Number 85-DET-90
49. Spitznogle F (1971) Improvements in the complex exponential signal analysis computational
algorithm, Report Number U1-829401-5, Representation and Analysis of Sonar Signals,
vol 1, p 37
50. Pappa R (1982) Some statistical performance characteristics of the ITD modal identification
algorithm. AIAA Paper Number 82-0768, p 19
42 R. J. Allemang

51. Fukuzono K (1986) Investigation of multiple-reference Ibrahim time domain modal parameter
estimation technique. M. S. Thesis, p 220
52. Longman RW, Juang J-N (1989) Recursive form of the eigensystem realization algorithm for
system identification. AIAA J Guid Control Dyn 12(5):647–652
53. Zhang L, Kanda H, Brown D, Allemang R (1985) A polyreference frequency domain method
for modal parameter identification. ASME Paper No. 85-DET-106, p 8
54. Lembregts F, Leuridan J, Zhang L, Kanda H (1986) Multiple input modal analysis of
frequency response functions based on direct parameter identification. In: Proceedings,
international modal analysis conference (IMAC)
55. Lembregts F (1988) Frequency domain identification techniques for experimental multiple
input modal analysis, Doctoral Dissertation
56. Lembregts F, Leuridan J, Van Brussel H (1989) Frequency domain direct parameter identifi-
cation for modal analysis: state space formulation. Mech Syst Sig Process (MSSP) 4(1):65–76
57. Coppolino R (1981) A simultaneous frequency domain technique for estimation of modal
parameters from measured data. SAE Paper No. 811046, vol 12
58. Craig R, Kurdila A, Kim HM (1990) State-space formulation of multi-shaker modal analysis.
J Anal Exp Modal Anal 5(3):169–183
59. Natke H (1988) Updating computational models in the frequency domain based on measured
data: a survey. Probab Eng Mech 3(1):28–35
60. Richardson M, Formenti D (1982) Parameter estimation from frequency response measure-
ments using rational fraction polynomials. In: Proceedings, international modal analysis
conference (IMAC), pp 167–182
61. Shih C, Tsuei Y, Allemang R, Brown D (1988) A frequency domain global parameter
estimation method for multiple reference frequency response measurements. Mech Syst Sig
Process (MSSP) 2(4):349–365
62. Shih C (1989) Investigation of numerical conditioning in the frequency domain modal
parameter estimation methods. Doctoral Dissertation, p 127+
63. Van der Auweraer H, Leuridan J (1987) Multiple input orthogonal polynomial parameter
estimation. Mech Syst Sig Processing (MSSP) 1(3):259–272
64. Van der Auweraer H, Snoeys R, Leuridan J (1986) A global frequency domain modal
parameter estimation technique for mini-computers. ASME J Vib Acoust Stress Reliab
Des 108
65. Vold H (1986) Orthogonal polynomials in the polyreference method. In: Proceedings,
international seminar on modal analysis (ISMA)
66. Vold H, Shih CY (1988) On the numerical conditioning of some modal parameter estimation
methods. In: Proceedings, international seminar on modal analysis. In: Proceedings, interna-
tional seminar on modal analysis
67. Vold H (1990) Numerically robust frequency domain modal parameter estimation. Sound Vib
Mag (SVM) 24(1):38–40
68. Vold H (1990) Statistics of the characteristic polynomial in modal analysis. In: Proceedings,
international seminar on modal analysis (ISMA), pp 53–57
69. Vold H, Napolitano K, Hensley D, Richardson M (2008) Aliasing in modal parameter
estimation, an historical look and new innovations. In: Proceedings, international modal
analysis conference (IMAC), vol 16, pp 12–17
70. Fladung W, Vold H (2016) An improved implementation of the orthogonal polynomial
modal parameter estimation algorithm using the orthogonal complement. In: Proceedings,
international modal analysis conference (IMAC)
71. Fladung W, Vold H (2016) An orthogonal view of the polyreference least-squares complex
frequency modal parameter estimation algorithm. In: Proceedings, international modal
analysis conference (IMAC)
72. Van der Auweraer H, Guillaume P, Verboven P, Vanlanduit S (2001) Application of a fast-
stabilizing frequency domain parameter estimation method. ASME J Dyn Syst Meas Control
123(4):651–658
1 Recent History of Experimental Structural Dynamics 43

73. Guillaume P, Verboven P, Vanlanduit S, Van der Auweraer H, Peeters B (2003) A

polyreference implementation of the least-squares complex frequency domain estimator. In:
Proceedings, international modal analysis conference (IMAC), p 12
74. Verboven P, Guillaume P, Cauberghe B, Parloo E, Vanlanduit S (2003) Stabilization charts
and uncertainty bounds for frequency domain linear least squares estimators. In: Proceedings,
international modal analysis conference (IMAC), pp 1–10
75. Verboven P, Cauberghe B, Vanlanduit S, Parloo E, Guillaume P (2004) The secret behind
clear stabilization diagrams: the influence of the parameter constraint on the stability of the
poles. In: Proceedings, society of experimental mechanics (SEM) annual conference, vol 17,
pp 1–17
76. Verboven P (2002) Frequency domain system identification for modal analysis. Doctoral
Dissertation
77. Cauberghe B (2004) Application of frequency domain system identification for experimental
and operational modal analysis. Doctoral Dissertation, p 259
78. Shih C, Tsuei Y, Allemang R, Brown D (1988) Complex mode indication function and its
application to spatial domain parameter estimation. Mech Syst Sig Process (MSSP) 2(4):
367–377
79. Harris CM, Piersol AG (eds) (2009) Shock and vibration handbook, 6th edn. McGraw-Hill,
New York
80. Kobayashi A (ed) (1994) Handbook on experimental mechanics, 2nd edn. Prentice Hall
81. Reese RT, Kawahara WA (eds) (1993) Handbook on structural testing. Society of Experimen-
tal Mechanics/The Fairmont Press, Inc., Lilburn
82. Enochson LD, Goodman NR (1965) Gaussian approximations to the distribution of sample
coherence, USAF Report AFFDL-TR-65-57, p 33
83. Goodman NR (1965) Measurement of matrix frequency response functions and multiple
coherence functions, USAF Report AFFDL-TR-65-58, p 38
84. Allemang RJ, Brown DL, Rost RW (1984) Multiple input estimation of frequency response
functions for experimental modal analysis, No. AFATL-TR-84-15
85. Allemang R, Brown D (1987) Volume I: Summary of technical work. Experimental modal
analysis and dynamic component synthesis
86. Allemang R, Brown D (1987) Volume II: Measurement techniques for experimental modal
analysis. Experimental modal analysis and dynamic component synthesis
87. Allemang R, Brown D (1987) Volume III: Modal parameter estimation. Experimental modal
analysis and dynamic component synthesis
88. Allemang R, Brown D (1987) Volume IV: System modeling techniques. Experimental modal
analysis and dynamic component synthesis
89. Allemang R, Brown D (1987) Volume V: Modal parameter estimation. Experimental modal
analysis and dynamic component synthesis
90. Allemang R, Brown D (1987) Volume VI: Software users guide. Experimental modal analysis
and dynamic component synthesis
91. Enochson L, Otnes R (1968) Programming and analysis for digital time series data SVM-3.
In: The shock and vibration information center (SAVIAC)
92. Agency DT (ed) (1996) Dynamic test agency handbook, vols 0–9. Dynamic Testing Agency
and Crown Publications
93. Ewins D (1984) Modal testing: theory and practice. Wiley, New York
94. Avitabile P (2017) Modal testing: a practitioner’s guide. SEM and John Wiley & Sons Ltd.,
West Sussex
95. Coppolino RN (2019) The integrated test analysis process for structural dynamic systems.
Morgan & Claypool Publishers, San Rafael
96. Frazer R, Duncan W, Collar A (1938) Elementary matrices (1965 Reprint). Cambridge
University Press, London
97. Wilkinson JH (1965) The algebraic eigenvalue problem. Oxford University Press, Oxford
98. Strang G (2006) Linear algebra and its applications, 4th edn. Centage Learning, Boston
44 R. J. Allemang

99. Bendat JS, Piersol AG (1966) Measurement and analysis of random data. Wiley, New York
City, New York
100. Bendat JS, Piersol AG (1971) Random data: analysis and measurement procedures, 1st edn.
Wiley-Interscience, New York
101. Otnes RK, Enochson L (1972) Digital time series analysis. Wiley, New York
102. Den Hartog J (1947) Mechanical vibrations, 3rd edn. McGraw-Hill Book Company, New
York/London
103. Craig, R. Jr., Kurdila A (1982) Fundamentals of structural dynamics. Wiley, New York City,
New York
104. Freeman M (1996) Pioneers of shock and vibration. SAVIAC/Booz-Allen and Hamilton Inc.,
Arlington
Sensors and Their Signal Conditioning for
Dynamic Acceleration, Force, Pressure, 2
and Sound Applications

Gary Foss, Jessica Meloy, Mark Valentino, and Patrick Walter

Contents
1 Sensing Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.1 Piezoelectricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.2 Metal Strain Gages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
1.3 Semiconductor, Piezoresistive, and MEMS Strain Gages . . . . . . . . . . . . . . . . . . . . . 53
1.4 Piezoelectric Strain Gages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
1.5 Capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2 Sensor Dynamic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.1 Strain Gage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 Pressure, Force, and Acceleration Transducers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3 Sensor Selection and Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1 Piezoelectric Accelerometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Force Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3 Impedance Heads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4 Pressure Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Pressure Transducers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4 Sensor Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1 Sensor System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5 Signal Conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

G. Foss
The Boeing Company, Chicago, IL, USA
e-mail: [email protected]
J. Meloy
Pensar Development, Seattle, WA, USA
e-mail: [email protected]
M. Valentino ()
PCB Piezotronics, Inc., Depew, NY, USA
e-mail: [email protected]
P. Walter
Texas Christian University & PCB Piezotronics, Inc., Fort Worth, TX, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 45

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_33
46 G. Foss et al.

5.1 Types of Amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5.2 Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3 Analog-to-Digital Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Other Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.1 Grounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.2 Cabling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.3 Ground Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.4 Triboelectric Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7 Data Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

Abstract

The continued demand for efficiency, reliability, and lower operating costs
together with the general increase in awareness of the effects of vibration and
noise in the workplace has demanded a better understanding of the causes and
characteristics of the vibration of machines and structures in many industries.
This, coupled with advances in transducer technology, electronics, and signal
processing power, has led to a wealth of commercial products for performing
structural dynamics measurements.
Structural dynamics is the response of a structural system to dynamically
imposed loads. Undesirable responses can cause suboptimum performance or
structural failure. Understanding the relationships between forces and responses
may involve in-service operational measurements or the creation of simulated
environments in the laboratory. Analytical modeling verified by experimental
modal analysis is a frequently employed tool to understand and/or adjust
structural dynamics before the introduction of a product into service.
Successful testing depends on the accurate measurement of the dynamic force
and pressure loading encountered by the structure as well as the response of
the structure. The common electromechanical transducers that are used for these
measurements are pressure transducers and microphones, load cells or force
transducers, strain gages, and accelerometers. The signals from these devices
are amplified and filtered in a way that preserves their fidelity over the required
bandwidth and amplitude levels. Proper use depends on wise selection, informed
by the knowledge of the required data and the limitations of the products. This
chapter provides a guide to the sensing technologies and electronics available to
make these measurements and their use.

Keywords

Sensor · Sensing · Piezoelectric · Strain gage · Microphone · Force sensor ·

Pressure sensor · Impedance · Signal conditioning
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 47

1 Sensing Technologies

1.1 Piezoelectricity

Piezoelectricity is a transduction technology particularly applicable to transducers

intended to measure time-varying force, pressure, and acceleration measurands.
Just as the research focus since the 1960s has been on silicon technology (soon
discussed), a similar effort existed in the 1920s through the 1950s on piezoelectric
technology. Polycrystalline ceramics and quartz were the principal piezoelectric
materials of focus during the latter part of this time span. The uses for piezoelectric
materials continue to grow, although piezoelectric technology itself is now fairly
mature.
Piezoelectricity is attributable to strain inducing a change in the shape of a crystal
that possesses no center of charge symmetry. An electric charge results from this
change in shape. Twenty-one of the 32 crystal classes lack this symmetry element,
and crystals in all but one of these classes can exhibit piezoelectricity.
Quartz (Fig. 1) is a common piezoelectric material that exists in nature. Today,
however, quartz is principally grown artificially. Quartz is highly stable, rugged,
and linear and operates over a wide temperature range. It is typically the technology
of choice for piezoelectric pressure transducers and compression type load cells.
Tourmaline is another piezoelectric found in nature and is acquired through
mining operations. It operates well in high-temperature applications and underwater
explosions. In addition, some ferroelectric polycrystalline ceramic materials can be
artificially manufactured to exhibit piezoelectricity. This latter group of materials is
of particular interest since the manufacturing process can control their mechanical

Fig. 1 A boule of quartz

48 G. Foss et al.

Fig. 2 Poling ceramics to

effect piezoelectric
properties. (Courtesy PCB
Piezotronics)

and electrical properties. This manufacturing process consists of the weighing and
proportioning of ceramic powders, calcining at high temperatures to produce a
chemical combination of the ingredients, mixing in a ball mill to repowder the raw
compound, adding a binder, granulating, compressing the powder into pellet form,
and firing the pellets in a controlled atmosphere in a kiln. This firing transforms the
pellets into ceramic elements. The elements are lapped and plated for subsequent
polarization. A high-voltage field is applied across each pellet under controlled
environmental conditions (Fig. 2). The minute crystal domains within the ceramic
are forced to align themselves with the applied field, and this alignment is retained
after the field is removed.
There are many important properties of piezoelectric materials. The piezoelectric
constant for a material expresses the amount of charge generated per unit applied
force or the deflection per unit applied voltage. Typical units are Coulombs/Newton
(C/N) or meters/volt (m/v). The coupling coefficient provides the energy conversion
efficiency of the piezoelectric material; a high value is desired. The dielectric
constant determines the capacitance of the material. A piezoelectric transduction
element is effectively a capacitor that produces a charge across its plates propor-
tional to the force applied to it. The material resistivity must be high to keep the
charge generated in the material from leaking off. The open-circuit voltage is the
voltage generated at the output of the piezoelectric element per unit of applied force.
Unlike silicon-based transducers, piezoelectric transducers do not have response to
dc or 0 Hz.
The Curie temperature is the temperature above which the crystal lattice modifies
its structure and the piezoelectric properties are lost. Good practice is to avoid
transducer use at greater than 1/2 the Curie temperature.
The following table provides some electrical and mechanical properties for a few
piezoelectric materials. These properties indicate why these materials are desirable
for application as transduction elements.
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 49

Young’s modulus (c11 ) Density Piezoelectric constant

(N/m2 ) (g/cm3 ) (10−12 ) (C/N)
Quartz 86.7 x 109 2.65 −2.3
Tourmaline 270 x 109 3.1 2.1
PZT-5A 121 x 109 0.75 374

Fig. 3 Manufactured
ceramic shapes

The moduli of elasticity of these materials are between 50% and 150% of
the values for steel and silicon; they are quite stiff. Similarly, their modulus-to-
density ratios are between 60% and 130% that of steel. If the properties of PZT
5A (lead zirconate titanate) are used as an example, the piezoelectric constant is
374 picocoulombs/Newton. It is apparent that a small force input to this ceramic
material would result in a large electrical output signal. Therefore, these materials
enable miniaturization of transducer design while still permitting operation over
large amplitude ranges. Figure 3 illustrates various piezoelectric ceramic shapes
dependent on application.

1.2 Metal Strain Gages

A strain gage is a device used to measure strain on the surface of an object. The
most common type of strain gage consists of an insulated flexible backing on which
a metal foil serpentine pattern is supported. The gage is attached to the object
by a suitable adhesive for the application. As the object is deformed, the foil is
also deformed, causing its electrical resistance to change. This resistance change,
usually measured using a Wheatstone bridge, is related to the strain by the quantity
known as the gage factor. Bonded metal foil strain gages (Fig. 4) are a mature
technology. Their importance is routinely encountered in our daily lives every time
that a weighing process occurs, the dynamic characteristics of a structural system are
50 G. Foss et al.

Fig. 4 Foil strain gage. (Courtesy of VPG Sensors)

assessed to monitor for material fatigue and structural design margin, the damaging
resonant frequencies contained in a structure are identified, and much more.
The detailed history of the evolution of the strain gage is presented in reference
[1]. On September 10, 1936, Ed Simmons, an electrical engineering graduate
student at the California Institute of Technology, suggested using bonded wire to
measure the dynamic forces generated by an impact testing machine. The professor
with whom he was working, Dr. Gottfried Datwyler, bonded 40-gage, cotton-
wrapped, insulated, constantan wire to a piece of clock spring with Glyptal cement.
The spring was mounted as a cantilever beam, and the wire change in resistance was
proven to be linear, repeatable, and hysteresis-free with applied strain. The bonded
wire strain gage was born. The experimental work was completed and presented at
an ASTM meeting in June 1938. Scant attention was given at this meeting to the
strain gage development that supported the experimental work.
History: In 1938, at MIT, Prof. Arthur Ruge was working on a research contract
with his graduate assistant Hans Meir to measure the stresses induced in water
towers under earthquake conditions. On April 3, 1938, Prof. Ruge unwound wire
from a precision resistor, bonded it to a test beam, and created a strain gage. The
importance of this discovery was immediately apparent to Prof. Ruge. He and
Mr. Meir spent the rest of their lives developing and commercializing the bonded
strain gage and transducers based on its operating principle. During a patent search
following the 1938 MIT discovery, Mr. Simmons’ earlier work was uncovered. As a
result, Simmons ultimately received patent number 2,292,549 on August 11, 1942,
as the recognized inventor of the bonded resistance strain gage. Today the bonded
wire strain gage has been replaced by foil-etched gages formed by printed circuit
techniques. Currently the vast majority of strain gage applications associated with
experimental stress analysis are performed using bonded metal foil strain gages.
Their manufacture is as follows.
Construction: The two most common materials comprising metal foil gages are
constantan (55% Cu 45% Ni) and karma (20% Cr 2.5% Al 2.5% Cu balance Ni).
Both these materials offer (1) atypical resistance versus temperature behavior, (2)
malleability sufficient to allow processing into foil less than 0.001 in (0.025 mm.)
 2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 51

thick, (3) ease of photochemical machining into accurate configurations, and (4)
reasonable cost. Considering gage resistance-temperature behavior, it is desired that
a strain gage eliminates false signals due to thermal expansion of the material on
which it is mounted. If the increase in gage resistance due to thermal expansion of
this material can be offset by a corresponding decrease in gage alloy resistivity,
the result will be zero change and no false signal. In reality, a finished gage
assembly includes its backing, sealant, and adhesive. All of these materials expand
at their own rate and contribute to this false signal. Thermal coefficient of resistance
(TCR) values required to achieve thermal compensation tend to range from −25 to
+5 p.p.m./o C. These values are well within the capability of cold rolled constantan
and karma. Optimum compensation is obtained by heat treatment of the two foils
dependent on the material on which they are mounted. Figure 5 shows one typical
thermal or false strain compensation that can be achieved by a given strain gage in
a temperature range around room temperature.
Strain gage manufacturing involves first putting stringently manufactured alloys
through a closely controlled melting process. Repeated stages of rolling result in
a thin foil that is bonded to a backing film, acting as a carrier. The foil with the
backing is then etched with photochemicals to form the desired gage geometry.

TEMPERATURE – °C
−100 −50 0 +50 +100 +150 +200 +250
+500
+400
+300
+200 +24°C K-ALLOY
THERMAL OUTPUT – µ⑀ (FI = 2.0)

+100
0
−100
−200 +75°F A-ALLOY
−300
−400
−500
−600
−700
−800
−900
TEST SPECIMEN – 1018 STEEL
−1000

−200 −100 0 +100 +200 +300 +400 +500

TEMPERATURE – °F

Fig. 5 Typical temperature compensation curves achievable with constantan (A) and karma (K)
strain gage materials (courtesy of VPG Sensors)
52 G. Foss et al.

Sheets of gages are then cut apart and packaged for sale. Reference [2] provides a
detailed description of this process.
To build effective strain sensing circuits, one must be aware of the interaction
between the gage and the surface of the flexure to which it is mounted. Mechanical
aspects of this interaction include the influence of backing material, size, orientation,
transverse sensitivity, distance from the surface, bonding, and installation.
Backing material: The purpose of the backing material used in constructing
strain gages is to provide support, dimensional stability, and mechanical protection
for the grid element. The backing material acts as a spring in parallel with the
parent material to which it is attached and can potentially modify mechanical
behavior. In addition, the temperature operating range of the gage can be constrained
by its backing material, typically polyimide. The standard temperature range for
polyimide backed gages is from −100 to +350F. Special construction can extend
this range from cryogenic to +500F. Some gages are encapsulated for chemical and
mechanical protection as well as extended fatigue life. Special purpose metal gages
can be welded. The frequency response of welded gages, due to uncertainties in
dynamic response, is a subject area still requiring investigation.
Size: The major factors to be considered in determining the size of strain gage
to use are available space for gage mounting, strain gradient at the test location,
and character of the material under test. The strain gage must be small enough to
be compatible with its mounting location and the concentrated strain field. It must
be large enough so that, on metals with large grain size, it measures average strain
as opposed to local effects. Grid elements greater than 0.125 in. (3 mm.) generally
have greater fatigue resistance.
Transverse sensitivity and orientation: Strain gage transverse sensitivity and
mounting orientation are concurrent considerations. Transverse sensitivity in strain
gages is important due to the fact that part of the geometry of the gage grid is
oriented in directions other than parallel to the principal gage sensing direction.
Values of transverse sensitivities are provided with individual gages but typically
vary between fractional and several percent. The position of the strain gage axis
relative to the numerically larger principal strain on the surface to which it is
mounted will have an influence on indicated strain.
Distance from the surface: The grid element of a strain gage is separated from
the structure under test by its backing material and cement. The grid then responds
to strain at a location removed from its mounting surface. The strain on structures
such as thin plates in bending can vary considerably from that measured by the strain
gage.
Bonding adhesives: Resistance strain gage performance is entirely dependent on
the bond attaching it to the parent material. The grid element must have the strain
transmitted to it undiminished by the bonding adhesive. Adhesive types break into
categories depending on the manner of cure: one part or two part mixes and room
or elevated temperature cure. Common strain gage adhesives are as follows:

Cyanoacrylate adhesives. Most widely used general-purpose one-part adhesive.

Easiest to handle. Room temperature cures in under 1 min. Long-term temper-
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 53

ature range from -25F to +150F. Creep-free, fatigue-resistant bond with strain
capability of 5% or more.
Epoxy adhesives. Useful over a temperature range from – 270 to +320 ◦ C. Two
classes are room temperature curing or thermal setting: both available with
various organic fillers to optimize performance for individual test requirements.
Polyimide adhesives. For high performance applications, polyimide adhesives are
a one-part thermal setting resin with a very thin, nearly creep-free bond layer.
Temperature range for one product is from −450 to +700F and is achieved by
step curing at four different temperatures.
Polyester adhesives. For rough or porous surfaces like concrete, mortar, and wood.
Two-part, room temperature cure. Temperature range from −20 to +350F.
Ceramic adhesives. For high-temperature applications to 2000F. Gages must be free
wire or foil. Single part elevated temperature cure.

Weldable strain gages are also available for spot welding to structures and
components. They are ideal for applications where test or environmental conditions
preclude clamping and curing an adhesively bonded gage.

1.3 Semiconductor, Piezoresistive, and MEMS Strain Gages

In the 1940s through the 1950s, it became recognized that, when geometrically
distorted, the resistance change of semiconductor materials could also be correlated
to strain [3]. The semiconductor strain gage became of interest because its sensitivity
to strain was about 50 to 200 times that of metal gages. However, whether using
p- or n-type silicon, gage sensitivity was discovered to be strongly influenced by
both temperature and strain level. For this reason, semiconductor strain gages find
principal application only in experimental stress analysis involving small strain
differences at controlled temperatures. Colloquially, the term “piezoresistive” strain
gage is used synonymously with silicon or semiconductor strain gages.
While not extensively used in experimental stress analysis work, because of their
higher strain sensitivity, piezoresistive strain gages do find significant application
in the construction of transducers (e.g., pressure, force, and acceleration) whose
output can be thermally compensated. Piezoresistive transducers manufactured in
the 1960s first used silicon strain gages fabricated from lightly doped ingots. These
ingots were sliced with respect to the crystal axes of the silicon to form small
bars or patterns which became gages. These gages were usually bonded directly
to the transducer flexure. Since the late 1970s, there has been a continual evolution
of microsensors into the market place. Piezoresistive transducers manufactured in
this manner use silicon both as their flexural element and as their transduction
element. The strain gages are diffused directly into the flexure. The most typical
fabrication process has the following sequence of events: the single crystal silicon
is grown; the ingot is trimmed, sliced, polished, and cleaned; diffusion of a dopant
into a surface region of the parent silicon wafer is controlled by a deposited film;
a photolithography process includes etching of the film at places defined in the
54 G. Foss et al.

developing process, followed by removal of the photoresist; and isotropic and

anisotropic wet chemicals are used for shaping the mechanical microstructure. Both
the resultant stress distribution in the microstructure and the dopant control the
piezoresistive coefficients of the silicon.
Electrical interconnection of various controlled surfaces formed in the silicon
crystal as well as bonding pads is provided by thin film metallization. The silicon
wafer is then separated into individual dies. The dies are bonded by various
techniques into the transducer housing, and wire bonding connects the metalized
pads to metal terminals in the transducer housing. Sensors fabricated in this manner
are known as microelectromechanical systems (MEMS) transducers. While metal
strain gage-based transducers typically provide 20 to 30 millivolts of unamplified
full-scale signal, by comparison, MEMS resistance-based transducers produce 100–
200 millivolts of unamplified signal. This provides a signal-to-noise advantage at
low strain levels. MEMS transducer technology is rapidly expanding in commercial
and military applications. Reference [4] provides an extensive chapter on strain
gage-based transducers.

1.4 Piezoelectric Strain Gages

Strain within a piezoelectric material displaces electrical charges within the strained
elements, and the charges accumulate on opposing electrode surfaces. Piezoelectric
strain gages do not have response to zero Hz. Therefore, their application in
experimental stress analysis is limited. However, modern gages have integral signal-
conditioning electronics (ICP® or IEPE) that greatly enhance the measurement
system’s signal-to-noise ratio. Five volts of unamplified signal can be provided
for 100 microstrain making these type gages very desirable for low-level, dynamic
strain measurements (Fig. 6).

Fig. 6 Piezoelectric strain

gage. (Courtesy PCB
Piezotronics)
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 55

1.5 Capacitance

Both microphones and accelerometers can be constructed using the principle

of changing capacitance. Microphones based on capacitance are usually called
“condenser microphones.”
Condenser microphones: Through the years there have been a number of
different microphone designs used to measure sound, or unwanted sound commonly
referred to as noise. Sound gets transmitted through either liquids, solids, or gasses
as oscillations of pressure waves and can either be detected by the human ear (the
audible range) or be in the form of infrasound or ultrasound. Infrasound represents
sounds below the human hearing range, and ultrasound is above the designated
human hearing range of 20 Hz to 20 kHz. Acoustics is the study of these waves.
Microphones allow us the capability to measure these sounds. The most common
test and measurement microphone for accurately measuring sound is the condenser
style.
Construction: A condenser microphone is an electromechanical sensor that
converts a change in capacitance into a measureable electrical voltage as discussed
in reference [5]. The condenser microphone contains two parallel plates which are
electrically isolated surfaces separated by a dielectric medium, which acts as an
insulator. The dielectric for condenser microphones is air. The two parallel surfaces
consist of a backplate and a diaphragm.
The backplate in a prepolarized microphone includes an electret material that
stores a fixed embedded charge to provide a difference in electrical potential across
the air gap. An externally polarized microphone uses a power supply to apply
a voltage across the backplate-diaphragm gap. The diaphragm is a thin material
stretched across the housing, which is maintained at a separate electrical potential
from the backplate. When these two parallel surfaces are combined with the air
dielectric, they form a capacitor with a very stable electrical charge.
Changes in pressure cause the diaphragm to deflect, causing variations in the
gap between the diaphragm and the backplate, and the result is an oscillating
electric field. As the distance between the microphone’s diaphragm and stationary
backplate changes, the capacitance changes proportionately to the sensitivity of
the microphone and its air gap as shown in Fig. 7. An output voltage can then be
obtained that represent the changes in sound or pressure waves that the diaphragm
was exposed to (Fig. 8).
   
Q0 ε·A
C= =
E0 − e D0 − d

Where:
A = Area of capacitor plate
C = Instantaneous capacitance between plates
Do = Distance between plates at rest position
d = Displacement of moveable plate (diaphragm) from rest position
Eo = Polarization voltage
56 G. Foss et al.

Fig. 7 Microphone Output.

(Courtesy PCB Piezotronics)

Fig. 8 Microphone output

voltage superposed on
polarization voltage (E0 ) e
Amplitude

E0

Time

e = Voltage change caused by plate displacement

Qo = Constant charge on plate capacitor
Epsilon = Dielectric constant of air

The primary components of a condenser microphone (seen in Fig. 9) consist of a

diaphragm, a backplate, a body, an insulator, and grid cap. The diaphragm is a thin
material and is tensioned to a desired amount which can either allow it to be very
sensitive or be less sensitive. In general, high tension on the diaphragm will increase
the high-frequency response capability and lower the sensitivity. Conversely, lower
tension on the diaphragm will degrade the high-frequency response. Inherent noise
is also tied to the sensitivity of the diaphragm. High sensitivity will provide lower
inherent noise or high signal-to-noise ratio, while low sensitivity will have higher
inherent noise. For enhanced low-frequency response, adjustments in both the
venting design and the preamplifier design need to be incorporated. The preamplifier
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 57

Fig. 9 Components of a
condenser microphone

provides the impedance conversion required to drive the signals across the cable to
the end devices.
The backplate is a metal component that may contain strategically placed holes in
it to properly allow for the desired damping characteristics required for the specific
microphone design. The damping is modified to meet the desired inherent noise
and sound field design requirements. For example, critically damped systems will
yield a smooth frequency response with a small roll-off. This is desired in pressure
fields and diffuse field applications. An overdamped system is desired for free-field
applications where a more significant roll-off is desired. The field response types
and their intended applications are discussed later in this chapter.
The body, also called the housing, is the outer component that encapsulates the
microphone and helps tension the diaphragm. Stress relieving the housing (along
with the diaphragm) is one of the critical processes necessary to ensure long-term
stability of the microphone. The insulator shields the backplate from the rest of
the microphone and external components. The grid cap is a cover that protects the
delicate diaphragm.
The above components are designed to allow a microphone to output a voltage
that correlates to its desired amplitude and a frequency. The displacement of the
microphone diaphragm provides the changes in the electrical field that represent
the amplitude of the pressure exerted upon it. The oscillations of the diaphragm
provide the frequency component. When multiple microphones are placed in a
predetermined pattern (also called an array) and combined with the proper software
and phase characteristics, they can be used to analyze particle velocity and the
direction of sound, allowing this grouping of microphones to be an excellent choice
for noise source location applications.
Condenser microphones are stable over time, temperature, and humidity and are
the microphone of choice for precision sound level meters and many of today’s
test and measurement applications. The two most common design types desired for
precision testing are externally polarized and prepolarized condenser microphones.
58 G. Foss et al.

Fig. 10 Externally polarized

microphone design

Externally polarized microphones: The externally polarized method of supplying

the necessary voltage to the backplate is accomplished by applying a fixed 200 V
via an external power supply (see Fig. 10). The electrical isolation between the
backplate and the diaphragm provides a difference in electrical potential across the
air gap, providing the basic requirements for the capacitor used to measure pressure
waves exerted on the microphone [13].
The externally polarized design is the traditional method which was originally
widely accepted by acousticians and test engineers. In addition to the 200 V
power supply, the externally polarized models require multiconductor cabling and
connectors. The most common is a seven-pin LEMO® connector. Most externally
polarized microphone cartridges have the capability to reach temperatures of 150 ◦ C
without significant sensitivity loss. This temperature specification can be limited by
the preamplifier operating temperature capability. The externally polarized design
is easier to build and, prior to the technological advances of prepolarized electret
microphones, offered better stability and noise floor specifications.
Electrets or prepolarized microphones: An electret or prepolarized design oper-
ates in a similar capacitance fashion. The difference between the externally polar-
ized design and the electret or prepolarized design is that an electrical charge is
embedded into a polymer material that sits on top of the backplate facing the
diaphragm (see Fig. 11). This eliminates the need for an external 200 V power
supply. With this new electret-based design, any 2–20 mA constant current supply,
for example, ICP® power supplies (also called signal conditioners), can now be used
to power the microphone.
The prepolarized design is a more modern design and allows for the use of cost-
effective coaxial cables and BNC, 10–32, or SMB connectors. The combination
of the cable savings and the use of signal conditioners in lieu of 200 V power
supplies offer significant equipment savings. The prepolarized design is also better
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 59

Fig. 11 Prepolarized or
electret microphone design

for humid applications and for portability. The 2–20 mA power supply and coaxial
cabling are the same power and cabling required by many other common test and
measurement sensors such as accelerometers and piezoelectric pressure sensors.
With a multichannel signal conditioner, setup time can be minimized, and multiple
vibration and sound tests can be run within the same setup. This design is
significantly growing in popularity due to its flexibility to be used with other sensors
and the low-cost per-channel savings it offers. Data acquisition manufacturers have
taken notice and are manufacturing a greater number of their products with the
same 2–20 mA power built-in, making this electret prepolarized design even more
portable and easy to use.
Capacitive accelerometers: Capacitive accelerometers sense a change in elec-
trical capacitance, with respect to acceleration, to vary the output of an energized
circuit. Typically they are structured with a mass loaded diaphragm that undergoes
flexure in the presence of acceleration. Two fixed plates sandwich the diaphragm,
creating two capacitors, each with an individual fixed plate and each sharing the
diaphragm as a movable plate. The flexure causes a capacitance shift by altering the
distance between two parallel plates, the diaphragm itself being one of the plates.
The two capacitors operate in an AC bridge circuit, along with two fixed capacitors,
and alter the peak voltage generated by an oscillator when the structure undergoes
acceleration. Detection circuits capture the peak voltage, which is then fed to a
summing amplifier that processes the final output signal. One design is shown in
Fig. 12.
This type of mechanism creates a very stable, accurate measurement device,
which is inherently insensitive to base strain and transverse acceleration effects. The
design permits operation to DC acceleration and is ideal for low-amplitude, low-
frequency requirements. Maximum useful frequency range depends on sensitivity:
60 G. Foss et al.

Fig. 12 Capacitive Accelerometer Design (courtesy PCB Piezotronics)

typical limit is between 300 and 1000 Hz. The DC response permits the measure-
ment of tilt and orientation commonly required for consumer products such as game
controllers and cellphones.

2 Sensor Dynamic Models

A transducer may be thought of as an energy conversion device. In the transducers

being considered in this work, mechanical energy will be converted into electrical
energy. The sensing technologies just described, when integrated into the structural
flexures of transducers, enable this energy conversion to occur. Electromechanical
transducers are able to measure pressure, force, and acceleration from fractional
psi, pounds, and Gs, to greater than 100,000 units of these same measurands.
When measuring the loads applied to or the response of structural systems via
electromechanical transducers, the dynamic performance of these transducers must
first be understood. Strain gages are unique in the fact that they can either be
integrated into transducer flexures or used in a stand-alone configuration to measure
structural response. For this reason, the dynamic response of the strain gage as a
stand-alone element will first be considered.

2.1 Strain Gage

While often mounted on the flexures of electromechanical transducers, the basic

dynamic response capability of the strain gage itself is considered here. It will be
assumed that mounting variables such as adhesive bond thickness are controlled
so that they do not influence gage frequency response. If these assumptions are
satisfied, the strain gage can be shown to act as a spatial averaging device whose
frequency response is a function of both its gage length and the sound velocity of
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 61

Fig. 13 Frequency response of a strain gage

the material on which it is mounted. Figure 13 portrays this relationship. Reference

[6] discusses the analysis from which Fig. 13 is extracted.
As an example of using Fig. 13, we can see that a strain gage with a 0.32 cm
gage length (1/8th of an inch) uniformly passes all frequencies to about 1.0/2.54
or 0.394 cycles/cm. The velocity of sound in steel is nominally 510,000 cm/sec.
If we multiply 0.394 times 510,000, we get approximately 200,000 Hz as an upper
frequency limit for “uniform” or “flat” frequency response for a 0.32 cm (1/8th inch)
properly mounted strain gage. Thus, strain gages are capable of measuring structural
response to very high frequencies.

2.2 Pressure, Force, and Acceleration Transducers

The simplified dynamic model of a pressure, force, or acceleration transducer is

typically presented as in Fig. 14. Figure 15 shows equally simplified conceptual
loading models. The farthest left item in Fig. 15 shows a piezoelectric element
(red, the spring in every case) being compressed by an applied force on both its
top and bottom. The central item in Fig. 15 shows the piezoelectric element being
compressed on its top surface by a force attributable to an applied pressure load.
62 G. Foss et al.

Fig. 14 Dynamic model of a

pressure, force, or
acceleration transducer

Fig. 15 Pictorial load models for piezoelectric transducers. (Courtesy PCB Piezotronics)

Fig. 16 Physical implementation (left to right) of piezoelectric force, pressure, and acceleration
transducers. (Courtesy PCB Piezotronics)

The right item in Fig. 15 shows an inertial force being applied to the piezoelectric
element attributable to base acceleration. Figure 16 shows an assortment of trans-
ducers representing the physical implementations of Fig. 14. As noted previously,
piezoelectricity is just one of several sensing technologies that can be implemented
in electromechanical transducers.
Figure 17 (left) illustrates strain gages mounted on a mechanical flexure to
form an accelerometer. Figure 17 (right) shows, attributable to the miniaturization
achievable with MEMS technology, how strain gage accelerometers (as well as
capacitive) can be manufactured from silicon in extremely small sizes. For all of
these examples, the frequency response characterization of the simplified dynamic
model of Fig. 14 is presented in Fig. 18.
All piezoelectric sensors have very small values of damping (ζ = damping ratio
≈ 0.03). Focusing on the top set of curves in Fig. 18, it can be seen that the frequency
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 63

Fig. 17 Strain gage accelerometer (left) and physical implementation with MEMS technology
(right). (Courtesy PCB Piezotronics)

Fig. 18 Amplitude vs. normalized frequency response and simplified dynamic model for pressure,
force, and acceleration transducers
64 G. Foss et al.

content of signals measured by a transducer compatible with this model is replicated

with fidelity to about 1/5th of the natural frequency (ω/ωn = 1) of the transducer.
The phase response (bottom set of curves) indicates that over this same range of
response frequencies, essentially no phase shift is introduced by the transducer. For
MEMS technology this same observation holds. Three exceptions should be noted:

(1) The electrical model of a piezoelectric transducer (force, pressure, and

accelerometer) is that of a voltage source with a series capacitor. A transducer
following this model cannot respond to 0 Hz (DC). The frequency response
of a piezoelectric transducer is then uniform or “flat” over an intermediate
band of frequencies. Its low-frequency limitation is attributable to the time
constant of the associated electrical high-pass circuit, and its high-frequency
constraint is attributable to the resonant frequency of the transduction element.
See Fig. 19. By contrast, MEMS-based strain gage transducers do not have this
low-frequency limitation.
(2) Some low G metal strain gage and MEMS-based sensing technology accelerom-
eters have adequate flexure motion to enable mechanical damping of 0.7 to be
incorporated into them. Formerly this damping was achieved with fluids, but
with the technology of today, it is accomplished with a gaseous medium. Gases
are preferred since their viscosity is essentially constant with temperature. A
value of 0.7 damping optimizes the range of the accelerometers maximum “flat”
frequency response and yields a nominally linear phase response vs. frequency.
(3) Force transducers respond according to the simplified model of Fig. 13 when
impulsively loaded. However, in application most force transducers become
physically integrated into a test stand. This test stand contains distributed mass

Fig. 19 Useful operating range of piezoelectric transducers. (Courtesy PCB Piezotronics)

2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 65

and elasticity of differing materials and interfaces. Thus, the force transducer
becomes a structural part of the test stand and does not retain its individual
identity. The signal output from the force transducer must then be analyzed as
part of the structural response of the entire test stand.

In general, the dynamic response of a piezoelectric transducer can be broken

down into regions by frequency, shown in Fig. 19. A low-pass filter limits the low-
frequency response. The useful range is customarily assumed to be between the
frequency at which amplitude response is down 5% and extends to the frequency
where the amplitude response is up 5% (at around 1/5 of resonant frequency). This
would be defined as the 5% frequency band and is usually found on the vendor’s
specification sheet.

3 Sensor Selection and Use

3.1 Piezoelectric Accelerometers

A piezoelectric accelerometer is an electromechanical transducer that generates an

electrical output when subjected to vibration or shock and is the most widely used
motion measurement transducer to be found throughout industry. It is used in a
variety of environments and applications to measure absolute vibration, directly as
acceleration or integrated to provide velocity and displacement data. There are many
different piezoelectric accelerometer types commercially available, all of which
benefit from most of the following characteristics:

(1) Wide frequency range of operation

(2) Self-generating signal, built-in charge amplifier
(3) Linear over a wide dynamic range of operation
(4) Compact, with high output-to-mass ratio
(5) Robust and easy to install
(6) Operational in a wide variety of environmental conditions

A simple representation of a piezoelectric accelerometer was shown in Fig. 15

on the right. The mass element is connected to the mounting body via the
piezoelectric material, which acts as a spring-damper and provides an electrical
output proportional to the relative displacement of the mass element with respect
to the mounting body. This is directly proportional to the acceleration experienced
at the accelerometer mounting body. An amplifier may be added for the conversion
of the charge signal to a low impedance voltage signal, hereafter referred to as IEPE.
Most accelerometers can be considered as base excited single-degree-of-freedom
systems and are purposely designed to be sensitive in only one axis. The sensitivity
of an accelerometer is the ratio of the electrical output charge (Coulombs) or voltage
to the applied acceleration. This sensitivity is dependent on the inertia mass and the
volume and characteristics of the piezoelectric material. With no internal charge
66 G. Foss et al.

converter, the sensitivity is defined in units of picocoulombs/g (pC/g). With internal

electronics, the sensitivity is specified in units of millivolts/g. Ideally the sensitivity
of an accelerometer remains constant throughout its working frequency range.
The reference information for the selection of a piezoelectric accelerometer
is the manufacturer’s specification. From this specification, the user can match
the accelerometer type to meet the requirements. These are the most important
considerations:

(1) Physical size and weight

(2) Useable frequency range
(3) Acceleration limits and sensitivity
(4) Type of mount
(5) Base strain and cross-axis sensitivity
(6) Resonant frequency
(7) Connector position on the case
(8) Operating temperature and humidity

Piezoelectric accelerometers are available in packages ranging from less than 0.5
gram to thousands of grams. As a rule, the larger packages permit a larger sensing
element which translates into higher sensitivity. As another rule, smaller sensing
elements offer higher resonant frequency and therefore higher useful frequency
range. Sensitivity should be chosen such that the expected test levels reach an
appreciable portion of the full dynamic range. “Appreciable” could be defined as
anywhere from 20% to 90%, depending on uncertainty of the test levels. Exceeding
the full-scale range can lead to loss of data, so overestimating levels is safer than
underestimation.
The lower-frequency limit of a piezoelectric accelerometer is marked by a
decrease in sensitivity determined by the high-pass cutoff of the electronics. It
may be as low as 0.2 Hz for larger high-sensitivity units and up to 5 Hz for
ultraminiature or shock accelerometers. The upper frequency limit of a piezoelectric
accelerometer is marked by an unacceptable increase in sensitivity due to the
resonance behavior shown in Figs. 18 and 19. Full-scale acceleration limits of
commercial accelerometers range from .5G to 100,000G.
There are a variety of established techniques for mounting an accelerometer to
a structure. Each method has its benefits and disadvantages, and each affects the
working upper frequency limit of measurement. The effect of mounting is to add an
additional mechanical system consisting of a spring (the stud or adhesive), the mass
of the entire accelerometer, and some additional damping. If this system resonance
is lower than the crystal resonance, the upper useable frequency limit will be less
than the vendor’s specification.
The best mounting employs a stiff threaded stud supplemented with a thin layer
of grease between smooth, flat mating surfaces. This will require a drilled and tapped
hole and possibly a milled, flat surface on the test article (not always permitted). It
will produce a mounted natural frequency that should be close to the accelerometer
specification. The next best mounting is with a good quality adhesive such as epoxy,
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 67

Fig. 20 Frequency response curves showing the effect of the type of various accelerometer
mounts. (Courtesy PCB)

with properly cleaned and prepared mating surfaces. For room temperature, non-
shock applications, cyanoacrylate adhesive is very quick to apply. For temporary
mounting, a thin layer of beeswax offers attachment in seconds, with some sacrifice
in frequency response. Another temporary method is double-sided tape, which
further compromises frequency response. Figure 20 shows the frequency response
of several mounting methods together with their characteristics and applications.
The notable effect of mounting is to decrease the resonant frequency (and hence the
1/5 flatness limit) and increase the damping. Note that the use of a separate triaxial
block also compromises the specification.
Base strain sensitivity is undesirable when accelerometers are mounted on
thinner structures experiencing strain. The strain in the structure is transmitted
to the crystal where it is indistinguishable from the measured acceleration. Large
errors can result. The solution is to select an accelerometer with very small strain
sensitivity or mount it through a block which attenuates the strain.
Large errors can also result if there is high-frequency energy in the structure
capable of exciting the mounted transducer resonance. The high-frequency energy
can be mechanically amplified, and the output can saturate the electronics. A check
of the unfiltered transducer signal will show any output due to transducer resonance.
Piezoelectric accelerometers are available (usually a special order) with built-in
mechanical filters to mitigate this problem.
68 G. Foss et al.

Some vendors offer a choice between a side connector and a top connector.
The side connector is preferred for permanent applications, but some slack in the
cable near the sensor should be provided for strain relief. A top connector makes
troubleshooting easier, but the loop of cable should be fastened near the sensor so
the cable motion doesn’t cause an unwanted signal.
A typical temperature range for an IEPE accelerometer is from −60 to +250F.
Special accelerometers are available for cryogenic applications to -320F. Higher
temperatures can be achieved with charge accelerometers containing no internal
electronics. For special applications such as engine testing, units are available for
operation at 1200F.
For operation in high humidity or exposure to liquids, hermetically sealed units
should be used, and a watertight connector should be employed.
Although piezoelectric accelerometers are generally robust, they are susceptible
to damage if dropped. If the surface is hard, the acceleration levels experienced may
exceed 1000 g, and permanent damage may be caused. It is prudent to protect a
piezoelectric accelerometer with a rubber sleeve or boot, usually available from the
vendor.

3.2 Force Sensors

Force transducers or force gages are most frequently used to measure dynamic
forces and to determine frequency response functions of receptance, mobility, or
inertance. There are two types commonly used in the dynamic testing field. The first
is the piezoelectric force transducer, which is widely used for frequency response
measurements where the frequency range of interest is wide but does not include the
static load. The second type is the strain gage force transducer, which is more limited
in frequency but does extend to DC. The force gage manufacturer’s specification
sheets provide all the information to enable the choice of force gage to be made for
the more common applications.
The piezoelectric force gage operates using the same principle as the piezo-
electric accelerometer but is in fact one stage simpler, as the deformation of the
piezoelectric material is a direct result of the applied force which is being measured.
Figure 15, left, shows a conceptual cross section of a typical piezoelectric force
gage.
When a force is applied in the direction of sensitivity, the piezoelectric element is
compressed and produces an electrical output proportional to the force transmitted
through it. The transducer can measure both tensile and compressive forces, as its
piezoelectric element is preloaded. A high overall stiffness ensures that it has a high
resonant frequency and also minimal effect due to deformation. The sensitivity of a
force gage is the ratio of the electrical output charge (Coulombs) or voltage to the
applied force. This sensitivity is dependent on the configuration and characteristics
of the piezoelectric material and is defined in units of picocoulombs/lb. (pC/lb) or
millivolts/lb., depending on whether an internal IEPE preamplifier is used. Ideally
the sensitivity of a force gage remains constant throughout its working frequency
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 69

and load range. If the preload is ever relaxed, the full-scale range in tension will
no longer match the full-scale range in compression. During calibration, load cells
should be checked for linearity in both tension and compression.
To minimize the error caused by the inertia properties of the case, the end of the
transducer with the smallest mass should be connected to the surface where the force
is to be measured. The frequency response of the system is dependent not only on
the force gage but also on how it is mounted and the stiffness of the location at which
is mounted. A typical upper frequency limit for a piezoelectric force transducer, stud
mounted to a stiff foundation, will be of the order of several kHz.

3.3 Impedance Heads

A device which combines an accelerometer and a force transducer in one assembly

for the purpose of point mobility measurement is traditionally called an “impedance
head.” The design is a combination based on the characteristics of both transducers.
Although mechanical impedance is defined as the complex ratio of force to velocity
taken at the same or different points in a mechanical system during simple harmonic
motion, the main application of an impedance head is to measure point accelerance
(acceleration/force). For lower frequencies of interest, the point accelerance is
sufficiently estimated by simply mounting the two sensors in close proximity.

3.4 Pressure Measurements

Most machinery causes pressure disturbance in the environment in the form of

acoustic noise, and many hydraulic or aerodynamic devices experience pressure
fluctuations in their working fluids which are associated with mechanical vibrations,
either as cause or effect. The range of pressures of interest is very large, ranging
from a few tens of μPa to a few tens of Pa for environment acoustic measurements
and from a few kPa to tens of MPa for measurements in the working fluids
of machines. Transducers covering the first range of pressures for environmental
acoustic measurements are classed as microphones. Those covering the second area
are described as pressure transducers.
Condenser microphones: Condenser microphones are the standard type of
microphone used for test and measurement and come in multiple sizes and response
fields. The microphone size is referred to by its diameter, with the most common
sizes being 1/8 (3 mm), ¼ (6 mm), ½ (12 mm), and 1 (25 mm) models. Smaller
diameter microphones lend themselves better to high-pressure and high-frequency
applications, while larger diameter microphones have a lower noise floor making
them excellent choices for low sound pressure or low noise measurements.
For optimum accuracy, the microphone sensitivity should remain consistent
(flat) over the desired frequency range and within its intended sound field and
environment. Internal resonance frequencies determine the usable frequency range.
Unlike accelerometers and pressure sensors, whose resonance frequency is beyond
70 G. Foss et al.

Fig. 21 Frequency response of a microphone. (Courtesy PCB Piezotronics)

the typical measuring capability, the microphone’s resonance frequency is within

the usable range. The microphone resonance is damped greater than a piezoelectric
accelerometer, resulting in a minimal effect on the response. Figure 21 is an example
of one response curve. This shows a resonance peak below 20 kHz which is typical
for a ½ high sensitivity condenser microphone. The effect of resonance on the
sensitivity in this case is about +1db.
An object, when placed in a test area, can impact the measurement. Depending
upon the mass of the object and the frequency value of the pressure wave, the
microphone itself can negatively impact the measurement. The microphones own
presence within the sound field can disturb the very pressure wave that it is
attempting to accurately measure. With frequencies below 1 kHz, a ½ diameter
microphone (or smaller) will have a minimal impact on the sound. The wave size is
too big to be affected significantly by the small microphone size. As the frequency
increases (in Hertz), the microphone will have a greater impact and disturbance on
the sound field, and it becomes more important to select the correct microphone
design in order to achieve the most accurate test results. The goal is to measure the
true sound as if the microphone was not present.
In order to compensate for the negative effect the microphone has on the
sound field, manufacturers implement design changes to the diaphragms reso-
nance and the damping of the microphone in order to tune the microphone and
account for the impact of the microphone in the sound field. The three common
condenser microphone designs that account for the different sound fields are free-
field, random-incidence (also referred to as diffuse field), and pressure-response
microphones.
Free-field microphone: The free-field microphone is the most common condenser
microphone. It is designed to be used in an area that is open or free of reflections.
Some examples of free-field environments are anechoic chambers that limit sound
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 71

Fig. 22 Free-field microphone

reflections, or outdoors in an open field or area which do not have objects located
close enough to the microphone which can cause a reflection to be picked up by
the microphone. A free-field microphone is optimized to have a flat frequency
response with sound primarily coming from a single source facing the microphones
diaphragm, and this style of microphone will compensate for its presence in the
sound field. Free-field microphones can still pick up sounds coming from multiple
directions, but the free-field microphone is optimized for a sound source arriving at
zero degrees incidence. See Fig. 22.
Random-incidence microphone: A random-incidence (also called diffuse field)
microphone is designed to measure sound coming from multiple directions. Exam-
ples of areas where a random-incidence microphone would be applied are rever-
beration chambers, or any room with hard surfaces located close by that will cause
reflections, similar to music halls or industrial floor settings. This style of micro-
phone is optimized to measure sound pressure which stems from multiple sources
or reflections. See Fig. 23.The Random-incidence microphone will compensate for
its own presence in the sound field.
Pressure microphone: A pressure microphone is designed to measure sound
pressure in a similar application setting to that of a piezoelectric or ceramic
pressure transducer. Examples of areas where a pressure microphone would be used
would be inside a coupler or cavity or in a flush mounted wall or panel setup.
This style of microphone is optimized for having a flat frequency response while
taking measurements from uniform sound pressure. See Fig. 24. Since the pressure
microphone is typically flush mounted, it is purposely designed to not compensate
for its own presence in the sound field.
Microphone standards: Externally polarized and prepolarized microphones are
governed by the same set of standards which ensure that both will work equally
as well in most applications. The International Electrotechnical Commission (IEC)
with input from the American National Standards Institute (ANSI) published the
standard for test and measurement microphones. The IEC 61094–4 standard applies
to precision “working-class” condenser microphones and dictates specifications for
sensitivity, frequency, dynamic range, long-term stability, and coefficients for the
effects of environmental factors such as humidity, temperature, and atmospheric
pressure. The standard includes dimensional tolerances ensuring interchangeability
among different manufacturers microphones.
72 G. Foss et al.

Fig. 23 Random-incidence microphone

Fig. 24 Pressure
microphone

Many condenser microphone applications involve either designing products to be

more pleasing to the human ear or to assess the effects that sound has on the human
ear to help preserve hearing. The human hearing range is 20 Hz to 20 kHz. Within
this range, the human ear perceives sound differently depending upon the frequency
and the loudness of the signal. In order to understand how sound is perceived by
the ear, Fletcher and Munson performed a study in 1933. Equal loudness curves
were generated to show the relationship between a measured sound and a perceived
sound. Points were plotted to represent how loud a sound pressure level would have
to be in decibels at different frequencies in order to equal that of a 1 kHz sound at
the same loudness level in phons. It was found that the human ear is most sensitive
at frequencies between 1 kHz and 6 kHz and naturally attenuates sound at other
frequencies.
Weighting: ANSI incorporated this data into the S1.4 standard for sound level
meters. Separate frequency weighting systems were established through the years.
The IEC 61052 standard contained A-, B-, C-, and D-weightings. The A-weighting
was established to represent how the human ear would perceive sound at low levels,
40 phons. Testing also verified that the human ear does not attenuate the sound
as much when low frequencies are combined with higher-pressure levels. The B-,
C-, and D-weightings are based on higher-pressure levels. Examples of higher SPL
applications include gunshot testing, blast detection, or aircraft noise.
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 73

Fig. 25 Weighting curves

In 2003 IEC 61672 discontinued the use of the B- and D-weighting designations
and added the Z-weighting scale designation which represented a flat or linear
unweighted scale. Today’s higher-quality sound level meters contain the A-, C-,
and linear (Z)-weighting data, shown above in Fig. 25.

3.5 Pressure Transducers

Pressure transducers make a direct measurement of the force generated at an

interface diaphragm between the working and reference fluids under the difference
in pressure across the diaphragm. The rear of the diaphragm (reference side)
may be vented to enable differential pressure measurements to be made or sealed
for absolute pressure measurements. Strain gages are very commonly used as
the transducing element and offer a response down to zero frequency, though
piezoelectric transducers may also be used, especially in applications not requiring
measurements to zero frequency.
It is good practice to check the calibration of a pressure transducer prior to use,
and it is much more easily done with a strain gage pressure transducer, where static
calibration can be made with a deadweight tester. For more demanding applications,
the calibration should also be dynamic, using a reference transducer of known
dynamic calibration over the frequency range of interest.
One common type of pressure transducer uses a diaphragm, often of stainless
steel, with strain gages fixed to the face on the reference side. Another common type
74 G. Foss et al.

uses a diaphragm constructed out of a single slice of silicon, with mechanical and
electrical features produced by methods developed in the semiconductor industry.
The silicon strain gages so formed give high sensitivity. These transducers tend to
be rather temperature sensitive, and most are offered with built-in compensation
to overcome this effect. It is possible to obtain pressure transducers with the
primary interfacing electronics built-in, leading to very simple installation and
signal conditioning.
The parameters which need to be determined by the user, possibly in consultation
with the manufacturer, include:

(1) Pressure range, accuracy, and resolution required; whether overpressure can
occur; and whether differential or absolute pressure is required
(2) Frequency range.
(3) Temperature range.
(4) Working fluid: some working fluids are chemically active, and the transducers
should be of an appropriate material to resist attack.
(5) Whether there is separate or integral signal conditioning.
(6) Mounting method: the user should ensure that the mounting is of sufficient
accuracy to avoid strain of the transducer when it is fitted; fitting strain can
alter the calibration; and the zero offset is often very sensitive to fitting strain.
(7) Special environmental conditions, such as thermal or mechanical shock, expo-
sure to high g levels, or vibration.

4 Sensor Systems

4.1 Sensor System Architecture

Generally, a sensor system can be broken into five functional blocks, as shown in
Fig. 26. These five blocks do not necessarily represent five different physical items
as multiple functions can be performed in the same package or by the same circuit.
For example, an IEPE sensor combines the sensor element with an impedance
converter (Fig. 26).

Sensor System

User Interface
Analog to
Physical Transducer Impedance (Computer,
Voltage Gain Analog Filter Digital
Phenomena (Sensor) Buffer Oscilliscope,
Converter
etc.)

Fig. 26 Block diagram of a sensor system indicating the components required between a
transducer and an end user
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 75

Transducer: This component of the sensor system is a measurement device or

array of devices previously described.
Impedance buffer: This portion of a sensor system protects the sensor from any
loading effects due to downstream circuitry. All data acquisition systems, amplifiers,
filters, and other such systems have a finite input impedance that can be represented
as a resistor in parallel with a capacitor, usually on the order of 106  and 10−12 pF.
A simplified example of the need for an impedance buffer is the measurement of
a voltage drop across a resistor. The input impedance of a standard multimeter
is approximately 1M. Placing this impedance in parallel with a simple resistive
divider consisting of R1 and R2 produces a reading on the multimeter as

R2 Rmeter
R2 +Rmeter
Vmeasured = Vsource ∗
R1 + RR22+R
Rmeter
meter

When the resistance of the multimeter is much greater than the resistor R2, this
reduces to the desired measurement of
R2
Vmeasured = Vsource ∗
R1 + R2

As the value of R2 gets closer to the equivalent input resistance of the multimeter,
the measurement is attenuated. An impedance buffer is designed with the sensor
impedance in mind, eliminating or greatly reducing the attenuation due to long cable
length or a low resistive load at the end of the cable.
Voltage gain: In most cases, data acquisition systems have a set range of full-
scale inputs that can be measured by the system on the order of tens of millivolts to
tens of volts. As previously discussed, the native output of many of the transducers
described in this section is on the order of 10E-6 to 10E-3 V. In order to match the
range of the data system to the output of the sensor, gain may be added. There are
many higher-order effects that the introduction of gain can influence, such as phase
shifts or other types of signal distortion. A carefully designed gain stage will allow
for maximum measurement resolution with minimal loss of fidelity.
Filtering: Some level of filtering is always required on a digital data acquisition
system in order to prevent aliasing. Aliasing is the contamination of a signal due
to insufficient sample rate and will be discussed in a subsequent section. Beyond
that, it may be desirable to have a filter in a sensor measurement system in order
to reduce contamination of known noise sources, such as electromagnetic coupling
from a motor drive or powerline noise. Again, there are phase and distortion effects
that require careful design when implementing a filter.
Analog-to-digital converter: All instrumentation-based sensor systems end in a
conversion to the digital domain. The approximation of a continuous time signal into
a discretized series of values can be achieved through many different methods, all
with differing accuracy, timing, and speed. The ultimate method which will be used
to analyze the data will largely drive the requirements for analog-to-digital (A/D)
conversion.
76 G. Foss et al.

5 Signal Conditioning

The term “signal conditioning” refers to the analog electronic circuitry between the
sensor and the A/D converter. It may combine impedance buffering, amplification
(gain), and any filtering required to optimize performance of the analog-to-digital
converter.

5.1 Types of Amplifiers

There are four basic kinds of amplifiers commonly used for dynamic signals,
depending on sensor output type: one for single-ended voltage, one for a differential
voltage, one for a charge, and one for constant current excitation, also called
integrated circuit piezoelectric (ICP).
Voltage amplifier: A voltage amplifier is a common immediate interface for
a sensor. It is generally constructed using an operational amplifier (or op amp)
which effectively isolates the sensor from the influence of downstream circuitry.
This stage can also contain gain, which amplifies the sensor output. A schematic
of a simple unity gain noninverting voltage amplifier is shown in Fig. 27 left. This
specific type of voltage amplifier is commonly referred to as a buffer or voltage
follower. An example of a voltage amplifier that inverts the signal and provides gain
or attenuation is shown on the right. The output voltage will be determined by the
ratio of the two resistors: Vout = R 2
R1 Vin .
Charge amplifier: A charge amplifier is a current integrator that produces a
voltage output proportional to the integrated value of the input current. The amplifier
offsets the input charge using a feedback reference capacitor and produces an output
voltage proportional to the total input charge flowing during a specified time period.
A simple charge amplifier is shown in Fig. 28. The resistor Rfb can be used to set the
DC operating point and limit the low-frequency response. Without it the amplifier
output would drift either up or down to one of the supply voltages. “Gain” is not
the best way of describing the transfer function because the output voltage will be

Fig. 27 Operational amplifier configurations

2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 77

Fig. 28 Charge amplifier

Fig. 29 Differential
“instrumentation” amplifier

proportional to the transducer charge, Q, developed across CIN. The output will be
Vout = CQf b :
Differential amplifier: A differential amplifier amplifies only the difference
between two input voltages, suppressing any voltage common to the two inputs.
This is frequently used for Wheatstone bridge circuits which have two mirrored
outputs relative to a reference voltage. The circuit also cancels noise common to
both inputs. A differential amplifier can be constructed from a single op amp,
but a better design is shownin Fig. 29. If the R values are equal, the gain is
Vout = (V2 − V1 ) 1 + R2R gain
. This configuration is commonly referred to as an
“instrumentation amplifier.”
IEPE: “Integrated electronics piezoelectric” (IEPE) is a method by which two
wires from the transducer to the signal conditioning conduct both the power and the
superimposed signal. IEPE is very common for piezoelectric sensors and will be
explained in some detail.
IEPE sensors offer many advantages over traditional charge output sensors,
including:
78 G. Foss et al.

(1) Fixed voltage sensitivity, independent of cable length or capacitance.

(2) Low output impedance (<100 ohms) allows signals to be transmitted over long
cables through harsh environments with virtually no loss in signal quality.
(3) Two-wire system accommodates standard low-cost coaxial or other two-
conductor cables.
(4) Intrinsic sensor self-test feature by monitoring sensor output bias voltage.
(5) Low per-channel cost.
(6) Direct operation into data acquisition instruments which incorporate IEPE
power.

IEPE is achieved by providing a constant current excitation to the transducer

and allowing the voltage to vary. The variations in this compliance voltage are the
transducer output. Typical fixed currents for IEPE data systems are 2 mA, 4 mA,
or 10 mA. This has been standardized so that the circuitry inside of the transducer
package can condition the current excitation into the necessary voltage.
The constant current excitation is achieved by one of several methods. The
simplest implementation uses a constant current diode to provide the excitation
current. These diodes provide a deterministic current path that provides the rest
of the circuitry a bias point to operate around. It is preferable to use one of these
diodes instead of a transistor or resistor as the process variation associated with the
diode production provides a more stable reference than a more complicated circuit.
An example of an IEPE circuit using a diode is shown in Fig. 30.
There are a few constraints on IEPE circuit design that need to be considered.
First, there is no DC component to the output of an IEPE system, since it is removed
by the coupling capacitor. Additionally, as frequency increases, the ability of the
IEPE circuit to drive a long cable is diminished. If an IEPE system is being evaluated
for systems in the multi-kHz range, the cable type, cable length, and expected output
voltage amplitude should be evaluated. A graph illustrating this design space is
shown below. Most modern dynamic data systems have an option for powering an
IEPE transducer. If the data system offers voltage input only, external IEPE signal
conditioning must be used.

Fig. 30 IEPE transducer and

signal conditioning (courtesy
PCB Piezotronics)
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 79

Fig. 31 IEPE sensor system

operating space

Linearity: IEPE sensor systems are linear over a wide amplitude and frequency
range. The limits of linearity are due to mechanical and/or electrical constraints. The
operating range may be thought of as the two-dimensional space shown in Fig. 31.
Mechanical considerations: The mechanical structure within the sensor most
often imposes a high-frequency limit on sensing systems. That is, the sensitivity
begins to rise rapidly as the natural frequency of the sensor is approached:

k
ω=
m
where ω = natural frequency
k = stiffness of sensing element
m = mass loading of the sensing element
This equation helps to explain why larger sensors with greater mass have a lower
resonant frequency.
Electrical considerations: When acquiring low-frequency information, the con-
straints are:

(1) The transducer discharge time constant of the sensor crystal.

(2) The time constant of the coupling capacitor used in the signal conditioner. (If
DC coupling is used, only #1 needs to be considered).

Either or both of these factors can set the low-frequency limit, and it is important
that they are readily understood.
Transducer time constant: The transducer discharge time constant is the more
important of the low-frequency limits, because it is fixed by the sensor design.
80 G. Foss et al.

Consider the IEPE sensor shown previously in Fig. 30. While the sensing element
will vary widely in physical configuration for the various types (and ranges) of
pressure, force, and acceleration sensors, the basic theory of operation is similar
for all. The sensing element, when acted upon by pressure, force, or acceleration,
produces a quantity of charge proportional to the mechanical input. In quartz
IEPE sensors, this charge accumulates in the total capacitance, Ctotal, which
includes the capacitance of the sensing element, plus amplifier input capacitance
and any additional stray capacitance. The result is a voltage according to the law of
electrostatics: V = q/Ctotal . This voltage is then amplified by a MOSFET voltage
amplifier to determine the final sensitivity of the sensor. In ceramic IEPE sensors, the
charge from the crystal is typically used directly by an integrated charge amplifier.
In this case, only the feedback capacitor (located between the input and output
of the amplifier) determines the voltage output and consequently the sensitivity
of the sensor. While the principle of operation is slightly different for quartz and
ceramic sensors, the schematic (Fig. 30) indicates that both types of sensors are
essentially resistor-capacitor (RC) circuits. The low-frequency cutoff of the sensor
will be determined by

1
Fc =
2π RC

where Fc is the cutoff frequency in Hz at 3db of attenuation, R is the internal sensor

resistor value shown in Fig. 30, and C is the total capacitance as above.
Coupling capacitor time constant: If the constant current signal conditioner
is DC-coupled, the low-frequency response of the system is determined only by
the sensor time constant. However, since most IEPE signal conditioners are AC-
coupled, the coupling capacitor may be the limiting factor for low-frequency
measurements. For example, a typical AC-coupling capacitor might be 1 μF.
Assuming a 1 megohm input impedance on the readout instrument, the coupling
time constant is R times C or 1 second. If the transducer itself has a longer time
constant, the coupling capacitor in the signal conditioning may limit the low-
frequency response. To get the full lower limit on frequency range, the coupling
time constant should be at least 10 times larger than the sensor time constant.
Some signal conditioners have a DC-coupled option. This is usually paired with
an adjustment to remove the bias to zero volts in the amplifier. A DC-coupled output
from a standard AC-coupled signal conditioner can be obtained by inserting a “T”
connector in the transducer cable. The signal from the “T” connector contains the
raw sensor data signal which includes the DC bias, typically 8 to 10 VDC. This bias
can be removed by the readout instrument if it has an offset capability.
IEPE precautions: These measures should be followed to reduce risk of damage
or failure in IEPE sensors:

(1) Do not apply more than 20 mA constant current to the sensors.

(2) Do not exceed 30 VDC supply voltage.
(3) Do not apply voltage without constant current protection.
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 81

(4) Do not subject standard sensors to temperatures above 250 ◦ F (121 ◦ C). Consult
the vendor to discuss testing requirements in higher-temperature environments.
(5) Most IEPE sensors have an all-welded hermetic housing. However, due to
certain design parameters, some models may be epoxy sealed. In such cases,
humidity from moist environments may penetrate the housing and degrade the
frequency range specification of the internal electronics. If this has occurred,
the sensors can be restored by baking at 250 ◦ F (121 ◦ C) for 2 h to expel the
contaminants.
(6) Many IEPE sensors are not shock-protected, and a drop on a hard surface can
generate thousands of volts in the piezoelectric material. This will damage the
internal amplifier. Care should be taken to avoid dropping IEPE devices.

IEPE cable length check: For coaxial cables, the cable capacitance and the circuit
resistance form a low-pass filter. If cable length is greater than 100 feet, high-
frequency response could be compromised. This can be checked with the equation
above or solved using the nomograph below. The ordinate is the circuit resistance:
roughly the maximum signal peak voltage divided by the supplied constant current,
minus one milliamp to account for the quiescent power drawn by the internal
electronics, or R = I c−1
V
. For example, if the IEPE current is 2 ma, R would be
5 volts divided by .002A–.001A or 5000 ohms. A 100-foot cable rated at 30pf/ft.
would have a total capacitance of 3000pf. The ordinate value in the nomograph is
scaled by 1000, so the intersection of 5.0 on the abscissa with the 3000 pf line would
occur at 10.6 kHz. The recommended practice is to have this frequency be 1.5 to 2
times the maximum frequency of interest. If it is not, the constant current to the
sensor should be increased (if possible) (Fig. 32).

5.2 Filters

A filter is a transmission path that alters a signal based on the frequency of that
signal. A filter can alter both the magnitude and phase of the signal. The frequencies
that pass through a filter without attenuation in magnitude are said to be in the
passband of the filter. Frequencies which are attenuated by a filter are said to be in
the stop band. The frequency or frequencies where the magnitude is attenuated to
half of its original power (3 dB) is called the corner frequency or cutoff frequency.
The rate at which attenuation increases with respect to frequency from the corner
frequency toward the stop band is related to the number of poles or order of the
filter. In general there are four typical filter implementations: low pass, high pass,
band pass, or band stop (also called notch). In addition to magnitude alteration,
phase is influenced by a filter. The overall phase shift due to a filter is related to the
placement of the filter poles and overall order of the filter. Idealized responses of
these filters are shown in Fig. 33 in the frequency domain.
These idealized responses are not achieved in practice but are approached
as the filter model order is increased. Realistic filters have nonideal passband
characteristics. The most common of these characteristics is pass-band and stop-
band ripple. An illustration of this behavior is shown in Fig. 34.
82 G. Foss et al.

Passband
magnitude

magnitude
Passband

Stopband
Stopband

frequency frequency

(a) Low-pass filter (b) High-pass filter

Passband Passband Passband

magnitude
magnitude

Stopband Stopband Stopband

frequency frequency

(c) Band-pass filter (d) Band-stop filter

Fig. 32 Effect of cable length on IEPE signals. (Courtesy PCB Piezotronics)

In the time domain, there are two basic characterizations of a filter: the impulse
response and the step response. The impulse response is defined as the output of a
system due to an infinitely short, infinite magnitude input. The step response is the
output of a system due to a step increase in the input. This is also the integral of
an impulse response by definition. These tests can help determine the settling time,
bandwidth, and time delay due to a filter.
Designing a filter requires creating a mathematical description of the desired
frequency response and then implementing it in either the analog or digital domains.
There are some closed form solutions that are commonly implemented in systems:
Butterworth, Chebyshev, and Bessel. The trade-offs between these different filter
designs are the amount of ripple in the pass band, the phase response, and the
achievable filter roll-off. These characteristics are summarized in the following table
(Fig. 35, Table 1):
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 83

Fig. 33 Common idealized filter implementations

These systems can be implemented in the analog domain, while the signal is
still a continuous time-varying voltage or in the digital domain as a mathematical
function to alter digital data.
Analog filters: Analog filters are constructed using reactive circuit elements,
which change impedance with frequency. For example, the impedance of a capacitor
is Zc = 1/jwC. At DC, the impedance is infinite, while at high frequencies, the
impedance approaches zero. Placing this capacitor into a voltage divider with a
resistor, a simple filter can be created.
There are two types of analog filters: active and passive. A passive filter is
composed of resistors, inductors, and capacitors and can only attenuate the signal.
An active filter contains an operational amplifier. Some simple examples of filters
84 G. Foss et al.

Fig. 34 Filter ripple

Passband

magnitude
Stopband

frequency

Analog Butterworth Filter Analog Chebyshev Filter

1 1
Magnitude Response

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 0 1 2 3

Analog Inverse Chebyshev Filter Analog Elliptic Function Filter

1 1
Magnitude Response

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 0 1 2 3
Normalized Frequency Normalized Frequency

Fig. 35 Response characteristics of some filters in Table 1

are shown in Table 2; however, many detailed design manuals for active and passive
filter design exist [7, 8].
Digital filters: Digital filters are algorithms that digitized data is passed through
in order to manipulate the frequency content of the data. They are generally
considered as more accurate as they do not rely on physical components that
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 85

Table 1 Common filter implementations

Filter type Pass/stopband flatness Roll-off Phase response
Butterworth No ripple Slow roll-off Phase change near corner
Chebyshev High ripple Sharp roll-off Phase change near corner
Bessel No ripple Very slow roll-off Linear phase

Table 2 Comparison of passive and active analog filters

Filter type Passive example Active example
Low pass
Input Output
Voltage Voltage Input
+ Output
Voltage
- Voltage

Input Output
High pass
Voltage Voltage Input
+
Voltage Output
- Voltage

Band pass
Input Output Input
Voltage Voltage
+ Output
Voltage
- Voltage

Band stop filter

Input Output Input
+ Output
Voltage Voltage Voltage
- Voltage

have finite tolerances and drift characteristics. Digital filters can be realized in a
post-processing script on a host machine or on an in-line digital signal processor
(DSP) chip. A DSP is a specialized microcontroller that operates at much higher
speeds than the A/D converter. This allows a filtering algorithm to run and save
a result before the next sample is taken. The speed of the A/D converter, speed
of the DSP chip, and complexity of the filtering algorithm all influence overall
bandwidth capability. Digital filters are described in terms of their impulse response
and fall into one of two fundamental types: finite impulse response (FIR) and infinite
impulse response (IIR).
FIR filters are used to implement virtually any frequency response with linear
phase. The FIR filter uses a delay and weighted sum approach to alter the frequency
content of the input data. The simplest version of an FIR filter is a moving average. A
moving average is implemented when an output is represented by the average of the
last N inputs, where N is the number of points for the moving average. Considering
a sine wave input, if N corresponds to the total period of the sine wave, a moving
average will produce zero output. The higher the order of the filter, the greater the
86 G. Foss et al.

number of delay stages that are necessary and the faster the DSP needs to be relative
to the data rate.
IIR filters are more analogous to analog filters producing amplitude and phase
responses described by Butterworth, Chebyshev, Bessel, and others. They are
represented by two-pole building blocks called bi-quads, which can be added in
series to create higher-order filters. Designing an IIR filter begins with finding
the transfer function of the desired filter response. That transfer function is then
translated to the discrete, or z-domain, form and implemented in code on a DSP
or in a post-processing algorithm. Computationally they are more complicated than
FIR filters and therefore require additional computing power to implement.

5.3 Analog-to-Digital Conversion

An analog-to-digital converter (A/D converter) is a device that converts an analog

voltage time history to a stream of digital values; see Fig. 26. In an extreme case,
consider a simple comparator as shown in Fig. 36. This circuit will produce 5 V
when the input voltage is above 2.5 V and will produce 0 V when the input voltage is
below 2.5 V. As the input voltage changes, the comparator creates a series of digital
pulses based on the input voltage. As a digital system is by definition a discrete-time
system, this pulse train is recorded at a regular interval, or sample rate, to create a
digital representation of the input analog signal. This simple example demonstrates
the fundamental requirements for A/D converters.
More resolution can be obtained from a system which combines many of these
comparators in parallel, all with a different reference to test the input voltage against.
The resolution of an A/D converter is equal to the number of evenly spaced reference
voltages, or “bits,” that the signal can be compared to.
The maximum voltage that the A/D can resolve is set by the ceiling of
this comparison system. The relationship between the maximum voltage an A/D
converter can resolve and the minimum voltage it can resolve is called the dynamic
Voltage
Input

Reference
Input Voltage
Voltage time

Reference
Voltage
Output

Voltage
time
(a) Comparator Circuit (b) Example output of comparator circuit

Fig. 36 Analog-to-digital conversion using a simple comparator

2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 87

Fig. 37 Two-bit resolution

with four quantization levels

range of the A/D converter. The minimum voltage step an A/D system can resolve
is equal to the range divided by the number of discrete bits:

Vbit = (Vmax − Vmin ) /2total number of bits

Figure 37 shows a simplified example of a sampled sine wave with two-bit

resolution representing four discrete values. If a change in the analog voltage is
smaller than the minimum resolvable voltage step, there may be no change in the
digital output. This error in output is referred to as quantization error. If a transducer
with a low-level voltage output is connected directly to an A/D converter, the signal
change may not be large enough for the output to adequately represent the signal.
This motivates designing an amplifier stage into a sensor system to match the
expected signal levels to the dynamic range of the A/D converter.
Aliasing: In addition to considering the amplitude range of the signal being
digitized, the rate of sampling is important. In order to properly represent the
frequency content of a signal, the sample rate must be greater than twice the highest
frequency being measured in order to satisfy the Nyquist criterion. If this rule is
not followed, the resulting signal may be aliased. This can be demonstrated when a
car wheel or propeller blade is shown in a movie or television show and appears to
move backward. The frame rate of the camera recording the moving object is much
slower than the rate at which it is moving, so the resulting image appears to move
backward. In a data system, aliasing can be visualized as shown in Fig. 38. A higher
frequency is misinterpreted as a lower frequency, due to an inadequate sample rate.
This motivates the need for an analog filter stage prior to the A/D converter to ensure
that no high-frequency content “folds back” into the digital data. These filers are
referred to as anti-aliasing filters.
A/D converter types: There are many ways to implement an analog-to-digital
converter, and as integrated circuit design improves and systems get faster and
more accurate, the philosophy of A/D system design will continue to evolve. A
few examples of A/D converters are briefly described here.
88 G. Foss et al.

Fig. 38 Aliasing of a signal

Input
Voltage

Reference
Voltage

Latch
Digital
and
Output
Decoder

Fig. 39 Flash converter

Flash: The simplest type of A/D converter is the one described above. The flash
A/D converter uses multiple comparators in parallel with incrementally different
comparison voltages to determine a digital output. This is shown in Fig. 39. In
general these types of A/D converters have high-power consumption and poor
performance at high frequencies.
Subranging: Given the example of a flash A/D converter, imagine if a reference
voltage could be adjusted to follow the input voltage. The resulting comparison
voltages would have greater resolution as the dynamic range of the converter is
maximized. A subranging A/D converter operates on this principle. As shown in
Fig. 40, the subranging converter first digitizes the input signal and then uses a
digital-to-analog converter to regenerate the analog voltage to use as a reference
signal for a second stage of digitization. The final digital output is a combination of
the original digitized value and the subranged value, producing an overall higher-
resolution estimate. In order for this kind of converter to work, the A/D converters
and digital-to-analog converter must have a higher number of bits than the expected
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 89

Input
Voltage A/D Converter
A/D Converter D/A Converter
Reference
Voltage

Output Register/Buffer

Digital
Output

Fig. 40 Subranging converter

Input
+
Voltage
- Reference
Voltage

Trigger
Counter Digital
Free Running Input
Counter Output
Clock

Fig. 41 Integrating converter

final answer. There are many methods by which the subranging estimation can occur
beyond the simplest implementation of a series of flash converters. Many other
references are available that describe these designs and the trade-offs between more
complex designs and overall system accuracy (Datel) [7, 9].
Integrating: In an effort to increase accuracy over single counting topologies like
flash or subranging A/D converters, integrating or multi-slope A/D converters were
developed. As shown in Fig. 41, the integrating A/D converter applies an unknown
analog input to an integrating circuit. The integrator builds charge until a known
threshold is passed, triggered by the comparator. The input is then switched to a
known reference voltage that is of opposite polarity, and the integrator output begins
to decay. Both of these time periods of integration are counted. The resulting count
generates the digital representation of the analog input voltage. The accuracy of this
type of A/D converter is higher than previous design efforts; however, each digital
step requires a substantial integration time to achieve that accuracy. Integrating
90 G. Foss et al.

Fig. 42 Delta-sigma converter

A/D converters are still in use for low-frequency applications but have largely been
replaced by delta-sigma converters for higher bandwidth systems [10].
Delta-sigma: A delta-sigma () converter is an extremely fast one-bit digi-
tizer. The digitizer runs hundreds of times faster than the input signal, and the
resulting one-bit data stream is digitally filtered by a process called decimation.
The decimation process changes the effective sampling rate by several orders of
magnitude. Another effect of the decimation process is that the underlying noise
floor is shifted up in frequency. This makes delta-sigma converters well suited
for low-frequency applications that require a low noise floor. Designing a delta-
sigma converter requires expertise in both analog circuit design and digital signal
processing techniques. There are many excellent sources that delve into the details
of these systems [11, 12] The delta-sigma converter is currently the industry
standard for analog-to-digital converters (Fig. 42).

6 Other Considerations

While the choice of transducer and data system usually gets the most thought, other
aspects of installing a sensor system are equally important. There are many types
of cabling and installation considerations. In this section some of these installation-
related design considerations are discussed.

6.1 Grounding

Grounding is the establishment of a path of electrical conduction between a circuit

and an arbitrary reference point. This point is usually at zero potential in reference
to all other system voltages. It is dependent on the system and could be the earth,
the test article, an equipment case, or a bus structure defined as “ground.” Ideally
it would be the zero reference point for both power and data signals. Frequently in
practice, separate circuit paths to ground are provided for power returns, digital
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 91

signals, and analog signals to minimize electromagnetic interference, which is

visible in systems as background noise. In the case of low level, single-ended
analog signals, it is important to establish a single-point ground. Because a remotely
grounded transducer may be at a different potential than the first gain stage, the
resulting “ground loop” could introduce background noise in the data. This influence
can be removed by grounding the cable shield at one end only, usually at the input
to the signal conditioning. For this to be achieved, the sensors must be electrically
isolated. If this is not provided in the sensor packaging, electrically isolated studs
and mounts should be used. For some applications it may be sufficient to apply
aluminum “speed” tape to the test article, to which the sensors are adhesively
attached. The transducer is effectively ground, isolated through several layers of
nonconducting adhesive.
Electromagnetic noise: Noise may be added to the transducer signal because of
the presence of an alternating magnetic field in the vicinity of the cable, such an
electric motor or other running machinery. This causes electromagnetic induction
in the cable and results in a corrupted measurement signal. Precautions against this
effect include:

1. The rerouting of cables

2. The use of short cables
3. Shielding

Rerouting may be a cost-effective method of reducing electromagnetic noise,

as long as the source of the noise is recognized. The use of shielded cable
is a preventative method which is commonly used by test engineers, although
it is practically impossible to eliminate powerline-related frequency influences.
Shielding is achieved by surrounding the wire by a conductive surface, which
maintains the enclosed area free from external fields.
Electrostatic noise: Noise may also arise from electrostatic fields close to cable
runs. Precautions against this effect include:

1. The rerouting of cables

2. The use of short cables
3. Shielding
4. Lowering of line impedance

Electrostatic shielding has different requirements from electromagnetic screen-

ing. Electrostatic shielding provides a conducting surface for the termination of
electrostatic flux lines and commonly consists of either copper mesh or aluminum
foil. A good magnetic shield is usually a good electrostatic shield, but the converse
is not always true.
Lowering of the line impedance is effective at reducing electrostatic noise but
is not effective at reducing electromagnetic noise. Electrostatic noise is directly
proportional to the circuit impedance.
92 G. Foss et al.

6.2 Cabling

The most common method to achieve signal transmission is the use of electrical
cables to connect directly between the transducer mounted on the vibrating structure
and the signal-conditioning equipment. These electrical cables are extremely impor-
tant elements in the total measurement system. The type of cable used varies with the
particular application and may be well-shielded coaxial cables, which traverse long
distances between a piezoelectric accelerometer and its charge amplification system,
or, alternatively, may be short unshielded twisted wires connecting IEPE sensors to
signal conditioning, or strain gage bridges to the balance and amplification system.
The selection of a particular cable form will depend on:

(1) The application

(2) The type of transducer
(3) The cable length
(4) The severity of the environmental conditions

The cables may in some applications be far from the passive devices they are
often.
considered to be and may attract additional voltage signals which have not orig-
inated from the transducer. There are several possible origins for these extraneous
signals:

(1) Electromagnetic fields

(2) Electrostatic fields
(3) Ground loops
(4) Triboelectric effects

All of these influences add unwanted signals to the true signal and are, in general,
referred to as noise.

6.3 Ground Loops

Ground loops are generally more important than electromagnetic induction and
occur because the measuring device is not at the same potential as the remainder
of the system, and thus a voltage drop is created along the length of the cable. It is
also possible that there may be a potential difference between nominal “grounds”
in a laboratory. This influence can be removed by grounding the cable shield at one
end only.
The user should be aware that in an industrial environment, different electrical
power points can have different ground potentials, and that this may cause problems
when several units are interconnected. This problem may be overcome by running
equipment from a power source with a common ground.
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 93

Fig. 43 Common transducer connectors for dynamic measurements

6.4 Triboelectric Effects

Triboelectric charges result when coaxial cable is subjected to distortion either from
tension, compression, or bending motions. The loss mechanism is friction between
the inner cables and outer screening which effectively rubs off the charge. This
influence can be greatly reduced by routing the cable in a way that avoids flexing
and distortion of the cable during testing. Low noise cable is also available which
can reduce this effect even further. The influence of triboelectric effects can be
reduced by:

1. Taping down the cable where it is in contact with stationary objects

2. Allowing strain relief cable loops where relative motion occurs
3. Transferring cable runs from vibrating structures at points with lower vibration
levels

6.4.1 Connectors
Connectors are required to attach cables to transducers, amplifiers, analyzers, etc.
They rely on metal-to-metal surface contact to pass the signal. Each connection
is a possible point of signal loss or noise, and the number should be kept to a
minimum. Connectors have particular problems in difficult environments such as
high humidity, corrosive atmospheres, or underwater applications. Manufacturers
supply adaptors to allow conversion between differing standards. The most Com-
mon transducer connectors for dynamic measurements are BNC, microdot, and
LEMO, shown in Fig. 43.

7 Data Validation

In the measurement process, we wish for the accurate conversion of physical

parameters into a stream of digital values, within the amplitude and frequency range
of interest. Unfortunately, there are many ways for dynamic data signals to become
corrupted or compromised. Signals should be checked for validity before proceeding
94 G. Foss et al.

with detailed analysis operations. This validation can start with an inspection of
the measured time histories to verify that the basic characteristics of the signal are
consistent with the physical mechanisms producing the signal.

7.1 Examples

The following are examples of anomalies that can be observed in time domain
signals. They include overloads, underloads, noise contamination, null data, electro-
magnetic interference, cable and connector-generated signals, cross talk, telemetry
problems, sensor resonance and nonlinearity, insufficient or excessive signal band-
width, sensor mounting, lack of sensor power, and sensor base strain sensitivity.
One of the most common errors in data acquisition is the failure to provide an
adequate upper limit for the dynamic range, due to a sensitivity that is too high. This
results in signal clipping and is shown in the first three figures for different types of
signal (Figs. 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62,
and 63).

11.00
Real
g

-11.00
0.00 s 0.50
(Time)

Fig. 44 Insufficient full-scale range results in clipping of a sine wave with both positive and
negative peaks clipped. The limiting is readily apparent
11.00
Real
g

-11.00
0.00 s 1.60
(Time)

Fig. 45 Insufficient full-scale range results in clipping of a random signal. Clipped tops are visible
but not as apparent as in Fig. 44

11.00
Real
g

-11.00
0.00 s 1.60
(Time)

Fig. 46 Insufficient full-scale range results in clipping of a transient signal. Because of the short
duration, clipping is difficult to detect
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 95

0.02 1.00

Amplitude

Amplitude
V

0.01 0.00
0.00 s 0.13

Fig. 47 Excessive full-scale range results in poor signal definition, shown here as bitnoise in the
A/D converter. This could be due to insufficient gain or a defective transducer

100.00 1.00

Amplitude
Real
g

-100.00 0.00
0.00 s 0.32
(Time)

Fig. 48 Random noise can be expected in a data channel at some low level, originating in the
sensor, signal conditioning, or environment. Whether this is an acceptable level depends on the
full-scale level and how the noise compares to the signal of interest

3.00 1.00
Amplitude
Real
g

-3.00 0.00
0.00 s 0.32
(Time)

Fig. 49 Narrowband noise at multiples of 60 Hz can also be expected at some low level if the
sensor and signal conditioning are not sufficiently shielded or isolated from the AC powerline.
Ground loops are also a common cause of narrowband noise
96 G. Foss et al.

0.10 1.00

Amplitude
Real
g

-0.10 0.00
0.00 s 2.60
(Time)

Fig. 50 Noise can be generated in a signal cable by sudden mechanical flexing, due to the
triboelectric effect. This noise was created in an ICP accelerometer circuit by impacting the cable
bundle against a hard surface

1.00 1.00
Real (RMS)

Amplitude
g

-1.00 0.00
0.00 V 2000.00
(Voltage)

Fig. 51 Null data can indicate a problem. A stream of exactly zero values is unlikely in a physical
measurement and probably indicates signal to range mismatch, A/D failure, or unintentional
scaling by zero

8.00e-3 1.00
Amplitude
Real
g

-0.08 0.00
0.00 s 11.00
(Time)

Fig. 52 Sometimes a test involves a nearby firing or actuation circuit with large momentary
electrical currents. This can induce a pulse in sensor cables. In this case, an IEPE accelerometer
signal appears to contain a large impulse, which is actually induced from a high current pulse
through a nearby cable
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 97

0.50 1.00

Amplitude
Real
g

-1.00 0.00
0.00 s 2.60
(Time)

Fig. 53 RF interference created from a walkie-talkie keyed near the signal cable. The RF is
rectified in the circuitry and appears at much lower frequencies. Sources of RF energy should
not be in proximity to data cables

110.00 1.00

Amplitude
Real
g

-10.00 0.00
0.00 s 2.60
(Time)

Fig. 54 The most common problem in instrumentation cabling is momentary interruption of the
circuit from loose or dirty connectors, broken wires, or faulty solder joints. In IEPE circuitry this
is characterized by short excursions to full scale and a small DC offset, while the circuit recovers
from the electrical overload

0.23
Real
g

-0.06
0.00 s 0.50
(Time)

Fig. 55 In an AC-coupled circuit, it may take time after connections are made to reach
equilibrium. Data should not be acquired until the spurious trend has settled. In some cases this
may take several minutes
98 G. Foss et al.

3.90 1.00

Amplitude
Real
V

-1.70 0.00
0.00 s 21.00

Fig. 56 In RF telemetry systems, dropouts can occur from temporary low signal levels due to
poor antenna positioning or signal cancellation due to multiple RF paths. Typical behavior is the
persistence of the last data value until the link is restored

50.00
Real
g

-50.00
0.01
Real
g

-0.01
0.00 s 2.60
(Time)

Fig. 57 The signals in a data system can couple between channels if there is not enough isolation
and shielding. In this case a 50 g signal in channel 1 couples into an open circuit channel 2. This is
called “cross talk.” In this case, the attenuation is acceptably high, 80 db or more

4.80 1.00
Amplitude
Real
V

-4.80 0.00
0.00 s 7.90e-3

Fig. 58 If the energy received by the sensor is high enough in frequency, ringing of sensor
components may result. In this case, the mass and piezoelectric crystal in an accelerometer
are excited into resonance near full-scale amplitude, making identification of other frequencies
problematic. This can only be observed by unfiltered viewing on a high bandwidth system such as
an oscilloscope
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 99

0.58 1.00

Amplitude
Real
g

-0.20 0.00
0.00 s 0.13
(Time)

Fig. 59 Poor bonding of an accelerometer to a test article may result in nonlinear contact. In this
case, the measured sinusoidal motion from a handheld calibrator should be a clean 100 Hz sine
wave. A partially broken bond may result in small nonlinear impacts. This could also be caused by
a loose mounting stud

50.00
Real
g

-50.00
50.00
Real
g

-50.00
0.00 s 4.00
(Time)

Fig. 60 IEPE accelerometers are powered internally by a constant current, typically supplied by
the data system signal-conditioning circuitry. If the operator neglects to turn the IEPE power on,
a weaker, uncalibrated, and likely distorted signal will still be output by the sensor. This could be
mistaken for a real signal. The upper trace measures tapping on a powered accelerometer mounted
to a beam; the lower trace is the same situation with the accelerometer unpowered

0.50
g

-0.50 Time Point1

0.50
Real
g

Time Point1
-0.50
0.00 2.60
s

Fig. 61 Accelerometer base strain sensitivity. Upper figure shows accelerometer directly mounted
to thin panel subjected to flexing. Lower figure shows same condition with an aluminum block
inserted between the accelerometer and panel. The strain influence is greatly attenuated by the
block
100 G. Foss et al.

0.50 1.00

Amplitude
Real
g

-1.20 0.00
0.00 s 11.00
(Time)

Fig. 62 Thermal drift. Some types of accelerometers are sensitive to small thermal changes in
the environment. This output was caused by repeatedly waving a warm airstream source across a
bare accelerometer case. Smaller effects can result from laboratory air turbulence or outdoor wind
action on a sensor

1.30 1.00

Amplitude
Real
V

-0.30 0.00
0.00 s 7.90e-3

Fig. 63 A step input to a data system can make the anti-alias filter ring. The original signal is the
red trace; the measured result is the blue trace. For analysis which focuses on time histories, the
data system should employ an anti-alias filter which minimizes ringing and overshoot

References
1. Stein PK (1936) A banner year for strain gages and experimental stress analysis-an historical
perspective. Exp Tech 30(1):23–41, (2006)
2. Robinson M (2006) Strain gage materials processing, metallurgy, and manufacture. Exp Tech
30(1):42–46
3. Semiconductor Strain Gage Handbook. Baldwin, Lima, Hamilton, Inc. (1973)
4. Mechanical Engineer’s Handbook, 3rd Edition, Part 1, Instrumentation, Systems, Controls, and
MEMS, Myer Kutz, editor, Chapter 3, Bridge Transducers, pages 69–130, Patrick L. Walter,
Wiley, 2006 (originally printed: Instrumentation and Control, Wiley, 1990)
5. Cady WG (2018) Piezoelectricity: Volume One: An Introduction to the Theory and Applica-
tions of Electromechanical Phenomena in Crystals, ISBN-13:978046835129, Dover Publica-
tions
6. Walter PL (1980) Deriving the transfer function of spatial averaging transducers. ISA Trans
19(3). https://www.journals.elsevier.com/isa-transactions
2 Sensors and Their Signal Conditioning for Dynamic Acceleration, Force,. . . 101

7. Analog Devices (2008) Chapter 8: analog filters. In: Zumbahlen H (ed) Linear circuit design
handbook
8. INTEGRATED M (n.d.) Analog filter design demystified. Retrieved from https://www.
maximintegrated.com/en/app-notes/index.mvp/id/1795
9. Kester W (n.d.) Analog Devices MT-024 TUTORIAL. Retrieved from ADC Architectures
V: Pipelined Subranging ADCs: http://www.analog.com/media/en/training-seminars/tutorials/
MT-024.pdf
10. Kester W, Bryant J (n.d.) Analog Devices MT-027 tutorial. Retrieved from ADC Archi-
tectures VIII: Integrating ADCs: http://www.analog.com/media/en/training-seminars/tutorials/
MT-027.pdf
11. Baker B (n.d.) Texas instruments incorporated: data acquisition. Retrieved from How delta-
sigma ADCs work, Part I: http://www.ti.com/lit/an/slyt423/slyt423.pdf
12. Kester W (n.d.) Analog Devices MT-022. Retrieved from ADC Architectures III: Sigma-Delta
ADC Basics: http://www.analog.com/media/en/training-seminars/tutorials/MT-022.pdf
13. Lawrenson C, Valentino M (2013) Microphone handbook. PCB Piezotronics
14. Himelblau H, Piersol AG, Wise JH (1994) Grundvig handbook for dynamic data acquisition
and analysis. Institute of Environmental Sciences
15. Datel (n.d.) Application note. Retrieved from Subranging ADCs operate at High Speed with
High Resolution: http://www.datel.com/data/ads/adc-an5.pdf
16. Maxim Integrated (n.d.) Tutorial 1870. Retrieved from Demystifying Delta-Sigma ADCs:
https://www.maximintegrated.com/en/app-notes/index.mvp/id/1870
17. PCB Piezotronics (2009) General signal conditioning guide. https://www.pcb.com/
techsupport/docs/pcb/PCB-G0001G-1209_Lowres.pdf
Laser Doppler Vibrometry Measurements
in Structural Dynamics 3
Paolo Chiariotti, Christian Rembe, Paolo Castellini, and Matt Allen

Contents
1 Theory of Vibrometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
1.1 Lasers Sources and Doppler Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
1.2 From Interferometers to Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
1.3 Demodulation of Doppler Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
1.4 Noise and Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
1.5 Critical Aspects in Laser Vibrometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
1.6 Signal Enhancement Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
1.7 Uncertainty and Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
1.8 Laser Safety and Standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
2 Instrumentation, Measurement Issues, and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 129
2.1 Single-Point Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
2.2 Optical Fiber Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2.3 Differential Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2.4 Rotational Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
2.5 In-Plane Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
2.6 Scanning Vibrometers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
2.7 Multi-beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

P. Chiariotti · P. Castellini ()

Università Politecnica delle Marche, Ancona, Italy
e-mail: [email protected]; [email protected]
C. Rembe
TU Clausthal, Clausthal-Zellerfeld, Germany
e-mail: [email protected]
M. Allen
University of Wisconsin-Madison, Madison, WI, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 103

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_4
104 P. Chiariotti et al.

Abstract

Laser vibrometry is a powerful tool for measurement of vibration on a variety

of structures. Lasers do not mass-load or otherwise change the dynamics of a
structure, and so they have enabled measurements from surfaces that are too
light, delicate, hot, etc. to allow conventional surface mounted sensors. The
position of the measurement point can also be changed readily. Thus, laser
vibrometry has also allowed acquisition of measurements over a dense grid of
points, to more completely characterize a structure its deformation shapes, the
evolution of stress waves, or the identification of structural damages than it
might be feasible with other methods. The chapter is divided in two sections:
the first one is intended to provide an insight about the theory behind laser
Doppler vibrometry (LDV), while the second section aims at giving an overview
of the different types of laser Doppler vibrometers that have been developed
so far. The chapter is not intended to give a comprehensive discussion of
laser Doppler vibrometry, but it provides sufficient details about potentials,
issues, and best practice approaches for successfully exploiting such technique
in structural dynamics testing. References are provided to direct the interested
reader to more detailed information as well as to examples of application
cases.

Keywords

Non-contact measurement · Laser Doppler vibrometry · Vibration analysis ·

Interferometry

Nomenclature

AEL Accessible emission limits

A/D Analog-to-digital converter
AOM Acousto-optic modulator
BS Beam splitter
CSLDV Continuous-scan laser Doppler vibrometry
DAQ Data acquisition
DOE Diffractive optical element
DoF Degree of freedom
FDM Frequency demodulation
HPF High-pass filter
I&Q In-phase and quadrature signal
GUM ISO Guide to the Expression of Uncertainty in Measurement
LDV Laser Doppler vibrometer
LASER Light amplification by stimulated emission of radiation
LPF Low-pass filter
PBS Polarized beam splitter
PDM Phase demodulation
PID Proportional-integrative-derivative controller
PLL Phase-locked Loop
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 105

PD Photodiode
QWP Quarter-wave plate
RSSI Received signal strength indication
SM Self-mixing effect
SNR Signal-to-noise ratio
THD Total harmonic distortion
VCO Voltage-controlled oscillator

1 Theory of Vibrometry

This section of the chapter aims at providing an insight about the theory behind laser
vibrometry. It is not intended to give a comprehensive discussion of such subject;
references are provided to direct the reader to more detailed information.

1.1 Lasers Sources and Doppler Effect

The acronym LASER, light amplification by stimulated emission of radiation,

addresses a specific light source developed in the 1960s. Hence, the word LASER
describes the emission of light, which is an electromagnetic radiation, by a
mechanism called stimulated emission. Three main characteristics differentiate light
emitted by a laser source from light emitted by other sources:

• Laser light is monochromatic (narrowband electromagnetic spectrum) and coher-

ent (fixed phase relationship between the electric field values in time, different
times, and space, different locations).
• Laser light is collimated (light is emitted as a low-divergence beam, i.e., the beam
radius does not undergo significant modifications within moderate propagation
distances).
• Laser light can be linearly polarized (the electric field oscillates in a direction –
horizontal/vertical – perpendicular to the propagation direction in a stable way).

Anyone who has used a laser pointer is familiar with the first and second
properties, which cause the laser to have a single color and the beam to spread
little with distance. However, to truly understand vibration measurements with
lasers, one must consider the properties of the beam more closely. The stimulated
emission giving birth to the laser beam takes place into a resonance cavity, a
cylindrical volume of lasing material put between two parallel mirrors, one of which
is partly transparent to make it possible the exit of the laser beam. The resonating
action of such optical cavity swamps out the random nature of the spontaneous
emission and the light passing through the exit-end mirror will then have high
spatial coherence. Indeed, because it is created in a cavity, both transverse modes
(nodes on the cross section of the cavity) and longitudinal modes (nodes along the
length of the cavity) do exist. The most important (and dominant) transverse mode
106 P. Chiariotti et al.

is the uniphase mode TEM00 . This mode induces the typical circular symmetry
and Gaussian profile in the beam light intensity. Higher transversal modes can
be suppressed by careful design. Longitudinal modes are directly related to the
spectral characteristics of the laser. Let’s reverse the perspective, and let L be
the length of a given optical cavity (distance between the mirrors corrected for
the refracting index of the lasing medium). The frequency gap fm between two
consecutive modes (m) is cn /2L, where cn represents the speed of light in the
medium with refraction index n. This means that assuming a certain number of
modes existing, it might happen that for certain L the spectral width of the laser
light (frequency difference between the lowest and highest modes) is too small to
be detected by common spectrometers and the laser appears to be monochromatic
even if it is not single mode. When this happens (number of wavelengths within
the laser length) such spread of frequencies induces a finite coherence length
(length over which coherence degrades significantly). A truly single mode laser
(true monochromatic laser) would have an indefinitely long coherence length. A
typical laser that is utilized in the majority of commercially available vibrometers
is a multi-mode helium-neon laser, in which at least two laser modes exist. The
interference of two modes leads to a finite coherence length and thus causes the
intensity of the optical signal to vary periodically along the optical path difference
between the interfering beams. Such phenomena, which will be further explained
in the section related to the backscatter issues, do induce alternating visibility
maxima and minima and therefore might significantly influence the quality of the
measurement.
The Doppler effect is a very well-known effect in physics. It is the key in laser
Doppler vibrometry because it allows the possibility of directly relating the velocity
of a target to a Doppler frequency shift in a beam of light. The Doppler effect
for sound waves is a phenomenon that one can experience in everyday life, for
example, when observing the sound of a passing automobile. As a sound source
such as this moves toward a stationary observer, the number of wavefronts emitted
by the source must be constant, but the travel distance decreases (assuming a
constant sound propagation speed at Mach<1). Hence, the number of wavefronts
impinging on the observer in the unit of time increases. Therefore, pitch of the
sound seems to increase for the observer. This phenomenon can be well explained
by the velocity superposition principle. In contrast, when considering light waves,
the simple model of the velocity superposition that supports the understanding of
the effect in the case of sound is not so straightforward. One can use a relativistic
approach, which takes into account a transformation of space and time coordinates,
as explained in [1]. However, recent studies on retardation effects in laser Doppler
measurements [2] demonstrated that it is not necessary to consider relativistic
effects in measurements of a two-beam interferometer, since the source and the
observer are in the same reference frame. This condition is the one depicted
in Fig. 1a.
This case might be considered as a double Doppler shift, one involving the source
S and the moving object P and the other one involving the receiver O and the moving
object. The frequency observed by P (fP ) can be expressed as
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 107

Fig. 1 Diagram for the calculation of the Doppler shift on scattering by a moving object P: source
S and observer O in different locations (a); source S and observer O in the same location (b)

f  v 
fP =   1 + cos ϑ1 (1)
1 − v 2 /c2 c

This holding, when the velocity is much less than the speed of light (v < < c),
which is almost always the case in standard applications, it might be possible to
expand in a power series in v/c and stop the expansion to the first order, so to obtain
the Doppler shift (f ) as shown below in Eq. (2). The equivalence c = fλ and the
prosthaphaeresis formulas were utilized to obtain the well-known expression for the
Doppler shift.

fν 2v ϑ1 + ϑ2 ϑ1 − ϑ2
f = fO − fP = (cos ϑ1 + cos ϑ2 ) = cos cos (2)
c λ 2 2

In most vibrometers, the source and the observer are located at the same place
(backscatter configuration – Fig. 1b). The Doppler shift in such configuration can
be expressed as reported in (3).

2v
f = cos ϑ (3)
λ

A vibrometer (Fig. 2) measures the motion of a target by exploiting

the built-in interferometer capability, which resolves the phase of the back-
reflected/backscattered light once the target is illuminated with coherent light. The
motion is measured along the direction of the light ray because the illuminating and
detecting rays have the same optical axis and polarization; thus the measurement
can be considered to be a one-dimensional propagation problem where the
electromagnetic wave equation simplifies to

∂ 2 E (x, t) 1 ∂ 2 E (x, t)
= . (4)
∂x 2 cn2 (x, t) ∂t 2
108 P. Chiariotti et al.

Fig. 2 Base optical configuration of a vibrometer with beam splitter (BS) and photodetector (PD)

In (4), E(x, t) represents the norm of the linearly polarized electric field vector,
cn (x, t) = c/n(x, t) is the speed of light in the medium having a refraction index
n (x, t), and c is the speed of light in vacuum. For x = 0 E(x, t) results in
E(x, t)x = 0 = E0 cos (2π ft).
For homogeneous media (no variation of n in space) and for an observation
time tm in which the time dependency of n can be neglected, if considering, for
the target motion st (t) = s0 + s(t), the following conditions |s0 | > > |s(t)|max and
|v(t)|max < < cn , and by defining sd as the detector position, the Doppler shift,
which represents the time derivative of the phase variation of the electric field of
the light wave at the photodetector, to an approximation in the order v/cn , with v
instantaneous velocity of the target, can be expressed as

n ≈ const
dφ(t) ds opt (t) ds(t)
fD = =k = 2kn = 2knv(t) (5)
dt dt dt

where k = 2π f /c is the wavenumber of the coherent light in vacuum and sopt

represents the optical path length. The Doppler shift is then generated through
an optical path length variation. The interested reader might find more detailed
information related to the mathematical derivation of (5) in [3].

1.2 From Interferometers to Vibrometers

1.2.1 Optical Homodyne and Heterodyne

The motion of the target induces a phase-modulation (and a frequency modulation
at the same time) on the electric field of the measurement beam. This modulation
is at optical frequencies (in the THz), and typical photodiodes cannot deal with
this frequency directly; therefore optical beating has to be created by mixing the
measurement beam with a reference beam (e.g., by using another coherent light
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 109

source of frequency f2 and fixed phase relation or by exploiting a second laser mode
of the same source). Independent of the mechanism that generates the beating, the
mixing of the light waves takes place on the sensitive surface of a photodetector, a
device that converts light intensity into an electric signal. Interference takes place
properly (uniform over the full photodetector area) if interfering beams are well
aligned and mode-matched. This has two major implications: their intensity profiles
overlap, and their wavefronts have the same curvature on the detector. Photode-
tectors show a limited bandwidth; therefore they act as a low-pass filter (LPF) for
optical frequencies. Saying that beating occurs on the surface of the photodetector
means nothing else than a photocurrent is generated that is proportional to the
intensity of light impinging on it. However, the photodetector is of significance only
as a measurer of light intensity.
The output photocurrent id (t) from the photodetector is proportional to
light intensity I(t), which depends on the square of the total electric field
Ed (t) = Em cos (2π f1 t − ϕ(t)) + Er cos (2π f2 − ϕ0 ), being Er the amplitude of
the reference beam, f1 the frequency of the light source generating the measurement
beam, ϕ(t) the time varying phase containing the Doppler information, and ϕ0 the
initial phase difference between the two beams. Considering that the photodetector
cannot. This holding, the photocurrent id (t) from the photodetector can be expressed
as in (6)

   
1
id (t) = A 2
Em + Er2 + Em Er cos (2π (f1 − f2 ) t − φ(t) + φ0 ) (6)
2

where A = 1 G
2 Z0 is defined as amplification with conversion parameter G = ηq/hν
and the impedance of free space Z0 = ε01c0 ≈ 120 π . Here, the parameters are
the quantum efficiency η, the elementary electric charge q, the Planck constant h,
the light frequency ν, the electric constant ε0 , and the speed of light c. The factor ½
takes the beam splitter into account always necessary to interfere two collimated
beams without generating interference fringes. Obviously, a constructively interfer-
ence requires destructive interference somewhere else to conserve the energy. The
2
power of an electromagnetic wave is P = E Z0 where Edenotes the mean electrical
field for the time constant of the considered photodetector. The output thus contains
a DC term, which is proportional to the total intensity and a beat frequency term
that is proportional to the product of the amplitudes Em Er (by a linear amplification
A). The beat term √ can also be expressed as a function of the intensities of the two
beams, i.e., as Im Ir .
An interferometer without frequency shift between reference and measurement
beam is called homodyne interferometer. In a homodyne configuration, the current
id (t) is a direct function of the phase dependency, as shown in (7).

   
1
id (t) = A 2
Em + Er2 + Em Er cos (φ(t) − φ0 ) (7)
2
110 P. Chiariotti et al.

However, there are limitations due to the imperfect nature of real photodetectors.
Homodyne vibrometers show high sensitivity to photo-receiver nonlinearity and
susceptibility to electrical hum and noise, and the optical alignment is critical
(in the motion direction configuration) as discussed in [4]. For such reasons a
heterodyne-based optical arrangement, where a frequency shift is induced on one
of the beams (typically the reference beam), is more typically adopted. This makes
the frequency difference |f1 − f2 | equal to the frequency of the shifting device (|fc |),
which turns out to be the carrier for the phase modulation generated through the
optical path length variation linked to the motion of the target. In a heterodyne
configuration, the signal of the photodetector can therefore be expressed as in (8),
where 0 < ε < 1 defines the heterodyning efficiency (which considers degradation
of the AC current for optical distortions and misalignment and depends on the light
intensities, coherence, and polarization).
   
1
id (t) = A Em + Er + εEm Er cos (2πfc t − φ(t) + φ0 )
2 2
(8)
2

In the next sections the beat signal will be computed for different interferometric
configurations.

1.2.2 Michelson Interferometer

The Michelson interferometer configuration that had been implemented the most in
commercial vibrometers is represented in Fig. 3. Such representation is far from the
basic implementation of a Michelson interferometer, which presents only the central
beam splitter, the mirror on the top and one photodetector (PD1). The importance
of the configuration reported in Fig. 3, however, is related to the possibility of

Fig. 3 Optical scheme of a vibrometer based on Michelson interferometer – two detectors

configuration
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 111

solving the directional ambiguity that is characteristic of the basic configuration

of Michelson interferometers.
The beam emitted by the laser source is split in two coherent parts at the
polarizing beam splitter PBS. The polarizing beam splitter is utilized to split
light efficiently to the photodetectors. One beam (reference beam) goes toward
a reflecting mirror placed inside the vibrometer, passing through a quarter-wave
plate (QWP1) that changes the polarization of the beam from linear to circular
and vice versa after the reflection. The second beam (measurement beam) exits
the vibrometers (before it passes through a quarter-wave plate – QWP2 – as well),
interacts with the moving target, and is backscattered to the vibrometer. The second
passage through the quarter-wave plate changes again the beam polarization from
circular to linear. A non-polarizing beam splitter (BS) splits the two beams on the
photodetectors PD1 and PD2, where interference between the beams takes place. A
quarter-wave plate (QWP3) is placed on the line of sight of PD1, in order to obtain
a signal in quadrature with the signal given by PD2 (π /2 out-of-phase). The laser
beams at each photodetector are perpendicularly polarized; therefore a polarizer (P)
is needed to obtain an optical beat. The signals coming out of the photodetectors in
such an arrangement are called sine and cosine signals, or iI and iQ signal, where I
stands for in-phase and Q stands for quadrature.
These two signals can be written in terms of light intensity as reported in (9).

iI = II cos (φ(t) + φ0 )
(9)
iO = IO sin (φ(t) + φ0 )

The velocity sign can then be found using (10).

+v ⇒ iI ∝ + cos, iO ∝ + sin
(10)
− v ⇒ iI ∝ + cos, iO ∝ − sin

The relative phase of such signals contains information on the sign of fD and thus
on the sign of velocity v.
Advances in low-cost electronic devices have made it possible to tackle the
issue of demodulation and direction ambiguity using, for instance, phase shifting
approaches such as the one reported in Fig. 4.
By multiplying the signals coming out the photodetectors (I&Q signals), which
carry the Doppler shift information (ωD = 2π fD ), respectively, with two carriers
(ωS ) that are π /2 radians out of phase, and summing the resulting signals, a
frequency modulated output that is sensitive to the target velocity can be obtained.
This can then be passed to a standard demodulator to get the velocity information.
The phase shifting approach was a quite common approach in oldest vibrometers.
Nowadays, the tendency is to digitize the I&Q signals and compute the phase with
sign by an arctangent operation. This latter approach will be further detailed in the
demodulation of Doppler signals section.
112 P. Chiariotti et al.

Fig. 4 Functional scheme of electronic phase shifting approach

Fig. 5 Optical scheme of a vibrometer based on Mach-Zehnder interferometer

1.2.3 Mach-Zehnder Interferometer

The Mach-Zehnder (MZ) interferometer, based on a heterodyne configuration, is
the most common scheme implemented in commercial vibrometers (Fig. 5). One of
the advantages of the MZ interferometer is the ability it gives to address the sign
ambiguity that was present in the Michelson configuration.
The most common solution employs a Bragg cell to produce a frequency shift
(f + fB ) in the reference beam. The Bragg cell is an acousto-optic modulator based
on a vibrating transparent crystal that produces a travelling train of parallel plane
acoustic wavefronts that move at a certain velocity across the crystal. The moving
sound waves are generated by a piezoelectric actuator, driven by a voltage signal at
frequency fB , that is placed on one plane surface of the crystal. An acoustic absorber
is mounted on the opposite side of the crystal, at a slanted angle, in order to avoid
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 113

acoustic reflections inside the crystal. The wave train acts as a moving diffraction
grating, since the acoustic waves, pressure waves, induce a spatial modulation on
the refraction index of the crystal (the index of refraction is related to the strain field
in the crystal).
The quarter-wave plate (QWP) and polarizing beam splitter (PBS2) placed on the
measurement beam path constitute an optical isolator that contribute in maximizing
light intensity at the photodetectors and at preventing stray light from travelling
toward the laser cavity. Interference takes place at the photodetector PD1 and will
therefore produce a photocurrent (iPD1 ) out of PD1 that can be expressed as in (11),
where ε represents the heterodyning efficiency.
   
1
iP D1 =A Er + Em + εEr Em cos (2πfB t − φ(t) + φ0 )
2 2
(11)
2

The Mach-Zehnder scheme makes it possible to implement a differential ampli-

fication scheme with the photodetectors. Such a configuration is useful to reject
common mode noise. Indeed, the signals at PD1 and PD2 are π radians out of
phase (the sign in the alternating component of Eq. (11) changes at the two PDs),
because of energy conservation. The use of a differential transimpedance amplifier
will then produce a low-passed-filtered signal at the amplifier output without DC
component that can be expressed as in (12), where K represents the amplification of
the transimpedance amplifier.

uAMP (t) = K A ε Er Em cos [2πfB t − φ(t) + φ0 ] (12)

This configuration therefore provides a relevant improvement in terms of signal-

to-noise ratio. Also, all of the DC components cancel, if they are equal to each other.
The term in (14) thus represents the heterodyne Doppler signal with a carrier
frequency fB and modulated phase angle ϕ(t). The third term within the cosine
function (ϕ0 ) represents the starting phase. When the target moves undergoing a
displacement s(t), i.e., s(t) = ŝ · sin (2πfvibr t + φs ), the Doppler shift in Eq. (5)
can be reformulated in terms of the modulated phase angle (13)

4π · s(t) 4π · ŝ
φ(t) = = · sin (2πfvibr t + φs ) (13)
λ λ

and frequency shift (14) with respect to the center frequency fB .

ω dφ(t) 2 · v(t) 2 · v̂
fD (t) = = = = · cos (2πfvibr t + φs ). (14)
2π dt λ λ

The terms 4πλ·ŝ and 2·λv̂ , respectively, represent the phase deviation (ϕ) and
the frequency deviation (f ) of the modulated carrier. It is clear, then, that the
instantaneous frequency of the heterodyne signal can resolve the sign of the velocity
114 P. Chiariotti et al.

vector as long as the condition fB > |fD | is respected. Note that fextracted > fB implies
the target is moving toward the interferometer.
In standard HeNe vibrometers (λHeNe ∼ = 632 nm), fB = 40 MHz, therefore the
maximum measureable peak velocity is about 10 m/s (as long as the vibration
frequency does not exceed a few MHz). This value is usually sufficient in many
applications; when it is not the case, a larger frequency shift, fB , should be adopted.
Indeed, acousto-optic modulators up to 80 MHz and more are commercially
available.
A lens is also present in the scheme reported in Fig. 5. This allows the possibility
of increasing the amount of light backscattered by diffusive surfaces.

1.2.4 Self-Mixing
The use of the self-mixing (SM) effect (also named optical-feedback interferometry,
or less frequently, backscattered-modulation) for developing optical vibrometers has
gained a large interest among researchers and producers of commercial systems
over the years. This is mainly due to the clear advantages offered by the optical
configuration of such systems. The first paper describing the use of “self-mixing” to
measure velocity is from Rudd [5].
The SM vibrometer is based on reintroducing the laser radiation that is reflected
back from the target directly into the laser cavity (this is usually avoided in
traditional LDV). This configuration makes it possible to have a very compact
optical setup and therefore to reduce manufacturing costs. In SM vibrometers, the
interference process takes place directly inside the laser cavity (typically a laser
diode cavity) due to the slight frequency difference between the frequency of the
radiation emitted and the frequency of the Doppler-shifted backscattered radiation.
The interference signal is collected by a photodiode that is usually installed in the
laser cavity for monitoring purposes. The SM interferometric configuration requires
neither the presence of a reference arm nor discrete optical components; hence the
setup, shown in Fig. 6, is extremely simple and inexpensive [6, 7].
The power emitted by the laser diode is amplitude-modulated by an interfero-
metric waveform F(ϕ) which is a periodic function of the back-injected field phase
ϕ = 2ks, where k = 2π /λ is the wavenumber and s is the distance from the laser
diode to target. The power emitted by the laser diode can be written as

P (φ) = P0 [1 + mF (φ)] (15)

where P0 is the power emitted by the unperturbed laser diode and m is a modulation
index. The modulation index m and the shape of the interferometric function F(ϕ)
depend on the so-called feedback parameter C:

√
s0 κ 1 + α 2 1 − R2
C= · · √ (16)
Aopt Llas nlas R2
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 115

Fig. 6 Conceptual design of a self-mixing vibrometer

Fig. 7 SM interferometric signals for different optical feedback regimes. Upper-left displacement
signal, (a) C < < 1, (b) C ≈ 1, (c) C > 1

where Aopt > 1 is the total roundtrip power attenuation in the external cavity; α (typ.
α = 3) is the laser diode linewidth enhancement factor; κ< 1 (typ. κ = 0.5), which
accounts for a mismatch between the reflected and the lasing modes; Llas is the laser
cavity length; nlas is the cavity refractive index; and R2 is diode output facet power
reflectivity.
The C parameter, which depends on both the amount of optical feedback and
on the target distances0 , is of great importance, because it determines different
feedback regimes. For C < < 1 (weak feedback) the function F(ϕ) is a cosine, just
like the usual interferometric waveform. When C approaches unity, the function
F(ϕ) resembles a distorted cosine. For C > 1 (moderate feedback regime, which
corresponds to usual optical attenuation and laser-to-target distance), the function
F(ϕ) becomes sawtooth-like and exhibits hysteresis. Examples of the waveforms
obtained in the different feedback regimes are reported in Fig. 7.
116 P. Chiariotti et al.

An interesting solution to process the interferometric signal is to use an electronic

feedback loop that acts on the laser diode injected current to achieve wavelength
modulation. In this way the interferometric phase is locked to half-fringe, and
an active phase-nulling technique makes it possible to obtain a wide dynamic
range [8].

1.3 Demodulation of Doppler Signals

The modulated heterodyne signal from the optical interferometer in a heterodyne

scheme usually shows a frequency range of 8 to 72 MHz (40 MHz Bragg cell
+ 10 m/s full scale for a HeNe laser). Different decoding schemes are needed
for meeting these ranges. Furthermore, one can use phase demodulation (PDM)
to obtain displacement measurements, while frequency demodulation (FDM) is
utilized for velocity measurements (Fig. 8).
Frequency demodulation can be performed using the same technology developed
for signal transmission as, for instance, radio-based chips. However, a more complex
technique seems to be required for vibration applications. Indeed, in vibrometers
demodulation is performed as a down mixing of the interference signal with the
reference signal driving the Bragg cell. This down-mixed signal is then filtered
depending on the frequency range chosen (the latter determining the sensitivity
and linearity parameters of the down mixed signal), in order to obtain the velocity
signal. To this basic structure, usually smarter filtering strategies (such as amplifiers,
signal conditioners, tracking filters, Butterworth, etc.) are added, in order to improve
the signal-to-noise ratio (SNR). Commercially available laser Doppler vibrometers
(LDVs) use analog frequency demodulation to convert the Doppler frequency into
an analog voltage proportional to the velocity of vibration. A widely used FM
demodulation approach is based on the use of phase-locked loops (PLLs). The
baseband signal, corresponding to the velocity signal, is extracted at the input of
the voltage-controlled oscillator (VCO). On the one hand this approach makes it
possible to achieve high linearity and resolution, with a good SNR; on the other
hand, the bandwidth is quite limited, and this scheme is sensitive to signal dropouts.
The measurement of target displacement can be performed by counting inter-
ference fringes, as is typically utilized in the low frequency range of vibration,

Decoder

Frequency
I& 2v O
Demodulation
Interferometer
I 2s
Phase O
Demodulation

Fig. 8 Demodulation of Doppler signal

3 Laser Doppler Vibrometry Measurements in Structural Dynamics 117

Analog demodulated signal

FM signal LP Filter
Phase Detector VCO
from SLDV head

Fig. 9 FM demodulator based on PLL

or through phase decoding, as in the case of ultrasonic vibration up to 20 MHz.

Such decoders work on the high-pass filtered signal (>50 kHz) in order to eliminate
disturbing noise and utilize PLL circuits to detect the phase of the signal in a
range of ±90 degrees. These limit the resolving capability to relatively small
displacements (±λ/8 or about 150 nm peak to peak for a He-Ne Laser). In order to
obtain an improvement in resolution, the interference signal is usually interpolated,
and resolution down to λ/80 is obtained. In contrast, analog velocity decoders can
be matched to a variety of processing requirements such as very high frequency,
DC response, and high velocity range. These are still the preferred choices when
it comes to measurements at high speeds (10 m/s or more) and high frequencies
(Fig. 9).
The reader should be aware, however, that analog electronic components are
sensitive to drift as well as aging, and there are certain limits in linearity. Moreover,
these components must be calibrated in order to guarantee accurate measure-
ments. Accuracy, repeatability, and traceability strongly depend on each component
(amplifier, resistances, and capacitance) in the analog signal processing chain, with
particular emphasis on the quality of phase and frequency demodulators. In addition,
analog demodulation does not make it possible to adapt processing parameters to the
actual signal characteristics, while thermal drift can induce modifications in settings.
All these aspects have induced LDV producers to move to digital demodulation,
since digital demodulation can indeed overcome some of the problems and draw-
backs of analog demodulators. In digital demodulation, the interference signal from
the interferometer is digitized using analog-to-digital (A/D) converters operating
at an adequate sampling rate. Computational algorithms then process the signal
numerically. Data can be acquired for a defined time interval and then processed
in offline mode, or, if the A/D converter is associated with a circular buffer, data can
be continuously sampled and transferred to the processing unit for quasi-real-time
demodulation. The main advantages of LDVs based on digital demodulation include
minimal drift from calibration, amplitude resolution only limited by optical noise,
sub-nanometer displacement resolution within the full velocity, and frequency range
from DC up to the MHz range.
It is worth highlighting that digital vibrometers such as these have been accepted
by the German Institute of Standards (PTB) for traceable calibrations according to
the ISO 16063-11 standard. In fact, in digital decoding the repeatability in frequency
demodulation is related only to the accuracy of the timing in the A/D conversion,
118 P. Chiariotti et al.

and this is a very stable parameter, easily verified with high accuracy. All that
remains then are the algorithms implemented in the software, which are completely
deterministic.
The main limitations of digital decoding are related to the very high sampling
frequency and to the large storage required when vibration velocity increases in a
heterodyne scheme. For this reason, digitally decoded vibrometers usually have a
lower full-scale velocity range than the corresponding analog versions even though
the interferometer is the same.
To overcome such problems the most common approach is the conversion of the
heterodyne signal into an I&Q homodyne signal by quadrature mixing. An analog
circuit mixes the Doppler signal with the signal from a Bragg-cell oscillator, to
obtain Doppler signals in the baseband. Indeed, the RF carrier itself, introduced by
the Bragg cell, does not contain any Doppler information and can be suppressed.
Mixing is done with two reference signals, π /2radians out of phase to each other
so as to obtain two quadrature signals, named I&Q as above. In this way the noise
rejection capability of the heterodyne optical scheme is coupled with the advantage
of using baseband signals (about 3 MHz for 1 m/s of range of HeNe laser), which
require a less critical signal sampling. After the quadrature conversion, the two
signals (I&Q) are sampled and further processing is performed as in the case of
the homodyne interferometer. One of the most common demodulation algorithms is
the arctangent phase demodulation.
The signal sampling and the data processing takes some time, and so it is no
longer possible for the digital system to operate in real time as the analog systems
do. Furthermore, the time delay caused by the digital demodulation can vary for
different settings of the digital processor; this needs careful consideration when
performing vibration measurements together with analog sensors as a reference
(e.g., a load cell for modal analysis).
Digital demodulation also makes it easier to implement additional algorithms
into the signal processing, such as dropout detection.
After demodulation, an output signal related to the vibration velocity can be
made available to the user in different forms. It can be a digital signal with the
samples stored in a mass memory or transferred to a PC through a digital line (e.g.,
a SPDIF or serial bus), or it can be converted to an analog signal. The first solution
is conceptually simpler and more accurate, but it makes it difficult to synchronously
acquire any other signals with standard data acquisition boards.
The interested reader might refer to [3, 4] for further information regarding
demodulation of Doppler signal in vibrometers.

1.4 Noise and Resolution

Vibration amplitudes much smaller than the diameter of an atom can be relevant in
a vibration spectrum. There are applications where picometer resolution is required
as, for example, ultra-high-frequency (UHF) resonators. In the frequency regime
above 100 MHz, the vibration amplitudes rarely exceed a nanometer. Resolution is
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 119

defined as the smallest detectable magnitude of the desired measurement quantity

which, in the case of laser vibrometry, is the smallest detectable vibration amplitude
in the examined vibration bandwidth. The root mean square amplitude of the noise
is usually defined as a noise metric and depends on the bandwidth of the considered
signal. The noise power depends linearly on the bandwidth for white photon shot-
noise which results in white displacement noise. Therefore, the noise level of a laser
vibrometer is usually defined in displacement per square root Hertz.
The significant noise sources of a laser vibrometer have to be analyzed in order
to understand the limits of resolution
 [3]. This thermal-noise contribution i th to the
2 dt
iD
mean detector current i D = T can be estimated with the following Equation

4 kB T B
u2th = K 2 (17)
Rd

with kB the Boltzmann constant, Rd the detector load resistance, and B the bandwidth
of the photodetector.
Besides such a technical noise source, the quantum nature of light defines the
ultimate limitation for any interferometer. The light power cannot be constant since
the light energy is transferred by quanta with energy E = h f. This effect leads to a
mean square voltage noise of

u2sh = K 2 Gη q B (Pm + Pr ) (18)

2
Em
by photon shot-noise with Pm = 2Z0 the power of the measurement beam at a
2
Er
single photodetector and Pr = the power of the reference beam at a single
2Z0
photodetector. The third noise contribution originates in the signal processing
and decoding. For example, analog down mixing to a frequency band that is
suitable for a certain analog-to-digital converter introduces phase noise from the
reference oscillator. The analog digital conversion itself contributes digitization
noise. Assuming a direct digitization without down-mixing of the heterodyne carrier
signal with n bits would result in a quantization interval of q = Ur /2n − 1 which
yields the mean square voltage noise contribution within the evaluated detector
bandwidth B
2
1 Ur 2B
u2qn = , (19)
12 2n−1 fS

where Ur is the full scale of the A/D-converter

√ and fS is the sampling frequency.
Considering a full scale of Ur = 4KGε Pm Pr , Eq. (8) results in the noise
contribution by quantization noise

4 (K G ε)2 Pm Pr 2B
u2qn = (20)
3 · 22 n−1 fS
120 P. Chiariotti et al.

However, as mentioned previously, the ultimate limit is defined by the photon

shot-noise. It can be concluded from Eqs. (8) and (12) that the power of the
heterodyne signal and the power of the shot-noise both depend linearly on the power
of the reference light. The amplification of the interference signal by the reference
light is called coherent amplification and is absolutely noise free as the shot-noise
power increases by the same factor as the signal power. A properly designed
photodetection in a laser Doppler vibrometer is, therefore, shot-noise limited by
adjusting the reference light power to a level where the photon shot-noise exceeds
the thermal noise. A controlled amplifier (limiter) usually amplifies the detector
signal to a constant signal strength. Thus, it can be concluded that the quantization
noise is full-scale and defined by Eq. (20). Usually, the noise level is defined by
the decoder when the signal strength of the detector is maximum. The SNR of the
shot-noise-limited detector signal is

2Gε2 Pm Pr 2η ε2 Pm Pr 2 λ η ε2 Pm Pr
SN R = = = . (21)
q B (Pm + Pr ) h f B (Pm + Pr ) h c B (Pm + Pr )

To realize a shot-noise limited demodulated signal the SNR of the digitized

carrier signal

fS
SNR dig = 3 · 22 n−1 · (22)
2B
needs to be higher than the shot-noise-limited SNR in Eq. (21). Heterodyne
detection enables measuring always at the optimal sensitivity point of the fringe
pattern. At this point it is
 2
 2 8π K Gε ŝ/λ
u2S = 2KGε φ(t) Pm Pr = (8π K Gε s(t)/λ ) Pm Pr =2
Pm Pr .
2
(23)

The shot-noise limited resolution limit is reached at

 2
8π KGε ŝ/λ
u2S = Pm Pr = u2sh = 2K 2 G q B (Pm + Pr ) (24)
2
and results in a smallest resolvable displacement amplitude of
 
λ qB (Pm + Pr ) λ h f B (Pm + Pr ) λ
ŝsrda = = = √ . (25)
4π 2
G ε Pm Pr 4π 2
η ε Pm Pr 2π · SN R

With (25) the noise-equivalent displacement amplitude is obtained in the order

of a few femtometers per square root Hertz. To reach this resolution possible with
the shot-noise-limited detector signal, it is required an effective bit resolution for
a certain SNR in respect to Eqs. (21) and (22). Assuming 100 μW measurement
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 121

light and 1 mW reference light, a bandwidth of 1 MHz, a wave length of 1550 nm,
a quantum efficiency of 0.8, and a heterodyning efficiency of 0.9, it is obtained a
SNR of the analog photodetector signal of approximately 90 dB which corresponds
to the SNR of approximately 14√effective bits. The resulting resolution would
be approximately 6 pm or 6 fm/ Hz for these numbers. √ Modern commercially
available decoders achieve noise levels below 20 fm/ Hz. In this example, a shot-
noise limited demodulation would require a minimum of 14 effective bits if the
sampling rate is just high enough to fulfill the Nyquist-Shannon sampling theorem.
Doubling the sampling rate corresponds to one more effective bit.
From Eq. (25) it can be concluded that low light-power levels have a proportional
dependence of the displacement noise on the inverse of the square root of the
measurement-light power. Consequently, a reduction of the detected measurement-
light power by a factor of 100 increases the noise just by a factor of 10. Thus, even
low detected light-power levels result in reasonable noise levels. For example, just 1
nW measurement light on the detector results in 10 MHz demodulation bandwidth
still in a displacement resolution of approximately 5.5 nm. Therefore, laser Doppler
vibrometry is a very robust technique for the analysis of structural dynamics.

1.5 Critical Aspects in Laser Vibrometry

1.5.1 Backscatter Issues

The amount of light collected by a vibrometer lens is of primary importance
to obtain high-quality measurements. Laser Doppler vibrometers collect the light
backscattered by the vibrating target to produce the interference with a reference
wave. If light scattered back is insufficient (this condition implies low signal
level), demodulation cannot work properly, and measurement results could be
compromised. The signs and symptoms of inadequate light impinging on the
photodetectors are, among others, high noise levels, signal dropouts, and sensor
over-range conditions also for low vibration velocity. Two factors can induce this
phenomenon: a laser-to-target distance close to a low coherence distance of the laser
and the target surface characteristics.
The laser-to-target distance (stand-off distance) must be selected to avoid
periodic minima of the laser coherence. Coherence due to relative interference is
high when the difference of the optical path is an integer multiple of 2 L, where L
represents the laser cavity length. For other optical path length differences, the
coherence reduces. In case of low signal level, a simple repositioning in the range 5–
10 cm can change significantly the signal level. The positions of low coherence can
be easily determined if the laser system warms up.
The target surface roughness deeply influences the amount of light backscattered
to the photodetectors. Mirror-like surfaces make it possible to get the highest amount
of light impinging on the photodetector, with up to the 95% of light back reflected.
However, mirror-like surfaces do need perfect parallelism between the normal-to-
target surface and the incident beam. Such an ideal condition can be achieved in
very few cases (vibration of hard disk drives, electronic devices) because alignment
122 P. Chiariotti et al.

is very critical. In standard applications, the surface tilting induced by the vibration
of the target might be large enough to make the measurement useless even if the
alignment were perfect.
It is clear that a rough surface is better suited for a LDV measurement, even
though the surface roughness needs to be related to the wavelength of laser
(Ra < < λ, typical of very polished surfaces give rise to a mirror-like behavior,
while very rough surfaces Ra > > λ produce better scatter). On real rough surfaces,
light scattering usually deviates from the ideal behavior (ideal diffuser). Indeed, a
narrower scattering appears around the direction of specular reflection.
Surface scattering power could be maximized in several ways:

• Change of surface roughness (e.g., using sand paper)

• Use of white diffusive paints (e.g., spray developers utilized for crack analysis in
non-destructive testing)
• Use of reflecting tapes (e.g., “cat-eyes” reflectors)

Reflecting tapes are made of microspheres (transparent polymers or glass)

deposited on a tape. The microspheres act like microscopic multidirectional reflec-
tors. In such a way they make it possible to increase diffusivity of light in space
and to increase the amount of light backscattered to the LDV, also when the angle
between the incident beam and the normal-to-target surface is large.

1.5.2 Speckle Noise

Speckle [9–13] is the optical effect that can be observed when a coherent light
source, like a laser, interacts with a rough surface. It is due to the constructive
and destructive interference of light wavelets diffused from different portions of
the surface illuminated. Mirror-like surfaces are the only surfaces that would not
exhibit a diffuse pattern.
Speckle exhibits a grainy pattern, apparently random, that is related to the surface
characteristics and moves with it. As an example, when retroreflective tape is
utilized, the resulting speckle pattern appears very bright, concentrated in a reduced
solid angle, and geometrically periodic. The latter characteristic is related to the
regular pattern of the microspheres fixed on the substrate of the tape. On a generic
surface, the speckle pattern appears less regular. Another interesting example is
the speckle pattern generated by a machined aluminum surface, i.e., a surface
with spatially structured roughness. In such surface grooves induce diffusion in the
direction orthogonal to the grooves themselves; indeed, microscopic grooves on the
surface act as a diffraction grating and produce a fan-shaped diffuse pattern, with a
superimposed speckle pattern (Fig. 10).
The front lens of a vibrometer collects part of the light diffused by the surface
and focuses it on the photodiode, wherein interference and photoelectric conversion
take place. Considering that:

• Speckle moves with the surface movement, different speckle reach the photodi-
ode in the time
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 123

Fig. 10 Pictures of the diffuse light pattern from different surfaces

• Each speckle has its own amplitude and phase, which vary randomly from one
speckle to the other
• Such variation is correlated to the spatial distribution of the light scattered

lateral movements of the speckle pattern with respect to the photodiode or evolutions
in time of the speckle pattern induce time dependency in the amplitude and phase
of the Doppler signal; as a consequence, the output signal of the vibrometer become
amplitude and phase modulated, this causing the so-called speckle noise [14]. An
example on how to practically deal with speckle noise can be found in [15].
As discussed previously, whenever the amplitude of the Doppler signal decreases
too much, a so-called dropout takes place. While speckle-noise has received
considerable attention from the scientific community, far less attention has been
given to dropout noise. This is due, in part, to the fact that dropout noise can be
reduced by improving the scattering characteristics of the target surface. However,
dropout noise still remains an issue on varnished surfaces or enameled steel sheets,
since dropout noise becomes pseudo-random in nature.

1.5.3 Measuring in Media and Through Windows

If the vibrating surface lies in a transparent medium having a refractive index
(na ) different from the one characterizing air in standard conditions (n), the laser
wavelength of the beam crossing the medium changes. This is a matter of influence
of the refractive index on the LDV sensitivity. Indeed, sensitivity changes, and the
measured velocity should be corrected taking into account Eqs. (26) and (27).

na · sin θa = n · sin ψ (26)

vmeas na
vreal = · (27)
cos ψ n

Such a condition can be met in several applications, as on objects operating

immersed in liquids (e.g., water) or in air or gasses when the temperature changes
significantly (e.g., combusting fluids [16]).
124 P. Chiariotti et al.

Other effects could occur when changes in the refractive index take place along
the beam rather than on the target object itself. Since vibrometers are sensitive to
changes in the optical path, and the change in the refraction index does change the
optical path, the signal measured by the sensor will reproduce such variations. This
effect is sometimes utilized to characterize air pressure variations induced by fluid-
dynamic or acoustic effects [17].
It might happen that the target vibrating surface is not directly accessible. This is
the case of structures operating in confined domains, where envelops are needed to
limit high temperatures or pressure, chemical reactions, or simply operating fluids.
In these cases [18], an optical access must be built in order to make it possible
for the laser beam to reach the surface to be measured. Optical windows introduce
a constant additional path length; however, they do not have any influence on the
deviation of ϕm as when a medium change takes place.
Even then, optical windows should be managed carefully, since their use does
present some issue:

• Optical windows could distort the wavefronts of the laser; the surface flatness
should be very good, in the order of λ/10 or smaller, to avoid this.
• The LDV beam will be partially reflected at the optical window, with an intensity
that can be larger than the intensity of the light backscattered from the target
surface. If that light is collected by the front lens of the vibrometer, the beat
signal produced by window reflection may be larger than the beat signal produced
from the light scattered from the surface. In such a case, the demodulated signal
represents the window vibration rather than the vibration of the target behind the
window.
• The dimension of the optical window should be large enough to avoid any
reduction in the scattered light that is collected by the LDV.

Some tricks have to be taken into account to correctly measure on targets lying
behind optical windows. First of all, it is necessary to avoid perfect orthogonality
between the laser beam and the optical window surface. Indeed, on the one hand,
the first light reflection at the window reduces the light scattered by the target and
collected by the vibrometer; on the other hand, the light reflected by the window
does not reach the front lens of the vibrometer, thus avoiding disturbances. It is
also good practice to keep the optical window out of the depth-of-field of the
receiving optics. In cases where the window is close to the target, a possible
solution might be to place the vibrometer close to the window and operate with
a short focal length. This setup will limit the amount of reflected light collected by
the LDV.
Windows must be optically neutral, usually made of BK7 fused silica glass,
quartz glass, or sapphire windows. Birefringency induced by residual stresses,
which could modify laser beam polarization, must be carefully avoided.
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 125

1.6 Signal Enhancement Approaches

1.6.1 Tracking Filter

Signal dropouts are generated by dark speckles impinging the receiving aperture of
the LDV objective if the signal power is smaller than the noise in the demodulation
bandwidth. Such dropouts can be rather short depending on the time the laser
beam persists on the particular position. For surfaces with a fast replacement of the
measurement positions, the dropouts can be quite short, even much shorter than the
basis period of the vibration under investigation. In such a case, the information of
the vibration is still completely available in the signal but might be distorted by the
signal dropouts. In such cases it is preferable to filter the signal in order to remove
spikes and noise. Such a filter is called a tracking filter.
Since modern vibrometers employ digital decoding techniques, tracking filters
are realized nowadays in the digital domain. Two tracking filtering approaches can
be distinguished: (1) methods which remove the noise from the demodulated signal
by analyzing the demodulated signal and (2) methods which remove the noise by
evaluating the original detector signal.
A signal dropout reduces in the first place the signal strength on the photodetec-
tor. After demodulation, the lower signal strength results in a higher noise level in
the displacement or velocity signal. A power-controlled amplifier (limiter) amplifies
the signal power to a certain level to provide a carrier signal with constant amplitude
and, in addition, provides a control signal that is related to the signal strength and
which is a voltage signal in a certain bandwidth. This signal is called received signal
strength indication RSSI. Since a signal dropout is generated by a dark speckle , the
signal strength is reduced, and the noise of the demodulated displacement or velocity
signal is increased.
Method (1) detects the increase of the noise in the demodulated signal. Usually a
dropout generates a spike in the velocity signal. Therefore, a threshold of maximum
acceleration can be defined. In such a case the demodulated signal can be replaced
by the last valid velocity value. This would correspond to a sample-hold approach.
As soon as the acceleration is below the threshold and the velocity is within a
reasonable deviation to the hold velocity value, the real-time velocity signal is again
provided at the output. Such a simple approach generates buckles and hops in the
output signal. A better approach is to use the available information and to continue
the signal, for example, by holding also the derivative of the output signal.
Method (2) employs the fact that the signal strength corresponds directly to the
information content and the noise level of the detected signal. Thus, the bandwidth
of the output signal can be controlled with the signal strength. Such an approach
would define a low-pass filter for the demodulated signal with a cutoff frequency
which is monotonically dependent on the signal strength. In the most simple and
common case the cutoff frequency of the output signal would depend linearly on
the signal strength.
126 P. Chiariotti et al.

1.6.2 Diversity Combining

An advanced approach for speckle treatment has been explored at Polytec [19, 20]
which is based on the basic concepts of signal combining. It has been demonstrated
that this technique is suitable to treat effects from rapidly changing speckles. Here,
the light scattered from the measurement spot is analyzed from different apertures
usually with the same numerical aperture as the impinging beam. However, the
optical setup for diversity is realized, the total scattered light power remains constant
for a given surface reflectivity whatever roughness topography the surface may have.
Therefore, the received power of two different channels is uncorrelated or in case of
a directional scatter with only a few speckles rather anticorrelated.
In any event, the probability of receiving dark speckles in all detection channels
at the same time is a much rarer event than receiving a dark speckle in just one
detection channel. Thus, it can be concluded that a weighted sum of the demodulated
signals of each channel


N
scom (t) = ai (RSSI i ) si (t) N
(28)
i= ai (RSSI i )
i=

will result in a substantially improved combined signal. The weighting factors

are strictly increasing functions from the current RSSI value of the ith channel.
Therefore, the stronger the signal, the higher the weight of that channel.
Since the computation is performed in real time, the combined signal is slightly
delayed but retains its full bandwidth. A signal dropout appears at a certain RSSI
value in dependence on the demodulation bandwidth. A reduction of the probability
for a dropout can be reduced by a factor of 30 with only 2 channels [19]. A factor
above 2000 can be achieved with 4 channels [20]. The laser-speckle effect not only
leads to signal dropouts; it also generates phase jumps at dark speckles. Since dark
speckles are suppressed by diversity combining, velocity spikes at laterally moving
objects are reduced, and, thus, phase noise by laser speckles is reduced dramatically.

1.7 Uncertainty and Calibration

Each measurement performed is affected by a certain value of inaccuracy: such

inaccuracy is quantifiable by the uncertainty parameter, defined in the ISO Guide to
the Expression of Uncertainty in Measurement (GUM) [21].
It is a common perception that laser Doppler vibrometers, due to their com-
plexity, are measurement systems able to provide almost “uncertainty-free” data,
in the sense that their accuracy is much higher than needed in the large majority of
applications. However, this is misleading. Indeed, vibrometers are complex systems,
even though they are apparently very user-friendly. Moreover, their tasks can be
challenging, and therefore experienced users are needed. The apparent ease of use
of such systems might induce the general user to underestimate the complexity
of the application, thus leading to incorrect measurements. The complexity of the
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 127

instrument, together with the complexity of the application, causes stakeholders to

have quite high expectations in terms of results and accuracy, which are expected to
be reliable and to comply with given requirements.
Under controlled conditions vibrometers are very accurate measurement instru-
ments. As a result, it can be difficult to assess their metrological performance,
because the definition of procedures and setups that are able to stress such small
uncertainties represents a difficult task [22].
An international standard (ISO 16063-41 [23]) for the calibration of laser
vibrometers was recently published. Such standard is intended to define the
instrumentation and the procedures to be adopted for calibrating laser vibrometers
in the frequency range 0.4 Hz-50 kHz.
Calibration is, as usual, obtained by comparing the vibrometer to be calibrated
with a reference sensor, in this case an interferometer of a national measurements
and standards laboratory.
Specific interferometric systems have been developed in several national metrol-
ogy institutes to obtain outstanding performances in laboratory conditions. Such
primary standards have been established using primary calibration methods speci-
fied in the ISO standards [24].
Specification on each component must be as good as possible to improve the
accuracy (Table 1). The laser must guarantee a known wavelength with a stability
of 10−5 over a period of 2 years minimum within a temperature interval of
(23 ± 5) ◦ C. Usually a helium-neon laser should preferably be utilized. Under
laboratory conditions (i.e., at an atmospheric pressure of 100 kPa, a temperature
of 23 ◦ C, and a relative humidity of 50%), the wavelength of a helium-neon laser
is 0.63281 μm. In order to make it possible that both systems are measuring
the same displacement, an optical arrangement is utilized that leads the two laser
light beams to be parallel and operating to the same point. In general, digital
techniques, according to ISO standards, make it possible an accurate and traceable
demodulation. Such calibration procedures in national metrological institutes are
at the highest level of the traceability chain, and reference vibrations are to be

Table 1 Metrological characteristics of state-of-the-art laser vibrometers

Laser vibrometer Laser vibrometer standard
Specification (state of the art) (minimum requirements)
Total measurement range 0.1 μm/s to >10 m/s 1 μm/s to 0.1 m/s
Frequency range 0 Hz to >10 MHz 0.1 Hz to 20 kHz
Demodulation Analog voltage Digital data
Calibration uncertainty 1% to 2% 0.5%
Amplitude frequency ±1% (20 Hz to 100 kHz) ±0.2% (0.1 Hz to 20 kHz)
response
Phase frequency response <0.1 deg./kHz to Delay time specified
>10 deg./kHz
Harmonic distortions < 1%THD 0.1%THD
SFDR > 86 dB (per meas. Range) 100 dB (total)
128 P. Chiariotti et al.

measured with the smallest measurement uncertainties (mostly between 0.05% and
0.5%). Shakers have to be very accurate, ensuring uniaxial and purely sinusoidal
motion. Disturbing quantities (e.g., harmonics from nonlinear distortion) act on the
measuring instrument (laser vibrometer) and contribute to signal noise.
When a vibrometer is applied in the field in the real word, the complexity of the
instrument can determine sneaky effects if the user do not have, or cannot have, the
full control of the experimental conditions.
Aspects to be taken into account in the design of the experiment (DoE) are:

1. Surface diffusivity and signal level

2. Speckle noise
3. Coherence distance
4. Index of refraction of medium
(a) Air temperature
(b) Other media
5. Optical windows
6. Angle of incidence
7. In-plane motion
8. Tracking filter
9. Phase delays in digital decoding
10. Vibrations of vibrometer base

In some real-life cases, the vibrometer measurement uncertainty is less important

with respect to the possibility of getting reliable result in difficult tasks; in such
conditions measurement uncertainties between 5% and 20% (frequently 10%) are
usually accepted.
Finally, when mode shapes must be observed, uncertainty in the measurement
with scanning LDV is frequently not an issue. In fact, in such a case, the assessment
of the absolute value of the vibration level is not needed, and only spatial distribution
of vibration is of interest. In this case, the main requirement is that the amplitude
and phase information in system remain stable in the short period.

1.8 Laser Safety and Standards

Whenever dealing with lasers, one should pay attention to safety [25]. In general,
laser vibrometers do not pose severe laser safety problems, because they use almost
low power level, but safety risks are still present. The level of risk for laser exposure
is given by the irradiance parameter, i.e., the radiant flux received by a surface
per unit area (W/m2 ). For instance, considering the case of a 1 mW laser beam
(collimated beam diameter 2 mm) impinging on area of the retina of 1.13 10−13 m2 ,
the irradiance is about 16 MW/m2 .
Considering the power levels of typical vibrometers, possible damage by laser
radiation concerns only the eyes. Effects on the eye are wavelength dependent,
because radiation is absorbed by different parts of the eye versus wavelength. Visible
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 129

and near-infrared radiation (400–1400 nm range) is partially absorbed by aqueous

humor and is transmitted through the eye and reaches the retina. Mid-infrared
(1400 nm – 1 mm) and far-ultraviolet (180–315 nm) are absorbed by the cornea.
Near-ultraviolet (315–400 nm) is absorbed by the lens. Far-infrared radiation can
cause corneal burns, while ultraviolet radiation (UV-C and UV-B) can cause dermal
photokeratitis. Corneal burns can be self-repaired through the regenerative process
of the epithelium layer. Damages to the stroma can cause edema and collagen
shrinkage with probable permanent impairment.
The classification of a laser product depends on the safety of the laser involved,
and it is based on the concept of accessible emission limits (AEL – maximum power
or energy that can be emitted in a specified wavelength range and exposure time
that passes through a specified aperture stop at a specified distance). The reference
document determining the classification of laser products is the IEC: 60825–1:2014
[26]. Such document is the revised version of [27] and aims at classifying laser
products that emit radiation in the wavelength range 180 nm to 1 mm according to
their degree of optical radiation hazard. Almost all commercial vibrometers do have
usually an eye-safe laser class (typically 2, 2 M and 3R – the latter in microscope
systems) in respect to the last IEC classification scheme.

2 Instrumentation, Measurement Issues, and Applications

This section of the chapter aims at giving an overview of the different types of laser
Doppler vibrometers that have been developed so far. An insight into the working
principles of each vibrometer configuration, together with references to application
cases and best practice approaches, is also given.

2.1 Single-Point Vibrometers

The main difference between other single-point single-wavelength interferome-

ters and the vibrometers considered here lies in the broad-bandwidth frequency-
demodulations of the interference detector-signal, which was discussed in this
chapter in the section From Interferometers to Vibrometers. As discussed above the
analog photodetector signal is conditioned as two homodyne signals in quadrature
or as single heterodyne carrier signal with, at least, twice the bandwidth. At next
the conditioned detector signal is analog-digital converted and decoded by digital
signal processing. With reference to Fig. 8, it should be clarified that the decoder
can be separated in a second box or it can be integrated in the sensor housing with
the interferometer. In addition to these basic components, the LDV has control and
communication electronics and optical elements to focus the measurement laser-
beam on the specimen.
A classical single point LDV with sensor head and separated controller is the
OFV 500 with OFV 5000 controller from Polytec. This system is shown in Fig. 11a.
An example of a LDV single-point sensor where the decoder is integrated with the
130 P. Chiariotti et al.

Fig. 11 Examples of single-point vibrometers: Polytec OFV 500 with the separated controller
OFV 5000 (a); Compact Industry sensor Polytec IVS 400 (b); IR-vibrometer RSV-150 for highest
resolutions over long working distances (c). (Courtesy Polytec)

interferometer in a compact sensor housing is the IVS 400 (see Fig. 11b), which has
been designed by Polytec for industrial production monitoring. Historically, laser
vibrometers have been equipped with a helium-neon gas laser. Such lasers need to
be stabilized to realize a single-frequency operation. Since such stabilizations are
expensive double longitudinal-mode lasers are typically installed. The interference
between the two longitudinal modes generates variations of the signal strength
in dependence on the working distance if accidentally two modes have close
amplitudes. Usually, manufacturers call the position of possible minimum signal
strengths visibility minima and the position of maximal signal strength visibility
maxima and specify it in the manual. Vibrometers with helium-neon lasers are
usually built with 1 mW measurement-light power and are classified as laser class 2
devices which corresponds to the class of laser pointers.
It can be seen from Eq. (22) that the SNR of the detector signal depends linearly
on the wavelength. The higher the wavelength for the same measurement power,
the higher the shot-noise-limited SNR of the heterodyne carrier signal due to the
higher number of photons for long wavelengths. Therefore, it can be obtained a
more robust behavior against signal-level variations for IR wavelengths. It can be
concluded from Eq. (26) that the resolvable displacement amplitude increases with
the square root of the wavelength. A favorable wavelength is the telecommunication
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 131

wavelength at 1550 nm. Here, erbium fiber lasers can be employed which are
single-frequency, narrow-linewidth light sources that do not show visibility minima.
In addition, laser light at 1550 nm does not penetrate the eye, and, therefore,
sensors with below 10 mW laser power can be classified as laser class 1, which
is absolutely eye safe. The higher possible laser power improves the SNR of an
√
IR-vibrometer even further up to more than 160 dB/ Hz and allows even better
amplitude-vibration resolutions as compared to a vibrometer with a HeNe-laser.
However, these lasers are much more expensive than helium-neon lasers, and,
therefore, this wavelength is currently mainly utilized for special applications. The
better sensitivity makes an IR vibrometer of this type especially well-suited for long-
distance measurements (e.g., measurements on bridges and buildings). An example
is the RSV-150 from Polytec which is shown in Fig. 11c.

2.2 Optical Fiber Vibrometers

Sometimes measurement points are located at positions that are difficult to access.
In this case, small optical probes are required. This demand can be achieved
with optical fibers which are utilized to parcel the light from the interferometer
to the optical probe. The fibers need to be realized as single-mode, polarization-
maintaining fibers in order to prevent phase differences in different modes of the
light field traveling through the fiber. Such phase variations would appear in multi-
mode fibers. The polarization of the measurement light at the photodetector needs to
be aligned in the same direction as the reference light in order to achieve maximum
interference contrast. Perpendicular polarization states do not interfere at all. The
interferometer can be assembled similar as the standard single-point vibrometer,
but the measurement light is coupled into the fiber by a launcher. The launcher
provides four mechanical degrees of freedom (two orthogonal displacements, two
orthogonal angles) and an adjustable focus to align the impinging field distribution
to the required single-mode field of the fiber. In addition, opto-mechanic means
are necessary to align the polarization by either rotating the fiber in the launcher
or to rotate the laser polarization at the launcher by half-wave plate. The probe at
the fiber end has a miniature lens system to image the fiber exit on the specimen
(corresponds to the best focus). Such a probe can be really small, and in some cases
the probe consists of a gradient index lens with a few millimeter diameter. A fiber
vibrometer with a multiplexer and IR-light from OptoMET is shown in Fig. 12.

2.3 Differential Vibrometers

The laser Doppler vibrometer measures the relative movement between sensor and
specimen. Usually, this does not affect the measurement because the vibration
amplitudes at the specimen at the investigated frequencies are much larger as the
parasitic vibrations of the sensor. However, especially at low frequencies it is
sometimes necessary to suppress influences of the sensor motion by differential
132 P. Chiariotti et al.

Fig. 12 Fiber technology

allows flexible adjustment of
the measurement spot and
multiplexing to different
probes with the dual-fiber
vibrometer of OptoMET.
(Courtesy OptoMET)

measurements. To measure the movement of a measurement point on the specimen

relative to another surface point at any object, the reference beam is also coupled
out of the interferometer, and the optical path length of the reference path is
altered by the movement of the reference point. If both beams measure in the same
direction, the alteration of the interference signal depends only on the relative path
length between measurement and reference light. Differential vibrometers are often
realized as optical fiber vibrometers in order to adjust the measurement positions
of the two laser beams independently. An interesting application of differential
vibrometers is reported in [28], wherein the vibrometer is utilized to assess the
kinematics of poppet valves in a combustion engine. Differential measurements
are especially important for measurement tasks on small structures with slow
movements as, for example, MEMS mirrors; differential measurements are often
realized in laser-vibrometer microscopes. Here, fiber vibrometers are also favorable
because they can be more easily coupled into a confocal laser scanning microscope.

2.4 Rotational Vibrometers

Impinging two parallel laser beams with distance d on a rotating shaft, the sum of
the frequency shifts depends only on the rotation speed if the two beams impinged
on different sides of the shaft. This can be understood by Fig. 13.
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 133

Fig. 13 Rotational vibrometer: working principle (a) and example of a commercial device (b)
Courtesy of Polytec

Assuming a rotating cylinder the Doppler shifts fD of the upper and lower beams
can be expressed as in (29) and (30), respectively.

2R · sin α1 2 · h1
fD1 = = . (29)
λ λ

2R · sin α2 2 · h2
fD2 = − =− . (30)
λ λ

Thus the sum of both Doppler shifts result in the rotational speed

(fD1 − fD2 ) λ
= (31)
2·d

and reveals that the combined signal is proportional to the rotation velocity and the
fixed beam distance, but it is independent on the exact positions of the laser beams.
Therefore, a rotational vibrometer as it is shown in Fig. 13 is a flexible rotation speed
meter. The rotational vibrometer is indeed a very powerful instrument for structural
dynamics [29, 30]; however, particular attention should be given to its setup in order
to obtain reliable results.
The measurement head must always be mounted in such a way that the plane
containing the laser beams lies orthogonally to the rotation axis of the measurement
object. If the angle of incidence is different from 90◦ degrees, the measurement
results decrease in relation to the cosine functions.
The relative position of the two measurement beams with respect to the rotation
axis also influences the quality of the results. Optimal results are achieved with
a symmetrical position of the two beams, in particular for thin shafts. Indeed, if
one beam hits the surface too tangentially, the scattering conditions become sig-
nificantly worse. In addition, an asymmetrical alignment reduces also the available
134 P. Chiariotti et al.

measurement range, as the range is fixed for each interferometer whose output is
mixed with opposite signs.
In many practical cases, the rotating body undergoes also bending or translational
motion. In theory, an additional translational motion, overlapping both velocities
on the two beams, does not influence the resulting Doppler frequency. When the
vibrometer is correctly installed and its arms are perfectly balanced, the effect of
the generic additional bending modes will also be null (v1, bend = v2, bend ). On the
contrary, if for some reasons a good balancing cannot be achieved, the rotational
vibrometer becomes sensitive also to bending modes. Possible causes of a not
perfect balancing (v1, bend = v2, bend ) could be the following:

• Optical (e.g., the local reflection conditions are different for the two spots; the
two beams are disposed highly asymmetrically with respect to the rotation axis),
electrical, or hardware (e.g., differences in the two photodetectors) causes.
• Not perfect parallelism between the two measurement beams.
• The structure or the mode has a particular complex shape.

These problems are minimized when the vibrometer is measuring on a rotating

shaft, while they become more important if a torsional modal analysis in steady
conditions is performed. In fact, in this case the object vibrates and rotates on
a steady condition and thus the terms (δf1 − δf2 ) and (v1, bend + v2, bend ) are
comparable. In this situation, because of the contribution of the above factors, the
frequencies of the larger bending modes are recorded in the rotational velocity
signal, as shown for a camshaft analysis in [31].

2.5 In-Plane Vibrometers

In-plane vibrometers (Fig. 14) have two laser beams which are focused on the same
spot of the specimen and generate an interference pattern with fringe distance s in
the intersection volume [32]. The angle between the two beams is φ. Light scattered
from the intersection volume to the angle bisector has a modulation frequency (36)
that is only proportional to the in-plane velocity.

vin-plane 2 · vin-plane · sin φ

νbisec = = (32)
s λ

Such a system is not sensitive to out-of-plane displacements. Coherent applica-

tion is not applicable to such an optical system, and the photodetector receives light
scattered in the direction of the angle bisector.
In-plane vibrometers enable interesting applications in structural dynamics.
However, some best practices should be taken into account. Among others, three
main conditions should be satisfied:
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 135

Fig. 14 Schematic representation of in-plane vibrometer

• The stand-off distance should make it possible to have the cross-point of the two
laser beams lying on the target surface.
• The optical axis of the instrument should be directed perpendicular to the surface
under test.
• The plane containing the two laser beams should also contain the desired velocity
component.

The optical configuration of the vibrometers can induce a lower SNR with
respect to vibrometers measuring out-of-plane vibrations. When measuring on
highly reflective surfaces, better results might be found by tilting the laser head
at a small angle (about 5◦ ) around the axis of the velocity vector to be measured.
This tilting avoids saturation of the photodetector from direct reflections. Once
measurement uncertainty sources are taken into account [33], in-plane vibrometers
can really be exploited in applications where out-of-plane vibrometers cannot be
utilized (see, e.g., [34].).

2.6 Scanning Vibrometers

2.6.1 Step Scan Vibrometers (Single DoF)

The idea of a scanning LDV is based on the possibility to sequentially and
automatically measure vibration at several points defined by the user and was
invented in 1968 by Massey [35]. In this way, the time evolution of the phenomenon
under analysis can be related to the spatial evolution of the same phenomenon. When
performing a vibration test, this means obtaining operational deflection shapes of the
structure.
136 P. Chiariotti et al.

Different solutions could be adopted to perform a scan, as the movement of the

single-point vibrometer through a motorized traversing gear or the orientation of
the laser beam by a couple of oscillating thick glasses. However, the most common
solution involves the use of a pair of oscillating mirrors. The mirrors are usually
moved by galvanometric motors. Typical rotation angles reached ranges between
±20 deg. Motors are driven by a closed loop controller that checks mirror position
using an angular position sensor (usually a capacitive sensor). The system is also
thermo-stabilized, for better accuracy. Apart from galvo-motors, other systems
involving stepper motors or piezoelectric-driven motors can be found as well.
However, the latter solutions are rarely implemented in commercially available
scanning vibrometers.
The rotation axes of the two mirrors are orthogonal to each other. Such a
configuration makes a two-direction scan possible. An example of such system is
reported in Fig. 15. The focusing optics are sometimes installed after the oscillating
mirrors to align laser beam and make it possible to scan with parallel beams (same
incidence angle with respect to the normal to target surface).
In order to reduce the distance ds between the two oscillating mirrors
(ds ≈ 30 mm), and to better pack the whole scanning system, the axes of the
mirrors are placed at an acute angle. In this way, even at smaller distances between
the mirrors axes, the risk of ruining the mirrors at typical scanning angles reduces.
Moreover, the instrument head is quite compact, and wide measurement grids can be

P(xs, ys, z0)

y

z0
z x

ds θSy Arbitrarily
Shaped Target

θSx

Fig. 15 Schematic representation of a double mirror scanning configuration

3 Laser Doppler Vibrometry Measurements in Structural Dynamics 137

scanned once a proper laser to target distance (z0 ) is set up. The main disadvantage
of the double mirror configuration (without the focusing optics after the mirrors)
has to be identified in the measuring of a vibration velocity that is a superposition
of in-plane and out-of-plane components. Indeed, since the beam direction changes
during the scan, different components are measured on different points. In case of
pure out-of-plane motions, and when P(xs , ys , z0 ) = P(x0 , y0 , z0 ) for ϑsx , ϑsy being
in their “home” position (Fig. 15), the user can compensate for the angle incidence
error. When this is not the case, the vibration velocity measured is sensitive to
out-of-plane and in-plane components in a different way whenever the laser spot
moves to a different point. All these aspects can be well understood referring to the
modelling framework developed by Halkon and Rothberg [36].
A typical measurement procedure, involving a computer-controlled scanning
systems to support the user, comprises the following steps:

• Focus the laser beam on the target surface

• Acquire the image of the object, for reporting the position of the grid on the
object
• Align the camera coordinate with the grid coordinate
• Define the measurement grid
• Manage (if required) the excitation
• Manage the acquisition parameters and strategy
• Move the laser spot on each point of the grid (settling time of about <10 ms), for
acquisition
• Acquire, for each measurement point, the vibrometer and the reference (or
references) signals
• Post-process data and display results
• Store data

A mandatory requirement for performing a vibration test using step scanning

vibrometry is the stationarity of the phenomenon under analysis. If this is not
the case, repeatability of the phenomenon, under an adequate trigger condition,
should be guaranteed at least. Whenever all these requirements are not met,
scanning vibrometry cannot be utilized to assess the global vibration behavior of
a structure, since vibration information on different points cannot be related to each
other.
It is also good practice to have also a reference signal when performing a scan.
Such reference signal, i.e., a signal given by a sensor measuring on the same
position, plays the same role that a reference sensor (e.g., a load cell) plays in
traditional modal analysis tests, that is:

• To verify that phenomena under analysis are stationary

• To obtain a phase reference among each point
• To calculate correlation functions (e.g., frequency response functions), if
needed
138 P. Chiariotti et al.

2.6.2 Special Solutions

Continuous-Scan LDV
Continuous-scan laser Doppler vibrometry (CSLDV) can greatly accelerate modal
testing by continuously sweeping the measuring laser over a structure, effectively
capturing its response at hundreds of points along the laser path in one measurement.
Scanning the laser can increase laser speckle noise, so this application seeks to find
a balance between the additional spatial information that CSLDV provides (in a
reduced time) with increasing speckle noise. CSLDV typically increases speckle
noise, but it has been shown to decrease the noise in some cases [37]. A schematic
of the CSLDV process is shown in Fig. 16, as well as a photograph from an actual
CSLDV test from [38].
Continuous-scan vibrometry also changes the nature of the signals that are
measured, so new methods are needed to process CSLDV measurements and extract
the structure’s modal natural frequencies, damping ratios and the mode shapes
along the laser scan path. Among the first and probably, the most prolific is the
method developed by Ewins, Stanbridge, Martarelli, and others, who modelled the
operating deflection shape as a continuous polynomial function of the laser position.
The recorded vibration shape using CSLDV was treated as being modulated by the
moving laser position, which leads to sideband harmonics in the spectrum that are
separated by the scan frequency. They showed that the amplitudes at those har-
monics are related to the polynomial coefficients by a transformation matrix. Once
the polynomial coefficients are extracted, the mode shapes can be reconstructed
using the known laser path. This method is very effective and straightforward and
has been used to reconstruct the mode shapes of a structure using sinusoidal [39],

Fig. 16 Schematic representation of continuous-scan LDV using a double mirror configuration.

(a) The laser sweeps continuously over the surface as the measurement is acquired. (b) Photograph
from an actual CSLDV test on a wind turbine blade with the laser path artificially colored
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 139

Fig. 17 Spectrum measured using CSLDV on a cantilever beam that is vibrating at 20 Hz. The
laser is scanning at 3 Hz and so the measurement shows sideband harmonics at 20 ± 3 Hz. The
side band amplitudes can be collected to estimate the mode shape using the method in [38]

impact [40], and pseudo-random excitation [41]. An application of the approach

for the assessment of timing-belt dynamics in running condition is reported in
[42]. If the operational shape along the laser path is complicated, then high-order
polynomial coefficients may be necessary to accurately describe the shape. Hence,
the precision of CSLDV is limited by the number of harmonics that stands out above
the noise floor in the measured spectrum. An example is shown in Fig. 17.
The CSLDV methods mentioned so far are all tailored to a certain class of
forcing input (e.g., sinusoidal has been most common) and require specialized
post-processing to extract the natural frequencies and mode shapes from the
measurements. Allen et al. proposed several alternatives that more closely resemble
conventional modal analysis. In [43] they proposed the discrete “lifting” approach,
in which the responses at the same location along the laser path are grouped together.
The reorganized responses then appear to be from a set of pseudo transducers
distributed along the scan path, and then standard routines can be used to curve
fit the measurements and extract the mode shapes. However, each pseudo-sensor
samples only once per scan period, and there is a constant time delay between the
measurements at each point. As a result, according to the sampling theorem, all the
modes of the system and their harmonics are aliased to the range from zero to half of
the scan frequency. In [44] Allen et al. used the lifting approach to extract the natural
frequencies and mass-normalized mode shapes of a free-free beam under impact
excitation. They demonstrated that, ideally, one would choose the scan frequency
140 P. Chiariotti et al.

to be larger than twice the maximum frequency of interest, but often this is not
possible. The maximum scan frequency is restricted by the mechanical scanners
(up to 500 Hz), and the laser speckle noise also increases with the scan frequency.
Instead, the scan frequency can be chosen to ensure that the modes do not overlap
after being aliased to the band of half of the scan frequency. Therefore, the lifting
approach is more suitable for measuring structures whose natural frequencies are
only a few hundreds of Hertz or lower.
The most general approach for treating CSLDV measurements was presented in
[37] and is based on the harmonic transfer function (HTF) concept developed by
Wereley and Hall [45]. The HTF extends the transfer function concept to systems
where the system parameters are periodic functions of time. In the case of CSLDV,
the measurement point is a periodic function of time. This concept can also be used
to address output only modal analysis, where the input is not measured but can
be assumed to be broadband and random. For example, Yang and Allen used this
approach in [38] to process the CSLDV measurements of a parked wind turbine
that was excited by a light wind, as shown in Fig. 18. Detailed mode shapes were
obtained from one 20-minute time record using a single laser; to obtain similar
resolution with conventional stepped-scanning LDV, one would have needed two
lasers and several hours to acquire the measurements.
Alternative use of CSLDV can be cited for vibro-acoustic applications [46],
damage detection [47], and biomedical applications [48]. A multi-beam CSLDV
approach was also proposed in [49] for landmine detection.

LDV Strategies on Rotating/Moving Structures

Scanning Laser Doppler Vibrometry with Optical Derotator
There are mirror or prism systems with an odd number of reflections which rotate the
image when the optical system is rotated [50, 51]. For example, watching through a
rotating Dove prism, one would see the image rotating with twice the rotation speed
of the prism. A laser beam moves on a circle if the beams shine through a rotating
Dove prism if the optical axis of the prism is absolutely coaxial to the rotation axis.
To receive a standing image of a rotating object, the prism needs to rotate with
exactly half the rotational speed of the specimen. In addition, the optical axis of the
prism needs to intersect the rotational axis of the specimen on the specimen surface.
Therefore, an optical derotator is well suited to measure on a rotating specimen with
a scanning laser Doppler vibrometer if suitable degrees of freedom for alignment are
available.
It has been demonstrated that rotations up to 12.000 revolutions per minute
(rpm) of the prism can be achieved with standard electronic motors and motor
controllers. It is necessary to realize a phase-locked loop between an encoder signal
of the prism rotation and an encoder signal from the rotating object in order to
synchronize both rotations. Thus, a controller is required capable implementing a
digital phase control loop. Usually the control applies first a frequency control to
synchronize the rotational speed of the prism to half the rotational speed of the
object. Then the control parameters are adapted automatically in order to achieve
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 141

Fig. 18 Spectrum measured using CSLDV on free-free beam excited by an impulsive force. The
lifting approach was used to expand the measurement to hundreds of pseudo-measurement points,
and the average spectrum shown (a) reveals that each mode occurs at a single frequency, although
aliased within the 25.5 Hz bandwidth. The mode shapes (b) can be extracted by curve fitting this
set of measurements using standard LTI modal analysis algorithms, from [41]

a steady phase relation for a steady image of the rotating parts. Such a system can
follow rotational accelerations up to 700 rpm/s.

Tracking Laser Doppler Vibrometry

Tracking laser Doppler vibrometry (TLDV) is basically conceived as an enhanced
SLDV; it is aimed at enhancing common scanning LDV. The aim of TLDV is non-
contact vibration testing on a point, or over a grid of points, fixed on a moving
142 P. Chiariotti et al.

target (i.e., Lagrangian approach) [52]. The system is based on a SLDV where the
mirrors, conventionally utilized to scan the beam across a grid of static points, are
now controlled to track the motion of an arbitrary moving structure, its trajectory
being known.
In the case of rotating machinery, the trajectory to be followed can be measured
using a position sensor, i.e., which can track the changes of the target angular
position. For instance, in case a digital encoder is connected to the rotating shaft,
the encoder signal provides the angular position and becomes the clock (i.e., it acts
as the clock of the D/A converters of the acquisition system DAQ board which
generates analog voltages to directly drive the LDV mirrors) to drive the mirrors so
to track the position of a certain point during the target revolution.
In case the target trajectory is not a priori known, a closed loop approach can be
utilized [53]. A camera, coaxial with the laser beam, observes the target (a bright
spot on a dark background) position with respect to the laser beam and provides an
“error” signal to a PID controller. The output of the controller feeds the scanning
mirror, making possible a constant tracking of the target.
These architectures allow one to track each visible point on the moving structure;
a complete grid of points, moving with the target, can therefore be explored, by
taking sequential measurement on the different rotating points which are part of the
grid [54] under steady conditions.
The tracking strategy has been applied in several kinds of rotating machinery, as
helicopter rotor blades [55] and naval propeller blades [56].
A further evolution of the tracking concept is to be found in the use of continuous-
scan LDV for assessing the vibration of a structure undergoing rigid body motion.
Such concept, addressed in literature as Tracking CSLDV (TCSLDV), consists in
performing a continuous scan over the target surface while tracking the target rigid
body motion. A deeper insight into the method, given by application cases, can be
found in [57–60].

2.6.3 3D Scanning Vibrometer

One of the most important applications of laser Doppler vibrometry is discrete 3D
scanning [61]. Here, three laser beams impinge a measurement spot from three
different linearly independent directions. Thus, such a system can obtain three
linearly independent velocity components from a measurement spot. For every
scanning vibrometer, the Doppler shift can be expressed by the vector formula

1 − −
→
δfi = ·→
v · ki (33)
π

−
→ −
→→ −
→→
with the wave vector k i of the three components kxi = k i − e x , kyi = k i −
e y,
−
→− →
and kzi = k i e z in a Cartesian coordinate system. Thus, the velocity vector can be
derived by the Equation
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 143

Fig. 19 3D scanning vibrometer

⎛ ⎞−1 ⎛ ⎞
kx1 ky1 kz1 δf1
−
→ 1
v = ⎝ kx2 ky2 kz2 ⎠ · ⎝ δf2 ⎠ . (34)
π
kx3 ky3 kz3 δf3

Thus, a 3D scanning vibrometer as it is shown in Fig. 19 makes possible the

detection of broadband 3D vibration spectra at every measurement spot.
Complete 3D operational deflection shapes can be obtained by scanning the three
beams over a surface and by receiving a reference signal of a periodic vibration to
synchronize the measurements.
Geometry data of the measurement spots can be either imported in the user
software or it can be measured with a time-of-flight measurement with a laser
distance sensor integrated in the PSV-500.
A recently published solution [62] allows 3D measurements with just a single
measurement beam by collecting light from three directions and evaluating the
directional Doppler effect. This solution is especially well suited for small structures
since the laser spot is just defined by the central impinging beam which can
be focused with a microscope objective. A spot diameter of 3.3 μm has been
demonstrated for 3D measurements.

2.7 Multi-beam

Scanning measurements are not possible if either a reference signal is not available
or if the vibration is aperiodic or not repeatable. This is the case for events like
operational vibrations during heating which is the case for any engine during startup
or run-up. Other example applications are impacts, processes with strong friction
(e.g., brakes), breaking structures, or explosions. Since the number of channels
144 P. Chiariotti et al.

Fig. 20 12-channel, laser Doppler vibrometer realized as demonstrator in the corporation project
Holovib funded by the BMBF under the grant number 13 N9338: (a) optical setup (a); (b) example
of a measurement on an impact wrench (Courtesy Polytec)

is limited due to the high cost of a vibrometer, the flexible adjustment of the
measurement is essential. Such a system, which is capable of adjusting an arbitrary
measurement grid, has been demonstrated by Haist et al. [63]. The multichannel
vibrometer has been realized as complex bulk-optics setup with holograms to split
the laser light to multiple channels with homodyne [64] or heterodyne [65]. Fibers
are employed to parcel the measurement light to flexible probes with adjustable
objectives [45, 66, 67] (Fig. 20).

3 Conclusions

When it comes the necessity of avoiding contact sensors for measuring vibration
response of structures (e.g., on lightweight, hot structures, etc.), laser Doppler
vibrometry surely represents a target technique. Indeed, LDV has gained a relevant
role within structural dynamics testing and is currently widely used in several
application fields. However, as it happens with every measurement technique, LDV
must be properly known to be properly exploited. The aim of this chapter was
to provide the reader with the fundamentals of LDV (from interferometry basics,
to laser safety) as well as with the main instrumentations, measurement issues,
and applications involving this technique. The chapter was not intended to give a
comprehensive picture of LDV, for which the interested reader is encouraged to
go for the references reported, but can be considered to be a sufficiently detailed
document enabling the experimenter to learn the potentials of this measurement
approach and some “best practice” gathered from the experience gained by the
authors during their careers.
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 145

References
1. Drain LE (1980) The laser Doppler technique. John Wiley & Sons Ltd, Chichester
2. Rembe C (2008) Retardation effects in Laser-Doppler Measurements. In: 8th International
conference on vibration measurements by laser techniques, advances and applications, SPIE,
vol 7098, p 13
3. Rembe C, Siegmund G, Steger H, Wörtge M (2006) Measuring MEMS in motion by laser-
Doppler vibrometry. In: Osten W (ed) Optical inspection of microsystems, pp 245–292
4. Bauer M, Ritter F, Siegmund G (2002) High-precision laser vibrometers based on digital
Doppler-signal processing. In: 5th International conference on vibration measurements by laser
techniques, advances and applications, vol 4827. SPIE, pp 50–61
5. Rudd MJ (1968) A laser Doppler velocimeter employing the laser as a mixer-oscillator. J Phys
E 1:723–726
6. Giuliani G, Norgia M, Donati S, Bosch T (2002) Laser diode self-mixing technique for sensing
applications. J Opt A Pure Appl Opt 4:S283–S294
7. Scalise L, Yu Y, Giuliani G, Plantier G, Bosch T (2004) Self-mixing laser diode velocime-
try: application to vibration and velocity measurement. IEEE Trans Instrum Meas 53(1):
223–232
8. Scalise L, Steenbergen W, de Mul F (2001) Self-mixing feedback in a laser diode for intra-
arterial optical blood velocimetry. Appl Opt 40:4608–4615
9. Dainty JC (1975) Laser speckle and related phenomena. Springer, Berlin
10. Strean RF, Mitchell LD, Barker AJ (1998) Global noise characteristics of a laser Doppler
vibrometer – I. theory. Opt Lasers Eng 30(2):127–139
11. Rothberg SJ, Baker JR, Halliwell NA (1998) Laser vibrometry: pseudo-vibrations. J Sound Vib
135(3):516–522
12. Rothberg S (2006) Numerical simulation of speckle noise in laser Vibrometry. Appl Opt
45(19):4523–4533
13. Martin P, Rothberg S (2009) Introducing speckle noise maps for laser vibrometry. J Optics
Lasers in Engineering 47:431–442
14. Dräbenstedt A. (2007) Quantification of displacement and velocity noise in vibrometer
measurements on transversely moving or rotating surfaces. In: Proceedings of SPIE 6616,
Optical measurement systems for industrial inspection, vol 6616 (32). https://doi.org/10.1117/
12.726115
15. Vass J, Smìd R, Randall RB, Sovka P, Cristalli C, Torcianti B (2008) Avoidance of speck-
le noise in laser vibrometry by the use of kurtosis ratio: application to mechanical fault
diagnostics. Mech Syst Signal Process 22:647–671
16. Paone N, Revel GM (1998) Modelling and experimental analysis of the performance of a laser
Doppler vibrometer used to measure surface vibrations through combustive flows. Opt Lasers
Eng 30:163–178
17. Castellini P, Martarelli M, Tomasini EP (2013) Sub-sonic jet pressure fluctuation characteriza-
tion by tomographic laser interferometry. Exp Fluids 54:12
18. Buligan G, Paone N, Revel GM, Tomasini EP (2003) The measurement of hermetic com-
pressor valve motion by laser vibrometry, IMAC-2003, Kissimmee
19. Dräbenstedt A (2014) Diversity combining in laser Doppler vibrometry for improved signal
reliability. AIP Conference, vol 1600, p 263
20. Rembe C, Dräbenstedt A (2015) Speckle-insensitive laser-Doppler vibrometry with adaptive
optics and signal diversity. Sensor 2015, Erlangen, Deutschland, 19–21 May
21. Evaluation of measurement data – Guide to the expression of uncertainty in measurement, First
edition September 2008 © JCGM 2008
22. Robinson DC, Serbyn MR, Payne BF (1987) A description of NBS calibration services in
mechanical vibration and shock. NBS Technical Note:1232
23. ISO 16063-41:2011, Methods for the calibration of vibration and shock transducers – Part 41:
Calibration of laser vibrometers
146 P. Chiariotti et al.

24. ISO 16063-15:2006, Methods for the calibration of vibration and shock transducers – Part 15:
primary angular vibration calibration by laser interferometry. International Organization for
Standardization (ISO), Geneva
25. Practical Laser Safety, 2nd edn. CRC Press. ISBN: 0824782402
26. IEC 60825-1:2014 – Safety of laser products Part 1: equipment classification and requirements
27. IEC 60825-1:2007 – Safety of laser products Part 1: equipment classification and re-quirements
28. Paone N, Santolini C, Tomasini EP (1994) Application of a laser Doppler vibrometer to
evaluate engine poppet valve kinematics. In: XII International. Modal analysis conference,
Hawaii
29. Hallowell NA (1996) The laser torsional vibrometer: a step forward in rotating machinery
diagnostics. J Sound Vib 190(3):399–418
30. Lewin AC, Roth V, Siegmund G (1995) Measurement of rotational vibrations using a novel
interferometric technique. Measurement 16(2):81–90
31. Revel GM, Tomasini EP (1999) Torsional Vibrations: a laser vibrometry approach. In:
Vibration, noise & structural dynamics ‘99, Venezia, pp 448–453
32. van Netten SM (1988) Laser interferometer microscope for the measurement of nanometer
vibrational displacements of a light-scattering microscopic object. J Acoust Soc Am 83:
1667–1674
33. Gasparetti M, Revel GM (1999) The influence of operating conditions on the accuracy of in-
plane laser Doppler velocimetry measurements, measurement, vol 26. Elsevier Science Ltd.,
Oxford, pp 207–220. ISSN 0263-2241
34. Castellini P, Chiariotti P, Martarelli M, Tomasini EP (2013) Valvetrain motion measure-ments
in firing conditions by laser Doppler vibrometer. In: IMAC XXXI International modal analysis
conference, Hyatt Regency Orange County, Garden Grove
35. Massey GA (1968) An optical heterodyne ultrasonic image converter. IEEE Sonics Ultrasonics
56:2157–2161
36. Halkon BJ, Rothberg SJ (2003) Vibration measurement using continuous scanning laser
Doppler vibrometry: theoretical velocity sensitivity analysis with applications. Meas Sci
Technol 14:382–393
37. Yang S, Allen MS (2014) Harmonic transfer function to measure translational and rotational
velocities with continuous-scan laser doppler vibrometry. ASME J Vib Acoust 136(2)
38. Yang S, Allen MS (2012) Output-only modal analysis using continuous-scan laser Doppler
vibrometry and application to a 20 kW wind turbine. Mech Syst Signal Process 31:228–245
39. Stanbridge AB, Ewins DJ (1999) Modal testing using a scanning laser Doppler vibrometer.
Mech Syst Signal Process 13(2):255–270
40. Ribichini, R et al. (2008) Impact testing with a continuously-scanning LDV. In: 26th Interna-
tional modal analysis conference (IMAC XXVI), Orlando
41. Di Maio D, Carloni G, Ewins DJ (2010) Simulation and validation of ODS measurements
made using a continuous SLDV method on a beam excited by a pseudo random signal. In: 28th
International modal analysis conference (IMAC XXVIII), Jacksonville
42. Chiariotti P, Martarelli M, Castellini P (2017) Exploiting continuous scanning laser Doppler
vibrometry in timing belt dynamic characterization, 86, Part B, pp 66–81
43. Allen MS, Sracic MW (2010) A new method for processing impact excited continuous-scan
laser Doppler vibrometer measurements. Mech Syst Signal Process 24:721–735
44. Yang S, Sracic MW, Allen MS (2012) Two algorithms for mass normalizing mode shapes from
impact excited continuous-scan laser Doppler vibrometry. J Vib Acoust 134(2):021004
45. Wereley NM (1991) Analysis and control of linear periodically time varying systems. Depart-
ment of Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge
46. Chiariotti P, Martarelli M, Revel GM (2014) Exploiting continuous scanning laser Doppler
vibrometry (CSLDV) in time domain correlation methods for noise source identification. Meas
Sci Technol 25(7):075204 (13pp). https://doi.org/10.1088/0957-0233/25/7/075204
47. Chiariotti P, Martarelli M, Revel GM (2017) Delamination detection by multi-level wavelet
processing of continuous scanning laser Doppler vibrometry data. Opt Lasers Eng 99:66–79
3 Laser Doppler Vibrometry Measurements in Structural Dynamics 147

48. Salman M, Sabra KG (2011) Synchronized vibrations measurements at multiple locations

using a single continuously scanning laser Doppler vibrometer. Applications to non-contact
sensing of human body vibrations. J Acoust Soc Am 130(4):2394–2394
49. Aranchuk V, Lal AK, Zhang H, Hess CF, Sabatier JM (2004) Acoustic sensor for landmine
detection using a continuously scanning multi-beam LDV, vol 5415. SPIE, Orlando, p 61
50. Boedecker S, Dräbenstedt A, Heller L, Kraft A, Leonhardt A, Pape C, Ristau S, Reithmei-er E,
Rembe C (2006) Optical derotator for scanning vibrometer measurements on rotating objects,
SPIE vol 6345, 63450M
51. Johansmann M, Fritzsche M, Schell J (2011) A new method for measurement of rotating
objects utilizing laser Doppler vibrometry combined with an optical derotator with focus on
automotive applications. SAE Technical Paper 2011-26-0043
52. Castellini P, Paone N (2000) Development of the tracking laser vibrometer: performance and
uncertainty analysis. Rev Sci Instrum 71(12):4639
53. Castellini P, Tomasini EP (2004) Image-based tracking laser Doppler vibrometer. Rev Sci
Instrum 75(1):222
54. Castellini P, Montanini R (2002) Automotive components vibration measurements by
tracking laser Doppler vibrometry: advances in signal processing. Meas Sci Technol 13:
1266–1279
55. Fioretti A, Di Maio D, Ewins DJ, Castellini P, Tomasini EP (2010) Deflection shape re-
contructions of a rotating five-blade helicopter rotor from SLDV measurements. In: 9th
International conference on vibration measurements by laser and noncontact techniques &
short course, Ancona
56. Castellini P, Santolini C (1998) Vibration measurements on blades of naval propeller rotating
in water with tracking laser vibrometer. Measurement 24:43–54
57. Martarelli M, Castellini P, Santolini C, Tomasini EP (2011) Laser Doppler vibrometry on
rotating structures in coast-down: resonance frequencies and operational deflection shape
characterization. Meas Sci Technol 22:115106
58. Martarelli M, Castellini P (2012) Performance analysis of continuous tracking laser Doppler
vibrometry applied to rotating structures in coast-down. Meas Sci Technol 23:065202. https://
doi.org/10.1088/0957-0233/23/6/065202
59. Di Maio D, Ewins DJ (2010) Applications of continuous tracking SLDV measurement methods
to axially symmetric rotating structures using different excitation methods. Mech Syst Signal
Process 24(8):3013–3036
60. Gasparoni A, Allen MS, Yang MS, Sracic MW, Castellini P, Tomasini EP (2010) Experimental
modal analysis on a rotating fan using tracking-CSLDV. In: 9th International conference on
vibration measurements by laser and noncontact techniques & short course, Ancona, ISBN
978–0–7354-0801-2
61. Bendel K, Fischer M, Schüssler M (2004) Vibrational analysis of power tools using a
novel three dimensional scanning vibrometer. In: Sixth International conference of vibration
measurement by laser techniques, vol 5503. SPIE, pp 177–184
62. Rembe C, Kowarsch R, Ochs W, Dräbenstedt A, Giesen M, Winter M (2014) Optical 3D-
vibrometer microscope with picometer-resolution in x, y, and z. Opt Eng 53(3):034108
63. Haist T, Lingel C, Osten W, Winter M, Giesen M, Ritter F, Sandfort K, Rembe C, Bendel K
(2013) Multipoint vibrometry with dynamic and static holograms. Rev Sci Instrum 84:121701
64. Cupido E, Morel S, Smith D (2003) Multipoint laser Doppler vibrometer for transient analysis.
In: IMAC-XXI: conference & exposition on structural dynamics, SEM
65. Lal AK, Hess C (2006) Array and matrix based laser Doppler vibrometers for measuring noise
and vibration. Inter-Noise, Honolulu
66. Fu Y, Guo M, Phua PB (2011) Multipoint laser Doppler vibrometry with single detector:
principles, implementations, and signal analyses. Appl Opt 50(10):1280–1288
67. Waz AT, Kaczmarek PR, Abranski KM (2009) Laser-fibre vibrometry at 1550 nm. Meas Sci
Technol 20:105301
Applied Digital Signal Processing
4
Robert B. Randall, Jerome Antoni, and Pietro Borghesani

Contents
1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
2 Deterministic Signals and Traditional Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 152
2.1 Periodic Signals and Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
2.2 Transient Signals and the Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
2.3 Discrete-Time Signals and Digital Implementation . . . . . . . . . . . . . . . . . . . . . . . . 153
3 Experimental Signals and Stochastic Signal Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 155
3.1 Time-Varying Distributions: Ensemble Versus Time Quantities . . . . . . . . . . . . . . 155
3.2 Stationary Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.3 Cyclostationary Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
3.4 Cyclo-non-stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.5 Transient Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4 Scaling and Dimensions for Various Versions of the Fourier Transform . . . . . . . . . . . . . 164
4.1 Periodic and Quasi-periodic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.2 Stationary Random Signals, Power Spectral Density (PSD) . . . . . . . . . . . . . . . . . 165
4.3 Deterministic Transient Signals, Energy Spectral Density (ESD) . . . . . . . . . . . . . 166
5 Choosing the Right Model for an Experimental Signal . . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.1 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.2 Condition Monitoring for Rotating/Reciprocating Machines . . . . . . . . . . . . . . . . 167

R. B. Randall ()
University of New South Wales, Sydney, NSW, Australia
e-mail: [email protected]
J. Antoni
Univ Lyon, INSA-Lyon, Villeurbanne, France
e-mail: [email protected]
P. Borghesani
University of New South Wales, Sydney, NSW, Australia
Queensland University of Technology, Brisbane, QLD, Australia
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 149

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_6
150 R. B. Randall et al.

6 Signal Extraction and Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

6.1 General Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.2 Signal Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.3 Reference-Based Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.4 Filtration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.5 Blind Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.6 Blind Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.7 Discrete-Random Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.8 Blind Source Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.9 Blind Separation of Structural Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7 Time-Frequency Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.1 STFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.2 Wigner-Ville Distribution (WVD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
7.3 Wigner-Ville Spectrum (WVS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.4 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
8 Specialized Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.1 Order Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.2 Second-Order Cyclostationary Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.3 The Envelope Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.4 The Cyclic Modulation Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.5 The Spectral Correlation Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.6 The Relationship Between Kurtosis, Envelope, and CS2 Indicators . . . . . . . . . . . 200
9 Typical Diagnostic Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.1 Rolling Element Bearings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.2 Gears . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.3 Reciprocating Machines and Engines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
10 Cepstrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
10.1 Background and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
10.2 Cepstrum Liftering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.3 Cepstrum for Machine Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
10.4 Cepstrum for Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

Abstract

This chapter contains a broad discussion of digital signal processing techniques

as applied to the solution of mechanical problems, primarily by analyzing the
vibration responses of a machine or structure. Such responses are always a
combination of a set of excitation functions and structural response or transfer
functions, and the aim of the analyst is usually to separate them and learn their
characteristics, for purposes such as structural analysis, primarily concerned
with changes in the latter, and machine condition monitoring and diagnostics,
primarily concerned with changes in the former, but possibly in both. Since
a number of other chapters are mainly concerned with structural analysis, the
reader is referred to those for some specialized treatments.
The chapter first introduces a number of idealized signal types, including their
definitions and basic analysis methods, then gives a guide to the optimum choice
of such models to apply in practical situations, such as for modal analysis and
condition monitoring.
4 Applied Digital Signal Processing 151

A very important section deals with the two types of blind separation required
for complete analysis; first the separation of the various independent sources
acting on the machine or structure, and then the identification of the different
transfer functions by which the responses to these different sources are modified
at the various measurement points. Topics include separation by filtering, blind
extraction, blind deconvolution, and separation of responses to different sources,
including those distinguished by different characteristics (e.g. deterministic or
random), or by virtue of statistical independence.
There is a comprehensive discussion of analysis in different domains, or
domain pairs, such as time, frequency, and joint time-frequency, but also the
recognition that with variable speed machines, it is often best to represent “time”
as rotation angle, with corresponding “frequency” in terms of harmonic order.
A topic that has become much more important in recent years is the recognition
that many machine signals are stochastic, but with random carrier signals that
are modulated by deterministic modulation functions, usually linked to machine
speed, which can be extracted and identified, even though seemingly hidden in
normal signal representations. With constant speed machines, such signals are
“cyclostationary”, but with varying speed are termed “cyclo-non-stationary”.
Most of these approaches are demonstrated by applying them to three quite
different, but very important examples of machine diagnostics, namely for rolling
element bearings, gears and reciprocating machines and engines.
Finally, a topic which is not widely known, cepstrum analysis, is presented
in some detail, because of its very powerful properties in both source separation
and structural analysis, with examples of application to machine diagnostics and
modal analysis.

Keywords

Signal processing of mechanical signals · Vibration-based health monitoring ·

SHM · Denoising · Change detection · Change detection · Blind extraction ·
Blind separation · Cepstrum · Cepstral analysis · Cyclostationary signals ·
Angle-time cyclostationary signals · Cyclo-non-stationary signals

1 Overview

Rotating machine condition monitoring and traditional modal analysis have been
among the earliest applications of vibration signal analysis. The traditional analyti-
cal modelling of such systems and related excitations originally led researchers and
engineers to focus mainly on deterministic signals: periodic in the case of rotating
machines and transient in the case of impact-test-based modal analysis (although
experimental modal analysis has incorporated stationary random signals for many
years). Despite being obviously in contrast with the randomness of all experimental
signals, the deterministic framework has been successful for decades in these fields.
However, the further extension of the field of application of signal processing
(e.g., operational modal analysis and diagnostics of complex systems) emphasized
152 R. B. Randall et al.

the limits of the deterministic framework and the necessity for stochastic (in
particular non-stationary stochastic) signal models and analysis techniques.
This chapter starts with a review of periodic and deterministic transient signals
and the basic Fourier analysis tools to describe them. Then the most successful
stochastic models for different families of signals are introduced, and finally
advanced techniques for their analysis are discussed with examples.

2 Deterministic Signals and Traditional Fourier Analysis

Traditional vibration analysis models vibration signals as deterministic, i.e., math-

ematical functions whose value is determined and exactly predictable in the future.
Two main classes of signals belong to the deterministic family:

• Periodic signals, which replicate the same waveform after a fixed period T:

xT (t) = xT (t + nT ) with n = 1, 2, . . . (1)

• Transient signals, whose energy is finite, i.e., with an event-like behaviour and
decaying to zero in a sufficiently long time:

 ∞
|x(t)|2 dt = a finite (2)
−∞

The first have dominated the field of rotating machine condition monitoring,
where most excitations are modelled as periodic and often sinusoidal (e.g., unbal-
ance) and systems are generally approximated as linear time invariant (LTI). The
second have been at the basis of the traditional impact-test approach for modal
analysis, where an LTI system is excited by an impulsive force.

2.1 Periodic Signals and Fourier Series

In the analysis of periodic vibration signals, the traditional Fourier series represents
a strong and relatively simple methodology for the separation of vibration compo-
nents due to different excitations and the identification of those induced by faults
thanks to their a priori known frequencies (e.g., 1 per rev frequency of unbalance).
The Fourier series decomposes a periodic signal xT (t) (with period T) into a
series of sinewaves with frequencies multiples of the fundamental frequency 1/T,
amplitude Ak , and phase (at time zero) φ k .

∞
  
2π k
xT (t) = Ak cos t + φk (3)
T
k=0
4 Applied Digital Signal Processing 153

A more versatile form of the Fourier series uses the Euler identity to decompose
the cosine functions of the previous equation into rotating complex exponentials:

∞
 2π k
xT (t) = Xk ei T t (4)
k=−∞

where i is the imaginary unit.

In this case both the amplitude and phase information are retained in the complex
coefficients Xk , with the condition that X−k and Xk are a complex conjugate pair for
real signals. They can be computed as:
 T
1 2π k
Xk = xT (t)e−i T t dt (5)
T 0

2.2 Transient Signals and the Fourier Transform

Transient deterministic signals, such as dynamic responses of structures to impul-

sive excitations require an extension of the Fourier series. In this case the frequency
decomposition of the signal is continuous rather than discrete:
 ∞
x(t) = X(f )ei2πf t df (6)
−∞

and the function X(f ) of x(t) is obtained from:

 ∞
X(f ) = x(t)e−i2πf t dt (7)
−∞

The transformations of Eqs. (6) and (7) are called, respectively, inverse Fourier
transform and Fourier transform. Note that the spectrum of Eq. (7) has different
dimensions from that in Eq. (5).

2.3 Discrete-Time Signals and Digital Implementation

Digital signal processing is performed after an analogue-to-digital conversion

(ADC) of finite length signals. The ADC involves a discretization in the time
domain, usually with a constant sample rate Fs (in samples per second). Signals
become therefore time series, and an analogue vibration x(t) is converted into
a discrete-time signal by sampling only at constant time intervals, multiples of
t = 1/Fs :

x [n] = x (nt) (8)

154 R. B. Randall et al.

Quantization (the discretization in the domain of values of the signal) is hereby

not treated as it is usually a design issue, which, if properly dealt with, has negligible
effect on the result of digital vibration signal analysis.
For periodic signals, the extension of Fourier series to discrete time is straight-
forward, given the correct selection of sampling rate and acquisition length. In this
case a simple discretization of the time variable t(tn = nt) in Eq. (7) and the
approximation of the integral with a Riemann sum produces:

N −1
1 
x [n] e−i2π N
nk
X [k] = (9)
N
n=0

The algorithm of Eq. (9) is referred to as the discrete Fourier transform (DFT).
For a generic periodic signal xT (t) with period T, it can be demonstrated that the
DFT coincides exactly with the continuous time coefficients Xk provided that the
sampling rate Fs is an exact multiple of the fundamental frequency f0 = 1/T of the
signal (Fs = mf0 with m positive integer), with the additional condition of Fs > 2nf0
(m > 2n), where n is the number of harmonics of the fundamental frequency present,
and the signal is observed for an integer number of periods, i.e., Nt = pT with
p ∈ Z+ .
The inverse discrete Fourier transform (IDFT) allows exact transformation back
to the time domain:


N −1
nk
x [n] = X [k] ei2π N (10)
k=0

For transient signals, the discretization of time in Eq. (7) leads to the discrete-
time Fourier transform (DTFT):

∞

X(f ) = t x [n] e−i2πf nt (11)
n=−∞

The expression of the DTFT is clearly impractical, given the finite length n = 0,
. . . , N − 1 of all experimental signals. Therefore it is often approximated by a
further discretization, in this case of the frequency variable f (fk = kf ). Taking
Δf = Fs /N reduces the DTFT to a simple DFT, in this case normalized with t:


N −1
nk
X [k] = t x [n] e−i2π N (12)
n=0

The inversion is again performed with the IDFT, in this case normalized by Fs /N:
4 Applied Digital Signal Processing 155

N −1
Fs  nk
x [n] = X [k] ei2π N (13)
N
k=0

The use of the DFT for both periodic and transient signals has led to a certain
degree of freedom of terminology, with DFT and IDFT defined with different
normalizations in different references (the product of the normalization coefficients
of DFT and IDFT must in any case be 1/N to ensure identity when subsequently
applying the direct and inverse transformation). In the most common software
packages, the DFT is defined simply as:


N −1
nk
X [k] = x [n] e−i2π N (14)
n=0

and normalization must be performed according to the application. Details of this

are given below in Sect. 4.

3 Experimental Signals and Stochastic Signal Modelling

Owing to the imperfect nature of measured systems and measuring devices, and
also the inherent stochastic nature of certain excitation signals, such as turbulent
fluid flow in turbines, experimental signals are always characterized by a certain
degree of unpredictability and therefore belong to the random signal family. These
signals cannot be expressed by mathematical functions, but can only be described
by their statistical properties.

3.1 Time-Varying Distributions: Ensemble Versus Time

Quantities

In a generic stochastic model, any measured signal x(t) is seen as a single realization
of a (generally) time-varying distribution Dx (t). In a discrete-time perspective, this
means that the time-sample x[n] = x(nt) is one ensemble sample extracted from
the distribution Dx (nt) while the next time-sample x[n + 1] = x((n + 1)t) is a
stochastic realization of a different distribution Dx ((n + 1)t). The probability that
the two consecutive time-samples jointly take given values is similarly reflected by
a joint distribution. The difference between time and ensemble domains is crucial
in understanding the stochastic signal processing approach and at the basis of
the sub-classification of random signals. In Fig. 1 this interpretation of stochastic
signals is graphically explained with an example. Green lines show numerically
defined distributions (in PDF terms), while the red circles (x1 ) and blue crosses
(x2 ) represent two discrete-time signals obtained from the same time-varying
156 R. B. Randall et al.

distributions. The two simultaneous samples x1 (t0 ) and x2 (t0 ) of the signals are two
ensemble samples from the same ensemble-distribution Dx(t0 ).
Information regarding structure or machine properties is often hidden within the
two first statistical moments of the signal’s time-varying distribution, the expected
value μx (t) and the covariance function Cx (t, τ ):

μx (t) = E{x(t)} (15)

Cx (t, τ ) = E {[x(t) − μx (t)] · [x(t + τ ) − μx (t + τ )]} (16)

where the symbol E denotes the ensemble average in the population “produced” by
distribution Dx in Eq. (15) and its joint version in Eq. (16).
In most practical applications, only one sample is available for each signal at
each time instant (so in the case of Fig. 1, the measurement will either observe x1 or
x2 ); thus the task of inferring properties of the two moments is in general arduous
and must exploit assumptions on the time evolution of the distributions.
Some sub-classes of the random family have been identified as recurrent in
structural and machine dynamics:

Ensemble distributions
Signal x 1
Signal x 2
PDF

Signal
0
value

-5 7 8 9
4 5 6
1 2 3
0
Time

Fig. 1 Numerical example of a discrete-time signal obtained from a time-varying distribution

4 Applied Digital Signal Processing 157

• Stationary signals, characterized by a time-constant distribution. In this case the

ensemble moments can be estimated directly by the time moments, since every
time sample is generated by the same distribution. Condition of convergence
of the time moments to quantities that are identical for all measurements is a
technical concept known as ergodicity.
• Cyclostationary signals, characterized by a periodic evolution of the statistical
moments. In this case the ensemble moments can be estimated using periodic
time subsets of the signal, exploiting the periodic recurrence of samples extracted
from the same distribution. Condition of convergence of the time moments to
quantities that are identical for all measurements is a technical concept known as
cycloergodicity.
• Transient random signals, characterized by a finite energy and concentrated
within a limited time window, usually triggered by a controllable event. In this
case the ensemble moments can be obtained only by replicating experiments to
obtain multiple realizations of the same time-evolving distribution.
• Cyclo-non-stationary signals, where many excitation forces are linked to
machine speed. At constant speed they would be cyclostationary, but with
varying speed the instantaneous frequency varies directly with the speed, and
in general the instantaneous amplitude also varies with the speed.

The use of these sub-classes to model experimental signals in practice is variable

and not consistent, even within the same engineering applications. This is mainly
due to the fact that real signals often show hybrid and imperfect characteristics,
and also often evidence a mixture of types rather than an affiliation to a single
class, which allows assignment of their components to certain families only with
a certain level of approximation. For this reason, the following sections will refer to
these families as models, which are “unrigorously” applied to experimental signals
to allow their description in a “rigorous” analytical framework. Therefore, the role
of the engineer should often focus on the choice of the most suitable model for
the signal, resulting in a series of available analytical tools whose mathematics has
already been developed and validated by signal processing experts and statisticians.
In order to understand the importance of this step, it is important to highlight the fact
that the same analytical techniques are often applied in different signal models and
that a proper implementation and interpretation of their outcomes is only ensured
if considered in the selected modelling framework. In the rest of this section, some
popular and effective signal models are introduced, while the following section will
discuss typical engineering problems and their relationship with these models.

3.2 Stationary Signal Model

Stationary signals are characterized by time-constant stochastic properties, i.e.,

every sample of the signal x(t) is generated from the same statistical distribution
Dx , which does not vary in time (Dx (t1 ) = Dx (t2 ) = Dx ). Therefore, the mean μx
and the covariance function Rxx have the following properties:
158 R. B. Randall et al.

μx (t1 ) = μx (t2 ) = μx for each t1 , t2 (17)

Cx (t1 , τ ) = Cx (t2 , τ ) = Cx (τ ) for each t1 , t2 , τ (18)

Despite more stringent requirement for strict stationarity (full time-independent

distribution), often the conditions on the first two moments of Eqs. (17) and (18) are
considered sufficient for a de facto stationarity, labelled as “wide-sense stationarity”
(WSS). A typical example of stationary signal modelling is encountered in OMA
under the assumption of random excitation (e.g., wind, vehicle and pedestrian
traffic, earthquake). This modelling is of course only valid if the time scale of
the observation is chosen so that the excitation can be modelled as stationary.
Measurement noise is also typically modelled as stationary. A stationary signal
model is often a matter of convenience to simplify the analysis.

3.3 Cyclostationary Signal Model

Cyclostationary signals are non-stationary random signals characterized by periodic

statistical properties. As for the strict definition of stationarity, strict cyclostation-
arity requires a periodicity of the full distribution generating the signal, but a
“wide-sense cyclostationarity” (WSCS) only requires the first two moments to be
periodic. For the sake of simplicity, WSCS signals are generally called “cyclosta-
tionary” (CS), implicitly recognizing the actual unfeasibility of testing for “strict
cyclostationarity.” Within CS signals a strong distinction is made between signals
which show periodicity in their mean (first-order cyclostationary or CS1) and signals
which have a periodic autocorrelation function (second-order cyclostationary or
CS2). T-periodic CS1 signals have a periodic mean:

μx (t) = μx (t + T ) ∀t ∈ R (19)

T-periodic CS2 signals have a periodic autocorrelation function:

Cx (t, τ ) = Cx (t + T , τ )∀t, τ ∈ R (20)

This cyclic recurrence of statistical properties allows the extension of some

characteristics of stationary signals to this class. In particular, if a cyclostationary
signal is characterized by the cyclic period T, the subset of samples {x(t + kT ), k ∈
Z} is a stationary process (for a fixed value of t):

μx (t) = μx (t + kT ) k ∈ Z, ∀t, τ ∈ R (21)

4 Applied Digital Signal Processing 159

Cx (t, τ ) = Cx (t + T , τ ) k ∈ Z, ∀t, τ ∈ R (22)

CS1 and CS2 signals can be seen as a composition of a stationary random

component r(t) and a deterministic periodic signal pT (t). In particular, a pure CS1
signal xCS1 and a pure CS2 signal xCS2 can be modelled as:

xCS1 (t) = r(t) + pT (t) (23)

xCS2 (t) = r(t)· pT (t) (24)

An example of this signal composition is given in Fig. 2.

Most signals from machines running at constant speed show cyclostationary
behaviour and are generally composed of a mix of CS1 and CS2 components.
Typical CS1 components are detectable in system responses to periodic excitations
(e.g., unbalance in turbomachinery, dominant components in gearmesh-induced
vibrations), while CS2 components arise in case of modulation/convolution of
random phenomena by cyclic phenomena (e.g., bearing fault-induced vibrations,
systems excited by pulsating random forces such as engine combustion). The

Deterministic Periodic Stationary CS1

0 0 0

0 0.5 1 0 0.5 1 0 0.5 1

time time time
Deterministic Periodic Stationary CS2

0 0 0

0 0.5 1 0 0.5 1 0 0.5 1

time time time

Fig. 2 Numerically generated examples of deterministic periodic, stationary random, CS1 and
CS2 signals. CS1 is obtained by summation of stationary random and deterministic periodic CS2
by multiplication (amplitude modulation) of stationary random and deterministic periodic (positive
only)
160 R. B. Randall et al.

extraction/estimation of the periodic statistical properties is the main objective of

cyclostationary analysis, both for CS1 and CS2 signals.

3.3.1 CS1 Model

In case of first-order cyclostationary signals (CS1), the identification of the periodic
part consists of a separation of the two terms of the summation in Eq. (23) and
a removal of the “noisy” part r(t), which does not carry information on the fault.
This is obtained by a series of methodologies which will be discussed in detail in
Sect. 6.6.

3.3.2 CS2 Model

In the case of second-order cyclostationarity (CS2), the multiplication of Eq. (24)
makes the random part r(t) a relevant component of the fault symptom: i.e., in
contrast to the CS1 case, the random part is the carrier of the information, and
its properties are often of physical relevance to identify the source of the fault.
To better understand the properties of CS2 signals, a further numerical example
is produced in Fig. 3. Each of the four CS2 signals carries characteristics from both
the stationary and periodic “parent signals.” In particular the short-time behaviour
is inherited from the “random parent”: CS2 signals in rows (a) and (b) preserve the
high correlation Cx over time lags τ longer than the examples of rows (c) and (d);
i.e., the random short-term variability of the CS2 signals is “smoother” in (a, b) than
in (c, d). On the other hand, the macroscopic behaviour of the signals is determined
by the characteristic period of the “periodic parent”: CS2 signals in rows (a) and (c)
show a cyclic power modulation slower than those of rows (b) and (d).
In an intuitive acoustical example of CS2 sound, the short-term variability of
the stationary random carrier r(t) is reflected in the “pitch” of the sound, while
the modulation introduced by pT (t) gives the rhythm with which the noise is
reproducing.

3.4 Cyclo-non-stationarity

Here only a very brief introduction to cyclo-non-stationarity is given, but reference

is made to a number of papers giving further detail. In particular, the description
here is based largely on that in Ref. [1], which was one of the first to develop
these ideas. As mentioned earlier, cyclo-non-stationary (CNS) signals are similar
to cyclostationary (CS) signals except that the statistics, rather than being periodic,
vary deterministically in a known way. The most common situation giving rise to
this is where the machine speed, which gives periodic statistics when it is constant,
varies as a known function of time, and this is the example used here, in Fig. 4 (from
[1]), to demonstrate the general case.
Fig. 4a shows a typical CNS signal arising from an impulse occurring every
revolution of a shaft, at the same rotation angle, and exciting an impulse response
that is independent of the shaft speed (i.e., of the spacing between the impulses).
4 Applied Digital Signal Processing 161

Fig. 3 Examples of different CS2 signals obtained from long-time (a, b) and short-time (c, d)
correlation stationary noise and long (a, c) and short (b, d) period deterministic modulations

The rotation angle of the shaft is θ (t), and its variation with time must be known.
Such a signal can be described by a Fourier series expansion:

x(t) = ck (t)eikθ(t) (25)
k

where the Fourier series coefficients (related to the impulse responses in this case)
are time dependent and the complex exponentials (related to the impulses) are
functions of angle.
It is shown in [1] that if the correlation length of the time-varying components is
short with respect to the cycle duration, such signals are characterized as angle-time
cyclostationary (AT-CS). If they are analyzed by standard CS methods, as shown in
Fig. 4b where both time lag τ and cycle period are in time units, the time domain
autocovariance function gives local autocorrelations non-uniformly spaced in time,
 162

Fig. 4 Example of cyclo-non-stationarity and spectral correlation diagrams (from [1])

R. B. Randall et al.
4 Applied Digital Signal Processing 163

whereas if the cyclic period is in terms of angle (Fig. 4c), the angle-time covariance
function in Eq. (26), defined as a function of position in angle, and time lag in time,
gives uniformly spaced autocorrelations.

C2x (θ, t) = E {x (t (θ )) x (t (θ ) − τ )} (26)

The resulting spectral correlation diagrams obtained by performing Fourier

transforms in the lag and cyclic directions, respectively, give the results shown in
Fig. 4d, e. The non-uniform spacing of the autocorrelation functions in Fig. 4b
smears the classic spectral correlation (SC) in Fig. 4d whereas the periodic spacing
in Fig. 4c means that the corresponding order-frequency spectral correlation (OFSC)
in Fig. 4e has the frequency responses corresponding to the impulse responses
located at the various orders of the basic cyclic periodicity. Thus, a complete
separation has been made between the time-frequency nature of the impulse
responses and the angle/order nature of the repetition rate defined by the machine
speed.
Ref. [1] make reference to a number of applications of this CNS theory, while
Ref. [2] give further insights and applications.

3.5 Transient Signal Model

Transient random signals are stochastic signals whose energy is mostly concentrated
in a finite
 time window. Since the instantaneous power of a signal is defined as
E x(t)2 = |μx (t)|2 + Rxx (t, 0), a signal cannot be both stationary and transient.
Pure transient signals have a finite energy, resulting in the properties:

 +∞
μx (t) dt = a with a ∈ R (27)
−∞

 +∞
Cx (t, τ ) dt = p (τ ) with p (τ ) ∈ R, ∀τ (28)
−∞

As for stationary signals, pure transients are rarely encountered in practice, where
even a small background noise would by itself compromise the condition of finite
energy. However, cases such as responses of damped structures to a single impulsive
excitation (e.g., impact tests) can be generally modelled as transients, with μx (t)
dominated by the impulse response of the structure. This modelling choice is often
coupled with the application of a time-windowing function, forcing a long-term
decay of the residual vibration signal.
164 R. B. Randall et al.

4 Scaling and Dimensions for Various Versions

of the Fourier Transform

It is not sufficiently widely realized that the different forms of the Fourier transform
in general give frequency spectra with different dimensions and units, and this
can give problems when attempting to represent the spectra of signals containing
mixtures of different signal types.
Unless otherwise specified, in this chapter the word “spectrum” will be used
to represent the result of an FFT analysis scaled in engineering units EU (i.e.,
scaled as for Fourier series and valid for discrete frequency components). In
section 4 only, for consistency with other chapters, asterisk * is used to represent
multiplication, but elsewhere it is used to represent convolution. If a signal also
contains stationary random or transient components, these will not be scaled
correctly in this “spectrum”. A “power spectrum” will be the squared amplitude of
a “spectrum”, scaled in EU*EU, and will also only be valid for discrete frequency
components. A “power spectral density” spectrum will have units of (EU * EU/Hz)
and is only valid for the spectra of stationary random signals. If obtained from a
“power spectrum”, it must be divided by the effective bandwidth of each line in the
FFT spectrum, which is the line spacing multiplied by a factor accounting for the
noise bandwidth of any window used (e.g., it is 1.5 for a Hanning window). Scaling
of log spectra in dB re r EU will be the same for spectra and power spectra, since
it equals 10 log10 ({|X|/r}2 ) = 20 log10 (|X|/r), where X is the spectrum value in EU
and r is the dB reference value. For further details and the scaling of other signal
types, such as transients, see the following sections.

4.1 Periodic and Quasi-periodic Signals

This category of signals is made up entirely of discrete frequency sinusoidal

components. Periodic signals contain only components with frequencies that are
integer multiples (harmonics) of a fundamental frequency, because after one period
of the fundamental frequency, all harmonics have run through an integer number of
cycles and their sum starts again with the same conditions, making them periodic
with this period. Quasi-periodic signals have at least two frequencies without a
lowest common multiple, so that no fundamental frequency can be found. In theory
this means that the ratio of two frequencies must be an irrational number, but in
practice quasi-periodic signals occur when there is a mixture of two or more periodic
components from independent sources, such as the various rotors of an aero engine,
connected only aerodynamically.
Both signal types can be analyzed using the Fourier series equations (3, 4, 5)
with the proviso that for quasi-periodic signals, the limit should be taken in Eq. (5)
as T → ∞.
Since the integral is normalized by dividing by T (Eq. (5), this is the only form
of the Fourier transform giving spectra with the same units and dimensions as the
4 Applied Digital Signal Processing 165

original signal. The individual sinusoids have instantaneous “power” equal to the
square of the instantaneous value, as does the sum of all components, and the mean
power is the mean square value averaged over a period of the fundamental frequency
or infinity. The signal strength is usually expressed in terms of its RMS (root mean
square) value, the square root of the mean square value (power), as this has the
same units as the original signal, and is independent of the phase relationships of
the components.
The word “power” is used to describe the square of a measured parameter with
units EU and is thus EU*EU. It can usually be related to physical power via some
kind of dimensioned scaling constant. For example, for a resistive electrical circuit,
with voltage V, current I, and resistance R, power W = IV = I2 R = V2 /R since
V = IR. Similarly, “energy” is used to describe the time integral of “power.”
Parseval’s theorem states that the total power (or energy, depending on the signal
type) of a signal is the same whether integrated over all time or all frequency. For the
Fourier series spectrum, the mean square value of the time signal is equal to the sum
of squares of the amplitude of all spectrum components. For the two-sided spectrum
of Eqs. (4, 5), the power is equally divided between positive and negative frequency
components for each frequency, and |Xk | = Ak /2 so |Xk |2 + |X−k |2 = Ak 2 /2, the
same as obtained for a single sinusoid in the time domain, since Ak 2 cos2 (ωt) =
Ak 2 1/2 + 1/2 cos (2ωt) , whose mean value is Ak 2 /2.
The spectrum of squared amplitude values |Xk |2 is called the power spectrum,
and each discrete frequency component will have units EU · EU. If the DFT of
Eq. (14) is used (e.g., the FFT in Matlab® ), the result must be divided by N to
give correct scaling, and the inverse transform (divided by N in Matlab® ) must be
multiplied by N.

4.2 Stationary Random Signals, Power Spectral Density (PSD)

For stationary random signals, the instantaneous power is still the squared value of
the time signal, but now the spectrum is continuous. Parseval’s theorem now states
that the total power in the time domain (the mean square value) must be equal to
the integral over all frequency of the squared amplitude of the spectrum, which thus
must be a spectral density with units EU*EU/Hz. It is called the “power spectral
density” or PSD. There is no direct form of the Fourier transform for stationary
random signals,√but if there were, the normalization of the integral would
√ have to be
by division by T (since the integral tends to a limit proportional to T ), giving a
√
direct spectrum with units EU* s, for which the amplitude squared spectrum would
have units EU*EU*s or EU*EU/Hz. In fact the spectrum of a stationary random
signal is usually obtained by Fourier transformation (Eq. 7) of the autocorrelation
function with units EU*EU, once again giving a spectrum with units EU*EU*s or
EU*EU/Hz. The autocorrelation function of a stationary random signal is a transient
with finite length.
166 R. B. Randall et al.

The power spectral density of stationary random signals can be estimated by

averaging the squared amplitude spectra of records of length N obtained by FFT. If
the FFT corresponds to Eq. (14), the power in each spectral line will be obtained by
division by N2 , and if the weighting of each record is rectangular (i.e., no weighting),
the PSD at each line (assumed constant over the interval) can be estimated by
dividing by the bandwidth per line, Fs /N, altogether a division of the original
squared values by NFs . If other than a rectangular window is used, this estimate must
also be divided by the relative bandwidth of the window (e.g., 1.5 for a Hanning
window). This will give a PSD spectrum with the correct units of EU*EU/Hz.
Alternative ways of calculating PSD spectra are given in  Chap. 5, “Introduction
to Spectral and Correlation Analysis: Basic Measurements and Methods.”

4.3 Deterministic Transient Signals, Energy Spectral Density

(ESD)

Transient signals with units EU have a finite amount of energy, equal to the integral
of their instantaneous power over all time, and having the units EU*EU*s. The
application of the Fourier transform (Eq. (7)) to them gives a spectrum with the
units EU*s, whose amplitude squared spectrum thus has the units EU*EU*s*s or
EU*EU*s/Hz. This is known as “energy spectral density” or ESD.
In this case, Parseval’s theorem states that the total integral over all time or all
frequency of the squared amplitude is the same, and is equal to the total energy, with
units EU*EU*s.
If the FFT is used as above, the power per line is estimated by dividing
the squared amplitude values by N2 , once again converted to a spectral density
by dividing by the line spacing (equivalent to multiplication by record length
T = N/Fs ) and then converted to energy (per record length) by multiplication again
by T = N/Fs . Overall, this corresponds simply to division by Fs 2 and gives a result
with units EU*EU*s/Hz.
Note that rectangular windowing is normally assumed here, as even though
windows are sometimes used, they genuinely affect the energy and ESD in a non-
predictable way, and cannot be easily compensated for.

5 Choosing the Right Model for an Experimental Signal

In order to assist the reader in the complex task of choosing the best model for a
specific vibration analysis problem, it is useful to start from the identification of the
information to be retrieved from the signal. The information available in a vibration
signal usually relates to two key dynamical aspects: excitations and system response.
The extraction of selective information on the two aspects (always present in
combination) is the aim of the main typologies of application in the field of vibration
signal processing: modal analysis and condition monitoring. Modal analysis aims
at identifying the characteristics of the system response (e.g., natural frequencies,
4 Applied Digital Signal Processing 167

damping), thus “removing” the excitation-induced characteristics from the signal.

Condition monitoring can focus either on the identification of variations of the
system response indicative of specific damages in the structure, or on the detection
of particular excitations arising from faults or malfunctions.

5.1 Modal Analysis

Analysis methods for conventional modal analysis, involving frequency response

functions (FRFs), impulse response functions, etc., are well covered in specialist
chapters dealing with these topics, for example,  Chap. 6, “Frequency Response
Function Estimation” and  Chap. 11, “Experimental Modal Analysis Methods,”
with additional insights on signal processing given in  Chap. 5, “Introduction to
Spectral and Correlation Analysis: Basic Measurements and Methods.”
Here we refer only to material not treated elsewhere, in particular cepstral
methods of (operational) modal analysis, discussed briefly in Sect. 10.4. A historical
survey of cepstral methods has recently been published [3].

5.2 Condition Monitoring for Rotating/Reciprocating Machines

Condition monitoring signals are often composed of a rich series of diverse compo-
nents, each most suitably described by a different signal model. The complexity of
machine signals is often the result of the presence of multiple rotating/reciprocating
elements, each characterized by a different kinematic and dynamic behaviour.
Typical signals are, for instance, composed of vibration components carrying
information on multiple shafts, bearings, gears, combustion chambers, and fluid
guide vanes/blades. In the case of rotating and reciprocating machines, faults or
malfunctions in each part are usually associated with a specific set of frequencies
related to the kinematics of the component itself. The cyclic nature of such
kinematics and the resulting vibration components makes cyclostationary models
the most suitable for this family of applications. The choice between CS1 and
CS2 models is usually made considering the degree of randomness of the fault-
related phenomenon and in particular whether the periodic component is additive or
modulates random components. It is discussed in the following subsections for the
most typical cases.

5.2.1 Typical Rotor Problems

Typical rotor problems, such as unbalance, misalignment, and gearmeshing, have
excitations theoretically phase-locked to the rotation of the corresponding shaft.
Slight non-linearities in the response (e.g., fluid film bearings) do not compromise
the “periodicity” of the phenomenon, and simply result in the generation of higher
harmonics in the vibration signal. Therefore, if the speed is constant, the only source
of randomness is represented by measurement noise and other additive uncertainties
(e.g., external sources of vibration). In this case, the expected value of the vibration
168 R. B. Randall et al.

signal is highly periodic and characterized by a priori known frequencies related to

the shaft speed, making CS1 the most suitable model. The most common CS1 tools
used for the analysis of these signals are Fourier series and synchronous averaging.
Both methods are meant to highlight the periodic content to the detriment of the
random part.
The behaviour of flexible rotors suspended in fluid dynamic bearings and running
at speeds above their first critical is quite complex and has been the subject of much
specialized research over the years. It is not covered in much detail here, but the
reader is referred to specialist publications on rotor dynamics.

5.2.2 Gear Faults

Gear faults have been studied over many years, and many diagnostic methods
developed, but primarily for machines with almost constant speed. The vibration
signals from meshing gears at almost constant speed are mainly deterministic, at
least in the rotation angle domain, and with constant load, since particular pairs of
teeth (even with faults on either or both) always mesh in the same way, and their
response signals should be periodic with some period. On the other hand, for gear
systems with several stages, and with so-called “hunting tooth” design (where the
numbers of teeth on a meshing pair have no common factors), the period may be
very long and the signal pseudo-random with that period. When the time scales of
the analysis records are shorter than the pseudo-random period, the signals appear
random, and can then be modelled as a mixture of CS1 and CS2, though usually
with the CS1 dominant.
In any case the signal associated with a particular gear can usually be treated
as CS1 and the dominant periodic part extracted using time synchronous averaging
(TSA). If the TSA is done using the speed of a particular shaft for synchronization,
the average will represent the gear(s) on this shaft, and the spectrum of the TSA
signal will contain only harmonics of this shaft speed. If there is more than one
gear mounted on the shaft, the harmonics coming from each will be present, but
often these are widely spaced in frequency (very different numbers of teeth) and
can largely be separated in the frequency domain. Another alternative for TSA is
to make the average for a particular gearmesh, by making the average over records
encompassing a number of toothmesh periods corresponding to the lowest common
multiple of the numbers of teeth on each gear. The spectrum should then only
contain the harmonics of the two gear speeds in that mesh, including the gearmesh
frequency (and its harmonics) which is a common harmonic of both gear speeds.
TSA is primarily used for extracting information from the time signals, but as
discussed in Sect. 9.2 further analysis can be performed such as removal of the
regular toothmeshing patterns, so as to reveal local faults, and demodulation of the
toothmesh frequency to detect sudden changes in both amplitude and phase of the
meshing pattern.
The other main analysis technique for gears is frequency analysis, often supple-
mented by cepstrum analysis as discussed in Sect. 9.2. Gear faults can be roughly
divided between uniformly distributed faults such as uniform wear and localized
faults such as tooth root cracks and spalls on individual teeth or groups of teeth.
4 Applied Digital Signal Processing 169

The first show up primarily as changes in the harmonics of the toothmesh frequency
(often starting with the second harmonic, because wear tends to concentrate on
either side of the pitch point where there is no sliding between teeth). Localized and
non-uniformly distributed faults tend to show up in modulation sidebands around
the toothmesh harmonics, with a spacing equal to the speed of the gear on which
the fault is located. Because these sideband patterns can be complex, and mixed for
the two gears, cepstrum analysis is a useful tool to separate and quantify them, as
shown in Sect. 10.3.
With varying speed, new considerations arise in that the forcing functions vary
directly with speed, whereas resonant responses are fixed in frequency. Order
tracking (Sect. 8.1) compensates for frequency variation, but not, for example,
for the amplitude variations caused by passage of gearmesh frequencies through
resonances. A potential solution of this problem using cepstral methods is given in
Sect. 10.3.

5.2.3 Bearing Faults

Bearing diagnostics is a widely explored yet challenging signal processing applica-
tion. Most typical bearing faults are localized spalls on the bearing races caused
by the surfacing of contact-fatigue sub-surface cracks [4]. Once in operation at
constant speed, the rolling elements will subsequently impact on the race spall
with approximate time regularity, made imperfect by the stochastic slip between
rollers and races. Each roller-spall impact will trigger vibrations in the system at the
natural frequencies of the structure (the case of a spalled rolling element results in
an analogous signal, with a different cyclic periodicity). These vibration bursts are
reproduced with regularity in the signal, as subsequent rollers impact with the spall,
and their cyclic repetition is characterized by a known almost-period 1/BCF, where
the so-called bearing characteristic frequency (BCF) depends only on the shaft
rotational speed, the geometry of the bearing, and the location of the spall (inner,
outer race or roller). This signal model is expressed by the following equation:

  
1
s(t) = Ai h t − i − τi + n(t) (29)
i BCF

where h(t) is the impulse response of the system characterized by a set of natural
frequencies, τ i is a small random delay/anticipation of the i-th impact due to slip
between rollers and races, Ai is the amplitude of the i-th impact (usually random
and in general variable with the rotation of the shaft), and n(t) is background noise
(Table 1).
A simple numerical simulation of this model is reproduced in Fig. 5, assuming a
single degree of freedom h(t), i.e., a single natural frequency fn , and a constant Ai .
In practice, the presence of multiple frequencies, the highly random behaviour of
Ai , and the time variation of h(t) with the rotation of the shaft increase the random
behaviour of the signal, already ensured by the non-null random “jitter” τ i . The
second-order cyclostationary model is obviously the natural candidate to describe
170 R. B. Randall et al.

Table 1 Bearing characteristic frequencies. f = shaft rotational freqeuncy, NR = number of

rollers, PD = pitch diameter, RD = roller diameter, α = contact angle
Location of the fault Bearing characteristic frequency BCF expression
Inner race Ball pass frequency inner race BP F I = N R f2 1 + PRDD cos α
f
Outer race Ball pass frequency outer race BP F O = N R  2 1 − RD
P D cos α
2
Roller Ball spin frequency BSF = 2RD f 1 − P D cos α
PD RD

f
Cage Fundamental train frequency FT F = 2 1 − P D cos α
RD

Fig. 5 Simplified numerical example of a faulty bearing vibration signal

this combination of random and cyclic nature of bearing signals, showing a power
(i.e., second-order statistics) which cyclically varies in time.
Note that the random “jitter” model is not strictly correct, although it is a good
approximation in most cases. In Refs. [5, 6] it is shown that the true random variable
is not a jitter in the length of each period, but the actual period itself, meaning
that the bearing signal is not truly cyclostationary, but has been termed “pseudo-
cyclostationary.” As shown in the same references, the differences for practical
purposes are very small. In addition, the actual average period can deviate (usually
by a few percentage points) from the theoretical values [7]. Finally, as for all
rotating machines at variable speed (see Sect. 3.4), the most correct model if speed
fluctuations are of non-negligible magnitude is a cyclo-non-stationary one [8].

5.2.4 Internal Combustion Engines

The signals from internal combustion (IC) engines, and in fact most reciprocating
machines, are a series of impulsive events, for example, combustion, piston slap,
bearing knock, and valves opening and closing, and a natural way to analyze them
is by some form of time-frequency analysis. This is analogous to the way in which
an experienced mechanic detects faults by ear, by hearing changes in both the cyclic
patterns of the events, and in the frequency content of the events. Combustion events
are cyclic but not exactly periodic, so the signals are a mixture of CS1 and CS2
cyclostationary components, with neither necessarily dominant.
Mechanical faults, such as piston slap and bearing knock, caused by increasing
clearance between piston and cylinder, and in the bearings of the connecting rod,
respectively, give impulsive impacts rather like rolling element bearing faults, with
4 Applied Digital Signal Processing 171

some random variation in the spacing, and are therefore dominantly CS2. They can
be analyzed in the same way as bearing faults, by envelope analysis.
Combustion faults such as misfires are a particular case where non-uniformity
between the cylinders gives rise to fluctuating torques, and corresponding variations
in the torsional vibration of the crankshaft. Very simple analysis of this torsional
vibration is a powerful technique for detecting, diagnosing, and quantifying such
faults. Rotational accelerations of the engine block are also sensitive to the
combustion faults and can be used as an alternative, though not quite so simple,
method.
Examples of a number of diagnostic methods for typical cases are given in Sect.
9.3.

6 Signal Extraction and Separation

6.1 General Introduction

One of the principal aims of signal processing is to extract information from

non-ideal measurements. This is typically the case with vibration and acoustic
signals which are composed of the responses to multiple sources of excitation
together with possible instrumentation and environmental interferences. In many
situations, the user will be interested in extracting or assessing a restricted fraction
of the total information contained in the measurement – this formally defines the
“signal” on the one hand and the “noise” on the other hand. The objective of signal
extraction/separation is purposely to recover the signal part from the noise part.
This general statement actually goes beyond classical filtering or denoising and
embodies a collection of techniques which have been developed in various domains
of applications. In health monitoring, signal extraction/separation is typically used
to extricate symptomatic sources that can be further analyzed or controlled for
diagnostic purposes. In structural dynamics, these techniques have also been
successfully used to identify the different modal contributions of a system.
Formally, the issue may be described by the general equation:

y(t) = x1 (t) + · · · + xn (t) + n(t) (30)

where y(t) stands for the current measurement, x1 (t), . . . , xn (t) for the components
which constitute the informative part – or the “signal” – and n(t) for the disturbance,
or the “noise.” It is generally accepted that “signal extraction” aims at isolating
one of the components of interest, while “signal separation” aims at recovering
simultaneously all the components in the decomposition of Eq. (30).

6.2 Signal Extraction

Signal extraction may be described as isolating one single component, say x(t), from
measurement y(t):
172 R. B. Randall et al.

y(t) = x(t) + n(t) (31)

Note that, without loss of generality, other components which are not of interest
have been relegated to the noise part. In signal extraction, component x(t) is also
referred to as the “signal of interest.” In many instances and especially when dealing
with vibration and acoustics, the objective is far from easy since the signal of interest
may be completely masked by the noise. This is typically the case with incipient
faults (such as in bearings) which are often characterized by a small signal-to-noise
ratio.
Different strategies for signal extraction are envisioned depending on the avail-
able information and statistical properties of the components in Eq. (31).

6.3 Reference-Based Filtering

The ideal situation is when the component of interest is indirectly sensed by an

extra measurement which returns a possibly filtered but noise-free image of it. This
is described by the following set of equations:

y(t) = x(t) + n(t)

(32)
r(t) = h(t) ∗ x(t)

where h(t) stands for a (unknown) linear filter, * for the convolution operation,
and r(t) for the “reference” on the source of interest. This scenario is typically
encountered when there are several transmission paths for the source of interest, one
of which can be singled out. Two classical solutions are provided by the Wiener and
the Kalman filters. The Wiener filter assumes an unknown time-invariant filter h(t)
and returns an estimate of x(t) by filtering y(t) with the linear filter whose frequency
gain is the ratio of the cross-spectrum Sxr (f ) to the autospectrum Sr (f ):

Sxr (f )
G(f ) = (33)
Sr (f )

This is in all points similar to the H1 estimation of the frequency response

functions as commonly practised in modal analysis. It is widely used for transfer
path analysis (TPA) and dereverberation in acoustics. The Kalman filter is a model-
based approach that accepts a time-varying filter but requires it to have a known
state-space description (at least in its original formulation).

6.4 Filtration

The simplest situation in the case where no reference is available is when the two
components in Eq. 31 have known and quasi-disjoint support sets in a specific
domain. Signal extraction then boils down to denoising by filtering.
4 Applied Digital Signal Processing 173

If support sets happen to be quasi-disjoint in the frequency domain, then linear

filtering will obviously extract signal components with a degree of precision that
directly depends on the overlap between the spectra of x(t) and n(t). Indeed, if
Sx (f ) and Sn (f ) denote the energy/power spectra of x(t) and n(t), then the optimal
separation filter (in the sense of minimizing the mean-square error) to be applied to
y(t) to extract x(t) has frequency gain:

Sx (f )
G(f ) = (34)
Sx (f ) + Sn (f )

which is close to one when the signal-to-noise Sx (f )/Sn (f ) is high and close to zero
when it is low.
Similar ideas can be transposed to other domains, sometimes with different
filtering strategies. If component x(t) has known and nearly disjoint support sets
in time, then it can be simply extracted by time gating (or “windowing”), a practice
largely used to filter out transients in reciprocating machine signals, for instance. In
this case the optimal window is given by the exact counterpart of Eq. (34) in the
time domain. The cepstrum liftering discussed in Sect. 10.2 is a special case where
signals have disjoint supports in the cepstral domain.
More generally, signals may have overlapping time and spectral contents but are
still separable in the joint time-frequency or time-scale domains. This has led to
several new filtering strategies based on invertible time-frequency transforms (e.g.,
the Gabor transform) and the wavelet transforms which are direct generalizations
of the former approaches. A simple example is the use of time-frequency masks to
gate the signal components of interest in the time-frequency plane or of time-scale
thresholds to clip out noise components in the wavelet transform.

6.5 Blind Extraction

Blind extraction basically aims at achieving a filtration of the signal of interest when
prior information is available in the form of a statistical property that uniquely
defines it and differentiates it from the noise part. This is similar to a marker
which makes possible to track the presence of the signal even in strong background
noise. As illustrated in Fig. 6, blind extraction consists in shaping a linear filter
to be applied to y(t) which maximizes the marker property of the output so that it
resembles as much as much as possible the expected signal x(t). Three markers
commonly used with machine signals are i) periodicity, ii) impulsivity, and iii)
cyclostationarity. Since the blind extraction of periodic signals is found particularly
important, it is reviewed in a dedicated subsection. The blind extraction of impulsive
and cyclostationary signals is briefly discussed hereafter.

6.5.1 Blind Extraction of Impulsive Signals

Impulsive signals pertain to the non-stationary class and usually refer to signals
made of series of impulses or transients which may occur on a repetitive basis
174 R. B. Randall et al.

Fig. 6 Principle of blind extraction of a signal: a linear filter is blindly learned so as to extract a
signal of which a given characteristic (e.g., impulsivity, cyclostationarity) is maximized

or not. They are often indicative of mechanical impacts, shocks, or sudden force
variations in the system. In particular, they constitute symptomatic signatures of
many incipient faults such as in gears and in bearings. Impulsive signals produce
heavy-tailed histograms with a wider spread and more peaked in the middle than
the Gaussian (given the same standard deviation). Therefore, impulsivity is often
measured as departure from Gaussianity, for instance, by means of the kurtosis
(which will take high positive values) or the entropy (various definitions are pro-
posed in signal processing, which usually lead to a maximal value when the signal
is Gaussian). The effect of (presumably Gaussian) additive noise on an impulsive
signal is to progressively drive its original distribution to a Gaussian shape. Blind
extraction explicitly exploits this fact in the reverse way: by automatically tuning
a filter that maximizes a measure of impulsivity of the signal, additive noise is
consequently removed as much as possible.
The direct approach which consists in learning a filter by maximizing the kurtosis
(or by minimizing the entropy) leads to sophisticated optimization algorithms.
A much simpler but suboptimal method is to test several filters in a dictionary
(typically narrowband filters with different central frequencies and bandwidths) and
to retain that one (or a combination of those ones) which maximizes the criterion.
The Fast-Kurtogram proposes such a solution based on an efficient implementation
by means of multirate filters.

6.5.2 Blind Extraction of Cyclostationary Signals

Cyclostationary signals have been introduced in Sect. 3.3. They represent an
important class because many signals – especially from rotating machines –
can be modelled as such. In principle, the blind extraction of cyclostationary
signals follows exactly the same lines as for impulsive signals. In particular, the
cyclostationary property of a signal is progressively attenuated by adding stationary
noise so that, proceeding in the reverse sense, a blind extraction filter is learned
such as to maximize a criterion of cyclostationarity. The absolute (or quadratic)
sum of the cyclic correlations (i.e., the Fourier coefficients of the instantaneous
autocorrelation function) or of the cyclic spectra (i.e., the Fourier transform of the
cyclic correlations) is a good candidate to define indicators of cyclostationarity.
However, in the case of cyclostationary signals, the spectral correlation (the
waterfall of all cyclic spectra represented as a two-variable function of the carrier
4 Applied Digital Signal Processing 175

frequency f and the cyclic frequency α) provides a simple tool to visually identify
the frequency bands where cyclostationarity is dominant and which are therefore
eligible to design an extraction filter.

6.6 Blind Deconvolution

Blind deconvolution goes one step further than blind extraction by seeking to
reconstruct the excitation source that produced the signal of interest. The problem
reads:

y(t) = x(t) + n(t) = h(t) ∗ s(t) + n(t) (35)

where s(t) is the excitation source to be recovered and h(t) is an unknown linear
filter such that x(t) = h(t) * s(t). Note that the problem has no unique solution
in general since it contains two unknowns for one noisy observation. It however
represents the ultimate goal in many applications and especially in acoustics and
vibration where the source of excitation is often the quantity of interest. In the
diagnostics of machines, it gives access to the indirect measurement of the fault and
therefore allows the assessment of its dimensions, severity, and of the remaining
useful life. The blind deconvolution problem cannot be solved unless strong prior
information on the excitation source is available. For instance, it often happens
that incipient faults in rotating machines produce excitation forces which are well
modelled by a series of delta functions (with possible random times of occurrence
and magnitudes). This puts strong prior information on s(t) in the form of a flat
spectrum and a leptokurtic probability distribution (i.e., with a high kurtosis). The
effect of the model of Eq. (35) is to transform this much contrasted source s(t) into
a diffuse signal x(t) = h(t) * s(t) (the convolution has the effect of spreading the
energy of the impulsive signal over the time axis and additive noise of making
y(t) more and more Gaussian). Blind deconvolution attempts to proceed in the
reverse direction by seeking an inverse filter to be applied to y(t) that maximizes
the supposed characteristics of the source; the output of the inverse filter is then
expected to resemble as much as possible the unknown source s(t).
Popular criteria that are optimized to blindly deconvolve a white and impulsive
excitation are the signal entropy, higher-order statistics (of which the kurtosis is a
particular case), and the Lp (pseudo)-norms (0 ≤ p < 2). If the excitation is impulsive
and repetitive at the same time (e.g., synchronized on a rotating component), it is
wise to maximize a measure of cyclostationarity. As shown in Sects. 10.3 and 10.4,
the cepstrum sometimes deconvolves forcing and transfer functions.

6.7 Discrete-Random Separation

In vibration and acoustics, periodic and random signals often reflect phenomena
of different origins. For instance, in rotating machines, strong periodic signals are
176 R. B. Randall et al.

emitted by the shaft and gear rotations, while bearings contribute to producing
random (though almost periodic) signals (because of random slip). Periodic and
random vibrations may also be distinguished according to the scale of the physical
phenomena they originate from, the former being more related to macro-phenomena
and the latter to micro-phenomena. From another point of view, periodic signals
relate to deterministic phenomena (i.e., which can be precisely described by
equations) in contrast to random signals. For all these reasons, the separation of
discrete and random components is of importance in many applications. If only the
former components are of interest, the problem is also known as “harmonic (or line)
enhancement.” As seen in Sect. 2.1, the spectral signature of periodic signals is a
series of discrete components (i.e., “harmonics”), while that of random signals is a
continuous density. This provides good hope to achieve their separation.
Two scenarios must be considered: the case when the periods of the periodic
components to be extracted are known and the more difficult situation when they
are unknown.

6.7.1 Extraction of Periodic Signals with Known Periods: Synchronous

Averaging
The scenario is described by a periodic signal of interest, xT (t) = xT (t + T), in the
model of Eq. (30). A natural solution to extract xT (t) from the noisy measurement
y(t) is from the synchronous average:

1 K−1
x̂T (t) = y (t + kT ) (36)
K k=0

where K is the number of periods of xT (t) observed in y(t). Since the average of Eq.
(36) is synchronized on period T, it is constructive for the T-periodic component
but destructive for the noise part. Obviously x̂T (t) tends to xT (t) as the number of
available periods K becomes large. The synchronous average of Eq. (36) has an
equivalent interpretation in terms of a comb filter, a filter whose frequency gain is
unity at multiples of the frequency f = 1/T and which tends to zero (in inverse
proportion to K) elsewhere.
In practice the application of synchronous averaging is rarely performed directly
on the time domain sampled signal due to small speed fluctuations or slight
imperfection in the selection of the sampling frequency. In fact, if the reference
shaft speed is not perfectly constant and the sampling frequency is not its exact
multiple, the time domain signal will not have the same number of samples per
revolution, thus making impossible the application of the synchronous average. This
issue is solved by the transformation of the signal from time to angular domain, thus
obtaining a constant sample rate in samples/revolution. The process most commonly
used for this purpose is computed order tracking, discussed in Sect. 8.1

6.7.2 Extraction of Periodic Signals with Unknown Periods

This scenario is a particular case of the blind extraction discussed in Sect. 6.5 and
relies on the definition on a good marker that singles out a periodic signal (without
4 Applied Digital Signal Processing 177

Fig. 7 Principle of the blind extraction of a periodic signal: a linear filter is learned that optimally
(in the mean square sense) maps the signal to its advanced version, which naturally tunes itself into
a comb filter

the knowledge of its period) from random noise. This is provided by the property of
a periodic signal to be perfectly predictable in the future from its partial observation
in the present in contrast to random noise which is not predictable farther than its
correlation length. This means there exists a linear filter which exactly maps xT (t) to
xT (t + ) for any  > 0. Such a filter can be constructed according to Eq. (33) where
xT (t) is seen as the reference of xT (t + ). Now, provided that the delay  is taken
greater than the correlation length of noise n(t), the filter will naturally tune itself
into a comb filter with unit gain at the harmonic frequencies and zero gain elsewhere.
Its application to measurement y(t) will then return an estimate of all the periodic
components it comprises, as illustrated in Fig. 7. One advantage of this approach (in
addition to not requiring the knowledge of the periods) is that it can extract multiple
periodic components at the same time without requiring their periods to be coded on
an integer number of samples; however the corresponding comb filter is generally
less sharp than the one obtained from the synchronous average (given an equivalent
signal length).

6.8 Blind Source Separation

Blind source separation (BSS) is more stringent than signal extraction in that it tries
to separate simultaneously all the components that make up the observed signal in
Eq. (30). For this, the components are seen as the responses to a limited number of
underlying sources – sometimes also called the “latent variables” – as illustrated in
Fig. 8. The adjective “blind” indicates that the ambition is to achieve the separation
without any reference signal or knowledge of the transfer functions that relate the
underlying sources to their measured contributions.
BSS is a difficult problem which was first solved theoretically in the 1990s;
since then, works on the subject have grown exponentially, initially in the field of
telecommunications and more recently in other disciplines such as biology, physics,
and mechanics.
178 R. B. Randall et al.

Fig. 8 Blind source separation is to recover the sources si (t) s or their individual contributions
xij (t) from a set of simultaneous measurements yj (t)

6.8.1 The Notion of a Source

To be a well-posed problem, BSS requires the notion of a source to be carefully
defined. The physical understanding of a source is a signal linked to an excitation
mechanism in the system under study; different sources refer to excitation mecha-
nisms of different physical origins. In a rotating machine, for instance, sources will
typically relate to excitations in the motor, in the gears, in the bearings, etc. Note
that a source does not necessarily have a point-like nature; indeed, its exact spatial
distribution is irrelevant as long as a model of the form of Eq. (35) holds.
From a statistical point of view, the fact the sources originate from different
physical origins is reflected by their mutual independence. This means that the
knowledge of one source does not help at all to predict the value of the other sources.
Mutual independence implies in particular – but not only – that sources are mutually
uncorrelated, i.e., their cross-spectra are all nil.

6.8.2 Problem Statement

The rationale behind BSS is that a set of measurements is produced by the
superposition of the responses to a limited number of sources. The objective is
to recover the original sources or, to a lesser extent, their individual contributions
in the measurements. The contribution of a source is defined as the signal that
would be measured as if all other sources were switched off. This is illustrated in
Fig. 8 where BSS is interpreted like the decomposition of a multiple-input multiple-
output (MIMO) system into a superposition of single-input multiple-output (SIMO)
systems.
4 Applied Digital Signal Processing 179

The difficulty in BSS is that all sources are necessarily operating concurrently
and that only mixtures of their contributions can be observed. This is the situation
typically encountered with machine signals, which are produced by a superposition
of responses to multiple sources of excitation which are activated by the operation
of the machines. The separation of competing sources usually requires multiple and
simultaneous measurements.
The BSS problem can be formulated rather generally as:

⎧
⎪
⎪ y1 (t) = h11 {s1 (t)} + · · · + h1N {sN (t)} + n1 (t)
⎪
⎪      
⎪
⎪
⎨ x11 (t) x1N (t)
.. (37)
⎪ .
⎪
⎪
⎪ yM (t) = hM1 {s1 (t)} + · · · + hMN {sN (t)} + nM (t)
⎪
⎪
⎩      
xM1 (t) xMN (t)

where the yi (t) ’s, i = 1, . . . M refer to M simultaneous measurements, the sj (t) ’s,
j = 1, . . . N to N competing sources, and the ni (t) ’s to additive noise. The operator
hij { . . . } denotes the effect of the transfer on source sj (t) to measurement yi (t), which
may take different forms as seen below. Equation (37) clearly indicates that the set
of measurements is completely explained by a limited number of underlying sources
whose contributions are shared by all channels plus possible additive noise which,
by definition, is independent throughout the channels.
Depending on the applications, the objective of BSS is either to recover the
original sources sj (t) ’s or solely their contributions xij (t) = hij {sj (t)} ’s. The second
case is obviously less ambitious than the first one and is actually sufficient in many
instances.
It is important to realize that all quantities are unknown in Eq. (35) except
for the measurements, the yi (t) ’s. Due to the many unknowns to be solved, it is
therefore advantageous to have the number M of measurements as large as possible
as compared to the number N of sources. The over-determined case M > N is
recommended in the presence of significant additive noise. This justifies a major
difference between signal extraction, which usually proceeds with a single channel,
and BSS, which usually requires multiple simultaneous measurements.

6.8.3 Types of Mixture

The BSS problem strongly depends on the type of mixture described by Eq. (37).
A fairly general configuration is when the operator hij { . . . } is a convolution, i.e.,
xij (t) = hij (t) * sj (t) – the so-called convolutive mixture – yet it is also the most
difficult case to solve. The simplest configuration is when the operator hij { . . . } is a
static gain, i.e., xij (t) = hij · sj (t); this is the so-called linear instantaneous mixture
which is conveniently expressed in matrix form as:

y(t) = H s(t) + n(t) (38)

180 R. B. Randall et al.

with y(t), s(t), and n(t) column vectors that contain the elements yi (t) ’s, si (t) ’s, and
nj (t) ’s, respectively, and H a matrix whose element (i, j) is hij .
It is noteworthy that the convolutive mixture can be transformed into an
“instantaneous” mixture in the Fourier domain.

6.8.4 General Principles

Once the type of mixture has been specified, BSS proceeds in three steps: the design
of a separation operator, the definition of a separation criterion, and the construction
of an optimization algorithm. These steps are briefly reviewed in the case of the
simple instantaneous mixture.

Separation Operator
The principle of BSS is to find a separation operator – a separation matrix W in the
linear instantaneous mixture – such that

ŝ(t) = W y(t) (39)

returns an estimate ŝ(t) of the sources s(t). Note that W is generally not the (pseudo)
inverse of H (i.e., such that WH = I, the identity matrix) since it will also try to
reduce the effect of additive noise n(t) to some extent. It is important to note that
the recovered sources are arbitrarily scaled and sorted. Indeed, since H is unknown,
it can exchange any column and gain with W (e.g., WH = (k W)(k−1 H), thus
leading to another separation matrix k W and another source estimate kŝ(t)). The
unknown amplitude and sorting of the sources are fundamental indeterminacies
of BSS. However, the former indeterminacy is actually fixed when estimating
the contribution of a source. Once an estimate is available on a source si (t), its
contribution xij (t) to the measurement yj (t) can be estimated from the reference-
based approach of Sect. 6.3, i.e.:
   
yj (t)ŝi (t) yj (t)y(t)T wTi
x̂ij (t) =    ŝi (t) =   wi y(t) (40)
ŝi (t)2 wi y(t)y(t)T wTi

where . . . denotes the time-average operation, wi the i-th row of matrix W, and
. . . T the transpose of a vector. It is clear from Eq. (40) that contribution x̂ij (t) is
properly scaled, even if the estimated source ŝi (t) is not. It is also noteworthy from
this equation that BSS is to be interpreted as a spatial filter.

Separation Criteria
The main challenge of BSS is to blindly estimate the above separation operator from
the output measurements only. Similarly to blind extraction and blind deconvolution
previously discussed in Sects. 6.5 and 6.6, this requires strong a priori assumptions
4 Applied Digital Signal Processing 181

about the statistical properties of the sources. The separation operator is thus sought
so as to enforce recovered sources ŝ(t) with the expected properties. Some typical
separation criteria are listed hereafter:

• Mutual decorrelation of the sources. Because of their statistical independence,

sources are mutually uncorrelated, i.e., with zero cross-spectra. In contrast,
correlation is introduced between the measured signals in Eq. (38) because of
the mixture of sources. The idea of BSS is to design a separation matrix W
that recovers sources whose correlation is minimized. This may be seen as an
extension of principal component analysis to signals. The working assumption
of this approach is that sources must have distinct spectra.
• Mutual decorrelation of cyclostationary sources. When sources are cyclostation-
ary, they are also expected to have zero cyclic cross-spectra. This provides more
constraints into the problem to better estimate the separation matrix, in particular
in the presence of strong stationary background noise. This separation criterion
is particularly well suited to machine signals which are often well described as
cyclostationary.
• Mutual independence of sources. Going one step further, the idea is to enforce
not only the source decorrelation but also their mutual independence. Here
again independence is lost in the mixing in Eq. (38) and will not be recovered
unless the correct separation matrix is found in Eq. (39). This is achieved by
requiring some non-linear measure of statistical dependence, such as higher-
order cross-correlations or cross-entropy, to be zero. Again this may be seen
as a generalization of principal component analysis referred to as independent
component analysis (ICA). The working assumption of this approach is that no
more than one source in the mixture has a Gaussian probability distribution.
• Maximum kurtosis of sources. In the case of sources with an impulsive
behaviour, their mixture tends to average out this characteristic. More generally, a
mixture of leptokurtic sources (i.e., with high kurtosis) tends to become Gaussian
(i.e., with zero kurtosis) by virtue of the central limit theorem of probabilities.
Therefore, a separation matrix is sought that restores maximum kurtosis of
the sources. Similar popular criteria have been devised based on minimizing
measures of entropy.
• Intermittent sources. In some instances, the sources occur intermittently in the
time domain so that there exist time intervals where only one source is active at
a time. This makes possible to grasp this source as a reference and then estimate
a separation matrix from classical reference-based techniques. Other approaches
directly proceed with masks in the time-frequency domain.
• Sparse sources. Sparse sources are a generalization of intermittent sources which
are described (or well approximated) by only a few non-zero coefficients in a
given transformed domain, such as the Fourier or wavelet domains. The objective
is to seek a separation matrix that effectively returns sources which can be
represented as sparsely as possible.
182 R. B. Randall et al.

Optimization Algorithm
Estimating a separation operator by enforcing a given statistical property of the
sources is achieved by means of an optimization algorithm. Several toolboxes are
nowadays available that implement the most popular BSS algorithms.

6.9 Blind Separation of Structural Modes

6.9.1 Context
A successful application of BSS in structural dynamics is for the separation of modal
contributions from output-only data. The aim is to decompose a set of simultaneous
measurements of a MIMO system into the individual contributions of underlying
SIMO systems. When the sources are represented by the modal coordinates of a
structure, BSS boils down to the separation of each modal contribution, that is,
the response of the structure when only one mode is active. In other words, the
response of a complex multiple-degree-of-freedom is then decomposed into the sum
of responses of single-degree-of-freedom systems. This finds considerable interest
in Operational Modal Analysis (OMA) where modal parameters can then easily
be identified from single-degree-of-freedom responses, without controlling the
excitation. The working assumption – which is common to all OMA approaches –
is that the excitation to the structure is broadband enough so that its spectrum can be
considered constant across the bandwidth of any modal resonance of interest. This
assumption is found to apply reasonably well in several instances, in particular under
ambient excitations where loading on the structure is produced by fluid-dynamic
forces (wind, waves), natural seismic activity, traffic, etc.
As compared to other identification algorithms dedicated to OMA (e.g., stochas-
tic subspace identification), BSS may be used as a useful pre-processing step or a
substitute which instead allows the use of simple degree-of-freedom identification
techniques.

6.9.2 Principle
The rationale beyond the application of BSS to OMA proceeds from the similarity
between the linear instantaneous mixture of BSS (see Eq. (39)) and the modal
expansion theorem which states that the response of a structure to any excitation
can always be expanded into a linear combination of modes weighted by the
modal coordinates. Namely, by denoting φi the i-th mode shape (a vector), qi (t)
the corresponding modal coordinate (a signal), and y(t) the vector that stacks the
responses of the structure measured at different points,


y(t) = φ qi (t) + n(t) (41)
i i

where n(t) stands for possible additive noise. This can be further cast in matrix form
as:
4 Applied Digital Signal Processing 183

y(t) = q(t) + n(t) (42)

where matrix  contains the active mode shapes in columns and q(t) stands for the
vector of modal coordinates. Comparing Eq. 42 with Eq. (38), it is seen that the
modal matrix  plays the role of the mixing matrix H and the modal coordinates
q(t) play the role of the sources s(t) in BSS. There is therefore good hope that BSS
can be applied to Eq. (42) in order to estimate jointly the modal matrix  and the
modal coordinates q(t) from the output-only measurements y(t).
In order to do so, the assumptions of BSS must first be repeated. Section 6.8.4
introduced several separation criteria usually used in BSS. Among them, the
criterion of mutual decorrelation of the modal coordinates has been found to apply
generally well in the case of lightly damped structures with low modal density. The
reason is that under these assumptions the spectra of the modal coordinates have
nearly disjoint supports so that their mutual correlations are indeed very small. The
extreme case is given by a conservative system where the modal coordinates then
look like pure sines: they are therefore exactly uncorrelated as soon as they oscillate
at different natural frequencies. In practice, the criteria of mutual decorrelation has
been found to be quite robust even in the case of modal overlap factors slightly
greater than one and damping factors on the order of 20%.
Several algorithms are conceivable to separate modal contributions based on their
mutual decorrelation. A popular one is SOBI (second-order blind identification)
which forces decorrelation at several time lags by jointly diagonalizing a set of
correlation matrices. As explained in Sect. 6.8, the output of a BSS algorithm
comprises the separation matrix W which, in the present context, provides an
estimate of the inverse of the modal matrix . Therefore mode shapes can be
recovered in the columns of the inverse of the separation matrix, W−1 . Similarly,
application of the separation matrix to the measurements returns an estimate of the
modal coordinates, q̂(t) = Wy(t), from which global modal properties (natural
frequencies and damping ratios) can be easily estimated by using single-degree-of-
freedom techniques.
A necessary condition for this strategy is to have at disposal at least as many
sensors as active modes to recover. If not satisfied from the onset, this condition
can easily be forced by applying BSS in frequency bands that contain no more
active modes than available sensors. A frequency domain version of BSS can be
implemented on this basis which forces joint decorrelation at several frequency
bins – i.e., nil cross-spectra – in a given frequency band.

6.9.3 Example of Application

The use of BSS for the separation of modal contributions is illustrated here on data
provided in the Operational Modal Analysis Modal Parameter Estimation Round
Robin organized at the IMAC XXVII 2011 conference. The structure under study is
a scaled model of a wind turbine blade. The structure is fixed at the root and excited
by means of random tapping for duration of about 5 min. The response is measured
in all 3 directions at 16 locations at a sampling rate of 512 Hz. The power spectra
184 R. B. Randall et al.

Fig. 9 (a) Power spectra vibration responses of a scaled wind turbine blade (16 channels).
(b) Power spectra of separated modal responses in four sub-bands

of the 16 responses in the frequency range of interest up to 200 Hz are displayed

in Fig. 9a. It is seen that apart from the first three or four modes which clearly pop
up below 50 Hz, the structure quickly exhibits a “mid-frequency behaviour” with
significant damping and a modal overlap factor that becomes greater than 1 above
100 Hz. As explained above, these observations depart to some extent from the
theoretical assumptions of mutually uncorrelated sources used here as a criterion
for BSS.
In order to cope with the large number of modes up to 200 Hz, the frequency band
of interest is divided into 4 bands which are expected to contain less than 16 active
modes each, so as to comply with the condition of having more sensors than modal
contributions to recover. The power spectra of the separated modal coordinates in
each band are displayed in Fig. 9b (here a frequency domain version of BSS based
on mutual decorrelation has been used). It is seen that six dominant modes could
be separated in the first band from 0 to 68 Hz, up to eight modes in the band from
68 Hz to 122 Hz (the last one in the band actually pertains to the next band), up
to ten modes in the band from 122 Hz to 170 Hz and at least seven modes in
the last band. It is remarkable that the algorithm has been able to achieve good
separation not only in the first band characterized by a clean modal behaviour but
also in the higher-frequency bands characterized by “mid-frequency behaviour.”
4 Applied Digital Signal Processing 185

Table 2 Identified modal parameters and modal overlap factors

Damp. Damp. Damp.
Freq. (Hz) (%) MOF Freq. (Hz) (%) MOF Freq. (Hz) (%) MOF
7.33 1.19 1% 106.52 1.13 119% 155.50 0.85 77%
27.27 1.35 5% 107.64 0.91 56% 158.89 1.01 121%
38.94 1.71 15% 113.55 0.94 42% 160.81 0.85 80%
44.85 0.78 13% 117.76 0.89 57% 165.78 0.89 47%
49.73 1.38 19% 120.91 1.00 68% 173.46 0.99 58%
59.38 1.25 21% 124.90 1.09 68% 177.56 0.97 101%
63.90 1.22 17% 128.87 0.86 38% 180.29 1.22 125%
77.27 0.81 11% 136.73 0.87 37% 184.58 1.00 92%
86.46 1.04 22% 141.62 0.72 47% 188.35 1.10 114%
93.23 0.86 43% 145.41 0.70 48% 191.84 1.63 302%
93.91 0.91 33% 150.17 0.83 75% 192.50 0.96 121%
103.62 1.01 33% 152.07 0.75 86% 197.92 1.21 89%

From these separation results, the natural frequencies and damping ratios could
be easily estimated by using single degree-of-freedom techniques (more advanced
curve-fitting techniques or subspace algorithms could also be used at this stage).
The values are displayed in Table 2, along with the estimated modal overlap factor.
The recovered mode shapes finally made possible the identification of combinations
of flexure and torsion of the wing.

7 Time-Frequency Representations

This material is based largely on the discussion in Ref. [9].

7.1 STFT

A relatively simple, but powerful, way of dividing up a signal in both time and
frequency is to move a fixed window function, for example, a Hanning window,
along a slowly varying time record, usually in overlapping steps, and record the
autospectrum of the corresponding windowed time signal for each displacement of
the window. This is illustrated in Fig. 10.
The formula for the Fourier spectrum of each windowed section is:
 ∞
S (t, f ) = x (τ ) w (τ − t) exp (−i2πf τ ) dτ (43)
−∞

where the H(t, f ) in Fig. 10 is equal to |S(t, f )|, and if the squared amplitude spectrum
is plotted, the diagram is usually called a “spectrogram.” The window overlap
186 R. B. Randall et al.

Fig. 10 Generation of STFT

is chosen to give a smooth visual transition and is smaller than the actual time
resolution.
The length of weighting function w(t) determines the frequency bandwidth and
resolution of each spectrum and must therefore be chosen carefully. For example, in
Fig. 10 it should be chosen to be shorter than the spacing of the impulse responses
(IRs) dominating the signal, but longer than the individual IRs to give meaningful
results, but this of course is data dependent. The product of resolution in time
and resolution in frequency is a constant, this being one aspect of the Heisenberg
uncertainly principle.

7.2 Wigner-Ville Distribution (WVD)

The Wigner-Ville distribution (WVD) seems to violate the uncertainty principle in

appearing to give better resolution than the STFT, but suffers from interference
components, with possible local negative values, between the actual components.
The original Wigner distribution [10] was modified by Ville [11] who proposed
the analysis of the corresponding analytic signal so as to eliminate interference
between positive and negative frequency components. The WVD is one of the so-
called Cohen’s class of time-frequency distributions [12], most of which have been
proposed to improve on the WVD in some way. Even the STFT falls into this class.
Cohen’s class may be represented by the formula:

Cx (t, f, φ) = Fτ →f {R (t, τ )} (44)

4 Applied Digital Signal Processing 187

where R(t, τ ) is a weighted autocorrelation-like function defined by:

∞  
τ τ
R (t, τ ) = x u+ x∗ u − φ ((t − u), τ ) du (45)
2 2
−∞

and φ(u, τ ), with φ((t − u), τ ) = w(t − u − τ /2)w(t − u + τ /2) being a kernel
function used to smooth the WVD (with φ((t − u), τ ) = δ(t − u), the WVD is
obtained). The “pseudo Wigner-Ville distribution” is a finite windowed version
of the WVD, with φ((t − u), τ ) being a function of τ only, and the “smoothed
pseudo Wigner-Ville distribution,” with φ((t − u), τ ) being a separable function in
(t − u) and in τ , suppresses interference in both the time and frequency directions.
The results of smoothing are data dependent but usually give better simultaneous
resolution in the time and frequency directions than the STFT.

7.3 Wigner-Ville Spectrum (WVS)

For second-order cyclostationary signals (CS2), with no admixture of CS1 (i.e.,

after removal of periodic signals), it is possible to remove interference components
by indirect averaging. A standard way of analyzing cyclostationary signals is to first
generate a two-dimensional autocorrelation function, with lag time τ on one axis
and cyclic time t on the other, according to the formula:

 
Rxx (t, τ ) = E x (t + τ/2) x ∗ (t − τ/2) (46)

where the difference from Eq. (45) is the ensemble averaging given by the
expectation operator. As described in Ref. [13], this can be transformed in both time
lag and cyclic time directions to give the spectral correlation diagram, to give a 3D
diagram with normal frequency f on one axis and cyclic frequency α on the other. If
the Fourier transformation is done first in the τ direction, the Wigner-Ville spectrum
(WVS) is obtained, giving frequency distribution vs time t. Since the interference
components have random phase for CS2 signals, they average to zero, so that the
WVS has the same resolution as the WVD, but with interference components largely
removed. This may appear to violate the uncertainty principle, but it should be kept
in mind that much longer data records are required to perform the averaging. By
contrast, averaging STFT diagrams does not improve resolution, but does smooth
noise spectra.
Figure 11 compares the results of performing STFT, WVD and WVS on the
vibration signal from a reciprocating compressor, with many separate events with
different frequency contents occurring at different times (i.e. crank angles) in the
compression cycle. This shows the improved resolution of the WVS compared with
the STFT, without the interference components of the WVD.
188 R. B. Randall et al.

Fig. 11 Diagnostics of a reciprocating compressor. (a) WVS. (b) STFT. (c) Pressure on forward
and backward strokes. (d) Accelerometer signal. (e) WVD for one cycle. (a) Forward stroke,
(b) backward stroke. 1, 2, Opening, closing of discharge valve. 3, 4, Opening, closing of suction
valve. (From [13])

7.4 Wavelets

Another approach to time-frequency analysis is to decompose the signal in terms of

a family of “wavelets” which have a fixed shape but can be shifted and dilated in
time. The formula for the wavelet transform is:

∞  
1 ∗ (t − b)
W (a; b) = √ x(t)ψ dt (47)
a a
−∞

where ψ(t) is the mother wavelet, translated by b and dilated by factor a. Since this
is a convolution, the wavelets can be considered as a set of impulse responses of
filters.
4 Applied Digital Signal Processing 189

7.4.1 Constant Percentage Bandwidth

Displacement by b results in a linear time axis, but scaling by a, often by factors of
2, tends to give an octave-based (constant percentage bandwidth) frequency scale,
and the decomposition is known as time-scale rather than time-frequency.
Wavelets can be orthogonal or non-orthogonal, and continuous or discrete [14].
Examples of orthogonal wavelets are the Daubechies dilation wavelets [15], which
are compact in the time domain but in principle infinite in the frequency domain.
They tend to have irregular shapes in the time domain. Newland [14] describes
complex harmonic wavelets, which are compact in the frequency domain but infinite
in the time domain. They have the appearance of windowed sinusoids (harmonic
functions) and are typically of one octave bandwidth, although they can be narrower.
The advantage of complex wavelets is that the imaginary part of the wavelet is
orthogonal to the real part (sine rather than cosine) and thus the overall result is not
sensitive to the position (phasing) of the event being transformed (it may be centered
on a zero crossing of the real part, but this would be a maximum of the imaginary
part). The local sum of squares of the real and imaginary parts is a smooth function.
Wavelets are subject to the uncertainty principle, but because of the constant
percentage frequency scale give a better time localization at high frequencies, and
for that reason can be useful for detecting local events in a signal. Figure 12 shows
the difference in time-frequency localization for STFT and (octave-based) wavelets,
showing how the latter give fine time localization at some scales, whereas if this is
chosen for the STFT, it detracts from frequency localization. In all cases the areas
of the boxes are constant to comply with the uncertainty principle.
Many authors have described the use of wavelets for machine diagnostics,
including detecting local faults in gears and bearings, and Ref. [16] provides a good
summary.
Orthogonal wavelets are particularly useful for applications involving analy-
sis/synthesis, but not so important for pure analysis, where the information is
extracted directly from the wavelet diagram. For machine diagnostics, a very
important class of wavelets are Morlet wavelets, where the mother wavelet is a
cosine weighted with a Gaussian function. Complex Morlet wavelets exist, where
the cosine is replaced by a complex exponential function, whose imaginary part is
the Hilbert transform of the real part (sine rather than cosine). The advantage of the

Fig. 12 Time-frequency paving of (a) the STFT with a fine time resolution, (b) the TFT with a
fine frequency resolution, (c) the octave wavelet transform
190 R. B. Randall et al.

Gaussian function is that its Fourier transform is also a Gaussian function, so the
wavelet envelope is of the same form in both time and frequency domains.
A particular application involving analysis/synthesis is wavelet denoising, where
all components in the wavelet diagram whose amplitude is below a certain threshold
are considered to be noise. In so-called hard thresholding, all components thus
defined as noise are set to zero, but the retained components are left unchanged. In
“soft thresholding,” the noise estimate (the threshold value) is subtracted from them
also (symmetrical treatment of positive and negative values). Wavelet denoising is
based on the original work of Donoho and Johnstone [17]. This type of denoising
has the advantage that noise in time and scale directions is equally removed.

7.4.2 Wavelet Packets

Another way of considering wavelet decomposition at each scale is as a division into
details and approximation components, by a highpass and lowpass filter, respec-
tively. The shortest possible scale components (highest frequency) are obtained,
for example, as the mean of adjacent pairs of samples (each frequency component
requires at least two samples). When the shortest scale components are subtracted,
the approximation contains all longer scale components, but this can again be
decomposed into detail and approximation components at the next scale, and so
on. This leads to the normal octave-based frequency scale of the wavelet transform.
With wavelet packets, even the upper octave at each scale is separated by highpass
and lowpass filters, so that a linear frequency axis is divided into 2, 4, 8, 16, and
other bands, resulting in a constant bandwidth frequency resolution at each scale.
It could be said that wavelet packets are equivalent to a series of STFTs with time
window lengths changing in 2:1 steps. Thus, there seems little justification for them
in machine diagnostics.
Much of the extensive theory developed for the wavelet transform can be carried
over to wavelet packets.

8 Specialized Analysis

8.1 Order Analysis

Many machine vibration signals can be considered phase-locked with the rotation
of a reference shaft. If small speed fluctuations of the shaft occur, the signals will be
cyclostationary in the angular domain of the shaft, rather than in time domain. For
instance, a CS1 vibration signal x(t) of this type will show a periodic expected value
in the shaft angular domain θ , rather than in time domain:

E {x (θ )} = E {x (θ + 2kπ )} (48)

Fourier analysis is therefore much more effective on these signals if performed in

the angular domain. Resulting spectral quantities are therefore defined in the shaft-
order domain rather than the traditional Hertz-frequency domain f. Orders are
4 Applied Digital Signal Processing 191

expressed in times-per-revolution of the reference shaft used for the definition of

the angular domain θ . The following subsections will describe some methodology
to obtain the transformations t → θ and f → .
Synchronous averaging is only meaningful if carried out in the shaft angular
domain, as practised over several decades for gear diagnostics.

8.1.1 Computed Order Tracking

Computed order tracking is aimed at transforming the signal from time to angular
domain. This procedure is often found as a preliminary step of synchronous
averaging algorithms (see dedicated section), in order to ensure a uniform number
of samples for every revolution of the reference shaft. Order tracking implements
a series of interpolations on the shaft rotation measurement (often a 1 × rev
tachometer) and on the vibration signal resulting in a resampled vibration signal
whose sample rate is defined (and constant) in the angular domain of the shaft, rather
than in time. The first interpolation works on each shaft rotation k and estimates the
times tk [m] at which the shaft is in the uniformly spaced angular positions 2π m/P
(P the new sample rate in samples per shaft revolution). Considering a 1 × rev
tachometer and a piecewise-linear approach interpolation, this operation can be
written as:
m
tk [m] = tk + tk+1 − tk (49)
P

where tk is the k th 1 × rev tacho reference pulse and tk [m] is therefore estimating
the times corresponding to the angle 2π k + 2π m/P. A uniformly angular resampled
signal (i.e., order tracked) can be obtained by a second interpolation of the original
signal x(nt) at the times tk [m]:

∼ ∼
  m
x (rα) = x 2π k + ≈ x (tk [m]) (50)
P

McFadden [18] showed that cubic spline interpolation is considerably better than
linear interpolation from two considerations:

1. The interpolation gives a lowpass filtration of components approaching half the

sampling frequency, and cubic spline cuts off at a higher frequency than linear,
which in turn cuts off higher than quadratic interpolation.
2. Some aliasing of high-order sidelobes is unavoidable but is much lower for cubic
spline than quadratic interpolation, which in turn is better than linear.

8.2 Second-Order Cyclostationary Indicators

Cyclostationary signals are characterized by a periodic non-stationarity. Detec-

tion of the presence and quantification of cyclostationarity is of interest in sev-
eral applications, either because it is symptomatic of a particular phenomenon
192 R. B. Randall et al.

(e.g., a mechanical fault) or because it allows advanced processing (e.g., signal

extraction/separation – see related subsections). In many instances, cyclostationarity
cannot be revealed by classical signal processing tools and thus requires specific
approaches. This is in particular the case for second-order cyclostationary signals, a
wide class of considerable practical interest characterized by a periodic autocovari-
ance function (see Sect. 3.3.2).
An illustrative example of a second-order cyclostationary signal is provided by
a (zero-mean) stationary random noise, r(t), modulated by a periodic envelope,
pT (t) = pT (t + T) (example already introduced in the Sect. 3.3.2):

x(t) = pT (t)r(t) (51)

Both the frequency characteristics of random carrier and those of the modulating
signal are therefore inherited by the CS2 signal. Figure 13 reports an example of the
hybrid nature of CS2 signals.

• On a short-time scale, the effect of the PSD of the random carrier r(t) is
reproduced: low frequency in case of Fig. 13 (a) and high frequency in case of
Fig. 13 (b).
• On a long-time scale, the period/frequency of the periodic signal pT (t) is
reproduced in the cyclic flow of power of the signal: higher frequency in case
of Fig. 13 (a) and lower frequency in case of Fig. 13 (b).

This double-frequency characterization of CS2 signals is critical in their separa-

tion and identification. The frequency axis of the short-time PSD characteristic is
usually described by the symbol f and referred to as spectral frequency (associated
to the properties of the random parent signal r(t)), while the long-term periodicity
in the signal’s power inherited from the modulating signal pT (t) is usually described
by the symbol α and referred to as cyclic frequency.
Comparing Fig. 13a, b and Fig. 3b, c, it is possible to intuitively recognize
the correspondence between the cyclic frequency α and the time variable t in the
autocorrelation function C(t, τ ) and the equivalent relationship τ - f. An example of
such CS2 signal is the random vibration of a flexible shaft modulated by its rotation
(α linked to shaft rotational speed). In this case the signal will also show specific
f characteristics, depending on the shaft resonances. Moreover, as in most practical
cases, an additional stationary noise n(t) will be present (e.g., to reflect the presence
of other vibration sources).

y(t) = pT (t)r(t) + n(t) (52)

This model well describes also the typical vibrations obtained from a faulty
rolling element bearing: with the characteristic spectral frequency (f ) corresponding
to the structural resonances excited by each impact of the rolling elements with the
fault and the phenomenological cyclic frequencies (α) linked to the time lag between
subsequent impacts (see simplified numerical example of Fig. 14).
4 Applied Digital Signal Processing 193

Fig. 13 Two numerical examples of CS2 signals generated by multiplying a random station-
ary carrier and a periodic modulating signal: (a) random carrier with low-frequency PSD
(fa = 0 − 100 Hz) and mono-harmonic modulation frequency α a = 2 Hz, (b) random carrier with
high-frequency PSD (fa = 400 − 500 Hz) and mono-harmonic modulation frequency α a = 1 Hz

Figure 15 displays one realization of another CS2 signal in time domain (a)
and the same signal with the addition of stationary background noise (b). The
corresponding power spectral densities are shown in Fig. 16 (a). A few observations
are noteworthy. Signal x(t) clearly evidences a random behaviour, yet with a periodic
modulation; this may be interpreted as a periodic flow of its energy along the
time axis with cyclic frequency α. At the same time, the presence of this “hidden
periodicity” (additional α-axis) is completely missed in the classical power spectral
density (see Fig. 16 (a)) which essentially reflects the continuous distribution of the
random carrier (f -axis description) and the absence in the signal of any additive
sinusoid. Besides, the presence of cyclostationarity is not detectable in the noisy
194 R. B. Randall et al.

Fig. 14 Simplified numerically generated bearing signal

Fig. 15 (a) Realization of a second-order cyclostationary signal (sampling frequency Fs = 1 Hz)

made of a narrowband stationary random carrier modulated by a periodic amplitude. The signal
envelope (magnitude of the analytical signal after bandpass filtering in band [0.05Fs 0.15Fs ]) is
displayed in red. (b) Same signal after addition of stationary random noise (SNR = -9 dB). Note
that the cyclostationary behaviour of the signal is completely masked in (b)

version of the signal, even in the time domain. This asks for the use of specific CS2
tools, which will be investigated in the following sections.

8.3 The Envelope Spectrum

The idea of the squared envelope spectrum is to reveal the presence of a periodic
flow of energy in a cyclostationary signal. The flow of energy is estimated by
computing the squared envelope of the signal, which is then Fourier transformed
to evidence a harmonic structure.
As exemplified by Fig. 15, it is important to carefully select the frequency band
that carries the periodic modulation in order to extract it from additive noise which
4 Applied Digital Signal Processing 195

Fig. 16 (a) Power spectral density of signals in Fig. 15a (red curve) and in Fig. 15b (black curve)
(frequency resolution = 5 × 10−4 Hz). (b) Squared envelope spectrum of signal in Fig. 1b after
bandpass filtering in band [0.05Fs 0.15Fs ] (frequency resolution = 1 × 10−5 Hz). In contrast to the
power spectral density which displays a continuous distribution, the squared envelope spectrum
clearly reveals a harmonic structure. Note that correctly bandpass filtering the signal in the band
where the SNR is maximized is crucial

may possibly mask it otherwise. This is achieved by a preliminary bandpass filtering

of the signal. The squared envelope is then conveniently computed as the squared
magnitude of the analytic signal. The algorithm to compute the squared envelope
spectrum is summarized in Fig. 17. Its application to the signal of the previous
subsection is illustrated in Fig. 16b. It is seen that cyclostationarity is easily detected
even though it was barely visible in the time domain due to the presence of strong
additive noise.
It is eventually remarkable to notice that the key point of the squared envelope
spectrum is to introduce a quadratic non-linearity: squaring a second-order cyclosta-
tionary signal produces additive sinusoids which can then be detected and quantified
by classical spectral analysis.

8.4 The Cyclic Modulation Spectrum

The success of the squared envelope spectrum largely depends on correctly selecting
the frequency band where to demodulate the signal. When the latter is not known, its
selection by trial and error may be a tedious and hazardous task. A simple solution
196 R. B. Randall et al.

Fig. 17 Steps for computing

the squared envelope
spectrum

is to display a cascade of squared envelope spectra at the output of a filter bank

that decomposes the signal into a set of sub-band components [19]. This defines
the cyclic modulation spectrum, which is displayed as function of the modulation
frequency α – also called cyclic frequency – and the carrier frequency f, also called
spectral frequency. A simple interpretation of the cyclic modulation spectrum (and
an efficient computational approach) is actually as the Fourier transform of the
spectrogram (the squared magnitude of the short-time Fourier transform) along the
time variable (see Fig. 18).
A significant value of the cyclic modulation spectrum at a given pair (α, f )
thus indicates the presence of a periodic flow of energy with frequency α in the
frequency band centered at f. In the case of a broadband cyclostationary signal,
the cyclic modulation spectrum evidences a symptomatic set of vertical lines at
discrete values of the modulation frequency and continuous values of the carrier
frequency. Following a numerical example analogous to Figs. 3 and 13, the CMS
of a CS2 signal xCS2 (t) composed by the product of a stationary random carrier
r(t) and a periodic modulation pT (t) is reported in Fig. 19f. In this case the peak at
α = 4 Hz and f = 0 − 100 Hz shows a second-order cyclostationary component
whose energy is mainly concentrated in the 0–100 Hz band and cyclically fluctuates
with a frequency of 4 Hz. As clearly visible comparing Fig. 19b, c, d, this CMS
peak shows:

• An f -axis content which corresponds to the PSD of the carrier r(t)

• An α-axis content which corresponds to the Fourier series of the squared periodic
modulation |p(t)|2

The same concepts are illustrated for the more complex signal of Fig. 15b
in Fig. 20a. In this case, the presence of cyclostationarity is clearly evident in
4 Applied Digital Signal Processing 197

Fig. 18 (Left) Schematics of the procedure to calculate the CMS using the spectrogram, (right)
corresponding initial intermediate and final results for a numerically generated CS2 signal

the frequency band from 0.05 Hz to 0.15 Hz which carries the modulation and
with cyclic frequency at 1/200 Hz. The number of cyclic harmonics reflects the
“roughness” (non-sinusoidal characteristic) of the modulation, and, as explained in
Sect. 3.3.2, it is actually related to the time domain of the signal.
One limitation of the cyclic modulation spectrum – and thus of the squared
envelope spectrum – is that it cannot detect modulations that are faster than the
bandwidth set in the filter bank: fast modulations will require a filter bank with
large bands, thus with a coarse spectral resolution; in contrast, a fine spectral
resolution in the cyclic modulation spectrum implies slow modulations. This is a
direct consequence of the uncertainty principle of time-frequency analysis.
This has consequences in practice. For instance, when trying to detect a bearing
fault, the expected signal is in the form of a series of repetitive transients with mean
period T0 (as explained in 5.2.3, this signal is close to cyclostationary rather than
periodic due to random fluctuation in the period and possibly in the amplitude). The
bandwidth of the cyclic modulation spectrum should then be set large enough so as
to include the “hidden harmonics” of the fault, that is, larger than 1/T0 .

8.5 The Spectral Correlation Density

The limitation inherent to the cyclic modulation spectrum can be bypassed by

defining the spectral correlation density, a fundamental tool of second-order cyclo-
198 R. B. Randall et al.

Fig. 19 Relationship between the cyclic modulation spectrum of a numerically generated CS2
signal with the spectral characteristics of its random carrier and modulating periodic signal.
(a) Stationary carrier r(t), (b) PSD of the carrier PSDr (f ), (c) squared
 value
 of periodic
modulation pT (t) (red), (d) Fourier series of the squared modulation F |pt (t)|2 , (e) CS2 signal
xCS2 (t) = r(t)pT (t) (blue) and modulation pT (t)(red dashed),
 (f) CMS of xCS2 (t) (surface plot) in
comparison with PSDr (f ) (black dotted) and F |pt (t)|2 (red dotted)

stationary processes. The introduction of the spectral correlation density requires a

more formal approach, based on the covariance function of the signal. As mentioned
in Sect. 3.3.2, the covariance function of a second-order cyclostationary signal is
a periodic function of time (see Eq. (22)); taking the example of Sect. 3.3.2, it
writes:

Cx (t, τ ) = m(t)m (t + τ ) Cε (τ ) = Cx (t + T , τ ) (53)

The squared envelope spectrum and its extended version, the cyclic modulation
spectrum, essentially investigates the periodicity of Cx (t, 0) (after filtration in
frequency bands), which reflects the flow of energy as a function of time. The
spectral correlation density considers the complete covariance function (not only
for τ = 0) through the double Fourier transform.
4 Applied Digital Signal Processing 199

Fig. 20 (a) Cyclic modulation spectrum and (b) spectral coherence density of signal in Fig. 13
(b) (frequency resolution f = 8 × 10−3 Hz; cyclic frequency resolution = 1 × 10−5 Hz). The
cyclic modulation spectrum has a cyclic frequency range on the order of f (a strong attenuation
is seen after 2 f ), whereas the spectral coherence density has no such limitation. The spectral
coherence density is normalized between 0 and 1 and usually shows a better contrast. The
presence of cyclostationarity is evidenced in the band [0.05Fs ;0.15Fs ] and with a fundamental
cyclic frequency of 1/200 Hz
200 R. B. Randall et al.

Sx (α, f ) = Ft→α Fτ →f {Cx (t, τ )} (54)

According to the t-periodicity of Cx (t, τ ), the spectral correlation density of

a second-order cyclostationary signal is expected to be discrete in the cyclic
frequency and continuous in the spectral frequency. The relationship between the
two frequency axes and the two time domain variables is crucial in understanding
the properties of CS2 signals and their bispectral representation. Referring to the
acoustic example already mentioned in Sect. 3.3.2, the characteristic high-energy
f band(s) of the signal corresponds to the “pitch” of a CS2 sound, while the
characteristic α frequency (or frequencies) indicates the rhythm with which the
sound is repeated. Note also that the spectral correlation density at α = 0 is
equivalent to the power spectral density of the signal, Sx (f ) (as it indicates the
“t-average” component of Cx (t, τ )). As a consequence, the contribution of stationary
noise is also completely and only confined to α = 0. Interestingly, the spectral
correlation density can be equivalently defined by means of the Fourier transform
XD (f ) of the signal on a time interval of duration D:

1  ∗

Sx (α, f ) = lim E XD (f )XD (f − α) (55)
D→∞ D

(where E stands for the expectation operator and * for the conjugation) which
provides a practical algorithm of estimation. This definition gives evidence that
spectral components spaced apart by the modulation frequency are correlated.
In other words, although the spectrum of a cyclostationary signal does not give
evidence of harmonics in general, the presence of hidden periodicities in it manifests
itself by a correlation of its spectral components. Therefore, a natural tool to
quantify cyclostationarity is the spectral coherence density (in all points similar to a
correlation coefficient):

Sx (α, f )
γx (α, f ) = √ (56)
Sx (f )Sx (f − α)

which returns a measure between 0 and 1. The spectral coherence density is

particularly useful to disclose weak cyclostationary signals in the presence of strong
coloured noise because it is independent of the actual magnitude of the signal and,
when coming to estimation issues, its variance is constant in the whole (α, f ) plane.
Refs. [20, 21], respectively, detail the theoretical foundation and efficient imple-
mentation of a fast algorithm for the calculation of the spectral correlation.

8.6 The Relationship Between Kurtosis, Envelope, and CS2

Indicators

The bispectral CS2 indicators described in the previous section expose the full
nature of CS2 signals: while α “sets the tempo” of the long-term cyclic pulsation of
4 Applied Digital Signal Processing 201

energy, the f -axis describes the velocity of the short-term random signal variation
constituting the signal power itself.
The traditional squared envelope spectrum (calculated with a broad pass-band)
is actually equivalent to an f -integral estimation of the spectral correlation, i.e., the
SES is representing the cumulated effect of a series of horizontal (α-parallel) slices
of the spectral correlation Sx (α, f ):
 fhigh
SESx α; flow , fhigh ≈ Sx (α, f ) df (57)
flow

where SESx (α; flow , fhigh ) represents the squared envelope spectrum obtained after
bandpass filtering with pass-band [flow , fhigh ]. Any peak in the squared envelope
spectrum SESx (α; flow , fhigh ) will therefore represent an (α-)periodicity in the cyclic
flow of energy carried by frequencies within the range [flow , fhigh ]. This sheds
some light on the true cyclostationary meaning of envelope analysis: the explicit
α-axis represents periodicity in the second-order moment of the signal (power,
autocorrelation), while the implicit dependency on f (through the preliminary
bandpass filtering operation) reflects the τ -dependecy of the second-order moment
Cx (t, τ ). When choosing a filter for the envelope analysis operation, the pass-
band therefore defines the components of the signal whose energy pulsation is
analyzed. In the example of Fig. 20, the best envelope analysis result (clearest peak
at α = 0.005 Fs ) is thus obtained using the bandpass filter 0.05–0.15 Fs before
enveloping, since this band is the one whose energy is actually fluctuating with a
period of 2000/Fs .
This relationship can be further extended to kurtosis by integrating the squared
envelope spectrum over the remaining α-axis [22]. In fact, using Parseval’s theorem
it is possible to demonstrate that for a zero-mean signal x(t):

SESx (α) = RMSx 4 · κx (58)
α

where SESx (α) is the SES calculated using the full band (no bandpass filtering) and
RMSx and κ x are the RMS and (sample) kurtosis of the signal. Moreover, since the
fourth power of the RMS is equivalent to the null cyclic frequency SES, the (sample)
kurtosis is equivalent to:

α SESx (α)
κx = (59)
SESx (0)

This equivalence allows a cyclostationary interpretation of the kurtosis, which is

equivalent to the sum of all the different α-contributions of the full f -spectrum SES,
normalized by the signal’s power. Since the SES of Eq. (57) is performed using the
raw signal, it is equivalent to a full-band integration of the spectral correlation, and
consequently the (sample) kurtosis is equivalent to a full (α, f )-plane integration
of the spectral correlation. These results show how kurtosis, envelope analysis,
202 R. B. Randall et al.

and advanced spectral correlation estimators are all aiming, with different levels of
specificity, at the identification of the cyclic variation of the autocorrelation function
(and power) characteristic of CS2 signals.
The strong relationship between CS2 and fourth-order statistics highlights an
unwanted sensitivity of CS2 indicators to impulsive noise, which has been shown
to potentially mask CS2 components [23]. A series of techniques have been
proposed to deal with this issue, usually including fractional-order or logarithmic
transformations of the envelope [24, 25].

9 Typical Diagnostic Examples

In this section a number of examples are given which draw on combinations of

techniques introduced in earlier sections. They demonstrate the wide diversity of
problems that can arise in machine diagnostics, and how they can be tackled.

9.1 Rolling Element Bearings

The first example illustrates a case where the speed of the machine is constant,
but the bearing signal is completely masked by stronger signals from gears, right up
until just before failure. The example was first published in [26], but most details are
repeated in [27]. A series of signals were taken from a helicopter gearbox test rig
at DSTO (Defence Science and Technology Organisation), Melbourne, Australia,
as it was run to failure. The signals were analyzed blind to detect the source of the
failure(s). Direct comparison of raw time signals and their spectra could detect no
change from the start to finish of the test. The time signal just before failure is shown
in Fig. 21a and has a kurtosis of −0.61, effectively the same as Gaussian noise.
The analysis procedure first applied was linear prediction, as a means of removing
the discrete frequency components from the gears and leaving the bearing signal
in the residual. The kurtosis increased from −0.61 to 2.2, but primarily because
some modulation was now apparent at the rate of the planet pass frequency; the
bearing signal was not detectable. A wavelet kurtogram was then used to determine
the optimum band for demodulation, and that is shown in Fig. 21b. By filtering
the residual signal with the optimum band (1/12-octave centered on 18.8 kHz),
the kurtosis increased to 14.1, and the bearing fault impulses can now be seen in
Fig. 21c.
Frequency analysis of the squared envelope of the processed signal in Fig. 21c in
two different frequency ranges gave the envelope spectra in Fig. 22.
The harmonics in Fig. 22a at 9.83 Hz corresponded to the cage frequency
(relative to the load zone) of the planet gear bearings in the planetary section of
the gearbox. This indicates a fault travelling at cage speed, and this is often a
sign of faulty rollers, though often accompanied with components at the roller
spin frequency (not found in this case). When the gearbox was dismantled, three
4 Applied Digital Signal Processing 203

a 10
50 3 b 9

Number of filters /octave

0 8
–50 7
6
6
5
5 c
12 4
3
0
2
24 1
–5 0
5 10 15 20
Frequency [kHz]

Fig. 21 Signal processing stages for a bearing fault in a helicopter gearbox (from [26]). (a)
Original signal, kurtosis −0.61. (b) Wavelet kurtogram. (c) Processed signal, kurtosis 14.1

adjacent rollers in one bearing were found to have some spalling, and this group
was apparently modulating the signals as they passed through the load zone.
The harmonics in Fig. 22b at 117.7 Hz corresponded to the inner race fault
frequency in the same planet gear bearings, and when the faulty bearing was
dismantled, severe spalling was found in the inner race of one planet bearing. It
should be noted that with planet bearings, it is the inner race that is fixed with respect
to the load direction, so no modulation sidebands were expected or found.
Even though this result is from near the end of life, the trend of the kurtosis of the
processed signal showed that the faults could have been detected about 30 h before
the end of a 160 h test.

9.2 Gears

A standard way of analyzing gear signals is to start with TSA (time synchronous
averaging), which in general extracts the signal for each gear from the total signal,
including those of other gears. Even when the gearbox is running at nominally
constant speed, it is necessary to first carry out order tracking and resampling of
the signal, to ensure that there are an integer number of samples per rotation of
the gear, and a fixed starting phase (e.g., the exact time of a once-per-rev tacho
pulse, which may occur between two of the original samples). Even though the
order tracking removes speed variations, it is generally still necessary to resample
for each successive shaft speed so as to satisfy the requirement of a fixed number of
samples per rev.
The TSA result will generally allow visualization of the vibration response of
each tooth on the gearmeshing with the mating gear, and changes can often be
recognized in both amplitude and variability. The regular gearmeshing pattern can
however disguise some local disturbances, so a standard method of revealing these
is to remove that pattern by two alternative basic procedures:
204 R. B. Randall et al.

Fig. 22 Envelope spectra in two different frequency ranges (from [26]). (a) Low frequency,
showing a series of harmonics of 9.83 Hz. (b) High frequency, showing a series of harmonics
of 117.7 Hz

• Remove harmonics of the gearmeshing frequency in the frequency domain and

transform back to the time domain. Since the primary aim is to detect local faults
on the gears, it has been found advisable to remove slow modulations as well
by removing one or even two pairs of sidebands around each harmonic as well.
Since only the information of one gear is left in the TSA signal, these sidebands
will also be harmonics of the rotational speed of the gear.
• Use linear prediction to predict and subtract the regular toothmeshing pattern,
but leave the unpredictable sudden changes which deviate from this pattern.
4 Applied Digital Signal Processing 205

(a) (b) (c) Cracked tooth

0 360 0 360 0 360

Rotation angle (degrees)

Fig. 23 Amplitude and phase demodulation of a gearmesh signal (from [28]). (a) Original TSA
signal. (b) Amplitude demodulation. (c) Phase demodulation

The linear prediction filter can be adjusted to remove slow changes but retain
sudden ones.

The following example is one where the TSA signal did not immediately reveal
a fault, and neither did the residual after removing the toothmesh signal.
The actual fault was a tooth root crack in a gear in a helicopter gearbox, and the
method proposed to reveal it in [28] consisted in performing a demodulation of the
toothmesh component, both amplitude and phase demodulation. Figure 23 shows
the original TSA signal and the results of amplitude and phase demodulation of the
(second harmonic of) gearmesh frequency. The second harmonic was demodulated
because it was stronger than the first.
Even though the TSA signal is not completely uniform, there is no definite
indication of a fault, and the same applies to the amplitude modulation signal. The
phase modulation signal does however reveal a sudden deviation in one location,
which corresponded to the location of the tooth root crack. At a later stage in
the development, the amplitude modulation signal also revealed the crack, but the
phase deviation was the first indication. Note that this method only works when the
fault manifests itself within the maximum frequency band for demodulation, which
corresponds to the gearmesh frequency. In this case the demodulation band was
from 1.5 to 2.5 times the gearmesh frequency. In [29] a case is reported where a
tooth root crack on a very slow component in a wind turbine gearbox (gearmesh
frequency 30 Hz) manifested itself most strongly around 11 kHz, and was not
detectable by either synchronous averaging or toothmesh demodulation. It could,
however, be detected using spectral kurtosis.
Further examples of gear diagnostics using the cepstrum are given in Sect. 10.3.

9.3 Reciprocating Machines and Engines

Since reciprocating machine signals tend to be a series of impulsive events, and

thus vary rapidly in both time and frequency, even at constant speed, a common
way of analyzing them is by some form of time-frequency analysis (TFA). A large
component is usually second-order cyclostationary, since individual combustion
cycles have considerable random variation but repeat cyclically. This means that a
206 R. B. Randall et al.

very powerful form of TFA is the Wigner-Ville spectrum (see Fig. 11), where cross
terms are eliminated for CS2 components, and better simultaneous time-frequency
resolution is possible than by a normal spectrogram (better than the uncertainty
principle, by virtue of the extra data required for averaging).
There are a couple of other fault types, however, which can be diagnosed by
relatively simple techniques, but enhanced considerably by using simulation models
of the machine.
The first is combustion faults in internal combustion (IC) engines, which can
be detected by changes in torsional vibrations. The latter can be measured very
simply by frequency demodulation of a shaft encoder signal, and that can be as
simple as a tooth-pass signal from a proximity probe detecting passage of the
teeth on the engine flywheel. For engines with a rigid crankshaft (lowest torsional
resonance frequency above a number of times the engine firing frequency), this
is very straightforward as the torsional acceleration from a given torque pulse
is independent of the cylinder from which it originates and insensitive to speed.
However, as shown in [30], a relatively simple torsional vibration model of a
flexible crankshaft can be made with lumped parameters and then updated using
a small number of actual measurements. Waveforms are very sensitive to actual
natural frequencies, so the model updating involved allowing the natural frequencies
and damping to adjust to give optimal matching of the waveforms, while using
the analytical eigenvectors (mode shapes) from the simple analytical model. This
was successful on two nominally identical engines with quite different waveforms.
Figure 24 compares the simulated waveform with the measured one for one engine.

Fig. 24 Simulation of the angular speed variations at the crankshaft free end (continuous line)
compared to the actual measurement (dashed line). (Courtesy M. Desbazeille)
4 Applied Digital Signal Processing 207

Very little “faulty” data was available, but an artificial neural network (ANN)
trained on outputs from this simulation model was able to detect the two faults fed
to it.
Reference [31] used a similar principle to train neural networks for a much wider
range of combustion faults purely using simulated data, but the engine was smaller
and had a rigid crankshaft. On the other hand, another simulation model was also
generated to simulate the rotational motions of the engine on its supports due to the
same combustion faults, and this has similar characteristics to the torsional model
of the flexible crankshaft, since the engine mounting resonances were constant
and within the frequency excitation range of the combustion faults, meaning that
responses varied with the faulty cylinder and engine speed. The ANNs trained
purely on simulated data in both cases were 100% successful in diagnosing the
fault severity and location.
The same principles were used to train ANNs to diagnose mechanical faults
(piston slap, bearing knock) in the same IC engine and were also 100% successful
[32, 33]. For the mechanical faults, the features used to train the ANNs were the
amplitude and phase of various harmonics of the cycle frequency of the engine,
measured on the squared envelope of an optimally filtered band, quite analogous to
bearing faults (since these signals were CS2).

10 Cepstrum

10.1 Background and Definitions

The cepstrum was first proposed in 1963 [34], being defined as the “power spectrum
of the log power spectrum.” The original application was to the detection of echo
delay times, it being much less sensitive than the autocorrelation function (inverse
Fourier transform of the power spectrum) to the color of the signal. The reason for
the definition was apparently that it was published 2 years before the FFT (though
with a common author, Tukey) and so software was readily available for power
spectra (via the autocorrelation function) but not for complex Fourier transforms.
Shortly after the publication of the FFT, the “power cepstrum” was redefined as
the inverse Fourier transform of the log power spectrum, which meant that it was
reversible to the (log) power spectrum after editing in the cepstrum. At about the
same time, work by Oppenheim and Schafer led to the definition of the “complex
cepstrum,” which retains the phase information in the log spectrum, so that it is
reversible to a time function after editing in the cepstrum. This can only be done
for spectra whose phase can be unwrapped to a continuous function of frequency.
Given a signal x(t), most important definitions are now as follows:
Complex cepstrum:

Cc (τ ) = F −1 {log (F {x(t)})} = F −1 {log (Ax (f )) + iφx (f )} (60)

208 R. B. Randall et al.

where Ax (f ) and φ x (f ) are, respectively, the amplitude and phase of the discrete
Fourier transform of the signal x(t):

F {x(t)} = Ax (f )· eiφx (f ) (61)

Power cepstrum:
  
Cc (τ ) = F −1 {log (Sx (f ))} = F −1 log |F {x(t)}|2 = F −1 {2 log (Ax (f ))}
(62)

The so-called real cepstrum is sometimes defined by setting the phase term in
Eq. (60) to zero, which is seen to be a scaled version of Eq. (62).
In the original paper [34], the authors coined the word “cepstrum” by reversing
the first syllable of “spectrum,” the justification being that it was a “spectrum of
a spectrum” (although the autocorrelation is also a spectrum of a spectrum, just
without the log transformation). Similarly, the word “quefrency” was obtained from
“frequency,” and the authors also suggested a number of other terms, including
“rahmonic” from “harmonic” and “lifter” from “filter,” and these four terms are
still used in the cepstrum literature. Note that because the original cepstrum was not
reversible, a lifter had to be applied as a convolutive filter to the log spectrum, but it
is now applied as a window in the cepstrum.
Using z-transforms to replace Fourier transforms for digitized signals, Oppen-
heim and Schafer showed [35] that the complex cepstrum of a general transfer
function
 M0
B Mi 1 − ai z−1 i=1 (1 − bi z)
H (z) = Ni=1 N
(63)
−1
i=1 1 − ci z i=1 (1 − di z)
i 0

could be expressed in terms of its poles and zeros in the z-plane as:

Ch (n) = log(B) , n=0

 ain  cin
Ch (n) = − i n + , n>0 (64)
 bin  dinn
Ch (n) = i n − n , n<0

where the ci and ai are poles and zeros inside the unit circle, respectively, and the di
and bi are the (reciprocals of the) poles and zeros outside the unit circle, respectively.
Since the cepstrum of minimum phase functions is thus causal, the log amplitude
and phase of the log spectrum are related by a Hilbert transform, meaning that the
phase does not have to be measured or unwrapped, and the complex cepstrum can
be obtained from the real cepstrum by setting negative quefrency components to
zero and doubling positive quefrency components. Many simple mechanical transfer
functions are minimum phase, and it is shown in [36] that the cepstrum of each pole
term can be expressed as:
4 Applied Digital Signal Processing 209

cn c∗n |c|n  in arg(c) e−σ nt

Csdof (n) = + = e + e−in arg(c) = 2 cos (n arg(c))
n n n n
(65)

where t is time sample spacing (so that t = nt) and σ is the damping constant
corresponding to the exponential decay |c|n . Zeros have the same form, but negative,
and maximum phase zeros are similar but at negative quefrency. Maximum phase
poles are not viable for stable systems.
The cepstrum thus has at least three useful properties for analysis of mechanical
systems:

• Echoes give a periodic structure to the log spectrum (log amplitude and phase),
resulting in discrete rahmonics (spaced at the echo delay time) in the cepstrum.
This was the original application and allows accurate measurement of the delay
time, and even echo removal.
• Periodic structures in the log spectrum, such as families of equally spaced
harmonics and sidebands, also give corresponding rahmonics in the cepstrum
(of which the first two or three are the most important), facilitating detection and
evaluation of such families and accurate measurement of their spacing.
• For single-input multiple-output (SIMO) systems, the output at each response
point is a convolution in the time domain, a product in the frequency spectrum,
a sum in the log spectrum and equally a sum in the cepstrum, often allowing
separation of forcing and transfer functions in the response cepstrum. Multiple-
input multiple-output (MIMO) responses have to be separated into a sum of
SIMOs, e.g., by blind source separation, before this cepstral separation can be
applied.

Exploitation of these properties is demonstrated in the following sections.

10.2 Cepstrum Liftering

Liftering (editing) in the cepstrum is done for four primary purposes:

• To remove uniformly spaced rahmonics in the cepstrum with a (comb) notch

lifter, so as to remove the corresponding families of harmonics or sidebands in
the (log) spectrum. Note that there must be a minimum of about eight uniformly
spaced components in the spectrum for there to be distinct rahmonics in the
cepstrum.
• To apply an exponential weighting to the cepstrum as a “shortpass” lifter to
emphasize the modal characteristics of the transfer path (while adding a small
known amount of damping to each mode), at the same time as removing many
forcing functions and disturbances at higher quefrency.
210 R. B. Randall et al.

• Possibly by subtraction of this modally weighted part of the cepstrum, to retain

the part of the signal dominated by higher quefrency forcing functions, with
suppression of modal information. So-called cepstral pre-whitening is an extreme
example of this where all cepstrum components except that at zero quefrency
are set to zero to make the spectrum completely white. This not only removes
modal information but also sets discrete frequency components to a fixed level
throughout the spectrum.
• To apply a “longpass” lifter at quefrencies below the modal ones, to remove the
effects of broadband excitations (impulsive or broadband random) at even lower
quefrencies, and to further enhance the modal characteristics in the absence of
higher quefrency excitations.

10.2.1 Comb Notch Lifter

As described in [37], there are two types of comb notch lifters, depending on the way
in which the rahmonics deviate from the ideal of a series of delta functions, which
would only be the case if the corresponding families of harmonics/sidebands had
uniform strength over the whole of the (two-sided) log spectrum, with the sampling
frequency being an integer multiple of the harmonic/sideband spacing. For Type 1,
the width of the notch remains constant over the whole range of quefrency, while for
Type 2, the width increases in proportion to quefrency. The first applies where there
is a fixed windowing of the spectral components or amplitude modulation, while the
second applies when the width of the spectral components increases with frequency,
as caused by small random frequency modulation of a basic periodicity. Note that
removal of such components cannot be achieved by synchronous averaging, which
assumes that all harmonics are discrete frequencies.
Figure 25 shows how a Type 1 lifter can be applied to the complex cepstrum to
remove echoes from a time signal, even when these overlap. This is made easier
by the fact that the cepstrum of the basic signal (in this case an SDOF response) is
shorter than the impulse response itself (Eq. (65)).
Figure 26 shows an example where a series of groups of sidebands spaced at
120 Hz (around non-commensurate carrier frequencies) have been removed by a
Type 1 lifter (with notch width ±15% of the rahmonic spacing). This is another case
where synchronous averaging cannot be used because it only removes harmonics.
The width of the notch was determined by trial and error but corresponds to the
reciprocal of the average number of sidebands in each group (6–8). Note that
the lifter was applied to the real cepstrum in this case, but time signals can be
regenerated for each record by combining the modified amplitude spectrum with
the original phase spectrum [36].
Further examples are given in Sects. 10.3 and 10.4.

10.2.2 Exponential Lifter

Since every pole (and zero) has the basic form of Eq. (65), it can be seen that
multiplying the cepstrum of a response signal by an exponential window e−σ0 t
will add (constant) damping σ 0 rad/s to every pole (and zero) term. This can in
principle be compensated for in the same way as an exponential window applied to
4 Applied Digital Signal Processing 211

Fig. 25 Echo removal using a notch lifter in the complex cepstrum. (Courtesy Brüel&Kjær)

the response to a hammer blow in modal analysis, but is not limited to transients. It
can be applied to each record of a continuous signal, once again in the real cepstrum
but with regeneration of time signals using the original phase spectrum of each
record.
Figure 27 shows the application of this approach to a signal from a gas turbine,
with excitation by multiple harmonics of the various shaft speeds. It is seen that
most of these are at high quefrency, and removed by the exponential lifter, so that
the liftered spectrum is dominated by the modal information of the dominant transfer
paths. The time constant of the lifter was made approximately equal to that of the
lowest frequency mode at around 2 kHz, with corresponding decreasing effect on
higher-frequency modes.
Further examples are given in Sects. 10.3 and 10.4.

10.2.3 Modal Suppression

If the residual cepstrum after subtraction of an exponentially windowed cepstrum
is used, it will be dominated by the high quefrency excitation components such as
shaft harmonics arising from gears, etc. and will be unaffected by the passage of
such components through resonances. This is useful for variable speed machines
as demonstrated in Sect. 10.3. Cepstrum pre-whitening was first done using the
cepstrum but can equally be achieved by dividing the (complex) spectrum of
each record by its modulus, without using the cepstrum. Even though discrete
frequency components are reduced to the same strength as adjacent noise, phase-
 212

Fig. 26 Removal of groups of sidebands by notch lifter in real cepstrum. (From [37])
R. B. Randall et al.
 4 Applied Digital Signal Processing

Fig. 27 Use of an exponential shortpass lifter to enhance modal information (from [36]). (a) Full cepstrum (of (d)). (b) Exponential lifter. (c) Liftered cepstrum.
(d) Original spectrum. (e) Liftered spectrum
213
214 R. B. Randall et al.

coherent harmonics and sidebands in a particular frequency band can make that
band dominant in terms of impulsiveness of time signals, with uniform weighting
over all frequency and no masking by particular resonances. Cepstral pre-whitening
has also been found valuable for variable speed machines as shown in Sect. 10.3.3.

10.2.4 Longpass Lifter

If an impulse such as a hammer blow is appreciably shorter than the period of an
excited natural frequency, its cepstrum will also be shorter, and the cepstrum of the
response will be dominated by the modal properties for all quefrencies above that
of the excitation. In an equivalent way, if the bandwidth of a broadband random
excitation is appreciably broader than the 3 dB bandwidth of a mode, its cepstrum
will be shorter and not interfere with the cepstrum of the modal properties. A
“longpass lifter” can then be applied in the cepstrum of the response to remove the
excitation without greatly affecting the cepstrum of the structural transfer function.
This condition is often described as the log spectrum being relatively “smooth and
flat,” although it may still be far from white, in particular on a linear amplitude scale.
The longpass liftered cepstrum can then be curve-fitted for the modal properties of
the transfer function, as described in Sect. 10.4.

10.3 Cepstrum for Machine Diagnostics

Machine diagnostics, or condition monitoring, is based on the analysis of machine

vibration responses, to determine any change in condition. Most often, such changes
are indicated by changes in the forcing functions, such as unbalance, misalignment,
gearmeshing forces, and impacts caused by bearing faults, but occasionally the
change in condition is because of a change in the structural properties, such
as a developing crack, and the cepstrum can be useful for separating these two
possibilities, as well as assisting diagnosis by giving accurate indications of forcing
and response frequencies, as expanded in the following sections.
Reference [36] is an extensive survey and history of the uses of cepstrum analysis
for machine diagnostics.

10.3.1 Changes due to Forcing Functions

Many faults in machine components, such as gears, give rise to series of impulsive
responses which give large numbers of harmonics in the frequency domain and quite
often families of sidebands due to modulation by these impulsive forces. Gear faults
can be divided into those which are uniform, such as uniform wear spread over all
teeth, and those which give deviations around the mean errors, such as local tooth
root cracks and spalls. The former give changes in a small number of harmonics of
the toothmesh frequency and are best detected in the spectrum, while the latter give
the sidebands which complicate the spectrum and are best detected and evaluated
using the cepstrum. Figure 28 (from [37]) shows an example where (log) spectra
and cepstra are compared for a gearbox before and after a deterioration giving
both distributed wear and local spalls. The gearbox is a multi-stage wind turbine
 4 Applied Digital Signal Processing

Fig. 28 Comparison of spectra and cepstra for a gear set in original and deteriorated condition. (a, b) Spectra. (c, d) Cepstra. (a, c) Deteriorated. (b, d).
(Original (from [37]))
215
216 R. B. Randall et al.

gearbox, with an overall ratio of about 100, but careful choice of the spectrum
frequency range allows individual gear sets to be examined by centering on their
mesh frequencies.
Comparing the spectra in Fig. 28a, b, it is seen that there has been a considerable
increase in the harmonics of the high-speed gearmesh (HSGM), in particular of
the second which has increased from 80 to nearly 110 dB. There has also been a
considerable increase of all the harmonics of the intermediate shaft (IS) gearmesh
(highlighted by the harmonic cursor). This indicates uniformly distributed wear. The
spectrum of Fig. 28a, however, is complicated by the growth in multiple sidebands,
mainly spaced at the speed of the high-speed shaft (HSS) around the harmonics of
the HSGM. Comparing the cepstra in Fig. 28c, d gives a much clearer picture of
the sideband structures, where spacings at both the HSS and ISS are apparent in
deteriorated condition. Localized spalls were found on the HSS gear, which greatly
increased the sidebands with this spacing. This situation is complicated by the very
poor design of this gear set with an exact 4:1 ratio (88:22 teeth), and the two sets
of sidebands would be even better separated in the more normal case of a hunting
tooth design.
Figure 29 demonstrates the use of a comb notch lifter to remove one set of
sidebands to make the diagnosis clearer. Despite the exact 4:1 ratio, which meant
that every fourth rahmonic of the HSS coincided with a rahmonic of the ISS, the
remaining rahmonics of HSS were removed by a specially designed notch lifter
(Fig. 29b), and the effect on the spectrum can be seen by comparing Fig. 29c, d.
All HSS harmonics (including sidebands) have been removed from the spectrum,
leaving only the harmonics of the ISS, including those corresponding to the ISGM,
highlighted by a harmonic cursor. It is now evident that these do not form a sideband
structure around the gearmesh harmonics, meaning that the wear on this gear set is
uniform rather than localized. This was confirmed by inspection and ascribed partly
to the fact that this gear set was hunting tooth (23:82, with no common factors). This
means that an incipient fault on one tooth is smeared over all teeth on the mating
gear (and vice versa).
It was shown in [37] that the separation by this means gave a somewhat better
result than synchronous averaging, which was implemented by dividing the SA for
the IS shaft into the four revolutions of the HS shaft, averaging them, and then
subtracting the repeated average from the original. Moreover, the SA required initial
order tracking, not required for the cepstral liftering.

10.3.2 Changes due to Structural Response

In condition monitoring it is important to know whether a change in the response
is due to a change in the forcing function or a change in the structural dynamic
properties, and the cepstrum provides a simple way of doing this. For example,
an increase in the second harmonic of a gearmeshing frequency would normally
be taken as indicating an increase in uniform wear. However, it is possible that
a resonance frequency, originally above the toothmesh harmonic, has decreased
because of a growing crack in the shaft or casing and now amplifies it without
any change in the forcing function. The prognosis would be completely different.
 4 Applied Digital Signal Processing

Fig. 29 Removing the effects of the HSS by comb liftering in the cepstrum. (a, b) Cepstra. (c, d) Spectra. (a, c) Unmodified. (b, d) Liftered. (From [37])
217
218 R. B. Randall et al.

3 80
a b
75
2
70
65
1
60
0 55
50
–1
45
40
–2
35
–3 30
0 128 0 2500
Time (s) Frequency (Hz)

80
0.5 d 75
c
0.4
70
0.3
65
0.2
0.1 60
0 55
–0.1 50
–0.2
45
–0.3
40
–0.4
–0.5 35
30
0 128 0 2500
Time (s) Frequency (Hz)

Fig. 30 Suppression of modal information in a variable speed gearbox. (a) Original time record,
(b) log spectrum, (c) liftered spectrum after suppression of modal information, (d) regenerated
time signal lowpass filtered above 3×GM frequency. (From [38])

Applying an exponential shortpass lifter, as in Fig. 30, it is possible to obtain a

spectrum showing the modal frequencies very clearly and thus to see if they have
changed over time. In complicated cases it would even be possible to curve-fit the
natural frequencies using the methods described in Sect. 10.4.
The separation of the modal information like this can also be used for other
purposes, such as to enhance or suppress its effects, in particular with variable
speed machines. Bearing fault information is often carried by resonances, so it could
be an advantage to weight the spectrum with the modal shape before performing
order analysis (which smears modal frequencies in the shaft angle domain), the
latter being necessary to identify the order-related bearing fault frequencies. This
reduces at the same time the disturbing influence of unrelated shaft orders. The
exact opposite is the case for gear diagnostics, where the diagnostic information is
usually in forcing functions directly related to shaft speed. Order analysis removes
frequency modulation, allowing components to be identified on an order scale,
but does not remove amplitude modulation, such as might be given by gearmesh
4 Applied Digital Signal Processing 219

frequencies passing through resonances. Synchronous averaging of amplitude-

modulated signals is no longer meaningful.
Figure 30 (from [38]) shows an example of a signal from a wind turbine gearbox
over a period where the speed varied over a range of 30%. The raw signal in 6(a)
shows considerable amplitude variation with speed, and even after order tracking,
the amount of amplitude variation was unchanged.
The dB spectrum in (b) has a basic modal structure which has been removed in
(c) by subtracting its cepstrum from that of the total signal. The order-related com-
ponents now protrude from a basically flat base noise spectrum. Regenerating the
liftered time signal corresponding to (c) (using the original phase spectrum) gives
the result shown in (d), This signal was lowpass filtered below the third harmonic
of the highest gearmesh frequency in order to concentrate on information relating to
the gear signals. It is seen to be roughly stationary over most of its length, indicating
that the effect of gearmesh harmonics passing through resonances has largely been
suppressed. The reason for the localized amplitude modulation near the end of the
record is not known but suspected to be caused by acceleration/deceleration as the
speed varied more rapidly here. This would change the forcing function independent
of the transfer function. Even so it would be possible to choose a section with little
amplitude variation in order to perform synchronous averaging. In this case there
was no known fault.

10.3.3 Example of Cepstrum Pre-whitening for the Diagnostic

of Rolling Element Bearings in Variable Speed Conditions
An effective example of application of cepstrum pre-whitening is obtained in the
field of bearing diagnostics. The example reported in this section is obtained on a
bearing diagnostics test rig installed in the laboratories of the Queensland University
of Technology, Brisbane (Australia) [23]. A long shaft between the test bearing and
the gearbox ensures that the vibrations recorded on the gearbox (Acc#1) and on the
test bearing (Acc#2) are almost entirely uncoupled, so that vibratory effects of the
gearbox scarcely affect the bearing and vice versa. A faulty bearing (outer race spall)
was installed on the main shaft, with a theoretical bearing pass frequency outer of
approximately 4.85 times the rotational speed of the shaft (Fig. 31).
In order to show the effectiveness of the cepstrum pre-whitening procedure,
vibration measurements were taken with the two accelerometers during a speed
transient (from 1.2 Hz to 2.4 in 50 s) and summed together, so as to reproduce
the effect of a typical hybrid gearbox-bearing signal (Fig. 32).
The raw signal DFT and SES of the composite signal (Acc#1 + Acc#2) were of
course of scarce usefulness, given the smearing of the bearing fault symptoms along
the frequency axis, due to the high variations of the shaft speed (see Fig. 33a, b). The
signal was therefore subject to a two-step procedure of order tracking-synchronous
averaging and cepstrum pre-whitening.
The order tracking and synchronous averaging was aimed at transforming the
signal to the angular domain and then removing the shaft-synchronous components
(i.e., the harmonics of the shaft speed). The signal’s DFT (Fig. 33c) still showed
220 R. B. Randall et al.

2.5

Shaft speed [Hz]

2

1.5

1
0 10 20 30 40 50
Time [s]

Fig. 31 Speed profile during the acquisition

strong spectral components resulting in a “noisy” SES (Fig. 33d) which did not
clearly allow the identification of the BPFO component.
The final step of cepstral pre-whitening allowed to reduce this effect, by flattening
the signal’s spectrum (Fig. 33e) and therefore highlighting the information carried
by the phase of the spectrum. Therefore, the SES of the pre-whitened signal clearly
showed a peak in proximity of the theoretical BPFO (Fig. 33f).
To verify that the peak was actually representing the bearing fault frequency, the
SES obtained with the two-step pre-whitening method is compared in Fig. 34 with
the SES of the signal Acc#2 collected directly on the accelerometer (no effect of the
gearbox).

10.4 Cepstrum for Modal Analysis

Because of its possibility of separating forcing and transfer functions in response

signals, the cepstrum has the potential to be used for operational modal analysis
(OMA) in two ways:

1. As a complete stand-alone procedure

2. As a pre-processing technique for other OMA procedures

A full description of these procedures has recently been published in [3] and only
a summary is given here.

10.4.1 Full OMA Procedure

In contrast to some other OMA procedures, the cepstrum can retrieve both poles
and zeros in transfer functions which thus gives some of the information contained
in the residues of the different poles, as shown below.
The first step is to curve-fit the auto spectrum of the response for its poles and
zeros. If the excitation is impulsive and single input, with little extraneous noise,
this might be done directly, as described in the original publication [39].
4 Applied Digital Signal Processing 221

Acc#1

0 10 20 30 40 50
Time [s] 40 40.2 40.4
Time [s]
Acc#2

0 10 20 30 40 50
Time [s] 40 40.2 40.4
Time [s]
Acc#1 + Acc#2

0 10 20 30 40 50
Time [s]
40 40.2 40.4
Time [s]

Fig. 32 Experimental signals measured by the gearbox accelerometer (Acc#1, top), the bearing
accelerometer (Acc#2, middle), and composite signal obtained by superposition of Acc#1 and
Acc#2 (bottom). The axes are to scale

Most recently [3], it has been shown that the zeros can be obtained more accu-
rately from transmissibilities rather than response autospectra, with the advantages
that noise, which otherwise disguises zeros, is removed by averaging and, for SIMO
situations, that the transmissibilities are independent of the forcing function, as long
as it is broadband.
222 R. B. Randall et al.

(a) (b)

(c) (d)

(e) (f)

Fig. 33 (a, b) DFT and SES of the raw signal, (c, d) DFT and SES of the residual after order
tracking ad synchronous averaging (OT-SA), (e, f) DFT and SES after OT-SA and cepstrum pre-
whitening

If the frequency response function (FRF) is regenerated using only the identified
in-band poles and zeros, there are two missing pieces of information:

• An equalization curve is required to compensate for the effects of out-of-band

poles and zeros. This is dependent mainly on the relative number of poles and
zeros, however, and only weakly on their exact placement.
• An overall scaling factor, since this is mixed with the forcing function at zero
quefrency.
4 Applied Digital Signal Processing 223

OT-SA+CPW SES (Acc#1+Acc#2) OT-SA residual SES (Acc#2)

4.886 Nx
4.886 Nx

0 0
2 4 6 8 10 2 4 6 8 10
Order [Nx] Order [Nx]

Fig. 34 Comparison between SES obtained after OT-SA and CPW from the composite signal
Acc#1 + Acc#2 (left) with the SES of the bearing accelerometer signal Acc#2 (only OT-SA is
applied)

In [40] it was shown that the equalization curve could be found using “phantom
zeros” in a similar manner to the rational fraction polynomial technique [41], which
also used a pole-zero model. More recently it has been shown that it can more easily
be found by smoothing the dB difference between the FRF generated on the basis of
the in-band poles and zeros, and a reference FRF, which can either be from an earlier
measurement, an EMA measurement, or from a finite element (FE) model of the
structure. As mentioned, the equalization curve is insensitive to small changes such
as a slow change in condition, or in-service operating conditions. The smoothing
is required because of differences in actual positions of poles and zeros between
the updated and reference FRFs, giving peak notches in the difference curves. It has
recently been shown that the smoothing is best achieved by polynomial curve-fitting,
making sure that the regenerated curves include the negative (and zero) frequency
poles and zeros derived from the inferred support conditions and measured positive
frequency poles and zeros [3]. Such a smoothed equalization curve can also correct
for absolute scaling if the reference FRF is also correctly scaled. For example, an
FE model of a free-free structure can easily be arranged to have correct rigid body
inertial properties, which determine the zero frequency values of FRFs, and even
for constrained structures, whose zero frequency values are determined by static
stiffness, the FE model can be updated to give agreement on the first couple of
elastic modes and will then be accurate at zero frequency.
Figure 35 (from [42]) shows an example of the application of such an equal-
ization process to the OMA of a free-free beam excited by a shaker with a pink
noise excitation, the pink color having no influence on the equalization process.
A comparison is made between a driving point FRF, with the same number of
poles and zeros, and an end-to-end FRF with no zeros. In the former case, little
equalization is required because of the balance of numbers of poles and zeros out-
of-band, whereas the absence of zeros in the second case requires a maximum of
equalization. Note that only five of the six modes were curve-fitted. This example
224 R. B. Randall et al.

Fig. 35 Regeneration of FRFs using an equalization and scaling curve. (a, c, e) Driving point
measurement. (b, d, f) End-to-end measurement. (a, b) Regenerated FRF (blue) from in-band poles
and zeros vs reference (black). (c, d) Unsmoothed difference curve vs polynomially smoothed
equalization curves. (e, f) True (measured) FRF (black) vs equalized estimate (blue) for five modes.
(From [42])

used only the positive frequency poles and zeros, but it is shown in [3] that a better
result is achieved by including the negative frequency components.
Normally, for OMA the spectrum will be contaminated by extraneous compo-
nents, such as discrete frequencies, and it will be an advantage to suppress these by
applying a notch comb lifter, an exponential shortpass lifter, or both, beforehand.
In the general MIMO case, it would in principle be necessary to separate the
response to a single source by BSS methods. A particular case of this where the
cepstrum has advantages over other methods is where there is only one CS2 source
with a particular cyclic frequency, as described in [43]. The practical case proposed
there is a diesel railcar, where the excitation from the combustion is broadband but
cyclic at the engine firing frequency. It is shown in [43] that the cepstrum of the
cyclic spectrum at the cyclic frequency contains only the structural path information
4 Applied Digital Signal Processing 225

Fig. 36 (a) Spectral correlation of railcar body excited by a CS2 force (b) resulting OMA mode
shape from curve-fitting such cyclic spectra for the indicated cyclic frequency (compared with
EMA mode shape). (From [43])

from that particular source to each response point. This was illustrated by exciting
the car by a shaker using a CS2 force signal (burst random), which was measured
and used for experimental modal analysis (EMA) at the same time. Figure 36 shows
the spectral correlation of the response, and the cyclic spectrum at the excitation
frequency is highlighted. The figure also shows one of the extracted mode shapes
(of the floor) obtained both by OMA and EMA. The two estimates compare very
well.
However, it is likely only in specialized situations, such as the very complex
excitation signals from rotating machines, that the cepstral method of modal analysis
would have advantages over the conventional OMA methods.

10.4.2 Cepstral Pre-processing for Other OMA Procedures

There is thought to be considerably more potential for using cepstral methods for
pre-processing response signals prior to OMA, in order to enhance the modal part of
the response and remove other disturbances such as extraneous discrete frequency
excitations. In a case such as illustrated in Fig. 27, it is obvious that the application
of a simple exponential shortpass lifter has removed almost everything except the
modal properties of the transfer paths, equivalent to a white excitation.
When there are only a small number of periodic contaminating components to
remove, it has been shown that they can be isolated by synchronous averaging (SA)
and then subtracted [44]. In principle this requires transformation to the angle/order
domain for the removal, followed by transformation back to the time domain for the
modal analysis, although if speed variation is very small, the difference between
the frequency and order axes is primarily one of scaling. In [44] it was shown
that a combination of a comb notch lifter and exponential shortpass lifter was able
to achieve a similar result to the SA method, in removing the harmonics of rotor
226 R. B. Randall et al.

110
100 a
90
Shaft orders
80
70
60
dB

50
40
30
20
10
0
0 200 400 600 800 1000
Frequency (Hz)
60
b
50
40
dB

30
20
10
0
0 100 200 300 400 500 600 700 800 900 1000
40
30 c
20
dB

10
0
–10
–20
–30
–40
0 100 200 300 400 500 600 700 800 900 1000

Fig. 37 Use of an exponential lifter for OMA of responses to a simulated gear signal with speed
variation ±5%. (a) Response autospectrum. (b) Corresponding exponentially liftered spectrum. (c)
FRF obtained using the measured force. (From [45])

frequency from response measurement on a helicopter in steady flight, except that

the fundamental rotor frequency was interpreted as a mode. The time constant of the
exponential lifter could not be reduced enough to make it less than the rotor period.
The cepstral method could be done without order tracking however.
As soon as the speed varies a little more, there is likely to be amplitude modu-
lation which cannot be removed by SA, and then the cepstral method is superior.
In [45] it is shown that in a signal simulating the response to gearmeshing in a
gearbox with speed varying by ±5% (but with simultaneous low-level broadband
noise excitation), the application of a simple exponential lifter resulted in a response
spectrum very similar to the equivalent FRF, and a group of signals treated in this
way gave a very similar OMA result to that obtained when the gearbox casing was
excited by a pure broadband noise signal. Figure 37 shows a typical result from this
paper. Later improvements on this approach are shown in [3].

References
1. Abboud D, Baudin S, Antoni J, Rémond D, Eltabach M, Sauvage O (2016) The spectral
analysis of cyclo-non-stationary signals. Mech Syst Signal Process 75:280–300
4 Applied Digital Signal Processing 227

2. Abboud D, Antoni J (2017) Order-frequency analysis of machine signals. Mech Syst Signal
Process 87:229–258
3. Randall RB, Antoni J, Smith WA (2019) A survey of the application of the cepstrum to
structural modal analysis. Mech Syst Signal Process 118:716–741
4. Ferreira JLA, Balthazar JC, Araujo APN (2003) An investigation of rail bearing reliability
under real conditions of use. Eng Fail Anal 10(6):745–758
5. Antoni J, Randall RB (2002) Differential diagnosis of gear and bearing faults. J Vib Acoust
124:165–171
6. Antoni J, Randall RB (2003) A stochastic model for simulation and diagnostics of rolling
element bearings with localized faults. J Vib Acoust 125(3):282–289
7. Pennacchi P, Borghesani P, Ricci R, Chatterton S (2011) An experimental based assessment of
the deviation of the bearing characteristic frequencies. In: 6th international conference acoustic
and vibratory surveillance methods and diagnostic techniques
8. Borghesani P, Ricci R, Chatterton S, Pennacchi P (2013) A new procedure for using envelope
analysis for rolling element bearing diagnostics in variable operating conditions. Mech Syst
Signal Process 38(1):23–35
9. Randall RB (2011) Vibration-based condition monitoring: industrial, aerospace and automotive
applications. Wiley, Chichester
10. Wigner E (1932) On the quantum correction for thermodynamic equilibrium. Phys Rev
40:479–760
11. Ville J (1948) Theorie et applications de la notion de signal analytique. Cables Transmissions
2:61–74
12. Cohen L (1995) Time frequency analysis theory and applications. Prentice Hall, Englewood
Cliffs
13. Antoni J (2007) Cyclic spectral analysis in practice. Mech Syst Signal Process 21(2):597–630
14. Newland DE (2008) Wavelet analysis of vibration signals, Chapter 49. In: Handbook of noise
and vibration control. Wiley, New York
15. Daubechies I (1990) The wavelet transform, time-frequency localization and signal analysis.
IEEE Trans Inf Theory 36(5):961–1005
16. Peng ZK, Chu FL (2004) Application of the wavelet transform in machine condition
monitoring and fault diagnostics: a review with bibliography. Mech Syst Signal Process 18:
199–221
17. Donoho DL, Johnstone JM (1994) Ideal spatial adaptation by wavelet shrinkage. Biometrika
81:425–455
18. McFadden PD (1989) Interpolation techniques for time domain averaging of gear vibration.
Mech Syst Signal Process 3(1):87–97
19. Antoni J, Hanson D (2012) Detection of surface ships from interception of cyclostationary
signature with the cyclic modulation coherence. IEEE J Ocean Eng 37(3):478–493
20. Antoni J, Xin G, Hamzaoui N (2017) Fast computation of the spectral correlation. Mech Syst
Signal Process 92:248–277
21. Borghesani P, Antoni J (2018) A faster algorithm for the calculation of the fast spectral
correlation. Mech Syst Signal Process 111:113–118, Oct
22. Borghesani P, Pennacchi P, Chatterton S (2014) The relationship between kurtosis- and
envelope-based indexes for the diagnostic of rolling element bearings. Mech Syst Signal
Process 43(1–2):25–43
23. Borghesani P, Shahriar MR (2016) Cyclostationary analysis with logarithmic variance stabili-
sation. Mech Syst Signal Process 70–71:51–72
24. Borghesani P, Antoni J (2017) CS2 analysis in presence of non-Gaussian background noise –
effect on traditional estimators and resilience of log-envelope indicators. Mech Syst Signal
Process 90:378–398
25. Antoni J, Borghesani P (2019) A statistical methodology for the design of condition indicators.
Mech Syst Signal Process 114:290–327
26. Sawalhi N, Randall RB (2007) Semi-automated bearing diagnostics -three case studies. In:
Proceedings of COMADEM 2007, Faro, Portugal
228 R. B. Randall et al.

27. Randall RB, Antoni J (2011) Rolling element bearing diagnostics – a tutorial. Mech Syst Signal
Process 25:485–520
28. McFadden PD (1986) Detecting fatigue cracks in gears by amplitude and phase demodulation
of the meshing vibration. J Vib Acoust Stress Reliab Des 108:165–170
29. Barszcz T, Randall RB (2009) Application of spectral kurtosis for detection of a tooth crack in
the planetary gear of a wind turbine. Mech Syst Signal Process 23(4):1352–1365
30. Desbazeille M, Randall RB, Guillet F, El Badaoui M, Hoisnard C (2010) Model-based
diagnosis of large diesel engines based on angular speed variations of the crankshaft. Mech
Syst Signal Process 24:1529–1541
31. Chen J, Randall RB (2015) Improved automated diagnosis of misfire in internal combustion
engines based on simulation models. Mech Syst Signal Process 64–65:58–83
32. Chen J, Randall RB, Peeters B (2016) Advanced diagnostic system for piston slap faults in IC
engines, based on the non-stationary characteristics of the vibration signals. Mech Syst Signal
Process 75:434–454
33. Chen J, Randall RB (2016) Intelligent diagnosis of bearing knock faults in internal combustion
engines using vibration simulation. Mech Mach Theory 104:161–176
34. Bogert BP, Healy MJR, Tukey JW (1963) The quefrency alanysis of time series for echoes:
cepstrum, pseudo-autocovariance, cross-cepstrum and saphe cracking. In: Rosenblatt M (ed)
Time series analysis, pp 209–243
35. Oppenheim AV, Schafer RW, Buck JR (1999) Discrete-time signal processing solutions. In:
Discrete time signals and systems. Prentice Hall, Upper Saddle River
36. Randall RB (2017) A history of cepstrum analysis and its application to mechanical problems.
Mech Syst Signal Process 97:3–19
37. Randall RB, Sawalhi N Cepstral removal of periodic spectral components from time signals.
In: Proceedings of CMMNO 2013. Springer, Ferrara
38. Randall R, Smith W, Coats M Gear diagnostics under widely varying speed conditions. In:
CMMNO 2014. INSA, Lyon
39. Gao Y, Randall RB (1996) Determination of frequency response functions from response
measurements – I. extraction of poles and zeros from response cepstra. Mech Syst Signal
Process 10(3):293–317
40. Gao Y, Randall RB (1996) Determination of frequency response functions from response
measurements – II. Regeneration of frequency response functions from poles and zeros. Mech
Syst Signal Process 10(3):319–340
41. Richardson M, Formenti D (1985) Global curve fitting of frequency response measurements
using the rational fraction polynomial method. In: Proceedings of the IMAC III
42. Smith WA, Randall RB (2016) Cepstrum-based operational modal analysis revisited: a
discussion on pole–zero models and the regeneration of frequency response functions. Mech
Syst Signal Process 79:30–46
43. Hanson D, Randall RB, Antoni J, Waters TP, Thompson DJ, Ford RAJ (2007) Cyclostationarity
and the cepstrum for operational modal analysis of MIMO systems-Part I: modal parameter
identification. Mech Syst Signal Process 21(6):2441–2458
44. Randall RB, Peeters B, Antoni J, Manzato S (2012) New cepstral methods of signal pre-
processing for operational modal analysis. In: Proceedings of the international conference on
noise and vibration engineering (ISMA 2012), Leuven, Belgium
45. Randall RB, Coats MD, Smith WA (2016) Repressing the effects of variable speed harmonic
orders in operational modal analysis. Mech Syst Signal Process 79:3–15
Introduction to Spectral and Correlation
Analysis: Basic Measurements and Methods 5
Anders Brandt and Stefano Manzoni

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
1.1 Properties of Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
1.2 Common Applications in Experimental Structural Dynamics . . . . . . . . . . . . . . . . . . 231
2 Signal Classes and Their Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
2.1 Periodic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
2.2 Random Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
2.3 Transient Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
2.4 Double-Sided Versus Single-Sided Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
3 Frequency Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
3.1 Spectrum Analysis Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
3.2 The Discrete Fourier Transform (DFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
4 Block-Based Spectrum and Correlation Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
4.1 The Linear (RMS) Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
4.2 The Phase Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
4.3 Welch’s PSD and CSD Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
4.4 Energy Spectral Density Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
4.5 Welch’s Correlation Function Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
5 Periodogram-Based Spectrum and Correlation Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 253
5.1 The Periodogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
5.2 The Smoothed Periodogram PSD Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
5.3 The Periodogram Correlation Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
5.4 Dealing with Harmonics in Correlation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

A. Brandt ()
Department of Mechanical and Electrical Engineering, University of Southern Denmark,
Odense, Denmark
e-mail: [email protected]
S. Manzoni
Department of Mechanical Engineering, Politecnico di Milano, Milan, Italy
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 229

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_7
230 A. Brandt and S. Manzoni

Abstract

Spectral analysis is one of the most important tools used in experimental

structural dynamics. This can be partly explained by the fact that the output
of a linear system in the frequency domain, at each frequency, is equal to the
product of the input spectrum at that frequency and the frequency response
at the same frequency. For random vibrations, correlation functions and their
frequency counterparts, spectral densities, are the tools used to describe the
frequency content of the vibrations. In this chapter, we start by briefly describing
the essential properties of linear systems. After this, we describe the three classes
of signals: periodic, random, and transient signals, and for each signal class, we
define a spectrum to describe its frequency content. We then go on to describe
the discrete Fourier transform, DFT, since this is by far the most common tool to
compute spectra, by the fast Fourier transform (FFT) algorithm. In this context,
leakage and time windowing are explained, after which we go into detail on
how to compute the spectra for each type of signal. Two different methods are
described: Welch’s method, based on averaging several shorter DFT blocks, and
the periodogram-based method which relies on making one, long DFT and then
averaging adjacent frequency bins. Finally, also correlation function estimates
are described using the same two techniques.

Keywords

Fourier transform · Spectrum analysis · Linear spectrum · Spectral density ·

Correlation function

1 Introduction

Spectral analysis is one of the most important tools in experimental structural

dynamics. This is for a very simple reason; the response of a linear system at
a particular frequency is only dependent on the input at that frequency and the
frequency response at the same frequency, as we will discuss in Sect. 1.1. Spectra
are thus very important functions, and the spectra of periodic, random, and periodic
functions will be presented in Sect. 2. Finding a solution to a vibration problem
often boils down to finding at what frequency the problem occurs and then finding
the structure’s dynamics (defined by the frequency response functions, FRFs) and/or
the loads, at this frequency.
Correlation functions are other important measurement functions, particularly
for operational modal analysis (OMA). Correlation functions are closely related to
spectra; more precisely they are inverse Fourier transforms of the spectral density
functions. As we will see in Sect. 2.2.1, correlation functions are closely related with
impulse response functions (IRFs) and can essentially be treated as such functions
when extracting modal parameters in OMA.
The contents of the present chapter are mostly textbook stuff. It is not so easy to
find easily readable books on applied signal analysis aimed for mechanical or civil
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 231

Fig. 1 Illustration of single-input/single-output linear system with input signal (e.g., force) x(t)
and output signal (response, e.g., acceleration) y(t)

engineers. There are, however, a few books that can be recommended; for general
reading, see [8, 5, 2], for more in-depth analysis of random signals, there are rather
a few books, for example [1, 7, 11], and also see the chapter on Random and Shock
Testing. A more specialized book on spectrum analysis that can be recommended
is [9].

1.1 Properties of Linear Systems

The simplest linear system, depicted in Fig. 1, has an input x(t) and an output
(or response) y(t). The system can, for example, describe the linear relationship
between the force input in a particular point and the response acceleration in
another point on a structure. The linear system can be characterized by its frequency
response function (FRF), H (f ), which is defined by

Y (f )
H (f ) = (1)
X(f )

where X(f ) and Y (f ) are the Fourier transforms of the input and response signals,
respectively. We can rewrite the equation as Y (f ) = X(f ) · H (f ) which clearly
shows that at a particular frequency, f , the response is only dependent on the input
and the frequency response at that particular frequency. Thus, knowing the spectra
X(f ) and Y (f ) are partial steps toward the description of the system in terms of
FRFs. It should be pointed out here that Eq. (1) is only conceptual and should not
be used in practice. To estimate FRFs, rather one of the methods described in the
chapter “Structural Measurements – FRF” should be used.

1.2 Common Applications in Experimental Structural Dynamics

The basic measurements we will describe in the present chapter are common to
many applications in the field of experimental structural dynamics. Some of the
most common applications will be mentioned here, together with the typical spectra
and other functions associated with them.
Operating deflection shape measurements, ODS, is a very common application
for troubleshooting vibration problems. ODSs may be based on any spectral
functions that contain amplitude and phase information, e.g., cross-spectra, or even
FRFs.
232 A. Brandt and S. Manzoni

Measurements of FRFs for experimental modal analysis measurements, EMA,

will be covered in the chapter on FRF measurements. Here, it is sufficient to explain
that EMA is usually based on measurements of input forces and responses in some
form (velocity or acceleration signals, typically). The frequency response functions
between force(s) and response(s) are then estimated from the spectra of the force
and response signals that we present in the present chapter.
Operational modal analysis measurements, OMA, are often, although not
always, based on correlation functions (sometimes also called covariance functions).
If correlation functions are used, OMA is sometimes referred to as NeXT, [6].

2 Signal Classes and Their Spectra

Now that we have established the need to experimentally obtain spectra, we will
have to discuss how this can be done for different types of signals. This is necessary,
as different signals have spectra with different properties. The main types of signals
in this respect can be divided into three signal classes

• periodic signals,
• random signals, and
• transient signals

Each of these types of signals needs to be described by its own type of spectrum as
we will describe in the following.

2.1 Periodic Signals

Periodic signals, or signals that can, at least, be approximated as periodic, occur in

applications where there are rotating or reciprocating parts, for example, turbines
and combustion engines. For periodic signals, we know from Fourier series theory
that they can be described as a discrete sum of harmonics. Furthermore, for vibration
signals we know they have to be zero mean, so in the following, we assume all
signals are zero mean. It is good practice to always remove means of all measured
vibration signals after data acquisition. Any such periodic signal, xp (t), with period
Tp seconds, can be written as the Fourier series

∞
    
2π kt 2π kt
xp (t) = ak cos + bk sin (2)
Tp Tp
k=1

where the Fourier coefficients ak and bk can be computed by

t0+Tp  
2 2π k
ak = xp (t) cos t dt k = 1, 2, . . . (3)
Tp Tp
t0
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 233

and

t0+Tp  
2 2π k
bk = xp (t) sin t dt k = 1, 2, . . . (4)
Tp Tp
t0

for an arbitrary starting point t0 .

Instead of the Fourier series presented in Eq. (2), the complex Fourier Series
is often used, particularly since it is more closely related to the discrete Fourier
transform; see Sect. 3.2. The complex Fourier series is thus defined by

∞
 j 2π k
Tp t
xp (t) = ck e (5)
k=−∞

where the coefficients ck are given by

a0
c0 =
2
t0+Tp
1 1 − j T2πp k t
ck = (ak − j bk ) = xp (t)e dt (6)
2 Tp
t0

and c0 = 0 since we assume no mean. The complex Fourier coefficients, ck , can

easily be shown (for real signals xp (t)) to be such that Re(c−k ) = Re(ck ) and
Im(c−k ) = −Im(c−k ). The negative values of k in Eq. (5) can be interpreted as
negative frequencies, which will be discussed more in Sect. 2.4.
The Fourier series, real or complex, show that a periodic signal only contains
discrete frequencies, fk = k/Tp . For physical signals, furthermore, the higher
frequencies will vanish, resulting in finite sums in Eqs. (2) and (5).

2.2 Random Signals

Random signals occur when the forces originate in some random phenomenon, for
example, wind loads, traffic loads, or the loads from the road surface acting on
road vehicles. For random signals, or with a fancier name, stochastic processes,
we need to use a rather different description of the spectral content. First, random
signals have continuous spectra, as opposed to the discrete spectra of periodic
signals. This means that the spectra of random signals must be described in terms of
density functions. Before exploring this, however, we need to deal with the random
(stochastic) nature of the signals, by defining the correlation functions.
234 A. Brandt and S. Manzoni

2.2.1 Correlation Functions

A random signal is stationary if its mean, variance, and higher-order moments
are constant in time. It is, further, ergodic if its statistical values (mean, variance,
etc.) can be replaced by time averages. In engineering applications, a signal which
is stationary can normally be regarded also to be ergodic. It should, however, be
stressed that most signals are not stationary, at least not over any particularly long
time. Stationarity is often something that needs to be investigated or controlled by
experimental design.
For two stationary and ergodic random signals, x(t) and y(t), we define the cross-
correlation, Ryx (τ ), by

Ryx (τ ) = E [y(t)x(t − τ )] (7)

where E denotes the expectation value. It should be noted that the cross-correlation
function is equal to the signal y(t) convolved by x(−t). For a single signal, x(t),
the autocorrelation is a special case of the cross-correlation between the signal and
itself, i.e.:

Rxx (τ ) = E [x(t)x(t − τ )] (8)

Ryx (τ ) thus tells if there is correlation (relationship) between the signal y(t) and
a shifted version of x(t) shifted τ seconds. Periodic signals thus have periodic
correlation functions, with period τ = Tp . For random signals, the correlation
function is usually an oscillating function with a maximum at, say, τ0 and which
decays as ±(τ − τ0 ) increases. For our purposes it can be of interest to look at two
different correlation functions, depicted in Fig. 2. In Fig. 2a, the autocorrelation of
a bandlimited random signal with 1 Hz bandwidth is plotted. It can be seen that
the autocorrelation function is of sin(x)/x type. In Fig. 2b, the cross-correlation
between the input and output of a single-degree-of-freedom (SDOF) system with
undamped natural frequency 1 Hz and relative damping of 1% is shown. Here it
can be seen that the cross-correlation resembles the impulse response function, for
positive lags τ . Actually, the positive lags in cross-correlation functions include the
same information about the system as the impulse response, which is why it can be
used for OMA parameter extraction; see the chapter on Operational Modal Analysis
Methods.

2.2.2 Spectral Density Functions

The spectral density of a random signal is defined by the so-called Wiener-
Khinchin(-Einstein) relationship, as the Fourier transform of the correlation func-
tion. Thus, the Fourier transform of an autocorrelation is referred to as a power
spectral density, (PSD) function, or sometimes an autospectral density (ASD)
function, thus defined by
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 235

(a) (b)

0.8 5

Cross-correlation
Autocorrelation

0.6

0.4 0

0.2

0 -5

-5 0 5 -10 0 10
Time lag, [s] Time lag, [s]

Fig. 2 Example correlation functions: in (a) autocorrelation function of bandlimited noise with
1 Hz bandwidth and in (b) cross-correlation function between the input and output of a single-
degree-of-freedom (SDOF) mechanical system with undamped natural frequency of 1 Hz and
relative damping of 1%. It can be seen that the cross-correlation in (b) is similar to an impulse
response function, for positive lags τ

∞
Sxx (f ) = Rxx (τ )e−j 2πf τ dτ (9)
−∞

The Fourier transform of a cross-correlation function, on the other hand, is

referred to as a cross-spectral density (CSD) function

∞
Syx (f ) = Ryx (τ )e−j 2πf τ dτ (10)
−∞

An example of a PSD of a velocity signal measured on a bridge is shown in

Fig. 3.
The units of a PSD are the measured unit squared per Hz, for example,
(m/s2 )2 /Hz if acceleration in metric (SI) units are used. This unit reflects the fact
that a PSD displays the distribution of the power of the signal, over frequency. The
area under the PSD between two frequencies thus equal the mean square (square of
the RMS level) of the signal in that frequency range.
It should be noted that the PSD is a real-valued, and positive, function, whereas
the CSD is complex-valued (i.e., includes phase information).
236 A. Brandt and S. Manzoni

10 -5
PSD [(Velocity) /Hz]

-6
10
2

10 -7

-8
10

-9
10
0 0.5 1 1.5 2 2.5 3 3.5 4
Frequency [Hz]

Fig. 3 Example power spectral density (PSD) of a velocity signal measured on a bridge

2.3 Transient Signals

A transient signal is a finite signal with limited duration. Examples of transient

signals are responses to pyroshock testing, for example, the sound of a car door
closing. Transient signals are usually, but not always, deterministic. An example
of a nondeterministic transient signal is a burst random signal; see the chapter
on FRF measurements. If we restrict the discussion here to deterministic transient
signals, then an appropriate spectrum of the signal is the direct, continuous Fourier
transform. The spectrum of a transient, deterministic, signal, x(t), is thus

∞
X(f ) = x(t)e−j 2πf t dt. (11)
−∞

The units of the transient spectrum of an acceleration in SI units are thus

[m/s2 · s] which can also be represented as [(m/s2 )/Hz]. Often, the energy spectral
density function (ESD) is used instead of the straight Fourier transform. The ESD is
equal to the magnitude squared of the spectrum X(f ).

2.4 Double-Sided Versus Single-Sided Spectra

The Fourier transform results in double-sided functions. The PSDs and transient
spectra in the previous sections are therefore double-sided, i.e., they contain both
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 237

negative and positive frequencies. The negative frequencies are a result of the
symmetry of the Fourier transform, and thus half the “energy” of the signal is
distributed on the negative frequency axis and the other half on the positive. For
practical use, this is not very convenient, so the spectra found in real life are usually
single-sided spectra. For random signals, we thus get the single-sided PSD, for
example, which is computed by discarding all negative frequencies and multiplying
all the positive frequencies (not the 0 frequency!) by a factor 2. The single-sided
PSD is denoted Gxx (f ) to distinguish it from the double-sided version.

3 Frequency Analysis

In this section we will first describe a basic principle of spectrum analysis as it is

fundamental to the understanding of the spectra we are going to cover in Sect. 3.1.
We will then describe the discrete Fourier transform (DFT) in some detail, since it is
the tool most often used for spectrum estimation in structural dynamics applications.

3.1 Spectrum Analysis Principle

There are two main groups of methods for spectrum analysis, either parametric or
nonparametric methods. The parametric methods rely on some a priori knowledge
about the spectrum to be estimated, e.g., that it is produced by a Gaussian, white
random force that has passed a structure having eight modes. The nonparametric
methods do not require any such knowledge but give a spectrum estimate without
any assumption. Experience has shown that the parametric methods usually perform
poorly on vibration signals, so what is used in practice, with only rare exceptions,
are the nonparametric methods.
The most common nonparametric methods for spectrum estimation in the field of
structural dynamics are octave (and 1/n octave) analysis, DFT/FFT-based estimates,
and wavelet spectra. It is very important to realize that all nonparametric methods
rely on the same principle, which we will now outline.
The principle of nonparametric spectrum estimation is schematically illustrated
in Fig. 4. The picture illustrates that each frequency in the estimated spectrum is the
result of a calculation of the RMS level of the output of a bandpass filter with a
particular center frequency, fc , and bandwidth, B (often alternatively denoted by
Δf ). It is important to understand that all nonparametric methods for spectrum
analysis rely on this simple principle. The only thing that can differ between the
methods is the properties of the bandpass filter: the center frequency, bandwidth,
and shape. In octave and 1/n octave analysis, the center frequencies and bandwidths
increase exponentially with frequency. This is often called constant relative band-
width because Δf/fc is kept constant. DFT/FFT-based spectra typically have fixed
frequency increment and bandwidth, although we will see in Sect. 5.2 that this is
not necessary. Wavelet spectra, finally, usually have constant relative bandwidth.
238 A. Brandt and S. Manzoni

Octave Spectrum
B
f
fc
63 125 250500 1k 2k 4k

Fig. 4 Principle for spectrum analysis. The content at each discrete frequency, where the estimated
spectrum contains information, is the result of the RMS evaluation of the output of a bandpass filter

Since DFT/FFT-based spectrum analysis is by far the most common in structural

dynamics, we will limit the discussion here to those types of estimates.

3.2 The Discrete Fourier Transform (DFT)

The discrete Fourier transform (DFT) is the most common tool used to compute
discrete spectra of measured signals. The fast Fourier transform (FFT) is an
algorithm that computes the DFT much faster than the direct formula implies. The
output of the FFT is thus the DFT, and for that reason, it is sufficient to discuss the
DFT and its properties here.
The DFT, X(k), at discrete frequencies k, of a signal x(n) sampled with N
samples (which we call the blocksize), is


N −1
X(k) = x(n)e−j 2π kn/N k = 0, 1, . . . , N − 1 (12)
n=0

with the inverse discrete Fourier transform, the IDFT, defined by

N −1
1 
x(n) = X(k)ej 2π kn/N n = 0, 1, . . . , N − 1 (13)
N
k=0

In order to understand the DFT, first of all the complex exponential e−j 2π kn/N
can be expanded to cos(2π kn/N ) − j sin(2π kn/N ). The DFT at frequency k is
thus obtained by multiplying the signal x(n) by a cosine and a sine and sum each of
these products to produce the real part and imaginary part of the DFT, respectively.
It is easily seen that the exponential is a cosine and sine with k periods during the
measurement time N · Δt, since the ratio n/N goes from 0 to 1 when n goes from 0
to N . The first frequency, k = 1, which is also the frequency increment of the DFT
output, is thus Δf = 1/(NΔt) = fs /N. This means that k = N corresponds to the
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 239

(a)
2

1
x(n)

-1

-2
0 5 10 15 20 25 30
Sample number, n

(b)
20

15
Real[X(k)]

10

0
0 5 10 15 20 25 30
Frequency number, k

Fig. 5 Time signal in (a) and corresponding DFT result in (b). The signal is a cosine, periodic
with three periods within the time window, and sampled with N = 32 samples. The DFT result is
entirely real-valued, because the cosine is an even function, and the peak is located on frequency
bin k = 3 because the cosine contains three periods in the time window

sampling frequency, and thus k = N/2 corresponds to half the sampling frequency,
known as the Nyquist frequency. It should also be noted that the measurement time
NΔt is actually one sampling distance longer than the actual time it takes to acquire
the N samples. It is apparent why this is so in the upper plot of Fig. 5, where the last
sample is one sample short of the periodicity of the cosine plotted.
In Fig. 5 the result of a DFT computation of a cosine with three periods within
the time window is shown. First of all, in the figure it is clear that the peaks in the
DFT are not immediately (at least seemingly) related to the amplitude of the cosine;
the DFT output is unscaled, as can be clearly seen from the value X(0) = x(n),
not the mean of the signal as it “should” be. Furthermore, since a cosine is an even
function, only the multiplication with the cosine in the complex exponential will
result in a nonzero sum. Thus the result of the DFT of the cosine is real-valued, and
therefore the imaginary part of X(k) is not plotted in the figure.
240 A. Brandt and S. Manzoni

Since the cosine in Fig. 5 contains three periods within the time window, the
peak in the spectrum X(k) occurs at k = 3. It can also be seen that the values for
k = N/2 + 1, N/2 + 2, . . . , N − 1 are the negative frequencies, translated to the
right. Therefore, after computing the DFT, we usually discard all frequencies above
k = N/2. For a block size of 1024 samples, for example, 513 values in the DFT
result should be kept. Due to the symmetry of the DFT, the discarded frequencies
can easily be computed by procedures that can be found in any textbook on the
topic.
A summary of the most important properties of the DFT are:

1. The DFT result X(k) is periodic with period N

2. The DFT is symmetric such that Re [X(−k)] = Re [X(k)] and Im [X(−k)]
= −Im [X(k)] (even real part, odd imaginary part)
3. The real part of X(k) is the DFT of the even part of x(n), and the imaginary part
of X(k) is the DFT of the odd part of x(n) 
4. The DFT is unscaled, which can be seen from X(0) = x(n), which is not the
mean but the sum of the signals. Thus, the DFT needs some scaling, which we
will address in the following sections.

3.2.1 Leakage and Windowing

Leakage is a phenomenon caused by the truncation in time that occurs for
continuous signals which are not periodic in N, when we limit the calculation of
the DFT to N samples. If a continuous signal is measured during a limited time
T = N Δt, then the time truncation corresponds to multiplying the continuous time
signal by a rectangular time window


1, |t| ≤ T /2
w(t) = (14)
0, |t| > T /2

Since multiplication in the time domain corresponds to convolution in the

frequency domain, the result of the truncation of the measurement time is that
the true spectrum of the signal x(t) is convolved with the spectrum of the time
window, W (f ). It is therefore essential to understand the properties of the spectrum
(Fourier transform) of the time window. This is illustrated in Fig. 6, where the time
and frequency domain plots of three common windows are shown: the rectangular,
Hanning, and flattop windows. The frequency axis of the spectra is showing number
of frequency increment (bins), Δf . In the frequency domain, all windows have a
main lobe (the center lobe) and side lobes. What differs between the window spectra
is essentially the width of the main lobe and the height and falloff rate of the side
lobes, and there is a trade-off between the two, the broader the main lobe, the lower
the side lobes.
The rectangular window, in Fig. 6b, has the narrowest main lobe of only ±1
frequency bin and the highest side lobes, the highest being approximately −18 dB
relative to the main lobe. The falloff rate of the rectangular window is very small
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 241

a) b)
0
1
−20
0.8
−40
0.6

dB
0.4 −60

0.2 −80

0 −100
0 N−1 −10−8 −6 −4 −2 0 2 4 6 8 10

c) d)
0
1
−20
0.8
−40
0.6
dB

0.4 −60

0.2 −80

0 −100
0 N−1 −10−8 −6 −4 −2 0 2 4 6 8 10

e) f)
5 0

4 −20
3
−40
dB

2
−60
1

0 −80

−1 −100
0 N−1 −10−8 −6 −4 −2 0 2 4 6 8 10
n k

Fig. 6 Some common time windows in time and frequency domain. In (a) and (b), time and
frequency domain representations of the rectangular window; in (c) and (d), the same for the
Hanning window and in (e) and (f) for the flattop window. It can be seen that the rectangular
window has a narrow main lobe and high side lobes, compared to the Hanning and flattop windows.
It can also be noted that a wider main lobe results in lower side lobes
242 A. Brandt and S. Manzoni

(6 dB/octave). The Hanning window, in Fig. 6d, has a main lobe of ±2 frequency
bins, and the first side lobe is approximately −32 dB under the main lobe, and the
falloff rate is considerably faster than for the rectangular window (18 dB/octave).
The flattop window, finally, has a very broad mainlobe, ±5 frequency bins, very
low side lobes, approximately −90 dB below the main lobe.
Consider now the convolution of one of the window spectra and an actual
spectrum of a structure with a resonance. Remember that the “ideal” convolution
is that of a δ-function, i.e., convolving with a δ-function results in the function it is
convolved with. The wider the main lobe is, the more smearing there will be in the
resulting spectrum. The higher the side lobes are, the more contribution there will
be at a particular frequency, from frequencies far away from that frequency. A good
all-round window for random signals and spectra that exhibit resonance peaks is the
Hanning window, which is therefore the standard window used for random signals
and spectral density estimates. The flattop window is good for periodic signals, with
well-separated harmonics. The rectangular window, finally, should only be used
when the measured signal does not have any leakage, i.e., transients that are entirely
captured by the N samples of the DFT or signals that are periodic inside the N
samples.
The maximum amplitude error caused by leakage when estimating the DFT of
a sine is approximately −40% for the rectangular window, −16% for Hanning
window, and −0.1% for the flattop window.

3.2.2 Cyclic Convolution and Zero Padding

Cyclic convolution is a phenomenon due to the periodicity of the DFT which
is important to understand in many cases of applying the DFT, for example,
for filtering purposes. We know from continuous Fourier transform theory that
multiplication in the frequency domain corresponds to convolution in the time
domain. But this is not immediately true for the DFT, since it is periodic in both
time domain and frequency domain. This is easy to show with an example.

Example 1. Consider the discrete sequence (signal)

x(n) = [1 1 1 1].
The true convolution of this sequence is clearly
y(n) = [1 2 3 4 3 2 1].
But try the following, for example, in MATLAB:
» x=[1 1 1 1];
» y = ifft(fft(x).*fft(x);
and the result will show
ans = 4 4 4 4



5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 243

This result is an effect of the cyclic nature of the DFT. For each shift in the
convolution, the sample(s) shifted out to the right are appearing on the left side.
Thus, for each value, the convolution result is the sum of the sequence

[1 1 1 1]

multiplied by itself and summed. The solution to this is zero padding. If we add as
many zeros to the original sequence as it is long, we get the following result, again
illustrated by an example in MATLAB.

Example 2. Now, include zero padding in the sequence x, with as many zeros as
the length of x

» x=[1 1 1 1 0 0 0 0];
» y = ifft(fft(x).*fft(x);
and the result will show
ans = 1 2 3 4 3 2 1 0 


which is clearly the correct result!

Zero padding should be used for proper results, whenever a convolution in

the time domain is computed by a multiplication in the frequency domain. Thus,
it should always be used when any kind of filtering operation is computed in
the frequency domain. We will also use it to compute correlation functions in
Sects. 4.5 and 5.3.

3.2.3 Window Scaling Factors

The time windows obviously remove part of the windowed signal, and this causes
the DFT of the windowed signal having lower values than the DFT of the
unwindowed signal. Luckily, this can be compensated for, however, we need to
discuss two different compensations, depending on what result we want.
If, in the first instance, we apply a window to a periodic signal and we want the
DFT of the windowed signal to show correct peak values (i.e., peaks corresponding
to the amplitude of the harmonic at the frequency bin k), then we need to find an
amplitude correction factor (ACF). We denote this factor Aw , and it can readily
be found. Assume that we have a complex sine with amplitude A, coinciding with
frequency bin k0 , which can then be written as

x(n) = Aej 2π k0 n/N , n = 0, 1, . . . , N − 1 (15)

If we compute the DFT of this sine windowed by window w(n), divided by N

(because we saw in Sect. 3.2 that the DFT has to be scaled by N ), we obtain
244 A. Brandt and S. Manzoni

N −1
1 
X(k) = w(n)Ae−j 2π k0 n/N ej 2π kn/N (16)
N
n=0

All of the values of X(k) will be zero, except the value for X(k0 ) which will be

N −1
A 
X(k0 ) = w(n) (17)
N
n=0

Since we want this frequency bin to equal A, then we have that the window
amplitude correction factor, Aw , is

N
Aw = (18)
N
−1
w(n)
n=0

so that Aw · DFT [w(n)x(n)] produces the correct amplitude A.

In the second instance, we will later see that often, instead of wanting a peak to
be properly scaled, we wish for the sum of the (square of) values of the DFT to be
correct. This factor is a little more complicated to deduce, and we therefore refer
to [1, 2] for a deduction. Here, it is sufficient to define the normalized equivalent
bandwidth, Ben , of a window w(n) by

N
−1
N w 2 (n)
n=0
Ben =  2
(19)
N
−1
w(n)
n=0

This factor is exactly 1.5 for the Hanning window and will be used in Sect. 4.3.

4 Block-Based Spectrum and Correlation Estimation

In this section we will look at estimators for the theoretical spectra that were
presented in Sect. 2. We thus assume that we have a discrete time signal, x(n),
equal to the continuous signal x(t) sampled at the equidistant sampling instances
tn = n · Δt, where the sampling increment is the reciprocal of the sampling
frequency, fs , i.e., Δt = 1/fs . We further assume that we have sampled L samples
of the signal x(n).
There are two different principles for spectrum and correlation estimation using
the DFT/FFT: either the time signal is divided into smaller segments, for which
windowed DFTs are computed and averaged in the frequency domain, or one, large
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 245

FFT
FFT

m=1 m=3 ...

m=2

Fig. 7 Schematic illustration of Welch’s method for spectrum and correlation function estimation.
The data are divided into a number of, possibly overlapping, segments, each of which is windowed,
after which the DFT is computed (by FFT) whereafter averaging is done in the frequency domain

DFT is computed using the entire time signal. The former method is referred to as
Welch’s method after [10] and the latter as a periodogram method, or sometimes
Daniell’s method, after [4]. In this section we will present Welch’s method for
computation of both spectral density estimates and correlation function estimates.
This method is usually the only method implemented in commercial systems for
structural dynamics, although, as we will see in Sect. 5, there can sometimes be
reasons for choosing the periodogram-based methods.
The principle of Welch’s method is illustrated schematically in Fig. 7. The data
are divided into M blocks of length N samples (the block size). The M segments
can be overlapped; see Sect. 4.3. For each segment, a time window is applied, the
DFT calculated, and averaging is then performed in the frequency domain, as we
will show in the next subsections.
In Sect. 3 we mentioned that the spectral density of a signal is the Fourier
transform of the correlation function. This is not a very practical solution for com-
puting spectra, however, since correlation functions are computationally intense.
The spectrum estimators are based on computing the FFT because it is considerably
faster than the direct DFT. Therefore, we need to see how to use the FFT/DFT to
obtain the Fourier transform of the correlation function. This can relatively easily be
seen, if we consider Eq. (7), where it was shown that the cross-correlation Ryx (τ )
essentially corresponds to the convolution of the signal y(t) by the signal x(−t) (it
actually corresponds to the convolution divided by the number of samples, as there
is an expectation value in Eq. (7)). We can now use the Fourier transform pair of
x(−t), which is X∗ (f ). Thus, the cross-spectral density between signals x and y is
proportional to

Syx (f ) = F Ryx (τ ) ∝ Y (f )X∗ (f ) (20)

246 A. Brandt and S. Manzoni

where F [ ] denotes the Fourier transform. To complete the equation, and to

compute the estimators by using the DFT, we need some scaling that we will address
below.

4.1 The Linear (RMS) Spectrum

We are now ready to address the spectrum estimator best suited for dealing with
periodic signals, without significant noise. Such signals can, for example, be
vibrations on a reciprocating engine or on power generators. The spectrum estimator
is the so-called linear spectrum or RMS spectrum. This estimator, which we denote
XL (k), is the square root of the autopower spectrum. The reason for using the
intermediate autopower spectrum is that if averaging is necessary to reduce the
variance of the estimate due to noise, then this averaging has to be applied to squared
spectral components, in order for the average to be a mean square average. This is
necessary because we always want the averaging result to display true RMS levels.
The autopower spectrum is produced by dividing the signal x(n) with length
L samples, into M blocks of length N that we denote xm (n), (if L is not an
integer number of blocks with length N , we simply discard the last few samples
of the signal). Each of the M blocks is windowed, the DFT of the windowed
block is computed, and the DFT results are averaged, frequency by frequency.
An intermediate step is thus to compute the DFT of the windowed block number
m = 1, 2, . . . , M, of x(n) that we denote Xw,m and which equals


N
Xw,m (k) = w(n)xm (n)e−j 2π kn/N k = 0, 1, · · · , N/2 (21)
n=0

where w(n) is the time window. Note that we discard the upper half of the DFT
blocks, i.e., the negative frequencies in the definition of Xw,m . We can now define
the autopower spectrum of the signal x, by

SA  SA 
M M
∗
Âxx (k) = Xw,m (k)Xw,m (k) = |X(k)|2 (22)
M M
m=1 m=1

where SA is a scaling constant to produce peaks in Ayx (k) equal to the square of
the RMS level of the corresponding harmonic component in the signal. The scaling
constant is
⎧ 2
⎪
⎪ 2A
⎨ 2w , k = 0
SA = N (23)
⎪
⎪ A2
⎩ w, k = 0
N2
where Aw is the window amplitude scaling factor from Eq. (18).
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 247

After we have estimated the autopower spectrum, the linear (RMS) spectrum can
be computed, as the square root of Axx

X̂L (k) = Âxx (k) (24)

A peak in the linear spectrum is thus interpreted as the RMS level of a harmonic
component at that frequency. Note that the linear spectrum does not have any phase.
The choice of scaling the linear spectrum to the RMS level of the harmonics is,
of course, not mandatory, although we certainly recommend it. In some fields it is
common to also scale linear spectra to the amplitude of the harmonics. Therefore, it
is important to include the scaling in the units of the plot, by specifying, e.g., “[m/s2
RMS]” if the measured signal was acceleration in the SI unit of m/s2 .
We are now ready to apply the linear spectrum to a real signal. The parameters
we need to consider are listed in Table 1. One setting is not listed in the table; the
bandwidth (highest frequency) of the measurement naturally needs to be chosen.
This is a setting that should normally be chosen based on some knowledge about
the signal to be measured, such as the RPM of the engine and how many harmonics
are required. Alternatively, one may try a high-frequency setting and, after looking
at the obtained spectrum, gradually reduce the bandwidth until the spectrum shows
a sufficient amount of higher frequencies for the purpose of the measurement.
To select the settings for the linear spectrum estimator, the time window is
usually set based on taste. A flattop window is sometimes preferred, as it guarantees
that the RMS levels of harmonic components are accurately determined (to within
0.1%). On the other hand, due to the wide main lobe of the flattop window, some
users prefer the Hanning window and accept the error in RMS readings of up
to −16%, which, in many cases in vibration applications, is certainly negligible.
An appropriate number of averages usually has to be found by a trial and error
procedure. If the RMS levels change significantly for repeated measurements
without averaging, the number of averages can be increased until stable RMS levels
are obtained. This usually requires a small number of averages, typically between

Table 1 Typically used settings for block-based spectrum estimators. Low number of averages
are typically 5–20; high can be 50–500
Spectrum type Window Number of averages Overlap, % Block size, N
Linear spectrum Flattop, Hanning Low 0–50 Large enough to
obtain a line
spectrum
PSD, CSD Hanning High 50 Large enough to
avoid bias
ESD Rectangular No∗ No Large enough to
capture entire
transient
∗ For transients, triggered acquisition can be used with averaging, for example, for impact testing,
see the chapter on FRF measurements
 248 A. Brandt and S. Manzoni

a) b)
1 1
Linear Spectrum Acc. [m/s2 RMS]

Linear Spectrum Acc. [m/s2 RMS]

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Frequency [Hz] Frequency [Hz]

Fig. 8 Example of linear spectrum estimates with two different block sizes; in (a) a frequency
increment of 5 Hz has been used; in (b) the fre8quency increment is 1.25 Hz. It is clear in (a)
that the frequency increment is not small enough to allow the spectrum to reach zero between the
peaks, whereas in (b) the peaks are distinct, and thus the spectrum is an appropriate line spectrum
as expected for a periodic signal

5 and 20. For periodic signals, overlap is often not used, although 50% can be
advantageously used to speed up the measurement. See more about averaging in
Sect. 4.3.
The last setting in Table 1 is the most important to set correctly. The block size
should be set such that a line spectrum is obtained, i.e., that the spectrum reaches
zero between every peak. If this is not the case, the frequency increment is too large.
The recommended procedure to find the optimal block size is thus to make repeated
measurement, starting with a small block size that does not create a line spectrum,
and then increase the block size until a line spectrum is obtained. This procedure
is illustrated in Fig. 8 where two linear spectra of the same signal, the acceleration
measured on a fan, are shown.

4.2 The Phase Spectrum

Sometimes, for periodic signals, a phase relative to some reference signal (a

tachometer signal or another response signal) is required, for example, for operating
deflection shape (ODS) measurements. There is really only one way to obtain such
a phase, and that is from a cross-spectrum. The phase spectrum is an “artificial”
spectrum where the linear spectrum is added a phase from the cross-spectrum
between the response signal and some reference signal. Assume we have a reference
signal x(t) and a response signal y(t) for which we wish to have the linear spectrum
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 249

with reference to x(t). For this purpose, first the cross-power spectrum Âyx (k) is
estimated by
SA 
M
∗
Âyx (k) = Yw,m (k)Xw,m (k) (25)
M
m=1

and the linear spectrum, ŶL (k), is estimated by Eqs. (22) and (24). The phase
spectrum Ŷpx is then constructed from these two spectra as

Ŷpx (k) = ŶL (k)ej arg [Âyx (k)] (26)

where arg[ ] denotes the phase in radians. Each spectral line of the phase spectrum
will thus contain the RMS level of the signal y and the phase of y related to the
reference x at that frequency. This allows software to pick amplitude and phase
information for animation of ODSs.

4.3 Welch’s PSD and CSD Estimates

Welch’s method for PSD estimation relies on the same procedure as the method for
estimating autopower spectra we discussed in Sect. 4.1. Similarly to Eq. 22, we thus
define the PSD estimate, Ĝxx (k) as

SP  SP   2
M M
Ĝxx (k) = ∗
Xw,m (k)Xw,m (k) = Xw,m (k) (27)
M M
m=1 m=1

where SP is a scaling constant to produce spectral density scaling, such that the area
under the estimate Ĝxx (k) equals the mean square of the signal x(t). The scaling
constant is
SA
SP = (28)
Ben Δf

where SA is the RMS scaling constant for autopower spectrum defined by Eq. (23)
and Ben is the normalized equivalent noise bandwidth defined by Eq. (19). Similarly
the cross-spectral density estimate between two signals, x(t) and y(t), is defined by

SP 
M
∗
Gyx (k) = Yw,m (k)Xw,m (k) (29)
M
m=1

The PSD estimate in Eq. (27) is rather different from the estimate of the
autopower spectrum, since the signal, x(t), is assumed to be random. Thus, the
spectral density is continuous, and there will be two errors in the estimate: a bias
error and a random error. The bias error is a result of a continuous spectrum being
250 A. Brandt and S. Manzoni

estimated by a discrete estimator and can be eliminated if necessary, by making the

frequency increment, Δf small enough, i.e., by using a block size, N , which is large
enough. This can be achieved by a similar approach as that described for obtaining
an appropriate frequency increment for the linear spectrum estimate, at the end of
Sect. 4.1.
The random error of a PSD estimate is defined by the standard deviation of
the difference between the estimated and true PSDs. Usually we talk about the
normalized random error, εr , which is the mentioned standard deviation divided
by the true value of the PSD or
  
  E (Ĝxx − Gxx )2
εr Ĝxx = (30)
Gxx

The random error is independent of frequency and is reduced as the number of

averages is increased. Usually 50% overlap is recommended for PSD estimation, as
this gives a significantly lower (almost half) random error, for a given measurement
time. Although rather complicated in detail, the random error of a Welch PSD
estimate is an effect of the time window, the overlap, and the number of averages.
For space reasons we refer to [2, 10] for a thorough treatment. It is sufficient here to
refer to Fig. 9 which shows the random error as a function of the number of averages
used, when the PSD is estimated by Welch method, a Hanning window, and 50%
Normalized random error, εr

−1
10

−2
10
1 2 3 4
10 10 10 10

Number of FFTs

Fig. 9 Random error of Welch PSD estimate estimated using Hanning window and 50% overlap,
as a function of the number of averages used for the estimate
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 251

overlap processing. As is evident from this plot, many averages are needed to obtain
a small random error; for an error of 5%, for example, 400 averages are needed.
With 50% overlap, that means that the data have to contain 200 non-overlapped
blocks.
In Table 1, in the row for PSD estimates, the recommended measurement system
settings for a PSD estimate are shown. A Hanning window should always be used
together with 50% overlap (slightly more overlap can be used, but hardly pays off
as it results in very slightly lower random error). The number of averages normally
has to be high, which means, say, between 50 and 500 averages. Most important,
however, is that the block size is made large enough to eliminate the bias error, at
least in cases where the PSD is going to be used for further processing, for example,
for modal parameter extraction.

4.4 Energy Spectral Density Estimates

The energy spectral density (ESD) function is usually used as the preferred spectrum
for (deterministic) transient signals. As we described in Sect. 2.3, the ESD is the
magnitude squared of the Fourier transform of the transient. Although the ESD is
not, strictly speaking, a block-based estimate, we still describe it here, as it is usually
found alongside the linear spectrum and PSD estimates in commercial software for
vibration analysis.
The ESD estimate Gxx (k) of the transient captured in x(n) of length L, is
therefore rather straightforwardly defined by

 2
 
L−1 
 
Gxx (k) = Δt · x(n)e−j 2π kn/L  (31)
 
n=0

where it should be especially noted that we do not use any time window. The typical
measurement system settings for an ESD measurement are listed in the last line in
Table 1.

4.5 Welch’s Correlation Function Estimates

To estimate correlation functions, the most common estimator used in commercial

measurement systems is Welch’s estimator. It is very important, however, to note
that the correlation function should not be estimated as the inverse DFT of a PSD
estimated by the procedure described in Sect. 4.3 as this will lead to a distorted
(biased) estimate, not suitable for parameter extraction of modal parameters, for
example. To see how to estimate the correlation function properly, we refer back
to the definition. A direct calculation of the cross-correlation of two random time
signals x(t) and y(t) as by Eq. (7) is
252 A. Brandt and S. Manzoni

T /2
1
Ryx (τ ) = E [y(t)x(t − τ )] = y(t)x(t − τ )dt (32)
T
−T /2

It is easily realized that this estimator will lead to a biased estimator, because we
divide by T , the length of the signal x(t), whereas for any lag τ = 0, there will
actually only be overlapping data that can be multiplied together of length T − τ .
Thus we obtain the unbiased cross-correlation estimator by

T /2
1
Ryx (τ ) = E [y(t)x(t − τ )] = y(t)x(t − τ )dt (33)
T − |τ |
−T /2

and for a single signal, x(t), as mentioned previously, the autocorrelation is obtained
by replacing y(t) by x(t).
Since, for structural dynamics applications, we are usually only interested in the
nonnegative lags τ ≥ 0, we will limit the discussion here to those lags. To find an
appropriate discrete estimator for the time integral in Eq. (33), we start by observing
that the discrete correlation function should be equal to the convolution of x(n)
by x(n − m), weighted by 1/(N − m) where we let m denote the discrete lag, in
number of samples, and N is the block size of the DFT. This is thus equal to a
multiplication in the frequency domain, of X(f ) by its complex conjugate X∗ (f ).
But, it is very important to use zero padding in the computation of the DFT, X(f ),
to obtain the intermediate spectra (DFT results) Xm,z and Ym,z , in order to avoid
cyclic convolution effects. If the data contains N samples, then it should be zero
padded by N zeros, prior to the DFT calculation.
Furthermore, using Welch’s approach, we use a block size, N , much smaller
than the total length L, thus averaging M DFT results in the frequency domain. We
denote the resulting, temporary, PSD estimate by Ŝyx C and obtain it by

1 
M
∗
C
Ŝyx (k) = Ym,z (k)Xm,z (k) (34)
M
m=1

C (k)
The next step is to calculate and properly scale the inverse DFT (IDFT) of Ŝyx
which results in the unbiased estimate of the cross-correlation function


2N −1
1
R̂yx (m) = C
Ŝyx (k)ej 2π km/(2N ) , m = 0, 1, . . . mmax < N/2
2N(N − m)
k=0
(35)
where the factor 2N in the denominator belongs to the inverse DFT. It is not neces-
C (k), since it does not significantly
sary to apply overlap processing to compute Ŝyx
contribute to a reduced random error in the correlation estimate.
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 253

5 Periodogram-Based Spectrum and Correlation Estimation

In this section we introduce the alternative to the block-based technique (Welch’s

method) for computing spectra and correlation functions. The methods we will
discuss are based on making only one, large DFT/FFT of each measurement
signal, followed by some processing and scaling to produce reliable estimates.
Although this requires more memory in the computer, with modern computers
and measurement systems, this is rarely a drawback, and as we will see, there are
some reasons for sometimes choosing the periodogram-based methods instead of
the block-based.

5.1 The Periodogram

The (auto)periodogram of a signal x(n) is defined as the magnitude squared of the

DFT of the signal, X(l), scaled by the length, L,

L−1 2
 
Δt  
P̂xx (l) =  x(n)e−j 2π ln/N  (36)
L  
n=0

We could, of course, also define a cross-periodogram, if we have an additional

signal y(n) (where x(n) is considered the reference) by

L−1  L−1 ∗
Δt  
P̂yx (l) = y(n)e−j 2π ln/N x(n)e−j 2π ln/N (37)
L
n=0 n=0

where ∗ denotes complex conjugation. Note that we have used the symbol l to
denote the discrete frequency of the periodogram. This is to distinguish this, usually
much finer, frequency index from the frequency index of the estimates using Welch’s
estimate, for the discussion in Sect. 5.2.
The periodogram in Eq. (36) is an estimator for spectral density but a rather poor
such estimator, as can be seen in Fig. 10a, b where periodograms of white Gaussian
noise are plotted, based on L = 512 samples in (a) and L = 4096 samples in
(b). The periodogram is an inconsistent estimator, in that it does not approach the
true PSD with increasing length L. On the contrary, as can be seen in the figure, it
behaves more and more wildly the more data it is based on. On the right-hand side
of Fig. 10, in (c) and (d), similar periodograms but of a signal containing a periodic
signal plus random noise are plotted. As can be seen in the plots, with more data,
the periodic components stand out of the noise. Thus, the periodogram is a good
estimator for finding sines hidden in noise, although there is a lot of leakage. This
can easily be avoided by applying a flattop window to the data before computing the
periodogram, however, thus computing a windowed periodogram.
254 A. Brandt and S. Manzoni

a) c)
Periodogram length N= 512

Periodogram length N= 512

10 0 10 0

-2 -2
10 10

-4 -4
10 10

-6 -6
10 10

-8 -8
10 10
0 200 400 0 200 400
Frequency [Hz] Frequency [Hz]

b) d)
Periodogram length N= 4096

Periodogram length N= 4096

0 0
10 10

-2 -2
10 10

-4 -4
10 10

-6 -6
10 10

10 -8 10 -8
0 200 400 0 200 400
Frequency [Hz] Frequency [Hz]

Fig. 10 Periodogram plots of random signal (left) and a periodic signal with noise (right). In (a)
and (b), periodograms of the same random signal are shown, with data length of N = 512 in (a)
and N = 4096 in (b) and in (c) and (d), similarly for periodograms of a periodic signal with noise

5.2 The Smoothed Periodogram PSD Estimate

An alternative to the PSD estimate by Welch’s method, which can offer some
advantages, is the so-called smoothed periodogram or sometimes the Daniell
method. This estimate is obtained by smoothing (averaging) the periodogram in
Eq. (36) around each frequency where an estimate is wanted. If we, for example,
choose the same frequencies, k · f s/N, k = 0, 1, . . . N/2 that Welch’s estimate
would yield, then the smoothed autoperiodogram estimator is

2 s /2
k·D+L
ĜSP
xx (k) = P̂xx (l) k = 1, 2, . . . (38)
(Ls + 1)
l=k·D−Ls /2
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 255

and similarly for the cross-periodogram estimator. The factor 2 in the numerator in
Eq. (38) is due to scaling for a single-sided PSD, ĜSP
xx (k), as the periodogram P̂xx (l)
is double-sided. We have assumed Ls is an even number, and the variable D is the
number of frequency bins l between each target frequency k. The straight average
in Eq. (38) can be replaced by weighted averaging, but it turns out to give very little
difference in the properties of the resulting estimates.
We will illustrate the smoothed periodogram estimator with an example.

Example 3. Assume we have data x(n) with length L = 100 · N and N = 1024, the
block size we use with Welch’s method, and with a sampling frequency of 1024 Hz.
This would yield frequencies for Welch’s estimate at kΔf = kfs /N = k · 1 Hz.
If we, for the sake of a direct comparison, assume that we do not use any overlap
processing, Welch’s method would result in 100 averages.
The corresponding smoothed periodogram estimator would be obtained by
selecting the same frequencies fk = kΔf for the resulting spectrum and setting
Ls = 100, the same as the number of averages used for Welch’s estimate. Since the
frequency increment of the periodogram is fs /L = fs /(Ls N ) = Δf/Ls = 1/100
in this case, then obviously D = Ls = 100. So, each frequency bin k in the Welch’s
estimate corresponds to the frequency bins l = kD in the periodogram. At each of
these latter bins, the periodogram is averaged over ±50 bins. 


The smoothed periodogram computed as described above is similar to Welch’s

estimate, in terms of its bias and random errors. It requires a much larger FFT than
Welch’s method but only one, as opposed to 100 FFTs of the Welch estimate in our
example. Welch’s estimate has the advantage of requiring much less memory, but
the computation speed for the periodogram estimate is similar to that for Welch’s
estimate for typical data lengths in structural dynamics applications. There are,
furthermore, at least two advantages that can make the smoothed periodogram PSD
estimator a preferred estimator:

• harmonics can easily be removed using the periodogram; see Sect. 5.4
• the smoothed periodogram estimate can easily be computed using constant rela-
tive bandwidth, Δf/fk by letting D and Ls grow exponentially with frequency

5.3 The Periodogram Correlation Estimate

The periodogram can also be used to compute the correlation functions, as described
in [3]. This is actually a more straightforward way than using Welch’s method
as described in Sect. 4.5. To obtain correlation functions without effects of cyclic
correlation, L zeros have to be added to the time signal of length L, prior to
computing the periodogram. In addition, we do not use the scaling by the time
increment and data length, as in Eq. (36), but instead use the magnitude square of
the DFT, without scaling. A straightforward computation of the cross-correlation
256 A. Brandt and S. Manzoni

estimate, R̂yx (τ ), is thus obtained by first computing the zero-padded DFTs to be

the 2L length FFTs to obtain Xz (l) and Yz (l). Then the unbiased cross-correlation
estimate for positive lags, m, is obtained by the inverse DFT, divided by L − m, i.e.:

1 
2L−1
1
P
R̂yx (m) = Yz (l)Xz∗ (l)ej 2π ml/(2L) = IDFT Yz Xz∗
(2L(L − m)) L−m
l=0
(39)

The autocorrelation Rxx (m) is obtained by replacing Yz in the equation by Xz .

Normally we are only interested in a number of lags m L, so the large part of the
IDFT in Eq. (39) can be discarded. This estimator requires a long DFT (FFT) to be
computed, but for most data lengths typically used for OMA, for example, modern
computers have sufficient performance that this estimate is computed in a similar
amount of time compared to the Welch’s estimate (and sometimes even faster). It
should be noted that the correlation function estimate in Eq. (39) is identical to the
direct time domain computation, as well as the computation by Welch’s method.

5.4 Dealing with Harmonics in Correlation Functions

Harmonics in the response signals is a large problem in applications of OMA on,

for example, wind power mills, ships, and power generators, since it violates the
assumptions for OMA. The periodogram offers a relatively easy way to remove
the harmonics in such cases which is particularly attractive if spectral densities or
correlation functions are going to be computed as described in Sects. 5.2 or 5.3.
As shown in Fig. 10, harmonics stand out from the periodogram; the longer the
data length, the more they stand out. Thus, after computing the long DFTs, Yr (l),
r = 1, 2, . . . , Nr for Nr responses used for OMA (see the chapter on OMA),
each auto and cross-periodogram is produced. The harmonics will typically only
appear as single, or very few, frequency bins in these periodograms and can easily
be edited away, by replacing them by linear interpolation between the remaining
frequency bins on the left- and right-hand sides. After this, the periodograms are
either smoothed, to produce PSDs and CSDs, or inverse Fourier transformed, to
produce auto- and cross-correlation functions. More information of this may be
found in [3].

6 Summary

In this chapter we have described the basic properties of spectra and correlation
functions and how these functions should be estimated. When calculating and
interpreting spectra, it is important to consider what type of signal is at hand, a
5 Introduction to Spectral and Correlation Analysis: Basic Measurements. . . 257

periodic, a random, or a transient signal. The most important results presented in

the chapter are the following.
A periodic signal should be described by a linear spectrum, also called RMS
spectrum, as described in Sect. 4.1. This spectrum should be scaled so that a peak
value in the spectrum corresponds to the RMS level of a harmonic component at
the frequency. The units of this spectrum, if the measured unit is EU (engineering
units), are [EU RMS], for example, [m/s2 RMS] for an acceleration signal. For the
linear spectrum, a time window has to be selected, as described in Sect. 4.1. Some
recommendations for settings can be found in Table 1.
A random signal should be represented by a power spectral density (PSD)
spectrum, which is interpreted such that the area under the PSD is the mean
square value (the square of the RMS) of the signal. PSDs may be estimated
in two different ways; in Sect. 4.3 we described how to do it with the Welch
estimator (dividing the signal into a number of segments, which are processed and
averaged in the frequency domain). Another method for estimating a PSD is the
periodogram method described in Sect. 5.2 which is based on one DFT (FFT) being
computed on the entire measured signal. The units of the PSD are [EU2 /Hz], for
example, [(m/s2 )2 /Hz] for an acceleration signal. For the PSD, a Hanning window
is recommended. For further recommendations, see Table 1.
A transient signal is most often represented by the energy spectral density (ESD)
spectrum, computed from a DFT (FFT) of the entire transient signal. The area under
the ESD is interpreted as the energy of the transient and has units of [EU2 /Hz · s].
The estimator is described in Sect. 2.3. For an ESD, no time window should be used;
see Table 1.
Correlation functions of random signals are mostly used for operational modal
analysis. Again, there are two ways to compute these functions in the frequency
domain: one block-based method (Welch’s method) and one method based on a
periodogram. How to estimate these functions are described in Sects. 4.5 and 5.3,
respectively.

References
1. Bendat JS, Piersol AG (2010) Random data: analysis and measurement procedures, 4th edn.
Wiley Interscience, Hoboken
2. Brandt A (2011) Noise and vibration analysis – signal analysis and experimental procedures.
Wiley, Chichester
3. Brandt A (2019) A signal processing framework for operational modal analysis in time and
frequency domain. Mech Syst Signal Process 115:380–393
4. Daniell PJ (1946) Discussion of ‘on the theoretical specification and sampling properties of
autocorrelated time-series’. J R Stat Soc 8(suppl.)(1):88–90
5. de Silva CW (2006) Vibration: fundamentals and practice, 2nd edn. CRC Press, Boca Raton,
Florida
6. James GH, Carne TG, Lauffer JP (1995) The natural excitation technique (next) for modal
parameter extraction from operating structures. Modal Anal Int J Anal Exp Modal Anal
10(4):260–277
258 A. Brandt and S. Manzoni

7. Newland DE (2005) An introduction to random vibrations, spectral, and wavelet analysis, 3rd
edn. Dover Publications Inc., Mineola/New York
8. Shin K, Hammond J (2008) Fundamental of signal processing for sound and vibration
engineers. Wiley, Chichester
9. Stoica P, Moses R (2005) Spectral analysis of signals. Prentice Hall, Upper Saddle River
10. Welch PD (1967) The use of fast Fourier transform for the estimation of power spectra:
a method based on time averaging over short, modified periodograms. IEEE Trans Audio
Electroacoust AU-15(2):70–73
11. Wirsching PH, Paez TL, Ortiz H (1995) Random vibrations: theory and practice. Wiley
Interscience, New York, NY
Frequency Response Function Estimation
6
A. W. Phillips and R. J. Allemang

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
2 Frequency Response Function Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
3 Frequency Response Function Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
3.1 Noise/Error Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
3.2 Single Input FRF Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
3.3 Multiple Input FRF Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
3.4 Coherence: Ordinary, Multiple, and Conditioned . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
3.5 Multiple Input Force Analysis/Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
4 Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
4.1 General Averaging Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
4.2 Estimation of Frequency Response Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
4.3 Special Types of Signal Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
5 Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
5.1 Excitation Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
5.2 Excitation Terminology and Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
5.3 Classification of Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
5.4 Random Excitation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
5.5 Deterministic Excitation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
5.6 Excitation Example: H-Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
6 Structural Testing Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
7 Practical Measurement Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347

A. W. Phillips () · R. J. Allemang

Structural Dynamics Research Laboratory, Mechanical and Materials Engineering, University of
Cincinnati, Cincinnati, OH, USA
e-mail: [email protected]; [email protected]

© The Society for Experimental Mechanics 2022 259

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_8
260 A. W. Phillips and R. J. Allemang

Abstract

For current approaches to experimental modal analysis, the frequency response

function is the most important measurement to be made. This chapter develops
the frequency response function from the perspective of experimentally measured
system excitations and responses. Experimental measurement and numerical
processing techniques are presented that allow minimization of the impact of
measurement noise and signal processing errors.

Keywords

Multiple input frequency response function · Ordinary coherence · Multiple

coherence · Conditioned coherence · Spectral averaging · Selective excitation

Nomenclature

N Number of degrees of freedom

Ni Number of inputs
No Number of outputs
Ns Number of spectral lines (frequencies)
Fmax Maximum frequency (Hz)
ω Frequency (rad/s)
f Frequency resolution (Hz)
T Observation period (s)
λ Complex eigenvalue
{V } Complex eigenvector
[H (s)] Transfer function matrix
[H (ω)] Frequency response function matrix
{X(ω)} Response vector
η Noise on response
{F (ω)} Excitation vector
υ Noise on excitation
[GF F ] Input-input power spectral matrix
[GXF ] Output-input cross power spectral matrix
[GXX ] Output-output power spectral matrix
OCOH Ordinary coherence
MCOH Multiple coherence
PCOH Partial coherence
CCOH Cumulative coherence
FCOH Fractional coherence
VCOH Virtual coherence
SV D Singular value decomposition
ED Eigenvalue decomposition
CD Cholesky decomposition

1 Introduction

For current approaches to experimental modal analysis, the frequency response

function is the most important measurement to be made. When estimating frequency
response functions, a measurement model is needed that will allow the frequency
6 Frequency Response Function Estimation 261

response function to be estimated from measured input and output data in the
presence of noise (errors). Some of the errors are:

• Digital signal processing errors

– Leakage
– Aliasing
• Noise
– Equipment problem (power supply noise)
– Cabling problems (RFI, EMI)
– Rattles, cable motion
• Calibration (operator error)
– Complete system calibration
– Transducer calibration

Since the frequency response function can be expressed in terms of system

properties of mass, stiffness, and damping, it is reasonable to conclude that in most
realistic structures, the frequency response functions are considered to be constants
just like mass, stiffness, and damping. This concept means that when formulating
the frequency response function using H1 , H2 , or Hv algorithms, the estimate of
frequency response is intrinsically unique, as long as the system is linear and the
noise can be minimized or eliminated. The estimate of frequency response is valid
whether the input is stationary, nonstationary, or deterministic. Therefore, several
important points to remember before estimating frequency response functions are
as follows:

• The system (with the boundary conditions for that test) determines the frequency
response functions for the given input/output locations.
• It is important to eliminate or at least minimize all errors (aliasing, leakage, noise,
calibration, etc.) when collecting data.
• If all noise terms are identically zero, the assumption concerning the
source/location of the noise does not matter (H1 = H2 = Hv = Hs = H ). Therefore,
concentrate on eliminating the source of the noise.
• Since modal parameters are computed from estimated frequency response
functions, the modal parameters are only as accurate as the estimated frequency
response functions.

There are at least four different testing configurations that can be compared.
These different testing conditions are largely a function of the number of acquisition
channels or excitation sources that are available to the test engineer. In general, the
best testing situation is the multiple input/multiple output (MIMO) configuration
since the data is collected in the shortest possible time with the fewest changes in
the test conditions.

• Single input/single output (SISO)

– Only option if two-channel data acquisition system.
262 A. W. Phillips and R. J. Allemang

– Longest testing time. Roving inputs. Roving outputs.

– Time invariance problems between measurements.
• Single input/multiple output (SIMO)
– Multiple channel system (three or more). (One ADC channel for each
response signal to be measured plus one ADC channel for an input signal.)
– Shorter testing time than SISO. Transducers not necessarily moved.
– Consistent frequency and damping for data acquired simultaneously.
– Time invariance problems between measurements from different inputs.
• Multiple input/single output (MISO)
– Multiple channel system required (three or more.). (One ADC channel for
each input signal to be measured plus one ADC channel for a response signal.)
– Long testing time. Roving response transducer.
– More than one input location per measurement cycle.
– Detects repeated roots. Maxwell reciprocity checks are possible.
– Time invariance problems between measurements from different responses
are possible.
• Multiple input/multiple output. (MIMO)
– Multiple channel system (up to 512 channels). Increased setup time. Large
amount of data to be stored and organized.
– Shortest testing time.
– Consistent frequency and damping for all data acquired simultaneously.
– Detects repeated roots. Maxwell reciprocity checks are possible.
– Best overall testing scheme.

2 Frequency Response Function Development

In this chapter, frequency response functions are developed from experimentally

measured input-output (force-response) relationships. However, prior to that devel-
opment, it is advantageous to show the relationship to a lumped parameter mass,
damping, and stiffness (M,C,K) model and the system transfer function matrix.
The transfer function representation of a general multiple degree of freedom
system can be formulated by starting with the differential equation of motion in
terms of mass, stiffness, and damping matrices:

[M] {ẍ(t)} + [C] {ẋ(t)} + [K] {x(t)} = {f (t)} (1)

Taking the Laplace transform of Equation 1, assuming all initial conditions are
zero, yields:
 
[M] s 2 + [C] s + [K] {X (s)} = {F (s)} (2)
6 Frequency Response Function Estimation 263

Letting:
 
[B (s)] = [M] s 2 + [C] s + [K] (3)

Then Equation 2 becomes:

[B (s)] {X (s)} = {F (s)} (4)

where [B (s)] is referred to as the impedance matrix or just the system matrix.
Premultiplying Equation 4 by [B (s)]−1 yields:

[B (s)]−1 {F (s)} = {X (s)} (5)

Defining:

[H (s)] = [B (s)]−1 (6)

Then:

[H (s)] {F (s)} = {X (s)} (7)

Equation 7 relates the system response {X (s)} to the system forcing functions
{F (s)} through the matrix [H (s)]. The matrix [H (s)] is generally referred to as
the transfer function matrix.
By observing the partitioned nature of the matrix equations and evaluating the
equation for a single output response Xp (s):


N
Hpq (s) Fq (s) = Xp (s) (8)
q=1

the familiar relationships for the transfer function (output over input) are obtained
by evaluating for a single input excitation Fq (s) with all other inputs zero Fk (s) =0:
k = q:

Xp (s)
Hpq (s) = (9)
Fq (s)

• p is the output degree of freedom (physical location and orientation).

• q is the input degree of freedom (physical location and orientation).

Thus, Hpq (s) is the transfer function which would be measured by exciting
the system with Fq (s) and measuring the response Xp (s). Unfortunately, it is not
possible to measure F (s) and X(s) directly.
264 A. W. Phillips and R. J. Allemang

However, since s is a general frequency variable (s = σ + j ω with units of

rad/s), by evaluating the transfer function matrix at s = j ω, the frequency response
function matrix results:

 −1
H (ω) = −ω2 [M] + j ω [C] + [K] (10)

Therefore, measuring a column of the frequency response function matrix

[H (ω)] can be accomplished by using a single, fixed input (exciter system) with
a roving response, and measuring a row can be accomplished by using a roving
input (hammer) and a single fixed response:

Xp (ω)
Hpq (ω) = (11)
Fq (ω)

It should be reiterated that the subscript notation of p or q refers to both a

physical location and also direction or orientation.

3 Frequency Response Function Estimation

Frequency response functions are normally used to describe the input-output (force-
response) relationships of any system. Most often, the system is assumed to be linear
and time invariant although this is not a necessary part of the definition. In the cases
where assumptions of linearity and time invariance are not valid, the measurement
of frequency response functions is also dependent upon the independent variables of
time and input. In this way, a conditional frequency response function is measured
as a function of other independent variables in addition to frequency. Note that the
different possible formulations listed in Table 1 can all be considered frequency
response functions since each of these formulations can be numerically manipulated
(synthetic differentiation, integration, etc.) into the equivalent displacement over
force relationship. This assumes that initial conditions can be ignored.

Table 1 Frequency response Acceleration

Receptance F orce
function formulations
F orce
Effective Mass Acceleration
V elocity
Mobility F orce
F orce
Impedance V elocity
Displacement
Dynamic Compliance F orce
F orce
Dynamic Stiffness Displacement
6 Frequency Response Function Estimation 265

The estimation of the frequency response function depends upon the transforma-
tion of data from the time to frequency domain. The Fourier transform is used for
this computation. Unfortunately, though the integral Fourier transform definition
requires time histories from negative infinity to positive infinity, since this is not
possible experimentally, the computation is performed digitally using a fast Fourier
transform (FFT) algorithm which is based upon only a limited time history. In this
way, the theoretical advantages of the Fourier transform can be implemented in a
digital computation scheme. The frequency response functions satisfy the following
single and multiple input relationships:

Single Input Relationship:

Xp = Hpq Fq (12)

Multiple Input Relationship:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X1 H11 · · · · · · · H1q F1
⎢X ⎥ ⎢H · ⎥ ⎢F ⎥
⎢ 2⎥ ⎢ 21 ⎥ ⎢ 2⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ · ⎥ =⎢ · · ⎥ ⎢ · ⎥ (13)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ · ⎦ ⎣ · · ⎦ ⎣ · ⎦
Xp N Hp1 · · · · · · · Hpq N Fq N ×1
o ×1 o ×Ni i

An example of a two-input, two-output case for Equation 13 is shown in

Equation 14 and Fig. 1.

X1 H11 H12 F1
= (14)
X2 H21 H22 F2

Fig. 1 Two-input,
two-output FRF concept
266 A. W. Phillips and R. J. Allemang

3.1 Noise/Error Minimization

The most reasonable, and most common, approach to the estimation of frequency
response functions is by the use of least squares (LS) or total least squares (TLS)
techniques [1, 2, 3, 4, 5]. This is a standard technique for estimating parameters in
the presence of noise. Least squares methods minimize the square of the magnitude
error and, thus, compute the best estimate of the magnitude of the frequency
response function but have little effect on the phase of the frequency response
function. The primary difference in the algorithms used to estimate frequency
response functions is in the assumption of where the noise enters the measurement
problem. The different assumptions of the source of the error is noted graphically in
Fig. 2.
Three algorithms, referred to as the H1 , H2 , and Hv algorithms, are commonly
available for estimating frequency response functions. Table 2 summarizes this
characteristic for the three methods that are widely used.
Consider the case of Ni inputs and No outputs measured during a modal test.
Based upon the assumed location of the noise entering the estimation process,
Equations 15, 16, and 17 represent the corresponding model for the H1 , H2 , and
Hv estimation procedures.

Fig. 2 Least squares concept

Table 2 Summary of Frequency response function models

frequency response function
Technique Solution Assumed location of noise
estimation models
Method Force inputs Response
H1 LS No noise Noise
H2 LS Noise No noise
Hv TLS Noise Noise
6 Frequency Response Function Estimation 267

H1 Technique:

[H ]No × Ni {F }Ni × 1 = {X}No × 1 − {η}No × 1 (15)

H2 Technique:

[H ]No × Ni {F }Ni × 1 − {υ}Ni × 1 = {X}No × 1 (16)

Hv Technique:

[H ]No × Ni {F }Ni × 1 − {υ}Ni × 1 = {X}No × 1 − {η}No × 1 (17)

Note that while not necessarily obvious yet, in all methods, the inversion of a
matrix will be involved. Therefore, the inputs (references) that are used must not
be fully correlated so that the inverse will exist. Extensive evaluation tools (using
eigenvalue decomposition) have been developed in order to detect and avoid this
condition [6, 7].

3.2 Single Input FRF Estimation

Figure 3 represents the model of the measurement situation for a single input, single
output frequency response function measurement.

Fig. 3 System model: single

input
268 A. W. Phillips and R. J. Allemang

With reference to Fig. 3 for a case involving only one input and one output (input
location q and response location p), the equation that is used to represent the input-
output relationship is:

X̂p − ηp = Hpq (F̂q − υq ) (18)

where:

• F = F̂ − υ = actual input
• X = X̂ − η = actual output
• X̂ = spectrum of the p − th output, measured
• F̂ = spectrum of the q − th input, measured
• H = frequency response function
• υ = Spectrum of the noise part of the input
• η = Spectrum of the noise part of the output
• X = Spectrum of the p − th output, theoretical
• F = Spectrum of the q − th input, theoretical

If υ = η = 0, the theoretical (expected) frequency response function of the

system is estimated. If η = 0 and/or υ = 0, a least squares method is used to
estimate a best frequency response function, in the presence of noise.
In order to develop an estimation of the frequency response function, a number of
averages Navg are used to minimize the random errors (variance). This can be easily
accomplished through the use of intermediate measurement of the auto and cross
power spectra. The estimate of the auto and cross power spectra for the model in
Fig. 3 can be defined as follows. Note that each function is a function of frequency.

Auto Power Spectra:

Navg

GF F qq = Fq Fq∗ (19)
1
Navg

GXXpp = Xp Xp∗ (20)
1

Cross Power Spectra:

Navg

GXF pq = Xp Fq∗ (21)
1
Navg

GF Xqp = Fq Xp∗ (22)
1
6 Frequency Response Function Estimation 269

where:

• F ∗ = complex conjugate of F (ω)

• X∗ = complex conjugate of X(ω)

3.2.1 H1 Algorithm: Minimize Noise on Output (η)

The most common formulation of the frequency response function, often referred to
as the H1 algorithm, tends to minimize the noise on the output. This formulation is
shown in Equation 23.

GXF pq
Hpq = (23)
GF F qq

Note that recently an FRF development based upon a Cholesky decomposition

has been developed [8]. This formulation is equivalent to the H1 formulation.

3.2.2 H2 Algorithm: Minimize Noise on Input (υ)

Another formulation of the frequency response function, often referred to as the H2
algorithm, tends to minimize the noise on the input. This formulation is shown in
Equation 24.

GXXpp
Hpq = (24)
GF Xqp

In the H2 formulation, an auto power spectrum is divided by a cross power

spectrum. This can be a problem since the cross power spectrum can theoretically
be zero at one or more frequencies. In both formulations, the phase information is
preserved in the cross power spectrum term.

3.2.3 Hv Algorithm: Minimize Noise on Input and Output (η and υ)

Rather than assuming that the noise is solely on the input or output, the noise can
be assumed to be on both the input and output. This results in a vector least squares
solution that is found via an eigenvalue decomposition. This approach to estimating
the frequency response function is referred to as the Hv algorithm.
The eigenvalue solution is used to find the optimal least squares solution of two
linear equations represented by the H1 algorithm and the H2 algorithm.
The H1 algorithm can be rewritten as:

GF F qq Hpq − GXF pq = 0 (25)

The H2 algorithm can be rewritten as:

GF Xqq Hpq − GXXpq = 0 (26)

270 A. W. Phillips and R. J. Allemang

Combining these two representations for Hpq in a matrix equation:

   
GF F qq GXF pq Hpq 0
= (27)
GF Xqp GXXpp 2×2
−1 2×1
0 2×1

Recognizing that in the presence of noise, the right hand side of the above
equation will not be equal to zero. Replacing the right hand side by a noise vector
{} with the following form:
   
ε1 Hpq
=λ (28)
ε2 2×1 −1 2×1
   
GF F qq GXF pq Hpq Hpq
=λ (29)
GF Xqp GXXpp 2×2
−1 2×1 −1 2×1

This is a recognizable eigenvalue equation. The solution associated with the

smallest eigenvalue will give an eigenvector, when normalized as above, that will
estimated the optimal value for Hpq .
Therefore, the solution for Hpq using the Hv algorithm is found by the eigenvalue
decomposition of a matrix of power spectra. For the single input case, the following
matrix involving the auto and cross power spectra can be defined:

  GF Fqq GXFpq
GF F Xp = (30)
GF Xqp GXXpp 2×2

The solution for Hpq is found by the eigenvalue decomposition of the [GF F Xp ]
matrix as follows:
 
GF F Xp = [V ]   [V ]H (31)

where:

•   = diagonal matrix of eigenvalues

Solution for the Hpq matrix is found from the eigenvector associated with the
smallest (minimum) eigenvalue (λmin ). The size of the eigenvalue problem is second
order resulting in finding the roots of a quadratic equation. This eigenvalue solution
must be repeated for each frequency, and the complete solution process must be
repeated for each response point Xp .
Note that the eigenvalue decomposition does not necessarily yield the eigenval-
ues in any particular order so the smallest or minimum eigenvalue must be found
and is not generally the first or last eigenvalue. However, since the decomposition
matrix [GF F X ] is Hermitian, the eigenvalues should be real valued (and, while not
required, the eigenvectors can be scaled to be orthonormal as shown in Equation 31.)
6 Frequency Response Function Estimation 271

Also note that the noises on the input and output are assumed proportional to
the eigenvector scaling. Other forms of scaling are possible by altering the above
equation giving rise to the definition of Hs .
Alternately, the solution for Hpq is found by the eigenvalue decomposition of the
following matrix of auto and cross power spectra:

  GXXpp GF Xqp
GXF Fp = (32)
GXFpq GF Fqq 2×2
 
GXF Fp = [V ]   [V ]H (33)

where:

•   = diagonal matrix of eigenvalues

The solution for Hpq is again found from the eigenvector associated with the
smallest (minimum) eigenvalue (λmin ).
The frequency response function is found from the normalized eigenvector
associated with the smallest eigenvalue. If [GF F Xp ] is used, the eigenvector
associated with the smallest eigenvalue must be normalized as follows:
 
Hpq
{V }λmin = (34)
−1

If [GXF Fp ] is used, the eigenvector associated with the smallest eigenvalue must
be normalized as follows:
 
−1
{V }λmin = (35)
Hpq

One important consideration of the three formulations for frequency response

function estimation is the behavior of each formulation in the presence of a bias error
such as leakage. In all cases, the estimate differs from the expected value particularly
in the region of a resonance (magnitude maxima) or antiresonance (magnitude
minima). For example, H1 tends to underestimate the value at resonance, while
H2 tends to overestimate the value at resonance. The Hv algorithm gives an answer
that is bounded by the H1 and H2 values for the single input case. The different
approaches are based upon minimizing the magnitude of the error but have no effect
on the phase characteristics.
There are a number of formulations for the Hv estimation of the FRFs that
appear to be different. Note that there are several decompositions of the [GF F X ]
and [GXF F ] matrices, including SVD and ED, that have been used in the develop-
ment of Hv . The original, historical presentation/development used the eigenvalue
decomposition as presented in this chapter.
272 A. W. Phillips and R. J. Allemang

3.2.4 Ordinary Coherence

In addition to the attractiveness of H1 , H2 , and Hv in terms of the minimization of
the error, the availability of auto and cross power spectra allows the determination
of other important functions. The quantity γpq 2 is called the scalar or ordinary

coherence function and is a frequency-dependent, real value between zero and one.
The ordinary coherence function indicates the degree of correlation in a frequency
response function. If the coherence is equal to one at any specific frequency, the
system is said to have perfect correlation at that frequency. In other words, the
measured response power is linearly related to the measured input power (or by
sources which are coherent with the measured input power). A coherence value
less than unity at any frequency indicates that the measured response power is
greater than that due to the measured input. This is due to some extraneous noise
also contributing to the output power. It should be emphasized, however, that low
coherence does not necessarily imply poor estimates of the frequency response
function but simply means that more averaging is needed for a statistically reliable
result. Note that the coherence will not necessarily improve (i.e., get closer to one)
unless the noise is eliminated, but the variance will mean that the FRF answer will
be statistically better. The ordinary coherence function is computed as follows:

 
GXF 2 GXFpq GF Xqp
OCOHpq = = =
2 pq
γpq (36)
GF Fqq GXXpp GF Fqq GXXpp

When the coherence is zero, the output is totally uncorrelated to the measured
input suggesting that the response is due to other unmeasured sources. In general,
then, the coherence can be a measure of the degree of noise contamination in a
measurement. Thus, with more averaging, the estimate of coherence may contain
less variance, therefore giving a better estimate of the noise energy in a measured
signal. This is not the case, though, if the low coherence is due to bias errors such as
nonlinearities, multiple unmeasured inputs, or leakage. A typical ordinary coherence
function is shown in Fig. 4 together with the corresponding frequency response
function magnitude. In Fig. 4, the frequencies where the coherence is lowest are
often the same frequencies where the frequency response function is at a maxima
in magnitude or at a minima in magnitude. This is often an indication of leakage
since the frequency response function is most sensitive to the leakage error at the
lightly damped peaks corresponding to the maxima. At the minima, where there is
little response from the system, the leakage error, even though it is small, may still
be significant.
Note that while the above development of ordinary coherence is historical and
was based upon the H1 FRF algorithm, this definition is also valid for both the H2
and Hv FRF algorithms.
In all of these cases, the estimated coherence function will approach, in the limit,
the expected value of coherence at each frequency, dependent upon the type of noise
present in the structure and measurement system. Note that with more averaging,
the estimated value of coherence will not increase; the estimated value of coherence
6 Frequency Response Function Estimation 273

Ordinary Coherence
1
0.9
0.8
0.7
Magnitude

0.6
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50
Frequency (Hz)

Frequency Response - Automotive Exhaust

200
150
100
Phase (deg)

50
0
-50
-100
-150
-200
0 10 20 30 40 50
Frequency (Hz)

Frequency Response - Automotive Exhaust

1
10

0
10
Magnitude

-1
10

-2
10

-3
10
0 10 20 30 40 50
Frequency (Hz)

Fig. 4 Ordinary coherence function and frequency response function

always approaches the expected value from the upper side. This is described in
Table 3 [1].
Note that a high value of coherence (0.9) after 16 averages has approximately
the same possible variance of the frequency response function as a low value of
coherence (0.3) after 256 averages.
Two special cases of low coherence are worth particular mention. The first
situation occurs when a leakage error occurs in one or both of the input and output
274 A. W. Phillips and R. J. Allemang

Table 3 Ordinary coherence relationship – averaging (±|H | dB, ±φ ◦ )

Measured value Number of averages of coherence function
16 32 64 128 256
+5.2 dB +3.8 dB +2.8 dB +2.1 dB +1.5 dB
0.2 −14.6 dB −7.1 dB −4.2 dB −2.7 dB −1.8 dB
(±54◦ ) (±34◦ ) (±23◦ ) (±16◦ ) (±11◦ )
+4.2 dB +3.1 dB +2.2 dB +1.6 dB +1.2 dB
0.3 −8.4 dB −4.8 dB −3.0 dB −2.0 dB −1.4 dB
◦
(±38 ) (±25 )◦ (±17◦ ) (±12◦ ) (±8◦ )
+3.5 dB +2.6 dB +1.8 dB +1.3 dB +1.0 dB
0.4 −6.0 dB −3.6 dB −2.3 dB −1.6 dB −1.1 dB
(±30◦ ) (±20◦ ) (±14◦ ) (±10◦ ) (±7◦ )
+3.0 dB +2.1 dB +1.5 dB +1.1 dB +0.8 dB
0.5 −4.5 dB −2.8 dB −1.9 dB −1.3 dB −0.9 dB
(±24◦ ) (±16◦ ) (±11◦ ) (±8◦ ) (±5◦ )
+2.5 dB +1.8 dB +1.3 dB +0.9 dB +0.7 dB
0.6 −3.5 dB −2.2 dB −1.5 dB −1.0 dB −0.7 dB
(±19◦ ) (±13◦ ) (±9◦ ) (±6◦ ) (±4◦ )
+2.1 dB +1.5 dB +1.0 dB +0.7 dB +0.5 dB
0.7 −2.7 dB −1.7 dB −1.2 dB −0.8 dB −0.6 dB
(±15◦ ) (±10◦ ) (±7◦ ) (±5◦ ) (±4◦ )
+1.6 dB +1.1 dB +0.8 dB +0.6 dB +0.4 dB
0.8 −2.0 dB −1.3 dB −0.9 dB −0.6 dB −0.4 dB
(±12◦ ) (±8◦ ) (±6◦ ) (±4◦ ) (±3◦ )
+1.1 dB +0.8 dB +0.5 dB +0.4 dB +0.3 dB
0.9 −1.3 dB −0.8 dB −0.6 dB −0.4 dB −0.3 dB
(±8◦ ) (±5◦ ) (±4◦ ) (±3◦ ) (±2◦ )
90% confidence limits on the measurement of the amplitude |H | and phase φ of transfer functions,
as a function of the measured value of coherence and the number of averages

measurements. This causes the coherence in the area of the peaks of the frequency
response to be less than unity. This error can be reduced by the use of weighting
functions or by cyclic averaging. The second situation occurs when a significant
propagation time delay occurs between the input and output as may be the case
with acoustic measurements. If a propagation delay of length t is compared to a
sample function length of T , a low estimate of coherence will be estimated as a
function of the ratio t/T . This propagation delay causes a bias error in the frequency
response and should be removed prior to computation if possible. Coherence will be
explained in more detail in Sect. 3.4 once the concept of multiple inputs is discussed.

3.3 Multiple Input FRF Estimation

Multiple input estimation of frequency response functions is desirable for several

reasons. The principal advantage is the increase in the accuracy of estimates of the
6 Frequency Response Function Estimation 275

frequency response functions. During single input excitation of a system, there may
exist large differences in the amplitudes of vibratory motion at various locations
because of the dissipation of the excitation power within the structure. This is
especially true when the structure has heavy damping. Small nonlinearities in the
structure will consequently cause errors in the measurement of the response. With
multiple input excitation, the vibratory amplitudes across the structure typically will
be more uniform, with a consequent decrease in the effect of nonlinearities.
A second reason for improved accuracy is the increase in consistency of the
frequency response functions compared to the single input method. When a number
of exciter systems are used, the elements from columns of the frequency response
function matrix corresponding to those exciter locations are being determined
simultaneously. With the single input method, each column is determined indepen-
dently, and it is possible for small errors of measurement due to nonlinearities and
time-dependent system characteristics to cause a change in resonance frequencies,
damping, or mode shapes among the measurements in the several columns. This is
particularly important for the polyreference modal parameter estimation algorithms
that use frequency response functions from multiple columns or rows of the
frequency response function matrix simultaneously.
An additional, significant advantage of the multiple input excitation is a reduction
of the test time. In general, using multiple input estimation of frequency response
functions, frequency response functions are obtained for all input locations in
approximately the same time as required for acquiring a set of frequency response
functions for one of the input locations, using a single input estimation method.
Another potential advantage of the simultaneous measurement of a number of
columns or rows of the frequency response function matrix is the ability to use a
linear combination of frequency response functions in the same row of the matrix
in order to enhance specific modes of the system. This technique is analogous to the
forced normal mode excitation experimental modal analysis in which a structure is
excited by a forcing vector which is proportional to the modal vector of interest. For
this analysis, the coefficients of a preliminary experimental modal analysis are used
to weight the frequency response functions, so that the sum emphasizes the modal
vector that is sought. The revised set of conditioned frequency response functions
is analyzed to improve the accuracy of the modal vector. A simple example of this
approach for a structure with approximate geometrical symmetry would be to excite
at two symmetric locations. The sum of the two frequency response functions at
a specific response location should enhance the symmetric modes. Likewise, the
difference of the two functions should enhance the antisymmetric modes.

3.3.1 Multiple Input Versus Single Input

Advantages:

• Better energy distribution reduces nonlinearities at input location.

• Better energy distribution excites the structure more evenly.
• Data collected simultaneously has consistent frequency and damping information
which is consistent with parameter estimation algorithms.
276 A. W. Phillips and R. J. Allemang

• Advances in hardware/software have kept data collection time the same for single
input/multiple output. More measurements per measurement cycle.
• Multiple input data permits the detection of repeated or closely spaced roots.

Disadvantages:

• Inputs must not be correlated.

• More equipment required.

The theoretical basis of multiple input frequency response function analysis is

well documented in a number of sources [1, 2, 3, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
19, 20]. While much had been written about multiple input theory, the application
of multiple input theory to experimental modal analysis apparently had not been
seriously investigated prior to 1980 [9,10,11,12,13,14,15,16,17,18,19,20]. It also
needs to be noted that this application of multiple input, multiple output (MIMO)
theory represents a very special case of multiple input, multiple output data analysis.
For this case, everything about the inputs is known or can be controlled. The
number of inputs, the location of the inputs, and the characteristics of the inputs are
controlled by the test procedure. For the general case, none of these characteristics
may be known.
Consider the case of Ni inputs and No outputs measured during a modal test on
a dynamic system as shown in Fig. 5. The model assumed for the dynamics is:

Ni
X̂p − ηp = Hpq ∗ ( F̂q − υq ) (37)
q=1

where:

• F = F̂ − υ = actual input
• X = X̂ − η = actual output
• X̂p = spectrum of the p − th output, measured
• F̂q = spectrum of the q − th input, measured
• Hpq = frequency response function of output p with respect to input q
• υq = spectrum of the noise part of the input
• ηp = spectrum of the noise part of the output
• Xp = spectrum of the p − th output, theoretical
• Fq = spectrum of the q − th input, theoretical

In order to develop an estimation of the frequency response function for the

multiple input case, a number of averages Navg will be used to minimize the random
errors (variance). This can be easily accomplished through use of intermediate
measurement of the auto and cross power spectra as defined in Equations 19, 20,
21 and 22. Additional matrices, constructed from the auto and cross power spectra
need to be defined as follows. Note that each function and, therefore, each resulting
matrix is a function of frequency.
6 Frequency Response Function Estimation 277

Fig. 5 System model: multiple inputs

Input/Output Cross Spectra Matrix:

⎧ ⎫
⎪ X1 ⎪ ⎡ ⎤
⎪
⎪ ⎪
⎪ GXF11 · · · GXF1Ni
⎪
⎪ ⎪
⎪
⎨ X2 ⎬   ⎢ ⎥
⎢ · · · ⎥
[GXF ] = {X} {F }H = · F1∗ F2∗ · · FN∗ i = ⎢ ⎥
⎪
⎪ ⎪ ⎣ · · · ⎦
⎪
⎪ · ⎪⎪
⎪
⎪
⎩X ⎪ ⎭ GXFNo 1 · · · GXFNo Ni
No
(38)

Input Cross Spectra Matrix:

⎧ ⎫
⎪ F1 ⎪ ⎡ ⎤
⎪
⎪ ⎪
⎪ GF F11 · · · GF F1Ni
⎪
⎪ ⎪
⎨ F2 ⎪⎬  ⎢ ⎥
⎢ · · · ⎥
[GF F ] = {F } {F }H = · F1∗ F2∗ · · FN∗ i = ⎢ ⎥
⎪
⎪ ⎪ ⎣ · · · ⎦
⎪ · ⎪
⎪ ⎪
⎪
⎪
⎩F ⎭ ⎪ GF F Ni 1 · · · GF FNi Ni
Ni
(39)
The frequency response functions can now be estimated for the three algorithms
as follows:

3.3.2 H1 Algorithm: Minimize Noise on Output ( η )

[H ]No × Ni {F }Ni × 1 = {X}No × 1 − {η}No × 1 (40)

[H ] {F } {F } H
= {X} {F }H
− {η} {F } H
(41)
278 A. W. Phillips and R. J. Allemang

Since the noise vector {η} is assumed to be independent of the excitation force
vector {F }, thus with averaging, the {η} {F }H cross term approaches zero:

[H ]No × Ni {F }Ni × 1 {F }H
1 × Ni = {X}No × 1 {F }1 × Ni
H
(42)

The above relationship can be concisely stated as:

[H ] [GF F ] = [GXF ] (43)

[H ] = [GXF ] [GF F ]−1 (44)

where:

• [ ]H = complex conjugate transpose (Hermitian matrix)

In the experimental procedure, the input and response signals are measured,
and the averaged cross spectra and auto spectra necessary to create the [GXF ] and
[GF F ] matrices are computed. If the computation of ordinary, multiple, or partial
coherence functions will be required, then the diagonal elements of the output cross
spectrum matrix [GXX ] must be computed also. Note that Equation 44 involves a
matrix inverse or equivalent numerical procedure on the [GF F ] matrix. This will
only be possible when the number of averages is equal to or greater than the number
of inputs.
Equation 43 is valid regardless of whether the various inputs are correlated.
Unfortunately, there are a number of situations where the input cross spectrum
matrix [GF F ] may be singular for specific frequencies or frequency intervals.
When this happens, the inverse of [GF F ] will not exist, and Equation 44 cannot
be used to solve for the frequency response function at those frequencies or in
those frequency intervals. A computational procedure that solves Equation 44 for
[H ] should therefore monitor the rank of the matrix [GF F ] that is to be inverted
and desirably provide direction on how to alter the input signals or the use of the
available data when a problem exists. The current approach for evaluating whether
the inputs are sufficiently uncorrelated at each frequency involves determining the
principal/virtual forces using principal component analysis [7]. This will be covered
later in Sect. 3.5.
Note that recently an FRF development based upon a Cholesky decomposition
has been developed [8]. This formulation is equivalent to the H1 formulation.

3.3.3 H2 Algorithm: Minimize Noise on Input (υ)


[H ]No × Ni {F }Ni × 1 − {υ}Ni × 1 = {X}No × 1 (45)
[H ] { {F } − {υ} } {X} H
= {X} {X} H
(46)
6 Frequency Response Function Estimation 279

Since the noise vector {υ} is assumed to be independent of the response vector
{X}, thus with averaging, the {υ} {X}H cross term approaches zero:

[H ]No × Ni {F }Ni × 1 {X}H

1 × No = {X}No × 1 {X}1 × No
H
(47)

One problem with using the H2 algorithm is that the solution for [H ] can only
be found directly using an inverse when the number of inputs Ni and number of
outputs No are equal. Then:

[H ] [GXF ] = [GXX ] (48)

[H ] = [GXX ] [GXF ]−1 (49)

3.3.4 Hv Algorithm: Minimize Noise on Input and Output (υ and η)


[H ]No × Ni {F }Ni × 1 − {υ}Ni × 1 = {X}No × 1 − {η}No × 1 (50)
[H ] { {F } − {υ} } = {X} − {η} (51)

The solution for [H ] is found by the eigenvalue decomposition of one of the

following two matrices:
 
  [GF F ] GXFq
GF F Xp =    (52)
GF Xp GXXp (Ni +1)×(Ni +1)

  
  G GF Xp
GXF Fp =  XXp  (53)
GXFq [GF F ] (Ni +1)×(Ni +1)

Therefore, the eigenvalue decomposition would be:

 
GF F Xp = [V ]   [V ]H (54)

Or:
 
GXF Fp = [V ]   [V ]H (55)

where:

•   = diagonal matrix of eigenvalues

Solution for the p − th row of the [H ] matrix is found from the eigenvector
associated with the smallest (minimum) eigenvalue. Note that the size of the
eigenvalue problem is Ni + 1 and that the eigenvalue solution must be repeated for
each frequency. Note also that the complete solution process must be repeated for
each response point Xp .
280 A. W. Phillips and R. J. Allemang

The frequency response functions associated with a single output p and all inputs
are found by normalizing the eigenvector associated with the smallest eigenvalue.
If [GF F Xp ] is used, the eigenvector associated with the smallest eigenvalue must be
normalized as follows:
⎧ ⎫
⎪
⎪ Hp1 ⎪⎪
⎪
⎪ ⎪
⎪
⎪
⎪ H ⎪
⎪
⎪
⎨
p2 ⎪
⎬
·
{V }λmin = (56)
⎪
⎪ · ⎪⎪
⎪
⎪H ⎪
⎪
⎪
⎪ pNi ⎪
⎪
⎪
⎩ ⎪
⎭
−1

If [GXF Fp ] is used, the eigenvector associated with the smallest eigenvalue must
be normalized as follows:
⎧ ⎫
⎪
⎪ −1 ⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎪ Hp1 ⎪⎪
⎪
⎪
⎨ ⎬
Hp2
{V }λmin = (57)
⎪
⎪ · ⎪⎪
⎪
⎪ ⎪
⎪
⎪
⎪ · ⎪⎪
⎪
⎪
⎩ ⎭
HpNi

Note that there are several decompositions of the [GF F X ] or [GXF F ] matrices,
including SVD and ED, that have been used in the development of Hv . The original,
historical presentation/development used the eigenvalue decomposition as presented
in this chapter.
Recent developments have shown that there is an alternative formulation, Hsvd ,
which gives equivalent results to the presented Hv form [21].

3.4 Coherence: Ordinary, Multiple, and Conditioned

The concept of the coherence function, as defined for single input measurement
(Sect. 3.2.4), needs to be expanded to include the variety of additional relationships
that are possible for multiple inputs.

3.4.1 Ordinary Coherence

Ordinary coherence is defined in this general sense as the correlation coefficient
describing the linear relationship between any two spectra. This is consistent with
the ordinary coherence function that is defined for single input, single output
measurements. Great care must be taken in the interpretation of ordinary coherence
when more than one input is present. The ordinary coherence of an output with
respect to an input can be much less than unity even though the linear relationship
between inputs and outputs is valid, because of the influence of the other inputs.
6 Frequency Response Function Estimation 281

The ordinary coherence function can be formulated in terms of the elements of

the matrices defined previously. The ordinary coherence function between the pth
output and the qth input can be computed from the following formula:

Ordinary Coherence Function

 
GXF 2
OCOHpq = γpq =
2 pq
(58)
GF Fqq GXXpp

where:

• GXXpp = auto power spectrum of the output p

• GF Fqq = auto power spectrum of the input q
• GXFpq = cross power spectrum between output p and input q

3.4.2 Multiple Coherence

Multiple coherence is defined as the correlation coefficient describing the linear
relationship between an output and all known inputs. There is a multiple coherence
function for every output. Multiple coherence can be used to evaluate the importance
of unknown contributions to each output. These unknown contributions can be mea-
surement noise, nonlinearities, or unknown inputs. Particularly, as in the evaluation
of ordinary coherence, a low value of multiple coherence near a resonance will
often mean that the leakage error is present in the frequency response function.
Unlike the ordinary coherence function, a low value of multiple coherence is not
expected at antiresonances. The antiresonances for different input and the same
response locations do not occur at the same frequency. Though one response signal
may have a poor signal-to-noise ratio at its antiresonance, other inputs will not have
at the same frequency.
The formulation of the equations for the multiple coherence functions can be
simplified from the normal computational approach to the following equation.

Multiple Coherence Function:

{GXF } [GF F ]−1 {GF X }T

MCOHp = (59)
GXXpp

Frequently, the multiple coherence is expressed in various references in terms of

the computed or estimated frequency response functions as in Equation 60.


Ni 
Ni ∗
Hpq GF Fqt Hpt
MCOHp = (60)
GXXpp
q=1 t=1
282 A. W. Phillips and R. J. Allemang

where:

• Hpq = frequency response function for output p and input q

• Hpt = frequency response function for output p and input t
• GF Fqt = cross power spectrum between input q and input t
• GXXpp = auto power spectrum of output p

However, what is not necessarily obvious is that the formulation expressed in

Equation 60 is only valid if the frequency response is computed using the H1
estimator, and therefore, Equation 60 should generally be avoided.
If the multiple coherence of the p −th output is near unity, then the p −th output
is well predicted from the set of inputs using the least squares frequency response
functions.
It should be noted that there is only one multiple coherence definition, Equa-
tion 59, which is always correct; however, the multiple coherence can also be
computed or estimated from conditioned coherences as shown in the following
section.

3.4.3 Conditioned Coherence

An important feature, characteristic, or premise of coherence in general is that
it identifies either the individual contribution of an input to an output (as in
ordinary coherence) or the collective input contribution to an output (as in multiple
coherence). In addition, its value must be bounded 0 to 1. [22]
Historically, the conditioned coherence, called partial coherence, was defined
as the ordinary coherence between a conditioned output and another conditioned
output, between a conditioned input and another conditioned input, or between a
conditioned input and a conditioned output. The output and input are conditioned
by removing contributions from other input(s). The removal of the effects of the
other input(s) is formulated on a linear least squares basis. The order of removal
of the inputs during conditioning has a definite effect upon the partial coherence if
some of the inputs are mutually correlated. Thus, there will be a partial coherence
function for every input/output, input/input, and output/output combination for all
permutations of conditioning.
Unfortunately, this variability in the order of conditioning for the partial coher-
ence, especially with respect to outputs, has limited its value and applicability to
experimental FRF estimation because the combination of these coherences for a
given output do not add up to be the multiple coherence for that output.
However, recent reformulations of the partial coherence where conditioning is
restricted to the input-input GF F matrix and the order of conditioning is determined
by the relative importance or significance of the input contributor to the output
have proved valuable. Particularly, since when appropriately conditioned, the partial
coherence can be computed using a traditional ordinary coherence calculation, and
the combination of these conditioned coherences (partial coherences in this case)
for a given output does add up to be the multiple coherence for that output.
6 Frequency Response Function Estimation 283

Table 4 Summary of Conditioned coherence models

conditioned coherence
Name Conditioning Reference DOF
models
Partial coherence CD Physical
Cumulative coherence CD Physical
Fractional coherence SVD Virtual
Virtual coherence SVD or ED Virtual

In the following sections, two different classes of conditioned coherences are

developed, the principal difference between them being the type of conditioning
applied to the GF F matrix: the first which results in coherence functions which
can be referenced back to physical input degrees of freedom and the second which
results in coherence functions which correspond to the virtual input degrees of
freedom which are found by principal component analysis (Sect. 3.5). A comparison
of the characteristics of the conditioned coherences is presented in Table 4.
While the individual coherence functions identified by these methods cannot be
compared between methods, for both approaches, the functions can be summed to
yield the traditional multiple coherence function defined in Sect. 3.4.2.

Conditioned Coherences Which Retain Physical Source Reference

The following two coherence formulations retain the connection to a physical source
reference. The order and value of the individual coherence components are not
necessarily the same, but each still sums to give the multiple coherence.

Partial Coherence
Thus, one way to study the linear correlation of the output with one of the inputs
in a MISO/MIMO system is to calculate the corresponding partial coherence by
arranging the inputs in a predetermined order, usually based on their significance.
They are then sequentially conditioned so that each new input is completely
uncorrelated from its preceding inputs. Once the order has been determined, the
partial coherence of the output with the first force is obviously the same as the
corresponding ordinary coherence. For the second force, the conditioned input auto
power is found as shown in Equation 63.

GF2 F1
GF2 F2 .F1 = GF2 F2 − GF1 F2 (61)
GF1 F1
GF2 F1
GF2 F2 .F1 = GF2 F2 − GF1 F2 (62)
GF1 F1
GF2 F2 .F1 = [1 − γF21 F2 ]GF2 F2 (63)

The subscript on the left hand side signifies that the auto power computed for
the second input in this fashion does not have any correlation with the first input.
Each subsequent input is conditioned in the same way, and the partial coherence
284 A. W. Phillips and R. J. Allemang

of the output with any arbitrary conditioned input can be computed as shown in
Equation 64.

|GFq Xp .Fq−1 ! |2
PCOHpq = γF2q Xp .Fq−1 ! = (64)
GFq Fq .Fq−1 ! GXp Xp

The term Fq−1 ! signifies that the uncorrelation takes place from input 1 through
q −1 to find the conditioned input q. Adding the partial coherences for all the inputs
results in the corresponding multiple coherence. The above formulation was shown
in [23] which is the fourth edition of this textbook. (In the first two editions, the
output term in the denominator on the right hand side used a conditioned output
spectrum for normalization instead. When a conditioned output spectrum is used,
the sum of the conditioned coherences is not the same as the corresponding multiple
coherence. This disambiguation is necessary, as several follow-up works [24, 8]
that made use of the older editions of the textbook as reference include the latter
development which is inconsistent.)
Note that the corresponding set of partial coherences is dependent on the ordering
of the inputs chosen.

Cumulative Coherence
In [8], a use of Cholesky decomposition is demonstrated to obtain the conditioned
inputs efficiently. In [25], the input auto power matrix is decomposed into a lower
triangular matrix L and a diagonal matrix C multiplied to the former’s Hermitian
as shown in Equation 65. Here, the diagonal matrix C is representative of the
conditioned inputs.

GF F = LF C C LH
FC (65)

A relation between the output spectra with the conditioned input spectra can be
computed as described in Equation 66.

−1 −1
LXC = GXF [LH
F C ] GCC (66)

The cross spectrum between the output and the conditioned inputs can then be
found as:

GXC = LXC GCC LH

XC (67)

The output-by-output cumulative coherence can then be found as shown in

Equation 68.

|GCq Xp |2
CCOHpq = γF2q Xp .Fq−1 ! = (68)
GCq Cq GXp Xp
6 Frequency Response Function Estimation 285

The corresponding set of cumulative coherences is dependent on the ordering

of the inputs chosen. A variant of this concept that includes restoring force terms
representing any nonlinear functional terms that describe the system behavior was
developed in [26] also named cumulative coherence.

Conditioned Coherences Which Utilize Virtual Source Reference

The following two coherence formulations do not retain the connection to a physical
source reference. Instead, they resolve the coherence to virtual source inputs.
While the order of the virtual sources is not necessarily the same between the two
formulations, each still sums to give the multiple coherence.

Fractional Coherence
Instead of depending on some a priori knowledge of the significance of the inputs,
a singular value decomposition [8] of the input auto power matrix allows for the
computation of the partial dependency of the output with the virtual forces thus
obtained. This metric does not retain the spatial information and always orders
the resulting virtual forces in the order of their significance and is, therefore,
independent of the ordering of the inputs chosen. In the same vein as the Cholesky
decomposition shown at the end of previous section, the relation between the output
and the virtual forces can be used to find the corresponding fractional coherence as
shown in Equation 71.
GF F = UF S GSS UFHS (69)
 −1
GXS = GXF UFHS G−1
SS (70)

|GSq Xp |2
FCOHpq = γF2q Xp .Fq−1 ! = (71)
GSq Sq GXp Xp

The discussion regarding the disambiguation of the normalization factor pre-

sented for partial coherence holds true here as well. If the fractional coherence is
computed as presented, it can be added up to compute the corresponding multiple
coherence.

Virtual Coherence
As a coherence computed in conjunction with, or as a correspondence to, the
virtual forces (or principal components) defined in Sect. 3.5, the virtual coherence
can be computed using either singular value decomposition [27] or eigenvalue
decomposition [28].
GF F = FV GV V H
FV (72)
H −1 −1
GXV = GXF [ F V ] GV V (73)
|GVq Xp |2
VCOHpq = γF2q Xp .Fq−1 ! = (74)
GVq Vq GXp Xp
286 A. W. Phillips and R. J. Allemang

If the virtual coherence is computed as presented, it can be added up to compute

the corresponding multiple coherence.
In summary, all the conditioned coherence formulations presented utilize some
form of orthogonal conditioning basis and as a result are consistent with the
fundamental coherence premise presented: bounded 0 to 1 and sum to be the
multiple coherence.

Two DOF Illustration

To begin to understand the size of the problem involved, start with the two-input,
one-output case (Fig. 6).

X̂p − ηp = Hp1 F1 + Hp2 F2 (75)

If more than one output is measured, the equations become:

 
 F1
Xp [F1∗ F2∗ ] = [Hp1 Hp2 ] [F1∗ F2∗ ] (76)
F2

Therefore, for input locations 1 and 2, each output is used with the two inputs
to compute two frequency response functions. Therefore, there will be 2 × No
frequency response functions to be computed:

Fig. 6 Two-input,
one-output model
6 Frequency Response Function Estimation 287

⎡ ⎤ ⎡ ⎤
H11 H12 GXF11 GXF12
⎢ H ⎥ ⎢ GXF22 ⎥
⎢ 21 H22 ⎥ ⎢ GXF21 ⎥
⎢ ⎥ ⎢ ⎥ −1
⎢ H31 H32 ⎥ ⎢ GXF31 GXF32 ⎥ GF F11 GF F12
⎢ ⎥=⎢ ⎥ (77)
⎢ · · ⎥ ⎢ · · ⎥ GF F21 GF F22
⎢ ⎥ ⎢ ⎥
⎣ · · ⎦ ⎣ · · ⎦
HNo 1 HNo 2 GXFNo 1 GXFNo 2

For each output location, one formulation of the equations to be solved can
be developed by replacing the inverse of the [GF F ] matrix with the equivalent
adjoint of the [GF F ] matrix divided by the determinant of the [GF F ] matrix. In
this way, it is clear that the frequency response functions can be found as long as the
determinant of the [GF F ] matrix is not zero:

GXFp1 GF F22 − GXFp2 GF F21

Hp1 = (78)
det[GF F ]

GXFp2 GF F11 − GXFp1 GF F12

Hp2 = (79)
det[GF F ]

where:

• det[GF F ] = determinant of [GF F ] matrix

• det[GF F ] = GF F11 GF F22 − GF F21 GF F12

For the two-input, one-output case, several possible coherence functions can be
formulated. While the ordinary coherence between the output and each input can be
formulated, these coherence functions may not provide useful information due to
the possible interaction between the two forces.

Ordinary coherence (output p and Input 1):

 
GXF 2
OCOHp1 =
p1
(80)
GF F11 GXXpp

Ordinary coherence (output p and input 2):

 
GXF 2
OCOHp2 =
p2
(81)
GF F22 GXXpp

The ordinary coherence between the two inputs is a useful function since this is
a measure of whether the forces are correlated. If the forces are perfectly correlated
at a frequency, the inverse of the [GF F ] matrix will not exist, and the frequency
response functions cannot be estimated at that frequency. In this case, the ordinary
coherence between the two forces cannot be unity, although values from 0.0 to 0.99
288 A. W. Phillips and R. J. Allemang

are theoretically acceptable. The limit is determined by the accuracy of the measured
data and the numerical precision of the computation.

Ordinary Coherence (Input 1 and Input 2):

 
GF F 2
OCOH12 = 12
(82)
GF F11 GF F22

Multiple coherence is always a good measure of whether the output response

is caused by the combination of the measured inputs. Multiple coherence is used
in multiple input situations in the same way that ordinary coherence is used in the
single input situations.

Multiple Coherence
{GXF } [GF F ]−1 {GF X }T
MCOHp = (83)
GXXpp

Summary of Methods
H1 Technique:

• Underestimates amplitude at resonances. Causes damping to be overestimated.

• Underestimates amplitude at antiresonances.

H2 Technique:

• Overestimates amplitude at resonances. Causes damping to be underestimated.

• Overestimates amplitude at antiresonances.

Hv Technique:

• Best estimate of amplitude at resonances. Causes damping to be estimated best.

• Best estimate of amplitude at antiresonances.
• Phase characteristics not altered.

3.5 Multiple Input Force Analysis/Evaluation

Of the variety of situations that can cause difficulties in the computation of the
frequency response functions, the highest potential for trouble is the case of coherent
inputs. If two of the inputs are fully coherent at one of the analysis frequencies, then
there are no unique frequency response functions associated with those inputs at that
analysis frequency. Unfortunately, there are a number of situations where the input
cross spectrum matrix [GF F ] may be singular at specific frequencies or frequency
intervals. When this happens, the inverse of [GF F ] will not exist, and Equation 44
cannot be used to solve for the frequency response function at those frequencies
6 Frequency Response Function Estimation 289

or in those frequency intervals. First, one of the input autospectra may be zero in
amplitude over some frequency interval. When this occurs, then all of the cross
spectra in the same row and column in the input cross spectrum matrix [GF F ] will
also be zero over that frequency interval. Consequently, the input cross spectrum
matrix [GF F ] will be singular over that frequency interval. Second, two or more of
the input signals may be fully coherent over some frequency interval. Although
the signals used as inputs to the exciter systems must be uncorrelated random
inputs, the response of the structure at resonance, combined with the inability to
completely isolate the exciter systems from this response, results in the ordinary
or conditioned partial coherence functions with values other than zero, particularly
at the system poles. For example, for the two-input case, as long as the coherence
function between the inputs is not unity at these frequencies, Equation 44 can be
solved uniquely for the frequency response functions.
Note that the auto and cross spectra involved in the calculation of the multiple
input case for the estimation of frequency response functions should be computed
from analog time data that has been digitized simultaneously. If data is not processed
in this manner, many more averages are required to reduce the variance on each
individual auto and cross spectrum, and the efficiency of the multiple input approach
to the estimation of frequency response functions will not be as attractive. Finally,
numerical problems, which cause the computation of the inverse to be inexact, may
be present. This can happen when an autospectrum is near zero in amplitude, when
the cross spectra have large dynamic range with respect to the precision of the
computer, or when the matrix is ill-conditioned because of nearly redundant input
signals.
Due to the form of the equations that must be solved to compute frequency
response functions in the presence of multiple inputs, special care must be taken
to assure that the input spectrum matrix is not singular. Therefore, techniques have
been investigated to evaluate the form of the input spectrum matrix before taking
any data. Singular, in this case, implies that:

• Input forces may not be coherent at any frequency.

– Independent, uncorrelated noise sources must be used (random, random
transient, periodic random).
– The impedance of the structure at the input locations may tend to correlate the
inputs at resonance.
• There are no zeros in the input spectrum matrix.

3.5.1 Ordinary and Partial Coherence Functions

The historical approach that was used to try to evaluate the correlation between the
forces utilized ordinary and partial coherence functions. The ordinary coherence
function measures the degree of linear dependence (or correlation) between the
spectra of two signals. The partial coherence function measures the degree of linear
dependence between the spectra of two signals, after eliminating, in a least squares
sense, the contribution of some other signals. Both functions can be used in a
systematic procedure to verify that the forces are not correlated or that the input
290 A. W. Phillips and R. J. Allemang

cross spectra matrix [GF F ] is not singular. For cases involving more than two inputs,
this approach is very difficult and requires considerable judgment. In reality, only
the ordinary coherence function, for the case of two inputs, is still used.
 
GF F 2
OCOHik = ik
(84)
GF Fii GF Fkk

where:

• GF Fik = cross power spectrum between inputs i and k

• GF Fii = auto power spectrum of input i
• GF Fkk = auto power spectrum of input k

3.5.2 Principal/Virtual Input Forces (Virtual Forces)

The current approach used to determine correlated inputs involves utilizing principal
component analysis to determine the number of contributing forces to the [GF F ]
matrix. In this approach, the matrix that must be evaluated is:
⎡ ⎤
GF F11 · · · GF F1Ni
⎢ · · ⎥
⎢ ⎥
⎢ ⎥
[GF F ] = ⎢ · · ⎥ (85)
⎢ ⎥
⎣ · · ⎦
GF FNi 1 · · · GF FNi Ni

where:

• GF Fik =  GF Fki ∗ (Hermitian matrix)

• GF Fik = Fi Fk∗
• GF F is the power spectrum of a given input.

Principal component analysis involves a singular value or eigenvalue decom-

position of the [GF F ] matrix [7]. Since the eigenvectors of such a decomposition
are unitary, the eigenvalues should all be of approximately the same size if each of
the inputs is contributing. If one of the eigenvalues is much smaller at a particular
frequency, one of the inputs is not present, or one of the inputs is correlated with the
other input(s):

[GF F ] = [V ]  [V ]H (86)

  represents the eigenvalues of the [GF F ] matrix. If any of the eigenvalues

of the [GF F ] matrix are zero or insignificant, then the [GF F ] matrix is singular.
Therefore, for a three-input test, the [GF F ] matrix should have three eigenvalues at
each frequency. (The number of eigenvalues is the number of uncorrelated inputs.)
6 Frequency Response Function Estimation 291

This concept is shown graphically in Fig. 7 for the auto power spectra for a three-
input case. It is difficult to determine if the inputs are mutually correlated from these
plots. Figure 8 shows the principal force plots for the same case. At the frequencies
where the third principal/virtual force drops (lowest curve), this indicates that the
inputs are mutually correlated at those frequencies. This is not apparent from Fig. 7.

3.5.3 Optimum Number Of Inputs

The location and number of inputs have a direct effect on the quality of frequency
response functions that are estimated. This is an area that has not been researched
completely and is still being reviewed. It is clear that beyond some number of inputs,
the return from the investment of more equipment, in the form of inputs, is not
warranted. Some considerations are:

• Two at symmetric locations. Frequency response functions can be added or

subtracted to enhance in-phase or out-of-phase modes.
• To excite as many modes as possible in one test configuration.
– Two vertical and one horizontal on a car.
– One on each wing and one on each horizontal stabilizer, all symmetric, on an
aircraft structure.
• To excite operating conditions.

4 Averaging

The averaging of signals is normally viewed as a summation or weighted summation

process where each sample function has a common abscissa. Normally, the desig-
nation of history is given to sample functions with the abscissa of absolute time,
and the designation of spectra is given to sample functions with the abscissa of
absolute frequency. The spectra are normally generated by Fourier transforming the
corresponding history. In order to generalize and consolidate the concept of signal
averaging as much as possible, the case of relative time can also be considered.
In this way, relative history can be discussed with units of the appropriate event
rather than seconds, and a relative spectrum will be the corresponding Fourier
transform with units of cycles per event. This concept of signal averaging is used
widely in structural signature analysis where the event is a revolution. This kind
of approach simplifies the application of many other concepts of signal relation-
ships such as Shannon sampling theorem and Rayleigh criterion of frequency
resolution.

4.1 General Averaging Methods

When comparing data taken with different equipment, care must be taken to be
certain that the averaging is being performed the same way. The terminology with
292 A. W. Phillips and R. J. Allemang

-4
Gxx : 1z,1z
10

-5
10
Magnitude

-6
10

-7
10

-8
10
0 50 100 150 200 250 300 350 400
Frequency, Hz

-4 Gxx : 2z,2z
10

-5
10
Magnitude

-6
10

-7
10

-8
10
0 50 100 150 200 250 300 350 400
Frequency, Hz

-4 Gxx : 3z,3z
10

-5
10
Magnitude

-6
10

-7
10

-8
10
0 50 100 150 200 250 300 350 400
Frequency, Hz

Fig. 7 Auto power spectrum of input forces

6 Frequency Response Function Estimation 293

-4 Principal Force Analysis

10

-5
10
Magnitude

-6
10

-7
10

-8
10

-9
10
0 50 100 150 200 250 300 350 400
Frequency, Hz

Fig. 8 Principal (virtual) force spectrum

regard to averaging is not always the same, so some sort of evaluation may be
required using test cases to be certain that the same form of averaging is being
used.

4.1.1 Linear Averaging

Linear averaging is the simplest form of averaging and is what most people think
of as averaging. Essentially, linear averaging is simply trying to find the mean
value in a set of numbers. Since measurements deal with information at different
times or frequencies, linear averaging refers to finding the mean value at each
time or frequency over a number of ensembles or averages. Linear averaging is
fundamentally calculated with the following formula:

N
avg
x(t)i
i=1
x̄(t)Navg = (87)
Navg
N
avg
X(ω)i
i=1
X̄(ω)Navg = (88)
Navg

Linear averaging is useful primarily when some form of initial trigger is available
in order for information that is synchronized with the trigger to be emphasized.
Note that if no trigger is available and the averages are collected in a free run data
acquisition mode, the phasing of any dynamic signals will be random from ensemble
to ensemble. With a large number of averages, only the DC signal content will be
preserved. Note that linear averaging gives the same result whether implemented
in the time or frequency domain. Linear averaging is normally implemented in the
frequency domain.
294 A. W. Phillips and R. J. Allemang

If some sort of initializing trigger is available, the averaged data reduce the
noise and enhance the signal, giving an improved signal-to-noise ratio (SNR). The
variance is reduced a function of √ 1 . This means that to reduce the variance to
Navg
10 percent of the variance on a single average, 100 averages must be taken.
The terminology time averaging refers to a special case of linear averaging when
the trigger of each average is synchronized with a specific position of a rotating
shaft (e.g., top dead center) in the time domain. In this case, each ensemble will
have a fixed number of rotations in the time history, and each data point in the time
history will be collected when the rotating system is in the same position as long
as the speed of rotation is constant. If the data is sampled at fixed intervals during
the rotation (e.g., by utilizing an encoder to give 32 samples per revolution), the
fixed speed is not required. The processing of data for this situation requires further
consideration and will not be presented here.

4.1.2 Magnitude Averaging

Magnitude, or amplitude, averaging involves finding the mean value of the absolute
values of the data at each time or frequency. While this form of averaging is
not very common, it has been used in some older digital signal analyzers. This
form of averaging has generally been replaced by RMS averaging in most current
digital signal analyzers. Magnitude averaging and RMS averaging will give nearly
identical results if there is little dynamic range in the data being averaged. Since the
absolute value is formed, the phasing provided by an initial trigger is not required for
magnitude averaging. However, magnitude averaging does not improve the signal-
to-noise ratio (SNR) since the noise magnitude accumulates in the same way as the
signal magnitude. Therefore, magnitude averaging does not reduce the variance in
the data with averages in the same way as linear averaging. Magnitude averaging is
fundamentally calculated with the following formula:

avg 
N
x(t)i × x(t)i
i=1
x̄(t)Navg = (89)
Navg
avg 
N
X(ω)i × X(ω)i ∗
i=1
X̄(ω)Navg = (90)
Navg

Magnitude averaging is generally not a concern unless historical data or data

acquisition procedures are involved. This can be a concern if data or data acquisition
procedures are specified, particularly in patents or other historical documents. Note
that magnitude averaging does not give the same result whether implemented in the
time or frequency domain. Magnitude averaging is normally implemented in the
frequency domain.
6 Frequency Response Function Estimation 295

4.1.3 Root Mean Square (RMS) Averaging

Root mean square (RMS) averaging is commonly used in many digital signal
analyzers and is the basis for estimating frequency response and coherence functions
from auto and cross power spectra. RMS averaging refers to the computational
procedure involving the mean squared value of the data at each time or frequency.
This is useful for determining the power, or energy, in the data which may contain
positive and negative values and/or real and complex values. Since the squared
value is formed, the phasing provided by an initial trigger is not required for RMS
averaging. However, RMS averaging does not improve the signal-to-noise ratio
(SNR) since the noise power accumulates in the same way as the signal power.
Therefore, RMS averaging does not reduce the variance in the data with averages
in the same way as linear averaging. In the estimation of the frequency response
function, the variance error is reduced by choosing the FRF estimation algorithm
based upon the location of the noise in the inputs and outputs. Magnitude averaging
and RMS averaging will give nearly identical results if there is little dynamic range
in the data being averaged. RMS averaging is fundamentally calculated with the
following formula:

N
 avg
 x(t)i × x(t)i

 i=1
x̄(t)Navg = (91)
Navg

N
 avg
 X(ω)i × X(ω)i ∗

 i=1
X̄(ω)Navg = (92)
Navg

Note that in the above equations, the units on the averaged data are the same
as the units of each average. Frequently, digital signal analyzers show the resulting
averaged data with units squared. This is simply indicating that the averaged data is
being displayed without the square root. This is simply a display issue and the user
can choose between units and units squared. Note that RMS averaging does not
give the same result whether implemented in the time or frequency domain. RMS
averaging is normally implemented in the frequency domain.
As a simple example of the difference between linear, magnitude, and RMS
averaging, Table 5 indicates the problem.

4.1.4 Exponential Averaging

Exponential averaging weights new data differently (typically more heavily) than
old data. This is useful for tracking time-varying characteristics in the data (not used
in data for experimental modal analysis). The weighting is generally dependent on
the number of averages chosen for the exponential averaging (typically a power of
2, either 4 or 8). Once the exponential averaging is started, the averaging continues
until it is stopped (it does not stop after the number of averages selected; the
296 A. W. Phillips and R. J. Allemang

Table 5 Comparison of Ensemble Linear Magnitude RMS

averaging methods
1 2 2 4
2 3 3 9
3 4 4 16
4 −2 2 4
5 −3 3 9
6 −4 4 16
Total 0 18 58
Average 0 3.000 3.109

number of averages determines the weighting or forgetting factor). For the first
few averages, linear and exponential averaging is nearly the same. Exponential
averaging is fundamentally calculated with the following formula:

X̄(ω)Navg = (1 − 2−Navg ) X̄(ω)Navg−1 + 2−Navg X(ω)i (93)

4.1.5 Stable Averaging

Stable averaging is not really a separate form of averaging but refers to a display
characteristic. If the averaging process is stopped before Navg are reached, stable
averaging always shows the display information with the appropriate correction
for the number of averages. Stable averaging means that the above equations are
reformulated in a recursive form so that the displayed can be updated, average by
average, so that the amplitude of the data is correct regardless of when the average
is stopped. The recursive form of the averaging equations weights old and new
data records appropriately to yield the appropriate arithmetic mean for the current
number of averages.

4.1.6 Peak Hold

Peak hold data collection is often included in the averaging selection of a digital
signal analyzer. Peak hold, as the name indicates, is not really a form of averaging
since only the peak value is retained at each time or frequency and no arithmetic
mean (in the common understanding of the term) is formed. Normally, the peak
magnitude or peak RMS information is retained over a number of ensembles. This
data collection approach is very useful for identifying the maxima that occur during
transients or general time-varying events.

4.2 Estimation of Frequency Response Functions

The process of signal averaging as it applies to frequency response function

estimation involves linear averaging of auto and cross power spectra. This is a bit
6 Frequency Response Function Estimation 297

confusing since the general averaging concepts always are explained in terms of a
single data signal. With frequency response function estimation, there are a number
of input and output signals that are all initiated (triggered) at the same start time for
each average and are all sampled at the same time. The reduction of noise on the
frequency response function estimates depends upon the noise model that is used
to describe where the noise enters the measurements, on the inputs, on the outputs,
or both. Based upon the noise model, the FRF estimation algorithm will reduce
the random noise according to a least squared error procedure. Note that since the
least squared error procedure minimizes the squared error by eliminating the phase
information, these procedures will affect the magnitude but not the phase.
It is important to realize that while the frequency response function is assumed
to be unique, the auto and cross power spectra used to estimate the FRF are not
unique unless the input is stationary and a sufficiently large number of averages are
taken. Generally, this is never the case. This is not a concern, however, since the
desired information is the frequency response function not auto and cross power
spectra. After a reasonable number of averages (5–100), the auto and cross power
spectra may still appear to be noisy. Some of the noise is due to random or bias
errors in the data, and some of the noise is simply due to the uneven excitation that
occurs with transient or random input(s) to the system being tested. The uneven
excitation in the auto and cross spectra is consistent between the input and output
power spectra (related by the FRF) and will cancel when the frequency response
function is estimated. The random portion of the noise will be minimized due to the
least squared error estimation of the frequency response function. The bias portion
of the noise will not generally be eliminated. Therefore, it is critical that bias errors,
such as leakage, be eliminated if possible.
The triggering issues relative to averaging of auto and cross power spectra that
will be used to estimate FRFs required some additional terminology in order to
clarify the measurement procedure. One of the two forms of linear averaging,
asynchronous and synchronous averaging, is always used to estimate FRFs depend-
ing upon the type of excitation. Additionally, cyclic averaging, a special case of
time domain linear averaging, may be used when leakage is a serious problem
in conjunction with traditional averaging methods (asynchronous or synchronous
averaging). Cyclic averaging reduces the leakage bias error by digitally filtering the
data to eliminate the frequency information that cannot be described by the FFT
(only integer multiples of f are retained) prior to the application of the FFT.
Cyclic averaging can always be used, together with asynchronous or synchronous
averaging, to reduce both the leakage error and the random errors.
Since the Fourier transform is a linear function, there is no theoretical difference
between the use of time or frequency domain data when averaging. Practically,
though, synchronous and asynchronous averaging is normally performed in the
frequency domain and cyclic averaging is normally performed in the time domain.
Therefore, these three classifications primarily refer to the initial trigger and
sampling relationships between averages and ensembles while collecting the auto
298 A. W. Phillips and R. J. Allemang

and cross power spectra used to estimate frequency response functions. This will be
explained further in the following sections.

4.2.1 Asynchronous Signal Averaging

The asynchronous classification of linear signal averaging of the auto and cross
power spectra represents the case where no known relationship exists each average.
The FRF is correctly estimated solely on the basis of the uniqueness of the frequency
response function [29, 30]. In this case, the auto and cross power spectra (least
squares) approach to the estimate of frequency response must be used since no other
way of preserving phase and improving the estimate is available. In this situation,
the trigger to initiate each average takes place in a random fashion dependent only
upon the data acquisition equipment timing. The triggering is said to be in a free
run mode.

4.2.2 Synchronous Signal Averaging

The synchronous classification of linear signal averaging adds the additional
constraint that each average must be initiated with respect to a specific trigger
condition (often the magnitude and slope of the excitation) [31, 32]. This means
that the frequency response function could be formed as X(ω)N avg divided by
F (ω)N avg since phase is preserved. Even so, linear averaging of the auto and cross
power spectra is still the preferred FRF estimation method due to the reduction of
variance and the ability to estimate the ordinary coherence function. The ability
to synchronize the initiation of digitization for each average allows for use of
nonstationary or deterministic inputs with a resulting increased signal-to-noise ratio.
The synchronization takes place as a function of a trigger signal occurring in the
input (internally) or in some event related to the input (externally). An example of
an internal trigger would be the case where an impulsive input is used to estimate
the frequency response. Each average would be initiated when the input reached
a certain amplitude and slope. A similar example of an external trigger would be
the case where the impulsive excitation to a speaker is used to trigger the estimate
of frequency response between two microphones in the sound field. Again, each
average would be initiated when the trigger signal reached a certain amplitude
and slope.

4.2.3 Cyclic Signal Averaging

The cyclic classification of signal averaging involves a special case of linear
averaging with the added constraint that the digitization is coherent between
averages [31, 33, 32, 34]. This means that the exact time between each average is
used to enhance the signal averaging process. Rather than trying to keep track of
elapsed time between averages, the normal procedure is to allow no time to elapse
between successive averages. This simple averaging procedure results in a digital
comb filter in the frequency domain, with the teeth (passbands) of the comb at
frequency increments that are integer multiples of the f = 1/T relationship.
6 Frequency Response Function Estimation 299

The result is an attenuation of the spectrum between the teeth not possible with
other forms of averaging. Cyclic signal averaging is generally performed in the time
domain and is coupled with asynchronous or synchronous averaging procedures in
the frequency domain.
This form of signal averaging is very useful for filtering periodic components
from a noisy signal since the teeth of the filter are positioned at harmonics of
the frequency of the sampling reference signal. This is of particular importance in
applications where it is desirable to extract signals connected with various rotating
members. This same form of signal averaging is particularly useful for reducing
leakage during frequency response measurements.
A very common application of cyclic signal averaging is in the area of analysis of
rotating structures. In such an application, the peaks of the comb filter are positioned
to match the fundamental and harmonic frequencies of a particular rotating shaft or
component. This is particularly powerful, since in one measurement it is possible
to enhance all of the possible frequencies generated by the rotating member from
a given data signal. With a zoom Fourier transform type of approach, one shaft
frequency at a time can be examined depending upon the zoom power necessary to
extract the shaft frequencies from the surrounding noise.
The application of cyclic averaging to the estimation of frequency response
functions can be easily observed by noting the effects of cyclic averaging on a single
frequency sinusoid. Figures 9 and 10 represent the cyclic averaging of a sinusoid
that is periodic with respect to the observation time period T . Figures 11 and 12
represent the cyclic averaging of a sinusoid that is aperiodic with respect to the
observation time period T . By comparing Figs. 10, 11, and 12, the attenuation of
the nonperiodic signal can be clearly observed.

Theory of Cyclic Averaging

In the application of cyclic averaging to frequency response function estimates,
the corresponding fundamental and harmonic frequencies that are enhanced are

1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0 0.5 1 1.5 2 2.5 3 3.5 4

Time Periods (T)

Fig. 9 Contiguous time records (periodic signal)

300 A. W. Phillips and R. J. Allemang

1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time Periods (T)

Fig. 10 Averaged time records (periodic signal)

1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0 0.5 1 1.5 2 2.5 3 3.5 4
Time Periods (T)

Fig. 11 Contiguous time records (nonperiodic signal)

1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time Periods (T)

Fig. 12 Averaged time records (nonperiodic signal)

6 Frequency Response Function Estimation 301

the frequencies that occur at the integer multiples of f . In this case, the spectra
between each f is reduced with an associated reduction of the bias error called
leakage.
The first observation to be noted is the relationship between the Fourier transform
of a history and the Fourier transform of a time shifted history. In the averaging
case, each history will be of some finite time length T which is the observation
period of the data. Note that this time period of observation T determines the
fundamental frequency resolution f of the spectra via the Rayleigh Criteria
(f = T1 ).
The Fourier transform of a history is given by:

+∞
X(ω) = x(t) e− j ωt dt (94)
−∞

Using the time shift theorem of the Fourier transform, the Fourier transform of
the same history that has been shifted in time by an amount t0 is [35]:

+∞
X(ω) e−j ω t0 = x(t + t0 ) e− j ωt dt (95)
−∞

For the case of a discrete Fourier transform, each frequency in the spectra is
assumed to be an integer multiple of the fundamental frequency f = T1 . Making
this substitution in Equation 95 (ω = k 2Tπ with k as an integer) yields:

2π
+∞
X(ω) e−j n T t0
= x(t + t0 ) e− j ωt dt (96)
−∞

Note that in Equation 96, the correction for the cases where t0 = N T with N is
an integer will be a unit magnitude with zero phase. Therefore, if each history that
is cyclic averaged occurs at a time shift, with respect to the initial average, that is
an integer multiple of the observation period T , then the correction due to the time
shift does not affect the frequency domain characteristics of the averaged result. All
further discussion will assume that the time shift t0 will be an integer multiple of the
basic observation period T .
The signal averaging algorithm for histories averaged with a boxcar or uniform
window is:

Nc − 1
1 
x̄(t) = xi (t) (97)
Nc
i=0
302 A. W. Phillips and R. J. Allemang

where:

• Na number of asynchronous averages

• Nc number of cyclic averages

For the case where x(t) is continuous over the time period Nc T , the complex
Fourier coefficients of the cyclic averaged time history become:

1 T
Ck = x̄(t) e− j ωk t dt (98)
T 0

Nc − 1
1 T 1 
Ck = xi (t) e− j ωk t dt (99)
T 0 Nc
i=0

Finally:

c −1
T N
1
Ck = xi (t) e− j ωk t dt (100)
Nc T 0 i=0

Since x(t) is a continuous function, the sum of the integrals can be replaced with
an integral evaluated from 0 to Nc T over the original function x(t). Therefore:

1 Nc T
Ck = x(t) e− j ωk t dt (101)
Nc T 0

The above equation indicates that the Fourier coefficients of the cyclic averaged
history (which are spaced at f = T1 ) are the same Fourier coefficients from the
original history (which are spaced at f = Nc T ). Note that the number of Fourier
coefficients for the cyclic averaged history will be N1c the number of coefficients of
the original history since the number and size of the frequency spacing change by
this factor. Also note that Parseval’s theorem, concerning the energy representation
of each Fourier coefficient, is not preserved by the cyclic averaging process since the
frequency information not related to the harmonics of f = T1 is removed [35].
The approach used to understand the frequency domain effects of windows on
digital data can be used to understand the effect of cyclic averaging [36, 37].
Since cyclic averaging yields the Fourier coefficients of an effectively larger
observation time (Nc T compared to T ), the effect of cyclic averaging results in
an effective frequency domain window characteristic that is a result of this longer
observation time. However, the f axis needs to be adjusted to account for the
actual frequencies that occur in the cyclic averaged spectra.
Figure 13 shows the two-sided frequency domain characteristic of the cyclic
averaged (Nc = 4) case with a uniform window. Likewise, Fig. 14 shows the two-
sided frequency domain characteristic of the cyclic averaged (Nc = 4) case with
6 Frequency Response Function Estimation 303

10 0
Normalized Amplitude

10 −1

10 −2

10 −3

10 −4

10 −5
−4 −3 −2 −1 0 1 2 3 4
Frequency Resolution: (Delta f)

Fig. 13 Cyclic averaging (Nc = 4) with uniform window

0
10
−1
Normalized Amplitude

10
−2
10
−3
10
−4
10
−5
10
−6
10
−4 −3 −2 −1 0 1 2 3 4
Frequency Resolution: (Delta f)

Fig. 14 Cyclic averaging (Nc = 4) with Hanning window

a Hanning window. Further detail of these characteristics is given in Figs. 15, 16,
17, 18, 19, and 20. Figures 15, 16, and 17 show the cyclic averaging effect in the
frequency domain for the cases of 1, 2, and 4 averages with a uniform window
applied to the data. Figure 15 essentially represents no cyclic averaging and is the
familiar characteristic of a uniform window [36, 37]. Figures 16 and 17 show how
cyclic averaging effects this window characteristic with respect to the f = T1
frequency spacing. Figures 18, 19, and 20 show the cyclic averaging effect in the
frequency domain for the cases of 1, 2, and 4 averages with a Hanning window
applied to the original contiguous data. Figure 18 essentially represents no cyclic
averaging and is the familiar characteristic of a Hanning window [36,37]. Figures 19
and 20 show how cyclic averaging effects this window characteristic with respect to
the f = T1 frequency spacing.
These figures demonstrate the effectiveness of cyclic averaging in rejecting
nonharmonic frequencies. Practically, these figures also demonstrate that based
upon effectiveness or the limitations of the dynamic range of the measured data,
304 A. W. Phillips and R. J. Allemang

0
10
Normalized Amplitude

−1
10
−2
10
−3
10
−4
10
−5
10
−6
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 15 Uniform window characteristics

0
10
Normalized Amplitude

−1
10
−2
10
−3
10
−4
10
−5
10
−6
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 16 Cyclic averaging (Nc = 2) with uniform window

0
10
Normalized Amplitude

−1
10
−2
10
−3
10
−4
10
−5
10
−6
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 17 Cyclic averaging (Nc = 4) with uniform window

6 Frequency Response Function Estimation 305

0
10
Normalized Amplitude

−1
10
−2
10
−3
10
−4
10
−5
10
−6
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 18 Hanning window characteristics

0
10
Normalized Amplitude

−1
10
−2
10
−3
10
−4
10
−5
10
−6
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 19 Cyclic averaging (Nc = 2) with Hanning window

0
10
−1
10
Normalized Amplitude

−2
10
−3
10
−4
10
−5
10
−6
10
−7
10
−8
10
0 1 2 3 4 5 6 7 8
Frequency Resolution: (Delta f)

Fig. 20 Cyclic averaging (Nc = 4) with Hanning window

306 A. W. Phillips and R. J. Allemang

0.1

0.05

−0.05

−0.1

−0.15
0 0.5 1 1.5 2 2.5 3 3.5 4
Time Periods (T)

Fig. 21 Contiguous time records

0.05
0.04
0.03
0.02
0.01
0
−0.01
−0.02
−0.03
−0.04
−0.05
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time Periods (T)

Fig. 22 Averaged time records

a maximum of 16 to 32 averages is recommended. Realistically, four to eight cyclic

averages together with a Hanning window provide a dramatic improvement in the
FRF estimate.
The results of cyclic averaging of a general random signal with the application
of a uniform window are shown in Figs. 21 and 22. Likewise, the results of cyclic
averaging of a general random signal with the application of a Hanning window are
shown in Figs. 23 and 24.

Practical Example
The implementation of cyclic signal averaging to frequency response function (FRF)
estimation is not easily applicable to many existing discrete Fourier transform
analyzers. The reason for this is that the user is not given control of the time data
acquisition such that the cyclic averaging requirements can be met. However, many
users currently are acquiring data with personal computer (PC) data acquisition
boards or the VXI-based data acquisition boards where control of the time data
6 Frequency Response Function Estimation 307

0.08
0.06
0.04
0.02
0
−0.02
−0.04
−0.06
−0.08
−0.1
0 0.5 1 1.5 2 2.5 3 3.5 4
Time Periods (T)

Fig. 23 Contiguous time records with Hanning window

0.03

0.02

0.01

−0.01

−0.02

−0.03

−0.04
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time Periods (T)

Fig. 24 Averaged time records with Hanning window

acquisition is more available to the user. In this environment, cyclic averaging is

simple to implement by acquiring long data records and breaking the long data
record into Nc contiguous time records which can be cyclic averaged.
The cyclic averaged inputs and outputs are normally computed by simply
summing successive time records. The important requirement of the successive time
records is that no data is lost. Therefore, these successive time records could be laid
end to end to create the original longer time data record (Nc T ). The cyclic averaged
records are then created by simply adding each time record of length T together in
a block mode.
While the basic approach to cyclic averaging involves using the data weighted
uniformly over the total sample time Nc T , the benefits that can be gained by
using weighting functions can also be applied. The application of a Hanning
window to the successive time records before the summation occurs yields an
even greater reduction of the bias error. Therefore, for frequency response function
measurements, Hanning weighted signal averaging should drastically reduce the
308 A. W. Phillips and R. J. Allemang
1
0.9
0.8
Coherence

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350 400 450 500
Frequency, Hz

101

100
Amplitude

10−1

10−2

0 50 100 150 200 250 300 350 400 450 500

Frequency, Hz

Fig. 25 Case I: asynchronous averaging

leakage errors which can exist when using broadband random excitation techniques
to measure frequency response.
In Figs. 25, 26, 27, and 28, four different measurement cases are documented for
the same FRF measurement. This data was acquired as typical data from a lightly
damped, cantilever beam. Each figure shows the amplitude of an FRF with the
associated ordinary coherence function shown as a measurement quality indicator.
In each case, the H1 FRF estimation algorithm was used; only the windowing and
the number of asynchronous (Na ) and the number of cyclic (Nc ) averages were
changed. The four cases are as follows:

• Case I: The FRF is computed from 64 asynchronous averages (Na = 64).

A uniform window (no additional window) is applied to the data. This is an
unacceptable measurement and represents poor measurement procedure.
• Case II: The FRF is computed from 64 asynchronous averages (Na = 64).
A Hanning window is applied to the data. This is a marginally acceptable
measurement and represents a common measurement procedure.
• Case III: The FRF is computed from 4 cyclic averages (Nc = 4) and 16
asynchronous averages (Na = 16). A uniform window (no additional window)
is applied to the data. This is a marginally acceptable measurement and compares
reasonably to Case II.
• Case IV: The FRF is computed from 4 cyclic averages (Nc = 4) and 16
asynchronous averages (Na = 16). A Hanning window is applied to the data.
This is a good measurement. Note particularly the increase in the FRF amplitude
at the peak frequency locations compared to the three previous cases.
6 Frequency Response Function Estimation 309

1
0.9
0.8
Coherence

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350 400 450 500
Frequency, Hz

101

100
Amplitude

10−1

10−2

0 50 100 150 200 250 300 350 400 450 500

Frequency, Hz

Fig. 26 Case II: asynchronous averaging with Hanning window

1
0.9
0.8
Coherence

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350 400 450 500
Frequency, Hz

101

100
Amplitude

10−1

10−2

0 50 100 150 200 250 300 350 400 450 500

Frequency, Hz

Fig. 27 Case III: cyclic averaging

310 A. W. Phillips and R. J. Allemang

1
0.9
0.8
Coherence

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350 400 450 500
Frequency, Hz

101

100
Amplitude

10−1

10−2

0 50 100 150 200 250 300 350 400 450 500

Frequency, Hz

Fig. 28 Case IV: cyclic averaging with Hanning window

The value of Nc indicates the number of cyclic time records averaged together,
and Na is the number of asynchronous auto and cross spectrum averages: a total
of Nc Na time records were sampled. This is done so that, statistically, the same
amount of independent information is available in each averaging case. Note that
in this example, the data for these cases was acquired only once. Each case results
from processing the original time data differently.
Clearly, the measurement using cyclic averaging with the Hanning window
(Fig. 28) shows a significant reduction of the bias error. An interesting point is that
the data near the antiresonance is also drastically improved due to the sharp roll off
of the line shape of the Hanning weighted averaging.
Cyclic averaging is a powerful digital signal processing tool that minimizes
the leakage error when FRF measurements are being estimated. While existing
discrete Fourier analyzers may not be able to include cyclic averaging for the FRF
estimation case, computer-based data acquisition common to personal computer or
workstation systems generally permits the user to apply cyclic averaging together
with asynchronous or synchronous averaging to effectively minimize both random
errors and the leakage bias error.

4.3 Special Types of Signal Averaging

There are at least two common averaging techniques that use histories which may or
may not overlap. In both cases, the averaging techniques involve processing random
data histories in order to enhance the data.
6 Frequency Response Function Estimation 311

4.3.1 Overlap Processing

The first case is that of overlap processing. Overlap processing involves using
individual sample histories which are not totally independent from one another.
The dependence that occurs results from each successive history starting before the
previous history ends. For the general case where the time data is not weighted in
any fashion, it should be obvious that this averaging procedure does not involve any
new data and, therefore, statistically does not improve the estimation process. In the
special case where weighting functions are involved, this technique can utilize data
that is otherwise ignored. Figure 29a is an example of a data record that has been
weighted to reduce the leakage error using a Hanning weighting function. The data
prior to 20 percent of each sample period and after 80 percent of each sample period

Fig. 29 Overlap processing. (a) Zero overlap. (b) Fifty percent overlap
312 A. W. Phillips and R. J. Allemang

is nearly eliminated by the Hanning window used. Using an overlap factor of at least
20–30 percent as in Fig. 29b involves this data once again in the averaging process.
The second case involving overlapping histories is that of random decrement
analysis [31, 33, 29, 32, 34]. This process involves the overlapping of histories in
order to enhance the deterministic portion of the random record. In general, the ran-
dom response data can be considered to be made up of two parts: a deterministic part
and a random part. Averaging in the time domain, the random part can be reduced
if a trigger signal with respect to the information of interest exists. In the previous
discussions, this trigger signal has been a function of the input (asynchronous or
synchronous averaging) or of the sampling frequency (cyclic averaging). More
generally, though, the trigger function can be any function with characteristics
related to the response history. Specifically, then, the random decrement technique
utilizes the assumption that the deterministic part of the random response signal
itself contains free decay step and impulse response functions and can be used as
the trigger function. Therefore, by starting each history at a specific value and slope
of the random response function, characteristics related to the deterministic portion
of the history will be enhanced.

4.3.2 Random Decrement

There are three specific cases of random decrement averaging that represent the
limiting results of its use. The first case occurs when each starting value is chosen
when the random response history reaches a specific constant level with alternating
slopes for each successive starting value. The random decrement history for this
case becomes the free decay step response function. An example of this case for the
first few averages is shown in Fig. 30.
The second case occurs when each starting value is chosen when the random
response history crosses the zero axis with positive slope. The random decrement
history for this case becomes the free decay positive impulse response function.
The third case occurs when each starting value is chosen when the random
response history crosses zero with negative slope. The random decrement history
for this case becomes the free decay negative impulse response function.
Therefore, in each of these cases, the random decrement technique acts like a
notched digital filter with pass bands at the poles of the trigger function. This tends
to eliminate spectral components not coherent with the trigger function.
If a secondary function is utilized as the trigger function, only the history
related to the poles of the secondary function will be enhanced by this technique.
If the trigger function is sinusoidal, the random decrement history will contain
information related only to that sinusoid. Likewise, if the trigger function is white
noise, the random decrement history will be a unit impulse function at time zero.
One useful example of this concept was investigated for conditioning random
response histories so that information unrelated to the theoretical input history is
removed. In this situation, the theoretical input history serves as the trigger function.
The random decrement history formed on the basis of this trigger function represents
the random response function that would be formed if the theoretical input history
were truly the system input.
6 Frequency Response Function Estimation 313

Fig. 30 Random decrement

averaging

In reality, the measured input history may vary due to noise, impedance
mismatch, and many other factors. However, the success of the technique is
highly dependent upon the identification of the precise trigger timing. The default
uncertainty of ±T in the timing of the actual trigger event means that the
procedure will take a very long time to converge, if at all. Thus, it is necessary
to over sample the data by a factor of 10X to 100X and then down sample to the
desired sampling rate, in order to make sure that the start of each average (ensemble)
is more precise and allow for more rapid convergence.

5 Excitation

Excitation includes any form of input that is used to create a response in a

mechanical system. This can include environmental or operational inputs as well
as the controlled force input(s) that are used in a vibration or modal analysis test.
In general, the following discussion will be limited to the force inputs that can
be measured and/or controlled in some rigorous way. With respect to frequency
response function measurements to be used in experimental modal analysis, the
excitation normally is applied using shakers or with impact devices (hammers). For
314 A. W. Phillips and R. J. Allemang

Fig. 31 Typical test configuration: shaker

those excitation signals that require the use of a shaker, Fig. 31 shows a typical test
configuration, while Fig. 32 shows a typical test configuration when an impact form
of excitation is to be used.
Single and multiple input estimation of frequency response functions (FRFs)
via shaker excitation has become the mainstay of most mechanical structure
measurements, particularly in the automotive and aircraft industries. While there
are appropriate occasions for the use of deterministic excitation signals (sinusoids),
the majority of these measurements are made using broadband (random) excitation
signals. These signals work well for moderate to heavily damped mechanical
structures which exhibit linear characteristics. When the mechanical structures
are very lightly damped, care must be taken to minimize the leakage error so
that accurate frequency response function (FRF) data can be estimated in the
vicinity of the modal frequencies of the system. Frequently, when random excitation
methods are compared to deterministic methods (sinusoids), the comparisons are
questionable since proper procedures for eliminating the leakage error have not been
followed.
Historically, a number of random excitation signals have been utilized, together
with appropriate digital signal processing techniques [1, 2, 3, 35, 38], to obtain
accurate FRF data. The most common random signal that is used in this situation is
the pure random signal together with a Hanning window. This signal is normally
6 Frequency Response Function Estimation 315

Fig. 32 Typical test configuration: impact hammer

generated by the data acquisition system utilizing built-in random signal gener-
ator(s) or via external random signal generator(s). While this approach does not
eliminate the source of leakage and the effect of applying the Hanning window must
be considered, this approach is normally considered as a baseline random excitation
method for estimating FRF measurements since this method is available with almost
any data acquisition system.
Other forms of random signals (pseudorandom, periodic random, burst random,
etc.) utilize more control or frequency shaping of the excitation signal(s) and
generally require digital-to-analog (DAC) converter(s). For this reason, some of
these alternate methods are infrequently available and therefore not used. This is
unfortunate since these methods often yield a superior FRF measurement in less
total testing time.
When FRFs are measured on lightly damped systems, great care must be taken
to eliminate the leakage error. Regardless of the type of excitation signal hardware
involved (random signal generator or DAC), there are random excitation methods
that can nearly eliminate the leakage error. In some cases, one approach will
be superior on the basis of minimizing the total test time, but on the basis of
accurate, leakage-free FRFs, one of the methods will always work if test time
can be sacrificed. Note that these alternate forms of random excitation focus on
eliminating the source of leakage by customizing the random signal to match the
requirements of fast Fourier transform (FFT) that is used in converting from the
316 A. W. Phillips and R. J. Allemang

time to frequency domain. The FFT requires that the time domain signal must either
be totally observed in the observation period (T) or be periodic in the observation
period (T). For leakage-free FRF measurements, all of the input and output signals
must match one of these two requirements. Burst random excitation is an attempt
to match the first requirement; pseudorandom and periodic random excitations are
attempts to match the second requirement.

5.1 Excitation Assumptions

The primary assumption concerning the excitation of a linear structure is that the
excitation is observable. Whenever the excitation is measured, this assumption
simply implies that the measured characteristic properly describes the actual input
characteristics. For the case of multiple inputs, the different inputs must often be
uncorrelated for the computational procedures to yield a solution. In most cases,
this means only that the multiple inputs must not be perfectly correlated at any
frequency. As long as the excitation is measured, the validity of these limited
assumptions can be evaluated.
Currently, there are a number of techniques that can be used to estimate modal
characteristics from response measurements with no measurement of the excitation.
If this approach is used, the excitation assumptions are much more imposing.
Obviously, if the excitation is not measured, estimates of modal scaling (modal
mass, modal A, residues, etc.) cannot be generated. Even under the assumption
that the estimation of these parameters is not required, all of these techniques have
one further restriction: an assumption has to be made concerning the characteristics
of the excitation of the system. Usually, one assumes that the autospectrum of the
excitation signal is sufficiently smooth over the frequency interval of interest.
In particular, the following assumptions about the excitation signal can be
used:

• The excitation is impulsive. The autospectrum of a short pulse (time duration

much smaller than the period of the greatest frequency of interest) is nearly
uniform, or constant in amplitude, and largely independent of the shape of the
pulse.
• The excitation is white noise. White noise has an autospectrum that is uniform
over the bandwidth of the signal.
• The excitation signal is a step. A step signal has an autospectrum that decreases
in amplitude in proportion to the reciprocal of frequency. The step signal can be
viewed as the integral of an impulsive signal.
• There is no excitation. This is called the free response or free decay situation. The
structure is excited to a condition of nonzero displacement, or nonzero velocity,
or both. Then the excitation is removed, and the response is measured during free
decay. This kind of response can be modeled as the response of the structure to
an excitation signal that is a linear combination of impulsive and step signals.
6 Frequency Response Function Estimation 317

When the excitation autospectrum is uniform, the autospectrum of the response

signal is proportional to the square of the modulus of the frequency response
function. Using the notation of a pole-zero model, the poles of the response
spectrum are the poles of the frequency response, which are the parameters of
the system resonances. If the autospectrum is not uniform, then the excitation
spectrum can be modeled as an analytic function to a precision comparable to typical
experimental error in the measurement of spectra. In this model, the excitation
spectrum has poles that account for the nonuniformity of the spectrum amplitude.
The response signal, therefore, can be modeled by a spectrum that contains zeros
at the zeros of the excitation and the zeros of the frequency response and contains
poles at the poles of the excitation and at the poles of the frequency response. It is
obviously important that the force spectrum should have no poles or zeros which
coincide with poles of the frequency response.
For transient inputs, such as an impact or step relaxation, the assumption of
smooth excitation spectra is generally true, but for operating inputs or inputs
generated by an exciter system, care must be taken to insure the input force spectrum
is smooth. This is especially true for tests performed using a hydraulic or an
electromechanical exciter, because the system being analyzed may load the exciter
system (the structure’s impedance is so low that the desired force level cannot be
achieved within the constraint of small motion), and this causes a nonuniformity in
the input force spectrum.
To determine the characteristics of the system from the response, it is necessary
that the response have the same poles as the frequency response or that the analysis
process corrects for the zeros and poles of the excitation. If the force input spectrum
has a zero in the frequency range of interest, the pole location measured from the
response spectrum will not match that of the frequency response. This potential
problem is demonstrated in Fig. 33 for the typical case of shaker excitation. The top
figure is the magnitude of the frequency response function. The middle figure is the
auto power spectrum of the input, and the lower figure is the auto power spectrum
of the response. Note that the estimates of modal parameters that would be derived
from the auto power spectrum of the response would be quite different from those
derived from the frequency response function.
Presently, there is a great deal of interest in determining modal parameters from
measured response data taken on operating systems (e.g., turbulent flow over an
airfoil, road inputs to automobiles, and environmental inputs to proposed large space
structures). For these cases, care must be taken not to confuse poles that are system
resonances with those that exist in the output spectrum due to unmeasured inputs.
In general, the poles of the response include those of the frequency response and
of the input spectrum. Therefore, if the force is not measured, it is not possible
without some prior knowledge about the input to determine if the poles of the
response are truly system characteristics. If no poles or zeros exist in the force
spectrum in the frequency range of interest, then any poles in the response in this
range must be a result of the system characteristics. Obviously, when the excitation
can be measured, it is prudent to do so.
318 A. W. Phillips and R. J. Allemang

Fig. 33 Input spectrum example

6 Frequency Response Function Estimation 319

5.2 Excitation Terminology and Nomenclature

Historically, a number of terminology and nomenclature issues have not been

rigorously defined when excitation methods have been described.
The following terminology is important to the explanation of different excitation
methods together with the associated digital signal processing requirements.

5.2.1 Signal Type

Signal type refers to the basic form of the signal, such as random, impact, sinusoidal,
or chirp.

5.2.2 Frequency Shaping

Frequency shaping refers to any frequency domain constraint or characteristic that
is applied to the specific signal type. With respect to random excitation, a common
frequency shaping is pseudorandom. Other frequency shaping is commonly applied
to sinusoids and chirps via the rate at which the change of frequency and/or
amplitude occurs. Impact excitation is commonly frequency-shaped by controlling
the tip characteristic of the hammer.

5.2.3 Contiguous Blocks

The number of contiguous blocks of excitation that takes place without the
associated input and output data being acquired is referred to as the delay blocks
(Nd ). This is normally associated with an excitation technique that is periodic in
nature. The delay blocks are needed in order to give the transient response to any
start or change in the periodic excitation to decay out of the response signal(s) so that
both the input(s) and output(s) are periodic with respect to any observation period
(T). It is this requirement that makes swept sinusoidal excitation methods (analog
swept or digitally stepped) so time-consuming, particularly on lightly damped
systems. Each delay block is equal in length to the observation period (T), and
the number of delay blocks is normally chosen as an integer. The number of delay
blocks does not have to be an integer for all excitation methods but, for the purposes
of this paper and in common usage, is normally chosen as an integer. The delay
blocks are not recorded and are not used in the estimation of the FRFs.

5.2.4 Capture Blocks

The number of capture blocks refers to the number of contiguous blocks of time
data (excitation (input) and response (output)) that are recorded or captured for each
average (Nc ). The number of capture blocks is also the number of cyclic averages
that will be used to reduce the leakage error. Each group of contiguous capture
blocks (Nc ) is used as the time domain data contributing to one power spectral
average that contributes to the estimate of the FRF measurements.
320 A. W. Phillips and R. J. Allemang

5.2.5 Window Function

The window function refers to the digital signal processing, time domain window
that is applied to the capture blocks. The application of the window function to the
capture blocks is on the basis of the group of contiguous capture blocks, not on each
capture block.

5.2.6 Ensemble or Average

The term ensemble (or average) refers to the total collection of contiguous time
blocks that contribute to each power spectral average. The total time of each
ensemble is equal to the sum of the number of delay blocks (Nd ) plus the number of
capture blocks (Nc ) times the observation period (T) which is the same for all delay
and capture blocks. The number of averages (Navg ) refers to the number of these
contiguous collections of time blocks and is, therefore, the same as the number
of power spectral averages. The number of capture blocks can also be thought of
as the number of cyclic averages (Nc ). Cyclic signal averaging is often used with
excitation characteristics in order to better match the time domain input and output
signals to the requirements of the FFT prior to the application of the FFT. Cyclic
signal averaging essentially digitally comb filters the time domain data to reduce the
amount of information in the data that is not periodic with the observation period
(T). This type of averaging reduces the effects of the leakage error. As long as the
Nc successive blocks of data are contiguous, the blocks of time domain data can
be averaged together, with or without windows, to achieve the benefit of leakage
reduction [39, 40].

5.2.7 Excitation Signal

If the excitation signal is repeated for each delay and capture block, the signal
is referred to as periodic. This classification is consistent with the definition of a
periodic function and includes typical examples of sinusoids and chirps as well as
a random signal that is repeated on the basis of the observation period (T). The
periodic classification does not define whether the same signal is repeated for each
successive group of contiguous delay and capture blocks.

5.2.8 Burst Length

Burst length is the percentage (0% to 100%) of the average or ensemble time that
the excitation signal is present. Burst length is normally adjusted in order to achieve
a signal that is a totally observed transient. The decay of the signal is a function of
the system damping and the characteristics of the excitation hardware. Burst length
can be defined as the percentage of the total number of contiguous delay and capture
blocks or of a percentage of just the capture blocks. For the purpose of this paper,
the burst length refers to the percentage of the total number of contiguous delay and
capture blocks.

5.2.9 Power Spectral Average

The number of power spectral averages (Navg or Na ) is the number of auto and cross
spectra that are averaged together to estimate the FRF measurements. The actual
6 Frequency Response Function Estimation 321

Fig. 34 Total contiguous

time per power spectral
average (ensemble) 0 100
Burst Length (%)

D D C C C C
0 1 2 3 4 5 6
Number of Contiguous Time Blocks (6T)

Window Function

amount of test time contributing to each power spectral average is a function of

the number of contiguous delay and capture blocks. The purpose of power spectral
averages is to eliminate the noise that is random with respect to the averaging
procedure in order to reduce the variance on the resulting FRF estimate. This type
of averaging does not reduce the effects of bias errors like the leakage error.

5.2.10 Excitation Terminology Illustration

In order to clarify the preceding terminology, Fig. 34 is a schematic representation
of the number of contiguous blocks of time domain data contributing to one power
spectral average. In this example, the two blocks marked D represent delay blocks,
and the four blocks marked C represent capture blocks. The total time for each
power spectral average is, therefore, six contiguous blocks of time data (6 × T
seconds of data).

5.3 Classification of Excitation

Inputs that can be used to excite a system in order to determine frequency response
functions (FRFs) belong to one of the two classifications, random or deterministic
[4, 5, 7].
The first classification is that of a random signal. Signals of this form can only be
defined by their statistical properties over some time period. Any subset of the total
time period is unique, and no explicit mathematical relationship can be formulated
to describe the signal. Random signals can be further classified as stationary or
nonstationary. Stationary random signals are a special case where the statistical
properties of the random signals do not vary with respect to translations with time.
Finally, stationary random signals can be classified as ergodic or non-ergodic. A
stationary random signal is ergodic when a time average on any particular subset
of the signal is the same for any arbitrary subset of the random signal. All random
signals which are commonly used as input signals fall into the category of ergodic,
stationary random signals.
322 A. W. Phillips and R. J. Allemang

The second classification of inputs which can be used to excite a system in order
to determine frequency response functions is that of a deterministic signal. Signals
of this form can be represented in an explicit mathematical relationship. Determinis-
tic signals are further divided into periodic and nonperiodic classifications. The most
common inputs in the periodic deterministic signal designation are sinusoidal in
nature, while the most common inputs in the nonperiodic deterministic designation
are transient in form.
The choice of input to be used to excite a system in order to determine frequency
response functions depends upon the characteristics of the system, upon the
characteristics of the modal parameter estimation, and upon the expected utilization
of the data. The characterization of the system is primarily concerned with the
linearity of the system. As long as the system is linear, all input forms should give
the same expected value. Naturally, though, all real systems have some degree of
nonlinearity. Deterministic input signals result in frequency response functions that
are dependent upon the signal level and type. A set of frequency response functions
for different signal levels can be used to document the nonlinear characteristics
of the system. Random input signals, in the presence of nonlinearities, result in
a frequency response function that represents the best linear representation of the
nonlinear characteristics for a given RMS level of random signal input. For systems
with small nonlinearities, the use of a random input will not differ greatly from the
use of a deterministic input.
The characterization of the modal parameter estimation is primarily concerned
with the type of mathematical model being used to represent the frequency
response function. Generally, the model is a linear summation based upon the
modal parameters of the system. Unless the mathematical representation of all
nonlinearities is known, the parameter estimation process cannot properly weight
the frequency response function data to include nonlinear effects. For this reason,
random input signals are prevalently used to obtain the best linear estimate of the
frequency response function when a parameter estimation process using a linear
model is to be utilized.
The expected utilization of the data is concerned with the degree of detailed
information required by any post-processing task. For experimental modal analysis,
this can range from implicit modal vectors needed for troubleshooting to explicit
modal vectors used in an orthogonality check. As more detail is required, input
signals, both random and deterministic, will need to match the system characteristics
and parameter estimation characteristics more closely. In all possible uses of
frequency response function data, the conflicting requirements of the need for
accuracy, equipment availability, testing time, and testing cost will normally reduce
the possible choices of input signal.
With respect to the reduction of the variance and bias errors of the frequency
response function, random or deterministic signals can be utilized most effectively
if the signals are periodic with respect to the sample period or totally observable
with respect to the sample period. If either of these criteria are satisfied, regardless
of signal type, the predominant bias error, leakage, will be minimized. If these
criteria are not satisfied, the leakage error may become significant. In either case,
6 Frequency Response Function Estimation 323

the variance error will be a function of the signal-to-noise ratio and the amount of
averaging.

5.4 Random Excitation Methods

Random signals are widely utilized for general single input and multiple input
shaker testing when evaluating structures that are essentially linear. Signals of this
form can only be defined by their statistical properties over some time period. Any
subset of the total time period is unique, and no explicit mathematical relationship
can be formulated to describe the signal. Random signals can be further classified as
stationary or nonstationary. Stationary random signals are a special case where the
statistical properties of the random signals do not vary with respect to translations
with time. Finally, stationary random signals can be classified as ergodic or non-
ergodic. A stationary random signal is ergodic when a time average on any particular
subset of the signal is the same for any arbitrary subset of the random signal. All
random signals which are commonly used as input signals fall into the category of
ergodic, stationary random signals.
Many signals are appropriate for use in experimental modal analysis. Some of
the most commonly used random signals, used with single and multiple input shaker
testing, are described in the following sections.

5.4.1 Pure Random Signal

The pure random signal is an ergodic, stationary random signal which has a
Gaussian probability distribution. In general, the frequency content of the signal
contains energy at all frequencies (not just integer multiples of the FFT frequency
increment (f = 1/T )). This characteristic is shown in Fig. 35. This is undesirable
since the frequency information between the FFT frequencies is the cause of the
leakage error. The pure random signal may be filtered (Fmin to Fmax ) to include
only information in a frequency band of interest. The measured input spectrum
of the pure random signal, as with all random signals, will be altered by any
impedance mismatch between the system and the exciter. The number of power
spectral averages used in the pure random excitation approach is a function of the
reduction of the variance error and the need to have a significant number of averages
to be certain that all frequencies have been adequately excited.

5.4.2 Pseudorandom Signal

The pseudorandom signal is an ergodic, stationary random signal consisting of
energy content only at integer multiples of the FFT frequency increment (f ).
The frequency spectrum of this signal is shaped to have a constant amplitude with
random phase. This characteristic is shown in Fig. 36. If sufficient delay time is
allowed in the measurement procedure for any transient response to the initiation
of the signal to decay (number of delay blocks), the resultant input and output
histories are periodic with respect to the sample period. The number of power
spectral averages used in the pseudorandom excitation approach is a function of
324 A. W. Phillips and R. J. Allemang

Power Spectrum − Pure Random

101

100
Magnitude

10−1

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 35 Signal energy content – pure random

Power Spectrum − Pseudo Random

101

100
Magnitude

10−1

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 36 Signal energy content – pseudorandom

the reduction of the variance error. In a noise-free environment, only one average
(per input) may be necessary.

5.4.3 Periodic Random Signal

The periodic random signal is an ergodic, stationary random signal consisting only
of integer multiples of the FFT frequency increment. The frequency spectrum of this
signal has random amplitude and random phase distribution. This characteristic is
shown in Fig. 37. For each average, input signals are created with random amplitude
and random phase. The system is excited with these inputs in a repetitive cycle until
the transient response to the change in excitation signal decays (number of delay
blocks). The input and response histories should then be periodic with respect to
the observation time (T) and are recorded as one power spectral average in the total
process. With each new average, a new history, random with respect to previous
input signals, is generated so that the resulting measurement will be completely
randomized. The number of power spectral averages used in the periodic random
6 Frequency Response Function Estimation 325

Power Spectrum − Periodic Random

101

100
Magnitude

10−1

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 37 Signal energy content – periodic random

Power Spectrum − Burst Random

100

10−1
Magnitude

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 38 Signal energy content – burst random

excitation approach is a function of the reduction of the variance error and the need
to have a significant number of averages to be certain that all frequencies have been
adequately excited.

5.4.4 Burst Random Signal

The burst random signal is neither a completely transient deterministic signal nor a
completely ergodic, stationary random signal but contains properties of both signal
types. The frequency spectrum of this signal has random amplitude and random
phase distribution and contains energy throughout the frequency spectrum. This
characteristic is shown in Fig. 38. The difference between this signal and the random
signal is that the random transient history is truncated to zero after some percentage
of the observation time (T). Normally, an acceptable percentage is 50–80 percent.
The measurement procedure duplicates the random procedure but without the need
to utilize a window to reduce the leakage problem as long as both the input and
output decay to zero in the observation time (T ).
326 A. W. Phillips and R. J. Allemang

The burst length (0–100%) is chosen so that the response history decays to
zero within the observation time (T). For moderate to heavily damped systems, the
response history will decay to zero very quickly due to the damping provided by the
system being tested. These systems do not cause a leakage error in the first place.
For lightly damped cases, burst random will force the response to decay to zero in
the observation time (T ) primarily due to the exciter system characteristics. Exciter
systems, particularly electromagnetic, attempt to match the excitation signal to some
physical characteristic of the exciter. Typically, this means that the displacement,
velocity, or acceleration of the armature of the shaker will attempt to match the
excitation signal. (Note that this is normally an open-loop control process; no
attempt is made to exactly match the excitation signal.) Electromagnetic shaker
systems work either in a voltage or current feedback configuration in order to control
the shaker according to the desired input signal. Voltage feedback refers to the type
of amplifier in the exciter system that attempt to match the voltage supplied to the
shaker to the excitation signal. This effectively means that the displacement of the
armature will follow the excitation signal. Therefore, if a zero voltage signal is
sent to the exciter system, the exciter will attempt to prevent the armature from
moving. This damping force, provided by the exciter/amplifier system, is often
overlooked in the analysis of the characteristics of this signal type. Since this
measured input, although not part of the generated signal, includes the variation of
the input during the decay of the response history, the input and response histories
are totally observable within the sample period, and the system damping that will be
computed from the measured FRF data is unaffected. Current feedback refers to the
type of amplifier in the exciter system that attempts to match the current supplied to
the shaker to the excitation signal. This effectively means that the acceleration of the
armature will follow the excitation signal. Therefore, if a zero voltage signal is sent
to the exciter system, the exciter will allow the armature to move, preventing any
force to be applied by the exciter system. The characteristic of a voltage feedback
exciter system for a burst random excitation is shown in the following figures.
Note the difference between the desired burst random signal and the actual force
measured (Figs. 39, 40, and 41).
For very lightly damped systems, the burst length may have to be shortened
below 20 percent. This may yield an unacceptable signal-to-noise ratio (SNR). The
number of power spectral averages used in the burst random excitation approach is
a function of the reduction of the variance error and the need to have a significant
number of averages to be certain that all frequencies have been adequately excited
plus the exciter/amplifier system trying to maintain the input at zero (voltage
feedback amplifier in the excitation system).

5.4.5 Slow Random Signal

The slow random signal is an ergodic, stationary random signal consisting only of
integer multiples of the FFT frequency increment. This signal behaves just like the
pseudorandom signal but without the frequency shaping of the amplitude. The slow
random signal is generated by cyclic averaging a random signal in order to produce
digitally comb filtered excitation signal(s) with the proper characteristics.
6 Frequency Response Function Estimation 327

1.5

0.5

−0.5

−1

−1.5
500 1000 1500 2000

Fig. 39 Burst random – signal to shaker

1.5

0.5

−0.5

−1

−1.5
500 1000 1500 2000

Fig. 40 Burst random – signal from load cell (voltage feedback)

x 10−3
6

−2

−4

−6
500 1000 1500 2000

Fig. 41 Burst random – signal from accelerometer

328 A. W. Phillips and R. J. Allemang

5.4.6 MOOZ Random Signal

The MOOZ random signal is an ergodic, stationary random signal consisting
only of integer multiples of the FFT frequency increment band limited to the
frequency band of a ZOOM fast Fourier transform (FFT)(Fmin to Fmax ). The MOOZ
(ZOOM spelled backward) random signal requires synchronization between the
data acquisition and the digital-to-analog converter (DAC). The MOOZ random
signal is essentially a slow random excitation signal adjusted to accommodate the
frequencies of a ZOOM FFT.
The relationships between delay blocks and averages for some of the most
commonly used random excitation methods are shown in Fig. 42.

5.4.7 Hybrid Random Signal

Several random excitation methods have recently been demonstrated that are hybrid
methods involving combinations of burst random and pseudorandom and burst
random and periodic random together with cyclic averaging.
Figure 43 shows the energy content of a hybrid excitation method that combines
pseudorandom with burst random. This excitation signal would be combined with
cyclic averaging.

Fig. 42 Excitation characteristics

6 Frequency Response Function Estimation 329

Power Spectrum − Burst Pseudo Random

101

100
Magnitude

10−1

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 43 Signal energy content – burst pseudorandom

Power Spectrum − Burst Periodic Random

101

100
Magnitude

10−1

10−2

10−3
0 5 10 15 20 25 30
Spectral line (bin)

Fig. 44 Signal energy content – burst periodic random

Figure 44 shows the energy content of a hybrid excitation method that combines
periodic random with burst random. This excitation signal would be combined with
cyclic averaging.

5.5 Deterministic Excitation Methods

Deterministic signals can be characterized directly by mathematical formula, and

the characteristic of the excitation signal can be computed for any instance in time.
While this is true for the theoretical signal sent to the exciter, it is only approximately
true for the actual excitation signal due to the amplifier/shaker/structure interaction
that is a function of the impedances of these electromechanical systems. Determinis-
tic signals can, nevertheless, be controlled more precisely and are frequently utilized
in the characterization of nonlinear systems for this reason.
330 A. W. Phillips and R. J. Allemang

5.5.1 Slow Swept Sine Signal

The slow swept sine signal is a periodic deterministic signal with a frequency that
is an integer multiple of the FFT frequency increment. Sufficient time is allowed in
the measurement procedure for any transient response to the changes in frequency
to decay so that the resultant input and response histories will be periodic with
respect to the sample period. Therefore, the total time needed to compute an
entire frequency response function will be a function of the number of frequency
increments required and the system damping.

5.5.2 Periodic Chirp Signal

The periodic chirp is a deterministic signal where a sinusoid is rapidly swept from
Fmin to Fmax within a single observation period (T). This signal is then repeated
in a periodic fashion. While this signal is not random in characteristic, it is often
included in discussions of random excitation since it has similar properties as
pseudorandom (Figs. 45 and 46).

Chirp Force Excitation

6

2
Amplitude

−2

−4

−6
0 1 2 3 4 5 6 7 8 9 10
Time (Seconds)

Fig. 45 Typical chirp signal – time domain

Chirp Force Excitation

103

102
Magnitude

101

100
0 1 2 3 4 5 6 7 8 9 10
Frequency (Hertz)

Fig. 46 Typical chirp signal – frequency domain

6 Frequency Response Function Estimation 331

5.5.3 Impact Signal

The impact signal is a transient deterministic signal which is formed by applying
an input pulse to a system lasting only a very small part of the sample period. The
width, height, and shape of this pulse will determine the usable spectrum of the
impact. Briefly, the width of the pulse will determine the frequency spectrum, while
the height and shape of the pulse will control the level of the spectrum. Impact
signals have proven to be quite popular due to the freedom of applying the input
with some form of an instrumented hammer. While the concept is straightforward,
the effective utilization of an impact signal is very involved [29, 41, 42].

Impact Testing
Impact testing is an attempt to match the input and output data to the requirement of
the discrete or fast Fourier transform that the data be a totally observed transient in
the observation time (T ). While the impact is almost always totally observable, the
response for lightly damped systems may not be. Special windows are often used
for impact testing that accommodate the characteristics of the transient input and
the response of the system to a transient input.
Impact excitation is widely used due to the minimal equipment required,
portability and low cost of the impact devices, and broad applicability to both
small-, medium-, and large-size structures. However, impact testing also suffers
from limitations imposed by the human control of the impact. Repeatability and
consistency of the impact (force and direction) cannot be guaranteed, particularly as
the test becomes long and repetitious. Care must be taken to ensure that the impact
and response are not too small and not too large (overload) and that there is only
one impact per observation period.
When impact testing is used, windows are generally required on both the force
and response data in order to minimize different errors. The force window is used to
eliminate the signal coming from the impact device after the short-duration impact
is over. This eliminates electrical noise and spurious output from the hammer during
data acquisition that is caused by motion of the impact device that does not put force
into the system. The response (exponential) window is used to force the response
closer to zero by the end of the observation period (T ) and should be used carefully.
If the response is already near zero at time T , no response window should be added.
To be theoretically correct and to allow for the effects of this response window to
be accounted for, the decay rate of the exponential must be recorded, and the same
window should also be applied to the input data, in addition to the force window.

Force Window
Force windows are used to improve the signal-to-noise problem caused by the
noise on the input channel measured after the impact is completed. Note that the
exponential window used on the response should also be applied to the input in
addition to the force window (Fig. 47).
332 A. W. Phillips and R. J. Allemang

Force Window

1
0.8
Amplitude

0.6
0.4
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1
Time: (0−T)

Force Window * Exponential Window (10%)

1
0.8
Amplitude

0.6
0.4
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1
Time: (0−T)

Fig. 47 Typical force windows

Response (Exponential) Window

Response (exponential) windows are used to minimize the leakage error for lightly
damped systems by attenuating the response so that it decays to zero within the
observation period. Normally, for lightly damped systems, a window that attenuates
to 1–5 percent at the end of the response is appropriate. For heavily damped systems,
a window that is similar to the decay of the system will attenuate any noise (Fig. 48).

Response (Exponential) Window Correction

The windows that are added to the force and response signals must be corrected.
Primarily, the response (exponential) window may add significant damping to the
resultant frequency response function. This can only be corrected after the modal
damping for each mode is found:
6 Frequency Response Function Estimation 333

Exponential Window (10%)

1
0.8
Amplitude

0.6
0.4
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1
Time: (0−T)

Exponential Window (5%)

1
0.8
Amplitude

0.6
0.4
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1
Time: (0−T)

Exponential Window (1%)

1
0.8
Amplitude

0.6
0.4
0.2
0
−0.2
0 0.2 0.4 0.6 0.8 1
Time: (0−T)

Fig. 48 Typical response windows

334 A. W. Phillips and R. J. Allemang


2N
• hpq (t) = Apqr eλr t
r=1

2N
• eβt hpq (t) = eβt Apqr eλr t
r=1

2N
• eβt hpq (t) = Apqr eβt eλr t
r=1

2N 
2N
• eβt hpq (t) = Apqr e(λr +β)t = Apqr eλ̂r t
r=1 r=1
• λ̂r = σ̂r + j ω̂r = (σr + β) + j ωr
• σ̂r = σr + β
• σr = σ̂r − β
• ωr = ω̂r

5.5.4 Step Relaxation Signal

The step relaxation signal is a transient deterministic signal which is formed by
releasing a previously applied static input. The sample period begins at the instant
that the release occurs. This signal is normally generated by the application of a
static force through a cable. The cable is then cut or allowed to release through a
shear pin arrangement [43].

5.5.5 Summary of Excitation Signal Characteristics

Table 6 summarizes the advantages and disadvantages for the most commonly used
excitation signals.

5.6 Excitation Example: H-Frame

The following example presents a single FRF measurement on an H-frame test

structure in a test lab environment as a representative example. The configuration
of the test involved two shaker locations (inputs) and eight response accelerometers
(outputs). The test results are representative of all data taken on the H-frame
structure. This H-frame test structure is very lightly damped and has been the
subject of many previous studies. For all FRF measurement cases, the same test
configuration was used. Sensors were installed and left in place; no additions or
changes were made to the test configuration other than altering the excitation,
averaging, and digital signal processing parameters. Therefore, any changes in the
FRF measurements are assumed to be due to the change in measurement technique
and not due to a test setup variation. The test results were repeated to be certain that
the results are representative.
All FRF measurements are estimated using the H1 estimation algorithm using
1024 spectral (frequency) lines of information. The frequency bandwidth is from
0 to 250 hertz for the 1024 spectral lines; only the first 80% of the spectral lines
6 Frequency Response Function Estimation 335

Table 6 Summary of excitation signals

Excitation signal characteristics
Steady Pure Pseudo Periodic Fast Impact Burst Burst
-state random random random sine sine random
sine
Minimize leakage No No Yes Yes Yes Yes Yes Yes
Signal-to-noise Very Fair Fair Fair High Low High Fair
ratio high
RMS-to-peak High Fair Fair Fair High Low High Fair
ratio
Test measurement Very Good Very Fair Fair Very Very Very
time long short short short short
Controlled frequency Yes Yesa Yesa Yesa Yesa No Yesa Yesa
content
Controlled amplitude Yes No Yesa No Yesa No Yesa No
content
Removes distortion No Yes No Yes No No No Yes
Characterize Yes No No No Yes No Yes No
nonlinearity
a Special hardware required

(0 to 200 Hertz) are displayed in order to exclude the data affected by the anti-
aliasing filters.
The FRF data is plotted with phase above log magnitude. The log magnitude
portion of the plot also contains the relevant multiple coherence plotted on a linear
scale in the background. The log magnitude scaling is annotated on the left side of
the plot, and the multiple coherence scaling is annotated on the right side of the plot.
Fourteen representative cases were measured on this structure. The relevant
excitation and digital signal processing characteristics of each case are shown in
Table 7.
Case 1 (Fig. 49) is considered a baseline case since this a very popular method
for making an FRF measurement and it can be easily made on all data acquisition
equipment. However, it is clear that in this measurement situation, there is a
significant drop in the multiple coherence function at frequencies consistent with
the peaks in the FRF measurement. This characteristic drop in multiple (or ordinary)
coherence is often an indication of a leakage problem. This can be confirmed if a
leakage reduction method reduces or eliminates the problem when the measurement
is repeated. In all subsequent cases, the test configuration was not altered in any way
– data was acquired simply using different excitation, averaging, and digital signal
processing combinations.
Case 2 (Fig. 50) demonstrates an improvement over Case 1 when the same
total measurement time is used, but cyclic averaging is used to reduce the leakage
error. Case 3 (Fig. 51) further demonstrates that burst random with cyclic averaging
336 A. W. Phillips and R. J. Allemang

Table 7 Test cases – excitation/averaging/DSP parameters

Case Signal Frequency Periodic Burst Window Nd Nc Navg Total
type shaping function length function blocks
Case 1 Random No No No Hanning 0 1 20 20
Case 2 Random No No No Hanning 0 5 4 20
Case 3 Random No No Yes (75%) Uniform 0 5 4 20
Case 4 Random Pseudo No No Uniform 4 1 4 20
Case 5 Random No Yes No Uniform 4 1 4 20
Case 6 Random Pseudo No No Uniform 3 1 5 20
Case 7 Random No Yes No Uniform 3 1 5 20
Case 8 Random Pseudo No Yes (75%) Uniform 0 5 4 20
Case 9 Random No Yes Yes (75%) Uniform 0 5 4 20
Case 10 Random No No Yes (75%) Uniform 0 8 12 20
Case 11 Random No No No Hanning 0 1 96 96
Case 12 Random No No No Hanning 0 8 12 96
Case 13 Random Pseudo No No Uniform 3 2 4 20
Case 14 Random No Yes No Uniform 3 2 4 20

Input: 1x Output: 1x
Phase (Deg)

180

0
0
10 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 49 Case 1: random excitation with Hanning window

improves the measurement further. Again the total measurement time remains the
same.
Cases 4 through 7 (Figs. 52, 53, 54, and 55) demonstrate the quality of FRF
measurements that can be achieved with pseudorandom and periodic random
excitation methods with very few power spectral averages.
6 Frequency Response Function Estimation 337

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 50 Case 2: Random excitation with Hanning window and cyclic averaging

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9

−1 0.8
10
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 51 Case 3: Burst random excitation with cyclic averaging

Cases 8 and 9 (Figs. 56 and 57) are hybrid techniques involving the combination
of burst random with pseudorandom and periodic random excitation together with
cyclic averaging.
Case 10 (Fig. 58) demonstrates that Case 3 can be marginally improved with
more averages, both cyclic and power spectral averages. However, Case 11 (Fig. 59)
338 A. W. Phillips and R. J. Allemang

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 52 Case 4: Pseudorandom excitation

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence

0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 53 Case 5: Periodic random excitation

demonstrates that Case 1 (random with Hanning window) cannot be improved by

adding power spectral averages. This is a popular misconception that adding power
spectral averages will improve the FRF estimate. This is clearly not true for this
case.
6 Frequency Response Function Estimation 339

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 54 Case 6: Pseudorandom excitation

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9

−1 0.8
10
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 55 Case 7: periodic random excitation

Case 12 (Fig. 60) demonstrates that additional cyclic averages, together with
power spectral averages, are an improvement over Case 2, but the improvement
is not significant considering the additional measurement time.
Finally, Cases 13 and 14 (Figs. 61 and 62) demonstrate that when pseudorandom
and periodic random excitation is coupled with cyclic averaging, a nearly perfect
340 A. W. Phillips and R. J. Allemang

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 56 Case 8: Burst pseudorandom excitation with cyclic averaging

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9

−1 0.8
10
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 57 Case 9: burst periodic random excitation with cyclic averaging

FRF measurement results. Note also that in almost every case where high-quality
FRF measurements have been achieved, window functions are not required, so
correction for the window characteristics is unnecessary.
It is clear that in many of the measurement cases, the multiple coherence can
be improved dramatically using simple excitation, averaging, and digital signal
6 Frequency Response Function Estimation 341

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 58 Case 10: burst random excitation with cyclic averaging

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9

−1 0.8
10
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 59 Case 11: random excitation with Hanning window

processing methods. Note that as the multiple coherence improves, dramatic

changes in the FRF magnitude accompany the improvement (factors of 2 to more
than 10). When estimating modal parameters, the frequency and mode shape would
probably be estimated reasonably in all cases. However, the damping and modal
scaling would be distorted (overestimating damping and underestimating modal
342 A. W. Phillips and R. J. Allemang

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 60 Case 12: random excitation with Hanning window and cyclic averaging

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9

−1 0.8
10
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10−4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 61 Case 13: pseudorandom excitation with cyclic averaging

scaling). Using these results for model prediction or FE correction would bias the
predicted results.
The most important conclusion that can be drawn from the results of this
measurement exercise on a lightly damped mechanical system is that accurate data
is an indirect function of measurement time or number of averages but is a direct
6 Frequency Response Function Estimation 343

Input: 1x Output: 1x
Phase (Deg)

180

100 1
0.9
0.8
10−1
0.7
Magnitude (g/lb)

Coherence
0.6
10−2 0.5
0.4
0.3
10−3
0.2
0.1

10 −4 0
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hertz)

Fig. 62 Case 14: periodic random excitation with cyclic averaging

function of measurement technique. The leakage problem associated with utilizing

fast Fourier transform (FFT) methodology to estimate frequency response functions
on a mechanical system with light damping is a serious problem that can be managed
with proper measurement techniques, like periodic and pseudorandom excitation or
cyclic averaging with burst random excitation. Hybrid techniques demonstrated in
this paper clearly show that a number of measurement techniques are acceptable,
but some commonly used techniques are clearly unacceptable.
It is also important to note that while ordinary/multiple coherence can indicate a
variety of input/output problems, a drop in the ordinary/multiple coherence function,
at the same frequency as a lightly damped peak in the frequency response function,
is often a direct indicator of a leakage problem. Frequently, comparisons are made
between results obtained with narrowband (sinusoid) excitation and broadband
(random) excitation when the ordinary/multiple coherence function clearly indicates
a potential leakage problem. It is important that good measurement technique be an
integral part of such comparisons.

6 Structural Testing Conditions

The test condition for any modal analysis test involves several environmental factors
as well as appropriate boundary conditions. First of all, temperature, humidity,
vacuum, and gravity effects must be properly considered to match with previous
analysis models or to allow the experimentally determined model to properly reflect
the system.
344 A. W. Phillips and R. J. Allemang

In addition to the environmental concerns, the boundary conditions of the

system under test are very important. Traditionally, modal analysis tests have been
performed under the assumption that the test boundary conditions can be made to
conform to one of four conditions:

• Free-free boundary conditions (impedance is zero)

• Fixed boundary conditions (impedance is infinite)
• Operating boundary conditions (impedance is correct)
• Arbitrary boundary conditions (impedance is known).

It should be obvious that except in very special situations, none of these

boundary conditions can be practically achieved. Instead, practical guidelines are
normally used to evaluate the appropriateness of the chosen boundary conditions.
For example, if a free-free boundary is chosen, the desired frequency of the highest
rigid body mode should be a factor of ten below the first deformation mode of
the system under test. Likewise, for the fixed boundary test, the desired interface
stiffness should be a factor of ten greater than the local stiffness of the system
under test. While either of these practical guidelines can be achieved for small test
objects, a large class of flight vehicle systems cannot be acceptably tested in either
configuration. Arguments have been made that the impedance of a support system
can be defined (via test and/or analysis) and the effects of such a support system
can be eliminated from the measured data. This technique is theoretically sound,
but due to the significant dynamics in the support system and limited measurement
dynamics, this approach has not been uniformly applicable.
In response to this problem, many alternate structural testing concepts have been
proposed and are under current evaluation. Active, or combinations of active and
passive, suspension systems are being evaluated, particularly for application to very
flexible space structures. Active inert gas suspension systems have been used in the
past for the testing of smaller commercial and military aircraft, and, in general, such
approaches are formulated to better match the requirements of a free-free boundary
condition.
Another alternate test procedure is to define a series of relatively conventional
tests with various boundary conditions. The various boundary conditions are chosen
in such a way that each perturbed boundary condition can be accurately modeled
(e.g., the addition of a large mass at interface boundaries). Therefore, as the
experimental model is acquired for each configuration and used to validate and
correct the associated analytical model, the underlying model will be validated and
corrected accordingly. This procedure has the added benefit of adding the influence
of modes of vibration that would normally occur above the maximum frequency
of the test into the validation of the model. For example, the inertial effect of the
addition of a mass at an interface will cause a downward shift in frequency of any
mode that is active at the interface (modes that are not affected by the interface
dynamics will not be shifted). Since this shift is measured and the analytical
model can accurately define the dynamics of the added mass, any inaccuracy in
6 Frequency Response Function Estimation 345

the analytical prediction of the frequency shifts as well as the corresponding effects
on the modal vectors will be due to the lack of fidelity of the underlying analytical
model.
Recently, other researchers have proposed multiple configurations of test con-
ditions as a methodology of utilizing practical test configurations in the testing
of flight vehicle systems. In a related research area, work is progressing on using
constrained testing together with direct parameter estimation methods to define the
characteristics of the unconstrained structure. In this test procedure, the excitation
forces and the constraint forces are measured together with appropriate response
information. The direct parameter estimation method produces a general matrix
model that describes the unconstrained (free-free) structural system. All of these
newer methods will increase the cost (time, financial, technical) of performing
structural tests with the attendant incremental increase in the accuracy of the test
results.

7 Practical Measurement Considerations

There are several factors that contribute to the quality of actual measured frequency
response function estimates. Some of the most common sources of error are due to
measurement mistakes. With a proper measurement approach, most of this type of
error, such as overloading the input, extraneous signal pickup via ground loops,
strong electric or magnetic fields nearby, etc., can be avoided. Violation of test
assumptions is often the source of another inaccuracy and can be viewed as a
measurement mistake. For example, frequency response and coherence functions
have been defined as parameters of a linear system. Nonlinearities will generally
shift energy from one frequency to many new frequencies, in a way which may be
difficult to recognize. The result will be a distortion in the estimates of the system
parameters, which may not be apparent unless the excitation is changed. One way to
reduce the effect of nonlinearities is to randomize these contributions by choosing
a randomly different input signal for each of the n measurements. Subsequent
averaging will reduce these contributions in the same manner that random noise
is reduced. Another example involves control of the system input. One of the most
obvious requirements is to excite the system with energy at all frequencies for which
measurements are expected. It is important to be sure that the input signal spectrum
does not have holes where little energy exist. Otherwise, coherence will be very low,
and the variance on the frequency response function will be large.
Assuming that the system is linear, the excitation is proper, and obvious
measurement mistakes are avoided, some amount of error (noise) will be present
in the measurement process. Five different approaches can be used to reduce the
error involved in frequency response function measurements in current fast Fourier
transform (FFT) analyzers. First of all, the use of different frequency response
function estimation algorithms (Hv compared to H1 ) will reduce the effect of the
leakage error on the estimation of the frequency response function computation.
346 A. W. Phillips and R. J. Allemang

The use of averaging can significantly reduce errors of both variance and bias and is
probably the most general technique in the reduction of errors in frequency response
function measurement. Selective excitation is often used to verify nonlinearities or
randomize characteristics. In this way, bias errors due to system sources can be
reduced or controlled. The increase of frequency resolution through the zoom fast
Fourier transform can improve the frequency response function estimate primarily
by reduction of the leakage bias error due to the use of a longer time sample.
The zoom fast Fourier transform by itself is a linear process and does not involve
any specific error reduction characteristics compared to a baseband fast Fourier
transform (FFT). Finally, the use of weighting functions (windows) is widespread,
and much has been written about their value [1, 2, 3, 44, 45]. Primarily, weighting
functions compensate for the bias error (leakage) caused by the analysis procedure.

8 Summary

The quality of results in modern experimental modal analysis is highly dependent

upon the quality and accuracy of the measured frequency response functions
(FRF). In this chapter, several different methods of computing the FRFs have been
presented. If there is no noise in the measurement process, then all perform equally
well. The differences between them lie in how the noise enters the system and how
the measurement errors are handled. These issues hold true regardless of whether
the measurements are single input or multiple input.
In addition, a number of different coherence formulations (measurement quality
metrics) have been presented. Under the basic premise that a coherence function
provides essentially explained output variation relative to the total output variation,
there are several coherence forms that, while not identical to one another, provide
comparable information. The differences are primarily in the ordering of the
functions and the ordering or assumed relative significance of the inputs. Of crucial
importance is that in the newer formulations, the functions sum to given the
traditional multiple coherence, thus while the underlying individual functions may
be different, they do provide essentially equivalent information.
Further influencing the quality of experimentally measured FRFs is the process
by which the data blocks are accumulated and averaged. Several different averaging
mechanisms were presented by which the effect of noise, in various forms and of
various origins, might be minimized.
Finally, the effect of measurement technique, particularly excitation choices, on
the measured FRFs was presented. As shown (Sect. 5.6), there are a great variety
of traditional and hybrid excitation techniques to address many different testing
scenarios. The proper use of which can mitigate, and essentially eliminate, one of
the most prevalent measurement errors, leakage.
In summary, through proper choice of selective excitation, averaging choices, and
FRF formulation, coupled with coherence indicators, it is possible to make quality
measurements suitable for use in modern parameter identification schemes.
6 Frequency Response Function Estimation 347

References
1. Bendat JS, Piersol AG (1971) Random data: analysis and measurement procedures. Wiley,
New York
2. Bendat JS, Piersol AG (1980) Engineering applications of correlation and spectral analysis.
Wiley, New York
3. Otnes RK, Enochson L (1972) Digital time series analysis. Wiley, New York
4. Strang G (1988) Linear algebra and its applications, 3rd edn. Harcourt Brace Jovanovich
Publishers, San Diego
5. Lawson CL, Hanson RJ (1974) Solving least squares problems. Prentice-Hall, Inc., Englewood
Cliffs
6. Wilkinson JH (1965) The algebraic eigenvalue problem. Oxford University Press, Oxford
7. Jolliffe IT (1986) Principal component analysis. Springer, New York
8. Smallwood DO (1996) Matrix methods for estimating the coherence functions from estimates
of the cross-spectral density matrix. Shock Vib 3:237–247
9. Allemang RJ (1980) Investigation of some multiple input/output frequency response function
experimental modal analysis techniques. Doctor of Philosophy Dissertation, Mechanical
Engineering Department, University of Cincinnati
10. Allemang RJ, Rost RW, Brown DL (1982) Dual input estimation of frequency response
functions for experimental modal analysis of aircraft structures. In: Proceedings, international
modal analysis conference (IMAC), pp 333–340
11. Carbon GD, Brown DL, Allemang RJ (1982) Application of dual input excitation techniques
to the modal testing of commercial aircraft. In: Proceedings, international modal analysis
conference (IMAC), pp 559–565
12. Allemang RJ, Brown DL, Rost RW (1982) Dual input estimation of frequency response
functions for experimental modal analysis of automotive structures, SAE Paper Number
820193
13. Allemang RJ, Brown DL, Rost RW (1984) Multiple input estimation of frequency response
functions for experimental modal analysis. U.S. Air Force Report Number AFATL-TR-84-15,
185pp
14. Allemang RJ, Brown DL, Rost RW (1987) Measurement techniques for experimental modal
analysis, experimental modal analysis and dynamic component synthesis, USAF Technical
Report, Contract Number F33615-83-C-3218, AFWAL-TR-87-3069, vol 2
15. Cobb RE, Mitchell LD (1988) A method for the unbiased estimate of system FRFs in the
presence of multiple correlated inputs. J Anal Exp Modal Anal 3:123–128
16. Mitchell LD, Cobb RE, Deel JC, Luk YW (1988) An unbiased frequency response function
estimator. J Anal Exp Modal Anal 3:12–19
17. Vold H, Crowley J, Rocklin G (1985) A comparison of H1, H2 and Hv frequency response
functions. In: Proceedings, international modal analysis conference (IMAC), pp 272–278
18. Rost R, Leuridan J (1985) A comparison of least squares and total least squares for multiple
input estimation of frequency response functions, ASME Paper Number 83-DET-105, 6pp
19. Rost R, Leuridan J (1985) Multiple input estimation of frequency response functions:
diagnostic techniques for the excitation, ASME Paper Number 83-DET-107, 5pp
20. Leuridan J, De Vis D, Van der Auweraer H, Lembregts F (1986) A comparison of some
frequency response function measurement techniques. In: Proceedings, international seminar
on modal analysis (ISMA), 11pp
21. Napolitano KL (2016) Using singular value decomposition to estimate frequency response
functions. In: Proceedings, international modal analysis conference (IMAC), pp 27–43
22. Patwardhan RS (2020) Frequency response and coherence function estimation methods. Master
of Science Thesis, University of Cincinnati, 73pp
23. Bendat JS, Piersol AG (2011) Random data: analysis and measurement procedures, 4th edn.
Wiley, New York
348 A. W. Phillips and R. J. Allemang

24. Potter RW (1977) Matrix formulation of multiple and partial coherence. J Acoust Soc Am
66:776–781
25. Coppolino RN (2000) Manual: ITAP Toolbox; User’s Guide, Measurement Analysis Corpora-
tion
26. Richards CM, Singh R (1999) Feasibility of identifying non-linear vibratory systems consisting
of unknown polynomial forms. J Sound Vib 220:413–450
27. (2016) Manual: Functions; LMS Test.Lab, Siemens Industry Software
28. Price SM, Bernhard RJ (1986) Virtual coherence: a digital signal processing technique for
incoherent source identification. In: Proceedings, international modal analysis conference
(IMAC), 4pp
29. Phillips AW, Allemang RJ, Zucker AT (1998) An overview of MIMO FRF excitation/averaging
techniques. In: Proceedings, international seminar on modal analysis (ISMA), vol 1, no. 12
30. Phillips AW, Zucker AT, Allemang RJ (1999) A comparison of MIMO FRF excita-
tion/averaging techniques on heavily and lightly damped structures. In: Proceedings, inter-
national modal analysis conference (IMAC), 10pp
31. Fladung WA, Zucker AT, Phillips AW, Allemang RJ (1999) Using cyclic averaging with impact
testing. In: Proceedings, international modal analysis conference (IMAC), 7pp
32. Allemang RJ, Phillips AW (1998) A new excitation method: combining burst random
excitation with cyclic averaging. In: Proceedings, international modal analysis conference
(IMAC), pp 891–899
33. Phillips AW, Zucker AT, Allemang RJ (1999) Frequency resolution effects on FRF estimation:
cyclic averaging vs large block size. In: Proceedings, international modal analysis conference
(IMAC), 14pp
34. Allemang RJ, Phillips AW (1996) Cyclic averaging for frequency response function estimation.
In: Proceedings, international modal analysis conference (IMAC), pp 415–422
35. Hsu HP (1970) Fourier analysis. Simon and Schuster, New York
36. Cole HA J (1971) Failure detection of a space shuttle wing flutter model by random decrement,
NASA TM X-62, 041, May 1971
37. Cole HA (1973) On-line failure detection and damping measurement of aerospace structures
by random decrement signatures, NASA CR-2205, Mar 1973
38. Dally JW, Riley WF, McConnell KG (1984) Instrumentation for engineering measurements.
Wiley, New York
39. Akaike H (1965) On the statistic estimation of the frequency response function of a system
having multiple inputs. Ann Inst Stat Math 17:185–210
40. Carter G, Knapp C, Nuttall A (1973) Estimation of the magnitude-squared coherence function
via overlapped fast fourier transform processing. IEEE Trans Audio Electroacoust 21:337–344
41. Halvorsen WG, Brown DL (1977) Impulse technique for structural frequency response testing.
Sound Vib Mag 8–21, Nov 1977
42. Van Karsen C (1987) A survey of excitation techniques for frequency response function
measurement. Master of Science Thesis, University of Cincinnati, 81pp
43. Stenger G (1979) Step relaxation method for structural dynamic excitation. Master of Science
Thesis, Department of Mechanical Engineering, University of Cincinnati, 54pp
44. Coleman AD, Driskill TC, Anderson JB, Brown DL (1988) A mass additive technique for
modal testing as applied to the space shuttle ASTRO-1 payload. In: Proceedings, international
modal analysis conference (IMAC), pp 154–159
45. Kuo C-P, Wada BK (1989) Multiple boundary condition test (MBCT): identification with mode
shapes, NASA JPL Report, NPO-17574
Random Vibration and Mechanical Shock
7
Thomas L. Paez, Norman F. Hunter, and David O. Smallwood

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
2 Mathematical Foundations of Structural Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
2.1 Single-Degree-of-Freedom Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.2 Multiple-Degree-of-Freedom Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
2.3 Random Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
2.4 Random Vibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
2.5 Mechanical Shock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
3 Random Vibration Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
3.1 Random Processes and the Autospectral Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
3.2 Random Test Control Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
3.3 Test Setup and Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
3.4 Conducting a Random Vibration Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
4 Shock Testing on Shakers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
4.1 Time History Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
4.2 Time History Reproduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
5 Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407

Abstract

All physical systems are exposed to structural dynamic environments, including

random vibration or mechanical shock, or both. These environments can cause
structural or component failure. The capability to analyze dynamic response
is critical not only for purposes of response prediction and design, but also
for specification of random vibration and shock tests. This chapter develops

T. L. Paez () · N. F. Hunter · D. O. Smallwood

Thomas Paez Consulting, Albuquerque, NM, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 349

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_9
350 T. L. Paez et al.

the ideas and the mathematics underlying the structural dynamics of linear
single-degree-of-freedom and multiple-degree-of-freedom structures, random
processes, random vibration, mechanical shock, random vibration testing, and
mechanical shock testing. Examples are provided and many recommendations
are given for the performance of random vibration and shock tests.

Keywords

Random vibration · Mechanical shock · Structural dynamics ·

Single-degree-of-freedom structure · Multiple-degree-of-freedom structure ·
Expected value function · Autocorrelation function · Spectral density
function · Shock response spectrum · Random vibration control · Transient
control

Nomenclature

In general, bold letters denote vectors and matrices, and non-bold letters denote scalar
quantities. Dots denote differentiation with respect to time.
a Acceleration
A Amplitude, Fourier transform of acceleration
Bw Bandwidth
c Viscous damping coefficient
d Displacement
Db Decibels
Df RMS duration of a function of frequency
Dt RMS duration of a function of time
e Estimation error
E Energy of a shock
E[•] Expectation
f Frequency (in Hertz), a function
ĝ One-sided spectral density estimator
G One-sided spectral density
h Impulse response function
H Frequency response function
k Stiffness coefficient
m Mass
mi A temporal moment
M Number of modes retained in modal analysis
N Number of averages used to form spectral density estimate
q Force
Q Fourier transform of q
RMS Root mean square
S Shock response spectrum
Sr Sample rate
v Velocity
x Absolute displacement
X Fourier transform of x
y A function
Y Fourier transform of y
z A function
Z Fourier transform of z
7 Random Vibration and Mechanical Shock 351

{X(t)} Random process

xb Base excitation
z Relative displacement
γ Modal coordinates
γ2 Coherence
 Fourier transform of γ
ζ Damping factor, A decay rate
θ Fourier transform of force
ξ Fourier transform of displacement
μ Mean
ϕ Matrix of mode shapes
ω Frequency (in radians/second)
ωn Natural frequency (in radians/second)
ωd Damped natural frequency (in radians/second)

1 Introduction

Mechanical systems are universally exposed to random vibration and mechanical

shock environments, and sometimes, those environments are important enough to
merit consideration during system modeling or testing. This chapter has three parts:
Section 2 reviews the mathematical foundations of structural dynamics, mechanical
shock, and random vibration. Section 3 introduces the ideas of random vibration
testing. Section 4 introduces the ideas of mechanical shock testing.

2 Mathematical Foundations of Structural Dynamics

Because mechanical shock and random vibration environments can cause structures
to fail, it is critical to understand the behaviors of structures that are dynamically
excited. The principles of mechanics can be used to approximately model the
behaviors of mechanical systems. This section summarizes the equations governing
structural dynamic behaviors of linear systems and the solutions to those equations
(Because space is limited, the behaviors of nonlinear systems are not considered
here.), random vibration, and mechanical shock. Section 2.1 presents the equations
governing motion of single-degree-of-freedom (SDOF) structures and their solu-
tions. Section 2.2 generalizes the governing equations and their solutions to the case
of multiple-degree-of-freedom (MDOF) structures. Section 2.3 presents the idea
of the random process as a representation of a mechanical excitation. It continues
to describe the spectral density as the fundamental characteristic of a stationary
random process. In Sect. 2.4, the fundamental relation of random vibration is written
and described. Section 2.5 presents the idea of mechanical shock as the application
of transient excitation to a structure and computation of structural response. The
important idea of shock response spectrum (SRS) is introduced, as well as the
framework in which it is applied to the specification of shock tests.
352 T. L. Paez et al.

2.1 Single-Degree-of-Freedom Structures

The response of a linear structure is unique; that is, when a specific excitation is
applied to a structure, there is one and only one response that the structure executes.
This fact arises from the characteristics of linear differential equations [1]. That
unique response can, however, be obtained and expressed in different ways. The
theory of structural dynamics can express the equations governing motion and their
solutions in both the time and frequency domains; under different circumstances,
each means of expression is preferable to the other. The following two subsections
present, first, results obtained in the time domain and then results obtained in the
frequency domain.

2.1.1 Structural Dynamics in the Time Domain

The single-degree-of-freedom (SDOF) structure idealization represents the most
basic of oscillatory systems. The idealization is shown in Fig. 1. The structure
consists of a rigid mass m that is constrained to move along a single coordinate;
the absolute displacement of the mass is x. The mass is attached to an immobile
reference via a linear spring with constant stiffness k and a linear, viscous damper
with constant c. The force applied directly to the mass is q(t), − ∞ < t < ∞. The
force may be known over a finite or semi-infinite time interval, only; when that is
so, the response may be evaluated over that interval, and into the infinite future, if
desired. When the time at which the response commences is finite, then the initial
conditions of the system at the start of the response must be specified in order to
evaluate the response.
Newton’s second law can be used to write the equation governing motion of the
SDOF structure. (Ref. [2], for example, has a more detailed development of the
equations and solutions to follow.) It is:

mẍ + cẋ + kx = q(t) −∞<t <∞ (1)

where dots denote differentiation with respect to time. We consider a transformed

version of the equation, one in which both sides are divided by the mass m:

q(t)
ẍ + 2ζ ωn ẋ + ωn2 x = −∞<t <∞ (2)
m

Fig. 1 The fixed-base,

force-excited
single-degree-of-freedom
structure
7 Random Vibration and Mechanical Shock 353

√
where ωn = k/m > 0 is the structure’s natural frequency (in radians/second),
and ζ = c/(2mωn ) ∈ (0, 1] is its damping factor. The meanings of these parameters
will become clear, in a moment.
The solution to the equation of motion has two parts: the homogeneous solution
and the particular solution. The former defines the response of the structure to initial
conditions, in the absence of a forcing function. The latter defines the response of
the structure to the forcing function, only. The superposition of the two solution
parts forms the response to both initial conditions and applied force.
The homogeneous solution is obtained by solving Equation (1) when q(t) = 0,
and initial conditions, x(0) = x0 , and ẋ(0) = v0 , are specified. (Here, we have
specified the initial conditions at time t = 0. The solution can easily be written for
the case where the initial conditions are specified at the arbitrary time, t = t0 .) For
lack of space in this chapter, we will not provide a detailed development of the
method for obtaining the homogeneous solution. Rather, we simply note that the
displacement and velocity responses to the initial conditions are:
 
−ζ ωn t ζ ωn x0 + v0
xh (t) = e x0 cos (ωd t) + sin (ωd t) t ≥0
ωd

 
−ζ ωn t ωn x0 + ζ v0
ẋh (t) = e v0 cos (ωd t) −  sin (ωd t) t ≥0 (3)
1 − ζ2

where ωd = ωn 1 − ζ 2 is the damped natural frequency of the SDOF structure.
The results clarify the terminology of ωn and ζ . The damped natural frequency is
the circular frequency of motion of the response when the response results from
initial conditions, only. The damping factor is required to occupy the interval (0, 1]
for structures to exhibit oscillatory response, but it usually is quite small, in the range
(0.002, 0.100]. Monolithic structures (i.e., structures with no joints, like a steel bar)
that are free-in-space may have dampings as low as a fraction of 1 percent; structures
composed of many parts connected by mechanical joints or concrete structures
excited to high levels may have dampings up to several percent, or even 10 percent.
Because realistic damping factors are near zero, ωd ∼ = ωn .
The particular solution to the equation of motion is the response of the SDOF
structure to the forcing function q(t). For a linear structure that response is unique,
but it may be expressed in several different ways. We start with the time domain
expression for forced response. Again, because there is not room, here, to develop
the response expression, we simply write it for absolute displacement:
 t
x(t) = hxq (t − τ ) q (τ ) dτ t ≥0 (4)
0

where hxq (t), t ≥ 0 is the impulse response function (IRF) of absolute displacement
response to force excitation; the subscripts indicate the response followed by the
354 T. L. Paez et al.

input of interest. This response expression is written for an input that starts at time
t = 0; however, it can be written for any starting time, including t → − ∞. When
the start time of the response is not zero, then the lower limit on the integral must
be changed to the start time. The form of Equation (4) is known as a convolution
integral.
The feature that makes the convolution integral an appropriate form for the
expression of a particular measure of response is the IRF. The reader can confirm
the mathematical expression for the IRF by substituting the convolution integral of
Equation (4) into Equation (2). He or she would obtain a second order differential
equation in hxq (t) whose solution is:

e−ζ ωn t
hxq (t) = sin (ωd t) t ≥0 (5)
mωd

A heuristic approach to the solution of Equation (2) would show that Equation
(4) is its proper solution where the IRF is the first of Equations (3) with x0 = 0 and
v0 = 1/m. The IRF of a causal system (i.e., a system that responds to an input only
at the time of, and after the input has been applied) must always be zero for t < 0.
Because of that, the upper limit on the integral of Equation (4) can be made ∞.
The IRFs of other measures of force-excited response can be written, as well.
For example, the IRF of absolute velocity response to force excitation for SDOF
structures is:
 
e−ζ ωn t ζ
hẋq (t) = cos (ωd t) −  sin (ωd t) t ≥0 (6)
m 1 − ζ2

Recognize that this is simply the derivative of Equation (5) with respect to time.
When this IRF is used in a convolution integral, like Equation (4), with the force,
q(t), the integral yields the absolute velocity response.
So far, we have considered the structure of Fig. 1, governed by Equation (2).
There is another important application, namely, the base-excited structure. Figure 2
shows an SDOF structure that is excited by forces transmitted to the mass through
the spring and damper that attach the mass to its base. The enforced displacement of

Fig. 2 SDOF structure

excited at its base
7 Random Vibration and Mechanical Shock 355

the base is xb (t), − ∞ < t < ∞. (As with externally applied force, base excitation
can be applied over a finite or semi-infinite time interval.) We assume that at least
two derivatives of the enforced base displacement exist; that is, there are a velocity
and an acceleration that correspond to the base displacement.
Define the relative displacement of the mass z(t) = x(t) − xb (t). The equation
governing relative motion of the SDOF system is:

mz̈ + cż + kz = −mẍb −∞<t <∞ (7)

The form is identical to Equation (1). We usually consider the normalized version
of the equation where both sides are divided by the mass m:

z̈ + 2ζ ωn ż + ωn2 z = −ẍb −∞<t <∞ (8)

Because of the similarity of Equation (8) to Equation (2), the homogeneous

solution for z(t) is identical to Equations (3) for displacement and velocity.
Further, the particular solution to Equation (8) takes the same form as the
particular solution to Equation (2), a convolution integral. When we use −ẍb (t) in
place of q(t)/m in Equation (4), we obtain the relative displacement response excited
by a base acceleration input. We modify the factors in the integrand, slightly, to
obtain:
 t
z(t) = hzẍb (t − τ ) ẍb (τ ) dτ −∞<t <∞ (9)
0

where the IRF of relative displacement response to base acceleration input is:

e−ζ ωn t
hzẍb (t) = − sin (ωd t) t ≥0 (10)
ωd

Likewise, we can obtain the IRF of relative velocity response to base acceleration
input. It is:
 
−ζ ωn t ζ
hżẍb (t) = −e cos (ωd t) −  sin (ωd t) t ≥0 (11)
1 − ζ2

When we use hżẍb (t) in a convolution integral with the base acceleration ẍb (t), the
result is the relative velocity response of the SDOF structure.
During experimental applications of mechanical shock and random vibration,
many excitation and response quantities can be measured, including accelerations,
strains, velocities, and relative displacements. However, for reasons related to
practicality and economics, the quantity most frequently measured is the absolute
acceleration at a point on a structure in a particular direction. Therefore, we are
interested in expressing the acceleration response to an acceleration input. When
356 T. L. Paez et al.

we rearrange the definition of relative displacement and take its second derivative,
we obtain:

ẍ(t) = z̈(t) + ẍb (t) t ≥0 (12)

Rearrange Equation (8) and use this expression in it to show:

ẍ(t) = −ωn2 z(t) − 2ζ ωn ż(t) t ≥0 (13)

Finally, substitute the appropriate convolution integrals for z(t) and ż(t), combine
integrands and simplify to obtain:
 t
ẍ(t) = hẍ ẍb (t − τ ) ẍb (t)dτ t ≥0 (14)
0

where the IRF for absolute acceleration response to base acceleration input is:
 
−ζ ωn t 1 − 2ζ 2
hẍ ẍb (t) = ωn e  sin (ωd t) + 2ζ cos (ωd t) t ≥0 (15)
1 − ζ2

The expression in Equation (15) can be used in Equation (14) to compute the
acceleration response of a linear, SDOF structure.

2.1.2 Structural Dynamics in the Frequency Domain

We are often most interested in the temporal behavior of structures and use the
approach of Sect. 2.1.1 to express responses. However, interpretation of structural
behavior is best accomplished in the frequency domain. Further, numerical analysis
of structural response is more efficient in the frequency domain, and the classical
analysis of random vibration depends on the frequency domain interpretation of
structural behavior.
Several approaches lead to the solution of the problems of linear structural
dynamics in the frequency domain [3]. We pursue the solution of this section by
taking the Fourier transform of each term in Equation (2):

1
ωn2 − ω2 + 2ζ ωn iω ξ0 (ω) = θ (ω) −∞<ω <∞ (16)
m
√
where i = −1 is the imaginary unit, and ξ 0 (ω) and θ (ω) are the Fourier transforms
of x(t) and q(t), respectively, defined as:
 ∞
ξ0 (ω) = x(t)e−iωt dt
−∞
 ∞
(17)
−iωt
θ (ω) = q(t)e dt −∞<ω <∞
−∞
7 Random Vibration and Mechanical Shock 357

The “ 0” subscript on ξ 0 (ω) indicates that it is the Fourier transform of the zeroth
derivative of the function x(t). The Fourier transforms of ẍ(t) and ẋ(t) are −ω2 ξ 0 (ω)
and iωξ 0 (ω), respectively.
Fourier transformation converts the differential relation of Equation (2) into an
algebraic relation. Equation (16) can be solved to obtain:

1
ξ0 (ω) =   θ (ω) −∞<ω <∞ (18)
m ωn2 − ω2 + 2ζ ωn iω

The coefficient on the right is known as the force input/absolute displacement

output, frequency response function (FRF), and is denoted as:

1
Hxq (ω) =   −∞<ω <∞ (19)
m ωn2 − ω2 + 2ζ ωn iω

The FRF is a fundamental descriptor of the behavior of a linear struc-

ture (SDOF in this case). If the structure governed by Equation (2) were
excited by the purely harmonic input q(t) = Qeiωt , it would excite the
purely harmonic response Hxq (ω)Qeiωt ; the response would be amplified
by the amount |Hxq (ω)| and its phase would be modified by the amount
phase[Hxq (ω)] = tan−1 (Im[Hxq (ω)]/ Re [Hxq (ω)]).
When we use Equation (19) in Equation (18), we obtain the concise expression:

ξ0 (ω) = Hxq (ω) θ (ω) −∞<ω <∞ (20)

Figure 3 shows a normalized version of Hxq (ω) for ω ≥ 0, and for several values
of damping factor ζ , in terms of its modulus and phase. As the damping factor
diminishes, the amplification reflected by the FRF increases at the natural frequency
of the SDOF structure. The phase passes through −π /2 at ω/ωn = 1.
Because Equation (20) represents the Fourier transform of the absolute dis-
placement response, and because the Fourier transform is invertible, the absolute
displacement response can be obtained in the time domain through inversion of
Equation (20):
 ∞
1
x(t) = Hxq (ω) θ (ω) eiωt dω −∞<t <∞ (21)
2π −∞

For some excitations q(t), − ∞ < t < ∞, this computation might be carried out in
closed form. However, when q(t) is known over a finite time interval and at equally
spaced, discrete times, the computation can be carried out numerically, using the fast
Fourier transform (FFT) [4]. Hxq (ω) can be evaluated at discrete frequencies using
Equation (19). The product in the integrand of Equation (21) can be formed. Then
the inverse Fourier transform of Equation (21) can be approximated using FFT. This
sequence of operations is much more efficient than the numerical approximation of
the convolution integral.
358 T. L. Paez et al.

Fig. 3 Modulus (top) and phase (bottom) of force input/absolute displacement output FRF of an
SDOF structure

Leibniz’s rule [5] can be used with Equation (21) to obtain velocity and
acceleration responses, and this leads to expressions for the force input – absolute
velocity, and absolute acceleration FRFs:

iω
Hẋq (ω) =   −∞<ω <∞ (22)
m ωn − ω2 + 2ζ ωn iω
2

−ω2
Hẍq (ω) =   −∞<ω <∞ (23)
m ωn2 − ω2 + 2ζ ωn iω

The time domain velocity and acceleration responses are:

 ∞
1
ẋ(t) = Hẋq (ω) θ (ω) eiωt dω −∞<t <∞ (24)
2π −∞
 ∞
1
ẍ(t) = Hẍq (ω) θ (ω) eiωt dω −∞<t <∞ (25)
2π −∞
7 Random Vibration and Mechanical Shock 359

When we Fourier transform the two equations, above, we obtain:

ξ1 (ω) = Hẋq (ω) θ (ω) −∞<ω <∞ (26)

ξ2 (ω) = Hẍq (ω) θ (ω) −∞<ω <∞ (27)

where the subscripts “1” and “2,” on the left indicate that ξ 1 (ω) and ξ 2 (ω) are the
Fourier transforms of the first and second derivatives of x(t), respectively. These two
equations and Equation (20) serve as the frequency domain relation between force
and the measures of response of an SDOF structure.
We can follow the procedure used to obtain the FRFs for a force-excited structure
to establish the FRFs for the responses of a base-excited structure. The preliminary
analysis would consider Equation (8) and obtain FRFs of relative displacement,
velocity, and acceleration to an enforced base excitation. The FRFs are Hzẍb (ω),
Hżẍb (ω), and Hz̈ẍb (ω) , −∞ < ω < ∞, respectively, where base acceleration is
the input. Then the relation z(t) = x(t) − xb (t) could be used to solve for the absolute
displacement x(t). The first and second derivatives of the expression can be evaluated
and then Fourier transformed to obtain the FRFs of absolute displacement, velocity,
and acceleration to enforced base excitation. The base acceleration input/absolute
acceleration response FRF is:

ω2 + 2iζ ωn ω
Hẍ ẍb (ω) =  2 n 2 −∞<ω <∞ (28)
ωn − ω + 2iζ ωn ω

The graph of this FRF, when normalized, is similar to the graphs in Fig. 3.
By analogy with Equation (27), the Fourier transform of the absolute acceleration
response to base acceleration input is:

ξ2 (ω) = Hẍ ẍb (ω) ξb2 (ω) −∞<ω <∞ (29)

where ξ b2 (ω), − ∞ < ω < ∞ is the Fourier transform of the base acceleration.
This is the frequency domain relation between base acceleration input and absolute
acceleration response. The acceleration response in the time domain is:
 ∞
1
ẍ(t) = Hẍ ẍb (ω) ξb2 (ω) eiωt dω −∞<t <∞ (30)
2π −∞

This expression provides the time domain, absolute acceleration response in terms
of the Fourier transform of the base acceleration input and the structure FRF.

Example 1 Computation of SDOF Structure Response in the Frequency

Domain
An input applied at its base is used to excite an SDOF structure with natural
frequency ωn = (2π )100 rad/sec and damping factor ζ = 0.02. The time-domain
360 T. L. Paez et al.

0
50 10

|2b(f)|, g-sec
d xb/dt , g
2

-1
0 10
2

-2
-50 10 1 2 3
0 0.5 1 10 10 10
Time, sec Frequency, Hz
2
10

|H(f)|, g/g

0
10

-2
10 1 2 3
10 10 10
Frequency, Hz
2
10 200

100

d x/dt , g
|2(f)|, g

0
10
2
2 0

10
-2 -100

1 2 3
-200
10 10 10 0 0.5 1
Frequency, Hz Time, sec

Fig. 4 Graphic showing the sequence of operations involved in computation of linear structure
response in the frequency domain

input is shown on the top left in Fig. 4; it is sampled at a rate of 2048 samples/sec
(i.e., Δt = 4.88 × 10−4 sec). We compute the absolute acceleration response in the
frequency domain. The modulus of the DFT of the input is shown at top center
in the figure at frequencies fk = kΔf, k = 0, . . . , 1024; Δf = 1 Hz. (All DFT
moduli are plotted at these frequencies. The phase is erratic and provides little
intuition regarding the structural response. The phase of the structural response is
the sum of the phases of the input and the FRF.) The modulus of the FRF of absolute
acceleration response to base acceleration input is shown in the center of the figure.
The modulus of the product of the input and the FRF is shown at bottom center;
this is the response DFT. The inverse DFT of the response DFT is shown at bottom
right; of course, computation of the inverse DFT requires use of the phase as well
as the modulus of the response DFT. Within numerical round-off, this is the result
that would be obtained via numerical convolution.
7 Random Vibration and Mechanical Shock 361

2.2 Multiple-Degree-of-Freedom Structures

The developments of Sect. 2.1 consider the responses of SDOF structures in the
time and frequency domains. Many real structures can be modeled, approximately,
as SDOF; however, most real structures are spatially continuous and reflect more
complex behavior. This section presents (without development) the equations
governing motion of more complex, more realistic structures and presents their
solutions.

2.2.1 Structural Dynamics in the Time and Frequency Domains

It is beyond the scope of this chapter to develop the methodology for analyzing
complex structures; however, the equations that govern the behavior of linear contin-
uous structures can be approximated with simultaneous, linear, ordinary differential
equations (ODE) [6]. Those equations describe what are known as multiple-degree-
of-freedom (MDOF) models. Finite element (FE) codes [6] develop such models for
specific structures, and solve the governing equations to establish MDOF structure
responses. Figure 5 shows, at left, the schematic of a (two-dimensional) continuous
structure. The idea behind FE modeling is that the continuous structure is divided
into elements; the schematic on the right in Fig. 5 shows the idea. Equations
that describe the equilibrium of each element are developed, and the collection
of elements is combined and connected to obtain a set of simultaneous equations
approximately governing the equilibrium of the overall structure. The modeler may
be interested in predicting motions at the location and in the directions indicated
by (xi , xi + 1 , xi + 2 ). When a model provides the capacity to execute those motions,
it is said to have those degrees-of-freedom (DOF). Indeed, the model has DOF
associated with all permissible motions at the nodes of the model; we consider
the nodes as the points where elements connect. (The motions are translations or
rotations and are referred to as generalized motions.)
The idea behind a dynamic model is that it can be used to approximate responses
(x, ẋ, ẍ) at all DOF and at all times of interest. Dots denote differentiation with

Fig. 5 Schematic of a continuous structure (left) and its spatially discretized approximation (right)
362 T. L. Paez et al.

respect to time. x denotes displacements at the model DOF; ẋ denotes velocities;

ẍ denotes accelerations. When the model has N DOF, then each of the vectors
(x, ẋ, ẍ) has dimension N × 1.
During the analysis of dynamic response, one or more externally applied forces,
qi (t), i = 1, 2, . . . , − ∞ < t < ∞, may be specified. (The forces may alternately be
applied over a finite or semi-infinite time interval.) The forces can be collected into
the N × 1 vector, q(t), − ∞ < t < ∞. There is a force in q(t) for each model DOF;
some or all the qi (t) in q(t) may be zero. The forces are generalized in the sense that
they may be translational or rotational, depending on the DOFs they excite.
Given information on the geometry of a structure, the material properties of
its various parts, and its boundary conditions, an FE code establishes the mass
matrix m of a structure to describe its inertial properties and the stiffness matrix
k of a structure to describe its stiffness. Both the mass and stiffness matrices
have dimension N × N. The matrix equation governing motion of a linear MDOF
structure is:

mẍ + cẋ + kx = q(t) −∞<t <∞ (31)

The N × N matrix c is the matrix of viscous damping coefficients; it is usually

specified based on experimental data or an understanding of typical dissipative
behavior of a class of structures.
Within the framework of FE analysis, there are multiple approaches to solving
the governing Equation (31). For example, when the force excitations are specified
at the discrete times tj = jΔt, = 0, 1, . . . , with Δt > 0 a time increment, the
forces might be denoted qj = q(tj ), j = 0, 1, . . . . Equation (31) can be solved,
approximately,
 at all the structural
 DOF and at all times to obtain the responses
x j, ẋ j, ẍ j = x tj , ẋ tj , ẍ tj , j = 0, 1, . . . . However, that approach does
not provide the most efficient method for obtaining responses. The more common
method for analyzing response involves modal analysis, and the modes of a structure
also characterize the MDOF structure.
Modal analysis starts by considering the eigenvalue problem associated with the
undamped, unforced form of the equation of motion:

mẍ + kx = 0 (32)

Solution of the eigenvalue problem obtained from Equation (32) yields a collec-
tion of up to N eigenvectors vk , k = 1, . . . , N and corresponding modal frequencies
ωk , k = 1, . . . , N. These have physical meaning: There are certain frequencies
ωk , where, in free, undamped vibration, a linear structure will execute harmonic
motion while maintaining a constant shape vk . The eigenvectors are referred to
as mode shapes. The mode shapes and modal frequencies are approximated via
iterative procedures, and, typically, the lowest M modes, only, are approximated.
(Typically, especially in the case where the model is large M  N.) The analyst
typically chooses M so that ωM ∼ = ωMax where ωMax is the highest frequency where
the input q(t) has substantial signal content.
7 Random Vibration and Mechanical Shock 363

Each mode shape can be normalized using ϕ k = bk vk , so that:

ϕ Tk mϕ k = 1 k = 1, . . . , M (33)

When the normalization of Equation (33) is established, it is also true that:

ϕ Tk kϕ k = ωk2 k = 1, . . . , M (34)

Further, the following relations hold:

ϕ Tk mϕ r = 0 k=r ϕ Tk kϕ r = 0 k=r (35)

The mode shapes with the properties in Equations (33) through (35) are referred
to as ortho-normal. When the modal vectors are gathered into an N × M matrix ϕ
whose columns are the ϕ k , k = 1, . . . , M, then:

ϕ T mϕ = I ϕ T kϕ = ω2 (36)

where I is the M × M identity matrix and ω2 is the diagonal M × M matrix with

elements ωk2 , k = 1, . . . , M. Typically, we also assume that:

ϕ T cϕ = 2ζ ω (37)

where ζ is the diagonal M × M matrix with elements ζ k , k = 1, . . . , M. (c can be

constructed so that Equation (37) holds, or if c is constructed so that Equation (37)
does not hold, we might ignore the off-diagonal terms.)
In view of Equations (36) and (37), we can define the transformation x = ϕ γ
and use it in Equation (31). The M × 1 vector γ (t) includes the modal coordinates
of motion. We then pre-multiply the result by ϕ T to obtain:

I γ̈ + 2ζ ωγ̇ + ω2 γ = ϕ T q(t) −∞<t <∞ (38)

Because the coefficient matrices on the left are all diagonal, Equation (38) represents
a sequence of M second order ODEs like Equation (2), each governing the motion
of an SDOF structure. Therefore, the effect of modal analysis is to reduce the
governing Equation (31) to a set of M equations, each one describing the behavior
of one mode. The kth modal equation is:

γ̈k + 2ζk ωk γ̇k + ωk2 γk = ϕ T q(t) = qmod,k (t) k = 1, . . . , M, −∞ < t < ∞

k
(39)

where (ϕ T q(t))k = qmod, k (t) is the kth element in the M × 1 column vector ϕ T q(t).
It is the kth modal excitation.
364 T. L. Paez et al.

The time-domain solution to Equation (39) was developed in Sect. 2.1.1. One
form of it is:
 ∞
γk (t) = hγk q (t − τ ) qmod,k (τ ) dτ k = 1, . . . , M, −∞ < t < ∞ (40)
−∞

where hγk q (t), t ≥ 0 is the force input/absolute displacement response IRF of an

SDOF structure. As shown in Sect. 2.1.1, expressions for the force-excited absolute
velocity and acceleration responses can also be obtained. As well, expressions for
the base-excited relative motion and absolute motion responses can be developed.
When the γ k (t), k = 1, . . . , M of Equation (40) are used to construct the vector
γ (t), we obtain:
 t
γ (t) = hγ q (t − τ ) ϕ T q (τ ) dτ −∞<t <∞ (41)
0

where hγ q (t), t ≥ 0 is the M × M, diagonal matrix of modal IRFs for the structure.
Finally, we obtain the response in physical coordinates by using the definition of
modal coordinates:
 t
x(t) = ϕγ (t) = ϕhγ q (t − τ ) ϕ T q (τ ) dτ −∞<t <∞ (42)
0

The leading triplet of terms inside the integrand is the matrix of IRFs of the
physical system; therefore, we write:

hxq (t) = ϕhγ q (t)ϕ T t ≥0 (43)

The ith − row, jth − column element of hxq is interpreted as the absolute
displacement response of the structure at DOF i to a force impulse applied at DOF j.
The dimension of hxq is N × N, but we rarely compute the entire matrix of functions.
Rather, the elements of hxq can be computed one-at-a-time, and they usually are
computed in this manner. For example, it makes no sense to compute the ith − row,
jth − column element of hxq for columns j corresponding to rows in q where the input
is zero; so those elements are always skipped. Likewise, we are normally interested,
only, in evaluating the response at certain DOF; so we evaluate the IRF at those rows
i only.
Section 2.1.2 developed the method for solving the equation governing motion
of SDOF structures in the frequency domain. Equation (39) can be solved using that
method to obtain the response in the frequency domain of the kth structural mode.
That solution is the Fourier transform of the solution written in Equation (40):

k (ω) = Hγk q (ω) Qmod,k (ω) k = 1, . . . , M, −∞ < ω < ∞ (44)

7 Random Vibration and Mechanical Shock 365

where Hγk q (ω) is the FRF of the kth structural mode, the Fourier transform of
hγk q (t), and Qmod, k (ω) is the Fourier transform of qmod, k (t). The M × 1 vector
of absolute, modal displacement responses can be constructed with Equation (44)
to obtain:

 (ω) = H γ q (ω) ϕ T Q (ω) −∞<ω <∞ (45)

where Q(ω) is the Fourier transform of q(t), and Hγ q (ω) is the M × M, diagonal
matrix of modal FRFs, the Fourier transform of hγ q (t).
The Fourier transform of the vector of physical responses can be denoted
X(ω), − ∞ < ω < ∞, and it can be recovered by using the definition of the modal
coordinates:

X (ω) = ϕ (ω) = ϕH γ q (ω) ϕ T Q (ω) −∞<ω <∞ (46)

We recognize the leading triplet on the right as the matrix of FRFs of the physical
structure; therefore, we write:

H (ω) = ϕH γ q (ω) ϕ T −∞<ω <∞ (47)

H(ω) has dimension N × N, and its ith − row, jth − column element is complex-
valued. When the excitation eiωt is input at DOF j (and all other inputs are zero), then
the response at DOF i is the ith − row, jth − column element of H(ω) times eiωt , i.e.,
Hij (ω)eiωt . The magnitude |Hij (ω)| describes the amplification or diminishment of
the response over the input; the phase of Hij (ω) describes the phase lag between the
input and the response. Of course, only those FRFs of interest in a given analysis
need be established with Equation (47).
Equation (47) indicates that the matrix of FRFs of a linear structure is a
superposition of its modal FRFs. Each modal FRF is a function that amplifies the
response motion at its modal frequency. Therefore, the magnitude of each function
in H(ω) reflects a sequence of peaks at the structural modes and troughs between
the peaks. The relative magnitudes of the peaks depend on the relative amplitudes
of the elements in the columns of the modal matrix ϕ.
The matrix FRF developed in Equation (47) relates input forces to response
motions. The specific form of H(ω) depends on what measure of response motion is
to be established. The base excitation input/structural motion output FRF could be
established as well. The FRFs developed here could also be expressed as a function
of cyclic frequency H(f ), f ≥ 0.

2.3 Random Processes

The bases for the study of random vibration are the theories of random processes and
structural dynamics. Some ideas of structural dynamics have been outlined in Sects.
2.1 and 2.2. Here, we introduce some ideas from the theory of random processes.
366 T. L. Paez et al.

There are multiple ways to define a random process, but it suits our purposes
to think of a random process as an abstract collection including an infinite
number of continuous-valued oscillatory signals. Denote the signals xj (t), j = 1,
2, . . . , − ∞ < t < ∞. The collection of signals is the ensemble of the random
process. When we measure a signal from a random source, it is a member of the
ensemble, and we are said to draw a realization from the random process. Typically,
in the practice of random vibration, we measure input excitations and responses of
mechanical systems. Five finite-duration segments of structural response random
process realizations are shown in Fig. 6. The independent variable of the random
process whose realizations are shown is continuous time; this is the random process
parameter. The ordinate has a measure of interest like acceleration, velocity, or
displacement; the signals shown in Fig. 6 are accelerations in units of g’s.
When the signals in the ensemble of a random process are in a random steady-
state, then the random process is said to be stationary. In principle, the signals in
the ensemble cover the interval (−∞, ∞), but, of course, in practice, no real source
maintains a steady-state for infinite time. When the realizations of a random process
maintain a steady-state for a time period that is large compared to the fundamental
period of the structure under consideration, then the excitation and response random
processes can be considered approximately stationary. For example, a store carried
on an aircraft wing might have a fundamental frequency of f1 = 50 Hz and a

0
x(t)

-5 1
0.9
0.8
5 0.7
0.6
4
0.5
0.4
3
0.3
2 0.2
0.1
Time, sec
Ensemble element 1 0

Fig. 6 Five realizations of a response random process

7 Random Vibration and Mechanical Shock 367

corresponding fundamental period of T1 = 1/f1 = 0.02 sec. When that system is

excited in a random steady state for 2 sec or more, then the excitation and response
might be considered stationary.
A stationary random process is often denoted as {X(t), − ∞ < t < ∞}. The
average or expected value of a random process is denoted as E[X(t)], − ∞ < t < ∞;
this defines the average of all members of the ensemble at time t:

1
n
E [X(t)] = lim xj (t) −∞<t <∞ (48)
n→∞ n
j =1

The conditions for weak stationarity of a random process are (1) that is has an
expected value that is constant as a function of time, i.e., E[X(t)] = E[X] = μX ,
and (2) that it possesses a frequency-domain decomposition known as the spectral
density. (When the expected value of an excitation random process is non-zero, we
usually consider the oscillatory portion, only, of the random process and evaluate
the response to the mean separately. This is how we treat random excitations in
the following.) The spectral density of a random process can be defined in several
equivalent ways; we provide an intuitive description of the one-sided spectral
density that is a function of cyclic frequency f, f ≥ 0. The spectral density of the
random process {X(t)} is denoted GXX (f ), f ≥ 0. To describe the spectral density,
we define the random process {X(t; f, Δf )} as a filtered version of {X(t)} with
signal content in the frequency interval [f, f + Δf ], only [3]. The realizations of
{X(t; f, Δf )} can be denoted xj (t; f, Δf ), j = 1, 2, . . . . The area under the spectral
density curve over the interval [f, f + Δf ] is defined as:

 f+ f 1 2
n
GXX (u)du = lim xj (t; f, f ) (49)
f n→∞ n
j =1

where the quantity on the right is the mean square of a realization of the filtered
random process. This defines the spectral density implicitly. The definition of the
spectral density can be converted into a time-average in which the spectral density
can be obtained from a single (infinite duration) random process realization. In
order for this to hold, the stationary random process must be ergodic; that is, any
realization of the random process must be representative of all the others.
When we denote the mean square of the random process {X(t)} as σX2 (a constant),
then the definition in Equation (49) implies:
 ∞
σX2 = GXX (f )df (50)
0

That is, the area under the spectral density curve equals the mean square of
the random process. Because the area under GXX (f ), f ≥ 0 is a mean square,
which cannot be negative, the function GXX (f ) must be non-negative. Further, the
368 T. L. Paez et al.

Fig. 7 A graphic
interpretation of spectral
density

amplitudes of the spectral density reveal how the mean square signal content of a
random process is distributed with frequency.
A graphic interpretation of the spectral density function is provided in Fig. 7.
The fundamental idea is that a random signal (first column) can be filtered (second
column) into a sequence of components (third column) with independent signal
content. The mean square of each component (fourth column) can be established
and plotted as a function of the center frequency of the filter. Each mean square can
be divided by the filter bandwidth to obtain the spectral density (fifth column). Until
the advent of digital analysis and control, spectral density was actually estimated in
the manner described in Fig. 7.
There is another important measure of the behavior of stationary random process
pairs known as the cross-spectral density [7]. The spectral density, described in the
previous equations, is also referred to as the autospectral density. The autospectral
density characterizes the mean square signal content of a random process in the
frequency domain. That is, it describes the signal content of a quadratic form of
a random process. The cross-spectral density describes the simultaneous signal
content of a pair of different random processes {X(t), − ∞ < t < ∞} and
{Y(t), − ∞ < t < ∞}; it is based on a quadratic form of the two random processes.
It is denoted as GXY (f ), f ≥ 0, and it is generally a complex function. When
the two random processes are strongly correlated (either positively or negatively,
see reference [7]) at a given frequency f, then the magnitude of GXY (f ) is great;
otherwise, the magnitude of GXY (f ) is small. The phase of GXY (f ) measures the
average phase difference between the phases of the random processes {X(t)} and
{Y(t)}.
7 Random Vibration and Mechanical Shock 369

The units of autospectral density are the units of the random process, squared,
per unit of frequency. For example, when a random process has units of g s and the
frequency measure of interest is Hz, then the spectral density units are g2 /Hz.

2.4 Random Vibration

The mathematical theory of random vibration has been developed over the course
of the last century (approximately); the first paper on the subject was written by
Einstein [8]. (Reference [8] is reprinted in Reference [9].). The idea that enables
treatment of random vibration with modern techniques is the spectral density, and
that was developed by Wiener in [10]. When a linear, stable, deterministic structure
is excited by one or more stationary random inputs, the responses at all DOF on the
structure converge to stationary random processes. When the responses reach the
stationary state, both the inputs and the responses possess auto- and cross-spectral
densities.
Consider a structure with the Nout × Nin matrix FRF H(f ), f ≥ 0; assume that
H(f ) is a force input/motion output FRF. The Nout rows of the FRF correspond
to DOF where we desire the response of the structure; the Nin columns of the FRF
correspond to DOF where there are force inputs. The matrix form of the fundamental
relation of random vibration (a descriptive name specified by Crandall [11, 12]) is:
 T
GXX (f ) = H (f )GQQ (f ) H ∗ (f ) f ≥0 (51)

where GQQ (f ), f ≥ 0 is the matrix of one-sided spectral densities of the stationary,

vector random force excitation {Q(t)}. The vector of random input forces is
 T
Q(t) = Q1 (t), . . . , QNin (t) ; the elements on the right are scalar, stationary,
force random processes. The matrix GQQ (f ) has dimension Nin × Nin . Its diagonal
terms are the autospectral densities of the elements of {Q(t)}, and its off-diagonal
terms are the cross-spectral densities between pairs of the elements in {Q(t)}. GXX (f )
is the Nout × Nout matrix of one-sided spectral densities of the stationary, vector
random response motions {X(t)}. The elements of {X(t)} and GXX (f ) are described
in terms analogous to the description of the random force input.
When the input is a base excitation applied in one direction, only, then the
dimension of the FRF is Nout × 1. The spectral density of the excitation is, in this
case, a scalar. The matrix, fundamental relation of random vibration takes the form:

T
GẌẌ (f ) = GẌb Ẍb (f )H ẌẌb (f ) H ∗ẌẌ (f ) f ≥0 (52)
b

In this particular case, the relation is written for a base acceleration input and
an absolute acceleration output. This is the idealized form that might be used,
for example, for computation of the behavior of a structure excited on a shaker
in the laboratory. (It is an idealization because, in the laboratory, base excitation
usually has inputs in three directions and rotations about three axes. When those
370 T. L. Paez et al.

additional inputs are known, then expressions like the one in Equation (52) can be
superimposed to obtain predictions of response spectral densities.)
The fundamental relation of random vibration for the single-input/single-output
(SISO) case is a special case of Equation (51). When stationary, random force input
to a structure is applied at one location, only, its spectral density can be denoted as
GQQ (f ), f ≥ 0. When we are interested in assessing the response spectral density
at one location, only, then the response spectral density can be denoted as GXX (f ),
f ≥ 0. The relation between excitation and response spectral densities is:

GXX (f ) = |H (f )|2 GQQ (f ) f ≥0 (53)

Because the magnitude of the FRF, |H(f )|, of an MDOF structure displays a
sequence of peaks at structural modes, this relation implies that if there is mean
square signal content in the stationary random force input at a modal frequency,
then that input is amplified in the spectral density of the response. The response
spectral density is not amplified, and may be diminished, between modes.

Example 2 Random Vibration of an MDOF Structure

The structure shown in Fig. 8 is a simple, spring-mass structure that has a series
configuration. All masses are equal with m = 8.64 × 10−4 lb − sec2 /in. The
stiffnesses are equal with k = 250 lb/in. The damping constants are equal with
c = 7.50 × 10−3 lb − sec/in. These parameters yield modal frequencies of
[38.1, 106.8, 154.3] Hz and damping factors of [0.020, 0.016, 0.019]. The structure
is excited at its base with a zero-mean, stationary random input; the input spectral
density is shown in Fig. 9. The input root-mean-square (RMS) is 2.29 g.

Fig. 8 Simple spring-mass

structure

-1
10
GXbXb(f), g /Hz
2

-2
10

-3
10 1 2
10 10
Frequency, Hz

Fig. 9 Spectral density of the base input applied to the structure of Fig. 8
7 Random Vibration and Mechanical Shock 371

2
10

|H2(f)|, g/g
0
10

-2
10 1 2
10 10
Frequency, Hz

Fig. 10 Modulus of the FRF for acceleration output at DOF 2 to base acceleration input

2
10
GXX(f), g /Hz

-1
2

10

-4
10
1 2
10 10
Frequency, Hz

Fig. 11 Spectral density at DOF 2 of the structure shown in Fig. 8 excited by the base-excitation
with spectral density shown in Fig. 9

The base acceleration input/absolute acceleration output FRF was computed for
the structure using the approach that led to Equation (47), but for a base-excited
structure. Then Equation (52) was used to compute response spectral densities at
the structural DOF. Figure 10 shows the modulus of the FRF. The spectral density
of absolute acceleration response at DOF 2 is shown in Fig. 11. The RMS of the
response is 4.64 g. Of course, because the structure has three DOF, there are three
peaks in the response spectral density, and they occur at the modal frequencies of
the structure. As long as the stationary random excitation has signal content at the
modal frequencies of the structure and response is measured at structural locations
where there are not nodes in mode shapes, peaks will occur at structural modes in
the response spectral density.

The theory of random vibration has many applications. To understand the

applications, we note that three items of information are involved in analysis of
stationary random vibration, namely, excitation spectral density, structure FRF, and
response spectral density. When two of the three pieces of information are available,
then, in principle, the third can be estimated. However, estimation of excitation
spectral density from the other two pieces of information requires, in some cases,
substantial caution.
372 T. L. Paez et al.

2.5 Mechanical Shock

Computation of structural response to mechanical shock follows the approach

developed in Sects. 2.1 and 2.2 when the structure under consideration is modeled
as linear. When a shock input is known, then one of the approaches to response
computation can be used to establish the desired measure of response. However,
if we were to limit our comments to this simple statement, we would omit the
fundamental measure of mechanical shock used in the laboratory and in analysis,
today. That measure is the shock response spectrum (SRS). The SRS aims to be a
measure of the severity – or damage-causing capability – of a shock. Its definition
and characteristics are developed in more detail in [13].
The SRS of a shock motion (usually analyzed for measured acceleration) is
defined as a frequency spectrum of peak responses excited in linear SDOF structures
by the shock motion. Let ẍb (t), t ≥ 0 denote a mechanical shock. Assume that it
is applied at the base of an SDOF structure with natural frequency ωn = 2π fn and
damping factor ζ and that it excites the absolute acceleration response ẍ(t), t ≥ 0.
The peak in the absolute value of the acceleration defines one ordinate in the SRS,
Sζ (f ), f > 0, of ẍb (t), t ≥ 0:

Sζ (fn ) = max |ẍ(t)| (54)

t≥0

By repeating this computation for SDOF systems whose natural frequen-

cies cover a range of frequencies [f0 , fmax ], we develop the entire SRS, Sζ (f ),
f0 ≤ f ≤ fmax , of ẍb (t), t ≥ 0. This form of the SRS is known as the absolute acceler-
ation maximax SRS; absolute acceleration because ẍ(t), t ≥ 0 is an absolute accel-
eration; maximax because it considers the maximum value of the absolute response
over all time. Other forms of the SRS can also be defined for alternate measures of
response, like displacement or velocity, or even for built-up measures of response
like ωn x(t), known as the pseudo-velocity. Of course, the response depends on
the damping factor of the SDOF structure involved in the SRS computation; that
dependence is denoted by the “ζ ” subscript on the SRS.

Example 3 Some SRS

Compute the absolute acceleration maximax SRS of:

1
[1 − cos (2π t/T )] 0≤t ≤T
ẍb (t) = 2 (55)
0 t >T

The shock signal ẍb (t), a classical pulse known as a haversine, is shown on the left
in Fig. 12. Its SRS is shown on the right.
The SRS of Fig. 12 shows a feature of all absolute acceleration maximax SRS.
They asymptotically approach the peak value of the time domain signal at high
frequencies.
7 Random Vibration and Mechanical Shock 373

1 0
10
0.8

Sζ (f), g
xb(t), g

0.6
0.4
0.2
-1
0 10 -1 0 1
0 1 2 10 10 10
Time, sec Frequency, Hz

Fig. 12 Haversine shock pulse (left) and its SRS (right)

3
80 10

40
Sζ (f)
xb(t)

2
0 10

-40
1
-80 10 1 2 3
0 0.5 1 10 10 10
Time, sec Frequency, Hz

Fig. 13 An oscillatory shock pulse (left) and its SRS (right)

Now compute the absolute acceleration maximax SRS of the shock shown on the
left in Fig. 13. This is known as an oscillatory shock, and may be the type measured
in practice. The SRS is shown on the right of Fig. 13.

The two shocks and their SRS, shown in Example 3, help to explain why the
SRS was developed in the first place (in the 1930s). Sometimes our laboratory test
equipment is not capable of reproducing shocks like the one shown on the left in
Fig. 13, yet we might have equipment capable of producing shocks like the one on
the left in Fig. 12; in instances like these, we need a method for establishing an
“equivalence” between two types of shock.
The method of shock response spectra specifies that if two shocks have “equiva-
lent” SRS, then the severities of the two shocks are equivalent. This holds even when
the shock time histories are dissimilar. The method of SRS permits replacement of
an oscillatory shock with a shock that has a classical pulse shape (or with another
oscillatory shock). This is useful when it is desired to obtain a test shock to represent
a collection of shocks measured in the field.
374 T. L. Paez et al.

Example 4 Application of the Method of SRS

In this example, we find a haversine test shock that is equivalent, in the SRS sense,
to the shock shown on the left in Fig. 13. To accomplish this task, we note that
[13] there are amplitude and time/frequency scaling laws that apply to all SRS.
Because the SRS is simply the responses of a sequence of linear SDOF structures
to a shock, the amplitude of every SRS scales linearly in the vertical direction;
i.e., as the amplitude of a shock is doubled, the amplitude of its SRS is doubled.
Further, for the same reason, when the duration of a shock is lengthened/shortened,
then the SRS shifts downward/upward along the frequency axis in proportion to the
lengthening/shortening of the shock signal.
Note that the peak in the SRS on the right in Fig. 13 occurs at an amplitude of
319 g and the peak in the SRS of the unit haversine on the right in Figure 12 is
1.59 g. Then a haversine with amplitude of 200 = 319/1.59 is called for. Further,
the peak in the SRS of the oscillatory pulse occurs at frequency 139 Hz, and the
peak in the SRS of the haversine occurs at 1 Hz. Then a haversine with duration
0.0072 = 1/139 is called for.
The test haversine is shown on the right in Fig. 14, and its SRS is shown with the
SRS of the oscillatory pulse on the left side of the figure.
The SRS of the test pulse is not a perfect match for the SRS of the oscillatory
pulse, but it cannot be because the time histories of the shocks are so far apart.
Indeed, to make the test pulse conservative with respect to the oscillatory pulse,
we normally enforce conservatism in the SRS “match.” The amplitude of the test
pulse is greater than the amplitude of the oscillatory shock on the left in Fig. 13. This
relation is typical; the SRS of classical pulses “equivalent” to oscillatory pulses tend
to be greater.

Of course, there are many caveats to the use of the method of SRS for
specification of test pulses; among the most important is that analysts and testers are

3
10
200

150
xtst(t)
Sζ (f)

2
10 100

50
1
10 1 2 3
0
10 10 10 0 0.005 0.01
Frequency, Hz Time, sec

Fig. 14 SRS of the oscillatory pulse from Example 3 and the SRS of a test pulse (left) and the
time history of the classical pulse (right)
7 Random Vibration and Mechanical Shock 375

encouraged to specify test with a duration that is comparable to or shorter than the
field-measured shock. This tends to assure that some devious method of establishing
a test pulse that is much lower than the oscillatory pulse and yet has an SRS that is
correctly specified cannot be developed.
The ideas of SRS and mechanical shock testing will be amplified in Sect. 4.

3 Random Vibration Testing

Vibration tests form a critical component of the environmental testing sequence. The
test structure, or test item, is exposed to a combination of oscillatory forces in the
laboratory for a specified time period. This laboratory testing simulates exposure
to a similar field environment. Such field vibration environments are induced
by road transportation, by aerodynamics, by propulsion systems, in shipboard
environments and sometimes by operation of the equipment itself. Vibration testing
was historically conducted using the most basic from of excitation, the sinusoidal
waveform a(t) = A0 sin (2π ft). Here A0 is the amplitude f is the frequency and t is
time.
Vibration amplitude is typically displayed in g s, where one g is the nominal
acceleration of earth’s gravity, 386.4 in/sec2 or 9.806 meters/sec2 . Vibration
amplitudes may also be quantified in terms of velocity or displacement. Tradition
favors acceleration because acceleration measurements are readily conducted using
piezoelectric accelerometers. Note that, as frequency increases, the peak velocities
and displacements which correspond to constant peak acceleration are defined by:

a(t) = A0 sin (2πf t)

A0 g
v(t) = cos (2πf t)
2πf

A0 g
d(t) = sin (2πf t) (56)
4π 2 f 2

Typical vibration tests cover some portion of the frequency range from 1 Hz
to 5000 Hz. For tests on an electrodynamic vibration machine, a typical test uses
frequencies from 10 Hz to 2000 Hz. At 10 Hz, 1 g corresponds to 0.2 inches peak-
to-peak displacement. At 2000 Hz, 1 g corresponds to 5 microinches peak-to-peak.
In general, lower frequencies imply higher velocities and displacements. Of course
pure sinusoidal testing is unusual today, but it remains true that lower frequency
vibrations, whatever their complex spectral content, imply greater velocities and
displacements.
Random vibration testing supplanted sinusoidal tests for two reasons, first
because random tests are much more realistic, and second because improved test
376 T. L. Paez et al.

technology allowed random vibration tests to be conducted expeditiously. The

advent of digital vibration control technology in the 1970s enhanced random
vibration test capabilities enormously.
Random vibration testing is a broad topic, encompassing the theory of random
processes (Sect. 2.3), practical test specification, differences between field and
laboratory tests, over-testing concerns, test setup, instrumentation, test conduct, data
analysis, and test reports.
Section 3.1 briefly reviews random processes, illustrates the autospectral density
(also referred to as the autospectrum), and shows some typical random process
realizations. Section 3.2 considers the essential differences between field and
laboratory tests with emphasis on boundary conditions and the problems involved
in test specification. Section 3.3 outlines the components of a random vibration test
in a block diagram format. Section 3.4 covers that major portion of a test, setup and
instrumentation. Section 3.5 describes test conduct.

3.1 Random Processes and the Autospectral Density

Section 2.3 provides a brief introduction to the theory of random processes. From
a practical testing standpoint, three features of a random signal are important: the
time history, the autospectrum, and the probability density [14]. The time history
is the time evolution of the random signal. The autospectrum is an estimate of the
normalized average squared magnitude of the frequency content as a function of
frequency, and the probability density is an estimate of the distribution of signal
magnitudes. A typical random vibration might expose a test item to an input
acceleration of 5 g s RMS (Root mean squared) over a frequency range from
20 Hz to 2000 Hz for 10 minutes. Figure 15 illustrates a typical realization of a
time history, the corresponding autospectrum, and the probability density function.
The time history is a typical broadband random signal. The autospectrum estimate
is computed from the average magnitude of the frequency lines in discrete Fourier
transforms (DFT) of successive blocks of the time history data [7]. (The estimate
is erratic because it is based on finite data.) The normalized estimation error is
approximately:

1
e= √ = estimation error
2N

N = Number of averages used to form estimate = Blocks of data (57)

In the example shown in Fig. 15, a total of 62, 1024 point data blocks are
averaged, so the expected error is about e = 0.09. This is consistent with the
variation in the estimated autospectrum, which shows about 10% variation about
a mean value of 0.0005 g2 /Hz. Autospectra are typically displayed on log-log plots.
7 Random Vibration and Mechanical Shock 377

Typical Random Process Time History

a 4 c Estimated Probability Density
0.45
3
0.4
2
0.35

PDF Amplitude
Amplitude in G

1 0.3

0 0.25

0.2
–1
0.15
–2
0.1
–3 0.05

–4 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 –5 –4 –3 –2 –1 0 1 2 3 4 5
Time in Seconds Signal Magnitude
Time History Estimated
Probability Density

Random Process Auto Spectral Density

b 10–2

10–3
G2/Hz.

10–4

10–5

10–6
101 102 103
Frequency in Hz.
Estimated Spectral Density

Fig. 15 Field tests and laboratory tests – implications for test specification

The desired test level is indicated by the red reference line. Tolerance bands of
+/ − 3 Db are shown by the cyan lines in Fig. 15. The definition of a decibel is:

Db = 20 log10 R (Ratio of two amplitudes)

(58)
Db = 10 log10 R (Ratio of powers)

where R is a ratio of two quantities, the denominator is a reference, and the

numerator is said to be Db above or below the reference. Since the autospectrum
is a squared quantity, the power ratio formula is used. Hence, 6 Db corresponds
to about 4 times the reference level and 3 Db to twice the reference level on the
g2 /Hz scale in Fig. 15. In Fig. 15, for example, the reference line is 0.0005 g2 /Hz.
The lower −3 Db tolerance line is at half this value, 0.00025 g2 /Hz, and the upper
+3 Db tolerance line is at twice 0.0005 g2 /Hz or 0.001 g2 /Hz. The root mean square
378 T. L. Paez et al.

√ to the upper +3 Db line is the square root of the g /Hz

(RMS) value corresponding 2

level, or approximately 2 × 1.013 g.

The probability density illustrated in Fig. 15c is typical for Gaussian random
noise. Lower amplitudes are more probable than higher amplitudes. The signal is
symmetric about the mean of zero. Magnitudes greater than about three times the
RMS value are improbable, so the typical peak value of Gaussian random noise is
often very roughly stated as +/− three times the RMS level.
A vibration test in the laboratory attempts to replicate the field environment.
Such replication means that a component failure in the laboratory should replicate
a similar failure in the field. Further, analytical models developed from laboratory
test measurements should represent field conditions and field loads. Even so, field
and laboratory conditions may differ substantially, as illustrated in Fig. 16.

Fig. 16 Test item in field (top, a) and laboratory (bottom, b)

7 Random Vibration and Mechanical Shock 379

In the field, Fig. 16a, the test item is potentially excited by a suite of forces, some
of which travel from a relatively discrete source through a supporting structure to the
test body and some of which may be distributed. Acoustic and aerodynamic forces
are examples of distributed forces. The forces act simultaneously in six degrees
of freedom, x, y, z translations, and rotations about x, y, z axes. The dynamic
loads on the test body are the result of a complex combination of the structural
characteristics of the test item, the mechanical impedance of the support structure,
and the excitation forces.
A conceptual laboratory environment is illustrated in Fig. 16b. The test item,
attached to a fixture, is excited by a vibration machine. The force path is through the
fixture interface into the test item. Distributed forces are not generally replicated.
Further, the fixture mechanical impedance differs from the field structure. In the
laboratory, the fixture is usually very rigid while in the field the structure is often
moderately flexible. In the laboratory, the input control is usually at the fixture-
test item interface where the field acceleration autospectrum is reproduced. Up to a
point, this is reasonable. However, in the field, test item resonances produce notches
in the acceleration spectrum at the control point. In the laboratory, the control
spectrum is an approximation to the complex field measurement. For conservatism,
this is usually accomplished by enveloping the measured field spectrum, thus
eliminating the notches and yielding some level of over-test.
In the field then, test item resonances reduce acceleration levels at some frequen-
cies. In the laboratory, test item resonances do not reduce control acceleration levels
because the vibration machine, acting with large force capability through a rigid
fixture, keeps the control acceleration spectra at the requested, enveloped, level.
This results in a substantial over-test at resonant test item resonant frequencies.
The over-testing issue has been recognized for many years. Several approaches
address the issue. One approach uses measurements of interface force in field and
laboratory. With combined force-acceleration control, the interface acceleration is
a function of required force, which mitigates the over-testing issue. This can work
well, but force measurements are often hard to implement in the laboratory and are
simply not available in the field.
A second approach to mitigate the over-testing issue is widely implemented. In
this scheme, the interface acceleration spectrum is notched at frequencies where
measured test item responses exceed field levels. This mitigates the over-testing
issue. Even so, it is evident that, across the test item, the range of acceleration values
experienced is quite different in field and laboratory environments, although the
laboratory environment in some sense approximates the field environment.
In the field, the test body experiences excitation in six degrees of freedom. These
are translations in the x, y, z directions, and the corresponding rotations about the
x, y, z axes. In the laboratory, accelerations are usually applied sequentially in x,
y, z axes. Clearly the test body responds quite differently in the multiaxis field
environment than in the single axis laboratory environment. A further complica-
tion is that, in general, rotational environments are only rarely simulated in the
laboratory. Recent advances in multi-shaker, multi-axis testing allow six degree of
freedom excitation, but to date these are fairly complex and expensive to implement.
380 T. L. Paez et al.

Research continues and some form of six degree of freedom excitation will likely
become common in future vibration tests.
In summary, then, current vibration tests expose a test item to random accelera-
tion with a specified spectral density, sequentially, in each of three orthogonal axes.
The spectrum is usually different for each axis. To mitigate over-testing, the spectral
response may be limited in certain frequency bands at some response locations.

3.2 Random Test Control Loop

A random vibration test exposes the test item to an acceleration time history.
This time history is a realization of a Gaussian random process with a specified
autospectral density. The desired spectral density is typically specified by a set
of straight lines on a log-log scale. The electrodynamic vibration machine and
power amplifier, when operating properly, are reasonably well approximated by a
set of linear differential equations. The combined dynamics of the power amplifier,
vibration machine, and test item result in a complicated spectral response at the
control point. In each frequency band, the system dynamics modify the gain and
phase characteristics of the voltage drive signal. This is compactly expressed in
the frequency response function (FRF) graph in Fig. 17, which illustrates the
acceleration response as a function of a constant voltage at each frequency. Below
about 5 Hz, the illustrated FRF is inaccurate. In any case, below 5 Hz, small
accelerations can exceed the maximum vibration machine displacement of between
1 and 3 inches peak-to-peak. 1 g peak at 3 Hz corresponds to a peak-to-peak
displacement of 2 inches. Between about 5 Hz and 100 Hz, the velocity response

Fig. 17 Typical control acceleration/voltage drive FRF

7 Random Vibration and Mechanical Shock 381

Random Vibration Control

Test Item

Signal
Power Conditioner
Amplifier
Vibration
Machine 1. Sample Rate
Time History
2. Blocksize

Inverse FFT
4. Control Gain Compare Autospectrum
Autospectrum to
Reference
3. Number of
Averages
Generate and Adjust Drive
Spectrum
Reference
Spectrum

Fig. 18 Closed-loop scheme used to control random vibration test

is proportional to applied voltage, so the acceleration response increases linearly as

a function of frequency. Above 100 Hz, the acceleration response is governed by the
electrical resonance of the shaker armature-power amplifier system combined with
the mechanical dynamics of the test item. Above about 500 Hz, the effects of test
item and fixture resonances dominate the acceleration response.
To produce a flat acceleration spectrum, the voltage applied to the power
amplifier must have the inverse of the frequency response function of Fig. 17. This
is accomplished using a closed loop control system like that illustrated in Fig. 18.
Historically, random vibration control was accomplished by passing a random
signal with a flat spectral density through a set of analog band pass filters. The
composite signal, the sum of the filter outputs, was applied to a power amplifier
driving the vibration machine. With the advent of digital vibration control in the
1970s, the analog filters were replaced by digital filters defined by the fast Fourier
transform (FFT). (The FFT is an efficient implementation of the DFT.) Using the
FFT a filter is effectively located at each discrete frequency line. Given a time
series sampled at a rate Sr , the time record is typically broken into N blocks, each
containing n − points; an FFT is performed on each block. Denote the blocks xjm ,
j = 0, . . . , n − 1, m = 1, . . . , N; within each block, the measurements are made
at times tj = j/Sr , j = 0, . . . , n − 1. Denote the DFTs of the blocks of random
process realizations Akm , k = 0, . . . , n − 1, m = 1, . . . , N. Within a constant, the
DFT represents the Fourier transform at frequency fk = kΔf, k = 0, . . . , n − 1,
where Δf = 1/(nΔt). Because the DFT is symmetric, the autospectrum is computed
using the frequency lines k = 0, . . . , n/2. The spectral density estimate is the
382 T. L. Paez et al.

scaled average of the squared magnitudes of the FFT values from each block. The
autospectrum estimator is:

2 
N
ĝ (fk ) = |Akm |2 k = 0, . . . , n/2 (60)
N nSr
k=1

In commercially available vibration control systems, typical sample rates range

from 2.56 to 4 times the highest control frequency. For example, for a test bandwidth
of 2 KHz, the sample rate ranges from 5120 to 8000 Hz. Block lengths are
traditionally a power of 2. For a block length of 4096 points, there are 20,489
spectral lines spaced 1.25 Hz apart.
The known response autospectrum is compared to the desired, or reference
autospectrum. The drive voltage spectrum defines the normalized, squared mag-
nitude of each frequency line in the drive voltage. If the response is too low at the
control point, then the drive voltage spectrum is increased at that frequency line,
and conversely if the response spectrum is high. To generate the drive voltage time
history, the phase of each frequency component is randomized and an inverse FFT
performed to generate a block of drive voltage. In practice, to prevent discontinuities
at block boundaries, a time windowing and overlap algorithm is used to combine
adjacent blocks.
The entire system in Fig. 18 runs in closed loop, with drive voltage continually
applied to the power amplifier and continuous autospectral measurements computed
from the time history response.
When a test is started at low level, drive voltage with a flat spectrum is applied to
the power amplifier. The control loop steadily modifies the drive voltage spectrum
until the response spectrum is within the desired bounds. These bounds are typically
+/ − 3 Db. The test level is then increased in steps until the full test level is reached.
At each step, a drive voltage magnitude correction is performed on the drive spectra
at each frequency line. Thus, ideally, the spectrum is just as specified at the full test
level.
If acceleration limits are desired, the limit autospectrum, derived from some
response accelerometer on the test item, is computed and compared to the desired
limit. If the limit is exceeded, the drive voltage in the affected bandwidth is
decreased until the response spectral density is within the desired limit. This
“notches” the spectrum at the control point.

3.3 Test Setup and Instrumentation

The most time-consuming part of a random vibration test is almost always the test
setup and instrumentation. Each test is unique, but there is a common thread. Table 1
provides a general guide to test setup and conduct.
7 Random Vibration and Mechanical Shock 383

Table 1 Test setup

Step Designation Comments
1 Instrument test unit
2 Bolt fixture and test unit to vibration table
3 Connect instrumentation to signal conditioning.
4 Set sensitivity
5 Setup control system test profile(s)
6 Check sensitivity and control profiles
7 Check that control transducer is in the test axis

Comments:

1. For optimum results, the test operator should follow a procedure, but not blindly,
and check each step in the process.
2. The test operator should check that the fixture attaches properly to the vibration
table and the test unit.
3. Instrument the test unit, typically with accelerometers. Use a robust adhesive
that is not prone to failure. Document the location of each transducer with text
and photos. Develop a chart listing transducer locations and sensitivities. Ensure
that the transducers are not inadvertently grounded to the test item. Grounding
tends to introduce excessive 60 Hz noise due to ground loops. Secure cable
attachments to the transducers. In some cases, the cable connection may be fixed
with adhesives. Two common problems causing signal loss in vibration tests are
failure of transducer mounting and poor cable connections. These both can occur
during the test and may be difficult to remedy.
4. Bolt the test unit to the vibration table. Ensure that the power amplifier is not
active. Over-testing has occurred in those (fortunately rare) cases where the
power was applied to the shaker during setup and an inadvertent signal resulted
in shocks or vibration applied to a partially mounted test system.
5. Torque all mounting bolts. Double check that all bolts are properly torqued. A
relatively common problem in test setup is the failure to properly torque all bolts.
One loose bolt can chatter, contaminating the test signals with accelerations
caused by impacts. These signals are usually high amplitude, high frequency
pulses.
6. Systematically connect all instrumentation to the signal conditioning system.
Document the connection order. Set and check all transducer sensitivity values.
Use a standard, documented output sensitivity. This sensitivity defines the volt-
age signal corresponding to a given acceleration value. An example sensitivity
might be 25 mv/g, meaning that 25 mv corresponds to 1 g. For a system with
10 volts full scale, this means a full scale acceleration level of 400 g. Setting the
sensitivity is a balancing act. Set a value that keeps the expected signals as large
as possible, but well below the system full scale range. For many vibration tests,
sensitivities ranging from 10 mv/g to 100 mv/g are optimal.
384 T. L. Paez et al.

7. Set up the control system. This means defining the maximum test frequency,
the desired spectrum, the number of frequency lines, and the control channel or
channels. Control may be defined by a single channel, but it may also be defined
by the average of the spectral density at several locations, or by a combination
of control and limit channels. Define the test duration and the initial test level. A
typical initial test level is 12 Db below the full level test. The RMS acceleration
at −12 Db is one-quarter that of full level and the corresponding spectral density,
in g2 /Hz, is one-sixteenth of full level.
8. Traditionally control autospectral values are defined using straight line segments
on a log-log plot. An example of a flat spectrum is shown in Fig. 4. Once
the autospectrum is input, check that the RMS level computed by the control
system matches the expected RMS value of the reference spectrum. Also
check that the displacement, velocity, and acceleration values estimated by
the control system for the desired reference spectrum are within the vibration
machine capabilities. This is somewhat problematic in many cases. Displacement
is computed from the lower frequency components present in the reference
autospectrum and is usually an accurate estimate. A peak-to-peak estimated
displacement significantly less than the vibration machine capability is a good
sign. Velocity estimates are also usually good. A problem arises with acceleration
limits since the actual acceleration achieved is a function of the combined effects
of the acceleration/voltage frequency response function of the combination of
test item and vibration machine. This may be available prior to testing. A rough
estimate can be made by dividing the vibration machine force capability by the
total test item plus fixture mass. For example, consider a vibration machine with
a force rating of 24,000 lb RMS. Suppose the armature weight is 100 lb, the
fixturing is 200 lb, and the test unit is 100 lb. Estimated maximum RMS g is then
24,000/400 = 60 g. In practice, this is a pretty rough estimate, but does provide
some guidance.

With the instrumentation complete and the control system setup, the next step is
conducting the test.

3.4 Conducting a Random Vibration Test

Conducting the random vibration test requires attention and vigilance. Even if one
has conducted hundreds of tests, it is important to treat each new test as a new
experience. While different vibration controllers have their own unique features,
the testing procedure for all is similar. Table 2 lists the typical sequence of steps
involved in test conduct. The most crucial item is the integrity of the control
loop. Common problems include incorrect transducer sensitivity and transducer
orientation. If the control system is set for a 100 mv/g sensitivity and the actual
transducer/instrumentation system sensitivity is 25 mv/g, then the controller will, to
achieve a 1 g response, increase the drive voltage until the 100 mv signal is reached.
This is then an overtest since the true level is 100/25 = 4 g. There are a couple of
7 Random Vibration and Mechanical Shock 385

Table 2 Conducting a random vibration test

Step Designation Comments
1 Start power amplifier and set gain Gain appropriate for test level
2 Check control loop integrity Control signal as expected, drive voltage
reasonable
3 Initiate data recording
4 Start test at low level Levels of in axis channels consistent
5 Check instrumentation responses All channels show reasonable response
6 Sanity check on responses Frequency response function of in-axis
channels unity at low frequency
7 Increase test level
8 Stabilize at full level
9 Conduct test for required duration Continue to check control and response
spectra and time signals
10 Reduce gain
11 Shut down power amplifier
12 Review test data Recorded? Reasonable?

ways to minimize this problem. First, double check sensitivity settings. Second,
have a “level check” transducer installed whose sensitivity is set to a constant
value for all tests. Compare the level check and control transducer levels at lower
frequencies where the test system approaches rigid body motion. They should be
approximately equal.
Another insidious problem in the control loop occurs when the control transducer
is inadvertently mounted in the wrong axis. This is especially easy to accomplish
when a triaxial control accelerometer is used. If, for example, the test item is driven
in the x axis, but the y axis is set up to read the control signal, then the measured
control is really the off-axis motion. The controller then applies a very large drive
voltage to make the off-axis response equal to the desired control level. This is
a huge over-test! Fortunately, the loop integrity check integral to startup of most
control systems will often read “open control loop” and thus avoid this problem.
Sometimes the cross motion excited by the loop check drive voltage is sufficient
to pass the loop check. Then the over-test occurs. There are a ways to minimize
this problem. First, double check the control accelerometer orientation. Second, for
critical tests, utilize a backup control accelerometer and set the control system to
control on the maximum of the two acceleration signals. Then, if one accelerometer
is incorrectly oriented, the control will, for most frequencies, be done using the
correctly oriented accelerometer. Once control is achieved, a sanity check confirms
accelerometer orientation and the controller can be set to control on the desired
control channel.
Be ready to abort the test if anything untoward occurs, like violent shaker motion.
The control autospectrum is often initially poor but should rapidly approximate
the desired spectral response. Check that the control and level check spectra are
in reasonable agreement (say, within 10%, +/ − 1 Db) at lower frequencies. This
386 T. L. Paez et al.

means that the autospectra nearly overlap. If these do not agree, it may indicate
problems with one of the accelerometers or instrumentation problems. “Lower
frequencies” usually mean frequencies from about 10– 30 Hz. In this frequency
range, most test systems respond nearly as rigid bodies and consequently all in-axis
accelerations are similar.
Observe the responses on all of the instrumentation channels. All channels should
have detectable random acceleration values. It is most unusual for a viable channel
not to show some response to random excitation. In-axis channels should, like the
level check channel, respond such that the low frequency autospectra overlay the
control autospectra. If any channels fail to respond, troubleshoot the instrumentation
system to isolate the issue.
Systematically increase the test level until full level is achieved. This is often
done in steps. For example, the test might be started at −12 Db and levels increased
in steps to −9 Db, −6 Db, −3 Db, and finally to full level. Once at full level, the test
is conducted for the desired duration. Continually monitor the test responses. It is
important to focus on the test, even with durations of hours. Failure to concentrate
on the test can lead to problems that could have been corrected but, undetected,
compromise the test. For example, if averaging or limiting is part of the test control,
loss of a limit acceleration channel could mean an over-test. Control systems are
designed to detect loss of signal, but this feature is by no means perfect.

4 Shock Testing on Shakers

Shock (transient vibration) testing became practical on electrodynamic and elec-

trohydraulic shakers in the late 1960s, as evidenced by a paper of Favour and
LeBrun [15]. The methods used have matured in the past decades, and several
commercial software packages are now available. In spite of the basic simplicity
of the technique, serious problems can arise. These are often due to nonlinearities
that can hamper successful tests. The purpose of this chapter is to review the basic
methods and explain why the tests don’t always work.
Shock testing conducted on shakers can be divided into two basic classes. The
first is reproduction of classical shock waveforms. By classical waveforms, we
mean such shapes as a terminal-peak sawtooth or a haversine, the first application
of shock testing on shakers. However, it was quickly realized that shakers had
inherent limitations that prevented faithful reproduction of these waveforms. Several
methods evolved to modify the classical waveforms so that they could be reproduced
on a shaker. Reproduction of classical waveforms has serious limitations because of
the velocity and displacement limits of the shakers. Except in special circumstances,
it is not recommended that classical shock testing be performed on shakers and
this method will not be further discussed. Shakers are better suited for reproducing
oscillatory waveforms of the type that are suitable for shock testing. It was also
realized that oscillatory shock environments are better simulated with an oscillatory
shock than with a classical shock test, like a drop test.
7 Random Vibration and Mechanical Shock 387

The most common method for measuring severity of a shock is the shock-
response spectrum (SRS, Sect. 2.5). This led to methods for preforming shock
tests on shakers: synthesizing a waveform to match an SRS and reproducing the
synthesized waveform on the shaker. The success of the test was the matching of the
SRS and not the faithful reproduction of the time history. There are several varieties
of the SRS, and there are many time histories that have essentially the same SRS.
The variations in the definition of the SRS will be discussed and the most common
methods to synthesize a shock time history to match an SRS will be reviewed.
The most common methods used today divide synthesis and reproduction into
separate phases of the test. This section is organized to reflect such division. In
some commercial software packages for shock testing on shakers, this separation is
not always apparent to the user.

4.1 Time History Synthesis

The ideas and limitations of time history synthesis are covered in this section.

4.1.1 Basic Shaker Limitations

Before starting the discussion of shock testing on shakers, we should consider
characteristics and limitations of shakers and their effect on the types of waveforms
we can reproduce.
Initial and final acceleration and velocity must be zero for electrodynamic and
electrohydraulic exciters. As with any testing machine, the maximum attainable
values of acceleration and velocity are limited by the design of the machine and
its power source. Flexures in electrodynamic exciters generate restoring forces
that return the exciter table to its original position (defined as zero). This is not a
requirement for electrohydraulic systems; however, by imposing this limitation, we
can take advantage of both the forward and return portion of the stroke to generate
the required transient.
The maximum acceleration that can be achieved is limited by two factors: the
physical strength of the shaker armature (the manufacturer supplies this informa-
tion) and the current and voltage capabilities of the power amplifier, combined with
the complex electrical impedance of the shaker with the test item attached. Voltage
and current being supplied by the amplifiers should be monitored. If peak voltage
and current capabilities of the amplifier (set by the manufacturer) exceed preset
maximum values, protection circuits shut down the amplifier. The acceleration limit
of a particular transient waveform reproduced on a shaker/power amplifier, with a
test item in place, is dependent on the system and is difficult to predict. Usually a
lowlevel test is run and the peak current and voltage are observed to determine this
limit. Another trick is to run a full-level transient through the power amplifier with
the shaker field off. This can be used to determine whether the amplifier will dump
without exposing the test item to a potentially damaging transient.
The peak-velocity capability of a shaker is determined by the voltage and current
drawn during the transient test. The usual velocity limits specified for a shaker are
388 T. L. Paez et al.

determined by steady-state cooling requirements and may not be representative

for transient testing. As for acceleration, the manufacturer sets velocity limits.
Experiments to test the transient limits can also be useful. The displacement limits
are usually rigidly determined by the shaker design.

4.1.2 Classification of Waveforms

Velocity and displacement requirements for a waveform are related to the initial
slope of the magnitude of the Fourier-spectrum of the waveform. This can be seen
by noting that velocity and acceleration are the first and second derivatives of the
displacement, respectively. Hence, the Fourier spectrum of velocity and acceleration
can be found from the displacement Fourier spectrum to within a constant by
multiplying by the frequency and the frequency squared, respectively. The SRS
at moderately low frequencies is dominated by the residual spectrum. Since the
undamped residual-shock spectrum and the Fourier spectrum are related, we would
expect the initial slope of a lightly damped SRS also to be related to velocity and
displacement requirements of a waveform. Many field measurements of the SRS of
oscillatory shocks are corrupted at the lowest frequencies resulting in an incorrect
slope of the SRS at low frequencies, suggesting a larger required velocity than
realistically possible. The slope of the required SRS for transient testing on shakers
should be carefully monitored to make sure the resulting synthesized waveforms
can be reproduced on the shaker.

4.1.3 Matching a Required Shock Response Spectrum

The concept of the SRS [13] is deeply ingrained in dynamic testing. We will not
develop the concept here except to say that the shock spectrum is a measure of the
peak response of a single-degree-of-freedom system to a given input, plotted as a
function of the natural frequency of the system. The shock spectrum is probably
the most used, misunderstood, and abused concept in dynamic testing. It is a useful
concept, and many shock environments are described and specified in terms of their
shock spectra. As a result, a large effort has been made to synthesize waveforms that
have some specified shock spectrum. At the outset, we note that this synthesis is not
unique; that is, many waveforms exist that have essentially the same shock spectrum.
It is not clear that all waveforms with the same shock spectrum will produce an
equivalent test or produce the same damage to a test structure. Test engineers should
have several waveforms available that will produce similar spectra so they can pick
the one that best satisfies test requirements. Listed below are some of the many
supplements to the shock spectrum that have been suggested to restrict the class of
waveforms that can be used for a particular test. None has been universally accepted.

Limit the duration of the transient. It has been suggested that limits could be placed
on the minimum and maximum allowable durations for the transient. It is felt
that if the shock spectrum is matched and the duration is comparable to the
field environment, the “damage” should be the same. For complex waveforms,
careful attention should be given to “defining” the duration: STD-MIL-STD-810
suggests this method. This is the method favored by the author.
7 Random Vibration and Mechanical Shock 389

Require the shock spectrum to be matched at two different values of damping.

Because damping is the least-known parameter in many systems, it is felt that if
the shock spectrum is matched at two different values of damping (for example,
a Q of 5 and 25), the resulting transient should be a reasonable simulation for all
values of damping. However, it is very difficult to find a transient that will match
a shock spectrum at two values of damping except for a limited class of functions.
In fact, it is not even clear that a solution always exists. It is true, however, that a
solution can exist. An example is the set of shock spectra with different damping
values for a known time history. But if those spectra are modified (for example,
smoothed, raised in level, or enveloped), it is not clear that a solution will still
exist.
Specify the allowable ratios of the peak shock response and the peak input level.
Encourage the use of oscillatory-type input as opposed to a single pulse (for
example, a half-sine) input. Specifically exclude certain methods.

4.1.4 Time History Synthesis Using Oscillatory Waveforms

Several methods have been developed to synthesize a time history that will match
a specified SRS. These methods have many similarities, and all are iterative to
converge to a satisfactory solution. One method will be discussed in detail and the
others more briefly since the basic procedures are the same for all. We will first
discuss using sums of exponentially decaying sinusoids, then several variations of
decaying sinusoids. Another method is WAVSYN (also called wavelets, [16]). The
basic waveform of a WAVSYN pulse is an odd number of half sine waves windowed
with a half sine. Other techniques, which include shaker optimized cosines (SHOC,
not used much anymore), fast sine sweeps (not recommended), modulated random
noise (frequently used), modification of field-time histories (used in the seismic
industry), and least-favorable responses, are discussed in [17].

Sums of Decaying Sinusoids

It has been recognized for years that many field environments can be adequately
represented by sums of decaying sinusoids [18, 19]. For example, the response of a
structure to an impulsive load is the impulse response of the structure, which can be
represented as a sum of decaying sinusoids. A component mounted on that structure
will see this impulse response as an input.

The basic waveform for an exponentially decaying sinusoid is given by

a(t) = Ae−ζ ωt sin ωt (ωt) t ≥0 (61)

Figure 19 is a plot of the basic waveform. The waveform has three parameters
that can be used to change its shock spectrum: the amplitude (A), to raise and lower
the entire curve; the frequency (ω), to change the location of the peak in the curve;
and the decay rate (ζ ), to change the shape of the curve. Note that the decay rate is
a parameter of the waveform and is not related to the fraction of critical damping
used to compute the shock spectrum. By forming a sum of the basic waveforms, each
390 T. L. Paez et al.

Fig. 19 Exponentially decaying sinusoidal acceleration and corresponding velocity and displace-
ment

with different amplitude, decay rate, and frequency, we should be able to synthesize
a waveform with almost any shock-response spectrum.
However, this waveform does not meet the requirements for a zero-velocity and
zero-displacement change. If an attempt is made to reproduce this waveform, it will
be distorted by the shaker system, removing the velocity and displacement change.
Because the exact distortion will be a function of the shaker, it is difficult to predict
the velocity and displacement requirements for the test.
Several modifications have been suggested to remove the objectionable velocity
and displacements. Nelson and Prasthofer [19] suggested velocity and displacement
compensation by adding two exponential pulses and a phase shift to the basic
waveform.

x(t) = Aωa 2 Ka e−at + Aωb2 Kb e−bt + A c2 + ω2 Kc e−ct sin (ωt + ϕa )

 
Ka = 1/ (a − b) (c − a)2 + ω2
 
Kb = 1/ (b − a) (c − b)2 + ω2
  
Kc = 1/ (c − b)2 + ω2 (c − a)2 + ω2
7 Random Vibration and Mechanical Shock 391

     
−1 −2cω −1 ω −1 ω
ϕa = tan − tan − tan (62)
c2 − ω2 a−c b−c

The first two terms are added for velocity and displacement compensation. The
phase shift ϕa is added to force the initial values to zero. Considerations in the
Laplace domain led to the above form. With this method, each individual component
is compensated. Fisher and Posehn [20] derived a different method to accomplish
the same objective (an initial and final acceleration, velocity, and displacement of
zero). They called the method ZERD (zero residual displacement).
A third method [18] is to add a highly-damped, low-frequency, delayed, expo-
nentially decaying sinusoid to compensate for velocity and displacement, CEDS
(compensated, exponentially decaying sinusoids). This method has the advantage
of correctly compensating for velocity and displacement when the acceleration
waveform is truncated. As a result, lower values of decay rates can be used than
for either Nelson’s or Fisher’s method. The time history of a single compensated
component using (1 Hz, 1 g amplitude, 3% damping, 0 delay, is shown as Figure
(EXPScTH.pdf). The compensated waveform is shown in Fig. 20. The SRS of the
un-compensated and compensated waveforms is shown in Fig. 21.

Compensatted 1 Hz exponentially decaying sinusoid

1
Acceler at ion ( g)

0.5

-0.5

-1
0 5 10 15 20 25

2
Velocity ( m/ s)

-1

-2
0 5 10 15 20 25

0.5
Displacement (m )

-0.5

-1
0 5 10 15 20 25
Time(s)

Fig. 20 A compensated, exponentially decaying sinusoid

392 T. L. Paez et al.

1 Relative Displacement SRS no compensation

10
Maximax
0 Residual
10
SRS ( g)

-1
10

-2
10

-3
10
-2 -1 0 1 2
10 10 10 10 10

2 Relative Displacement SRS with compensation

10

0
10
SRS ( g)

-2
10

-4
10
-2 -1 0 1 2
10 10 10 10 10
Natural Frequency (Hz)

Fig. 21 SRS of uncompensated and compensated exponentially decaying sinusoids

Each of the three previous methods should give similar results. They are all based
on exponentially decaying sinusoids (Figure 20) and have three free parameters for
each component: amplitude, frequency, and decay rate. A fourth parameter, delay,
is usually included. The delay is used to change the start time of a component. In
this way, the components can be delayed with respect to each other and with respect
to the start of the data frame.
It is probable that all three methods are still being used in some software
packages. We will illustrate matching an SRS using Smallwood’s method.
The iteration used is straightforward:

1. The shock pulse is a sum of indexed terms, each of which mimics Equation (61).
2. The amplitude of each component is changed a small amount, but not allowed to
change sign.
3. The effect on the shock spectrum at the component’s frequency is calculated.
4. A linear extrapolation of the amplitudes to produce the desired shock spectrum
is then made.
5. For stability, at each iteration only a user-specified fraction of the correction is
made.
7 Random Vibration and Mechanical Shock 393

The sample rate of 10,000 samples/second was used. English units were used
with g = 386.088 in/sec2 . The SRS was the absolute acceleration model, with
12 natural frequencies/octave. The frequency of the compensating pulse was 40 Hz.
The resulting acceleration waveform was 1024 points long. The frequencies of the
decaying sinusoids are spaced at 1/3 octaves from 100 to 2000 Hz rounded to 2
significant digits.
The reference SRS has the break points provided in Table 3.
The decay rates were set so the components decay three time constants at the end
of the pulse. The component delays were all set to zero. The initial amplitudes of
the components were set to 1 and then iterated to the final amplitudes.
The final results are provided in Table 4.
The last component (15) is the compensating pulse to force the final velocity and
displacement to zero.
Figure 22 shows the acceleration, velocity, displacement, and SRS of the final
result.

Table 3 Break points for Frequency (Hz) SRS (g)

reference SRS
100 10
1000 1000
1500 2000
2000 2000
10,000 1600

Table 4 Components in the sum of compensated, exponentially decaying sinusoids

Decay Reqd. Actual
Index(k) Freq.(Hz) Ampl.(g) Rate(ζ ) Delay(sec) SRS(g) SRS(g) Percenterror
1 100 1 0.047 0 10 10.69 6.9
2 130 −2.44 0.036 0 16.9 16.78 −0.7
3 160 2.6 0.029 0 25.6 24.64 −3.7
4 200 −4 0.023 0 40 41.06 −2.6
5 250 6.2 0.019 0 62.5 61.42 −1.7
6 320 −11.45 0.015 0 102.4 99.75 −2.6
7 400 13.08 0.012 0 160 165.2 3.2
8 500 −25 0.009 0 250 248.7 −0.5
9 630 32.91 0.007 0 396.9 395.7 −0.3
10 800 −52.22 0.006 0 640 633.7 −1
11 1000 85 0.005 0 1000 1003 0.3
12 1300 −127.7 0.004 0 1566 1581 1
13 1600 194.3 0.003 0 2000 2016 0.8
14 2000 −176.5 0.002 0 2000 2069 3.5
15 40 2.71 1 −0.004 0.01 1.64 0
394 T. L. Paez et al.

test 1
500 20

10
acceler at ion ( g)

velocity (in/s)
0 0

-10

-500 -20
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
time (s) time (s)

-3
x 10 4 ζ = 0.05
6 10

4 SRS, peak response (g)

displacem ent (in)

2
2 10

0
0 SRS
-2 10
Reference
-4 ±tol
-2
-6 10
0 0.05 0.1 0.15 0.2 0.25 1 2 3 4
10 10 10 10
time (s)
Natural frequency (Hz)

Fig. 22 Acceleration (top left), velocity (top right), displacement (bottom left), and SRS (bottom
right) of the example

4.1.5 Problems with Synthesis

When any of the above-mentioned methods are used, problems are sometimes
encountered. Some of these problems are now discussed.
Iteration will not converge – The components of the synthesized waveform
are assumed to be independent from each other during iterations when in fact they
are not. If the components are too close together, the damping is too large, or if the
slope of the shock spectrum is too large, a solution may not exist. A small increase in
the amplitude of one component can sometimes reduce the shock spectrum at that
frequency because of the interaction between adjacent components. The absolute
amplitude of a component cannot be reduced to less than zero even if the shock
spectrum is high at that frequency; the component amplitude must not be allowed to
go through zero. Changing the sign of the component amplitude will not, in general,
lower the shock spectrum. The iteration software must be carefully written to avoid
problems of convergence when these anomalies occur.
If the shock spectrum at high frequencies is less than the spectrum at some other
frequency, a solution may not exist because the peak amplitude required to achieve
the correct spectrum at the lower frequency may be larger than the required spectrum
at the high frequencies. Remember that the shock spectrum at high frequencies will
7 Random Vibration and Mechanical Shock 395

approach the value of the peak acceleration. If convergence is a problem, you must
determine why. Suggested solutions include: lowering the decay rates; reducing
the number of components; changing the component frequencies; and changing the
signs of the amplitudes of some of the components.

The spectrum matches at the component frequencies but is too low between the
frequencies – To solve this problem, the number of components can be increased,
placing the frequencies of the added components near the low places in the
spectrum. The decay rates can be increased, filling in the low spots. The sign
of the component amplitudes can be changed; however, the components interact
in unpredictable ways when the sign of one of two nearby components is
changed.
The spectrum matches at the component frequencies but is too high between some
frequencies – You can lower the decay rates for components near the problem,
change the sign of the amplitude of one of the nearby components, or drop a
component.
The resulting waveform is not within the shaker capabilities – If the acceleration
is too high, you must look for ways to reduce the peak acceleration. Delays
of some of the components will sometimes work, as will changing the signs
of some of the component amplitudes. A different waveform type could give a
lower peak input. Lowering the decay rates will allow more resonant buildup
during SRS computation, resulting in lower peak inputs. Sometimes, the fre-
quency range over which the spectrum is matched will have to be narrowed to
achieve an acceptable peak acceleration. If the velocity is too large, the low-
frequency components are usually the cause, and a compromise at the low
frequencies is sometimes required. A large displacement is also usually caused
by the low-frequency components. A different waveform will have different
velocity and displacement requirements; a change of waveforms sometimes
helps. If compromises required to achieve a waveform within the shaker capa-
bilities are not acceptable, the test should be moved to some other shock
facility.

Historically, field shock spectra are often incorrectly high at low frequencies.
Incorrect algorithms have been and are being used. Small errors in the zero line
or waveform truncation can cause big errors in the low-frequency end of the shock
spectrum. The SRS of a field environment that involves a very small velocity change
should roll off between 12 and 18 dB/octave at low frequencies. Specifications that
require large velocity changes should be questioned: Is a large-velocity change part
of the environment being simulated? If a large-velocity change is involved, is a
shaker the appropriate place for the test?

4.2 Time History Reproduction

The idea and limitations of time history reproduction are covered in this section.
396 T. L. Paez et al.

4.2.1 Classical Theory

Time history reproduction on shakers rests on the Fourier transform theory. More
specifically the discrete Fourier transform (DFT). A complete discussion of the
theory is given in several standard texts [13].
In this section, a lower-case variable is generally a function of time, and its
Fourier transform (or DFT) is given by the corresponding upper-case variable.
An important theorem connected to the Fourier transform relates the Fourier
transforms of functions involved in a convolution (see Sect. 2.1.2). Let x(t), y(t),
z(t), − ∞ < t < ∞ be functions of time, and let X(ω), Y(ω), Z(ω), − ∞ < ω < ∞
be their respective Fourier transforms. If the convolution:
 ∞
z(t) = x(t) ∗ y(t) = x (τ ) y (t − τ ) dτ − ∞ < t < ∞ (63)
−∞

is defined, then:

Z (ω) = X (ω) Y (ω) − ∞ < ω < ∞ (64)

These equations state that a multiplication in the frequency domain is equivalent

to a convolution in the time domain. The former equation is important because it
can be shown that for a single-input/single-output linear system, the output can
be found from the convolution of the system’s input and the impulse response of
the system (for example, Equation (20)). The Fourier transform of the impulse
response function is called the system frequency response function (FRF). There-
fore, the Fourier transform of the output can be found from the product of the FRF
and the Fourier transform of the input.
The shaker’s transient control problem is a variation of the classical convolution
problem. In our case, if we can determine the frequency response function, the
required shaker drive (the input) can be determined from the desired output and
the measured FRF:

Y (ω)
X (ω) = −∞<ω <∞ (65)
H (ω)

This class of problems is called deconvolution and presents more problems than
convolution, as will be explained later.
The continuous Fourier transform is never calculated when testing on shakers.
The FFT (fast Fourier transform) is used, which is an algorithm for computing
the DFT. The differences and similarities between the DFT and the continuous
transform are discussed in several texts [21, 22]. The three basic problems caused
by using the DFT are aliasing, circular convolution, and leakage.

Aliasing is caused by using an inadequate sampling rate. High frequencies appear

as low frequencies; therefore, analog, anti-aliasing filters should always be used.
Filters with linear phase in the frequency band of interest should be used to
7 Random Vibration and Mechanical Shock 397

measure the response of the system. It is not essential to use linear phase filters
with the drive waveforms.
Circular convolution is caused by the implicit assumption of periodic waveforms
in the DFT. The convolution of two waveforms results in a waveform that has
duration equal to the sum of the durations of the two individual waveforms. This
can be seen by looking at the equation that defines a convolution. We will perform
convolutions by taking the product of two DFTs. The two original time histories
must contain enough zeros in the frame to assure that the convolved sequence
will fit into the data block. If this is not done, circular convolution errors result.
Leakage arises because of the periodic assumption of a DFT. A DFT of length N
assumes that the (N + 1)st point in the data sequence is the same as the first
point. If this is not true, a discontinuity exists at the frame boundary. Extra
frequencies will be generated to match this discontinuity; this is called leakage.
Discontinuities in the derivatives at the frame boundary also cause leakage. If
we multiply the original time history by a window to eliminate the discontinuity
at the frame boundary, we are still performing a convolution in the frequency
domain of the window and the original time history. This will smear the original
frequencies over the bandwidth of the window. Leakage will be reduced but not
eliminated. Leakage and the effects of windows are important because of the
methods discussed later to measure the FRF.

4.2.2 Duration of a Transient Waveform

The ith temporal moment of the function f (t), − ∞ < t < ∞ about a is defined as

 ∞
mi (a) = (t − a)i f 2 (t)dt i = 0, 1, . . . (66)
−∞

The zero moment is called the energy (E) and is independent of a. The value of
a, which makes the first moment zero, is called the centroid (τ ). The square root of
the second moment about the centroid (τ ) normalized by the energy E is called the
RMS duration [13] Dt .

 ∞
m2 (τ ) 1
Dt2 = = (t − τ )2 [f (t)]2 dt (67)
E E −∞

The RMS duration is a convenient way to define the duration of a transient.

One of the flaws of transient testing using the SRS is that there are no restrictions
on the duration of the waveform. It is not clear that waveforms with the same SRS
but with different durations have the same damage potential. One way to avoid this
issue is to specify the duration, with tolerances, of the test waveform.
There is an interesting relationship between the RMS duration and the RMS
bandwidth of a transient. The RMS bandwidth of the Fourier transform of f (t),
namely F(f ), is computed about its centroid, c, and is defined as:
398 T. L. Paez et al.

 ∞
2
Df2 = (f − c)2 |F (f )|2 df (68)
E 0

The product of the RMS bandwidth and the RMS duration must satisfy the
inequality:

1
Df Dt ≥ = 0.0796 (69)
4π

This implies that we cannot have a transient with a short duration and a narrow
bandwidth.

4.2.3 Measurement of the System Frequency Response Function

The first step in reproducing a transient is to measure the FRF. Figure 23 outlines
the basic procedure. A low-level calibration waveform is applied to the shaker with
the test item attached. It is usually essential to have the test item mounted because
the dynamic loading of the test item can greatly influence the FRF. Many forms of
a calibration waveform can be used (Table 5). The only strong requirement is that
the waveform has nonzero frequency content at all frequencies of interest. The RMS
level of the waveform should be large with respect to the RMS noise in the system to
maximize the signal-to-noise ratio. However, consideration must be made to assure
the calibration procedure is not more damaging than the test waveform.
The first waveform used was an exponential, decaying pulse [15]. This waveform
has good frequency content. The frequency content is flat to a cutoff frequency
defined by the decay rate of the exponential and then gently rolls off. The faster
the waveform decays (the shorter the pulse), the wider the frequency content of the
pulse. A short pulse degrades the signal-to-noise ratio as the average value of the
pulse over the whole data frame decreases. Thus, a conflict arises: a short pulse

Fig. 23 Basic procedure for obtaining the FRF

7 Random Vibration and Mechanical Shock 399

Table 5 Waveforms that can be used for system identification

TYPE S/N Frequency content Noise rejection
Fast sine sweeps Good Good Fair
Random burst Good Good Good
Exponential decaying pulse Poor Good Poor
Reference waveform Fair Poor Fair/poor

has better frequency content but increases the noise in the FRFs. The same pulse is
repeated for averaging; therefore, this wave form cannot reject nonlinearities.
The next waveform used was a fast sine sweep. This waveform had a better
signal-to-noise ratio but still could not reject nonlinearities. To prevent leakage, the
sine sweep must be short enough for the response to decay before the end of the
frame.
Another frequently tried waveform uses the reference waveform as the calibra-
tion pulse. This is often a poor choice because the reference waveform may have
near zeros in its frequency content.
Another popular transient waveform is a random burst. The burst is usually re-
randomized between averages to reject nonlinearities. To prevent leakage errors, the
burst must be short enough for the response to decay before the end of the frame.
Steady-state excitation is also used. Steadystate random excitation is popular.
This waveform has a good signal-to-noise ratio and can reject nonlinearities.
However, a window has to be used to prevent leakage, causing a smearing of the
frequency lines and distorting the FRF.
Pseudo-random can be used. It is usually defined as a waveform that appears to be
random, but is really one period of a very complicated periodic signal. If a pseudo-
random signal is repeatedly output and the period of the digitized waveform is the
same as the pseudo-random period, it is not necessary to window the data to prevent
leakage. The estimate of the FRF is thus not biased by the window. The signal-to-
noise ratio is good and, while uncorrelated related noise is rejected, nonlinearities
are not.
The Fourier transforms of the calibration waveform C(ω) and the system
response R(ω) are calculated. In the simplest form, the FRF is estimated by taking
the ratio of these transforms. The FRF is then stored for later use. In practice, the
FRF that is found from the ratio of the transforms of a single calibration pulse tends
to be noisy and is of limited use. Some averaging is usually required to produce a
satisfactory FRF. A convenient form for estimating the FRF when averaging is given
by:

Gcr (ω)
H2 (ω) = (70)
Gcc (ω)

where Gcr (ω) = Ave[C∗ (ω)R(ω)], the cross-spectrum between the input and output,
and
Gcc (ω) = Ave[C∗ (ω)C(ω)], the auto(power) spectrum of the input
400 T. L. Paez et al.

For one average, this formulation reduces to the simple form:

Gcr (ω) C ∗ (ω) R (ω) R (ω)

H2 (ω) = = ∗ = (71)
Gcc (ω) C (ω) C (ω) C (ω)

Another advantage of using the above formulation is the availability of the

coherence function. The coherence is a measure of the quality of the linear FRF
estimate. The magnitude of the FRF can be found from

Gcr G∗cr
|H2 (ω)|2 = (72)
Gcc Gcc

For a noise free linear system, another estimate of the FRF is given by

Grr (ω)
|H1 (ω)|2 = (73)
Gcc (ω)

where Grr (ω) = the autospectrum of the response.

The coherence is defined as

Gcr G∗cr |H2 (ω)|2

γ 2 (ω) = = (74)
Gcc Grr |H1 (ω)|2

A coherence of less than one indicates the presence of uncorrelated response.

Thus, a coherence of almost one indicates a linear system with little noise.
The coherence will drop when nonlinearities are present in the system if the
excitation is varied in some random way between data frames. If both input and
output are deterministic, the coherence will stay almost at one even if the system
is nonlinear. To detect nonlinearities, the phase relationships between different
frequency components must change from one excitation pulse to the next.
These equations illustrate why it is important that the input waveform have a
non-zero frequency content at all frequencies of interest. If the Fourier spectrum is
near zero at any frequency, the FRF will be the ratio of two small numbers, resulting
in a poorly defined value at that frequency.
The formulation above can reject uncorrelated noise from the estimate of the
FRF. If the system is nonlinear, the formulation will reject the nonlinearities if the
input waveform is varied in some random way between averages.
The level that should be used to determine the FRF is never clear. A high level
is desired to minimize effects of the assumption of linearity between the calibration
level and the test level. This is particularly important with electrohydraulic shakers
because these shakers are nonlinear. However, if the level is too high, the calibration
pulse to determine the FRF can be more damaging than the actual test. Some of
7 Random Vibration and Mechanical Shock 401

Fig. 24 Procedure for performing a shock test

the waveforms require a long excitation time of the test item. Periodic random is an
example. The test time as well as the level must then be considered.
Also, the excitation used to determine the FRF is frequently broader in frequency
content and has a higher RMS-to-peak ratio than the test pulse. The customer should
always be aware of the requirement to excite the system to determine the FRF, and
the levels used should be agreed on.
The system FRF H(ω) is estimated from the input and response of the system to
the calibration waveform. The drive waveform x(t) required to reproduce the desired
waveform y(t) is then computed in the frequency domain using:

X (ω) = Y (ω) /H (ω) (75)

Careful attention must be used when preforming this calculation to avoid the
errors discussed above. Many times it is necessary to window the impulse response
of the FRF inverse to prevent leakage.
The computed drive waveform is then used to drive the shaker, and the result
is observed in Fig. 24. Typically several reduced level transients are output to
determine if the results are satisfactory before the full level test is performed. The
FRF and the computed drive may be updated during this phase.
Often the success of the test is measured by the SRS of the control accelerometer,
and not the exact reproduction of the desired waveform.

4.2.4 Why Things Do Not Always Work

Deconvolution is not easily and accurately accomplished; many times it is ill-
conditioned, and a good solution is difficult. The software needs a bag of tricks
to help make things as easy as possible. If the FRF has zeros or notches at any
frequency, they become maximums when the FRF is inverted. If these notches
402 T. L. Paez et al.

are noisy (random amplitude and phase errors), the noise is amplified. The noise
causes ripple in the amplitude and phase of the FRF, which tends to spread the
time history of the inverse impulse response (IIR) as the RMS duration discussion
illustrated earlier. Typically, the FRFs found in the test laboratory do have poorly
defined notches. The FRF tends to roll off at low frequencies, resulting in a notch
at zero frequency. Often the high frequencies in the FRF have several very noisy
notches. Even if the desired waveform does not have much energy content at the
high frequencies, the computed pulse will, because of the noisy FRF at the high
frequencies. The system is trying to say that if the FRF is poorly defined in a
notch, a large, poorly defined input is required to achieve the desired response at
that frequency.

4.2.5 Why Things Do Not Always Work, an Extreme Example

The desired waveform (a sum of exponentially decaying sinusoids) is shown in
Fig. 25. A measured FRF is shown in Fig. 26. The resulting reproduced waveform is
shown in Fig. 27. The result is satisfactory. A noisy notch is introduced into the FRF
(Fig. 28). The desired waveform along with the simulated reproduced waveform is
shown as Fig. 29. As can be seen, any resemblance to the desired waveform is
coincidental.

4.2.6 Improving Your Chances for a Good Test

Several things can be done to improve your chances of a good test (if your software
will let you). Always preview the FRF before running the test. A little experience
will frequently identify a bad FRF. Preview the drive waveform. Again, experience
will indicate problems before, rather than after, the test. Delete part of the FRF.
Often a band of frequencies (usually the high frequencies) will have little effect
on the resulting drive. If the drive at these frequencies is poorly defined, they can
be deleted. For example, consider the case of a desired waveform composed of
exponentially decaying sinusoids, with a highest frequency component of 1000 Hz.

Fig. 25 Desired time history

7 Random Vibration and Mechanical Shock 403

Fig. 26 FRF based on

accurate measurements

Fig. 27 Test time history

obtained during accurate
analysis

Fig. 28 FRF with noisy

notch

Assume that a sample rate of 10,000 samples/sec was used. The FRF could be
calculated to 5000 Hz in some systems. The frequencies above 1500 Hz are not
necessary for good reproduction of the waveform. If a noisy FRF is causing a
404 T. L. Paez et al.

Fig. 29 Test time history obtained using FRF with noisy notch

problem, the FRF could be deleted above 1500 Hz. This example also indicates
how an ideal notch (deleted frequency lines) can sometimes have less effect than a
noisy notch.
Multiply the inverse FRF by the coherence; some authors have called this
function H2 . The modified FRF will deemphasize those frequencies where the
uncertainty is the greatest.
Smooth the FRF; a Hanning smoothing in the frequency domain often gives good
results.
Some of the transient control packages in use offer an option of correcting the
drive, based on an error signal derived from the difference between the actual return
waveform and the desired waveform. These options frequently do not work very
well because the error (the difference between the actual return waveform and the
desired waveform) can be divided into two parts. The first is deterministic – the part
for which an identical input will produce the same output. This part is created by
two factors. The estimated FRF may not be the true FRF but contains some errors
and nonlinearities. This factor can be corrected by dividing the Fourier transform of
the error by the FRF and then transforming the correction to the time domain and
subtracting it from the drive waveform. For stability, sometimes only a fraction of
the correction is used. The correction should produce a response equal and opposite
to the error, thus canceling the error. However, part of the error is not deterministic.
We have seen how this error can be amplified by the deconvolution process. The
correction calculated will not improve this error, but will make it worse. After a few
iterations the error can grow enough to dominate the response.
Another iteration method uses the drive waveform and the response of the control
point to update the FRF. In concept, multiple applications of the test waveform
7 Random Vibration and Mechanical Shock 405

will improve the FRF, and the difference between the desired waveform and the
actual shaker response at the control point should be reduced. In practice, the drive
waveform will usually have notches in the frequency domain. These notches result
in poor estimates of the FRF, and the error between the desired and actual waveforms
will degenerate instead of improving.
The frequency content of the waveform can be changed to improve the estimate
of the FRF. For example, the lower frequencies are often difficult to measure because
of the low response at low frequencies. Additional energy in the calibration pulse
can improve the results. The mechanical setup can also be improved. Joints that
open or slip can cause problems. If these joints are part of the setup, they can be
improved.
The control accelerometer can be moved since its location is often poorly defined.
For example, the control could be specified on a flange; moving it a few inches can
result in significantly different responses in the kilohertz frequency ranges.
The largest cause of poor results in transient testing on shakers is nonlinearities.
Some of the most common forms of nonlinearities follow:

Rattling loose parts – The FRF is sensitive to amplitude and the response is
nonsymmetrical. For example, when the input is <1 g, the FRF looks normal.
When the input exceeds this amount, the nonlinearities become apparent. The
only solution to this problem is to remove the loose parts.
Gapping joints – The FRF will be sensitive to amplitude, but the input level required
is not necessarily 1 g, for the loose-part problem. It can vary depending on the
load required to open the joint. Typically, the FRF will look normal because the
levels used to measure the FRF were not high enough to open the joint. When
the test is run at low levels, the results are good. But at some higher level, the
waveform reproduction looks good part-way into the transient; however, it then
breaks up, showing increased high-frequency content from impacts as the joint
closes. Little can be done except to tighten the joint. If the joint is within the test
item, the test engineer frequently cannot change the joint preload.
Sliding joints – This is similar to the gapping joint except that the joint slides instead
of opening. The sliding joint chatters or impacts when the shear gaps close. The
symptoms are also similar to the gapping problem.

4.3 Conclusion

This section should provide the reader with some insight into the strengths and
weaknesses of transient testing on shakers. The insight should provide users of
transient software packages with a better understanding of their test results and give
them ideas for improving lessthan-ideal tests. The customers of dynamic testing
services can use the insight to understand that a poor test is not always due to the
test methods but can also be due to the less than ideal capriciousness of Mother
Nature. The methods assume linear systems; the shakers are usually almost linear;
but the test items are sometimes very nonlinear.
406 T. L. Paez et al.

5 Closure

Practically all real-world structures are subjected to random vibration and mechan-
ical shock environments during their design lives. The time domain and frequency
domain analyses in Sect. 2 illustrate methods for estimating structural responses to
applied dynamic loads. For either single- or multiple-degree-of-freedom systems,
the analytical approach uses differential equations to estimate response time
histories and the Fourier transforms of response time histories. Each mode of a
multiple degree of freedom structural response is governed by the equations for a
single-degree-of-freedom structure. In the field where a system is excited by forces
derived from transportation, aerodynamic or other environments, the analytical
methods provide a useful, robust means of estimating dynamic behavior based on
first principles.
From the experimental perspective outlined in Sect. 3, strains, displacements,
velocities, and accelerations are measured variables. The contrasting conditions
characteristic of field and laboratory tests are noted. Laboratory modal and vibration
tests measure dynamic responses under controlled conditions.
In field or laboratory, in analysis and experiment, time history responses, spectra,
and modes are basic features of the dynamics. For approximately linear systems,
exposed to non-destructive environments, the analytical and experimental results
are reasonably comparable, especially at frequencies corresponding to the lower
eigenvalues (modes). Four major sources of difference between analytical and
experimental results are:

1. The difficulty of properly defining the numerical functions governing damping

2. Problems specifying the actual characteristics of physical joints
3. The inevitable presence of some amount of nonlinearity
4. The statistical nature of the prosperities of nominally identical physical structures

The analytical model often provides an excellent approximation to the measured

physical response, but its utility is far reaching because it also provides a way of
thinking about and quantifying the basic principles underlying the often exceedingly
complex temporal responses of real systems.
From a testing standpoint, many vibration environments, whether field or labo-
ratory, deal with as noted in Sect. 3. Section 4 dealt with mechanical shock, another
common field environment. The same analysis concepts of spectra, frequency
response functions, modes, temporal responses, and frequency domain response
apply to both random vibration and mechanical shock. Analytically the approaches
to response estimation are quite similar. Experimentally, for shock tests a specific
response waveform is the desired response. For random tests, the autospectral
density characterizes the desired response. Of course, shock spectrum applies almost
exclusively to shock tests.
In general, the less averaging involved in the measure considered the more
problematic the comparison between analytical and experimental results. Autospec-
tral responses and lower frequency mode frequencies and shapes are often very
7 Random Vibration and Mechanical Shock 407

comparable. In contrast comparisons of temporal response are a best vaguely

similar.
Analytical first principles models and experimental test results are mirror aspects
of the same phenomena. Both address the forces and accompanying strains to
which real systems are exposed. The intelligent use of both is essential to full
understanding of structural dynamics.
New developments continue in both analysis and experiment. The very sub-
stantial increase in computational power in the past three decades means that first
principles models can now deal with finer physical features and can be applied
to a range of statistically similar systems. In the experimental world, new testing
procedures like multi-shaker and multi-axis tests do a better job of approximating
the real field environment. In both venues, there is a real need for new ways of
viewing nonlinear behavior and the statistical nature of real structures. The time is
ripe for the emergence of new measures to supplement the tried and true spectral
analysis, shock spectra, frequency response, and modal characterizations.

References
1. Murphy GM (2011) Ordinary differential equations and their solutions. Dover Publications,
Mineola
2. Inman DJ (2000) Engineering vibration, 2nd edn. Prentice Hall, Upper Saddle River
3. James JF (2011) A student’s guide to Fourier transforms: with applications in physics and
engineering, 3rd edn. Cambridge University Press, Cambridge, UK
4. Bendat JS, Piersol AG (2010) Random data, analysis and measurement procedures, fourth
edition. Wiley, New York
5. Abramowitz M, Stegun I (2014) Handbook of mathematical functions with formulas, graphs,
and mathematical tables. National Bureau of Standards, Applied Mathematics series 55. U.S.
Government Printing Office, Washington, DC
6. Bathe K-J (1996) Finite element procedures. Prentice Hall, Upper Saddle River
7. Wirsching PH, Paez TL, Ortiz K (1995) Random vibrations, theory and practice. Dover
Publications, Mineola
8. Einstein A (1905) On the movement of small particles suspended in a stationary liquid
demanded by the molecular kinetic theory of heat. Annalen der Pyhsik 17:549. Also, reprinted
in Einstein [9]
9. Einstein A (1956) Investigations on the theory of the Brownian movement. Dover Publications,
New York, Edited by R. Furth
10. Wiener N (1930) Generalized harmonic analysis. Acta Math 55:118
11. Crandall S (1958) Statistical properties of response to random vibration, Chapter 4. In: Crandall
S (ed) Random vibration. Technology Press of MIT and Wiley, New York
12. Crandall S (ed) (1958) Random vibration. Technology Press of MIT and Wiley, New York
13. Piersol AG, Paez TL (2010) Harris’ shock and vibration handbook, 6th edn. McGraw-Hill,
New York. (See Chapter 8)
14. Ang AH-S, Tang WH (2006) Probability concepts in engineering: emphasis on applications to
civil and environmental engineering, 2nd edn. Wiley, New York
15. Favour JD, LeBrun JD, Young JP (1969) Transient waveform control of electrodynamic test
equipment. In: Shock and vibration bulletin 40. Shock and Vibration Exchange, Arvonia
16. Yang RC (1970) Safeguard BDM system – development of a waveform synthesis technique,
SAG-64. Ralph M. Parsons, New York
408 T. L. Paez et al.

17. Smallwood DO (1975) Time history synthesis for shock testing on shakers. In: Seminar on
understanding digital control and analysis in vibration test systems, Part II. A Publication of
the S&V Information Center, Shock and Vibration Exchange, Arvonia, pp 23–42
18. Smallwood DO, Nord AR (1974) Matching shock spectra with sums of decaying sinusoids
compensated for shaker velocity and displacement limitations. In: Shock and vibration bulletin
44, Part 3. Shock and Vibration Information Center, Naval Research Laboratory, Washington,
DC, pp 43–56
19. Nelson DB, Prasthofer PH (1974) A case for damped oscillatory excitation as a natural
pyrotechnic shock simulation. In: Shock and vibration bulletin 44, Part 3. Shock and Vibration
Exchange, Arvonia, pp 57–71
20. Fisher DK, Posehn MR (1977) Digital control system for a multiple-actuator shaker. In: Shock
and vibration bulletin 47, Part 3. Shock and Vibration Exchange, Arvonia, pp 79–96
21. Stearns SD (1975) Digital signal analysis. Hayden Book Company, Rochelle Park
22. Rabiner LR, Gold B (1975) Theory and application of digital signal processing. Prentice-Hall,
Englewood Cliffs
DIC and Photogrammetry for Structural
Dynamic Analysis and High-Speed Testing 8
Christopher Niezrecki, Phillip L. Reu, Javad Baqersad,
and Daniel P. Rohe

Contents
1 Introduction and Relation to Prior Related Work on DIC and Point Tracking . . . . . . . . 411
1.1 Photogrammetry Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
1.2 DIC Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
1.3 DIC Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
1.4 Patterning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
1.5 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
1.6 Measurement and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
2 Overview of Modal Testing and Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
2.1 Frequency Response Function Measurement Considerations . . . . . . . . . . . . . . . 420
2.2 Fourier Transformation and Leakage Considerations . . . . . . . . . . . . . . . . . . . . . 422
2.3 Curve Fitting Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
3 The Distinction Between Operating Shapes and Mode Shapes . . . . . . . . . . . . . . . . . . . . 425
4 DIC Measurement Resolution in Relation to Structural Dynamic
Testing/Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427

Sandia National Laboratories is a multimission laboratory managed and operated by National

Technology and Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell
International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration
under contract DE-NA0003525.

C. Niezrecki ()
Department of Mechanical Engineering, University of Massachusetts Lowell, Lowell, MA, USA
e-mail: [email protected]
P. L. Reu · D. P. Rohe
Sandia National Laboratories, Albuquerque, NM, USA
e-mail: [email protected]; [email protected]
J. Baqersad
Department of Mechanical Engineering, Kettering University, Flint, MI, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 409

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_3
410 C. Niezrecki et al.

5 DIC Measurement Range in the Context of Structural Motion and Frequency . . . . . . . . 427
6 Analysis in the Temporal Versus Frequency Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
7 Identifying the Number of Images Needed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
7.1 Sampling Theory Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
7.2 Selecting Proper Sampling Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
7.3 Dealing with Long Sampling Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
8 Sources of Measurement Error and Best Practices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
8.1 Modes of the Stereo System Hardware and iIts Measurement Effect . . . . . . . . . 431
8.2 Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
8.3 Artificial Aliasing to Enhance Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
8.4 Lighting Requirements, Shutter Time, and Lens Adjustment . . . . . . . . . . . . . . . 436
9 Excitation Strategies for Modal Testing and Application to DIC Measurements . . . . . . 440
9.1 Impact Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
9.2 Shaker Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
9.3 Recommended Inputs for DIC Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
10 DIC and Photogrammetry Measurement Range and Noise Floor . . . . . . . . . . . . . . . . . . 444
10.1 Pre-measurement Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
10.2 Image Correlation and Data Processing Parameters . . . . . . . . . . . . . . . . . . . . . . . 448
11 Strain Mode Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
12 Projected Patterns Pros and Cons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
12.1 Projected Speckle Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
12.2 Deflectometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
13 Rotating Optical Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
13.1 Frequency of Measurement, Duration, and Shutter Time . . . . . . . . . . . . . . . . . . 452
13.2 Camera Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
13.3 Rigid Body Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
13.4 Mode Extraction Challenges and Effects of Harmonics . . . . . . . . . . . . . . . . . . . 454
14 Some Experimental Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
14.1 Comparison bBetween 3D Scanning Laser Doppler
Vibrometry and 3D Stereo-DIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
15 DIC Comparison to Traditional Modal Analysis Sensing . . . . . . . . . . . . . . . . . . . . . . . . . 458
16 DIC for High Rate Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
16.1 Definition of High-Rate Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
16.2 High-Rate Camera Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
16.3 2D Versus 3D Stereo-DIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
16.4 Environmental Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
16.5 Camera Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
16.6 Camera Protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
16.7 Extended Noise Floor Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
16.8 Camera Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
16.9 Lighting Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
16.10 Camera Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
16.11 Painting Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
16.12 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474

Abstract

This chapter provides an overview and some important considerations to be

made when making optical and stereophotogrammetry measurements on struc-
tures for dynamic applications. In particular, the chapter focuses on leveraging
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 411

those measurements to perform digital image correlation (DIC) to extract

dynamic parameters (e.g., strain, deflection, operating shapes, and mode shapes).
Structural dynamic testing and analysis in the context of performing optical
measurements is described. Information on optical high rate testing is also
presented along with lessons learned and best practices.

Keywords

Stereophotogrammetry · Digital image correlation · High-speed testing ·

Structural dynamics · Modal analysis

1 Introduction and Relation to Prior Related Work on DIC

and Point Tracking

Photogrammetry is a method of identifying geometry, displacement, and deforma-

tion of objects using photographs or digital images. This non-contacting measure-
ment approach was initially used aerially in order to create maps. The fundamental
approach in photogrammetry involves correlating an object’s dimensions to its
dimensions within a photograph (see Fig. 1). With advances in digital cameras and
computing systems, the accuracy of the photogrammetry technique has improved
significantly. These advances have allowed engineers and scientists to use close-
range photogrammetry for laboratory measurements with a very high degree of
precision. Recently, photogrammetry has received particular attention due to its
nonintrusive, noncontact measurement capabilities and its robustness for in situ
measurements. This technique provides accurate full-field geometrical measurement

Fig. 1 The fundamentals of photogrammetry showing how the sensor size and pixel size can be
correlated to the field of view (FOV) and object sample size, respectively
412 C. Niezrecki et al.

Camera
distance
Left pinhole aperture Right pinhole aperture
or perspective center Image points or perspective center

Left image Right image

plane plane

Object point

Object surface

Fig. 2 A schematic of the stereophotogrammetry technique showing the two cameras and how the
location of a point is identified using stereo-triangulation [27]

results as well as graphically meaningful images and videos of structural shape,

motion, deformation, and strain.
The background of the photogrammetry technique goes back to Leonardo da
Vinci who studied the perspective concept [15]. A single camera can measure the
displacement of objects in a two-dimensional (2D) planar motion. However, three-
dimensional (3D) measurements are performed using stereo or multiple cameras.
In the 1900s, Pulfrich and Fourcade were among the pioneers who studied the
analytical aspects of the stereovision for 3D measurement [15]. A triangulation
technique using a ray-tracing process is used to determine the coordinates of the
features in the test structure (see Fig. 2). In the 1980s, Peters and Ranson developed
a technique to use the photogrammetry technique to measure the deformations of
objects. They proposed an approach to measure the 2D deformation of a structure
by measuring the displacement of points on the surface of a structure [52]. This
new technique was later expanded to measure full-field strain on structures and was
called digital image correlation (DIC) [38]. Using this technique, a single camera
can measure the displacement of objects in a 2D plane. The DIC approach has been
widely used in solid mechanics to measure semi-static deformations. With current
advances in camera technology, cameras are able to record high-speed phenomena.
Thus, researchers have started to use this technology to measure vibrations and
transient phenomena in structures [7].

1.1 Photogrammetry Techniques

Point-tracking, DIC, and target-less approaches are three common photogrammetry

techniques used in structural dynamics.
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 413

Fig. 3 An image showing optical targets mounted to a single wind turbine blade; the optical targets
are identified by using a point tracking technique [16]

1.1.1 Point Tracking

The point tracking technique uses cameras to identify discrete points mounted
to a structure. For this method, a series of optical targets (usually high contrast
circular points) are mounted to the structure. Depending on the type of target
finding algorithms, software programs may suggest using a specific type of target
for optimal measurement. For circular optical targets, the centers of the targets
are found using an ellipse finding algorithm, and the coordinates of the points
in space are identified using a triangulation technique. 3D point tracking (3DPT)
determines the displacements of the targets by tracking the optical targets at different
time stages and comparing the coordinates to the reference stage (usually the first
stage). Therefore, only the displacement of the targets at a handful of points can be
measured by using the 3DPT method (see Fig. 3). The size of these optical targets
is usually greater than eight pixels as a rule of thumb.

1.1.2 Digital Image Correlation (DIC)

DIC is another technique for image-based optical measurement. This technique
works based on detecting gray-scale variations of a contiguous pattern. A high-
contrast speckle pattern with dot sizes of 3–7 pixels is applied to the surface of
the structure (the speckle size can be made much bigger for modal testing when
global mode shapes of the structure are desirable and local deformations are not of
interest); for more information interested readers are referred to the “iDIC Good
Practice Guide” [28]. After surface preparation, a series of pictures is recorded
from the object under deformation. In every digital picture, a specific gray-level
is assigned to each pixel. Tracking a single pixel in several digital photographs
recorded during the deformation is impossible because many pixels might have a
similar gray-level. However, a neighborhood of pixels (i.e., subset or facet) can
create a unique light intensity variation that can be used to track a point on the
surface of a structure. The red rectangles in Fig. 4a show the location of the subsets.
414 C. Niezrecki et al.

Fig. 4 (a) An image showing overlapping subsets (facets) and the subset size; (b) and (c) images
showing how DIC tracks the subsets after deformation; (d) a photo showing the patterned area of
a wind turbine blade; with the computed strain of the blade overlaid on the patterned blade surface

To be effective, a subset must contain several gray-level variations and its length
and width is usually five times larger than speckle size (15–25 pixels for speckle
size of 3–5 pixels). The distance between these facets is called step size. The step is
usually selected as 3/4 of the subset size but can have a wide range of values. DIC
achieves spatially distributed measurement by creating an array of overlapping (or
nonoverlapping) subsets over the total area of interest and correlating each subset to
the subset in the undeformed structure. In the photo taken from the deformed model,
a subset with the highest values of similarity to the reference subset is set as the new
deformed subset (see Fig. 4b, c).
Correlating the deformed subset to the undeformed subset can be performed
using mathematical correlation functions. Equation 1 shows a sample 2D correlation
function used to identify the possible matches of subsets in the deformed photos to
a reference subset in the original photos using least square minimization [90].


n/2
 2
C (x, y, u, v) = I (x + i, y + j ) − I ∗ (x + u + i, y + v + j ) (1)
i,j =−n/2

In this equation, C is the correlation function, which is a function of the reference

pixel coordinate (x,y) and displacement (u, v). I and I* are related to the image before
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 415

and after motion, respectively. The differences between pixels in a subset after the
motion and before the motion are summed on the entire subset (n = subset size).
The minimum value of the correlation function (C) shows that a correct deformed
subset has been identified and can be tracked. It should be noted that Eq. 1 is a
sample correlation equation and DIC packages might use other correlation functions
to identify the optimum correlation values.
To perform a DIC analysis, usually an initial start point is manually defined in
the reference image (or images for a stereo system). Afterwards, the remainder
of the field of view (FOV) that is visible to both cameras is covered by an array
of subsets. After finding the deformed subsets, the displacement of each subset
is calculated. The full-field strain can be computed using continuum mechanics
equations (strain = ∇Displacement). This calculation is usually performed using
triangular meshes and applying linear or spline strain computations. It should be
noted that DIC can only identify the strain in the tangential plane of the surface
unless volumetric DIC is performed (e.g., Digital Volume Correlation using CAT or
MRI scans).

1.1.3 Target-Less Approaches

Computer vision has enabled researchers to identify deformations of an object using
features within the object or at the edges of the object. This approach does not need
patterning or mounting of optical targets. This is an emerging approach and has been
used in structural dynamics. In this approach, edge detection, pattern matching, and
blob detection algorithms are used to identify the areas that may be tracked. This
technique has been used to measure deformations of a traffic signal structure [9],
a cylindrical structure [73], and a solar array mounted to a space station [24] in
addition to other applications. It should be noted that the non-target approaches have
approximate resolutions of fractions of pixels (e.g., edge detection = 0.5 pixels and
circular fiducials = 0.02 pixels) while DIC has been shown to have better resolution
(DIC patterns for in-plane measurements = 0.01 pixels) [81, 85].

1.2 DIC Hardware

A photogrammetry system consists of several components (1) camera (s) to record

photos, (2) camera tripods, camera bar or stands to hold the cameras in fixed
positions, (3) a controller to start the measurement and to synchronize the cameras,
and (4) a computer to possibly record and process the data.
Charged coupled device (CCD) digital cameras, CMOS cameras, and infrared
cameras have been used to record images. For a 3D measurement, at least two
cameras are required. The pair of cameras needs to be held in fixed positions with
respect to each other during the measurement. The relative displacement of the
cameras during a measurement in laboratory conditions should ideally be less than
sub-micron [54]. Thus, strong stands with high stiffness are typically used. When
high-magnification lenses are used, vibration isolation is required to prevent image
416 C. Niezrecki et al.

blurring. Synchronizing the cameras can be performed using a control box. Current
high-speed cameras are able to synchronize using their built-in control system.

1.3 DIC Software

Currently, many software packages for DIC and point tracking are available. Most
of the commercial software programs have been integrated with their associated
cameras and are capable of controlling the cameras and calibrating them. They
are also equipped with post-processing features such as rigid-body correction,
coordinate transformation, and result presentation. Nowadays, 2D DIC versions of
software packages are available at no charge. A list of some of the software packages
used for photogrammetry along with their capabilities is shown in Table 1.

1.4 Patterning

For a DIC measurement, a random or nonuniform speckle pattern needs to be

applied to the surface of the test object. This pattern should deform with the object
during testing without affecting the mechanical integrity of the host structure. Thus,
special attention must be paid to choose a flexible pattern for the test cases in which
large deformation occurs. The pattern must also be thin enough to prevent adding
stiffness or mass to the test object. The paint for the pattern should be matte (flat
or non-glossy) in order to prevent reflection; using a reflective background for the
pattern would reduce the accuracy of the measurement. It should be noted that using
a dark background and white dots or the opposite would lead to similar accuracy
level in the results. More information about patterning in DIC can be found in Reu
[61, 62] and IDIC [28] and is discussed further in Sect. 16.

1.5 Calibration

A calibration process for the photogrammetric cameras needs to be performed

before each measurement series. The objective of the calibration process is to
identify the intrinsic (focal length, image center, distortion, skew) and extrinsic
(camera pair’s relative position) parameters, and the scale of the units. Prior to
calibration, a measurement volume is determined. This volume should be large
enough to fit the object before and during motion or deformation. Care should be
taken to ensure that the test structure fills as much of the FOV as possible, so that
the entire sensor resolution can be effectively used to record the area of interest.
Camera calibration is usually performed using bundle adjustment [84], which is a
robust process commonly used in commercial DIC packages. In this calibration pro-
cess, a series of photos are taken of a pattern or an array of optical targets (fiducials)
applied to a rigid calibration object that is placed in different positions/orientations.
There is no need for a knowledge of the calibration object, but the distance between
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 417

Table 1 An example of some DIC and point tracking software packages as of 2018a
Software Country
name Company of origin Capabilities Website
Aramis/ GOM Germany 2D and 3D DIC, www.gom.com
Pontos 3DPT, FFT
calculator
Correli LMT Cachan France 2D and 3D DIC, FFT www.correli-stc.com
and Airbus calculator
DICe Sandia USA 2D DIC https://dice.sandia.gov
National Lab
DPA AICON Germany 3DPT http://aicon3d.com/
Limess Limess Germany 2D and 3D DIC www.limess.com
MatchID MatchID Belgium 2D and 3D DIC http://matchidmbc.be
The Catholic USA 2D and 3D DIC www.opticist.org
University of
America
Ncorr Georgia USA 2D DIC www.ncorr.com
Institute of
Technology
Q-400 Dantec Germany 2D and 3D DIC, FFT www.dantecdynamics.com
Dynamics calculator, operating
shape extraction
(using Scilab)
Strain LaVision Germany 2D and 3D DIC www.lavision.de/en
Master
TEMA Imagesystems Sweden 3D PT, 2D and 3D www.imagesystems.se
Motion DIC, and FFT
/TEMA calculator
DIC
VIC-3D Correlated USA 2D and 3D DIC, FFT www.correlatedsolutions.
Solutions calculator, operating com
shape extraction
ProAnalyst Xcitex USA 2D and 3D point http://www.xcitex.com
tracking
a Note:The authors attempted to prepare a comprehensive list, but some vendors may have been
unknowingly omitted and over time it is likely that additional software packages will become
available

the reference points must not change during the calibration. This pattern can include
coded targets, uncoded targets, or a combination of each. The optical targets are
typically recognized by the two cameras using an ellipse finder algorithm. The
location of the targets is identified in the software using a triangulation algorithm.
A minimum of two scale bars (measurements between two points in the calibration
pattern) is needed to be introduced to the software to finalize the calibration. The
calibrating parameters are extracted using a least-square minimization approach.
The calibration results from the DIC software packages are usually a residual from
the calibration (a comparison between what the extracted points are and what the
418 C. Niezrecki et al.

model shows after minimization) for the measurement volume. The test structure
must remain in this measurement volume in order to extract accurate results.
After calibration, a DIC system can stay calibrated as long as the camera’s
relative position is unchanged even if both cameras are moved. The parameters that
can cause a loss in calibration include: movement of the cameras with respect to
each other, zooming or refocusing of the lenses, and changes in temperature. The
epipolar line status in calibration can be examined with a correlation check using
static photos. In this pretest, a handful of stereo-photos are taken from the test object.
An arbitrary point in the first image is selected. The ability of the camera system to
locate the corresponding point in the second image of the stereo system shows how
accurate the camera calibration is. If this epipolar constraint is violated, the software
should give a warning or an error to show that the measurement is not accurate. The
noise floor of the measurement also can be monitored using these photos. If all the
parameters stay fixed, it has been shown that the system can maintain the calibration
for days. It should be noted that when camera calibration can be readily performed,
the system should be re-calibrated before each measurement.
Commercial calibration objects for photogrammetry are usually made in the
shape of either panels or crosses (see Fig. 5). These objects are able to calibrate
FOVs in limited ranges. However, in many experiments, the desired FOV that
needs to be calibrated exceeds the size of available calibration objects. Extended
calibration and large-area calibration are two approaches to calibrate large fields of
view.
One example of an extended calibration process uses a standard calibration
object to calibrate a FOV larger than the calibration object. The normal calibration
procedure involves recording a series of photos from the calibration panel in
different orientations and at different distances away from the cameras. The
extended calibration is performed after performing a normal calibration. A series
of extra pictures is taken from the calibration panel placed close to the edges of the
larger FOV in order to cover the whole FOV. Taking photos on the edges of the

Fig. 5 A sample calibration panel (left) and a calibration cross (right)

8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 419

extended FOV helps to accurately identify the distortion of the lenses near the edges
and corners of the FOV [2].
Another approach to calibrate large areas uses large area calibration. In this
approach, there is no need for a calibration panel or object that is the same scale
as the FOV. Instead, a series of coded targets and uncoded targets are mounted to a
fixed structure [53]. Several photos are recorded from the targets with the cameras
located in different locations and orientations. The final photo is taken when the
cameras are fixed in the final position. Some software packages use a DIC speckle
pattern for calibration and establish a scale bar between two points. At least two
scale bars (the displacement of two targets measured by a tape measure or another
device) are introduced to the software. For example, for a measurement on a rotor
with large rotations and deformations, distances between two targets along a blade
can be assumed to be fixed (in the non-rotating state) and used to calibrate the
cameras and monitor the calibration status.

1.6 Measurement and Applications

After the cameras are calibrated, the measurement can be triggered using a control
unit. The DIC system can be synchronized with other machines that provide a trigger
signal (e.g., a mechanical testing machine or stroboscope). The reference image is
usually taken of the undeformed object or at some initial point in time. However, if
the undeformed shape of the object is not available, a photograph of the deformed
structure might be used as the reference. The other photographs are compared to the
reference to extract relative displacement, deformation, or strain.
The photogrammetry system has wide areas of applications including: material
testing, finite-element model validation, strain computation, component testing,
bioengineering, structural health monitoring, crash analysis, metal forming, and
vibration measurement. The applications vary from nanometers in scanning electron
microscopes [79] to hundreds of meters as in the case of wind turbines [49].

2 Overview of Modal Testing and Requirements

An important aspect of experimental modal analysis is to derive the modal

parameters of the test article (i.e., natural frequencies, damping ratios, and mode
shapes) from the measured test data. These parameters are intrinsic properties of
the test article and its boundary conditions, and as such they are often of interest for
model validation and updating. If an analyst can match natural frequencies and mode
shapes between their model and experiment, a good deal of confidence is gained in
that model. Additionally, in many situations the damping of a system cannot be
computed analytically, so the analyst may rely on experimental estimates of the
damping ratio for their model. In addition to model validation, modal parameters
derived from experiments can be useful for other situations such as instrumentation
placement [33], force reconstruction [10], or experimental substructuring [86].
420 C. Niezrecki et al.

Modal tests are often designed to approximate a free boundary condition for the
part, using soft bungee cords or foam to support the test article. This facilitates
easier comparison with an analytical model of the test article, because it is often
easier to quantify and model the effect of a few soft springs on the dynamics of the
test article than a potentially complicated interface to the ground, which may require
tuning bolt stiffness, contact patches, and other parameters.
To perform a traditional modal test, the test article is excited, and the responses
to the excitation are measured. Historically, accelerometers mounted to the test
article’s surface have been used to measure the responses, although these have
the disadvantage of adding mass to the test article that can alter its dynamics.
More recently, methods such as laser Doppler vibrometry [17] and 3D DIC or 3D
point tracking [27, 45, 88] have enabled noncontact, full-field measurement of the
response of the test article. The considerations listed in this chapter will focus on the
capabilities and important considerations when modal testing is performed using a
DIC approach.
The most common types of excitation used for experimental modal analysis
are impact testing using an instrumented hammer or shaker testing using an
electrodynamic shaker attached to a force transducer. Alternatively, operational or
output-only modal analysis is a testing strategy that attempts to identify modal
parameters of the system without a measured excitation force. Here the system may
be excited by its natural operating conditions (e.g., a wind turbine being excited by
the wind), or a more novel excitation method such as focused or unfocused acoustic
pressure and magnetic excitation [29]. For the most accurate results, however, the
forces used to excite the structure must be measured by a calibrated force transducer,
and that force input should be used to compute the frequency response functions
from which the modal parameters can be estimated.

2.1 Frequency Response Function Measurement Considerations

In analytical modal analysis, the natural frequencies and mode shapes can be found
from the eigensolution of the general eigenvalue problem formed by the mass
and stiffness matrices of the structure. For experimental modal analysis, no such
matrices exist a priori. The modal parameters will, therefore, be estimated from the
test data.
For experimental modal analysis the frequency response function (FRF) H(ω),
which relates the input force spectra F(ω) to the output displacement spectra X(ω)
(or velocity or acceleration) at each frequency line ω, is of key importance. For
standard modal test hardware and software packages, FRF creation is generally
integrated into the software and performed automatically for the test engineer based
on specified parameters. However, since DIC is a relatively new technology in
structural dynamics testing, FRF computation capabilities may not be included
in the software, and the test engineer may need to create FRFs externally. The
structural dynamic input-output relation is governed by:
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 421

X (ω) = H (ω) F (ω)

Through modal substitution, it can be shown that the FRF values depend on the
modal parameters of the system,

N N
modes
{}k {}Tk modes
Ak A∗k
H (ω) = = +
ωk2 + 2iζk ωk ω − ω2 iω − λk iω − λ∗k
k=1 k=1

where ωk , ζ k , and {}k are the natural frequency, damping ratio, and mass-
normalized mode shape vector for the mode k, and λk and Ak are the system poles
and residues, respectively. The asterisk operator denotes the complex conjugate.
Therefore, if the FRF matrix H is known, modal parameters can be fit to that data
using various strategies.
Because we cannot measure the frequency response functions directly, they
must be estimated from the time domain data measured by the data acquisition
system. For DIC measurements, this is often a displacement time history at each
measurement point supplemented with one or more force time histories measuring
the excitation applied to the test article. The time domain data is then transformed
into frequency domain data via the Fourier Transform, and cross power spectra
GXFjk and GFXjk and auto power spectra GFFjj and GXXjj can be computed:

Navg Navg
 
GXFj k = Xj Fk∗ GFXj k = Fj Xk∗
l=1 l=1

Navg Navg
 
GFFjj = Fj Fj∗ GXXjj = Xj Xj∗
l=1 l=1

To reduce the noise in the computed power spectra, several sets of data are
typically measured, and the resulting spectra are averaged. This averaging smooths
out some of the random errors inherent in any measured signal. The frequency
response functions are then computed from the power spectra. Several formulations
are available depending on where the noise is expected in the measurement.

Estimator Formulation Reduces noise on:

GXFij
H1 Hij = GF Fjj Outputs/responses
GXXii
H2 Hij = GF Xj i Inputs/forces

Note that for multiple-input, multiple-output (MIMO) cases, GFX, and GFF will
be matrices that will need to be inverted. This limits the H2 estimator because it
422 C. Niezrecki et al.

requires that GFX be square, meaning the number of inputs must equal the number
of outputs. Many other FRF estimators exist in the literature, for example, the
HV [41], Hc [42], and Hs [89] estimators, and each have their own strengths and
weaknesses.

2.2 Fourier Transformation and Leakage Considerations

The signals measured with DIC or other data acquisition systems will be discretely
sampled and have finite length. These signals are then transformed into the
frequency domain via a discrete Fourier transform (DFT). Due to the nature of the
DFT, the measured signals must either be fully contained within the measurement
window (i.e., go to zero at the beginning and end of the measurement frame) or
consist of only signals that are periodic in the measurement frame length.
If the signal is not periodic or the transient is not completely captured within
the measurement time frame, a bias error commonly referred to as leakage may
occur, resulting in the frequency content of a particular sinusoid being smeared
across many frequency bins in the DFT. Figures 6 and 7 show two examples
of this phenomenon. Note that the frequency resolution in this example is quite
low which exacerbates the effects of leakage; this was done to more clearly
illustrate the effects. Effects of leakage may not be as severe in a typical appli-
cation.
Leakage can generally be reduced by increasing the frequency resolution of
the FFT (which effectively lengthens the measurement frame), and may also be
improved with averaging. However, with DIC measurements, the camera memory

Measurement Frame Repeated Frame

4 3

2
2
1
Signal

Signal

0
0
-1

-2 -2
0 0.02 0.04 0.06 0.08 0.1 0 0.05 0.1 0.15 0.2 0.25 0.3
Time Time
0
Discrete Fourier Transformation
10
|FFT|

-10
10

-20
10
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)

Fig. 6 Signal made from unit-amplitude sine waves at 30, 80, and 110 Hz with random phases.
The signal is periodic in the measurement frame, and no leakage occurs
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 423

Measurement Frame Repeated Frame

4 4

2 2
Signal

Signal
0 0

-2 -2

-4 -4
0 0.02 0.04 0.06 0.08 0.1 0 0.05 0.1 0.15 0.2 0.25 0.3
Time Time
0
Discrete Fourier Transformation
10

-2
10
|FFT|

-4
10

-6
10
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)

Fig. 7 Signal made from unit-amplitude sine waves at 28.1, 86.9, and 112.5 Hz with random
phases. The signal is not periodic in the measurement frame, and significant leakage is seen in the
frequency domain spectra

may be limited; thus simply measuring for a longer duration may not be an option.
One common method to reduce the effects of leakage (but not eliminate it entirely) is
to apply a window or weighting function to the time data before computing spectra.
This can improve the spectra quality; however, we note that applying a window does
affect the results. The window function should be viewed as a last resort when other
methods to reduce leakage such as increasing the frequency resolution, averaging,
or applying a more periodic excitation are not feasible. A common window for
random vibration testing is the Hann or Hanning window. Figure 8 shows the
results of applying a Hann window (red curve) to the time signal, and the resulting
improvement in the estimation of the amplitudes of the sine waves.
When to use a window is sometimes obvious, for example, when shaker-testing
using pure random signal, as there is no guarantee that the signal will be periodic
within the measurement frame. Other times, for example when impact testing, it
may not be obvious if the response to the impact will decay within the measurement
frame. Previewing the data is then advantageous to ensure that the transient is
entirely contained in the measurement frame. However, DIC does not provide
a practical way to preview the data to iterate on measurement parameters such
as exponential window properties or frame length due to the time required to
download images from the camera and then process them into time histories. We
suggest supplementing the DIC measurement with another measurement technique,
an accelerometer or single point laser Doppler vibrometer, from which data can be
compared and evaluated more quickly.
424 C. Niezrecki et al.

Measurement Frame Repeated Frame

4 4

2 2
Signal

Signal
0 0

-2 -2

-4 -4
0 0.02 0.04 0.06 0.08 0.1 0 0.05 0.1 0.15 0.2 0.25 0.3
Time Time
0
Discrete Fourier Transformation
10

-2
10
|FFT|

-4
10

-6
10
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)

Fig. 8 Signal from Fig. 7 with Hann window applied. The Hann window forces the signal to zero
at the start and end of the measurement frame, so that it can be repeated without discontinuity. The
estimates of the sine wave amplitudes are significantly improved, but the estimated frequency of
each of the sine waves become less precise

2.3 Curve Fitting Considerations

With the FRFs created, modal parameters can be estimated and fit to the data.
There are a number of different strategies and algorithms used (for more detail,
interested readers are encouraged to refer to  Chap. 11, “Experimental Modal
Parameter Evaluation Methods”). When fitting modes to the data, it is important to
understand what data the mode fitting software is expecting. For example, because
of the popularity of accelerometers for measuring modal response, many software
packages expect Receptance or Accelerance FRFs which relate the acceleration at
response points to forces at the inputs. Because the output from DIC measurements
is not acceleration but displacement, the data may need to be converted to
acceleration frequency response functions if required by the curve fitting software.
This differentiation can easily be performed in the frequency domain simply by
multiplying all displacement FRFs by −ω2 at each frequency line, where ω is the
angular frequency (in radians per second) of that frequency line.
An additional consideration to make when curve fitting DIC data is that in general
many more degrees of freedom (DoF) (or data points) are measured during a DIC
test than during a standard modal test using accelerometers. Software packages used
to analyze standard modal test data seldom need to handle more than a few hundred
degrees of freedom, while high-resolution DIC may measure 1000 points or more
during a given test. Code that runs sufficiently fast for standard testing conditions
may not be optimized for the large volume of data supplied by DIC, so analysis time
may suffer, especially if any iteration on the curve fitter settings is necessary. The
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 425

DIC data may need to be decimated or spatially averaged and down sampled to a
manageable size to improve turnaround time in this case.

3 The Distinction Between Operating Shapes and Mode

Shapes

In modal analysis, the common parameters of interest include natural frequencies,

damping ratios, and mode shapes. The mode shapes are very valuable because they
form a “basis” for the deflections that the part undergoes and can be visualized and
animated to create an intuitive picture of how the structure vibrates. Closely related,
but not the same, are operating deflection shapes, often called “operating shapes” or
“deflection shapes.” These shapes are more general in that they are any shape that a
structure may take as it vibrates during an arbitrary excitation (e.g., single frequency,
multiple frequencies, or broadband). Operating shapes are often formed from a row
or column of the FRF matrix corresponding to a given frequency line, and are often
animated by applying the magnitude and phase of the FRF to a deflection of the
corresponding node in a display model.
Confusion often arises between the two terms because operating shapes can
appear very similar to mode shapes, especially at frequency lines in the vicinity of a
mode if the modes are well-spaced in frequency. This often occurs when performing
a rudimentary form of curve-fitting called peak-picking, where the analyst estimates
the natural frequency and mode shape from an FRF at a single frequency line that
is a local maximum in the FRF curve. Though the shape values extracted from the
FRFs may approximate a mode shape if the modes are spaced appropriately, the
shape at any given frequency line is actually a summation of many mode shapes
(though it is possible that a single mode dominates the summation).
We illustrate this point with an example. A free mass-spring system with four
degrees of freedom has natural frequencies, damping ratios, and mass-normalized
mode shapes shown in Tables 2 and 3. Figure 9 shows the FRFs of this system
from an input at DoF 4. From Fig. 9 it can be seen that the two modes are closely
spaced and the third is fairly far from the other two. Drawn on the figure are each
mode’s contribution to the FRFs. Table 4 compares the actual values of the mode
shapes used to create the FRFs in Fig. 9 to the deflection shape constructed from the
imaginary part of the FRFs at the peak. These values have been normalized to unit
length.

Table 2 Natural frequencies and damping ratios for the example problem
Mode Natural frequency (Hz) Damping ratio (%)
Rigid body (RB) mode 0 0
1 14.6 1.5
2 15.3 1.5
3 19.7 1.5
426 C. Niezrecki et al.

Table 3 Mass-normalized RB mode Mode 1 Mode 2 Mode 3

mode shapes for the example
DoF 1 −0.4264 0.6559 0.5222 0.3395
problem
DoF 2 −0.4264 0.1893 −0.5222 −0.0980
DoF 3 −0.4264 −0.6187 0.0000 0.6598
DoF 4 −0.4264 −0.2772 0.3482 −0.5355
Phas e

Phas e
Magnitude DoF 1

Magnitude DoF 2
Phas e

Phas e
Magnitude DoF 3

Magnitude DoF 4

Summation
Mode 1
Mode 2
Mode 3

Fig. 9 FRF of acceleration with respect to force showing three modes with each mode’s
contribution to the FRF superposed on the figure

Table 4 Mode shapes versus peak-picked deflection shapes normalized to unit length
Mode 1 Mode 2 Mode 3
Mode shape Defl. shape Mode shape Defl. shape Mode shape Defl. shape
DoF #1 0.6817 0.6258 0.6396 −0.5782 0.3689 0.3686
DoF #2 0.1968 0.2663 −0.6396 0.6681 −0.1065 −0.1045
DoF #3 −0.6431 −0.65 0 −0.0615 0.7169 0.7159
DoF #4 −0.2881 −0.339 0.4264 −0.4643 −0.5819 −0.5836

Notice that the deflection shape near the peak corresponding to Mode 3 is
actually a very good approximation for the mode shape. This is because Mode 3 is
sufficiently far away from other modes so that the contributions from those modes
are several orders of magnitude lower than the contributions from Mode 3. Similarly,
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 427

it can be seen that the operating deflection shapes from the closely spaced modes do
not approximate the mode shapes as well because the tails of neighboring modes are
still a significant fraction of the peak of the mode of interest. Note that for all three
modes, the deflection shapes would not be scaled properly, and therefore would not
be able to be used in computations such as substructuring or modal effective mass;
however, they may still be adequate for providing a qualitative understanding of the
mode shapes.

4 DIC Measurement Resolution in Relation to Structural

Dynamic Testing/Modal Analysis

Prior to embarking on a measurement campaign or test using DIC or point tracking

for modal analysis or structural dynamic testing, it is important to consider the
measurement resolution of the sensors being used. As with any sensor, the minimum
measurable quantity is limited by the physics of the sensing element and the
associated external influences that contribute to additive noise. For a given set of
cameras, one of the most important considerations that influences the resolution of
the measurement is the FOV. As the FOV increases, the measurement resolution
will decrease proportionally. Likewise, DIC is a displacement measurement and a
practitioner needs to be careful that the minimum structural motion of interest is
sufficiently large to obtain a useful measurement. This topic is discussed in detail
in Sects. 5 and 6. Lastly, there are many factors that influence and contribute to
noise in the measurement (e.g., calibration, atmospheric turbulence, lighting and
glare, patterning, post-processing choices). Many factors that influence noise are
presented in Sect. 10.

5 DIC Measurement Range in the Context of Structural

Motion and Frequency

For typical structural dynamic measurement, accelerometers are used to measure

acceleration and laser vibrometers are used to measure velocity. However, for DIC
or point tracking measurements made in conjunction with stereo-photogrammetry,
the measurand is displacement. This presents some benefits as well as challenges
in order to extract a useful measurement. The most common structural dynamic
measurement that most engineers are familiar with utilize accelerometers and the
harmonic acceleration, velocity, and displacement can be respectively expressed as:

a(t) = Aeiωt

1
v(t) = Aeiωt
iω
428 C. Niezrecki et al.

−1 iωt
x(t) = Ae
ω2

where A, ω, and t represent the acceleration amplitude, angular frequency, and

the time, respectively. Compared to the acceleration amplitude, the displacement
amplitude is decreased to the second power as the frequency is increased. This
effect results in DIC being less effective for measuring a structure’s vibration
at higher frequencies because the motion of the structure may fall below the
noise floor of the measurement. Conversely, as the frequency is reduced (e.g.,
<∼50 Hz) accelerometer measurements can have diminished sensitivity and laser
vibrometers are poor at measuring low-frequency vibration especially when the
amplitude of vibration is very large (e.g., > a few centimeters). Therefore, for
structural dynamic measurement compared to other measurement approaches, DIC
performs best in applications when the operating frequencies are relatively low,
and the displacements are large. However, the useful frequency range is very much
dependent on the size of the FOV, the camera resolution (i.e., pixel count), and the
camera sample rate. As the FOV decreases, the DIC measurement resolution will
improve extending the measurable frequency range.

6 Analysis in the Temporal Versus Frequency Domains

The smallest time-domain displacement that a DIC system can resolve is approxi-
mately 0.01 pixels [81] at the detector, meaning the smallest physical displacement
measurable is dependent on the FOV and the camera resolution. Given the typically
low level of vibration used when modal testing (especially at high frequencies),
it may seem that DIC measurement techniques may not be adequate to capture
the corresponding responses. However, preliminary results indicate that processing
results in the frequency domain reveal that inherent measurement noise is reduced
compared to time domain processing of data. This is likely due to the displacement
noise being spread over multiple frequency bins, but the reason is still under
investigation by the research community.
Figure 10 shows an example measurement taken where due to the FOV of
the cameras, 0.01 pixels corresponds to approximately 2 μm of out-of-plane
displacement [68]. However, it can be seen that the frequency domain noise floor
is significantly lower than the time-domain noise floor (approximately 10 nm) and
modes of the system are still visible in this spectrum.

7 Identifying the Number of Images Needed

The determination of the number of images/samples needed for a particular modal

measurement is a function of the measurement bandwidth and number of frequency
lines required, as well as the number of averages required to reduce the noise in the
measured data. For DIC, there can be a great deal of overhead involved in processing
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 429

Fig. 10 Example test data showing the typical noise floor for a time-domain DIC measurement,
as well as acquired data displayed in the frequency domain [68]. Note the extremely low strain
noise floor. 0.02 με

the images recorded in a test, and the memory of the camera may be limited to a
certain number of images (see Fig. 30 in Sect. 16). Therefore, it is often important
to optimize the test to minimize the number of images required while still achieving
good data.

7.1 Sampling Theory Relationships

Sampling theory defines a number of relationships between time-domain parameters

of the measurements such as sample time T, sampling frequency fsample , time
between samples t, and number of samples N, and frequency domain parameters
such as maximum frequency Fmax of interest, frequency bin spacing f :

1 1
fsample = = 2 × Fmax T =
t f

N
fsample = = N f
T

We note that many data acquisition systems sample at a rate higher than 2 × Fmax
in order to provide additional bandwidth for anti-aliasing filters to help reduce the
contamination caused by higher frequency content. For DIC, it may be advantageous
to sample at a higher rate as well: though DIC generally does not have anti-aliasing
filters, displacement magnitudes do tend to decrease naturally as the frequency
increases [11], and this extra bandwidth may allow any aliased frequency content
430 C. Niezrecki et al.

to negligibly influence the true spectrum. If the signal is adequately oversampled, it

may be possible to use a digital anti-aliasing filter, though oversampling will result
in more images to download and process. It is therefore better to control the input
such that there is no frequency content above the bandwidth of interest (see Sect.
8.2 for additional details).

7.2 Selecting Proper Sampling Parameters

Selecting the sampling parameters of the system may be an iterative process. For
example, the damping of a lightly damped system might drive the measurement
time required to capture the entire transient response to an impact excitation, but the
damping will generally not be known a priori, so the initial measurement may not be
long enough. Due to the processing overhead associated with DIC measurements, it
may be advantageous to set up the sampling parameters using another measurement
technique such as an accelerometer or single point laser Doppler vibrometer. This
can also help the user prevent aliasing. Shortening the time spent on each iteration
will allow the test engineer to more rapidly identify the appropriate sampling
parameters.
Many different properties of the test may determine what sampling parameters
should be used. Damping is perhaps the property of a test article that is least
understood, but it is often the most critical for setting up sampling parameters. If
a mode is lightly damped, the corresponding peak in an FRF will be very sharp, and
it may require closely spaced frequency bins to adequately resolve the peak (f will
be small). The equations in Sect. 7.1 indicate that the measurement time T will grow
large, and if a certain bandwidth is required for the test, the number of samples N
will correspondingly grow large.
Similarly, in the time domain, especially for impact testing, the damping directly
affects the time it takes for the responses of the part to decay. If the decay time is
long, the test engineer may increase the testing time T, and for a given frequency
bandwidth this will involve increasing the number of samples N. Alternatively, the
test engineer may utilize a window function (e.g., exponential window) to help damp
out the response by the end of the measurement.
Another test parameter that can significantly influence the number of images
required is the number of averages. Averaging is utilized when constructing FRFs
from time domain data. By averaging data, noise is reduced in the measured transfer
functions at the expense of longer testing times and more samples. Because of the
overhead involved with processing DIC measurements, it is advantageous to utilize
an excitation technique that does not require a large number of averages.

7.3 Dealing with Long Sampling Requirements

While high-speed cameras record very quickly to their internal memory, transfer off
of the camera to a computer for processing can take a significant amount of time.
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 431

This can create a severe testing bottleneck if long-running tests are recorded. For a
traditional modal test, the continuity of the test is often not a significant concern,
and while tedious, pausing the testing every handful of averages to download the
camera memory onto an external data storage and processing system will not prevent
a successful test. However, one must use care to maintain the boundary conditions
of the test, which are more likely to change over a long-running test. For example,
bungee cords that support the part in a free-free boundary condition may sag (due
to stress relaxation), and this sagging may cross-load a shaker stinger causing a
stiffening effect that can cause natural frequencies to shift significantly; inconsistent
data such as this can impede curve fitting efforts.
A more significant concern arises for more general structural dynamics and
vibration testing, perhaps on a shaker table, where the environment is to be a specific
spectrum profile for a specific amount of time. A test like this might need to run to
completion in order to satisfy some requirement, for example a fatigue life test may
run until a component fails, and this leaves no time to download data off of the
camera. The utility of DIC for this type of measurement may need to be evaluated
on a case-by-case basis.
If the excitation profile is stationary over the period of the test, one or more
DIC measurements taken during the test might be considered representative of the
response to the applied excitation. If the excitation is transient, the DIC system
may not be able to record all of the various portions of the test. The test engineer
may need to pick which portions are measured, for example, if one portion of the
excitation profile is expected to be more damaging to the component under test.

8 Sources of Measurement Error and Best Practices

8.1 Modes of the Stereo System Hardware and iIts Measurement

Effect

All mechanical structures that have stiffness, mass, and finite boundary conditions,
when excited will have standing waves that are referred to as modes of the system.
The specific modes (e.g., bending, in-plane, torsional) are dependent on the unique
structural geometry and the boundary conditions. One of the fundamental assump-
tions for DIC testing is that once a camera pair is calibrated, the cameras’ relative
position should not change. For DIC testing, the cameras are typically placed on
an aluminum or steel camera bar that can exhibit modal vibration influencing the
measurement. Several things should be considered by the practitioner: (1) although
the rigid body motion of the camera bar can be compensated for during post-
processing, it is generally best to keep them fixed in space and eliminate all
vibration that is transmitted to the camera bar mounts or to the tripods supporting
the camera pair; (2) the tester needs to be aware of the vibrational modes of the
camera bar and avoid structural modes that coincide with the external excitation
through the flanking path from the excitation source. A list of natural frequencies
and mode shapes of common structures can be found in [13]; (3) for some cameras,
432 C. Niezrecki et al.

a cooling fan is located within the camera and the cameras themselves can be the
source of mechanical excitation. If possible, the camera bar should be designed
such that the modes of the camera bar or tripod are not coincident with the blade
passage frequency of the fan and its harmonics. If possible, the camera cooling fans
need to be off during the measurement and if the practitioner is concerned about
self-excitation, a measurement can be made on a stationary object to help quantify
the measurement errors.

8.2 Aliasing

8.2.1 Description of Temporal Aliasing for Image Processing

When conventional data acquisition is performed using typical sampling hardware
(e.g., A/D converters), anti-aliasing filters are applied to the input signal to eliminate
aliasing (see Sect. 7.1). The filters eliminate frequency content past the Nyquist
frequency and therefore the sample rate is typically 2.56 times the desired bandwidth
for many data acquisition systems. The anti-aliasing filters reduce the amplitude of
the higher frequency content and prevent the influence of higher frequency signals to
fold about the Nyquist frequency into the lower frequency spectrum. Unfortunately,
when most cameras sample data, there is no simple way to filter out structural
motion that has frequency content above the Nyquist frequency (half the sample
rate). For image processing, no robust anti-aliasing filters currently exist. However,
as described in Sect. 5, as the frequency of a structure’s motion increases, the dis-
placements will decrease. Therefore, in practice the higher frequency displacements
are measured by the camera but their relative magnitude is typically negligible. As
a rule of thumb, the sample rate of the cameras should be at a minimum higher
than the twice the highest frequency of interest (bandwidth) and any displacements
occurring past half the camera sample rate should be very small or below the noise
floor of the DIC system. A single point laser vibrometer or accelerometer can be
used in this case to ensure there is negligible motion above the Nyquist frequency
of the cameras.

8.2.2 Mitigating Aliasing with a Single Point Measurement

To identify if aliasing is present in the optical measurements, the following approach
can be taken. A single point laser Doppler vibrometer or accelerometer sampling at
a higher frequency will help the test engineer determine if there is any frequency
content in the optically sampled images above the bandwidth of interest that may
alias and contaminate the data.

8.3 Artificial Aliasing to Enhance Measurement

Digital camera technology has gone through rapid advancement in the last three
decades and has been built on a ∼ hundred years of film camera motion picture
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 433

evolution. Some of the lessons learned from prior generations can be adapted to DIC
measurement to enhance structural dynamic measurement and are now discussed.

8.3.1 Stroboscope Lights and High-Speed Measurements

Dynamic motion events (rigid body or structural dynamic) have the potential to have
blurred images and may require cameras to be triggered in correspondence with the
motion at a frame rate that is less than the frequency of operation. A short shutter
time does not allow enough light to be recorded by the camera’s pixel array under
normal lighting conditions. A stroboscope solves both the problems of needing a
sufficient amount of light in a short shutter time because a quick and bright flash
will both illuminate the object over a brief moment in time preventing motion blur.
A stroboscope can be used for events which require maximum illumination with the
shortest shutter time since the flash duration of a strobe is often less than the smallest
possible shutter time of a camera. Cameras can be triggered from the stroboscope
external pulse signal, off of a voltage signal from a variety of sensors, or from a
function generator if the precise frequency of a cyclic event is known.
Strobe lights require less power to illuminate a test object compared to typical
lights. For example, a 2 kW generator can provide the power for LED strobe
lights to illuminate a 2.5 MW utility-scale wind turbine [49]. These lights can
be synchronized with the camera shutter trigger mechanism so that they only
work for short intervals of time. In general, if at all possible, strobe times should
be longer than the exposure time (e.g., ∼twice the shutter time duration). Other
benefit of using a strobe light is a saving in energy consumption because of the
discontinuous power output and the reduced heating of the test specimen as opposed
to continuously illuminated lighting. More discussion of strobe lighting for high-
speed applications can be found in Sects. 16.9 and 16.10.
As an example, a bench test was designed to determine if measurements could
be made on a rotating blade using the DIC approach [26]. The test object used
was a variable speed commercially available desk fan with three blades. A speckle
pattern was applied to the surfaces of the blades and the hub, and a Tenma 72-7601
stroboscope was used. The cameras and the stroboscope were simultaneously
triggered by a pulse from a function generator. The setup is shown in Fig. 11.
The stroboscope allows for an intense flash of light which is less than that of the
minimum shutter time of the cameras and had a flash duration of approximately
60 microseconds. The desk fan was rotated at a speed of 18 Hz (1080 RPM).
The fastest possible camera speed is 12 fps. In order to record the shape of the
fan at all rotation angles, the stroboscope and cameras were triggered at a rate of
8.96 Hz. This translates to one image taken every 363.25 degrees of actual fan
rotation.
From the captured images (an example is shown in Fig. 12a), a rendering of the
fan’s surface at one point in time can be created (see Fig. 12b). Three complete
rotations of the fan blades were captured where approximately 40 image pairs were
taken of each rotation. Surface displacements were calculated by subtracting the
measured surface shape of the fan in a static condition from the measurements
of the fan made while the blades were rotating. The measurements can be de-
434 C. Niezrecki et al.

Fig. 11 Experimental setup

showing the use of DIC and a stroboscope
stroboscope for measuring the
shape of a rotating fan [26]

fan
computer

stereo camera

a) b)
location 1

location 2

location 3

Fig. 12 (a) One of the images captured by DIC using the stroboscope technique to capture the
position of the rotating fan blades; (b) rendering of the fan blades using the data captured while
the blades were rotating with the location of three points whose displacements were plotted as a
function of rotation angle [26]

rotated allowing them to be displayed in a rotating coordinate system. One point

on each blade was chosen to demonstrate the cyclic displacements of the fan blades.
Displacement in the out-of-plane direction at the chosen point locations are plotted
as a function of the angle of the fan for approximately three full cycles of the
fan (see Fig. 13). These plots show that the fan had a rigid body displacement
of approximately 3.5 mm during operation as compared to its initial position.
Moreover, the results indicate that two of the points are moving approximately in
phase with each other while the third point is moving essentially out of phase.
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 435

-2

-2.5
out-of-plane dis plac ement (mm)

-3

-3.5

-4

-4.5
loc 1
-5 loc 2
loc 3
-5.5
0 50 100 150 200 250 300 350
angle of rotation (degrees)

Fig. 13 Measured displacements of the points shown in Fig. 12b as a function of rotation angle
[26]

0.1
dis plac ement [mm]

actual displacement
measured displacement
0

-0.1
0 0.5 1 1.5 2 2.5 3 3.5
time [s]

Fig. 14 Example of how low-frequency measurements can be made to capture a high-frequency

periodic structural motion

8.3.2 Phase Stepping

Generally speaking, aliasing should be avoided for dynamic measurements. How-
ever, when a structure is moving in a periodic manner, aliasing can be exploited to
enable the use of low-speed cameras to measure the motion of a structure whose
operating frequency exceeds the sample rate of the cameras. The term “phase-
stepping” refers to a measurement in which the time period of the camera shutter
opening (or image sample) is an integer number of periods of the periodic structural
vibration plus a small additional time duration. For each image sample the time
duration between image samples is increased a slight amount or phase-stepped
so that the cameras record the motion of the structure throughout its periodic
oscillation. An example of a phase-stepped measurement to capture a high-speed
cyclic event with a low-speed camera is shown in Fig. 14.
As an example, a test was conducted on a base-upright structure in which forced
normal mode testing was conducted to drive the structure at resonance (26 Hz)
[88]. Shakers were mounted near the base of the upright, as shown in Fig. 15.
The camera pair then captured a series of images throughout several cycles using
a phase-stepping approach. Figure 15 shows an example of the motion captured at
the 8 measurement points for the 26 Hz structural mode using 3D point-tracking
436 C. Niezrecki et al.

Fig. 15 Base-upright structure with shaker orientation and measurement points (left) and sample
output of point-tracking obtained using phase-stepping (right) [88]

measurements with the cameras sampling at ∼11 fps. Note that each signal is in-
phase, which is to be expected when measuring what is essentially the first bending
mode of the structure. The three points along the top of the upright are moving at
roughly the same amplitude while those along the midline are moving in phase
with approximately half the amplitude as the top points. Likewise, the bottom
measurement points display very little motion since they are located near the root of
the upright portion of the structure.

8.4 Lighting Requirements, Shutter Time, and Lens Adjustment

8.4.1 Lens Adjustment

Camera lenses need to be adjusted before every measurement in order to provide
a proper focal length, focus, aperture, and shutter speed. The focal length is the
distance between the optical center and image sensor (see Fig. 1). The focal length
can determine the dimensions of the FOV (b) by using the equation below.

d
b=s (2)
f

In this equation, s, f, and d represent the sensor size, the focal length, and the
distance between the camera and the object, respectively (see Fig. 1). Longer focal
length lenses (e.g., telephoto lenses) provide an equivalent FOV at greater standoff
than a shorter focal length lens. When using these lenses, special attention must
be paid to the camera fixtures because any camera movement is greatly magnified.
On the other hand, for some measurements, there might be a space limitation, and
the cameras must be installed close to the object (e.g., measurement inside a wind
tunnel). For these measurements, short focal length lenses (e.g., 10 or 12 mm lenses)
can be used to provide a wider FOV. However, using very short focal length lenses
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 437

Fig. 16 Effects of aperture

on the FOV; smaller aperture
size can increase the depth of
field

(wide angle lenses) can significantly increase distortions in the images particularly
when the sample has large translations in the images. Fixed focal length lenses
are more common in photogrammetry than zoom lenses because they create fewer
optical distortions in the photos.
The depth of field of photographs is critical when large rigid body motions occur.
The depth of field is affected by three parameters: (1) focal length – the shorter the
focal length, the deeper depth of field; (2) distance from the object to the camera –
the closer to the camera, the shallower depth of field; and (3) aperture – the smaller
the aperture (f/stop), the deeper depth of field you will have.
The relationship between the aperture size, shutter speed, and depth of field is
graphically shown in Fig. 16. The aperture specifies the amount of light that is
transmitted to the sensor. A wider aperture allows more light to enter the camera
box and sensor; thus, it would need a shorter shutter time. However, using a wide
aperture would also create a shallow depth of field. On the other hand, when a
smaller aperture (a large f-stop) is used, a longer shutter time or more light is needed
to allow enough light to transmit to the sensor. This setup creates a deeper depth of
field. Selecting a proper aperture size depends on the lighting condition in the test
space, the required shutter speed, and the distance that the object is moving away or
toward the cameras. When measuring large rigid body motions, a higher f-stop (a
smaller opening of the lens) allows one to track the object while the subject is still
in focus; however, extra illumination might be necessary.
It should also be noted that when the aperture is wide open, the cameras are very
sensitive to focus. Thus, one should focus the cameras when the aperture is wide
open. Afterwards, the aperture can be closed to the desired aperture size.

8.4.2 Blurring
In dynamic conditions, moving objects may create blurring effects (see Fig. 17). In
other words, the shutter time sometimes is not negligible with respect to the velocity
of the targets. The shutter time must be fast enough to prevent image smearing.
When an object has a velocity of v (mm/s) and the movement is recorded using
438 C. Niezrecki et al.

Fig. 17 (a) An image showing a pattern that has no blurring, (b) the same photo when it contains
blurring because of horizontal movement, (c) a photograph showing that the photogrammetry
software cannot identify some optical targets due to high blur of the targets near the tip of the
blades [8]

a lens with an image magnification of k (pixel/mm), the displacement during the

exposure time of t is calculated as:

w(pixel) = k.v.t (3)

As an example, for an object with an in-plane velocity of 10 m/s, which can be

seen in rotating structures, and the exposure time of 8 microseconds, and image
magnification of 7 pixel/mm, the displacement during the exposure is 0.48 pixel
of motion during the image capture. This corresponds to a very large motion that
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 439

creates blurry photos. The displacement during exposure time value should ideally
be very small to create focused photos. It is important to note that for point tracking,
ellipse finding algorithms can still find the center of the moving circular optical
targets with reasonable accuracy even when the images are slightly blurry. It should
be noted that the blurring is more critical for dot targets than speckle patterns. Also,
change in the blur is more critical than the absolute blur value. More discussion
about blurring is found in Sects. 10.1 and 16.

8.4.3 Lighting
Uniform illumination of the test object is a crucial part of generating high-quality,
high-contrast photographs with minimum noise. Lighting might be required because
of the high frame rates used for vibration measurements and the small apertures
used for increased depth of field. A high contrast between the pattern/optical target
and the background is critical for accurate measurements. Diffuse lighting leads to
uniform illumination. Thus, when appropriate, sunlight can be used for illumination.
Indoor measurements need to be performed using several lights or strobe lights
to create a uniform illumination. Uniform lighting can also be provided using
photography shades. Halogen lights have been conventionally used but LED lights
have recently been more popular. The LED lights require less power and provide
light with less heat than using a halogen light.
For DIC patterns, a flat or matte paint can be easily illuminated with a uniform
light. Illuminating a gloss paint is very challenging because of reflections that
create highlights in the photos. However, for point tracking measurements on
large test objects in dark environments, using retro-reflective targets and a lighting
system can limit the background light required. Researchers have suggested using
monochromatic lights to reduce the effects of radiations from tests object at high
temperatures [18] and to perform measurements in varying ambient light [51].

8.4.4 Heating Effects Due to Lighting

During an experiment, the lighting system can increase the temperature of the
object. This temperature increase can be detected by touching the specimen or using
infrared thermometer. The heat may change the material property of the structure or
induce artificial strain due to thermal expansion. Furthermore, the heat waves may
influence the camera sensors. This is more critical when the specimen is patterned
with a dark background. Thus, proper measures, such as using a fan or switching off
the lights (when not used), should be adopted. LED lights generate less heat and are
becoming the more popular choice for illuminating a test object. More discussion
about heating effects is found in Sect. 10.1.
Additional discussion of lighting, strobe lights, and heating effects is provided
in the section on DIC for High-rate Testing located at the end of this chapter (see
Sect. 16).
440 C. Niezrecki et al.

9 Excitation Strategies for Modal Testing and Application

to DIC Measurements

9.1 Impact Testing

Impact testing is widely used in experimental modal analysis and is performed with
an instrumented modal hammer. It is generally quick to set up compared to shaker
testing: no shaker alignment needs to be performed and no amplifiers are needed for
excitation signals. There is some additional effort involved when hammer testing in
that a trigger needs to be set so the data acquisition system (cameras for DIC and
whatever additional data acquisition system will measure the hammer input) records
at the correct time.
For impact testing, there are few parameters that are controllable. The test
engineer can generally control the force applied when impacting, though perhaps
with a lesser degree of precision due to human factors involved. The mass of the
hammer and the stiffness of the tip can also be varied and will both influence the
length of the impact and thus control the frequency content of the impact. A more
massive hammer will generally remain in contact with the part for a longer period
of time, resulting in a wider pulse which has lower frequency content. Similarly,
a softer hammer tip will also result in the hammer remaining in contact with the
part for a longer period of time. Conversely, utilizing a lightweight hammer with a
very stiff tip will result in a very short impact pulse and therefore higher frequency
content imparted to the test article. Figure 18 shows three hammer hits with varying
pulse widths and corresponding frequency content imparted to the structure.
The hammer spectra shown in Fig. 18 clearly demonstrate that there is not a sharp
cut-off in the frequency content exciting test article. This means that modes outside
the bandwidth of interest may be excited by the hammer impact if the spectrum is
not tuned correctly. This is especially important for DIC measurements which have
limited options for dealing with frequency domain aliasing: out-of-band modes may
alias down into the bandwidth of interest and contaminate the data. When impact
testing with DIC, it is strongly suggested that the hammer tip and mass be tuned
so that the hammer input spectrum rolls off sufficiently past the frequency band of
interest, but within the sampling rate of the cameras.

9.2 Shaker Testing

Shaker testing is advantageous to hammer testing in that the test engineer can often
specify the exact signal sent to the shaker, and therefore can appropriately tune the
force spectrum that the component receives. Additionally, shaker testing can often
provide inputs at multiple locations on the test article which can shorten the testing
time (MIMO) compared to exciting each location sequentially. However, shakers
have some disadvantages compared to impact testing. The shaker is physically
attached to the test article, which can change the dynamics of the system due to mass
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 441

3/4 lb hammer, rubber tip

1/4 lb hammer, plastic tip
0.1 lb hammer, metal tip
Force

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035

Time (s)

3/4 lb hammer, rubber tip

1/4 lb hammer, plastic tip
Log auto-power

0.1 lb hammer, metal tip

0 1000 2000 3000 4000 5000 6000 7000 8000 9000

Frequency (Hz)

Fig. 18 Normalized force-time histories and auto-power spectra from three hits from different
modal hammers

loading and can also introduce spurious modes of the stinger and shaker hardware
if they couple with the dynamics of the system.

9.2.1 Sine Excitation

One of the simplest signals that can be supplied to a shaker is a single sinusoid. This
signal excites the system at a single frequency, and if the test article is sufficiently
linear, it will respond at that frequency as well. Sine testing provides the highest
signal-to-noise ratio of any excitation technique, but that signal is limited to only
one spectral line. A sine dwell excitation can provide large deformations at a
resonant frequency; this can make it easier for a DIC system to capture the mode
shape.

9.2.2 Swept Sine/Chirp Excitation

A slowly sweeping or stepped sine excitation can be used to measure responses
frequency line by frequency line, but one must be careful to allow sufficient time
between measurements for any transients due to the change in frequency to die
442 C. Niezrecki et al.

out. This type of testing results in long testing times and long acquisition periods,
resulting in a potentially large number of DIC images which would then need to be
downloaded and processed. The measurement may need to be paused and the images
downloaded if the camera memory is not sufficient. Using a chirp input is similar
to sine testing in that the signal is deterministic. However, instead of measuring
one frequency line per measurement frame, the sinusoid sweeps from a starting
frequency to an ending frequency over the course of the measurement frame. A
chirp excitation is typically quick and generally provides very high signal-to-noise
ratios and FRFs with high coherence.

9.2.3 Pure Random Excitation

Pure random excitation is generally implemented as a stationary random signal
with a Gaussian probability distribution. Because it is random, there is generally
frequency content over the entire bandwidth, not just the discrete frequency lines
in the FFT: the signal is not periodic within the measurement frame. Leakage is
a concern for this excitation method so windows are generally necessary. Pure
random excitation can be advantageous in systems where nonlinearities are present
if the test engineer wishes to obtain a linear representation of the system for a
specified level of excitation. Often a large number of averages are required to ensure
that all frequencies are adequately excited. This can be a disadvantage for DIC
measurements due to the large number of images that would need to be downloaded
and processed. The measurement may need to be paused and the images downloaded
if the camera memory is not sufficient.

9.2.4 Pseudo Random Excitation

Pseudo random excitation is a sum of sinusoids with constant amplitude and
randomized phase having frequencies equivalent to the discrete frequency values in
the FFT. If enough delay cycles are run for the startup transients to die out, pseudo
random excitation is then periodic in the measurement frame so it does not suffer
from leakage: no window is necessary. Because there is a constant amplitude at
each frequency line, this input strategy can use fewer averages than a pure random
excitation, making it more attractive for a DIC measurement. The downside of
the pseudo random excitation occurs for significantly nonlinear systems where the
distortion caused by those nonlinearities will not be removed: the excitation level
at each frequency line is constant, so the nonlinear effects are not averaged over a
range of amplitudes, as they are in a pure random input.

9.2.5 Periodic Random Excitation

Periodic random excitation is similar to pseudo random in that the frequency
content is tailored to the discrete frequency values in the FFT, but instead of
only randomizing the phases of the frequency content, the amplitudes are also
randomized. If enough delay cycles are run for the startup transients to die out,
periodic random excitation is periodic in the measurement frame, so similar to
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 443

pseudo random excitation, it should not require any windowing. However, like
the pure random excitation, a large number of averages are required to ensure
that all frequencies are adequately excited. This can be a disadvantage for DIC
measurements due to the large number of images that would need to be downloaded
and processed. The measurement may need to be paused and the images downloaded
if the camera memory is not sufficient.

9.2.6 Burst Random

The burst random excitation is a transient random signal that after a specified
percentage of the measurement frame is set to zero. When the signal to the shaker
cuts off, the structure then rings down and the responses as well as the forces
applied to the structure decay to zero. If this entire transient is captured within the
measurement frame, no window is needed to reduce leakage, but if the transient dies
out to quickly in the measurement frame, the signal-to-noise ratio may be reduced.
Similar to pure random excitation, a large number of averages are required to ensure
that all frequencies are adequately excited. This can be a disadvantage for DIC
measurements due to the large number of images that would need to be downloaded
and processed. The measurement may need to be paused and the images downloaded
if the camera memory is not sufficient.

9.3 Recommended Inputs for DIC Testing

Because of the large amount of data gathered per sample when using DIC, the
number of samples that can be taken before the cameras need to have their memories
dumped to a computer may be limited. For this reason, excitation techniques requir-
ing many measurement samples may be difficult or time-consuming to perform.
Impact testing and pseudo random excitation are two good candidates for excitation
that can be used that generally require few averages. Additionally, because pseudo
random is periodic in the measurement frame, low-speed cameras could potentially
be used to measure relatively high frequency content if the triggering is set up
appropriately (see Sect. 8.3.2 on phase-stepping).
However, none of these signals can be applied blindly. In impact testing the
frequency content is not easily controllable; the test engineer may have only a
handful of hammers and hammer tips to choose from to tune the frequency content
of the impact. Improper frequency content in the impact can lead to serious issues
such as frequency domain aliasing. Additionally, due to its transient nature, impact
testing tends to have a poorer signal-to-noise ratio than the shaker inputs that
last the entire measurement frame. A low-level modal test attempting to linearly
excite a structure may already require a low-level force, and the tail end of the
response decay may not be resolvable or may be very noisy. Pseudo random
testing will not adequately remove the distortion from the test article due to
nonlinearities, so if a system is not sufficiently linear, it may not be a good excitation
strategy.
444 C. Niezrecki et al.

10 DIC and Photogrammetry Measurement Range and Noise

Floor

Similar to other measurement systems, DIC results are subjected to bias and noise.
Bias is the systematic deviation of the results from correct values while noise is
random and has a zero mean. Proper setup and calibration can significantly reduce
the bias. On the other hand, noise can be minimized but never removed.
As a rule of thumb, with current camera technology, the nominal noise floor for
the displacement measured with DIC in the time domain is as low as 1% of the
object sample size (see Fig. 1) for in-plane and 3% for out of plane motions [88].
For example, if the pixel within in the image spans 1 mm of the FOV, the in-plane
and out-of-plane displacement will have a resolution of no better than ∼0.01 mm
and ∼ 0.03 mm, respectively. A well-calibrated DIC system can also measure strain
with a noise floor as low as 5 microstrain [65] but more typical results are on
order of 50 microstrain. It should be noted that by using proper data processing
techniques, better accuracies than the nominal noise floors can be reached. The
parameters that influence the noise can be categorized as “pre-measurement” and
“post-measurement parameters.” A list of these parameters is shown in Table 5.

10.1 Pre-measurement Parameters

10.1.1 Camera Setup and Calibration

The camera type and setup can change the accuracy of the DIC system. Using high-
resolution cameras can improve the accuracy of DIC for measuring displacements
(because it increases the resolution of the obtained images). However, strain
accuracy stays constant when the camera resolution changes. Using high-speed
cameras that can provide the proper frame rate to capture the vibrations can affect
the noise floor [20]. For a stereo camera system, the cameras should be synchronized

Table 5 Parameters that affect the noise floor of DIC

Pre-measurement parameters Post measurement parameters
Camera calibration Subset size
Camera resolution and sample size Subset step
Camera angle Shape function
Shutter speed and sample motion Strain calculation size
Camera synchronization Filtering
Illumination
Speckle pattern size
Speckle pattern density
Aliasing
Air turbulence
Thermal radiation
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 445

to within a small fraction of the exposure time. The camera calibration can also
affect the accuracy of the DIC measurement. Selecting a proper calibration volume
and calibrating the entire space helps to ensure that the cameras are accurately
calibrated (see Sect. 1).

10.1.2 Blurring
Due to the fast motion of objects in vibrating structures, blurring can be one of
the sources of uncertainty in dynamic measurements. Selecting a proper shutter
speed (see Sect. 8.4) with minimum image blurring can improve the accuracy of
the system. The graphs shown in Fig. 19 can be used to quantify the uncertainty of
DIC due to blurring. The results are extracted by adding motion blur in the form of
Gaussian noise to the images. In this figure, w (x-axis) is defined as the length of
the path covered by a target during the exposure time (see Eq. 3 in Sect. 8.4). The y-
axis shows the mean and standard deviation of displacement and strain uncertainty
of the DIC system. As can be seen, the uncertainty significantly increases when the
imposed motion blur increases.
It should be noted that blurring creates more uncertainty when there is a change
in blur amount. For example, the velocity of the object due to the impact made by a
modal impact hammer changes rapidly and the blur is different between before and

Fig. 19 Mean and standard deviation of displacement in case of rigid motion with zero displace-
ment after adding Gaussian Noise (V and eyy are the displacement and strain in the direction of
motion and U and exx are in the orthogonal direction, respectively) [90]
446 C. Niezrecki et al.

after impact. However, with shaker testing – the blur stays approximately constant
during the entire event.
An image pre-processing technique has been suggested to minimize the effects
of blurring in dynamic applications [91]. In this approach, a blurry target can be
compared to the original target in the reference frame to identify the mathematical
function that can convert the original target to the blurry target. When this part of the
blur is identified, the pure displacement can be measured with higher accuracy. This
technique can reduce the bias error and the uncertainty of measurements in dynamic
applications of DIC.

10.1.3 Speckle Pattern and Target Shape

The DIC speckle pattern is another parameter that can influence its accuracy. A
proper speckle size (3–7 pixels) and contrast can improve the accuracy of the results.
Using soft-edge speckles has shown better results than sharp-edged speckles [62].
The soft-edge pattern can be generated by using inkjet printers, sharpies, stochastic
sprays, rollers, pattern brushes, or pattering stencils. Printed adhesive mask foils
are also used to create patterns. Post filtering of the images can also create soft
edges for the pattern. It should be noted that contrast is the most important factor
for patterning and should be optimized – even if the speckle edge is sharp. On the
other hand, images can be post-processed to create soft-edge speckles.
Optical targets used for point tracking may have different shapes (e.g., square,
circle, or ellipse). It has been shown that some point tracking algorithms are more
accurate using sharp angles targets (e.g., square shape targets) [20].

10.1.4 Camera Angle

The shooting angle of the cameras and the stereo angle can also affect the accuracy
of DIC measurement. To obtain a high accuracy for in-plane displacement and strain
measurement, the cameras need to shoot perpendicular to the object surface. For 3D
measurements, the minimum stereo angle between cameras should be ∼11 degrees
to satisfy triangulation requirement. However, a larger stereo angle (e.g., ∼35
degrees) can result in a better out-of-plane displacement accuracy. Thus, selecting an
appropriate stereo angle depends on whether out-of-plane displacement or in-plane
displacement and strain is desired. It should also be noted that the larger stereo angle
may cause difficulties when curved objects and objects with complex geometries are
tested. In these cases, the two cameras might not have line of sight on all areas of
the object.
The stereo angle is very critical for the cameras with wide-angle lenses. Figure 20
shows captured images using 8-mm and 75-mm lenses when they are placed in
narrow and wide stereo angles. The uncertainty of the stereo-correlation is shown
with the red squares (i.e., Matching Error in Fig. 20). The green region is captured
using the 75-mm lenses and the blue region is recorded using the 8-mm lenses. The
green and blue lines intersecting with rays of light represent the sensor planes for
the 75-mm and 8-mm lenses, respectively. As can be seen, the wide angle lenses
move the sensor plane toward the cameras; this can increase the effects of matching
error. Figure 20 shows that the out-of-plane error for the cameras used with wide
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 447

Fig. 20 Uncertainty as a function of stereo angle and focal length [56]. The figure shows the
uncertainty in DIC when a 75-mm (results shown in green) and an 8-mm (results shown in blue)
lenses are both used in narrow and wide stereo angles

stereo angle is smaller than the when the cameras are mounted with narrow stereo-
angle. Thus, it is best to use wide-angle lenses with wide stereo angles to reduce
the matching error (minimum stereo angle of 25 degrees). Furthermore, because
distortion increases on the corners of the lenses, noise is highest in the image
boundaries. Thus, to have less distortion when wide-angle lenses or small stereo
angles are used, it is recommended to set the area of interest near the optical axis
(center of the images).

10.1.5 Air Turbulence

The air turbulence and thermal fluctuations can also increase the uncertainty of DIC
measurements [32]. Camera fans can create turbulence in the air and motions in
cameras which are sources of noise; using cooled cameras that do not have fans
significantly reduces the noise floor. It should be noted that the air turbulence is
less problematic in the frequency domain because usually it is distributed in the 1
to 50 Hz range. On the other hand, the lighting system can heat up the cameras
and specimen. As such, the lighting system should be installed at a higher-level
position with respect to the cameras and as far from the cameras as possible. It has
also been shown that air turbulence for experiments at elevated temperatures can
also distort the images. Using a fan during measurement or testing in a vacuum
are two suggested approaches to reduce the noise in these cases [81]. Under no
circumstances should the lights be placed directly below or in front of the cameras as
448 C. Niezrecki et al.

the thermal convective turbulence in the air will cause air density variations leading
to light refractions and distortions in the images taken.

10.2 Image Correlation and Data Processing Parameters

The image correlation parameters can effectively influence the accuracy of DIC.
A subset should include an adequate number of grayscale variations. The spatial
resolution (i.e., the spacing between each of the data points) of DIC is directly
related to the step size. Similar to finite-element analysis, one may use subsets
of small sizes to extract more accurate results with finer resolution. In order to
use smaller subsets, the speckle sizes must also be small so that the small subset
can contain enough grayscale variations to perform accurate correlation. Another
approach to improve the spatial resolution of DIC is to increase overlapping parts
of subsets up to 50% of the subset size (using smaller subset steps). However, this
also increases the computation time.
The displacement and strain measured by DIC will virtually always contain some
noise. The noise can be reduced using low-pass filters and spatial filtering. A median
or Gaussian spatial filter is usually used to serve this purpose (median is best when
outliers exist in the data median while Gaussian can reduce the noise). In this
approach, the median value in a matrix of unfiltered data around a point is assigned
to this point. The size of this matrix can be adjusted based on the measurement.
Because strain is related to the spatial derivative of displacement values, it contains
more noise than displacement. A suggested filter size of seven, with a count of three
to measure strain, is used as a default in some software packages to reduce the noise
in strain data. Furthermore, the computation size specifies the number of data points
that are used in the strain calculations (it defines the virtual strain-gage size) [63].
For more information interested readers are referred to the “iDIC Good Practice
Guide” [28].

10.2.1 Data Processing Parameters

Postprocessing techniques for DIC data can significantly influence the accuracy.
In quasi-static configuration, several photos can be acquired from the object in a
single position and the results are averaged to reduce the noise. However, this is
not possible for vibration measurement. The results of DIC in the time domain
are dependent on the noise floor of a DIC system. This noise floor might not be
low enough considering small magnitudes of displacements and strain for vibrating
structures. However, the noise in DIC results can be assumed as Gaussian noise.
Therefore, when the results are transferred to the frequency domain, the noise is
quite evenly spread on the entire bandwidth. On the other hand, the response of the
structure in the frequency domain is concentrated at a few resonant frequencies.
Therefore, higher accuracies can be achieved when the data is processed in the
frequency domain (see Sect. 6).
Researchers have also used finite-element smoothing and the least-square method
to improve the accuracy of the DIC system [39, 50]. In dynamic conditions, a
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 449

linear combination of the mode shapes can be used to smooth and expand the
measured data and reduce the noise [5, 6]. Similar to conventional modal analysis
measurements that need several averages for a single measurement, an averaging
approach on the final results can be used to reduce the noise floor [25].

11 Strain Mode Shapes

Conventional modal analysis techniques use accelerometers to measure the response

of a structure and eventually extract displacement mode shapes of the structure.
However, in many applications such as damage prognosis, durability and fatigue
analysis, and structural health monitoring, strain data can be a better representative
of the structure performance and many of these structures are monitored using
strain-gages. Strain mode shapes have recently received more attention and some
modal software packages (e.g., LMS) have added strain mode shape modules. For
these types of measurements, strain-gages are used to measure strain at a few
locations, and modal analysis is performed on the strain data to identify the modes.
Strain mode shapes can be extracted using numerical and finite-element models.
DIC offers new capabilities to extract full-field strain mode shapes with no need to
develop a finite-element model.
To extract the equations for strain mode shapes, we start with displacement mode
shapes. Equation 4 shows the basic theory of modal analysis. This theory states that
the response of a structure to excitations can be expressed as a linear combination
of the mode shapes.


n
 
u(t) = pi (t){φ}i (4)
i=1

In this equation, u is the response of the structure, and {∅}i is the i-th
displacement mode shape, while pi (t) shows the contribution of each mode in the
response. Using the theory of elasticity, the strain in the x-direction can be calculated
as:

∂u
εx = (5)
∂x

A similar equation exists between the displacement and strain mode shapes.

∂{φ}i
{ψ}i = (6)
∂x

where {ψ}i shows the i-th strain mode shape. It can be shown that the strain response
of a structure (ε) can also be represented using a linear combination of its strain
mode shapes.
450 C. Niezrecki et al.


n
ε(t) = pi (t){ψ}i (7)
i=1

The contribution of each mode in the response depends on the excitation force
(F).

−1
pi = i {φ}i F, and i = −mi ωi2 + j ci ωi + ki (8)

In this equation, mi , ci , ki , and ωi , respectively represent, mass, stiffness, damping,

and natural frequency associated with a mode of structure. The equations for the
Strain Frequency Response Function [Hε ] are:

ε⎡ ε ... Hε ⎤ ⎡ ⎤
H11 H12 1n ψ1i φ1i ψ1i φ2i . . . ψ1i φni
 ε ⎢ ε ε
⎢ H21 H22 . . . H2n
ε ⎥ 
⎥
n
−1 ⎢
⎢ ψ2i φ1i ψ2i φ2i . . . ψ2i φni ⎥
⎥
H =⎢ . .. .. .. ⎥= i .⎢ .. .. .. .. ⎥
⎣ .. . . . ⎦ ⎣ . . . . ⎦
i=1
ε
Hm1 Hm2ε ε
. . . Hmn ψmi φ1i ψmi φ2i . . . ψmi φni
(9)

In this equation, m and n represent the number of strain sensors and excitation
points, respectively. The strain frequency response function matrix (SFRF) contains
information about both strain and displacement mode shapes. Eq. 9 shows that
unlike an FRF matrix, an SFRF matrix is not symmetric. Any column of the
SFRF matrix represents unscaled strain mode shapes of the structure (mode shapes
multiplied by a constant number) [21]. A column of this matrix is extracted by
exciting the structure at a single location and using multiple strain sensors to record
the response. Using a roving modal impact hammer with a single strain sensor
results in a row of the SFRF matrix (an unscaled displacement mode shape). It
should be noted that for an operational modal analysis, the excitation force is not
measured; thus, unscaled strain mode shapes are extracted.
Both strain-gages and DIC can measure SFRF of a specimen. SFRF can be
extracted using a test with strain-gages mounted to the structure. However, these
modes may not be graphically shown because the results are discrete strain values
at a few locations. On the other hand, a DIC system can present both the full-field
displacement and strain mode shapes with a graphical interface.
Many equations and relationships used for displacement mode shapes can also be
applied to strain mode shapes. For example, modal reductions/expansion techniques
have conventionally been used for correlation purposes or for real-time monitoring
of structures [5]. The System Equivalent Reduction and Expansion Process [47] is a
technique that uses displacement mode shapes to reduce or expand the displacement
data. It has been shown that strain data can also be reduced/expanded using strain
mode shapes [4, 12]. This technique can be applied when the full-field strain data
for an operating system needs to be monitored while only a few strain-gages can
be mounted to the sample (e.g., wind turbine blades or helicopter rotors). The
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 451

strain mode shapes for the expansion can be extracted using a numerical model.
Developing an accurate finite-element model can be very challenging. On the other
hand, DIC enables us to extract strain mode shapes without any need to develop
a finite-element model. Using this approach, the strain mode shapes for a structure
can be extracted in a testing facility. An in situ measurement can be performed using
a limited set of strain-gages or fiber optic sensors. The limited set of measurements
can be expanded using the strain mode shapes to extract full-field results.

12 Projected Patterns Pros and Cons

12.1 Projected Speckle Patterns

In order to perform DIC or point tracking measurement, a pattern or optical targets

must be placed onto the surface of interest in order to track changes in deformation
and strain. Without a pattern that is adhered to the structure, it is still possible to
obtain a shape measurement over time but not mechanical strain. Several static shape
measurements using projected patterns were successfully performed to measure
bridge spalling [46] (see Fig. 21) and railroad tie deflection [69]. For structural
dynamic measurement, it is possible to measure the shape of the structure of interest
at each image sample. In order to determine the structural deflection, the shape
needs to be compared to some reference state at an instance in time. Again, it is not
possible to determine strain directly because the projected pattern will not respond
to changes in mechanical strain.

Fig. 21 Contour plot of deviation between damaged and undamaged bridge surfaces, and
photograph of damaged surface; red circles denote the locations of induced damage [46]
452 C. Niezrecki et al.

12.2 Deflectometry

Deflectometry is another measurement technique that can be used to investigate

modal shapes with high-speed cameras. It uses a mirror finish on the surface to
reflect a grid pattern to the camera. By analyzing the motion of the grid image, the
local slope of the surface can be measured. Extremely sensitive measurements of
the slope are possible because of the magnification effect of the reflected grid. For
more information, see Surrel et al. [77].

13 Rotating Optical Measurements

Measuring the dynamics of rotating structures has been a challenge due to wiring
and data transmission issues. However, the noncontact capability of photogramme-
try makes it useful for measuring the dynamics of rotating structures. Furthermore,
this technique is not sensitive to large displacements and rotations. This makes
photogrammetry very desirable for monitoring the dynamics of rotating structures.
Optical targets can be readily mounted to the structure; thus, point tracking has
been frequently used to measure the dynamics of rotating structures [35, 37, 49].
Another advantage of using point tracking is the similarity of the data from 3DPT
to conventional measurement systems (e.g., accelerometers). The measured data in
the time domain can be transferred to a modal software package such as LMS, for
further processing and for the extraction of the mode shapes. Many modal packages
have special tools for analyzing dynamics of rotating structures. This is usually
performed by organizing data in MATLAB and creating Universal File Format
(UFF) files that can then be used as inputs to modal analysis packages.

13.1 Frequency of Measurement, Duration, and Shutter Time

Measuring the vibrations of rotating structures usually needs high-speed cameras

due to the fast displacements that occur at the tip of the rotors and the high frequency
vibrations. The frame rate of cameras should be adjusted based on the desired
information. As a rule of thumb, the frame rate of the cameras should be set to
approximately 20–30 times the rotation frequency, but is dependent on the frequency
range of the modes of interest. In order to perform an effective operational modal
analysis on the data, the measurement duration should be 200 times greater than
the period of the lowest order modes in the data. However, for high-damping cases,
this number might be greater than 200 [48]. Similar to vibration measurements,
the phase stepping technique can be used when low-speed cameras are used for
measurements in rotating structures [26] (see Sect. 8.3.2).
The shutter time for the cameras needs to be based on the speed at the tip of the
blades. The tips of the rotor travel tangentially faster than the rest of the rotor, and
the targets and pattern at the tip of the blades are more prone to blurring. It should
be noted that a small amount of blurring might be visible when the rotation speed
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 453

is constant. In these cases, one of the images with a blurred pattern can be selected
as the reference image; the rest of the photos will likely contain the same amount of
smearing.

13.2 Camera Setup

A single stereo system can be used to measure the vibrations of a rotor. To measure
the vibrations of a rotating structure, a stereo camera can be installed above, below,
or in front of the rotor (see Fig. 22a) in order to cover full the motion over the
entire rotation. In some measurements, it might not be possible to locate the cameras
in front of the rotor. When the cameras are located underneath the rotor, a single
stereo system can only cover an azimuth angle of the entire rotation. However, this
measured data can also be used for operational modal analysis and for the extraction
of the mode shapes of single blades.
Another approach involves mounting several stereo systems to record a complete
rotation or large azimuth angles. This technique was used to measure the dynamics
of a helicopter’s rotor in a wind tunnel [1]. In this measurement, cameras were
placed underneath the helicopter. Each camera pair would only cover a quarter of
the rotation; thus, eight cameras were used to cover a complete rotation.

Fig. 22 (a) A photogrammetric measurement on a hovering helicopter when the cameras are
looking from the top of the rotor [37], (b) a DIC measurement on a rotor when the cameras are
mounted to the rotor and rotates with it, (c) progressively patterned blade used for the case with
rotating blade [76]
454 C. Niezrecki et al.

Another approach for measuring the dynamics of a rotating structure is to use

cameras that rotate with the blades (see Fig. 22b). These cameras need to have their
own batteries and can be controlled by means of a wireless system. If the cameras
are installed very close to the rotor with a high inclination angle, the patterns close
to the cameras might appear very large while the points far from the sensor are very
small in the images. This issue can be resolved by using a special type of cauliflower
pattern, or a pattern that increases its size when moving away from the center along
with elliptical optical targets (Fig. 22c). Researchers have used a tachometer as well
as a laser sensor to synchronize the cameras and strobes with the rotation speed [75].
By using this approach, images can be recorded at exactly the same angle in each
rotation to perform a more consistent measurement.

13.3 Rigid Body Correction

Rigid body correction is a technique to extract actual deformations of a structure

from measured displacements. If a group of optical targets attached to a structure is
translating or rotating together, the flexible deformation of the structure is calculated
by subtracting the rigid body motion of the entire structure from the measured
displacement. Rigid body correction in fact changes the coordinate system from the
camera to the rotor. A set of points (minimum of three) that are ideally fixed with
respect to each other during the test are considered as reference points to perform a
rigid body correction. For the rigid body correction, the measured displacements of
all points or facets are subtracted by the reference points’ translation and/or rotation.
This method is very useful when the deformations of blades in a rotating turbine are
desirable. The out-of-plane measurement is very small compared to the rigid body
displacement of the entire structure during rotation. In this case, the displacement
of the structure that is caused by rotation of the turbine might not be desirable.
Using the points on the hub, the measured displacement can be de-rotated and the
actual deformation of the blades can be extracted [36]. If the hub is not fixed, an
“instrumentation pod” can be attached to the rotor to add a few points that are fixed
with respect to each other.
Identifying the rotation plane (or surface) and rotation center is critical for
processing the data for rotating structures. The center of rotation is usually extracted
using the trajectory of markers. If a circle is fit to this path, the center of the circle
is the center of rotation. Using more targets in this process can result in a more
accurate estimate of the center of the circle.

13.4 Mode Extraction Challenges and Effects of Harmonics

Analyzing rotating data to extract mode shapes for rotating structures are chal-
lenging endeavors. These challenges also exist when other techniques are used
for measurement. An important factor that hampers the processing of data for
rotating structures is that these structures are not usually in a steady state, and their
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 455

dynamics may change during operation. Having a steady-state system is one of the
assumptions used in operational modal analysis. The support stiffness for a rotor
may change during operation based on the blade’s azimuth angle. Furthermore,
aerodynamic effects may also vary during measurement.
Aeroelastic damping is another factor that makes the dynamic measurement of
the rotating structures challenging. The aerodynamic damping can be as high as 10%
to 30% of critical damping while for most of the structures the structural damping is
less than 1% [48]. This high aerodynamic damping does not allow for the accurate
extraction of the structural modes and can mask the structural damping.
The effects of harmonics of rotation speed on the dynamics of rotating structures
are significant. In the operational modal analysis, there is an assumption that the
input of the system has a Gaussian distribution to excite all resonant frequencies
of the structure. The excitations due to the harmonics of rotation speed violate this
assumption. Many of these harmonics might coincide with the resonant frequencies
of the system; thus, it might be difficult to verify the mode shapes. The rotating data
modified by harmonics can be analyzed using multiple approaches. One technique
is to measure the operating data with different rotating speeds. This can show
the structural modes that are likely to be inherent in the results (if the centrifugal
effects are negligible) compared to harmonics of the rotation speed, which are very
dependent on the rotation speeds. Other approaches are based on using harmonic
filters. These filters remove the effects of harmonics on the data. There are many
OMA approaches that can identify the modal parameters when the system is
influenced by harmonics [43] and some modal analysis packages are equipped with
harmonic removal tools.

14 Some Experimental Case Studies

14.1 Comparison bBetween 3D Scanning Laser Doppler

Vibrometry and 3D Stereo-DIC

Some of the first quantitative comparisons of test/analysis/correlation of data

measured using the DIC approach to traditional accelerometers, a scanning laser
vibrometer, and a finite element model were presented by Helfrick et al. The results
indicated that all three approaches correlated well with the finite element model and
provided validation for the DIC approach for full-field vibration measurement [3,
27]. Another related work obtained frequency response function measurements for a
vibrating structure using 3DPT while simultaneously validating the shape of the two
optically based measurements to the two other traditional vibration measurements
as well as to a finite element model [88] (see Fig. 23).
In Reu et al. [68], a comparison between 3D Scanning Laser Doppler Vibrometry
(3D SLDV) and 3D Stereo DIC was performed using a modal test on a small
plate structure as a test problem (shown in Fig. 24). Both systems allow full-field
measurements over a surface and have the ability to compute strains; however,
the systems use vastly different methods to measure responses of the test article.
456 C. Niezrecki et al.

Fig. 23 Correlation of the mode shapes predicted by the FEM to the various sensing methods, for
modes 1 (26 Hz) and 3 (78 Hz) of a base-upright structure from Fig. 15 [88]

The authors of that work sought to determine each system’s advantages and
disadvantages.
For the test, the DIC system and 3D SLDV system were set to sample at
3906.25 Hz. About 1600 frequency lines were measured with a frequency resolution
of 0.977 Hz. Because pseudo random excitation was supplied to the shaker, only
five frames were measured and averaged to form frequency response functions.
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 457

Fig. 24 Experimental setup

for a comparison of 3D
SLDV and 3D DIC

Fig. 25 Overlay of the

subset (facet) size on the
speckle pattern over the entire
plate. Note that in the
solution there was an overlap
of approximately ½ the subset
size

Note that this is the maximum number of averages that could be recorded with
the DIC cameras used for this test. The DIC system measured 715 points on
the surface of the test article (see Fig. 25), and the 3D SLDV system measured
545 (see Fig. 26). Because the 3D SLDV scans point sequentially, the 3D SLDV
measurement took significantly longer than the DIC measurement which measures
all points simultaneously. However, due to the significant download and processing
time involved to turn DIC images into time histories, the time difference between
the two measurements ended up being negligible.
Both the DIC and 3D SLDV data needed to be transformed to accelerations via
frequency domain differentiation before curve fitting could be performed. There
was good agreement between the modal parameters extracted from the measured
data. Table 6 and Fig. 27 show the natural frequencies and mode shapes. It should
be noted that the speckle patterned surface, which was necessary for the DIC
measurement, was not optimal for the 3D LDV system, since dark surfaces do not
458 C. Niezrecki et al.

Fig. 26 Measurement point

locations for the 3D SLDV
system

Table 6 Comparison of modal parameters extracted from the test data

Frequency Damping
Mode (Hz) (%)
DIC LDV LDV w/RET DIC LDV LDV w/RET
1 529.4 529.5 524.8 0.042 0.042 0.104
2 883.0 883.1 875.0 0.022 0.024 0.095
3 891.7 891.7 883.3 0.015 0.026 0.098
4 948.3 948.4 938.8 0.019 0.019 0.071
5 1420.1 1420.4 1406.6 0.015 0.020 0.096

return adequate light to the laser head for a good measurement signal. Therefore
the 3D SLDV shapes have some erroneous points. These errors disappeared
when retroreflective tape (RET) was applied to the surface of the test article, but
application of the tape required taking down the test article, so some of the natural
frequencies and damping ratios have shifted slightly due to small changes in the test
setup. Strain shapes computed from a sine dwell test at the natural frequencies were
also compared between the two tests, and are shown in Fig. 28.
In this test it was shown that the DIC and 3D SLDV can achieve similar results for
traditional modal testing. FRFs measured with the 3D SLDV were significantly less
noisy (see e.g., Fig. 29), but it seemed that the DIC strain computation algorithms
were more robust. Table 7 shows a comparison between the two measurement
methods for the various test parameters performed in this example. A broader
comparison between measurement approaches is shown in Table 8.

15 DIC Comparison to Traditional Modal Analysis Sensing

Traditionally, modal analysis has been performed by using modal impact hammers
to excite the structure and by using accelerometers to record the response. Strain-
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 459

Fig. 27 Mode shapes for out-of-plane motion for the DIC and different SLDV measurement
approaches

gages, LVDTs, and other contact-based measurement systems have been used to
measure the response of the structure. All of these sensors measure the response
at discrete locations and need instrumentation for data transmission. Fiber optic
sensors have been recently used to measure vibration to monitor structural health.
However, they need to be integrated into the structure in the manufacturing
stage and are only capable of one-dimensional measurement. Optical measurement
techniques are used in structural dynamics due to their non-contact features. These
460 C. Niezrecki et al.

3D DIC 3D SLDV

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02
εxx

0 0

-0.02 -0.02

-0.04 -0.04

-0.06 -0.06

-0.08 -0.08
-0.05 0 0.05 -0.05 0 0.05

-39.7388 -27.9018 -16.0648 -4.22777 7.60924 19.4463 31.2833 43.1203 54.9573

Microstrain

3D DIC 3D SLDV

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02
εyy

0 0

-0.02 -0.02

-0.04 -0.04

-0.06 -0.06

-0.08 -0.08
-0.05 0 0.05 -0.05 0 0.05

-39.7162 -29.4315 -19.1467 -8.862 1.42273 11.7075 21.9922 32.2769 42.5616

Microstrain

3D DIC 3D SLDV

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02
εxy

0 0

-0.02 -0.02

-0.04 -0.04

-0.06 -0.06

-0.08 -0.08
-0.05 0 0.05 -0.05 0 0.05

-72.9311 -58.9165 -44.9018 -30.8871 -16.8725 -2.85779 11.1569 25.1715 39.1862

Microstrain

Fig. 28 Strain from a sine-dwell test at 530 Hz for both SLDV (25-mm filter) and DIC. Top is εxx ,
middle is εyy , and bottom is εxy
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 461

360 360
270 270
180 180
90 90
0 0
200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400

2 2
10 10
X-direc tion FRF

Z-direction FRF
0 0
10 10

-2 -2
10 10
200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400
Frequency (Hz) Frequency (Hz)

DIC LDV LDV w/ Retro-reflective Tape

Fig. 29 Noise floor comparison between DIC and LDV; phase (top) and magnitude (bottom). The
z-direction corresponds to an out-of-plane measurement, while the x-direction corresponds to an
in-plane measurement

Table 7 Comparison between LDV and DIC for the tests performed by Reu et al. [68]
Comparison metric LDV DIC
Cost ≈$650 k ≈$350 k
Setup time 2h 2h
Acquisition time Hours Seconds
Analysis time Seconds Hours
Disp. resolution ≈ picometers ≈ nanometers
Strain resolution ? 5 microstrain
Strain calculation Integrated – but researchy Seamlessly integrated
Anti-aliasing Included Not possible at the moment
Data volume Small (Mbytes) includes only Large (Gbytes) but includes time
frequency data history
Software Designed for structural dynamics In its infancy
testing

optical techniques include pattern interferometry, laser Doppler vibrometry, and

photogrammetry.
The interferometry technique measures vibrations of structures by using an
interference fringe pattern created by superposing two coherent light patterns and
measuring the displacement between them. Electronic Speckle Pattern Interferome-
try (ESPI) and Digital Speckle Shearography (DSS) use the phase-shift between the
references and reflect beam waves in order to measure the vibrations of a structure.
462 C. Niezrecki et al.

Table 8 A comparison between different dynamic measurement techniques

Point-wise sensors
(accelerometers, Interferometry
strain-gages, and techniques (ESPI, Laser Doppler Photogrammetry and
LVDTs) DSS) vibrometer DIC
Needs wiring; may Noncontact Noncontact Noncontact
induce mass or
stiffness loading
Easy to obtain Needs post Easy to obtain Needs post
transfer functions processing to obtain transfer functions processing to obtain
transfer functions transfer functions
Has anti-aliasing No anti-aliasing filter Has anti-aliasing No anti-aliasing filter
filter filter
Sensitivity for each Sensitive; sensitivity Very sensitive; Sensitive; sensitivity
point depends on the goes down as the sensitivity is related improves as the field
transducer fringe gets larger to frequency shift of view gets smaller
measurement
Wide frequency range Better for Very wide frequency Better for
low-frequency range low-frequency
measurements measurements
Not sensitive to rigid Calibration is highly Alignment/calibration Calibration is less
body motions sensitive (for ESPI) are highly sensitive sensitive to changes
and sensitive (for to changes in setup in setup condition and
DSS) to changes in conditions and rigid rigid body motions
setup body motions
Inexpensive Expensive Very expensive Very
expensive/expensive
depending on camera
requirements for the
application
Low special High spatial High spatial High spatial
resolution resolution resolution resolution (DIC)
Due to sensitivity Hard to extract rigid Hard to extract rigid Easy to extract rigid
issues, it is very body modes body modes body modes
challenging to extract Applicable for Applicable for Appropriate to be
both rigid body simpler spatial simpler spatial used for highly
modes and flexible deformation patterns deformation patterns flexible structure
modes using one type and not applicable and not appropriate Very accurate for
of transducer for highly flexible for very high spatially complex
structures deflections or deformations
deformations
Very fast Very fast Very Very fast
measurement unless measurement time-consuming measurement
roving hammer or measurement
sensor techniques are
employed
Very fast data Fast data processing Very fast data Time-consuming data
processing processing processing (DIC); fast
data processing (PT)
(continued)
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 463

Table 8 (continued)
Can be real-time Can be real-time Can be real-time Usually works
off-line
Small data volume Large data volume Small data volume Large data volume
Depending on the Measures both Measures velocity; Measures
type of the sensor, displacement and can measure strain displacement (DIC
only measures strain with additional and PT) and strain is
displacement, post-processing readily obtained for
acceleration, or strain DIC;
Difficult to use for Difficult to use for Difficult to use for Very appropriate to
rotating structures rotating structures rotating structures use for rotating
due to wiring structures

They can perform fast, real-time, and accurate dynamic measurements, but they are
sensitive to ambient vibrations.
Similar to the other interferometry techniques, a Laser Doppler Vibrometer
(LDV) compares the reference beam and the beam reflected from the object.
However, an LDV measures the frequency shift between the reference and reflected
beam waves. The Scanning Laser Doppler Vibrometer (SLDV) is frequently used
for full-field vibration measurements. The laser vibrometer typically records data
sequentially, has a wide frequency range, and can measure very high frequency
vibrations. However, laser vibrometer measurements are very time-consuming and
the results might not be consistent because during the measurement the structure
or the excitation may change. Furthermore, the laser vibrometer is very sensitive to
large motions and cannot effectively measure structural dynamics when a large rigid
body motion occurs.
Photogrammetry is a noncontact measurement approach that has a distributed
sensing capability allowing for spatially simultaneous measurement. This technique
can measure the true dynamics of structures without adding mass or stiffness effects.
Photogrammetry can be used for in situ measurements and in testing conditions in
which none of the interferometry techniques can operate. Photogrammetry is also
able to measure large deformations and rotations. A comparison between commonly
used dynamic measurement techniques is shown in Table 8.

16 DIC for High Rate Testing

The flexibility of DIC and the incredible improvement in high-speed imaging have
opened a new world of research and have opened up a range of experiments that
were heretofore inconceivable. However, the move into high-rate DIC testing is not
as simple as replacing traditional machine vision cameras with their more expensive
cousins. There are a number of complications that need to be addressed before
working in this regime. These complications and potential pitfalls are the subject
of this section.
464 C. Niezrecki et al.

16.1 Definition of High-Rate Testing

For the purposes of this chapter, it is assumed that high-rate testing is any experiment
that requires frame rates that are approximately 2-kHz and above. This includes
the modal testing [11, 22, 68, 72, 87] covered in the previous sections, as well as
Hopkinson bar testing [44, 70], drop tables, explosive loading [23, 71, 74, 83], crash
testing [14], and the like. Each of these examples includes complications that may
be unique, but generally there are similar hurdles in setting up a DIC experiment for
these types of tests, and include:

1. Camera selection for high-speed testing

2. 2D versus stereo-DIC: What do I choose and why?
3. The testing environment and its complications
4. Unique camera calibration situations
5. Lighting for high-rate testing
6. Camera synchronization
7. Patterning, painting and speckling

In addition to these considerations – it is also important to consider all of the

traditional rules of DIC testing that are covered elsewhere in this volume, including:
image correlation algorithms, speckle attributes, camera calibration, and so forth.
For more information about DIC for shape and deformation measurement and best
practices for testing readers can refer to Sutton [78] and IDIC [28]. We discuss in
this section the unique problems associated with high-speed (HS) and ultra-high-
speed (UHS) DIC testing.

16.2 High-Rate Camera Selection

The high-speed camera market has expanded significantly since 2005. The number
of vendors, the frame rate, and record time have all been improving very rapidly.
While the cost of these cameras remains relatively high, the increased competition
in the market segment has held prices in check and it is hoped that they will become
even more affordable in the future. Figure 30 contains a market survey of the existing
high-speed cameras as of 2016, and for interest can be compared with the original
chart published in 2008 [66]. The cameras can be broken into two groups: HS
and UHS cameras. HS cameras have a single detector that is read out into HS on-
board memory. They have a fundamental frame rate limited by the throughput of the
detector to the memory. Because the memory rate (GBit/second) is fixed, to increase
the frame-rate, the resolution of the camera needs to be decreased. Therefore, at rates
above approximately 100 KHz in the current cameras, the usable number of pixels
becomes very small. The UHS cameras use a variety of techniques to avoid this
problem, and older generation UHS cameras use beam-splitters or rotating mirrors
to distribute the image onto different detectors, which are then read out slowly.
 8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . .

Fig. 30 High-speed testing camera survey. Highest frame rate and resolution; for a circa 2008 version see [66]
465
466 C. Niezrecki et al.

These camera types have the drawback of a limited number of images available,
usually 16 or less, and have a complicated optical path making them difficult to use
for 2D-DIC and nearly impossible for stereo-DIC. For a review of these cameras
and their pitfalls, a paper by Tiwari is an excellent source of information [82].
Fortunately for DIC, a new UHS detector architecture has arrived that includes a
single detector with fixed resolution at all frame rates and a reasonable record time.
The first such camera was the Shimadzu HPV series of cameras, followed by the
improved HPV-X and a competitor in the Specialised-Imaging Kirana camera. The
importance of the single-chip architecture of the camera for stereo-DIC cannot be
overstated. Before this advancement, stereo-DIC was virtually impossible at a MHz
or higher frame capture rate.

16.3 2D Versus 3D Stereo-DIC

Because of the camera cost, the decision to do 2D DIC may be a hard budget
requirement. However, if the UHS camera rates needed are not available in the
single detector architecture (e.g., greater than 5 MHz), or if an intensified camera is
needed, 2D may also be a practical requirement. 2D-DIC places some fundamental
and important limitations on the experimental setup that must be heeded. These
requirements include having a flat specimen and the sample must remain in the same
plane during the entire test. The first requirement is often easy to meet in material
testing with traditional dog-bone samples, but meeting the second requirement (i.e.,
remaining in plane) is extremely difficult to ensure with nearly any practical DIC
setup. The pitfalls and errors associated with out-of-plane motion are covered in
Sutton et al. [80] where the strain error is found to be directly related to the out-of-
plane motion divided by the sample stand-off distance. A few simple calculations
should suffice to convince one that even small motions (sub-mm) will lead to
unacceptably large strain errors. However, the math immediately suggests a solution
to this problem: Increase the sample stand-off. This can be accomplished by either
using a long focal length lens or better by using a bi-telecentric lens. A bi-telecentric
lens only accepts light rays arriving along the axis of the lens and therefore has
fixed magnification. This increases the “effective” stand-off significantly reducing
the out-of-plane errors. However, this advantage is problematic in terms of flexibility
because the lens then also has a fixed sample size it can image and a fixed standoff
distance, making the lens dedicated for a particular type of experiment. The bi-
telecentric lens is often large, as the lens has to be the size of the FOV. However,
they are the most practical approach to helping eliminate out-of-plane errors and
their use is strongly encouraged in situations when stereo-DIC is not practical.

16.4 Environmental Concerns

The test environment for HS and UHS testing is often dynamic and highly energetic.
Because of this there are often unique requirements of the experimental setup,
including eliminating camera motion and protecting the camera.
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 467

16.5 Camera Motion

Highly energetic testing, out-door testing, and vibration and shock testing are all
environments where the cameras are likely to move. This is far more problematic
at the high-speed regimes, where the acquisition rates are in the kHz, as the modal
frequencies of the stereo-rig are likely to overlap with the frequency content of the
test in question [11]. If modal testing is being done as described in the previous
sections, the frequency domain can be used to isolate the camera motion error;
however, for most other testing, the data must remain in the time-domain. In this
case, the camera motion will be an added error source that must be considered.
In the MHz acquisition realm, the cameras do not have time to move appreciably
before the experiment is over and camera motion is generally not a problem.
The first approach is to mitigate the camera motion by careful camera mounting
that is rigid and/or isolated from the experiment. It is particularly important that
the mounting of the stereo-pair hold the two cameras rigid relative to each other,
because the motion of the entire stereo-rig is a rigid-body-motion of the test item
and may often be removed from the data relatively easily (e.g., by performing rigid
body correction or calculating the strain on the sample). However, relative motion
of the stereo-pair leads to issues with the triangulation and will lead to bias errors
in the results. The first approach to mitigation is to move the cameras far enough
away that the shock from the experiment will be delayed in reaching the cameras
until after the data of interest has been captured. Often this is either impractical or
inherently impossible, in which case the camera motion must either be accounted
for in the uncertainty analysis and/or removed from the images [40]. Both of these
techniques require that there be a “stationary” speckle pattern somewhere in the
FOV that can be used to track the motion of the cameras.
The strategy of moving the cameras further from the test sample must be pursued
with caution. The longer the focal length of the lens, the greater the stand-off
from the sample; however, this comes with two important downsides, including
increased influence of “heat waves” and an amplified effect of camera motion.
The heat wave effect is caused by imaging through air that has variable index of
refraction along the optical path. The longer the path, the more likely there is to be
a problem. In HS imaging, refraction issues can come from simple environmental
effects, such as unevenly heated air, or from the experiment itself, with a shock
wave due to detonation for example. Again, these errors are more problematic in
the kHz imaging regime because the frequency content of the heated air may be of
the same order as the experiment. In the MHz regime, the heated air is stationary
relative to the experiment and will not cause problems. This is not true of shock
waves, which will be at a high enough rate to cause issues. The use of long focal
length lenses for either camera protection, mitigation of camera motion, or limiting
out-of-plane errors in 2D has the negative effect of amplifying a number of errors.
This can be visualized by considering the viewing angle subtended by a single pixel
on the detector (see Fig. 31). With a long focal length lens, this angle will be much
smaller than with a short focal length for an equivalent FOV. Therefore, even small
camera motions will be amplified when using a long focal length lens. That is, a
468 C. Niezrecki et al.

Fig. 31 Camera motion illustration. For 20-μm pixels focused at infinity. Pixel on object not to
scale

Fig. 32 Camera protection using Lexan and heavy steel plates

small motion of the camera will cause the camera pixel to travel a larger distance
across the sample surface than for a shorter focal length lens.

16.6 Camera Protection

The first approach to protecting the camera was traditionally to move the camera a
long distance from the test item and use a long telephoto lens. However, this comes
with the important downsides of increased index of refraction issues and increased
camera motion errors. It is therefore typically better to avoid using lenses longer
than 150 mm and provide other methods of protecting the equipment. Two choices
include optical grade Lexan (bullet proof glass) and first-surface mirrors. Both of
these options provide good camera protection but at the cost of introducing optical
distortions into the measurement that are difficult to correct. At a minimum, these
added errors should be quantified and included in the uncertainty estimates (see the
following extended noise floor section).
For direct viewing of the sample, an optical grade Lexan can be used that
provides very good camera protection (see Fig. 32). The thickness of the Lexan
will need to be minimized, without compromising the protection because thicker
material will have larger distortions. Thick Lexan or glass will also cause imaging
problems and cause problems with focusing the lens if longer focal length lenses are
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 469

used. This can be somewhat mitigated by using a smaller aperture, but at the cost
of needing more light on the sample to compensate. If the shockwave is going to
impact on the Lexan before the experiment is over, flexing of the viewport should
also be considered. The typical rule of thumb is to put as little distance between the
camera and the sample as is safely possible.
A safer option, although optically somewhat more complicated, is to use mirrors.
This allows the cameras to not be directly in the path of the hazard, while allowing
an inexpensive component to be sacrificed on each test. Mirrors of good optical
quality are generally inexpensive and easily obtained. Mounting of the mirrors must
be done with some care, as any added bending or lack of flatness will cause issues
with the calibration and triangulation. The other consideration with using mirrors is
that there is another component in the optical path that must be held in position and
stationary. This is often hard to do because the mirrors will need to be placed away
from the cameras and independent of them. As noted previously with the camera
motion, this is less of a problem during the test if the cameras are running at MHz
frequencies, but will definitely cause issues during the calibration, which acquires
the images slowly, and will result in a lower quality calibration. The same issue
will occur between the calibration time and the actual test, where it is likely that
there will be relative motion between the mirrors and the cameras. Some DIC codes
allow for an extrinsic parameter calibration correction. To do this more accurately,
a known dimension is required on the sample itself to provide scale. As a best
practice, it is good to provide some fiducials (or targets) on the surface with known
dimensions for correcting the calibration using the reference frame of the test image
series.
A final important note for using mirrors is that the images must be flipped
to correct for the left/right switch of viewing through a mirror. This can eas-
ily be automated with many after market software packages including MAT-
LAB, LabVIEW, Python, etc. and is built into many HS camera control software
packages.

16.7 Extended Noise Floor Measurements

A very useful, and indeed required technique when using Lexan or mirrors, is
to do what is called an “extended noise floor” measurement (see [28]). This is
accomplished by having a sample with a known shape (flat can be used, or maybe
better a curved surface of known radius) and translating the sample through the
measurement volume. Five to ten images that span the expected range of motion
of the sample where accurate DIC results are desired should be taken. Then, the
extended noise floor images using the DIC software can be analyzed and the rigid
body motion from the results can be removed (most DIC software provides a
function to do this). Any changes from the original shape are a result of uncorrected
distortions in the optical system, and the residual noise is the estimated measurement
resolution. This quick and intuitive top-down uncertainty approach should be done
on all experiments.
470 C. Niezrecki et al.

16.8 Camera Calibration

Camera calibration is often more difficult when doing HS and UHS DIC because
the optical system is usually more complicated due to camera protection, mirrors,
Lexan, and long focal length lenses. Because of this, a reduced calibration accuracy
can be expected, and care should be taken to understand the effect of this on the
final results, by using an extended noise floor measurement for example.

16.9 Lighting Techniques

As with the camera technologies, lighting has improved exponentially over the last
decade. The introduction of high-intensity LED lights, strobed LED lights, and
commercial strobe lights is incredibly helpful for HS imaging. We look at the pros
and cons of these light sources in this section.
Strobe lights will provide more light over a shorter period of time that can
be synchronized with the camera framing. This is helpful for minimizing the
heat coming from the light source while maximizing the efficiency of the lights.
Additionally, most HS cameras have an output specifically designed for triggering
strobes synchronously with the camera framing. There can be limits though on the
strobe rate that can be obtained and they will not work with the higher frame rates
of the newer UHS cameras.
High-intensity LED lights are another viable solution. The increased brightness
of these lights together with the improved sensitivity of the modern HS cameras
makes LED lights easy to use and an excellent solution for all but the shortest
exposures.
Flash lamps have traditionally been the preferred lighting source for ultra-high-
speed imaging. These lamps discharge a high voltage into a gas which provides
a short duration (milliseconds) pulse of high intensity light and are often a good
solution for UHS imaging. Flash lamps will require pre-triggering of the lamps to
provide time for the lights to reach full intensity before imaging. Norman D24 flash
lamps are a commercial solution that provides adequate illumination of modest FOV
at up to 5 MHz.
A last class of pulsed lights includes the CAVILUX or SI-LUX incoherent laser
illumination. They provide very short duration pulses (50 ns) with high power
and can pulse up to the MHz range. Even though they are laser driven, they are
incoherent so they do not create laser speckle. Unfortunately, applicable FOVs are
only up to 10’s of millimeters.
“Diffuse” light is the optimum for all DIC imaging [55]. Highlights and shadows
are to be avoided at all costs. This often means that diffusers, polarizers or indirect
lighting will be needed to flatten and improve the light. These techniques work
by removing any path for the light to go directly from the light source to the
camera. The problem with this for HS and UHS imaging is that you need a lot of
light and all of these techniques “waste” a large amount of the light. Therefore,
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 471

Fig. 33 (a) Specular reflections from light source, compromising DIC in those regions. (b)
Polarizers installed but not correctly aligned. (c) Correctly aligned polarizers cleaning up the
lighting for ideal DIC images [19]

to obtain adequate intensity, the lights are nearly always aimed directly at the
object. Choosing “flat” paint and using polarizers are two solutions that will help
in obtaining quality images using direct illumination.

16.9.1 Polarization
When lighting high-speed experiments, diffuse light is often very difficult to achieve
because of the required intensity for the short exposures. One helpful possibility is
the use of cross-polarization techniques [19, 34]. This removes specular or reflective
highlights from the surface because the linearly polarized light from the source
maintains its polarization during specular reflection, but loses its polarization with
diffuse reflection. A cross-polarizer on the camera then eliminates the specular
reflections and removes the highlights from the image. An example of this from
a stereomicroscope UHS experiment is shown in Fig. 33. While this technique is a
quick and easy way to clean up the images and simplifies the lighting, it should be
noted that the cross-polarization technique will result in a loss of approximately 2–3
f-stops of light. That is, four to eight times the amount of light or a similar increase
in exposure time to obtain the same contrast in the image will be needed.

16.9.2 Motion Blur

Traditional thinking in DIC was that motion blur was always a problem and should
be minimized to below the DIC displacement resolution of approximately 0.01
pixels. This criterion is overly restrictive and can be hard to meet in many high-
speed applications, where exposures cannot be made short enough to completely
eliminate motion blur. Current research [90] indicates that motion blur is acceptable
if the average position of the object will be calculated and for many applications this
is acceptable. It should be noted, however, that there are two types of motion blur:
constant and varying. If the motion blur is constant, that is, the velocity of the object
remains fairly constant over the test period, the blurred speckles will appear the same
in all images and there will be no correlation problems. However, if the velocity is
varying, say in an impact test, the blurred and unblurred image segments may be so
dissimilar that decorrelation results. At this point, the analysis will need to be done
in an incremental mode, or a new reference image chosen after the impact event
472 C. Niezrecki et al.

where the blur has been eliminated or changes drastically. The ideal situation is to
remove all blur if possible, or certainly keep it subpixel, but if that is not possible,
DIC can still be performed.

16.10 Camera Synchronization

An assumption for stereo-DIC is that the images were taken at the same moment,
that is, they are synchronized. With most HS cameras, we cannot assume exact
synchronization, even when being run in a “synced” mode (FSync for example in
the Phantom cameras). The sync error is caused by the timing architecture of the
camera, due to delays between the camera syncing pulse and when the frame is
actually taken. Fortunately, most manufacturers also supply a strobe signal that is
synchronous with the framing. This signal should be inspected with an oscilloscope
to ensure that both cameras are acquiring their images at exactly the same time.
Because of the variability in the delay, the camera software provides the option to
include a frame delay to achieve camera synchronization. When the correct delay
has been determined it is possible to synchronize the cameras to within the internal
timing frequency of the camera, usually in the 10’s of nanoseconds. By taking the
velocity of the sample multiplied by the timing error, you can calculate the relative
motion of the sample between the two stereo frames. For nearly all mechanical
experiments, the nanosecond synchronization of the cameras is more than adequate.
If the cameras are not synchronized, a bias error will occur in the data. More
information on this topic can be found in Reu and Miller [67].

16.10.1 IRIG (Inter-Range Instrumentation Group) Timing

Most HS cameras also provide synchronization with each other as well as other
data acquisition systems IRIG-B [30] timecode via the timing signals received from
Global Positioning System (GPS) satellites. Specially designed receivers convert
Coordinate Universal Time (UTC) data in to various IRIG timecodes in both digital
and amplitude modulated formats. The IRIG-B timecode [31] is the most common
format and is typically connected to each instrument (e.g., camera, data recorder,
oscilloscope) using the amplitude modulated version, especially when distributing
the signal at a distance greater than 50 m.
A few benefits of using a common timing signal such as IRIG-B are: (a)
synchronization between asynchronous data recorders and cameras, (b) introduction
of a known time scale, and (c) ability to coordinate results captured from instruments
separated by very large distances (kilometers).

16.11 Painting Techniques

For nearly all DIC tests it is required that the sample surface be painted and speckled
to provide a high contrast pattern that moves with the underlying surface. High-rate
DIC imposes the same constraints on the speckles as traditional DIC and these are
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 473

covered in [58–62, 64, 81]. The ideal speckles as outlined in these documents are
at least 3-pixels in size and have an even distribution across the surface. The DIC
technique requires that a region of the speckle pattern be analyzed in the subset or
facet, and this region must contain 3-speckles (significant gray level variations) as
a rule of thumb. To have the highest spatial resolution then, the speckles must be
carefully applied to meet these rules, without being too large or too small. Beyond
these general rules for all DIC, there are a few other considerations for high rate
testing:

1. Very careful attention should be paid to the surface preparation. The surface
should be cleaned and prepared to receive the paint according to the manufac-
turer’s recommendations. This involves having a clean and grease-free surface.
2. Use an appropriate paint. A flat white or black paint for the first coat should
be used. There are also “primer” paints that are made to specifically bond with
different metals. These will often adhere better than a cheap spray paint. Some
examples include Rust-Oleum™ and SEM brand paints.
3. The painted surface should be “fresh.” That is, you should test the sample within
24 hours or less of painting. Paint that has cured too long becomes brittle and
will fail, coming off the surface leaving nothing to track.

The last point raises an important consideration: We are assuming that the paint
surface is following the contour of the substrate surface and it is important to be sure
to investigate whether this is true in your experiment. Another approach is to not
use a painted surface. This can work in situations where the sample itself contains
contrast within it that is of an appropriate scale and contrast for DIC. Unfortunately,
having structure or test panel that already has an appropriate pattern is very rare.

16.12 Conclusion

HS and UHS DIC follow all of the same rules as for quasi-static testing, with a few
important added challenges. Most of the complications are not due to the cameras,
which behave image wise very similarly to traditional machine vision cameras, but
more because of experimental issues derived from the test itself. Most high rate DIC
testing is performed in “highly” dynamic events where camera motion and camera
protection are required. Even with these problems, nearly always the most difficult
aspect is speckling and lighting the sample. Obtaining enough light that is well
distributed and diffused can take practice and time to set up. Timing and triggering
the cameras with the experiment are also problematic for UHS experiments, as there
is a limited record time available to take the data. Tight triggering synchronization is
nearly always required to have the event of interest within the camera memory. The
final difficulty is cost. Camera equipment is expensive and unfortunately there are
no easy substitution;, however as technology advances, camera costs are continually
becoming cheaper with higher resolution and improved frame capture rate.
474 C. Niezrecki et al.

Best practices for HS and UHS DIC testing include acquiring high contrast
images with a well-calibrated DIC system taking into account the possibly com-
promised optical path and any issues with motion or timing. Because of these
complications, it is strongly encouraged that efforts be made to ensure that
adequate consideration is given to the uncertainty. Uncertainty recommendations
and information are found in Reu [57, 65].

References
1. Abrego AI, Olson LE, Romander EA, Barrows DA, Burner AW (2012) Blade displacement
measurement technique applied to a full-scale rotor test. American Helicopter Society 68th
annual forum proceedings, Fort Worth, 1–3 May 2012
2. Aramis – User Manual-Software; Aramis v6.1 and higher, 2018. https://www.gom.com
3. Avitabile P, Niezrecki C, Helfrick M, Warren CP, Pingle P (2010) Noncontact measurement
techniques for model correlation. Sound and Vibration Magazine, January, pp 8–12
4. Baqersad J, Bharadwaj K (2018) Strain reduction expansion approach. Mech Syst Signal
Process 101:156–167. https://doi.org/10.1016/j.ymssp.2017.08.023
5. Baqersad J, Niezrecki C, Avitabile P (2015a) Full-field dynamic strain prediction on a wind
turbine using displacements of optical targets measured by stereophotogrammetry. Mech Syst
Signal Process 62:284–295. https://doi.org/10.1016/j.ymssp.2015.03.021
6. Baqersad J, Niezrecki C, Avitabile P (2015b) Extracting full-field dynamic strain on a wind
turbine rotor subjected to arbitrary excitations using 3D point tracking and a modal expansion
technique. J Sound Vib 352:16–29. https://doi.org/10.1016/j.jsv.2015.04.026
7. Baqersad J, Poozesh P, Niezrecki C, Avitabile P (2016a) Photogrammetry and optical methods
in structural dynamics – a review. Mech Syst Signal Process. https://doi.org/10.1016/j.ymssp.
2016.02.011
8. Baqersad J, Poozesh P, Niezrecki C, Avitabile P (2016b) A noncontacting approach for full-
field strain monitoring of rotating structures. J Vib Acoust 138:031008. https://doi.org/10.1115/
1.4032721
9. Bartilson DT, Wieghaus KT, Hurlebaus S (2015) Target-less computer vision for traffic signal
structure vibration studies. Mech Syst Signal Process 60–61:571–582. https://doi.org/10.1016/
j.ymssp.2015.01.005
10. Bateman VI, Mayes RL, Carne TG (1997) Comparison of force reconstruction methods for a
lumped mass beam. Shock Vib 4(4):231–239
11. Beberniss TJ, Ehrhardt DA (2017) High-speed 3D digital image correlation vibration measure-
ment: recent advances and noted limitations. Mech Syst Signal Process 86:38–48
12. Bharadwaj K, Baqersad J, Poozesh P (2016) Modal expansion using strain mode shapes. In:
IMAC XXXV, Garden Grove
13. Blevins RD (1979) Formulas for natural frequency and mode shape. Van Nostrand Reinhold.
ISBN 10: 0442207107/ISBN 13: 9780442207106
14. Burguete RL, Lampeas G, Mottershead JE, Patterson EA, Pipino A, Siebert T, Wang W (2013)
Analysis of displacement fields from a high-speed impact using shape descriptors. J Strain
Anal Eng Design
15. Burtch R (2004) History of photogrammetry. Notes of the Center for Photogrammetric Training
16. Carr J, Baqersad J, Niezrecki C, Avitabile P (2015) Full-field dynamic strain on wind turbine
blade using digital image correlation techniques and limited sets of measured data from
photogrammetric targets. Exp Tech. https://doi.org/10.1111/ext.12129
17. Castellini P, Martarelli M, Tomasini EP (2006) Laser Doppler vibrometry: development
of advanced solutions answering to technology’s needs. Mech Syst Signal Process 20(6):
1265–1285
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 475

18. Chen X, Xu N, Yang L, Xiang D (2012) High temperature displacement and strain measure-
ment using a monochromatic light illuminated stereo digital image correlation system. Meas
Sci Technol 23:125603
19. Cooper MA, Skaggs MN, Reu PL (2016) High-speed stereomicroscope digital image correla-
tion of rupture disc behavior. In: Advancement of optical methods in experimental mechanics,
vol 3. Springer, Cham, pp 19–26
20. D’Emilia G, Razzè L, Zappa E (2013) Uncertainty analysis of high frequency image-based
vibration measurements. Measurement 46:2630–2637. https://doi.org/10.1016/j.measurement.
2013.04.075
21. dos Santos F, Peeters B, Lau J, Desmet W, Góes L (2014) An overview of experimental strain-
based modal analysis methods. In: Proceedings of the international conference on noise and
vibration engineering (ISMA), Leuven
22. Ehrhardt DA, Yang S, Beberniss TJ, Allen MS (2014) Mode shape comparison using
continuous-scan laser Doppler vibrometry and high speed 3D digital image correlation. In:
Special topics in structural dynamics, vol 6. Springer, Cham, pp 321–331
23. Gagliardi FC, Cunningham B (2010) The use of digital image correlation in explosive
experiments. In: 14th Annual detonation symposium, Coeur d’Alene, 11–16 April 2010
24. Gilbert MG, Welch SS, Pappa RS, Demeo ME (1997) STS-74/Mir photogrammetric
appendage structural dynamics experiment preliminary data analysis. In: Proceedings of
the 1997 38th AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics, and materials
conference. Part 4 (of 4), 7–10 April 1997. AIAA, Kissimmee, pp 566–576
25. Ha NS, Vang HM, Goo NS (2015) Modal analysis using digital image correlation technique:
an application to artificial wing mimicking beetle’s hind wing. Exp Mech 55:989–998. https://
doi.org/10.1007/s11340-015-9987-2
26. Helfrick MN, Pingle P, Niezrecki C, Avitabile P (2009) Optical non-contacting vibration
measurement of rotating turbine blades. In: 27th Conference and exposition on structural
dynamics 2009, IMAC XXVII, 9–12 February 2009. Springer/Society for Experimental
Mechanics (SEM), New York
27. Helfrick MN, Niezrecki C, Avitabile P, Schmidt T (2011) 3D digital image correlation methods
for full-field vibration measurement. Mech Syst Signal Process 25:917–927. https://doi.org/10.
1016/j.ymssp.2010.08.013
28. IDIC (2018) iDIC good practice guide. International Digital Image Correlation Society
website. www.idics.org. Accessed 2018
29. Ind P (2004) The non-intrusive modal testing of delicate and critical structures. PhD thesis,
Imperial College of Science, Technology & Medicine, University of London, London
30. IRIG (2018a) Inter-Range Instrumentation Group. https://en.wikipedia.org/wiki/Inter-Range_
Instrumentation_Group. Accessed 2018
31. IRIG (2018b) Time code. Accessed 2018. https://en.wikipedia.org/wiki/IRIG_timecode
32. Jones EMC, Reu PL (2017) Distortion of digital image correlation (DIC) displacements and
strains from heat waves. Exp Mech. https://doi.org/10.1007/s11340-017-0354-3
33. Kammer DC (Mar 1991) Sensor placement for on-orbit modal identification and correlation of
large space structures. J Guid Control Dyn 14:251–259
34. LePage WS, Daly SH, Shaw JA (2016) Cross polarization for improved digital image
correlation. Exp Mech 56:1–17
35. Lundstrom T, Baqersad J, Niezrecki C, Avitabile P (2012) Using high-speed stereophotogram-
metry techniques to extract shape information from wind turbine/rotor operating data. In: 30th
IMAC, a conference on structural dynamics, 30 January 2012–2 February 2012. Springer New
York/Jacksonville, pp 269–275. https://doi.org/10.1007/978-1-4614-2419-2_26
36. Lundstrom T, Niezrecki C, Avitabile P (2013) Appropriate rigid body correction of stereopho-
togrammetry measurements made on rotating systems. Exp Tech. https://doi.org/10.1111/ext.
12030
37. Lundstrom T, Baqersad J, Niezrecki C (2015) Monitoring the dynamics of a helicopter main
rotor with high-speed stereophotogrammetry. Exp Tech. https://doi.org/10.1111/ext.12127
476 C. Niezrecki et al.

38. Luo PF, Chao YJ, Sutton MA, Peters WH III (1993) Accurate measurement of three-
dimensional deformations in deformable and rigid bodies using computer vision. Exp Mech
33:123–132
39. Meng LB, Jin GC, Yao XF (2007) Application of iteration and finite element smoothing
technique for displacement and strain measurement of digital speckle correlation. Opt Lasers
Eng 45:57–63. https://doi.org/10.1016/j.optlaseng.2006.04.012
40. Miller TJ, Schreier HW, Reu P (2007) High-speed DIC data analysis from a shaking camera
system. Society for Experimental Mechanics, Springfield
41. Mitchell LD (1982) Improved methods for the FFT calculation of the frequency response
function. J Mech Des 104:277–279
42. Mitchell LD, Deel JC (1987) An unbiased frequency response function estimator. In Proceed-
ings of the 5th international modal analysis conference, London
43. Mohanty P, Rixen DJ (2004) Operational modal analysis in the presence of harmonic
excitation. J Sound Vib 270:93–109
44. Moulart R, Pierron F, Hallett SR, Wisnom MR (2011) Full-field strain measurement and
identification of composites moduli at high strain rate with the virtual fields method. Exp Mech
51(4):509–536
45. Niezrecki C, Avitable P, Warren C, Pingle P, Helfrick M (2010) A review of digital image
correlation applied to structural dynamics. In: Proceedings of the 9th international conference
on vibration measurements by laser and noncontact techniques, Ancona, 22–25 June 2010
46. Nonis C, Niezrecki C, Yu T, Ahmed S, Su C, Schmidt T (2013) Implementation of digital
image correlation for structural health monitoring of bridges. In: Proceedings of 9th inter-
national workshop on structural health monitoring, Stanford/Palo Alto, 10–12 September
2013
47. O’Callahan J, Avitabile P, Riemer R (1989) System equivalent reduction expansion process
(SEREP). In: Proceedings of the 7th international modal analysis conference, Union College
Schnectady, pp 29–37
48. Ozbek M, Rixen DJ (2013) Operational modal analysis of a 2.5 MW wind turbine using optical
measurement techniques and strain gauges. Wind Energy 16:367–381. https://doi.org/10.1002/
we.1493
49. Ozbek M, Rixen DJ, Erne O, Sanow G (2010) Feasibility of monitoring large wind turbines
using photogrammetry. Energy 35:4802–4811. https://doi.org/10.1016/j.energy.2010.09.008
50. Pan B, Asundi A, Xie H, Gao J (2009) Digital image correlation using iterative least squares
and pointwise least squares for displacement field and strain field measurements. Opt Lasers
Eng 47:865–874. https://doi.org/10.1016/j.optlaseng.2008.10.014
51. Pan B, Wu D, Xia Y (2012) An active imaging digital image correlation method for
deformation measurement insensitive to ambient light. Opt Laser Technol 44:204–209. https://
doi.org/10.1016/j.optlastec.2011.06.019
52. Peters WH, Ranson WF (1982) Digital imaging techniques in experimental stress analysis. Opt
Eng 21:213427. https://doi.org/10.1117/12.7972925
53. Poozesh P, Baqersad J, Niezrecki C, Avitabile P, Harvey E, Yarala R (2016) Large-area
photogrammetry based testing of wind turbine blades. Mech Syst Signal Process. https://doi.
org/10.1016/j.ymssp.2016.07.021
54. Reu P (2012) Stereo-rig design: creating the stereo-rig layout – part 1. Exp Tech 36:3–4. https://
doi.org/10.1111/j.1747-1567.2012.00871.x
55. Reu P (2013a) Stereo-rig design: lighting – part 5. Exp Tech 37(3):1–2
56. Reu P (2013b) Stereo-rig design: stereo-angle selection – part 4. Exp Tech 37:1–2. https://doi.
org/10.1111/ext.12006
57. Reu P (2013c) A study of the influence of calibration uncertainty on the global uncertainty for
digital image correlation using a Monte Carlo approach. Exp Mech 53(9):1661–1680
58. Reu P (2014a) All about speckles: aliasing. Exp Tech 38(5):1–3
59. Reu P (2014b) Speckles and their relationship to the digital camera. Exp Tech 38(4):1–2
60. Reu P (2014c) All about speckles: speckle size measurement. Exp Tech 38(6):1–2
61. Reu P (2015a) All about speckles: contrast. Exp Tech 39(1):1–2
8 DIC and Photogrammetry for Structural Dynamic Analysis and High-Speed. . . 477

62. Reu P (2015b) All about speckles: edge sharpness. Exp Tech 39:1–2. https://doi.org/10.1111/
ext.12139
63. Reu P (2015c) Virtual strain gage size study. Exp Tech 39:1–3. https://doi.org/10.1111/ext.
12172
64. Reu P (2015d) All about speckles: speckle density. Exp Tech 39(3):1–2
65. Reu P (2016) A realistic error budget for two dimension digital image correlation. In: Jin H
et al (eds) Advancement of optical methods in experimental mechanics, vol 3. Proceedings
of the 2015 annual conference on experimental and applied mechanics. Springer International
Publishing, Cham, pp 189–193
66. Reu PL, Miller TJ (2008) The application of high-speed digital image correlation. J Strain Anal
Eng Des 43(8):673–688
67. Reu P, Miller TJ (2009) Synchronization errors in high-speed digital image correlation. Society
for Experimental Mechanics, Bethel
68. Reu PL, Rohe DP, Jacobs LD (2017) Comparison of DIC and LDV for practical vibration and
modal measurements. Mech Syst Signal Process 86:2–16
69. Sabato A, Niezrecki C (2017) Feasibility of digital image correlation for railroad tie inspection
and ballast support assessment. Measurement 103. https://doi.org/10.1016/j.measurement.
2017.02.024
70. Schmidt T, Tyson J (2009) 3D and 2D high speed image correlation for dynamic testing.
Society for Experimental Mechanics Annual Meeting, Albuquerque
71. Schmidt T, Tyson J, Galanulis K, Revilock D, Melis M (2005) Full-field dynamic deformation
and strain measurements using high-speed digital cameras. In: 26th International congress on
high speed photography and photonics, vol 5580, pp 174–185
72. Siebert T, Wang WZ, Mottershead JE, Pipino A (2011) Application of high speed image
correlation for measurement of mode shapes of a car bonnet. Appl Mech Mater 70:45–50
73. Son K-S, Jeon H-S, Park J-H, Park JW (2015) Vibration displacement measurement technology
for cylindrical structures using camera images. Nucl Eng Technol 47:488–499. https://doi.org/
10.1016/j.net.2015.01.011
74. Spranghers K, Vasilakos I, Lecompte D, Sol H, Vantomme J (2012) Full-field deformation
measurements of aluminum plates under free air blast loading. Exp Mech 52(9):1371–1384
75. Stasicki B, Boden F (2009) Application of high-speed videography for in-flight deformation
measurements of aircraft propellers. In: 28th International congress on high-speed imaging and
photonics, 9–14 November 2008. SPIE, Canberra. https://doi.org/10.1117/12.822046
76. Stasicki B, Boden F (2015) In-flight measurements of aircraft propeller deformation by means
of an autarkic fast rotating imaging system. In: International conference on experimental
mechanics 2014, ICEM 2014, 15–17 November 2014. SPIE, Singapore. https://doi.org/10.
1117/12.2081393
77. Surrel Y, Fournier N, Grédiac M, Paris P-A (1999) Phase-stepped deflectometry applied
to shape measurement of bent plates. Exp Mech 39(1):66–70. https://doi.org/10.1007/
BF02329303
78. Sutton MA (2008) Digital image correlation for shape and deformation measurements, Chapter
20. In: Handbook of experimental solid mechanics. Springer Science+Business Media, New
York, pp 565–600. ISBN: 978-0-387-26883-5
79. Sutton MA, Li N, Garcia D, Cornille N, Orteu JJ, McNeill SR, Schreier HW, Li X (2006)
Metrology in a scanning electron microscope: theoretical developments and experimental
validation. Meas Sci Technol 17:2613
80. Sutton MA, Yan JH, Tiwari V, Schreier HW, Orteu JJ (2008) The effect of out-of-plane motion
on 2D and 3D digital image correlation measurements. Opt Lasers Eng 46(10):746–757
81. Sutton MA, Orteu JJ, Schreier H (2009) Image correlation for shape, motion and deformation
measurements: basic concepts, theory and applications. Springer Science+Business Media,
New York
82. Tiwari V, Sutton MA, McNeill SR (2007) Assessment of high speed imaging systems for
2D and 3D deformation measurements: methodology development and validation. Exp Mech
47(4):561–579
478 C. Niezrecki et al.

83. Tiwari V, Sutton MA, McNeill SR, Xu S, Deng X, Fourney WL, Bretall D (2009) Application
of 3D image correlation for full-field transient plate deformation measurements during blast
loading. Int J Impact Eng 36(6):862–874
84. Triggs B, McLauchlan PF, Hartley RI, Fitzgibbon AW (1999) Bundle adjustment – a modern
synthesis. In: International workshop on vision algorithms. Springer, London, pp 298–372
85. Valley MT, Shields RW, Reed JM (2004) TrackEye tracking algorithm characterization. In:
Target-in-the-loop: atmospheric tracking, imaging, and compensation. International Society
for Optics and Photonics, pp 179–189
86. Voormeeren SN, van der Valk PLC, Rixen DJ (2011) Generalized method for assembly and
reduction of component models for dynamic substructuring. AIAA J 49(5):1010–1020
87. Wang W, Mottershead JE, Siebert T, Pipino A (2012) Frequency response functions of shape
features from full-field vibration measurements using digital image correlation. Mech Syst
Signal Process 28:333–347
88. Warren C, Niezrecki C, Avitabile P, Pingle P (2011) Comparison of FRF measurements and
mode shapes determined using optically image based, laser, and accelerometer measurements.
Mech Syst Signal Process 25:2191–2202. https://doi.org/10.1016/j.ymssp.2011.01.018
89. Wicks A, Vold H (1986) The Hs frequency response function estimator. In Proceedings of the
4th international modal analysis conference
90. Zappa E, Mazzoleni P, Matinmanesh A (2014a) Uncertainty assessment of digital image
correlation method in dynamic applications. Opt Lasers Eng 56:140–151. https://doi.org/10.
1016/j.optlaseng.2013.12.016
91. Zappa E, Matinmanesh A, Mazzoleni P (2014b) Evaluation and improvement of digital image
correlation uncertainty in dynamic conditions. Opt Lasers Eng 59:82–92. https://doi.org/10.
1016/j.optlaseng.2014.03.007
 Part II
Modal Model Development
Design of Modal Tests
9
Thomas Carne, Ralph Brillhart, Daniel Kammer,
and Kevin Napolitano

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
2 Excitation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
2.1 General Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
2.2 Artificial Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
2.3 Natural or Operational Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
3 Response Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
3.1 Transducers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
3.2 Number of Degrees of Freedom to Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
3.3 Acquisition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
3.4 Geometry Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
4 Support Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
4.1 Approximating Free Boundary Conditions, and the Resulting Compromises . . . . . 505
4.2 Suspension System Design, Low Spring Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
4.3 Constrained Support, Built-In, and Other Boundary Conditions . . . . . . . . . . . . . . . . 512
4.4 Operating Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
5 Measurement Quality Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
6 Modal Tests for Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
6.1 Selecting Response Locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
6.2 Selecting Input Locations, Directions, and Number . . . . . . . . . . . . . . . . . . . . . . . . . . 526
6.3 Planning the Criteria for “Test Exit” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
7 Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529

T. Carne ()
Sandia National Laboratories, Retired, Albuquerque, NM, USA
e-mail: [email protected]
R. Brillhart · K. Napolitano
ATA Engineering, Inc., San Diego, CA, USA
e-mail: [email protected]; [email protected]
D. Kammer
University of Wisconsin – Madison, Madison, WI, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 481

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_11
482 T. Carne et al.

Abstract

This chapter examines a number of issues that require consideration when a

modal test is being planned or designed. As with any engineering procedure, a
modal test needs to be designed; otherwise, objectives may not be fulfilled or time
and effort may be poorly used. The issues discussed in this chapter include the
purpose of the test, excitation considerations, response measurements, support
conditions, measurement quality criteria, and considerations for model valida-
tion. When a modal test is to be performed to validate a finite element model,
one needs to design the test so that the resulting measurements will provide the
data required for the correlation of modeling results with those from the test.
From a correlation perspective, one would like to select the response locations to
allow a definitive, one-to-one correspondence between the measured modes and
the predicted modes. Further, the excitation must be designed to excite all the
modes of interest at a sufficient level so that the modal estimation algorithms can
accurately extract the modal parameters.

Keywords

Test design · Test plan · Support conditions · Free boundary conditions ·

Model validation · Modal kinetic energy · Effective independence ·
Min-Mac · Cross-orthogonality · TAMs

Acronyms

DOFs Degrees of freedom

FFT Fast Fourier transform
ω Frequency (radian/sec)
FEM Finite element model
m Mass element of a 2 DOF system
kt True stiffness of structure
ks Stiffness of the support
ct True damping coefficient
cs Damping coefficient of the support
cm Measured damping coefficient
ωt True frequency of structure
ωs Frequency due to supports only
ωm Measured frequency of structure
ζt True modal damping
ζs Modal damping of support system
ζm Measured modal damping
γt True structural modal damping
γs Structural damping of supports
γm Measured structural modal damping
ω Change in frequency due to support
φj value of mass-normalized mode shape at position j
ψ Mode shape of interest
Cs Damping matrix of support system
TAM Test-analysis model
φi The i th mass-normalized mode shape
9 Design of Modal Tests 483

keij The contribution of the j th DOF to the i th mode generalized mass

KE The fractional importance of each FEM DOF to the modal mass
kei Column vector containing contribution of DOF j to the kinetic energy of the i th
mode
EfI Effective independence sensor selection
ys Response at sensor locations
φ fs Matrix of FEM target mode shapes partitioned to sensor DOFs
q Target modal response
w The sensor noise
Q Fisher information matrix, FIM
G A matrix whose columns sum to the eigenvalues of FIM Q
Ψ Eigenvectors
λ Eigenvalue matrix
FE Matrix representing fractional contribution of i th sensor to the j th information
matrix eigenvalue
ED Effective independence distribution
{1}n A column vector of 1’s length n
Min-Mac Minimizing MACs-based sensor selection
MACij Modal assurance criterion between mode shape vectors i and j
ϕi Mode shape vector for i th mode
Φ Mode shape matrix, existing set of sensors
ψ Mode shape matrix, remaining set of sensors
aij Elements of A = Φ T Φ
(MACij )+k MAC value between modes i and j with added DOF k
O12 Orthogonality between modes 1 and 2

1 Introduction

One might ask, why would one design a test? You can design hardware or a
manufacturing process, but you just perform a test. However, a test costs time and
resources, just as any other process or hardware, and there are expectations of the
information that will be measured or estimated from the test data, so the test requires
a similar level of attention. Obviously, if the test will be a one-hour procedure to
measure the frequency of the first mode, there may be little reason to spend much
time on test planning or design. But suppose you are testing a satellite whose current
value is 50 million dollars, and every day spent testing is costing the project 25,000
dollars. Further, if the measured results will be used to validate or modify the finite
model that will be used to validate coupled loads, the design of the modal test clearly
requires careful consideration taking into account the test issues discussed in this
chapter.
In designing a modal test, the first issue that must be clarified is the purpose of
the test. There are many possible purposes – such as model validation, diagnosis,
troubleshooting, structural modification, comparison with operating conditions,
design input, or control system evaluation – and one particular design will not
satisfy them all. Additionally, the test design from a previous program may not
necessarily fulfill the requirements for a new and different purpose. The second issue
is identifying who has a vested interest in the results of the test. This may not be at
all clear when the modal test is first conceived. The finite element modeler, controls
484 T. Carne et al.

person, vibration test engineer, hardware owner, management personnel – all of

these individuals may have preconceived ideas about what results the modal test
will provide. To be successful, one needs a clear understanding of both the purpose
of the test and the expectations of interested individuals. With a full understanding
of all the desired outputs, one can think about the implications for the test design.
These might include the number of modes to be measured, the frequency range of
interest, what excitation is best, and what response measurements are required.
Of course, there are always constraints on any engineering process, including
the costs, human resource time, test article availability, modal analysis tools, data
acquisition hardware, and so on. However, with a well-conceived test design, the
various constraints and conflicts can be evaluated based on the expected output of
the test. Professor Carlos Ventura of The University of British Columbia proposed a
mnemonic for the key elements of test design: SMART, for Specific, Measurable
objectives, Agreed, Realistic, and Time constraint. These five items comprise a
general list of aspects to be considered; the more specific considerations and some
of their consequences are addressed more thoroughly in sections below.

2 Excitation Techniques

Just as there are many types of disturbances or excitations of structures in nature,

there are a number of types of excitations that can, and should, be considered
in designing a modal survey of a structure. Historically, early modal testing was
conducted with periodic, harmonic (sinusoidal) excitation, which led to observations
that were readily understood. However, as modal testing methods evolved, it has
become clear that there are various different excitation methods that lend themselves
to excellent modal tests without any sacrifice in the clear definition of the modal
behavior. As a result, modal test design should include an appropriate assessment
of the type or multiple types of excitation to be employed for the best chance
of success. The following discussion provides insight into excitation technique
selection and how it is important for developing a successful modal test design.

2.1 General Considerations

In selecting an appropriate excitation type, early consideration should be given

to the structure that will be tested and the overriding objectives of the test. The
ultimate data quality from the modal test is highly dependent on the excitation
approach used. Keep in mind that it is common that a single excitation technique
will not be adequate or sufficient, and the use of multiple types of excitation can
provide better insight into the structural behavior. This is particularly true when the
structure will be subjected to different excitation environments in service. As such,
just as qualification testing of structures might employ random vibration, sinusoidal
vibration, and shock environments, modal testing can often be most effectively
conducted using a combination of excitation types.
9 Design of Modal Tests 485

2.1.1 Frequency Range

The modal test frequency range of interest is typically a parameter that is quantified
at the early stages of modal test preparation. It is often selected based on a finite
element model (FEM) which is used to predict the natural frequencies and mode
shapes of interest. In addition, the expected excitation environment can be used to
help in selecting an appropriate modal test excitation type and frequency range. In
preparation for any modal test, there should be accommodations to allow frequency
range adjustments as needed during the test once data becomes available. Initial
investigations with an expanded frequency range may be conducted, followed by
detailed testing after frequency range adjustments are made.
Launch vehicles are typically exposed to random (both vibration and acoustic
environments) and some harmonic excitation sources. Similarly, satellites and
launch vehicle payloads are exposed to the launch vehicle environment types as
well as harmonic tones once they are “on-orbit.” Automotive vehicles experience a
variety of harmonic inputs as well as random and shock environments transmitted
through their suspensions.
The modal test frequency range of interest is often the result of an evaluation
of which modes of the structure have “significant” modal mass, which in turn
can produce loads on the structure or its interfaces which will be important in the
operational environment.
In the aerospace industry, slightly different objectives can be used to select the
frequency range of interest. For aircraft, the modes that can develop significant
dynamic loads as well as those that may result in dynamic instability, or flutter,
define the frequency range of interest. For large aircraft, this frequency range can
be quite low – often below 30 Hz. Modest-sized aircraft typically exhibit higher
frequencies, and the range of interest might well be above 100 Hz. Drones and other
smaller aircraft may yield frequencies of interest upward of 500 Hz. Very small
components (engine turbine blades, for example) may have frequencies of interest
in the kilohertz range. Giving appropriate consideration to actual environments, per-
formance behavior required, and possible model predictions will lead to successful
modal testing.

2.1.2 Excitation Level

The next consideration involves identifying the proper excitation level to make sure
that the modes of the structure will be well excited and that all modes of interest
can be characterized. The magnitude of the excitation required in a modal test can
be affected by the amount of damping or energy dissipation in the structure and can
also be influenced greatly by the size of the test structure and the number and types
of excitation sources. Large structures with light or modest damping might be easily
excited using only modest excitation levels; for example, wind turbine blades and
towers, which typically are lightly damped, can be excited with only a few hundred
pounds of excitation force.
Of course, the excitation type and frequency range of interest also affects the
amount of force required. Impulsive excitation such as step relaxation can focus
the frequency content at low frequencies, and peak forces may be thousands of
486 T. Carne et al.

Fig. 1 Good shaker location selection is important for successful modal testing

pounds. On the other hand, impulsive excitation using small load-cell-instrumented

hammers can provide frequency content covering high frequencies using small to
large force levels.
Excitation level requirements are strongly influenced by the excitation location
on the structure and the direction. To excite a mode, locations that have good modal
participation need to be selected. As seen in Fig. 1, during testing of an H-frame,
shaker 1 and shaker 2 do a good job of exciting both modes of interest, but shaker
3 only excites one of the two modes. This does not mean that locations of high
modal displacement need to be selected; locations with lower response can be used
effectively by increasing the excitation input level, as seen in Fig. 2.
The excitation levels can also be described differently depending on the exci-
tation type [14]. Peak force might be used to assess impact and step relaxation,
whereas peak and root-mean-square (RMS) might be used to describe the input
from random and sine excitation signals.
Excitation levels can be constrained in a modal test by the structural limitations of
the test article, but damage to the test article can be avoided by selecting structural
locations that can withstand a wide range of excitation levels, or excitation level
limits might be imposed to ensure that no damage is done. Mechanical fuses and
software limit controls might be implemented to limit the maximum excitation
levels. Hardware can also be added to the structure to spread the load over a larger
9 Design of Modal Tests 487

Fig. 2 Excitation at highest modal displacement locations is not required

area and lower the local pressure to reduce the chance of damage. Peak force levels
that occur in impact testing may need to be evaluated to determine whether there is
any likelihood of local structural damage.

2.1.3 Linearity of Structure

Many modal tests require an investigation of linearity . These linearity tests require a
range of excitation levels to be applied. Multiple excitation levels might be applied
as a matter of course in any modal test so that the test program can define how
linearly the structure behaves within the excitation range applied. Linearity can
of course be dependent on the excitation level and type – almost any structure
will exhibit nonlinearity at significant enough excitation levels. The intent in the
modal test is to provide engineering information to help characterize this so that the
importance of any nonlinearity can be determined and computer modeling efforts
can take this information into account. In almost all modal test programs, the
linearity should be investigated using multiple levels or multiple types of excitation
with the intent of observing how the behavior changes.
If the modal test article is highly nonlinear, the test should investigate this to
make sure that the behavior is well documented. An example of this is an aircraft
control surface, which is expected to have a nonlinear response. Characterization
of this behavior is important to fully understand how the system will behave in
operation and is critical to making control system interaction and flutter predictions.
Controlled tests using excitation level variations should be used in these cases to
document the type of nonlinear behavior that can be expected. Excitation levels
required for this type of testing should be evaluated prior to the test to ensure
488 T. Carne et al.

that expected levels can be achieved and that the test structure can be adequately
exercised. Automotive vehicle suspensions also typically exhibit a high degree of
nonlinear behavior due to the types of springs or geometric behavior and dampers.

2.1.4 Damping of Structure

Structural damping has a key influence on the excitation levels required in the modal
test program. The damping may also vary with excitation level, so using variable
excitation levels can help define this effect. The type of damping present will
influence the behavior observed. Most modal evaluations treat damping as though it
is distributed and can be characterized as equivalent viscous damping; since this is
not always the case, a series of tests can be planned to provide a better understanding
of the behavior. Damping may be distributed or discrete, viscous or hysteretic, and
targeted modal testing can help evaluate the damping characteristics.
During selection of shaker locations using a simulation study, a parametric study
of estimated damping levels should be applied to allow reasonable estimation of the
required test excitation levels. The expected damping levels should be selected from
historical data when possible; similar structures provide reasonable damping ranges
that can be used in this process. When similar structural data is not available, a range
of damping values can be used to perform a parametric evaluation.

2.1.5 Simulation of Operational Loads

The types (random, sinusoidal, transient) of operational loads should be used in
the appropriate selection and definition of the excitation used for a modal test. In
some cases, the operational loading environment or an approximation can, and
should, be used in the modal test; the magnitude of the environment may limit
the actual excitation selected, but the spectrum applied can mimic the operational
environment. This is clearly not always the case, as an operating environment
may be discrete frequency tones, and modal testing is used to quantify the overall
structural dynamic characteristics. The modal test engineer must therefore use
judgment in selecting whether the operational environment is applicable to the
modal test.
Shaping the frequency spectrum to more closely match the operating conditions
can be very useful in obtaining modal properties that closely match what will be
observed in service, as it can allow the magnitude of the excitation levels to be
adjusted to higher levels in frequency bands than might be achieved if a flat force
spectrum were applied.

2.2 Artificial Input

In discussing the excitation applied to a system to generate responses, a distinction

should be made between artificial inputs and natural or operational inputs. Natural
inputs are generally considered those that occur in nature; for example, they could
be wind, seismic, or traffic. Operational inputs are the excitation resulting from
operating the system. An example of operational inputs could be aerodynamic
9 Design of Modal Tests 489

input on a flying aircraft or input from a rocket motor burning its fuel. Natural
and operational inputs can be somewhat related. In contrast, artificial inputs are
specifically designed and controlled excitations applied to the structure to measure
the resulting responses and then extract information about the structure’s modal
parameters. Examples include a random force applied by an attached shaker, an
acoustic input from speakers, or a short transient force. These inputs and their
advantages and disadvantages for use in modal testing are discussed in the following
sections, starting with a discussion of artificial inputs.

2.2.1 Impulsive Inputs

Impulsive or transient excitation is an input of short duration relative to the measured
time record, in contrast to random or sine inputs. The versatility of transient
excitation techniques provides several advantages over typical vibration shaker
input: it enables quick diagnostics of structures with short setup times, the input
frequency spectrum is generally flat and can range from less than ten hertz to
a million hertz, and high-amplitude inputs are achievable as well. Finally, the
scalability of transient excitation is quite good; tested structures can range from
a fraction of a centimeter to hundreds of meters. Impulsive excitation can consist of
much more than hammer impacts, although that is by far the most common approach
to apply a short-duration input: the excitation could also include projectile impacts,
explosive inputs, step relaxation, or many other short-duration inputs. For a more
complete description of various transient excitations, see Carne and Stasiunas [9].

Impact Inputs
The most commonly used method of transient excitation for modal testing is the
impact hammer. The idea of exciting a structure with an impact hammer is actually
very simple: one strikes the structure at a particular location and in a particular
direction with an impact hammer. The impact hammer, which is instrumented with
a force transducer located behind the tip, measures the force used to excite the
structure. Figure 3 shows a small impact hammer with an integral load cell being
used to apply a transient forcing function to a test article. The force input and
corresponding responses are then used to compute the frequency response functions
(FRFs) [7, 24]. The response to a hammer impact is an approximation of the impulse
response function, although in actual applications, the impact is not assumed to
be instantaneous but is measured and transformed into the frequency domain and
used to compute the FRFs. An example force transient and corresponding frequency
content is shown in Fig. 4.
The most important issue when utilizing impact hammer excitation is the choice
of hammer mass and tip stiffness. These parameters determine the impact duration,
which consequently determines the frequency content of the input. When impacting
the test structure, the input frequency content should be sufficient to excite only
the modes in the frequency range of interest. Exciting the structure above the
frequency range of interest should be avoided, as the undesired response will fill the
dynamic range of the measurement system and more than likely obscure the desired
490 T. Carne et al.

Fig. 3 Impact hammer with

integral load cell to measure
the applied force

Fig. 4 Impact hammer force transient (upper) and spectral content (lower)

information. Therefore, an estimate of the desired frequency range of interest is

important in the test design process.
Hammer impact testing has some distinct advantages. The input spectrum from
the impact is flat out to the roll-off frequency, typically with no holes in the
spectrum. The technique can be very efficient and portable compared to aligning
and moving shakers and their associated control systems. Relatively small numbers
of averages – as low as three to six – are sufficient to reduce noise in the calculated
FRFs. Finally, for lightly damped, linear test structures, hammer impact testing can
result in very high-quality FRFs.
9 Design of Modal Tests 491

Step Inputs
Step relaxation is a very powerful, but seldom used, excitation technique for formal
modal analysis, although it is used extensively for informal tests. A step relaxation
input is basically a pluck to the structure: an initial deformation is enforced upon
a test structure and then rapidly released. Every time a guitar string or cantilever
beam is plucked, the step relaxation technique is performed. In contrast to impact
excitations, which have very flat Fourier spectra out to their roll-off frequency, the
Fourier spectrum of the step input rolls off similar to 1/ω. Consequently, step input
is extremely well suited to use with low-frequency systems [9].
Applying a step input to a structure is simple in design. Typically, a wire or
string is attached to the test structure, with the free end anchored to the ground
and tensioned, enforcing the required deformation. An example schematic of such
an arrangement is shown in Fig. 5, and a simple gravity load release mechanism
is shown in Fig. 6. Once preloaded, the wire is abruptly released or cut, resulting
in step input excitation. In a test of a wind turbine, a 2-cm-diameter steel cable
was used to provide a deformation to a 110-meter-tall wind turbine. For release,
explosive cutters were used on the cable [10].
One requirement of step relaxation is the use of a force transducer as an element
of the mechanism that applies the enforced deformation. This input force must
be measured to compute the input–output FRF. Furthermore, the transducer needs
to have a frequency response down to DC if it is to monitor the applied static
force. A major advantage of step relaxation is that it is basically a noncontact
excitation method; after the enforced deformation is released, there is no attachment

Fig. 5 Step relaxation hardware arrangement

492 T. Carne et al.

Fig. 6 Gravity loaded quick release step relaxation system

to the excitation system. Step relaxation also scales very easily to large or small
structures.
Applying step relaxation results in a unique signal processing issue. The force
from a step input is ideally a classic step function, with a constant force beginning
in the past and then suddenly dropping to zero within the test acquisition window as
seen in Fig. 7. The data processing issue lies in using the Fast Fourier transform
(FFT) to process the input-force time-domain signal: to the FFT, the step input
appears to be a square wave, which results in very large holes or zeros in the
input spectrum as seen in Fig. 8. This situation makes the original input step force
unsuitable for typical FRF calculation. This signal processing issue can be solved
easily but requires a bit of forethought: if both the original step input and the
resulting responses are passed through matched high-pass filters with a low cut-
off frequency, perhaps at 0.3 Hz (e.g., using an AC coupling circuit), the input
step function converts to a negative spike, similar to an impact test as seen in
Fig. 9. The resulting FFT, seen in Fig. 10, is very well behaved and is consequently
easy to use for FRF computation and the averaging of repeated tests. Despite the
drastic change in the time-domain signal caused by filtering the step, the resulting
frequency content will be unchanged above 0.3 Hz. In fact, this representation of
the force signal is compatible with all signal processing techniques used for impact
testing including force windows, negative exponential windows, and pre-triggering.
9 Design of Modal Tests 493

Fig. 7 Step relaxation force transient

Fig. 8 FFT of step relaxation transient force

494 T. Carne et al.

Fig. 9 Step relaxation transient force after filter application

Fig. 10 FFT of step relaxation transient force after filter application

9 Design of Modal Tests 495

Fig. 11 Step relaxation transient force using piezoelectric load cell transducer

Another approach which does not require additional filtering involves the use of
a piezoelectric load sensor. These sensors do not exhibit the static load and only
yield the transient load when the preload is released. This results in a waveform and
frequency content as shown in Fig. 11, which can be used directly in computing
the desired FRF. Combining both types of load cells (static and dynamic) in the test
design allows for optimal signal processing using this excitation technique.
In summary, step relaxation results in an input spectrum whose magnitude is
proportional to 1/ω, making this technique particularly applicable to low-frequency
structures. As with impact testing, step inputs work best with lightly damped, linear
structures; the input does not work well for structures with significant nonlinearities
or slipping joints. Scalability is possible with step relaxation, enabling excitation for
small or massive structures, but the technique is particularly useful for the testing
of large structures due to the ability to apply significant forces with a pre-tensioned
cable. Finally, the FRFs measured from step relaxation can be of very high quality
with very little noise.

Damping and Nonlinearities

In many cases, impulsive excitation works quite well, resulting in high-quality
measured FRFs. One feature that enables the measurement of quality FRFs is
low damping and inherent linearity. With high damping, the response due to the
impact is quickly diminished, reducing the information available in the response
496 T. Carne et al.

measurement. Highly damped structures can be difficult to test utilizing impacts or

step relaxation, although good results can still be obtained under certain conditions.
Significant structural damping may indicate nonlinearities in the structure as
well. For example, a common cause of high damping in structures is joint interaction
or joint sliding, which is one source of nonlinear behavior in structures. In this case,
the application of impulsive inputs is not advisable. In fact, even small nonlinearities
can result in poor FRF measurements when impulsive inputs are used, as the input
excites the structure initially to high response levels and then the response damps
down to very low levels, resulting in data from multiple structural states all in one
FRF measurement. This can create very confusing data, with the observed modal
frequencies varying within one data sample. For structures with significant nonlinear
behavior, random excitation or sine testing, provided with a shaker, will yield the
best results for the linear model of the system. Random excitation is discussed in a
subsequent section.

2.2.2 Controlled Inputs

The most common type of controlled excitation source is a modal shaker. Usu-
ally modal shakers are voice coil actuators that impart a measured force on a
structure that is proportional to the provided electrical signal. An electromagnetic
(electrodynamic) shaker as shown in Fig. 12 is coupled to the structure using a
mechanical stinger rod and a load cell to measure the applied excitation force.
However, the actuator can also be hydraulic, pneumatic, piezoelectric, or another
actuator type that imparts a force appropriate for the type of structure and objectives
of the test. Selection of the type and size of shaker is dependent on the structure

Fig. 12 Electromagnetic
shaker attached to the test
article with a mechanical
stinger rod
9 Design of Modal Tests 497

size, required excitation force, frequency range of interest, number of excitation

locations, as well as other factors such as power requirements. Inputs can also
be acoustic using speakers (another voice coil actuator which is coupled to the
structure through an acoustic media), or even accelerations such as those provided
by vibration tables. The inputs imparted to the test article can be deterministic
or random. Deterministic signals are those that can be described at any point in
time by a mathematical function [2], whereas random signals cannot. Examples
of deterministic signals include sinusoidal dwells, stepped sine, sinusoidal sweeps,
and pseudo-random. Examples of random signals include burst random, periodic
random, and true random signals.
Stepped-sine excitation is a special case of sinusoidal excitation which dwells
at discrete frequencies for some amount of time while data is collected. The sine
frequency is then stepped to the next frequency. Both the amount of time spent at
each frequency and the frequency spacing can be selected during the test planning
process, and they can also be adjusted during the test. The sinusoidal signal used to
drive the structure is controlled in amplitude and phase by each excitation actuator.
Sinusoidal sweeps, which can consist of a long sweep or a “chirp,” involve
continuously changing the sinusoidal frequency of an input over a frequency range.
The difference between a chirp and a long sweep is the amount of time the sweep
lasts. A chirp is defined here as a sweep that lasts less than a frame of data. A long
sweep is a sweep that lasts over all frames of data.
The total amount of time required using stepped-sine is dependent on the
frequency range being covered, the frequency spacing or resolution, the dwell time,
and the number of excitation phase combinations to be used. These parameters
can be evaluated during test planning to assess the total time required using this
method. In general, stepped-sine excitation is considered a somewhat slow method,
particularly when high frequency resolution is required. However, it can be quite
effective when performing detailed testing covering small frequency ranges and can
provide high-resolution FRFs to allow detailed focus on a particular mode or modes
of interest.

2.2.3 Multiple Inputs

There are two major considerations that should drive the definition of the inputs for
a modal survey:

• The input must be able to excite the modes (at least a subset of modes)
• The combined inputs must be able to uniquely excite all the modes of interest
(this is related to mode independence and force appropriation)

Multiple inputs are typically used in modal testing to ensure that a good
distribution of energy to the system is achieved. It is also important to develop a set
of FRFs that provides either a full or partial matrix which can be used to describe
the structural behavior. Multiple inputs allow development of the multi-reference
FRFs needed for parameter estimation, such as the well-known Polyreference
parameter estimation tools. Multiple inputs are not limited to shakers. Various
498 T. Carne et al.

impact excitation techniques have been employed (Multiple Reference Impact

Testing – MRIT) to develop FRF matrices that can be used in multi-reference
parameter estimation techniques. Further, single-input excitation techniques can
be used (and were for many years) to develop a partial FRF matrix for multi-
reference parameter estimation by simply moving the excitation input to a number
of locations on the structure. However, in most cases – and what is of most concern –
the implementation of multiple, simultaneous inputs to the test article will yield a
more consistent modal characterization and better energy distribution through the
structure.
The number of excitation sources or inputs used in a modal test is often a matter
of convenience as well as availability. However, there are specific considerations
that must be taken into account in determining the number of excitation sources
to be used in a modal test. Single inputs can be used to develop a set of multiple-
input FRFs. However, except in the case of stationary, linear structures, there is
a likelihood that the FRFs will not be consistent due to the changes in time and
physical boundary conditions. It is better to configure all inputs simultaneously so
that a consistent data set is obtained in the test. The capacity of any shakers to be
used and where they can be attached to the structure should be considered as part of
the planning process.
One key consideration in setting the number of excitation sources is whether
the inputs are linearly independent. This is required so that the reference auto-
spectral matrix, SXX , can be inverted in developing the desired FRFs. For random
input signals, the number of frames of test data must be greater than the number of
inputs used. For deterministic input signals, the number of phasing cases between
the different input signals must be greater than or equal to the number of exciters.

2.3 Natural or Operational Inputs

In contrast to the artificial excitations discussed above, there are natural or oper-
ational excitations that can be considered as a viable excitation to a structure for
measuring its modal parameters. In some cases, a structure can be so immense
that it is very difficult to apply an adequate artificial excitation; examples include
large wind turbines, buildings, and bridges. Frequently for structures that have
been assembled in their operational environment, there exists substantial natural
excitation and responses that would interfere with an artificial excitation test. Gen-
erally, natural inputs are those existing in the environment, such as wind, seismic
inputs, and vehicular or foot traffic. Somewhat contrasting to natural inputs are
the operational inputs that result from operating the system or structure. Examples
here would be a rotating wind turbine, a flying aircraft, or a moving vehicle on
a road surface. The two terms – natural and operational inputs – are frequently
interchanged with the currently more common term “operational modal analysis.”
The use of natural inputs for wind turbine testing was first introduced at the
1988 International Modal Analysis Conference (IMAC) [11], and the concept of
operational modal analysis (OMA) followed quickly. Numerous improvements
9 Design of Modal Tests 499

and refinements and the theoretical grounding have been subsequently published.
James et al. [19] discussed the theoretical development of the NExT (Natural
Excitation Technique) approach for using natural inputs for modal analysis. In
that development, it was shown that the cross-correlation functions arising from
broadband random inputs can be represented as a sum of decaying sinusoids of the
same form as the impulse response functions. This was a significant development as
it revealed that any modal extraction algorithm that uses impulse response functions
could be used to extract modal parameters from the cross-correlation functions.
Various time-domain extraction algorithms have been used to estimate the measured
modal parameters, including ERA and Polyreference. In practice, cross-correlations
can be much noisier than FRFs due to the lack of noise cancellation that is obtained
in computing FRFs. Consequently, it is recommended that many more averages be
used than one would use to compute FRFs; 100–200 averages is not an unreasonable
number.
There have been a tremendous number of modal analysis applications using natu-
ral or operational inputs – including wind turbines, bridges, buildings, monuments,
stadiums, aircraft, rockets, vehicles, and other civil structures – and there is even
a biannual conference, the International Operational Modal Analysis Conference
(IOMAC), dedicated to applications and developments in OMA.

3 Response Measurements

3.1 Transducers

To measure structural responses, one clearly requires some sort of transducer that
will create an electrical signal proportional to the response being observed . Ninety
percent of modal tests use accelerometers with specific characteristics chosen
depending on the size, weight, geometry, and other properties of the structure under
test. Other chapters in this handbook discuss in detail various characteristics one
must consider in selecting an accelerometer. However, there are many applications
in which accelerometers are not the best choice for response measurement. In those
cases, one needs to be aware of the other means that can be used to measure
responses.
High-output strain gauges might serve well as a transducer when weight or
temperature is a particular concern. Strain gauges do not have the cross-axis
sensitivity that accelerometers do, and they can be geometrically combined to
sense only axial strain or bending strains. Velocity probes, interferometry devices,
or displacement gauges might also have application for a particular object under
test. Interferometry can be particularly effective for objects of extremely low mass,
such as micromechanical systems, or any object where contact needs to be avoided
or frequencies are extremely high. Additionally, active materials attached to the
structure could act as a strain gauge or a force transducer in some applications.
500 T. Carne et al.

3.2 Number of Degrees of Freedom to Measure

Clearly, when a modal test is designed, the number of response degrees of freedom
(DOFs) needs to be considered. By measuring more responses, one obtains more
detail; however, the number of measured responses also increases the cost and
complexity of the test. These considerations are discussed more in Sects. 3.4 and 6,
which discusses requirements for model validation. However, just because one can
measure 300 responses in a test does not mean that one should. One needs to be
conscious of the costs and time involved in the installation and operation of the
sensors as well.
One aspect that needs to be decided is the number of DOFs to be measured at
each location. For some structures, one DOF is just fine; but others may require
two or three – or one may even need to measure rotations as well as translations.
If a high-fidelity model is available, some responses can be inferred or computed
using the model, such that they do not need to be explicitly measured. Again,
further discussion of the number of DOFs and which DOFs are required for modal
validation is included in Sect. 6.

3.3 Acquisition Methods

Independent of the types of transducers used in conducting the modal test, a key
consideration to be addressed before starting the instrumentation is how the sensors
will be installed and how many will be used or collected at a time. The number of
sensors available for installation can provide an upper limit here, as can the number
of input channels on the data acquisition system. However, potential measurement
effects should be taken into account regardless of the approach taken. There are
primarily four approaches which can be taken when acquiring the modal data.

• Simultaneous collection of all channels with all sensors installed

• Sequential collection of all channels with all sensors installed
• Sequential collection of channels with roving sensors
• Sequential collection of channels with roving input, using reciprocity

In the ideal case, all sensors should be installed and measured simultaneously
using a high-performance, high-channel data acquisition system. This ensures that
a consistent data set is acquired using the same data acquisition conditions and
with the same excitation inputs. Of course, this means that the sensors and data
collection system must be matched for the total number of channels; if the test
involves measuring 100 sensors, the data system used must be able to collect all 100
sensors and the excitation input signals simultaneously. This can be a problem if the
data collection system does not have enough channels to acquire all of the sensors, or
if there are not enough sensors to be installed at every desired measurement location.
This is true whether the sensor types are accelerometers or lasers or some other
9 Design of Modal Tests 501

type. The reason this is so important is that if all test measurements are not made
simultaneously, structural variations may occur and may change the modal behavior.
These changes in the modal behavior then make parameter estimation difficult if not
impossible. For example, an airplane being tested was instrumented with over 100
sensors, but the data system could only collect a subset (16 channels) of these at
one time. Since the sensors had to be collected in subsets, the measurements were
spread out over some time, during which there were thermal effects that caused
the aircraft structure to expand, yielding inconsistent sensor orientation and modal
behavior over time. Even though the excitation input was essentially the same for
each subset of the measurements, the structural changes resulted in a poorly defined
set of FRFs, which made modal parameter extraction much more difficult.
Nonetheless, there are many cases where it is not possible to make all of
the measurements simultaneously, and roving accelerometer and laser vibrometer
measurements are common methods used to make FRF measurements in these
cases. If this approach is to be used, attention must be paid to keeping the test article
as consistent as possible throughout the measurement cycle. If accelerometers
are being moved (roved), mass simulators should be installed at all locations to
represent the physical accelerometer mass at each measurement location. These
masses can be removed as the sensors are installed and then reinstalled when the
sensors are moved to another location. In all cases, even with laser measurements,
the environment around the test article should be well controlled. This includes
temperature and humidity controls as well as things like environmental noise.
This may mean utilizing a special facility to house the test article. The same
considerations should be applied when deciding to use a roving hammer with a
few fixed sensor measurements. With this approach, the roving input is applied at
all of the measurement points of interest and reciprocity of the FRF measurements
is used in development of the FRF matrix. The number of locations to be measured
and thus the amount of time required to do so may have an effect on how consistent
the structure remains during the test.
Many different approaches can be used in collecting the test measurements.
With careful planning of the approach to be used and control of the test article,
the test should yield a consistent set of test data. The amount of time required
for the testing activity will also be strongly influenced by the approach selected. If
multiple test configurations will be required, roving techniques will typically require
considerably more time due to the setup required, even though the initial setup time
may be less. In general, if more than one test configuration will be studied, it will
require less overall time to install all sensors at the beginning of the testing than
would be taken using a roving technique. Of course, the total number of sensors to
be installed will also need to be evaluated.

3.4 Geometry Definition

Assignment of sensors to specific measurement locations on the test article structure

are important to allow mode shape visualization and for comparisons to FEM
502 T. Carne et al.

predictions. This geometric or physical placement of the sensors is an important

step in the planning and setup phase of conducting a modal test. When FEMs are
available, geometry can be extracted from the model. But since these models are
often simplifications of the real structure, the actual geometry may be different. The
test engineer should ask how important this geometry is and determine the amount
of effort that should go into defining specific measurement locations.
Many modal tests are performed without any prior knowledge or geometric
information from FEMs or geometry models. In these cases, the geometry of the
structure can be measured where the sensors are installed. Or geometric assumptions
can be made to define measurement locations and then validated as the sensors
are placed. This is often the case when modal testing is performed as part of a
troubleshooting effort and measurement selection is defined as the testing effort
progresses.

3.4.1 Accuracy of Location Measurements

When measuring where the modal test sensors are to be installed on a structure,
an important question that arises is what accuracy is required for the location
measurements. There is no specific answer to this question, but it is important to
consider the modes of interest and whether the geometry measurement accuracy will
influence the ability to quantify the shapes being measured. A way of demonstrating
this is to consider the measurement of a sinusoidal waveform. If the wavelength
being considered is shorter than the accuracy of the measurement location, then
there will be a high degree of uncertainty associated with the corresponding shape.
Of course, this will vary greatly depending on the structure under test and the
frequency range of interest; accuracy of ±1 inch on a Boeing 747 aircraft would
have much less impact than on a Cessna 172 aircraft.
During preparations for a modal test, the expected mode shape wavelengths and
frequency range should be evaluated to determine whether placement accuracy of
the sensors will have an influence on the accuracy of the mode shapes, as the
answer may influence the method used to make the measurements. Laser systems
are available now which can accurately measure the physical locations of the sensors
(before or after installation). On the other hand, a simple tape measure can enable
good geometry accuracy, provided there are well-defined reference points on the
structure.

3.4.2 Layout and Documentation

Much of the layout for sensor placement can be planned before the test. The
geometry definition, if made available prior to the test, can be used to develop
methods to quickly locate the sensor placement on the structure. There are many
methods that can be used in laying out the measurement locations on the test
article structure. New methods enable accurate measurement of the locations where
sensors will be installed. For instance, a laser tracking system can be used to define
locations that are visible from outside of the test structure; for interior locations,
other methods must be used to properly verify where measurements are to be made.
Structural landmarks can often be used as reference points or physical locations
9 Design of Modal Tests 503

where sensors might be installed; these might be apparent as structural seams, weld
lines, rivet lines, mold lines, and so on. As new methods such as 3-D printing
become more widely used, however, some of these physical attributes that can be
used for geometric definition may become less predominant, and new geometry
definition methods may need to be employed.
If geometry is developed during the test planning stage, this information can be
entered into the laser tracking system with appropriate reference to allow automated
sensor placement. Laser systems can also be used to measure or validate sensor
placement even when prior geometry is not available. Similarly, optical methods
such as photogrammetry can be used to record where sensors are located.
Many test articles can be laid out for the measurement locations using simple tape
measures. Good reference points should be identified from which measurements can
be made. Since sensors are typically placed on the exterior surface of the test article,
the surface geometry is taken to correspond to models or drawings that define the
surface. Measurements along the surface then define the location where sensors are
to be placed.
Templates can also be used as a reliable method to define where measurements
are to be made. Preliminary geometry can be located on the template for accurate
placement, and then the template can be aligned to reference positions on the test
article to show where sensors are to be placed.
As locations are identified, it is a good approach to mark the sensor installation
location before the sensor is installed. Since some surfaces are sensitive and marking
the surface may not be desired, tape can be placed on the surface and the location
marked on the tape. The sensor can then be attached to the tape at the marked
location. It is important to consider the frequency range of interest and the mass
of the sensor being installed when using tape, since the attachment method can
affect the measurements made. An approach that has minimal effect on the expected
measurement should be selected.
The measurements of geometry should be documented showing three-DOF
coordinates in the appropriate coordinate systems. If a local coordinate system
is used, those measurements should be recorded along with any transformed
coordinates, such as global coordinates.

3.4.3 Test Display Model

A test display model is an important tool for conducting a modal test and evaluating
the quality of the data as the test proceeds. This model, in its simplest form, is
a geometric 3-D representation of where the test measurements are being made.
Nodes (which are analogous to geometry grids and finite element nodes) are used to
define the locations on the structure where sensors are located. Connective elements
between the measurement nodes can be made using line traces or by employing var-
ious finite element types (typically linear beams, triangular and quadrilateral shells,
etc.) to yield a 3-D display that is analogous to the test article. As measurements are
made, they are directly associated with the display nodes and directions. The test
display model, in addition to allowing the visualization of the mode shapes, allows
visual comparisons to be made with a FEM as shown in Fig. 13.
504 T. Carne et al.

Fig. 13 A test display model (left) is important for mode shape visualization

The test display model can be expanded beyond the sensor measurement
locations to provide improved mode shape displays and animation. During prepa-
ration for the test, if a FEM is available, model-based transformations should be
considered; these enable the test measurement DOFs to be expanded to other
unmeasured locations, which provide improved visualization of mode shapes. While
it is not a necessary step, this process can provide better interpretation of modal
behavior without changing any of the original measurement data. In some cases, it
is also possible to use a variety of constraint equations, such as rigid body behavior,
to compute shape behavior where measurements were not made in the test. Care
should be taken when using these approaches, however, as the physical structure
may behave differently than the assumed constraint equations.
In the extreme case, modal results from test can simply be expanded to the full
FEM, which can then be used for visualization. While this might be done for some
simpler models, detailed FEMs can require significant time to convert and display,
thus limiting the effectiveness of this approach for evaluation of test results during
the performance of the test. Also, any errors in the model can lead to distorted mode
shape information in the expansion of the test data. It may prove more effective
to use this method at the completion of the testing while model updates are being
performed.

4 Support Conditions

There is a maxim used in the aerospace industry which applies to many types
of testing, and certainly to modal testing: “Test Like You Fly.” This has many
meanings, but in modal testing, it is meant to imply that the test article should
be representative of the flight hardware and the boundary condition should be
representative of the way the article will be constrained when it is in service.
Interestingly, this can lead to multiple tests, each with different boundary conditions:
a satellite has a constrained boundary condition when it is being launched and
delivered to orbit, yet it has a free boundary condition when it is in service.
9 Design of Modal Tests 505

However, most satellite modal tests are conducted with a fixed or constrained
boundary condition, since that is the condition for which large loads can develop
at the boundary as experienced during launch. Similarly, components mounted on a
satellite have a constrained boundary condition and would typically be tested with
a fixed condition boundary.
On the other hand, aircraft and land vehicles are operated in either a free or a
softly suspended condition, and modal tests conducted in that configuration should
be considered first. Aircraft are most often tested with low-frequency suspensions
intended to approximate a free-flight boundary condition and isolate the aircraft
rigid body modes from the aircraft flexible modes. Automobiles are typically tested
using their actual suspension systems, which properly capture the rigid body vehicle
suspension modes along with the suspension characteristics.
With all these considerations, if one is performing a modal test to validate a
FEM, either the model must include the support conditions from the test or the test
must adequately approximate the boundary conditions imposed by the model. This
requires coordination between the modal test design and the modeling effort. In any
case, it is virtually impossible to achieve either a totally free boundary condition
or a completely fixed boundary condition. Consequently, suspension development
for a modal test must take into account the expected changes to the test article
behavior to achieve a suspension system that meets the objectives of the test. The
effectiveness of the suspension, whether fixed or free, can be evaluated through
parametric studies performed prior to the test using FEMs, if available. Designing
the test with suspension components imposed also ensures that the appropriate
support locations are included in the model if suspension elements may need to
be adjusted for model updating.

4.1 Approximating Free Boundary Conditions, and the Resulting

Compromises

When modal testing a structure for the purpose of model validation, free boundary
conditions are frequently approximated in the lab to compare with free boundary
condition analyses. Free conditions are used because they are normally easy to simu-
late analytically and easier to approximate experimentally than boundary conditions
with fixed conditions. However, the free conditions can only be approximated in the
lab because the structure must be supported in some manner. To properly design the
modal test, one needs to be able to quantify the effects of the support conditions
on both the measured modal frequencies and damping factors. As will be shown,
the measured modal damping is significantly more sensitive to the support system
(stiffness and damping) than the measured modal frequency. There are simple
formulas that can be used to predict the effect on the measured modal parameters
given the support stiffness and damping.
Historically, there has been concern for support stiffness and its effect on
measured modal frequencies. As early as 1955, Bisplinghoff et al. [4] discuss the
effects of support stiffness and mass on measured modal frequencies. Wolf [33]
506 T. Carne et al.

discusses the effects of support stiffness with regard to modal testing of automotive
bodies. He reports that the rule of thumb to simulate free boundary conditions is
to design the support system so that the rigid body modes – that is, the modes that
would be at zero frequency except for the support conditions – are no more than one-
tenth the frequency of the lowest elastic mode. But it is seldom possible to achieve
this separation for vehicle tests, with test engineers frequently using a 1:3 to 1:5
separation ratio between the rigid body modes and the lowest elastic mode. Wolf
[33] shows that such stiff supports can lead to significant errors in the measured
modal frequencies. Ewins [16], in the test planning chapter of his second edition
of Modal Testing, briefly discusses the issue of the location of suspensions for free
boundary conditions. More recently, Brillhart and Hunt [6] presented an exposition
of many of the practical difficulties involved in designing good fixtures for a modal
test, and Avitabile [3] briefly discusses this issue in a “back to basics” article.
Many authors have subsequently added to this topic, and Carne et al. [12]
developed relatively simple formulas for predicting the effect of support stiffness
and damping on the measured modal frequencies and modal dampings. They applied
these techniques to the testing of a wind turbine blade in the freely supported
condition. The formulas presented here are taken from Carne et al. [13].
Perhaps the best way to develop an understanding of the effects of support
conditions is to examine the very simple two-DOF system. Let us consider a simple
model, pictured in Fig. 14, of an unconstrained structure (free boundary conditions),
consisting of two masses connected by a linear spring and a viscous damper with
motion restricted to a single direction.
This system could be support in several ways, but let us add a support system
symmetrically as diagrammed in Fig. 15. Here, kt and ct designate the true stiffness
and damping of the structure, and ks and cs designate the added support stiffness
and damping. Using the symmetry, the modal parameters of the supported system
can be solved by inspection. There are two modes for this system, φ 1 = [1 1]T ,
and φ 2 = [1 −1]T . The first mode is referred to as the support mode or the rigid
body mode because there is no deformation in the original structure (Fig. 14), only
deformation of the support system. The second mode is the elastic mode because it
involves elastic deformation of the original structure as well as the support system.
The undamped natural frequencies for the two modes are

Fig. 14 Freely supported

two-degree-of-freedom kt /2
system

m m
ct /2
9 Design of Modal Tests 507

ks kt /2 ks

m m
cs ct /2 cs

Fig. 15 Two-degree-of-freedom system with added support stiffness and damping

 
ωs = ks /m ωm = (kt + ks ) /m (1)

where ωs indicates the rigid body mode due to the support system, and ωm indicates
the measured frequency of the elastic mode, which includes the support. Similarly,
the damping factors can be derived from inspection and are

ζs = cs /2mωs ζm = (cs + ct ) /2mωm (2)

Now define symbols for the true natural frequency and damping factor of the
structure, if it had no supports, as

ωt = kt /m (3a)

ζt = ct /2mωt . (3b)

Combining Eqs. (1) and (3a), a very simple, and easily remembered, expression
results, relating the measured frequency, ωm , to the true frequency, ωt , and the
support frequency, ωs :

2
ωm = ωt2 + ωs2 (4)

Or the true frequency can be expressed as

 1/2
ω2
ωt = ωm 1 − 2s (5)
ωm

And if ωs /ωm << 1.0, then

 
ω ωm − ωt ∼ 1 ωs 2
= = (6)
ωm ωm 2 ωm

From Eq. (6), it is easy to see the effect of added support stiffness on the
measured frequency of the test item. If the support stiffness is such that the ratio
508 T. Carne et al.

of the rigid body frequency, ωs , to the measured frequency, ωm , is 1:10, the true
frequency would be approximately one half of 1% different from the measured
frequency, so the 1:10 ratio is a good rule of thumb for most applications for
reasonable accuracy. If the ratio were 1:3, the error would be over 5%, which
generally would be unacceptable.
However, this is just a rule of thumb. For example, envision freely supporting a
horizontal beam with two vertical, soft bungee cords and then measuring the first
bending mode of the beam. If the supports are attached at the extreme ends of the
beam, the supports would have a much greater effect on that modal frequency. In
fact, the effect would be four times greater than that shown in Eq. (5). One would
need to insert a multiplier of 4.0 in front of the ω2 s term. In contrast, if the supports
are attached at the node points of the bending mode, the supports would have zero
effect on that particular modal frequency.
Let us now turn our attention to the measured damping ratio. Following the
example of the frequency analysis above, combining Eqs. (2) and (3b), another
simple formula relating the damping factors results.

ζm ωm = ζt ωt + ζs ωs (7)

The above expression can now be solved for the true damping ratio in terms of
the measured and rigid body damping ratios.
 
ωm ωs ζs
ζt = ζm 1− (8)
ωt ωm ζm

This expression has similarities to that for the frequencies, Eq. (5), except that
the frequency ratio inside the brackets is no longer squared, and it is also multiplied
by the ratio of the damping ratios. If there is a frequency ratio of 1:10, as the rule
of thumb suggests, and if the support and measured damping ratios, ζ s and ζ m , are
equal, there would be a 10% error if the true damping were assumed equal to the
measured damping.
However, suppose there is a moderately damped structure and the frequency ratio
is still 1:10, but the support damping is 5% and the measured damping is 1%. Now
the ratio of dampings in the bracket is 5.0, and it has a large effect. The true damping
would only be 0.5%, so one would have 100% error if one assumed the measured
damping to be equal to the true damping. Lastly, let us now consider the case in
which the frequency ratio is 1:3. If the true damping ratio is again 0.5% and the
support damping ratio is 5%, the measured damping ratio would be 2.59%, resulting
in 400% error if one assumed the measured damping to be equal to the true damping.
From these examples and Eq. (8), one can see that the situation for the measured
damping ratios is significantly different from that for the measured modal frequen-
cies. Assuming the true damping ratio is the same as the measured damping ratio can
result in huge errors as compared to those for the frequencies. Unfortunately, most
FEMs do not include damping, so one cannot validate a damping model with test
9 Design of Modal Tests 509

data and then remove the support damping. Frequently, test-derived modal damping
is used in the model to create the damping model.
There is one saving factor in the measurement of modal damping. The viscous
damping model (one that is independent of frequency) is frequently not a good
model for support structures such as bungee cords and airbags. A discussion of
damping models is beyond the scope of this present work, but many authors would
model the damping as structural damping or solid damping, at least in part. See,
for example [30]. Using structural damping model, then, the typically measured
viscous damping factor, ζ , at the resonant frequency is approximately equal to γ /2
[25]. And Eq. (8), which relates the viscous damping factors, would be modified
for structural damping models instead of viscous damping models. The measured
structural damping coefficient, γ m , can be expressed as a weighted sum of the
support damping, γ s , and the true damping, γ t , as

ks γs + kt γt
γm = (9)
kt + ks

Note that structural damping elements do not add in the same way viscous
damping elements do. Dividing the numerator and denominator of the fraction in Eq.
(9) by the mass, we obtain a relationship similar in form to Eq. (8), but significantly
different due to the squares on the frequencies:

2
ωm γm = ωs2 γs + ωt2 γt (10)

This expression can be solved for the true structural damping in terms of the
measured and support damping coefficients:

2 
ωm ωs2 γs

γt = γm 1 − (11)
ωt2 2 γ
ωm m

This Eq. (11) is much more forgiving of the support system damping than Eq. (8),
due to the squares on the frequency ratio. Also, recall that the viscous damping
factor, ζ , at the resonant frequency is approximately equal to γ /2. So, for example,
suppose again there is a moderately damped structure and the support frequency
to measured frequency ratio is still 1:10, but the structural support damping is
five times that of the structural measured damping: the measured damping would
contain only a 5% error compared to the true damping. In the extreme case when
the frequency ratio is 1:3 and the support damping is five times that of the measured
structure, the measured damping is 50% in error and would be unacceptable, but
not nearly as bad as the viscous damping model. Even for the structural damping
model, one can still vastly overestimate the modal damping in a structure if the true
damping is small. So for applications with low damping in the structure, one must
be particularly attentive to the added support damping, regardless of the damping
model.
510 T. Carne et al.

The discussion, up to this point, has examined the very simple two-DOF model
as diagrammed in Fig. 15. The two-DOF model is very useful as it can be analyzed
exactly and is very helpful for creating intuition and very simple formulas. As
Winston Churchill once said, “Out of intense complexities, intense simplicities
emerge.” That is the beauty of the two-DOF system. However, the multi-DOF
problem is the concern for most modal testing. Typically, the most concern is with
the lowest mode of the dynamic system because it will be most affected by the
support system, so one might think that the two-DOF model should be sufficient.
Frequently, that is true. However, with a multi-DOF system, the placement of the
supports relative to the mode shape can be accounted for, and that can be a very
important aspect of the support problem. This aspect was mentioned in the previous
section with the example of the softly supported beam, where the frequency shift
depends whether the supports are attached at the beam ends or the nodes of the first
mode: if the supports are attached at the beam extremities, the effect of the support
conditions is four times more than the two-DOF formulas provided; and if they are
attached at the nodes of the mode, the effects are reduced to zero.
The derivation for the multi-DOF system is somewhat more complicated than
that for the two-DOF system, so only the final formulas are included in this text.
For the full derivation, refer to [12]. Referring back to Eq. (6), a similar result can
be derived for the multi-DOF system. The change in the frequency, ω, due to the
added support stiffness, kii , is given by the formula below:

1  i 2
ω ∼
= ϕ kii (12)
2ωt

where the change in frequency, ω, is the difference between the measured and
true frequencies. kii is the added or support stiffness at point i on the structure,
and φ i is the value of the mass-normalized mode shape at position i. This formula
for the change in frequency is quite simple, and it is straightforward to evaluate
using the stiffness of the support system and the value of the mode shape at the
attachment position. With more than one support element, one would simply add the
contributions from the elements, including rotational constraints, if any. Comparing
Eq. (12) with Eq. (6), one can reduce Eq. (12) to be exactly the same as Eq. (6),
keeping in mind that the mode shape in Eq. (12) has been mass normalized and
there are two added support stiffnesses. Also note that ωt is in the denominator
on the right-hand side of Eq. (12), so at higher modal frequencies, ω/ωt varies
proportionally as (1/ωt )2 .
Turn now to the issue of damping in a multi-DOF softly supported structure.
For the support damping, the situation is more complicated than for the modal
frequency because one will typically not have an analytical model of the damping
in the structure, in contrast to the stiffness of the support system. But the issue is the
same as for the stiffness: given a measurement of the damping for a particular mode,
how can we determine the change in that modal damping due to the support system?
It has been shown [12] that an approximation to the change in modal damping can
be computed if the support system makes a negligible change to the mode shape
9 Design of Modal Tests 511

of the structure. Further, the mode shape components at the support connections
for mass-normalized mode shapes are required, as well a damping model for the
support system. From the analysis of the complex eigenvalue equation, the result
for the measured damping is
ψ  Cs ψ
ζm ωm = ζt ωt + (13)
2
which is very similar to Eq. (7) derived for the single-DOF system. Here, the
contribution due to the support damping has just been generalized to include the
mode shapes. Cs is the damping matrix for the support system, and ψ is the mass-
normalize mode shape for the particular mode of interest. ψ  is the transpose of the
column vector ψ. Solving for ζ t , results in
 
ωm ψ  Cs ψ
ζt = ζm 1− (14)
ωt 2ωm ζm

This formula for the true damping ratio, as the frequency formula in Eq. (12),
is a fairly simple expression. Given the measured modal damping, the measured
modal frequency, the components of the mass-normalized mode shape at the support
DOFs, and the damping model, the true damping ratio of the unsupported structure
can be calculated. Equation (14) can also be compared to Eq. (8) for the single-DOF
case. Again, these equations are very similar, and Eq. (14) reduces to Eq. (8) for the
single-DOF case.
Equation (14) reveals some important features, just as Eq. (8) did. Because the
quantity in the brackets is the difference between one and a positive number, the
difference between the true damping ratio and the measured damping ratio can be
significant if the last term in the brackets is not close to zero. In conclusion, when
attempting to simulate a free body in the lab for modal testing, one must be careful
that the supposedly soft support system does not significantly affect the measured
modal frequency or the measured modal damping. Highly damped support systems
can significantly affect the measured damping.

4.2 Suspension System Design, Low Spring Rates

When free boundary conditions are desired, the most common means to achieve this
is to design a system that provides a low spring rate in all six DOFs of the structure.
However, “low spring rate” is a relative phrase. For a structure with modes of interest
in the hundreds of hertz, a spring rate that provides rigid body modes below 20 Hz
may be quite acceptable, whereas for a structure with modes below 10 Hz, it might
be desirable to have a suspension system yield rigid body modes below 1 Hz. This
can be a difficult design goal and is often a challenge for large aircraft. As a result,
the suspension systems used may not yield any pure rigid body modes and the lowest
modes may reveal coupled behavior between the desired rigid body modes and the
flexible body modes.
512 T. Carne et al.

The design goal in developing a low-frequency suspension system is to achieve

good isolation between the six-DOF rigid body modes and the first flexible modes
of the structure. If these frequencies are too close together, the suspension system
modes will influence the responses in the flexible modes of the system, and in
extreme cases they will couple, combining rigid body and flexible body behavior
in the mode shapes.
Some simple evaluation of desired isolation frequency, required spring rates, and
the capability to support the test article weight can help in the suspension design
process. In a single-masscase, the static spring force F = W = Kx = Ma, and the
natural frequency fn = 2π 1 K
M . This helps define the ratio between K and M,
which can yield the desired frequency and show the static deflection requirements
for a linear spring. A good rule of thumb is that static deflection, x, is equal to 10/fn2
in inches.
The suspension design can be evaluated using a rigid body model of the test
article with appropriate suspension system attachment locations. After adjustments
are made to the spring rates to achieve the suspension frequencies, the suspension
spring rates can be applied to a flexible model of the structure and further analysis
completed.
For many test articles, good suspension system isolation can be achieved by using
rubber bands or bungee cord to support the test article. If a single-point support
location can be identified, such a system allows five DOFs of the suspension system
to be characterized by pendulum behavior, which can yield very low frequencies.
The spring rate of the bungee then defines the sixth DOF suspension frequency,
which can be adjusted while accommodating the static weight of the article.
Other test articles can be placed on a suspension system made of foam. This is
often done when the test article is light weight and the flexible frequencies are quite
high. Many new composite structures can be characterized in this fashion. A foam
suspension provides good isolation between the rigid body modes and the flexible
body modes so that they do not exhibit coupled behavior.
Large aircraft present a variety of suspension challenges. It is difficult to achieve
low spring rates with the capacity to carry this large weight and still achieve a
sufficiently low suspension frequency. Often the first flexible mode of a large aircraft
is near 1 Hz, making suspension isolation difficult. As a result, the suspensions used
are often the aircraft landing gear tires. This is analogous to the tires for automotive
vehicles, and these cannot typically achieve the low spring rates desired. Suspension
system effects can be accounted for in these cases by using FEMs that include
the suspension system or by making accurate modal models, which can be used
to subtract the suspension system’s influence.

4.3 Constrained Support, Built-In, and Other Boundary

Conditions

FEMs are effective in providing estimates of completely fixed boundary conditions.

These models can be used to plan for the test and perform parametric studies of
9 Design of Modal Tests 513

the influence of the planned constraint method. Also the model can include any
flexibility in future updates.
In cases where a constrained or fixed boundary condition is desired, it is often
difficult to achieve a completely fixed condition for actual hardware. Instead, the
intent is to develop a boundary condition stiff enough to achieve constraint loads
similar to those that would be imposed by a fixed boundary condition. This may be
achieved by attaching the test article to a plate or mass that is much larger than the
test article. Unfortunately, in many cases, the selected mass may not be large enough
to impose a truly fixed base condition. In these cases, a model of the boundary
condition and supporting hardware should be included in the FEM.
The boundary condition should be validated by making measurements around
the base of the test article to ensure that there is no unexpected motion or excessive
participation in the structural test modes. In some cases, the interface loads at the
connection between the test article and the boundary constraint can be made to
validate the actual interface loads imposed.

4.4 Operating Environments

In planning for the type of boundary condition to be used in a modal test, the
operating environment should be strongly considered. The structure’s operating
environment will influence the modal behavior of the structure, and it is important
to capture those effects in the test boundary condition to make sure that similar
modal behavior is developed. This will ensure that the strain energy in the test
structure is similar to that observed in the operating environment. Since free and
fixed boundary conditions can completely alter the modal parameters (including
the shapes) of the test article, the strain energy distribution will be substantially
different. Therefore, the test boundary condition should be carefully considered in
planning for the modal test.

5 Measurement Quality Criteria

There are numerous quality checks that should be planned for in conducting a modal
test. These checks will help ensure high-quality FRF measurements so that accurate
modal models can be achieved. The quality checks should be performed at various
stages over the course of the modal test, and they should be included in test planning
and procedure documents to verify that they are properly completed.
During test performance, validation that measurement sensors are providing
proper, clean signals is important to all subsequent data processing. All data
channels can be assessed visually for signs of problems with connections or signal
conditioning. These visual checks can be accomplished by examining the time-
domain responses and evaluating the FFT signals. All channels should be reviewed.
All time-domain data should be stored for further review. While storing all time-
domain responses was not practical in the past due to storage constraints and data
processing capabilities, this is now easily accomplished. Storing all time-domain
514 T. Carne et al.

data allows many data quality checks to be performed immediately after measuring
the data to determine whether more data should be acquired. It also allows the data
to be processed using alternative processing techniques, which can also prove useful
in the evaluation of data quality.
The review of time-domain data should include an assessment of fundamental
statistics, including the peak-to-RMS ratio for the measured data. The peak-to-
RMS ratio can provide insight into whether any channel of data experienced
intermittent signals or “spikes” during acquisition, as can be caused by loose cables,
connections, or structural rattles. Assessing this behavior for each channel allows
problems to be identified and corrected before further data is collected.
Review of the time-domain waveforms can also be done to identify when any
channel clipping may have occurred. Channel clipping may occur if the signal
levels are too high for the data acquisition system input configuration; in this case,
the setup can be corrected. Clipping or saturation may also occur in the sensor
if the electronics voltage levels are exceeded. This can occur if a sensor is used
whose sensitivity is too high for the response levels observed. Changes to the
instrumentation or the excitation input levels may be required to account for these
issues. Conversely, if the excitation levels are too low, or the sensors have low
sensitivity, or the data system input voltage range is set too high, the data may
exhibit signs of digitization bit noise. Once the reason for the bit noise is identified,
corrections can be made before further testing proceeds.
Following review of time-domain signals, results of data processing can be
evaluated using frequency-domain data. Plotting the auto spectra of all channels
together can help identify any channels that appear to be out of range of all the other
channels, as this can be indicative of a calibration issue or a channel ranging issue.
Coherence is also a standard quality metric that is applied to data channels
to evaluate whether the response measured resulted primarily from the applied
excitation input. Ranking data channels using overall coherence allows channels
with poor coherence to be flagged for further review. Additionally, weight coherence
can be computed to bias results toward those channels with low response levels, and
order ranking again can be used to identify channels that may need further review.
Drive point reciprocity should be reviewed to assess all excitation inputs and
identify any input issues. For these reciprocity comparisons, response measurements
must be made at each of the input locations in the same direction as the input,
yielding drive point FRFs. Drive point FRFs and drive point reciprocity FRFs help
confirm that all shakers are performing properly and yielding high-quality FRFs.
At the same time, the power spectra of the force signals should be compared
to verify that the excitation levels and frequency range of interest are being
achieved.
As the data quality checks are completed and all data channels are confirmed to
be operating properly, further quality checks of the modal survey can be conducted
by adjusting excitation input levels to check for structural linearity. This check is
not so much a check of data quality as a check of the structural behavior observed,
and it can also indicate whether high excitation levels result in excessive amounts
of noise due to structural rattles.
9 Design of Modal Tests 515

6 Modal Tests for Model Validation

When a modal test is to be performed for purposes of validating an analytic or finite

element model, one needs to design the test so that the resulting measurements
will provide the required data for validating or updating the model. There are three
key issues to consider in the design of a model validation test: the number and
location of response sensors; the number, location, and orientation of the inputs;
and the criteria that will be used to judge the validity of the model. Typically, one
compares the modal frequencies and mode shapes from the model with those of the
test. However, for example, one may want to compare particularly important FRFs,
or examine the cross orthogonality between the test mode shapes and the analytical
mode shapes [5]. This latter criterion is described more thoroughly in Sect. 6.1.4.
Regardless of the validation criteria, if agreement between the model and the
test is not acceptable, uncertain structural parameters may need to be identified and
adjusted to improve the agreement. One must measure quantities that are affected by
these uncertain parameters, whether the quantities are frequencies, components of
the modal vector, rotations, strains, gradients, damping, or other factors. Frequently,
it is the uncertain structural parameters one wishes to identify in the modal test. For
example, if there were uncertainty about the stiffness of a bolted joint connecting
two elements of the structure, one would want to choose instrumentation that would
reveal the relative motions across the joint so that a comparison with the model
could be conducted. Due to these considerations, one must design the test so as to
be relatively certain of making the appropriate measurements, but one should also
keep in mind the effort required for testing and avoid making more measurements
than necessary.
Fortunately, if validating the model is our objective, then the model is typically
available, and it can be used as a tool in the design process. As mentioned above,
two of the key issues are sensors and inputs. Regarding the sensors, the number,
direction, and location must be chosen so that a correspondence between the
analysis and the measured mode shapes can be obtained. One cannot assume that
the numerical order of the modes measured in the test will align with those from the
analysis, so being able to differentiate between mode shapes is critical, and all the
mode shape vectors must be relatively independent; otherwise, differentiation will
be very difficult. Consequently, the inner products between all the shape vectors
need to be relatively small. The discussion that follows in Sect. 6.1 addresses only
the first issue – the response sensors – with a review of various approaches for
selecting an optimal set of measurement locations and directions.
Regarding the input excitation, the various approaches to developing an adequate
excitation design are discussed in Sect. 6.2. For all the approaches, the primary
requirement is that an excitation must be designed so that all the target modes are
excited to the degree that they are fairly easily (and thus accurately) identified and
estimated from the measurements. However, this brings up another issue, that of the
target modes. As part of the design process, one must determine the number and the
character of the target modes. In some tests and analyses, one may be concerned
with only a few modes that dominate the dynamic behavior. In other tests, one
516 T. Carne et al.

may need to examine many modes to characterize all the important dynamics of
the structure.
One more issue involved in the design of any modal test is the ability to visualize
the mode shapes once they have been estimated. This is most important for the
test engineer so that he or she can visualize the shapes and ascertain the accuracy
of the mode shape estimation. All too frequently, a sensor will have the wrong
direction, position, or polarity associated with it, but the visualization of the shapes
will frequently resolve these issues. Consequently, in selecting a set of measurement
locations, one needs to start with a visualization or intuition set. Then, locations can
be added to that set to optimize the design requirement, as discussed in the sections
below.

6.1 Selecting Response Locations

FEMs in general contain many more DOFs than could ever be instrumented with
accelerometers during a modal test. Therefore, a subset of DOFs must be selected
for measurement in the test. The selection of these DOFs must be based on the figure
of merit for the modal survey test. In many cases, the figure of merit may be mode
shape visualization or the ability to identify and quantify sources of compliance
of a load path in the model, such as compliance across joints. In these cases,
manual selection of a preliminary set of response locations may be very useful. In
cases where the figure of merit includes a mathematical objective function, such as
mode shape independence, automated methods are useful to supplement the manual
selection.
Automated procedures for selecting appropriate accelerometer locations are most
often required in the cases of complex structures or when only a small number of
accelerometers are available. In any case, it is important to remember that automatic
sensor placement techniques can never totally supplant engineering judgment. The
test and analysis engineers must still be proactive in the placement of sensors; the
DOFs from an automatically selected sensor set must be further studied to determine
whether the set makes sense and whether any DOF has been included at which it
is physically impossible to place a sensor. Also, rotational DOFs are usually not
included in the sensor set unless there exists a large mass with a large rotational
inertia, because rotational acceleration cannot easily be measured.
Goals in selection of the test response DOFs are as follows:

1. Ability to visualize all unique target mode shapes

2. Ensuring the independence of the mode shapes evaluated at those response
DOFs, that is, small shape-vector inner-products
3. Identifying responses at key structural locations and joints where needed

6.1.1 Modal Kinetic Energy

The first automatic technique that is discussed, modal kinetic energy, is for use
with the static reduction method [17]. See, for example, Shah and Udwadia [29],
Udwadia and Garba [31], and Salama et al. [28]. For an accurate static test-analysis
9 Design of Modal Tests 517

model (TAM), we must retain in the sensor set all the DOFs that have large
deformations in the target modes or possess a lot of mass. In addition, the sensor
set should be somewhat evenly distributed over the structure. The sensor DOFs
must render the sensor partition of the FEM target modes linearly independent
or the TAM will not predict all the target modes. Assuming that each target
mode of interest is mass normalized, the FEM mode shapes satisfy the relation
φiT Mφi = 1.0, and we need to determine the fractional contribution of each
candidate sensor DOF to the modal generalized mass. Let us consider just the i th
mode. The contribution of the j th DOF to this i th mode generalized mass is given by

n
keij = φij Mj k φik = φij Mj φi (15)
k=1

in which φ ij is the j th row of the i th target FEM mode shape corresponding to the
j th DOF, Mjk is the entry in the j th row and k th column of the FEM mass matrix,
Mj is the j th row of the FEM mass matrix, and n is number of FEM DOFs. This
contribution is a fractional one because if the value is computed for each FEM DOF
and summed over all the DOFs, the total must naturally be 1.0, corresponding to the
value of the i th generalized mass.
Using the above computation, one can generate a column vector kei for each
mode shape listing the contribution of each DOF. Ultimately, one wants a single
measure of importance for each DOF over all the target modes. Note that modal
kinetic energy gives us no information as to the importance of one mode with respect
to another, so we will treat them all equally. One way of coming up with a single
measure of goodness is to average each DOF’s importance over all the target modes:

nm
1
KE = kei (16)
nm
i=1

where kei is a column vector containing the contribution of each of the degrees of
freedom j to the kinetic energy of the i th mode.
The vector KE provides a listing of the fractional importance of each FEM
DOF. This vector can then be sorted from high to low. This computation is usually
performed for a large candidate set of sensor DOFs. The modal kinetic energy is then
used to either truncate the low-energy DOFs all at once, or one candidate sensor at a
time can be deleted. At each iteration, the lowest-energy candidate sensor is deleted
from the candidate sensor set. That DOF is statically reduced out of the mass matrix,
and then the kinetic energy distribution is recomputed and sorted.

6.1.2 Effective Independence

The modal kinetic energy approach to sensor placement is dictated by the require-
ments of the accurate use of the static reduction technique for TAM generation.
In contrast, the Effective Independence (EfI) technique [20] approaches the sensor
518 T. Carne et al.

placement problem from the standpoint of a structural dynamicist who must use the
modal parameters identified during a modal test to perform test-analysis correlation
[18] and FEM updating. It is vital that the targeted test mode shapes are linearly or
spatially independent. As in the case of modal kinetic energy, the sensor placement
process begins by designating a large set of candidate sensor locations from which
the smaller final sensor configuration will be selected. The modal kinetic energy
distribution, presented in Eq. (16), can be used to help determine a good candidate
sensor set.
The objective of the EfI sensor placement strategy is to select sensor locations
that render the FEM target mode shape partitions as linearly independent as possible
while at the same time retaining as much information as possible about the target
modal responses within the sensor data. Independence of the target mode partitions
is required such that the test data can be used in test-analysis correlation as
mentioned previously. The sensor placement problem can be alternatively cast in
the form of a state estimation problem. The target mode independence requirement
implies that at any time t, the sensor output can be sampled and the target modal
response can be estimated. A static Fisher model is assumed for the output equation
of the form

ys = φfs q + w (17)

in which ys is the response at the sensor locations, φ fs is the matrix of FEM target
mode shapes partitioned to the sensor degrees of freedom, q is the target modal
response, and w is the sensor noise. It can be shown that the best estimate of
the modal response is obtained by minimizing the error covariance matrix, which
is equivalent to maximizing the Fisher information matrix (FIM). Therefore, the
sensors should be placed such that the FIM is maximized in the appropriate matrix
norm. Details can be found in Kammer [20].
Assuming we know nothing a priori about the sensor noise, the corresponding
FIM is given by

Q = φfsT φfs (18)

Maximization of the FIM determinant has been a commonly used criterion

for optimal parameter estimation. It is suggested here that maximization of the
information matrix determinant is also appropriate for optimal sensor placement.
The form of the information matrix presented in Eq. (18) indicates that if the
target modal partitions are not linearly independent, the determinant will be zero.
Therefore, maximizing the information matrix determinant will maximize the
spatial independence of the target modal partitions as desired. It will also maximize
the signal strength of the target modal responses in the sensor output, which is very
desirable in the presence of noise. Thus, for the purpose of sensor placement using a
state estimation formulation, the determinant is the appropriate measure of the size
of the FIM.
9 Design of Modal Tests 519

The EfI method approaches the problem by determining how much each sensor
location contributes to the eigenvalues of the FIM. For nm target modes, the positive
definite and symmetric information matrix possesses a set of nm orthonormal
eigenvectors and positive eigenvalues λ. The eigenvectors represent orthogonal unit
directions in an nm -dimensional identification space. In the matrix G formed by the
product

G = φfs ˆ2 (19)

in which the symbol ˆ2 indicates a term-by-term multiplication, each row contains

the squares of the components of the corresponding row of φ fs in terms of the
coordinates defined by the eigenvectors . Each column of matrix G sums to the
corresponding eigenvalue of FIM Q. Therefore, the i th term in the j th column of
G represents the contribution of the i th sensor location to the j th eigenvalue of Q.
Normalizing G by post-multiplying by the inverse of the eigenvalue matrix produces

FE = φfs ˆ2 λ−1 (20)

where now the directions in identification space are all of equal importance and the
ij th term in FE represents the fractional contribution of the i th sensor location to the
j th information matrix eigenvalue. Adding the terms within the rows of FE produces
an nc dimensional vector called the Effective Independence Distribution, which is
given by

ED = φfs ˆ2 λ−1 {1}nm (21)

in which {1}nm is a column vector of 1’s of dimension nm . The sum of the terms
in column vector ED is nm , which is the rank of the FIM and the target modal
partition φ fs . It can be shown that individual values EDi of vector ED satisfy the
relation 0 ≤ EDi ≤ 1.0, where a value of 0.0 indicates that the i th sensor contributes
nothing to the linear independence of the target modes or even their observability,
and a value of 1.0 indicates that the corresponding sensor is absolutely vital to the
independence of the target modes and thus cannot be deleted from the candidate set.
As mentioned previously, the initial candidate sensor set is selected such that φ fs
is full column rank implying that Q is positive definite. Entries in vector ED are
sorted by magnitude, and the lowest-ranked sensor is deleted from the candidate
set. Remaining sensor locations are then ranked and sorted again. In an iterative
fashion, the initial candidate set of sensor locations is rapidly reduced to the number
allotted for the modal test. It is important to note that candidate sensor locations
must be deleted in an iterative fashion and not all at once to obtain the desired
number, because as sensor locations are deleted, their EfI value EDi changes such
that the sum of the EfI distribution vector ED remains at nm . Therefore, deletion of
large numbers of sensors in one iteration may lead to the removal of a vital sensor
location, resulting in the loss of linear independence of the target mode partitions.
520 T. Carne et al.

As suggested, the determinant is the appropriate measure of the size of the FIM
for the state estimation problem in Eq. (17) and thus also the sensor placement
problem for modal identification. It is a direct measure of the amount of target
modal response information contained in the sensor data, and it can be used to
determine the goodness of one sensor set with respect to another. It can be shown
that the EfI of the i th sensor location EDi represents the fractional change in
the determinant of the FIM when the corresponding sensor location is removed
from the candidate set [27]. It is important to note that whenever a sensor is
removed from the candidate sensor set, the determinant decreases. When the lowest-
ranked sensor is removed at each iteration, the remaining sensor locations produce
the largest possible FIM determinant. Therefore, even though the EfI method of
sensor placement is suboptimal due to its iterative nature, it tends to maintain the
determinant of the FIM, which leads to a smaller error covariance matrix and better
estimates of the target modal response.
As an example, consider sensor placement for a simple unconstrained beam
representation of a large space structure with a concentrated mass at the midpoint
equal to three quarters of the mass of the beam itself. The finite element represen-
tation, illustrated in Fig. 16, was constructed using 22 grid points and 21 elements.
Each grid possesses a transverse and a rotational DOF. The first ten mode shapes
and frequencies were computed, which include two rigid body modes and eight
elastic modes. The first seven elastic mode shapes were selected as target modes
to be identified during a modal test. All 22 transverse displacement DOFs were
considered in an initial candidate sensor set. Figure 17 presents each of the initial
22 sensor locations’ fractional contributions to the target mode independence; this
is essentially a ranking of the importance of prospective sensor locations to the

Concentrated Mass

- Indicates Selected Sensor Location

108

111

114

117

120
101

123
104

107

110

113

116

119
100

122
103

106

109

112

115

118

121
102

105

Candidate Sensor Location ID

Fig. 16 Simple beam model with central concentrated mass

9 Design of Modal Tests 521

Fig. 17 Fractional contribution of sensor locations to linear independence

success of the modal survey. As expected for this case, the beam endpoints are
the most important locations. Note, however, that none of the locations is vital to
independence for the initial set.
Fourteen iterations were used to reduce the initial candidate sensor set to a final
configuration of eight sensors which maintains the determinant of the FIM and
the linear independence of the target modes. The eight selected sensor locations
are illustrated in Fig. 16, and the corresponding EfI distribution for the selected
configuration is pictured in Fig. 17. Note that the EfI values for the sensor locations
in the final configuration are larger than their initial values in the original candidate
set; as sensors are deleted, the remaining locations become more important and, in
some cases, the relative importance of sensors changes. For instance, in the initial
EfI distribution, sensor 116 was more important that sensor 120; however, as sensors
were deleted, location 120 surpassed 116 in importance and was selected for the
final configuration while 116 was deleted.
The EfI method for sensor placement has another advantage in that the distribu-
tion for the final configuration indicates the cost of losing a sensor. This information
can be used to place backup sensors when necessary, such as in on-orbit modal
identification. The EfI technique of sensor placement provides sensor configurations
that produce independent target mode partitions that an analyst can use to perform
test-analysis correlation and then model updating. It has also been shown that this
approach to sensor placement enhances the actual identification of the target modes
during a modal survey [21]. The method can also be extended to include DOF mass
weighting [23] and the placement of triaxial sensors [22].
522 T. Carne et al.

6.1.3 Min-MAC
The Min-MAC approach to sensor selection is quite straightforward [8]: response
sensor locations and directions are simply chosen to minimize the off-diagonal ele-
ments of the modal assurance criterion (MAC) matrix [1, 15]. To easily distinguish
one mode shape vector from another, the inner product between the two vectors,
normalized by the vector lengths, must be relatively small. The normalized inner
product is merely the cosine of the angle between the vectors. If the cosine is near
±1, the vectors are close to parallel; if the cosine is small, the vectors are very
independent and easily distinguishable. The commonly used MAC matrix can be
utilized to evaluate the square of the cosine between the shape vectors as given
below.
2
ϕiH ϕj
MAC ij =   (22)
ϕiH ϕi ϕjH ϕj

where ϕi and ϕj are the mode shape vectors for the i th and j th modes. The superscript
H indicates the Hermitian of the vector or just the complex-conjugate transpose. For
real vectors obtained from a FEM, the Hermitian reduces merely to the transpose,
and the MAC is simply computed using vector transposes as a real quantity. By
examining the off-diagonal terms in the MAC matrix, using modal vectors computed
from a detailed FEM, that are evaluated (or partitioned) only at a reduced set of
DOFs, we can determine whether a set of sensors at those DOFs will be a good
design for obtaining independence of the vectors. Thus, our design criterion is to
require all the off-diagonal elements of the MAC matrix to be relatively small.
The MAC matrix uses an identity-weighting matrix (not the mass matrix), so one
should not expect the off-diagonal terms to be zero, which implies orthogonality.
Orthogonality of the vectors is not necessary; their MAC values need only be small
(e.g., 0.2) so that vectors are easily distinguishable. Mathematically, any MAC value
less than 1 would reveal two unique independent vectors. However, the MAC values
between the test vectors must be small because when the analysis shape vectors are
aligned or compared with the test shape vectors with the MAC computation as a
guide, the test-analysis cross-MAC values may actually not be near unity if there is
significant error in the FEM or significant experimental error in the measurement or
estimation of the shape vectors.
In the introduction to this section, the visualization or intuition set was discussed.
This set of response locations and directions is important for enabling the test
engineer to visualize the deformed mode shapes and identify any measurement
errors and to characterize the various modes. This visualization set is then used
to initialize the selection process in Min-MAC. For a complex multidimensional
structure like a satellite or car body, one might use only triaxial sensors in the
visualization set. This allows the shapes to be immediately visualized without
depending upon any transformation based on the model to fill in the omitted DOFs
for a full three-dimensional visualization. This does increase the number of sensors
used, but it significantly increases the robustness of the design for visualization.
9 Design of Modal Tests 523

Alternatively, one can intuitively select single-direction sensors, which will still
reveal the fundamental shapes of the measured modes.
After the small visualization set has been defined, a procedure is used to add
one sensor at a time, choosing from the DOFs remaining in the FEM to reduce the
off-diagonal MAC values to an acceptable magnitude. This procedure is described
subsequently. However, this approach is in significant contrast to many other design
techniques. Here, the sensor set in initialized using a small visualization set and
additional DOFs are added one at a time, whereas other approaches start with the
entire set of available DOFs and attempt to reduce that number to a manageable
set. It will be subsequently shown that there is a tremendous numerical advantage in
starting small and increasing rather than starting with a large set of DOFs and paring
down: with this procedure, a very large FEM can be used without computational
difficulties or excessive time required.
This numerical procedure is next described in detail. As discussed earlier, the
goal of this technique for sensor selection is to reduce the magnitude of off-diagonal
elements of the MAC matrix by adding one sensor at a time to the initial set or the
current set. Thus, a method is required for evaluating the effect of adding a single
additional sensor on the MAC matrix.
Let (n × m) and (p × m) denote the mode shape matrices for the existing set
and the remaining set of DOFs from the FEM, respectively, where n is the number
of DOFs in the current set, p is the number of remaining DOFs from which to select,
and m is the number of modes. The MAC value between modes i and j is

aij 2
MAC ij = (23)
aii ajj

where aij are the elements of A = T . When a new response is added to the 
from the Ψ , for instance row k of , the MAC value between modes i and j with the
added DOF k becomes

2
 aij + ψki ψkj
MAC ij +k
=  (24)
(aii + ψki ψki ) ajj + ψkj ψkj

Examination of Eq. (24) shows that evaluating a candidate sensor only requires
five multiplications, three additions, and one division per MAC matrix element. The
simplicity of this equation makes it possible to efficiently evaluate the influence of
a large number of candidate DOFs on the MAC matrix.
After a response DOF is chosen to be added to the existing set, the matrices ,
, and A are updated to reflect the changes. One can then continue this process
for evaluating candidate DOFs. Of all the DOFs available, the DOF leading to the
smallest maximum above-diagonal MAC value in the entire matrix is then included.
This method requires the use of Eq. (24) m(m − 1)p/2 times for each added sensor.
Of course, one can limit the range of p to be a subset of all available DOFs in the
524 T. Carne et al.

FEM; realistically, one might choose 1000 or so candidate DOFs from the entire
FEM to be considered for sensor placement.
To illustrate the use of Min-MAC technique, an example application is described
below for which the technique was applied to an automotive body structure. As
previously described, an initial visualization sensor set was selected based on
intuition. For the car body, 34 nodes were selected with all three DOFs, as triaxial
sensors were to be used for the modal test, creating an initial set of 102 DOFs. This
is more than one would ordinarily use, but it did provide excellent visualization.
Now using the FEM, the 11 mode shape vectors with the lowest frequencies were
partitioned down to these 102 DOFs, and the MAC matrix was formed to determine
the off-diagonal values. (The diagonal values of the self-MAC are unity, of course,
in contrast to a cross-MAC between the FEM modes and the test modes.) Table 1
shows the full MAC matrix (11 by 11), which is symmetric, for the shapes using
this initial visualization set of DOFs. In this matrix, one can see that two of the off-
diagonal elements are fairly large. In particular, element (3,5) has a value of 0.51 and
element (5,7) has a value of 0.60. This is an unacceptable design, as mode shapes
5 and 7 will appear similar, as will mode shapes 3 and 5. An estimated test mode
shape could easily have a MAC value of 0.5 with both modes 5 and 7, and there
would be no way of determining the correspondence between the test and analysis
modes.
Now, the Min-MAC algorithm is applied using the analysis mode shapes from
the FEM to reduce the off-diagonal MAC values by adding one DOF at a time. The
result of adding DOFs to the existing set is shown in Table 2.

Table 1 MAC matrix for initial set based on mode visualization and intuition
Mode Number 1 2 3 4 5 6 7 8 9 10 11
1 1.00 0.01 1.00 0.00 0.00 0.16 0.00 0.00 0.02 0.00 0.01
2 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00
3 1.00 0.16 0.51 0.00 0.25 0.02 0.00 0.01 0.00
4 1.00 0.21 0.00 0.00 0.21 0.00 0.00 0.00
5 1.00 0.00 0.60 0.23 0.00 0.09 0.05
6 Symmetric values 1.00 0.00 0.00 0.14 0.06 0.10
7 1.00 0.00 0.00 0.24 0.15
8 1.00 0.00 0.10 0.06
9 1.00 0.00 0.00
10 1.00 0.04
11 1.00

Table 2 Maximum off-diagonal MAC values for the added sensors

Added sensors 0 1 2 3 4 5 6 7 8 9 10
Largest MAC 0.598 0.232 0.203 0.185 0.175 0.179 0.163 0.159 0.157 0.152 0.149
values
9 Design of Modal Tests 525

Notice that with ten additional DOFs (or sensors), the maximum off-diagonal
MAC value has been reduced to below 0.15. There were actually four MAC values
of 0.15 with this design. For purposes of comparison, the MAC matrix using all the
DOFs in the entire FEM was computed. The maximum off-diagonal MAC value
was 0.12 – virtually the same as the design with just 112 sensors.
In conclusion, the Min-MAC procedure for selecting sensors is initialized with
a small intuition or visualization set, and then sensors are added to reduce the
off-diagonal MAC elements. This procedure addresses both the visualization and
the vector correspondence issues. The procedure is very efficient numerically, as it
starts with a small set and adds sensors one at a time, and it confers a tremendous
numerical advantage that can be used when computing off-diagonal MAC elements,
given you already have the MAC matrix for the vectors with one less component
included in the vectors. With this procedure, one can evaluate an extremely large set
of DOFs from a FEM without difficulty and directly obtain a sensor set that reduces
the off-diagonal MAC elements.

6.1.4 Aerospace Cross-Orthogonality; TAMs

It is important to establish the goals and instrumentation response locations during
the test planning stage when the test mode shapes will be used to help validate a
FEM. The test mode shapes can be most successful and easily compared to those
of the FEM when a TAM is developed from the FEM during test preparation. This
TAM provides the specific DOFs that will be compared between the test and analysis
modal models. In the test planning activity, this TAM can be used to determine
how many DOFs or response measurements are needed and where they need to
be located to meet the comparison goals that will be applied. This determination
of DOFs establishes the FEM mass matrix that will be used in orthogonality and
cross-orthogonality comparisons, as well as the partitioned analysis mode shapes to
be used in the cross-orthogonality checks [5].
The possible success of the test sensor set is evaluated by using the full FEM
mode shapes, partitioned to the test DOFs, and computing the pseudo-orthogonality
relative to the TAM mass matrix. The pseudo-orthogonality can be computed
with the mode shapes unnormalized and then normalized for a unit diagonal
orthogonality matrix, just as would be done during the test. These computations
allow assessment of whether all of the system mass has been captured at the test
DOFs and whether the test DOFs provide an independent set of modes. The modes
acquired from the partitioned FEM shapes are checked for orthogonality using the
TAM analytical mass matrix. The orthogonality matrix [O12 ] is computed from the
following matrix product.

[O12 ] = [1 ]T [MAA ] [2 ] (25)

When checking the DOFs for orthogonality as would be obtained in the test, the
terms [ϕ1 ] and [ϕ2 ] are mode shape matrices (normalized to [MAA ] to produce 1.0
on the [O12 ] diagonal), and [MAA ] is the TAM analytical mass matrix. The pseudo-
526 T. Carne et al.

orthogonality of the partitioned FEM mode shapes relative to the TAM mass matrix
is computed from Eq. (25) with [ϕ1 ] = [ϕ2 ].
It is important to note that the DOFs preserved in the mass matrix match the
DOFs of the analysis TAM, which also match the response locations to be used
in the test. Proper selection of these DOFs is important to the ability to achieve
good orthogonality. A good set of DOFs for the TAM does not require that three
DOFs be measured at each test node described; instead, it is usually more effective
to distribute the response measurements around the structure with measurements
at key locations that have high mass weighting. Sometimes these will yield three-
DOF measurements at individual points, but usually numerous locations will be
developed with one-DOF and two-DOF responses. In this way, pretest evaluation of
the selected DOFs and resulting possible orthogonality results help in defining the
most effective response measurement set for the test.

6.2 Selecting Input Locations, Directions, and Number

Good test planning should include taking steps to identify the specific locations,
number, and directions of the inputs to be used in the modal test. Evaluation
and selection of input locations can involve engineering judgment; however, more
rigorous selection steps can and should be taken particularly when a FEM is
available for pretest planning. This selection is an important step in the test planning
effort as using a good set of input locations can strongly influence the quality of the
modal test data.

6.2.1 General Guidelines

Candidate sets of possible input locations can be assembled by reviewing locations
that are accessible and show acceptable modal participation. These candidate
locations can supplement the other measurement DOFs that have been defined for
the test; however, the input locations are often co-located with other measurement
locations. Alignment with the local and global coordinate systems is not required,
and it is frequently beneficial to use inputs that are not aligned with any of the
coordinate directions. Inputs need to be defined in number and orientation such that
the most modes are effectively excited. It is not necessary to use locations that are
part of the measurement set or part of the test display model. However, drive point
accelerometers co-located and aligned with the input directions should always be
installed, as doing so ensures that reciprocity comparisons can be made during data
quality checks.
In general, the number and locations of exciters should be selected so as to be
able to excite all modes of interest. If possible, they should be placed at stiff points
on the structure that are designed to carry static loads so that the vibration energy
can be distributed throughout the structure. Using multiple input locations increases
the likelihood of exciting all of the modes. Pretest planning with the FEM can help
9 Design of Modal Tests 527

in identifying the appropriate locations: FRFs can be computed to reveal whether a

location is predicted to excite many of the modes.
It is also important that expected response levels be acceptable. Again, response
levels can be predicted using the FEM. This does require some assumptions of
the amount of damping in the structure, but this can be reasonably estimated
based on other test articles, where available, and other engineering information.
Predicted response levels can be used to match sensor measurement capabilities and
sensitivities to ensure that proper modal signal levels can be observed.

6.2.2 Selection of Locations

It is possible to select modal input locations using ad hoc approaches simply by
looking at the structure and estimating which locations might be effective for
exciting the structure, and these input locations can often be selected with regard
to convenience. However, it is usually best to perform some evaluation of the drive
point FRFs to assess whether input locations selected in this fashion will provide
acceptable results.
As an initial step, impact testing approaches can be used to assess input locations
when shakers are going to be installed. Measuring the drive point FRFs using
hammer impacts will show whether acceptable FRFs can be obtained from the
specific locations and specific directions evaluated, and numerous drive points can
be evaluated and compared. From these, the best locations can be selected with
an appropriate number of locations that indicate excitation of all modes in the
frequency range of interest.
A valuable tool for selecting the input locations is the mode-indicator function
(MIF) [8, 26, 32]. This tool can be used whether the input locations are identified
analytically using the FEM or experimentally using the measured results from a
candidate input. The FRFs are used to compute the MIFs associated with each
input location, which allows assessment of whether all modes of interest are
excited. With MIFs computed using FRFs from the FEM, an automated procedure
can be developed to identify an acceptable set of input locations and directions.
Minimum criteria of the MIF can be defined and evaluated across all of the possible
input locations, and candidate inputs can then be selected to determine the best
combinations of input locations and directions.

6.3 Planning the Criteria for “Test Exit”

Specific goals and test evaluation metrics should be included in the test planning
for a modal test. These provide tools to assess whether sufficient testing has been
completed to meet the overall objectives of the test. Test objectives are not always
the same, so they need to be clearly stated during the planning stage. There are
numerous objectives that can be listed in terms of the modal parameters to be
identified, such as the following:
528 T. Carne et al.

• Mode independence checks

– Modal Assurance Criteria (MAC)
• Criteria for specific off-diagonal values
• Criteria applied to only specified target modes
– Orthogonality
• Mass weighting derived from FEM TAM
• Criteria to achieve specific off-diagonal values
• Criteria applied to only specified target modes; effective mass weighting is
often used
• Mode completeness
– All target modes identified
– Comparison to analysis target modes
• MAC
• Orthogonality
• Root Sum Square of orthogonality terms

The target modes can be defined a number of ways, such as the following:

• All modes within a specified frequency range

• First n modes
• Modes defined using various criteria such as effective mass weighting

Establishing specific goals and metrics to be used in assessing the modal results
during test planning enables a clear evaluation path to be used in checking the modal
results. Some or all of these criteria may be applied depending on the overriding
reason for conducting the modal test. In cases where simple exploratory testing is
being performed, it may be sufficient to identify the first three flexible structural
modes, for example, without any other criteria. This criterion seems rather simple
but may serve the purpose of the test program. An even simpler criterion might be
to identify the first three frequencies, with no mode shapes. While this might seem
an oversimplification, it too may satisfy the test objectives in certain cases.
In most cases, however, the criteria applied are more extensive and require an
assessment of the modal parameters extracted from the test. MAC is often used as a
tool to check test mode shapes, particularly when an analysis model is not available
for comparison. Even if the model is available, cross-MAC might be a good check
if the model has not been used to develop a TAM and corresponding mass matrix, as
it helps in assessing whether the test mode shapes are comparable to those predicted
by analysis for both frequency and shape.
Orthogonality provides a more complete and rigorous assessment of the test
mode shapes than simply looking at MAC. This is particularly true for aerospace
structures where the mass normalization of the mode shapes helps ensure that the
modes’ mass-weighted behavior is checked, since the mass of the elements in the
mode shapes can skew the overall comparisons. Small, lightweight components can
have large mode shape coefficients, which cause mode shapes to appear similar even
though they do not produce substantial loads. Orthogonality (test results against
9 Design of Modal Tests 529

themselves) and cross-orthogonality (test versus analysis) are typical criteria applied
in detailed aerospace modal tests. For orthogonality, mass weighting is applied to
provide unit diagonal terms, and then the off-diagonal terms are evaluated to see
whether there is any shape coupling. Typical criteria are 10% or less for all off-
diagonal terms for all target modes.
As a result of test shape normalization, cross-orthogonality does not yield a
unit diagonal matrix. Instead, the terms closest to 100% provide an indication of
similarity between test and analysis shapes. Since differences between the model
and the physical test structure can result in various combinations of modal behavior
and shapes, it is often necessary to also check for the Root Sum Square of the cross-
orthogonality terms to determine whether various combinations of extracted test
shape behavior match those predicted by analysis.
In summary, developing an appropriate list of test criteria during the test planning
effort ensures that the right tools are in place for conducting these final checks
to ensure that the test has been successfully completed. Evaluating the criteria
before the test is started enables a good understanding of the methods that will be
used to assess test result quality and to determine when testing can be considered
complete.

7 Closure

In conclusion, planning or designing a modal test is just as important as the data

collection or the data analysis. One needs to clearly identify the objectives of the
test and the stake holders; then develop a test design that supports the realization
of those objectives. When a modal test is to be performed to validate a finite
element model, one needs to design the test so that the resulting measurements will
provide the data required for the correlation of modeling results with those from the
test. From a correlation perspective, one needs to select the response locations to
allow a definitive, one-to-one correspondence between the measured modes and the
predicted modes. Further, the excitation must be designed to excite all the modes of
interest at a sufficient level so that the modal estimation algorithms can accurately
extract the modal parameters.

References
1. Allemang RJ, Brown DL (1982) A correlations coefficient for modal vector analysis. In:
Proceedings of the 1st international modal analysis conference, pp 110–116
2. Avitabile P (1998) Which shaker excitation is best? Is there any difference? SEM Experimental
Techniques. https://www.uml.edu/docs/dec98_tcm18-189813.pdf
3. Avitabile P (2006) Modal space, back to basics. Exp Tech 30(3):19–20
4. Bisplinghoff RL, Ashley H, Halfman RL (1955) Aeroelasticity. Addison-Wesley Publishing
Company, Cambridge, MA, pp 771–779
5. Blelloch PA (1992) Direct correlation of test-analysis cross-orthogonality. In: Proceedings of
the 10th international modal analysis conference, pp 999–1003
530 T. Carne et al.

6. Brillhart R, Hunt D (2005) Lessons learned in modal testing—part 1: the pitfalls, pratfalls, and
downfalls of fixturing. Exp Tech 29(6):58–61
7. Brown DL, Allemang RJ, Phillips AW (2015) Forty years of use and abuse of impact testing:
a practical guide to making good FRF measurements. In: De Clerck J (ed) Experimental
techniques, rotating machinery, and acoustics, vol 8. Conference proceedings of the society
for experimental mechanics series
8. Carne TG, Dohrmann CR (1995) A modal test design strategy for model correlation. In:
Proceedings of the 13th international modal analysis conference, pp 927–933
9. Carne TG, Stasiunas EC (2006) Lessons learned in modal testing—part 3: transient excitation
for modal testing, more than just hammer impacts. Exp Tech:69–79
10. Carne TG, Lauffer JP, Gomez AJ (1988a) Modal testing of a very flexible 110m wind turbine
structure. In: Proceedings of the 6th international modal analysis conference, pp 845–855
11. Carne TG, Lauffer JP, Gomez AJ (1988b) Modal testing of an immense flexible structure using
natural and artificial excitation. Int J Anal Exp Modal Anal 3(4):117–122
12. Carne TG, Griffith DT, Casias ME (2007a) Support conditions for free boundary-condition
modal testing. In: Proceedings of the 25th international modal analysis conference
13. Carne TG, Griffith DT, Casias ME (2007b) Support conditions for experimental modal
analysis. Sound and Vibration Magazine, pp 10–15
14. Deiters T, Hunt D, Osborne E (1994) A comparison of excitation methods to the Taurus vehicle
modal test. In: Proceedings of the 12th international modal analysis conference
15. Ewins DJ (1984) Modal testing: theory and practice. Wiley, New York, p 225
16. Ewins DJ (2000) Modal testing: theory and practice and application, 2nd edn. Research Studies
Press, Letchworth
17. Guyan RJ (1965) Reduction of mass and stiffness matrices. AIAA J 3:380
18. Hasselman TK, Coppolino RN, Zimmerman DZ (2000) Criteria for modeling accuracy –
a state-of-the-practice survey. In: Proceedings of the 18th international modal analysis
conference
19. James GH, Carne TG, Lauffer JP (1995) The natural excitation technique (NExT) for modal
parameter extraction from operating wind turbines. Int J Anal Exp Modal Anal 10(4):260–277
20. Kammer DC (1991) Sensor placement for on-orbit modal identification and correlation of large
space structures. J Guid Control Dyn 14:251–259
21. Kammer DC (1996) Optimal sensor placement for modal identification using system-
realization methods. J Guid Control Dyn 19(3):729–731
22. Kammer DC (2004) Optimal placement of triaxial accelerometers for modal vibration tests.
Mech Syst Signal Process 18:29–41
23. Kammer DC, Peck J (2008) Mass weighting for sensor placement using sensor set expansion
techniques. Mech Syst Signal Process 22:1515–1525
24. Klosterman AL (1971) On the experimental determination and use of modal representations
of dynamic characteristics. Doctor of philosophy dissertation, Department of Mechanical
Engineering, University of Cincinnati
25. Maia NMM, Silva JMM (1997) Theoretical and experimental modal analysis. Research Studies
Press, Baldock, p 32
26. Napolitano, KL , Blelloch PA (2003) Automated selection of shaker locations for modal tests.
In: Proceedings of the 21st international modal analysis conference on structural dynamics
27. Poston WL, Tolson RH (1992) Maximizing the determinant of the information matrix with the
effective independence method. J Guid Control Dyn 15:1513–1514
28. Salama M, Rose T, Farga J (1987) Optimal placement of excitation and sensors for verification
of large dynamical systems. In: Proceedings of the 28th AIAA/ASME structure, structural
dynamics, and material conference, pp 1024–1031
29. Shah PC, Udwadia FE (1978) Methodology for optimal sensor locations for identification of
dynamic systems. J Appl Mech 45:188–196
30. Slater JC, Belvin WK, Inman DJ (1993) A survey of modern methods for modeling frequency
dependent damping in finite element models. In: Proceedings of the 11th international modal
analysis conference, pp 1508–1512
9 Design of Modal Tests 531

31. Udwadia FE, Garba JA (1985) Optimal sensor locations for structural identification. In:
JPL Proceedings of the workshop on identification and control of flexible space structures,
pp 247–261
32. Williams R, Crowley J, Vold H (1985) The multivariate mode Indicator function in modal
analysis. In: Proceedings of the international modal analysis conference, pp 66–70
33. Wolf JA Jr (1984) The influence of mounting stiffness on frequencies measured in a vibration
test. SAE Paper 840480, Society of Automotive Engineers
Experimental Modal Analysis Methods
10
R. J. Allemang and D. L. Brown

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
2 Modal Parameter Estimation: Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
2.1 Assumptions, Definitions, and Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
2.2 Analytical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
2.3 Experimental Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
3 Single Degree of Freedom Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
3.1 SDOF Algorithms: Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
3.2 Operating Vector (Peak-Pick) Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
3.3 Complex Plot (Circle Fit) Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
3.4 Two-Point Finite Difference Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
3.5 Least-Squares (Local) SDOF Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
3.6 Least-Squares (Global) SDOF Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
3.7 Other SDOF Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580
4 Multiple Degree of Freedom Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580
4.1 General (Two-Stage) Solution Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
4.2 Current MPE Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
4.3 Kernel Equations: Time Domain Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
4.4 Kernel Equations: Frequency Domain Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
4.5 Residue (Modal Vector) Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
5 Differences in MPE Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
5.1 Polynomial Coefficient Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
5.2 Generalized Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
5.3 Data Sieving/Filtering/Decimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
5.4 Coefficient Condensation (Virtual DOFs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602
5.5 Equation Condensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610

R. J. Allemang () · D. L. Brown

Structural Dynamics Research Laboratory, Mechanical and Materials Engineering, University of
Cincinnati, Cincinnati, OH, USA
e-mail: [email protected]; [email protected]

© The Society for Experimental Mechanics 2022 533

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_36
534 R. J. Allemang and D. L. Brown

Abstract

This chapter provides the basis and background of all experimental modal
analysis (EMA) methods that have been developed over the last fifty years.
In this context, modal parameters refer to complex valued modal frequencies,
complex valued modal vectors and complex valued modal scaling. The chapter
focusses on modal parameter estimation (MPE) methods that have been or
are commercially available but includes many related MPE methods that have
been developed and presented in research journals and articles as well. The
methods are mostly based upon experimentally measured frequency response
function (FRF) or impulse response function (IRF) data. MPE methods that are
fundamentally single input, single output (SISO) methods finding one single
mode are included through modern multiple input, multiple output (MIMO)
methods that find all modal parameters for all modes simultaneously (in one
or two passes). Discussion includes the theoretical background of all methods
along with the kernel equations for each method. The mathematical development
utilizes a central concept of matrix coefficient polynomials that provide the
basis of the unified matrix polynomial approach (UMPA). Basic definitions are
included as concepts are developed and a complete set of historical references is
provided.

Keywords

Modal parameter estimation · Experimental modal analysis · Modal parameter

estimation methods · SDOF methods · MDOF methods · Polynomial models ·
Unified matrix coefficient polynomial approach (UMPA) · Partial fraction
models · Residue and residual estimation · Complex z mapping

Nomenclature

Ni = Number of inputs
No = Number of outputs
NS = Short dimension (min(Ni , No ))
NL = Long dimension (max(Ni , No ))
Nf = Number of frequencies (spectral lines)
Nt = Number of times
Ne = Number of effective modal frequencies
N = Number of modal frequencies
Fmax = Maximum frequency (Hz)
ωi = Frequency (rad/sec)
ωmax = Maximum frequency (rad/sec)
f = Frequency resolution (Hz)
λr = Complex modal frequency
T = Observation period (sec)
si = Generalized frequency variable
m = Model order for denominator polynomial
n = Model order for numerator polynomial
Apqr = Residue, output DOF p, input DOF q, mode r
10 Experimental Modal Analysis Methods 535

RIpq = Residual inertia, output DOF p, input DOF q

RFpq = Residual flexibility, output DOF p, input DOF q
[C] = Companion matrix
[α] = Denominator polynomial matrix coefficient
[β] = Numerator polynomial matrix coefficient
[I ] = Identity matrix
[H (ωi )] = Frequency response function matrix. (No × Ni )
[T ] = Transformation matrix
[U ] = Left singular vector matrix
[] = Singular value matrix (diagonal)
[] = Eigenvalue matrix (diagonal)
[V ] = Right singular vector, or eigenvector, matrix
MPE-1 = First stage of modal parameter estimation
MPE-2 = Second stage of modal parameter estimation

1 Introduction

Experimental modal analysis is the process of determining the modal parameters

of a structural system (frequencies, damping factors, modal vectors, and modal
scaling values) from experimental input-output data, normally frequency response
functions (FRFs) or impulse response functions (IRFs). This is in contrast to
analytical modal analysis where the modal parameters are found from a theoretical,
continuous, or discrete model of the structural system. This could be a closed form
solution for simple structures or, more commonly, a discrete model like a finite
element model for structural systems with more complex geometry. Analytical
modal analysis generally does not include damping factors as part of the solution.
Finally, operational modal analysis (OMA), sometimes called response-only
modal analysis, attempts to determine modal parameters from data taken without
measured inputs. The inputs come from natural excitation (wind, waves, traffic,
etc.) or from the operational inputs provided to a structural system. Operational
modal analysis gives a subset of modal parameters that depends upon whether the
unmeasured inputs excite the modal parameters and may mix modal parameters with
forced vibration information. Since operational modal analysis does not measure the
inputs, modal scaling cannot be determined without additional testing. Operational
modal analysis is the subject of the “Operational Modal Analysis Methods” chapter
in this Handbook. The following discussion is concerned with experimental modal
analysis. Analytical and operational modal analysis is covered in other related
material.
One important continuing focus of experimental modal analysis is the presenta-
tion of modal parameter estimation algorithms in a single, consistent mathematical
formulation with a corresponding set of definitions and unifying concepts [1, 2, 3,
4, 5, 6]. In particular, a matrix coefficient polynomial approach can be used to unify
the presentation with respect to current algorithms such as the least-squares complex
exponential (LSCE), polyreference time domain (PTD), polyreference least-squares
complex frequency (LSCF), Ibrahim time domain (ITD), Eigensystem realization
algorithm (ERA), rational fraction polynomial (RFP), orthogonal polynomials
536 R. J. Allemang and D. L. Brown

(OP), polyreference frequency domain (PFD), and the complex mode indication
function (CMIF) methods. Using this unified matrix polynomial approach (UMPA)
encourages a discussion of the similarities and differences of the commonly used
methods, as well as a discussion of the numerical characteristics. Some of the
different numerical methods that are used in different methods are the least squares
(LS), total least squares (TLS), double least squares (DLS), and singular value
decomposition (SVD) methods (to take advantage of redundant measurement data)
and the eigenvalue and singular value decomposition transformation methods (to
reduce the effective size of the resulting eigenvalue-eigenvector problem).
The unified matrix polynomial approach (UMPA) is an attempt to place most
commonly used experimental modal parameter estimation algorithms within a
single educational framework. The goal of the UMPA presentation is to highlight the
similarity between the different algorithms rather than differences. This approach
does not attempt to explain the detailed development of the authors who originated
each method but attempts to present a common framework so that different
algorithms can be easily compared and contrasted.

2 Modal Parameter Estimation: Background

Modal parameter estimation is a special case of system identification where the a

priori model of the system is known to be in the form of modal parameters. Over
the past 40 years, a number of algorithms have been developed to estimate modal
parameters from measured multiple input, multiple output (MIMO), frequency
response function (FRF), or impulse response function (IRF) data. While most
of these individual algorithms, summarized in Table 1, are well understood, the
comparison of one algorithm to another has become one of the thrusts of current
research in this area. Comparison of the different algorithms is possible when the
algorithms are reformulated using a common mathematical structure.
This reformulation attempts to characterize different classes of modal param-
eter estimation techniques in terms of the structure of the underlying matrix
coefficient polynomials rather than the physically based models used historically.
Since the modal parameter estimation process involves a greatly overdetermined
problem (more data than independent equations), this reformulation is helpful
in understanding the different numerical characteristics of each algorithm and,
therefore, the slightly different estimates of modal parameters that each algorithm
yields. As a part of this reformulation of the algorithms, the development of a
conceptual understanding of modal parameter estimation technology has emerged.
This understanding involves the ability to conceptualize the measured data in terms
of the concept of characteristic space, the data domain (time, frequency, spatial),
the dimension of the measured data, the evaluation of the order of the problem,
the condensation of the data, and a common parameter estimation theory that can
serve as the basis for developing any of the algorithms in use today. The following
sections review these concepts as applied to the current modal parameter estimation
methodology.
10 Experimental Modal Analysis Methods 537

Table 1 Acronyms – experimental modal parameter estimation algorithms

Modal parameter estimation algorithms
CEA Complex exponential algorithm [7, 8]
LSCE Least-squares complex exponential [8]
PTD Polyreference time domain [9, 10]
ITD Ibrahim time domain [11, 12]
MRITD Multiple reference Ibrahim time domain [13]
ERA Eigensystem realization algorithm [14, 15, 16]
PFD Polyreference frequency domain [17, 18, 19, 20]
FDPI Frequency domain direct parameter identification [19, 20]
SFD Simultaneous frequency domain [21]
MRFD Multi-reference frequency domain [22]
RFP Rational fraction polynomial [23, 24, 25]
OP Orthogonal polynomial [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36]
PLSCF Polyreference least-squares complex frequency [37, 38, 39, 40, 41, 42]
CMIF Complex mode indication function [43]

2.1 Assumptions, Definitions, and Concepts

A number of assumptions, basic definitions, and general concepts are essential to

understanding the experimental modal analysis process. Since there is considerable
history in the development of modal parameter estimation methodology, it is helpful
to provide a structure that provides a common basis for all modal parameter
estimation algorithms. Some of these definitions and concepts are briefly introduced
in the following subsections.

2.1.1 Assumptions
By its very nature, modal analysis is one way to describe the dynamic characteristics
of a structural system. For modal analysis to be considered, there are several
assumptions involved. Generally, the structural system is assumed to be linear,
time invariant, and reciprocal. Other assumptions can be involved but are not
a requirement. An example would be proportional or Rayleigh damping which
restricts the form of the modal vectors that are found.

2.1.2 Definition: Modal Parameters

While some situations do not require all modal parameters, a complete set of
modal parameters is needed if a complete model of the input-output relationships
is desired. Modal parameters include the complex valued modal frequencies (λr ),
the associated complex valued modal vectors ({ψr }), and the complex valued modal
scaling (Modal Ar ). Note that the complex valued modal frequencies are of the
form (λr = σr + j ωr ) where σr is the damping factor and ωr is the damped natural
frequency for the r − th mode. Additionally, most current multiple input, multiple
output (MIMO) algorithms estimate modal participation (weighting) vectors ({Lr })
and residue vectors ({Ar }) as part of the overall process of estimating the complex
538 R. J. Allemang and D. L. Brown

valued modal vectors. Modal participation vectors are a result of MIMO modal
parameter estimation algorithms and relate how well each modal vector is excited
from each of the reference locations included in the measured data. The combination
of the modal participation vector ({Lr }), the modal vector ({ψr }), and the Modal A
(MAr ) for a given mode yields the residue matrix ([Ar ]) for that mode.
In general, modal parameters are considered to be global properties of the system.
The concept of global modal parameters simply means that there is only one
answer for each modal parameter and that the modal parameter estimation solution
procedure enforces this constraint. Most of the current modal parameter estimation
algorithms estimate the modal frequencies and damping in a global sense but very
few estimate the modal vectors in a global sense. This is due to various modal vector
scaling normalization methods which, together with the modal scaling value, give
equivalent results but not unique, mathematical values.

Modal Vector Normalization

While there is a unique answer for the modal frequencies, the complex valued
modal vectors represent the relative pattern of motion associated with each complex
valued modal frequency. This relative pattern results from a rank-deficient system of
equations where the modal vector is found via an eigenvalue-eigenvector solution.
The complex valued modal vectors, together with the modal scaling, represent the
unique characteristic for each modal vector. For this reason, the normalization of
the complex valued modal vector is important when modal vectors or the associated
modal scaling values are compared numerically.
While there are a number of acceptable normalization schemes for modal vectors,
choosing the largest element in each complex valued modal vector equal to unity is
the most common and most useful in a physical sense. This method of normalization
will be used in all following discussions. Recognizing that an arbitrarily scaled
modal vector is complex valued means that the largest element will in general be
complex valued. When the arbitrarily scaled modal vector is divided (normalized)
by the largest element, this will force all elements of the scaled modal vector to be
bounded by the complex unit circle and, for most cases, to lie along the real axis of
the unit circle.

2.1.3 Definition: Degrees of Freedom (DOFs)

Degrees of freedom (DOFs) refer to the physical location and direction of all of the
potential inputs to, or outputs from, a structural system. For the theoretical problem,
the number of inputs and the number of outputs are always the same and are equal
to the number of modal parameter sets of information (typically designated N ) that
will be found. The general notation that is often used involves inputs in the form of
forces and outputs in the form of displacements, velocities, and/or accelerations. The
generalized concept of inputs and outputs could be used to represent translational or
rotational outputs (or their derivatives) or translational or rotational inputs.
10 Experimental Modal Analysis Methods 539

In an experimental sense, the DOFs are where the input and output sensors
are located (physical location and direction). Today that includes response mea-
surements from scanning laser vibrometers and digital image correlation (DIC)
photogrammetry methods. Equation 1 is a representation of an FRF where the output
DOF utilizes the notation p and the input DOF utilizes the notation q. Note that the
X and F notation represents the output and input in a general way and does not imply
displacement or force.

Xp (ω)
Hpq (ω) = (1)
Fq (ω)

• p is the output degree of freedom (physical location and orientation).

• q is the input degree of freedom (physical location and orientation).

For the experimental case, the number of input DOFs (Ni ) and output DOFs
(No ) is not the same and in general cannot be directly linked to the N sets of modal
parameters that will be estimated.

2.1.4 Concept: Experimental Modal Parameter Estimation

Experimental modal parameter estimation involves estimating the modal parameters
of a structural system from measured input-output data. The experimental approach
originally involved methods that are referred to as phase resonance methods using
sinusoidal excitation (narrowband) and mode by mode tuning with a forcing
vector to balance the damping effect, assuming the modal vectors are normal
modes. These methods gave way to phase separation methods that involve a
range of frequency information (broadband) where the effects of several modes
are separated using mathematical models for the experimental data. Most current
modal parameter estimation are phase separation methods, based upon the measured
data being the frequency response function or the equivalent impulse response
function, typically found by inverse Fourier transforming the frequency response
function.
The current approach is to use numerical techniques to separate the contributions
of individual modes of vibration in measurements such as frequency response
functions. The concept involves estimating the individual single degree of freedom
(SDOF) contributions to the multiple degree of freedom (MDOF) measurement.
Then, the modal parameters are found from the SDOF contributions.


N
[Ar ]NL ×NS [A∗r ]NL ×NS 
2N
[Ar ]NL ×NS
[H (ωi )]NL ×NS = + = (2)
j ωi − λr j ωi − λ∗r j ωi − λr
r=1 r=1
540 R. J. Allemang and D. L. Brown

Equation 2 represents a mathematical problem that, at first observation, is

nonlinear in terms of the unknown modal parameters. Once the modal frequencies
(λr ) are known, the mathematical problem is linear with respect to the remaining
unknown modal parameters ([Ar ]). For this reason, the numerical solution in many
modal parameter estimation (MPE) algorithms frequently involves two linear stages
that utilize least-squares (LS) solution methods. Typically, the modal frequencies
and modal participation vectors are found in a first stage (MPE-1), and residues,
modal vectors, and modal scaling are determined in a second stage (MPE-2).
This concept, involving a summation of partial fraction terms, is represented
mathematically in Eq. 2 and graphically in Fig. 1.
While the model stated in Eq. 2 is fundamental to the linear superposition of
individual SDOF contributions, this model is normally limited to being used as the
basis for estimating the residues Apqr once the modal frequencies (λr ) are known.
Based upon the speed and memory of modern personal computers, these two stages
are often executed together giving the appearance of a single stage solution.

20
0
−20
−40
Phase (Deg)

−60
−80
−100
−120
−140
−160
−180
0 50 100 150 200 250 300 350 400 450 500
−40

−50

−60
Log Magnitude (dB)

−70

−80

−90

−100
0 50 100 150 200 250 300 350 400 450 500

Frequency (Hertz)

Fig. 1 MDOF – superposition of SDOF (positive frequency poles)

10 Experimental Modal Analysis Methods 541

Equation 2 can be represented in the time domain in terms of impulse response

functions in a similar summation form as shown in Eq. 3:


N
∗ 
2N
[h(ti )]NL ×NS = [Ar ]NL ×NS e λr ti
+ [A∗r ]NL ×NS eλr ti = [Ar ]NL ×NS eλr ti
r=1 r=1
(3)

2.1.5 Concept: Data Domain

Modal parameters can be estimated from a variety of different measurements that
exist as discrete data in different data domains (time and/or frequency). These
measurements can include free decays, forced responses, power spectra, covariance,
frequency response functions (FRFs), or impulse response functions (IRFs). These
measurements can be processed one at a time or in partial or complete sets simul-
taneously. The measurements can be generated with no measured inputs, a single
measured input, or multiple measured inputs. The data can be measured individually
or simultaneously. There is a tremendous variation in the types of measurements
and in the types of constraints that can be placed upon the testing procedures used
to acquire this data. For most measurement situations, FRFs are utilized in the
frequency domain, and IRFs are utilized in the time domain. When IRFs are utilized,
they are generally formed from the inverse Fourier transform of the measured
FRFs. For the purpose of the following discussions, force-displacement data is
considered the basis for the equations recognizing that force-displacement data can
be numerically synthesized from measured force-acceleration data or force-velocity
data.

2.1.6 Concept: Characteristic Space

From a conceptual viewpoint, the measurement space of a modal identification
problem can be visualized as occupying a volume with the coordinate axes defined
in terms of the three sets of characteristics. Two axes of the conceptual volume
correspond to spatial information and the third axis to temporal information. The
spatial axes are in terms of the input and output degrees of freedom (DOF) of the
system. The temporal axis is either time or frequency depending upon the domain
of the measurements. These three axes define a 3D volume which is referred to as
the Characteristic Space (Figs. 2, 3, and 4).
This space or volume represents all possible measurement data. This conceptual
representation is very useful in understanding what data subspace has been mea-
sured. Also, this conceptual representation is very useful in recognizing how the
data is organized and utilized with respect to different modal parameter estimation
algorithms (3D volume to stacked 2D matrices). Information parallel to one axis
consists of a superposition of the characteristics defined by that axis. The other two
characteristics determine the scaling of each term in the superposition.
Any structural testing procedure measures a subspace of the total possible data
available. Modal parameter estimation algorithms may then use all of this subspace
or may choose to further limit the data to a more restrictive subspace via sieving
542 R. J. Allemang and D. L. Brown

Fig. 2 Characteristic space: columns of measurements. (a) Single input reference. (b) Multiple
input references

Fig. 3 Characteristic space: rows of measurements. (a) Single output reference. (b) Multiple
output references

and/or filtering. It is theoretically possible to estimate the characteristics of the total

space by measuring any subspace which samples all three characteristics. Measure-
ment data spaces involving many planes of measured data are the best possible
modal identification situations since the data subspace includes contributions from
temporal and spatial characteristics. The particular subspace which is measured
and the weighting of the data within the subspace in an algorithm are the main
10 Experimental Modal Analysis Methods 543

Fig. 4 Characteristic space: times/frequencies of measurements. (a) Single time/frequency. (b)

Multiple times/frequencies

differences between the various modal identification procedures which have been
developed.
It should be obvious that the data which defines the subspace needs to be acquired
in a consistent measurement process in order for the algorithms to estimate accurate
modal parameters. This fact has triggered the need to measure all of the data
simultaneously and has led to recent advancements in data acquisition, digital signal
processing, and instrumentation designed to facilitate this measurement problem.
While it is not always obvious, most modal parameter estimation methods
assume that the measured data includes one or more pairs of matched input and
output DOFs in what is commonly referred to as driving point FRFs.

2.1.7 Concept: Data Dimensionality

In the following discussion, the use of notation involving the number of input
degrees of freedom (DOFs), (Ni ), and the number of output DOFs, (No ), is replaced
by an alternate notation. Since the FRF matrix is always assumed to be reciprocal,
it is more important to note the smaller and larger of the number of DOFs with
respect to the inputs and outputs when the data is utilized by modern MIMO modal
parameter estimation algorithms. For this reason, the dimension that is larger is
referred to as the long dimension, (NL ), and the dimension that is smaller is referred
to as the short dimension, (NS ). The short dimension is often referred to as the
number of references, and the long dimension is often referred to as the number of
responses. The dimension of the third axis (temporal axis) is either the number of
frequencies (Nf ) or the number of times (Nt ).
544 R. J. Allemang and D. L. Brown

2.1.8 Concept: Generalized Frequency

In the following development, all frequency domain models will be presented
as a function of the generalized frequency variable s. This variable is a general
complex valued variable that is most often thought of as s = j ω, representing the
independent frequency variable associated with the FRF. There are several other
scaled or mapped versions of the generalized frequency variable s which are useful
for minimizing the numerical conditioning associated with different frequency
domain MPE algorithms.

Xp (ωi ) βn (si )n + βn−1 (si )n−1 + · · · + β1 (si )1 + β0 (si )0

Hpq (ωi ) = =
Fq (ωi ) αm (si )m + αm−1 (si )m−1 + · · · + α1 (si )1 + α0 (si )0
(4)
One example is a simple scaled version of s where s = ωjmax ω
. When the
generalized frequency variable is scaled or mapped in this fashion, any frequency
estimate that results must be corrected accordingly. This concept will be explained
further in the section on high-order, frequency domain algorithms.

2.1.9 Concept: Kernel Equations

Each experimental modal parameter estimation method has a basic equation that is
repeated for different powers or subspaces of the independent variable (t, s, ω, z)
to get additional sets of solutions. The base equation for each method is referred
to as the kernel equation. Consistency between the different sets of solutions is one
common method of determining the most likely modal parameters in the data.

2.1.10 Concept: Overdetermined Linear Models

Most current modal parameter estimation methods utilize linear models, sometimes
in several successive solution steps, that have more equations than unknowns in each
solution step. This will be true for simple SDOF models as well as complicated
MIMO MDOF models. The number of equations is larger than the unknowns since
an equation can be formed at each measured frequency or time, while the number
of unknowns is limited to a function of the number of modal parameters N. The
overdetermination factor is the ratio of the number of equations to the number of
unknowns. The overdetermination factor can easily be a number greater than 2-5.
Overdetermined sets of linear equations are most often solved using least-squares
(LS) techniques, particularly when the noise on the data is expected to be random.
The least-squares solution will yield the best solution in the presence of random
noise. With respect to the estimation of frequency response functions (FRFs) as the
measured data for most modal parameter estimation, this means that bias errors on
the FRF data are much more concerning than the random errors.

2.1.11 Concept: General (Two-Stage) Solution Procedure

Most modern modal parameter estimation (MPE) algorithms are implemented in
two stages that each involves the solution of overdetermined linear equations. The
first stage, referred in this text as MPE-1, is where the modal frequencies (λr ) are
estimated along with associated modal weighting vectors ({Lr }). The second stage,
10 Experimental Modal Analysis Methods 545

referred in this text as MPE-2, is where the modal vectors ({ψr }) and modal scaling
are estimated. These two stages can generally be briefly summarized as follows:

• MPE-1: Utilize measured FRF or IRF data with a matrix coefficient, polynomial
model to find multiple estimates of the modal parameters. Select the best set of
modal frequencies and modal participation vectors. This stage is often performed
in two fundamental steps: (1) solve a set of overdetermined linear equations and
(2) solve for the roots of a polynomial.
• MPE-2: Utilize measured FRF or IRF data, with selected complex valued
modal frequencies and modal participation vectors to find complex valued modal
vectors from an overdetermined set of linear equations.

Additional steps are often added to the above procedure to enhance the numerical
solution. This concept will be described in greater detail in a later section.

2.1.12 Concept: Equation Normalization

Many of the overdetermined sets of linear equations involve a null space problem in
which the right-hand side (RHS) of the equation is zero or null. In these cases, the
unknown coefficients of the equation are non-unique but can be found by choosing
one of the coefficients, typically, equal to the identity matrix. The numerical solution
of the overdetermined sets of linear equations will yield different answers depending
on which coefficient is chosen. While theoretical data would yield only one answer,
experimental data will have random and bias noise, and this noise is what yields
somewhat different answers. Historically, two solutions have been used in these
cases, setting the lead coefficient and the last coefficient to the identity matrix to
give two answers that bracket the solution. This is true for both frequency domain
methods and time domain methods. This choice of choosing a coefficient to be the
identity is referred to as equation normalization.
 m   n  
[αk ](si )k [H (ωi )] = [βk ](si )k [I ] (5)
k=0 k=0

Equation 5 is an example of such a null space (rank deficient) equation. The

unknowns in the equation are the [α] and [β] coefficients. Since every term has an
unknown coefficient matrix, the equation can be pre-multiplied by the inverse of any
one of the [α] coefficients to derive a solvable (full rank) base equation.
Lowest [α0 ] Normalization: [α0 ] = [I ]
m 
  n 
   
[αk ](si )k [H (ωi )] − [βk ](si )k [I ] = − [α0 ](si )0 [H (ωi )] (6)
k=1 k=0

Highest [αm ] Normalization: [αm ] = [I ]


m−1  n 
   
[αk ](si )k [H (ωi )] − [βk ](si )k [I ] = − [αm ](si )m [H (ωi )] (7)
k=0 k=0
546 R. J. Allemang and D. L. Brown

It is important to note that this equation normalization must occur in the initial
step of choosing the base equation. Then the unknown coefficients are solved in
an overdetermined set of these linear equations. It is not sufficient to manipulate
the final solution of polynomial coefficients into a different normalized form. The
difference that equation normalization provides is determined by the LS solution for
the coefficients [39, 41].

2.2 Analytical Models

While analytical models can be continuous, represented in closed form, in terms of

time and/or frequency, most analytical models of structures are discrete, represented
by mass [M], damping [C], and stiffness [K] matrices in a second-order system of
equations. An equivalent discrete analytical model in terms of [A] and [B] matrices
in a first-order model provides a more direct relationship to the modal parameters
found in experimental approaches.

2.2.1 [M] [C] [K] Models

Generally, most structures are more complicated than the single degree of freedom
mass, spring, and damper system. The general case for a multiple degree of freedom
system will be used to show how the frequency response functions of a structure are
related to the modal vectors of that structure.

[M] {ẍ(t)} + [C] {ẋ(t)} + [K] {x(t)} = {f (t)} (8)

 
[M] s 2 + [C] s + [K] {X(s)} = {F (s)} (9)

The transfer function representation of a general multiple degree of freedom

system can be formulated by starting with generalized frequency domain equation
of motion in terms of mass, stiffness, and damping matrices.
 
[M] s 2 + [C] s + [K] {X (s)} = {F (s)} (10)
 
[M] s 2 + [C] s + [K] {X (s)} = [B (s)] {X (s)} = {F (s)} (11)

 
where [B (s)] = [M] s 2 + [C] s + [K] and [B (s)] is referred to as the system
impedance matrix or just the system matrix. Pre-multiplying both sides of the
previous equation by [B (s)]−1 yields:

[B (s)]−1 {F (s)} = [H (s)] {F (s)} = {X (s)} (12)

where [H (s)] = [B (s)]−1 and is referred to as the transfer function matrix. By

evaluating the transfer function matrix at s = j ω, the frequency response function
matrix results.
10 Experimental Modal Analysis Methods 547

 −1
[H (ω)] = −ω2 [M] + j ω [C] + [K] (13)

2.2.2 [A] [B] Models

A more general representation of the matrix model can be formed from the [M] [C]
[K] model. This model is in the form of a first-order differential equation utilizing
matrices that are combinations of the [M] [C] [K] matrices and is more desirable
due to its direct relationship to the generalized eigenvalue-eigenvector form. This
methodology is often referred to as a state space expansion due to its final form.
However, this methodology was documented in the 1930s by Frazer, Duncan, and
Collar [44] to put the [M] [C] [K] model into Hamilton’s dynamical equations and
has since been put in this final form by others [45, 46]. This development begins
with the [M] [C] [K] differential equation:

[M] {ẍ(t)} + [C] {ẋ(t)} + [K] {x(t)} = {f (t)} (14)

This system of equations can be augmented by the following identity:

[M] {ẋ(t)} − [M] {ẋ(t)} = {0} (15)

The two above equations can be combined as follows to yield a new system of
2N equations in a classical eigenvalue solution form.
   
[0] [M] ẍ(t) −[M] [0] ẋ(t) 0
+ = (16)
[M] [C] ẋ(t) [0] [K] x(t) f (t)

[A] {ẏ(t)} + [B] {y(t)} = f  (t) (17)

where
   
[0] [M] −[M] [0] ẋ(t)
• [A] = [B] = {y(t) } =
[M] [C] [0] [K] x(t)

The generalized frequency domain equation is therefore:

[[A] s + [B]] {Y (s)} = F  (s) (18)

The advantage of representing the [M], [C], [K] model in the [A], [B] form is
that the equation becomes a first-order problem that can be solved with standard
eigenvalue-eigenvector solution methods. This form gives diagonalized matrices
when the weighted orthogonality concept is applied, similar to the diagonalized
matrices when the weighted orthogonality concept is considered for the undamped
and proportionally damped cases.
548 R. J. Allemang and D. L. Brown

Also note that the well-known Hamilton’s dynamical equations result when
Eqs. 14 and 15 are first pre-multiplied by the inverse of the mass matrix.
   
[0] [I ] ẍ(t) −[I ] [0] ẋ(t) 0
+ =
[I ] [M]−1 [C] ẋ(t) −1
[0] [M] [K] x(t) [M]−1 {f (t)}
(19)

Note that Eq. 19 is just a normalized form of Eq. 16.

2.2.3 [A] [B] [C] [D] Models

Another theoretical methodology that is used by some researchers to determine
modal parameters is based upon a system identification approach that involves a
four matrix form. The model uses input-output data and a MIMO state space process
that includes inputs, outputs, noise on the inputs, noise on the outputs, and noise on
the state.

˙ = [A] {y(t)} + [B] {f (t)}

y(t) (20)

{x(t)} = [C] {y(t)} + [D] {f (t)} (21)

This time domain approach is commonly used in control theory, system identifi-
cation modeling. With respect to the [M], [C], [K] model, the [A] matrix in the above
equation is proportion to the [A] matrix in Eq. 19, but the other matrices are not
directly related to any other matrices defined in the previous sections [47,48,49,50].
This method has been adapted to utilize frequency domain data in the form of
frequency response functions using an algorithm referred to as N4SID.
This method, however, is not commonly used by any commercial algorithms and
will not be discussed further in this text.

2.2.4 Eigen-Solutions, Orthogonality and Modal Scaling

The modal parameters for analytical models can be found directly from the homo-
geneous form of Eq. 10 or from Eq. 18 using a number of numerical methods. Both
equations represent a rank-deficient system of equations where the complex valued
frequencies are found first, followed by a solution for the complex valued modal
vectors. This is complicated for a generally damped system of more than two or
three degrees of freedom. However, the general solution of any size problem can be
easily found using eigenvalue-eigenvector methods when the [A] [B] matrix model
is used. Any specialized form of damping, including undamped and proportionally
damped models, is likewise found using this approach.
The solution of the homogeneous equation for the [A] [B] matrix model yields
the complex valued natural frequencies (eigenvalues) and complex valued modal
vectors (eigenvectors) once an augmented 2N equation system is formed. Note that
in this mathematical form, the complex valued eigenvalues will directly yield the
10 Experimental Modal Analysis Methods 549

complex valued modal frequencies, while the complex valued modal vectors will
be found from the 2N eigenvectors, which are each 2N in length. Rearranging
the homogeneous form of Eq. 18 puts the equation into a standard generalized
eigenvalue problem form.

[B] {Y (s)} = −s [A] {Y (s)} (22)

   
−[M] [0] [0] [M]
{Y (s)} = − λ {Y (s)} (23)
[0] [K] [M] [C]

where

λ {X(s)}
• Y (s) =
{X(s)}

The standard form of the eigenvalue-eigenvector solution method can cause some
confusion since the standard mathematical notation is similar to, but different from,
Eq. 18 as follows:
 
à {Y (s)} = λ [I ] {Y (s)} (24)

   
à {Y (s)} = λ B̃ {Y (s)} (25)

Note that the [Ã] and [B̃] matrices in Eq. 25 are in different positions in the
equation compared to the [A] and [B] matrices in Eq. 22. Also note the difference
in the eigenvector notation.
The 2N eigenvalues give the 2N complex valued modal frequencies, which result
in N complex conjugate pairs of complex valued modal frequencies. The exact form
of the eigenvectors can be seen from the associated modal matrix for the 2N × 2N
equation system. Note that the notation {φ} is used for an eigenvector in the 2N ×2N
equation system and that the notation {ψ} is used for the modal vector of the N × N
equation system.
The modal matrix or modal state matrix [φ] for this generally damped system
can be represented as follows:
 
[φ] = {φ}1 {φ}2 {φ}3 ...{φ}r ...{φ}2N (26)
 
λ1 {ψ}1 λ2 {ψ}2 λ3 {ψ}3 ... λr {ψ}r ... λ2N {ψ}2N
[φ] = (27)
{ψ}1 {ψ}2 {ψ}3 ... {ψ}r ... {ψ}2N

Note that this matrix involves the N × 2N matrix of modal vectors in the lower
half of the matrix and the N × 2N matrix of the modal vectors, each modified by
the appropriate complex valued modal frequency, in the upper half of this matrix.
550 R. J. Allemang and D. L. Brown

These sub-matrices are not square since both the modal vector and the complex
conjugate of the modal vector are involved. Also note that the upper half of this
matrix is often viewed as the derivative of the lower half of the matrix (each upper
vector is the derivative of the corresponding lower vector). Finally, each column with
information associated with one modal vector is often described as a state vector.

Orthogonality

Since the [A] and [B] matrices form an eigenvalue problem, the associated
eigenvectors are weighted orthogonal with respect to these matrices.

[φ]T [A] [φ] = MA  (28)

[φ]T [B] [φ] = MB  (29)

Weighted orthogonality means that the [MA ] and [MB ] matrices are diagonal
with scaling terms, for each of the associated 2N modal eigenvectors, on the
diagonal. The characteristics (magnitude and phase) of the scaling terms will depend
on how the eigenvectors are normalized/scaled.
Note that for the generally damped problem, the weighted orthogonality concept
will not extend to the mass and stiffness matrices and the lower half of the
eigenvector matrix noted in Eq. 25 even when the 2N eigenvectors are reduced to
the N modal vectors associated with the N positive modal frequencies.

[ψ]T [M] [ψ] = M̃ (30)


[ψ]T [K] [ψ] = K̃ (31)

If, and only if, the damping matrix represents some form of proportional damping
(which includes no damping as a trivial case), then the weighted orthogonality
concept will extend to the mass and stiffness matrices.

[ψ]T [M] [ψ] = M̃ (32)

[ψ]T [K] [ψ] = K̃ (33)

Modal Scaling

When weighted orthogonality applies, the diagonal terms become what is

referred to as modal scaling. For the generally damped problem, the terms on the
diagonal of the [MA ] and [MB ] matrices are defined as the Modal A (MAr ) and
Modal B (MB r ) scaling terms, for each of the associated 2N modal eigenvectors.
10 Experimental Modal Analysis Methods 551

Note that the actual value of the Modal A (MAr ) and Modal B (MB r ) terms will
depend upon the normalization/scaling of the eigenvectors.

{φr }T [A] {φ} = MAr (34)

{φr }T [B] {φ} = MB r (35)

Since modal vectors are often compared and modal scaling is sometimes
compared, a standardized scaling of the modal vectors is desirable. Unfortunately,
this has not been done to this point in time. For several reasons, scaling each modal
vector such that each modal vector is dominantly real valued and such that the
largest coefficient in each modal vector is unity length, works optimally most of
the time. Note, however, this means that the complete eigenvector will need to be
scaled so that the modal vector, the lower half of the eigenvector, takes on these
characteristics.
When the damping matrix does represent proportional damping, then the
weighted orthogonality concept extends to the mass, damping and stiffness matrices.

{ψr }T [M] {ψr } = M̃r (36)

{ψr }T [C] {ψr } = C̃r (37)
{ψr }T [K] {ψr } = K̃r (38)

Note that in Eqs. 36 through 38, the modal mass (M̃r ), modal damping (C̃r ),
and modal stiffness (K̃r ) definitions are sometimes referred to as generalized mass,
damping, and stiffness. Like Modal A and Modal B, these modal scaling terms are
sensitive to the normalization/scaling of the modal vectors. In all cases, the normal-
ization and scaling of the modal vectors relate to the force-displacement model.

2.3 Experimental Models

Experimental models are used to directly relate to the measured data that is esti-
mated in the time and frequency domains. These models are generally physics-based
models that originate from the general [M] [C] [K] models but are represented in
frequency response function (FRF), impulse response function (IRF), or general
polynomial matrix models that utilize the relationships between input and output
data in either the frequency or time domains. The general polynomial matrix model
can be directly related to the measured FRF and IRF models and are most often the
core of modern modal parameter estimation (MPE) methods. The following models
can also be applied to measured power spectra data (frequency domain) or measured
covariance data (time domain), but direct scaling is not possible if input and output
data is not included. While this idea is very useful, it will not be explored further in
this text.
552 R. J. Allemang and D. L. Brown

2.3.1 Polynomial Models

Frequency Domain

Rather than using a physically based mathematical model, the common

characteristics of different modal parameter estimation algorithms can be more
readily identified by using a matrix coefficient polynomial model. One way of
understanding the basis of this model can be developed from the polynomial model
used historically for the frequency response function. Note the nomenclature in the
following equations regarding measured frequency ωi and generalized frequency
si . Measured input and response data are always functions of measured frequency,
but the generalized frequency variable used in the model may be altered to improve
the numerical conditioning. This will become important in a later discussion of
generalized frequency involving normalized frequency, orthogonal polynomials,
and complex Z mapping.

Xp (ωi ) βn (si )n + βn−1 (si )n−1 + · · · + β1 (si )1 + β0 (si )0

Hpq (ωi ) = =
Fq (ωi ) αm (si )m + αm−1 (si )m−1 + · · · + α1 (si )1 + α0 (si )0
(39)
This can be rewritten:


n
βk (si )k
Xp (ωi ) k=0
Hpq (ωi ) = = m (40)
Fq (ωi )  k
αk (si )
k=0

Further rearranging yields the following equation that is linear in the unknown α
and β terms:


m 
n
αk (si )k Xp (ωi ) = βk (si )k Fq (ωi ) (41)
k=0 k=0

This model can be generalized to represent the general multiple input, multiple
output case as follows:


m 
n
(si )k [αk ] Xp (ωi ) = (si )k [βk ] Fq (ωi ) (42)
k=0 k=0

Note that the size of the coefficient matrices [αk ] will normally be NS × NS or
NL × NL and the size of the coefficient matrices [βk ] will normally be NS × NL or
NL × NS when the equations are developed from experimental data.
Rather than developing the basic model in terms of force and response infor-
mation, the models can be stated in terms of power spectra or frequency response
information. First, post multiply both sides of the equation by {F }H :
10 Experimental Modal Analysis Methods 553


m
H

n
H
(si )k [αk ] Xp (ωi ) Fq (ωi ) = (si )k [βk ] Fq (ωi ) Fq (ωi )
k=0 k=0
(43)

Now recognize that the product of {X(ωi )} {F (ωi )}H is the output-input cross-
spectra matrix ([GXF (ωi )]) for one ensemble and {F (ωi )} {F (ωi )}H is the input-
input cross-spectra matrix ([GF F (ωi )]) for one ensemble. With a number of
ensembles (averages), these matrices are the common matrices used to estimate the
FRFs in a MIMO case. This yields the following cross-spectra model:


m 
n
(si ) [αk ] [GXF (ωi )] =
k
(si )k [βk ] [GF F (ωi )] (44)
k=0 k=0

The previous cross-spectra model can be reformulated to utilize frequency

response function (FRF) data by post multiplying both sides of the equation by
[GF F (ωi )]−1 :


m 
n
(si )k [αk ] [GXF (ωi )] [GF F (ωi )]−1 = (si )k [βk ] [GF F (ωi )] [GF F (ωi )]−1
k=0 k=0
(45)
Therefore, the multiple input, multiple output FRF model is:


m 
n
(si )k [αk ] [H (ωi )] = (si )k [βk ] [I ] (46)
k=0 k=0

Additional equations can be developed by repeating Eq. 46 at many frequencies

(ωi ) until all data or a sufficient overdetermination factor is achieved. Note that
both positive and negative frequencies are required in order to accurately estimate
conjugate modal frequencies.
In terms of sampled data, the frequency domain matrix polynomial coefficients
result from a set of linear equations (repeated application of Eq. 46 where each
equation is formulated by choosing a different frequency from the FRF data. From
a numerical perspective, Eq. 46 is generally not well-formed, and the condition
number associated with this system of equations will be extreme when the model
order (m) exceeds five or six. This issue will require special consideration in order
to obtain reasonable answers for the modal frequencies. This is discussed further in
Sect. 5.2.
Once the alpha ([α]) and beta ([β]) coefficients have been found, the modal
frequencies can be found from the roots of the alpha ([α]) matrix coefficient
polynomial.


m
[αk ] s k = 0 (47)
k=0
554 R. J. Allemang and D. L. Brown

The roots of this matrix coefficient polynomial are in terms of the generalized
frequency variable s and will be the complex modal frequencies (λr ) directly.

Time Domain

Paralleling the development of Eqs. 39 through 46, a time domain model

representing the relationship between a single response degree of freedom and a
single input degree of freedom can be stated as follows:


m 
n
αk x (ti+k ) = βk f (ti+k ) (48)
k=0 k=0

For the general multiple input, multiple output case:


m 
n
[αk ] {x (ti+k )} = [βk ] {f (ti+k )} (49)
k=0 k=0

The above model, in the time domain, is also known as an autoregressive-

moving-average (ARMA(m,n)) model when developed from a set of discrete
time equations in the time domain. More properly, this model is known as the
autoregressive with exogenous inputs (ARX(m,n)) model.
If the discussion is limited to the use of free decay or impulse response function
data, the previous time domain equations can be simplified by noting that the forcing
function can be assumed to be zero for all time greater than zero. If this is the case,
the [βk ] coefficients can be eliminated from the equations.


m
[αk ] [h (ti+k )] = 0 (50)
k=0

Additional equations can be developed by repeating Eq. 50 at different time shifts

(initial times ti ) into the data until all data or a sufficient overdetermination factor is
achieved. Note that at least one time shift is required in order to accurately estimate
conjugate modal frequencies.
In terms of sampled data, the time domain matrix polynomial coefficients result
from a set of linear equations (repeated application of Eq. 50 where each equation is
formulated by choosing various distinct initial times. From a numerical perspective,
Eq. 50 is generally well-formed, and the condition number associated with this
system of equations will not be extreme even for high order (m).
In contrast, the frequency domain matrix polynomial developed in Eqs. 39
through 46 results from a set of linear equations (repeated application of Eq. 46)
where each equation is formulated at one of the frequencies of the measured FRF
data. This distinction is important to note since the roots of the matrix characteristic
equation formulated in the time domain are in a mapped complex z domain (zr ),
which is similar but not identical to the z-domain familiar to control theory. Once the
10 Experimental Modal Analysis Methods 555

alpha ([α]) coefficients have been found, the roots of a polynomial in the complex
z domain can be found since all of the measured time domain data utilize the same
t spacing.


m
[αk ] zk = 0 (51)
k=0

Equation 51 is developed from the fixed time spacing (t) and derivative
relationships associated with discrete time data. It is important to note that this
development is theoretically exact and no approximations are involved.
These mapped complex z values (zr ) must be converted back to the generalized
frequency domain (λr ), while the roots of the matrix characteristic equation
formulated in the frequency domain (λr ) are already in the desired domain [1, 2, 3].
Note also that the roots that are estimated in the time domain are limited to
maximum values determined by the Shannon sampling theorem relationship (due
to the discrete time steps).

zr = eλr t
λr = σr + j ωr (52)
   
ln zr ln zr
σr = Re ωr = I m (53)
t t

In light of the above discussion, it is now apparent that the general matrix
coefficient polynomial model is a way in which most modal parameter estimation
methods, both for the time and for the frequency domain, generate functionally
similar matrix coefficient polynomial models. In this way, all of these methods can
be viewed in a similar framework. For that reason, the unified matrix polynomial
approach (UMPA) terminology is used to describe both domains since the time
domain, autoregressive-moving average (ARMA) terminology is already connected
with only the time domain.
For the frequency domain data case, this yields:
 
 
[αm ] s m + [αm−1 ] s m−1 + [αm−2 ] s m−2 + · · · · · · · · · + [α0 ] = 0 (54)

For the time domain data case, this yields:

 
 
[αm ] zm + [αm−1 ] zm−1 + [αm−2 ] zm−2 + · · · · · · · · · + [α0 ] = 0 (55)

Once the matrix coefficients ([α]) have been found, the modal frequencies can
be found from the roots of the associated matrix coefficient polynomial. Originally,
this was accomplished by forming a companion matrix from the matrix coefficients
and using an eigenvalue solution method on the companion matrix. Most modern
matrix solution packages now use the matrix coefficients directly in a root solving
algorithm (e.g., polyeig in Matlab® ) where the algorithm forms the companion
matrix and then solves the associated eigenvalue problem.
556 R. J. Allemang and D. L. Brown

2.3.2 Companion Matrix

Once the matrix coefficients ([α]) have been found, the modal frequencies (λr or zr )
can be found using a number of numerical techniques. While in certain numerical
situations, other numerical approaches may be more robust, a companion matrix
approach yields a consistent concept for understanding the process. Therefore, the
roots of the matrix characteristic equation can be found as the eigenvalues of the
associated companion matrix. The companion matrix can be formulated in one of
several ways.
Highest [α] Matrix Normalization:
One common formulation for finding the modal frequencies from a matrix
coefficient polynomial comes from the eigenvalue solution of a companion matrix
formed from the normalization of the highest order [α]m coefficient matrix to the
identity matrix [I ] yielding the following companion matrix:
⎡ ⎤
−[α]m−1 −[α]m−2 ··· ··· ··· −[α]1 −[α]0
⎢ [I ] [0] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎢ [0] [I ] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎢ [0] ··· ··· ··· [0] ⎥
⎢ [0] [0] ⎥
⎢ ⎥
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
[C] = ⎢ ⎥ (56)
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
⎢ ⎥
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
⎢ ⎥
⎢ [0] [0] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎣ [0] [0] ··· ··· ··· [0] [0] ⎦
[0] [0] ··· ··· ··· [I ] [0]

Note again that the numerical characteristics of the eigenvalue solution of the
companion matrix will be different for low-order cases compared to high-order
cases for a given data set. The companion matrix can be used in the following
eigenvalue formulation to determine the modal frequencies for the original matrix
coefficient equation:

[C] {X} = λ [I ] {X} (57)

⎡ ⎤
[I ] [0] · · · · · · · · · [0] [0]
⎢ [0] [I ] · · · · · · · · · [0] [0] ⎥
⎢ ⎥
⎢ [0] [0] · · · · · · · · · [0] [0] ⎥
⎢ ⎥
⎢··· ··· ··· ··· ··· ··· ···⎥
⎢ ⎥
⎢ ⎥
[I ] = ⎢ · · · · · · · · · · · · ··· ··· ···⎥ (58)
⎢ ⎥
⎢··· ··· ··· ··· ··· ··· ···⎥
⎢ ⎥
⎢ [0] [0] · · · · · · · · · [0] [0] ⎥
⎢ ⎥
⎣ [0] [0] · · · · · · · · · [I ] [0] ⎦
[0] [0] · · · · · · · · · [0] [I ]
10 Experimental Modal Analysis Methods 557

Lowest [α] Matrix Normalization:

Another common formulation for finding the modal frequencies from a matrix
coefficient polynomial comes from the eigenvalue solution of a companion matrix
formed from the normalization of the lowest-order [α]0 coefficient matrix to the
identity matrix [I ] yielding the following companion matrix:
⎡ ⎤
−[α]m−1 −[α]m−2 ··· ··· ··· −[α]1 −[I ]
⎢ [I ] [0] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎢ [0] [I ] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎢ [0] ··· ··· ··· [0] ⎥
⎢ [0] [0] ⎥
⎢ ⎥
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
[C] = ⎢ ⎥ (59)
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
⎢ ⎥
⎢ ··· ··· ··· ··· ··· ··· ··· ⎥
⎢ ⎥
⎢ [0] [0] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎣ [0] [0] ··· ··· ··· [0] [0] ⎦
[0] [0] ··· ··· ··· [I ] [0]

Note again that the numerical characteristics of the eigenvalue solution of the
companion matrix will be different for low-order cases compared to high-order
cases for a given data set. The companion matrix can be used in the following
eigenvalue formulation to determine the modal frequencies for the original matrix
coefficient equation:
⎡ ⎤
[α]m [0] ··· ··· ··· [0] [0]
⎢ [0] [I ] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎢ [0] ··· ··· ··· [0] ⎥
⎢ [0] [0] ⎥
⎢ ···⎥
  ⎢ ··· ··· ··· ··· ··· ··· ⎥
⎢ ⎥
I˜ = ⎢ · · · ··· ··· ··· ··· ··· ···⎥ (60)
⎢ ⎥
⎢ ··· ··· ··· ··· ··· ··· ···⎥
⎢ ⎥
⎢ [0] [0] ··· ··· ··· [0] [0] ⎥
⎢ ⎥
⎣ [0] [0] ··· ··· ··· [I ] [0] ⎦
[0] [0] ··· ··· ··· [0] [I ]
 
[C] {X} = λ I˜ {X} (61)

The eigenvectors that can be found from the eigenvalue-eigenvector solution

utilizing the companion matrix may, or may not, be useful in terms of modal
parameters. The eigenvector that is found, associated with each eigenvalue, is
of length model order m times matrix coefficient size, NS or NL . In fact, the
unique (meaningful) portion of the eigenvector is of length equal to the size of the
coefficient matrices, NS or NL , and is repeated in the eigenvector m times. For each
repetition, the unique portion of the eigenvector is repeated, multiplied by a different
complex scalar which is a successively larger, integer power of the associated modal
558 R. J. Allemang and D. L. Brown

frequency. Therefore, the eigenvectors of the companion matrix have the following
form:

⎧ ⎫ ⎧ ⎫
⎪
⎪ λrm−1 {ψ}r ⎪⎪ ⎪
⎪ zrm−1 {ψ}r ⎪⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ · ⎪
⎪ ⎪
⎪ · ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎨ · ⎪
⎬ ⎪
⎨ · ⎪
⎬
{φ}r ∝ · ∝ · (62)
⎪
⎪ ⎪ ⎪ 2 ⎪
⎪
⎪ λ2r {ψ}r ⎪⎪
⎪
⎪
⎪
⎪ zr {ψ}r ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ λ 1 {ψ} ⎪
⎪ ⎪
⎪ z 1 {ψ} ⎪
⎪
⎪
⎩ r r ⎪
⎭ ⎪
⎩ r r ⎪
⎭
λr {ψ}r
0
r
z 0
r {ψ} r r

These vectors are often referred to as m-th order state vectors. Note that the
zr mapped version of the λr is commonly used to avoid numerical problems with
weighting that occurs when high frequencies are involved. Note that, unless the size
of the coefficient matrices is at least as large as the number of measurement degrees
of freedom, only a partial set of modal coefficients, the modal participation coeffi-
cients (Lqr ) will be found. For the case involving scalar polynomial coefficients, no
meaningful modal coefficients will be found.
If the size of the coefficient matrices (NS ), and therefore the modal participation
vector ({Lr }), is less than the largest spatial dimension of the problem (NL ), then
the modal vectors are typically found in the second stage solution (MPE-2) process
using one of Eqs. 63 through 68. Even if the complete modal vector ({ψ}) of the
system is found from the eigenvectors of the companion matrix approach, the modal
scaling and modal participation vectors for each modal frequency are normally
found in this second stage formulation.

2.3.3 Partial Fraction Models: Residues and Residuals

Residues

The following sections review the relevant theoretical concepts and equations
required for discussing the estimation of final scaled modal vectors. The final scaled
modal vectors are often derived from the residues of the partial fraction model of
the MIMO FRF data matrix [6]. Alternatively, the final scaled modal vectors can
be estimated as a vector proportional to the residue vector with associated modal
scaling, such as Modal A (MAr ).

Residues from Single Reference FRFs

The equations that relate the complex modal frequencies, complex valued residues,
and complex valued modal vectors to the single reference FRF data are well-known
and are restated in the following equations for discussion purposes [6, 51]:
10 Experimental Modal Analysis Methods 559


N
Apqr A∗pqr
Hpq (ω) = + (63)
j ω − λr j ω − λ∗r
r=1
 

N Apqr N ×1 A∗pqr
NL ×1
Hpq (ω) NL ×1
= L
+ (64)
j ω − λr j ω − λ∗r
r=1

The above equations represent a partial fraction, residue model for the force-
displacement FRF relationships. At the point where modal vectors are estimated,
the FRF measurements, the measured frequencies, and the complex valued modal
frequencies, λr , are known. The above model assumes that the residues (the
numerator terms) occur in complex conjugate pairs. This model is often generalized
as follows:


2N
Apqr
Hpq (ω) = (65)
j ω − λr
r=1


2N Apqr NL ×1
Hpq (ω) NL ×1
= (66)
j ω − λr
r=1

In the above equations, no assumption is made regarding the complex conjugate

pair relationship among the residues. Once the residues are estimated, the estimates
are evaluated to determine if there are complex conjugate pairs as a measure of
quality.

Residues from Multiple Reference FRFs

Likewise, the equations that relate the complex modal frequencies, complex valued
modal participation vectors, complex valued modal vectors, and complex valued
residues in the multiple reference and multi-input, multi-output (MIMO) FRF data
situation are well-known and are restated in the following equations for discussion
purposes [6, 52]:


2N
[Ar ]N S ×NL
[H (ω)]NS ×NL = (67)
j ω − λr
r=1

The above equation is the natural extension of the single reference, partial
fraction residue model in the previous section. When this equation is applied to
the estimation of modal vectors, the relationships between the references (known as
the modal participation vectors ([L])) and the complex valued modal frequencies,
λr , are already known. This allows the equation to be rearranged to take this a-priori
information into account.
560 R. J. Allemang and D. L. Brown

[[H (ω)]]NS ×NL ×Nf = [L]NS ×2N [[(ω) ]]2N×2N ×Nf [ψ]T2N ×NL (68)

Noting that:

Apqr = Lpr ψqr (69)

In the above equation, note that an unconventional format is used to clarify the
structure of the equation. The double bracket notation [[ ]] is used to note that the
term within the brackets is repeated across all frequencies (Nf ). This makes the
matrix products ambiguous since a matrix product between a 2D and a 3D matrix is
not uniquely defined. This will be explained further in a later section.
Also in the above equations, note that the columns of the [[H (ω)]] matrix are
uncoupled as are the columns of the [ψ] matrix. This allows this equation to be
implemented for NS FRFs, associated with one response DOF. This solution then
proceeds for each of the NL response DOFs, one at a time as in the single reference
situation.
In all equations to this point in the chapter, it should be noted that the residues
should be purely imaginary for a normal mode case (proportional damping) utilizing
displacement over force FRF data. For the anticipated normal mode situation, there
is no constraint on the numerical characteristics of either the modal participation
coefficient or the modal vector coefficient individually as long as the product of
these two terms yields the correct residue characteristic.

Residues from IRFs

Most current modal parameter estimation algorithms utilize frequency or impulse

response functions as the data, or known information, to solve for modal parameters.
If the data is utilized in the form of impulse response functions, a damped complex
exponential model is appropriate. Impulse response functions are rarely directly
measured but are calculated from associated frequency response functions via the
inverse FFT algorithm.
The damped complex exponential model is essentially the Fourier or Laplace
transform of the partial fraction models already discussed. Since the damped
complex exponential model is numerically well-formed, this model has also been
used extensively. In terms of the same notation used previously, the damped complex
exponential model can be formed as either a summation of terms or product of terms,
just as the partial fraction model is formed in the frequency domain.
The general equation that can be used to represent the relationship between the
measured impulse response function matrix and the modal parameters is shown as
a summation of terms in Eq. 70 or, in the equivalent matrix product form, in Eqs. 71
and 72.
10 Experimental Modal Analysis Methods 561


N
∗
[h(t)]NL ×NS = [Ar ]NL ×NS eλr t + [A∗r ]NL ×NS eλr t
r=1


2N
= [Ar ]NL ×NS eλr t (70)
r=1

[[h(t)]]NL ×NS ×Nt = [ψ] eλr t ]]2N ×2N ×Nt [L]T2N ×NS (71)

[[h(t)]]TNS ×NL ×Nt = [L] eλr t ]]2N ×2N ×Nt [ψ ]T 2N ×NL (72)

Again noting that:

Apqr = Lpr ψqr (73)

Many modal parameter estimation algorithms have been originally formulated

from Eqs. 63 through 73. However, a more general development for all algorithms
is based upon relating the above equations to a general matrix coefficient polynomial
model.

Residuals from FRFs

Continuous systems have an infinite number of degrees of freedom but, in general,

only a finite number of modes can be used to describe the dynamic behavior
of a system. The theoretical number of degrees of freedom can be reduced by
using a finite frequency range. Therefore, for example, the frequency response
can be broken up into three partial sums, each covering the modal contribution
corresponding to modes located in the frequency ranges. Note how the mode at 2.2
Hertz affects the modes below and above in the frequency range plotted in Fig. 5.
In the frequency range of interest, the modal parameters can be estimated to be
consistent with Eq. 63. For a situation where modes below and above the frequency
range of interest cannot be ignored, residual terms can be included to account for
the effect that these modes will have when estimating modal parameters within
the frequency range of interest. In this case, Eq. 63 can be rewritten for a single
frequency response function as:


N
Apqr A∗pqr RIpq
Hpq (ω) = RFpq + + ∗
+ (74)
j ω − λr j ω − λr (j ω)2
r=1

where

• RFpq = Residual flexibility

• RIpq = Residual inertia
 562 R. J. Allemang and D. L. Brown

10

-10
Log Magnitude, (dB)

-20

-30

-40

-50

-60

-70
-8 -6 -4 -2 0 2 4 6 8
Frequency, (Hertz)
200
150
Phase, (Degrees)

100
50
0
-50
-100
-150
-200
-8 -6 -4 -2 0 2 4 6 8

Frequency, (Hertz)

Fig. 5 Residual effects on FRF frequency band

The oldest and simplest approach is to include two frequency domain terms to
account for the inertia effect of lower frequency modes of vibration, below the
minimum frequency, and the stiffness effect of higher frequency modes of vibration,
above the maximum frequency. These two terms are referred to as residual inertia
(RIpq) and residual flexibility (RFpq).
Note that using the form of the above equation, residual flexibility (RFpq) should
be a positive constant, and residual inertia (RIpq) should be a positive constant
representing the magnitude of the second-order frequency term (Fig. 6).
The lower residual is a term reflecting the inertia or mass of the lower modes
and is an inverse function of the frequency squared. The upper residual is a term
reflecting the flexibility of the upper modes and is constant with frequency. In many
cases the lower residual term is called the inertia restraint.
Residuals are a function of each individual frequency response function mea-
surement and are not global properties of the frequency response function matrix.
Therefore, residuals cannot be estimated unless the frequency response function has
been measured.
10 Experimental Modal Analysis Methods 563

0
10

−1
10
Magnitude (in/lb)

−2
10

−3
10

−4
10
0 50 100 150 200 250 300
Frequency (Hz)
50

0
Phase (deg)

−50

−100

−150

−200
0 50 100 150 200 250 300

Frequency (Hz)

Fig. 6 Influence of residuals on nearby modes

Note that in this common formulation of residuals, both terms are real valued
quantities. Also note that in general this is a simplification: the residual effects of
modes below and/or above the frequency range of interest cannot be completely
represented by such simple relationships.
In this case, the form of the residual is based upon a physical concept of how the
system poles below and above the frequency range of interest will affect the data
in the range of interest. As the system poles below and above the range of interest
are located in the proximity of the boundaries, these effects are not the simple real
valued quantities noted in Eq. 74.
In these cases, residual modes may be included in the model to partially account
for these effects. When this is done, the modal parameters that are associated with
these residual poles have no physical significance but may be required in order
to compensate for strong dynamic influences from outside the frequency range of
564 R. J. Allemang and D. L. Brown

interest. Using the same argument, the lower and upper residuals can take on any
mathematical form that is convenient as long as the lack of physical significance is
understood. Mathematically, power functions of frequency (zero, first, and second
order) are commonly used within such a limitation. In general, the use of residuals
is confined to frequency response function models. This is primarily due to the
difficulty of formulating a reasonable mathematical model and solution procedure
in the time domain for the general case that includes residuals.
The above models do not show additional mathematical terms, referred to as
residuals, that account for the influence of modes with modal frequencies outside of
the range of the frequencies used in the modal parameter estimation, or even outside
the frequency range of the FRF measurements. This approach can be generalized to
a set of five or more frequency domain, polynomial terms that are mathematical in
structure but simply an extension of the residual inertia and flexibility concept. This
approach better accounts for modes with frequencies above and below the maximum
and minimum frequency of interest for cases where the modal frequencies are close
to the frequency boundaries.

Hpq (ω) = RFpq + R1pq (j ω) + R2pq (j ω)2


N
Apqr A∗pqr R3pq RIpq
+ + ∗
+ + (75)
j ω − λr j ω − λr (j ω) (j ω)2
r=1

The above two approaches to estimating residuals are most often a part of
the LS and WLS solutions. The concept is that the addition of residuals reduces
the contamination of the modal vectors that are estimated from the modes with
frequencies outside the frequency band of interest.
Another form of residuals is to allow extra computational poles (terms that have
the same form as complex modal frequencies) to augment the physical complex
modal frequencies that have been identified. These additional terms are of similar
mathematical structure to the physical complex modal frequencies and are easy to
implement in the solution procedure. Again the concept is that these additional
computational terms allow the modal vectors to be estimated with less influence
or contamination from modes that are nearby in frequency or even noise within the
frequency band of interest. These computational poles may take on any complex
valued frequency value and are found from the modal parameter estimation process.
In normal situations, these poles are not included as physical, complex modal
frequencies since they do not satisfy requirements of complex modal frequencies
(consistency of estimates, realistic damping or frequency, etc.).
In all cases, the value of the residuals are of little practical value except to
improve the estimates of the modal vectors. The residual information is generally
not retained after the modal vectors have been estimated.

2.3.4 Modal Vectors and Modal Scaling from Residues

The modal vectors are most often determined by normalizing the residues. There are
a number of normalization methods that can be used. Once the modal vectors are
10 Experimental Modal Analysis Methods 565

determined, the modal scaling can be determined. The modal vector normalization
is up to the user but most commonly is chosen to give modal vectors that are
dominantly real valued (close to normal modes). The choice of normalization may
also consider how the comparable modal vectors from analytical methods have been
normalized. For this reason, it is common to choose the modal vector normalization
such that the realities of experimental estimation of the modal vectors (noting that
the residues will not be perfectly imaginary valued) are considered. One example
of a useful normalization is to choose to make the largest modal coefficient in each
modal vector unity.

Apqr NL ×1
{ψr }NL ×1 = (76)
max(Apqr )

In the above equation, the residue vector is divided by the residue with the largest
complex magnitude, mode by mode. This yields a modal vector that will generally
be dominantly real valued and limited in magnitude to plus/minus unity.
The modal A scaling term (MAr ) can then be found to provide the absolute
scaling associated with this modal vector normalization/scaling. Note that if the
normalization/scaling of this modal vector is the same as that used for an analytical
solution, then the modal A scaling terms are comparable.

ψpr ψqr ψpr ψqr

Apqr = MAr = (77)
MAr Apqr

If the modal vector is completely real valued, the modal mass scaling term (Mr )
can be estimated.

ψpr ψqr ψpr ψqr

Apqr = Mr = (78)
j 2ωr Mr j 2ωr Apqr

Note that if the normalization/scaling of this modal vector is the same as that
used for an analytical solution, then the modal mass scaling terms are comparable.
Finally, from the above equations, modal mass can be estimated from modal A
as long as the modal vector is a normal (real valued) mode.

MAr
Mr = (79)
j 2ωr

If the modal vector is not a real valued normal mode, the modal vector can
be normalized to an equivalent real valued normal mode and Eq. 78 can be used.
Alternatively, an effective modal mass can be estimated from Eq. 80 as long as the
modal vector is dominantly real valued.

MAr
Mr  =   (80)
j 2ωr
566 R. J. Allemang and D. L. Brown

2.3.5 Other Experimental Model Methods

There are several other models and methods that researchers use to solve for, or
enhance, modal parameter estimates up through methods that are used to solve for
data originating from a nonlinear system. Many times, these techniques are used
to improve modal parameter estimates that originated with data that has noticeable
noise. One method would be a maximum likelihood method [53, 54, 55, 56, 56]
that begins with a modal parameter solution from one of the methods discussed
here, integrating a procedure that can allow one or more of the modal parameters
to be varied to improve the comparison between measured data and synthesized
data. Another method that would be applied to input-output data that has expected
nonlinear characteristics is the time domain NARMAX methods or the frequency
domain reverse path methods. These methods are not covered in any further detail
in this text.

3 Single Degree of Freedom Methods

3.1 SDOF Algorithms: Overview

For any real system, the use of single degree of freedom algorithms to estimate
modal parameters is always an approximation since any realistic structural system
will have many degrees of freedom. Nevertheless, in cases where the modes are
not close in frequency and do not affect one another significantly, single degree
of freedom algorithms are very effective. Specifically, single degree of freedom
algorithms are quick, rarely involving much mathematical manipulation of the
data, and give sufficiently accurate results for most modal parameter requirements.
Naturally, most multiple degree of freedom algorithms can be constrained to
estimate only a single degree of freedom at a time if further mathematical accuracy
is desired.

3.2 Operating Vector (Peak-Pick) Estimation

Technically, when many single degree of freedom approaches are used to estimate
modal parameters, sufficient simplifying assumptions are made such that the results
may not be modal parameters in a rigorous sense. In these cases, the results are
often referred to as operating vectors or operating deflection shapes (ODS) rather
than modal vectors. This term refers to the fact that if the structural system is
excited at this frequency, the resulting motion will be a linear combination of the
modal vectors rather than a single modal vector. If one mode is dominant, then the
operating vector will be approximately equal to the modal vector. Modal parameters
can be estimated, but, generally, methods that use a model and several frequencies
are preferred. The term operating deflection shapes (ODS) is generally used when
data other than FRFs, such as cross power spectra (CPS), is the experimentally
measured data.
10 Experimental Modal Analysis Methods 567

The most commonly used single degree of freedom algorithms involve using the
information at a single frequency as an estimate of the modal vector. These methods
have historically been referred to as peak-pick methods. Figure 7 is an example of
using the information at the peak frequency location (positive or negative peak in
the imaginary part of the X/F frequency response functions) as an estimate of the
modal vectors of a simple beam.

Summary: Operating (Peak-Pick) Vector Methods

• Methods are simple, approximate methods

• Modal frequency
– Peak in imaginary part of FRF
– Zero crossing in real part of FRF
– Peak in magnitude or log magnitude of FRF
• Modal damping: half-power method
• Residue: SDOF equation
– Complex qmplitude of FRF
– Imaginary part of FRF – X/F or A/F
– Real part of FRF – V/F
– Magnitude of FRF (+/− sign depending on phase angle)
• Separation of close modes is generally not possible
• No residuals

If modal parameters are desired, the approximate relationships that are used in
these cases are represented in the following two equations.

Apqr A∗pqr
Hpq (ωr ) ≈ + (81)
j ωr − λr j ωr − λ∗r

Apqr
Hpq (ωr ) ≈ Apqr ≈ − σr Hpq (ωr ) (82)
− σr

For these less complicated methods, the damped natural frequencies (ωr ) are
estimated by observing the maxima in the frequency response functions. The
damping factors (σr ) are estimated using half-power methods [45]. The residues
(Apqr ) are then estimated from Eq. 82 using the frequency response function data at
the damped natural frequency.
Figures 8, 9, and 10 show frequency response function measurements in different
formats to emphasize that the data can be determined in several ways. Note the
characteristics of the FRF data, in each format, in the frequency range around the
damped natural frequency (ωr ).
568 R. J. Allemang and D. L. Brown

Fig. 7 Modal vectors from the imaginary part of the FRF

The damped natural frequency ωr is estimated in one of three ways:

• The frequency where the magnitude of the FRF reaches a maximum

• The frequency where the real part of the FRF crosses zero
• The frequency where the imaginary part of the FRF reaches a relative minima (or
maxima)

Of these three methods, the last approach gives the most reliable results under all
conditions.
Once the damped natural frequency ωr has been estimated, the real part of the
modal frequency, the damping factor σr , can be estimated. The damping factor σr
can be estimated by using the half-power bandwidth method or possibly the log
decrement method.
 10 Experimental Modal Analysis Methods 569

200
Phase (Degrees)
150
100
50
0
-50
-100
-150
-200
0 10 20 30 40 50 60 70 80 90 100
-5 Frequency (Hertz)
x10
7

5
Magnitude

0
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hertz)

Fig. 8 Single degree of freedom method – magnitude/phase

200
Phase (Degrees)

150
100
50
0
-50
-100
-150
-200
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hertz)

-80

-90

-100
Log Magnitude (dB)

-110

-120

-130

-140

-150

-160
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hertz)

Fig. 9 Single degree of freedom method – log magnitude/phase

 570 R. J. Allemang and D. L. Brown

x10-5
4
3
Amplitude (Real)

2
1
0
-1
-2
-3
-4
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hertz)
x10-5
8
Amplitude (Imag)

6
4
2
0
-2
-4
-6
-8
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hertz)

Fig. 10 Single degree of freedom method – real/imaginary

3.2.1 Half-Power Bandwidth Method

The half-power bandwidth method [45] is a frequency domain method that can give
good results if only one mode is present in the frequency domain data and sufficient
frequency resolution is available. For lightly damped situations, failure to meet this
second criteria is quite common and can lead to significant error.
This method uses the data from the FRF in the region of the damped natural
frequency to estimate the fraction of critical damping. Note that as long as the modes
are separated in frequency, each mode can be treated as a SDOF. If this is the case,
the fraction of critical damping can be found from the following formula:

ωb 2 − ωa 2
ζr = (83)
(2 ωr )2

In the above equation, ωr is the damped natural frequency as previously

estimated. ωa is the frequency, below ωr , where the magnitude is 0.707 of the peak
magnitude of the FRF. This corresponds to a half-power point. ωb is the frequency,
above ωr , where the magnitude is 0.707 of the peak magnitude of the FRF. This also
corresponds to a half-power point. These half-power points are also referred to as
the −3 dB points, relative to the dB value at the peak.
For lightly damped systems, the above equation can be approximated by the
following:
10 Experimental Modal Analysis Methods 571

ωb − ωa
ζr ≈ (84)
(2 ωr )

Once ζr is estimated, the damping factor σr can be estimated from the following
equation.

σr = − ζr Ωr (85)

Again, assuming that the system is lightly damped, Ωr ≈ ωr , the damping factor
can be estimated from the following equation:

σr ≈ − ζr ωr (86)

Figure 11 shows the frequency region around a damped natural frequency where
no other nearby modes are present. The half-power bandwidth method can be
applied to this data to estimate damping, but the lack of frequency resolution
near the peak will mean that the true “full-power” amplitude cannot be observed.
Therefore, the frequencies associated with the half-power amplitudes will not be
found accurately. The fraction of critical damping will not be accurately found for
this case.

0.04

0.035

0.03

0.025
Magnitude

0.02

0.015

0.01

0.005

0
11.5 12 12.5 13 13.5
Frequency (Rad/Sec)

Fig. 11 SDOF FRF: magnitude plot with poor frequency resolution

572 R. J. Allemang and D. L. Brown

Note that while this approach can yield reasonable results, most of the partial
fraction methods discussed in upcoming Sects. 3.4 through 3.7 give superior results
using the same data without the frequency resolution issues.

3.2.2 Logarithmic Decrement Method

An alternate approach to estimating the damping is from the time domain response
using the log decrement method [45]. This method is rarely used since it will be
difficult to obtain the time response of one mode when many modes are present.
Nevertheless, the log decrement will be presented for completeness.
If time domain data for one mode of vibration can be obtained, looking like the
free decay response in Fig. 12, the log decrement, defined as δ, can be found from
two successive peak amplitudes in the following relationship:

x(tn ) x(tn )
δ = loge ( ) = ln( ) (87)
x(tn+T ) x(tn+T )

where T is the length of one period of the oscillation.

The log decrement value can be used to estimate the fraction of critical damping:

SDOF - Free decay response.

5

1
Amplitude

-1

-2

-3

-4

-5
0 2 4 6 8 10
Time (seconds)

Fig. 12 SDOF Time Domain: free decay response

10 Experimental Modal Analysis Methods 573

2π ζ
δ = ≈ 2π ζ f or ζ 0.1 (88)
1 − ζ2

Note that the damped natural frequency ωr can be found from the time of one
period of this oscillatory response as:
Note that while this approach can yield reasonable results, most of the partial
fraction methods discussed in the upcoming Sects. 3.4 through 3.7 give superior
results without the frequency resolution limitation, using the same data.

3.3 Complex Plot (Circle Fit) Method

The Circle Fit method is based upon a technique first reported by Kennedy and
Pancu in 1947 [57]. This method utilizes the concept that the data curve in the
vicinity of a modal frequency looks circular. In fact, the diameter of the circle is
used to estimate the residue once the damping factor is estimated. More importantly,
though, Kennedy and Pancu noted that the distance along the curve between
data points at equidistant frequency maximized in the neighborhood of the modal
frequency. In this way, the circle fit method was the first method to detect closely
spaced modes.
This method can give erroneous answers when the modal coefficient is near
zero. This occurs essentially because when the mode does not exist in a particular
frequency response function (either the input or response degree of freedom is at
a node of the mode), the remaining data in the frequency range of the mode will
be strongly affected by the next higher or lower mode. Therefore, the diameter of
the circle that will be estimated will be a function of the modal coefficient for the
next higher or lower mode. This can be detected visually but is somewhat difficult
to detect automatically.

Summary: Complex Plot – Circle Fit Method

• Simple historical method (Kennedy and Pancu, 1947)

• Modal frequency: frequency where largest spacing between data points
occurs (frequency at bottom or top of complex plot)
• Modal damping: half-power method
• Residue: SDOF equation
A
– Diameter of the circle ≈ −pqrσr
• Separation of close modes possible
• Residuals possible
– Center of the circle not on axis
• Complex modal coefficients
• Problem: Error when modal coefficient should be zero
574 R. J. Allemang and D. L. Brown

x10 -5
8

2
Amplitude (Imag)

-2

-4

-6

-8
Amplitude (Real)

Fig. 13 SDOF FRF Complex Plot: real versus imaginary

The approximate relationship that is used in this case is represented in the

following equation (Fig. 13).

Apqr A∗pqr
Hpq (ωr ) ≈ Rpq + + (89)
j ωr − λr j ωr − λ∗r

General equation of a circle

x2 + y2 + a x + b y + c = 0 (90)
a x + b y + c = − x2 − y2 (91)

Center of circle

a b
xc = − yc = − (92)
2 2
10 Experimental Modal Analysis Methods 575

Radius of circle
! $ %2 &0.5
" a #2 b
Radius = + − c (93)
2 2

Data points for circle

xi = Real(Hpq (ωi )) yi = I mag(Hpq (ωi )) (94)

Direct solution (three data points)

a x1 + b y1 + c = − x12 − y12 (95)

a x2 + b y2 + c = − x22 − y22 (96)

a x3 + b y3 + c = − x32 − y32 (97)

⎡ ⎤⎧ ⎫ ⎧ ⎫ ⎧ ⎫
x1 y1 1 ⎨ a ⎬ ⎨ −x12 − y12 ⎬ ⎨a⎬
⎣ x2 y2 1 ⎦ b = −x 2 − y 2 or [T ] b = {S} (98)
⎩ ⎭ ⎩ 22 2⎭ ⎩ ⎭
x3 y3 1 c −x3 − y32 c
⎧ ⎫
⎨a⎬
b = [T ]−1 {S} (99)
⎩ ⎭
c

Least-squares solution (more than three data points)

⎡ ⎤ ⎧ ⎫
x1 y1 1 ⎪
⎪ −x12 − y12 ⎪
⎪
⎢x ⎪
⎪ −x 2 − y 2 ⎪
⎢ 2 y2 1 ⎥
⎥⎧a ⎫ ⎪ ⎪ 2⎪
⎪
⎪ ⎧ ⎫
⎢ ⎥⎨ ⎬ ⎪ ⎨ 22 ⎪
⎬ ⎨a⎬
⎢ x3 y3 1 ⎥ −x3 − y3 2
⎢ ⎥ b = or [T ] b = {S} (100)
⎢ x4 y4 1 ⎥⎩ ⎭ ⎪ ⎪ −x42 − y42 ⎪
⎪ ⎩ ⎭
⎢ ⎥ c ⎪
⎪ ⎪
⎪ c
⎣··· ··· ···⎦ ⎪
⎪ · · · ⎪
⎪
⎪
⎩ 2 ⎪
⎭
xn yn 1 −xn − yn2
⎧ ⎫
⎨a ⎬
[T ]T [T ] b = [T ]T {S} (101)
⎩ ⎭
c
⎧ ⎫
⎨a⎬  −1
b = [T ]T [T ] [T ]T {S} (102)
⎩ ⎭
c
576 R. J. Allemang and D. L. Brown

3.4 Two-Point Finite Difference Formulation

The difference method formulations are simple methods that are based upon
comparing adjacent frequency information in the vicinity of a resonance frequency
[58]. When a ratio of this information, together with information from the derivative
of the frequency response function at the same frequencies, is formed, a reasonable
estimation of the modal frequency and residue for each mode can be determined
under the assumption that modes are not too close together. This method can
give erroneous answers when the modal coefficient is near zero. This problem
can be detected by comparing the predicted modal frequency to the frequency
range of the data used in the finite difference algorithm. As long as the predicted
modal frequency lies within the frequency band, the estimate of the residue (modal
coefficient) should be valid.

Summary: Finite Difference Method

• Common method implemented on many two channel digital signal process-

ing analyzers.
• Solves for both modal frequency (λr ) and residue (Apqr ).
• Uses the frequency response function information at two frequencies in the
vicinity of a single mode.
• Several combinations of two frequencies can be used and the results
averaged.
• Approximate method (ignores complex conjugate contribution).
• No residuals.
• Problem: Error when modal coefficient should be zero.

The approximate relationships that are used in this case is represented in the
following three equations. The frequencies noted in these relationships are as
follows: ω1 is a frequency near the damped natural frequency ωr , and ωp is the
peak frequency close to the damped natural frequency ωr .

Apqr
Hpq (ω1 ) ≈ (103)
j ω1 − λr
Apqr
Hpq (ωp ) ≈ (104)
j ωp − λr

The finite difference relationships are formulated as follows:

Apqr (j ω1 − j ωp )
1 = Hpq (ωp ) − Hpq (ω1 ) ≈ (105)
(j ω1 − λr )(j ωp − λr )
10 Experimental Modal Analysis Methods 577

Apqr (j ω1 − j ωp )λr
2 = j ωp Hpq (ωp ) − j ω1 Hpq (ω1 ) ≈ (106)
(j ω1 − λr )(j ωp − λr )
j (ω1 − ωp )Apqr Apqr
3 = j (ω1 − ωp )Hpq (ω1 )Hpq (ωp ) ≈ (107)
(j ω1 − λr )(j ωp − λr )

Modal frequency (λr )

2 j ωp Hpq (ωp ) − j ω1 Hpq (ω1 )

λr ≈ = (108)
1 Hpq (ωp ) − Hpq (ω1 )

Residue (Apqr )

3 j (ω1 − ωp )Hpq (ω1 )Hpq (ωp )

Apqr ≈ = (109)
1 Hpq (ωp ) − Hpq (ω1 )

Least-Squares Solution
Since both of the equations that are used to estimate modal frequency λr and
residue Apqr are linear equations, a least-squares solution can be formed by using
other frequency response function data in the vicinity of the resonance. For this
case, additional equations can be developed using Hpq (ω2 ) or Hpq (ω3 ) in the above
equations instead of Hpq (ω1 ). Starting with Eqs. 108 and 109 for any frequency (ω1 )
in the vicinity of the peak frequency (ωp ) (rearranging Eqs. 108 and 109 slightly):

(Hpq (ωp ) − Hpq (ω1 )) λr = j ωp Hpq (ωp ) − j ω1 Hpq (ω1 ) (110)

(Hpq (ωp ) − Hpq (ω1 )) Apqr = j (ω1 − ωp )Hpq (ω1 )Hpq (ωp ) (111)

Two more equations, involving the same two unknowns, can now be written
for any other frequency in the vicinity of the peak frequency (ωp ). Putting these
equations into matrix form yields:

⎧ ⎫ ⎧ ⎫
⎪
⎪ Hpq (ωp ) − Hpq (ω1 ) ⎪
⎪ ⎪
⎪ j ωp Hpq (ωp ) − j ω1 Hpq (ω1 ) ⎪
⎪
⎪ H (ω ) − H (ω ) ⎪
⎪ ⎪ ⎪ j ω H (ω ) − j ω H (ω ) ⎪
⎪ ⎪
⎪
⎨ pq p pq 2 ⎪ ⎬ ⎪
⎨ p pq p 2 pq 2 ⎪ ⎬
Hpq (ωp ) − Hpq (ω3 ) {λr }1×1 = j ωp Hpq (ωp ) − j ω3 Hpq (ω3 )
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ··· ⎪
⎪ ⎪
⎪ ··· ⎪
⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
Hpq (ωp ) − Hpq (ωs ) N j ωp Hpq (ωp ) − j ωs Hpq (ωs ) N ×1
s ×1 s
(112)
578 R. J. Allemang and D. L. Brown

⎧ ⎫ ⎧ ⎫
⎪
⎪ Hpq (ωp ) − Hpq (ω1 ) ⎪
⎪ ⎪
⎪ j (ωp − ω1 )Hpq (ωp ) Hpq (ω1 ) ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎨ Hpq (ωp ) − Hpq (ω2 ) ⎪
⎪ ⎬ ⎨ j (ωp − ω2 )Hpq (ωp ) Hpq (ω2 ) ⎪
⎪ ⎬
Hpq (ωp ) − Hpq (ω3 ) Apqr = j (ωp − ω3 )Hpq (ωp ) Hpq (ω3 )
⎪
⎪ ⎪
⎪
1×1 ⎪
⎪ ⎪
⎪
⎪
⎪ ··· ⎪
⎪ ⎪
⎪ ··· ⎪
⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
Hpq (ωp ) − Hpq (ωs ) N j (ωp − ωs )Hpq (ωp ) Hpq (ωs ) N ×1
s ×1 s
(113)

The above equations represent overdetermined sets of linear equations that can
be solved using any pseudoinverse or normal equations approach.

3.5 Least-Squares (Local) SDOF Method

The least-squares local SDOF formulations are simple methods that are based upon
using an SDOF model in the vicinity of a resonance frequency [58]. A reasonable
estimation of the modal frequency and residue for each mode can be determined
under the assumption that modes are not too close together. This method can give
erroneous answers when the modal coefficient is near zero. This problem can be
detected by comparing the predicted modal frequency to the frequency range of the
data used in the algorithm. As long as the predicted modal frequency lies within the
frequency band, the estimate of the residue (modal coefficient) should be valid.

Summary: Least-Squares (Local) SDOF Method

• Common method implemented on many two channel digital signal process-

ing analyzers.
• Solves for both modal frequency (λr ) and residue (Apqr ).
• Uses the frequency response function information in the vicinity of a single
mode.
• Approximate method (ignores complex conjugate contribution).
• No residuals.
• Problem: Error when modal coefficient should be zero.

The approximate relationship that is used in this case is represented in the

following equation. The frequency ω1 is a frequency near the damped natural
frequency ωr .

Apqr
Hpq (ω1 ) ≈ (114)
j ω1 − λr
Hpq (ω1 ) (j ω1 − λr ) = Apqr (115)
Hpq (ω1 ) λr + Apqr = (j ω1 )Hpq (ω1 ) (116)
10 Experimental Modal Analysis Methods 579

Repeating the above equation for several frequencies in the vicinity of the peak
frequency:
⎡ ⎤ ⎧ ⎫
Hpq (ω1 ) 1 ⎪
⎪ (j ω1 )Hpq (ω1 ) ⎪
⎪
⎢ H (ω ) ⎪ ⎪
⎢ pq 2 1 ⎥
⎥
⎪
⎪
⎪
⎪ (j ω2 )Hpq (ω2 ) ⎪
⎪
⎪
⎪
⎢ ⎥ ⎨ ⎬
⎢ Hpq (ω3 ) 1 ⎥ λr (j ω3 )Hpq (ω3 )
⎢ ⎥ = (117)
⎢ Hpq (ωp ) 1 ⎥ Apqr ⎪ (j ωp )Hpq (ωp ) ⎪
⎪ ⎪
⎢ ⎥ 2×1 ⎪
⎪ ⎪
⎪
⎣ ··· ···⎦ ⎪
⎪ · · · ⎪
⎪
⎪
⎩ ⎪
⎭
Hpq (ωs ) 1 N (j ωs )Hpq (ωs ) N
s ×2 s ×1

The above equation again represents an overdetermined set of linear equations

that can be solved using any pseudoinverse or normal equation approach.

3.6 Least-Squares (Global) SDOF Method

The least-squares, global SDOF formulations are simple methods that are based
upon using an SDOF model in the vicinity of a resonance frequency for all
measurements in a row or column of the FRF matrix [58]. A reasonable estimation
of the modal frequency and residue for each mode can be determined under the
assumption that modes are not too close together. This method can give erroneous
results for a specific residue when the modal coefficient is near zero.

Summary: Least-Squares (Global) SDOF Method

• Solves for both modal frequency (λr ) and residue (Apqr ).

• Uses the frequency response function information in the vicinity of a single
mode.
• Approximate method (ignores complex conjugate contribution).
• No residuals.
• Problem: Error when modal coefficient should be zero still possible but
normally avoided.

The approximate relationship that is used in this case begins with the result of
the least-squares local SDOF method.
⎡ ⎤ ⎧ ⎫
Hpq (ω1 ) 1 ⎪
⎪ (j ω1 )Hpq (ω1 ) ⎪
⎪
⎢ H (ω ) ⎪ ⎪
⎢ pq 2 1 ⎥
⎥
⎪
⎪
⎪
⎪ (j ω2 )Hpq (ω2 ) ⎪
⎪
⎪
⎪
⎢ ⎥ ⎨ ⎬
⎢ Hpq (ω3 ) 1 ⎥ λr (j ω3 )Hpq (ω3 )
⎢ ⎥ = (118)
⎢ Hpq (ωp ) 1 ⎥ Apqr ⎪
⎪ (j ωp )Hpq (ωp ) ⎪
⎪
⎢ ⎥ 2×1 ⎪
⎪ ⎪
⎪
⎣ ··· ···⎦ ⎪
⎪ · · · ⎪
⎪
⎪
⎩ ⎪
⎭
Hpq (ωs ) 1 N (j ωs )Hpq (ωs ) N
s ×2 s ×1
580 R. J. Allemang and D. L. Brown

Note that the above equation can be written for each measurement in a column
or row of the frequency response function matrix. When this is done, the modal
frequency (λr ) is the same for each measurement, while the residue (Apqr ) changes
with each measurement. This is described by the matrix version of the above
equation:
⎡ ⎤ ⎧ ⎫
Hpq (ω1 ) [I ] ⎪
⎪ (j ω1 ) Hpq (ω1 ) ⎪
⎪
⎢ ⎪ ⎪
⎢ Hpq (ω2 ) [I ] ⎥
⎥
⎪
⎪
⎪
⎪ (j ω2 ) Hpq (ω2 ) ⎪
⎪
⎪
⎪
⎢ ⎥ ⎨ ⎬
⎢ Hpq (ω3 ) [I ] ⎥ λr (j ω3 ) Hpq (ω3 )
⎢ ⎥ =
⎢ Hpq (ωp ) [I ] ⎥ Apqr ⎪ (j ωp )
⎪ Hpq (ωp ) ⎪
⎪
⎢ ⎥ No +1×1 ⎪
⎪ ⎪
⎪
⎣ ··· ···⎦ ⎪
⎪ ··· ⎪
⎪
⎪
⎩ ⎪
⎭
Hpq (ωs ) [I ] N (j ωs ) Hpq (ωs )
o Ns ×2 No Ns ×1
(119)

In the above equation, the size of the frequency response function column
( Hpq (ω1 ) ) determines the number of residues that will be estimated as well as the
size of the identity matrix. The above equation again represents an overdetermined
set of linear equations that can be solved using any pseudoinverse or normal
equations approach.

3.7 Other SDOF Methods

There are numerous other SDOF methods that may yield improved results when
special conditions exist. One example would be when errors in the data require
special attention like the case where the FRF data has frequency shift errors due
to sensor mass loading. In this case, any kind of least squares estimate of the
modal frequency across all FRFs, as with the previous least-squares (global) SDOF
method, will yield compromised answers.
Some errors that are mostly random, or even systematic, bias errors from modes
that are close in frequency, can be minimized by adding residuals to the SDOF
model. The least-squares (global) SDOF method can be improved with the addition
of residuals [59, 60] to account more completely for the nearby modes.

4 Multiple Degree of Freedom Methods

Modern multiple degree of freedom (MDOF) methods normally will include the
general case of solving for multiple modal parameters that represent all measure-
ments in a multiple input, multiple output (MIMO) set of data. While some of these
methods may have originated as single measurement or single reference methods,
all have been generalized to include those cases as a subset of the total method.
All methods utilize a kernel equation in the time or frequency domain with
square matrix coefficients that are used to find modal frequencies based upon the
size (dimension) of the square matrix coefficient and the model order.
10 Experimental Modal Analysis Methods 581

4.1 General (Two-Stage) Solution Procedure

Based upon Eqs. 41 through 46 and Eqs. 47 through 50, most modern modal
identification algorithms can be outlined briefly as in Sect. 2.1.11. This two stage
procedure can be elaborated as follows:

• MPE-1: Utilize measured FRF or IRF data with a matrix coefficient, polynomial
model to find multiple estimates of the modal parameters. Select the best set of
modal frequencies and modal participation vectors.
– Choose polynomial or subspace order (m).
– Load measured data into over-determined linear equation form.
– Solve for unknown matrix coefficients ([αk ]).
* Add sufficient block kernel equations (for different starting times or
frequencies) until there are more block kernel equations than there are
block unknowns ([αk ]).
* Solve for matrix coefficients ([αk ]) in a least squares (LS) sense.
– Solve for complex valued modal frequencies for (λr or zr )
* Formulate companion matrix.
* Obtain eigenvalues of companion matrix. (λr or zr ).
· Time Domain: convert eigenvalues from (λr or zr ).
· Frequency Domain: compensate for generalized frequency.
* Obtain modal participation vectors (Lqr ) or modal vectors (ψr ) from
eigenvectors of the companion matrix, normalize as needed
– Iterate over different polynomial orders or subspace orders in the kernel
equation to get multiple estimates of the modal parameters.
– Select one set of complex valued modal frequencies, with associated partici-
pation vectors, from the multiple sets. Manual or automated selection methods
can be used.
• MPE-2: Utilize measured FRF or IRF data, with selected complex valued modal
frequencies and modal participation vectors, to find complex valued modal
vectors and modal scaling from an overdetermined set of linear equations.
– Solve for modal vectors using a weighted least squares solution method
involving the fixed set of modal frequencies and modal participation (weight-
ing) vectors. Solution process often involves the estimation of residues.
– Find modal vectors and modal scaling from Equations 75 through 79.
– Normalize modal vectors as needed.
– Solve for modal scaling associated with normalized modal vectors.

4.1.1 Consistency Diagrams

The MPE-1 stage is generally repeated for different matrix coefficient dimensions,
model order, and/or equation normalizations to find statistically consistent sets
of modal parameter results (sometimes referred to as clusters). Different kernel
equations can also be combined to provide a statistical set of answers across all
methods. Normally, 20 to 40 solution iterations are combined in a single diagram
referred to as a consistency diagram. The consistency diagram shows the consistent
582 R. J. Allemang and D. L. Brown

Fig. 14 Consistency diagram example

modal parameter results via a set of symbols denoting consistency in: modal
frequency, modal frequency and damping, and modal frequency, damping and
vectors. Historically, the consistency diagram was referred to as a stability diagram.
Figure 14 is an example of a consistency diagram. More examples and further
explanation of such validation methods are given in  Chap. 11, “Experimental
Modal Parameter Evaluation Methods” of this Handbook.
Once the final choice of modal frequency, modal damping, modal participation
vectors, and modal scaling is found from the consistency diagram or other validation
methods, the MPE-2 stage is implemented with the known information to determine
the final scaled modal vectors.

4.2 Current MPE Methods

Using the concepts developed in the previous section, the most commonly used
modal identification methods can be summarized as shown in Table 2. The high-
order model is typically used for those cases where the system is under-sampled in
the spatial domain. For example, the limiting case is when only one measurement is
made on the structure. For this case, the left-hand side of the general linear equation
corresponds to a scalar polynomial equation with the order equal to or greater than
the number of desired modal frequencies.
10 Experimental Modal Analysis Methods 583

Table 2 Modal vector estimation methods: overview

Domain Matrix polynomial order Coefficients
Algorithm Time Freq Zero Low High Scalar Matrix
Complex exponential • • •
algorithm (CEA)
Least-squares complex • • •
exponential (LSCE)
Polyreference time domain • • NS × NS
(PTD)
Ibrahim time domain (ITD) • • N L × NL
Multiple reference Ibrahim • • NL × NL
time domain (MRITD)
Eigensystem realization • • NL × NL
algorithm (ERA)
Polyreference frequency • • NL × NL
domain (PFD)
Simultaneous frequency • • NL × NL
domain (SFD)
Multi-reference frequency • • NL × NL
domain (MRFD)
Rational fraction polynomial • • • NS × NS
(RFP)
Orthogonal polynomial (OP) • • • N S × NS
Polyreference least-squares • • • NS × NS
complex frequency (PLSCF)
Complex mode indication • • NL × NS
function (CMIF)

The low-order model is used for those cases where the spatial information is
complete. In other words, the number of physical coordinates (NL ) is greater than
the number of desired (positive) modal frequencies (N ). For this case, the order (m)
of the left-hand side of the general linear equation, Eq. 46 or 50, is equal to two.
The zero-order model corresponds to a cases where the temporal information is
neglected and only the spatial information is used. These methods directly estimate
the eigenvectors as a first step. In general, these methods are programmed to
process data at a single temporal condition or variable. In this case, the method
is essentially equivalent to the single degree of freedom (SDOF) methods which
have been used with frequency response functions. In others words, the zeroth-order
matrix polynomial model compared to the higher-order matrix polynomial models is
similar to the comparison between the SDOF and MDOF methods used historically
in modal parameter estimation.
Modal parameter estimation algorithms are similar in more ways than they are
different. Fundamentally, all algorithms can be developed beginning with a linear,
constant coefficient, symmetric matrix model involving mass, damping, and stiff-
ness. The common goal in all algorithms, therefore, is the development of a charac-
584 R. J. Allemang and D. L. Brown

teristic matrix coefficient equation that describes a linear, time invariant, reciprocal
mechanical system consistent with this theoretical background. This is the rationale
behind using the Unified Matrix Polynomial Approach as the educational basis for
demonstrating this common kernel for all modal parameter estimation algorithms
[1, 2, 3, 4, 5, 6]. The following sections discuss the similarity of the kernel equations
common to all widely used modal parameter estimation algorithms.

4.3 Kernel Equations: Time Domain Algorithms

The following are the kernel equations for low- to high-order time domain algo-
rithms. While both first- and second-order kernel equations are included to be
consistent with historical development, research has shown that the two approaches,
first order and second order, give exactly the same values in the estimated coefficient
matrices when exactly the same data is utilized [61].

4.3.1 High-Order Methods

Summary: High-Order Time Domain Methods

Typical algorithms

• Complex exponential (CE)

• Least-squares complex exponential (LSCE)
• Polyreference time domain (PTD)

General linear equation formulation

• Model order (m ≥ 2N/NS ) where N is the number of desired positive

frequency roots
• Number of roots (m × NS )
• Data matrix size (NS × NL )
• Alpha coefficient matrix size (NS × NS )

Kernel equation: high-order coefficient normalization

⎡ ⎤
[h(ti+0 )]
  ⎢ [h(ti+1 )] ⎥  
[α0 ] [α1 ] · · · [αm−1 ] N ×mN ⎢ ⎥ = − [h(ti+m )]
S S⎣ ··· ⎦ NS ×NL
[h(ti+m−1 )] mN
S ×NL
(120)
10 Experimental Modal Analysis Methods 585

Kernel equation: zero-order coefficient normalization

⎡ ⎤
[h(ti+1 )]
  ⎢ [h(ti+2 )] ⎥  
[α1 ] [α2 ] · · · [αm ] N ×mN ⎢ ⎥ = − [h(ti+0 )] (121)
S S⎣ ··· ⎦ NS ×NL
[h(ti+m )] mN
S ×NL

4.3.2 Low-Order Methods: First Order

Summary: Low-Order (First) Time Domain Methods

Typical algorithms

• Ibrahim time domain (ITD)

• Multiple reference time domain (MRITD)
• Eigensystem realization algorithm (ERA)

General linear equation formulation

• Model order (m = 1)
• Number of roots (1 × 2NL )
• Data matrix size (NL × NS )
• Alpha coefficient matrix size (2NL × 2NL )

Kernel equation: high-order coefficient normalization

   
[h(ti+0 )] [h(ti+1 )]
[[α0 ]]2NL ×2NL =− (122)
[h(ti+1 )] 2NL ×NS
[h(ti+2 )] 2N
L ×NS

Kernel equation: zero-order coefficient normalization

   
[h(ti+1 )] [h(ti+0 )]
[[α1 ]]2NL ×2NL =− (123)
[h(ti+2 )] 2NL ×NS
[h(ti+1 )] 2N
L ×NS
586 R. J. Allemang and D. L. Brown

4.3.3 Low-Order Methods: Second Order

Summary: Low-Order (Second) Time Domain Methods

Typical algorithms

• Time domain equivalent to polyreference frequency domain

General linear equation formulation

• Model order (m = 2)
• Number of roots (2 × NL )
• Data matrix size (NL × NS )
• Matrix coefficients (NL × NL )

Kernel equation: high-order coefficient normalization

   
  [h(ti+0 )]
[α0 ] [α1 ] NL ×2NL
= − [h(ti+2 )] (124)
[h(ti+1 )] 2NL ×NS
NL ×NS

Kernel equation: zero-order coefficient normalization

   
  [h(ti+1 )]
[α1 ] [α2 ] NL ×2NL
= − [h(ti+0 )] (125)
[h(ti+2 )] 2NL ×NS
NL ×NS

4.4 Kernel Equations: Frequency Domain Algorithms

The following are the kernel equations for low- to high-order frequency domain
algorithms. While both first- and second-order kernel equations are included to be
consistent with historical development, research has shown that the two approaches,
first order and second order, give exactly the same values in the estimated coefficient
matrices when exactly the same data is utilized [61].

4.4.1 Generalized Frequency

High-order, frequency domain (rational fraction polynomial) methods all have
distinctive numerical problems that make the solution process problematic. Low-
order, frequency domain (rational fraction polynomial) methods do not have this
numerical problem.
The numerical problem associated with high-order, frequency domain methods
can be highlighted by looking at the characteristics of the data matrix involved in
estimating the matrix coefficients. These matrices involve power polynomials that
are functions of increasing powers of the frequency, typically si = j ωi . These
10 Experimental Modal Analysis Methods 587

matrices are of the Van der Monde form and are known to be ill-conditioned for
cases involving wide frequency ranges and high-ordered models.

Van der Monde Matrix Form:

⎡ ⎤ ⎡ ⎤
(s1 )0 (s1 )1 (s1 )2 · · · (s1 )m (s1 )m (s2 )m (s3 )m ··· (si )m
⎢ (s )0 m⎥
· · · (s2 ) ⎥ ⎢ ··· ⎥
⎢ 2 (s2 )1 (s2 )2 ⎢ ··· ··· ··· ··· ⎥
⎢ ⎥ ⎢ ⎥
⎢ (s3 )0 (s3 )1 (s3 )2 · · · (s3 )m ⎥ or ⎢ (s1 )2 (s2 )2 (s3 )2 ··· (si )2 ⎥ (126)
⎢ ⎥ ⎢ ⎥
⎣ ··· ··· ··· ··· ··· ⎦ ⎣ (s1 )1 (s2 )1 (s3 )1 ··· (si )1 ⎦
(si )0 (si )1 (si )2 · · · (si ) m
(s1 )0 (s2 )0 (s3 )0 ··· (si )0

The above two matrices are just transposes of one another; both have the same
ill-conditioning, and both are still considered Van der Monde matrix forms. The
matrix on the right will be used in upcoming graphical examples.
The ill-conditioned characteristic of matrices that are of the Van der Monde form
can be reduced, but not eliminated, by the following:

• Minimizing the frequency range of the data

• Minimizing the order of the model
• Normalizing the frequency range of the data (-1,1) or (-2,2)
• Use of orthogonal polynomials
• Use of complex z mapping

The last three methods involved the concept of generalized frequency where
the actual FRF complex data is not altered in any way but is remapped to a new
generalized or virtual frequency which eliminates or reduces the ill-conditioning
caused by the weighting of the FRF data by the physical frequency (si = j ωi ) in
the linear equations. These concepts are explained further in Sect. 5.2. Note that the
use of generalized frequency concepts does not change the kernel equation for the
high-order, frequency domain but does improve the solution as different frequency
weighting appears in the kernel equation.

4.4.2 High-Order Methods

Summary: High-Order Frequency Domain Methods

Typical algorithms

• Rational fraction polynomial (RFP)

• Orthogonal polynomial (OP)
• Polyreference least-squares complex frequency (PLSCF)

(continued)
588 R. J. Allemang and D. L. Brown

General linear equation formulation

• Model order (m ≥ 2N/NS ) where N is the number of desired positive

frequency roots
• Number of roots (m × NS )
• Data matrix size (NS × NL )
• Alpha coefficient matrix size (NS × NS )

Kernel equation: high-order coefficient normalization

 
[α0 ] [α1 ] · · · [αm−1 ] [β0 ] [β1 ] · · · [βn ] N ×(mN +(n+1)N )
S S L
⎡ ⎤
(si ) [H (ωi )]
0
⎢ (si )1 [H (ωi )] ⎥
⎢ ⎥
⎢ ··· ⎥
⎢ ⎥
⎢ ⎥
⎢ (si )m−1
[H (ωi )] ⎥
×⎢ ⎥ = −(si )m [H (ωi )]NS ×NL
⎢ −(si ) [I ] ⎥
0
⎢ ⎥
⎢ −(si )1 [I ] ⎥
⎢ ⎥
⎣ ··· ⎦
−(si ) [I ]
n
(mNS +(n+1)NL )×NL
(127)

Kernel equation: zero-order coefficient normalization

 
[α1 ] [α2 ] · · · [αm ] [β0 ] [β1 ] · · · [βn ] N ×mN +(n+1)N
i i o
⎡ ⎤
(si ) [H (ωi )]
1
⎢ (si )2 [H (ωi )] ⎥
⎢ ⎥
⎢ ··· ⎥
⎢ ⎥
⎢ ⎥
⎢ (si ) [H (ωi )] ⎥
m
×⎢ ⎥ = −(si )0 [H (ωi )]Ni ×No
⎢ −(si ) [I ] ⎥
0
⎢ ⎥
⎢ −(si )1 [I ] ⎥
⎢ ⎥
⎣ ··· ⎦
−(si ) [I ]
n
mNi +(n+1)No ×No
(128)
10 Experimental Modal Analysis Methods 589

4.4.3 Low-Order Methods: First Order

Summary: Low-Order (First) Frequency Domain Methods

Typical algorithms

• Simultaneous frequency domain (SFD)

• Multiple reference simultaneous frequency domain (MRSFD)
• Frequency domain equivalent to ITD, MRITD, ERA

General linear equation formulation

• Model order (m = 1)
• Number of roots (1 × 2NL )
• Data matrix size (NL × NS )
• Alpha coefficient matrix size (2NL × 2NL )

Kernel equation: high-order coefficient normalization

⎡ ⎤
(si )0 [H (ωi )]
  ⎢ (si )1 [H (ωi )] ⎥
[α0 ] [β0 ] 2N ×(2N +2N ) ⎢   ⎥
o o i ⎣ −(si )0 [I ] ⎦
−(si ) [I ]
1
(2No +2Ni )×Ni
 
(s )1 [H (ωi )]
=− i 2 (129)
(si ) [H (ωi )] 2N
o ×Ni

Kernel equation: zero-order coefficient normalization

⎡ ⎤
(si )1 [H (ωi )]
  ⎢ (si )2 [H (ωi )] ⎥
[α1 ] [β0 ] 2N ×(2N +2N ) ⎢   ⎥
o o i ⎣ −(si )0 [I ] ⎦
−(si ) [I ]
1
(2No +2Ni )×Ni
 
(si )0 [H (ωi )]
=− (130)
(si )1 [H (ωi )] 2N
o ×Ni
590 R. J. Allemang and D. L. Brown

4.4.4 Low-Order Methods: Second Order

Summary: Low-Order (Second) Frequency Domain Methods

Typical algorithms

• Polyreference frequency domain (PFD)

• Frequency domain direct parameter identification (FDPI)

General linear equation formulation

• Model order (m = 2)
• Number of roots (2 × NL )
• Data matrix size (NL × NS )
• Alpha coefficient matrix size (NL × NL )

Kernel equation: high-order coefficient normalization

⎡ ⎤
(si )0 [H (ωi )]
  ⎢ (si )1 [H (ωi )] ⎥
[α0 ] [α1 ] [β0 ] [β1 ] N ×(2N +2N ) ⎢ ⎥
o o i ⎣ −(si )0 [I ] ⎦

−(si )1 [I ] (2No +2Ni )×Ni

= −(si )2 [H (ωi )]No ×Ni (131)

Kernel equation: zero-order coefficient normalization

⎡ ⎤
(si )1 [H (ωi )]
  ⎢ (si )2 [H (ωi )] ⎥
[α1 ] [α2 ] [β0 ] [β1 ] N ×(2N +2N ) ⎢ ⎥
o o i ⎣ −(si )0 [I ] ⎦

−(si )1 [I ] (2N o +2Ni )×Ni

= −(si )0 [H (ωi )]No ×Ni (132)

4.5 Residue (Modal Vector) Estimation

Modal vectors are normally found directly when a weighted least-squares solution is
utilized, where the weighting comes from the modal participation vectors found in
the first stage of the modal parameter estimation (MPE-1). This is the case for most
multiple reference methods. If residues are desired to complete a partial fraction
model, they can be found from combinations of the modal participation vector
coefficients and the modal vector coefficients.
10 Experimental Modal Analysis Methods 591

For the single reference case, residues are first found directly, and then the modal
vectors are found by normalizing the residue vectors. Then the modal scaling can
be estimated.
Historically, residues and modal vectors were found with both time and fre-
quency domain estimations. Since it is easiest to involve residuals in frequency
domain methods, only frequency domain methods are utilized in most solution
procedures today.
As residues or modal vectors are estimated, the number (N ) of modal frequencies
and modal participation vectors are already known. This means that the solution
procedure will always involve a least squares (LS), weighted least squares (WLS)
or total least-squares (TLS) approach.

4.5.1 Time Domain Methods

Time domain estimation (single reference)
⎧ ⎫ ⎡ ⎤ ⎧ ⎫
⎪
⎪ hpq (t1 ) ⎪
⎪ eλ1 t1 eλ2 t1 eλ3 t1 · · · eλ2N t1 ⎪
⎪ Apq1 ⎪⎪
⎪
⎪ ⎪
⎪ ⎢ ⎥ ⎪
⎪ ⎪
⎪
⎨ hpq (t2 ) ⎪
⎪ ⎬ ⎢ eλ1 t2 eλ2 t2 eλ3 t2 · · · eλ2N t2 ⎥ ⎪
⎨ Apq2 ⎪⎬
⎢ ⎥
hpq (t3 ) = ⎢ eλ1 t3 eλ2 t3 eλ3 t3 · · · eλ2N t3 ⎥ Apq3
⎪
⎪ ··· ⎪ ⎪
⎪ ⎢ ⎥ ⎪
⎪ ⎪
⎪
⎪ ⎪ ⎣ ··· ··· ··· ··· ··· ⎦ ⎪
⎪ ··· ⎪
⎪
⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
h (t )
pq Nt Nt ×1
eλ1 tNt eλ2 tNt eλ3 tNt · · · eλ2N tNt Nt ×2N
Apq2N 2N ×1
(133)
where

• Nt = Number of time points

• N = Number of modal frequencies (positive and negative)
• Nt ≥ 2N

Time domain estimation (multiple references)

' (
{h(ti )}T Ni ×1 = [L]Ni ×2N eλr ti 2N ×2N {ψ}T 2N ×1 (134)

where

• Nt = Number of time points

• N = Number of modal frequencies (positive and negative)
• Nt ≥ 2N

Above equation is repeated Nt times

⎧ ⎫ ⎡ ⎤
⎪
⎪ hp1 (ti ) ⎪
⎪ eλ1 ti 0 0 ··· 0
⎪
⎪ ⎪
⎪ ⎢ 0 eλ2 ti 0
⎪
⎨ hp2 (ti ) ⎪
⎬ ' λt ( ⎢ ··· 0 ⎥ ⎥
⎢ ⎥
hpq (ti ) = hp3 (ti ) e i =⎢ 0 0 eλ3 ti ··· 0 ⎥ (135)
⎪
⎪ ⎪
⎪ ⎢ ⎥
⎪
⎪ ··· ⎪ ⎪ ⎣ ··· ··· ··· ··· ··· ⎦
⎪
⎩ h (t ) ⎪ ⎭
pq i 0 0 0 · · · eλ2N ti
592 R. J. Allemang and D. L. Brown

The residues are calculated from the modal participation vector coefficients (L)
and the modal coefficients (ψ); (Apqr = Lqr ψpr ). Note that if one column q of
the modal participation matrix ([L]) is normalized to unity, the modal coefficients
that are found will be equal to the residues for that reference (Apqr ).
⎧ ⎫ ⎡ ⎤
⎪
⎪ ψp1 ⎪⎪ L11 L12 L13 · · · L12N
⎪
⎪ ⎪
⎪ ⎢L
 
⎪
⎨ ψ p2 ⎪
⎬   ⎢ 21 L22 L23 · · · L22N ⎥ ⎥
⎢ ⎥
ψpr = ψp3 Lrq = ⎢ L31 L32 L33 · · · L32N ⎥ (136)
⎪
⎪ ⎪ ⎢ ⎥
⎪
⎪ ··· ⎪⎪
⎪ ⎣ ··· ··· ··· ··· ··· ⎦
⎪
⎩ψ ⎪
⎭
p2N LNi 1 LNi 2 LNi 3 · · · LNi 2N

4.5.2 Frequency Domain Methods

Frequency domain estimation (Single Reference with Residuals)
 
1
Hpq (ω) N ×1 = Apqr (2N +2)×1 (137)
s j ω − λr Ns ×(2N +2)

where

• Ns = Number of spectral lines (frequency points)

• N = Number of modal frequencies (positive and negative)
• Ns ≥ 2N + 2
⎡ −1 ⎤
1
j ω1 −λ1
1
j ω1 −λ2
1
j ω1 −λ3 ··· 1
j ω1 −λ2N 1
ω12
⎢ −1 ⎥
  ⎢ 1 1 1
··· 1
1 ⎥
⎢ j ω2 −λ1 j ω2 −λ2 j ω2 −λ3 j ω2 −λ2N ω22 ⎥
1 ⎢ −1 ⎥
=⎢ 1
j ω3 −λ1
1
j ω3 −λ2
1
j ω3 −λ3 ··· 1
j ω3 −λ2N 1 ⎥ (138)
j ω − λr ⎢ ω32 ⎥
⎢ ··· ··· ··· ··· ··· ··· ···⎥
⎣ ⎦
−1
1 1 1
j ωNs −λ1 j ωNs −λ2 j ωNs −λ3 ··· 1
j ωNs −λ2N ω 2 1
Ns
⎧ ⎫
⎪
⎪ Apq1 ⎪ ⎪ ⎧ ⎫
⎪
⎪ ⎪
⎪
⎪ Apq2 ⎪ ⎪
⎪ ⎪
⎪ Hpq (ω1 ) ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎨ Apq3 ⎪
⎪ ⎬ ⎪
⎨ Hpq (ω2 ) ⎪
⎬
Apqr = ··· Hpq (ω) = Hpq (ω3 ) (139)
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎪ Apq2N ⎪⎪
⎪
⎪
⎪
⎪ ··· ⎪
⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎩ H (ω ) ⎪ ⎭
⎪
⎪ R ⎪
pq ⎪
⎪
⎩
I ⎪
⎭
pq Ns
RFpq

Frequency domain estimation (multiple references)

) *
1
{H (ωi )}T Ni ×1 = [L]Ni ×2N {ψ}T 2N ×1 (140)
j ωi − λr 2N ×2N
10 Experimental Modal Analysis Methods 593

where

• Ns = Number of spectral lines (frequency points)

• N = Number of modal frequencies (positive and negative)
• Ns ≥ 2N
• Above equation is repeated Ns times

⎧ ⎫
⎪
⎪ Hp1 (ωi ) ⎪
⎪
⎪
⎪ H (ω ) ⎪ ⎪
⎪
⎨ p2 i ⎪ ⎬
Hpq (ωi ) = Hp3 (ωi ) (141)
⎪
⎪ ⎪
⎪
⎪ ··· ⎪ ⎪
⎪
⎪
⎩ H (ω ) ⎪ ⎭
pq i

⎡ ⎥
1
··· ⎥
j ωi −λ1 0 0 0 ⎥
⎢ ⎥
) * ⎢ 0 1
0 ··· 0 ⎥
1 ⎢ j ωi −λ2 ⎥
=⎢
⎢ 0 0 1
j ωi −λ3 ··· 0 ⎥
⎥ (142)
j ωi − λr ⎢ ⎥
⎢ ··· ··· ··· ··· ··· ⎦
⎢
⎢ 0 0 0 ··· 1
j ωi −λ2N

The residues are calculated from the modal participation vector coefficients (L)
and the modal coefficients (ψ); (Apqr = Lqr ψpr ). Note that if one column q of
the modal participation matrix ([L]) is normalized to unity, the modal coefficients
that are found will be equal to the residues for that reference Apqr .
⎧ ⎫ ⎡ ⎤
⎪
⎪ ψp1 ⎪ ⎪ L11 L12 L13 · · · L12N
⎪
⎪ ⎪
⎪ ⎢L ⎥
⎪ ⎪ 21 L22 L23 · · · L22N ⎥
  ⎨ ψp2 ⎬   ⎢ ⎢ ⎥
ψpr = ψp3 Lrq = ⎢ L31 L32 L33 · · · L32N ⎥ (143)
⎪
⎪ ⎪
⎪ ⎢ ⎥
⎪
⎪ · · · ⎪
⎪ ⎣ · · · · · · · · · · · · · · · ⎦
⎪
⎩ ⎪
⎭
ψp2N LNi 1 LNi 2 LNi 3 · · · LNi 2N

The details of how the data is positioned into the MIMO FRF model for the
estimation of modal vectors often require a little manipulation to see the matrix
conformal relationships. The following development expands each term so that the
overdetermined linear equation is easier to visualize.
The primary equation used for multiple reference modal vector estimation can be
restated as:

[[H (ω)]]NS ×NL ×Nf = [L]NS ×2N [[(ω)]]2N ×2N ×Nf [ψ]T 2N ×NL (144)
594 R. J. Allemang and D. L. Brown

where
⎡ ⎤ ⎡ ⎤
[H (ω1 )] [L] [(ω1 )]
⎢ [H (ω )] ⎥ ⎢ [L] [(ω )] ⎥
⎢ 2 ⎥ ⎢ 2 ⎥
⎢ ⎥ ⎢ ⎥
[[H (ω)]] = ⎢ [H (ω3 )] ⎥ [L] [[(ω)]] = ⎢ [L] [(ω3 )] ⎥
⎢ ⎥ ⎢ ⎥
⎣ ... ⎦ ⎣ ... ⎦
   
H (ωNf ) N [L] (ωNf ) N
S Nf ×NL S Nf ×2N
(145)

The modal participation matrix ([L]) is of the following form:

⎡ ⎤
L1,1 L2,1 L3,1 ... L2N,1
⎢L L2N,2 ⎥
⎢ 1,2 L2,2 L3,2 ... ⎥
⎢ ⎥
[L] = ⎢ L1,3 L2,3 L3,3 ... L2N,3 ⎥ (146)
⎢ ⎥
⎣ ... ... ... ... ... ⎦
L1,NS L2,NS L3,NS ... L2N,NS

Note that in the above notation, the modal participation vectors span the row
space with each column representing the participation information for another
mode. The modal vector matrix ([ψ]) is of the following form:
⎡ ⎤
ψ1,2 ψ1,2 ψ1,3 ... ψ1,NL
⎢ψ ψ2,NL ⎥
⎢ 2,1 ψ2,2 ψ2,3 ... ⎥
⎢ ⎥
[ψ]T = ⎢ ψ3,1 ψ3,2 ψ3,3 ... ψ3,NL ⎥ (147)
⎢ ⎥
⎣ ... ... ... ... ... ⎦
ψ2N,1 ψ2N,2 ψ2N,3 ... ψ2N,NL

Note that in the above notation, the transpose of the modal vector matrix span
the column space with each row representing another mode. Note that the notation
for both the [L] and the [ψ] matrices is consistent with the academic definition of a
modal matrix, but the solution method is actually finding the [ψ]T matrix.
The diagonal [(ωi )], at each frequency, is of the following form:
⎡ ⎤
1
⎢ j ωi − λ1 ⎥
⎢ ⎥
⎢ 1 ⎥
⎢ i2 ⎥
⎢ j ωi − λ2 ⎥
⎢ 1 ⎥
(ωi ) = ⎢
⎢ i3 ⎥
⎥ (148)
⎢ j ωi − λ3 ⎥
⎢ ⎥
⎢ .. ⎥
⎢ . ⎥
⎣ 1 ⎦
j ωi − λ2N
10 Experimental Modal Analysis Methods 595

Finally, the combined matrix of the [L] matrix and the [] matrix can be formed
at each frequency:

⎡ L1,1 L2,1 L3,1 L2N,1 ⎤

...
⎢ j ωi − λi j ωi − λ1 j ωi − λ1 j ωi − λ1 ⎥
⎢ L L2N,2 ⎥
⎢ 1,2 L2,2 L3,2 ⎥
⎢ ... ⎥
⎢ j ωi − λ2 j ωi − λ2 j ωi − λ2 j ωi − λ2 ⎥
⎢ L2N,3 ⎥
[L] [(ωi )] = ⎢ L1,3 L2,3 L3,3 ⎥ (149)
⎢ jω − λ j ωi − λ3 j ωi − λ3
...
j ωi − λ3 ⎥
⎢ i 3 ⎥
⎢ ⎥
⎢ ... ... ... ... ... ⎥
⎣ L1,NS L2,NS L3,NS L2N ,NS ⎦
...
j ωi − λNS j ωi − λNS j ωi − λNS j ωi − λNS

This combined matrix can now be stacked vertically so that the row space
matches the row space of the [[H (ω)]] matrix, as in Eq. 144. Now the solution
for the modal vectors can be found by multiplying both sides of Eq. 144 by the
pseudoinverse of the [L] [[(ω)]] matrix.
Note that in the above development, if the short dimension (NS ) of the FRF
matrix is one, the matrix in Eq. 144 will be the same as the matrix in Eq. 146.

5 Differences in MPE Algorithms

Modal parameter estimation algorithms typically give slightly different estimates of

modal parameters due to the way the FRF data is weighted and processed in the
computation of the matrix coefficients. Some of the most common variations are
discussed in the following sections.

5.1 Polynomial Coefficient Estimation

The polynomial coefficients that are used to define the matrix polynomial char-
acteristic equation are always found from an overdetermined set of block linear
equations. The answers for the modal parameters will always change if any degree of
overdetermination is changed. These changes will often be small if the overdetermi-
nation factor is four or more. The method of solving for the polynomial coefficients
when forming an overdetermined problem can also be changed. Originally and
historically, the approach was to form the direct solution by pre-multiplying the
equation with its transpose. Currently, this type of overdetermined problem is
generally solved with a singular value decomposition (SVD) approach. Any of
these differences in the implementation of the solution procedure will involve small
changes in ultimate estimate of the modal parameters. If the dynamic range of the
data in the overdetermined equations becomes very large, the impact on the modal
parameter estimates could become significant.
596 R. J. Allemang and D. L. Brown

5.2 Generalized Frequency

The fundamental problem with using a high-order, frequency domain method (e.g.,
rational fraction polynomial) can be highlighted by looking at the characteristics
of the data matrix involved in estimating the matrix coefficients. These matrices
involve power polynomials that are functions of increasing powers of the frequency,
typically si = j ωi . These matrices are of the Van der Monde form and are known
to be ill-conditioned for cases involving wide frequency ranges and high-ordered
models.

Van der Monde matrix form:

⎡ ⎤
(s1 )0 (s1 )1 (s1 )2 ··· (s1 )2N −1
⎢ (s )0 (s )1 (s2 )2 ··· (s2 )2N −1 ⎥
⎢ 2 2 ⎥
⎢ ⎥
⎢ (s3 )0 (s3 )1 (s3 )2 ··· (s3 )2N −1 ⎥ (150)
⎢ ⎥
⎣ ··· ··· ··· ··· ··· ⎦
(si )0 (si )1 (si )2 ··· (si )2N −1

The ill-conditioning problem can be best understood by evaluating the condition

number of the Van der Monde matrix. The condition number measures the sensi-
tivity of the solution of linear equations to errors, or small amounts of noise, in the
data. The condition number gives an indication of the accuracy of the results from
matrix inversion and/or linear equation solution. The condition number for a matrix
is computed by taking the ratio of the largest singular value to the smallest singular
value. A good condition number is a small number close to unity; a bad condition
number is a large number. For the theoretical case of a singular matrix, the condition
number is infinite.
As mentioned in Sect. 4.4.1, the ill-conditioned characteristic of matrices that are
of the Van der Monde form can be reduced, but not eliminated, by the following:

• Minimizing the frequency range of the data

• Minimizing the order of the model
• Normalizing the frequency range of the data (-1,1) or (-2,2)
• Use of orthogonal polynomials
• Use of complex z mapping

The last three methods involved the concept of generalized frequency where
the actual FRF complex data is not altered in any way but is remapped to a new
generalized or virtual frequency which eliminates or reduces the ill-conditioning
caused by the weighting of the FRF data by the physical frequency (si = j ωi ) in
the linear equations. These concepts are briefly explained in the following sections.

5.2.1 Normalized Frequency

The simplest method of using the generalized frequency concept is to normalize the
power polynomials by utilizing the following equation:
10 Experimental Modal Analysis Methods 597

Order: 8 Condition Number: 5.48e+002

1
Power Polynomial Functions − Real Part

0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec)

Imaginary: Order: 8 Condition Number: 5.48e+002

1
Power Polynomial Functions − Imag Part

0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec) - Imag part

Fig. 15 Van der Monde matrix – normalized frequency – orders 0–8

si = j ∗ (ωi /ωmax ) (151)

This gives a generalized frequency variable that is bounded by (-1,1) with much
better numerical conditioning than utilizing the raw frequency range (-ωmax ,ωmax ).
When the modal frequencies are estimated, the corrected modal frequencies must
be determined by multiplying by the normalizing frequency (ωmax ). All frequency
domain algorithms, at a minimum, will use some form of this frequency normal-
ization. The graphical plot of this Van der Monde matrix for orders 0 through 8 is
shown in Fig. 15.

5.2.2 Orthogonal Polynomials

In the past, the only way to avoid the numerical problems inherent in the frequency
domain methods (Van der Monde matrix), even when normalized frequencies are
598 R. J. Allemang and D. L. Brown

implemented, is to use a transformation from power polynomials to orthogonal

polynomials [25, 26, 28, 29, 30, 31, 32, 33, 36]. Any power polynomial series can
be represented by an equivalent number of terms in an orthogonal polynomial
series. Several orthogonal polynomials have been applied to the ill-conditioning
problem, such as Forsythe polynomials [25] and Chebyshev polynomials [26, 27].
The orthogonal polynomial concept is represented by the following relationship.


m 
m
(si )k αk = Pk (si )γk (152)
k=0 k=0

The orthogonal polynomial series can be formed by the following relationships:

P0 (si ) = 1.0 (153)

Pj (si∗ ) = Pj∗ (si ) (154)

n
Pn+1 (si ) = an si Pn (si ) − bn,k Pk (si ) (155)
k=0

This orthogonal polynomial series can be formulated in matrix form by utilizing

two weighting matrices involving the coefficients an and bn,k as follows:
⎡ ⎤ ⎡ ⎤
an 0 0 ··· 0 bn,n bn,n−1 bn,n−2 ··· bn,0
⎢ 0 a ··· 0 ⎥ ⎢ 0 b ··· bn−1,0 ⎥
⎢ n−1 0 ⎥ ⎢ n−1,n−1 bn−1,n−2 ⎥
⎢ ⎥ ⎢ ⎥
[Wa ] = ⎢ 0 0 an−2 ··· 0 ⎥ [Wb ] = ⎢ 0 0 bn−2,n−2 ··· bn−2,0 ⎥
⎢ ⎥ ⎢ ⎥
⎣··· ··· ··· ··· ···⎦ ⎣ ··· ··· ··· ··· ··· ⎦
0 0 0 ··· a0 0 0 0 ··· b0,0
(156)

Different orthogonal polynomials are generated using different weighting coef-

ficients and are orthogonal over different ranges. For example, Forsythe orthogonal
polynomials are orthogonal over the (-2,2) range, while Chebyshev orthogonal
polynomials are orthogonal over the (-1,1) range. In the orthogonal polynomial
approach, the original complex valued FRF data is used together with the orthogonal
polynomial coefficients Pk (si ) in place of the generalized frequency (si )k where si is
the properly normalized generalized frequency (e.g., Eq. 152). The unknown matrix
coefficients of the matrix orthogonal polynomial (γk ) are substituted in place of the
original matrix coefficients (αk ). These matrix orthogonal polynomial coefficients
are then loaded in the companion matrix ([C]) as before.
When this orthogonal polynomial transformation is used to generate a new
generalized frequency, the corrected modal frequencies are determined from a
generalized form of the companion matrix solution. The companion matrix ([C])
is determined in the same way as always, but the solution for the modal frequencies
is found from the following equation (note that the following form originates from
a high-order coefficient normalization):
10 Experimental Modal Analysis Methods 599

Order: 8 Condition Number: 3.22e+000

1
Chebyshev Polynomial Functions − Real Part

0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec)

Imaginary: Order: 8 Condition Number: 3.22e+000

Chebyshev Polynomial Functions − Imag Part

1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec) - Imag part

Fig. 16 Van der Monde matrix – Chebyshev orthogonal polynomials – orders 0–8

[[C] + [Wb ]] {X} = λ [Wa ] {X} (157)

The above equation can also be formed for the low-order coefficient normaliza-
tion using the same process as that shown in Eqs. 56 through 61.
The graphical plot of the Van der Monde matrix for orders 0 through 8 for a set
of Chebyshev orthogonal polynomials is shown in Fig. 16.

Discrete Orthogonal Polynomials

Most orthogonal polynomial methods are based upon continuous polynomials.

Since the measurements are made at discrete frequencies, this means that the
orthogonality is approximate. Discrete orthogonal polynomials are based upon the
discrete frequencies involved and provide a more exact orthogonality with added
600 R. J. Allemang and D. L. Brown

accuracy. Both continuous and discrete orthogonal polynomials are summarized in

a recent publication [35].

5.2.3 Complex Z Mapping

The important contribution behind the development of the polyreference LSCF
[37, 38, 39, 40, 41, 42] method is the recognition of a new method of frequency
mapping. The generalized frequency in this approach is a trigonometric mapping
function (complex z) that has superior numerical conditioning to orthogonal
polynomials without the added complication of solving a generalized companion
matrix eigenvalue problem. This approach can be applied to any frequency domain
method, low-order frequency domain methods, as well as high-order frequency
domain methods, although the numerical advantage is not as profound for the low-
order methods.
The basic complex Z mapping function, in the nomenclature of this presentation,
is as follows:

si = zi = ej ∗π ∗(ωi /ωmax ) = ej ∗ωi ∗t (158)

sim = zim = ej ∗π ∗m∗(ωi /ωmax ) (159)

This mapping function maps the positive frequency range to the positive unit
circle in the complex plane and the negative frequency range to the negative unit
circle in the complex plane. This is graphically represented in Fig. 17.
This effectively yields a real part of the mapping functions which are cosine
terms and an imaginary part which are sin functions. Since sin and cos functions

200
200
100
Phase

Phase

0
0
−200
−100 1 1
0 0
−200 Imag −1 −1
−100 −50 0 50 100 Real
Frequency [Hz]

102
Magnitude

40
Magnitude

101 20

0
1
1
0 0
100 Imag −1 −1
−100 −50 0 50 100
Real
Frequency [Hz]

Fig. 17 Mapping the frequency response function onto the unit circle in the complex plane
10 Experimental Modal Analysis Methods 601

Order: 8 Condition Number: 1.01e+000

1
0.8
Z Mapping Functions − Real Part

0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec)

Imaginary: Order: 8 Condition Number: 1.01e+000

1
0.8
Z Mapping Functions − Real Part

0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Normalized Frequency (Rad/Sec) − Imag Part

Fig. 18 Van der Monde matrix – complex Z mapped frequency – orders 0–8

at different frequencies are mathematically orthogonal, the numerical conditioning

of this mapping function is quite good. The graphical plot of this Van der Monde
matrix for orders 0 through 8 is shown in Fig. 18. The condition number for this
example matrix is 1.01 (Fig. 18) compared to a condition number of 548 for the
normalized frequency example (Fig. 15).

5.3 Data Sieving/Filtering/Decimation

For almost all cases of modal identification, a large amount of redundancy or

overdetermination exists. This means that the number of equations available
compared to the number required to form an exactly determined solution, defined
as the overdetermination factor, will be quite large. Beyond some value of overde-
602 R. J. Allemang and D. L. Brown

Frequency Response Function

0
Magnitude 10

−5
10

−10
10
0 500 1000 1500 2000 2500
Frequency (Hz)

−4 Impulse Response Function

x 10
4

2
Magnitude

−2

−4
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Time (sec)

Fig. 19 Filtering example

termination factor, the additional equations contribute little to the result but may
add significantly to the noise and, thus, to the solution time. For this reason, the data
space is often filtered (limited within minimum and maximum temporal axis values),
sieved (limited to prescribed input DOFs and/or output DOFs) and/or decimated
(limited number of equations from the allowable temporal data) in order to obtain a
reasonable result in the minimum time. The above figure is an example of choosing
a frequency band of the frequency response function (filtering) and working with
the inverse FFT of that frequency banded data in the time domain impulse response
function (Fig. 19).

5.4 Coefficient Condensation (Virtual DOFs)

For the low-order modal identification algorithms (model order 2), the number
of physical coordinates (typically NL ), which dictates the size of the coefficient
matrices ([αk ]), is often much larger than the number of desired modal frequencies
(N). For this situation, the numerical solution procedure is constrained to solve for
NL or 2NL modal frequencies. This can be very time consuming and is unnecessary.
One simple approach to reducing the size of the coefficient matrices is to sieve the
physical DOFs to temporarily reduce the dimension of NL . Beyond excluding all
10 Experimental Modal Analysis Methods 603

physical DOFs in a direction or those that are part of a subcomponent, this is difficult
to do in an effective manner that will retain the correct information from the FRF
data matrix.
The number of physical coordinates (NL ) can be reduced to a more reasonable
size (Ne ≈ N or Ne ≈ 2N ) by using a decomposition transformation from physical
coordinates (NL ) to the approximate number of effective modal frequencies (Ne ).
These resulting Ne transformed coordinates are sometimes referred to as virtual
DOFs. Currently, singular value decompositions (SVD) or eigenvalue decompo-
sitions (ED) are used to condense the spatial information while preserving the
principal modal information prior to formulating the linear equation solution for
unknown matrix coefficients [19, 62, 63].
It is important to understand that the ED and SVD transformations yield a
mathematical (linear) transformation that, in general, contains complex valued
vectors as part of the transformation matrix, [T ]. Conceptually, the transformation
will work well when these vectors are estimates of the modal vectors of the
system, normally a situation where the vectors can be scaled to real valued vectors.
Essentially, this means that the target goal of the transformation is a transformation
from physical space to modal space. As the modal density increases and/or as the
level of damping increases, the ED and SVD methods give erroneous results, if the
complete [H ] matrix is used. Generally, superior results will be obtained when the
imaginary part of the [H ] matrix is used in the ED or SVD transformation, thus
forcing a real valued transformation matrix, [T ]. Another option is to load both the
real and imaginary portions of the complex data into a real matrix which will also
force a real valued transformation matrix [64].
In order to form a consistency diagram using virtual DOFs, several approaches
can be used. In most cases, when the long dimension is large, results from different
subspaces can be combined in a consistency diagram to look for statistically
significant results. This is normally done by using a subspace of dimension two
and then successively increasing the size of the subspace to some limit that will
give 30 to 40 solution iterations up to the limit of N modal frequencies (30-40). The
largest subspace size is roughly equal to the number of desired modal frequencies.
An alternate approach is to choose a relatively small size for the dimension of the
subspace and then using model order iteration, beginning with model order 2 up to
a model order that will yield the number of required modal frequencies.

[H  ] = [T ] [H ] (160)

where

• [H  ] is the transformed (virtual or condensed) frequency response function

matrix.
• [T ] is the transformation matrix.
• [H ] is the original FRF matrix.
604 R. J. Allemang and D. L. Brown

The difference between the two techniques lies in the method of finding the
transformation matrix, [T ]. Once [H ] has been condensed, however, the parameter
estimation procedure is the same as for the full data set. Because the data eliminated
from the parameter estimation process ideally corresponds to the noise in the data,
the poles of the condensed data are the same as the poles of the full data set.
However, the participation factors calculated from the condensed data may need
to be expanded back into the full space.

[H  ] = [T ] [H ]
(161)
[] = [T ]T [  ]

where

• [] is the full-space participation matrix.

• [  ] is the condensed-space participation matrix.

While linear decomposition methods known as eigenvalue decomposition (ED)

and singular value decomposition (SVD) are most often used to find the transforma-
tion matrix ([T ]), a number of other methods have been developed in various areas
of mathematics and science that find linear orthogonal subspaces of a data matrix.
These methods are now grouped into a class of decomposition methods referred to
as principal component analysis [65].

5.4.1 Eigenvalue Decomposition

In the eigenvalue decomposition method, the [T ] matrix is composed of the
eigenvectors corresponding to the Ne largest eigenvalues of the power spectrum
of the FRF matrix as follows:

[H (ω)]NL ×NS Nf [H (ω)]H

NS Nf ×NL = [V ]  [V ]
H
(162)

The eigenvalues and eigenvectors are then found, and the [T ] matrix is con-
structed from the eigenvectors corresponding to the Ne largest eigenvalues:
 T
[T ]Ne ×NL = {v1 } {v2 } · · · {vk } · · · vNe (163)

where

• {vk } is the NL × 1 eigenvector corresponding to the k − th eigenvalue.

This technique may be adapted for condensing on the input space, as well.
The power spectrum matrix is again found, but the FRF matrix must be reshaped
(transposed) so that it is an NS × NL matrix for each spectral line:

[H (ω)]NS ×NL Nf [H (ω)]H

NL Nf ×NS = [V ]  [V ]
H
(164)
10 Experimental Modal Analysis Methods 605

The eigenvalues and eigenvectors are again found as before, and the transforma-
tion matrix [T ] becomes:

 T
[T ]Ne ×NS = {v1 } {v2 } · · · {vk } · · · vNe (165)

where

• {vk } is the NS × 1 eigenvector corresponding to the k − th eigenvalue.

5.4.2 Singular Value Decomposition

The singular value decomposition condensation technique is similar to the
eigenvalue-based technique but operates on the FRF matrix directly instead of
the power spectrum of the FRF matrix. The basis for this technique is the singular
value decomposition [66, 64] by which the matrix [H ] is broken down into three
component parts, [U ], [], and [V ]:

[H ]NL ×NS Nf = [U ]NL ×NL NL ×NL [V ]H

NL ×NS Nf (166)

The left-singular vectors corresponding to the Ne largest singular values are the
first Ne columns of [U ]. These become the transformation matrix [T ]:

 T
[T ]Ne ×NL = {u1 } {u2 } · · · {uk } · · · uNe (167)

where

• {uk } is the k − th column of [U ], which corresponds to the k − th singular value.

This technique may also be adapted for condensing the input space, as long as
the FRF matrix [H ] is reshaped (transposed) to an NS × NL matrix at each spectral
line. The SVD operation then becomes:

[H ]NS ×NL Nf = [U ]NS ×NS NS ×NS [V ]H

NS ×NL Nf (168)

The transformation matrix [T ] is still composed of the left singular vectors

corresponding to the Ne largest singular values,

 T
[T ]Ne ×Ni = {u1 } {u2 } · · · {uk } · · · uNe (169)

where

• {uk } is again the k − th column of [U ], which corresponds to the k − th singular

value.
606 R. J. Allemang and D. L. Brown

5.4.3 Virtual FRFs

One way of visualizing the effects of condensation on the FRF matrix is formulating
the condensed frequency response functions based on the reduced dimension of
the virtual DOFs. These FRFs are referred to as virtual FRFs or principal response
functions [64,67]. For a given decomposition, the virtual FRFs represent a subspace
of the full FRF matrix space. Using Eq. 160 for a given subspace, typically between
two and some fraction of the long dimension, the following figures represent virtual
FRFs for all of the subspaces above and below the chosen subspace. Note that the
short dimension remains the same for each subspace.
Figure 20 represents the virtual FRFs for a specific decomposition of the long
dimension of the FRF matrix. Note that the dynamics in the data are clearly shown
and that, therefore, the use of this decomposition (subspace) will give good results
for the estimated modal frequency information.
Figure 21 represents the virtual FRFs for all of the decompositions of the long
dimension of the FRF matrix that are more dominant than those shown in Fig. 20.
Again note that the dynamics of the data are clearly shown. The use of any of the
decompositions (subspaces) in these sets of virtual FRFs will give good results.
Figure 22 represents the virtual FRFs for all of the decompositions of the long
dimension of the FRF matrix that are less dominant than those shown in Fig. 20.
Again note that the dynamics of the data are clearly shown in a few of these virtual

Fig. 20 Virtual FRFs at a specific SVD

10 Experimental Modal Analysis Methods 607

Fig. 21 Virtual FRFs at smaller SVD dimensions

FRFs. Most of the virtual FRFs in these subspaces represent the noise on the FRF
data. The use of most of the decompositions (subspaces) in these sets of virtual
FRFs will give poor results.

5.5 Equation Condensation

Equation condensation methods are used to reduce the number of equations

generated from measured data to more closely match the number of unknowns in the
modal parameter estimation algorithms. There are a large number of condensation
algorithms available. Based upon the modal parameter estimation algorithms in use
today, the three types of algorithms most often used are:

• Least squares: Least squares (LS), weighted least squares (WLS), total least
squares (TLS), or double least squares (DLS) is used to minimize the squared
error between the measured data and the estimation model.
• Transformations: The measured data is reduced by approximating the data by the
superposition of a limited (reduced) set of independent vectors. The number of
significant, independent vectors is chosen equal to the maximum number modes
that are expected in the measured data. This set of vectors is used to approximate
608 R. J. Allemang and D. L. Brown

Fig. 22 Virtual FRFs at larger SVD dimensions

the measured data and used as input to the parameter estimation procedures.
Singular value decomposition (SVD) is an example of one of the more popular
transformation methods.
• Coherent averaging: Coherent averaging is another popular method for reducing
the data. In the coherent averaging method, the data is weighted by performing a
dot product between the data and a weighting vector (spatial filter). Information
in the data which is not coherent with the weighting vectors is averaged out of
the data.

The least squares and the transformation procedures tend to weight those modes
of vibration which are well excited. This can be a problem when trying to extract
modes which are not well excited. The solution is to use a weighting function for
condensation which tends to enhance the mode of interest. This can be accomplished
in a number of ways:

• In the time domain, a spatial filter or a coherent averaging process can be used to
filter the response to enhance a particular mode or set of modes.
• In the frequency domain, the data can be enhanced in the same manner as the
time domain plus the data can be additionally enhanced by weighting the data in
a frequency band near the natural frequency of the mode of interest.
10 Experimental Modal Analysis Methods 609

Obviously, the type of equation condensation method that is utilized in a modal

identification algorithm has a significant influence on the results.

6 Summary

Experimental modal analysis methods are probably one of the most misunderstood
aspects of the experimental modal analysis process. Since most modal parameter
estimation methods are mathematically intimidating, many users do not fully
understand the ramifications of the decisions made during the measurement stages
as well as later in the modal parameter estimation process. Ideally, by consolidating
the conceptual approach and unifying the theoretical development of modal iden-
tification algorithms, increased understanding, with respect to general advantages
and disadvantages of different algorithms, can be achieved. This sort of overview
of modal parameter estimation methods can be used simply as a guide toward
further study and understanding of the details of the individual modal identification
algorithms.
The two-stage modal parameter estimation process can be updated further to
include some of the details discussed in previous sections.

• MPE-1: Utilize measured FRF or IRF data with a matrix coefficient, polynomial
model to find multiple estimates of the modal parameters. Select the best set of
modal frequencies and modal participation vectors.
– Choose polynomial or subspace order (m).
– Load measured data into over-determined linear equation form.
– Solve for unknown matrix coefficients ([αk ]) Least squares (LS) approach.
– Solve for complex valued modal frequencies for (λr or zr ) Companion matrix
approach.
– Iterate over different polynomial orders or subspace orders in the kernel
equation to get multiple estimates of the modal parameters.
– Select one set of complex valued modal frequencies, with associated partici-
pation vectors, from the multiple sets. Manual or automated selection methods
can be used.
• MPE-2: Utilize measured FRF or IRF data, with selected complex valued modal
frequencies and modal participation vectors, to find complex valued modal
vectors and modal scaling from an overdetermined set of linear equations.
– Solve for modal vectors using a weighted least squares solution method
involving the fixed set of modal frequencies and modal participation (weight-
ing) vectors. Solution process often involves the estimation of residues.
– Find modal vectors and modal scaling from Equations 75 through 79.
– Normalize modal vectors as needed.
– Solve for modal scaling associated with normalized modal vectors.

Experimental modal methods will continue to evolve to add information that can
be gleaned from the redundancy in the MIMO data. One obvious addition will be
610 R. J. Allemang and D. L. Brown

the inclusion of more statistical information for the measured modal parameters.
This will help in the quantification of margin and uncertainty (QMU) for the
modal parameters, important in verification and validation evaluations.  Chap. 11,
“Experimental Modal Parameter Evaluation Methods” in this Handbook explains
many of the current methods used to evaluate and validate the modal parameters
estimated in this chapter.
Work will continue on several topics that are not viewed as complete topics,
including situations that cannot be adequately solved today. For example, specific
attention is being given to methodology needed to estimate modal parameters
for heavily damped systems, particularly systems with significant modal density.
Generally systems with high modal density (structures with many coupled panels)
are difficult to both model and test. Work is ongoing on autonomous methods
and methods that begin to evaluate nonlinearities, including both nonlinear signal
processing issues as well as nonlinear structure issues.

References
1. Leuridan J, Brown D, Allemang R (1985) Time domain parameter identification methods for
linear modal analysis: a unifying approach, ASME paper number 85-DET-90
2. Allemang RJ, Brown D (1987) Volume III: modal parameter estimation, USAF report:
AFWAL-TR-87-3069, Experimental modal analysis and dynamic component synthesis, no. 3,
pp 130
3. Allemang R, Brown D, Fladung W (1994) Modal parameter estimation: a unified matrix
polynomial approach. In: Proceedings, International Modal Analysis Conference (IMAC),
pp 501–514
4. Allemang R, Brown D (1996) A unified matrix polynomial approach to modal identification,
pp 379–390, Mar 1996
5. Allemang R, Brown D (1998) A unified matrix polynomial approach to modal identification.
J Sound Vib 211(3):301–322
6. Allemang R, Phillips A (2004) The unified matrix polynomial approach to understanding
modal parameter estimation: an update. In: Proceedings, International Seminar on Modal
Analysis (ISMA)
7. Spitznogle F (1971) Improvements in the complex exponential signal analysis computational
algorithm, Report number U1-829401-5, Representation and analysis of sonar signals, vol 1,
pp 37
8. Brown D, Allemang R, Zimmerman R, Mergeay M (1979) Parameter estimation techniques
for modal analysis. SAE Trans 88:828–846. SAE paper number 790221
9. Vold H, Rocklin T (1982) The numerical implementation of a multi-input modal estimation
algorithm for mini-computers. In: Proceedings, International Modal Analysis Conference
(IMAC), pp 542–548
10. Vold H, Kundrat J, Rocklin T, Russell R (1982) A multi-input modal estimation algorithm for
mini-computers. SAE Trans 91(1):815–821
11. Ibrahim S, Mikulcik E (1977) A method for the direct identification of vibration parameters
from the free response. Shock Vib Bull 47:183–198
12. Pappa R (1982) Some statistical performance characteristics of the ITD modal identification
algorithm, AIAA paper number 82-0768, p 19
13. Fukuzono K (1986) Investigation of multiple-reference Ibrahim time domain modal parameter
estimation technique, M. S. thesis, p 220
10 Experimental Modal Analysis Methods 611

14. Juang J, Pappa R (1985) An eigensystem realization algorithm for modal parameter identifica-
tion and model reduction. AIAA J Guid Control Dyn 8(4):620–627
15. Juang J (1987) Mathematical correlation of modal parameter identification methods via system
realization theory. J Anal Exp Modal Anal 2(1):1–18
16. Longman RW, Juang J-N (1989) Recursive form of the eigensystem realization algorithm for
system identification. AIAA J Guid Control Dyn 12(5):647–652
17. Zhang L, Kanda H, Brown D, Allemang R (1985) A polyreference frequency domain method
for modal parameter identification, ASME paper No. 85-DET-106, vol 8, pp 8+
18. Lembregts F, Leuridan J, Zhang L, Kanda H (1986) Multiple input modal analysis of frequency
response functions based on direct parameter identification. In: Proceedings, International
Modal Analysis Conference (IMAC)
19. Lembregts F (1988) Frequency domain identification techniques for experimental multiple
input modal analysis, Doctoral dissertation, Katholieke University of Leuven, Belgium,
vol 213
20. Lembregts F, Leuridan J, Van Brussel H (1989) Frequency domain direct parameter identifi-
cation for modal analysis: state space formulation. Mech Syst Signal Process (MSSP) 4(1):
65–76
21. Coppolino R (1981) A simultaneous frequency domain technique for estimation of modal
parameters from measured data, SAE paper No. 811046, vol 12
22. Craig R, Kurdila A, Kim HM (1990) State-space formulation of multi-shaker modal analysis.
J Anal Exp Modal Anal 5(3):169–183
23. Leuridan J (1984) Some direct parameter model identification methods applicable for multiple
input modal analysis, Doctoral dissertation, vol 384, pp 384+
24. Natke H (1988) Updating computational models in the frequency domain based on measured
data: a survey. Probab Eng Mech 3(1):28–35
25. Richardson M, Formenti D (1982) Parameter estimation from frequency response measure-
ments using rational fraction polynomials. In: Proceedings, International Modal Analysis
Conference (IMAC), pp 167–182
26. Shih C, Tsuei Y, Allemang R, Brown D (1988) A frequency domain global parameter
estimation method for multiple reference frequency response measurements. Mech Syst Signal
Process (MSSP) 2(4):349–365
27. Shih C (1989) Investigation of numerical conditioning in the frequency domain modal
parameter estimation methods, Doctoral dissertation, vol 127, pp 127+
28. Van der Auweraer H, Leuridan J (1987) Multiple input orthogonal polynomial parameter
estimation. Mech Syst Signal Process (MSSP) 1(3):259–272
29. Van der Auweraer H, Snoeys R, Leuridan JM (1986) A global frequency domain modal
parameter estimation technique for mini-computers. ASME J Vib Acoust Stress Reliab Des 10
30. Vold H (1986) Orthogonal polynomials in the polyreference method. In: Proceedings, Interna-
tional Seminar on Modal Analysis (ISMA)
31. Vold H, Shih CY (1988) Numerical sensitivity of the characteristic polynomial. In: Proceed-
ings, International Seminar on Modal Analysis (ISMA)
32. Vold H (1990) Numerically robust frequency domain modal parameter estimation. Sound Vib
Mag (SVM) 24(1):38–40
33. Vold H (1990) Statistics of the characteristic polynomial in modal analysis. In: Proceedings,
International Seminar on Modal Analysis (ISMA), pp 53–57
34. Vold H, Napolitano K, Hensley D, Richardson M (2008) Aliasing in modal parameter
estimation, an historical look and new innovations. In: Proceedings, International Modal
Analysis Conference (IMAC), vol 16, pp 12–17
35. Fladung W, Vold H (2016) An improved implementation of the orthogonal polynomial
modal parameter estimation algorithm using the orthogonal complement. In: Proceedings,
International Modal Analysis Conference (IMAC)
36. Fladung W, Vold H (2016) An orthogonal view of the polyreference least-squares complex fre-
quency modal parameter estimation algorithm. In: Proceedings, International Modal Analysis
Conference (IMAC)
612 R. J. Allemang and D. L. Brown

37. Van der Auweraer H, Guillaume P, Verboven P, Vanlanduit S (2001) Application of a fast-
stabilizing frequency domain parameter estimation method. ASME J Dyn Syst Meas Control
123(4):651–658
38. Guillaume P, Verboven P, Vanlanduit S, Van der Auweraer H, Peeters B (2003) A polyreference
implementation of the least-squares complex frequency domain estimator. In: Proceedings,
International Modal Analysis Conference (IMAC), p 12
39. Verboven P, Guillaume P, Cauberghe B, Parloo E, Vanlanduit S (2003) Stabilization charts
and uncertainty bounds for frequency domain linear least squares estimators. In: Proceedings,
International Modal Analysis Conference (IMAC), pp 1–10
40. Verboven P, Cauberghe B, Vanlanduit S, Parloo E, Guillaume P (2004) The secret behind clear
stabilization diagrams: the influence of the parameter constraint on the stability of the poles. In:
Proceedings, Society of Experimental Mechanics (SEM) Annual Conference, vol 17, pp 1–17
41. Verboven P (2002) Frequency domain system identification for modal analysis, Doctoral
dissertation
42. Cauberghe B (2004) Application of frequency domain system identification for experimental
and operational modal analysis, Doctoral dissertation, p 259
43. Shih C, Tsuei Y, Allemang R, Brown D (1988) Complex mode indication function and
its application to spatial domain parameter estimation. Mech Syst Signal Process (MSSP)
2(4):367–377
44. Frazer R, Duncan W, Collar A (1938) Elementary matrices (1965 Reprint). Cambridge
University Press, Cambridge
45. Tse F, Morse I, Hinkle R (1978) Mechanical vibrations, theory and applications, 2nd edn. Allyn
and Bacon, Inc., Boston
46. SDRC (unknown) (1974) Structural dynamics theory, structural dynamics research corporation
47. Ljung L (1987) System identification: theory for the user. Prentice-Hall, Inc., Englewood Cliffs
48. van Overschee P, De Moor B (1996) Subspace identification of linear systems: theory,
implementation, applications. Springer Publishing, Cham
49. McKelvey T, Akcay H, Ljung L (1996) Subspace-based multivariable system identification
from frequency response data. IEEE Trans Autom Control 41(7):960–979
50. Vakilzadeh M, Yaghoubi V, McKelvey T, Abrahamsson T, Ljung L (2015) Experimental design
for improved frequency domain subspace system identification of continuous-time systems.
IFAC PapersOnLine 48(28):886–891
51. Allemang R, Phillips A (2014) Spatial information in autonomous modal parameter estimation.
Shock Vib J 2014:1–18. Special ICEDYN Issue
52. Allemang R (2003) The modal assurance criterion (MAC): twenty years of use and abuse.
Sound Vib Mag (SVM) 37(8):14–23
53. Cauberghe B, Guillaume P, Verboven P, Parloo E, Vanlanduit S (2004) A poly-reference
implementation of the maximum likelihood complex frequency-domain estimator and some
industrial applications. In: Proceedings, International Modal Analysis Conference (IMAC),
p 12
54. De Troyer T, Guillaume P, Pintelon R, Vanlanduit S (2009) Fast calculation of confidence
intervals on parameter estimates of least-squares frequency-domain estimators. Mech Syst
Signal Process (MSSP) 23:261–273
55. El-Kafafy M, Accardo G, Peeters B, Janssens K, De Troyer T, Guillaume P (2015) A fast
maximum likelihood-based estimation of a modal model. In: Proceedings, International Modal
Analysis Conference (IMAC), pp 133–156
56. El-Kafafy M, Peeters B, Guillaume P, De Troyer T (2016) Constrained maximum likelihood
modal parameter identification applied to structural dynamics. Mech Syst Signal Process
(MSSP) 72–73:567–589
57. Kennedy C, Pancu C (1947) Use of vectors in vibration measurement and analysis. J Aeronaut
Sci 14:603–625
58. Phillips A, Allemang R (1996) Single degree-of-freedom modal parameter estimation algo-
rithms. In: Proceedings, International Modal Analysis Conference (IMAC)
59. Fladung W (1994) Multiple reference impact testing, MS thesis, University of Cincinnati
10 Experimental Modal Analysis Methods 613

60. Fladung W, Phillips A, Allemang R (2003) Application of a generalized residual model to

frequency domain parameter estimation. J Sound Vib (JSV) 262:677–705
61. Allemang R, Phillips A, Brown D (2011) Combined state order and model order formulations
in the unified matrix polynomial method (UMPA). In: Proceedings, International Modal
Analysis Conference (IMAC)
62. Dippery KD, Phillips AW, Allemang RJ (1996) Spectral decimation in low-order frequency
domain modal parameter estimation. J Modal Anal 11:145–154
63. Phillips A, Allemang R (2004) The impact of measurement condensation and modal partici-
pation vector normalization on the estimation of modal vectors and scaling. In: Proceedings,
International Modal Analysis Conference (IMAC), p 12
64. Dippery K, Phillips A, Allemang R (1996) An SVD condensation of the spatial domain in
modal parameter estimation. Int J Exp Anal Modal Anal 11
65. Jolliffe IT (1986) Principal component analysis. Springer, New York and Inc
66. Strang G (2006) Linear algebra and its applications, 4th edn. Cengage Learning
67. Pickrel CR (1996) Numerical assessment of test data using SVD. In: Proceedings, International
Modal Analysis Conference (IMAC), Oct 1996
Experimental Modal Parameter Evaluation
Methods 11
R. J. Allemang and A. W. Phillips

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
2 Background: Modal Parameter Estimation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
2.1 Assumptions, Definitions, and Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619
3 Modal Frequency Evaluation/Validation Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
3.1 Model Order Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
3.2 Auto Moment Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
3.3 Mode Indication Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
3.4 Consistency Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
3.5 Pole Surface Consistency Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
3.6 Modal Parameter Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
4 Modal Vector Evaluation/Validation Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645
4.1 Modal Vector Conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645
4.2 Modal Vector Validation: eFRF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650
4.3 Weighted Modal Vector Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
4.4 Weighted Pseudo Orthogonality of Modal Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 657
4.5 Modal Vector Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
5 Autonomous Modal Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 672
5.1 Current Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
5.2 Common Statistical Subspace Autonomous Mode Identification (CSSAMI) . . . . . 674
6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675

Abstract

Modern experimental modal analysis (EMA) methods provide a number of

modal parameter solutions based upon different models, different model orders

R. J. Allemang () · A. W. Phillips

Structural Dynamics Research Laboratory, Mechanical and Materials Engineering, University of
Cincinnati, Cincinnati, OH, USA
e-mail: [email protected]; [email protected]

© The Society for Experimental Mechanics 2022 615

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_12
616 R. J. Allemang and A. W. Phillips

and different numerical processing of the redundant data and/or results. Eval-
uation of the modal parameter solutions provides a way of obtaining a single
unique set of modal parameters that best represents the measured experimental
data. The early portion of this chapter is a review of some of the experimental
modal analysis (EMA) methods covered in detail in Chap. 10, “Experimental
Modal Analysis Methods” in this handbook. This is followed by presenting a
number of numerical tools that are used in connection with the EMA methods to
evaluate and validate the number of modal parameters that can be estimated from
a multiple input, multiple output (MIMO) set of measured data. Some tools like
complex and multivariate mode indication functions (CMIF and MvMIF) can be
used to determine the model order and/or number of modal frequencies that can
be estimated from the experimental data. These tools can be applied independent
of the EMA method that is used and are particularly useful when close or
repeated modal frequencies are present in the experimental data. Additionally,
various consistency diagrams, pole surface plots and modal parameter clustering
methods are defined that become part of, and enhance, the EMA method used to
estimate the modal parameters. Finally, the last portion of this chapter overviews
methods that are primarily post processing tools to evaluate and validate the
modal parameters that have been estimated. Methods include techniques for
normalizing, conditioning and presenting the modal vectors, like the modal
vector complexity plot (MVCP) along with techniques for using the estimated
modal vectors to estimate other functions like the enhanced frequency response
function (eFRF) which can be used to validate the physical validity of the
estimated modal vectors. Orthogonality of modal vectors along with consistency
of modal vectors, as measured by the modal assurance criterion (MAC), also
falls into this category of evaluation and validation tools that are applied after the
modal parameters have been estimated. The chapter finishes with a brief example
of how several of the evaluation and validation tools can be combined into an
autonomous modal parameter estimation method.

Keywords

Experimental modal analysis · Modal parameter evaluation · Modal parameter

validation · Modal assurance criterion · Enhanced frequency response function
(eFRF) · Complex mode indication function (CMIF) · Multivariate mode
indication function (MvMIF) · Consistency diagram · Pole surface diagram ·
Autonomous modal parameter estimation

Nomenclature

Ni = Number of inputs
No = Number of outputs
NS = Short dimension (min(Ni , No ))
NL = Long dimension (max(Ni , No ))
Nf = Number of spectral lines (frequencies)
11 Experimental Modal Parameter Evaluation Methods 617

Ne = Number of effective modal frequencies

N = Number of modal frequencies
Fmax = Maximum frequency (Hz)
ωi = Frequency (rad/sec)
ωmax = Maximum frequency (rad/sec)
Δf = Frequency resolution (Hz)
λr = Complex modal frequency
T = Observation period (sec)
si = Generalized frequency variable
m = Model order for denominator polynomial
n = Model order for numerator polynomial
Apqr = Residue, output DOF p, input DOF q, mode r
RIpq = Residual inertia, output DOF p, input DOF q
RFpq = Residual flexibility, output DOF p, input DOF q
[C] = Companion matrix
[α] = Denominator polynomial matrix coefficient
[β] = Numerator polynomial matrix coefficient
[I ] = Identity matrix
[H (ωi )] = Frequency response function matrix. (No × Ni )
[T ] = Transformation matrix
[U ] = Left singular vector matrix
[Σ] = Singular value matrix (diagonal)
[Λ] = Eigenvalue matrix (diagonal)
[V ] = Right singular vector, or eigenvector, matrix
MP E-1 = First stage of modal parameter estimation
MP E-2 = Second stage of modal parameter estimation

1 Introduction

Experimental modal analysis is the process of determining the modal parameters of

a structural system (frequencies, damping factors, modal vectors, and modal scaling
values) from experimental input-output data, normally frequency response functions
(FRFs) or impulse response functions (IRFs). This is in contrast to analytical modal
analysis where the modal parameters are found from a theoretical (continuous,
or discrete) model of the structural system. This could be a closed-form solution
for simple structures or, more commonly, a discrete model like a finite element
model for structural systems with more complex geometry. Analytical modal
analysis generally does not include damping factors as part of the solution. Finally,
operational modal analysis (OMA), sometimes called response-only modal analysis,
attempts to determine modal parameters from data taken without measured inputs.
The inputs come from natural excitation (wind, waves, traffic, etc.) or from the
operational inputs provided to a structural system. Operational modal analysis gives
a subset of modal parameters that depends upon whether the unmeasured inputs
excite the modal parameters and may mix modal parameters with forced vibration
618 R. J. Allemang and A. W. Phillips

information. Since operation modal analysis does not measure the inputs, modal
scaling cannot be determined without additional testing. Operational modal analysis
is the subject of the chapter “Operational Modal Analysis” in this handbook. The
following discussion is concerned with experimental modal analysis. Analytical and
operational modal analysis is covered in other related material.
One important continuing focus of experimental modal analysis is the presenta-
tion of modal parameter estimation algorithms in a single, consistent mathematical
formulation with a corresponding set of definitions and unifying concepts [1, 2, 3,
4, 5, 6]. In particular, a matrix coefficient polynomial approach can be used to unify
the presentation with respect to current algorithms such as the least squares complex
exponential (LSCE), polyreference time domain (PTD), polyreference least squares
complex frequency (LSCF), Ibrahim time domain (ITD), Eigensystem realization
algorithm (ERA), rational fraction polynomial (RFP), orthogonal polynomials
(OP), polyreference frequency domain (PFD), and the complex mode indication
function (CMIF) methods. Using this unified matrix polynomial approach (UMPA)
encourages a discussion of the similarities and differences of the commonly
used methods, as well as a discussion of the numerical characteristics. Some
of the different numerical methods that are used in different methods are the
least squares (LS), total least squares (TLS), double least squares (DLS), and
singular value decomposition (SVD) methods (to take advantage of redundant
measurement data) and the eigenvalue and singular value decomposition transfor-
mation methods (to reduce the effective size of the resulting eigenvalue-eigenvector
problem).
The unified matrix polynomial approach (UMPA) is an attempt to place most
commonly used experimental modal parameter estimation algorithms within a
single educational framework. The goal of the UMPA presentation is to highlight the
similarity between the different algorithms rather than differences. This approach
does not attempt to explain the detailed development of the authors who originated
each method but attempts to present a common framework so that different
algorithms can be easily compared and contrasted.

2 Background: Modal Parameter Estimation Methods

Modal parameter estimation is a special case of system identification where the a

priori model of the system is known to be in the form of modal parameters. Over
the past 40 years, a number of algorithms have been developed to estimate modal
parameters from measured multiple input, multiple output (MIMO) frequency
response function (FRF) or impulse response function (IRF) data. While most
of these individual algorithms, summarized in Table 1, are well understood, the
comparison of one algorithm to another has become one of the thrusts of current
research in this area. Comparison of the different algorithms is possible when the
algorithms are reformulated using a common mathematical structure.
This reformulation attempts to characterize different classes of modal parameter
estimation techniques in terms of the structure of the underlying matrix polynomials
11 Experimental Modal Parameter Evaluation Methods 619

Table 1 Acronyms – Modal parameter estimation algorithms

experimental modal
CEA Complex exponential algorithm [7, 8]
parameter estimation
Algorithms LSCE Least squares complex exponential [8]
PTD Polyreference time domain [9, 10]
ITD Ibrahim time domain [11, 12]
MRITD Multiple Reference Ibrahim Time Domain [13]
ERA Eigensystem realization algorithm [14, 15, 16]
PFD Polyreference frequency domain
[17, 18, 19, 20]
FDPI Frequency Domain Direct Parameter
Identification [19, 20]
SFD Simultaneous frequency domain [21]
MRFD Multi-reference frequency domain [22]
RFP Rational fraction polynomial [23, 24, 25]
OP Orthogonal polynomial
[25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36]
PLSCF Polyreference least squares complex frequency
[37, 38, 39, 40, 41, 42]
CMIF Complex mode indication function [43]

rather than the physically based models used historically. Since the modal parameter
estimation process involves a greatly overdetermined problem (more data than
independent equations), this reformulation is helpful in understanding the different
numerical characteristics of each algorithm and, therefore, the slightly different
estimates of modal parameters that each algorithm yields. As a part of this reformu-
lation of the algorithms, the development of a conceptual understanding of modal
parameter estimation technology has emerged. This understanding involves the
ability to conceptualize the measured data in terms of the concept of characteristic
space, the data domain (time, frequency, spatial), the dimension of the measured
data, the evaluation of the order of the problem, the condensation of the data, and
a common parameter estimation theory that can serve as the basis for developing
any of the algorithms in use today. The following sections review these concepts as
applied to the current modal parameter estimation methodology.

2.1 Assumptions, Definitions, and Concepts

A number of assumptions, basic definitions, and general concepts are essential to

understanding the experimental modal analysis process. Since there is considerable
history in the development of modal parameter estimation methodology, it is helpful
to provide a structure that provides a common basis for all modal parameter
estimation algorithms. Some of these definitions and concepts are briefly introduced
in the following subsections.
620 R. J. Allemang and A. W. Phillips

2.1.1 Assumptions
By its very nature, modal analysis is one way to describe the dynamic characteristics
of a structural system. For modal analysis to be considered, there are several
assumptions involved. Generally, the structural system is assumed to be linear,
time invariant, and reciprocal. Other assumptions can be involved but are not
a requirement. An example would be proportional or Rayleigh damping which
restricts the form of the modal vectors that are found.

2.1.2 Definition: Modal Parameters

While some situations do not require all modal parameters, a complete set of
modal parameters is needed if a complete model of the input-output relationships
is desired. Modal parameters include the complex-valued modal frequencies (λr ),
the associated complex-valued modal vectors (ψr ), and the complex-valued modal
scaling (modal Ar ). Note that the complex-valued modal frequencies are of the
form (λr = σr + j ωr ) where σr is the damping factor and ωr is the damped natural
frequency for the r − th mode. Additionally, most current multiple input, multiple
output (MIMO) algorithms estimate modal participation (weighting) vectors and
(Lr ) and residue vectors (Ar ) as part of the overall process of estimating the
complex-valued modal vectors. Modal participation vectors are a result of MIMO
modal parameter estimation algorithms and relate how well each modal vector is
excited from each of the reference locations included in the measured data. The
combination of the modal participation vector (Lr ), the modal vector (ψr ), and the
modal A (MAr ) for a given mode yields the residue matrix ([A]r ) for that mode.
In general, modal parameters are considered to be global properties of the system.
The concept of global modal parameters simply means that there is only one
answer for each modal parameter and that the modal parameter estimation solution
procedure enforces this constraint. Most of the current modal parameter estimation
algorithms estimate the modal frequencies and damping in a global sense, but very
few estimate the modal vectors in a global sense. This is due to various modal vector
scaling normalization methods which, together with the modal scaling value, give
equivalent results but not unique, mathematical values.

Modal Vector Normalization

While there is a unique answer for the modal frequencies, the complex-valued
modal vectors represent the relative pattern of motion associated with each complex-
valued modal frequency. This relative pattern results from a rank-deficient system of
equations where the modal vector is found via an eigenvalue-eigenvector solution.
The complex-valued modal vectors, together with the modal scaling, represent the
unique characteristic for each modal vector. For this reason, the normalization of the
complex-valued modal vector is important when modal vectors, or the associated
modal scaling values, are compared numerically.
While there are a number of acceptable normalization schemes for modal vectors,
choosing the largest element in each complex-valued modal vector equal to unity is
the most common and most useful in a physical sense. This method of normalization
11 Experimental Modal Parameter Evaluation Methods 621

will be used in all following discussions. Recognizing that an arbitrarily scaled

modal vector is complex valued means that the largest element will in general be
complex valued. When the arbitrarily scaled modal vector is divided (normalized)
by the largest element, this will force all elements of the scaled modal vector to be
bounded by the complex unit circle and, for most cases, to lie along the real axis of
the unit circle.

2.1.3 Definition: Degrees of Freedom (DOFs)

Degrees of freedom (DOFs) refer to the physical location and direction of all of the
potential inputs to, or outputs from, a structural system. For the theoretical problem,
the number of inputs and the number of outputs are always the same and are equal
to the number of modal parameter sets of information (typically designated N ) that
will be found. The general notation that is often used involves inputs in the form of
forces and outputs in the form of displacements, velocities, and/or accelerations. The
generalized concept of inputs and outputs could be used to represent translational or
rotational outputs (or their derivatives) or translational or rotational inputs.
In an experimental sense, the DOFs are where the input and output sensors
are located (physical location and direction). Today that includes response mea-
surements from scanning laser vibrometers and digital image correlation (DIC)
photogrammetry methods. Equation 1 is a representation of an FRF where the output
DOF utilizes the notation p and the input DOF utilizes the notation q. Note that the
X and F notation represents the output and input in a general way and does not imply
displacement or force.
Xp (ω)
Hpq (ω) = (1)
Fq (ω)

• p is the output degree of freedom (physical location and orientation).

• q is the input degree of freedom (physical location and orientation).

For the experimental case, the number of input DOFs (Ni ) and output DOFs
(No ) is not the same and in general cannot be directly linked to the N sets of modal
parameters that will be estimated.

2.1.4 Concept: Experimental Modal Parameter Estimation

Experimental modal parameter estimation involves estimating the modal parameters
of a structural system from measured input-output data. The experimental approach
originally involved methods that are referred to as phase resonance using sinusoidal
excitation (narrowband) and mode by mode tuning with a forcing vector to balance
the damping affect assuming the modal vectors are normal modes. These methods
gave way to phase separation methods that involve a range of frequency information
(broadband) where the effects of several modes are separated using mathematical
models for the experimental data. Most current modal parameter estimation is phase
separation methods, based upon the measured data being the frequency response
622 R. J. Allemang and A. W. Phillips

function or the equivalent impulse response function, typically found by inverse

Fourier transforming the frequency response function.
The current approach is to use numerical techniques to separate the contributions
of individual modes of vibration in measurements such as frequency response
functions. The concept involves estimating the individual single degree of freedom
(SDOF) contributions to the multiple degree of freedom (MDOF) measurement.
Then, the modal parameters are found from the SDOF contributions.


N
[Ar ]N L ×NS
[A∗r ]NL ×NS 
2N
[Ar ]N L ×NS
[H (ωi )]NL ×NS = + = (2)
j ωi − λr j ωi − λ∗r j ωi − λr
r=1 r=1

Equation 2 represents a mathematical problem that, at first observation, is

nonlinear in terms of the unknown modal parameters. Once the modal frequencies
(λr ) are known, the mathematical problem is linear with respect to the remaining
unknown modal parameters ([Ar ]). For this reason, the numerical solution in many
modal parameter estimation (MPE) algorithms frequently involves two linear stages
that utilize least squares (LS) solution methods. Typically, the modal frequencies
and modal participation vectors are found in a first stage (MPE-1), and residues,
modal vectors, and modal scaling are determined in a second stage (MPE-2).
This concept, involving a summation of partial fraction terms, is represented
mathematically in Eq. 2 and graphically in Fig. 1.
While the model stated in Eq. 2 is fundamental to the linear superposition of
individual SDOF contributions, this model is normally limited to being used as the
basis for estimating the residues Apqr once the modal frequencies (λr ) are known.
Based upon the speed and memory of modern personal computers, these two stages
are often executed together giving the appearance of a single stage solution.
Equation 2 can be represented in the time domain in terms of impulse response
functions in a similar summation form as shown in Eq. 3:


N
∗ 
2N
[h(ti )]NL ×NS = [Ar ]NL ×NS eλr ti + [A∗r ]NL ×NS eλr ti = [Ar ]NL ×NS eλr ti (3)
r=1 r=1

2.1.5 Concept: Experimental Modal Parameter Methods

Using the concepts developed in the previous section, the most commonly used
modal identification methods can be summarized as shown in Table 2. The high-
order model is typically used for those cases where the system is under-sampled in
the spatial domain. For example, the limiting case is when only one measurement is
made on the structure. For this case, the left-hand side of the general linear equation
corresponds to a scalar polynomial equation with the order equal to or greater than
the number of desired modal frequencies.
The low-order model is used for those cases where the spatial information is
complete. In other words, the number of physical coordinates (NL ) is greater than
11 Experimental Modal Parameter Evaluation Methods 623

20

−20
Phase (Deg)

−40

−60

−80

−100

−120

−140

−160

−180
0 50 100 150 200 250 300 350 400 450 500
−40

−50

−60
Log Magnitude (dB)

−70

−80

−90

−100
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hertz)

Fig. 1 MDOF – superposition of SDOF (positive frequency poles)

the number of desired (positive) modal frequencies (N ). For this case, the order (m)
of the left-hand side of the general linear equation, Eqs. 5 or 6, is equal to two.
The zero-order model corresponds to cases where the temporal information is
neglected and only the spatial information is used. These methods directly estimate
the eigenvectors as a first step. In general, these methods are programmed to
process data at a single temporal condition or variable. In this case, the method
is essentially equivalent to the single degree of freedom (SDOF) methods which
have been used with frequency response functions. In other words, the zeroth-order
matrix polynomial model compared to the higher-order matrix polynomial models is
similar to the comparison between the SDOF and MDOF methods used historically
in modal parameter estimation.
Modal parameter estimation algorithms are similar in more ways than they
are different. Fundamentally, all algorithms can be developed beginning with a
linear, constant coefficient, symmetric matrix model involving mass, damping, and
stiffness. The common goal in all algorithms, therefore, is the development of
624 R. J. Allemang and A. W. Phillips

Table 2 Modal vector estimation methods: overview

Domain Matrix polynomial order Coefficients
Algorithm Time Freq Zero Low High Scalar Matrix
Complex exponential algorithm • • •
(CEA)
Least squares complex • • •
exponential (LSCE)
Polyreference time domain • • NS × NS
(PTD)
Ibrahim time domain (ITD) • • NL × NL
Multi-reference Ibrahim time • • NL × NL
domain (MRITD)
Eigensystem realization • • NL × NL
algorithm (ERA)
Polyreference frequency domain • • NL × NL
(PFD)
Simultaneous frequency domain • • NL × NL
(SFD)
Multi-reference frequency • • NL × NL
domain (MRFD)
Rational fraction polynomial • • • NS × NS
(RFP)
Orthogonal polynomial (OP) • • • N S × NS
Polyreference least squares • • • NS × NS
complex frequency (PLSCF)
Complex mode indication • • NL × NS
function (CMIF)

a characteristic matrix coefficient equation that describes a linear, time-invariant,

reciprocal mechanical system consistent with this theoretical background. This is
the rationale behind using the unified matrix polynomial approach as the educational
basis for demonstrating this common kernel for all modal parameter estimation
algorithms [1,2,3,4,5,6]. The following sections discuss the similarity of the kernel
equations common to all widely used modal parameter estimation algorithms.

2.1.6 Concept: Data Domain

Modal parameters can be estimated from a variety of different measurements that
exist as discrete data in different data domains (time and/or frequency). These
measurements can include free decays, forced responses, power spectra, covariance,
frequency response functions (FRFs), or impulse response functions (IRFs). These
measurements can be processed one at a time or in partial or complete sets simul-
taneously. The measurements can be generated with no measured inputs, a single
measured input, or multiple measured inputs. The data can be measured individually
or simultaneously. There is a tremendous variation in the types of measurements
and in the types of constraints that can be placed upon the testing procedures used
11 Experimental Modal Parameter Evaluation Methods 625

to acquire this data. For most measurement situations, FRFs are utilized in the
frequency domain, and IRFs are utilized in the time domain. When IRFs are utilized,
they are generally formed from the inverse Fourier transform of the measured FRFs.
For the purpose of the following discussions, force-displacement data is considered
the basis for the equations recognizing that force-displacement data can be numeri-
cally synthesized from measured force-acceleration data or force-velocity data.

2.1.7 Concept: Characteristic Space

From a conceptual viewpoint, the measurement space of a modal identification
problem can be visualized as occupying a volume with the coordinate axes defined
in terms of the three sets of characteristics. Two axes of the conceptual volume
correspond to spatial information and the third axis to temporal information. The
spatial axes are in terms of the input and output degrees of freedom (DOF) of the
system. The temporal axis is either time or frequency depending upon the domain
of the measurements. These three axes define a 3D volume which is referred to as
the characteristic space.
This space or volume represents all possible measurement data. This conceptual
representation is very useful in understanding what data subspace has been mea-
sured. Also, this conceptual representation is very useful in recognizing how the
data is organized and utilized with respect to different modal parameter estimation
algorithms (3D volume to stacked 2D matrices). Information parallel to one axis
consists of a superposition of the characteristics defined by that axis. The other two
characteristics determine the scaling of each term in the superposition.
Any structural testing procedure measures a subspace of the total possible data
available. Modal parameter estimation algorithms may then use all of this subspace
or may choose to further limit the data to a more restrictive subspace via sieving
and/or filtering. It is theoretically possible to estimate the characteristics of the total
space by measuring any subspace which samples all three characteristics. Measure-
ment data spaces involving many planes of measured data are the best possible
modal identification situations since the data subspace includes contributions from
temporal and spatial characteristics. The particular subspace which is measured
and the weighting of the data within the subspace in an algorithm are the main
differences between the various modal identification procedures which have been
developed.
It should be obvious that the data which defines the subspace needs to be acquired
in a consistent measurement process in order for the algorithms to estimate accurate
modal parameters. This fact has triggered the need to measure all of the data
simultaneously and has led to recent advancements in data acquisition, digital signal
processing, and instrumentation designed to facilitate this measurement problem.
While it is not always obvious, most modal parameter estimation methods
assume that the measured data includes one or more pairs of matched input and
output DOFs in what is commonly referred to as driving point FRFs.
626 R. J. Allemang and A. W. Phillips

2.1.8 Concept: Data Dimensionality

In the following discussion, the use of notation involving the number of input
degrees of freedom (DOFs), (Ni ), and the number of output DOFs, (No ), is replaced
by an alternate notation. Since the FRF matrix is always assumed to be reciprocal,
it is more important to note the smaller and larger of the number of DOFs with
respect to the inputs and outputs when the data is utilized by modern MIMO modal
parameter estimation algorithms. For this reason, the dimension that is larger is
referred to as the long dimension, (NL ), and the dimension that is smaller is referred
to as the short dimension, (NS ). The short dimension is often referred to as the
number of references, and the long dimension is often referred to as the number of
responses. The dimension of the third axis (temporal axis) is either the number of
frequencies (Nf ) or the number of times (Nt ).

2.1.9 Concept: Generalized Frequency

In the following development, all frequency domain models will be presented
as a function of the generalized frequency variable s. This variable is a general
complex-valued variable that is most often thought of as s = j ω, representing the
independent frequency variable associated with the FRF. There are several other
scaled or mapped versions of the generalized frequency variable s which are useful
for minimizing the numerical conditioning associated with different frequency
domain MPE algorithms.

Xp (ωi ) βn (si )n + βn−1 (si )n−1 + · · · + β1 (si )1 + β0 (si )0

Hpq (ωi ) = = (4)
Fq (ωi ) αm (si )m + αm−1 (si )m−1 + · · · + α1 (si )1 + α0 (si )0

One example is a simple scaled version of s where s = ωjmax ω

. When the
generalized frequency variable is scaled or mapped in this fashion, any frequency
estimate that results must be corrected accordingly. This concept will be explained
further in the section on high-order, frequency domain algorithms.

2.1.10 Concept: Kernel Equations

Each experimental modal parameter estimation method has a basic equation that
is repeated for different model orders (m) powers or subspaces of the independent
variable (t, s, ω, z) to get additional sets of solutions. The base equation for each
method is referred to as the kernel equation. Consistency between the different
sets of solutions is one common method of determining the most likely modal
parameters in the data.

Frequency Domain
m 
n
(si ) [αk ] [H (ωi )] =
k
(si )k [βk ] [I ] (5)
k=0 k=0

Additional equations can be developed by repeating Eq. 5 at many frequencies

(ωi ) until all data or a sufficient overdetermination factor is achieved. Note that
11 Experimental Modal Parameter Evaluation Methods 627

both positive and negative frequencies are required in order to accurately estimate
conjugate modal frequencies.
In terms of sampled data, the frequency domain matrix polynomial coefficients
result from a set of linear equations (repeated application of Eq. 5) where each
equation is formulated by choosing a different frequency from the FRF data. From a
numerical perspective, Eq. 5 is generally not well-formed, and the condition number
associated with this system of equations will be extreme when the model order (m)
exceeds five or six. This issue will require special consideration in order to obtain
reasonable answers for the modal frequencies. This has been discussed further in
Section 5.2.3 of the  Chap. 10, “Experimental Modal Analysis Methods” in this
handbook. This approach involves mapping the generalized frequency si to the z
independent variable zi as noted in the two following equations:

si = zi = ej ∗π ∗(ωi /ωmax ) = ej ∗ωi ∗Δt (6)

sim = zim = ej ∗π ∗m∗(ωi /ωmax ) (7)

Once the alpha [α] and beta [β] coefficients have been found, the modal
frequencies can be found from the roots of the alpha [α] matrix coefficient
polynomial.


m
[αk ] s k = 0 (8)
k=0

The roots of this matrix coefficient polynomial are in terms of the generalized
frequency variable s and will be the complex modal frequencies (λr ) directly.

Time Domain
If the discussion is limited to the use of free decay or impulse response function
data, the previous time domain equations can be simplified by noting that the forcing
function can be assumed to be zero for all time greater than zero. If this is the case,
the [βk ] coefficients can be eliminated from the equations.


m
[αk ] [h (ti+k )] = 0 (9)
k=0

Additional equations can be developed by repeating Eq. 9 at different time shifts

(initial times ti ) into the data until all data or a sufficient overdetermination factor is
achieved. Note that at least one time shift is required in order to accurately estimate
conjugate modal frequencies.
In terms of sampled data, the time domain matrix polynomial coefficients result
from a set of linear equations (repeated application of Eq. 6) where each equation is
formulated by choosing various distinct initial times. From a numerical perspective,
628 R. J. Allemang and A. W. Phillips

Eq. 6 is generally well-formed, and the condition number associated with this
system of equations will not be extreme even for high order (m).
In contrast, the frequency domain matrix polynomial developed in Eqs. 3
through 5 results from a set of linear equations (repeated application of Eq. 4)
where each equation is formulated at one of the frequencies of the measured FRF
data. This distinction is important to note since the roots of the matrix characteristic
equation formulated in the time domain are in a mapped complex z domain (zr ),
which is similar but not identical to the z-domain familiar to control theory. Once
the alpha [α] coefficients have been found, the roots of a polynomial in the complex
z domain can be found since all of the measured time domain data utilize the same
Δt spacing.


m
[αk ]zk = 0 (10)
k=0

Equation 10 is developed from the fixed time spacing (Δt) and derivative
relationships associated with discrete time data. It is important to note that this
development is theoretically exact and no approximations are involved.

2.1.11 Concept: Overdetermined Linear Models

Most current modal parameter estimation utilizes linear models, sometimes in
several successive solution steps, that have more equations than unknowns in each
solution step. This will be true for simple SDOF models as well as complicated
MIMO MDOF models. The number of equations is larger than the unknowns since
an equation can be formed at each measured frequency or time while the number
of unknowns is limited to a function of the number of modal parameters N. The
overdetermination factor is the ratio of the number of equations to the number of
unknowns. The overdetermination factor can easily be a number greater than 2–5.
Overdetermined sets of linear equations are most often solved using least squares
(LS) techniques, particularly when the noise on the data is expected to be random.
The least squares solution will yield the best solution in the presence of random
noise. With respect to the estimation of frequency response functions (FRFs) as the
measured data for most modal parameter estimation, this means that bias errors on
the FRF data is much more concerning than the random errors.

2.1.12 Concept: General (Two-Stage) Solution Procedure

Most modern modal parameter estimation (MPE) algorithms are implemented in
two stages that each involves the solution of overdetermined linear equations. The
first stage, referred in this text as MPE-1, is where the modal frequencies (λr ) are
estimated along with associated modal weighting vectors (Lr ). The second stage,
referred in this text as MPE-2, is where the modal vectors (ψr ) and modal scaling
are estimated. These two stages can generally be briefly summarized as follows:
11 Experimental Modal Parameter Evaluation Methods 629

• MPE-1: Utilize measured FRF or IRF data with a matrix coefficient, polynomial
model to find multiple estimates of the modal parameters. Select the best set of
modal frequencies and modal participation vectors. This stage is often performed
in two fundamental steps: (1) Solve a set of overdetermined linear equations, and
(2) solve for the roots of a polynomial.
– Choose polynomial or subspace order (m)
– Solve for unknown matrix coefficients (αk )
* Add sufficient block kernel equations (for different starting times or
frequencies) until there are more block kernel equations than there are
block unknowns ([αk ]).
* If using a frequency domain method, utilize one or more of the generalized
frequency approaches to improve the numerical conditioning.
* Solve for matrix coefficients ([αk ]) in a least squares (LS) sense.
* Solve for complex-valued modal frequencies for (λr or zr ) utilizing the
eigenvalues from an eigenvalue-eigenvector solution method to solve the
matrix coefficient polynomial (companion matrix approach).
* If time domain methods have been used, correct from zr to λr
* If generalized frequency methods have been used, correct for the frequency
conversion.
* Using the eigenvectors from the above eigenvalue-eigenvector solution as
modal participation (weighting) vectors Lqr . Normalize as needed.
– Iterate over different polynomial orders or subspace orders in the kernel
equation to get multiple estimates of the modal parameters.
– Select one set of complex-valued modal frequencies, with associated partici-
pation vectors, from the multiple sets. Manual or automated selection methods
can be used.
• MPE-2: Utilize measured FRF or IRF data, with selected complex-valued modal
frequencies and modal participation vectors, to find complex-valued modal
vectors from an overdetermined set of linear equations.
– Solve for modal vectors or residues using a weighted least squares solution
method involving the fixed set of modal frequencies and modal participation
(weighting) vectors.
– Normalize modal vectors as needed.
– Solve for modal scaling associated with normalized modal vectors and
previous modal participation vectors.

2.1.13 Concept: Equation Normalization

Many of the overdetermined sets of linear equations involve a null space problem in
which the right-hand side (RHS) of the equation is zero or null. In these cases, the
unknown coefficients of the equation are non-unique but can be found by choosing
one of the coefficients, typically, equal to the identity matrix. The numerical solution
of the overdetermined sets of linear equations will yield different answers depending
on which coefficient is chosen. While theoretical data would yield only one answer,
630 R. J. Allemang and A. W. Phillips

experimental data will have random and bias noise, and this noise is what yields
somewhat different answers. Historically, two solutions have been used in these
cases, setting the lead coefficient and the last coefficient to the identity matrix to
give two answers that bracket the solution. This is true for both frequency domain
methods and time domain methods. This choice of choosing a coefficient to be the
identity is referred to as equation normalization.

m 
  n 
 
[αk ](si )k [H (ωi )] = [βk ](si )k [I ] (11)
k=0 k=0

Equation 11 is an example of such a null space (rank deficient) equation. The

unknowns in the equation are the [α] and [β] coefficients. Since every term has an
unknown coefficient matrix, the equation can be pre-multiplied by the inverse of any
one of the [α] coefficients to derive a solvable (full rank) base equation.

Lowest [α0 ] Normalization [α0 ] = [I ]

m 
  n 
   
[αk ](si )k [H (ωi )] − [βk ](si )k [I ] = − [α0 ](si )0 [H (ωi )] (12)
k=1 k=0

Highest [αm ] Normalization [αm ] = [I ]


m−1  n 
   
[αk ](si )k [H (ωi )] − [βk ](si )k [I ] = − [αm ](si )m [H (ωi )] (13)
k=0 k=0

It is important to note that this equation normalization must occur in the initial
step of choosing the base equation. Then the unknown coefficients are solved in
an overdetermined set of these linear equations. It is not sufficient to manipulate
the final solution of polynomial coefficients into a different normalized form. The
difference that equation normalization provides is determined by the least squares
solution for the coefficients [39, 41].

2.1.14 Concept: Modal Vectors, Modal Scaling, Residues

The modal vectors are most often determined by normalizing the residues. There are
a number of normalization methods that can be used. Once the modal vectors are
determined, the modal scaling can be determined. The modal vector normalization
is up to the user but most commonly is chosen to give modal vectors that are
dominantly real valued (close to normal modes). The choice of normalization may
also consider how the comparable modal vectors from analytical methods have been
normalized. For this reason, it is common to choose the modal vector normalization
such that the realities of experimental estimation of the modal vectors (noting that
the residues will not be perfectly imaginary valued) are considered. One example
11 Experimental Modal Parameter Evaluation Methods 631

of a useful normalization is to choose to make the largest modal coefficient in each

modal vector unity.

 
Apqr NL ×1
{ψr }NL ×1 = (14)
max(Apqr )

In the above equation, the residue vector is divided by the residue with the
largest complex magnitude, mode by mode. This yields a modal vector that
will generally be dominantly real valued and limited in magnitude to plus/minus
unity.
The modal A scaling term (MAr ) can then be found to provide the absolute
scaling associated with this modal vector normalization/scaling. Note that if the
normalization/scaling of this modal vector is the same as that used for an analytical
solution, then the modal A scaling terms are comparable.

ψpr ψqr ψpr ψqr

Apqr = MAr = (15)
MAr Apqr

If the modal vector is completely real valued, the modal mass scaling term (Mr )
can be estimated.

ψpr ψqr ψpr ψqr

Apqr = Mr = (16)
j 2ωr Mr j 2ωr Apqr

Note that if the normalization/scaling of this modal vector is the same as

that used for an analytical solution, then the modal mass scaling terms are
comparable.
Finally, from the above equations, modal mass can be estimated from modal A
as long as the modal vector is a normal (real-valued) mode.

MAr
Mr = (17)
j 2ωr

If the modal vector is not a real-valued normal mode, the modal vector can
be normalized to an equivalent real-valued normal mode, and Eq. 17 can be used.
Alternatively, an effective modal mass can be estimated from Eq. 18 as long as the
modal vector is dominantly real valued.

MAr
Mr  =   (18)
j 2ωr
632 R. J. Allemang and A. W. Phillips

3 Modal Frequency Evaluation/Validation Tools

3.1 Model Order Relationships

From a theoretical consideration, the number of characteristic values (number of

modal frequencies, number of roots, number of poles, etc.) that can be determined
depends upon the size of the matrix coefficients involved in the model and the
order of the polynomial terms in the model. The characteristic matrix polynomial
equation, Eq. 5 or Eq. 9, has a model order of m, and the number of modal
frequencies or roots that will be found from this characteristic matrix polynomial
equation will be m times the size of the coefficient matrices [α].
For a given algorithm, the size of the matrix coefficients is normally fixed;
therefore, determining the model order is directly linked to estimating N , the
number of modal frequencies that are of interest in the measured data. As has
always been the case, an estimate for the minimum number of modal frequencies
can be easily found by counting the number of peaks in the frequency response
function in the frequency band of analysis. This is a minimum estimate of N since
the frequency response function measurement may be at a node of one or more
modes of the system, repeated roots may exist, and/or the frequency resolution of
the measurement may be too coarse to observe modes that are closely spaced in
frequency. Several measurements can be observed, and a tabulation of peaks existing
in any or all measurements can be used as a more accurate minimum estimate of N .

3.2 Auto Moment Functions

A more automated procedure for including the peaks that are present in several
frequency response functions is to observe the summation of frequency response
function power (Fig. 2). This function represents the auto power or auto moment of
the frequency response functions summed over a number of response measurements
and is normally formulated as follows for the p-th reference:


NL
Hpower−p (ω) = Hpq (ω)Hpq ∗ (ω) (19)
q=1

Note in the above figure that the mode slightly above 500 Hz is not well observed.
If all FRFs from all NS references are included, all the modes are observed as in
Fig. 3.


NS 
NL
Hpower (ω) = Hpq (ω)Hpq ∗ (ω) (20)
p=1 q=1
11 Experimental Modal Parameter Evaluation Methods 633

FRF Autopower
−3
10

−4
10

−5
10
Magnitude

−6
10

−7
10

−8
10
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 2 Auto moment function-FRFs from reference p

FRF Autopower
−1
10

−2
10

−3
10
Magnitude

−4
10

−5
10

−6
10
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 3 Auto moment function-FRFs from all references

634 R. J. Allemang and A. W. Phillips

Reference Autopower
−2
10

−3
10

−4
10
Magnitude

−5
10

−6
10

−7
10

−8
10
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 4 Auto moment function-FRFs from all references

The above figure, however, does not indicate which modes are observed from the
different references. If the above figure is plotted reference by reference, the figure
provides more discernible information regarding which modes are well excited by
each reference (Fig. 4).
These simple techniques are extremely useful but do not provide an accurate
estimate of model order when repeated roots exist or when modes are closely spaced
in frequency. The above figure is found from FRFs of a circular object where
numerous repeated modal frequencies exist. Note that this plot looks very much
like a complex mode indication function (CMIF) for the FRF data but does not
discriminate close modes properly. The CMIF is explained in Sect. 3.3.1. For these
reasons, an appropriate estimate of the order of the model (function of the number
of modes) is of prime concern and is the single most important problem in modal
parameter estimation. A CMIF plot or the equivalent will be needed to properly
determine the order of the model required.
In order to determine a reasonable estimate of the model order for a set of
representative data, a number of techniques have been developed as guides or aids
to the user. Much of the user interaction involved in modal parameter estimation
involves the use of these tools. Most of the techniques that have been developed
allow the user to establish a maximum model order to be evaluated (in many cases,
this is set by the limits of the algorithm or the computer memory). Data is acquired
based upon an assumption that the model order is equal to this maximum. In a
sequential fashion, this data is evaluated to determine if a model order less than
11 Experimental Modal Parameter Evaluation Methods 635

the maximum will describe the data sufficiently. This is the point that the user’s
judgment and the use of various evaluation aids become important. Some of the
commonly used techniques are mode indication functions, consistency (stability or
stabilization) diagrams, and pole surface density plots.

3.3 Mode Indication Functions

Mode indication functions (MIF) are normally real valued, frequency domain
functions that exhibit local minima or maxima at the natural frequencies of the
modes. One mode indication function can be plotted for each reference available
in the measured data. The primary mode indication function will exhibit a local
minimum or maximum at each of the natural frequencies of the system under test.
The secondary mode indication function will exhibit a local minimum or maximum
at repeated or pseudo repeated roots of order two or more. Further mode indication
functions yield local minima or maxima for successively higher orders of repeated
or pseudo repeated roots of the system under test.

3.3.1 Complex Mode Indication Function

An algorithm based on singular value decomposition (SVD) methods applied to
multiple reference FRF measurements, identified as the complex mode indication
function (CMIF), was first developed for traditional FRF data in order to identify the
proper number of modal frequencies, particularly when there are closely spaced or
repeated modal frequencies [43]. Unlike the multivariate mode indication function
(MvMIF), which indicates the existence of normal (real-valued) modes, CMIF
indicates the existence of real-valued or complex-valued modes and the relative
magnitude of each mode. Furthermore, MvMIF yields a set of force patterns that
can best excite the normal (real-valued) mode, while CMIF yields the corresponding
mode shape and modal participation vector.
The CMIF is defined as the economical singular values, computed from the FRF
matrix at each spectral line. The CMIF is the plot of these singular values on a
log magnitude scale as a function of frequency. The peaks detected in the CMIF
plot indicate the existence of modes, and the corresponding frequencies of these
peaks give the damped natural frequencies for each mode. In the application of
CMIF to traditional modal parameter estimation algorithms, the number of modes
detected in the CMIF determines the minimum number of degrees of freedom of the
system equation for the algorithm. A number of additional degrees of freedom may
be needed to take care of residual effects and noise contamination.

[H (ω)] = [U (ω)]Σ(ω)[V (ω)]H (21)

Most often, the number of input points (reference points), Ni , is less than the
number of response points, No . In Eq. 21, if the number of effective modes is less
636 R. J. Allemang and A. W. Phillips

than or equal to the smaller dimension of the FRF matrix, i.e., Ne ≤ Ni , the singular
value decomposition leads to approximate mode shapes (left singular vectors) and
approximate modal participation factors (right singular vectors). The singular value
is then equivalent to the the scaling factor Qr divided by the difference between
the discrete frequency and the modal frequency j ω − λr . For a given mode, since
the scaling factor is a constant, the closer the modal frequency is to the discrete
frequency, the larger the singular value will be. Therefore, the damped natural
frequency is approximately the frequency at which the maximum magnitude of
the singular value occurs. If different modes are compared, the stronger the modal
contribution (larger residue value), the larger the singular value will be. The peak
in the CMIF indicates the location on the frequency axis that is nearest to the pole.
The frequency is the estimated damped natural frequency, to within the accuracy of
the frequency resolution Δf (Fig. 5).
Since the mode shapes that contribute to each peak do not change much
around each peak, several adjacent spectral lines from the FRF matrix can be
used simultaneously for a better estimation of mode shapes. By including several
spectral lines of data in the singular value decomposition calculation, the effect of
the leakage error can be minimized. If only the quadrature (imaginary) part of the
FRF matrix (for force-displacement or force-acceleration FRFs) is used in CMIF,
the singular values will be much more distinct. Note that if force-velocity FRFs are
used (as in laser-based response measurements), the coincident (real) part of the
FRF is used (Fig. 6).

Complex Mode Indication Function

−3
10

−4
10

−5
10
Magnitude

−6
10

−7
10

−8
10

−9
10
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 5 Complex mode indication function (CMIF)

11 Experimental Modal Parameter Evaluation Methods 637

Complex Mode Indication Function

−4
10

−5
10

−6
10
Magnitude

−7
10

−8
10

−9
10

−10
10
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 6 Quadrature complex mode indication function (CMIF)

3.3.2 Multivariate Mode Indication Function

The multivariate mode indication function (MvMIF) is based upon finding a force
vector F that will excite a normal mode at each frequency in the frequency range of
interest [44]. If a normal mode can be excited at a particular frequency, the response
to such a force vector will exhibit the 90◦ phase lag characteristic. Therefore, the
real part of the response will be as small as possible particularly when compared
to the imaginary part or the total response. In order to evaluate this possibility, a
minimization problem can be formulated as follows:

{F }T [HReal ]T [HReal ] {F }
min =λ (22)
||F ||=1 {F }T [HReal ]T [HReal ] + [HI mag ]T [HI mag ] {F }

This minimization problem is similar to a Rayleigh quotient, and it can be shown

that the solution to the problem is found by finding the smallest eigenvalue λmin and
the corresponding eigenvector {F }min of the following problem:

[HReal ]T [HReal ] {F } = λ [HReal ]T [HReal ] + [HI mag ]T [HI mag ] {F } (23)

The above eigenvalue problem is formulated at each frequency in the frequency

range of interest. Note that the result of the matrix product [HReal ]T [HReal ] and
[HI mag ]T [HI mag ] in each case is a square, real-valued matrix of size equal to the
638 R. J. Allemang and A. W. Phillips

Multi−Variate Mode Indication Function

1

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500
Frequency (Hz)

Fig. 7 Multivariate mode indication function (MvMIF)

number of references in the measured data Ni × Ni . The resulting plot of a mode

indication function for a seven reference case can be seen in Fig. 7.

3.4 Consistency Diagrams

One of the most common methods for determining the number of modes present
in the measurement data is the use of consistency diagrams, formerly referred to
as stability or stabilization diagrams. The consistency diagram is developed by
successively computing different model solutions (utilizing different model orders
for the characteristic polynomial, different normalization methods for the character-
istic matrix coefficient polynomial, different equation condensation methods, and/or
different algorithms) and involves tracking the estimates of frequency, damping,
and possibly modal participation vectors as a function of model solution iteration.
Symbols are used to identify consistency characteristics between the modal param-
eter solutions found in each iteration, comparing the estimates to those found in
the previous iteration. Normally, the consistency levels are evaluated sequentially
beginning with whether (1) the modal parameter estimates are physically reasonable
(damping), (2) the modal parameters are found as conjugate pairs, (3) the frequency
estimates are consistent, (4) the frequency and damping estimates are consistent,
and (5) the frequency, damping, and modal participation vectors are consistent.
Consistency at each level is defined by a user tolerance that defines how close
11 Experimental Modal Parameter Evaluation Methods 639

statistically that the estimates, from one model iteration to the next, must be to be
classified as consistent. Note that when modal participation vectors are compared,
the modal assurance criterion (MAC) is utilized (see Sect. 4.5 for more specific
information).
As the number of model solution iterations is increased, more and more modal
frequencies are estimated, but, hopefully, the estimates of the physical modal
parameters will be consistently determined as the correct model parameters are
found. Ideally, the nonphysical (computational) modes will not be estimated in a
consistent way during this process and can be sorted out of the modal parameter data
set more easily. Note that inconsistencies (frequency shifts, leakage errors, etc.) in
the measured data set will obscure this consistency and render the diagram difficult
to use. Normally, a tolerance, in percentage, is given for the consistency of each of
the modal parameters that are being evaluated.
Figures 8 through 9 demonstrate two different presentations of consistency
diagrams based upon two common normalizations of the characteristic matrix coef-
ficient polynomial, omitting the lower-level consistencies associated with conjugate
only poles and frequency only consistency in Fig. 9. The omission of the lower-
level consistencies results in a clear consistency diagram. In all of the following
figures, the MIMO FRF data was estimated from an impact test with seven reference
accelerometers and a frequency resolution of 5 Hz (Δf = 5 Hz.). For these cases,
the polyreference time domain (PTD) algorithm was used, but the results shown
are typical for all algorithms evaluated. In all figures, a CMIF plot is plotted on the
consistency diagram in the background for reference.

Fig. 8 Consistency diagram – max order 30 – both normalizations

640 R. J. Allemang and A. W. Phillips

Fig. 9 Consistency diagram – max order 30 – both normalizations, clear

If only model order is evaluated from iteration to iteration, research [40]

has shown that the correct choice of normalization of the characteristic matrix
coefficient polynomial (lowest order coefficient normalized to unity) has a distinct
effect on producing clear consistency diagrams. However, the biggest effect comes
from the solutions that are excluded, or included, based upon the tolerance levels
and consistency evaluations that are in the associated consistency diagram. Figure 9
compared to Fig. 8 distinctly shows that a clear consistency diagram is achieved
when the lowest levels of consistency evaluation are excluded.
Note that the only difference between Figs. 8 and 9 is the omission of the lower-
level consistency evaluations (omitting conjugate pair evaluation and frequency
consistency evaluation). Figure 9 clearly shows the benefits of a clear consistency
diagram. Similar results are obtained when the tolerance levels are changed for some
of the consistency evaluations.
Note that in Figs. 8 and 9, the icons used to represent consistency form a vertical
line, referred to as a trail, which helps identify valid results. Also note that the trails
in these figures are often so close together as to be viewed as a single trail when
more than one trail is actually present. When the consistency diagram is viewed
interactively, the frequency scale can be expanded to make this evident. This is
the case for these figures which involve FRF data from a circular plate with very
close modal frequencies, well within the frequency resolution (Δf = 5 Hz). Most
of the trails in these consistency diagrams represent two or more very close modal
frequencies.
11 Experimental Modal Parameter Evaluation Methods 641

3.4.1 Alternate Consistency Diagram

Another version of a consistency diagram has been developed to overcome some of
the sequential nature of the consistency evaluation. The process is the same as in
the more common consistency diagram, but the evaluation from one order iteration
to the next is what is changed. The modal frequency and damping is combined with
the modal participation vector estimate to form a state vector, sometimes referred to
as a pole-weighted vector. Then the state vectors from one iteration to the next are
evaluated via the modal assurance criterion (MAC) of the state vectors. The modal
assurance criterion (MAC) and the state vector concept are explained in more detail
in Sect. 4.5. The result of this change in presentation is shown in Fig. 10 which
represents the same information as that in Fig. 8.

3.5 Pole Surface Consistency Plots

One common problem with consistency diagrams is that consistency is estimated

based upon a statistical comparison to the previous model solution iteration,
evaluating the different levels of consistency in sequential order. While this works
in most cases, it will sometimes yield a consistency diagram that is difficult to
interpret. Furthermore, if model iterations from different algorithms are combined
in one consistency diagram, the sequential nature of the consistency diagram can
pose interpretation problems. In this case, since different model characteristics
are being evaluated at the same time, the order that the model solution iterations
are presented may affect the presentation of consistency (or stability). In general,

Fig. 10 Consistency diagram – max order 30 – both normalizations

642 R. J. Allemang and A. W. Phillips

Fig. 11 Pole surface consistency – max order 30 – both normalizations

Fig. 12 Pole surface consistency – max order 30 – both normalizations, clear

a clearer estimate of the modal frequencies will be determined by plotting the

consistency (or density) of poles found from all model solution iterations, presented
in the second quadrant of the complex plane [45]. Figures 11, 12, and 13 represent
pole surface consistency plots for the same previous two consistency diagrams. Note
11 Experimental Modal Parameter Evaluation Methods 643

Fig. 13 Pole surface consistency – max order 30 – both normalizations

that a black square represents the statistical solution for the one modal parameter set
that best represents the associated cluster of modal parameters.

3.6 Modal Parameter Clustering

The modal parameter clusters shown in Figs. 11, 12, and 13 can be grouped together
based on the closeness of the modal frequencies and damping along with the
similarity in the associated modal participation vectors or associated state vector
form of the modal participation vectors (see Sect. 4.5.3). Once grouped in this
fashion, the modal parameter cluster can be autonomously identified, and statistics
can be estimated to quantify the variance in the modal parameters found from the
associated consistency diagram.
The two modal parameter clusters represented in Fig. 13 are individually shown
in Figs. 14 and 15. The rectangular box and associated circle represent one sigma
of standard deviation in the modal frequency pole in terms of frequency units of
Hertz. All of the identified solutions in the figure have been used to identify the best
answer for modal frequency and establish the statistics, modal frequency by modal
frequency. Note that in both of these figures, 54 modal frequency estimates from
the consistency diagram trail have been utilized. The vertical axis in these figures
has been converted from fraction of critical damping to damping in units of Hertz to
allow the standard deviation to have square aspect ratio.
644 R. J. Allemang and A. W. Phillips

Fig. 14 Modal parameter cluster – 362.3 Hz

Fig. 15 Modal parameter cluster – 363.6 Hz

Please note that the identified modal frequency and damping information is very
consistent within a cluster despite the frequency resolution of the FRF data being
5 Hz.
11 Experimental Modal Parameter Evaluation Methods 645

4 Modal Vector Evaluation/Validation Tools

4.1 Modal Vector Conditioning

Vector conditioning generally applies to the three vectors (modal participation

vectors, residue vectors, and modal vectors) that are estimated as part of any
modal parameter estimation method. Vector conditioning involves methods that are
designed to improve the final modal vector estimates and may vary from user to user
and may be available at different levels in commercial software. In general, vector
conditioning, first and foremost, is designed to remove potential contamination from
the final modal vector estimates.
Vector conditioning begins with the recognition that the modal participation
vector is the result of an eigenvalue problem where the eigenvectors are arbitrarily
scaled and even a normal mode might be represented as a set of real-valued
numbers modified by a complex-valued scalar. This means that some sort of
normalization is needed to remove the complex-valued scalar. Furthermore, if some
sort of real normalization is desirable (to match up well with, e.g., an undamped,
analytical model), the understanding of the contamination that is being removed is
a prerequisite to any procedure. Random contamination may simply be ignored,
smoothed, or averaged out, but if the contamination is related to a bias from
nearby modes, it may indicate that the modal parameter estimation may need further
evaluation or that more data from additional reference DOFs is required.

4.1.1 Vector Normalization

Vector conditioning begins with some form of normalization. Normalization can
simply mean that each vector will be normalized with respect to one element within
the vector. This is often done as is shown in Eq. 24 where the residue vector is
divided by the largest element within the residue vector.

 
Apqr NL ×1 ψpr ψqr
{ψr }NL ×1 = Apqr = (24)
max(Apqr ) MAr

The meaning of largest in this sense is that the magnitude of the complex-valued
element of the vector is the maximum absolute value found within the vector but
the division occurs on a complex-valued basis. This means that the largest element
of the normalized vector will be unity. In general, though, the normalized vector
will still be complex valued at all other positions in the normalized vector. Other
normalization choices are possible (unity vector length, etc.), but dividing by the
maximum provides a simple and reliable method. This approach will also yield
modal scaling values that can be more directly related to the physical properties of
the system.
Normalization as defined above can be applied to either the modal participation
vectors or the modal vectors. Since the residue vectors have physical meaning and
646 R. J. Allemang and A. W. Phillips

units, normalization is generally not applied except to estimate the modal vector
from the residue vector.

4.1.2 Real Normalization

Once a vector has been normalized to be dominantly real valued, the vector
can be limited to only real values if desired. This limitation can be a simple
truncation (delete imaginary parts) or some other procedure that preserves the
complex magnitude with the appropriate 0 or 180 degree phase. The complex-
valued vectors should always be evaluated, visually or numerically, before any
real normalization is performed to be certain the process does not have unintended
consequences.
In the case of modal participation vectors, while these vectors are theoretically
proportional to the final modal vectors at the included DOFs, their function in the
modal parameter estimation process is one of weighting the least squares solution
in the second stage of modal parameter estimation (MPE-2). Performing a real
normalization of the modal participation vectors has been found to be beneficial
always and does not affect the final modal vector estimates adversely [46].

4.1.3 Central Axis Rotation

The normalization described in the previous section guarantees that one element
in the vector will become unity but the vector itself may statistically not be best
aligned with the real axis. A statistical estimate of the central axis of the vector
can be performed to make sure that the mean phase of the vector is along the real
axis. This is in addition to making sure that the largest magnitude in this statistical
solution is unity (in a complex magnitude sense).
In order to establish a uniform procedure for normalizing vectors that are related
to modal vectors (modal vectors, modal participation vectors, residue vectors), each
potentially complex-valued vector must first be rotated to an orientation where
the dominant information of the vector in complex space is aligned with the real
axis. This is required since the complex-valued vector is generally the result of
the solution of an eigenvalue-eigenvector problem involving the complex-valued
MIMO FRF data as stated above.
Given an original modal vector for mode r, a central axis rotation method is
utilized to determine the dominant axes. These dominant axes can be identified via
the singular value decomposition of the relationship between the real part, {ψR }r ,
and the imaginary part, {ψI }r , of the vector for mode r as follows:

 T  
[U, Σ, V ] = SV D {ψI }r {ψR }r {ψI }r {ψR }r (25)

Recognize that this decomposition is an attempt to locate the two dominant

axes of a 2D ellipse that encompasses the modal vector data in the complex
plane. Following the decomposition, the central axis angle is estimated using the
true (four-quadrant) arctangent of the right singular vector relationship. Note that,
regardless of the number of DOF positions represented in the modal vector, the right
11 Experimental Modal Parameter Evaluation Methods 647

singular vector matrix will always be two by two in dimension representing the 2D
characteristics of the ellipse.


−1 V22
φ̄r = tan (26)
−V12

After identifying the central axis angle, the original complex-valued modal vector
is rotated by multiplying by the complex rotational phasor.

 
ψ̂ = cos (φ̄r ) − j sin (φ̄r ) {ψ}r (27)
r

This rotation, followed by a normalization by the magnitude of the maximum

element in the resulting vector, results in a normalization procedure that assures
that the resulting vector is dominantly real, with a maximum value of unity,
based upon all of the vector information, rather than a single DOF that is chosen
arbitrarily (e.g., rather than selecting the DOF associated with the largest modal
vector coefficient).

4.1.4 Vector Complexity

Vector complexity is often defined in terms of mean phase deviation (MPD) as an
indication of how the phase deviates from 0 and/or 180 degrees. This definition
allows some ambiguity in what is meant by a complex mode. It may simply mean
that the elements of the estimated modal vector contain complex values. For this
case, the elements of the modal vector may be rotated by an angle in the complex
plane but are otherwise co-linear. Or it may mean that the modal vector contains
complex values that cannot be made real by a simple complex phasor rotation. For
this case, the modal coefficients are not all co-linear in the complex plane. For this
development, it is the second definition that is used. Note that if the central axis
rotation procedure is performed, the mean phase will always be zero degree.
The mean phase deviation (MP Dr ) for modal vector r has been defined
historically as a number between zero and unity where zero represents a real-
valued modal vector (normal mode) and where unity represents a complex-valued
modal vector with no recognizable dominant real or imaginary characteristic, once
an attempt has been made to rotate the vector to a dominant central axis position.
This fraction is often multiplied by 100 to represent the percentage of complex-
valued modal vector characteristics. In terms of the definitions utilized in the
previous section, assuming that the modal vector has already been rotated to its
most dominant real orientation, the MP Dr is defined as the norm of the imaginary
part of the rotated vector divided by the norm of the real part of the rotated vector,
as shown in Eq. 28. Thus, the MP Dr gives a dispersion ratio around the central axis
of the rotated modal vector bounded between zero and one.
648 R. J. Allemang and A. W. Phillips

Mean Phase Deviation (MP Dr )

  
 
 ψ̂I 
MP Dr =   
r
(28)
 
 ψ̂R 
r

Once the MP Dr is defined in terms of the fraction between zero and unity,
the associated mean phase correlation (MP Cr ) for modal vector r is defined as
in Eq. 29.

Mean Phase Correlation (MP Cr )

MP Cr = 1 − MP Dr (29)

The mean phase correlation can then be interpreted as an indicator of normalcy,

from a purely normal mode (1.0) to a purely complex mode (0.0).

4.1.5 Modal Vector Complexity Plots

A modal vector complexity plot (MVCP) is often used to represent the character-
istics of modal vectors with respect to MPC and vector complexity. The vectors in
this presentation are always scaled to a maximum value of unity. Figure 16 is an
example of such a plot.
For these case studies, the entire frequency range from 220 to 2500 Hz was fit
by the RFP-Z algorithm and the ERA algorithm using traditional complex-valued
participation vector weighting as well as real-valued participation vector weighting.
For most cases, the complete set of modal participation vectors was utilized although
two cases show the deficiencies of single-reference approaches. One case with two
references shows the power of multiple reference modal parameter estimation. The
CSSAMI autonomous modal parameter estimation procedure [47] was utilized to
remove user bias from the selected results.

Case 6
This modal vector complexity plot (MVCP) demonstrates that, while most of the
modes appear to have been fit for residues in the MPE-2 stage reasonably well,
simply converting the complex-valued participation to real does not guarantee the
quality of fit. The three modes between 2320 and 2340 Hz represent a repeated
root pair and an additional close mode. The MPC values show that the fit for two
of the modes retains significant modal complexity that is not anticipated in this
structure.
11 Experimental Modal Parameter Evaluation Methods 649

Modal Vector Complexity Plots

362.4 Hz 363.6 Hz 557.0 Hz 761.2 Hz 764.2 Hz

MPC 99.4% MPC 99.2% MPC 99.4% MPC 99.0% MPC 99.1%
M 1.10e+04〈−88.5°〉 M 1.53e+04〈−89.8°〉 M 4.13e+04〈−89.0°〉 M 2.10e+04〈−90.1°〉 M 2.06e+04〈−91.4°〉
A A A A A

1223.0 Hz 1224.1 Hz 1328.0 Hz 1328.7 Hz 2019.1 Hz

MPC 97.4% MPC 97.3% MPC 96.0% MPC 97.6% MPC 98.0%
M 5.30e+04〈−89.6°〉 M 5.84e+04〈−90.2°〉 M 3.51e+04〈−87.7°〉 M 3.43e+04〈−91.1°〉 M 4.51e+04〈−87.7°〉
A A A A A

2024.0 Hz 2321.6 Hz 2322.8 Hz 2337.5 Hz

MPC 95.2% MPC 94.7% MPC 66.1% MPC 81.2%
MA4.47e+04〈−86.2°〉 MA5.89e+04〈−85.1°〉 MA2.32e+05〈−59.3°〉 MA1.84e+05〈−99.7°〉

Fig. 16 Case 6: Modal vector complexity plot, multiple references, all DOFs, ERA method

Case 8
The data for this modal vector complexity plot (MVCP) is identical to Case 6 with
the addition of reprocessing the normalized eigenvector from the ERA method to
residue scaling and then performing the central axis rotation and decimation of the
long dimension NL vectors to the short dimension NS . This is followed by real
normalization of the short dimension participation vectors ([L]) which yields a very
physical, realistic result. The improvement on the estimation can be noted for the
last two modes (Fig. 17).

Case 10
This modal vector complexity plot (MVCP) shows a comparable result to Case 8
even though a short dimension MPE algorithm is utilized. Case 10 shows that
reprocessing the normalized eigenvector from the RFP-Z method to residue scaling
and then performing the central axis rotation, followed by real normalization of the
short dimension participation vectors ([L]), yields a very physical, realistic result.
This result has a slight improvement in the MPC value for the last two modes
(Fig. 18).
650 R. J. Allemang and A. W. Phillips

Modal Vector Complexity Plots

362.4 Hz 363.6 Hz 557.0 Hz 761.2 Hz 764.2 Hz

MPC 99.4% MPC 99.2% MPC 99.4% MPC 99.0% MPC 99.2%
MA1.10e+04〈−88.5°〉 MA1.54e+04〈−89.9°〉 MA4.14e+04〈−89.0°〉 MA2.11e+04〈−90.2°〉 MA2.06e+04〈−91.4°〉

1223.0 Hz 1224.1 Hz 1328.0 Hz 1328.7 Hz 2019.1 Hz

MPC 97.3% MPC 97.3% MPC 95.9% MPC 97.7% MPC 98.4%
M 5.30e+04〈−89.6°〉 M 5.87e+04〈−90.2°〉 M 3.39e+04〈−87.7°〉 M 3.43e+04〈−91.2°〉 M 4.47e+04〈−88.0°〉
A A A A A

2024.0 Hz 2322.2 Hz 2322.3 Hz 2337.5 Hz

MPC 95.8% MPC 97.3% MPC 94.8% MPC 94.9%
M 4.25e+04〈−86.3°〉 MA4.72e+04〈−85.9°〉 M 6.32e+04〈−86.3°〉 MA2.19e+05〈−99.2°〉
A A

Fig. 17 Case 8: Modal vector complexity plot, multiple references, all DOFs, ERA method

4.2 Modal Vector Validation: eFRF

A virtual measurement, known as the enhanced frequency response function

(eFRF), can be used to identify the modal frequencies and scaling of a single degree
of freedom characteristic that is associated with each peak in the CMIF [43]. If
only one peak is dominant in the eFRF, this can be used as a validation of the
particular modal vector. The eFRF is developed based upon the concept of physical
to modal coordinate transformation and is used to manipulate frequency response
functions so as to enhance a particular mode of vibration. The left singular vectors,
associated with the peaks in the CMIF, are used as an estimate of the modal filter
which accomplishes this.
4.2.1 eFRF: Theoretical Definition
Starting with the basic FRF partial fraction equation:


2N
Apqr 
2N
Qr ψpr ψqr
Hpq (ω) = = (30)
(j ω − λr ) (j ω − λr )
r=1 r=1
11 Experimental Modal Parameter Evaluation Methods 651

Modal Vector Complexity Plots

362.6 Hz 363.9 Hz 557.1 Hz 761.2 Hz 764.2 Hz

MPC 98.8% MPC 99.2% MPC 99.5% MPC 99.1% MPC 99.3%
MA1.08e+04〈−87.2°〉 MA1.45e+04〈−87.5°〉 MA4.10e+04〈−88.7°〉 MA2.03e+04〈−89.9°〉 MA2.02e+04〈−92.4°〉

1223.0 Hz 1224.1 Hz 1328.0 Hz 1328.8 Hz 2019.4 Hz

MPC 97.8% MPC 96.3% MPC 96.3% MPC 97.7% MPC 98.1%
M 5.50e+04〈−89.0°〉 M 5.85e+04〈−89.7°〉 M 3.27e+04〈−91.0°〉 M 3.21e+04〈−90.7°〉 M 4.37e+04〈−87.0°〉
A A A A A

2023.9 Hz 2322.0 Hz 2322.5 Hz 2338.2 Hz

MPC 95.1% MPC 97.4% MPC 97.5% MPC 97.0%
M 4.14e+04〈−86.9°〉 MA8.61e+04〈−90.3°〉 M 9.75e+04〈−91.9°〉 MA1.98e+05〈−89.8°〉
A A

Fig. 18 Case 10: Modal vector complexity plot, multiple references, all DOFs, RFP-Z method

Noting that the scaling term Qr can represent either modal A or modal mass
scaling, the FRF relationship can be redefined:


2N
Qr ψqr
Hpq (ω) = ψpr (31)
(j ω − λr )
r=1

The enhanced frequency response function (eF RFr (ω)) for mode r can now be
defined:

Qr ψqr
eF RFr (ω) = (32)
(j ω − λr )

The eFRF has only to do with the reference location q and is constant for a given
column of the FRF matrix. The FRF model for a column of FRFs can be rewritten as:
652 R. J. Allemang and A. W. Phillips

  
2N
  Qr ψqr 
2N
 
Hpq (ω) = ψpr = ψpr eF RFr (ω) (33)
(j ω − λr )
r=1 r=1

4.2.2 eFRF: Historical Development

The eFRF can be formulated from measured frequency response function data in
the following manner:

  
2N
  Qr ψqr Qs ψqs
{ψsr }T [M] Hpq (ω) = {ψsr }T [M] ψpr = Ms
(j ω − λr ) (j ω − λs )
r=1
(34)
 
{ψsr }T [M] Hpq (ω) = Ms eF RFs (ω) (35)

The last equation indicates that an estimate of the modal mass scaling (Ms )
is needed to develop the enhanced frequency response function as well as a
mass matrix whose dimension has been reduced to the DOFs of the FRF column
of measurements. In many situations, when the mass distribution is adequately
represented by the measured degrees of freedom (good spatial representation), the
modal vector can be used directly to estimate an eFRF that is proportional to within
a complex constant representing the mass scaling.
 
{ψsr }T Hpq (ω) ≈ eF RFs (ω) (36)

In reality, the goal of the eFRF is to allow a simple SDOF modal parameter
estimation algorithm to be used to estimate modal frequency and scaling from the
eFRF. If other modes are still observable in the eFRF, these modes can be handled
with residuals in the modal parameter estimation. The fact that the resultant eFRF
looks dominantly like an SDOF FRF is used as a validation that the modal vector
ψr is correct.

4.2.3 eFRF: FRF SVD Development

The above estimation of the eFRF was used in the original development [48] which
was based upon orthogonality. The original modal vectors used in the eFRF esti-
mation were simply quadrature FRF estimates taken from the FRF measurements
(SDOF estimate). Modal vectors from any modal parameter estimation algorithm
can be used.
Subsequently, the eFRF was revised to take advantage of the singular vectors
of the FRF matrix which changes the formulation so that it is based upon the linear
expansion theorem. This simplifies some of the scaling issues and does not require a
reduced mass matrix to get properly scaled eFRFs. In the newer technique, utilized
in the following examples, the left and right singular vectors, associated with the
singular values of the FRF matrix (found at the peaks of the CMIF), are used as
in the following equations. The frequency response function matrix is assumed to
11 Experimental Modal Parameter Evaluation Methods 653

have been corrected for transducer orientation. Although this is not required for
the computation of damped natural frequency and damping, this will simplify the
phasing issues associated with the eFRFs and any modal scaling estimation.
The singular value decomposition (SVD) of the measured FRF matrix can be
represented as:

[H (ω)] = [U (ω)]Σ(ω)[V (ω)]H (37)

Note that the dimensions of the FRF matrix are assumed to be NL × NS at

each frequency. Using lower case notation to represent the individual left and right
singular vectors, the eFRF can be represented by:

eF RFr (ω) = α {ur }T [H (ω)] {vr } (38)

The eFRF is typically used, together with single degree of freedom (SDOF)
modal parameter estimation methods, to estimate the frequency and damping of
the associated modal frequency. The validation of the modal vector is based upon
whether the associated eFRF is dominantly an SDOF function. In order for the
eFRF to also be used to estimate the modal scaling (modal mass and/or modal A),
the correct scaling (correct magnitude and phase) of the eFRF must be accounted
for (the complex-valued scaling term α in the above equation). Since the left and
right singular vectors in the singular value decomposition are unitary and scaled
consistently as a set, both left and right singular vectors must be involved to preserve
the physical characteristics of the eFRF.
For the general case, where the modal vector used in the eFRF is estimated from
the left singular vector associated with a peak in the singular values in the CMIF
plot, the eFRF is scaled by utilizing the values of the left and right singular vectors,
associated with the significant singular value, at the driving point locations. Note
that it is probable that v may not be a strict subset of u; in this case a scale factor
must be estimated from the common subset of the input/output degrees of freedom
(i.e., the driving point degrees of freedom). Complete details of the development
can be found in [49].
Figure 19 shows the eFRF for one of the modes around 360 Hz. The modal vector
used for the enhancement (weighted averaging) was a complex-valued estimate of
the modal vector. The result is clearly very close to a SDOF FRF plot with small
deviations around the frequencies where other modes were present in the data.
Figure 20 shows the eFRF for the same mode around 360 Hz. The modal vector
used for the enhancement (weighted averaging) was a real-valued estimate of the
modal vector. The result is clearly very close to an SDOF FRF plot with small
deviations around the frequencies where other modes were present in the data. Some
small differences are noted in the modal parameters when compared to Fig. 19. Both
figures clearly validate that the modal vectors are appropriate for this set of FRF
data.
654 R. J. Allemang and A. W. Phillips

Fig. 19 Enhanced FRF – complex-valued modal vector (362.32 Hertz 0.884 % critical damping)

Fig. 20 Enhanced FRF – real-valued modal vector (362.32 Hertz 0.884 % critical damping)

Figures 21 and 22 show the eFRF results for two very closely spaced modal
frequencies in the FRF data. The modal parameter information for the two eFRFs
is notably different, and the two modal vectors are validated for this FRF data set.
Note that the frequency resolution in the FRF data is 5 Hz, so the SVD information
that is the basis of the left and right singular vectors is likewise limited to 5 Hz
11 Experimental Modal Parameter Evaluation Methods 655

Fig. 21 Enhanced FRF – Repeated Modal Frequency 1 (362.32 Hertz 0.884 % critical damping)

Fig. 22 Enhanced FRF – Repeated Modal Frequency 2 (363.53 Hertz 0.569 % critical damping)

resolution. Nevertheless, the SDOF modal parameter estimation method is clearly

able to estimate modal frequencies within the 5 Hz resolution. These results compare
favorably with modal parameter estimation results from polyreference algorithms.
The concept of an enhanced frequency response function (eFRF) can be used
anytime a column/row of FRF data is available along with an estimate of a modal
vector.
656 R. J. Allemang and A. W. Phillips

4.3 Weighted Modal Vector Orthogonality

The primary method that has historically been used to validate an experimental
modal model involves a weighted orthogonality check comparing measured modal
vectors, analytical modal vectors, and an appropriately sized, square weighting
matrix. The size of the square weighting matrix must match the length and spatial
dimension (in terms of DOFs) of the analytical and experimental modal vectors and
is generally formed from an estimated mass or stiffness matrix. A common approach
uses analytical modal vectors together with experimental modal vectors and the
appropriately sized mass or stiffness matrix. This latter comparison is normally
referred to as a pseudo orthogonality check (POC).
In the traditional weighted orthogonality check, the modal vectors are used
together with a mass matrix, normally derived from a finite element model, to
evaluate orthogonality of the modal vectors. In the pseudo orthogonality check,
the experimental modal vectors are used together with a mass matrix, normally
derived from a finite element model, and the analytical modal vectors, normally
derived from the same finite element model, to evaluate orthogonality between the
experimental and analytical modal vectors.
The experimental and analytical modal vectors are scaled so that the diagonal
terms of the modal mass matrix are unity. With this form of scaling, the off-diagonal
values in the modal mass matrix are expected to be less than 0.1 (10 percent of the
diagonal terms). Theoretically, for the case of proportional damping, each modal
vector of a system will be orthogonal to all other modal vectors of that system when
weighted by the mass, stiffness, or damping matrix. In practice, these matrices are
made available by way of a finite element analysis, and normally the mass matrix
is considered to be the most accurate. For this reason, any further discussion of
weighted orthogonality will be made with respect to mass matrix weighting. As a
result, the weighted orthogonality relations can be stated as follows:

4.3.1 Weighted Orthogonality of Modal Vectors

For r = s:

{ψr }T [M] {ψs } = 0 (39)

For r = s:

{ψr }T [M] {ψs } = Mr (40)

{ψr }T   {ψs }
√ M √ = 1.0 (41)
Mr Mr

Weighted cross orthogonality refers to Eq. 39, which means that the vectors have
no projection on each other in the dimensionality of the modal vector space.
11 Experimental Modal Parameter Evaluation Methods 657

4.4 Weighted Pseudo Orthogonality of Modal Vectors

Weighted pseudo orthogonality refers to the case where one side of Eqs. 39
through 41 is replaced by experimental modal vectors. Note that this process
involves making sure that the dimensionality of the analytical modal vectors, the
experimental modal vectors, and the mass matrix is the same and the DOFs utilized
are consistent.
Experimentally, the result of zero for the weighted pseudo cross orthogonality
calculations (r = s) (Eq. 39) can rarely be achieved, but values up to one tenth of the
magnitude of the generalized mass of each mode are considered to be acceptable.
It is a common procedure to form the modal vectors into a normalized set of mode
shape vectors with respect to the mass matrix weighting. The accepted criterion in
the aerospace industry, where this confidence check is made most often, is for all of
the generalized mass terms to be unity and all weighted pseudo cross orthogonality
terms to be less than 0.1. Often, even under this criteria, an attempt is made to
adjust the modal vectors so that the weighted pseudo cross orthogonality conditions
are satisfied.
Note that, in general, experimental modal vectors are not always real valued and
Eqs. 39 through 41 are developed based upon normal (real-valued) modal vectors.
This complication has to be resolved by a process of real normalization of the
measured modal vectors prior to utilizing Eqs. 39 through 41 or by applying an
equivalent procedure involving the state-space form of the weighted orthogonality
relationship.
In Eqs. 40 and 41, the mass matrix must be an NL × NL matrix corresponding
to the measurement locations on the structure. This means that the finite element
mass matrix must be modified from whatever size and distribution of grid locations
required in the finite element analysis to the NL ×NL square matrix corresponding to
the measurement locations. This normally involves some sort of reduction algorithm
as well as interpolation of grid locations to match the measurement situation [50,
51, 52, 53, 54, 55, 56].
When Eqs. 39 is not sufficiently satisfied, one (or more) of three situations may
exist. First, the modal vectors can be invalid. This can be due to measurement errors
or problems with the modal parameter estimation algorithms. This is a very common
assumption and many times contributes to the problem. Second, the mass matrix
can be invalid. Since the mass matrix does not always represent the actual physical
properties of the system when it is built or assembled, this probably contributes
significantly to the problem. Third, the reduction of the mass matrix can be invalid
[50, 51, 52, 53, 54, 55]. This can certainly be a realistic problem and cause severe
errors. The most obvious example of this situation would be when a relatively
large amount of mass is reduced to a measurement location that is highly flexible,
such as the center of an unsupported panel. In such a situation, the measurement
location is weighted very heavily in the orthogonality calculation of Eqs. 39 but
may represent only incidental motion of the overall modal vector. In all probability,
all three situations contribute to the failure of orthogonality or pseudo orthogonality
criteria on occasion. When the orthogonality conditions are not satisfied, this result
658 R. J. Allemang and A. W. Phillips

does not indicate where the problem originates. From an experimental point of view,
it is important to try to develop methods that indicate confidence that the modal
vector is, or is not, part of this problem.

4.5 Modal Vector Consistency

Since any evaluation of weighted orthogonality requires an appropriately sized

weighting matrix and real-valued modal vectors, these two conditions pose prob-
lems as they are not often met by experimentally measured data. Beginning around
1975, the experimental focus changed to estimating a MIMO FRF set of data where
redundancy of the modal parameters, reference by reference, was expected. Since
a MIMO FRF set of data will have an estimate of the modal parameters for each
reference in the MIMO set, evaluating the redundancy, or consistency, became
desirable. Consistency in the MIMO FRF data set means that the modal parameters
are the same from each references. Consistency therefore means that the complex-
valued modal frequencies are expected to be the same and the modal vectors are
the same to within a complex-valued constant (same shape with different scaling).
Eventually, MIMO MPE methods became available that took advantage of this
concept to estimate a single set of modal parameters directly from MIMO FRF data.
Consistency, with respect to modal vectors, means that the vectors associated
with the same modal frequency from different references should always be linearly
related. This problem was a familiar least squares estimate of the linear complex-
valued scaling constant with an associated squared correlation coefficient that
directly paralleled the least squares estimate of the FRF with the associated ordinary
coherence function. The scaling constant is referred to as the modal scale factor
(MSF), and the squared correlation coefficient is referred to as the modal assurance
criterion (MAC).
The development of the modal assurance criterion [48, 57, 58] came about due
to the focus on modal vector consistency (linearity) as opposed to modal vector
orthogonality. The fact that modal vectors estimated from different references are
expected to be the same to within a complex-valued scaling constant revealed a
consistency methodology that could easily be explored experimentally, independent
of analytical methods and associated weighting matrices.
The development of the modal assurance criterion has led to a number of
similar assurance criteria used in the area of experimental and analytical structural
dynamics. It is important to recognize the mathematical similarity of these varied
criteria in order to be certain that conclusions be correctly drawn from what is
essentially a squared, linear regression correlation coefficient. The modal assurance
criterion is a statistical indicator, just like ordinary coherence associated with FRF
estimation, which can be very powerful when used correctly but very misleading
when used incorrectly. In Sect. 4.5.4, other similar assurance criteria are identified
although the list is not intended to be comprehensive. Typical uses of the modal
assurance criterion are discussed, and, finally, typical abuses are identified.
11 Experimental Modal Parameter Evaluation Methods 659

4.5.1 Modal Vector Linearity or Consistency

When the modal assurance criterion was originally developed, MIMO modal
parameter estimation was just beginning to be developed. Therefore, when MIMO
FRF data was acquired, the primary benefit was to make sure a complete set of
modal parameters was estimated. The big concern was that if a reference was too
close to a nodal position in a modal vector, that mode would be missed or estimated
poorly. If redundant modal vectors were estimated from the references, an average
of the modal vectors could improve the modal vector estimate.
The function of the modal scale factor (MSF) is to provide a means of nor-
malizing all estimates of the same modal vector, taking into account magnitude and
phase differences. Once two different modal vector estimates are scaled similarly,
elements of each vector can be averaged (with or without weighting), differenced,
or sorted to provide a best estimate of the modal vector or to provide an indication
of the type of error vector superimposed on the modal vector. In terms of modern,
multiple reference modal parameter estimation algorithms, the modal scale factor is
a normalized estimate of the modal participation factor between two references for
a specific mode of vibration. The function of the modal assurance criterion (MAC)
is to provide a measure of consistency (degree of linearity) between estimates of
a modal vector. This provides an additional confidence factor in the evaluation of
a modal vector from different excitation (reference) locations or different modal
parameter estimation algorithms. The modal scale factor and the modal assurance
criterion also provide a method of easily comparing estimates of modal vectors
originating from different sources. The modal vectors from a finite element analysis
can be compared and contrasted with those determined experimentally as well as
modal vectors determined by way of different experimental or modal parameter
estimation methods. In this approach, methods can be compared and contrasted
in order to evaluate the mutual consistency of different procedures rather than
estimating the modal vectors specifically. If an analytical and an experimental vector
are deemed consistent or similar, the analytical modal vector, together with the
modal scale factor, can be used to complete the experimental modal vector if some
degrees of freedom could not be measured.
The modal scale factor is defined, according to this approach, as follows:


NL
∗
ψcqr ψdqr  ∗
q=1 {ψcr }T ψdr
MSFcdr = =  ∗ (42)

NL
∗ {ψdr }T ψdr
ψdqr ψdqr
q=1

Since the modal vectors are in general complex valued, this is also equivalent to:

{ψdr }H {ψcr }
MSFcdr = (43)
{ψdr }H {ψdr }
660 R. J. Allemang and A. W. Phillips

Equations 42 and 43 imply that the modal vector d is the reference to which the
modal vector c is compared. In the general case, modal vector c can be considered
to be made of two parts. The first part is the part linearly correlated with modal
vector d. The second part is the part that is not linearly correlated with modal vector
d and is made up of contamination from other modal vectors and of any random
contribution. This error vector is considered to be noise.
The modal assurance criterion is defined as a squared, scalar correlation coef-
ficient constant, bounded zero to one, relating the degree of consistency (linearity)
between one modal vector and another reference modal vector as follows:

Consistency of Modal Vectors

 2
 
 NL ∗ 
 ψcqr ψdqr    
q=1  {ψcr }T ψ ∗ 2
dr
MACcdr = =     (44)

NL 
NL {ψcr }T ψcr∗ {ψ }T ψ ∗
∗ ∗ dr dr
ψcqr ψcqr ψdqr ψdqr
q=1 q=1

Since the modal vectors are complex valued, the modal assurance criterion is a
real-valued scalar equivalent to:

 
{ψdr }H {ψcr }2 {ψdr }H {ψcr } {ψcr }H {ψdr }
MACcdr = = (45)
{ψdr }H {ψdr } {ψcr }H {ψcr } {ψdr }H {ψdr } {ψcr }H {ψcr }

Note that the definition of MAC involves the Hermitian (conjugate transpose)
of the associated modal vectors. A frequent error is made when the Hermitian
is replaced by the transpose. If all the vectors involved are real valued, no error
is generated. However, when any of the vectors are complex valued, the MAC
will be estimated incorrectly. The proper approach is to always use the Hermitian
(conjugate transpose).
The modal assurance criterion takes on values from zero, representing no
consistent correspondence, to one, representing a consistent correspondence. In this
manner, if the modal vectors under consideration truly exhibit a consistent, linear
relationship, the modal assurance criterion should approach unity, and the value
of the modal scale factor can be considered to be reasonable. Note that, unlike
the orthogonality calculations, the modal assurance criterion is normalized by the
magnitude of the vectors and, thus, is bounded between zero and one.
The modal assurance criterion can only indicate consistency, not validity or
orthogonality. If the same errors, random or bias, exist in all modal vector estimates,
this is not delineated by the modal assurance criterion. Invalid assumptions are
normally the cause of this sort of potential error. Even though the modal assurance
criterion is unity, the assumptions involving the system or the modal parameter
estimation techniques are not necessarily correct. The assumptions may cause
11 Experimental Modal Parameter Evaluation Methods 661

consistent errors in all modal vectors under all test conditions verified by the modal
assurance criterion.

Modal Assurance Criterion (MAC) Zero

If the modal assurance criterion has a value near zero, this is an indication that the
modal vectors are not consistent. This can be due to any of the following reasons:

• The system is non-stationary. This can occur if the system is nonlinear, and
two data sets have been acquired at different times or excitation levels. System
nonlinearities will appear differently in frequency response functions generated
from different exciter positions or excitation signals. The modal parameter
estimation algorithms will also not handle the different nonlinear characteristics
in a consistent manner.
• There is noise on the reference modal vector. This case is the same as noise on
the input of a frequency response function measurement. No amount of signal
processing can remove this type of error.
• The modal parameter estimation is invalid. The frequency response function
measurements may contain no errors, but the modal parameter estimation may
not be consistent with the data.
• The modal vectors are from linearly unrelated mode shape vectors. Hopefully,
since the different modal vector estimates are from different excitation positions,
this measure of inconsistency will imply that the modal vectors are orthogonal.

Obviously, if the first three reasons can be eliminated, the modal assurance criterion
can be interpreted in a similar way as an orthogonality calculation.

Modal Assurance Criterion (MAC) Unity

If the modal assurance criterion has a value near unity, this is an indication that the
modal vectors are consistent. This does not necessarily mean that they are correct.
The modal vectors can be consistent for any of the following reasons:

• The modal vectors have been incompletely measured. This situation can occur
whenever too few response stations have been included in the experimental
determination of the modal vector.
• The modal vectors are the result of a forced excitation other than the desired
input. This would be the situation if, during the measurement of the frequency
response function, a rotating piece of equipment with an unbalance is present in
the system being tested.
• The modal vectors are primarily coherent noise. Since the reference modal vector
may be arbitrarily chosen, this modal vector may not be one of the true modal
vectors of the system. It could simply be a random noise vector or a vector
reflecting the bias in the modal parameter estimation algorithm. In any case, the
modal assurance criterion will only reflect a consistent (linear) relationship to the
reference modal vector.
662 R. J. Allemang and A. W. Phillips

• The modal vectors represent the same modal vector with different arbitrary
scaling. If the two modal vectors being compared have the same expected value
when normalized, the two modal vectors should differ only by the complex-
valued scale factor which is a function of the common modal coefficients between
the rows or columns.

Therefore, if the first three reasons can be eliminated, the modal assurance criterion
indicates that the modal scale factor is the complex constant relating the modal
vectors and that the modal scale factor can be used to average, difference, or sort the
modal vectors.
Under the constraints mentioned previously, the modal assurance criterion can
be applied in many different ways. The modal assurance criterion can be used to
verify or correlate an experimental modal vector with respect to a theoretical modal
vector (eigenvector). This can be done by computing the modal assurance criterion
between Ne modal vectors estimated from experimental data and Na modal vectors
estimated from a finite element analysis evaluated at common stations. This process
results in a Ne × Na rectangular modal assurance criterion matrix with values that
approach unity whenever an experimental modal vector and an analytical modal
vector are consistently related.
Once the modal assurance criterion establishes that two vectors represent the
same information, the vectors can be averaged, differenced, or sorted to determine
the best single estimate or the potential source of contamination using the modal
scale factor. Since the modal scale factor is a complex scalar that allows two vectors
to be phased the same and normalized to the same mean value, these vectors can be
subtracted to evaluate whether the error is random or biased. If the error appears
to be random and the modal assurance criterion is high, the modal vectors can
be averaged, using the modal scale factor, to improve the estimate of a modal
vector. If the error appears to be biased or skewed, the error pattern often gives an
indication that the error originates due to the location of the excitation or due to an
inadequate modal parameter estimation process. Based upon partial but overlapping
measurement of two columns of the frequency response function matrix, modal
vectors can be sorted, assuming the modal assurance function indicates consistency,
into a complete estimate of each modal vector at all measurement stations.
The modal assurance criterion can be used to evaluate modal parameter estima-
tion methods if a set of analytical frequency response functions with realistic levels
of random and bias errors is generated and used in common to a variety of modal
parameter estimation methods. In this way, agreement between existing methods
can be established, and new modal parameter estimation methods can be checked
for characteristics that are consistent with accepted procedures. Additionally, this
approach can be used to evaluate the characteristics of each modal parameter
estimation method in the presence of varying levels of random and bias error.
The concept of consistency in the estimate of modal vectors from separate testing
constraints is important considering the potential of multiple estimates of the same
modal vector from numerous input configurations and modal parameter estimation
algorithms. The computation of the modal scale factor and modal assurance criterion
11 Experimental Modal Parameter Evaluation Methods 663

results in a complex scalar and a correlation coefficient which does not depend
on weighting information outside the testing environment. Since the modal scale
factor and modal assurance criterion are computed analogous to the frequency
response function and coherence function, both the advantages and limitations of the
computation procedure are well understood. These characteristics, as well as others,
provide a useful tool in the processing of experimental modal vectors. Typical
presentations of the modal assurance criterion are shown in Figs. 23 and 24.

Auto−MAC of Reference File

1
2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.9

2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00
0.8
2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
0.7
2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00
Mode Frequency, Hz

1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.6

MAC Value
1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
0.5
1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.4

764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.3
761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.2
557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15

363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00 0.1

362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00
0
362.38 363.68 557.05 761.39 764.04 1222.99 1224.07 1328.05 1328.77 2019.26 2023.80 2321.99 2322.54 2337.97
Mode Frequency, Hz

Fig. 23 Modal assurance criterion

Auto−MAC of Reference File

1
2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00
0.9
2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00
1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.8
1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.7
764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Mode Frequency, Hz

761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15 0.6
MAC Value

363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00
−362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00 0.5
−363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
−557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15
−761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.4
−764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.3
−1327.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.25 0.00 0.00 0.00 0.00 0.02
−1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00
−1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.2
−2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00
−2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00
−2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.1
−2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
−2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
0
362.38 363.68 557.05 761.39 764.04 1222.99 1224.07 1328.05 1328.77 2019.26 2023.80 2321.99 2322.54 2337.97
Mode Frequency, Hz

Fig. 24 Modal assurance criterion – conjugate modal vectors

664 R. J. Allemang and A. W. Phillips

Auto−MAC of Reference File

1
2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00
0.9
2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00
1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.8
1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.7
764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Mode Frequency, Hz

761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15 0.6

MAC Value
363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00
−362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00 0.5
−363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
−557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15
−761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.4
−764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.3
−1327.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.25 0.00 0.00 0.00 0.00 0.02
−1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00
−1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.2
−2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00
−2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00
−2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.1
−2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
−2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
0
362.38 363.68 557.05 761.39 764.04 1222.99 1224.07 1328.05 1328.77 2019.26 2023.80 2321.99 2322.54 2337.97
Mode Frequency, Hz

Fig. 25 Modal assurance criterion

Figure 24 shows a variation in the presentation to include both the comparison

between the modal vectors associated with the positive frequencies and themselves
and also with the modal vectors associated with the conjugate modal (negative)
frequencies. This can be useful if the conjugate modal parameters are estimated
independently. Users have found that a good MAC value between the positive
and negative frequency modal vectors is a quality check on the estimated modal
parameters. Note that the upper diagonal in Fig. 24 theoretically must have MAC
values equal to 1.0 but this is not the case for the lower diagonal comparing modal
vectors to their conjugate counterparts.
Note that Fig. 25 uses a different colorbar as compared to Figs. 23 and 24 to
visually downplay the similarity between different modal vectors.

4.5.2 Cross Modal Assurance Criterion (Cross MAC)

Rather than comparing a set of modal vectors to themselves to look for cross-
contamination between modal vectors, two different sets of modal vectors can be
compared. This is useful when comparing experimental results to analytical results,
or when comparing different experimental results. Unlike the case when comparing
a modal vector set to itself, the cross MAC does not necessarily yield unity along the
diagonals and may not find a common modal vector between the two modal vector
sets. This is common when comparing analytical modal vector sets to experimental
modal vector sets. This is very useful when the modal frequencies occur in different
numerical order between the analytical (e.g., FEA) and the experimental modal
vector sets of data.
Figure 26 is a cross MAC between two modal vectors sets generated with
different modal parameter estimation (MPE) algorithms, PTD on the x-axis and
11 Experimental Modal Parameter Evaluation Methods 665

Fig. 26 Modal assurance criterion

RFP-Z on the y-axis. It is clear in this cross MAC that the two different MPE
methods give nearly identical modal vector results.

4.5.3 Pole-Weighted or State Vector MAC

Since the modal assurance criterion (MAC) is based upon the sensor degrees of
freedom (DOFs), there can be insufficient spatial information to describe similar
modal vectors adequately. This can lead to a situation where the MAC value between
different modal vectors is higher than it should. This is referred to as spatial
aliasing. Spatial aliasing can be seen in each of Figs. 23, 24, 25, and 26.
This problem can be somewhat minimized by replacing the modal vector in the
MAC equation by its equivalent state vector of model order m. This is referred to as
a pole-weighted MAC or state vector MAC.

⎧ ⎫ ⎧ ⎫
⎪
⎪
⎪ r {ψ}r ⎪
λm ⎪
⎪
⎪
⎪
⎪
zrm {ψ}r ⎪
⎪
⎪
⎪
⎪ · ⎪
⎪ ⎪
⎪ · ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ · ⎪
⎪ ⎪
⎪ · ⎪
⎪
⎨ ⎬ ⎨ ⎬
{φ}r = · or {φ}r = · (46)
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ λr {ψ}r ⎪
⎪
2 ⎪ ⎪ zr2 {ψ}r ⎪
⎪ ⎪
⎪
⎪
⎪ 1 {ψ} ⎪
⎪ ⎪
⎪ 1 {ψ} ⎪
⎪
⎪
⎪ λ ⎪
r ⎪
⎪
⎪ z ⎪
r ⎪
⎪
⎩ 0 r ⎪
⎭ ⎪
⎩ 0 r ⎪
⎭
λr {ψ}r r zr {ψ}r r

The second form of the state vector given in Eq. 46 is preferred for numerical
reasons. The state vector can be visualized in the graphical plots shown in Figs. 27
and 28.
666 R. J. Allemang and A. W. Phillips

Fig. 27 Modal assurance

criterion

Fig. 28 State vector modal assurance criterion

11 Experimental Modal Parameter Evaluation Methods 667

Auto−MAC of Reference File

1
2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00
0.9
2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00
1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.8
1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.7
764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Mode Frequency, Hz

761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15 0.6

MAC Value
363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00
−362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.06 0.00 0.5
−363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.04 0.00
−557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15
−761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.4
−764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.3
−1327.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.25 0.00 0.00 0.00 0.00 0.02
−1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00
−1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.2
−2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00
−2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00
−2321.99 0.04 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.1
−2322.54 0.06 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
−2337.97 0.00 0.00 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
0
362.38 363.68 557.05 761.39 764.04 1222.99 1224.07 1328.05 1328.77 2019.26 2023.80 2321.99 2322.54 2337.97
Mode Frequency, Hz

Fig. 29 Modal assurance criterion (no weighting or pole weight = 1)

Auto−pwMAC of Reference File [pw=5]

1
2337.97 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00
2322.54 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00
2321.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00
0.9
2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00
1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.8
1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
1224.07 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1222.99 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.7
764.04 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Mode Frequency, Hz

761.39 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
557.05 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.6

MAC Value
363.68 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
362.38 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−362.38 0.10 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.5
−363.68 0.00 0.10 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−557.05 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−761.39 0.00 0.00 0.00 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.4
−764.04 0.00 0.00 0.00 0.00 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1222.99 0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1224.07 0.00 0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.3
−1327.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
−1328.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.00
−1328.77 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.2
−2019.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00
−2023.80 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00
−2321.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.1
−2322.54 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.03 0.00
−2337.97 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.03
0
362.38 363.68 557.05 761.39 764.04 1222.99 1224.07 1328.05 1328.77 2019.26 2023.80 2321.99 2322.54 2337.97
Mode Frequency, Hz

Fig. 30 State vector modal assurance criterion (pole weight = 5)

The MAC results compared to the state vector MAC results are shown in Figs. 29
and 30.

4.5.4 Other Similar Assurance Criteria

The following brief discussion highlights assurance criteria that utilize the same
linear, least squares computation approach to the analysis (projection) of two vector
spaces as the modal assurance criterion. The equations for each assurance criterion
668 R. J. Allemang and A. W. Phillips

are not repeated unless there is a significant computational difference that needs to
be clarified or highlighted. This list is by no means comprehensive, nor is it in any
particular order of importance but includes most of the frequently cited assurance
criterion found in the literature.

Weighted Modal Analysis Criterion (WMAC)

A number of authors have utilized a weighted modal assurance criterion (WMAC)
without developing a special designation for this case. WMAC is proposed for
these cases. The purpose of the weighting matrix is to recognize that MAC is not
sensitive to mass or stiffness distribution, just sensor distribution, and to adjust
the modal assurance criterion to weight the degrees of freedom in the modal
vectors accordingly. In this case, the WMAC becomes a unity normalized weighted
orthogonality, or weighted pseudo orthogonality, check where the desirable result
for a set of modal vectors would be ones along the diagonal (associated with
the same modal vectors) and zeros off diagonal (associated with different modal
vectors) regardless of the scaling of the individual modal vectors. Note that the
weighting matrix is normally found by first estimating the modal assurance criterion
and then using the MAC values to define the weighting matrix.

Partial Modal Analysis Criterion (PMAC)

The partial modal assurance criterion (PMAC) [59] is developed as a spatially
limited version of the modal assurance criterion where a subset of the complete
modal vector is used in the calculation. The subset is chosen based upon the user’s
interest and may reflect only a certain dominant sensor direction (X, Y, and/or Z) or
only the degrees of freedom from a component of the complete modal vector.

Modal Assurance Criterion Square Root (MACSR)

The square root of the modal assurance criterion (MACSR) [60] is developed to be
more consistent with the orthogonality and pseudo orthogonality calculations using
an identity weighting matrix. Essentially, this approach utilizes the square root of
the MAC calculation which tends to highlight the cross terms (off-diagonal) which
are generally the MAC values that are very small.

Scaled Modal Assurance Criterion (SMAC)

The scaled modal assurance criterion (SMAC) [61] is essentially a weighted modal
assurance criteria (WMAC) where the weighting matrix is chosen to balance the
scaling of translational and rotational degrees of freedom included in the modal
vectors. This development is needed whenever different data types (with different
engineering units) are included in the same modal vector to normalize the magnitude
differences in the vectors. This is required since the modal assurance criterion
minimizes the squared error and is dominated by the larger values.

Modal Assurance Criterion Using Reciprocal Vectors (MACRV)

A reciprocal modal vector is defined as the mathematical vector that, when
transposed and pre-multiplied times a specific modal vector, yields unity. When
11 Experimental Modal Parameter Evaluation Methods 669

the same computation is performed with this reciprocal modal vector and any other
modal vector or any other reciprocal modal vector, the result is zero. The reciprocal
modal vector can be thought of as a product of the modal vector and the unknown
weighting matrix that will produce a perfect orthogonality result. Reciprocal modal
vectors are computed directly from measured frequency response functions and the
experimental modal vectors and are, therefore, experimentally based. The modal
assurance criterion using reciprocal modal vectors (MACRV) [62] is the comparison
of reciprocal modal vectors with analytical modal vectors in what is very similar to
a pseudo orthogonality check (POC). The reciprocal modal vectors are utilized in
controls applications as modal filters, and the MACRV serves as a check of the
mode isolation provided by each reciprocal modal vector compared to analytical
modes expected.

Modal Assurance Criterion with Frequency Scales (FMAC)

Another extension of the modal assurance criterion is the addition of frequency
scaling to the modal assurance criterion [63, 64]. This extension of MAC offers
a means of displaying simultaneously the mode shape correlation, the degree of
spatial aliasing, and the frequency comparison in a single plot. This development
is particularly useful in model correlation applications (model updating, assessment
of parameter variation, etc.)

Coordinate Modal Assurance Criterion (COMAC)

An extension of the modal assurance criterion is the coordinate modal assurance
criterion (COMAC) [65]. The COMAC attempts to identify which measurement
degrees of freedom contribute negatively to a low value of MAC. The COMAC
is calculated over a set of mode pairs, analytical versus analytical, experimental
versus experimental, or experimental versus analytical. The two modal vectors
in each mode pair represents the same modal vector, but the set of mode pairs
represents all modes of interest in a given frequency range. For two sets of
modes that are to be compared, there will be a value of COMAC computed for
each (measurement) degree of freedom. The coordinate modal assurance criterion
(COMAC) is calculated using the following approach, once the mode pairs have
been identified with MAC or some other approach:

L 
 
ψqr φqr 2
r=1
COMACq = (47)

L
∗

L
∗
ψqr ψqr φqr φqr
r=1 r=1

Note that the above formulation assumes that there is a match for every modal
vector in the two sets and the modal vectors are renumbered accordingly so that
the matching modal vectors have the same subscript. Only those modes that match
between the two sets are included in the computation.
670 R. J. Allemang and A. W. Phillips

Enhanced Coordinate Modal Assurance Criterion (ECOMAC)

One common problem with experimental modal vectors is the potential problem
of calibration scaling errors and/or sensor orientation mistakes. The enhanced
coordinate modal assurance criterion (ECOMAC) [66] was developed to extend the
COMAC computation to be more aware of typical experimental errors that occur in
defining modal vectors such as sensor scaling mistakes and sensor orientation (plus
or minus sign) errors.

Mutual Correspondence Criterion (MCC)

The mutual correspondence criterion (MCC) [67] is the modal assurance criterion
applied to vectors that do not originate as modal vectors but as vector measures
of acoustic information (velocity, pressure, intensity, etc.). The equation in this
formulation utilizes a transpose and will only correctly apply to real-valued vectors.

Modal Correlation Coefficient (MCC)

One of the natural limitations of a least squares-based correlation coefficient like
the modal assurance criterion is that it is relatively insensitive to small changes in
magnitude, position by position, in the vector comparisons. The modal correlation
coefficient (MCC) [68, 69] is a modification of MAC that attempts to provide a
more sensitive indicator. This approach is particularly important when using modal
vectors in damage detection situations where the magnitude changes of the modal
vectors being measured are minimal.

Inverse Modal Assurance Criterion (IMAC)

An alternative approach to increasing the sensitivity of the modal assurance criterion
to small mode shape changes is the inverse modal assurance criterion (IMAC) [70].
This approach uses essentially the same computational scheme as MAC but utilizes
the inverse of the modal coefficients. Therefore, small modal coefficients become
significant in the least squares-based correlation coefficient computation. Naturally,
this computation suffers from the possibility that a modal coefficient could be
numerically zero.

Frequency Response Assurance Criterion (FRAC)

Any two frequency response functions representing the same input-output rela-
tionship can be compared using a technique known as the frequency response
assurance criterion (FRAC) [71]. The simplest example is a validation procedure
that compares the FRF data synthesized from the modal model with the measured
FRF data. The basic assumption is that the measured frequency response function
and the synthesized frequency response function should be linearly related (unity
scaling coefficient) at all frequencies. Naturally, the FRFs can be compared over the
full or partial frequency range of the FRFs as long as the same discrete frequencies
are used in the comparison. This approach has been utilized in the modal parameter
estimation process for a number of years under various designations (parameter
estimation correlation coefficient (PMAC) [59], synthesis correlation coefficient
(SCC) [2,72], and response vector assurance criterion (RVAC) [73]). This procedure
11 Experimental Modal Parameter Evaluation Methods 671

is particularly effective as a modal parameter estimation validation procedure if the

measured data was not part of the data used to estimate the modal parameters. This
serves as an independent check of the modal parameter estimation process.
 2
  
 ω2 ∗ (ω)
 Hpq (ω)Ĥpq 
ω=ω1 
F RACpq = (48)

ω2 
ω2
Hpq (ω)Hpq ∗ (ω) ∗ (ω)
Ĥpq (ω)Ĥpq
ω=ω1 ω=ω1

Complex Correlation Coefficient (CCF)

A significant variation in the frequency response assurance criterion is the com-
plex correlation coefficient (CCF) [73] which is computed without squaring the
numerator term, thus, yielding a complex-valued coefficient. The magnitude of
the coefficient is the same as the FRAC computation, but the phase describes any
systematic phase lag or lead that is present between the two FRFs. In situations
where analytical and experimental FRFs are compared, the CCF will detect the
common problem of a constant phase shift that might be due to experimental signal
conditioning problems, etc.

Frequency Domain Assurance Criterion (FDAC)

A similar variation in the frequency response assurance criterion is the frequency
domain assurance criterion (FDAC) [74] which is a FRAC type of calculation
evaluated with different frequency shifts. Since the difference in impedance (FRF)
model updating is often an FRF that is in question due to frequencies of resonances
or anti-resonances, the FDAC is formulated to identify this problem. A related
criterion, the modal FRF assurance criterion (MFAC) [74], combines analytical
modal vectors with measured frequency response functions (FRFs) in an extension
of FRAC and FDAC that weights or filters the FRF data based upon the expected,
analytical modal vectors.

Coordinate Orthogonality Check (CORTHOG)

The coordinate orthogonality check (CORTHOG) [75] is a normalized error
measure between the pseudo orthogonality calculation, comparing measured to
analytical modal vectors, and the analytical orthogonality calculation, comparing
analytical to analytical modal vectors. Several different normalizing or scaling
methods are used with this calculation.

4.5.5 Uses of the Modal Assurance Criterion

Most of the potential uses of the modal assurance criterion are well-known, but a few
may be more subtle. A partial list of the most typical uses that have been reported
in the literature is as follows:

• Validation of experimental modal models

• Correlation with analytical modal models (mode pairing)
672 R. J. Allemang and A. W. Phillips

• Correlation with operating response vectors

• Mapping matrix between analytical and experimental modal models
• Modal vector error analysis
• Modal vector averaging
• Experimental modal vector completion and/or expansion
• Weighting for model updating algorithms
• Modal vector consistency/stability in modal parameter estimation algorithms
• Repeated and pseudo repeated root detection
• Structural fault/damage detection
• Quality control evaluations
• Optimal sensor placement

4.5.6 Misuse/Abuse of the Modal Assurance Criterion

Many of the alternate formulations of the modal assurance criterion were developed
to address some of the shortcomings of the original modal assurance criterion
formulation. When users utilize the original modal assurance criterion in these
situations, a poor result will often follow. For the purposes of this discussion, this
is referred to as misuse or abuse. The misuse or abuse of the modal assurance
criterion generally results due to one of the five issues. These issues can be
summarized as:

• The modal analysis criterion is not an orthogonality check.

• The wrong mathematical formulation for the modal assurance criterion
is used.
• The modal assurance criterion is sensitive to large values (wild points?) and
insensitive to small values.
• The number of elements in the modal vectors (space) is small.
• The modal vectors have been zero padded at unmeasured DOFs.

5 Autonomous Modal Parameter Estimation

Autonomous modal parameter estimation often involves sorting a large number

of possible solutions to develop one consistent estimate of the modal parameters
(frequency, damping, modal vector, and modal scaling). Once the final, consistent
estimate of modal parameters is established, this estimate can be compared to related
solutions from the larger set of solutions to develop statistical attributes for the final,
consistent set of modal parameters. These attributes include sample size, standard
deviation, and other familiar variance estimates. New variance estimates can be
introduced to categorize the modal vector solution. These modal vector statistics
are based upon the residual contributions in a set of correlated modal vectors that
are used to estimate a single modal vector.
11 Experimental Modal Parameter Evaluation Methods 673

5.1 Current Approaches

The interest in automatic modal parameter estimation methods has been documented
in the literature since at least the mid-1960s when the primary modal method was
the analog, force appropriation method [76, 77, 78]. Following that early work,
there has been a continuing interest in autonomous methods [79, 80, 81, 82, 83,
84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97] that, in most cases, have been
procedures that are formulated based upon a specific modal parameter estimation
algorithm like the Eigensystem realization algorithm (ERA), the polyreference
time domain (PTD) algorithm or more recently the polyreference least squares
complex frequency (PLSCF) algorithm, or the commercial version of the PLSCF,
the PolyMAX method.
Each of these past procedures has shown some promise but has not yet been
widely adopted. In many cases, the procedure focused on a single modal parameter
estimation algorithm and did not develop a general procedure. Most of the past
procedural methods focused on pole density but depended on limited modal vector
data to identify correlated solutions. Many of these current and past methods
are commercially developed, and, due to proprietary concerns, details of the
autonomous procedure may not be provided.
Currently, due to increased computational speed and larger availability of
memory, procedural methods can be developed that were beyond the computational
scope of available hardware only a few years ago. These methods do not require any
initial manual evaluation of the solution sets and rely upon correlation of the vector
space of thousands of potential solutions as the primary identification tool. With
the addition to any modal parameter estimation algorithm of the concept of pole-
weighted state vector, the length, and therefore sensitivity, of the extended vectors
provides an additional tool that appears to be very useful.
The larger question concerning autonomous modal parameter estimation is the
intended user. Is the autonomous modal parameter estimation procedure expected
to give results sufficiently robust for the novice user? This implies that the user
could have no experience with modal analysis and, therefore, have no experiential
judgment to use in assessing the quality of the results. The use of the term wizard is
very common in the commercial autonomous methods, and this description implies
that this is the desired situation. In contrast, the user could be very knowledgeable
in the theory and experienced in the practice of extracting modal parameters from
experimental data. For this case, the autonomous modal parameter estimation
procedure is simply an efficient mechanism for sorting a very large number of
solutions into a final set of solutions that satisfies a set of criteria and thresholds
that are acceptable to the user. Statistical parameters are often part of the solution of
these autonomous methods so that users can make judgments concerning the quality
of the results. Such statistical parameters are designed to give all users additional
information that will allow both experienced and novice users to successfully
identify the modal parameters, within the limits of the information provided by the
measured data.
674 R. J. Allemang and A. W. Phillips

5.2 Common Statistical Subspace Autonomous Mode

Identification (CSSAMI)

One autonomous modal parameter estimation (MPE) method that is somewhat

general and can be applied to any MPE method that generates a consistency
diagram or something similar is the Common Statistical Subspace Autonomous
Mode Identification (CSSAMI) method [47, 98, 99]. The reader is directed to these
references in order to get an overview of one methodology and to view application
results for several cases using that methodology.
As a brief summary, the CSSAMI methods begin by forming an m-th order state
vector associated with every solution represented in consistency diagram. The user
does not need to define which solutions to use. Note that a number of MPE methods
can be included in one consistency diagram using techniques noted in [100, 101].
Thus, the number of possible modal parameter solution sets can be in the thousands.
All of the state vectors are combined in a single matrix, and a state vector (pole-
weighted) MAC is computed between each state vector and all others. A threshold
of some value for the state vector MAC is used to find groups of solution sets
that represent the same modal vector. Note that this approach sorts all common
state vectors together with separate groups of solutions for modal frequencies and
conjugate modal frequencies.
Each group of state vectors that represents a common state vector, determined
by the MAC threshold, is then subjected to a singular value decomposition (SVD)
to find the dominant state vector characteristic for the group as well as checking
statistics to determine if more than one state vector is present. The zeroth-order
portion of the dominant state vector, compared to the first-order portion of the
dominant state vector, is then used to determine the modal frequency and the modal
vector. Statistics are computed and assigned to each state vector group to monitor
how well the autonomous MPE procedure meets the criteria. Criteria are represented
by a small number of thresholds at different points in the procedure.
During the development of any new autonomous modal parameter estimation
procedure, it becomes obvious that, since a large number of possible solutions
are being evaluated, this development was a natural way to introduce statistical
evaluations into the modal analysis estimation process.
Note that much of the background of the CSSAMI method is based upon the
unified matrix polynomial algorithm (UMPA) described in a number of papers
[3, 4, 5, 6, 102, 101]. However, any modal parameter estimation method can utilize
the process outlined in the CSSAMI method.

6 Summary

Modal parameter estimation is probably one of the most misunderstood aspects of

the experimental modal analysis process. The methods reviewed in this chapter
serve as tools for estimating valid modal parameters from the various modal
11 Experimental Modal Parameter Evaluation Methods 675

parameter estimation (MPE) algorithms available. These MPE algorithms were

summarized in  Chap. 10, “Experimental Modal Analysis Methods” in this
handbook. Since most modal parameter estimation methods are mathematically
intimidating, many users do not fully understand the ramifications of the decisions
made during the measurement stages as well as later in the modal parameter
estimation process. Tools presented in this chapter aid the user in validating the
modal parameters estimated from the different algorithms. Much of the basis for the
newer autonomous MPE methods utilizes the tools presented in this chapter.
Over the last 40 years, the modal assurance criterion and the consistency diagram
have been the primary tools used by most users and are available in all modal
parameter estimation software. The modal assurance criterion has demonstrated
how a simple statistical concept can become an extremely useful tool in the field
of experimental modal analysis and structural dynamics. The use of the modal
assurance criterion (MAC), and the development and use of a significant number of
related criteria, has been remarkable and is most likely due to the overall simplicity
of the concept. The consistency diagram was developed around the same time as the
modal assurance criterion and is widely used, providing a mechanism for evaluating
very large sets of modal parameter estimates. The consistency diagram is also very
simple and is just a repeated solution procedure over a range of modal parameter
algorithm model orders.
Certainly, in the next few years, the increased use of other statistical methods
and further development of singular value/vector decomposition methods are related
technical areas that will generate useful tools in this area. Currently, many users
are utilizing more statistical approaches to understand the meaning and bounds
of experimental modal parameters [45, 103, 104, 105, 106, 107]. This approach
extends again to the modal assurance criterion as well as to the consistency
diagram. Examples are the bootstrap and jackknife approaches [105,106,107] to the
evaluation of the mean and standard deviation of discrete sets of experimental data.
These approaches remove and/or replace portions of the computation (bootstrap
uses replicative resampling; jackknife uses sequential elimination) to evaluate
the bounds or limits on the MAC values. In this way, the sensitivity of any
statistical computation can be more effectively evaluated than with the current single
number indicating the degree of linearity between two modal vectors that are being
compared.

References
1. Leuridan J, Brown D, Allemang R (1985) Time domain parameter identification methods for
linear modal analysis: a unifying approach. ASME Paper Number 85-DET-90
2. Allemang RJ, Brown, D (1987) Volume III: modal parameter estimation. USAF Report:
AFWAL-TR-87-3069. Experimental modal analysis and dynamic component synthesis, no 3,
pp 130
3. Allemang R, Brown, D, Fladung W (1994) Modal parameter estimation: a unified matrix
polynomial approach. In: Proceedings, International Modal Analysis Conference (IMAC),
pp 501–514
676 R. J. Allemang and A. W. Phillips

4. Allemang R, Brown D (1996) A unified matrix polynomial approach to modal identification,

Allied Publishers Limited pp 379–390
5. Allemang R, Brown D (1998) A unified matrix polynomial approach to modal identification.
J Sound Vib 211(3):301–322
6. Allemang, R, Phillips A (2004) The unified matrix polynomial approach to understanding
modal parameter estimation: an update. In: Proceedings, International Seminar on Modal
Analysis (ISMA)
7. Spitznogle F (1971) Improvements in the complex exponential signal analysis computational
algorithm. Report Number U1-829401-5, Representation and Analysis of Sonar Signals,
vol 1, pp 37
8. Brown D, Allemang R, Zimmerman, R, Mergeay M (1979) Parameter estimation tech-
niques for modal analysis. SAE Paper Number 790221, SAE Transactions, vol 88,
pp 828–846
9. Vold, H, Rocklin T (1982) The numerical implementation of a multi-input modal estimation
algorithm for mini-computers. In: Proceedings, International Modal Analysis Conference
(IMAC), pp 542–548
10. Vold H, Kundrat J, Rocklin, T, Russell R (1982) A multi-input modal estimation algorithm
for mini-computers. SAE Trans 91(1):815–821
11. Ibrahim, S, Mikulcik E (1977) A method for the direct identification of vibration parameters
from the free response. Shock Vib Bull 47:183–198
12. Pappa R (1982) Some statistical performance characteristics of the ITD modal identification
algorithm. AIAA Paper Number 82-0768, p 19
13. Fukuzono K (1986) Investigation of multiple-reference ibrahim time domain modal parameter
estimation technique. M. S. Thesis, p 220
14. Juang, J, Pappa R (1985) An eigensystem realization algorithm for modal parameter
identification and model reduction. AIAA J Guid Control Dyn 8(4):620–627
15. Juang J (1987) Mathematical correlation of modal parameter identification methods via
system realization theory. J Anal Exp Modal Anal 2(1):1–18
16. Longman RW, Juang J-N (1989) Recursive form of the eigensystem realization algorithm for
system identification. AIAA J Guid Control Dyn 12(5):647–652
17. Zhang L, Kanda H, Brown, D, Allemang R (1985) A polyreference frequency domain method
for modal parameter identification. ASME Paper No. 85-DET-106, vol 8, pp 8+
18. Lembregts F, Leuridan J, Zhang, L, Kanda H (1986) Multiple input modal analysis of
frequency response functions based on direct parameter identification. In: Proceedings,
International Modal Analysis Conference (IMAC)
19. Lembregts F (1988) Frequency domain identification techniques for experimental multiple
input modal analysis. Doctoral Dissertation, Katholieke University of Leuven, Belgium,
vol 213
20. Lembregts F, Leuridan, J, Van Brussel H (1989) Frequency domain direct parameter identifi-
cation for modal analysis: state space formulation. Mech Syst Signal Process 4(1):65–76
21. Coppolino R (1981) A simultaneous frequency domain technique for estimation of modal
parameters from measured data. SAE Paper No. 811046, vol 12
22. Craig R, Kurdila A, Kim HM (1990) State-space formulation of multi-shaker modal analysis.
J Anal Exp Modal Anal 5(3):169–183
23. Leuridan J (1984) Some direct parameter model identification methods applicable for multiple
input modal analysis. Doctoral Dissertation, vol 384, pp 384+
24. Natke H (1988) Updating computational models in the frequency domain based on measured
data: a survey. Probab Eng Mech 3(1):28–35
25. Richardson, M, Formenti D (1982) Parameter estimation from frequency response measure-
ments using rational fraction polynomials. In: Proceedings, International Modal Analysis
Conference (IMAC), pp 167–182
26. Shih C, Tsuei Y, Allemang, R, Brown D (1988) A frequency domain global parameter
estimation method for multiple reference frequency response measurements. Mech Syst
Signal Process (MSSP) 2(4):349–365
11 Experimental Modal Parameter Evaluation Methods 677

27. Shih C (1989) Investigation of numerical conditioning in the frequency domain modal
parameter estimation methods, Doctoral Dissertation, vol 127, pp 127+
28. Van der Auweraer, H, Leuridan J (1987) Multiple input orthogonal polynomial parameter
estimation. Mech Syst Signal Process (MSSP) 1(3):259–272
29. Van der Auweraer H, Snoeys R, Leuridan JM (1986) A Global frequency domain modal
parameter estimation technique for mini-computers. ASME J Vib Acoust Stress Reliab
Des 10, vol 11
30. Vold H (1986) Orthogonal polynomials in the polyreference method. In: Proceedings,
International Seminar on Modal Analysis (ISMA-11)
31. Vold, H, Shih CY (1988) Numerical sensitivity of the characteristic polynomial. In: Proceed-
ings, International Seminar on Modal Analysis (ISMA-13)
32. Vold H (1990) Numerically robust frequency domain modal parameter estimation. Sound Vib
Mag (SVM) 24(1):38–40
33. Vold H (1990) Statistics of the characteristic polynomial in modal analysis. In: Proceedings,
International Seminar on Modal Analysis (ISMA), pp 53–57
34. Vold H, Napolitano K, Hensley, D, Richardson M (2008) Aliasing in modal parameter
estimation, an historical look and new innovations. In: Proceedings, International Modal
Analysis Conference (IMAC), vol 16, pp 12–17
35. Fladung, W, Vold H (2016) An improved implementation of the orthogonal polynomial
modal parameter estimation algorithm using the orthogonal complement. In: Proceedings,
International Modal Analysis Conference (IMAC)
36. Fladung, W, Vold H (2016) An Orthogonal view of the polyreference least-squares complex
frequency modal parameter estimation algorithm. In: Proceedings, International Modal
Analysis Conference (IMAC)
37. Van der Auweraer H, Guillaume P, Verboven, P, Vanlanduit S (2001) Application of a fast-
stabilizing frequency domain parameter estimation method. ASME J Dyn Syst Meas Control
123(4):651–658
38. Guillaume P, Verboven P, Vanlanduit S, der Auweraer V H and Peeters B (2003) A
polyreference implementation of the least-squares complex frequency domain estimator. In:
Proceedings, International Modal Analysis Conference (IMAC), p 12
39. Verboven P, Guillaume P, Cauberghe B, Parloo, E, Vanlanduit S (2003) Stabilization charts
and uncertainty bounds for frequency domain linear least squares estimators. In: Proceedings,
International Modal Analysis Conference (IMAC), pp 1–10
40. Verboven P, Cauberghe B, Vanlanduit S, Parloo, E, Guillaume P (2004) The secret behind
clear stabilization diagrams: the influence of the parameter constraint on the stability of
the poles. In: Proceedings, Society of Experimental Mechanics (SEM) Annual Conference,
vol 17, pp 1–17
41. Verboven P (2002) Frequency domain system identification for modal analysis. Doctoral
Dissertation
42. Cauberghe B (2004) Application of frequency domain system identification for experimental
and operational modal analysis. Doctoral Dissertation, p 259
43. Shih C, Tsuei Y, Allemang, R, Brown D (1988) Complex mode indication function and
its application to spatial domain parameter estimation. Mech Syst Signal Process (MSSP)
2(4):367–377
44. Williams R, Crowley, J, Vold H (1985) The multivariable mode indicator function in modal
analysis. In: Proceedings, International Modal Analysis Conference (IMAC), pp 66–70
45. Phillips AW, Allemang RJ, Pickrel CR (1997) Clustering of modal frequency estimates from
different solution sets. In: Proceedings, International Modal Analysis Conference (IMAC),
pp 1053–1063
46. Phillips, A, Allemang R (2014) Normalization of experimental modal vectors to remove
modal vector contamination. In: Proceedings, International Modal Analysis Conference
(IMAC), p 12
47. Phillips A, Allemang, R, Brown D (2011) Autonomous modal parameter estimation: method-
ology. In: Proceedings, International Modal Analysis Conference (IMAC), p 22
678 R. J. Allemang and A. W. Phillips

48. Allemang R (1980) Investigation of some multiple input/output frequency response function
experimental modal analysis techniques. Doctoral Dissertation, pp 141–214
49. Phillips, A, Allemang R (1998) The enhanced frequency response function (eFRF): scaling
and other issues. In: Proceedings, International Seminar on Modal Analysis (ISMA)
50. Guyan R (1965) Reduction of stiffness and mass matrices. AIAA J 3(2):380
51. Irons B (1965) Structural eigenvalue problems: elimination of unwanted variables. AIAA J
3(5):961–962
52. Downs B (1979) Accurate reduction of stiffness and mass matrices for vibration analysis and
a rationale for selecting master degrees of freedom. ASME Paper Number 79-DET-18, p 5
53. Sowers J (1978) Condensation of free body mass matrices using flexibility coefficients. AIAA
J 16(3):272–273
54. O’Callahan J, Avitabile, P, Riemer R (1989) System equivalent reduction expansion process
(SEREP). In: Proceedings, International Modal Analysis Conference (IMAC), pp 29–37
55. Kammer D (1987) Test analysis model development using an exact modal reduction. J Anal
Exp Modal Anal 2(4):174–179
56. Kammer D (1990) Sensor placement for on-orbit modal identification and correlation of large
space structures. In: Proceedings, American Control Conference, pp 2984–2990
57. Allemang, R, Brown D (1982) A correlation coefficient for modal vector analysis. In:
Proceedings, International Modal Analysis Conference (IMAC), pp 110–116
58. Allemang R (2002) The modal assurance criterion (MAC): twenty years of use and abuse. In:
Proceedings, International Modal Analysis Conference (IMAC), pp 397–405
59. Heylen W (1990) Extensions of the modal assurance criterion. J Vib Acoust (JVA)
112:468–472
60. O’Callahan J (1998) Correlation considerations – Part 4 (Modal vector correlation tech-
niques). In: Proceedings, International Modal Analysis Conference (IMAC), pp 197–206
61. Brechlin E, Bendel, K, Keiper W (1998) A New Scaled Modal Assurance Criterion for
Eigenmodes Containing Rotational Degrees of Freedom. In: Proceedings, International
Seminar on Modal Analysis (ISMA), vol ISMA23, pp 7
62. Wei J, Wang, W, Allemang R (1990) Model correlation and orthogonality criteria based on
reciprocal modal vectors. In: Proceedings, SAE Noise and Vibration Conference, pp 607–616
63. Fotsch, D, Ewins D (2000) Applications of MAC in the frequency domain. In: Proceedings,
International Modal Analysis Conference (IMAC), pp 1225–1231
64. Fotsch, D, Ewins D (2001) Further Applicatios of the FMAC. In: Proceedings, International
Modal Analysis Conference (IMAC), pp 635–639
65. Lieven, N, Ewins D (1988) Spatial correlation of mode shapes, the coordinate modal
assurance criterion (COMAC). In: Proceedings, International Modal Analysis Conference
(IMAC), pp 690–695
66. Hunt D (1992) Application of an enhanced coordinate modal assurance criterion (ECOMAC).
In: Proceedings, International Modal Analysis Conference (IMAC), pp 66–71
67. Milecek S (1994) The use of modal assurance criterion extended. In: Proceedings, Interna-
tional Modal Analysis Conference (IMAC), pp 363–369
68. Samman M (1996) A modal correlation coefficient (MCC) for detection of kinks in mode
shapes. ASME J Vib Acoust 118(2):271–272
69. Samman M (1997) Structural damage detection using the modal correlation coefficient
(MCC). In: Proceedings, International Modal Analysis Conference (IMAC), pp 627–630
70. Mitchell L (2001) Increasing the sensitivity of the modal assurance criteria to small
mode shape changes: the IMAC. In: Proceedings, International Modal Analysis Conference
(IMAC), pp 64–69
71. Heylen, W, Lammens S (1996) FRAC: a consistent way of comparing frequency response
functions. In: Proceedings, International Conference on Identification in Engineering,
Swansea, pp 48–57
72. Allemang R (1990) Vibrations: experimental modal analysis, UC-SDRL-CN-20-263-
663/664, pp 7-1–7-2
11 Experimental Modal Parameter Evaluation Methods 679

73. Van der Auweraer H, Iadevaia M, Emborg U, Gustavsson M, Tengzelius, U, Horlin N (1998)
Linking test and analysis results in the medium frequency range using principal field shapes.
In: Proceedings, International Seminar on Modal Analysis (ISMA), p 8
74. Pascual R, Golinval, J, Razeto M (1997) A Frequency domain correlation technique for model
correlation and updating. In: Proceedings, International Modal Analysis Conference (IMAC),
pp 587–592
75. Avitabile, P, Pechinsky F (1994) Coordinate orthogonality check (CORTHOG). In: Proceed-
ings, International Modal Analysis Conference (IMAC), pp 753–760
76. Hawkins F (1965) An automatic resonance testing technique for exciting normal modes of
comples structures. In: Symposium IUTAM, Recent Progress in the Mechanics of Linear
Vibrations, pp 37–41
77. Hawkins F J. (1967) GRAMPA – an automatic technique for exciting the principal modes of
vibration of complex structures. Royal Aircraft Establishment, vol RAE-TR-67-211
78. Taylor GA, Gaukroger, D, Skingle C (1967) MAMA – a semi-automatic technique for
exciting the principal modes of vibration of complex structures. Aeronautical Research
Council, vol ARC-R/M-3590, pp 20
79. Takahashi, K, Furusawa M (1987) Development of automatic modal analysis. In: Proceedings,
International Modal Analysis Conference (IMAC), pp 686–692
80. Yam Y, Bayard D, Hadaegh F, Mettler E, Milman M and Scheid R (1989) Autonomous
frequency domain identification: theory and experiment. NASA JPL Report JPL Publication
89-8, p 204
81. Lim T, Cabell, R, Silcox R (1996) On-line identification of modal parameters using artificial
neural networks. J Vib Acoust (JVA) 118(4):649–656
82. Pappa R, Woodard S E, Juang J (1997) A benchmark problem for development of autonomous
structural modal identification. In: Proceedings, International Modal Analysis Conference
(IMAC), pp 1071–1077
83. Pappa R, James, G, Zimmerman D (1998) Autonomous modal identification of the space
shuttle tail rudder. J Spacecr Rocket 35(2):163–169
84. Chhipwadia K, Zimmerman, D, James III G (1999) Evolving autonomous modal parameter
estimation. In: Proceedings, International Modal Analysis Conference (IMAC), pp 819–825
85. James III G, Zimmerman, D, Chhipwadia K (1999) Application of autonomous modal
identification to traditional and ambient data sets. In: Proceedings, International Modal
Analysis Conference (IMAC), pp 840–845
86. Vanlanduit S, Verboven P, Schoukens, J, Guillaume P (2001) An automatic frequency
domain modal parameter estimation algorithm. In: Proceedings, International Conference on
Structural System Identification, pp 637–646
87. Verboven P, Parloo E, Guillame, P, Overmeire M (2001) Autonomous modal parameter
estimation based on a statistical frequency domain maximum likelihood approach. In:
Proceedings, International Modal Analysis Conference (IMAC), pp 1511–1517
88. Verboven P, Parloo E, Guillaume, P, Van Overmeire M (2002) Autonomous structural health
monitoring, part i: modal parameter estimation and tracking. Mech Syst Signal Process
(MSSP) 16(4):637–657
89. Parloo E, Verboven P, Guillaume, P, Van Overmeire M (2002) Autonomous structural health
monitoring – Part II: vibration-based in-operation damage assessment. Mech Syst Signal
Process (MSSP) 16(4):659–675
90. Vanlanduit S, Verboven P, Guillaume, P, Schoukens J (2003) An automatic frequency domain
modal parameter estimation algorithm. J Sound Vib (JSV) 265:647–661
91. Lanslots J, Rodiers, B, Peeters B (2004) Automated Pole-Selection: Proof-of-Concept and
Validation. In: Proceedings, International Seminar on Modal Analysis (ISMA)
92. Mevel L, Sam, A, Goursat M (2004) Blind modal identification for large aircrafts. In:
Proceedings, International Modal Analysis Conference (IMAC), p 8
93. Lau J, Lanslots J, Peeters, B, der Auweraer V (2007) Automatic modal analysis: reality or
myth?. In: Proceedings, International Modal Analysis Conference (IMAC), p 10
680 R. J. Allemang and A. W. Phillips

94. Liu J, Ying H, Shen, S, Dong S (2007) The function of modal important index in autonomous
modal analysis. In: Proceedings, International Modal Analysis Conference (IMAC), p 6
95. Mohanty P, Reynolds, P, Pavic A (2007) Automatic interpretation of stability plots for analysis
of a non-stationary structure. In: Proceedings, International Modal Analysis Conference
(IMAC), p 7
96. Poncelet F, Kerschen, G, Golinval J (2008) Operational modal analysis using second-order
blind identification. In: Proceedings, International Modal Analysis Conference (IMAC), p 7
97. Rainieri C, Fabbrocino, G, Cosenza E (2009) Fully automated OMA: an opportunity for smart
SHM systems. In: Proceedings, International Modal Analysis Conference (IMAC), p 9
98. Allemang R, Phillips, A, Brown D (2011) Autonomous modal parameter estimation: statisti-
cal considerations. In: Proceedings, International Modal Analysis Conference (IMAC), p 17
99. Brown D, Allemang, R, Phillips A (2011) Autonomous modal parameter estimation: applica-
tion examples. In: Proceedings, International Modal Analysis Conference (IMAC), p 17
100. Phillips A, Allemang, R, Pickrel C (1998) Estimating modal parameters from different
solution sets. In: Proceedings, International Modal Analysis Conference (IMAC), p 10
101. Allemang, R, Phillips A (2014) Spatial Information in Autonomous Modal Parameter
Estimation, Shock and Vibration Journal: Special ICEDYN Issue, pp 1–18
102. Allemang R, Phillips, A, Brown D (2011) Combined state order and model order formulations
in the unified matrix polynomial method (UMPA). In: Proceedings, International Modal
Analysis Conference (IMAC)
103. DeClerck, J, Avitabile P (1996) Development of several new tools for modal pre-test
evaluation. In: Proceedings, International Modal Analysis Conference (IMAC), pp 1272–
1277
104. Cafeo J, Lust, R, Meireis U (2000) Uncertainty in Mode Shape Data and its Influence on
the Comparison of Test and Analysis Models. In: Proceedings, International Modal Analysis
Conference (IMAC), pp 349–355
105. Paez, T, Hunter N (1998) Fundamental concepts of the bootstrap for statistical analysis of
mechanical systems. Exp Tech 21(3):35–38
106. Hunter, N, Paez T (1998) Applications of the bootstrap for mechanical system analysis. Exp
Tech 21(4):34–37
107. Efron, B, Gong G (1983) A leisurely look at the bootstrap, the jackknife and cross-validation.
Am Stat 37(1):36–48
Damping of Materials and Structures
12
Lothar Gaul and André Schmidt

Contents
1 Classification and Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
1.2 The Notion of Damping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
1.3 Classification of Damping Phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
1.4 Notes on Modern, Computer-Based Analytical and Measurement Programs . . . . . 686
2 Damping of Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
2.1 Physical Phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
2.2 Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 689
2.3 Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
3 Damping of Assemblies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
3.1 From Material Description to Complete Homogeneous Component . . . . . . . . . . . . 712
3.2 Laminated Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
3.3 Damping in Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
3.4 Damping Due to Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 720
3.5 Damping by Displacement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
3.6 Assemblies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 740
4 Models for Damped Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
4.1 Basic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
4.2 Structures with a Finite Number of Degrees of Freedom . . . . . . . . . . . . . . . . . . . . . . 747
5 Experimental Techniques for the Determination of Damping Characteristics . . . . . . . . . . 757
5.1 Experimental Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757

Lothar Gaul: deceased.

L. Gaul
Institut für Nichtlineare Mechanik, Universität Stuttgart, Stuttgart, Germany
e-mail: [email protected]
A. Schmidt ()
Institut für Nichtlineare Mechanik, Universität Stuttgart, Stuttgart, Germany
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 681

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_19
682 L. Gaul and A. Schmidt

5.2 Experimental Techniques and Types of Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . 758

5.3 Special Experimental Techniques for Determining Damping
Under Difficult Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
5.4 Experimental Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
5.5 Experimental Techniques for Measuring Soil Damping . . . . . . . . . . . . . . . . . . . . . . 778
6 Application of Fractional Calculus to Viscoelastically Damped
Structures in the Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779
6.1 Grünwald Definition of Fractional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779
6.2 Numerical Calculation of Fractional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
6.3 Fractional-Order Constitutive Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785
6.4 Finite Element Formulation and Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785
6.5 Parameter Identification: A Case Study with DelrinT M . . . . . . . . . . . . . . . . . . . . . 787
6.6 Finite Element Calculations and Comparison of the Different Concepts . . . . . . . . . 788
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 790
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793

Abstract

Damping is a phenomenon that can be observed in connection with all kind

of materials: solid, liquid, or gaseous. Any kind of time-dependent change
in stresses or strains of the material results in a loss of mechanical energy,
which in most cases is transformed into thermal energy. However, all the other
mechanisms such as the conversion into electrical energy or any kind of radiation
over the system’s boundaries play a role. Typical observations that can be made in
connection with damping are the occurrence of creep and relaxation processes or
hysteresis curves in the case of cyclic loadings. The overall damping is influenced
by a variety of mechanisms, especially for structures assembled from different
components.
No matter whether the presence of damping is sought or should be avoided in
technical applications, for any kind of tuning or optimization of a system under
consideration, a basic understanding of the underlying physics is needed. This is
especially true if calculations or simulations have to be run in order to predict the
dynamical behavior of a system.
This chapter intends to introduce the reader into the subject and provide an
extensive overview on the different aspects of damping regarding the fundamen-
tals, mathematical, and numerical models as well as experimental techniques
for the detection of damping properties. It shall give an overview of the state
of knowledge and experience gathered in various fields of application and
research. For further information, the reader is referred to various publications
and textbooks whenever needed. This chapter is organized as follows: Sect.
1 provides an extensive overview on the topic, the classification of damping
phenomena, and some remarks on computer-based programs. Section 2 refers
to the damping of solids, while Sect. 3 extends the view on structures assem-
bled from different components. Section 4 deals with different mathematical
models toward the description of damping and relevant numerical approaches.
Experimental techniques for the detection of the damping parameters needed for
calculations are described in Sect. 5. This includes possible instrumentation as
well as analytical methods. Finally, in Sect. 6, an application of the whole subject
12 Damping of Materials and Structures 683

covering the detection of damping properties, its mathematical representation,

and parameter identification along with a numerical simulation is presented as an
example. Conclusions from this chapter are drawn in Sect. 7.

Keywords

Damping · Creep · Relaxation · Hysteresis · Dissipation · Experimental

techniques · Joints · Assemblies · Finite element method · Boundary element
method · Rheological models · Fractional derivatives

Nomenclature

α Order of derivative
β Order of derivative
 Matrix of eigenvectors
B Spatial derivatives of matrix of shape functions
H Matrix of shape functions
K Stiffness matrix
M Mass matrix
u Relative displacement vector
χ Material loss factor
γ̇ Shear speed
γ Pure shear distortion
κ Curvature
λ Wavelength
ν Poisson’s ratio, order of derivative
ω0 Natural frequency
ωd Natural damped frequency
(t) shift function
ρ Mass density
σ Normal stress
τ shear stress
σ Complex stress
ε Complex strain
E Complex elastic modulus
G Complex shear modulus
K Complex bulk modulus
ε Strain
 Angular frequency
ϑ damping ratio
ζ Angular phase difference
E Young’s modulus, spring constant
e Deviatoric strain
E(t) Relaxation function
E Storage modulus
684 L. Gaul and A. Schmidt

E  Loss modulus
F Force
Hk Coulomb element
i Unit imaginary number
J (t) Creep compliance function
K(t) Decay function
P Active power
p Spring-pot coefficient
Rk damping constant
So Sommerfeld number
T Temperature, time interval
t Time
U Potential energy
V Volume
W Mechanical work
WD dissipated mechanical energy

1 Classification and Survey

1.1 Introduction

All dynamic processes in mechanic systems are more or less damped. Consequently,
damping is highly relevant in those fields of technology and applied physics which
deal with dynamics and vibrations. These include

• machine-, building-, and structural dynamics,

• system dynamics,
• control engineering, and
• technical acoustics,

because damping in these cases often has a considerable effect on the time history,
intensity, or even the existence of vibrations. Important applications are:

• transient vibrations (transient effects associated with the onset or decay of

vibrations, shock-induced vibrations, reverberation effects)
• resonance vibrations (unavoidable with random excitation)
• wave propagation
• dynamic-stability problems

Accordingly, a multitude of scientific publications dealing with damping, or taking

it into account at least, are found in technical literature. Due to different theory
approaches, objects, and task definitions in the applications listed above, the
designations, the characterization of damping, the experimental techniques, and the
analytical and numerical methods are not harmonized.
12 Damping of Materials and Structures 685

The dynamic behavior of damped structures can, in special cases, be calculated

using generally valid material laws for inelastic materials based on continuum
mechanics taking into account boundary effects (e.g., joints). In general, this
approach is too elaborate or expensive or not at all practicable. In most cases,
therefore, phenomenological equivalent systems or mathematical models tailored
to the task definition are used which are only valid assuming a special state
of stresses and/or a special time history. Harmonic (sinusoidal) time histories
are a preferred special case where complex quantities describe the elastic and
damping properties. These depend on a number of parameters: material data, rate
of deformation, frequency, temperature, number of load cycles, etc. In the case of
nonlinear behavior, there is also a dependence on the amplitude.
For certain problems, it is sufficient to state, for one deformation cycle, the
energy dissipated in a unit volume or within the system, or the energy released into
the environment at the system boundaries, often related to a conveniently chosen
elastic energy in a unit volume or in the system as a whole. However, in structural
dynamics, the use of modal damping ratios has proven useful, which do no longer
contain detailed information about the damping.

1.2 The Notion of Damping

Damping in mechanical systems is understood to be the irreversible transition

of mechanical energy into other forms of energy as found in time-dependent
processes. Damping is mostly associated with the change of mechanical energy into
thermal energy. Damping can also be caused by releasing energy into a surrounding
medium. Electromagnetic and piezoelectric energy conversion can also give rise to
damping if the energy converted is not returned to the mechanical system.

1.3 Classification of Damping Phenomena

The physical causes of damping are multifarious. In addition to friction, wave

propagation or flow effects, other possible causes are phase transitions in mate-
rials or energy conversion by piezoelectric, magnetostrictive or electromechanical
processes.
Forces associated with damping are non-conservative. They can be internal or
external forces. If both action and reaction forces in a free body diagram are effective
within the system boundaries, the effect is said to be an internal damping effect.
Whereas if the reaction force is effective outside the system boundaries, the effect
is an external damping effect.
Examples of internal damping are:

• material damping due to nonelastic material behavior

• friction between components, e.g., in slide ways, gears, etc.
686 L. Gaul and A. Schmidt

• conversion of mechanical vibration energy into electrical energy by means of the

piezoelectric effect and dissipation due to dielectric losses

Examples of external damping are:

• friction against the surrounding medium

• airborne-sound radiation into the environment
• structureborne-sound radiation into the ground

Phenomenological contributions of damping to a mechanical system:

• Material damping
The energy dissipation within a material, due to deformation and/or displace-
ment, is called material damping. Its physical causes are, in essence:
For solids:
– heat flows induced by deformation (thermomechanical coupling)
– slip effects
– microplastic deformations
– diffusion processes
For fluids:
– viscos flow losses
• Contact-surface damping: relative motion, friction
Contact-surface damping is caused by relative motions in the contact surfaces
of joined components such as screwed, riveted, and clamped joints. The physical
causes are:
– friction due to relative motions in the contact surface
– pumping losses in the enclosed medium due to relative motion in a direction
normal to the contact surface (e.g., gas pumping)

Structural damping refers to the following cases:

• damping in guides This includes energy dissipation in longitudinal guides (e.g.,

slides) and circular guides (e.g., journal bearings)
• electromechanical damping. Electromechanical damping can be caused by piezo-
electric, magnetostrictive, or electromagnetic effects
• energy release to the surrounding medium. This includes causes such as air, fluid,
and bedding damping.

1.4 Notes on Modern, Computer-Based Analytical and

Measurement Programs

Whereas the mass and stiffness matrices of relatively complex structures can be
readily determined nowadays using three-dimensional CAD drawings, automatic
12 Damping of Materials and Structures 687

grid generation, and subsequent FEM analysis, an appropriate calculation model

cannot usually be established with sufficiently precise information on damping.
More precise damping parameters can be determined experimentally.
“Experimental Modal Analysis” (EMA) has become established as the suitable
tool worldwide. It uses measured frequency-response curves between appropriately
chosen excitation points and measuring points and modern curve-fitting techniques
for identifying the modal parameters: natural frequencies, eigenmodes, and modal
damping ratios. In the case of simple structures, the system can be excited by
means of a hammer impact. In the case of complex components and considerable
damping, excitation using one or several exciters has proven convenient, allowing
to control exciter amplitudes and energy distribution for selected frequency ranges.
The system response is often measured by means of piezoelectric accelerometers or
laser-optical sensors.
Modern measurement and analytical systems offer the possibility to identify
discrete damping couplings provided that the substructures have been separately
investigated beforehand.
Link modules allow to establish the connection between the results of experimen-
tal modal analysis and the calculated FEM analysis (e.g., matching of nodal points
and coordinate axes through interpolation).
Quality criteria such as MAC (Modal Assurance Criterion) compare the relations
(such as orthogonality) between the eigenmodes found in terms of the scalar product
of the eigenvectors. Additional normalization using the mass or stiffness matrix
allows a quantitative assessment.
After model updating on the modal level, including damping ratios determined
by experiment, operation vibrations can be calculated for any load function. The
simulation model which was developed step by step can thus be verified under
practical conditions.

2 Damping of Solids

2.1 Physical Phenomena

It is observed experimentally that solid materials do not exhibit behavior as

predicted by a model of ideal elasticity. Such materials are therefore referred
to as inelastic. The two standard tests depicted in Fig. 1 illustrate the typical
characteristics of inelastic material behavior. When the material is subjected to a
step-like change in stress as shown in the following Eq. (1), it responds by time-
dependent creep, ε(t). If the strain, on the other hand, is kept constant after a
step-like change in strain shown in Eq. (2), the stress, σ (t), will decrease due to
reordering at the atomic, molecular, crystalline, and intercrystalline levels. This is
known as recovery or relaxation.
688 L. Gaul and A. Schmidt

Fig. 1 Creep and relaxation behavior (source: VDI guideline 3830)

σ (t) = σ0 1(t) (1)

ε(t) = ε0 1(t) (2)

In case of harmonic stressing of the material, inelastic behavior manifests itself

in the form of a loss of usable mechanical energy (hysteresis). Figure 2 shows two
such hysteresis cycles.
The mechanical energy dissipated due to damping effects per unit volume during
one cycle of period Tp and also in the case of arbitrary time history, which affects
the damping can be expressed as follows:
  Tp
WD = σ dε = σ (t)ε̇(t)dt. (3)

The material loss factor is a relative damping characteristic defined by:

WDh
χ= . (4)
2π Uref

where WDh is the damping work per unit volume per cycle in case of a harmonic
strain curve, and Uref is the reference energy per unit volume at maximum strain
(ε̂) shown as an area in Fig. 2. In the linear case, Fig. 2, Uref is the maximum stored
potential energy Umax per unit volume. The material loss factor depends to a great
extent on the amplitude and the time period, i.e., χ : χ = χ (Tp , ε̂)
For a component, the loss factor is defined by analogy to Eq. (4):

WSh
χS = . (5)
2π US
12 Damping of Materials and Structures 689

Fig. 2 Stress-Strain curves, cyclic loading, hysteresis loops for linear and nonlinear materials
(source: VDI guideline 3830)

The straightforward description of the characteristic damping work per cycle has
the following shortcomings:

• The characteristic can only be given for harmonic stresses or distortions and only
for hysteresis curves without sub-loops.
• The damping work does not contain any information on the rigidity of the
material.

2.2 Linear Models

A viscoelastic model is the easiest manner in which a linear inelastic material

may be described. The behavior can be characterized by employing a network of
massless springs and dampers, known as rheological models. The most fundamental
configuration is a 3-parameter model; it allows to describe the results of the standard
tests as per Fig. 1. A better conformity to the curves can be achieved by using several
combinations of the said models. However, this leads to difficulties when model
parameters are required to be associated with experimental behavior.

2.2.1 Three-Parameter Models

Figure 3 shows two possible configurations of springs and dampers in a 3-parameter
model for solids. Other spring and damper configurations in 3-parameter models
describe the behavior of fluids.
For model (a) in Fig. 3, E0a , E1a , and R1a are the parameters of the springs and
the damper, ε is the total strain, and ε1a is the spring excursion. For equilibrium, we
have:

σ = E0a ε + E1a ε1a (6)

E1a ε1a = R1a (ε̇ − ε̇1a ). (7)
690 L. Gaul and A. Schmidt

σ σ σA

E1a E0b
ε 1a ε − ε 1b l0
E0a
( 1 + ε ) l0
ε
R1a E1b R1b
ε − ε 1a ε 1b ε 1b
εl0

σ σ σA

Model a Model b Sample of material

Fig. 3 Three-parameter models for linear inelastic materials

The equilibrium conditions in Fig. 3 lead to the respective viscoelastic material law:

σ + p1 σ̇ = q0 ε + q1 ε̇ (8)

where,

R1a
p1 = (9)
E1a
q0 = E0a (10)
E0a + E1a
q1 = R1a . (11)
E1a

Considering model (b) in Fig. 3 using subscript “b” we have,

R1b
p1 = (12)
E0b + E1b
E1b E0b
q0 = (13)
E0b + E1b
R1b E0b
q1 = . (14)
E0b + E1b
12 Damping of Materials and Structures 691

2.2.2 Three-Parameter Models in the Standard Test

Considering Fig. 1 when subjected to a creep test the three-parameter model behaves
by responding with a strain as follows:

ε(t) = σ0 J (t) (15)

where,
   
1 p1 q0 − qq0 t
J (t) = 1− 1− e 1 (16)
q0 q1

is known as the creep compliance.

Furthermore, considering a relaxation test in the context of Fig. 1, we have the
stress response:

σ (t) = ε0 E(t) (17)

where,
   
q1 − p1 t
E(t) = q0 − q0 − e 1 (18)
p1

is the relaxation function.

It can be inferred from Fig. 1 that both 3-parameter models, as depicted in Fig. 4,
will initially respond to a step-like change in stress and will then, subsequently
exhibit asymptotically elastic behavior:

σ0
σ

σ∞

0 t
ε E R
ε ε

ε∞
ε1

σ
0 t

Fig. 4 Standard Test on the Kelvin-Voigt 2 parameter model

692 L. Gaul and A. Schmidt

p1 σ0
ε0 = ε(0+ ) = σ0 = (19)
q1 E0

and
σ0 σ0
lim ε(t) = = (20)
t→∞ q0 E∞

where E0 = E(0+ ) denotes the initial modulus and E∞ = E(∞) is the equilibrium
modulus. As long as the strain is held constant, the material will continue to relax
up until the point when σ∞ = ε0 E∞ ≤ σ0 = ε0 E0 . This can be seen in Fig. 1
above.
Figure 4 above graphically shows the temporal stress and strain relations in the
two parameter Kelvin-Voigt model for comparison purposes. Also, the material law
as per Eq. (8) has the following factors in this case:

p1 = 0 (21)
q0 = E (22)
q1 = R. (23)

The creep compliance and relaxation functions J (t) and E(t) are given by the
following relations:
 q

1 − 0t
J (t) = 1 − e q1 (24)
q0
E(t) = q0 . (25)

The two-parameter model does not allow for a spontaneous elastic strain ε0 = 0.
However, the relaxation process is spontaneous with an initial modulus of E0 → ∞.

2.2.3 N-Parameter Model

A more accurate description of the actual system behavior can be achieved by
modifying models described previously to incorporate more parameters. This of
course, entails more effort.
Figure 5 shows a rheological N -parameter model for a given material. For a given
set of parameters, E0 , Ek and Rk (k = 1, . . . , K = (N − 1)/2) of N springs and
dampers, equilibrium requires that the total strain ε and the spring extensions εk
satisfy the following relation in the model:


K
σ = E0 ε + Ek εk (26)
k=1

and at the nodes between the springs and dampers

12 Damping of Materials and Structures 693

σ σA

≈
Ek EK
εk εK
E0
( 1 + ε ) l0
ε
Rk RK
ε − εk ε − εK
≈

σ σA
N-parameter model Sample of material

Fig. 5 Rheological model of viscoelastic material behavior

Ek εk = Rk (ε̇ − ε̇k ). (27)

From the point of view of continuum mechanics, the material behavior is

characterized using two variables, viz., the external variable ε and the internal
variables εk . The internal variables εk satisfy the relaxation equations given by
Eq. (27). We can now eliminate the internal strains εk and then differentiate with
respect to time, t, yields the material law as shown below:

σ + p1 σ̇ + . . . pm σ (m) = q0 ε + q1 ε̇ + . . . + qn ε(n) (28)

N −1
where m = n = .
2

2.2.4 Operator Notation

Other rheological models have different orders m = n of the derivatives in the
associated material equations. When using the differential operators


m
dk 
n
dk
P = pk Q= qk (29)
dt k dt k
k=0 k=0

Equation (28) in the generalized form is given below:

694 L. Gaul and A. Schmidt

P σ = Q ε. (30)

A more generalized form of the above equation replaces the integer-order time
derivatives by fractional-order derivatives:


m
dαk 
n
dβk
P = pk Q= qk (31)
dt αk dt βk
k=0 k=0

where αk ∈ R0+ and βk ∈ R0+ , [1, 36, 54]. The definition of a fractional-order
derivative is as given below:

dα x(t) 1 dn t x(t − τ )
= dτ (32)
dt α (n − α) dt n 0 τ α−n+1

where
 ∞
n ∈ N, n − α > 0 (1 − α) = e−x x −α dx. (33)
0

The gamma term in Eq. (33) is called Gamma Function.

2.2.5 Creep and Relaxation

Substituting the integrals of the relaxation equations in Eq. (27) for the internal
strains, εk , in Eq. (26) results in a description of the material behavior by means
of the so-called memory integrals. Applying the boundary condition εk (−∞) = 0,
the integrals are given by

 τ
∞ −
εk (t) = T
e k ε̇(t − τ )dτ. (34)
0

The quantity Tk = Rk /Ek is the relaxation time corresponding to the k th spring-

damper element. At this point, the material equation can be obtained by inserting
Eq. (34) into Eq. (26). This yields:
 ∞
σ (t) = E(∞) ε(t) + Ẽ(τ )ε̇(t − τ )dτ (35)
0

where the function Ẽ(τ ) is given by:

τ

K −
Ẽ(τ ) = T
Ek e k . (36)
k−1
12 Damping of Materials and Structures 695

Equation (36) describes the monotonous fading of the material memory for the
past deformation events, ε(t). Ẽ(τ ) shows asymptotic behavior, i.e., lim Ẽ(τ ) = 0.
t→∞
Equation (35) also describes relaxation of the stress σ (t) after a step-like change in
strain, ε(t) = ε0 1(t) according to the equation:

σ (t) = (E(∞) + Ẽ(t))ε0 = E(t) ε0 . (37)

It can be seen from the above figure that the material parameters E(∞) , Ek , Rk
can be determined from measured relaxation function E(t). Equation (35) may be
expressed as a relaxation type equation by utilizing the relaxation function E(t) as
follows:
 t
σ (t) = E(t − τ ) ε̇(τ ) dτ. (38)
−∞

Also on the other hand, solving the equation for strain yields a creep type material
law given by the following equation:
 t
ε(t) = J (t − τ ) σ̇ (τ ) dτ. (39)
−∞

The creep function J (t) describes creep after a step-like change in stress, σ (t) =
σ0 1(t), as shown in Fig. 6. The creep and relaxation functions J (t) and E(t) are
related to each other via the following linear integral relation:
 t
d
J (t − τ ) E(τ ) dτ = 1. (40)
dt 0

If the strain is initially zero, i.e., ε(t) = 0 for t < 0 we can rewrite Eq. (38) as
follows:

ε/σ0 σ/ε 0
J (t)

J (∞) E (0)
J (0)

0 t 0 t

Fig. 6 Creep and Relaxation Functions

696 L. Gaul and A. Schmidt

 t
σ (t) = E(t) ε(0) + E(t − τ ) ε̇(τ ) dτ. (41)
0

Also, a consequence of Eq. (39) is that for σ (t) = 0 and for t < 0
 t
ε(t) = J (t) σ (0) + J (t − τ ) σ̇ (τ ) dτ. (42)
0

2.2.6 Harmonic Stress and Strain Function

When the strain is harmonic and time dependent:

ε(t) = ε̂ cos(t + ϕ0ε ) = Re ε̂ eit (43)

the material equation, i.e., Eq. (38) yields the following phase-shifted expression for
steady-state stress:

σ (t) = σ̂ cos(t + ϕ0σ ) = Re σ̂ eit (44)

with the complex amplitudes ε̂ = ε̂ eiϕ0ε and σ̂ = σ̂ eiϕ0σ . It is now possible

to define a complex modulus. The real part is denoted as E  () and is called the
storage modulus. The imaginary part on the other hand is known as the loss modulus,
and is denoted by the E  (). Therefore, we have:

E() = E  () + iE  (). (45)

The relation between the complex stress and strain is given by the following
relation:
σ̂ = E() ε̂. (46)

Equation (38) can be utilized to obtain a relation between the complex modulus,
E() and the relaxation function, E(t). For this, we use the approach of Eq. (44):
 ∞  ∞
 
E () = E0 +  Ẽ(τ ) sin τ dτ E () =  Ẽ(τ ) cos τ dτ. (47)
0 0

The complex modulus can also be formulated using the so-called loss factor,
χ ()

E  ()
χ () = . (48)
E  ()

The complex modulus thereby is given by the following expression:

12 Damping of Materials and Structures 697

E() = E  () [1 + i χ ()]. (49)

Equation (43) into Eq. (49) results in the expression:

χ () = tan δ() (50)

where δ() is essentially the phase difference between the stress and the strain:
δ() = ϕ0σ − ϕ0ε . The rheological model shown in Fig. 5 comprises the storage
and the loss modulus as per the following relations:


K
2 Tk 2
E  () = E0 + Ek (51)
k=1
1 + 2 Tk 2

and


K
Tk

E () = Ek . (52)
k=1
1 + 2 Tk 2

In the case of a low-frequency harmonic process, an approximation for the

complex modulus can be used, which is given as follows and is equivalent to the
2-parameter Kelvin-Voigt model shown in Fig. 4:


K
E() = E0 + iR with R = Rk . (53)
k=1

The damping work per unit volume per cycle is given by:

E  () 2
WDh () = π E  ()ε̂2 = π χ ()E  ()ε̂2 = π χ () σ̂ . (54)
|E()|2

With the exception of the normalizing constant π ε̂2 , the loss modulus E  () and
the damping work per unit volume per cycle Wdh (), are identical.
The frequency dependence of the measured damping work under harmonic
excitation can be approximated in a given frequency range by appropriate choice
of the model parameters, E0 , Ek , and Rk .
The storage and loss moduli are shown in Fig. 7 for the 3-parameter model
(E0 , E1 , and R1 ) and the 2-parameter Kelvin-Voigt model. Figure 8 shows the
characteristics in the hysteresis curve, assuming a linear-viscoelastic material.

2.2.7 Three-Dimensional Stress State

Analogous to the 1-D case, the mathematical and rheological models may be
extended to the three-dimensional stress state. The symmetric stress tensor (σj l )
698 L. Gaul and A. Schmidt

Fig. 7 Storage and loss

moduli of the 3-parameter
model and the 2-parameter
Kelvin-Voigt model (source:
VDI guideline 3830)

ε̂

1 + χ2

√
 E ε̂2 − ε2
σ̂ = E 1 + χ2 ε̂
E ε̂
E ε̂
E ε

ε
WDh

χ
 ε̂
1 + χ2

ε̂

Fig. 8 Hysteresis characteristics for linear viscoelastic material behavior

can be considered to be comprised of a spherically symmetric, i.e., hydrostatic

component (s δj l )and the stress deviator (sj l )

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
σ11 σ12 σ13 100 s11 s12 s13
⎝σ21 σ22 σ23 ⎠ = s ⎝0 1 0⎠ + ⎝s21 s22 s23 ⎠ (55)
σ31 σ32 σ33 001 s31 s32 s33

where the mean stress s is:

12 Damping of Materials and Structures 699

1
s= (σ11 + σ22 + σ33 ) (56)
3

Analogous to the stress, the strain tensor (εj l ):

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
ε11 ε12 ε13 100 e11 e12 e13
⎝ε21 ε22 ε23 ⎠ = e ⎝0 1 0⎠ + ⎝e21 e22 e23 ⎠ (57)
ε31 ε32 ε33 001 e31 e32 e33

with the mean strain, as with stress given by the relation,

1
e= (ε11 + ε22 + ε33 ). (58)
3

The strain deviator (ej l ) characterizes the deformation in the system. The volume
dilatation, i.e., the change in volume is 3e. It is to be noted that only five of the six
deviator coordinates are independent in each case. This is a result of the condition
that

s11 + s22 + s33 = 0 and e11 + e22 + e33 = 0. (59)

Using the differential operator, the following statements are true for the case of
isotropic materials:

• for the deviator components (j, l = 1, 2, 3)

P1 sj l = Q1 ej l (60)

• and for the mean stress and mean strain

P2 s = Q2 e. (61)

For integer time derivatives, the differential operators are defined as follows:


m1
dk 
n1
dk
P1 = p1k Q1 = q1k (62)
dt k dt k
k=0 k=0


m2
dk 
n2
dk
P2 = p2k Q2 = q2k . (63)
dt k dt k
k=0 k=0

As for fractional derivatives,

700 L. Gaul and A. Schmidt


m1
dα1k 
n1
dβ1k
P1 = p1k Q1 = q1k (64)
dt α1k dt β1k
k=0 k=0


m2
dα2k 
n2
dβ2k
P2 = p2k Q2 = q2k . (65)
dt α2k dt β2k
k=0 k=0

Because of isotropy, P1 and Q1 are identical for all deviator coordinates, and all
quantities sj l have the same relaxation times, Tk . From Eq. (60), we can see that
simulation of the material model requires a unified rheological model for the six
deviator coordinates. In addition to this, an additional model is required for volume
dilatation as per Eq. (61).
Three cases are relevant:

(a) For elastic materials with shear modulus G, bulk modulus K, and Poisson’s
ratio ν, the differential operators reduce to constant factors:

P1 = P2 = 1 (66a)
Q1 = 2G (66b)
2(1 + ν)
Q2 = 3K = G . (66c)
1 − 2ν

(b) In case of elastic volume change, energy is lost via deformations. However, there
are no losses due to change in volume. Here

P2 = 1 (67a)
Q2 = 3K. (67b)

(c) When stress is mono-axial, the operators P and Q in Eq. (29) are related to
P1 , Q1 , P2 , and Q2 as follows:

P = P2 Q1 + 2Q2 P1 and (68a)

Q = 3 Q2 Q1 . (68b)

When written down in terms of memory integrals, the material equations of the
relaxation type are as given below:
 t
sj l (t) = E1 (t − τ )ėj l (τ ) dτ for j, l = 1, 2, 3 (69a)
−∞
 t
s(t) = E2 (t − τ )ė(τ )dτ (69b)
−∞
12 Damping of Materials and Structures 701

where E1 (t) and E2 (t) are the memory functions. If we solve for strains, we
obtain the material equations of creep type:
 t
ej l (t) = J1 (t − τ )ṡj l (τ ) dτ for j, l = 1, 2, 3 (70a)
−∞
 t
e(t) = J2 (t − τ )ṡ(τ )dτ (70b)
−∞

where J1 (t) and J2 (t) are the creep functions.

In case of harmonic time histories:

sj l (t) = Re( ŝ j l eit ) for j, l = 1, 2, 3 (71a)

ej l (t) = Re( êj l eit ) for j, l = 1, 2, 3 (71b)

s(t) = Re( ŝ eit ) (71c)

e(t) = Re( ê eit ) (71d)

the complex shear modulus

G() = G () + iG () (72)

relates the complex deviator amplitudes, ŝ j l and êj l as follows

ŝ j l = 2G() êj l for j, l = 1, 2, 3 (73)

and the complex bulk modulus

K() = K  () + iK  () (74)

relates the complex amplitudes ŝ and ê of the mean stress and strain, respectively

ŝ = 3K () ê. (75)

Using the Equations (61) through (65) in Eq. (59), the relation between the
modulli G (), G (), K  (), K  (), and the memory functions E1 (t) and E2 (t)
can be deduced as follows:

 1  ∞
G () = E1 (∞) + [E1 (τ ) − E1 (∞)] sin τ dτ (76a)
2 2 0

 ∞
G () = [E1 (τ ) − E1 (∞)] cos τ dτ (76b)
2 0
702 L. Gaul and A. Schmidt


 1  ∞
K () = E2 (∞) + [E2 (τ ) − E2 (∞)] sin τ dτ (76c)
3 3 0

 ∞
K  () = [E2 (τ ) − E2 (∞)] cos τ dτ. (76d)
3 0

It thus implies that

G () K  ()

χ1 () = and χ2 () = (77)
G () K  ()

are the loss factors for deviatoric deformations and changes in volume.

2.2.8 Temperature Dependence of Viscoelastic Material Properties

Up until now, we discussed isothermal visco-elastic changes of state or neglected
temperature dependence of material properties when dealing with non-isothermal
phenomena. Material properties of thermoplastics are highly temperature dependent
especially near the glass transition temperature.
Creep and relaxation test under conditions of varying temperature show similar
dependencies of mechanical properties. This can be seen in Fig. 9, shown below:
Storage modulus E  and the material loss factor χ are also temperature dependent
as evident form Fig. 10. The short-term domain in Fig. 11 and the respective high-
frequency domain in Fig. 10 pertain to the vitreous state. The rubbery state in
the long-term and low-frequency domain is marked by a borderline of a narrow
temperature range around the glass transition temperature Tg . Tg is a reference
quantity. Its corresponding quantity in the frequency axis is ωg . As evident from

Fig. 9 Stress relaxation of a rubber material (Vulcollan A) at four temperatures (source: VDI
guideline 3830)
12 Damping of Materials and Structures 703

Fig. 10 Examples of the temperature and frequency dependence of storage modulus and material
loss factor (source: VDI guideline 3830)

Fig. 11 Change in volume as 3e

a function of temperature

Change in volume

Temperature
Tg T

Fig. 11, the glass transition temperature is mostly determined via the change in
volume.

2.2.9 Thermo-Rhoelogical Simple Materials

Experiments show [8, 26] that a wide range of viscoelastic materials exhibit a
thermo-rheological simple behavior [55, 56]. The relaxation function, E(t, T0 )
and Creep function, J (t, T0 ) at a given reference temperature, T0 , meet the below
mentioned conditions:

E(t, T ) = E(tϕ(T ), T0 ) and (78a)

J (t, T ) = J (tϕ(T ), T0 ). (78b)

This means that the material functions for the temperature T are obtained from those
at the temperature T0 by substituting the reduced time, tϕ(T ), for the time t. The
shift function ϕ(T ) accounting for the time distortion fulfills the conditions:
704 L. Gaul and A. Schmidt

dϕ(T )
ϕ(T0 ) = 1 and > 0. (79)
dT

This material property must be determined experimentally. The temperature-

dependent complex moduli E(, T ) at temperature T result from the moduli
E(, T0 ) at the reference temperature T0 by substituting /ϕ(T ) for . The
frequency shift function a(T ), therefore, is the reciprocal of the time shift function
ϕ(T ).
Time and frequency shifts will not change the asymptotic values of the mechan-
ical material properties but will affect the transition range. A simple approach for
the time shift function ϕ(T ) is given as [26]:

C1 (T − T0 )
log ϕ(T ) = . (80)
C2 + (T − T0 )

This function is called the WLF function after the names of Williams, Landel,
and Ferry. It is often used with numerical values (Fig. 12).

8.86(T − T0 )
log ϕ(T ) = where T0 ≈ Tg + 50K. (81)
101.5K + (T − T0 )

More details can be found in the literature, see, e.g., [26, 12]

2.3 Nonlinear Models

Under conditions of large deformations, all materials show nonlinear behavior, and
the stresses and strains cannot be related to each other by linear equations and also
superposition principle fails to apply [21, 36].

Fig. 12 Time shift function

(WLF function) (source: VDI
guideline 3830)
12 Damping of Materials and Structures 705

Mathematical models of continuum mechanics exist. It allows conditions of

triaxial stress states and also arbitrary time histories. They result in nonlinear
material laws in the form of an expansion of Equations (69) and (70) and Equations
(38) and (39), respectively. Series expansions of the memory function terms lead to
sums of simple and multiple integrals which are of little practical use as they hardly
find any application in vibration problems [46].
In addition, rheological models or calculation models exist, each with a limited
range of application. The parameters pertaining to these models are identified
in measured results such as hysteresis curves, memory, or creep functions. This
chapter will give examples of such models, which apply to the uniaxial stress
condition.

2.3.1 Models for Static Hysteresis

The hysteresis curve of many materials (mostly metals), when subjected to harmonic
tensile strain ε(t) = ε̂ cos(t +ϕ0ε ), will have the distorted shape shown in Fig. 2. In
case the stress amplitude exceeds about 1/20 of the fatigue strength under alternating
stress, the shape of the hysteresis curve and the damping work per unit volume have
almost no dependence on frequency, WDh = WDh (ε̂) and therefore called static
hysteresis or velocity independent behavior. While damping work WDh per unit
volume for linear-viscoelasticity is proportional to ε̂2 , under nonlinearity, this is not
the case.

Point-Symmetrical Hysteresis Without Reversal Points

If the hysteresis curve measured under harmonic strain is point symmetrical and has
no reversal points, the Masing model as shown in Fig. 13 can be used. This is the
simplest model for such an application. It consists of massless linear spring elements
and massless Coulomb friction elements.

Fig. 13 Phenomenological σ̄k σ

model for static hysteresis
≈

εk εk εK
Ek Ek EK
ε
E0

Hk Hk HK
≈

σ̄k σ
Jenkin element Masing model
706 L. Gaul and A. Schmidt

Fig. 14 Characteristic of the σ̄k

kth Coulomb element
Hk

ε̇ − ε̇ k

− Hk

A combination of one linear spring and a Coulomb element in series is termed as

a Jenkins element. Each Coulomb element transmits a “stress” σ̄k whose magnitude
is Hk or less. Refer Fig. 14
The total stress is given by:


K
σ = E0 ε + σ¯k . (82)
k=1

If the kth Coulomb element slips, in that case:

σ¯k = Hk sgn ε̇ and ε˙k = 0 (83)

The end of the slip phase is marked by the change of sign of the strain velocity, ε̇ and
then the kth element sticks. If ε+ and ε̇+ denote the strain and strain time derivative,
immediately at the beginning of the stick range, then the following can be written:

σ̄k = Ek (ε − ε+ ) − Hk sgn ε̇+ and ε̇k = ε̇ (84)

as long as |σ¯k | = Hk . The next slip phase starts when, |σ¯k | = Hk with |ε| growing at
that point in time. The stress-strain curves of the overall model consist of piecewise
linear polygon lines.
Figure 15 shows the hysteresis curve for a strain ε(t) = ε̂ cos(t + ϕ0ε ) in
the steady state which, depending on the initial conditions, is only reached after an
instationary phase. In this case, only the Coulomb element with parameter ratios

Hk
< ε̂ (85)
Ek

dissipate energy; the others always stick. The damping work per unit volume per
period, WDh , is:
12 Damping of Materials and Structures 707

Fig. 15 Qualitative
representation of the σ
hysteresis of the Masing
model shown in Fig. 13
assuming harmonic strain

−ε̂
ε̂ ε

Fig. 16 Qualitative curve of WDh

loss work, WDh , as a function
of the strain amplitude, ε̂,
under harmonic strain

ε̂

∗

k
Hk
WDh (ε̂) = 4 Hk (ε̂ − ). (86)
Ek
k=1

Contributions to the above sum only come from elements that fulfill the condition
in Eq. (85). The function WDh (ε̂) forms a polygon line linking the nodes ε̂ =
Hk /Ek , k = 1, . . . , k ∗ . As the number of elements increases, the quality of the
approximation improves (Fig. 16).
The Masing model, however, has two disadvantages: stiffness discontinuities and
the need to distinguish between phases of slip and stick.
Nonetheless, we can modify the model. The modification replaces Eqs. (83) and
(84), while retaining Eq. (82), by evolution equations for the internal variables [5].

  m 
1  σ¯k 
σ˙¯k = E0 ε̇ 1 − (1 + sgn (ε̇ σ̄k ))   . (87)
2 Hk

In the above equation, m is always a positive value. As m increases the quality of

approximation of Eqs. (83) and (84) improve.
708 L. Gaul and A. Schmidt

The curvilinear method [43] for pure shear distortion, γ , separates the shear
stress, τ , into a linear part Gγ , and a frictional part, τR :

τ = Gγ + τR . (88)

After each reversal point, γU , of the strain, the frictional part, τR , is approximated
using the two model parameters R and ρ, by the following equation:

τR = R sgn γ̇ ln(1 + ρ|γ (t) − γU |). (89)

This method dispenses with the integration of evolution equations. Instead,

the reversal points, γU i , of the distortion, i.e., the zero crossings of γ̇ , must be
determined over and over, and the part τR must be determined by consecutive
appending of line elements.

General Shape of Hysteresis Curves

The characterization of the behavior of materials whose hysteresis curves have
reversal points, or which are unsymmetrical, requires more general models. One
possible method consists in the generalization of Eqs. (82) and (87) [5]:


K
σ = E(ε)ε + σ̄k (90a)
k=1
   
 σ̄k m
σ̄˙ k = ε̇ Ak (ε) − (βk + αk sgn(ε̇ σ̄k ))   . (90b)
Hk (ε) 

The calculation model contains distortion dependent stiffnesses, E(ε) > 0,

Ak (ε) > 0 and stick forces Hk (ε) as well as the real constants αk , βk , and m.
This ensures a good potential for matching to measured results.

2.3.2 Models for Nonlinear Viscoelasticity

There are a deviation from the elliptical shape of the hysteresis curve of a viscoelas-
tic solid subjected to a tensile test with harmonic strains and large amplitudes, ε̂, see
Fig. 17.
In the general case, the resulting hysteresis curve will not even be a closed
loop. Where closed hysteresis loops exist, the damping work, WDh , depends on
the angular frequency, , and the strain amplitude, ε̂, i.e., WDh = WDh (, ε̂). The
loss work per period for a viscoelastic solid is:

lim WDh (, ε̂) = 0. (91)

→0
12 Damping of Materials and Structures 709

Fig. 17 Stress Strain curve

of a nonlinear viscoelastic σ
material under harmonic
strain

−ε̂
ε̂ ε

The following models apply to arbitrary time histories of monoaxial stress

conditions and are a modification of the integral distortion laws used in the theory
of linear viscoelasticity [54, 60]:
If the longitudinal strain, ε(t), is given:

 t 
K

− t−τ dε(τ )
σ (t) = h0 (ε(τ ))E0 + h1 (ε(τ )) Ek e τk
dτ (92)
0 dτ
k=1

where E0 and Ek are constant material parameters having the dimension of Young’s
moduli, τk are the discrete relaxation times, and h0 and h1 are the strain dependent
nonlinearity functions. If the longitudinal stress, σ (t), is given:

 t 
K 

 dσ (τ )
− t−τ
ε(t) = g0 (σ (τ ))J0 + g1 (σ (τ )) Jk 1 − e τk dτ (93)
0 dτ
k=1

where J0 and Jk are constant material parameters, τk are the discrete relaxation
times, and g0 and g1 are the stress-dependent nonlinearity functions. Equations (92)
and (93) encompass the special case of linear viscoelasticity for h0 = h1 = 1, and
g0 = g1 = 1.

2.3.3 Models for Static Hysteresis and Viscoelasticity

The models in Sect. 2.3 are two limiting cases of nonlinear material behavior. In
general, the characteristics of static hysteresis emerge at the same time as the
dependence on distortion velocity, as the effects of different damping mechanisms
are superimposed. In this case, the hysteresis curves measured in tensile tests under
harmonic strain, ε(t) = ε̂ cos(t + ϕ0ε ), will have peaks, and the area and shape
of the hysteresis curve changes with the angular frequency  as shown in Fig. 18.
Filled polymers such as technical rubber materials show this behavior.
710 L. Gaul and A. Schmidt

Fig. 18 Hysteresis curves of Ω  = Ω1

a material exhibiting static σ
hysteresis and
frequency-dependent
damping under harmonic
strain

−ε̂ Ω1
ε̂ ε

Fig. 19 Rheological model σ

of nonlinear material
behavior

≈
Ek EK
εk εK

ε
Rk RK
E0

Hk HK
uk uK
≈

Rheological Models
Filled polymers may be described by Rheological models combining velocity-
dependent behavior and velocity-independent behavior. The resulting combination
models comprise Coulomb elements, springs, and dampers in varying configura-
tions depending on the material and the load condition, see Fig. 19 [42]. If σ̄k
denotes the partial stress transmitted along the line consisting of Coulomb element
Hk , damper Rk , and spring Ek , the total stress is:


K
σ = E0 ε + σ̄k . (94)
k=1
12 Damping of Materials and Structures 711

Fig. 20 Dependence of the σ̄k

frictional stress, σ̄k , on the
distortion velocity, u̇k
Hk

u̇k

− Hk

Distinctions between the cases of stick and slip for the kth Coulomb element can be
avoided by using the approximation:

2 Hk u̇k
σ̄k = arctan (95)
π vk

for the dependence of the frictional stress σ̄k on the distortion velocity of the
Coulomb element, with the sufficiently small reference velocity vk > 0 (Fig. 20).
From Eq. (95), follow the evolution equations for the internal variables uk :

π σ̄k
u̇k = vk tan for k = 1, . . . , K. (96)
2 Hk

Taking into account the characteristic curves of the damper Rk and the spring Ek ,
and Eq. (96), the evolution equations for the partial stresses result as:
 
σ̄k π σ̄k
σ̄˙ k = Ek ε̇ − − Ek vk tan (97a)
Rk 2 Hk
for k = 1, . . . , K. (97b)

The above equations, viz., (94), (96), and (97) describe the material behavior for
the superposition of viscoelasticity and static hysteresis.

Mathematical Model
The following mathematical model is an updated version of the “curvilinear
method” of Sect. 2.3 [43], assuming that the linear elastic part, Gγ (t), remains
unchanged and the frictional part, τR , is replaced by a relaxing viscoelastic part τν :

τ (t) = Gγ (t) + τν (t). (98)

The following applies to the relaxing, viscoelastic part:

712 L. Gaul and A. Schmidt

 t Rρ γ̇ (s)
τν (t) = K(t − s)ds (99)
−∞ 1 + ρ|γ (s) − γU |

where γU is the shear stress in the preceding reversal point of the distortion
direction, R and ρ are the constant material parameters, and K(t) is the decay
function.

3 Damping of Assemblies

In this chapter, we will deal with the effects of damping on assembled structures and
the mathematical models used to it. We will review material damping as discussed
in Sect. 2 and move further to damping of components and assemblies
A component can be defined as the smallest coherent, jointless unit of an
assembly. Components that have a positive fit yield laminated components (or
composites). An assembly on the other hand is “composed” of components, e.g.
a clamped, bolted or screwed connection. A machine or system is in turn formed via
various assemblies, or an assembly could be a system in itself. But this is a matter
of definition.
In this chapter, we will also consider the following special damping effects:

• damping in joints between components and assemblies

• damping due to fluids surrounding the oscillator
• damping by local displacements of liquids between components

The damping properties are “assembled” in a hierarchical fashion beginning with

the material and moving up via components eventually to the assembly. The descrip-
tion takes into account the possibilities of measuring damping characteristics, i.e.,

• We accomplish this in the time domain by analyzing the time histories for linear
deformation (material) laws, majorly by an assessment of the amplitude decay of
damped vibrations.
• In the frequency domain, we use relations between the input and output
quantities.
• We can also determine the energy dissipation, particularly for steady state
vibrations.

3.1 From Material Description to Complete Homogeneous

Component

If material properties are uniform over the entire volume, it is called homogeneous.
The damper model chosen will have the same parameters, for any element and
any point in the component. However, stress and distortion conditions are position
dependent in a component – as opposed to differentially small volume elements.
12 Damping of Materials and Structures 713

For example, beams, shafts, plates, and shells. As per Eqs. (100) and (101), the loss
factor of a component χS and loss factor for the material will differ. For χS , we
integrate the damping energy over the volume of the component, V , and divide by
the maximum stored potential energy, US , of the vibration:

V WD (x, y, z)dV WS
χS = = . (100)
2π V U (x, y, z)dV 2π US

Here, WD is the damping energy per unit volume and cycle. Displacements
are assumed to have a harmonic time history. U is the potential energy per
unit volume at the point of maximum excursion. Using the loss factor of the
material χ :

WD
χ= . (101)
2π U

This leads to the loss factor of the component:


χ U dV
χS = V . (102)
V U dV

Within a given zone, the specific damping energy and the stress amplitude, σ̂ can be
related as in the following power law:
 n
σ̂
WD = J . (103)
σˆ0

J and n are real material parameters determined via experiments for a given
reference stress, cf. Lazan [45]. Also, if the effect of stress is understood, the
damping afforded by the component can be obtained from the specific damping
energy, WD as shown below:
WD = WD (σ̂ ). (104)

One can deduce from Eq. (103): therefore that the damping energy per period of
the component is as follows:
   n
σ̂
WS = WD dV = J dV . (105)
V V σˆ0

Example. Beam with one end fixed and a mass on the other end, in steady-state
vibration in its fundamental mode (Fig. 21):
In the above, we will neglect the mass of the beam. It follows that M̂y (x) =
F̂ x, where F̂ is the amplitude of force at x = 0. Therefore the position-dependent
normal stress amplitude is, then:
714 L. Gaul and A. Schmidt

L b

EIy x h

F (t) = F̂ cos(Ωt)

Fig. 21 Beam exposed to forced vibration

F̂ x
σˆx (x, z) = z. (106)
Iy

The damping energy is:

   n
L h/2 F̂ x z
Ws = 2J b dx dz (107a)
x=0 z=0 Iy σˆ0
 n
F̂ Lh bLh
Ws = J . (107b)
2Iy σˆ0 (n + 1)2

Here we have the bending stiffness at the fixed end as the reference quantity:
 
F̂ L h
σˆ0 = σˆL = . (108)
2Iy

The volume V and the damping energy of the component WS are as follows:

V = bLh (109)
JV
WS = . (110)
(n + 1)2

The maximum deformation energy can be expressed as:

 L 2  L
M̂y (x) (F̂ x)2
Us = dx = dx (111a)
x=0 2EIy x=0 2EIy
F̂ 2 L3 V σ̂L2
Us = = . (111b)
6EIy 18 E
12 Damping of Materials and Structures 715

If we now use Eq. (100) i.e., the definition of loss factor, we have:

9 JE
χS = σ̂ −2 . (112)
π (n + 1)2 L

3.2 Laminated Components

Laminated components are essentially composites comprising homogeneous com-

ponents. Each component typically has different material properties, and no relative
motion is permitted between these components. If at all any relative motion is
possible, it needs to be accounted for. This is dealt with in the next section.
Calculation of the loss factor, χs , in laminates as in Eq. (100) is:
 
WDi (x, y, z)dVi
i Vi Ws
χs =   = . (113)
2π i Vi Ui (x, y, z)dVi 2π Us

Analogous to the previous section, displacement is assumed to have a harmonic time

history.

3.3 Damping in Joints

Compared to Material damping, damping at the contact surfaces of joints such as

rivets, joints or shrink fits, etc. is greater. It occurs mainly due to the relative motion
of the mating surfaces, micro-slip being one of the mechanisms. Micro slip is the
interfacial slip of small areas. With an increase in tangential loading or the decrease
in contact pressure, a transition occurs from micro to macro level slip. This means
that the entire face of the joint now has relative motion. Coulomb’s Law promises an
approximation for macro-slip. Apart from this, a surface normal relative motion is
also possible. This is mainly associated with elastoplastic deformation at the rough
surfaces [16].
The above mechanisms have a combined effect at the joints and thereby result
in nonlinearity. However, there are cases where nonlinearity is not of significant
importance, viz., where “pumping losses” as in gas pumping are the dominant
factor. This is described in Sect. 3.5. There is a strong relation coupling between
the dissipation mechanisms and the contact pressure, surface condition of mating
surfaces, and the mating materials.
If the wavelength of the vibrations of a joined component is shorter or similar
to the joint dimensions, an FEM simulation can allow for inferences about the
transmission characteristics of the joints in the structure [18,67]. This is particularly
true in the case of large surface joints and for high-frequency phenomena.
716 L. Gaul and A. Schmidt

Fig. 22 Measurement of longitudinal force and relative displacement at an isolated bolted lay
joint as described in [29] (source: VDI guideline 3830)

Alternatively, if the wavelength of the vibration is much larger than the

joint dimensions, local modeling is possible. One can experimentally investigate
individual joints and in a phenomenological manner, develop a mechanical model
of the combined effects [52]. Figure 22 shows one such experimental setup. It is a
set-up for testing a lay joint loaded in the longitudinal direction [29]. The joint is a
bolted lay joint between two masses. The bending compliance of the beam inserted
between the left and right mass causes a system capable of vibrating, which, for
identifying the parameters, is investigated under resonance conditions. This allows
creating large periodic longitudinal forces and easy measurement of the friction
hysteresis in the joint.
It is found from experiments that for a given harmonic load, force/displacement
hysteresis are more or less frequency independent (Sect. 2.3). It is the stress and
strain amplitudes that majorly influence the shape of the hysteresis. They have peaks
at the zeros of the velocity, which is a hint toward Coulomb friction as shown in
Fig. 23.
Transmission behavior of joints may be described in many ways.

3.3.1 Description by a Functional Equation

Frictional hysteresis can be described by the theory of viscoplasticity [62]. By
means of a transform of the variables, an ordinary differential equation is obtained
[31], where the first time derivative of the force Ḟ is a function of the relative
displacement u, the relative velocity u̇, and the force F and is given as per the
following relation. The equation holds true for longitudinal and transverse forces
and thereby also moments:
12 Damping of Materials and Structures 717

F0 Et

E0 u

κ = 0.99
E0
κ < 0.99

Fig. 23 Joint hysteresis with the parameters of the Valanis model as described in [30]

 
E0 u̇ 1 + E0 sgn u̇(Et u − F )
λ
Ḟ (u, u̇, F ) = . (114)
1 + κ Eλ0 sgn u̇(Et u − F )

The above equation consists of four parameters. Static stiffness E0 which char-
acterizes behavior without slip, the sliding stiffness, Et , a dimensionless parameter
κ to denote the curvature of transition from E0 to Et :

0 ≤ κ < 1. (115)

The parameter λ is determined as per the following equation:

E0
λ=  . (116)
Et
F0 1 − κ E 0

Here F0 is the equivalent limit of static friction. It characterizes the force at which
the transition from stiffness E0 to stiffness Et occurs. The parameters E0 , Et , κ, and
F0 can be located on the hysteresis curve is shown in Fig. 23. κ = 1 denotes the
trivial case of linear elastic behavior and therefore is not included in the inequality
(115). The mathematical model, i.e., Eq. (114), does not distinguish between static
and dynamic friction and therefore can be easily integrated into dynamic systems
where rate-independent friction forces must be considered [29, 30].

3.3.2 Description in Terms of Springs and Coulomb Elements

Another method is to use models comprising springs and Coulomb elements [45].
The transmission behavior can be characterized for any arbitrary load history using
the material models. This was already discussed in Sect. 3.1. Figure 24 shows as to
how one could accommodate a joint model into a bar structure. Figure 25 shows the
718 L. Gaul and A. Schmidt

c0

mk m1
c1
H1

cm
Hm

uk u1
Fig. 24 Nonlinear joint model with flange masses

Fig. 25 Hysteresis loop of a

joint model with seven
parameters (source: VDI
guideline 3830)

hysteresis for cyclic relative displacements, u, with the velocity zero crossings at û
and −û for a seven-parameter model. As in Eq. (86), the loss energy per cycle, WD
is determined as a function of û:
In contrast to an equivalent discrete model, the dependence of the loss energy
WD on the relative displacement amplitude, û is not approximated by a polygon
12 Damping of Materials and Structures 719

line as in Fig. 16. We use here Eq. (104) via a smooth curve such as:

WD (û) = 4H ûm (117)

where H and m are real parameters in the continuous equivalent model as per [52]
and [58]:
In [28], we can see the integration of a joint model into a structure discretized
by the FEM using a bar as an example. In case the dimensions of the joint prevent
representing it as a single local model, several such models can be arranged along a
line or in a surface [67].
If one encounters weak nonlinearities, additional effort is required for numerical
computation of structural dynamics. Nonlinear joints may be included by reduction
via equivalent linearization. Joint models with linear springs and viscous dampers
with experimentally determined parameters ignore nonlinearities and static hystere-
sis [64].

3.3.3 Description in Terms of Equivalent Spring and Equivalent

Damper
In the frequency domain, using equivalent spring stiffness and damping coefficients
is one of the options to describe linearized stationary periodical transmission
behavior of joints. It is based on the analysis of an experimental hysteresis curve
measured under harmonic relative displacement, f (u) and g(u) being functions
related to loading and unloading. The longitudinal force is given by:

1 1
F = [f (u) + g(u)] + [f (u) − g(u)] sgn u̇. (118)
2 2

Using harmonic balance method, the equivalent spring stiffness and damping
coefficient can be determined as follows [18]:

 2π
1
c(, û) = F (û cos t, −û sin t) cos t d(t) (119)
π û 0
 2π
1
d(, û) = F (û cos t, −û sin t) sin t d(t). (120)
π û 0

The elliptical hysteresis loop of the linearized longitudinal force with the com-
plex stiffness c(, û) = c(, û) + id(, û) can be compared to the hysteresis
polygon of the nonlinear model. Usually, the nonlinear amplitude and the frequency
dependence of the parameters of the equivalent system requires iterative solutions
of the nonlinear algebraic functions of the dynamic transmission behavior [27].
Using an equivalent linearization of the joint model of Fig. 24, αi =
arccos(1 − 2Hi /ci û), we get the spring stiffness as:
720 L. Gaul and A. Schmidt


m  
ci sin 2α
c(, û) = c0 + αi − . (121)
π 2
i=1

Also, the equivalent damping coefficient is:


m
ci
d(, û) = (1 − cos 2αi ). (122)
2π 
i=1

Implementation of the above joint model in a discretized structures can be found

in [53].

3.3.4 Description Using Finite Element Models

One could also use FEM to model the joint. Convenient optimization of damping
of joints or comparing different variants is possible as these changes are easier to
implement in a numerical simulation than in an experiment.
To obtain good predictions, physically reasonable contact formulae are needed
that describe macro- and micro-slip behavior. By taking into account the rough
surfaces and deformation behavior, such formulae can be achieved. This kind
of detailed modeling requires a large amount of effort, and therefore, statistical
averaging methods are used [65]. Based on a description of the rough surface as
a random process, relations for the elastoplastic behavior of the roughnesses are
obtained as averages over the contact surface [66].
Outputs from these relations such as load dependent contact stiffnesses in the
normal and tangential directions or limit of static friction can be used as input
parameters in FEM models to obtain a realistic joint behavior [32].

3.4 Damping Due to Fluids

3.4.1 Interaction Between a Structure and the Surrounding Medium

Excursions normal to the surface of a vibrating structure cause displacements in the
surrounding medium which, in turn, acts back on the structure, exerting a pressure
on the surface. The normal velocity and the pressure variations are related via the
Kirchhoff-Helmholtz integral. For harmonic vibrations in a loss-free medium, we
have:

p(r, t) = p̂(r)eit (123)

where r is a position vector specifying a particular point in space. Also, the normal
velocity is given by:

v n (r S , t) = v̂(r S )eit (124)

12 Damping of Materials and Structures 721

where r S is the position vector of a point on the surface of the vibrating structure.
p̂(r) is given by the following integral:

    
1 ∂ e−ikR e−ikR
p̂(r) = p̂(r S ) + iρ v̂ n (r S ) dS (125)
4π S ∂n R R

where the pressure amplitude, p̂(r S ), at the surface of the structure, is determined
by substituting on p̂(r) the left-hand side of Eq. (125) by p̂(r S ) and solving
the integral. However, this involves high computational costs. Equation (125)
shows:

R = |r − r s | distance between the structure surface and the point under

consideration in the medium
ρ density of the medium
k = /c = 2π/λ angular wavenumber of the medium
∂/∂n spatial derivative in the surface normal direction
S total surface area of boundary surfaces of the medium

• Pressure amplitudes in the medium and at the surface depend on the normal
velocities of all points of the surface.
• Pressure variations and normal velocities in general do not vary in phase.

An often made assumption that p(r S , t) and vn (r S , t) are proportional is in

contradiction to Eq. (125). The following assumes a linear, infinite, and loss-free
medium. The theoretical basis of the structure surrounding medium are found in
[23, 25, 40].

3.4.2 Radiation Efficiency, Logarithmic Radiation Efficiency, and

Radiation Loss Factor
The radiation behavior of bodies and structures, as per technical acoustics [22, 24],
is characterized by the radiation efficiency:

PS
σ = (126)
ρcAv̄ 2

or the logarithmic radiation efficiency given by

σ  = 10 log σ dB (127)
722 L. Gaul and A. Schmidt

PS sound power radiated

c and ρ speed of sound in, and density of, the medium
A total area of radiating surface
ṽ root mean square of the vibration velocity normal to the surface,
time-averaged over the area

In cases where the acoustic wavelength is much smaller than the dimensions of
say, a plate vibrating in phase:

2π c c
λ= = (128)
 f

the radiation efficiency is σ = 1. The vibration frequency and the vibration mode
significantly affect the radiation efficiency. The energy radiated by a vibrating
structure per cycle (harmonic with angular frequency of )

2π PSh 2π σρcAv̄ 2
WS = = . (129)
 

Also the radiation loss factor is given by the following equation [22]:

WS PS
χS = = . (130)
2π US US

In this equation, US is the vibration energy of the structure as shown in Eq. (113).
Vibration energy of the surrounding however is not taken into account in this
equation. χS plus the loss factors due to other damping mechanisms give the
complete loss factor.

3.4.3 Elementary Radiators

This section deals with some idealized vibrating systems which one could use to
estimate power and radiation efficiency of real systems.

Monopole or Zero-Order Spherical Radiator (Breathing Sphere)

It is essentially a sphere with a fixed center and an oscillating radius r(t) = r0 +
r(t) where r0 is the mean position about which the oscillation occurs. The surface
velocity, vS = d(r)/dt, is assumed to be known. A spherically symmetrical sound
field is created around the radiator. A harmonic velocity:

v S = v̂S eit (131)

induces a pressure variation on the surface, which is given by:

pS = p̂S eit . (132)

12 Damping of Materials and Structures 723

Here the complex amplitude is given by:

(2π r0 /λ)
p̂S = iρcv̂S (133)
1 + i(2π r0 /λ)

where λ is the acoustic wavelength as described in Eq. (128). As v̂ 2 = v̂S2 /2, the
expression for sound power is:
(2π r0 /λ)2
PS = 4π r0 2 ρcṽ 2 (134)
1 + (2π r0 /λ)2

and the radiation efficiency as:

(2π r0 /λ)2
σ = . (135)
1 + (2π r0 /λ)

As can be seen from Fig. 26, the only quantity that affects the magnitude of the
logarithmic radiation efficiency is the ratio (2π r0 /λ). The following borderline
cases can be seen:

1. (2π r0 /λ) << 1 (very low frequencies)

PS ≈ 4π r0 2 ρcṽ 2 (2π r0 /λ)2 (136a)

σ ≈ (2π r0 /λ)2 . (136b)

2. (2π r0 /λ) >> 1 (very high frequencies)

PS ≈ 4π r0 2 ρcṽ 2 (137a)
σ ≈ 1. (137b)

Fig. 26 Logarithmic
radiation efficiencies, σ  , of
spherical radiators as a
function of the ratio (2π r0 /λ)
(source: VDI guideline 3830)
724 L. Gaul and A. Schmidt

Fig. 27 First-order spherical

radiator ϑ
M

υM Axis of radiation

2r0

Dipole or First-Order Radiator (Vibrating Rigid Sphere)

As can be seen from Fig. 27, a rigid sphere with radius r0 with center M moves
harmonically on a straight line with a velocity v M = v̂M eit . The surface normal
velocity of the surface elements vS depends on the reaction angle ϑ given by the
relation:

v S = v̂M cos ϑeit (138)

and the sound power is as follows:

2
ṽM (2π r0 /λ)4
PS = 4π r0 2 ρc . (139)
3 4 + (2π r0 /λ)4

In the above equation, ṽM is the root mean square (RMS) value of the vibration
velocity of the sphere. The relation between ṽS2 averaged over the surface and ṽM
2 is:


1
ṽS2 = ṽ 2 cos2 ϑdS (140a)
4π r0 2 S M
 π
1 2 1 2
ṽS = ṽM
2
sin ϑ cos2 ϑ dϑ = ṽM . (140b)
2 0 3

The above result in conjunction with the sound power, i.e., Eq. (139) can be used
with Eq. (126) that the radiation efficiency is:

(2π r0 /λ)4
σ = (141)
4 + (2π r0 /λ)4

which depends only on the ratio (2π r0 /λ) as seen in Fig. 26. Yet again, we observe
the following cases:

1. (2π r0 /λ) << 1 (very low frequencies)

v˜M 2
PS ≈ 4π r0 2 ρc (142a)
12
12 Damping of Materials and Structures 725

1
σ ≈ (2π r0 /λ)4 . (142b)
4

2. (2π r0 /λ) >> 1 (very high frequencies)

v˜M 2
PS ≈ 4π r0 2 ρc (143a)
3
σ ≈ 1. (143b)

Plane Radiator (Piston) [24]

A rigid plane with a surface normal velocity vK can be idealized as a piston. In this
case, the precise form of the surface does not matter. Only the surface area AK is of
interest which can be given by its mean radius:

rK = AK /π . (144)

The following two differentiations can be made here:

• (2π rK /λ) << 1 (low frequency)

If one side of the radiator is acoustically screened such that there is no acoustic
short circuiting of the front and rear of the radiator surface – for harmonic
vibrations, v K = v̂K eit the sound power is given by:
 2
1 2π rK
PS ≈ AK ρcṽK
2
. (145)
2 λ

This can be used in cases of in-phase vibration of multiple membranes or plates

2 in place of
that are supported over the entire perimeter. In such cases, we use ṽK
2
ṽ and the radiation efficiency turns out as:
 2
1 2π rK
σ ≈ . (146)
2 λ

However, due to short circuiting in radiators without acoustic screening, both

radiation efficiency and sound power are diminished as seen below:
 
1 2π rK 4
PS ≈ 2
AK ρcṽK (147a)
24 λ
 4
1 2π rK
σ ≈ . (147b)
24 λ
726 L. Gaul and A. Schmidt

• (2π rK /λ) >> 1 (high frequency)

In this case, acoustic short circuits play no role and

PS ≈ AK ρcṽK
2
(148a)
σ ≈ 1. (148b)

3.4.4 Damping of Bending Vibrations of Plates

Radiation loss Factor of Homogeneous, Constant Thickness Plates

For a homogeneous plate with density ρp , thickness h, and surface area Ap the
vibration energy is given by:

US = ρp hAp ṽ 2 (149)

and the loss factor

σ ρ c
χS = . (150)
2π ρp f h

Infinite Homogeneous Plate of Constant Thickness

We consider a thin plate which has, in its undeformed state, its center plane in the
x, y plane. The equation of motion neglecting the

• shear deformation and rotational inertia (Kirchhoff’s theory of a plate)

• internal damping
• interaction with the surrounding medium

is given by:
 
 ∂ 4w ∂ 4w ∂ 4w ∂ 2w
B + 2 + + m =0 (151)
∂x 4 ∂x 2 ∂y 2 ∂y 4 ∂t 2

w Deflection of the plate center plane along the z direction

 Eh3
B = bending stiffness of the plate
12(1 − v 2 )
h Plate thickness
E, v Young’s modulus, Poisson’s ratio
ρp Density of plate material
m = ρp h Plate mass per unit area

For a harmonic plane bending wave propagating in the x direction:

w = ŵ cos(t − kB x) (152)
12 Damping of Materials and Structures 727

the wavenumber follows from the Eq. (150):


√ 4 m
kB =  . (153)
B

This directly gives us the wavelength:


2π 4 B
λB = √ . (154)
 m

Here we introduce a term cut-off frequency or coincidence frequency denoted by ωg

or fg . At this frequency, the acoustic wavelength equals the bending wavelength.
Therefore from Eq. (153):

4 m
ωg = c 2
(155a)
B

c2 4 m
fg = . (155b)
2π B

As the plate thickness increases, the cut-off frequency decreases (Fig. 28).

σ =0 for f < fg (156a)

σ = [1 − (fg /f )]−0.5 for f > fg . (156b)

This can also be represented in terms of angular wavenumbers, kB and k = ω/c

σ =0 for k < kB (157a)

σ = [1 − (kB /k)2 ]−0.5 for k > kB . (157b)

A graphical illustration of Eqs. (155) and (156) is given in Fig. 29 where

σ ≈ 1 for f >> fg or k >> kB (158)

In case of infinite plates, there is no radiation for frequencies f < fg , and the
vibrations in the medium decay very quickly as distance from the plate increases.

Homogeneous Rectangular Plate of Constant Thickness

Apart from the ration of f/fg or k/kg , the radiation efficiencies also depend on:

• the geometry
• the vibration mode
• the boundary conditions of the sound field outside the plate
728 L. Gaul and A. Schmidt

Fig. 28 Cut-off frequencies,

fg , as a function of the plate
thickness, h, for radiation into
air under normal conditions
by plates, as per [22] (source:
VDI guideline 3830)

Fig. 29 Radiation efficiency

of an infinite plate as a
function of the frequency
ratio f/fg or of the
wavenumber ratio k/kB
(source: VDI guideline 3830)
12 Damping of Materials and Structures 729

Eigenmodes of bending vibrations are important and have two qualitative

differences compared to the infinite plate:

1. For f = fg the radiation efficiency remains finite.

2. Even if the frequency is less than the cut-off frequency fg , radiation occurs and
therefore damping takes place.

Also, for all eigenmodes:

σ ≈1 for f >> fg (k << kB ). (159)

Rectangular Plates Supported on All Sides

For rectangular plates fitted in an infinite acoustic screen as in Fig. 30, there is
extensive numerical information available [37,63]. The eigenmodes, wm,n of a plate
jointed on all sides having side lengths a and b, are of the type:
 x  y
wm,n = ŵm,n sin mπ sin nπ m = 1, 2, 3 . . . ; n = 1, 2, 3 . . . . (160)
a b

In the range k < kB (f < fg ), radiation efficiencies vary strongly as a function

of the mode number (m, n) and the aspect ration b/a of the rectangular plate. As
Fig. 31 shows the results calculated for a square plate demonstrating the effect of
mode numbers m, n on the radiation efficiency, σm,n . Similarly, Fig. 32 illustrates
the effect of the aspect ratio b/a of a rectangular plate for (1,1) and (2,2) modes.
When the vibration mode numbers are odd, the radiation efficiency is the highest and
on the other hand, lowest when they are even. Mode (1,1), which is the fundamental
mode, has the highest radiation efficiency among all others. For a given value of

Fig. 30 Rectangular plate x acoustical screen

fitted in an infinite acoustical
screen
z

a
x
b

y
730 L. Gaul and A. Schmidt

Fig. 31 Radiation efficiencies, σm,n , of the eigenmodes wm,n of a square plate supported on all
sides, as a function of the wavenumber ratio k/kB , as per [63] (source: VDI guideline 3830)

Fig. 32 Radiation
efficiencies, σ1,1 and σ2,2 , of
the eigenmodes w1,1 and w2,2
of a rectangular plate
supported on all sides, for
selected aspect ratios b/a, as
a function of the wavenumber
ratio k/kB , as per [63]
(source: VDI guideline 3830)
12 Damping of Materials and Structures 731

(m,n) and k/kB , vibration modes whose nodal lines (approximately) enclose square
areas radiate little sound; the more the ratio na/mb deviates from unity, the more
intense the radiation from the plate. This can also be seen in [63].

Other Boundary Conditions

As per the investigations by Gomperts [37] for more general boundary conditions,
radiation efficiency can increase or decrease under stronger restraint to fixed plate
edges.
For a specific eigenmode, a simple analytical expression for radiation efficiency
cannot be given. However, average modal logarithmic radiation efficiencies as a
function of the frequency, f can be estimated via the diagram developed by Holmer
and Ver [39], see Fig. 33. An average is taken over the eigenmodes with eigen
frequencies in a frequency band centered around the frequency f . This is a problem
as the applicability of the diagram is then limited. It is applicable in cases with
broadband excitation forces exciting several resonances at the same time, if the
responses of the eigenmodes are then summed up.
At frequencies below the cut-off frequency, the plate edges are the main
contributor to sound radiation and therefore, damping. As an assumption in rigid
and infinite acoustic screens, we consider prevention of flow of the surrounding
medium around the plate edges. With no screen, acoustic short circuit occurs and
therefore reduces sound radiation and damping.

Fig. 33 Theoretical mean values for the logarithmic radiation efficiencies of eigenmodes of
rectangular plates (having surface area A and perimeter length U ) in a rigid acoustical screen;
λg = c/fg is the wavelength at the cut-off frequency as per [39]. In the range (1), add 3 dB if
edges are fixed (source: VDI guideline 3830)
732 L. Gaul and A. Schmidt

Ribbed Plates
Stiffening ribs on thin plates effectively act like the subdivision of the total area
into many sub-areas. They act as internal edges and thus increases the radiation
efficiency below the cut-off frequency. If the total length of the ribs is L and U
is the perimeter of the plate, then, Maidanik [48] recommended to increase the
radiation efficiency in this range by a factor of 1 + (2L/U ).

3.4.5 Damping of Vibrating Pipes

We consider here homogeneous, cylindrical pipes with constant cross sections.

Infinite Regular Cylindrical Pipe

The radial velocity of the surface v of an infinite regular cylindrical pipe with outer
radius a as shown in Fig. 34 due to a harmonic wave in axial direction, expressed in
cylindrical coordinates is as follows:

vn (ϕ, z, t) = v̂ n cos nϕ exp[i(t − kz z)] n = 0, 1, 2 . . . . (161)

Figure 35 shows pipe deformations in the lowest three-mode numbers:

• n = 0 : breathing
• n = 1 : bending
• n = 2 : ovalizing

ϕ
2ai
2a

Fig. 34 Regular-cylindrical pipe; dimensions and cylindrical coordinate system

+
+

+ +

n=0
n=1
n=2

Fig. 35 First eigenmodes of a pipe

12 Damping of Materials and Structures 733

The infinite tube will only radiate sound if the axial wavenumber kz is less than the
acoustical wavenumber k, kz < k [23]. The mode number n has a strong influence
on the radiation behavior. In the limiting case where kz a << ka << 1, the radiated
sound power per unit length is governed by the following:

1 2
P0  = π a(ρc)(ka)|v̂ 0 |2 n=0 (162a)
2
1
P1  = π 2 a(ρc)(ka)3 |v̂ 1 |2 n = 1 (162b)
4
1 2
P2  = π a(ρc)(ka)5 |v̂ 2 |2 n = 2. (162c)
32

In the entire range ka < 1 in a uniformly vibrating pipe (kz = 0) as shown in Fig. 36
- as n increases, the sound radiation decreases. ka here, is the ratio of pipe perimeter
to acoustical wavelength. In addition to this, to avoid acoustic short circuits at the
surface, the following condition is required:

kz 2 + (n/a)2 < k 2 . (163)

Fig. 36 Radiation efficiency, σ  , of a uniformly vibrating, regular cylinder as a function of the

ratio ka = outer perimeter/acoustical wavelength, as per [23] (source: VDI guideline 3830)
734 L. Gaul and A. Schmidt

This is analogous to the requirement of infinite plates that kB < k. This means
that for a given value of n, cylinders with large outer radius can fulfill the criteria
in Eq. (162), while ones with small outer radius cannot. For efficient radiation in
case of bending vibrations of regular cylinders (n = 1), the following must be
assured:

(k 2 − kz 2 ) a 2 > 1 (164)

where kz is the axial wavenumber of bending vibrations given by


4ρr 2
kz =
4
(165)
E(a 2 + ai 2 )

where ρr is the density of the pipe and ai is the inner radius.

Bending Vibrations of Long Regular Cylindrical Pipes

As evident from the discussion, so far we know that the finite length of long regular
cylinders only has an effect far below the cut-off frequency. Also this was true if the
pipes are poor radiators anyway. This provides enough reason to neglect boundary
conditions and calculate radiation loss factor only in an approximate manner [41]
as shown in Fig. 37.

Bending Vibrations of Long Pipes with Elliptical or Rectangular Cross Section

For rotationally nonsymmetric cross sections, several geometrical parameters influ-
ence the radiation behavior. In case of elliptical cross sections, Fig. 38 shows that
the radiation loss factor for a range of semiaxes ratios, (0.1 < a/b < 10),

Fig. 37 Diagram for

estimating the radiation loss
factor, χS , of long regular
cylindrical pipes in
accordance with [41] (source:
VDI guideline 3830)
12 Damping of Materials and Structures 735

Fig. 38 Diagram for approximate determination of the radiation loss factor, χS , of long pipes
with elliptical and rectangular cross sections, as per [41]. x: calculation results for rectangular
cross sections with various aspect ratios (source: VDI guideline 3830)

approximately lies on a curve. An acceptable estimate of the radiation loss factor for
rectangular cross sections with various aspect ratios also fit into this diagram [15].

3.4.6 References to Nonlinearities

In bending vibrations of strip specimens, considerable nonlinearity of air damping,
increasing with the vibration amplitude, is observed. This is, as used for measure-
ments of material damping refer [13, 14]. Having said that, nonlinearities can also
occur in other cases.

3.5 Damping by Displacement

3.5.1 Damping by Air Displacement

In screwed on or riveted plates and stiffeners, there are gaps that will open and
close. This will result in the so-called gas pumping and involves suction and
ejection of air as shown in Fig. 39. As per the experiments in [20, 49], this will
contribute significantly to the total damping in a frequency range far above the
lowest eigenfrequency.
Known theoretical models cannot predict the magnitude of such damping. Exten-
sive experimentation on the decay of vibrations nonetheless allows the estimation
of loss factors for vibration frequencies in a given one-third octave band [20, 61].
Typical dynamic loads do not entail any nonlinearity.
For plates without stiffeners, damping depends on the wavelength λB of the
harmonic bending waves. The longitudinal wave velocity for a homogeneous plate
with thickness h, density ρ, Young’s modulus E, and Poisson’s ratio ν is given by:
736 L. Gaul and A. Schmidt

Fig. 39 Damping by air

displacement in plates with
stiffeners (source: VDI
guideline 3830)


E
cL = . (166)
ρ

For steel, aluminum, titanium, and magnesium, cL ≈ 5100m/s. Additionally, we

also have the relation:
√ 
π hcL
λB =  . (167)
4
3(1 − ν )
2 f

The bending wavelength may also be expressed in terms of the absorption coeffi-
cient and the loss factor [38]:
LλB
χ =γ . (168)
π 2A
The absorption coefficient γ has roughly a linear behavior with respect to the width
of the overlap w. Other influencing factors are the ratio of spacing d between the
fixing points, to bending wavelength λB and on the ambient pressure p. Figure 40
shows recommended values (“modified”- γm0 ) of the absorption coefficient. Fig-
ure 41 shows the dependence of absorption factors.
The following can be used for conversion purposes to other conditions:
 2
d h0 cL0
fm = f (169)
d0 h cL

and
w
γ = γm . (170)
w0

If there are many joints, the following summation needs to be employed to find the
loss factor of the entire component:
12 Damping of Materials and Structures 737

Fig. 40 Average value of the

modified absorption
coefficient γm0 as a function
of the modified
one-third-octave band center
frequency fm or of the ratio
d/λB , as per [20] (source:
VDI guideline 3830)

Fig. 41 Dependence of the

absorption coefficient, γ , on
the ambient pressure, p, as
per [20] (source: VDI
guideline 3830)

λB
χS = Σγn Ln . (171)
π 2A

3.5.2 Journal Bearings, Squeeze Film Dampers

In journal bearings and squeeze film dampers, the damping effect results from the
buildup of pressure as a consequence of displacement flows. The resulting damping
force is proportional to the vibration velocity and progressively increases with the
eccentricity of the journal, in plain journal bearings and radial dampers. As an
approximation, total carrying force for a given static eccentricity is linearized, and
therefore, the damping force can be expressed in terms of damping constants.

Journal Bearings
Figure 42 shows a plain journal bearing. As per simple bearing theory [44], the
relative eccentricity, ε = e/δ, and the angle α depend on the Sommerfeld number
and on the ratio of the widths, B/D:
738 L. Gaul and A. Schmidt

u2 , F2

FS
Ω

e
R u1 , F1
r
α

Fig. 42 Circular Bearing

FS 2
So = (172)
BDη

where

η dynamic viscosity
D = 2R bearing diameter
B bearing width
r journal radius
δ =R−r bearing play
= δ/R relative bearing play
e journal eccentricity
ε = e/δ relative eccentricity
FS static bearing force
α attitude angle
 angular velocity of rotor

If the journal executes small amplitude transverse movements, u1 (t) and u2 (t),
around the equilibrium position determined by the static bearing force, FS , the
components of the resulting additional force are approximately linear functions of
the movements:

F1 (t) = k11 u1 + k12 u2 + d11 u̇1 + d12 u̇2 (173a)

F2 (t) = k21 u1 + k22 u2 + d21 u̇1 + d22 u̇2 . (173b)

In the above equations, k and d are the stiffness and damping coefficients of the oil
film. Nondimensional coefficients for stiffness and damping are:
12 Damping of Materials and Structures 739

δ δ
γik = kik βik = dik . (174)
FS FS

A more detailed tabulations can be found in [33, 47, 59] and DIN 31 657-1. For
B/D < 0.3 analytical expressions for γik and βik can be found [50]. Dependence of
γik and βik are shown in Figs. 43 and 44. In case of large journal amplitudes or large
dynamic amplitudes a nonlinear approach needs to be taken to adequately describe
the vibration behavior. In practice, static bearing loads and the excitation forces
acting on the rotor are given, and the vibration movements of rotor and journal are
investigated. Rotor movements can be known via the solution to Reynolds equation.
The forces F1 (t) and F2 (t) can be calculated via an iterative procedure [19, 34].

3.5.3 Squeeze Film Dampers

The working principle of squeeze film dampers is shown in Fig. 45. The design
details of the squeeze film dampers can be seen in Fig. 1 of [35] and [57]. Squeeze
film bearings can be considered as journals with a nonrotating journal with the

Fig. 43 Relative stiffness

coefficients of the short
circular bearing (source: VDI
guideline 3830)

Fig. 44 Relative damping

coefficients of the short
circular bearing (source: VDI
guideline 3830)
740 L. Gaul and A. Schmidt

Housing
y, Fy Liquid
x, Fx Ring (non rotating)
Bearing
R r
r
R
e
Housing
Ring

Fig. 45 Squeeze film damper

medium in the gap. Using a linear approximation we may write:

Fx (t) = dx ẋ (175a)
Fy (t) = dy ẏ. (175b)

The above is valid even for an eccentric position. As opposed to a journal bearing
squeeze film dampers are open at the sides and have no pockets and have no coupling
terms in the principle frame of reference. For a regular cylindrical design and centric
position dx = dy :

2Bη 2Bη
dx = 3
βx∗ , dy = 3
βy∗ . (176)

Using Eq. (176) the damping coefficients dx , dy can be calculated. The dependence
of βx∗ , βy∗ on eccentricity and the ratio of widths B/D are shown in Figs. 46
and 47. For high vibration velocities, inertial characteristics must also be taken into
account [57]. As expected from Eq. (175), until the load limit, the vibration velocity
amplitude will increase linearly after which cavitation will occur in the squeeze film
leading to reduced damping performance or even complete failure. Load limits may
be altered by changing the static oil pressure in the damper.

3.6 Assemblies

In assemblies, many components are joined together between which energy dissipa-
tion can occur due to relative motion. The total loss factor can be given as follows:
Σi WDi + Σj WDFj + Σk WDAk
χS = . (177)
2π Uges
12 Damping of Materials and Structures 741

Fig. 46 Relative damping β∗x , β∗y

coefficients of a squeeze film
damper as a function of
relative eccentricity, as per
[35] 3
B/D = 0.5

2 β∗x

β∗y

0 ε = e/δ
0 0.2 0.4 0.6

Fig. 47 Relative damping β∗x

coefficients of a squeeze film
damper as a function of
relative eccentricity, as per
[35] 3
e=0
β∗y = β∗x
2

0 B/D
0 0.2 0.4 0.6 0.8

In the above equation the following dissipation energies occur:

• component damping WD
• joint damping WDF
• external damping WDA

A synthetic determination of χS is problematic and can only be successful in simple

cases. The loss factor corresponding to the rth eigen vibration is calculated via a
weighted summation [17]:

N (r)
j =1 χSj Uj
χS(r) = N (r)
. (178)
j =1 Uj
742 L. Gaul and A. Schmidt

4 Models for Damped Structures

In Sects. 2.1 through 2.3, we dealt with deformations in structures as a whole and
also deformations on the surface. It is clear that the usability of a particular model is
subject to the problem formulation. In the models presented below, we assume linear
damping. In light of what was said in the earlier paragraph, the linearity condition
should be checked ascertained in case of application.

4.1 Basic Model

A standard model in vibration engineering is the 1-DOF oscillator. This may be

used to model oscillatory systems such as mechanical, electrical, acoustic, or fluidic
systems. In a mechanical system, there is the exchange of energy from potential
energy to kinetic energy. Due to damping effects, energy is dissipated, and the free
vibrations decay. In case of an externally applied force, there is also a phase-shift
between excitation and response resulting in hysteresis. The governing equation of
the model, see Fig. 48, is as follows:

mü + d u̇ + ku = F (t) (179)

where,
u displacement at time t
F (t) external force
m mass
k spring stiffness
d damping coefficient

Fig. 48 Mechanical system

with translatory vibrations

u(t)
m F (t)

k
12 Damping of Materials and Structures 743

4.1.1 Free Vibrations with F(t) = 0

With a natural frequency of:

k
ω0 = (180)
m

damping ratio,
d
ϑ= √ (181)
2 km

the natural angular frequency ωd of the damped system,


ωd = ω0 1 − ϑ 2 (182)

the decay coefficient,

δ = ϑω0 (183)

and the abbreviation


ωd∗ = ω0 ϑ 2 − 1 (184)

the time dependent solution of the free vibrator with F (t) = 0 and initial conditions
u(0) = u0 and u̇(0) = ν0 :

• for 0 ≤ ϑ < 1:
 
−δt δu0 + ν0
u(t) = e u0 cos ωd t + sin ωd t (185a)
ωd
u(t) = e−δt û sin(ωd t + ϕ) (185b)

where,

 2
δu0 + ν0
û = u20 + (185c)
ωd
u0 ωd
tan ϕ = (185d)
δu0 + ν0

• for ϑ = 1:

u(t) = e−ω0 t [u0 (1 + ω0 t) + ν0 t] (186)

744 L. Gaul and A. Schmidt

• for ϑ > 1:
 
δu0 + ν0
u(t) = e−δt u0 cosh ωd∗ t + sin ω ∗
t . (187)
ωd∗ d

When no damping is present the above equations reduce to harmonic vibrations with
angular velocity ω0 . When damping is very high, i.e., ϑ ≥ 1 the time history curve
passes through a zero at most, once.
The most common case in engineering is when damping is weak, i.e., ϑ < 1,
there is a succession of zero crossings at a time interval of:

2π 2π
Td = > T0 = . (188)
ωd ωd

The amplitude decay occurs on a logarithmic scale as given below:

u(t) ϑ
 = ln = δTd = 2π √ . (189)
u(t + Td ) 1 − ϑ2

If damping is very weak:

 = 2π ϑ. (190)

4.1.2 Forced Vibrations Where F(t)=0

If the external force is a harmonic excitation force:

F (t) = F̂ cos(t) (191)

the displacement can be written as:

u(t) = û cos(t − ζ ) (192)

where,

F̂
û = α(η, ϑ). (193)
k

In the above equation, α is the frequency-dependent amplitude with the amplitude-

frequency characteristic

1
α(η, ϑ) =  (194)
(1 − η )2 + (2ϑη)2
2

where the frequency ratio is:

12 Damping of Materials and Structures 745


η= . (195)
ω0

ζ is the angle of lag and is given by:

2ϑη
tan ζ = where, 0 ≤ ζ < π. (196)
1 − η2

Dissipation of energy by the damper can be computed using:

WDh = π d  û2 . (197)

Using the term for loss factor Sect. 3.1, we get:

1 2
Umax = k û (198)
2
WDh d
χ= = . (199)
2π Umax k

In case of resonance  = ω0 or η = 1 results in:

d ω0
χ0 = = 2ϑ. (200)
k

In complex notation:

F (t) = F̂ cos(t + γ ) = Re[F̂ eit ] (201)

u(t) = û cos(t + β) = Re[û eit ]. (202)

F̂ and û are complex amplitudes as can be seen from Fig. 49:

F̂ = F̂ eiγ (203)

û = ûeiβ . (204)

Therefore, the angular phase difference can be given by:

ζ = γ − β. (205)

In complex notation, the following frequency responses occur:

746 L. Gaul and A. Schmidt

Fig. 49 Forced harmonic Im

vibration on the complex
plane. The pointers rotate at F̂eiΩt
the angular velocity .

ûeiΩt
ζ

Ωt + γ

Ωt + β
0
F (t) u(t) Re

4.1.3 Dynamic Compliance (Receptance)

û α 1 1
= e−iζ = (206a)
F̂ k k (1 − η ) + i2ϑη
2

 
û 1 1 − η2 −2ϑη
= +i (206b)
F̂ k (1 − η2 )2 + (2ϑη)2 (1 − η2 )2 + (2ϑη)2

4.1.4 Dynamic Stiffness

F̂ k
= eiζ = k[(1 − η2 ) + i2ϑη] (207)
û α

4.1.5 Mobility (Admittance)

u̇ˆ û α
= i  = i e−iζ (208a)
F̂ F̂ k
 
u̇ˆ η 2ϑη 1 − η2
=√ + i (208b)
F̂ km (1 − η2 ) + (2ϑη)2 (1 − η2 ) + (2ϑη)2
12 Damping of Materials and Structures 747

4.1.6 Mechanical Impedance

F̂ F̂ k iζ 2ϑη − i(1 − η2 )
= = e =d (209)
u̇ˆ iû iα 2ϑη

4.1.7 Accelerance
üˆ û α
= −2 = −2 e−iζ (210a)
F̂ F̂ k
 
üˆ η2 1 − η2 2ϑη
=− + i (210b)
F̂ m (1 − η2 )2 + (2ϑη)2 (1 − η2 ) + (2ϑη)2

4.1.8 Dynamic Mass or Inertance

F̂ 1 F̂ 1 (1 − η2 ) + i2ϑη
= − 2 = −m 2 eiζ = −m (211)
üˆ  û η α η2

It is possible to represent complex frequency responses in the form of:

• Amplitude and phase frequency response

• Real- and imaginary-part frequency response
• Locus diagram in the complex plane
• Logarithmic plotting in the Bode diagram

4.2 Structures with a Finite Number of Degrees of Freedom

For a finite element or boundary element implementation, we discretize the domain

so as to have a finite number of degrees of freedom. Damping can be accounted for
in the following ways:

• viscoelastic material behavior

• structural damping with given frequency dependence
• local energy dissipation

With the help of special damping matrices and nodal transformation, equations of
motion can be decoupled and solved. If such decoupling is not reflective of the actual
system, coupling effects need to be incorporated. If decoupling is not possible,
bimodal calculation [75] needs to be carried out.
748 L. Gaul and A. Schmidt

4.2.1 N-Parameter Model for Viscoelastic Material Behavior

In a discretized domain, let the displacement vector be u = {u1 , u2 , . . . un }T .
Assuming linear behavior and no damping, the equation of motion is given by:

M ü + Ku = f (t) (212)

Here M is the symmetrical, positive definite mass matrix, and K is the symmetrical,
positive definite or semi-definite elastic stiffness matrix. K and f (t) may be
determined using a finite element model with ue (t) as the nodal displacements and
N (x, y, z) as the trial or shape functions, we obtain the displacement vector as:

v(x, y, z, t) = N (x, y, z) ue (t). (213)

The strain vector:

ε = {εxy , εyz , εzx , εxx , εyy , εzz }T (214)

can be obtained via a differential operator C as shown below:

⎛ ⎞
1 ∂ 1 ∂
⎜ 0 ⎟
⎜ 2 ∂y 2 ∂x ⎟
⎜ ⎟
⎜ 1 ∂ 1 ∂ ⎟
⎜ 0 ⎟
⎜ ⎟
⎜ 2 ∂z 2 ∂y ⎟
⎜ ⎟
⎜1 ∂ 1 ∂ ⎟
⎜ 0 ⎟
⎜ 2 ∂x ⎟
C = ⎜ 2 ∂z ⎟ (215)
⎜ ∂ ⎟
⎜ 0 ⎟
⎜ 0 ⎟
⎜ ∂x ⎟
⎜ ⎟
⎜ ∂ ⎟
⎜ 0 0 ⎟
⎜ ∂y ⎟
⎜ ⎟
⎝ ∂ ⎠
0 0
∂z

ε = C v = C N ue = CN ue. (216)

The elemental stiffness matrix is given by [68, 79]:


T
Ke = CN E CN dV . (217)
V

Stress and strain are related by:

σ = E ε. (218)
12 Damping of Materials and Structures 749

The global stiffness matrix K is obtained by assembling the element stiffness

matrices Ke which link the nodal degrees of freedom to the nodal forces fe :

Ke ue = fe . (219)

The case of viscoelasticity is treated by means of the correspondence principle.

Starting point are the constitutive equations of the homogeneous isotropic contin-
uum in the form:

σ = [2GEG + 3KEK ]ε (220)

where,

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
σxy 1 000 0 0 0000 0 0
⎜σ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ yz ⎟ ⎜ 1 0 0 0 0⎟ ⎜ 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜σ ⎟ ⎜ 1 0 0 0⎟ ⎜ 0 0 0 0⎟
σ = ⎜ zx ⎟ EG = ⎜
⎜ 2 1 1⎟
⎟ ⎜
EK = ⎜ 1 1 1⎟
⎟. (221)
⎜σxx ⎟ ⎜ 3 3 3⎟ ⎜ 3 3 3⎟
⎜ ⎟ ⎜
⎝σyy ⎠ ⎝ 2 1⎟⎠
⎜
⎝ 1 1⎟⎠
3 3 3 3
σzz (symm) 2 1
3 3

Differential Operator Formulation

It should be noted here that the elastic constitutive law as shown in Eq. (220) is
related to the viscoelastic constitutive law in the differential operator formulation as
shown in Sect. 2.2.7. Shear modulus G and the bulk modulus K have the following
relations:
 
1 Q1
G= (222)
2 P1

 
1 Q2
K= . (223)
3 P2

To avoid undefined divisions, we multiply by P1 P2 and obtain the following:

P1 P2 σ (t) = (EG Q1 P2 + EK Q2 P1 ) ε(t). (224)

Replacing stiffness matrix 2GEG + 3KEK by the operator EG Q1 P2 + EK Q2 P1

gives us the system of differential equations:

(KG Q1 P2 + KK Q2 P1 ) ue = P1 P2 fe (225)
750 L. Gaul and A. Schmidt

where,

T
KG = CN EG CN dV (226a)
V

T
KK = CN EK CN dV . (226b)
V

On the elemental level therefore, the equation of motion, in differential form, is

written as:

Me P1 P2 ue + (KG Q1 P2 + KK Q2 P1 )ue = P1 P2 fe . (227)

4.2.2 Memory Integral Formulation

The memory integral formulation as shown in Sect. 2.2.5 for the constitutive law
in the case of uniaxial loading corresponds to the elastic constitutive law as per
Eq. (220):
 ∞
σ (t) = [2G0 EG + 3K0 EK ]ε(t) + [2G̃(τ )EG + 3K̃(τ )EK ]ε̇(t − τ )dτ.
0
(228)

The shear and compression relaxation functions are given as below, respectively,

G(t) = G0 + G̃(t) (229)

K(t) = K0 + K̃(t). (230)

Using the constitutive law as in Eq. (228) and the matrices in Eqs. (226), Eq. (219)
yields:
 ∞
[2G0 KG + 3K0 KK ] ue + [2G̃(τ )KG + 3K̃(τ )KK ]u̇e (t − τ )dτ = fe . (231)
0

Finally, the finite element equation of motion may be expressed as an integro-

differential equation:
 ∞
Me üe +[2G0 KG +3K0 KK ]ue + [2G̃(τ )KG +3K̃(τ )KK ]u̇e (t −τ )dτ = fe (t).
0
(232)
In the literature solutions from the difference operator formulation can be found.
Methods for determining u(t) primarily use:

• numerical integration
• finite time elements
• integral transformations (Laplace, Fourier)
12 Damping of Materials and Structures 751

• modal analysis
• matrix functions

Integration of the viscoelastic constitutive equations in order to determine the stress

states can be found in the references [69, 76].

4.2.3 2-Parameter Model According to Kelvin-Voigt, Viscous Damping

The equation of motion for the 2-parameter Kelvin-Voigt model is given by:

M ü + D u̇ + Ku = f (t). (233)

For the solutions to and the properties of displacements u(t) as given in Eq. (233)
under free and forced vibrations can be found in the references [75, 78].
By means of the modal transformation:

u=Φq (234)

we can achieve a decoupling of the n equations of motion from Eq. (233). This is
applicable however, only in cases when the following relation holds (commutative
property of the system of equations):

K M −1 D = D M −1 K. (235)

Φ is an (n × n) modal matrix given by Φ = [u1 . . . ui . . . un ] where ui are the

eigenvectors of the undamped system (D = 0 in Eq. (233)). The diagonalization
condition leads to the decoupled equations as given below:

diag(mi ) q̈ + diag(di ) q̇ + diag(ki ) q = g (236)

with the following diagonal matrices:

diag(mi ) = Φ T M Φ of the modal masses mi

diag(di ) = Φ T D Φ of the modal damping coefficients di
diag(ki ) = Φ T K Φ of the modal stiffness coefficients ki

and the modal forces g = Φ T f . Real eigenvectors can be scaled arbitrarily. In

practical applications, often the quantitively largest component of the eigenvector
is set equal to one, while for theoretical studies, the modal masses are frequently
set such that mi = 1 ∀ i. From this, since ωi2 = ki /mi , the corresponding modal
stiffness ki = ωi2 follows.
In the case where there is Rayleigh-type damping, the damping matrix can be
split into a mass proportional and stiffness proportional parts:

D = a0 M + a1 K. (237)
752 L. Gaul and A. Schmidt

For a more general case refer [70, 71]. As in the case referring to Eq. (237), the
modal damping ratios:

di
ϑi = √ (238)
2 kt mt

assigned to the damping coefficients bi satisfy the relation:

 
1 a0
ϑi = + a1 ωt . (239)
2 ωt
√
Here ωt = kt /mt is the natural undamped angular frequency. a0 and a1 can be
calculated from Eq. (240) where ϑn and ϑm are determined experimentally:
⎛ ⎞
  ωn −ω m
 
a0 ωm ωn ⎝ ⎠ ϑm .
=2 2 1 1 (240)
a1 ωn − ωm
2 − ϑn
ωn ωm

This means that the damping ratios of the remaining eigenforms will have the
frequency dependence of Eq. (239), which is shown in Fig. 50. In the case of more
than two known damping ratios, one should determine the parameters in Eq. (237)
by matching.
When modal dampings are known, D can be calculated from:
 n 
 2ϑt ωt
−T −1
D=Φ diag(di )Φ =M ut uTt M. (241)
mt
i=1

When modal decoupling is not possible, Eq. (233) must be used. The damp-
ing matrix can be calculated from Eqs. (237) and (241). If D does not satisfy
diagonalization condition as in Eq. (235) secondary diagonal elements dij∗ (i = j )
will occur signifying the so-called Damping coupling. This is further discussed in
the literature, refer [77, 78].
When component damping is predominant, the diagonalization condition as in
Eq. (235) is warranted, but care should be taken not to ignore coupling from discrete
dampers. If there is no diagonalization, the eigenvectors of the damped system are
complex. A decoupling of the first-order system assigned to Eq. (233):

A1 ż + A0 z = b(t) (242a)

where,    
u f
z= b(t) = (242b)
u̇ 0
   
D M K 0
A1 = A1 = (242c)
M 0 0 −M
12 Damping of Materials and Structures 753

ϑi

 
1 a0
ϑi = + a1 ωi
2 ωi

ϑn
ϑm 1
a ω stiffness - proportional part
2 1 i

1 a0
mass - proportional part
2 ωi

0 ωm ωn ωi

Fig. 50 Connection of damping ratio and associated natural angular frequencies with Rayleigh
damping

can always be performed by a bimodal transformation with the right-hand and left-
hand eigenvectors of the corresponding eigenvalue problems A1 ż + A1 z = 0 and
AT1 ż + AT0 z = 0. Further reference to this can be found in [75, 80].

4.2.4 Damping with Given Frequency Dependence

The complex modulus can be represented in terms of the frequency-dependent
damping coefficients R as:

E() = E + iR. (243)

This is the representation of the Kelvin-Voigt model as discussed in Sect. 2. Refer

Fig. 51.
The member loss factors χ () can be determined as follows:

R||
χ () = >0 (244)
E
and

E() = E[1 + iχ ()sgn ]. (245)

Regardless of the experimentally determined damping coefficients R(), Eq. (246)

can be used. However, as χ () will not be proportional to , one can no longer use
754 L. Gaul and A. Schmidt

χ
experimentally obtained damping factor
l
o de
tm
-Voig |
n | Ω
lvi R
Ke ≡
E
Ω)
model of constant hysteresis
χ(Ω) ≡ const. χ(

0 Ω

Fig. 51 Frequency dependence of the loss factors

the rheological model. Equation (246) only applies within the frequency range – in
other words, to Fourier-transformed state variables:
 ∞
U (i) = u(t) eit dt (246)
−∞

or to the steady state of a harmonic motion (248)

u(t) = Re(û eit ). (247)

4.2.5 Calculation of Viscoelastic Components by the Boundary

Element Method
The boundary element method can be used especially for 3D problems involving
compact members. Using the formulations mentioned in Sect. 2.2.7, BEM may be
formulated for viscoelastic problems. The BEM reduces the problem dimension by
one as calculation is done on the basis of the boundary integral equation.
The problem shown in Fig. 53 describes a viscoelastic domain ∗ with the
boundary , where a mixed boundary value problem is formulated with displace-
ments ūi prescribed on u and tractions p̄i prescribed on p . Initial conditions also
apply to displacements and velocities.
With the finite element method, trial functions in the element are selected which
do not exactly satisfy the field equations. In contrast, the boundary element method
uses weighting functions which satisfy the field equations exactly in the domain. At
the boundary, approximations are formulated.
In BEM, the fundamental solution is used as weighting functions which are
exact solutions of the field equations of an elastic/viscoelastic unbounded space (full
space) at a field point x with a single load acting at load point ξ . This acts either
harmonically time dependent and spatially as a Dirac delta function (the solution
in the frequency range of the Fourier transformation, or in the steady state with
12 Damping of Materials and Structures 755

Fig. 52 Region and pi = σij n j = p̄i

boundary of a continuum ni
with initial and boundary
conditions

Γu

Γp

ui = ūi

Boundary Γ = Γ p + Γu Boundary conditions

ui ( x, t) = ūi at Γu
pi ( x, t) = σij n j = p̄i at Γp
Initial conditions
ui ( x, 0) = ui0
in Ω∗ ∪ Γ
u̇i ( x, 0) = vi0

B3∗ B2∗
ξ3 B1∗
Load point ξ
x2 r = x−ξ
x3 ξ2 Field point x

ξ1
x1

Bi = Bi∗ eiΩt
Bi∗ = δ( x − ξ )ei

Fig. 53 Load application and geometry for the fundamental solution

harmonic motion, is independent of the initial conditions) or as a spatial and Dirac

impulse (we will not treat here the solution in the time domain; see [73]).
Coordinate Bi∗ of the load (body force) in the i direction at the source or load
point x gives rise to the stress vector coordinates Pj∗i (x, ξ , ) and the displacement
vector coordinates Uij∗ (x, ξ , ) in the j direction at the receiver or field point x (as
shown in Fig. 53).
756 L. Gaul and A. Schmidt

The fundamental solution of the viscoelastic full space satisfies the constitutive
equations in Sect. 2.2.7 discussed in Sect. 2. In the fundamental solution of the
displacement field Uij∗ (xi , ξ ) in the frequency range, the complex moduli of shear
G(i ) and compression K(i ) and the distance r = |x − ξ | are contained in the
functions Ψ and κ (see [72, 81]).
 
1 ∂r ∂r
Uij∗ (x, ξ , ) = Ψ δij − κ (248)
4π G ∂xi ∂xj

As per Maxwell and Betti’s theorem, the work associated with Pi , Bi , and Uij∗ must
equal to the work associated with Bi∗ , Pj∗i , and Ui . With no body forces, we have
the equation:

Ui (ξ ) = [(Pj (x)Uij∗ (x, ξ )) − Pij∗ (x, ξ )Uj (x)]d (249)


Referring to Fig. 52, the boundary values Ūt at u and P̄t at p are known. In the
direct formulation of the BEM, the unknown displacement field Ui and the stress
field Pi at the boundary are determined from a boundary integral equation which
is derived from the integral Eq. (249) by relocating the load point ξ (see Fig. 53)
from a point within the region to the boundary. If the solution fields Ui and Pi on
the boundary are known, then it will be possible to calculate from Eq. (249) the
displacement field at any point in the interior of the domain.
For a numerical solution apart from the geometry of the component boundary,
displacements and stress fields must be approximated. For this purpose, the
boundary is discretised into the so-called boundary elements within which the
approximation is done using the isoparametric concept. Refer Figs. 54 and 55 and
[72] for a more detailed discussion. For each load point, we get two or three
equations depending whether the problem is two or three dimensional, respectively.
These are then numerically integrated using say, the Gauss Quadrature.
Due to the usage of approximations, the boundary integral equations are not
completely satisfied. This results in a defect or residual. According to the method
of weighted residuals, this residual must disappear when weighted with a function
and integrated over the boundary. If we chose each node to be a load point –

Fig. 54 Discretisation of a surface into boundary elements

12 Damping of Materials and Structures 757

Fig. 55 Two-dimensional trial functions for boundary variables and geometry, The points repre-
sent nodes

the number of equations is the same as the number of unknowns. The solution
results in boundary displacements and stress. These can be calculated using the
constitutive laws. Using inverse Fourier transform, one can convert the frequency
domain solution to one in the time domain [74].

5 Experimental Techniques for the Determination of

Damping Characteristics

5.1 Experimental Techniques

5.1.1 Basic Procedures

Many of the variables that constitute the mathematical description of materials,
components or systems require experimental determination. The excitation and
corresponding response are measured and linked or an energy balance is applied.
Material properties are determined from component or system level experimenta-
tion.
If the nonlinearity of the system and the damping is small, a linear viscoelastic
model is a reasonable approximation. However, the diverse approximations in
conjunction with measurement errors have a cumulative effect on the end result and
thus give rise to a great deal of scatter in the results of identified damping measures.

5.1.2 External Damping

Damping measurement is influenced by the loss of mechanical energy due to
external factors (external damping). The following are some of these factors:
758 L. Gaul and A. Schmidt

• ambient air
• bearing points
• losses due to clamping
• parts of the measure instrument itself

In measurement of cases with low damping special care must be taken to avoid
external damping.

5.1.3 Applicability of Results

Factors influencing the damping characteristics can be classified into three cate-
gories:

• material parameters such as chemical composition and cross-linking in polymers

and anisotropies induced via manufacturing processes such as heat treatment
• mechanical loads and deformations such as the magnitude of the deformations or
the damage history
• ambient influences and test conditions

Measured damping characteristics are highly specific to testing conditions, and

therefore global application is problematic especially in the case of nonlinear
damping.

5.2 Experimental Techniques and Types of Apparatus

5.2.1 Survey of Experimental Techniques

With powerful computers, these days, complex numerical identification algorithms
can be used. This section however deals with fundamental evaluation methods which
make possible a fast estimation of parameters and checking of the numerical results.
There is a diverse array of applications of damping and its measurement that can
be converted from one to the other via calculation. Evaluation of measurement
results in order to determine damping characteristic values can be made easier if
experimental techniques are used which come close to the operational conditions
which apply in the real world.
For example, in solids, determination of creep or relaxation functions, is done
via a creep or relaxation tests, respectively. In the case of fluids, evaluation of flow
behavior is also required. For a more detailed discussion, refer [94].

5.2.2 Quasi-Static Methods for the Determination of Material

Properties
Quasi-static methods include the creep and relaxation tests for solids and the
determination of its viscosity in the case of fluid. Loads and displacements are
measured temporally, while stress and strain are determined by geometry factors
called cgeom . Creep and relaxation functions can be determined by following the
guidelines based on Sect. 2.
12 Damping of Materials and Structures 759

5.2.3 Experimental Determination of Damping in Solid Bodies with a

Low Shear Modulus
Materials with a shear modulus in the range of 0.1 N/mm2 ≤ G ≤ 100 N/mm2
create some issues regarding external damping, but this can be overcome by using
sufficiently rigid testing equipment. Specimens used to determine damping charac-
teristics from creep tests are usually simple shear specimens, tension specimens, and
torsion specimens with a rectangular cross-section as shown in Fig. 56.
The following equations for converting measured loads (forces F and torsional
moments MT ) and deformations (extensions  l and torsions  ϕ) into stresses
(σ and τ ) and distortions (elongations ε and slip angle γ ) apply to these.

τ 1 F bl0
Shear specimen : = where cgeom = (250a)
γ cgeom  l a
σ 1 F ba
Tension specimen : = where cgeom = (250b)
 cgeom  l l0
τ 1 MT ba 3 c1
Torsion specimen : = where cgeom = . (250c)
γ cgeom  ϕ l0

a) b) c)
F F MT

l0 l0 l0
cross-section

F F MT

Fig. 56 Solid specimens for creep and relaxation tests (a) Specimen under pure shear (in
central longitudinal cross-section) (b) Specimen with evenly distributed tensile stress (outside the
clamping zone) (c) Torsional specimen with rectangular cross-section (inhomogeneous distribution
of the shear stress and cross-sectional camber)
760 L. Gaul and A. Schmidt

For a more detailed description of the methodology for tests on the above mentioned
specimens, refer [94].

5.2.4 Experimental Determination of Damping in Solid Bodies with a

High Shear Modulus
For specimens of high stiffness, testing machines with high stiffness are required.
Such machines for creep and stress relaxation under torsion have been developed.
Further reading can be done in [85]. Solids whose viscoelastic behavior can be
reproduced by a Kelvin-Voigt model as a parallel arrangement of spring and damper
only exhibit signs of creep and negative creep but not of relaxation as shown in
Fig. 57. Basic equations with regard to storage modulus E  and damping coefficient
R are:
σ∞ σ0
E = R= . (251)
ε0 (dε/dt)t=0

5.2.5 Experimental Determination of Damping in Viscous Liquids

The standardized measurement equipment used for determining the dynamic vis-
cosity η(T , γ̇ ) of Newtonian liquids is the flow viscosimeter, where, T is the
temperature and γ̇ is the shear speed given by γ̇ = dν/dh. Many different types

a) σ b) ε

σ0 ε0

jump in stress jump in strain

0 t 0 t

ε σ
σ0

creeping relaxation
ε∞
σ∞
σ0 J (t) ε 0 E(t)
ε0
0 q1 /q0 t 0 t

Fig. 57 Stress and strain plotted against time in the creep test and in the relaxation test with
tension specimens (ideal curve with Kelvin-Voigt model) (a) Creep Test (b) Relaxation Test
12 Damping of Materials and Structures 761

of viscosimeters are available – a few rotational viscosimeters are shown in Fig. 58.
Major equations involved in the experimental set up are as follows:

τ 1 MT
η= = . (252)
γ̇ cgeom ϕ̇

For a cone plate viscosimeter shown in Fig. 58a:

MT
a) b)
ϕ̇
MT

ϕ̇

r
r1

α r2
l

c) d)
MT MT

ϕ̇ ϕ̇
length of action

r1 r1 h
l
r2

1 h

air bubble h 2
α=
r1

Fig. 58 Types of rotation viscosimeters (a) Cone-plate viscometer (b) Simple shear flow between
coaxial cylinders (Couette flow) (c) Shear flow between coaxial cylinders with flat end faces (d)
Shear flow between coaxial cylinders and between cone end face and plate (h r1 )
762 L. Gaul and A. Schmidt

2π r 3
cgeom = . (253)
3a

For the Couette flow shown in Fig. 58b (exact only for Newtonian liquids) this
equation applies:

4π/(r1 r2 )2
cgeom = . (254)
r22 − r12

In the case of the viscosimeter shown in Fig. 58c, Eq. (254) corresponds to Fig. 58b
provided corrections for the ultimate effect of the internal cylinder are included.
Corresponding to Fig. 58d, we have:
 
1
2π r13 l 1 + r1 / l
3
cgeom = . (255)
h

The following referencing have a detailed discussion on the above mentioned topics:
DIN 1342-3, DIN 53018-1 and -2, DIN 51563, DIN 53017, and DIN 53018, [85]
and [94].
With silicone oils or melts, it is recommended, as a way of checking the test
results, that the so-called reduced viscosity η(T , γ̇ )/η0 where η0 = η(T , 0) be
formed and plotted against the reduced shear speed γ̇ η0 . In a graph of this kind, the
measured values of silicone oils of different characteristic viscosities νsn = ηs0 /ρ
and temperatures T will lie on approximately the same curve (Fig. 59).

Fig. 59 Reduced viscosity as a function of reduced shear speed (Note: 1cSt = 10−6 m2 /s)
(source: VDI guideline 3830)
12 Damping of Materials and Structures 763

l
Ω
Ω x

ν
δ
h
F

Fig. 60 Method of rotating bending

5.2.6 Determination of Damping in Uniformly Rotating Specimens

In rotating, rotationally symmetrical and homogeneous specimens, constant forces
with constant direction generate harmonically changing stresses within the com-
ponent. This means that on the basis of measured, time-independent forces and
deformations, it is possible to deduce the material characteristic values for harmonic
time responses. During the test, a test rod of circular cross-section with fixed
diameter d0 is used which is clamped at one end [83], rotated at the constant angular
velocity , and loaded with the constant, directionally true force F (Fig. 60).
Here, we have two displacements, ν due to the applied force and h resulting from
material damping. The resultant displacement therefore depends on the rotational
velocity . For viscous absorption and with the stress strain equation σ = Eε + R ε̇
for the displacement components, the following equations apply where the axial
area moment of inertia: I = π d04 /64:
F l3 1
ν() = (256a)
3EI [1 + (R/E)2 ]
F l3 R/E
h() = . (256b)
3EI [1 + (R/E)2 ]

From Eqs. (256)

F l3
E= (257a)
3 νI [1 + (h/ν)2 ]
h
R= E (257b)
ν

where the storage modulus E  = E and the loss modulus E  = Eh/ν. The angle δ
follows the following relation:

h
tan δ = X = . (258)
ν
764 L. Gaul and A. Schmidt

For a detailed explanation with regard to other material types and loading patterns,
refer [94].

5.2.7 Determination of Damping in the Case of Free Vibrations with

One Degree of Freedom
In a free vibration test, the magnitude of the damping is determined in the form of
the logarithmic decrements  of the time-decreasing vibration amplitude A(t) or
the peak-to-valley value A∗ (t). In the case of flexural beams supported on bearings
which are free of damping, the free-vibration test is carried out in a vacuum;
torsional vibrations are, however, applied by preference. Figures 61 and 62 show the
arrangements used for standardized torsional vibration devices. As per Fig. 63, the
measured variables are the maximum torsional vibration deflections An in the points
of maximum excursion or the oscillation peak-to-valley values A∗n . The procedural
details which need to be observed when carrying out a torsional vibration test and
related calculations are summarized in DIN EN ISO 6721-2, DIN 1311-2. Important
equations are summarized below.

   
An−1 A∗n−1
 = ln ≈ ln . (259)
An A∗n

For the damping ratio


ϑ=√ . (260)
4π 2 + 2

For weak damping i.e. ϑ 1 we also have ϑ ≈ /(2π ) and therefore

Fig. 61 Torsional vibration

apparatus as per DIN EN ISO 1
6721-2, Method A (1) Upper
fixed clamping point (2)
Specimen (3) Heating 2
chamber (4) Lower moving
clamping point (5)
Connection rod (rigid) (6) 3
Inertial element
4

6
12 Damping of Materials and Structures 765

Fig. 62 Torsional vibration

apparatus as per DIN EN ISO
6721-2, Method B (1)
Counterweight to compensate
for vibration body weight (2)
1
Flexible wire with low
torsional stiffness (3) Inertial
element (4) Connection rod 2
(rigid) (5) Upper moving
clamping point (6) Tempering
chamber (7) Specimen (8) 3
Lower fixed clamping point
4

6
7
8

 
1 An
= ln . (261)
m An+m

In case the system behavior is non-linear then ∗n = (A∗n ). For procedure B,

fB B − f0 0
=  . (262)
f0 0
fB2 − f02 + (fB B − f0 0 )
2π 2

Here variables 0 and f0 relate to a test not using specimens, variables B and
fB relate to a test with a wire and specimen. Note that care must be taken to
account for the logarithmic dependence of  in case of non linear viscous damping.
Free vibration tests under flexural stress can in principle be carried out using
freely suspended or beams clamped at one end or both ends (Fig. 64). For specific
advantages, disadvantages and further related reading refer [94].

5.2.8 Determination of Damping via Specification of Harmonic

Deformations
Here, the testing apparatus forces onto the specimen undergoing harmonic displace-
ments of amplitude ŝ with a mean displacement sm . The set up can be seen in Fig. 65.
The deformation enforced is s ≈ sm + ŝ sin t. The force F is measured at a
specified s which has a steady component Fm and harmonics with whole number
multiples of the angular frequency . It has the form:

F1 = F̂1 sin(t + δG ). (263)

766 L. Gaul and A. Schmidt

deflection

An A∗n

time

Td

Fig. 63 Time curve for oscillatory deflection with damped free vibrations

F (t)
a)

c)

F (t)
0.22 l 0.22 l
b) l

vacuum, excitation orthogonal to the plane of presentation

Fig. 64 Bearing supports for flexural rods for decay tests

force F

a) crank drive (aspect ratio λ = r/l ≤ 1/10) b)

l WD
two specimens force sensor F̂1
Ω

Fm
r

a+s F̂1

dislocation s
sm
ŝ ŝ

Fig. 65 Shear test with approximate harmonic forced drive (a) Test apparatus (b) Force-
deformation diagram
12 Damping of Materials and Structures 767

For cross section and thickness, A0 and L0 , respectively – the complex shear
modulus is

F̂1 L0
|G(, ŝ)| = . (264)
2A0 ŝ

Also, energy loss per cycle is given by:

Ws = π F̂1 ŝ sin δG . (265)

From this, the loss factor results as χG = tan δG . |E| and χE are determined using
the strain specimen although this is limited to low frequencies.

5.2.9 Measurement of the Oscillation Amplitude in Vicinity of

Resonance (Determination of Halfwidth Value)
If a system only has one degree of freedom or if the natural frequency of one eigen-
form differs notedly enough from the adjacent, the resonance test for determining
damping parameters is the most commonly used method in the frequency range of
20 Hz to 1 kHz.
The damping of a separated degree of freedom can be obtained from the
resonance curve by means of the “half-width value” [82] (0 − u ) where care is
taken to ensure that resonance maximum is included in the resonance curve. Errors
may occur especially if steady state is not achieved or, when the measurement is
done if no frequency sampling point coincides with the resonance frequency. In the
case of a resonance curve of the vibration velocity resulting from force or spring
excitation as shown in Fig. 66, double the damping ratio is expressed as below:
   
1 0 ω0 u ω0
2ϑ = − − − . (266)
2 ω0 0 ω0 u

Due to the symmetry, 0 /ω0 = ω0 / u the damping ratio is given by the relation:

0 − u 1
ϑ= = (η0 − ηu ). (267)
2ω0 2

Given above is a basic outline of the idea. Further possibilities, interpretations, and
procedures can be found in [94].

5.2.10 Measurement of Amplitudes and Phase Angles

In case a real system can be sufficiently approximated by a 1 DOF oscillator, the
following possibilities exist for the determination of the damping ratio. Figure 67
shows the Nyquist curve for the displacement amplitude û(). a and b are the
excitation frequencies at which the tangent to the Nyquist curve is parallel with the
imaginary axis. With this, in theory, the damping ration is expressed as:
768 L. Gaul and A. Schmidt


smax

amplitude of the vibration velocity 

s
√
2
2

smax

0 Ω u ω0 Ω o Ω

Fig. 66 Amplitude of the vibration velocity with spring or force excitation with resonance
maximum and “half-value ordinates” as well as the corresponding angular frequencies 0 , u
and ω0

Im û Im û

Re û Re û
−45◦

−135◦
Ωo
Ωb Ωa
Ωu

û(Ω)

ω0 Ω ω0 Ω

Fig. 67 Nyquist curve of the displacement amplitude with spring force excitation

1 2b − 2a
ϑ= (268)
2 2b + 2a

where u is the excitation frequency at the intersection of the −45◦ line and the
Nyquist curve and 0 the excitation frequency at the intersection with the −135◦
line. From this we have, the damping ratio as:
12 Damping of Materials and Structures 769

Fig. 68 Phase frequency

ϕ
response with force or
displacement excitation

Δϕ
π
2
ΔΩ

0
0 ω0 Ω

   
1 1 1 ω0 u
ϑ= − ηu = − (269)
2 ηu 2 u ω0

and
   
1 1 1 0 ω0
ϑ= η0 − = − . (270)
2 η0 2 ω0 0

Another possibility to determine damping is to use the phase frequency response.

When  = ω0 :

dϕ 1
= (271)
d ω0 ϑ

and as per Fig. 68 we have:

1 Δ
ϑ= . (272)
ω0 Δϕ

Refer [94] for details about the assumption and applicability of the above
mentioned procedure.

5.2.11 Determination of Damping via Thermal Energy Balances

Irreversible deformational work in most cases results in production of measurable
heat. Satisfactory results can be obtained for quasi-stationary temperature states.
Older studies [86] used specimens immersed in flowing water in the vibration
test with measurements of inlet and outlet water temperatures. However, there still
remained a large number of potential sources of errors in integral heat quantity
determination for the material regions under load.
770 L. Gaul and A. Schmidt

5.2.12 Energy Balances at the Subsystem Boundaries of

Multicomponent Systems
Here, we have the assumption that in the subsystem as shown in Fig. 69 for which
the damping is to be measured, the force F can be applied in a harmonic manner
and also measured (also the oscillatory moment MT ) along with the displacements
u (also the angle of oscillation ϕ). So we have:

u1 = û1 cos t (273a)

u2 = û2 cos(t − β) (273b)

F1 = F̂1 cos(t + δ1 ) (273c)

F2 = F̂2 cos(t − β + δ2 ). (273d)

Active power applied at boundary 1 is given as

1
P1 = F̂1 û1 sin δ1 (274)
2

and the active power transmitted at boundary 2

1
P2 = F̂2 û2 sin δ2 . (275)
2

Therefore, the energy lost in 1 cycle

WS = π(F̂1 û1 sin δ1 − F̂2 û2 sin δ2 ). (276)

Ignoring the inertial forces, the amplitude of the displacement difference is:

ŝ = û21 + û22 − 2û1 û2 cos β. (277)

Fig. 69 Subsystem within a

multipart total system
d

u(t)
m F (t)

k
12 Damping of Materials and Structures 771

The loss angle can be calculated from:

WS = π F̂ ŝ sin δ. (278)

5.2.13 Force and Displacement Measurements at Subsystem

Boundaries
Any energy dissipating element is usually a part of a larger system, and its
corresponding parameters can be measured without knowing the parameters of the
entire system. The structure can be referred in Fig. 70. From the motion equations
or directly from a consideration of the complex dynamic stiffness in the subsystem,
we obtain the two relations:

F̂ + m1 2 û1 + (k + id)(û2 − û1 ) = 0 (279)

(k + id)(û2 − û1 ) − m2 2 û2 = 0. (280)

Parameters k and d can be obtained from the complex force and/or displacement
amplitudes [82]. For the four complex amplitudes F̂ , û1 , û2 and ŵ, it must be
possible to measure the ratio of any two in quantity and phase. The other complex
amplitudes can be eliminated from the two equations (279) and (280). Depending
on which variables are measured, formulations can be made for k and d. Further to
this, force and distortion amplitudes can also be measured. For a detailed treatment
of this, refer [94].

5.3 Special Experimental Techniques for Determining Damping

Under Difficult Conditions

5.3.1 Systems with High Damping

In structures where damping is very high, resonance vibration tests may not be
accurate enough. This is especially true if stiffness and damping depend on the
frequency of excitation or if the excitation force is not large enough to produce
sufficiently large oscillations. In such a case, a known damping-free spring with

Fig. 70 Subsystem (m1 , m2 )

with in-line spring damper
element (source: VDI
guideline 3830)
772 L. Gaul and A. Schmidt

Fig. 71 Reduction in the

effective damping due to an
additional spring kz
connected in series (source:
VDI guideline 3830)

real stiffness kz can, as shown in Fig. 71, be connected in series with the damped
specimen with complex stiffness k + id. The stiffness and the damping coefficient
can be determined from the measured complex frequency response of the expanded
system. uA (t), um (t), uP (t) have the following relations between them:

k(um − uP ) + d(u̇m − u̇P ) = kz (uP − uA ) (281a)

müm = −kz (uP − uA ). (281b)

For the complex amplitudes ûA , ûP , ûm the following relations hold:

uA = Re{ûA eit } uP = Re{ûP eit } um = Re{ûm eit } (282)

(kz + k + id)ûP = (k + id)ûm + kz ûA (283a)

kz ûP = m2 ûm + kz ûA . (283b)

Elimination of ûP leads to the complex stiffness

(ûm /ûA )m2

(k + id) = . (284)
(ûm /ûA )(1 − m2 /kz ) − 1

Separating the real and imaginary part, stiffness and damping can be found which
needs to be determined for every frequency of interest .
This method assumes that the mass m and the stiffness kz of the additional spring
are known. These parameters can be determined experimentally beforehand, with
the aid, for example, of a resonance test which does not include the specimen to be
tested. The resonance frequency of the entire arrangement can be suitably chosen
by selection of mass m. If the stiffness k and damping coefficient d of the specimen
are required as a function of the frequency , tests should be carried out with
different excitation frequencies and, if necessary, with different additional springs. If
k and d can be regarded as frequency independent, taking measurements at different
excitation frequencies will minimize random error creating influences.
12 Damping of Materials and Structures 773

5.3.2 Flexural Vibrations of Lamellar Specimens

In the field of acoustics, complex moduli or loss factors of elastomers are of much
interest in the frequency range of 10 Hz or higher. Flexural vibrations of lamellar
specimen with an assumption of the Euler Bernoulli beam theory can be useful for
parameter identification [89]. The thickness h must be small as compared to the
distance between the adjacent nodes ak , i.e., h ≤ ak /6. The following have proved
useful as boundary conditions:

• rigid clamp at on end, the other end unrestrained

• rigid clamps at both ends
• both ends unrestrained

Homogeneous Strips
Homogeneous strips are used for measuring the complex modulus of materials
which, having a modulus of elasticity of E > 5 e9 N/m2 , are so stiff that the natural
frequencies of the strips fall within the frequency range in which the modulus is to
be determined. Which measurement technique to use will depend on the magnitude
of the loss factor. Some possibilities are:

• resonance test
• free-vibration test

Laminated Strips
For materials which are relatively soft and/or exhibit very heavy damping lamnated
strips may be used. For this purpose, stiff and comparatively damping-free strips are
bonded together as shown in Fig. 72. A detailed mathematical representation of the
above mentioned subheadings can be found in [94].

Fig. 72 Examples of cross-sections of layered strips (a) two layers (topmost beam [92]) (b) three
layers, symmetrical (Nashif beam [89]) (c) and (d) sandwich arrangements (source: VDI guideline
3830)
774 L. Gaul and A. Schmidt

5.3.3 Longitudinal Waves in Bars

Complex moduli of elasticity can also be determined from longitudinal oscillations
in bars. The method involving progressive, damped extensional waves [88] is less
advantageous in practical terms – the resonance method is more suitable. One end of
the bar is excited; the other end is unrestrained [87] or bears a rigid mass [91]. The
complex transfer function between the exciting force and the longitudinal movement
supplies the information for identifying the complex modulus.

5.4 Experimental Modal Analysis

Real mechanical problems are problems in continua which are approximated to N

(sufficiently large) pointwise degrees of freedom – collocation points. Mathematical
formulations exist that describe phenomena of this equivalent model. The model
parameters are adjusted so that it gives the best approximation to the real-world
problem evaluated in terms of a specific error criterion [84, 90].

5.4.1 Discrete Equivalent Model

In the equation of motion:

M ü + D u̇ + Ku = f (t) (285)

where u and f are vectors of displacements and excitations at the collocation points.
M, D and K are the symmetrical, positively (semi-) definite mass, damping and
stiffness matrices of the equivalent model. Apart from the symmetry, also,

KM −1 D = DM −1 K. (286)

The frequency range Equation (285) reads:

û(i) = H (i) fˆ(i) (287)

H (i) is the matrix of transfer functions that contains the complex frequency-
dependent compliances H ln which links the complex Fourier spectra fˆ(i) and
û(i). The Fourier spectrum of a transformable time function is defined by:
 ∞
F {x} = X(i) = x(t)e−it dt. (288)
−∞

Eq. (287) also applies in particular even to harmonic excitation and vibratory
response:

f = Re [fˆeit ] and u = Re [ûeit ] (289)

12 Damping of Materials and Structures 775

The assumptions Eqs. (285) and (286) make it possible for the elements of the
matrix of the transfer function to be expressed by the N real eigenvectors uj of
the associated undamped system H (i):

M ü + Ku = 0 (290)

and the modal characteristics


N
uj l uj n
H ln = = H nl (291)
kj − mj 2 + idj
l=j

where

mj = uTj Muj are the modal masses

dj = uTj Duj are the modal damping coefficients
kj = uTj Kuj are the modal stiffnesses


with natural angular frequency ωj = kj /mj and modal damping ratios ϑj =
dj /(2mj ωj ) we have


N
uj l uj n /mj
H ln = = H nl . (292)
ω2
j =l j
− 2 + i2ϑj ωj 

In experimental modal analysis, a distinction is drawn between time and fre-

quency range evaluations. Time-domain procedures fall back on measured impulse
responses; frequency-domain procedures on measured transfer functions. Both
types of approach are equivalent from the theoretical point of view since impulse
responses hln (t) and transfer functions H ln (i) are Fourier transform pairs:

H ln (i) = F −1 {hln (t)} ⇔ hln (t) = F {H ln (i)}. (293)

The following section will therefore be restricted to dealing with the frequency
domain.

5.4.2 Basic Principles in the Measurement of Complex Frequency

Responses
Due to symmetry, the matrix of transfer functions is also symmetric – which means
that the excitation and the response measurement points can be interchanged. Also,
a column or a row of the matrix of the transfer functions already contains all
eigenvalues. Therefore, measurement of N transfer functions is sufficient (instead
of N 2 ). For this purpose, the structure is either
776 L. Gaul and A. Schmidt

• excited only at one point n and the responses measured at all N response points
• successively excited at all N collocation points and the response measured only
at one and the same point n

The mathematical procedure describing this process in detail can be found in [94].

5.4.3 Evaluation of Measured Frequency Responses at an Isolated

Resonance Point
The following methods exclude closely adjacent natural frequencies and evaluate
the frequency response in each case in the vicinity of an isolated resonance
frequency ωGr ≈ ωr (SDOF identification method). The damping ratio ϑr must
on the one hand be small enough (approximately ϑ ≤ 0.1) for the influence of the
r-th eigenform in the frequency response to dominate at the resonance point. On the
other hand, it should not be so small that accurate measurements are not possible.

Idealization as Vibrator with One Degree of Freedom

A simple method completely ignores the eigenforms j = r at ωGr chooses the
approximation of a system with one DOF. Therefore, in Eq. (292), only:

url urn /mr

H lnr = = H ln (294)
ωr2 − 2 + i2ϑr ωr 

is taken into account (Fig. 73).

Fig. 73 Value of the frequency response of a vibrating system in the environment of the natural
frequency ωr (source: VDI guideline 3830)
12 Damping of Materials and Structures 777

Approximative Inclusion of the Other Degrees of Freedom

The sum in Eq. (292) can be split into H lnr as in Eq. (294) and the component
covering other eigenforms where j = r

 uj l uj n /mj
Alnr (i) = . (295)
j =l,j =r
ωj2 − 2 + i2ϑj ωj 

Thus, we have:

H̃ ln (i) = H̃ lnr (i) + Ãlnr (i). (296)

Here only a short description of the basic idea of the above two subheadings is
covered. Refer [94] for an exhaustive guideline.

5.4.4 Approximation of Measured Frequency Responses in an Interval

with Several Resonance Points
The above mentioned methods are not applicable in many cases such as when
resonance frequencies lie close to each other. The accuracy obtained is not sufficient
if the measured frequency response has several resonance points within a frequency
interval. Cases of this call for methods which simultaneously take into account the
dynamic behavior of several natural modes of vibration – so-called MDOF (multiple
degree of freedom) methods. Many methods for such phenomenon have already
been developed [84, 90]. Here a brief outline of a select few methods is presented.
A common assumption will be that there exist multiple natural frequencies ωj where
j ∈ [J1 , J2 ].

Incomplete Equivalent Model

Instead of the incomplete frequency response:

J2
uj l uj n /mj
H ln (iΩ) = (297)
ω2
j =J1 j
− 2 + i2ϑj ωj 

a better approximation


J2
uj l uj n /mj 1 1
H ln (i) = + − (298)
ω2
j =J1 j
− 2 + i2ϑj ωj  Kln Mln 2

with the constants Kln and Mln is used for the equivalent model. This approximation
covers all modal DOF where ωj > ωJ2 .

Generalization of the Method for Isolated Resonance Points

The accuracy of the method can be improved by incorporating a corrected frequency
response to cover isolated resonance points:
778 L. Gaul and A. Schmidt

⎡ ⎤
cor 
J2
uj l uj n /mj 1 1 ⎦
H̃ nr (i) = H̃ ln (i) − ⎣ + − .
j =J1 ,j =r
ωj2 − 2 + i2ϑj ωj  Kln Mln 2
(299)

General Approximation of the Frequency Response

The general approximation fits the frequency response H ln (i) of the computa-
tional model to the measured frequency response H ln (i). The values H̃ ln (ip )
at the frequencies p where p = 1, . . . , P are known from measurement. For the
equivalent model the following applies at these points:


J2
uj l uj n /mj 1 1
H ln (ip ) = + − . (300)
j =J1
ωj2 − 2p + i2ϑj ωj p Kln Mln 2p

In an iterative manner, the quantities Kln , Mln , ωj , ϑj , and uj l uj n /(2ϑj kj ) are

determined ∀j ∈ [J1 , J2 ]. Then the error is determined and minimized as per the
following equations:

ep = |H̃ ln (ip ) − H ln (ip )|2 (301)


P
e= wp ep (302)
p=1

∂e
=0 → Kln , Mln , ωj , ϑj uj l uj n /(2ϑj kj ). (303)
∂x

5.5 Experimental Techniques for Measuring Soil Damping

Soil damping is composed of the material damping of the ground and the energy
extraction due to wave radiation (geometric damping). As there are no methods to
determine energy extraction due to wave radiation, total soil damping and material
damping are separately defined. Here, a brief outline of testing methods is given. For
a more detailed description refer [94]. Among individual tests, boundary conditions
should be ensured to be the same as material damping parameters greatly depend
on the

• deformation amplitude
• mean hydrostatic pressure
• density
• void ratio
• load cycles
12 Damping of Materials and Structures 779

Geometric damping, on the other hand, depends only on the frequency and is
calculated as a difference between total and material damping.
Measurement techniques broadly fall into the category of either seismic, where
the propagation of an impulse or equivalent signals is employed or stationary
methods, where a wave source is used to generate a standing wave field.
In the field, the following methods are suitable for determining soil damping:

• cross-hole measurement
• Rayleigh wave dispersion measurement
• measurement with model foundations

Laboratory measurement techniques for determining the material damping of soil

samples are limited to the analysis of stationary vibrations using the:

• triaxi apparatus
• resonant column apparatus
• simple shear apparatus
• ring shear apparatus

In these experiments, care should be taken to remove soil samples such that the
mechanical properties of the subsoil are unchanged. Hydrostatic pressure can be
simulated in the test equipment.

6 Application of Fractional Calculus to Viscoelastically

Damped Structures in the Finite Element Method

All materials show some amount of material damping. Under time periodic or Heav-
iside stress/strain loads – hysteresis or creep/relaxation responses can be observed,
respectively. Damping in the case of many rubbers and polymers can be high and
as a result may not be ignored, while material modeling. Material damping may
be modeled by differential operators or hereditary integral viscoelastic constitutive
equations. It is generally sufficient to use a linear stress-strain relationship – which
is often realized via linear sprin and viscous dashpots. These result in constitutive
equations of integer-order differential operator type. However, when applied to large
time or frequency intervals, these models are cumbersome. Thermodynamically
consistent and more robust models can be formulated using fractional derivatives.
Refer [95,96,97,98,99,101]. An implementation of fractional constitutive equations
into finite element and parameter identification is given in [103, 105]. Boundary
element implementations can be found in [100, 100].

6.1 Grünwald Definition of Fractional Derivatives

Using backward difference, the first (integer order) derivative is given by:
780 L. Gaul and A. Schmidt

d1 f (t) 1
1
= lim [f (t) − f (t − t)]. (304)
dt t→0 t

By the same rule, we may define any integer-order derivative by the following
relation:
⎡ ⎤
n n  
d f (t) 1 n
= lim ⎣ (−1)j f (t − j t)⎦ (305)
dt n t→0 (t)n j
j =0

 
n
where are binomial coefficients. If we replace the time step t by a fraction
j
t
with N = 1, 2, 3. . . , we can rewrite the above equation as:
N
⎡ ⎤
 −n N −1  
dn f (t) t n t
= lim ⎣ (−1)j f (t − j )⎦ (306)
dt n N →∞ N j N
j =0

with the condition that

 
n
= 0 for j > n. (307)
j

A detailed treatment of the above equation can be found in [102, 104]. In order to
deduce a formulation that is valid for any real-order derivative, we use the extended
definition of the binomial coefficient:
⎧
  ⎨ a(a − 1)(a − 2) . . . (a − j + 1)
a for j > 0
= j (308)
j ⎩
1 for j = 0

where a is a real number and j is a natural number. For j > 0 therefore:

j factors
  ' () *
n n(n − 1)(n − 2) . . . (n − j + 2)(n − j + 1)
(−1)j = (−1)j
j j!
(j − n − 1)(j − n − 2) . . . (−n + 1)(−n)
=
j! (309)
 
j −n−1
=
j
(j − n)
≡
(−n)(j + 1)
12 Damping of Materials and Structures 781

such that  is the gamma function. Using the above equation in Eq. (306), we have:
⎡ ⎤
 −n N
−1
dn f (t) t (j − n) t
= lim ⎣ f (t − j )⎦ , n > 0.
dt n N →∞ N (−n)(j + 1) N
j =0
(310)
If we assume now, n to be any real number ν, we obtain the Grünwald definition of
fractional derivatives and integrals given by:
⎡ ⎤
 −ν N
−1
dν f (t) t t
= lim ⎣ Aj +1 f (t − j )⎦ (311)
dt ν N →∞ N N
j =0

where

(j − ν)
Aj +1 = (312)
(−ν)(j + 1)

are the so-called Grünwald coefficients. Here, as long as ν is not a positive integer,
Aj +1 is nonzero. If, e.g., ν = −1, then Aj +1 = 1 for all j , according to the Riemann
sum for integer-order integration. If ν is a natural number n, only the first n + 1
Grünwald coefficients Aj +1 are nonzero, indicating a local operator. Conversely,
fractional derivatives are non-local operators if ν is a positive non-integer. The limits
of the summation here are the terminal values. The lower limit j = 0 corresponds
−1)t
to A1 f (t), and the upper limit j = N − 1 corresponds to AN f (t − (N N ) =
t
AN f ( N ). Thus, the interval (0; t] is divided into N sections of equal size. Here, we
always assume the lower terminal to be 0. As in differential operator notation, we
may write:
⎡ ⎤
ν f (t)  −ν N
−1
d t t
0 Dt f (t) =
ν
= lim ⎣ Aj +1 f (t − j )⎦ = D ν . (313)
dt ν N →∞ N N
j =0

6.2 Numerical Calculation of Fractional Derivatives

Approximate numerical evaluations can be done by replacing the infinite by a finite

sum:
 −ν N
−1
t t
D f (t) ≈
ν
Aj +1 f (t − j ). (314)
N N
j =0

For reasons of numerical stability, calculation of Aj +1 is carried out using the

following recursive relationship:
782 L. Gaul and A. Schmidt

(j − ν) j − 1 − ν (j − 1 − ν) j −1−ν

Aj +1 = = = Aj . (315)
(−ν)(j + 1) j (−ν)(j ) j

It is also possible to show that:

lim |Aj +1 | = 0. (316)

j →∞

With growing j , the values Aj +1 are weighting function values that are situated
further in the past. This is why the influence of the past is faded out as time elapses,
which is attributed to the “fading memory” property. We use here, a time integration
scheme with fractional derivatives being evaluated at each increment. The time-
discrete function values f (t − j Nt ) that are needed to evaluate the fractional
derivative are then computed from the history of the time integration. Also, time step
for the fractional derivative is set equal to the time step size for the time integration.
This means that in the beginning of the evaluation (N = 1), there are no history
values to be taken into account. But as time elapses, the calculation of the fractional
derivative slows down as it has to take more and more history data into account.
Also, as a result of more history data, storage requirements increase. Shown below
are three concepts that may be used to accelerate the computational process and
reduce the storage requirements.

Concept 1. Due to the property of fading memory, the history data beyond a given
point in time is not of much significance and therefore motivate the truncation of
Eq. (314) such that:

 −ν 
Nl
t t
D f (t) ≈
ν
Aj +1 f (t − j ), Nl > N − 1. (317)
N N
j =0

We see that only Nl sample points of the past are taken into consideration. Therefore,
computational time and storage requirement first increase and then remain constant
for rest of the time integration.

Concept 2. Another alternative is to choose time steps for fractional derivatives as

a multiple of the time step for time integration.
t
= ct, c > 1. (318)
N

If c = 2 then compared to the number of evaluations for the time integration, the
number of evaluations for the fractional derivatives is half and consequently, also the
storage requirements. It is to be noted that the accuracy of the fractional derivatives
drops as higher values of c is chosen. However, stronger requirements exist on t
than on the time step for fractional derivatives; refer [103]. Therefore c = 2 is a
reasonable compromise.
12 Damping of Materials and Structures 783

Concept 3. In this case, we approximate newer history with high resolution while
the older one with lower resolution. We start with Eq. (314) and use the recursive
relation as in Eq. (315). Therefore, at time t0 , we have:
⎡
 −ν i 
t0 ⎣ i−ν+1
D ν f (t) ≈ Aj +1 fj + Ai+2 fi+1 + f1+2
N i+2
j =0

(i − ν + 1)(i − ν + 2)
+ fi+3 + . . . (319)
(i + 2)(i + 3)
⎡ ⎤
 −ν i
! t0 ⎣
= Aj +1 fj + Ai+2 Ti+1 ⎦ .
N
j =0

t0
In the above equation, fj is the abbreviation for f (t0 − j N ). Also, Ti+1 represents
the whole interval as one contribution to the factional derivative as shown in
Fig. 74. Keeping the upper and lower limit constant for the time integration part
(i = const.)- n time increments later, i.e., at time t0 + nt the fractional derivative
has the form:
⎡
 −ν  i+n 
t0 ⎣ i+n−ν+1
D ν f (t) ≈ Aj +1 fj −n + Ai+n+2 fi+1 + fi+2
N i+n+2
j =0

(i + n − ν + 1)(i + n − ν + 2)
+ fi+3 + . . .
(i + n + 2)(i + n + 3)
⎡ ⎤
 −ν i+n
t0 ⎣
= Aj +1 fj −n + Ai+n+2 Ti+n+1 ⎦ .
N
j =0
(320)

For viscoelastic problems, ν is positive. For i > ν − 1, the value of all weighting
factors in T range between 0 and 1. In addition, all weighting factors tend to unity
for increasing elapsed time, while at the same time, the weighting factor Ai+n+2
tends to zero (fades out). We can approximate the time dependent quantity Ti+n+1
as we know the starting value Ti+1 at time t0 , the value R∞ for t → ∞, and the
unity function f 1 = 1. Thus, we have:

Fig. 74 Time axis; definition of the interval

784 L. Gaul and A. Schmidt

1
Ti+n+1 − Ti+1
1
!
Ti+n+1 = Ti+1 + (T∞ −Ti+1 ) = Ti+1 +wi+n+1 (T∞ −Ti+1 ). (321)
1 −T1
T∞ i+1

The bold face indices on the upper right indicate that the value is calculated using
the unity function.
This concept is especially beneficial in methods utilizing spatial discretization
such as FEM. In the following section, we will calculate the fractional derivative
of stresses and strains – the fractional derivatives are needed for all nodal displace-
ments and the stresses at all integration points. Weighting factors w are calculated
once for each increment as it depends only on time. For each nodal displacement
and stress state at each integration point, only the values of Ti+1 and T∞ have to
be calculated and stored in order to approximate the influence of the time interval
under consideration on the fractional derivative. Also, when we have large numbers
of time increments, many intervals can be setup successively.
The quality of Approximation (Eq. (321)) is demonstrated by applying it to three
different functions:

1
f1 (t) = 1 f2 (t) = at f3 (t) = sin(π at) where a = (322)
s

and comparing it to Eq. (314). Figure 75 shows the comparison. The time step used
t0
is N = t = 0.05s i = 20 time increments, and the order of the derivative is
ν = 0.5. The value of T is then computed for the next 500 time increments. Values
from Eqs. (320) and (321) are depicted as solid and dashed lines, respectively.

Fig. 75 Approximation of fractional derivatives using concept C

12 Damping of Materials and Structures 785

Fig. 76 Rheological elements of viscoelasticity

6.3 Fractional-Order Constitutive Equations

The constitutive equations of linear viscoelastic can be deducend from rheological

models consisting of springs and ‘spring-pots’ (Fig. 76). The stress-strain relation
of the latter may be written as:

σ = pD ν ε. (323)

In the above equation, if ν = 0, we get a spring with stiffness p. On the other hand,
for ν = 1 we obtain a dashpot with viscosity p. This means a fractional derivative
denotes a material behavior between that of a spring and a dashpot. This is denoted
as a Rhombus and was first introduced by Koeller [101]. The hybrid element is
called a “spring pot”. By replacing the dashpots in rheological models by spring-
pots, fractional rheological models are derived. Application to the 5-parameter
model (two Maxwell elements and a spring in parallel) results in the “fractional
5-parameter model,” see Fig. 77. The constitutive equation is given by:

p1 ν1 p2 ν2 p1 p2 ν1 +ν2
σ+ D σ+ D σ+ D σ
E1 E2 E1 E2
E0 + E1 ν1 E0 + E2 ν2
= E0 ε + p1 D ε + p2 D ε (324)
E1 E2
E0 + E1 + E2 ν1 +ν2
+ p1 p2 D ε.
E1 E2

Here, all fractional order initial conditions are assumed to be zero referring to a
material that is totally relaxed at t = 0. A three-dimensional extension to Eq. (324)
can be referred to in [105].

6.4 Finite Element Formulation and Implementation

The displacement-type formulation of the FEM is given by:

u = H û (325)
786 L. Gaul and A. Schmidt

Fig. 77 5-parameter model

where u denotes the displacement field of an element, û is the vector of the nodal
displacements, and H is the matrix of shape functions. The strain is given by:

ε = B û. (326)

The matrix B consists of the spatial derivatives of H . From the principle of virtual
work, we have the equation of motion as:


B T σ dR + M û¨ = r (327)
R

where R is the region in which the element is defined, and r defines the external and
the body forces. M is the consistent mass matrix given by the relation:


M= H T ρH dR. (328)
R

Here ρ is the mass density of the material. For convenience and readability ˆ· notation
is dropped, and the equation of motion at time t (as upper left index) is given by:

B T t σ dR + M t ü = t r. (329)
V

tσ is the stress vector derived from Eq. (324). If we apply the time discrete
Grünwaldian fractional derivatives to Eq. (324), we obtain:
12 Damping of Materials and Structures 787

       −1
t σ = 1 + p1 A(ν1 ) t −ν1 p2 (ν2 ) t −ν2 p1 p2 (ν1 +ν2 ) t −ν1 −ν2
+ A + A .
E1 1 N E2 1 N E1 E2 1 N
⎡
   
t −ν1  (ν1 ) t−j t E0 + E2 t −ν2
N
⎣E0 t ε + p1 E0 + E1 Aj +1 N ε + p2
E1 N E2 N
j =0
 
 t −ν1 −ν2  (ν1 +ν2 ) t−j Nt
N N
(ν2 ) t−j Nt E + E1 + E2
Aj +1 ε + p1 p2 0 Aj +1 ε
E1 E2 N
j =0 j =0
   −ν2 
t −ν1  (ν1 ) t−j nt
N N
p1 p t (ν2 ) t−j Nt
− Aj +1 σ− 2 Aj +1 σ
E1 N E2 N
j =1 j =1
⎤
 
t −ν1 −ν2  (ν1 +ν2 )t−j Nt ⎦
N
p1 p2
− Aj +1 σ .
E1 E2 N
j =1
(330)

In Eq. (330), the upper-right indices in brackets indicate the dependence of the Grün-
wald coefficients on the order of respective fractional derivative. Note, Eq. (330)
depends on the actual strain, the strain history, and the stress history. If we insert
Eq. (330) into Eq. (329) and replace the strains ε by Eq. (326), the resulting equation
of motion can be transformed into:

M t ü + K ∗t u = t r ∗ (331)

where K ∗ and r ∗ are the modified stiffness matrix and the modified force vector,
respectively. Further reference to this can be seen in [105]. Regarding the form of
Eq. (331), it can be solved with any elastic FE solver in conjunction with either
implicit and explicit integration schemes.

6.5 Parameter Identification: A Case Study with DelrinT M

Here, a parameter identification for the polymer DelrinT M is carried out in the time
domain and in the frequency domain simultaneously. The time-dependent behavior
is given in terms of the creep modulus Ec in the range from 10 s up to 10,000 h.
Besides the measurements of the manufacturer, own measurements have been car-
ried out to cover the short time period smaller than 360 s. In addition, free decay tests
of a cantilever made of DelrinT M have been carried out at 12 different frequencies
in the range from 50 Hz up to 500 Hz. The oscillations were measured by a laser
vibrometer, and the frequency-dependent complex modulus was calculated.
The fractional 5-parameter model as shown in Fig. 77 has seven material
properties which are determined by a least-squares fit method. While the complex
modulus of the fractional rheological models can be calculated analytically [105],
788 L. Gaul and A. Schmidt

Fig. 78 Identified parameters in the time and frequency domain

Fig. 79 Comparison between measured data and fractional 5-parameter model

the time-dependent behavior is evaluated by numerical time integration in each

iteration step. The material parameters are given in Fig. 78, while the time- and
frequency-dependent material behavior is compared to the measurements in Fig. 79.

6.6 Finite Element Calculations and Comparison of the Different

Concepts

A finite element implementation is realized via 8-noded isoparametric brick ele-

ments in MARC. Numerical and computational costs are reduced by using concepts
1, 2, and 3 as described in Sect. 6.2.
As an example, the free decay of a cantilever made of DelrinT M is calculated,
where the identified fractional 5-parameter model is used. The FE model is shown
in Fig. 80. At the left side, fixed displacement type of boundary conditions is applied
to the first two rows of nodes in order to model the fixed support. The free length
of the cantilever is 100 mm, while its cross-section measures 2.2 mm × 10 mm. The
model is initially at rest and in the first increment; a steady force in y-direction is
applied to the last row of nodes at the right-hand side. Thus, an oscillation about
the position of equilibrium is excited, while the position of equilibrium moves with
time due to material creep. Since the height of the model has no influence on either
the vibration and then creep, only one row of elements is modeled. The calculation
is continued for 2000 time increments of 0.001 s using the Newmark integration
method without numerical damping. Calculations are performed with the “original”
approximation of fractional derivatives Eq. (314) as the “reference calculation”
and with the concepts 1, 2, and 3. The results of the decaying oscillation can be
compared to the respective measurements, while the superposed creep behavior can
be compared to the theoretical curve from numerical integration.
12 Damping of Materials and Structures 789

Fig. 80 FE model and calculated tip deflection

Fig. 81 Computational requirements for the different concepts

Fig. 82 Creep behavior of

the cantilever

The CPU time needed and the storage requirements are summarized in the
table labeled Fig. 81. The creep behavior shows significant deviation when different
concepts are compared with each other. The equilibrium state is available from the
calculation by means of values of each two consecutive maxima and a polynomial is
fitted. This is depicted in Fig. 82. While concepts 2 and 3 show very good agreement
to the theoretical calculations, concept 1 does not. The reason for this is that only the
most recent part of the history information is considered. After Nl time increments,
the errors accumulate during time integration.
Thus, from the above results, it can be inferred that for pure oscillations, concepts
1 and 3 result in considerable saving on the computational requirements while
being both stable and accurate. Concept 2, on the other hand, is susceptible to
numerical instability consequently giving rise to inaccuracies after a few hundred
time increments. Having said that, concept 1 performs badly when the creep
790 L. Gaul and A. Schmidt

behavior is examined. Concept 3, therefore, is the only candidate providing an

advantage in computational requirements without compromising on accuracy.

7 Conclusion

The term “damping” indicates the loss of mechanical energy of a system beyond its
physical boundaries. In most cases, damping implies the conversion of mechanical
energy into heat energy; however, other mechanisms such as the transformation into
electric energy by piezoelectric materials, propagating waves through the system’s
boundaries, or radiation into surrounding media may also cause energy losses. This
chapter mainly dealt with damping effects of materials and assembled structures by
means of phenomenological effects, modeling aspects, and experimental techniques.
The measurement and modeling of damping are challenging tasks that are still
current research topics. As long as the material damping has to be determined,
the damping properties are mainly linear resulting in somewhat simple damping
models which do not have to account for any local effects. Additionally, a variety of
techniques are available to accurately determine the amount of damping depending
on the desired parameters such as frequency or temperature. Special materials,
such as rubbers or plastics, show an extensive dependency on the temperature,
which might be modeled using a temperature-frequency shift function that holds
for thermal-rheologically simple materials. Therefore, instead of measuring the
frequency-dependent damping properties over a broad frequency range, one can
restrict to a limited frequency band which is used for different temperatures
and results in a master curve, consequently covering an extremely broad range.
As long as the amount of damping is equally distributed within the material
under consideration, a modal analysis leads to complex eigenvalues indicating the
presence of damping but real-valued eigenmodes.
As soon as the damping properties of a structure assembled from different
components are considered, its occurrence and extent usually are mainly influenced
by its joints instead of the respective material damping. In fact, the amount of
material damping typically ranges between 1% and 10% relative to the overall
loss of energy. In contrast to material damping, dissipative effects in the assembled
structures mainly occur locally which consequently leads to complex eigenmodes.
Thus, respective damping models must account for the location of the underlying
damping effects. In addition, linearity is not fulfilled anymore, and the applicability
of linear techniques such as the modal analysis must be verified carefully. The
physical background of the nonlinearities is given by the occurrence of macro slip
in the interface of the different components. As long as the contact pressure of all
joints is large enough to guarantee the exclusive existence of micro slip for a given
load level, the assumption of linearity is justified. However, the contact pressure of
most interfaces to high extent depends on the distance between the origin of the
contact force (e.g., the location of bolts, rivets, weld points). As a consequence,
regions of micro slip and macro slip exist in parallel whereupon their extension
changes with the current load level. Therefore, respective mechanical models have
12 Damping of Materials and Structures 791

to be solved in the time domain making use of suitable friction models resulting in
a high computational effort. A general problem for the detection of damping values
of assembled structures is given by the lack of reproducibility. After disassembling
and reassembling a structure under consideration, its damping properties in general
show a substantial deviation.

Technical Standards

DIN 1311-1: 2000-02 Schwingungen und schwingungsfähige Systeme – Teil

1: Grundbegriffe, Einteilung ((Mechanical) vibrations, oscillation and vibration
systems – Part 1: Basic concepts, survey). Berlin: Beuth Verlag
DIN 1311-2: 2002-08 Schwingungen und schwingungsfähige Systeme – Teil
2: Lineare, zeitinvariante schwingungsfähige Systeme mit einem Freiheitsgrad
((Mechanical) vibrations, oscillations and vibration systems - Part 2: lineartime-
invariant vibration systems with single degree of freedom). Berlin: Beuth Verlag
DIN 1311-3: 2000-02 Schwingungen und schwingungsfähige Systeme – Teil
3: Lineare, zeitinvariante schwingungsfähige Systeme mit endlich vielen Freiheits-
graden ((Mechanical) vibrations, oscillation and vibration systems – Part 3: linear
time-invariant vibration systems with a finite number of degrees of freedom). Berlin:
Beuth Verlag
DIN 1311-4: 1974-02 Schwingungslehre – Schwingende Kontinua, Wellen
(Vibration; vibrating continua, waves). Berlin: Beuth Verlag
DIN 5487: 1988-07 Fourier-, Laplace- und Z-Transformation – Zeichen und
Begriffe (Fourier-, Laplace- and Z-transformation; symbols and concepts). Berlin:
Beuth Verlag
DIN 13 342: 1976-07 Nicht-newtonsche Flüssigkeiten – Begriffe, Stoffgesetze
(Non-newtonian liquids; concepts, rheological equations). Berlin: Beuth Verlag
DIN 13 343: 1994-04 Linear-viskoelastische Stoffe – Begriffe, Stoffgesetze,
Grundfunktionen (linear viscoelastic materials; concepts, constitutive equations,
basic functions). Berlin: Beuth Verlag
DIN 51 563: 1976-12 Prüfung von Mineralölen und verwandten Stoffen – Bes-
timmung des Viskosität-Temperatur-Verhaltens – Richtungskonstante m (Testing
of Mineral Oils and Related Materials; Determination of Viscosity Temperature
Relation; Slope m). Berlin: Beuth Verlag
DIN 53 017: 1993-11 Viskosimetrie – Bestimmung des Temperaturkoeffizienten
der Viskosität von Flüssigkeiten (Viscosimetry; determination of the temperature
coefficient of viscosity of liquids). Berlin: Beuth Verlag
DIN 53 018-1: 1976-03 Viskosimetrie – Messung der dynamischen Viskosität
newtonscher Flüssigkeiten mit Rotationsviskosimetern – Teil 1: Grundlagen (Vis-
cosimetry; Measurement of the Dynamic Viscosity of Newtonian Fluids with
Rotational Viscosimeters; Part 1: Principles). Berlin: Beuth Verlag
DIN 53 018-2: 1976-03 Viskosimetrie – Messung der dynamischen Viskosität
newtonscher Flüssigkeiten mit Rotationsviskosimetern – Teil 2: Fehlerquellen und
Korrektionen bei Zylinder-Rotationsviskosimetern (Viscosimetry; Measurement of
792 L. Gaul and A. Schmidt

the Dynamic Viscosity of Newtonian fluids with Rotational Viscosimeters; Part 2:

Sources of Errors and Corrections concerning Cylinder; Rotation Viscosimeters).
Berlin: Beuth Verlag
DIN 53 019-1: 1980-05 Viskosimetrie – Messung von Viskositäten und Fließkur-
ven mit Rotationsviskosimetern mit Standardgeometrie – Teil 1: Normalausführung
(Viscosimetry; Determination of viscosities and flow curves using standard design
rotary viscosimeters with a standard geometry measuring system). Berlin: Beuth
Verlag
DIN 53 019-2: 2001-02 Viskosimetrie – Messung von Viskositäten und Fließkur-
ven mit Rotationsviskosimetern mit Standardgeometrie – Teil 2: Viskosimeterkalib-
rierung und Ermittlung der Messunsicherheit (Viscosimetry – Measurement of
viscosities and flow curves by means of rotation viscosimeters – Part 2: Viscosimeter
calibration and determination of the uncertainty of measurement). Berlin: Beuth
Verlag
DIN 53 513: 1990-03 Prüfung von Kautschuk und Elastomeren – Bestimmung
der viskoelastischen Eigenschaften von Elastomeren bei erzwungenen Schwingun-
gen außerhalb der Resonanz (Determination of the viscoelastic properties of
elastomers on exposure to forced vibration at non-resonant frequencies). Berlin:
Beuth Verlag
DIN 53 535: 1982-03 Prüfung von Kautschuk und Elastomeren – Grundlagen
für dynamische Prüfverfahren (Testing of rubber; general requirements for dynamic
testing). Berlin: Beuth Verlag
DIN EN ISO 6721-1: 2003-01 Kunststoffe – Bestimmung dynamisch- mechanis-
cher Eigenschaften – Teil 1: Allgemeine Grundlagen (ISO 6721-1:2001). Deutsche
Fassung der EN ISO 6721-1:2002 (Plastics – Determination of dynamic mechanical
properties – Part 1: General principles (ISO 6721-1:2001); German version EN ISO
6721-1:2002). Berlin: Beuth Verlag
DIN EN ISO 6721-2: 1996-12 Kunststoffe – Bestimmung dynamisch- mech-
anischer Eigenschaften – Teil 2: Torsionspendel-Verfahren. (ISO 6721-2:1994)
Deutsche Fassung der EN ISO 6721-2:1996 (Plastics – Determination of dynamic
mechanical properties – Part 2: Torsion-pendulum method (ISO 6721-2:1994,
including Technical Corrigendum 1:1995); German version EN ISO 6721-2:1996).
Berlin: Beuth Verlag
DIN EN ISO 6721-3: 1996-12 Kunststoffe – Bestimmung dynamisch- mechanis-
cher Eigenschaften – Teil 3: Biegeschwingung, Resonanzkurven- Verfahren. (ISO
6721-3:1994) Deutsche Fassung der EN ISO 6721-3:1996 (Plastics – Determination
of dynamic mechanical properties – Part 3: Flexural vibration; resonance-curve
(ISO 6721-3:1994, including Technical Corrigendum 1:1995); German version EN
ISO 6721-3:1996). Berlin: Beuth Verlag
ISO/DIS 18 437-2: 2003-04 Mechanical Vibration and shock – Characterisation
of the dynamic mechanical properties of resilient materials – Part 2: Resonance
method
ISO/DIS 18 437-3: 2003-04 Mechanical Vibration and shock – Characterisation
of the dynamic mechanical properties of resilient materials – Part 3: Single
cantilever beam method
12 Damping of Materials and Structures 793

ISO/DIS 18 437-4 Mechanical Vibration and shock – Characterisation of the

dynamic mechanical properties of resilient materials – Part 4: Impedance method
VDI Richtlinie 3830 Blatt 1: 2004-08 Werkstoff- und Bauteildämpfung – Ein-
teilung und Übersicht (Part 1: Damping of materials and members – Classification
and survey). Berlin: Beuth Verlag
VDI Richtlinie 3830 Blatt 2: 2004-10 Werkstoff- und Bauteildämpfung –
Dämpfung in festen Werkstoffen (Part 2: Damping of materials and members –
Damping of solids). Berlin: Beuth Verlag
VDI Richtlinie 3830 Blatt 3: 2004-07 Werkstoff- und Bauteildämpfung –
Dämpfung von Baugruppen (Part 3: Damping of materials and members – Damping
of assemblies). Berlin: Beuth Verlag
VDI Richtlinie 3830 Blatt 4: 2005-05 Werkstoff- und Bauteildämpfung –
Modelle für gedämpfte Strukturen (Part 4: Damping of materials and members –
Models for damped structures). Berlin: Beuth Verlag
VDI Richtlinie 3830 Blatt 5: 2005-11 Werkstoff- und Bauteildämpfung –
Versuchstechniken zur Ermittlung von Dämpfungskenngrößen (Part 5: Damping of
materials and members – Experimental techniques for the determination of damping
characteristics). Berlin: Beuth Verlag
VDI Richtlinie 3833 Blatt 1: 2003-12 Schwingungsdämpfer und Schwingungstil-
ger – Schwingu- ngsdämpfer – Begriffe, Kenngrößen, Realisierung, Anwendung
(Dynamic damper and dynamic vibration absorber – Dynamic damper – Terms,
characteristics implementation, application). Berlin: Beuth Verlag
VDI Richtlinie 3833 Blatt 2 (Entwurf): 2004-04 Schwingungstilger und
Schwingungsdämpfer – Schwingungstilger und Schwingungstilgung (Dynamic
damper and dynamic vibration absorber – Dynamic vibration absorber and dynamic
vibration absorption). Berlin: Beuth Verlag

References
1. Bagley RL, Torvik PJ (1986) On the fractional calculus model of viscoelastic behavior. J Rheol
30(1):133–155
2. Christensen RM (1971) Theory of viscoelasticity – an introduction. Academic Press, New
York/San Francisco/London
3. Ferry JD (1980) Viscoelastic properties of polymers. J. Wiley & Sons, New York/
London
4. Göldner H (1978) Übungsbuch der höheren Fertigkeitslehre, Elastizitätstheorie – Plastizitäts-
theorie – Viskoelastizitätstheorie. Physik-Verlag, Weinheim
5. Kolsch H (1993) Schwingungsdämpfung durch statische Hysterese; Modellierung von
Bauteilen, Parameteridentifikation, Schwingungsberechnungen. Fortschr.-Ber. VDI Reihe 11
Nr. 190. Düsseldorf: VDI Verlag
6. Kümmlee H (1986) Ein Verfahren zur Vorhersage des nichtlinearen Steifigkeits- und Dämp-
fungsverhaltens sowie der Erwärmung drehelastischer Gummikupplungen bei stationärem
Betrieb. Fortschr.-Ber. VDI Reihe 1 Nr. 136. VDI Verlag, Düsseldorf
7. Lambertz S (1995) Spannungs-Dehnungsverhalten technischer Gummiwerkstoffe. Antrieb-
stechnisches Kolloquium ATK’95, S 1–35
8. Leaderman H (1943) Elastic and creep properties of filamentous materials and other high
polymers. Textile Foundation, Washington, p 175
794 L. Gaul and A. Schmidt

9. Lockett FJ (1972) Nonlinear viscoelastic solids. Academic Press, New York

10. Schapery RA (1969) On the characterization of nonlinear viscoelastic materials. Polym Eng
Sci 9(4):295–310
11. Steiner H (1997) Untersuchungen zum nichtlinear-viskoelastischen Materialverhalten von
Kunststoffen. Herbert Utz Verlag Wissenschaft, München
12. Tobolsky AV (1960) Properties and structure of polymers. J. Wiley & Sons, New York/London
13. Adams RD, Bacon DGC (1973) Measurement of the flexural damping capacity and dynamic
Young’s modulus of metals and reinforced plastics. J Phys D Appl Phys 6:27–41
14. Baker WE, Woolam WE, Young D (1967) Air and internal damping of thin cantilever beams.
Int J Mech Soc 9:743–766
15. Barr ADS, Johnston RA (1969) Acoustic and internal damping of uniform beams. J Mech Eng
Soc 11(2):117–127
16. Beards CF (1981) The damping of structural vibration by controlled interfacial slip in joints.
ASME Paper, No. 81-DET-86, pp 1–5
17. Beck CJ, Ikegami R, Johnson DW, Walker WJ (1985) Application of viscoelastic passive
damping to satellite equipment support structures. J Vib Acoust Stress Reliab Des 107(4):
367–374
18. Bohlen S, Gaul L (1984) Zum nichtlinearen Übertragungsverhalten von Fügestellen. ZAMM
64:T45–T48
19. Butenschön HJ (1986) Das hydrodynamische, zylindrische Gleitlager endlicher Breite unter
instationärer Belastung. Diss. TH Karlsruhe
20. Carbonell JR, Ungar EE (1966) On panel vibration damping due to structural joints. AIAA J
4(8):1385–1390
21. Christensen RM (1971) Theory of viscoelasticity – an introduction. Academic Press, New
York/San Francisco,London
22. Cremer L, Heckl M (1996) Körperschall, 2. Auflage. Springer, Berlin
23. Fahy F (1985) Sound and structural vibration; radiation, transmission and response. Academic
Press, London
24. Fallen M, Henn H, Sinambri GR (1984) Ingenieurakustik. F. Vieweg & Sohn, Wiesbaden,
Braunschweig
25. Feit D, Junger MC (1972) Sound, structures and their interaction. MIT Press, Cambridge
26. Ferry JD (1980) Viscoelastic properties of polymers. J. Wiley and Sons, New York/London
27. Gaul L (1981) Zur Dämmung und Dämpfung von Biegewellen an Fügestellen. Ing Archiv
51:101–110
28. Gaul L (1983) Zum Einfluss von Schraub-, Niet- und Klemmverbindungen auf die Dynamik
von Maschinen und Strukturen. In: Maschinenbau FB (Hrsg) Festschrift 10 Jahre Hochschule
der Bundeswehr Hamburg. Fachbereich Maschinenbau, S 49–58
29. Gaul L, Lenz J (1997) nonlinear dynamics of structures assembled with bolted joints. Acta
Mech 125:169–181
30. Gaul L, Lenz J, Sachau D (1998) Active damping of space structures by contact pressure
control in joints. Mech Struct Mach 26:81–100
31. Gaul L, Lenz J, Willner K (1994) Modellbildung des dynamischen Übertragungsverhaltens
von Fügestellen. Techn. Report, Arbeitsbericht zum DFG-Projekt Az. Ga209/11-1
32. Gaul L, Willner K (1995) A penalty approach for contact description by FEM based on
interface physics. In: Proceedings of CONMEC’95, Ferrara, pp 257–264
33. Glienicke J (1966) Feder-Dämpfungskonstanten von Gleitlagern für Turbomaschinen und
deren Einfluss auf das Schwingungsverhalten eines einfachen Rotors. Dissertation, Technische
Hochschule Karlsruhe, Karlsruhe
34. Glienicke J (1993) Nichtlineare Schwingungen allgemeiner gleitgelagerter Rotoren. VDI-
Berichte, Nr 1082. VDI Verlag, Düsseldorf, S 299–316
35. Glienicke J, Schwer M (1985) Kennlinien realer Quetschöldämpfer. Konstruktion 37(8):
301–308
36. Göldner H (1978) Übungsbuch der höheren Festigkeitslehre, Elastizitä tstheorie – Plastizitäts-
theorie – Viskoelastizitätstheorie. Physik-Verlag, Weinheim
12 Damping of Materials and Structures 795

37. Gomperts MC (1977) Sound radiation from baffled, thin, rectangular plates. Acustica
37(2):93–102
38. Heckl M (1962) Measurement of absorption coefficients on plates. J Acoust Soc Am
34(6):803–806
39. Holmer CI, Vér IL (1971) Interaction of sound waves with solid structures. In:
Beranek LL (ed) Noise and vibration control, pp 270–361. McGraw-Hill Book Company,
New York
40. Ingard KU, Morse PM (1968) Theoretical acoustics. McGraw-Hill Book Company, New York
41. Jeyapalan RK, Richards EJ, Wescott MW (1979) On the prediction of impact noise; Part II:
ringing noise. J Sound Vib 65(3):417–451
42. Kümmlee H (1986) Ein Verfahren zur Vorhersage des nichtlinearen Steifigkeits- und Dämp-
fungsverhaltens sowie der Erwä rmung drehelastischer Gummikupplungen bei stationärem
Betrieb. Fortschr.-Ber. VDI Reihe 1, Nr 136. VDI Verlag, Düsseldorf
43. Lambertz S (1995) Spannungs-Dehnungsverhalten technischer Gummiwerkstoffe. Antrieb-
stechnisches Kolloquium ATK ’95, S 1–35
44. Lang OR, Steinhilper W (1978) Gleitlager. Springer, Berlin/Heidelberg/New York
45. Lazan BJ (1968) Damping of materials and members in structural mechanics. Pergamon Press,
Oxford
46. Lockett FJ (1972) Nonlinear viscoelastic solids. Academic Press, New York
47. Lund JW (1979) Evaluation of stiffness and damping coefficients of fluid-film bearings. Shock
Vibration Digest 11(1):5–10
48. Maidanik G (1962) Response of ribbed panels to reverberant acoustic fields. J Acoust Soc Am
34(6):809–826
49. Maidanik G (1966) Energy dissipation associated with gas-pumping in structural joints. J
Acoust Soc Am 40(5):1064–1072
50. Merker H-J (1981) Über den nichtlinearen Einfluss von Gleitlagern auf die Schwingungen von
Rotoren. Fortschr.-Ber. VDI Z, Reihe 11, VDI-Forschungsheft, Nr 40. VDI Verlag, Düsseldorf
51. Morland LW, Lee EH (1960) Stress analysis for linear viscoelastic materials with temperature
variation. Trans Soc Rheol 4:233
52. Ottl D (1981) Schwingungen mechanischer Systeme mit Strukturdämpfung. VDI-
Forschungsheft, Nr 603. VDI Verlag, Düsseldorf
53. Petersmann N (1986) Substrukturtechnik und Kondensation bei der Schwingungsanalyse.
Fortschr.-Berichte VDI Reihe 11, Nr 6. VDI Verlag, Düsseldorf
54. Schapery RA (1969) On the characterization of nonlinear viscoelastic materials. Polym Eng
Sci 9(4):295–310
55. Schwarzl F, Staverman AJ (1952) Time-temperature dependence of linear viscoelastic behav-
ior. J Appl Phys 23:838–843
56. Schwarzl FR (1990) Polymermechanik. Springer, Berlin
57. Schwer M (1986) Eigenschaften von Quetschöldämpfern. Diss. TH Karlsruhe. TH Karlsruhe
58. Siegl G (1981) Das Biegeschwingungsverhalten von Rotoren, die mit Blechpaketen besetzt
sind. Diss. TU Berlin. TU Berlin
59. Someya T (1989) Journal-bearing handbook. Springer, Berlin
60. Steiner H (1997) Untersuchungen zum nichtlinear- viskoelastischen Materialverhalten von
Kunststoffen. Herbert Utz Verlag Wissenschaft, München
61. Ungar EE (1980) Damping of panels due to ambient air. In: Torvik PJ (ed) Damping applica-
tions in vibration control, vol. 38, pp 73–81. American Society of Mechanical Engineers, New
York
62. Valanis K (1971) A theory of viscoplasticity without a yield surface. Arch Mech 23(4):517–551
63. Wallace CE (1972) Radiation resistance of a rectangular panel. J Acoust Soc Am 51(3):
946–952
64. Weck M, Petuelli G (1981) Steifigkeits- und Dä mpfungskennwerte verschraubter Fügestellen.
Konstruktion 33:241–245
65. Willner K (1995) Ein statistisches Modell für den Kontakt metallischer Körper. Diss. Univ. der
Bundeswehr Hamburg, Univ. der Bundeswehr Hamburg
796 L. Gaul and A. Schmidt

66. Willner K (1997) Elasto-plastic contact of rough surfaces. In: Aliabadi M, Samartin A (ed)
Computational methods in contact mechanics III. Computational Mechanics Publications,
Madrid
67. Wissbrock H (1985) Untersuchungen zur Fugendämpfung zusammengesetzter Bauteile.
Fortschr.-Berichte VDI Reihe 11, Nr 68. VDI Verlag, Düsseldorf
68. Bathe KJ (1996) Finite element procedures. Prentice-Hall, New Jersey
69. Carpenter WM (1972) Viscoelastic stress analysis. Int J Numer Methods Eng 4:357–366
70. Caughey TK (1962) Vibration of dynamic systems with linear hysteretic damping. In:
Proceedings of the fourth U.S. National Congress of Applied Mechanics 1, pp 87–97
71. Caughey TK, O’Kelly ME (1965) Classical normal modes in damped linear dynamic systems.
J Appl Mech E 32:583–588
72. Gaul L, Fiedler C (1997) Methode der Randelemente in Statik und Dynamik. Friedrich Vieweg
& Sohn, Braunschweig/Wiesbaden
73. Gaul L, Schanz M (1999) A comparative study of three boundary element approaches to
calculate the transient response of viscoelastic solids with unbounded domains. Comput
Methods Appl Mech Eng 179:11–123
74. Gaul L, Wirnitzer J (1999) Calculation of transient wave propagation problems in semi-infinite
media by BEM in frequency domain. In: Proceedings of the 13th International Conference on
Boundary Element Technology (BETECH99), Las Vegas, pp 185–194
75. Krämer E (1984) Maschinendynamik. Springer, Berlin
76. Krings W (1976) Beitrag zur Finiten Element Methode bei linearem, viskoelastischem
Stoffverhalten. Mitteilungen aus dem Institut für Mechanik der Ruhr-Universität Bochum, Nr
3, Bochum
77. Link, M (1984) Finite Elemente in der Statik und Dynamik. Teubner-Verlag, Stuttgart
78. Natke HG (1992) Einführung in Theorie und Praxis der Zeitreihen-und Modalanalyse. 3.
Auflage. Vieweg-Verlag, Braunschweig
79. Zienkiewicz OC (1984) Methode der Finiten Elemente, 2. Auflage. Carl Hanser Verlag,
München/Wien
80. Zurmühl R, Falk S (1974) Matrizen und ihre Anwendungen, 6. Auflage. Springer-Verlag,
Berlin
81. Gaul L, Kögl M, Wagner M (2003) Boundary element methods for engineers and scientists.
An introductory course with advanced topics. Springer, Berlin
82. Broede, J (1983) Experimentelle Untersuchungen und analytische Betrachtungen zur Deutung
der Drehschwingungs-Eigendä mpfung in Kolbenmaschinen-Triebwerken. Dissertation, Tech-
nische Universität Berlin, Berlin
83. Dubbel (1990) Taschenbuch für den Maschinenbau, 17. Auflage. Springer, Berlin
84. Ewins DJ (1989) Modal testing: theory and practice, 5. Auflage. John Wiley & Sons,
London/New York
85. Ferry JD (1980) Viscoelastic properties of polymers. J. Wiley and Sons, London/New York
86. Lehr, E (1925) Die Abkürzungsverfahren zur Ermittlung der Schwingungsfestigkeit von
Materialien. Dissertation, Technische Hochschule Stuttgart, Stuttgart
87. Madigosky W (1984) Improved extensional modulus measurements for polymers and metal
matrix composites. In: Rogers L (ed) Proceedings Vibration damping 1984 Workshop Pro-
ceedings, Nr. AFWAL-TR-84-3064 in Q1–Q12
88. Madigosky, WM, Lee GF (1979) Automated dynamic Youngs’s modulus and loss factor
measurements. J Acoust Soc Am 66(2):345–349
89. Nashif AD (1967) New method for determining damping properties of viscoelastic materials.
Shock Vib Bull 36(Part 4):37–47
90. Natke HG (1992) Einführung in Theorie und Praxis der Zeitreihen-und Modalanalyse, 3.
Auflage. Vieweg-Verlag, Braunschweig
91. Norris DM, Young W-C (1970) Complex- modulus measurement by longitudinal testing. Exp
Mech 10:93–96
12 Damping of Materials and Structures 797

92. Oberst H, Frankenfeld K (1952) Über die Dämpfung der Biegeschwingungen dünner Bleche
durch festhaftende Beläge. Acustica 2, Akustische Beihefte, Heft 4, S 181–194
93. Pourabdolrahim R (1979) Entwicklung und systematische Untersuchung einer
Drehschwingungsprüfmaschine für umlaufende Maschinenelemente. Dissertation, Technische
Universität Berlin, Berlin
94. Verein Deutscher Ingenieure: damping of materials and members – Experimental techniques
for the determination of damping characteristics. VDI 3830 Blatt 5:2005–2011
95. Bagley RL, Torvik PJ (1979) A generalized derivative model for an elastomer damper. Shock
Vib Bull 49:135–143
96. Bagley RL, Torvik PJ (1983) Fractional calculus – a different approach to the analysis of
viscoelastically damped structures. AIAA J 21(5):741–748
97. Bagley RL, Torvik PJ (1983) A theoretical basis for the application of fractional calculus to
viscoelasticity. J Rheol 27(3):201–210
98. Bagley RL, Torvik PJ (1986) On the fractional calculus model of viscoelastic behavior. J Rheol
30(1):133–155
99. Caputo M, Mainardi F (1971) Linear models of dissipation in anelastic solids. Rivista del
Nuovo Cimento 1(2):161–198
100. Gaul L (1999) The influence of damping on waves and vibrations. Mech Syst Signal Process
13(1):1–30
101. Koeller RC (1984) Application of fractional calculus to the theory of viscoelasticity. J Appl
Mech 51:299–307
102. Oldham KB, Spanier J (1974) The fractional calculus. Academic Press, New York/London
103. Padovan J (1987) Computational algorithms for FE formulations involving fractional opera-
tors. Comput Mech 2:271–287
104. Podlubny I (1999) Fractional differential equations. Academic Press, San Diego/London
105. Schmidt A, Gaul L (2001) FE implementation of viscoelastic constitutive stress-strain rela-
tions involving fractional time derivatives. In: Constitutive Models for Rubber II. A.A.Balkema
Publishers, Tokyo, S 79–89
Modal Analysis of Nonlinear Mechanical
Systems 13
G. Kerschen and A. F. Vakakis

Contents
1 Nonlinear Normal Modes: A Brief Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . 800
2 Nonlinear Normal Modes: What Are They? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
2.1 Definition of a Nonlinear Normal Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
2.2 Fundamental Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805
3 Nonlinear Normal Modes: How to Compute Them? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
3.1 Analytical Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
3.2 Numerical Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819
3.3 Assessment of the Different Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
4 Nonlinear Normal Modes: Why Are They Useful? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
4.1 “Linear” Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
4.2 Nonlinear Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 821
4.3 Reduced-Order Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 826
5 Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 828
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 828

Abstract

The objective of this chapter is to introduce nonlinear normal modes (NNMs)

to structural dynamicists who are not acquainted with them. Specifically, this
chapter describes how the concept of modes can be extended to the nonlinear

G. Kerschen ()
Space Structures and Systems Laboratory, Department of Aerospace and Mechanical
Engineering, University of Liège, Liege, Belgium
e-mail: [email protected]
A. F. Vakakis
Department of Mechanical Science and Engineering, University of Illinois at Urbana-Champaign,
Champaign, IL, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 799

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_35
800 G. Kerschen and A. F. Vakakis

case. It also describes, in simple terms, the fundamental properties of NNMs,

including frequency-energy dependence, harmonics, bifurcation, and stability.

Keywords

Nonlinearity · Vibration modes · Natural frequencies

1 Nonlinear Normal Modes: A Brief Historical Perspective

The concept of a normal mode is central in the theory of linear vibrating systems.
Besides their obvious physical interpretation, the linear normal modes (LNMs) have
interesting mathematical properties. They can be used to decouple the governing
equations of motion; i.e., a linear system vibrates as if it were made of independent
oscillators governed by the eigensolutions. Two important properties that directly
result from this decoupling are:

1. Invariance: if the motion is initiated on one specific LNM, the remaining LNMs
remain quiescent for all time.
2. Modal superposition: free and forced oscillations can conveniently be expressed
as linear combinations of individual LNM motions.

In addition, LNMs are relevant dynamical features that can be exploited for
various purposes including model reduction (e.g., substructuring techniques [1]),
experimental modal analysis [2], finite element model updating [3], and structural
health monitoring [4].
Clearly, though, linearity is an idealization, an exception to the rule; nonlin-
earity is a frequent occurrence in real-life applications [5]. For instance, in an
aircraft, besides nonlinear fluid-structure interaction, typical nonlinearities include
backlash and friction in control surfaces, hardening nonlinearities in engine-to-
pylon connections, saturation effects in hydraulic actuators, plus any underlying
distributed nonlinearity in the structure. Furthermore, the next generations of
aircraft are using materials such as glass-fiber or carbon-fiber composites to a
greater extent for structural weight reduction. These materials entail new challenges
for performance prediction, because they exhibit a structural behavior deviating
significantly from linearity. Their increased use also creates more interfaces between
different materials, which are further sources of nonlinear behavior.
Any attempt to apply traditional linear analysis to nonlinear systems results, at
best, in a suboptimal design. Thus, there is a need for efficient, analytically rigorous,
broadly applicable analysis techniques for nonlinear structural dynamics. In this
context, nonlinear normal modes (NNMs) offer a solid theoretical and mathematical
tool for interpreting a wide class of nonlinear dynamical phenomena, yet they have
a clear and simple conceptual relation to the LNMs, with which practicing structural
engineers are familiar.
NNMs were pioneered in the 1960s thanks to the seminal work of Rosenberg
[6, 7, 8]. They were further studied in the 1970s by Rand [9, 10, 11] and Manevitch
13 Modal Analysis of Nonlinear Mechanical Systems 801

and Mikhlin [12]. They were regarded as a theoretical curiosity until the beginning
of the 1990s when they were given a new impetus through the efforts of Vakakis
et al. [13, 14, 15, 16, 17, 18] and Shaw and Pierre [19, 20, 21, 22]. Since then, a
large body of literature has addressed, with notable success, the qualitative and
quantitative analysis of nonlinear phenomena using NNMs (see, e.g., [23, 24, 25,
26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43]).
However, most structural engineers still view NNMs as a concept that is foreign
to them. There are at least two reasons supporting this statement:

1. Nonlinear systems can exhibit extremely complex behaviors which linear sys-
tems cannot. These phenomena include jumps, bifurcations; saturation; subhar-
monic, superharmonic, and internal resonances; resonance captures; limit cycles;
modal interactions; and chaos.
2. NNMs have two important limitations compared to their linear counterpart. First
and foremost, the principle of superposition, which is the cornerstone of linear
theory, does not apply to nonlinear systems. Second, the lack of orthogonality
relations satisfied by the NNMs complicates their exploitation as bases for order
reduction of the nonlinear dynamics.

The objective of the present chapter is thus to describe, in simple terms,

the fundamental properties of NNMs, including frequency-energy dependence,
harmonics, bifurcation, and stability. The chapter is organized as follows. In the
next section, the two main definitions of NNMs are provided. In addition, their
fundamental properties are described, and their representation in a frequency-energy
plot is introduced. In Sect. 3, the different means of computing the NNMs are briefly
reviewed and assessed. The potential applications of NNMs to “linear” and nonlin-
ear modal analysis and nonlinear model reduction are then discussed in Sect. 4.

2 Nonlinear Normal Modes: What Are They?

To illustrate the different concepts, a two-degree-of-freedom (2DOF) system with a

cubic stiffness is chosen. The system is depicted in Fig. 1, and its motion is governed
by the equations

1 1 1

1 1

0.5 x1 x2

Fig. 1 Schematic representation of the 2DOF system example

 802 G. Kerschen and A. F. Vakakis

ẍ1 + (2x1 − x2 ) + 0.5 x13 = 0

ẍ2 + (2x2 − x1 ) = 0 (1)

For comparison purposes, the underlying linear system

ẍ1 + (2x1 − x2 ) = 0
ẍ2 + (2x2 − x1 ) = 0 (2)

is also considered in this study. The time series corresponding to in-phase and
out-of-phase normal mode motions of the linear system (2) are depicted in Fig. 2.
Motion in the configuration space (i.e., in the plane of the displacements x1 (t) and
x2 (t)) is given in Fig. 3. Obviously, LNM motions correspond to straight lines in
this plane.

1 1

0.5 0.5
Displacement(m)

Displacement(m)

0 0

−0.5 −0.5

−1 −1
0 2 4 6 8 10 12 0 2 4 6

Time (s) Time (s)

Fig. 2 Time series of LNM motions of system (2) (——: x1 (t); − − −: x2 (t)). Left
plot: in-phase LNM ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [1 1 0 0]); right plot: out-of-phase LNM
([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [1 − 1 0 0])

1 1
Displacement x2 (m)

Displacement x2 (m)

0.5 0.5

0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Displacement x1 (m) Displacement x1 (m)

Fig. 3 LNM motions of system (2) in the configuration space. Left plot: in-phase LNM; right plot:
out-of-phase LNM
13 Modal Analysis of Nonlinear Mechanical Systems 803

2.1 Definition of a Nonlinear Normal Mode

There exist two main definitions of the NNMs in the literature, due to Rosenberg
[6,7,8] and Shaw and Pierre [19,20,21,22]. We note that a new, interesting definition
has been proposed recently [44], but it is not discussed herein.

2.1.1 Rosenberg’s Definition

During the normal mode motion of a linear conservative system, each system
component moves with the same frequency and with a fixed ratio among the dis-
placements of the components. Targeting a straightforward nonlinear extension of
the LNM concept, Rosenberg defined an NNM as a vibration in unison of the system
(i.e., a synchronous oscillation). This definition requires that all material points of
the system reach their extreme values and pass through zero simultaneously and
allows all displacements to be expressed in terms of a single reference displacement.
For illustration, time series corresponding to the free response to specific initial
conditions of system (1) are depicted in Fig. 4. These time series correspond
to synchronous motion, and, hence, to NNM motions according to Rosenberg’s
definition. The associated modal shapes in Fig. 5 are curves, resulting from the
nonlinear relationship between the coordinates during the periodic motion. These
curved NNMs, termed nonsimilar NNMs by Rosenberg, are generic in nonlinear
systems, and their existence certainly complicates the concept of orthogonality
between modes. When special spatial symmetries exist, the NNMs may degenerate
into (energy-invariant) straight modal lines (see [14] and Sect. 2.2.3).
At first glance, Rosenberg’s definition may appear restrictive in two cases:

1. This definition, as such, cannot be easily extended to nonconservative systems.

2. In the presence of modal interactions (i.e., when two NNMs interact), some
coordinates may have a dominant frequency component different than that of

15 15

10 10
Displacement(m)

Displacement(m)

5 5

0 0

−5 −5

−10 −10

−15 −15
0 2 4 6 8 0 0.5 1 1.5

Time (s) Time (s)

Fig. 4 Time series of NNM motions of system (1) (——: x1 (t); − − −: x2 (t)). Left plot: in-
phase NNM ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [3.319 11.134 0 0]); right plot: out-of-phase NNM
([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [−10.188 0.262 0 0])
 804 G. Kerschen and A. F. Vakakis

15 0.3

10 0.2
Displacement x2 (m)

Displacement x2 (m)
0.1
5
0
0
−0.1
−5
−0.2

−10 −0.3

−15 −0.4
−4 −2 0 2 4 −15 −10 −5 0 5 10 15

Displacement x1 (m) Displacement x1 (m)

Fig. 5 NNM motions of system (1) in the configuration space. Left plot: in-phase NNM; right
plot: out-of-phase NNM

60 60

40 40
Displacement x2 (m)
Displacement(m)

20 20

0 0

−20 −20

−40 −40

−60 −60
0 2 4 6 8 −10 −5 0 5 10
Time(s) Displacement x1 (m)

Fig. 6 Internally resonant NNM (3:1 internal resonance; [x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] =
[8.476 54.263 0 0]). Left plot: time series (——: x1 (t); − − −: x2 (t)); right plot: configuration
space

the other coordinates. In this case, the system may no longer be synchronous.
This is illustrated in Fig. 6 for a 3:1 internally resonant NNM of system (1).

However, these two limitations can be circumvented. On the one hand, the
damped dynamics can often be interpreted based on the topological structure of
the NNMs of the underlying undamped system. The concept of periodic motion
was also extended to damped systems in [45]. On the other hand, realizing that
the motion is still periodic in the presence of modal interactions, Rosenberg’s
definition can be extended to a (non-necessarily synchronous) periodic motion of the
system, as proposed in [29]. This extended definition is particularly attractive when
targeting a numerical computation of the NNMs. It enables the nonlinear modes to
be effectively computed using algorithms for the continuation of periodic solutions,
which are really quite sophisticated and advanced.
13 Modal Analysis of Nonlinear Mechanical Systems 805

2.1.2 The Invariant Manifold Approach

Shaw and Pierre proposed a generalization of Rosenberg’s definition that provides
a direct and elegant extension of the NNM concept to damped systems. Based on
geometric arguments and inspired by the center manifold technique, they defined
an NNM as a two-dimensional invariant manifold, i.e., an invariant surface, in
phase space. Such a manifold is invariant under the flow (i.e., orbits that start out
in the manifold remain in it for all time), which extends the invariance property
of LNMs to nonlinear systems. In order to parametrize the manifold, a single
pair of state variables (i.e., both the displacement and the velocity) are chosen as
master coordinates, the remaining variables being functionally related to the chosen
pair. Therefore, the system behaves like a nonlinear single-DOF system on the
manifold.
Geometrically, LNMs are represented by planes in phase space, and NNMs
are two-dimensional surfaces that are tangent to them at the equilibrium point.
For illustration, the manifolds corresponding to in-phase and out-of-phase NNMs
motions of system (1) are given in Fig. 7.

2.2 Fundamental Properties

NNMs have intrinsic properties that are fundamentally different from those of
LNMs. They are reviewed and illustrated in what follows.

2.2.1 Frequency-Energy Dependence

One typical dynamical feature of nonlinear systems is the frequency-energy depen-
dence of their oscillations. One important consequence is that the frequency
response functions (FRFs) of nonlinear systems are no longer invariant. For
illustration, the (unscaled) FRFs of system

Fig. 7 Two-dimensional invariant manifolds of system (1) with the corresponding LNMs. Left
plot: in-phase LNM and NNM; right plot: out-of-phase LNM and NNM
806 G. Kerschen and A. F. Vakakis

2.5 5

2 4
Amplitude(m)

1.5 3

1 2

0.5 1
F F
0 0
0.12 0.14 0.16 0.18 0.2 0.22 0.12 0.14 0.16 0.18 0.2 0.22
Frequency(Hz)

Fig. 8 Nonlinear frequency response functions close to the first resonant frequency (5 different
forcing amplitudes: 0.002N, 0.01N, 0.05N, 0.1N, 0.2N). Left plot: x1 ; right plot: x2

3.5 1.4

3 1.2

2.5 1
Amplitude(m)

2 0.8

1.5 0.6

1 0.4

0.5 0.2
F F
0 0
0.2 0.25 0.3 0.35 0.4 0.45 0.2 0.25 0.3 0.35 0.4 0.45

Frequency(Hz)

Fig. 9 Nonlinear frequency response functions close to the second resonant frequency (5 different
forcing amplitudes: 0.002N, 0.01N, 0.05N, 0.1N, 0.2N). Left plot: x1 ; right plot: x2

ẍ1 + (0.02ẋ1 − 0.01ẋ2 ) + (2x1 − x2 ) + 0.5 x13 = F cos ωt

ẍ2 + (0.02ẋ2 − 0.01ẋ1 ) + (2x2 − x1 ) = 0 (3)

are depicted in Figs. 8 and 9 for F varying between 0.002 N and 0.2 N.
The modal shapes and frequencies of oscillation of NNMs also depend on the
total energy in the system. In contrast to linear theory, this energy dependence
prevents the direct separation of space and time in the governing equations of
motion, which complicates the analytical calculation of the NNMs.
Returning to the undamped system (1), Fig. 10 shows the time series, the config-
uration space, the power spectral density (PSD), and two-dimensional projections
of the phase space of three in-phase NNM motions of increasing energies. The
NNM motion at low energy resembles that of the in-phase LNM of the underlying
linear system (2). The modal shape is a straight line, there is one main harmonic
13 Modal Analysis of Nonlinear Mechanical Systems 807

0.02 5 20
x1 , x2 (m)

x1 , x2 (m)

x1 , x2 (m)
−0.02 −5 −20
0 t (s) 12.56 0 t (s) 10.93 0 t (s) 9.51
0.02 5 20
x2 (m)

x2 (m)

x2 (m)
−0.02 −5 −20
−0.02 x1 (m) 0.02 −2 x1 (m) 2 −5 x1 (m) 5
−200
P SD (dB)

P SD (dB)

P SD (dB)
−300
0 Frequency (Hz) 4 0 Frequency (Hz) 4 0 Frequency (Hz) 4
0.02 2 5
ẋ1 (m)

ẋ1 (m)

ẋ1 (m)

−0.02 −2 −5
−0.02 x1 (m) 0.02 −2 x1 (m) 2 −5 x1 (m) 5

Fig. 10 Frequency-energy dependence of NNMs. From left to right: low-

([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [0.105 0.105 0 0]), moderate- ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] =
[1.476 2.100 0 0]) and high- ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [3.319 11.134 0 0]) energy in-phase
NNMs. From top to bottom: time series (——: x1 (t); − − −: x2 (t)); configuration space; power
spectral density of x1 (t); two-dimensional projections of the phase space

component in the system response, and the motion in phase space is a circle. For
the motion at moderate energy, the NNM is now a curve, and the presence of two
harmonic components can be detected. A clear departure from the LNM (harmonic)
motion is observed. At high energy, this is even more visible. For instance, the
motion in phase space is a strongly deformed ellipse. When moving from the low-
to the high-energy NNM, the motion period decreases from 6.28 s to 4.755 s, due
to the hardening characteristic of the cubic spring. Another noticeable characteristic
of the NNMs is that the modes at higher energies are not the geometric continuation
of those at lower energies. For illustration, Fig. 11 superposes the three-in-phase
NNMs in the configuration space.
To further illustrate the frequency-energy dependence of the NNMs, the har-
monic balance method can be applied to system: (1). This method expresses the
periodic motion of a system by means of a finite Fourier series. For simplicity, a
series with a single harmonic component is considered

x1 (t) = A cos ωt, x2 (t) = B cos ωt (4)

 808 G. Kerschen and A. F. Vakakis

15 0.03

10 0.02
Displacement x2 (m)

Displacement x2 (m)
5 0.01

0 0

−5 −0.01

−10 −0.02

−15 −0.03
−4 −2 0 2 4 −0.02 −0.01 0 0.01 0.02

Displacement x1 (m) Displacement x1 (m)

Fig. 11 Superposition of low-, moderate- and high-energy NNM motions in the configuration
space. The right plot is a close-up of the left plot

This expression is plugged into the equations of motion (1). Expanding cos3 ωt in
terms of cos ωt and cos 3ωt, and balancing all the coefficients of the cos ωt terms
yields

3A3
− Aω2 + (2A − B) + 0.5 =0
4
−Bω2 + (2B − A) = 0 (5)

Analytic expressions for coefficients A and B are then readily obtained


8(ω2 − 3)(ω2 − 1)
A=± (6)
3(ω2 − 2)
A
B= (7)
2 − ω2

The square root exists in the two frequency intervals

√ √
ω1 ∈ [1, 2] and ω2 ∈ [ 3, +∞] (8)
√
noting that ω = 1 rad/s and ω = 3 rad/s are the two natural frequencies of the
underlying linear system (2). In the first (second) frequency interval, B has the same
(opposite) sign as A; an in-phase (out-of-phase) NNM motion is observed for initial
conditions [x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [A B 0 0]).
The (conserved) total energy during the free response of system (1) is

A2 (B − A)2 B 2 A4
Total Energy = Kinetic Energy+Potential Energy = + + +0.5
2 2 2 4
(9)
13 Modal Analysis of Nonlinear Mechanical Systems 809

0.9

0.8

0.7

0.6
Frequency (Hz)

0.5

0.4

0.3

0.2

0.1

0 −6 −4 −2 0 2
10 10 10 10 10
Energy (log scale)

Fig. 12 Frequency-energy plot of system (2). LNM motions depicted in the configuration space
are inset

which, according to equations (6) and (7), demonstrates the frequency-energy

dependence of NNM motions.
An appropriate graphical depiction of the NNMs is a frequency-energy plot
(FEP). An NNM is represented by a point in the FEP, which is drawn at a frequency
corresponding to the minimal period of the periodic motion and at an energy equal
to the conserved total energy during the motion. A branch, represented by a solid
line, is a family of NNM motions possessing the same qualitative features (e.g., the
in-phase NNM motions of a 2DOF system).
As a point of comparison, the FEP of the linear system (2) is shown in Fig. 12.
Because the modal parameters of a linear system do not depend on energy, this FEP
comprises two horizontal lines at the two natural frequencies of the system. The
LNM motions represented in the configuration space are inset and are also unaltered
by the energy level.
The FEP of the nonlinear system (1) was computed using the method proposed
in [46]; it is shown in Fig. 13. The backbone of the plot is formed by two branches,
which represent in-phase (S11+) and out-of-phase (S11−) synchronous NNMs.
The letter S refers to symmetric periodic solutions for which the displacements and
velocities of the system at half period are equal but with an opposite sign to those at
time t = 0. As shown in the next section, unsymmetric periodic solutions may also
be encountered and are denoted by a letter U . The indices in the notations are used
810 G. Kerschen and A. F. Vakakis

0.9

0.8 S11−

0.7
Frequency (Hz)

0.6

0.5

0.4

0.3
S11+
0.2

0.1

0 −5 0
10 10
Energy (log scale)

Fig. 13 Frequency-energy plot of system (1). NNM motions depicted in the configuration space
are inset. The horizontal and vertical axes in these plots are the displacements of the first and
second DOFs, respectively; the aspect ratio is set so that increments on the horizontal and vertical
axes are equal in size to indicate whether or not the motion is localized to a particular DOF

to mention that the two masses vibrate with the same dominant frequency. The FEP
clearly shows that the nonlinear modal parameters have a strong dependence on the
total energy in the system:

1. The frequency of both the in-phase and out-of-phase NNMs increases with the
energy level, which reveals the hardening characteristic of the system.
2. The modal curves change for increasing energies. The in-phase NNM tends to
localize to the second DOF (i.e., it resembles a vertical curve), whereas the out-
of-phase NNM localizes to the first DOF (i.e., it resembles an horizontal curve).

The comparison between Figs. 12 and 13 also reveals that NNMs have a clear
and simple conceptual relation to the LNMs.

2.2.2 Modal Interactions: Internally Resonant Nonlinear Normal

Modes
Another salient feature of nonlinear systems is that NNMs may interact during a
general motion of the system. Nonlinear modal interactions have been studied exten-
sively in the literature (see, e.g., the monograph [47]). A case of particular interest is
13 Modal Analysis of Nonlinear Mechanical Systems 811

0.9
0.25

0.8 S11−

0.7 S31
Frequency (Hz)

0.6

0.5
0.2 2 3 4
10 10 10

0.4

0.3

0.2
S11+

0.1

0 −8 −6 −4 −2 0 2 4
10 10 10 10 10 10 10
Energy (log scale)

Fig. 14 Frequency-energy plot of system (1) featuring a 3:1 internal resonance between the in-
phase and out-of-phase NNMs

when the linear natural frequencies are commensurate or nearly commensurate. An

energy exchange between the different modes involved may therefore be observed
during the internal resonance. For instance, exciting a high-frequency mode may
produce a large-amplitude response in a low-frequency mode. Vibration absorbers
exploiting these energy transfers have been studied in [48].
The FEP in Fig. 13 does not seem to feature internally resonant NNMs. However,
when carrying out the NNM computation at higher energy levels, Fig. 14 shows that
another branch of periodic solutions, termed a tongue, emanates from the backbone
branch S11+. On this tongue, denoted S31, there is a 3:1 internal resonance/modal
interaction between the in-phase and out-of-phase NNMs. √
Surprisingly, the ratio of the linear natural frequencies of system (1) is 3. Due
to energy dependence, a 3:1 ratio between the two frequencies can still be realized,
because the frequency of the in-phase NNM increases less rapidly than that of
the out-of-phase NNM. This shows that NNMs can be interact without necessarily
having commensurate linear natural frequencies.
To better illustrate the resonance mechanism, the branch S11− is represented in
the FEP of Fig. 15 at the third of its frequency. This is relevant, because a periodic
solution of period T is also periodic with period 3T , and the resulting branch is
therefore denoted S33−. It turns out that a smooth transition from S11+ to S33−
occurs on tongue S31. This transition is also depicted in Fig. 16 where the evolution
812 G. Kerschen and A. F. Vakakis

0.9
0.25

0.8 S33− S11−

0.7 S31
Frequency (Hz)

0.6
S11+

0.5
0.2 2
10 10
3
10
4 S33−
0.4

0.3
S11+
0.2

0.1

0 −8 −6 −4 −2 0 2 4
10 10 10 10 10 10 10
Energy (log scale)

Fig. 15 Frequency-energy plot of system (1). − − −: S33−, that is S11− represented at the third
of its dominant frequency

of the configuration space and of the Fourier coefficients is shown for several points
on S31 or in its vicinity. Starting from NNM (a), an in-phase motion characterized
by two perceptible harmonic components is observed. From (a) to (d), the relative
importance of the third harmonics grows, as clearly confirmed by the motion in
the configuration space. Moving from (d) to (e) corresponds to a drastic qualitative
change in the dynamics. Firstly, the first harmonics has completely disappeared for
both oscillators. Secondly, the signs of the coefficients of the third harmonics are
opposite. Overall, this means that an out-of phase motion with a three times as large
frequency is realized. Eventually, through a 3:1 internal resonance, the motion ends
up on S33− or, equivalently, on S11−. From (f) to (h), the relative importance of
the third harmonics diminishes, and a motion qualitatively similar to that at (a) is
observed. However, the configuration space of NNM (h) reveals the presence of a
fifth harmonics, which is a precursor to the gradual development of tongue S51.
This indicates that other resonance scenarios exist in this seemingly simple
system. The frequency of the out-of-phase NNM motions on S11− steadily
increases for increasing energies, whereas the NNM motions √ on S11+ have their
frequency asymptotically approaching a value close to 3 rad/s. Following this
reasoning, we expect the existence of a countable infinity of internal resonance
cases (e.g., 2:1, 4:1, 5:1, etc.). To confirm this conjecture, additional tongues were
computed numerically and are represented in Fig. 17. These tongues emanate from
13 Modal Analysis of Nonlinear Mechanical Systems 813

S11+ and coalesce into S11− following a mechanism similar to that described
above (Fig. 16). To illustrate the rich dynamics, a few representative modal shapes
of system (1) are depicted in Fig. 18.
We note that interactions between NNMs were observed in real-world structures
[49, 50, 51, 52].

2.2.3 Mode Bifurcations and Stability

A third fundamental property of NNMs is that their number may exceed the number
of DOFs of the system. Due to mode bifurcations, not all NNMs can be regarded
as nonlinear continuation of normal modes of linear systems [17]. Internally
resonant NNMs are one example. Another possible example corresponds to the
NNM bifurcations of the system

ẍ1 + x1 + x13 + K(x1 − x2 )3 = 0

ẍ2 + x2 + x23 + K(x2 − x1 )3 = 0 (10)

for variations of the coupling spring K [14]. This system possesses similar NNMs
that obey to the relation x2 (t) = cx1 (t). Eliminating x2 from equations (10) yields
 
ẍ1 + x1 + 1 + K(1 − c)3 x13 = 0

1 
ẍ1 + x1 − K(1 − c)3 + c3 x13 = 0 (11)
c

Because both equations must lead to the same solution, it follows

K(1 + c)(c − 1)3 = c(1 − c2 ), c = 0 (12)

Equation (12) means that system (10) always possesses two modes characterized by
c = ±1 that are direct extension of the LNMs. However, this system can possess two
additional similar NNMs that cannot be captured using linearization procedures. At
K = 0.25, these NNMs bifurcate from the out-of-phase mode, as shown in Fig. 19.
Another important characteristic of NNMs is that they can be stable or unstable,
which is in contrast to linear theory where all modes are neutrally stable. In this
context, instability means that small perturbations of the initial conditions that
generate the NNM motion lead to the elimination of the mode oscillation. Therefore,
unstable NNMs are not physically realizable. The NNM stability analysis can be
performed numerically or analytically. In Fig. 20, stability is computed numerically
through the eigenvalues of the monodromy matrix. In other studies, analytical
results are obtained through Floquet theory after adding small perturbations to the
periodic solutions. For a detailed stability analysis of the NNMs, the reader can refer
to [14, 16, 17, 53, 54].
Bifurcations and stability are interrelated concepts, because a change in stability
occurs through a bifurcation. For instance, the bifurcation in system (10) generates a
814 G. Kerschen and A. F. Vakakis

0.226
S31 (d)
0.224 (e)
(c)
(f)
0.222
(h)
S33−
0.22
Frequency (Hz)

(b) (g)
0.218

0.216

0.214

0.212

0.21 (a)
S11+

0.208
2 3 4
10 10 10
Energy
10 7

(a) 8 (a) (e) 6

5
(e)
6
4
Ai
Ai

4 3

2
2
1
0
0

−2 −1
0 1 2 3 4 5 0 1 2 3 4 5
Harmonic i Harmonic i
30
10

(b) 25
(b) (f) 5

0
(f)
20
−5
15
−10
Ai

Ai

10 −15

−20
5
−25
0
−30

−5 −35
0 1 2 3 4 5 0 1 2 3 4 5
Harmonic i Harmonic i

70 10

(c) 60

50
(c) 0

−10

40 −20
Ai

Ai

30 −30

20 −40

10 −50

0 (g) −60 (g)

−10 −70
0 1 2 3 4 5 0 1 2 3 4 5
Harmonic i Harmonic i

35
20

(d) 30

25
(d) −20
0

20 −40
Ai

Ai

15 −60

10 −80

5 −100

0 (h) −120 (h)

−5 −140
0 1 2 3 4 5 0 1 2 3 4 5
Harmonic i Harmonic i

Fig. 16 (Continued)
13 Modal Analysis of Nonlinear Mechanical Systems 815

pair of stable/unstable NNMs (Fig. 19). Returning to system (1), another illustration
of NNM stability is shown in the FEP of Fig. 20. When the tongue U 21 bifurcates
from S11+, the NNMs on this latter branch lose stability. A detailed description
of this tongue and the related dynamical mechanisms (e.g., symmetry-breaking
bifurcation) is beyond the scope of this paper. This figure also shows that stability
can be lost when a turning point is encountered.

3 Nonlinear Normal Modes: How to Compute Them?

Different methods for computing NNMs of discrete and continuous systems are
briefly described in this section. They are classified in two categories, according to
whether the computation relies on analytical or numerical methods.

3.1 Analytical Techniques

Rosenberg was the first to develop constructive techniques for computing NNMs of
discrete conservative oscillators. The book by Vakakis et al. [17] summarizes the
developments until the 1990s. In what follows, two techniques directly inspired by
the two NNM definitions discussed previously are briefly reviewed.

3.1.1 An Energy-Based Formulation

This formulation relies on Rosenberg’s work [8] and expresses an NNM as a modal
curve in the configuration space. It was further developed by Manevitch and Mikhlin
for discrete conservative oscillators [12] and exploited in a few other studies [15,17].
To illustrate the method, it is applied to system (1)

ẍ1 + (2x1 − x2 ) + 0.5 x13 = 0

ẍ2 + (2x2 − x1 ) = 0 (13)

When the system vibrates along an NNM, the displacement x2 is linked to x1

through the expression of the modal curve x̂2

x2 = x̂2 (x1 ) (14)

The objective of the method is to eliminate the time derivatives from the equations
of motion (13). To compute the second time derivative of x2 , relation (14) is
differentiated twice using the chain rule


Fig. 16 Internally resonant NNMs (3:1 resonance). Top plot: close-up of the tongue S31 in the
frequency-energy plot. Bottom plots: configuration space (horizontal axis: x1 ; vertical axis: x2 )
and Fourier coefficients of a series containing cosine terms only (grey: x1 ; black: x2 )
816 G. Kerschen and A. F. Vakakis

0.23
U 21 S31 U 41 S51
0.22

0.21
Frequency (Hz)

0.2

0.19

0.18

0.17

S11+
0.16

0.15 −2 0 2 4 6
10 10 10 10 10
Energy (log scale)

Fig. 17 Close-up of S11+ at higher energy levels

Fig. 18 A few representative NNMs of system (1) in the configuration space

13 Modal Analysis of Nonlinear Mechanical Systems 817

0.8

0.6
K

0.4

0.2

0
−15 −10 −5 0 5
c
Fig. 19 NNM bifurcations of system (10) [14] (——: stable NNMs; − − −: unstable NNMs)

S31
U 21
0.22

0.21
Frequency (Hz)

0.2

0.19

0.18

0.17

0.16 S11+
−2 0 2
10 10 10
Energy (log scale)

Fig. 20 Close-up of S11+ with stability results (——: stable NNMs; • • • : unstable NNMs)
818 G. Kerschen and A. F. Vakakis

ẍ2 = x̂2 ẋ12 + x̂2 ẍ1 (15)

where prime denotes differentiation with respect to x1 . This expression involves the
second time derivative of x1 , which is readily obtained from the equations of motion

ẍ1 = −2x1 + x̂2 − 0.5x13 (16)

It then remains to compute the first time derivative of x1 appearing in equation (15).
To this end, a first integral of motion expressing explicitly the conservation of energy
during the motion is written by multiplying equation (16) by ẋ1 and integrating
 ẋ1  x1  
ẋ12 = 2 ẋ1 dẋ1 = −2 2u − x̂2 (u) + 0.5u3 du (17)
0 X1

where X1 is the maximum amplitude attained by x1 ; i.e., when ẋ1 = 0. The

derivatives are substituted into the second of equations (13), which yields the
equation governing the modal curve:
  x1     
x̂2 −2 2u − x̂2 (u) + 0.5u du +x̂2 −2x1 + x̂2 − 0.5x13 +(2x̂2 −x1 ) = 0
3
X1
(18)
Because the coefficient of the highest derivative vanishes when x1 = X1 , this
functional equation is singular at the maximum equipotential surface. It must
therefore be supplemented by a boundary condition
  
x̂2 −2x1 + x̂2 − 0.5x13 + (2x̂2 − x1 ) =0 (19)
x1 =X1

which expresses that the nonlinear mode intersects orthogonally the maximum
equipotential surface in the configuration space. Equation (18) does not depend on
the time variable, and its solution is amenable to a power series expansion:

(0) (1) (2)

x̂2 (x1 ) = x̂2 (x1 ) +  x̂2 (x1 ) +  2 x̂2 (x1 ) + O( 3 ) (20)

This formulation was extended to undamped continuous systems in [16]. The

displacement of any point of the system is expressed in terms of a single reference
displacement x0 (t) = x(s0 , t) by the functional relation

x(s, t) = X [s, x0 (t)] (21)

where s is the spatial coordinate and X is a modal function characterizing the

considered NNM. Then, an integral equation expressing the conservation of energy
during the motion is used in conjunction with equation (21) to eliminate the time
derivatives from the equations of motion. Eventually, the equation governing the
modal function X is obtained and is solved using power series.
13 Modal Analysis of Nonlinear Mechanical Systems 819

In the presence of internal resonances, the folding of the NNMs in the configura-
tion space may result in multivalued relationship among the various coordinates (see
Fig. 6). This was nicely addressed in [55] by considering NNMs in an appropriately
defined modal space.

3.1.2 The Invariant Manifold Approach

The invariant manifold approach [19, 20, 21, 22] is similar in spirit to the energy-
based formulation. The difference with the previous approach is that a pair of
state variables (i.e., both the displacement and the velocity) are chosen as master
coordinates, the remaining variables being functionally related to the chosen pair:

x(s, t) = X1 [s, x0 (t), ẋ0 (t)] and ẋ(s, t) = X2 [s, x0 (t), ẋ0 (t)] (22)

These relations define a two-dimensional invariant manifold in phase space. By

taking the time derivative of these constraint equations and using the chain rule
differentiation, the explicit time dependence from the equations of motion can be
eliminated. Eventually, this yields a set of partial differential equations governing
the modal functions X1 and X2 . These equations are as difficult to solve as the
original problem, but the solution can be approximated using power series.
For systems with internal resonances, a multi-mode invariant manifold is consid-
ered in [56] to account for the influence of several modes. For instance, when two
modes are resonant, the master coordinates comprise two pairs of state variables,
and the resulting invariant manifold is four-dimensional.

3.2 Numerical Techniques

One of the first approaches was proposed by Slater in [57]. Based on Rosenberg’s
definition, the procedure integrates directly the governing equations of motion
over one period using numerical algorithms (e.g., Runge-Kutta and Newmark). It
comprises two steps:

1. An isolated periodic solution corresponding to a specific energy level is com-

puted by modifying iteratively the initial conditions governing the free response
of the system. This is carried out using optimization algorithms that minimize a
periodicity condition (i.e., a cost function representing the lack of periodicity of
the current iterate).
2. Low-energy modal curves and the corresponding periods of oscillation are first
computed, taking the normal modes and natural frequencies of the underlying
linear system as initial guesses. The energy is then gradually increased with the
previously computed NNM as an initial guess for the next NNM.

This step-wise type of continuation of periodic solutions is called sequential

continuation. A more advanced continuation scheme, i.e., pseudo-arclength contin-
uation, was used in [46] and led to the development of a computationally effective
820 G. Kerschen and A. F. Vakakis

method for the calculation of NNMs. Another sophisticated continuation method is

the so-called asymptotic-numerical method [58]. It is a semi-analytical technique
that is based on a power series expansion of the unknowns parameterized by a
control parameter. It is utilized to follow the NNM branches in conjunction with
the harmonic balance method in [59] or with finite difference methods in [60].

3.3 Assessment of the Different Methodologies

Analytical methodologies have the advantage that NNMs can be constructed

symbolically, which is certainly useful for gaining insight into the dynamics and
for performing parametric studies. Among other things, they clearly highlight the
frequency-energy dependence of the NNMs. The fundamental drawbacks of these
techniques is that (i) they are quite analytically involved and require a careful
treatment in the presence of internal resonances; (ii) the resultant dynamics are only
accurate for small-amplitude motions; and (iii) the upper bound for these motions is
not known a priori.
In this context, numerical methods have certainly the potential to make nonlinear
modal analysis more accessible to the practicing structural engineer. Most of them
provide an exact solution to the NNM calculation. But their key advantage is that
they can calculate NNMs of real-world structures in strongly nonlinear regimes of
motion. For instance, pseudo-arclength continuation was exploited in [50, 52] to
compute NNMs of real-world spacecraft and aircraft, respectively, and in [37] for
structures with distributed nonlinearity.

4 Nonlinear Normal Modes: Why Are They Useful?

The objective of this section is to describe several applications where NNMs rep-
resent a useful framework for the structural dynamicist. Specifically, we highlight
how useful the NNMs are for modal analysis and system identification and how they
may be exploited in conjunction with time-frequency analysis in order to extend
the existing linear methodologies [2]. Nonlinear model reduction is also briefly
discussed.

4.1 “Linear” Modal Analysis

Modal analysis and testing of linear mechanical structures has been developed over
the past 40–50 years, and the techniques available today are mature and advanced
[2]. Clearly, though, linearity is an idealization, an exception to the rule; nonlinearity
is a frequent occurrence in real-life applications. In the presence of nonlinear
phenomena, the structural dynamicist should therefore ask the question: can I still
use the linear modes ? Obviously, the answer depends on the type of the nonlinearity
and on the excitation level.
In this context, we believe that the computation of the NNMs and their
representation in a FEP is a robust and accurate tool to decide whether or not the
13 Modal Analysis of Nonlinear Mechanical Systems 821

linear framework is still applicable. It can be used to determine which modes (and to
what extent) are sensitive to the nonlinearity. Going back to Fig. 13, it is clear that,
until an energy of 10−1 , the mode shapes and natural frequencies are unaffected by
the nonlinearity and can safely be used. Beyond this critical energy level, both the
in-phase and out-of-phase modes show a significant departure from the LNMs and
become dependent on the total energy in the system.
As another example, the FEP of system

ẍ1 + (2x1 − x2 ) = 0
ẍ2 + (2x2 − x1 − x3 ) + 0.5 x23 = 0 (23)
ẍ3 + (2x3 − x2 ) = 0

is depicted in Fig. 21. The linear modal parameters remain unchanged until approx-
imately an energy of 10−1 . Another interesting finding is that the nonlinearity has
no influence either on the frequency or on the mode shape of the second mode.

4.2 Nonlinear Modal Analysis

When it is certain that the system is excited in the nonlinear range, the linear
framework should be abandoned in favor of a nonlinear modal analysis. Any attempt

0.55

0.5

0.45

0.4
Frequency (Hz)

0.35

0.3

0.25

0.2

0.15

0.1 −6 −4 −2 0 2
10 10 10 10 10
Energy (log scale)

Fig. 21 Frequency-energy plot of system (23). NNMs represented by bar graphs are inset; they
are given in terms of the initial displacements that realize the periodic motion (with zero initial
velocities assumed)
822 G. Kerschen and A. F. Vakakis

to apply traditional linear analysis in this context results, at best, in a suboptimal

design.
Considering again system (1) as a first example, its FEP in Fig. 13 greatly helps
to understand how the modal curves deform under the action of the nonlinearity. The
in-phase NNM tends to localize to the second DOF, whereas the out-of-phase NNM
localizes to the first DOF. Regarding the corresponding frequency of oscillation,
both modes are characterized by a hardening behavior due to the presence of the
cubic spring.
As a second example, a planar cantilever beam discretized by 20 finite elements
and with a cubic spring at the free end is now considered (see Table 1 for the
geometrical and mechanical properties). This models a real nonlinear beam that
was used as a benchmark for nonlinear system identification during the European
action COST F3 [61]. The first two modes are plotted in the FEPs of Figs. 22
and 23, respectively. Considering the same energy level, the first modal curve
seems somewhat more affected by the nonlinearity compared to the second modal
curve. Their frequencies of oscillation undergo a strong increase with increasing
energy levels. The FEPs also highlight the presence of two tongues, revealing the
existence of internal resonances. The tongue in Fig. 22 corresponds to a 5:1 internal
resonance between the first and second modes of the beam. When the energy
gradually increases along the tongue, a smooth transition from the first mode to the
second mode occurs following a dynamical mechanism similar to that described in
Sect. 2.2.2. Similarly, a 5:1 internal resonance between the second and fourth modes
is observed in Fig. 23. These internal resonances occur despite that the linear natural
frequencies are not commensurate, as also discussed in Sect. 2.2.2.
These two examples demonstrate that such a nonlinear modal analysis is an
important tool for thoroughly understanding the system’s vibratory response in the
nonlinear regime. Clearly, this cannot be achieved using linearization procedures.
However, because the general motion of a nonlinear system cannot be expressed
as a superposition of individual NNM motions and because the modes in all these
figures are computed based on the underlying undamped system, the practical utility
of the nonlinear modal analysis might appear, at first, questionable.
A first motivation to compute and exploit the NNMs is that forced resonances in
nonlinear systems occur in their neighborhoods. The knowledge of the NNMs can
therefore provide valuable insight into the structure of the resonances, a feature of
considerable engineering importance [17]. For illustration, system (3) is considered.
In Figs. 24 and 25, the backbone of the FEP of Fig. 13 is superposed to the nonlinear
frequency response functions of Fig. 8 and 9. It can be observed that the backbone
of the FEP traces the locus of the frequency response peaks for both the in-phase

Table 1 Geometrical and mechanical properties of the planar cantilever beam

Length Width Thickness Young’s modulus Density Nonlinear coeff.
(m) (m) (m) (N/m2 ) (kg/m3 ) (N/m3 )
0.7 0.014 0.014 2.05e11 7800 6 109
13 Modal Analysis of Nonlinear Mechanical Systems 823

50
31

30.5 0

45 30

29.5

29
Frequency (Hz)

40
28.5

28 −2
10
35

30

0 0
0

25

20 −6 −5 −4 −3 −2 −1
10 10 10 10 10 10

Energy (log scale)

Fig. 22 Frequency-energy plot of the cantilever beam; close-up of the first mode

and out-of-phase modes. Furthermore, Fig. 26 compares the forced response of the
system close to the first resonance (for F = 0.1, see the square in Fig. 24) to the
free response of the corresponding point of the backbone. An excellent agreement
is obtained between the two types of motion.
A second motivation is that the damped dynamics closely follows the NNMs of
the underlying undamped system. To demonstrate this, a time-frequency analysis
method, the continuous wavelet transform (CWT) is used. The CWT can track the
temporal evolution of the instantaneous frequencies, which makes it an effective tool
for analyzing nonlinear signals. The usual representation of the transform is to plot
its modulus as a function of time and frequency in a three-dimensional or contour
plot. To use the CWT in conjunction with the FEP, a different representation is
proposed herein. The CWT is represented in a frequency-energy plot by substituting
the instantaneous energy in the system for time.
The free response of system

ẍ1 + 0.03ẋ1 + (2x1 − x2 ) + 0.5 x13 = 0

ẍ2 + 0.01ẋ2 + (2x2 − x1 ) = 0 (24)
824 G. Kerschen and A. F. Vakakis

180
165
0

175 164

163

170 162
Frequency (Hz)

161

165 160 0 1
10 10

160 0

155

150

145 −6 −4 −2 0 2
10 10 10 10 10

Energy (log scale)

Fig. 23 Frequency-energy plot of the cantilever beam; close-up of the second mode

3 6

2.5 5
Amplitude (m)

2 4

1.5 3

1 2

0.5 F 1 F
0 0

−0.5 −1
0.12 0.14 0.16 0.18 0.2 0.22 0.12 0.14 0.16 0.18 0.2 0.22

Frequency (Hz)

Fig. 24 Nonlinear frequency response functions close to the first resonant frequency (5 different
forcing amplitudes: 0.002N, 0.01N, 0.05N, 0.1N, 0.2N). The dashed line is the backbone S11+ of
the frequency-energy plot. Left plot: x1 ; right plot: x2
13 Modal Analysis of Nonlinear Mechanical Systems 825

5 1.4

1.2
4
1
Amplitude (m)

3
0.8

2 0.6

0.4
1
0.2
F F
0 0
0.2 0.25 0.3 0.35 0.4 0.45 0.2 0.25 0.3 0.35 0.4 0.45

Frequency (Hz)

Fig. 25 Nonlinear frequency response functions close to the second resonant frequency (5
different forcing amplitudes: 0.002N, 0.01N, 0.05N, 0.1N, 0.2N). The dashed line is the backbone
S11− of the frequency-energy plot. Left plot: x1 ; right plot: x2

2
Displacement x2 (m)

−1

−2

−3
−2 −1 0 1 2
Displacement x1 (m)
Fig. 26 Free (F = 0) and forced responses (F = 0.1) of system (3) in the configuration space.
——: forced response; - - -: free response

is depicted in Figs. 27 and 28 for an excitation of an in-phase and out-of-phase

NNM, respectively. The left plot is the theoretical FEP that is the FEP computed
from the equations of motion. The right plot is the “experimental” FEP, calculated
directly from the time series: (i) the backbone is provided by the CWT, and (ii) the
modal curves are obtained by representing the time series in the configuration space
for one oscillation around a specific energy level. For comparison, the theoretical
backbone is represented by a solid line in the experimental FEP. A perfect agreement
is obtained between the two FEPs, which shows that the undamped NNMs are
826 G. Kerschen and A. F. Vakakis

0.5 0.5
0.45 0.45
0.4 0.4
Frequency (Hz)

0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
−4 −3 −2 −1 0 1 2 −4 −3 −2 −1 0 1 2

Energy (log scale)

Fig. 27 Frequency-energy plot of system (1). Left plot: theoretical FEP; right plot: experimental
FEP for an excitation of an in-phase NNM ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [2.500 5.895 0 0])

0.6 0.6

0.5 0.5
Frequency (Hz)

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
−4 −3 −2 −1 0 1 2 3 −4 −3 −2 −1 0 1 2 3

Energy (log scale)

Fig. 28 Frequency-energy plot of system (1). Left plot: theoretical plot; right plot: experimental
plot for an excitation of an out-of-phase NNM ([x1 (0) x2 (0) ẋ1 (0) ẋ2 (0)] = [−6.842 0.389 0 0])

attractors for the damped trajectories. In the present case, the modal damping ratios
are 1% and 0.6%, but we note that this result holds for higher damping ratios.
The combined use of the FEP and the CWT represents a suitable framework for
developing a nonlinear system identification method, which could be viewed as a
practical nonlinear analog of experimental modal analysis.

4.3 Reduced-Order Modeling

In a recent series of works [31, 62, 63, 64], it was shown that NNMs can provide
effective bases for constructing reduced-order models of the dynamics of discrete
and continuous nonlinear oscillators.
Specifically, Touzé et al. performed a comparative study of reduced-order models
of large-amplitude vibrations of shell structures of different configurations using
either LNMs or NNMs [64]. They showed that one or two NNMs were sufficient
for accurately capturing the shell dynamics, and even the bifurcation structure of the
13 Modal Analysis of Nonlinear Mechanical Systems 827

Fig. 29 Frequency response curve of an hyperbolic paraboloid panel: reference (exact) computa-
tional solution compared to reduced-order models based on the leading LNM and NNM for varying
forcing amplitudes. (a) 2.84N; (b) 4.37N and (c) 6.66N. (Taken from Touze et al. [64])

dynamics that resulted from the nonlinear interaction of two shell modes in internal
resonance. By contrast, multiple linear modes were necessary to achieve the same
accuracy. For illustration, a specific application taken from [64] is shown in Fig. 29.
It depicts the frequency response curve of the nondimensionalized amplitude of the
transverse displacement of a hyperbolic paraboloid panel under harmonic excitation.
The harmonic excitation is applied at the center of the panel, and its frequency
is in the vicinity of the first eigenfrequency. Comparing the reference (exact)
computational solution to reduced-order models obtained using the leading NNM
and LNM, respectively, the accuracy of the NNM-based model and its superiority
over the LNM-based model are established. In this example, 15 LNMs were required
to obtain results of similar accuracy.
These results demonstrate that NNMs hold promise for low-order reduction of
structural models with many DOFs (e.g., finite element computational models).
Even though NNMs do not possess orthogonality properties (as do the LNMs), the
resulting models are still expected to be much more accurate compared to their
linear counterpart.
828 G. Kerschen and A. F. Vakakis

5 Closure

To robustly and accurately model nonlinearity in realistic vibrating structures is one

of the greatest challenges in structural engineering. In this context, NNMs have a
clear conceptual relation to the linear normal modes, yet they can highlight nonlinear
phenomena that are unexpected (and unexplainable) from a linear viewpoint.
By combining algorithms for the numerical calculation of NNMs with a signal
processing tool such as the wavelet transform and with a relevant graphical depiction
such as the frequency-energy plot, the damped dynamics of a nonlinear system
can be interpreted based on NNMs. These tools should help extend experimental
modal analysis, which is well-established for linear systems, to a practical nonlinear
analog. An experimental illustration of the NNM concept can be found in [65].

References
1. Craig R, Bampton M (1968) Coupling of substructures for dynamic analysis. AIAA J 6:1313–
1319
2. Ewins DJ (2000) Modal testing: theory, practice and application, 2nd edn. Research Studies
Press LTD, Hertfordshire
3. Friswell MI, Mottershead JE (1995) Finite element model updating in structural dynamics.
Kluwer Academic Publishers, London
4. Doebling SW, Farrar CR, Prime MB, Shevitz DW (1996) Damage identification and
health monitoring of structural and mechanical systems from changes in their vibration
characteristics: a literature review, Los Alamos National Laboratory Report LA-13070-MS
5. Kerschen G, Worden K, Vakakis AF, Golinval JC (2006) Past, present and future of nonlinear
system identification in structural dynamics. Mech Syst Signal Process 20:505–592
6. Rosenberg RM (1960) Normal modes of nonlinear dual-mode systems. J Appl Mech 27:
263–268
7. Rosenberg RM (1962) The normal modes of nonlinear n-degree-of-freedom systems. J Appl
Mech 29:7–14
8. Rosenberg RM (1966) On nonlinear vibrations of systems with many degrees of freedom.
Adv Appl Mech 9:155–242
9. Rand R (1971) Nonlinear normal modes in two-degree-of-freedom systems. Journal of
Applied Mechanics 38:561.
10. Rand R (1971) A higher-order approximation for nonlinear normal modes in two-degree-of-
freedom systems. Int J Nonlinear Mech 6:545–547
11. Rand R (1974) A direct method for nonlinear normal modes. Int J Nonlinear Mech 9:363–368
12. Manevitch LI, Mikhlin YV (1972) On periodic solutions close to rectilinear normal vibration
modes. PMM 36:1051–1058
13. Vakakis AF (1990) Analysis and identification of linear and nonlinear normal modes in
vibrating systems. Ph.D. Dissertation, California Institute of Technology
14. Caughey TK, Vakakis AF, Sivo JM (1990) Analytical study of similar normal modes and their
bifurcations in a class of strongly nonlinear systems. Int J Nonlinear Mech 25:521–533
15. Vakakis AF (1992) Non-similar normal oscillations in a strongly non-linear discrete system.
J Sound Vib 159:341–361
16. King ME, Vakakis AF (1994) An energy-based formulation for computing nonlinear normal-
modes in undamped continuous systems. J Sound Vib 116:332–340
17. Vakakis AF, Manevitch LI, Mikhlin YV, Pilipchuk VN, Zevin AA (1996) Normal modes and
localization in nonlinear systems. John Wiley & Sons, New York
13 Modal Analysis of Nonlinear Mechanical Systems 829

18. Vakakis AF (1997) Non-linear normal modes and their applications in vibration theory: an
overview. Mech Syst Signal Process 11:3–22
19. Shaw SW, Pierre C (1991) Non-linear normal modes and invariant manifolds. J Sound Vib
150:170–173
20. Shaw SW, Pierre C (1992) On nonlinear normal modes. ASME Winter Annual Meeting
21. Shaw SW, Pierre C (1993) Normal modes for non-linear vibratory systems. J Sound Vib
164:85–124
22. Shaw SW, Pierre C (1994) Normal modes of vibration for non-linear continuous systems.
J Sound Vib 169:319–347
23. Jezequel L, Lamarque CH (1991) Analysis of nonlinear dynamic systems by the normal form
theory. J Sound Vib 149:429–459
24. Yabuno H, Nayfeh AH (2001) Nonlinear normal modes of a parametrically excited cantilever
beam. Nonlinear Dyn 25:65–77
25. Xie WC, Lee HP, Lim SP (2003) Nonlinear dynamic analysis of MEMS switches by nonlinear
modal analysis. Nonlinear Dyn 31:243–256
26. Mazzilli CEN, Soares MES, Baracho Neto GP (2004) Non-linear normal modes of a
simply supported beam: continuous system and finite-element models. Comput Struct 82:
2683–2691
27. Gendelman OV (2004) Bifurcations of nonlinear normal modes of linear oscillator with
strongly nonlinear damped attachment. Nonlinear Dyn 37:115–128
28. Lacarbonara W, Camillacci R (2004) Nonlinear normal modes of structural systems via
asymptotic approach. Int J Solids Struct 41:5565–5594
29. Lee YS, Kerschen G, Vakakis AF, Panagopoulos PN, Bergman LA, McFarland DM (2005)
Complicated dynamics of a linear oscillator with a light, essentially nonlinear attachment.
Physica D 204:41–69
30. Dick AJ, Balachandran B, Mote CD (2006) Nonlinear vibration modes in micro-resonator
arrays. In: Smart Structures and Materials 2006: Modeling, Signal Processing, and Control,
Proceedings of the SPIE, vol 6166, pp 206–217
31. Touzé C, Amabili M (2006) Nonlinear normal modes for damped geometrically nonlinear
systems: Application to reduced-order modelling of harmonically forced structures. J Sound
Vib 298:958–981
32. Srinil N, Rega G (2007) Two-to-one resonant multi-modal dynamics of horizontal/inclined
cables. Part II: Internal resonance activation, reduced-order models and nonlinear normal
modes. Nonlinear Dyn 48:253–274
33. Lenci S, Rega G (2007) Dimension reduction of homoclinic orbits of buckled beams via the
non-linear normal modes technique. Int J Nonlinear Mech 42:515–528
34. Lacarbonara W, Paolone A, Vestroni F (2007) Non-linear modal properties of non-shallow
cables. Int J Nonlinear Mech 42:542–554
35. Cochelin B, Vergez C (2009) A high order purely frequency-based harmonic balance
formulation for continuation of periodic solutions. J Sound Vib 324:243–262
36. Laxalde D, Thouverez F, Complex non-linear modal analysis for mechanical systems:
application to turbomachinery bladings with friction interfaces. J Sound Vib 322 (2009),
1009–1025.
37. Kuether RJ, Allen MS (2014) A numerical approach to directly compute nonlinear nor-
mal modes of geometrically nonlinear finite element models. Mech Syst Signal Process
46:1–15
38. Hill TL, Cammarano A, Neild SA, Wagg DJ (2015) Interpreting the forced responses of
a two-degree-of-freedom nonlinear oscillator using backbone curves. J Sound Vib 349:
276–288
39. Renson L, A. Gonzalez-Buelga, Barton DAW, Neild SA, Robust identification of backbone
curves using control-based continuation. Journal of Sound and Vibration 367 (2016),
145–158.
40. Hill TL, Cammarano A, Neild SA, Barton DAW (2017) Identifying the significance of
nonlinear normal modes. Proc R Soc A 473:20160789
830 G. Kerschen and A. F. Vakakis

41. Scheel M, Peter S, Leine RI, Krack M (2018) A phase resonance approach for modal testing
of structures with nonlinear dissipation. J Sound Vib 435:56–73
42. Ribeiro P, Thomas O (2017) Nonlinear modes of vibration and internal resonances in nonlocal
beams. J Comput Nonlinear Dyn 12:031017
43. Song M, Renson L, Noel JP, Moaveni B, Kerschen G (2018) Bayesian model updating of
nonlinear systems using nonlinear normal modes. Struct Control Health Monit 25:e2258
44. Haller G, Ponsioen S (2016) Nonlinear normal modes and spectral submanifolds: existence,
uniqueness and use in model reduction. Nonlinear Dyn 86:1493–1534
45. Krack M (2015) Nonlinear modal analysis of nonconservative systems: extension of the
periodic motion concept. Comput Struct 154:59–71
46. Peeters M, Viguié R, Sérandour G, Kerschen G, Golinval JC (2009) Nonlinear normal modes,
Part II: toward a practical computation using numerical continuation. Mech Syst Signal
Process 23:170–194
47. Nayfeh AH (2000) Nonlinear interactions: analytical, computational and experimental meth-
ods. Wiley-Interscience, New York
48. SS. Oueini, Chin CM, Nayfeh AH (1999) Dynamics of a cubic nonlinear vibration absorber.
Nonlinear Dyn 20:283–295
49. Noel JP, Renson L, Kerschen G (2014) Complex dynamics of a nonlinear aerospace structure:
experimental identification and modal interactions. J Sound Vib 333:2588–2607
50. Renson L, Noel JP, Kerschen G (2015) Complex dynamics of a nonlinear aerospace structure:
numerical continuation and normal modes. Nonlinear Dyn 79:1293–1309
51. Claeys M, Sinou JJ, Lambelin JP, Todeschini R (2016) Modal interactions due to friction in the
nonlinear vibration response of the “Harmony” test structure: experiments and simulations.
J Sound Vib 376:131–148
52. Kerschen G, Peeters M, Golinval JC, Stephan C (2013) Nonlinear modal analysis of a full-
scale aircraft. AIAA J Aircr 50:1409–1419
53. Pak CH (2006) Synge’s concept of stability applied to non-linear normal modes. Int J
Nonlinear Mech 41:657–664
54. Recktenwald G, Rand R (2007) Stability of strongly nonlinear normal modes. Commun
Nonlinear Sci Numer Simul 12:1128–1132
55. King ME, Vakakis AF (1995) An energy-based approach to computing resonant nonlinear
normal modes. J Appl Mech 63:810–819
56. Boivin N, Pierre C, Shaw SW (1995) Non-linear modal analysis of structural systems
featuring internal resonances. J Sound Vib 182:336–341
57. Slater JC (1996) A numerical method for determining nonlinear normal modes. Nonlinear
Dyn 10:19–30
58. Cochelin B, Damil N, Potier-Ferry M (1994) Asymptotic numerical methods and Padé
approximants for nonlinear elastic structures. Int J Numer Methods Eng 37:1187–1213
59. Pérignon F (2004) Vibration forcées de structures minces, élastiques, non-linéaires. Ph.D.
Thesis, Université de la Méditerrané, Marseille
60. Arquier R, Bellizzi S, Bouc R, Cochelin B (2006) Two methods for the computation of
nonlinear modes of vibrating systems at large amplitudes. Comput Struct 84:1565–1576
61. Thouverez F (2003) Presentation of the ECL benchmark. Mech Syst Signal Process 17:
195–202
62. Pesheck E (2000) Reduced-order modeling of nonlinear structural systems using nonlinear
normal modes and invariant manifolds. Ph.D. Thesis, University of Michigan, Ann Arbor
63. Touzé C, Thomas O, Chaigne A (2004) Hardening/softening behaviour in non-linear oscilla-
tions of structural systems using non-linear normal modes. J Sound Vib 273:77–101
64. C. Touzé, Amabili M, Thomas O (2007) Reduced-order models for large-amplitude vibrations
of shells including in-plane inertia. In: Proceedings of the EUROMECH Colloquium on
Geometrically Nonlinear Vibrations, Porto
65. Peeters M, Kerschen G, Golinval JC (2011) Modal testing of nonlinear vibrating structures
based on nonlinear normal modes: experimental demonstration. Mech Syst Signal Process
25:1227–1247
 Part III
Analytical/Experimental Modeling
Applications
Substructuring Concepts and Component
Mode Synthesis 14
Daniel Rixen

Contents
1 Model Reduction: General Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834
1.1 Reduction by Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834
1.2 The Guyan–Irons Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835
1.3 Model Reduction Through Substructuring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 837
2 Numerical Techniques for Model Reduction of Substructures . . . . . . . . . . . . . . . . . . . . . . 839
2.1 The Hurty/Craig–Bampton Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 839
2.2 Substructure Reduction Using Free Interface Modes . . . . . . . . . . . . . . . . . . . . . . . . . 842
2.3 Numerical Examples of Different Substructure Reduction Techniques . . . . . . . . . . 848
2.4 Other Reduction Techniques for Substructures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851
3 Interface Reduction with the Hurty/Craig–Bampton Method:
Characteristic Constraint Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 852
3.1 Interface Reduction Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 852
3.2 System-Level Characteristic Constraint (S-CC) Modes . . . . . . . . . . . . . . . . . . . . . . . 853
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854

Abstract

In this chapter, the basic principle of model reduction for linear models in
structural dynamics is explained. In particular, the principle of substructuring
is outlined following different common approaches.

Keywords

Model reduction · Representation modes · Component dynamics

D. Rixen ()
Technische Universität München, Garching, Germany
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 833

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_16
834 D. Rixen

1 Model Reduction: General Concepts

1.1 Reduction by Projection

Often, a finite element model is first built for static analysis in order to assess static
deformations and stress levels which might be concentrated around small features
of the structure (stress concentration areas). Therefore, it is very common to find
static models that have a very refined mesh, hence a high number of degrees of
freedom (typically up to several millions). Such very large static models can be
solved by means of efficient solvers, but when it comes to computing free vibration
modes, harmonic responses and/or transient responses, it is required to solve many
static-like problems (i.e., in the inverse iterations of the eigensolvers or in the time-
stepping loop) and the computing time required is often unacceptable.
Fortunately, for most dynamical problems, such highly refined meshes are not
needed to capture the interesting dynamic behavior. For instance, if we want to
compute the fundamental modes, we know that the first free vibration modes have a
rather smooth deformation and thus a coarser mesh would be sufficient to compute
these modes. Similarly, when computing the dynamic response to external loads
which are no shocks (i.e., which have no high spectral content compared to the
eigenfrequency spectrum of the structure), a coarse mesh yields in most cases
accurate results. In summary, using a coarser representation of the displacement
field in the dynamic model is often acceptable in terms of accuracy and is required
in order to perform dynamic analyses in a reasonable time.
The structural analyst should thus in principle build two models, one for the
static simulations, one for the dynamic analysis. Knowing that building a model is
a significant part of the entire study, it would be very useful to device a procedure
that reduces the size of the dynamic problem without modifying the mesh grid.
Such procedures indeed exist and are known as reduction methods. Similarly
to the way one approximates continuous fields by a set of shape functions in a
Rayleigh–Ritz approach (e.g., in Finite Elements), the driving idea in reduction
techniques is to replace the degrees of freedom (DOF) by a set of global variables
representing the amplitudes of possible displacement modes:

u(t)  Tq(t) (1)

where u are all the n DOF of the system, T a reduction matrix of dimension n × r,
r < n, and q of dimension r × 1 is a set of reduced generalized degrees of freedom.
The columns of T define possible displacement shapes for the DOF u, and q are the
corresponding amplitude.
In general, replacing u by Eq. (1), only an approximate solution can be found and
the accuracy of the approximation will depend on how well the assumed modes in T
can represent the exact solution. Introducing Eq. (1) in the linear dynamic equation
of motion of the structure

Mü + Cu̇ + Ku = f(t) (2)

14 Substructuring Concepts and Component Mode Synthesis 835

one obtains

MTq̈ + CTq̇ + KTq̇ = f(t) + r (3)

where r is a residual force that is a remainder for the dynamic equilibrium equation;
since the reduction subspace T can in general not represent the exact solution, there
will always be an equilibrium error, whatever the choice of q. Indeed, one now has
only r unknowns q to satisfy n > r equations. Then following the idea of virtual
work, one requires that the reduced DOF q be chosen such that the residual force r
does not contribute to the dynamics in the representation space T or, in other words,
that the residual force does not produce any work for the possible shapes of motion
contained in T:

TT r = 0. (4)

With this condition, we can find the equations to determine q by projecting the
dynamic equilibrium equations (3) onto the subspace T:

TT MTq̈ + TT CTq̇ + TT KTq = TT f (5)

which is usually written as

M̃q̈ + C̃q̇ + K̃q = f̃ (6)

where the tilde superscript indicates that the matrices and vectors pertain now to a
representation in a reduced space.
After having solved the reduced problem (6) for q, one can build the solution
u for the physical DOF by substituting in Eq. (1). The residual force r for the full
problem can be computed by substitution in the original problem (3); it provides a
way to monitor, a posteriori, the error on the equilibrium.
In what follows, we will drop the damping term in order to clarify the presen-
tation. Reduction of the damping matrix, in particular for substructure reduction, is
discussed in Refs. [15, 16] and the references therein.

1.2 The Guyan–Irons Method

The cornerstone of every reduction method consists in finding a representation space

T of the solution that allows finding a good approximation and that can be obtained
with a computational cost significantly smaller than the one involved in solving the
full problem. One very common way to find a reasonably good representation space
was proposed in 1965 by Guyan [18] and Irons [23], as described next.
Let us consider the matrix equation that governs the system dynamics (2). To
reduce the size of matrices K and M, let us eliminate a subset of DOF. The
eliminated (condensed) and remaining DOF are written, respectively, u1 and u2 .
836 D. Rixen

Assuming that no forces are applied on u1 , the equation can be partitioned as

follows:
       
M22 M21 ü2 K22 K21 u2 f
+ = 2 (7)
M12 M11 ü1 K12 K11 u1 0

or

M22 ü2 + M21 ü1 + K22 u2 + K21 u1 = f2 (8)

M12 ü2 + M11 ü1 + K12 u2 + K11 u1 = 0. (9)

One may imagine separating the condensed coordinate u1 into two contributions

u1 = u1,stat + u1,dyn (10)

with the “static” part deduced from

u1,stat = −K−1
11 K12 u2 . (11)

This is obtained by neglecting the inertia forces in Eq. (9).

The static condensation algorithm consists in neglecting u1 ,dyn and in building
the reduction
   
u2 I
= T u2 =
GI
u2 (12)
u1 S

where TGI stands for the Guyan–Irons reduction matrix and where

S = −K−1
11 K12 (13)

is the static condensation matrix, the columns of which contain the so-called static
modes representing the static response of u1 for unit u2 displacements. In other
words, they represent the static deformation induced on u1 when a unit displacement
is given to one of the DOFs u2 . The retained DOFs u2 are sometimes called “master
DOFs” whereas the condensed ones, u1 , are often called “slave DOFs” since they
are statically enslaved to u2 .
The reduced matrices are then

KGI = TT KT = K22 − K21 K−1

11 K12 (14)

MGI = TT MT = M22 − M21 K−1 −1 −1 −1

11 K12 − K21 K11 M12 + K21 K11 M11 K11 K12 .
(15)
14 Substructuring Concepts and Component Mode Synthesis 837

We observe that the reduced stiffness matrix is the stiffness matrix statically
condensed on u2 . The reduced mass matrix is the mass matrix associated to u2
and augmented by the inertia of u1 assumed to respond quasi-statically.
The dynamic problem is then reduced to

MGI ü2 + KGI u2 + fGI

2 . (16)

If the static condensation algorithm is applied to static problems, the exact

solution is found. But when applied to dynamic problems, an approximation is
introduced by neglecting the dynamic response of the interior of the substructure
and thereby assuming that all condensed DoFS u1 respond quasistatically to the
u2 displacements. The validity of the condensation algorithm thus depends on the
extent to which the correction u1,dyn is negligible. It is possible to show (see, for
instance, Ref. [13]) that the static condensation technique is valid if

ω2 = μ21 (17)

where ω is the highest eigenfrequency that one wants to compute for the complete
structure and μ1 is the lowest eigenfrequency of the structure when u2 are clamped.
The complete analysis allows to show that static condensation always leads to
overestimating the eigenvalues compared to the full model. This is natural if we
recall that a model obtained by applying consistently a Rayleigh–Ritz approach is
always stiffer due to the restrictions imposed to the model by the discretization field
(see, for instance, Ref. [13]).
The algorithm described above is very frequently used in the context of finite
element structural analysis. Although originally proposed independently by Guyan
and Irons, it is commonly known as the Guyan reduction method. In commercial
codes, this method is often implemented and used to reduce the complexity of the
problem when no substructuring approach is used. The choice of an appropriate set
of master DOF u2 is sometimes done automatically by the software. One simple
heuristics consists in ranking the DOFs u according to their individual pseudo-
frequency computed by the ratio of the diagonals of the stiffness and mass matrices,
namely, for a DOF k, vk2 = Kkk /Mkk , and choosing as master DOF u2 the ones with
the lowest vi2 .

1.3 Model Reduction Through Substructuring

A very interesting way of applying reduction technique is found by applying the

following procedure:

1. Define subparts of the structure, called substructures. These substructures cor-

respond for instance to parts of the model that are analyzed and designed by
838 D. Rixen

Fig. 1 Example of a
substructured system. The red
nodes suggest the location of
interface nodes between
subcomponents. (Courtesy of
ESA/ESTEC)
ub

different teams. For an aircraft, this might be the wings, fuselage, stabilizers,
and tail. For a launcher system, it could be different stages of the rocket and
the payload (see the schematic in Fig. 1). For a vehicle, one could define as
substructures the engine block, the accessories, the suspension, and the car body.
We will denote every substructure by Ω (s) . The stiffness and mass matrices
corresponding to the nonassembled substructures will be denoted by K(s) and
M(s) , respectively. The DOF per substructure are denoted as u(s) .
2. For every subpart, define a reduction matrix T(s) that retains at least the DOF on
the interface boundary, called u(s)
b , such that

 
(s)
ub
u(s)
=T
(s)
(18)
ζ (s)

where ζ (s) are generalized degrees of freedom denoting the amplitudes of

(s)
representation modes additional to the ones governed by ub . The substructures
are now seen as macroelements (also called superelement) whose stiffness and
mass matrices are given by

T T
K̃(s) = T(s) K(s) T(s) M̃(s) = T(s) M(s) T(s) . (19)

These reduced matrices of the substructures can easily be shared between

different design teams. Such methods are sometimes called component mode
synthesis or CMS.
14 Substructuring Concepts and Component Mode Synthesis 839

(s)
3. The interface boundary degrees of freedom ub are assembled on the interface
of the substructures exactly as if one would assemble the macroelement.
If one first assumes that the entire reduction of a substructure is performed
(s)
by keeping only the interface DOF ub (i.e., no additional reduced DOF ζ (s)
per substructure), one can resort to the method of Guyan–Irons (see previous
section) where the interface DOF are the master DOF (previously called u(s) 2
when the Guayn–Irons reduction was applied to a nondecomposed problem).
With a simple Guyan–Irons reduction on the interface DOF of the substructures,
the dynamics of the DOF inside the substructures is neglected, which could be
a crude assumption if the eigenfrequencies of the substructures fixed on the
interface are not small (see criteria (17)). Hence additional information about
the substructure vibrations should be added to the reduction basis as explained in
the next section.

2 Numerical Techniques for Model Reduction

of Substructures

2.1 The Hurty/Craig–Bampton Method

The substructure reduction method discussed in this section is one of the most
commonly used substructuring technique in engineering practice. It was proposed
by Roy Craig in 1968 [6], writing in a more intuitive form ideas previously published
by Hurty [22]. Hence, although most commonly known as the Craig–Bampton
method, we will call it the Hurty/Craig–Bampton method.
The FE discretized system of equations of motion for a substructure reads

M(s) ü(s) + K(s) u(s) = f(s) , (20)

where M(s) and K(s) are the mass and stiffness matrices, respectively, u(s) is the
displacement vector, f(s) is the forcing vector, and the superscript s denotes the sth
substructure. The substructure equations of motion are partitioned into interior and
boundary (or interface) DOF (denoted by subscripts i and b, respectively) as
       
M(s) (s)
ii Mib ü(s) K(s) (s)
ii Kib u(s) 0
(s) (s)
i
(s) + (s) (s)
i
(s) = (s) (21)
Mbi Mbb üb Kbi Kbb ub fb

where the subscripts b and i are indexes referring to the boundary and internal
component of the matrices and vectors.
The representation modes are formed by computing the static response of the
interior of the substructure when one interface DOF is given a unit displacement
and all the other DOF are held fixed. The set of interface static modes for the entire
interface is expressed as
840 D. Rixen

 
(s)−1 (s)
 (s)
= −Kii Kib . (22)
I

The resulting basis Ψ (s) is used to statically eliminate all interface DOF from the
model, retaining only the boundary DOF. The resulting reduced system is usually of
small size since only boundary DOF are remaining unknowns. As mentioned earlier,
these static modes can be seen as a specific case of Guyan–Iron modes (see previous
section) for the case where the master DOFs are chosen on the interface.
In order to capture the dynamics of the system, the static modes are augmented
with a set of dynamic modes which are obtained by fixing the interface DOF and
solving the following eigenvalue problem

(s) (s) (s)
Kii − ωr2 Mii φi,r = 0. (23)

The eigenvectors obtained from this equation are referred to as fixed-interface

modes. A truncated set of m of these mass-normalized eigenvectors are collected
into a fixed-interface mode matrix
   
φ (s) , . . . , φ (s) (s)
i
 =
(s) i,1 i,m = . (24)
0 0

These modes provide a normal basis for the interior DOF of the substructure.
The fixed-interface modes and the interface static modes are combined to form the
HCB reduction matrix as

T(s)HCB = (s)  (s) , (25)

which provides a transformation from the substructure physical DOF to the HCB
generalized DOF,
   
(s) (s)
ui HCB(s) qi
≈T , (26)
u(s)
b u(s)
b

(s)
where qi represents the modal coordinate vector associated with the fixed-interface
modes. The meaning of the static modes and of the fixed interface modes as used in
the HCB method is illustrated in Fig. 2.
The uncoupled substructure reduced mass and stiffness matrices are now formed
by applying the HCB transformation to these matrices as
 T  T
M(s)HCB T(s)HCB M(s) T(s)HCB , K(s)HCB = T(s)HCB K(s) T(s)HCB . (27)

which, using Eq. (25), are given by

14 Substructuring Concepts and Component Mode Synthesis 841

Static mode Vibration mode

Fig. 2 Static modes and internal vibration modes of a substructure

 2
  
(s) 0 I M̃(s)
K (s)HCB
= m
(s) and M (s)HCB
= (s)
ib
(s) (28)
0 K̃bb M̃bi M̃bb

(s)2
with m the diagonal matrix of the eigenfrequencies of the fixed interface modes
and with the full submatrices

(s) (s) (s) (s)−1 (s)

K̃bb = Kbb − Kbi Kbb Kbi (29)

(s) (s) (s) (s)−1 (s) (s) (s)−1 (s)

M̃bb = Mbb − Mbi Kii Kib − Kbi Kii Mib
(30)
(s) (s)−1 (s) (s)−1 (s)
+ Kbi Kii Mii Kii Kib

T (s) (s) (s)−1 (s)
M̃ib = (s) Mib − Mii Kii Kib = M̃Tbi . (31)

The HCB reduced order models are typically coupled using a primal assembly
by defining a transformation between the uncoupled and coupled DOF that selects
the substructure boundary DOF from a unique global set of boundary DOF. This
assembly is very similar to the assembly of finite elements, except that now
macroelements with many nodes (the interface nodes) are assembled, the internal
(s)
DOFs qi remaining unassembled.
The choice of the number of fixed interface modes kept in the reduction basis for
each substructure can be made based on different criteria:

• A criterion often used in practice is based on the eigenfrequencies of the kept

modes. Typically, one chooses all fixed interface modes having an eigenfre-
quency lower than 1.8 or 2 times the highest frequency of interest in the assembly.
• One can also choose the fixed interface modes based on how easily they can be
excited through the support, for instance by defining a measure of the interface
reaction force associated to a mode. Following a similar reasoning, one evaluates
how complete the basis of the fixed interface modes needs to be by specifying
842 D. Rixen

how much of the mass of the substructure must be represented in its reduced
matrix. This can be evaluated using the concept of effective modal mass (see, for
instance, Ref. [13]).
• The number of modes can also be chosen based on a posteriori error estimators
[24], later reformulated in a less mathematical form and used for adaptive
selection strategies in Ref. [41].

2.2 Substructure Reduction Using Free Interface Modes

The discussion in this section is based on the overview in Ref. [14].

Considering the equation of motion of substructure s,

M(s) ü(s) + C(s) u̇(s) + K(s) u(s) = f(s) + g(s) . (32)

every substructure can be seen as being excited by the interface connection forces
and the external forces (contrary to the paradigm underlying the HCM method,
where the substructures are considered as being excited by interface displacements).
This indicates that the displacements of each substructure u(s) can be expressed
(s)
in terms of local static solutions ustat and in terms of eigenmodes associated to the
entire substructure matrices K and M(s) (hence the free interface modes):
(s)

n(s) −r (s)
(s) (s) (s)
u(s)
= ustat + θ j ηj (33)
j =1

where n(s) and r(s) are the number of DOF and of rigid body modes for substructure
s, and where the free interface modes are solutions of the eigenvalue problem

(s)2 (s)
K(s) − ωj M(s) θ j = 0. (34)

The static solution is written as follows (assuming no external forces are applied
for simplicity):

r (s)
(s) (s)+ (s) (s) (s)
ustat = −K g + rj αj . (35)
j =1

The static solution arises from solving Eq. (32) under the assumption that there
are no inertia forces and no external forces acting on the substructure. The notation
+
K(s) denotes the inverse of K(s) when there are enough boundary conditions to
prevent the substructure with a free interface from floating [32]. If a substructure is
+ (s)
floating, K(s) is the generalized inverse of K(s) and rj are the rigid body modes
14 Substructuring Concepts and Component Mode Synthesis 843

of s (see, for instance, Ref. [13] on how to compute a generalized inverse and the
rigid body modes).
If only a limited number of free interface modes are used for the substructure
dynamics equations (33), and (35) result in the approximation

+ T
u(s) ≈ −K(s) B(s) λ + R(s) α (s) + (s) (s)
η . (36)

The vector α (s) contains the amplitudes of the rigid body modes and the
(s)
vector η(s) contains the amplitudes of the retained nθ local free interface modes
(s) (s)
eigenmodes. The matrices R and Θ contain all rigid body modes and the
retained eigenmodes.
+
Since a part of the subspace spanned by Θ (s) is already included in K(s) , the
(s) +
residual flexibility matrix Gr can be used instead of K(s) , which is defined by

T (s) T
n(s) −r (s) θ (s) (s) nθ
θ (s) (s)
j θj (s)+ j θj
G(s)
r = =K − . (37)
(s)2 (s)2
(s)
j =nθ +1
ωj j =1 ωj

T
Note that, by construction, G(s)
r = Gr
(s)
, and it is computed using the second
equality in Eq. (37). For further properties of G(s)
r see, for instance, Ref. [32]. As a
result, the approximation of one substructure writes
⎡ ⎤
α (s)
u(s) ≈ R(s) (s)
G(s)
r A
(s)T ⎣ η(s) ⎦ . (38)
  g (s)
(s) b
T1

(s) T
Gr A(s) is the matrix containing the residual flexibility attachment modes of
T
substructure s, since the Boolean localization matrix A(s) simply picks the columns
(s)
of Gr associated to the boundary DOF. In other words,
 
(s)T 0
A gb(s) = = g(s) .
g(s)
b

The approximation (38) can now be used to reduce the substructure DOF. Using
the orthogonality properties of the modes in Eq. (38), the equation of motion of one
substructure becomes
⎡ ⎤ ⎡ (s) ⎤
α̈ (s) α 
(s)T
Mf ree ⎣ η̈(s) ⎦ + Kf ree ⎣ η(s) ⎦ = T1
(s) (s)
f(s) + g(s) (39)
(s) (s)
g̈b gb
844 D. Rixen

with the matrices

⎡ ⎤
0 0 0
(s) (s)T (s) ⎢ 2 ⎥
Kf ree = T1 K(s) T1 = ⎣ 0 (s) 0 ⎦ (40)
(s)
0 0 Gr,bb

where

(s) T
Gr,bb = A(s) G(s)
r A
(s)
(41)

and
⎡ ⎤
I0 0
(s) (s)T (s) ⎢ ⎥
Mf ree = T1 M(s) T1 = ⎣0 I 0 ⎦. (42)
(s)
0 0 Mr,bb

where
T
M(s)
r,bb = A Gr M Gr A
(s) (s) (s) (s) (s)
. (43)

(s) (s)
Gr,bb is the residual flexibility and Mr,bb is the interface inertia associated to
the residual flexibility related to the boundary DOF, respectively, and Ω (s) being a
(s) (s)
diagonal matrix containing the retained nθ eigenvalues ωj .

2.2.1 Rubin Method (RM)

The Rubin method was proposed in 1975 in Ref. [35] and, with the definitions
above, can be explained in a simple manner following the derivation proposed in
Ref. [30] as explained next.
In order to assemble in a primal manner the substructure equation of motion (39)
in the global system, a second transformation is applied by the RM. The force DOF
(s) (s)
gb are transformed back to the boundary displacements ub using Eq. (38):

(s) (s) (s) (s) (s) (s)

ub = A(s) u(s) = Rb α(s) + b η + Gr,bb gb (44)

(s) (s)
Rb and b are the subparts of R(s) and Θ (s) related to the boundary DOF,
respectively. From this equation, the interface force DOF can be solved as

(s) (s) (s) (s) (s) (s)
gb = Kr,bb ub − Rb α(s) − b η (45)

(s) (s)−1 (s) (s)

with Kr,bb = Gr,bb . The transformation matrix T2 from force DOF gb back to
(s)
the boundary displacements ub leaving α(s) and η(s) unchanged writes then
14 Substructuring Concepts and Component Mode Synthesis 845

⎡ ⎤
I 0 0
(s) ⎢ ⎥
T2 = ⎣ 0 I 0 ⎦. (46)
− K(s) (s) (s)
r,bb Rb −Kr,bb
(s)
b K(s)
r,bb

The RM approximation for one substructure writes therefore

⎡ ⎤
α(s)
= TR ⎣ η(s) ⎦
(s) (s) (s) (s)
u(s) where TR = T1 T2 (47)
(s)
ub

Application of this transformation to the matrices of Eqs. (42) and (40) gives the
RM reduced matrices of one substructure:
(s) (s)T (s) (s) (s)T (s)
Kred,R = T2 Kf ree T2 = TR K(s) TR (48)

(s) (s)T (s) (s) (s)T (s)

Mred,R = T2 Mf ree T2 = TR M(s) TR (49)

These matrices can be directly assembled using primal assembly to get the RM
reduced matrices Kred,R and Mred,R of the global system. The RM applies the
(s)
reduction matrix TR consistently to the mass and stiffness matrix resulting in a
true Rayleigh–Ritz method as was observed in Ref. [10].

2.2.2 MacNeal Method (MNM)

The MNM [29] is nearly identical to the RM except for a small change. First, we will
(s) (s)
derive the preliminary MNM reduced matrices K̃red,MN and M̃red,MN following
the derivation of the RM to show the similarities between these two methods. The
reduced stiffness matrix of both the RM and the MNM are identical (given in Eq.
(48))
(s) (s)
K̃red,MN = Kred,R (50)

but the MNM reduced mass matrix M̃(s)

red,MN is obtained differently. The residual
(s) (s)
mass term Mr,bb of the matrix Mf ree in Eq. (42) is neglected, resulting in a modified
matrix [40].
⎡ ⎤
I 00
M(s)
f ree,MN = ⎣0 I 0⎦. (51)
000

The MNM reduced mass matrix writes now

(s) (s)T (s) (s) (s)
M̃red,MN = T2 Mf ree,MN T2 = Mf ree,MN . (52)
846 D. Rixen

This gives in fact inconsistent equations of motion since the mass and stiffness
matrices are not reduced with the same basis. The assembly of the MNM reduced
(s) (s)
matrices K̃red,MN and M̃red,MN in the global system proceeds in the same manner
as for the RM. Observing that the boundary DOF ub have no associated inertia in
Eq. (52), those DOF can be condensed out of the equation of motion of the
assembled problem and the final MNM reduced matrices Kred,MN and Mred,MN are
obtained [29]. Thus, the size of the assembled MNM system is reduced further by
the number of DOF of ub .

2.2.3 Dual Craig–Bampton Method (DCBM)

The reduction transformation (Eq. 38) can also be used directly, without transform-
(s)
ing gb in interface DOF (as done in Eq. (45) for the RM or for the MNM), if one
considers the dually assembled problem. The dual assembly of the full (original)
problem writes
⎡ .. ⎤ ⎡ .. ⎤ ⎡ . . .. ⎤⎡ . ⎤ ⎡ . ⎤
.. .. ..
⎢
. 0 . ⎥⎢ . ⎥ ⎢ . 0 .
⎥ ⎢ (s) ⎥ ⎢ (s) ⎥
⎢ 0 M(s) ⎥ ⎢ (s) ⎥
0 ⎥ ⎢ ü ⎥ ⎢ 0 ⎢ K(s) B(s) T
⎥⎢u ⎥ ⎢f ⎥
⎢ + ⎥⎢ ⎥=⎢ ⎥
⎢ . . .. ⎥ ⎢ . ⎥ ⎢ .. . ⎥ ⎢ .. ⎥ ⎢ .. ⎥ .
⎣ . . ⎦ ⎣ .. ⎦ ⎣ . .. ⎦⎣ . ⎦ ⎣ . ⎦
... 0 ... 0 λ̈ . . . B(s) . . . 0 λ 0
(53)

Here, the compatibility constraints between the substructures are written explic-
itly and are enforced using Lagrange multipliers λ. Those Lagrange multipliers can
be interpreted as the internal force intensities that are required between matching
interface nodes in order to assemble them. The matrices B(s) are signed Boolean
matrices that express the compatibility condition between interface DOFs. For a
more detailed discussion on the dual assembly and its formulation, see for instance
[26, 32, 39].
Comparing the equations of motion of the substructures in Eq. (53) with the
formulation used earlier in Eq. (32), it is clear that
T (s) T
g(s) = A(s) gb = −B(s) λ. (54)

Assembling all substructures N in a dual fashion, the entire problem unknowns

can consequently be approximated by rewriting Eq. (38) as
⎡
⎤
α (1)
⎢ η(1) ⎥
⎢ ⎥
  ⎢ . ⎥
u ⎢ .. ⎥
≈ TDCB ⎢
⎢
⎥
⎥ (55)
λ ⎢ α (N ) ⎥
⎢ (N ) ⎥
⎣η ⎦
λ
14 Substructuring Concepts and Component Mode Synthesis 847

with the DCBM reduction matrix TDCB :

⎡ T
⎤
(1)
R(1) (1)
0 0 −Gr B(1)
⎢ .. ⎥
⎢ .. .. ⎥
TDCB =⎢
⎢
. . . ⎥. (56)
(N ) (N )T ⎥
⎣ 0 0 R(N ) (N )
−Gr B ⎦
0 0 0 0 I

The reduced form of the dynamic equations of the dual assembled system (53) is
then found as
⎡ ⎤ ⎡ ⎤
α̈ (1) α (1)
⎢ η̈(1) ⎥ ⎢ η(1) ⎥
⎢ ⎥ ⎢ ⎥
⎢ . ⎥ ⎢ . ⎥
⎢ . ⎥ ⎢ . ⎥
Mred,DCB ⎢ . ⎥ ⎢ . ⎥
⎢ (N ) ⎥ + Kred,DCB ⎢ (N ) ⎥ = TDCB f
T
(57)
⎢ α̈ ⎥ ⎢α ⎥
⎢ (N ) ⎥ ⎢ (N ) ⎥
⎣ η̈ ⎦ ⎣η ⎦
λ̈ λ

with the DCBM reduced mass and stiffness matrix

   
M0 I 0
Mred,DCB = TTDCB TDCB = (58)
0 0 0 Mr

 
K BT
Kred,DCB = TTDCB TDCB (59)
B 0

with

N
T
Mr = B(s) G(s) (s) (s) (s)
r M Gr B . (60)
s=1

Mred,DCB and Kred,DCB are diagonal for the parts related to the different sub-
structures. The coupling between the substructures is only achieved by the rows and
columns related to λ. The DCBM applies the reduction matrix TDCB consistently to
the mass and stiffness matrix resulting in a true Rayleigh–Ritz method.
The DCBM enforces only a weak compatibility between the substructures
and does not enforce a strong displacement compatibility between the interfaces
compared to many other common reduction methods [32]. Considering the system
of Eqs. (53) and (32) multiplied by the reduction matrix TTDCB , the last row of
Eq. (57) results from
848 D. Rixen

⎡ T ⎤
M(1) ü(1) + K(1) u(1) + B(1) λ = f(1)
⎢ .. ⎥
⎢ ⎥
⎢ . ⎥
⎢ (N ) (N ) T ⎥ (61)
⎢ M ü + K(N ) u(N ) + B(N ) λ = f(N ) ⎥
⎢ ⎥
⎣ 
N ⎦
B(s) u(s) = 0
s=1

multiplied from left by the last row of TTDCB which is

−B(1) G(1)
e ··· − B
(N ) (N )
Gr I . (62)

Replacing the strong interface compatibility condition of Eq. (53) by the weak
form according to the multiplication of Eq. (61) by Eq. (62) can be interpreted as
follows. Denote f(s) the residual forces of substructure s resulting from the weak
satisfaction of the local equilibrium of the substructure approximating the dynamics
(s)
by a small number of free interface normal modes. Name u(s) = Gr f(s)
(d)
the displacements these residual force f would create locally. Then the weak
interface compatibility condition (Eqs. 61 and 62) states that a compatibility error
(i.e., an interface displacement jump) equal to the incompatibility of u(s) is
permitted [32].
The fact that a weak interface compatibility is allowed in the DCBM implies
that the infinite eigenvalues related to the Lagrange multipliers λ in the nonreduced
problem (53) are now becoming finite and negative [34]. In practice those negative
eigensolutions will appear only in the higher eigenvalue spectrum if the reduction
space is rich enough [34]. Nevertheless, the reduction basis has to be selected
with care avoiding potential nonphysical effects of the possibly occurring negative
eigenvalues. Using the dually reduced problem can nevertheless be used for time
integration as shown in Ref. [17] when using an appropriate modal superposition.
If Mr in Eq. (58) is neglected, strong interface compatibility is enforced again
and the DCBM reduced system with Mr = 0 is equivalent to the MNM [32]. Then
static condensation can be applied again to remove λ (as it was done for ub at the
end of the derivation of the MNM) from the assembled system since no mass is
associated to them.

2.3 Numerical Examples of Different Substructure Reduction

Techniques

The Benfield truss [8] of Fig. 3 is used to compare the results obtainable by
the HCB, the MNM, the RM, and the DCBM. The planar truss consists of 2
substructures having uniform bay section whereas all members have constant
area and uniform stiffness and mass properties. The left component consists
of 5 equal bays and has a total of 18 joints and the right component consists of
4 equal bays and has a total of 15 joints [8]. The lowest eigenfrequencies ω of
14 Substructuring Concepts and Component Mode Synthesis 849

Fig. 3 Benfield truss [8]

Fig. 4 Relative error εrel, j of

eigenfrequency j using five
normal modes per
substructure for the
approximation of the lowest
eigenfrequencies of the
Benfield truss

the entire structure shall be approximated by the different methods. The relative
error εrel, j = | ωred, j − ωfull, j | /ωfull, j of the j-th eigenfrequency is used as a
criterion to assess the accuracy of the different methods. Thereby, ωfull, j is the
j-th eigenfrequency of the full (nonreduced) system and ωred, j represents the j-th
eigenfrequency of the reduced system obtained by each method.
Using five elastic (fixed or free interface normal modes) per substructure, the
resulting relative errors εrel are depicted in the semilog graph in Fig. 4. Since all
methods give the correct rigid body modes, only the relative errors of the elastic
modes are plotted. All methods give a relative error less than 1% for the first six
eigenfrequencies. Comparing the free interface methods for this example, the RM
performs always better than the DCB and the DCB performs again always better as
the MNM. The HCB and the DCB result in similar frequency errors.
The sparsity pattern of the reduced stiffness matrix Kred and reduced mass matrix
Mred of the HCB (Fig. 5), the MNM (Fig. 6), the RM (Fig. 7), and the DCB (Fig. 8),
respectively, illustrate the differences of the assembled reduced structures. Both
the reduced stiffness matrix Kred and the reduced mass matrix Mred applying the
HCB and the DCB, respectively, have only diagonal entries for the subparts of each
substructure. On the one hand, the coupling between the substructures using the
HCB is entirely achieved by the last rows and last columns in the mass matrix
Mred,CB (Fig. 5b) and the remaining part is diagonal [6]. On the other hand, the
coupling applying the DCB is entirely achieved by the last rows and last columns
in the stiffness matrix Kred,DCB (Fig. 8a) and again the remaining part is diagonal
850 D. Rixen

Fig. 5 Sparsity pattern of the

reduced matrices applying the
HCB using five normal
modes per substructure

Fig. 6 Sparsity pattern of the

reduced matrices applying the
MNM using five normal
modes per substructure

Fig. 7 Sparsity pattern of the reduced matrices applying the RM using five normal modes per
substructure

[32]. The corresponding degrees of freedoms are either the interface displacements
ub or the interface forces λ but no direct coupling between the modal parameters of
adjacent substructures occurs which ensures the sparse structure.
In contrast, the sparsity pattern of stiffness matrix Kred and the reduced mass
matrix Mred obtained but the MNM and the RM, respectively, show full matrices.
The MNM gives indeed an entirely diagonal reduced mass matrix Mred,MN (Fig. 6b)
but causes always a full coupling between all DOF of all substructures via the
reduced stiffness matrix Kred,MN (Fig. 6a). This makes the reusability of reduced
14 Substructuring Concepts and Component Mode Synthesis 851

Fig. 8 Sparsity pattern of the reduced matrices applying the DCB using five normal modes per
substructure

Table 1 Number n of nonzero elements in the reduced matrices obtained by the different methods
for the Benfield truss using five normal modes per substructure
HCB MNM RM DCB
n in Kred 40 216 314 196
n in Mred 118 16 354 50
ntotal 158 232 668 246

models obtained by the MNM very inefficient and therefore nearly impossible from
a practical point of view. The RM also causes a coupling between the substructures
via interface displacements ub in the reduced stiffness matrix Kred,R (Fig. 7a) as
well as in the reduced mass Mred,R (Fig. 7b). Moreover, all DOF belonging to one
reduced substructure are coupled with all other DOF of the same substructure which
is why the reduced matrices of the RM are full for the substructure blocks and not
diagonal.
This result concerning the sparsity of the reduced matrices is outlined in Table 1
which shows the number n of nonzero elements in the reduced matrices Kred and
Mred , and the sum ntotal of both obtained by the different methods for this example.
The reduced matrices of the HCB, the MNM, and the DCB contain a similar number
of entries while the RM causes even for such a simple example a remarkable high
number of entries. The number of entries of the MNM are comparable to the
HCB and the DCB but will increase significantly if the number of substructures
is increased since Kred will always be completely full.

2.4 Other Reduction Techniques for Substructures

In the previous sections, we have outlined the classical substructure reduction

methods (or component mode synthesis) using either free or fixed interface modes.
852 D. Rixen

Other methods are not outlined here, like for instance the Craig–Chang approach
which uses also free interface modes, but computes the interface forces from a global
(assembled) problem [10].
Many other variants of CMS methods were published over the last years and they
can be classified as follows:

• Loaded Interface Modes: Some authors have proposed to compute the qua-
sistatic modes around a central frequency [36] or to use vibration modes with
an impedance attached to the interface. In certain cases, such modifications can
improve the accuracy of the reduced model.
• Modal Truncation Augmentation and Moment Matching: The vibration
modes used for instance in the HCB or in the RM approaches are not specifically
tuned for the excitations coming through the interface, whereas the static modes
are. It is possible to enrich the static modes with higher-order static contribution
computed over a Krylov series. This leads to a method originally called modal
truncation augmentation (or MTA) [12] and was generalized to higher orders
corrections in the HCB in Ref. [31]. Later, techniques called moment matching
[38], that basically used the same ideas as MTAs, were applied for model
reduction mainly in the control community (see, for instance, Refs. [2, 20, 37]).
The idea of MTA was also applied for the DCBM [25, 33].
• Balanced Truncation: In the control community, reduction is seen from the
point of view of controllability and observability. This leads to representation
modes derived from so-called Grammians that were used for reducing substruc-
ture models in Refs. [20, 37].
• Mixed Methods: Several combined methods were, where different types of
vectors (attachment modes and static modes were combined). One of the most
commonly used in major finite element software (although not cited much in
the community, maybe because the explanations in the publication are poor)
is the one from Herting [19]. Note also that it is possible to mix primal and
dual assembly for different DOF on an interface (which can be advantageous
depending on the stiffness ratios across the interface) [40].
• Finally, let us mention that methods have been proposed where the reduction is
performed in an iterative manner, computing the residual force on the full model
after a reduction (see Eq. (3)) and enriching the reduction base with, for instance,
the static response of the substructures to that residual force [5, 7].

3 Interface Reduction with the Hurty/Craig–Bampton

Method: Characteristic Constraint Modes

3.1 Interface Reduction Approaches

While the number of degrees of freedom in a model may be reduced dramatically

using substructuring techniques such as the Hurty/Craig–Bampton (HCB) method,
that method retains all of the DOF at the interfaces between substructures, and so the
14 Substructuring Concepts and Component Mode Synthesis 853

resulting model may still be unacceptably large. Furthermore, since the minimum
timestep for time integration is limited by the distance between the two closest
nodes, the HCB model may still be almost as expensive to integrate as the original
model. Several methods have been proposed over the years to address this problem,
and while none has been widely implemented as has the HCB method, they do seem
to be effective in many scenarios. This section presents a brief review of methods
for interface reduction. For further details and a comparison of each method on the
W-bracket, see, for instance, the review in Ref. [27].
Craig and Chang [11] seem to have been the first to propose methods to reduce
the interface DOF, presenting three methods in 1977. However, they do not appear
to have been used subsequently until Castanier et al. [9] rediscovered the modal
method by applying a secondary eigenvalue analysis on the interface partition
of the assembled CMS model and obtaining what they called the characteristic
constraint (CC) modes. The assembled, system-level HCB matrices were used to
compute these modes, so this method is referred to as the system-level characteristic
constraint (S-CC) mode method. While the method is effective, it is often undesir-
able to have to assemble the system before reduction and so this has inspired the
investigation of other methods that perform the reduction before assembly.
Hong et al. [21] subsequently proposed an interface reduction technique that
performs an eigenvalue analysis on the HCB interface DOF prior to assembly. They
then concatenate the shapes obtained for each interface and enforce compatibility
between the local-level characteristic constraint (L-CC) modes obtained from each
interface. This method is referred to in Ref. [27] as the “exact compatibility L-CC
method,” because it is possible to enforce exact compatibility using that method.
However, in practice some of the shapes are typically truncated and there is some
level of approximation.
Alternatively, Kuether et al. [28] proposed a method that weakens the compati-
bility at the interface. The L-CC modes are combined to minimize the compatibility
error between connecting substructure interfaces. Aoyama et al. [3] presented a
method similar to S-CC except that each interface is assembled and then reduced
separately. Balmès [4] explored a CMS basis that defined arbitrary interface
deformations to describe a set of generalized DOF along an interface. A few other
less commonly known methods are reviewed in Ref. [27].
In the next section, the method based on so-called system-level characteristic
constraint (S-CC) modes is shortly described.

3.2 System-Level Characteristic Constraint (S-CC) Modes

As done by Craig and Chang [11], and detailed by Castanier et al. [9], the HCB
models for all subcomponents are first assembled, the interface partition of the
assembled system is extracted and then an eigenvalue analysis is performed

bb − ωr Mbb
KHCB 2 HCB
φ Sr -CC = 0. (63)
854 D. Rixen

The eigenvectors obtained are called S-CC modes and the first n vectors are
collected into a matrix

S-CC = φ S1 -CC , . . . , φ Sn-CC . (64)

The matrix Φ S-CC is used to replace the interface DOF of the HCB model into
amplitudes of the S-CC modes

ub ≈ S-CC qb . (65)

The S-CC reduced mass (and similarly stiffness) matrices are obtained by pre-
and post multiplying the assembled HCB matrices in Eq. (28) with the S-CC
transformation matrix to obtain
 T   
S-CC I 0 MHCB MHCB I 0
M = ii ib . (66)
0 S-CC Mbi MHCB
HCB
bb 0 S-CC

This reduces the size of the interface partition of the HCB model from the number
of physical DOF on the interface to the number of S-CC modes in the truncated
eigenvector set nb . It is interesting to note that static reduction modes were also
proposed for the interface of a dual Criag–Bampton model in Ref. [34].

References
1. Allen MS, Rixen D, van der Seijs M, Tiso P, Abrahamsson T, Mayes TL (2020) Substructuring
in engineering dynamics. In: CISM international centre for mechanical sciences. Vol 594.
Springer
2. Antoulas AC (2005) Approximation of large-scale dynamical systems. Society for Industrial
and Applied Mathematics, Philadelphia
3. Aoyama Y, Yagawa G (2001) Component mode synthesis for large-scale structural eigenanal-
ysis. Comput Struct 79(6):605–615
4. Balmès E (1996) Use of generalized interface degrees of freedom in component mode
synthesis. In: Proceedings of the 14th international modal analysis conference, 1
5. Balmès E (1997) Modèles expérimentaux complets et modèles analytiques réduits en
dynamiquedes structures. PhD thesis, Université Pierre et Marie Curie, Paris VI, ONERA,
Directions des Structure, DRET/AC
6. Bampton MCC, Craig JRR (1968) Coupling of substructures for dynamic analyses. AIAA J
6(7):1313–1319
7. Bathe KJ, Dong J (2014) Component mode synthesis with subspace iterations for controlled
accuracy of frequency and mode shape solutions. Comput Struct 139:28–32
8. Benfield W, Hruda R (1971) Vibration analysis of structures by component mode substitution.
AIAA J 9(7):1255–1261
9. Castanier MP, Tan YC, Pierre C (2001) Characteristic constraint modes for component mode
synthesis. AIAA J 39(6):1182–1187
10. Craig RR, Chang CJ (1977) On the use of attachment modes in substructure coupling
for dynamics analysis. In: 18st structures, structural dynamics and material conference.
AIAA/ASME, San Diego, pp 89–99
14 Substructuring Concepts and Component Mode Synthesis 855

11. Craig RR Jr, Chang CJ (1977) Substructure coupling for dynamic analysis and testing. Report,
National Aeronautics and Space Administration
12. Dickens JM, Stroeve A (2000) Modal truncation vectors for reduced dynamic substructure
models. In: 41st AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics, and materials
conference and exhibit, Atlanta
13. Géradin M, Rixen D (2015) Mechanical vibrations. Theory and application to structural
dynamics, 3rd edn. Wiley, Chichester
14. Gruber FM, Rixen DJ (2016) Evaluation of substructure reduction techniques with fixed and
free interfaces. Stroj Vestn–J Mech E 62(7–8):452–462
15. Gruber F, Rixen D (2018) A comparison of Craig-Bampton approaches for sys-
tems with arbitrary viscous damping in dynamic substructuring. In: IMAC-XXXVI:
international modal analysis conference, Orlando. Society for Experimental Mechanics,
Bethel
16. Gruber FM, Rixen DJ (2018) Dual Craig-Bampton component mode synthesis method for
model order reduction of nonclassically damped linear systems. Mech Syst Signal Process
111:678–698
17. Gruber FM, Gille M, Rixen DJ (2019) A strategy to stabilize the transient analysis and increase
the approximation accuracy of dual Craig-Bampton reduced systems. Finite Elem Anal Des
160:32–45
18. Guyan R (1965) Reduction of mass and stiffness matrices. AIAA J 3(2):380
19. Herting D (1985) A general purpose, multi-stage, component modal synthesis method. Finite
Elem Anal Des 1(2):153–164
20. Holzwarth P, Eberhard P (2015) SVD-based improvements for component mode synthesis in
elastic multibody systems. Eur J Mech A Solids 49:408–418
21. Hong SK, Epureanu BI, Castanier MP (2013) Next-generation parametric reduced-order
models. Mech Syst Signal Process 37(1–2):403–421
22. Hurty WC (1965) Dynamic analysis of structural systems using component modes. AIAA J
3(4):678–685
23. Irons B (1965) Structural eigenvalue problems – elimination of unwanted variables. AIAA J
3(5):961–962
24. Jakobsson H, Bengzon F, Larson MG (2011) Adaptive component mode synthesis in linear
elasticity. Int J Numer Methods Eng 86:829–844
25. Kim JH, Kim J, Lee PS (2017) Improving the accuracy of the dual Craig-Bampton method.
Comput Struct 191:22–32
26. Klerk DD, Rixen DJ, Voormeeren SN (2008) General framework for dynamic substructuring:
history, review and classification of techniques. AIAA J 46(5):1169–1181
27. Krattiger D, Wu L, Zacharczuk M, Buck M, Kuether RJ, Allen MS, Tiso P, Brake MRW (2019)
Interface reduction for Hurty/Craig-Bampton substructured models: review and improvement.
Mech Syst Signal Process 114:579–603
28. Kuether RJ, Allen MS, Hollkamp JJ (2017) Modal substructuring of geometri-
cally nonlinear finite element models with interface reduction. AIAA J 55(5):1695–
1706
29. MacNeal RH (1971) Special issue on structural dynamics: a hybrid method of component mode
synthesis. Comput Struct 1(4):581–601
30. Martinez D, Carne T, Gregory D, Miller A (1984) Combined experimental/analytical modelling
using component mode synthesis. In: 25th structures, structural dynamics and materials
conference, pp 140–152
31. Rixen D (2002) High order static correction modes for component mode synthesis. In: Fifth
world Congress on computational mechanics, Vienna
32. Rixen D (2004) A dual Craig-Bampton method for dynamic substructuring. J Comput Appl
Math 168(1–2):383–391
33. Rixen DJ (2009) Dual Craig-Bampton with enrichment to avoid spurious modes. In:
IMACXXVII: international modal analysis conference, Orlando. Society for Experimental
Mechanics, Bethel
856 D. Rixen

34. Rixen DJ (2011) Interface reduction in the dual Craig-Bampton method based on dual interface
modes. In: IMAC-XXIX: international modal analysis conference, Jacksonville. Society for
Experimental Mechanics, Bethel
35. Rubin S (1975) Improved component-mode representation for structural dynamic analysis.
AIAA J 13(8):995–1006
36. Shyu WH, Ma ZD, Hulbert GM (1997) A new component mode synthesis method: quasi-static
mode compensation. Finite Elem Anal Des 24(4):197–298
37. Stoev J, Möhring J (2010) Hybrid reduction of non-symmetric mimo systems. In: Paul Sas
et al. (ed) ISMA. KU Leuven, Leuven
38. Su TJ, Craig R (1991) Model reduction and control of flexible structures using Krylov vectors.
J Guid Control Dyn 14(2):260–267
39. Voormeeren S (2012) Dynamic substructuring methodologies for integrated dynamic analysis
of wind turbines. PhD thesis, Delft University of Technology, Delft
40. Voormeeren SN, van der Valk PLC, Rixen DJ (2011) Generalized methodology for assembly
and reduction of component models for dynamic substructuring. AIAA J 49:1010–1020
41. Voormeeren S, Nortier B, Rixen D (2013) Error estimation and adaptive model reduction
applied to offshore wind turbine modeling. In: IMAC-XXXI: international modal analysis
conference, error estimation and adaptive model reduction applied to offshore wind turbine
modeling. Society for Experimental Mechanics, Bethel
Finite Element Model Correlation
15
Peter Avitabile and Michael Mains

Contents
1 Introduction/Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
2.1 Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 862
2.2 Model Expansion (Vector Expansion) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 870
2.3 Test Data Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 876
2.4 Vector Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 879
3 Closing Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 889
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 891
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 891

Abstract

Correlation of a finite element model with test data is commonly performed.

In order to perform these correlation studies, the finite element model may
require reduction due to the large size of the model, or the test data may be
expanded to the size of the finite element model. Model reduction and model
expansion techniques are presented first. Correlation tools typically deployed
are then presented. Some additional commentary related to the test data and the
correlation process is also provided to give insight into some of the issues that
must be faced.

P. Avitabile ()
Structural Dynamics and Acoustic Systems Laboratory, University of Massachusetts Lowell,
Lowell, MA, USA
e-mail: [email protected]
M. Mains
University of Cincinnati, Cincinnati, OH, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 857

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_17
858 P. Avitabile and M. Mains

Keywords

Correlation · Modal correlation · Vector correlation · Orthogonality · Cross

orthogonality · Pseudo orthogonality · Model reduction · Reduced order
model · Vector expansion

Acronyms

FEM Finite element model

EMA Experimental modal analysis
SEREP System equivalent reduction expansion process
IRS Improved reduced system
MAC Modal assurance criteria
CoMAC Coordinate modal assurance criteria
POC Pseudo orthogonality check
CORTHOG Coordinate orthogonality check
FRAC Frequency response assurance criteria
RVAC Response vector assurance criteria
MACCO Modal assurance criteria contribution
FRF Frequency response function
DOF Degree of freedom
NDOF Full set of “n” finite element DOF
ADOF Reduced set of “a” test DOF
DDOF Remaining set of “d” deleted or omitted DOF

Nomenclature

Matrix
[M] Analytical mass matrix
[C] Analytical damping matrix
[K] Analytical stiffness matrix
[U] Analytical modal matrix
[I] Diagonal modal mass matrix
[2 ] Diagonal modal stiffness matrix
[T] Transformation matrix
[E] Experimental modal vectors
[D] Dynamic matrix

Vector
{ẍ} Acceleration
{ẋ} Velocity
{x} Displacement
{F} Force
{p} Modal displacement
{u} Modal vector
{ei } ith experimental modal vector
{uj } jth analytical modal vector
Scalars
ek kth degree of freedom of the experimental modal vector
up pth degree of freedom of the analytical modal vector
15 Finite Element Model Correlation 859

Subscript
n Full set of finite element DOF
a Tested set of experimental DOF (also master/active DOF)
d Deleted (omitted) set of DOF
S Static condensation
I IRS condensation
f Dynamic condensation
U SEREP condensation
H Hybrid condensation
k,p Degree of freedom identifiers
i,j Mode identifiers

Superscript
T Transpose
g Generalized inverse
−1 Standard inverse
* Conjugate

1 Introduction/Background

Correlation of a finite element model with test data is commonly performed to

identify the suitability of the model for response predictions. This may be required
contractually or performed to improve the overall system performance based on
design requirements or internally generated specifications to assure good modeling
techniques are utilized in the design and analysis process.
The finite element model is typically very large with many degrees of freedom,
whereas the test data acquired typically has very few degrees of freedom. This
mismatch then requires that the finite element model be reduced to the size of the
test model or the test data expanded to the size of the finite element model; both of
these can introduce errors in the correlation process. Model reduction and expansion
techniques commonly employed are presented. The typical correlation tools often
deployed are also presented.
The biggest problem to overcome in the correlation of vectors is the model
reduction and vector expansion process which can have a direct effect on the results
depending on the particular techniques utilized. One other difficulty is that all of
the correlation tools may provide some indication of discrepancies and hint to
where problems may exist, but it must be clearly understood that the test data and
the finite element model can both have deficiencies, so the correlation tools may
provide conflicting results and may not always provide the same overall indication
of the model adequacy based on the measured data. In fact, there can be cases where
the test data (which is not perfect by any means) has errors, and this complicates
the correlation process. The user of any of these correlations tools must be very
careful in any conclusion obtained from each tool. Engineering judgment and a
clear understanding of both the finite element modeling theories and experimental
testing approaches is of paramount importance in any correlation analysis.
860 P. Avitabile and M. Mains

Some of the relevant theory is presented for model reduction and model expan-
sion along with the commonly used correlation tools. The references are provided
at the end of this section to assist the reader in finding additional information related
to this and are provided in a bibliography for the different tools and techniques.

2 Theory

The basic matrix equation of motion is generally written as

[Mn ] {ẍn } + [Cn ] {ẋn } + [Kn ] {xn } = {Fn (t)} (1)

where the [M], [C], and [K] are the mass, damping, and stiffness matrices of
the system and {F} is the vector of applied forces on the system and {x} is the
vector of displacements of the system (with appropriate derivatives of velocity
and acceleration); the subscript “n” denotes the full dimensionality of the set of
equations. Assuming that the damping matrix is proportional to either the mass or
stiffness matrix (which is often assumed to be the case), the eigensolution can be
written as

[[Kn ] − λ [Mn ]] {xn } = {0} (2)

with ωi {ui } as the resulting eigenvalue and eigenvector, respectively. The eigen-
vectors can be arranged in column fashion to form the modal matrix [U]. Using this
notation and noting the eigenvalues can be assembled into a diagonal matrix, the
eigen problem can be restated as
 
[Kn ] [Un ] = [Mn ] [Un ] 2 (3)

where
⎡ ⎤ ⎡ ⎤
\ ω21
⎣ 2 ⎦ = ⎣ ω22 ⎦ and [U] = [{u1 } {u2 } · · · ] (4)
\ \

Using the modal matrix, a transformation can be made from physical space to
modal space using the relationship:
⎧ ⎫
⎪
⎨ p1 ⎪
⎬

{x} = [U] {p} = {u1 } {u2 } · · · p2 (5)
⎪
⎩ .. ⎪
⎭
.

Substituting Eq. (5) into the equation of motion in Eq. (1) and premultiplying by
the transpose of the projection operator to put the equations into normal form give
15 Finite Element Model Correlation 861

the standard modal space representation:

[Un ]T [Mn ] [Un ] {p̈n } + [Un ]T [Kn ] [Un ] {pn } = [Un ]T {Fn (t)} (6)

Due to the orthogonality of the modal vectors with respect to the system mass and
stiffness matrices, the highly coupled set of physical mass, damping (proportionality
assumed), and stiffness matrices become uncoupled in modal space to form the
modal mass, modal damping (with proportional damping shown for completeness),
and modal stiffness of the system in modal space given as

⎡ ⎤ ⎧ p̈ ⎫ ⎡ ⎤ ⎧ ṗ ⎫
m1 ⎪
⎨ 1⎬ ⎪ c1 ⎪
⎨ 1⎪ ⎬
⎣ m2 ⎦ p̈2 + ⎣ c2 ⎦ ṗ2
⎪
⎩ .. ⎪
⎭ ⎪
⎩ .. ⎪⎭
\ . \ .
⎡ ⎤ ⎧p ⎫ ⎧ ⎫ (7)
⎨ {u1 } {F} ⎪
T
k1 ⎪
⎨ 1⎪ ⎬ ⎪ ⎬
+ ⎣ k2 ⎦ p2 = {u2 }T {F}
⎪
⎩ .. ⎪⎭ ⎪
⎩ .. ⎪
⎭
\ . .

or as
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
\ \ \
⎣ M ⎦ {p̈} + ⎣ C ⎦ {ṗ} + ⎣ K ⎦ {p} = [U]T {F} (8)
\ \ \

where the bar overscore denotes modal space.

When the mode shapes are scaled to unit modal mass, this relationship reduces to

[In ] {p̈n } + [n ] {pn } = [Un ]T {Fn (t)} (9)

It is important to note that the diagonal modal mass matrix is given as

[Un ]T [Mn ] [Un ] = [I] (10)

and the diagonal modal stiffness matrix is given as

 
[Un ]T [Kn ] [Un ] = 2 (11)

(Note: Subscript “n” here denotes normalized and elsewhere denotes size of
matrix.)
862 P. Avitabile and M. Mains

2.1 Model Reduction

This section presents some of the basic approaches to the reduction of finite element
models. Model reduction is typically performed to obtain a reduced model for
other structural dynamic applications such as forced response analysis, system
modeling, and component model synthesis techniques. However, for this work,
model reduction is specifically used to form a mapping between the very large set
of finite element degrees of freedom and the relatively small set of tested degrees of
freedom necessary for model correlation studies.
Most reduction or condensation techniques affect the dynamic character of the
resulting reduced model. Model reduction is performed for a number of reasons,
but the technique is used primarily as a mapping technique. A schematic of the
reduction process is shown in Fig. 1.
In general a relationship between the full set of analytical or finite element DOF
and the reduced set of active or condensed DOF can be written as
 
xa
{xn } = = [T] {xa } (12)
xd

The “n” subscript denotes the full set of analytical DOFs, the “a” subscript
denotes the active set of DOF (sometimes referred to as master DOF and for
correlation studies referred to as test DOF), and the subscript “d” denotes the deleted
DOF (sometimes referred to as embedded or omitted DOF); the [T] transformation
relates the transformation between these two sets of DOFs.
The reduced mass matrix and reduced stiffness matrix can be expressed as

[Ma ] = [T]T [Mn ] [T] (13)

and

FEA ‘N’ DOF ACTIVE OR MASTER OR

TEST ‘A’ DOF

Reduction

Fig. 1 Schematic of reduction process

15 Finite Element Model Correlation 863

[Ka ] = [T]T [Kn ] [T] (14)

Using these new mass and stiffness matrices in “a” space, the equation of motion
becomes

[Ma ] {ẍa } + [Ka ] {xa } = {Fa (t)} (15)

with a corresponding eigensolution given by

[[Ka ] − λ [Ma ]] {xa } = {0} (16)

Depending on the reduction scheme utilized, the eigenvalues of the reduced

system will generally be greater than, or at most equal to, the eigenvalues of the
full system.

2.1.1 Guyan Reduction

For a static system, the equation of motion can be written in partitioned form as
    
[Kaa ] [Kad ] xa Fa
= (17)
[Kda ] [Kdd ] xd Fd

Again the “a” subscript denotes the master or active set of DOF, and the “d”
subscript denotes the embedded or deleted DOF. Assuming that the forces on the
deleted DOF are zero, the equations can be partitioned and manipulated to obtain
the equation written in terms of the active DOF as

[Kaa ] {xa } + [Kad ] [Kdd ]−1 [Kda ] {xa } = {Fa } (18)

Therefore, a relationship is available relating the active DOF to the full set of
DOF as
 
[I]
{xn } = {xa } = [Ts ] {xa } (19)
− [Kdd ]−1 [Kda ]

Using this transformation matrix, the reduced system stiffness matrix can be
written as
 
a = [Ts ] [Kn ] [Ts ]
KG T
(20)

This transformation is exact in the static sense. Guyan [1]/Irons [2] proposed that
the same system transformation matrix used to modify the stiffness matrix be used
to modify the system mass matrix:
864 P. Avitabile and M. Mains

 
a = [Ts ] [Mn ] [Ts ]
MG T
(21)

This transformation attempts to convert the system mass to the set of active
DOFs. However, because this technique is based solely on the static stiffness of the
system, there is no guarantee that the reduced matrix will be accurate for dynamic
applications. The solution of the reduced problem will contain eigenvalues and
eigenvectors that are similar to the eigenvalues and eigenvectors of the full system
model. The degree of similarity is heavily dependent on the selection of the set of
“a” DOF – both the total number of DOF and the distribution of the DOF. In general,
the relative difference increases as the mode number increases with the lower-order
modes generally having less discrepancy than the higher-order modes.

2.1.2 Improved Reduced System (IRS)

As an extension of the Guyan reduction process, the improved reduced system
(IRS) [3] attempts to account for some of the effects of the deleted DOFs that
cause distortion in the Guyan reduction process. The development is based on the
fact that the static structural model containing distributed forces can be condensed
producing a reduced system and solution. The displacements of the reduced system
are then expanded and adjusted for the deleted forces producing an exact statistical
solution of the complete system. A first-order approximation of the eigensystem
is formed using a Guyan/Irons reduced model approach which is based on the
static condensation process with no adjustment for the deleted distributed inertia
forces. The modal vectors of the approximate solution can be adjusted in a similar
fashion as in the static solution to produce an improved set of eigenvectors. Finally,
an estimate of the transformation matrix from full space to reduced space can be
formed for the IRS system. The resulting transformation equation is summarized as

 
[I]
[Ti ] = + [ti ] (22)
[ts ]

where

[ts ] = −[Kdd ]−1 [Kda ]

  (23)
[0]  [0] 
[ti ] = [Mn ] [Ts ] [Ma ]−1 [Ka ]
[0] K−1dd

The IRS technique generally produces a better reduced eigensystem when

compared with the Guyan/Irons approach, because an estimate of the inertia
associated with the deleted DOF is developed as part of the reduction process. The
IRS technique has also been deployed in an iterative fashion to further improve the
reduced matrices.
15 Finite Element Model Correlation 865

2.1.3 Dynamic Condensation

A dynamic implementation of the Guyan reduction process is the dynamic con-
densation [4] process which is often used in correlation studies, particularly for
expansion of mode shapes. A shift value, f, is introduced into the set of equations
describing the dynamic system, thus

[[Kn ] − (λ − f) [Mn ]] {xn } = {0} (24)

and rearrange terms to group the constant shift frequency term f times the mass
matrix with the stiffness matrix to yield

[[[Kn ] + f [Mn ]] − λ [Mn ]] {xn } = {0} (25)

Now let

[Dn ] = [Kn ] + f [Mn ] (26)

Using the same approach in Guyan condensation, these equations can be written
for the active DOF as

[Daa ] {xa } + [Dad ] [Ddd ]−1 [Dda ] {xa } = {Fa } (27)

Therefore, the relationship between the active DOFs and the full set of analytical
DOFs can be written as
 
[I]
{xn } = {xa } = [Tf ] {xa } (28)
− [Ddd ]−1 [Dda ]

So the reduced mass and stiffness matrices can be written as

 
Mfa = [Tf ]T [Mn ] [Tf ] (29)

 
Kfa = [Tf ]T [Kn ] [Tf ] (30)

Due to the formulation of the dynamic condensation process, the eigensolution of

the reduced matrices will result in one eigenvalue which will correspond to the shift
value used for the reduction process. If the shift value happens to correspond exactly
to one of the eigenvalues of the system, then this eigenvalue will be preserved
accurately in the reduced model and will also produce an expanded eigenvector
which will be exactly the same as the corresponding eigenvector from the full finite
element model relating to the shifted eigenvalue. None of the other eigenvalues will
correspond exactly to any of the eigenvalues of the full system.
866 P. Avitabile and M. Mains

2.1.4 System Equivalent Reduction Expansion Process (SEREP)

As done with the other reduction schemes, there is a relationship between the tested
or active “a” DOF and the deleted “d” DOF which can be written in general form as
 
xa
{xn } = = [T] {xa } (31)
xd

The modal transformation in Eq. (5) can be rewritten using this notation as

   
xa Ua
{xn } = = {p} (32)
xd Ud

Note that the modal matrix can also be partitioned into the “a” active and “d”
deleted set of DOF. Looking at just the relationship for the “a” set of DOF, then

{xa } = [Ua ] {p} (33)

The inverse specification of this equation involves a generalized inverse because

the number of unknowns is not equal to the number of equations that need to be
solved. There are two possible solutions to this situation:

• When the number of equations “a” are greater than or equal to the number of
solution variables “m” (an overspecification of the system)
• When the number of equations “a” are less than the number of solution variables
“m” (an underspecification of the system)

Least squares solution

{xa } = [Ua ] {p}

[Ua ]T {xa } = [Ua ]T [Ua ] {p}

 −1  −1
[Ua ]T [Ua ] [Ua ]T {xa } = [Ua ]T [Ua ] [Ua ]T [Ua ] {p}
 −1
{p} = [Ua ]T [Ua ] [Ua ]T {xa } = [Ua ]g {xa }

Average solution
 −1
{p} = [Ua ]T [Ua ]T [Ua ] {xa } = [Ua ]g {xa }

For most structural dynamic applications in dynamic testing, the least squares
solution is used because the number of master DOF (or tested DOF) is far greater
than the number of modes in the system and then the generalized inverse is
15 Finite Element Model Correlation 867

 −1
{p} = [Ua ]T [Ua ] [Ua ]T {xa } = [Ua ]g {xa } (34)

This equation for the modal displacement can be substituted into the modal
transformation equation to give

{xn } = [Un ] [Ua ]g {xa } = [TU ] {xa } (35)

where

[TU ] = [Un ] [Ua ]g (36)

This is the SEREP [5] transformation matrix that is used for either the reduction
of the finite element mass and stiffness matrices or the expansion of the measured
experimental modal vectors. The system equivalent reduction expansion process
(SEREP) relies on a finite element model or analytical model from which an
eigensolution is obtained for developing the mapping between the full set of “n”
finite element DOF and the reduced set of “a” DOF. The eigensolution of the full set
of system matrices yields a set of modal vectors which can be partitioned into those
degrees of freedom that correspond to the active set of “a” DOF and the inactive set
of “d” DOF. This is shown pictorially in Fig. 2. (Note that it is not required to have
a sequential set of modes for the SEREP reduction process.)
Using this SEREP transformation matrix, the reduced mass and stiffness matrices
can then be written as

Un Ua

SELECTED
ADOF

SELECTED DOF

Fig. 2 Schematic of Ua partition of Un

868 P. Avitabile and M. Mains

 
MSa = [TU ]T [Mn ] [TU ]
  (37)
KSa = [TU ]T [Kn ] [TU ]

The equation of motion for the “a” set of DOF can be written as
   
MSa {ẍa } + KSa {xa } = {Fa (t)} (38)

Substituting in the SEREP transformation matrix in Eq. (36) into the reduced
mass matrix in Eq. (37) gives

T
[Ma ] = [Ua ]g [Un ]T [Mn ] [Un ] [Ua ]g (39)

From mass orthogonality in Eq. (10), Eq. (39) can be written for the reduced
mass as
T
[Ma ] = [Ua ]g [Ua ]g (40)

Note that the original system mass matrix is not needed in order to compute
the reduced mass matrix. Similarly for the reduced stiffness matrix, the SEREP
transformation matrix can be substituted into the reduced stiffness matrix in Eq.
(37) to give

T
[Ka ] = [Ua ]g [Un ]T [Kn ] [Un ] [Ua ]g (41)

From stiffness orthogonality in Eqs. (11) and (41) can be written for the reduced
stiffness matrix as
T
 
[Ka ] = [Ua ]g 2 [Ua ]g (42)

Note that the original system stiffness matrix is not needed in order to compute
the reduced stiffness matrix.
While the size of these reduced mass and stiffness matrices is “a” by “a,” the rank
of the reduced matrices is only “m.” Therefore, use of these matrices must be done
with caution. Due to this rank deficiency, an alternate form of the SEREP reduction
process which invokes an exact solution can be obtained by using a = m for the
reduction.

2.1.5 Modal TAM

The modal TAM is very similar to SEREP, except that the a-set modes are
unmodified (not smoothed). This is accomplished by realizing there is a direct rep-
resentation of {xa } like we see in Guyan reduction, IRS, and dynamic condensation:
15 Finite Element Model Correlation 869

{xa } = [I] {xa } (43)

The next observation is to look at the lower partition of Eq. (32) given as

{xd } = [Ud ] {p} (43)

and substituting Eq. (34) into Eq. (43) providing the relationship between the a-set
and d-set DOFs as

{xd } = [Ud ] [Ua ]g {xa } (44)

Referring back to Eq. (32), the modal TAM is written as

   
xa [I]
{xn } = = {xa } = [TM ] {xa } (45)
xd [Ud ] [Ua ]g

where
 
[I]
[TM ] = (46)
[Ud ] [Ua ]g

When using this transform to reduce mass and stiffness matrices, “m” is typically
less than “a”, resulting in rank deficient matrices, but retaining the full accuracy of
the original “m” FEM modes in frequency and full mode shape accuracy. When this
transform is used for test shape expansion the modes at the a-set are unmodified.
As the number of modes are increased, and the a-set is well chosen, then SEREP
and Modal TAM become equivalent. This occurs when and if

[Ua ] [Ua ]g = [I] (47)

2.1.6 Hybrid
Another method for reducing of the system matrices utilizes the exactness of the
SEREP or modal TAM process and overcomes the rank deficiency by incorporating
the effects of Guyan condensation into the process and is referred to as the hybrid
reduction [6] technique and is formulated as

[TH ] = [Ts ] + [[TM ] − [TS ]] [P] (48)

The reduced matrices of the hybrid TAM are not rank deficient with the
eigenvalue decomposition giving the mode shapes and natural frequencies that the
modal TAM would provide for the m target modes included in the aforementioned
modal TAM plus the additional information that the Guyan reduction would provide
for the remaining model order (a-set minus the target modes (a-m)). The projector
matrix [P] is given in both its oblique and orthogonal forms as
870 P. Avitabile and M. Mains

Table 1 Summary of hybrid Hybrid Projector TAM combinations

TAM combinations
1 Oblique Modal Guyan
2 Orthogonal Modal Guyan
3 Oblique SEREP Guyan
4 Orthogonal SEREP Guyan
5 Oblique Modal IRS
6 Orthogonal Modal IRS
7 Oblique SEREP IRS
8 Orthogonal SEREP IRS


Poblique = [Ua ] [Ua ]T [TM ]T [M] [TM ] (49)

[Portho ] = [Ua ] [Ua ]g (50)

Observation shows that in addition to the oblique and orthogonal forms of the
hybrid TAM, SEREP can be substituted for the modal TAM, and IRS can be used
instead of the Guyan TAM. Table 1 summarizes the eight different hybrid TAM
combinations that are possible.

[TH ] = [Ts ] + [[TM ] − [TS ]] [P]

[TH ] = [Ts ] + [[TU ] − [TS ]] [P]

[TH ] = [TI ] + [[TM ] − [TI ]] [P]

[TH ] = [TI ] + [[TU ] − [TI ]] [P]

2.2 Model Expansion (Vector Expansion)

Experimental mode shapes only exist at the DOF associated with the test points
(“a” DOF). Because the mass and stiffness matrices are described at the full set of
finite element DOFs (“n” DOF), the system mass and stiffness matrices need to be
reduced to the set of experimental DOF for correlation studies. However, there is
also a need to expand the measured experimental mode shape over the full set of
finite element DOF for further correlation studies. Therefore, expansion techniques
are necessary for other studies. A schematic of the expansion process is shown in
Fig. 3.
15 Finite Element Model Correlation 871

TEST ‘A’ DOF FEA ‘N’ DOF

Expansion

Fig. 3 Schematic of the expansion process

Early expansion techniques evolved around using spline fits and polynomial
expansion based on geometry and measured data. While in concept they are useful,
in practice, using these approaches for general structural systems is not feasible.
Most expansion techniques utilized today involve the use of the finite element model
as a mechanism to complete the unmeasured DOF from the experimental modal
model. In essence, the finite element model is used as a high-order polynomial curve
fitter to estimate the experimental mode shapes at the deleted DOF. The majority
of the expansion techniques use the model reduction transformation matrix as an
expansion mechanism.
Recall that the basic relationship relating the “a” DOF to the “n” DOF is given
by
 
xa
{xn } = = [T] {xa } (51)
xd

Using this expansion concept along with measured experimental modal data, then
 
Ea
[En ] = = [T] [Ea ] (52)
Ed

The measured experimental modal vectors at “a” DOF are expanded over all the
finite element “n” DOF using the transformation matrix [T]. This transformation
matrix will take on various forms depending on which technique is utilized.
At times, the expansion is broken down into two parts, namely, completion and
smoothing. The completion of the mode shape refers to unmeasured DOFs which
are obtained in the expansion process. The smoothing of the mode shape refers to
the possible least squares fitting of the measured DOF. There are differences of
opinion as to whether or not smoothing should be performed, and there is no clear
answer on this. On one hand, the data measured is what was measured, and many
872 P. Avitabile and M. Mains

feel that this should not be manipulated in the expansion process. On the other hand,
the measured data is clearly never perfect and does have variance which is why
smoothing might be performed. One additional concern when using unsmoothed
data is that there is a mathematical inequity between the measured data and the
completed data that can have an effect on the correlation results.
Each of the reduction techniques discussed above is also useful for expansion.

2.2.1 Guyan Expansion

The Guyan expansion technique uses the static condensation transformation matrix
to expand the measured DOF over all the finite element DOF. The expansion is given
by

   
Ea [I]
[En ] = = [Ts ] [Ea ] = [Ea ] (53)
En − [Kdd ]−1 [Kda ]

Of course, the Guyan condensation process will not produce acceptable results
unless there are sufficient DOF to describe the mass of the system (as previously
discussed in the reduction section). If sufficient DOFs are available, then the Guyan
process will produce reasonably good results but will never produce exact results
because the inherent formulation of the reduction matrix is approximate. The Guyan
reduction process is still widely used for model reduction applications due to its long
historical background but is not widely used for expansion of mode shapes because
other more accurate techniques have been developed.

2.2.2 Improved Reduced System (IRS)

The IRS expansion technique uses the static condensation transformation matrix
along with adjustment terms to compensate for the inertia associated with the
deleted DOF to expand the measured DOF over all the finite element DOF. The
expansion is given by

 
Ea
[En ] = = [Ti ] [Ea ]
Ed
     (54)
[I] 0 0 −1
= + [Mn ] [Ts ] [Ma ] [Ka ] [Ea ]
− [Kdd ]−1 [Kda ] 0 K−1
dd

Of course, the IRS technique will improve on the Guyan expansion process, but
will not produce acceptable results unless there are sufficient DOFs to describe
the mass of the system (as previously discussed in the reduction section 2.1). If
sufficient DOFs are available, then the IRS process will produce reasonably good
results (which are improved over the Guyan results) but will never produce exact
results because the inherent formulation of the reduction matrices is approximate.
15 Finite Element Model Correlation 873

2.2.3 Dynamic Expansion

The dynamic expansion technique is very similar to the static expansion process
except that the stiffness matrix is modified to include the effects of the mass of
the system at a particular frequency. This is accomplished by adding an adjustment
term of the reference frequency times the system mass to the stiffness matrix as was
shown in the development of the model reduction equations. In essence, this matrix
is exact for this one particular frequency, and the transformation matrix will be exact
in regard to expanding a mode shape at that particular frequency. Of course, the shift
frequency must correspond to one of the eigenvalues of the system. The expansion
is given by
   
Ea [I]
[En ] = = [Tf ] [Ea ] = [Ea ] (55)
Ed − [Ddd ]−1 [Dda ]

The dynamic expansion process will produce exact results for one frequency and
only one frequency. Providing that the shift frequency corresponds exactly to one of
the eigenvalues of the system, then the expansion will produce an exact mode shape
for this one eigenvalue. If additional eigenvectors need to be expanded, then separate
shift values need to be processed. While many matrices need to be processed for
each eigenvector that needs to be expanded, the exactness of the process warrants
the additional processing.

2.2.4 System Equivalent Reduction Expansion Process (SEREP)

The SEREP expansion technique uses the SEREP transformation matrix to expand
the measured DOF over all the finite element DOF. The expansion is given by
  

Ea Ua
[En ] = = [TU ] [Ea ] = [Un ] [Ua ] [Ea ] =
g
[Ua ]g [Ea ] (56)
Ed Ud

Notice that the “a” DOF may be changed as seen by the upper partition of this
equation:

[Ea ]’ = [Ua ] [Ua ]g [Ea ] (57)

and that the deleted DOFs are estimated by

[Ed ] = [Ud ] [Ua ]g [Ea ] (58)

When the “a” DOFs are expanded, there is the possibility that the initially
measured DOF may be modified by the expansion process. This is referred to as
smoothing of the measured DOF. This occurs because the SEREP process is based
on a generalized inverse using a least squares error minimization. Therefore, the
measured data is smoothed as part of the process. While much controversy exists
over whether or not to smooth the actual measured data, this is the proper way
874 P. Avitabile and M. Mains

to process the data, from a mathematical standpoint. Otherwise, a mathematical

mismatch exists in the expanded vector between the “a” DOF and the “d” DOF.

2.2.5 Modal
The modal TAM expansion technique uses the modal TAM transformation matrix to
expand the measured DOF over all the finite element DOF. The expansion is given
by

   
Ea [I]
[En ] = = [TM ] [Ea ] = [Ea ] (59)
Ed [Ud ] [Ua ]g

Notice that the “a” DOFs are unchanged as seen by the upper partition of this
equation and the deleted DOFs are estimated in the same manner as the SEREP
process.

2.2.6 Hybrid
The Hybrid TAM expansion technique uses a combination of the SEREP or modal
TAM and Guyan or IRS transformation matrix to expand the measured DOF over
all the finite element DOF. The expansion is given by

 
Ea
[En ] = = [TH ] [Ea ] = [[Ts ] + [[TM ] − [TS ]] [P]] [Ea ] (60)
Ed

2.2.7 Model Reduction Considerations for Sensor Locations

The transformation matrices need to be understood when selecting DOF for
measurement locations. This is an important consideration because there may be
a limited budget for measurements on the structure to be made or portions of the
structure, which are dynamically active, may not be instrumented. Typically, the
transformation matrices are not plotted or reviewed, and there is useful information
to be seen in these matrices.
In order to illustrate some of the typical transformation matrix characteristics,
two simple academic structures are shown. The first structure has a very well
distributed set of “a” DOF for a two-beam system where there is modal energy in
each of the beams but distributed differently for each system model (as is typically
seen in any practical structure). For this structure, there are four DOFs on the upper
beam and eight DOFs on the lower beam which is a proper distribution for the
modes of this structure. The mode shapes are shown in Fig. 4 for reference for
the first several modes. The Guyan, IRS, and SEREP transformation matrices are
plotted in a 3D plot in Fig. 5. Notice that there is very little difference between the
different transformation matrices and that IRS slightly adjusts Guyan and SEREP
provides the most accurate representation for the matrix. Note that all of the system
frequencies and mode shapes are represented well with SEREP producing the exact
frequencies and mode shapes for the system; Guyan provides good estimates of the
15 Finite Element Model Correlation 875

Fig. 4 Schematic showing mode shapes for two-beam system

frequencies and mode shapes and IRS improves on Guyan, but neither has the exact
representation of the full system captured in the reduced model.
Now in the second case, the DOFs are only distributed on the lower beam which
contains most of the modal information, but no DOFs are distributed on the upper
beam which only has a few dynamically important modes for the system. The
DOF may not be distributed to the upper beam because it may be inaccessible for
instrumentation, it may not be considered as one of the important target modes of
the system, and there may be a lack of instrumentation budget for enough sensors,
or a number of other reasons. But the important item is that the portion of the
structure is not included in the reduction process; essentially an inadequate number
of transducer measurement locations have been identified. The Guyan, IRS, and
SEREP transformation matrices are plotted in a 3D plot in Fig. 6. Notice that the
SEREP transformation matrix is substantially different than the Guyan and IRS
matrices. Remember that the SEREP process is exact and the transformation matrix
876 P. Avitabile and M. Mains

Fig. 5 Transformation matrix [T] shown for Guyan, IRS, and SEREP for a well-selected set of
DOF showing similarity in the project matrix

for SEREP accurately illustrates the true transformation needed to accommodate

the poor selection of DOF for this reduced model. Notice that the Guyan and
IRS transformation matrices appear similar indicating that IRS can only make
useful adjustments to the Guyan model based on the starting DOF used for Guyan
reduction. Both the Guyan and IRS transformation matrices are not adequate to map
the system accurately with the DOF selected.

2.3 Test Data Considerations

There must be serious consideration given to the data that is measured and to be used
for the correlation process. All too often this is overlooked by the finite element
or correlation engineer performing the analysis with the assumption that the test
engineer has done an extensive evaluation of the data collected; this should have
been done but may not have been evaluated with the degree and depth needed when
performing correlation. Certainly, this is a critical step in the process, but only a few
of the more important considerations can be provided here.
In actuality, a detailed pretest analysis really needs to be performed prior to
actually starting the data collection. Issues related to target mode selection and
sensor placement are of critical concern. For target mode selection, many tools
15 Finite Element Model Correlation 877

Fig. 6 Transformation matrix [T] shown for Guyan, IRS, and SEREP for a poorly selected set of
DOF showing dissimilarity in the project matrix

such as effective modal mass, kinetic energy fraction, and mode participation need
to be evaluated for the test. As for sensor placement, many tools such as modal
displacement method, nodal kinetic energy, and sensor identification (effective
independence and MAC contribution) are commonly used.
But from the practical side, there are many test data issues that are often missed,
overlooked, or not clearly understood by the test engineer or the analyst performing
the correlation. There are always the considerations for time invariance, linearity,
reciprocity, repeatability, and measurement adequacy to name a few that can distort
measured data as well as quantization, frequency resolution, leakage, and windows.
Table 2 lists a few other general concerns for measured data.
Beyond the measurement aspects, the excitation utilized can also have an effect
on the measurement adequacy. Techniques such as burst random, pseudorandom,
sine chirp, and digital stepped sine to name a few of the more popular techniques
commonly used are very important to the success of the data extracted.
The actual setup of the test can have an impact on the results that are extracted
from the test data collected. Issues related to simple items such as the boundary
conditions used for the test setup can have an effect on the modes extracted. It is
imperative that the finite element model be constructed in a fashion that attempts to
replicate the actual test setup conditions so as to maintain a proper representation of
the boundary conditions as much as possible. All too often the finite element model
878 P. Avitabile and M. Mains

Table 2 Instrumentation Test instrumentation – amplifiers/filters

considerations • Calibration
• Magnitude and phase distortion
• Magnitude and phase mismatch
Test instrumentation – transducer mounting
• Accuracy
• Repeatability
• Local effects
Test instrumentation – transducers
• Calibration
• Magnitude and phase distortion
• Mass loading effects/stiffening effects
• Sensitivity to environment
• Cross axis sensitivity

is constructed with either a free-free or fully built-in condition which is impossible

to achieve in any test that is performed.
Often times there are never enough measurement transducers to provide the
proper amount of tested DOF to assure a good representation of the modes of
the system. In larger tests, there may need to be a multiple-input multiple-output
data collection performed. When too few shakers (or only one shaker) are used, the
data collection must be repeated several times with subsets of the total number of
multiple shakers deployed. This measurement scenario is usually not acceptable,
and data inconsistencies may exist due to test setup changes and the time required
to collect the entire data set when the test environment cannot be guaranteed to
remain constant. Too few accelerometers can result in missed (or poorly understood)
modes and may be inappropriate for the reduced model comparison using Guyan
and IRS techniques. Many times more effective measurement points are attempted
by roving accelerometers around the structure, but this can potentially cause mass
loading effects which can completely distort the experimental data extracted. Even
when using dummy masses at all measurement locations, the time required to remap
channels often results in a staggering in time when the data is collected and the test
environment may not be stable enough to accommodate these minor changes. The
most critical item is that the test data needs to be collected simultaneously in a
consistent fashion so that the data reduction requirements are not violated; often
this is unknown to the analyst when receiving data.
But one prevailing statement can be made which is not often understood by so
many that are involved in these test-analysis correlation studies.

All the measured degrees of freedom are not measured with the same
accuracy for each mode of the system. This means that for a correlation
study, a certain set of degrees of freedom may be measured very accurately
for one particular mode and be measured very inaccurately for another
mode. This raises the question as to whether to use all the degrees of

(continued)
15 Finite Element Model Correlation 879

freedom when correlating each mode of the system and is a critically

important item to understand when undertaking a test-analysis correlation
study.

2.4 Vector Correlation

The tools available for correlation can generally be broken down into global
assessment (vector based) or local assessment (DOF based). Each of the techniques
is identified below. Each generally has some metric, but the actual value of
acceptability is not an exact value. The specific level of acceptance is very dependent
on the design specification and intent of the correlation study. However, in some
industries, there are specific contractual values that must be achieved, but this does
not necessarily translate to all industries and applications. But often times due to the
lack of understanding of what specific target values are needed, many often revert to
use of some of these industry-mandated values, but there is no specific requirement
or need to do so. The level of correlation needed should really be dictated by the
design specification and overall system level performance necessary to achieve the
stated goals or design objectives.

2.4.1 Modal Assurance Criteria (MAC)

The modal assurance criteria [7] (MAC) is an extremely useful technique which
gives a first indication of the level of correlation that exists between the analytical
and experimental modal vectors. The MAC, for real valued vectors, is given by

   H
{ui }H ej ej {ui }
MACij =  H   (61)
{ui }H {ui } ej ej

While the original MAC is written using complex notation, the MAC is generally
used with real normal vectors when utilized in finite element correlations.
In this formulation, the values of MAC range between “0.0” and “1.0,” where
“0.0” indicates that there is little or no correlation between the vectors and “1.0”
indicates that there is a high degree of similarity between the modal vectors. MAC is
ideal for identifying those analytical modes corresponding to experimental modes,
and it is very useful when identifying mode switching. MAC is very sensitive to
the DOFs that are largest in value and is very insensitive to very small DOF in the
mode shape vector. Mass weighting is not used in this formulation, which has the
advantage that mass reduction is not needed. A typical schematic for a MAC matrix
is shown in Fig. 7.
880 P. Avitabile and M. Mains

Fig. 7 Typical modal assurance criteria matrix

2.4.2 Orthogonality Checks

Two orthogonality checks [8] are often made when evaluating vector correlation –
the cross orthogonality check and the pseudo orthogonality check. The obvious
hurdle to overcome is whether to reduce the finite element mass and stiffness
matrices to the set of tested DOF or to expand the measured experimental modal
vectors to the full space of the finite element model. Alternately, some combination
of both reduction and expansion, in order to compute the orthogonality, could be
used. A schematic of the reduction and expansion process for correlation is shown
in Fig. 8 along with an orthogonality matrix.
The cross orthogonality check is an orthogonality check where the modal vector
matrices are obtained from the experimentally measured modal data:

?
CROSS = [E]T [M] [E] = [I] (62)

The pseudo orthogonality check is essentially an orthogonality check where one

of the modal vector matrices is replaced with the experimentally measured modal
data:

?
POC = [U]T [M] [E] = [I] (63)
15 Finite Element Model Correlation 881

Fig. 8 Overall schematic of the model reduction/expansion process for correlation

882 P. Avitabile and M. Mains

Fig. 9 Schematic of
expansion for orthogonality
checks

EXPANSION

TEST
ADOF

EXPANDED
NDOF

REDUCTION

NDOF
BANDED REDUCED
FEA ADOF
MATRIX

Fig. 10 Schematic of reduction for orthogonality checks

As mentioned above, in order to accomplish this triple product, the matrices must
conform. Therefore, either the system mass matrix must be reduced to the set of
tested DOF (and corresponding “a” DOF from the test model), or the experimentally
measured modal vectors must be expanded to the full space of the finite element
model. This is schematically shown in Figs. 9 and 10. Of course, the results of either
of these checks will be dependent on the type of reduction, or expansion, utilized,
except for the SEREP process which preserves the dynamics of the system in the
reduced model.
In reviewing the results of the cross orthogonality check and the pseudo
orthogonality check, there are similarities that exist because one check basically
uses the experimentally measured vectors twice in the computation of orthogonality,
15 Finite Element Model Correlation 883

and, therefore, the resulting terms are larger. However, there is no benefit of using
these vectors twice in the cross orthogonality check, and no additional insight is
gained in this process. Thus, there is no benefit in using the cross orthogonality
check rather than the pseudo orthogonality check for general orthogonality using the
usual model reduction and expansion processes such as Guyan and IRS. However,
when using the SEREP process, there are significant computational benefits to be
gained when using the pseudo orthogonality check.
In all of the reduction/expansion techniques, there is some numerical processing
necessary to either reduce the mass matrix or expand the experimental modal
vectors. Due to its formulation, the SEREP process has some very important
characteristics. The POC at the set of tested “a” DOF is exactly equal to the POC
at the full set of “n” DOF. Also, the POC can be performed without the use of any
system matrices:

[Ua ]T [Ma ] [Ea ] = [Un ]T [Mn ] [En ] = [Ua ]g [Ea ] (64)

First, the pseudo orthogonality check will provide exactly the same results
whether the check is done at the full space of the finite element model or at the
reduced space of the test model. Therefore, the computation is most efficiently per-
formed in the reduced space of the test model. Second, tremendous computational
and procedural benefits are obtained for the pseudo orthogonality check using the
SEREP process because the mass matrix is not needed for this computation [9].

2.4.3 Coordinate Modal Assurance Criteria (CoMAC)

The coordinate modal assurance criteria [10] follows the same formulation as
MAC in that a correlation coefficient is developed to determine the degree of
correspondence that exists for a particular DOF over a set of correlated mode pairs.
CoMAC is useful in determining how well correlated each individual DOF may be
over a set of modes and provides some insight into where discrepancies may exist:

 m  2
  (c) (c) 
uk · ek 
c=1
CoMAC (k) = m   m   (65)
 (c) 2  (c) 2
uk · ek
c=1 c=1

However, without mass scaling to properly weight the DOFs, at times it is

difficult to determine the degree of correlation that exists. Another drawback of
the CoMAC is that it can only be used for correlated mode pairs. This implies
that only the diagonal related terms of the MAC correlation matrix can be assessed
(once the vectors are arranged in proper correlated order if necessary). The CoMAC
is shown schematically in Fig. 11. Of course it is important to mention that there
884 P. Avitabile and M. Mains

CoMAC
COORDINATE
MODAL
ASSURANCE
Experimental Analytical
CRITERIA

DOF CORRELATION

FINITE ELEMENT EXPERIMENTAL

Fig. 11 Schematic of CoMAC for vector correlation

are variations of this CoMAC formulation such as the enhanced CoMAC [11] and
modulus difference.

2.4.4 Frequency Response Assurance Criteria (FRAC)

The frequency response assurance criteria evaluates each DOF based on the FRF
comparison of the analytical and experimentally derived functions. The formulation
is very similar to the MAC function and has a similar interpretation. The FRAC is
given by
 2
{htest } {hfem (β)}∗
FRAC (β) =    (66)
{htest } {htest }∗ {hfem (β)} {hfem (β)}∗

The FRAC is a useful tool for evaluating FRFs. However, the main drawback
is that the analytical model FRF may have similar shape characteristics but differ
slightly in frequency which can cause significantly low FRAC values. Therefore,
the function is constructed with a shifting function to allow for some frequency
adjustment due to global stiffness differences. The FRAC is mainly used for corre-
lation in frequency response-based model updating studies. One critical concern is
the damping selected for the development of the finite element FRF which will have
an effect on the results. The FRAC is shown schematically in Fig. 12.

2.4.5 Response Vector Assurance Criteria (RVAC)

A companion to the FRAC is the response vector assurance criteria (RVAC) which
compares a specific spectral line of the FRF for the analytical and experimental
15 Finite Element Model Correlation 885

FRAC
FREQUENCY
RESPONSE
ASSURANCE
CRITERIA

FINITE ELEMENT EXPERIMENTAL

DOF CORRELATION

Fig. 12 Schematic of FRAC for vector correlation

FRF for a number of measurement points. Essentially, RVAC is a MAC correlation

technique for the analytical and experimental vectors (approximated using a peak
pick technique) and is given as

RVAC (ω) = MAC ({Etest (ω)} , {Ufem (ωβ)}) (67)

The RVAC is shown schematically in Fig. 13.

2.4.6 Test Response Assurance Criteria (TRAC)

Another correlation sometimes used is the test response assurance criteria (TRAC)
which is another variant of MAC that is used to compare time responses from a finite
element model and measured time response data. It follows the same interpretation
as MAC and is given as

 2
{Xn1 }T {Xn2 }
TRAC =   (68)
{Xn1 }T {Xn1 } {Xn2 }T {Xn2 }

The TRAC is schematically shown in Fig. 14a along with a MAC to illustrate
the two tools often used in correlation studies; Fig. 14b shows the two traces plotted
886 P. Avitabile and M. Mains

RVAC
RESPONSE
VECTOR
ASSURANCE
CRITERIA

FINITE ELEMENT EXPERIMENTAL

VECTOR CORRELATION

Fig. 13 Schematic of RVAC for vector correlation

against each other and illustrates that the TRAC (and MAC) are really nothing more
than the R2 value from a regression analysis.

2.4.7 CORTHOG
The coordinate orthogonality check [12] (CORTHOG) is a mass scaled degree
of freedom correlation that is conceptually similar to CoMAC but can be used
for checking any mode pair and not just correlated mode pairs. But the most
important feature for CORTHOG is the fact that mass scaling is included in the
formulation.
The simplest statement summarizing the coordinate orthogonality check is as fol-
lows: The coordinate orthogonality check is simply the comparison of what should
15 Finite Element Model Correlation 887

Fig. 14 (a, b) Schematic of the TRAC and MAC correlation tools

888 P. Avitabile and M. Mains

CORTHOG
COORDINATE
ORTHOGONALITY
CRITERIA Experimental Analytical

DOF CORRELATION

FINITE ELEMENT EXPERIMENTAL

Fig. 15 Schematic of the coordinate orthogonality check

have been obtained analytically for each degree of freedom in an orthogonality

check to what was actually obtained for each degree of freedom in a POC from test.
While many different formulations exist, the most basic form is the simple
difference given as


m
CORTHOGkij = eki mkl ulj − uki mkl ulj (69)
l=1

The CORTHOG is schematically shown in Fig. 15.

If the SEREP process is used for the formulation, then the CORTHOG can
be very easily computed without the system mass matrix. Remember that the
orthogonality can be written as

ORT = [Un ]T [Mn ] [Un ] = [Ua ]T [Ma ] [Ua ] = [Ua ]g [Ua ] (70)

and the pseudo orthogonality check can be written as

POC = [Un ]T [Mn ] [En ] = [Ua ]T [Ma ] [Ea ] = [Ua ]g [Ea ] (71)

Therefore, for any particular mode pair, the CORTHOG can be written as shown
schematically in Fig. 16.
15 Finite Element Model Correlation 889

POCij = u ig1e1 j + u gi 2 e 2 j + u ig3e 3 j + + u gia e aj

CORTHOG ij
ORTij = u ig1u1 j + u gi 2 u 2 j + uig3 u 3 j + + u gia u aj

dof 1 dof 2 dof 3

Fig. 16 Schematic of efficient coordinate orthogonality check computation

3 Closing Remarks

Any or all of these correlation tools can be used to help identify the similarities
or differences that exist between the finite element model and the measured data.
While in some industries (i.e., aerospace) there may be specific target values
to achieve to meet contractual requirements, there really is no hard and fast
rule as to acceptable levels of correlation. This really should be customized
for particular applications based on design specification needs for a particular
design and intended function (in the same way that tolerancing on a mechanical
part will vary depending on how the part mates to other parts where some
tolerances need to be much tighter than others depending on the design). It
is also very important to note that some tools may provide better informa-
tion for certain applications and other tools may be more useful for different
applications.
For reference, some typical MAC and orthogonality are shown for a few
academic structures to illustrate the use of these tools as seen in Fig. 17.
890 P. Avitabile and M. Mains

Fig. 17 Correlation of FEA model and test data along with MAC matrix and orthogonality check
for two components (top and middle) used to form an assembly (bottom)
15 Finite Element Model Correlation 891

References
1. Guyan, R.J., “Reduction of Stiffness and Mass Matrices”, AIAA Journal, Vol. 3, No 2, 1965,
p. 380
2. Irons, B, “Structural Eigenvalue Problems: Elimination of Unwanted Variables”, AIAA
Journal, Vol 3, No 5, 1965„ pp. 961–962
3. O’Callahan,J.C., “A Procedure for an Improved Reduced System (IRS)”, Seventh International
Modal Analysis Conference, Las Vegas, Nevada, February 1989, pp. 17–21
4. Paz, M, “Dynamic Condensation”, AIAA Journal, Vol22, No 5, May 1984, pp. 724–727
5. O’Callahan,J.C., Avitabile, P., Riemer,R., “System Equivalent Reduction Expansion Process”,
Seventh International Modal Analysis Conference, Las Vegas, Nevada, February 1989,
pp. 29–37
6. Kammer,D.C., “A Hybrid Approach to Test Analysis Model Development for Large Space
Structures”, Journal of Acoustics and Vibrations,Vol 113, No 3, July 1991, pp. 325–332
7. Allemang,R.J. and Brown,D.L., “A Correlation Coefficient for Modal Vector Analysis”, First
International Modal Analysis Conference, Orlando, Florida, November 1982, pp. 110–116
8. Avitabile, P., O’Callahan,J.C., Milani,J. “Model Correlation and Orthogonality Criteria”, Sixth
International Modal Analysis Conference, Orlando, Florida, February 1988, pp. 1039–1047
9. Avitabile, P., O’Callahan,J., “Mass and Stiffness Orthogonality Checks without a Mass or
Stiffness Matrix”, Thirteenth International Modal Analysis Conference, Nashville, Tenn,
February 1995, pp. 1515–1519
10. Lieven,N.A.J., Ewins,D.J., “Spatial Correlation of Mode Shapes, the Coordinate Modal
Assurance Criteria (COMAC)”, Sixth International Modal Analysis Conference, Orlando,
Florida, February 1988, pp. 690–695
11. Hunt,D., “Application of an Enhanced Coordinate Modal Assurance Criterion”, Tenth Interna-
tional Modal Analysis Conference, San Diego, California, February, 1992, pp. 66–71
12. Pechinsky,F., Avitabile, P., “Evaluation of DOF Correlation using the Coordinate Orthogonality
Check”, Thirteenth International Modal Analysis Conference, Nashville, Tennessee, February
1995, pp. 1313–1320

Bibliography
13. Allemang,R.J. and Brown,D.L., “A Correlation Coefficient for Modal Vector Analysis”, First
International Modal Analysis Conference, Orlando, Florida, November 1982, pp. 110–116
14. Avitabile,P., O’Callahan,J.C., Milani,J. “Model Correlation and Orthogonality Criteria”, Sixth
International Modal Analysis Conference, Orlando, Florida, February 1988, pp. 1039–1047
15. Avitabile,P., Li,P., Foster,T., Effects of DOF Selection on Orthogonality using Reduced System
Matrices”, Eleventh International Modal Analysis Conference, Orlando, Florida, February
1992, pp. 1073–1078
16. Avitabile,P., Pechinsky,F., O’Callahan, J.C., “Study of Modal Vector Correlation using Various
Techniques for Model Reduction”, Tenth International Modal Analysis Conference, San Diego,
California, February, 1992, pp. 572–583
17. Avitabile,P., O’Callahan,J., “Mass and Stiffness Orthogonality Checks without a Mass or
Stiffness Matrix”, Thirteenth International Modal Analysis Conference, Nashville, Tenn,
February 1995, pp. 1515–1519
18. Avitabile, P., Modal Model Correlation, University of Massachusetts Lowell, Copyright TX
4764-126 (1998 Dissertation)
19. Baker, M, “Review of Test/Analysis Correlation Methods and Criteria for Validation of Finite
Element Models for Dynamic Analysis”, Tenth International Modal Analysis Conference, San
Diego, February 1992, pp. 984–990
20. Blades, EL, Craig, RR, “A Craig-Bampton Test-Analysis Model”, Fifteenth International
Modal Analysis Conference, Orlando, Florida, February 1997, pp. 1386–1391
892 P. Avitabile and M. Mains

21. Blaschke,PG, Bosch,R, Ewins, D, “The MAC Revisited and Updated”, Fifteenth International
Modal Analysis Conference, Orlando, Florida, February 1997147–154
22. Blelloch, PA, “Direct Correlation of Test-Analysis Cross-Orthogonality”, Tenth International
Modal Analysis Conference, San Diego, February 1992, pp. 998–1002
23. Blelloch,P., “Shape- and Matrix Based Sensor Selection Algorithms”, Twenty-First Interna-
tional Modal Analysis Conference, Orlando, FL, February 2003
24. Brillhart, RD, Kammer, DC, “Evaluation of Optimal Modal Sensor Placement Using a System
Realization Method”, Twelfth International Modal Analysis Conference, Honolulu, Hawaii,
February 1994, pp. 1401–1407
25. Carne,T.G., Dohrmann,C.R., “A Modal Test Design Strategy for Model Correlation, Thir-
teenth International Modal Analysis Conference, Nashville, Tennessee, February 1995,
pp. 927–933
26. Chung, Y.T., Sernaker M.L., “Assessment of Target Mode Selection Criteria for Payload Modal
Survey”, 12th IMAC, 1994, Honolulu, Hawaii, pp. 272–279.
27. DeClerck,J.P., Avitabile,P., Development of Several New Tools for Pre-Test Modal Evalua-
tion”, Fourteenth International Modal Analysis Conference, Dearborn, Mich-igan, February
1996, pp. 1272–1277
28. Flanigan, CC, Hunt, DL, “Integration of Pretest Analysis and Model Correlation Methods
for Modal Surveys”, Eleventh International Modal Analysis Conference, Orlando, Florida,
February 1993, pp. 444–448
29. Flanigan C. C. and Hunt D. L., 1993, “Integration of Pretest Analysis and Model Correlation
Methods for Modal Surveys”, Eleventh International Modal Analysis Conference, Orlando,
FL, February 1993, pp. 444–448.
30. Flanigan, CC, Freymiller, JE, “Using Dynamic Expansion to Improve Test-Analysis Correla-
tion”, Twelfth International Modal Analysis Conference, Honolulu, Hawaii, February 1994,
pp. 198–204
31. Foster,T., “Evaluation of DOF Selection for Modal Vector Correlation and Expansion Studies”,
Masters Thesis, University of Massachusetts Lowell, Lowell, Massachusetts, July 1993
32. Freed,A.M., Flannigan,C.C., “A Comparison of Test Analysis Model Reduction Methods”,
Eighth International Modal Analysis Conference, Orlando, Florida, February 1990, pp. 1344–
1351
33. Friswell, MI, Garvey, SD, Penny, JET, “Model Reduction Using an Iterated IRS Technique”,
5th International Conference on Recent Advances in Structural Dynamics, Southhampton, UK,
July 1994, pp. 879–889
34. Goh, EL, Mottershead, JE, “On the Model Reduction Techniques for Finite Element Model
Updating”, NAFEMS/DTA International Conference on Structural Dynamics, Modeling, Test,
Analysis and Correlation, Milton Keynes, July 1993, pp. 421–432
35. Gordis, J, “An Analysis of the Improved Reduced System (IRS) Model Reduction Procedure”,
Tenth International Modal Analysis Conference, San Diego, February 1992, pp. 471–479
36. Guyan, R.J., “Reduction of Stiffness and Mass Matrices”, AIAA Journal, Vol. 3, No 2, 1965,
p. 380
37. Gysin, HP, “Comparison of Expansion Method for FE Modeling Error Localization”, Eighth
International Modal Analysis Conference, Orlando, Florida, February 1990, pp. 195–204
38. Gysin, HP, “Expansion: The Achilles Heel of Fe Model Updating”, Proceedings of the 15th
International Seminar on Modal Analysis, September 1990, pp. 919–937
39. Hemez, FM, Farhat, C, “Comparing, Mode Shape Expansion Methods for Test-Analysis
Correlation”, Twelfth International Modal Analysis Conference, Honolulu, Hawaii, February
1994, pp. 1560–1567
40. Humbad, NG, Mallick, S, “Modal Comparison Criteria”, Fifth International Modal Analysis
Conference, London, England, February 1987, pp. 1191–1197
41. Hunt, DL, Peterson, EL, Vold, H, Williams, R, “Optimal Selection of Excitation Methods for
Enhanced Modal Testing”, AIAA paper, No. 84–1068, AIAA Dynamics Conference, Palm
Springs, California, May 1984, pp. 549–553
42. Hunt,D., “Application of an Enhanced Coordinate Modal Assurance Criterion”, Tenth Interna-
tional Modal Analysis Conference, San Diego, California, February, 1992, pp. 66–71
15 Finite Element Model Correlation 893

43. Imamovic, N, Ewins, D, “Optimization of Excitation DOF Selection for Modal Tests”,
Fifteenth International Modal Analysis Conference, Orlando, Florida, February 1997, pp.
1945–1951
44. Imregum, M, Ewins, D, “An Investigation into Mode Shape Expansion Techniques”, Eleventh
International Modal Analysis Conference, Orlando, Florida, February 1993, pp. 168–175
45. Irons, B, “Structural Eigenvalue Problems: Elimination of Unwanted Variables”, AIAA
Journal, Vol 3, No 5, 1965„ pp. 961–962
46. Irons, B, “Structural Eigenvalue Problems: Elimination of Unwanted Variables”, AIAA
Journal, Vol 3, No 5, 1965, pp. 961–962
47. J.O’Callahan, C.M.Chou, I.W.Lieu, “Determination of Rotational Degrees of Freedom for
Moment Transfers in Structural Modifications”, Third International Modal Analysis Confer-
ence, Orlando, Florida, January 1985, pp. 465–470
48. J.O’Callahan, P.Avitabile, R.Madden, I.W.Lieu, “An Efficient Method of Determining Rota-
tional Degrees of Freedom from Analytical and Experimental Modal Data”, Fourth Interna-
tional Modal Analysis Conference, Los Angeles, California, February 1986, pp. 50–58
49. J.O’Callahan, “Comparison of Reduced Model Concepts”, Eighth International Modal Analy-
sis Conference, Orlando, Florida, January 1990, pp. 422–430
50. Kammer, DC, “Test-Analysis Model Development Using the Exact Modal Reduction”,
The International Journal of Analytical and Experimental Modal Analysis, Oct 1987,
pp. 174–179
51. Kammer,D.C., “A Hybrid Approach to Test Analysis Model Development for Large Space
Structures”, Journal of Acoustics and Vibrations, Vol 113, No 3, July 1991, pp. 325–332
52. Kammer D. C., “Sensor Placement for On-orbit Modal Identification and Correlation of Large
Space Structures.” Journal of Guidance, Control and Dynamics, V. 15. No. 2. pp. 251–259,
1991.
53. Kammer, DC, Triller, MJ, “Efficient Sensor Placement on On-Orbit Modal Identification
of Sequentially Assembled Large Space Structures”, Tenth International Modal Analysis
Conference, San Diego, February 1992, pp. 954–964
54. Kammer, DC, “Effects of Noise on Sensor Placement for On-Orbit Modal Identification of
Large Space Structures:, Journal of Dynamic System, Measurement and Control, Vol 114,
September 1992, pp. 436–443
55. Kammer, DC, “Effects of Noise on Sensor Placement for On-Orbit Modal Identifi-cation of
Large Space Structures:, Journal of Dynamic System, Measurement and Control, Vol 114,
September 1992, pp. 436–443
56. Kammer D. C., Cessna J, Kostuch A., “An Effective Mass Measure for Selecting Free-
Free Target Modes”, Twenty-Third International Modal Analysis Conference, Orlando, FL,
February 2005.
57. Kidder, RL, “Reduction of Structure Frequency Equations”, AIAA, Vol 11, No 6, June 1973,
p. 892
58. Kientzy,D., Richardson,M., Blakely,K., “Using Finite Element Data to Setup Modal Tests”,
Sound & Vibration, June 1989
59. Lallement,G. et al., “Optimal Selection of the Measured Degrees of Freedom and Application
to a Method of Parameter Correlation”, Ninth International Modal Analysis Conference, San
Diego, California, 1991.
60. Lang,G.F., “Including Rotational Terms in a Modal Model...An Experimental Prelude”, Sound
& Vibration, November 1989
61. Larson C. B.. Zimmerman D. C.. and Marek, E. L., “A Comparison of Modal Test Planning
Techniques: Excitation and Sensor Placement Using the NASA 8-Bay Truss”, Twelfth
International Modal Analysis Conference, Honolulu, HI, February 1994, pp. 205–211.
62. Larson C. B., Zimmerman D. C., and Marek, E. L., “A Comparative Study of Metrics for
Modal Pretest Sensor and Actuator Selection Using the JPL/MPI Testbed Truss”, Proceedings
of the AIAA Dynamic Specialists Conference, Hilton Head, SC, 1994.
63. Levine-West, MB, Kissil, A, Milman, M, “Evaluation of Mode Shape Expansion Techniques
on the Micro-precision Interferometer Truss”, Twelfth International Modal Analysis Confer-
ence, Honolulu, Hawaii, February 1994, pp. 212–218
894 P. Avitabile and M. Mains

64. Lieven, NAJ, Ewins, D, “Expansion of Modal Data for Correlation”, Eighth International
Modal Analysis Conference, Orlando, Florida, February 1990, pp. 605–609
65. Lieven,N.A.J., Ewins,D.J., “Spatial Correlation of Mode Shapes, the Coordinate Modal
Assurance Criteria (COMAC)”, Sixth International Modal Analysis Conference, Orlando,
Florida, February 1988, pp. 690–695
66. Lipkins, J., Vandeurzen, U., “The Use of Smoothing Techniques for Structural Modification
Applications”, Proceedings of the 12th International Seminar on Modal Analysis, Leuven
Belgium, 1987, S1–3
67. Mains, M.L., “Orthogonality for Modal Vector Correlation: The Effects of Removing
Degrees-of-Freedom”, Thirty-first International Modal Analysis Conference, Orange County,
California, 2013.
68. Mains, M.L., “Orthogonality for Modal Vector Correlation: The Effects of Removing
Degrees-of-Freedom”, Thirty-first International Modal Analysis Conference, Orange County,
California, 2013.
69. Mains, M.L., “Orthogonality for Modal Vector Correlation: The Effects of Removing
Degrees-of-Freedom”, Thirty-first International Modal Analysis Conference, Orange County,
California, 2013.
70. Mains, M, Vold, H, “Investigation of the Effects of Transducer Cross-Sensitivity and Mis-
alignment Error on Modal Vector Correlation”, Thirteenth International Modal Analysis
Conference, Nashville, Tennessee, February 1995, pp. 1048–1056
71. Mains, M.L., “Orthogonality for Modal Vector Correlation: The Effects of Removing
Degrees-of-Freedom”, Thirty-first International Modal Analysis Conference, Orange County,
California, 2013.
72. O’Callahan, JC, Lieu, IW, Chou, CM, “Determination of Rotational Degrees of Freedom
for Moment Transfers in Structural Modifications”, Third International Modal Analysis
Conference, Orlando, Florida, February 1985, pp. 465–470
73. O’Callahan,J.C., “A Procedure for an Improved Reduced System (IRS)”, Seventh International
Modal Analysis Conference, Las Vegas, Nevada, February 1989, pp. 17–21
74. O’Callahan,J.C., Avitabile,P., Riemer,R., “System Equivalent Reduction Expansion Process”,
Seventh International Modal Analysis Conference, Las Vegas, Nevada, February 1989,
pp. 29–37
75. O’Callahan, JC, “Comparison of Reduced Model Concepts”, Eighth International Modal
Analysis Conference, Orlando, Florida, February 1990, pp. 422–430
76. O’Callahan, JC, Li, P, “A Non-Smoothing SEREP Process for Modal Vector Expan-
sion”, Twelfth International Modal Analysis Conference, Honolulu, Hawaii, February 1994,
pp. 232–238
77. O’Callahan, JC, Li, P, “An Automatic Selection of Reduced Degrees of Freedom”,
Twelfth International Modal Analysis Conference, Honolulu, Hawaii, February 1994,
pp. 481–485
78. O’Callahan, JC, Li, P, “An Automatic Selection of Reduced Degrees of Freedom”, Twelfth
International Modal Analysis Conference, Honolulu, Hawaii, February 1994, pp. 481–485
79. O’Callahan, JC, Li, P, “The Effects of Modal Vector Expansion on Finite Element Model
Updating”, Thirteenth International Modal Analysis Conference, Nashville, Tennessee, Febru-
ary 1995, pp. 599–607
80. O’Callahan, JC, “Development of a General Pseudo Orthogonality Correlation Procedure”,
Thirteenth International Modal Analysis Conference, Nashville, Tennessee, February 1995,
pp. 1013–1021
81. Park, YS, Kim, HB, “Sensor Placement Guide for Model Comparison and Improvement”,
Fourteenth International Modal Analysis Conference, Dearborn, Michigan, February 1996, pp.
404–409
82. Paz, M, “Dynamic Condensation”, AIAA Journal, Vol22, No 5, May 1984, pp. 724–727
83. P.Avitabile, J.O’Callahan, E.Pan, “Effects of Various Model Reduction Techniques on Com-
puted Response”, Seventh International Modal Analysis Conference, Las Vegas, Nevada,
February 1989, pp. 777–785
15 Finite Element Model Correlation 895

84. P.Avitabile, J.O’Callahan, J.Milani, “Comparison of System Characteristics Using Various

Model Reduction Techniques”, Seventh International Modal Analysis Conference, Las Vegas,
Nevada, February 1989, pp. 1109–1115
85. P.Avitabile, Modal Model Correlation, University of Massachusetts Lowell, Copyright TX 4-
764-126 (1998 Dissertation)
86. P.Avitabile, J.O’Callahan, C.M.Chou, V.Kalkunte, “Expansion of Rotational Degrees of Free-
dom for Structural Dynamic Modification”, Fifth International Modal Analysis Conference,
London, England, April 1987, pp. 950–955
87. P.Avitabile, Modal Model Correlation, University of Massachusetts Lowell, Copyright TX 4-
764-126 (1998 Dissertation)
88. P.Avitabile, T.Foster, J.O’Callahan, “Comparison of Reduction vs. Expansion for Orthogonal-
ity Checks”, Twelfth International Modal Analysis Conference, Honolulu, Hawaii, January
1994, pp. 225–231
89. P.Avitabile, Modal Model Correlation, University of Massachusetts Lowell, Copyright TX 4-
764-126 (1998 Dissertation)
90. P.Avitabile, Modal Model Correlation, University of Massachusetts Lowell, Copyright TX 4-
764-126 (1998 Dissertation)
91. Pechinsky,F., Avitabile,P., Coordinate Orthogonality Check (CORTHOG)”, Twelfth Interna-
tional Modal Analysis Conference, Honolulu, Hawaii, February 1993, pp. 753–760
92. Pechinsky,F., Avitabile,P., “Evaluation of DOF Correlation using the Coordinate Orthogonality
Check”, Thirteenth International Modal Analysis Conference, Nashville, Tennessee, February
1995, pp. 1313–1320
93. Penny, JET, Friswell, MI, Garvey, SD, “The Automatic Choice of Measurement Locations
for Dynamic Testing”, Tenth International Modal Analysis Conference, San Diego, February
1992, pp. 30–36
94. Penny,JE, Friswell,MI, Garvey,SD, “The Automatic Choice of Meas-urement Location for
Dynamic Testing”, Tenth International Modal Analysis Conference, San Diego, CA, February
1992.
95. Pickrel C. R., “A Practical Approach to Modal Pretest Design”, Mechanical Systems and
Signal Processing, Vol. 13. No. 2, pp. 271–295, 1999.
96. Roy, N, Girard, A, Bugeat, LP, “Expansion of Experimental Mode Shapes - An Improvement
of the Projection Technique”, Eleventh International Modal Analysis Conference, Orlando,
Florida, February 1993, pp. 152–158
97. Schedlinski, C, Link, M, “An Approach to Optimal Pick-up and Exciter Placement”, Four-
teenth International Modal Analysis Conference, Dearborn, Michigan, February 1996, pp.
376–382
98. Urgueira, A, Lieven, NAJ, Ewins, D, “A Generalized Reduction Method for Modal Test-
ing”, Eighth International Modal Analysis Conference, Orlando, Florida, February 1990,
pp. 222–227
99. Yap Keng C., Zimmerman D.C., “Optimal Sensor Placement for Dynamic Model Correlation”,
Eighteenth International Modal Analysis Conference, San Antonio, TX, February 2000.
Model Updating
16
John E. Mottershead, Michael Link, Michael I. Friswell,
and Carsten Schedlinski

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 899
2 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 902
3 Modeling Errors and Measurement Inaccuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904
4 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 906
4.1 Undamped Eigenvalue Residual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 906
4.2 Undamped Mode-Shape Residual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 907
4.3 Frequency-Domain Displacement Response Residual . . . . . . . . . . . . . . . . . . . . . . . 908
5 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 909
5.1 Example: Two Degree-of-Freedom Statically Loaded System . . . . . . . . . . . . . . . . 911
6 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914
6.1 Mass, Damping, and Stiffness Matrix Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . 915
6.2 Material Properties, Thicknesses, and Sectional Properties . . . . . . . . . . . . . . . . . . . 916
6.3 Offset Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 917
6.4 Generic Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 918
7 Stochastic Model Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923
7.1 Example: Stochastic Model Updating of a Three
Degree-of-Freedom System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925
7.2 Example: Parameter Selection for Stochastic Model Updating . . . . . . . . . . . . . . . . 930

J. E. Mottershead ()
School of Engineering, University of Liverpool, Liverpool, UK
e-mail: [email protected]
M. Link
Department of Civil and Environmental Engineering, University of Kassel, Kassel, Germany
e-mail: [email protected]
M. I. Friswell
College of Engineering, Swansea University, Swansea, UK
e-mail: [email protected]
C. Schedlinski
ICS Engineering GmbH, Dreieich, Germany
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 897

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_18
898 J. E. Mottershead et al.

8 Validation of Updated Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 931

8.1 Benchmark Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 934
8.2 Summary of Model Validation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
9 Industrial Example Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
9.1 Automotive Example Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 937
10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946

Abstract

The term “model updating” describes the process of adjusting the parameters of
a finite element model in order that its predictions, in terms of eigenvalues and
eigenvectors, are in agreement with measurements obtained by modal testing.
The sensitivity method described in this chapter has been implemented numerous
times in commercial codes and applied successfully in industry. It has become
a mature technology in regular use in the automotive and aerospace industries
worldwide. However, there are various subtleties surrounding the application
of model updating that are discussed here for the benefit of potential users.
Firstly there must be an awareness of the frequency range in which the updated
model is to be applied. The available data is generally insufficient to define
the system parameters without the use of additional information provided by
regularization. And the choice of parameters is of critical importance: it is not
only a matter of choosing sensitive parameters; they should also be chosen as part
of an engineering understanding of the dynamics of the system. Careful choice
of parameters, together with regularization, will lead to validated models that
predict the behavior of the system beyond the scope of the original test data.

Keywords

Model updating · Sensitivity · Parameterization · Regularization · Stochastic

model updating · Validation

Nomenclature

f Vector of forces
x Displacement vector
z Vector of outputs
C Damping matrix
G Matrix of sensitivities
H Matrix of frequency response functions
K Stiffness matrix
M Mass matrix
P Matrix of stiffness-matrix eigenvalues
W Weighting matrix
λ Eigenvalue
θ Vector of parameters
16 Model Updating 899

ϕ Eigenvector
 Matrix of stiffness-matrix eigenvectors
Cov(•, •) Covariance matrix
(•) Mean

1 Introduction

Modern and highly sophisticated finite element (FE) procedures are available for
structural analysis, yet practical application often reveals considerable discrepancy
between analytical prediction and test results. The way to reduce this discrepancy
is to modify the modeling assumptions and parameters until the correlation of
analytical predictions and experimental results satisfies practical requirements.
Classically, this is achieved by a trial-and-error approach, which is generally time-
consuming and may not be feasible in some cases. Thus computational procedures
have been developed to update the parameters of analytical models using test data.
In particular, modal data (natural frequencies and mode shapes) extracted from
measured frequency response data have found broad application as a target for
model parameter adjustment. This procedure was described in detail by Natke [52]
and by the present authors [20, 45, 49] and in recent years has developed into a
mature technology applied successfully for the correction of industrial-scale FE
models.
One of the first attempts to address the problem of updating or “correcting”
finite element models by using vibration measurements was made by Collins, Hart,
Hasselman, and Kennedy [11]. This paper has since proven to be extraordinarily
influential in providing the common basis for modern model updating codes
and techniques using the sensitivity method and probabilistic model updating –
including Bayesian model updating. A review of model updating methods was
carried out by Mottershead and Friswell [45], and this was followed by the
research monograph Finite Element Model Updating in Structural Dynamics [20].
In the review paper, the techniques were separated into two categories: Lagrange
multiplier methods and penalty function methods. The penalty methods were then
divided according to the type of data used: natural frequencies and mode shapes or
frequency response functions (FRFs). The research monograph gives more detail.
The Lagrange multiplier methods are called direct methods: they have the advantage
that closed-form solutions are often available, but their disadvantages include (i)
lack of physical meaning of the updated model, (ii) failure to represent known
connectivities, and (iii) the need for model reduction or eigenvector expansion
techniques because of the mismatch of dimensions between the FE model and
the measurements. Various approaches were tried to overcome these problems,
including Kenigsbuch and Halevi [35] and Smith [63], both of which appeared in
a special issue of Mechanical Systems and Signal Processing on model updating.
Yuen [71] developed a direct method suitable for use with incomplete modal
measurements. Formulation as an inverse eigenvalue problem also leads to direct
model updating solutions [69], with quadratic orthogonality constraints [13] and
900 J. E. Mottershead et al.

model structure preservation [68]. In active vibration control, the concept of

assigning certain chosen poles while the other poles of the system remain unchanged
can be attractive in many applications because instability due to spillover is made
impossible. The same principle was applied to model updating by Carvalho et al.
[10], Mao and Dai [42], and Kuo and Datta [38] to prevent the occurrence of
spurious modes.
The penalty function methods in the review paper [45] are called iterative meth-
ods in the research monograph [20]. They are also known as sensitivity methods
and sometimes called model-based methods. An important part of the sensitivity
method is the parameterization of the finite element model and regularization
of the ill-conditioned updating equations [3, 21, 22, 47]. Further contributions
include Hua et al. [31] who developed an adaptive procedure for adjusting the
regularization parameter between Newton iterations. Webber et al. [67] addressed
the problem of linear regularization and Newton iterations affecting each other in
an undesirable way. Their approach includes nonlinear updating algorithms with
consistent regularization of the updating parameters relative to an a priori estimate
using line search and stopping criteria together with generalized cross-validation
for estimating the optimal regularization parameter. Goulet et al. [26] proposed a
model parameterization method based on model falsification – i.e., on the principle
that in science, data cannot truly validate a hypothesis; it can only be used to falsify
it. A space of possible models (a combination of parameters) was then generated,
and an error-domain falsification procedure was used to reject instances that have
unlikely differences (residuals) between predictions and measurements. Wang et al.
[66] carried out nonlinear model updating in the frequency domain by a sensitivity-
based approach that firstly localized and characterized discrete nonlinearities before
identifying the nonlinear parameters using a semi-analytical output residual. A
tutorial on the sensitivity method in model updating was presented by Mottershead
et al. [49].
The area of greatest research activity in very recent times has been in the develop-
ment of probabilistic (and non-probabilistic) methods for uncertainty quantification
in model updating, reviewed in detail by Simoen et al. [62] covering sensitivity,
Bayesian, interval, and fuzzy methods, and with particular emphasis on damage
detection. Among many notable contributions, Jacquelin et al. [33] introduced a
direct method using random matrix theory. Adhikari and Friswell [1] represented
randomized beam-bending properties using the Karhunen-Loève expansion. Goller
et al. [24] proposed a robust method using multidimensional Gaussian kernel
density functions in the case of insufficient data, thereby enabling the quantification
of design insensitivity. Mthembu et al. [50] used Bayesian evidence for model
selection, essentially selection from a finite set of candidate parameter groups based
on plausibility. Batou [7] considered uncertainty in the placement and orientation of
sensors and actuators, optimizing the sensitivity of measured data and the robustness
of updating parameters.
A distinction should be made regarding the interpretation of uncertainty. Epis-
temic uncertainty can be reduced by applying additional data. This applies, for
example, to the case in model updating when measurements are noisy and the
16 Model Updating 901

parameters are distinct but are estimated using probabilistic methods in terms of
mean values and covariances. The standard deviations may be used as a measure
of confidence in the estimated means. The other viewpoint on uncertainty is the
frequentist one, known as aleatory uncertainty, which is irreducible, and represents,
for example, the distribution on the depth of a finite set of nominally identical
beams. In this case the distribution on the depths of the beams is physically
meaningful, as well as being useful to a designer wishing to understand the variation
in performance resulting from the distribution. Bayesian approaches are inherently
epistemic. Early examples include Beck and Katafygiotis [8] and Katafygiotis and
Beck [34], whereby experimental data is used to progressively revise the updating
parameters expressed by a posterior probability density function. One problem
with the Bayesian approach has been the requirement for large computation, now
largely overcome as demonstrated by Goller et al. [25] using parallelization of the
updating code together with the transitional (Markov chain Monte Carlo) MCMC
algorithm, which identifies parameter regions with the highest posterior probability.
Behmanesh et al. [9] used a hierarchical Bayesian method to account for inherent
variability due to temperature changes, temperature gradient, wind speed, and traffic
flow in civil engineering structures. The reader is referred to Yuen [70] for a detailed
exposition of Bayesian inference in model updating.
Mares et al. [43] and Mottershead et al. [48] were the first to use the term
stochastic model updating. Hua et al. [30] and Khodaparast et al. [36] developed
efficient perturbation methods, applicable to the aleatory (frequentist) problem
of multiple, nominally identical test structures. Govers and Link [27] extended
the classical sensitivity-based model updating procedure for the determination of
parameter mean values and covariances; similar to the perturbation approach, this
technique was based on an assumption of small variability. It was demonstrated
very convincingly using data obtained by repeated disassembly and reassembly of
the DRL AIRMOD structure [28]. Hua et al. [32] developed a reliability index to
assess the quality of a model updated using a perturbation approach. A thorough
comparison of sensitivity and Bayesian methods using test data from the AIRMOD
structure was given by Patelli et al. [54]. Au [5] considered the connections between
the Bayesian and frequentist approaches and concluded that results obtained by the
two methods were very similar in the case of little or no modeling error.
Uncertainty quantification without the restriction of small variability generally
demands multiple runs of deterministic finite element code, and the expense of
doing so has led to the extensive use of surrogate models. The simplest of these are
polynomial input-output response surfaces, as described, for example, by Fang et al.
[16] with Monte Carlo simulation (MCS) and significance evaluation using analysis
of variance (ANOVA). Zhang et al. [72] used the polynomial chaos expansion as a
surrogate as well as an evolutionary MCMC algorithm where a population of chains
is updated by mutation to avoid being trapped in local basins of attraction. Non-
probabilistic methods have also been applied in model updating. Khodaparast et al.
[37] used a Kriging model in interval updating to construct a bounding hypercube
over the space of parameter uncertainty. This has the advantage of generating not
merely a fitted surrogate but the most probable input-output representation that
902 J. E. Mottershead et al.

reproduces the finite element model exactly at the training points. Fang et al. [17]
developed an interval response surface by removing the interaction terms from a
second-order polynomial response surface and completing the square. This had
the advantage, not only of efficiency, but also it also avoided the overestimation
frequently encountered with interval arithmetic.
This chapter provides a basic introduction to the most important procedures of
computational model updating and includes simple tutorial examples to reinforce
the reader’s understanding together with a typical model updating example taken
from the automotive industry, in Sect. 6.

2 Parameter Estimation

Parameter updating techniques aim to fit the parameters of a given initial analytical
model in such a way that the model behavior corresponds as closely as possible to
the measured behavior. The resulting parameters represent estimated values rather
than true values since the test data are unavoidably polluted by unknown random
and systematic errors. Also the mathematical structure of the initial analysis is not
unique depending on the idealizations made by the analyst for the real structure.
The residuals containing the test/analysis differences may be formed by force
and response equation errors, by eigenfrequency and mode shape errors, and by
frequency response errors.
The first step in parameter estimation is the definition of a residual containing the
difference between analytical and measured structural behavior

εz = zm − z (θ) (1)

where zm denotes the measurement. The analytical prediction is z(θ) and θ

represents the parameters to be updated. The analytical predictions are calculated
from the equation of motion in the frequency domain:

 
f = Z (iω) x; Z (iω, θ) = −ω2 M (θ) + iωC (θ) + K (θ) (2)

where Z ∈ CN × N is the dynamic stiffness matrix, with the finite element mass
matrix M, the damping matrix C, and the stiffness matrix K.√The excitation
frequency is denoted by ω, the excitation force vector is f, and i = −1.
Typically the eigenvalue λ and the mode shape ϕ are used as the analytical
predictions. These are calculated from the undamped eigenvalue problem
[−λ M(θ) + K(θ)]ϕ = 0 while the complex frequency response functions
H(ω) = Z−1 may be determined by inversion of Eq. (2).
16 Model Updating 903

The objective is to minimize

J (θ) = εTz Wε εz = εTz WTv Wv εz (3)

where the symmetric weighting matrix Wε has been included to account for the
importance of each individual term in the residual vector. Wε is difficult to estimate,
although at the very least this weighting should include scaling to equalize the effect
of amplitude and a reasonable choice is Wε = [diag(zm )]−2 . In general the model
response vector z(θ) represents a nonlinear function of the parameters resulting in
a nonlinear minimization problem. One of the techniques to solve this nonlinear
optimization problem is to expand the model response vector into a Taylor series
about the current parameter estimate, θ = θi , truncated after the linear term and
leading to the linearized expression:

εz ≈ ri − Gi (θ − θi ) = zm − zi − Gi θi (4)

where ri = zm − zi denotes the residual, the difference between the measured and
analytically predicted outputs, zm and zi = z(θi ), at the ith iteration. The sensitivity
matrix Gi is given by
 
∂zj
Gi = (5)
∂θk θ=θ i

where j = 1, 2, . . . , q denotes the output data points and k = 1, 2, . . . , p is the

parameter index. The sensitivity matrix Gi is computed at the current value of the
complete vector of parameters θ = θi . The error, εz , is assumed to be small for
parameter θ in the vicinity of θi .
The minimization ∂J/∂θi = 0 of the weighted objective function (3) together
with Eq. (4) yields the linear equation system:

Wv Gi θi = Wv ri (6)

which at each iteration step i is solved for θi and the model is then updated to give

θi+1 = θi + θi . (7)

This procedure continues until consecutive estimates θi and θi + 1 are sufficiently

converged.
The case when fewer measurements than parameters are available in Eq. (6)
(q < p) leads to an underdetermined system, whose solution is not unique. Indeed,
if rank(Gi ) = q and ri ∈ range(Gi ), then the model is able to reproduce the mea-
surements, i.e., εz = 0. Even if a minimum norm or a minimum parameter change
solution is selected, the resulting parameters will in general not retain their physical
meaning. In parameter updating the number of measurements should always be
made larger than the number of parameters (q > p) which yields overdetermined
904 J. E. Mottershead et al.

equation systems. A very effective way of choosing suitable parameters is by the

subset selection described by Lallement and Piranda [39] and Friswell et al. [21].
The data should of course be sensitive to the selected parameters, which must be
justified by physical understanding of the structure and the test arrangement. For the
overdetermined case, the solution of Eq. (6) with respect to θi gives an improved
parameter estimate as
 −1
θi = GTi Wε Gi GTi Wε ri . (8)

Even in the overdetermined case, the condition of the sensitivity matrix G plays
an important role for the accuracy and the uniqueness
 ofthe solution. A fundamental
requirement to obtain a solution is that rank GTi Wε Gi = p.

3 Modeling Errors and Measurement Inaccuracy

Model updating is essentially a process of adjusting certain parameters of the

finite element model. The user should be aware of numerous sources of modeling
error and make necessary adjustments particularly those aspects of the model
that cannot the corrected by changing the values of selected model updating
parameters. Examples of such errors, listed in categories (1) and (2) below, are
related to the mathematical structure of the model and generally referred to as model
structure errors. These errors cannot be eliminated by model updating and should be
eliminated or reduced by careful interrogation of the model before the application of
model updating techniques. The errors listed under category (3) are typical of those
that can be corrected by model updating:

1. Idealization errors resulting from the assumptions made to characterize the

mechanical behavior of the physical structure. Such errors typically arise from:
• Simplifications of the structure, for example, when a plate is treated like a
beam, which might or might not be erroneous depending on the length-to-
width ratio of the plate and the frequency range to be covered.
• Inaccurate assignment of mass properties, for example, when distributed
masses are modeled with too few lumped masses or when rotational inertia
is disregarded.
• When the finite element formulation neglects particular properties, for exam-
ple, when the influence of transverse shear deformation or warping due to
torsion in beam elements is neglected.
• Errors in the connectivity of the mesh, i.e., some elements are not connected
or are connected to a wrong node.
• Erroneous modeling of boundary conditions, for example, when an elastic
foundation is assumed to be rigid.
16 Model Updating 905

• Erroneous modeling of joints, for example, when an elastic connection is

assumed to be rigid (clamped) or when an eccentricity of a beam or a plate
connection is omitted from the model.
• Erroneous assumptions for the external loads.
• Erroneous geometrical shape assumptions.
• A nonlinear structure assumed to behave linearly.
2. Discretization errors introduced by numerical methods such as those inherent in
the finite element method, for example:
• Discretization errors when the finite element mesh is too coarse so that the
modal data in the frequency of interest is not fully converged.
• Truncation errors in order reduction methods such as static condensation.
• Poor convergence and apparent stiffness increase due to element shape
sensitivity.
3. Erroneous assumptions for model parameters, for example:
• Material parameters such as Young’s modulus or mass density
• Cross section properties of beams such as area moments of inertia
• Shell/plate thicknesses
• Spring stiffnesses
• Nonstructural mass

When the model includes idealization and discretization errors, it may only be
updated in the sense that the deviations between test and analysis are minimized.
The same happens when the selected correction parameters are not consistent with
the real source and the location of the error. The parameters in such cases may lose
their physical meaning after updating. A typical result of updating such inconsistent
models is that they may be capable of reproducing the test data but may not be useful
to predict the system behavior beyond the frequency range used in the updating.
Similarly, they may not be able to predict the effects of structural modifications or
to serve as a substructure model to be assembled as part of a model of the overall
structure.
The aim of all structural analyses to predict the structural response can only be
achieved if all three kinds of modeling errors are minimized with respect to the
given purpose of the structural analysis. Models that fulfil these requirements shall
be called validated models. Model quality must therefore be assessed in three steps:

Step 1: Assessment of idealization and numerical method errors (model structure

errors) prior to parameter updating.
Step 2: Correlation of analytical model predictions and test results and selection of
correction parameters.
Step 3: Assessment of model quality after parameter updating. Since a unique
solution cannot be expected, this requirement must be related to the intended
purpose for which the model is used, for example:
906 J. E. Mottershead et al.

• To predict the system behavior to types of load or response other than those
used in the test
• To predict the system behavior beyond the frequency range and/or at degrees
of freedom other than those used for updating
• To predict the effects of structural modifications
• To check if the model, when used as a substructure within an assembled
complete structure, will improve the response of the whole model

The validation of updated finite element models is discussed in more detail in

Sect. 9.

4 Sensitivity Analysis

In practice, the first step is the definition of a residual. In this section we consider
the following residuals, representing probably the most widely used techniques: real
eigenvalues, real mode shapes, and the frequency-domain displacement residual.
A comprehensive selection of these and other residuals with special consideration
of statistically based weighting and the statistical properties of the parameter
estimates is given by Natke et al. [53].

4.1 Undamped Eigenvalue Residual

The linearized undamped eigenvalue residuals are defined by the differences

between the vector of measured eigenvalues λm and their analytical counterparts
λ(θ), which being undamped are entirely real. The eigenvalues in this case are
defined as the squares of the system natural frequencies, λj = ωj2 , j = 1, 2, . . . ,
at the linearization point, “i.” Thus the eigenvalue residual and sensitivity are given
by Eqs. (4), (5), and (6) when z(θ) = λ(θ); zm = λm . It is necessary to ensure
that the analytical and measured eigenvalues correspond to the same physical mode.
This process, usually referred to as mode pairing, may be achieved by carrying out
a modal correlation using the modal assurance criterion (MAC) [20]. Special care
has to be taken for systems with repeated roots (e.g., axisymmetric systems) where
arbitrary combination of mode shapes (analytically and experimentally) may cause
significant degradation of MAC values. To overcome this, a general method for
transforming analytical eigenvectors is, for instance, described in Schedlinski and
Staples [58], which allows for an easy and especially automated compensation of
the effect.
The terms in the sensitivity matrix may be determined analytically [18] by
differentiation of the undamped eigenvalue equation:

 
∂λj ∂M ∂K
= ϕTj −λj + ϕj (9)
∂θk ∂θk ∂θk
16 Model Updating 907

where M, K ∈ N × N are the finite element mass and stiffness matrices, respec-
tively, and ϕj is the jth mass normalized mode shape. It is seen that only the jth
eigenvalue and eigenvector are needed to calculate all the jth eigenvalue sensitivities.
As well as the analytical approach to calculating the sensitivity matrix terms, it
is also possible to obtain numerical approximations by the simple procedure of
perturbing the parameters in turn by a suitably small quantity and determining
numerically the change in the predicted eigenvalues and eigenvectors. Although
efficient procedures exist in finite element codes such as MSC.NASTRAN, there is
a considerable advantage in using analytically determined sensitivities when large-
scale structures are to be updated.

4.2 Undamped Mode-Shape Residual

The linearized undamped mode-shape residuals are the differences between the
measured mode shapes at a restricted number of degrees of freedom corresponding
to the location of sensors and the analytical mode shapes at the same coordinates.
The differences are determined at Nm < N measured degrees of freedom. Thus
the mode-shape residual and sensitivity are given by Eqs. (4), (5), and (6) when
z = ϕ(θ); zm = ϕm , where the vectors ϕ(θ) and ϕm may contain many concatenated
vectors of the different analytical and measured mode shapes, respectively. The
analytical and experimental mode shapes should be normalized in the same way.
The determination of mode-shape sensitivities is a significant computational task,
and several approaches are available. In this chapter we consider only the method of
Fox and Kapoor [18]. The method, based on expanding the gradients into a weighted
sum of the eigenvectors, is widely used due to its simplicity of implementation:

∂ϕj 
H
= aj kh ϕh ; H ≤N (10)
∂θk
h=1

which, after substitution of Eq. (10) into the derivative of the eigenvalue equation,
produces the factor ajkh in the form:

ϕTh −λj ∂M
∂θk +
∂K
∂θk ϕj
aj kh = ; h = j (11)
λj − λh

and

1 ∂M
aj kj = − ϕTj ϕj . (12)
2 ∂θk

This expansion is exact if H=N modes are used. For H < N the expansion
represents an approximation depending on the number of modal terms. Corrections
908 J. E. Mottershead et al.

to this approach have been investigated by several authors. Equations (10) and (11)
show that the expansion contains large factors ajkh for neighboring eigenvalues
λj ≈ λh that can cause convergence problems. Nelson’s method [20] has the
advantage that only the mode shape of interest is required but can be very time-
consuming in the case of large-scale finite element models. The antiresonance
residual may be used as an alternative to the undamped mode-shape residual, as
was considered by D’Ambrogio and Fregolent [12].

4.3 Frequency-Domain Displacement Response Residual

The frequency response error linearized at θ is obtained from the differences of the
measured and the analytical frequency response at the measured degrees of freedom
nm < n. The analytical frequency response is given by

x (ω, θ) = H (ω, θ) fm (ω) (13)

at those degrees of freedom that coincide with the measured ones.

The sensitivity matrix Gr of the frequency response may be obtained by using
the identity

∂H/∂θk = −H (∂Z/∂θk ) H (14)

where
 −1
H (ω) = Z−1 = −ω2 M + iωC + K . (15)

Thus,

∂x
Gr =
∂θk
  (16)
∂M ∂C ∂K
= −H (ω, θ) −ω2 + iω + x (ω, θ) ; k = 1, 2, . . . , p.
∂θk ∂θk ∂θk

The frequency response function matrix H (ω, θ) = −ω2 M (θ) + iωC (θ) +
+K (θ)]−1 may be expressed either by


q
H (ω, θ) = ϕj ϕTj / ωj2 − ω2 + i2ωωj ζj (17)
j =1

in the case of proportional or modal damping or by

16 Model Updating 909


2q
H ω, θ(i) = ψj ψTj / iω − ηj (18)
j =1

for the general case of nonproportional damping where ηj and ψj denote the com-
plex eigenvalues and eigenvectors. In both cases q represents the number of modes
included in the sum, and the terms in the eigenvectors are those corresponding to
the measured coordinates.
One problem with the frequency response formulation is the measured, and
analytical FRF peaks do not coincide leading to large frequency-domain displace-
ment errors. This problem is particularly pronounced when close modes cross over
each other such that the orders of the test and analytical modes are different. A
possible solution is to firstly update the M and K matrices using a different residual.
The peaks then become closely aligned, and the frequency-domain displacement
response residual can be used to update the damping matrix C. This approach is
described in more detail in Sect. 9.

5 Regularization

The treatment of ill-conditioned, noisy systems of equations is a problem central

to finite element model updating [3, 22, 51, 65]. Such equations often arise in the
correction of finite element models by using vibration measurements. The classical
weighted least squares method described above can be extended in cases where it is
difficult to obtain a convergent solution because of an ill-conditioned sensitivity
matrix. The objective function, Eq. (3), is extended by the requirement that the
parameter changes θ should be minimized, to give

J (θ) = εTz Wε εz + μ2 θTi Wθ θi . (19)

The parameter weighting matrix Wθ should be chosen to reflect the uncertainty in

the initial parameter estimates [44, 55, 67]. This may be formally related to Bayesian
methods, where the optimum matrices Wε and Wθ are the inverse of the output
and parameter variances, respectively [11, 19]. However, the variance of the initial
parameter estimates are rarely known in practice, and an alternative given by Link
(1993) relates the choice of weighting matrix, Wθ , to the inverse of the squared
sensitivity matrix according to

mean (diag ()) −1

Wθ =  ; (20)
mean diag  −1
 
 = diag GTi Wε Gi . (21)
910 J. E. Mottershead et al.

This definition allows the parameter changes to be constrained according to their

sensitivity. In consequence the parameters remain unchanged if their sensitivity
approaches zero. Wθ = I represents the classical Tikhonov regularization [64] used
to solve ill-conditioned systems of equations.
Equation (3) is easily extended to penalize differences between the updated
parameters and the corresponding initial estimates or to penalize differences
between nominally identical parameters [3, 22]. For example, in an experimental
frame structure, a number of “T” joints may exist that are nominally identical. Due
to manufacturing tolerances, the parameters of these joints will be slightly different,
although these differences should be small. Therefore a side constraint is placed
on the parameters, so that both the residual and the differences between nominally
identical parameters are minimized.
Minimizing J in Eq. (19) gives the solution

 −1
θi = GTi Wε Gi + μ2 Wθ GTi Wε ri = Ti ri (22)
  
Ti

where Ti is the generalized pseudo-inverse of the sensitivity matrix.

The question remains how to choose the regularization parameter μ that provides
a balance between the measurement residual, Jε (θ) = εT Wε ε, and the side constraint
(or parameter change), Jθ (θ) = θTi Wθ θi . Link [40] suggested that the factor μ2
lies in the range between 0 (no regularization) and 0.3. High μ values are used
if there are many insensitive parameters, and the matrix GTi Wε Gi is strongly ill-
conditioned. If ill-conditioning is not too strong, μ2 = 0.05. The higher the value
of μ, the higher is the necessary number of iteration steps to achieve convergence.
This regularization approach is very closely related to the optimization of
multiple objective functions. From Eq. (19) it is clear that the residual and side
constraint are functions of μ: Jε (μ) and Jθ (μ). The way in which these two terms
are balanced depends on the size of the regularization parameter μ. If μ is too
small, then the problem will be too close to the original ill-posed problem, but if
μ is too large, then the problem solved will have little connection with the√original
problem. A useful approach is to √ plot the norm of the side constraint, Jθ (μ),
against the norm of the residual, Jε (μ), for different values of μ. For multi-
objective function optimization, this is called the Pareto front and in regularization
is called the L-curve. Hansen [29] showed that the norm of the side constraint is a
monotonically decreasing function of the norm of the residual. He pointed out that
for a reasonable signal-to-noise ratio and the satisfaction of the Picard condition,
the curve is approximately vertical for μ < μopt and soon becomes a horizontal
line when μ > μopt , with a corner near the optimal regularization parameter μopt .
The curve is called the L-curve because of this behavior. The optimum value of the
regularization parameter, μopt , corresponds to the point with maximum curvature at
the corner of the log-log plot of the L-curve. This point represents a balance between
confidence in the measurements and the analyst’s intuition.
16 Model Updating 911

One difficulty in model updating is that the relationship between the parameters
and the measurements is nonlinear and the estimation problem is solved by
constructing the linearized model and iterating until convergence. At each iteration
regularization may be applied, although often the corner of the L-curve disappears
as the iterations progress [65], and thus the value of the regularization parameter
is difficult to determine. Hence, it is often convenient to set the regularization
parameter at the first iteration and retain this value until convergence.

5.1 Example: Two Degree-of-Freedom Statically Loaded System

We now consider the two degree-of-freedom static example shown in Fig. 1. In the
initial model, the spring constants are all 1 N/m, so that k1 = k2 = k3 = 1N/m.
The measured data are taken from a system with k1 = 1.2N/m, k2 = 0.8N/m, and
k3 = 1.0N/m. The purpose of the example is to show how the stiffness correction
may be determined from measured static displacements u1 and u2 . Two cases will
be considered, namely, the underdetermined and overdetermined cases.
The updating equation is formulated using a force residual which is obtained by
introducing the measured displacement response into the equation of motion (2). In
the static load case, ω = 0, this results in an analytical force vector f(θ) = K(θ)um
which for the example of Fig. 1 is given by
   
f1 k + k2 −k2 um1
f= = 1 .
f2 − k2 k2 + k3 um2

The residual defined in Eq. (1) is then calculated from

εz ≈ fm − f − Gθ (23)
 ∂f 
1 ∂f1 ∂f1
where fm denotes the measured force vector and G = ∂k1 ∂k2
∂f2 ∂f2
∂k3
∂f2 =
  ∂k1 ∂k2 ∂k3
um1 um1 − um2 0
is the sensitivity matrix. This latter matrix doesn’t depend
0 −um1 + um2 um2

Fig. 1 Two degree-of-freedom discrete example

912 J. E. Mottershead et al.

on the current stiffness value so that there is no iteration necessary in this simple
case. Thus the updating Eq. (23) is of the same form as Eq. (4).

Case 1: Underdetermined System

Suppose that the force appliedto the springs
 is fm1 = [ 1 0 ]T and the measured
0.60811
static deflection is um1 = . The vector of stiffness changes that
0.27027
reproduces the measured displacement, for the minimum norm of the stiffness
change, may be obtained as k1 = 0.1398, k2 = − 0.0916, and k3 = 0.1354.
Figure 2 shows the results obtained by including the side constraint of minimum
stiffness change, for various values of regularization parameter, μ. In this case the
matrix G is not ill-conditioned, and hence the side constraint yields a solution that
is a weighted average of the minimum norm solution and the zero change solution,
as shown by the smooth variation in the spring stiffnesses with μ in Fig. 2. This is
also the reason the residual plot does not have the L-curve shape. Note also that the
simulated stiffness change is never recovered by the estimation procedure.

Case 2: Overdetermined System

Suppose that a second force, fm2 = [0 1 ]T , is applied to the springs, giving the
0.27027
measured static deflection of um2 = . If this data is combined with the
0.67568
data for the first load case fm1 = [ 1 0 ]T , then the estimation is overdetermined.
Since there is no noise or modeling errors, the exact spring stiffnesses are obtained:
k1 = 0.2, k2 = − 0.2, and k3 = 0.0.

Fig. 2 The estimation results for the underdetermined case. Dotted lines represent the simulated
stiffness change. (Reproduced by kind permission of Elsevier)
16 Model Updating 913

Regularization is most useful when the coefficient matrix is relatively ill-

conditioned and noise is present. In this simple example, ill-conditioning is
obtained by considering measurements taken at two force levels, namely, fm1 =
[ 1 0 ]T and fm3 = [ 1.05 0 ]T . The exact “measured” displacement
 for
 fm3 is
0.63851 −0.0002
um3 = . Noise is added to the measurements: to um1
0.28378 0.0003
 
0.0010
and to um3 . Figure 3 shows the results obtained by including the
− 0.0002
side constraint of minimum stiffness change, for various values of regularization
parameter, μ. The L-curve now has the classical shape, with a defined corner. For
low values of regularization parameter, the estimation is clearly ill-conditioned
leading to large estimated stiffness changes. The condition number of G in this
example is 880. At the corner of the L-curve, the stiffness changes are approximately
those obtained from the minimum norm solution for the underdetermined case. For
a regularization parameter slightly less than that at the corner of the L-curve, the
stiffness changes are very close to those simulated, although this is impossible to
determine solely from the estimated results. Figure 4 shows the results when using
the parameter weighting matrix given by Eqs. (19), (20), and (21). Although similar
to the results of Fig. 3, the updated parameters at the corner of the L-curve are now
much closer to the simulated values, as shown in Table 1.

Fig. 3 The estimation results for the overdetermined case with the minimum norm parameter
change. Dotted lines represent the simulated stiffness change. Circles denote the corner of the
L-curve and the associated values of residuals and stiffness changes. (Reproduced by kind
permission of Elsevier)
914 J. E. Mottershead et al.

Fig. 4 The estimation results for the overdetermined case with parameter weighting matrix given
by Eq. (20). Dotted lines represent the simulated stiffness change. Circles denote the corner of
the L-curve and the associated values of residuals and stiffness changes. (Reproduced by kind
permission of Elsevier)

Table 1 Updated parameter estimates for the overdetermined two degree-of-freedom static
example
μ k1 (N/m) k2 (N/m) k3 (N/m)
No regularization 0 1.779 −0.243 −0.304
L-curve corner, Fig. 3 0.0093 1.149 0.889 1.112
L-curve corner, Fig. 4 0.0179 1.179 0.836 1.046
“Exact” 1.200 0.800 1.000

Examples of regularization using experimental data may be found in the

literature. Examples include Link [40] on a five degree-of-freedom laboratory
test structure using Eq. (22) and Ahmadian et al. [3] who used the L-curve
to determine the regularization parameter for the experimental frame structure,
mentioned previously, with nominally identical welded joints.

6 Parameterization

The amount of information that can be obtained from vibration test data is limited,
and therefore taking more measurements in the same frequency range won’t nec-
essarily result in more information. Neither will the additional measurements allow
more parameters to be estimated necessarily. The number of parameters should be
considerably smaller than the number of measurements. The objective should be that
16 Model Updating 915

the model updating problem will be overdetermined. Often the resulting equations
are ill-conditioned, and it is then necessary to apply additional information in the
form of a side constraint by regularization as described previously. The parameters
should be justified by physical understanding of the structure under test and the test
setup. Ideally the chosen parameters should have a physical meaning directly, but
this is not always possible in practice. Equivalent models and their parameters often
lead to improved models when “physical” parameters cannot be found. The data
should be sensitive to small changes in the parameters. Difficult features, such as
joints, may be made more or less sensitive by choosing different types of parameters.
A study that includes several different parameterizations of the same joint in a space
frame structure is described by Mottershead et al. [47].
When choosing parameters it is always advisable to try to understand the
behavior of the structure globally and locally in those regions where local modeling
inaccuracies might be responsible for discrepancies in predictions. For example,
close study of finite element mode shapes is able to reveal the motion of joints
at each of the measured natural frequencies. Parameters can then be chosen that
influence this motion and their significance in model updating easily confirmed by
sensitivity analysis and subset selection. In regions of high strain energy, one would
usually choose stiffness parameters, whereas mass parameters would be useful in
regions of high kinetic energy. Stiffness is generally more difficult to model than
mass, and it is therefore more likely that errors in stiffness modeling are responsible
for inaccurate predictions than mass errors. Damping is in many respects a special
case. Whereas finite element mass and stiffness matrices may be readily derived
from variational or energy principles, similar derivations for damping are generally
not available. Joints and boundary conditions are particularly difficult to model
closely. In principle it is possible to design tests that increase the sensitivity of
chosen parameters, but this is extremely difficult to achieve in practice.

6.1 Mass, Damping, and Stiffness Matrix Multipliers

Probably the simplest parameters for model updating are nondimensional scalar
multipliers applied at the element or substructure level. The updated model takes
the form

M = M0 + α1 M1 + · · · + αr Mr + · · · + αR MR (24)

C = C0 + β1 C1 + · · · + βs Cs + · · · + βS CS (25)

K = K0 + γ1 K1 + · · · + γu Ku + · · · + γU KU (26)

where in this case U stiffness parameters, S damping parameters, and R mass

parameters are chosen for updating. The subscript “0” denotes the analytical model
before updating. Of course the parameters α r , β s , or γ u may be applied to more
916 J. E. Mottershead et al.

than just one element. In this way the parameters may be applied to substructures,
when those elements sharing the same updating parameter are connected, or there
may be elements dispersed through the mesh that for some physical reason are to be
updated in the same way. One reason why different parts of the structure might be
updated using the same parameter would be the sensitivity of the eigenvalues, (and
eigenvectors) in the frequency range of interest to small changes in the parameters
is very similar. In that case separating the elements whose changes have a similar
effect would be a bad choice, possibly leading to ill-conditioning of the sensitivity
matrix.
It is a good practice to scale the updating equations so that the parameters α 1 ,
. . . , α r , . . . , α R , β 1 , . . . , β s , . . . , β S , and γ 1 , . . . , γ u , . . . , γ U take similar
numerical values. One way to do this is by dividing the parameter correction by
the initial parameter values, which is done implicitly in Eqs. (24), (25), and (26).
Returning to the specific discussion of matrix multiplier parameters, it is seen that
∂M ∂C ∂K
the terms ∂α r
, ∂β ,
s ∂γu
are simply the rth element mass matrix, Mr ; sth element
damping matrix, Cs ; and the uth element stiffness matrix, Ku , a fact which makes
the computation of the various sensitivities especially simple.

6.2 Material Properties, Thicknesses, and Sectional Properties

The most common material property parameters are Young’s modulus and mass
density, identical to the γ u and α r above since the element stiffness and mass
matrices are linear in E and ρ.
In the case of a beam having a rectangular cross section, the element matrices are
⎡ ⎤
12 6l −12 6l
EI ⎢ 6l 4l 2 −6l 2l 2 ⎥ bt 3
Ke = 3 ⎢
⎣
⎥, I= (27)
l − 12 −6l 12 −6l ⎦ 12
6l 2l 2 −6l 4l 2

and
⎡ ⎤
156 22l 54 −13l
ρAl ⎢ 4l 2 −3l 2 ⎥
Me = ⎢ 22l 13l ⎥, A = bt (28)
420 ⎣ 54 13l 156 −22l ⎦
− 13l −3l 2 −22l 4l 2

where l denotes the element length and the breadth and thickness of the cross
section are denoted by t and b, respectively. Young’s modulus and mass density
are represented by E and ρ.
The choice of parameters E and I independently would lead to redundancy
and ill-conditioning since they lead to identical eigenvalue sensitivities (except
for a scaling factor). I and A independently would be a difficult choice to justify
16 Model Updating 917

physically. But, both Ke and Me depend upon b and t, so that the choice of these two
parameters would allow the correction of a beam cross section meaningfully. Other
useful parameters include the thicknesses and dimensions of thin-walled sections
and plate thicknesses.
The material parameters, thicknesses, and cross-sectional dimensions tend to be
powerful updating parameters because they often apply throughout a finite element
mesh affecting a large number of elements. Thus a small change in these parameters
often affects the natural frequencies very considerably.

6.3 Offset Nodes

Joints and boundary conditions are difficult to represent accurately, and it is in these
regions of the model that assumptions are often made. Probably the most common
assumption is that the connection made at a joint or boundary is rigid when in fact
there is flexibility. The problem of introducing flexibility into joints and boundaries
assumed to be rigid can be tackled in a number of different ways, all resulting in
equivalent models with parameters that cannot be justified on physical grounds.
However, the dynamic behavior of the model and its physical usefulness will most
definitely be improved by this approach.
One approach, useful in many applications, is to make use of offset finite element
nodes and to use the offset dimensions to correct the model [46]. Lengthening or
shortening an offset dimension usually corresponds to making the joint more flexible
or to stiffening it, and in this way it is possible to reconcile the modification with
engineering understanding of the structure.

6.3.1 Example: Parameterization of a “T” Joint

The “T” joint under consideration is shown in Fig. 5. It consists of three elements,
two horizontal elements and one vertical element, each of which is inextensible so
that x1 = x4 , x2 = x5 and y3 = y6 .
The shaded region may be considered rigid in which case the degrees of freedom
at nodes 1, 2, and 3 are referred to node 3 according to the connection matrix:

Fig. 5 “T” joint.

(Reproduced by kind
permission of Elsevier)
918 J. E. Mottershead et al.

⎧ ⎫ ⎡ ⎤
⎪
⎪ y1 ⎪
⎪ 1 −β
⎪
⎪ ⎪ ⎢ ⎥⎧
⎪
⎪
⎪ ϑ1 ⎪
⎪
⎪ ⎢ 1 ⎥ x ⎫
⎪ ⎪ ⎪ ⎢ ⎥⎪ ⎪
⎪ ⎪ ⎥⎪ ⎪
3
⎪ y ⎪ ⎢ 1 β ⎪ ⎪
⎪
⎪ ⎪ ⎥⎪ ⎪
2
⎪ ⎪ ⎢
⎪ ⎪
⎪ y3 ⎪
⎪
⎪ ϑ ⎪ ⎢ 1 ⎥⎪ ⎪
⎪
⎪ 2⎪ ⎪ ⎢ ⎥⎪⎪ ϑ3 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎢ ⎥⎪⎪ ⎪
⎪
⎪
⎪ x
⎨ ⎪
3 ⎪
⎬ ⎢
1 α ⎥⎪⎪
⎨ y4
⎪
⎪
⎬
ϑ3 ⎢ 1 ⎥
=⎢
⎢
⎥ ϑ4
⎥⎪ . (29)
⎪
⎪ y4 ⎪
⎪ ⎢ 1 ⎥⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎢ ⎥⎪⎪ y5 ⎪
⎪
⎪
⎪ ϑ4 ⎪
⎪ ⎪ ⎢
⎪ 1 ⎥⎪⎪
⎪
⎪
⎪
⎪
⎪ ⎪ ⎢ ⎥ ⎪ ϑ5 ⎪
⎪
⎪ y5 ⎪
⎪ ⎪ ⎢
⎪ 1 ⎥⎪⎪
⎪
⎪
⎪
⎪
⎪ ⎪ ⎢ ⎥ ⎪ x6 ⎪
⎪
⎪ ϑ5 ⎪
⎪ ⎪ ⎢
⎪ 1 ⎥⎪⎩ ⎪
⎭
⎪
⎪ ⎪ ⎢ ⎥
⎪ x6 ⎪
⎪ ⎪ ⎣
⎪ 1 ⎦
ϑ6
⎪
⎩ ⎭ ⎪
ϑ6 1

Construction of the overall stiffness matrix of the “T” joint by the usual methods
results in a matrix that contains the offsets α and β. If the overall dimensions of
the joint remain unchanged, then increasing the offsets causes the joint to become
stiffer while reducing them makes the joint more flexible. Parameters such as α
and β have been found to be extremely effective in the correction of inaccurately
modeled joints.

6.4 Generic Elements

Generic elements offer perhaps the most sophisticated and most general way to
parameterize a finite element model [2, 4, 23]. The basic idea is to change the
element formulation within the limitations imposed by the number of nodes and the
degrees of freedom available. A number of different methods are available, based
either on eigenvalue decomposition or the application of constraints (not discussed
in the present chapter). Whichever of the different methods are chosen, generic
elements are applied at the element (or substructure) level.
Quite often the mass matrix is considered to be accurate, and only the stiffness
matrix needs to be updated. Then an eigenvalue decomposition of the stiffness may
be carried out:

00
K =
e
T (30)
0P

where the matrix of eigenvectors is orthogonal, since Ke is a real symmetric matrix:

 T  = I. (31)

The eigenvalues and eigenvectors of the stiffness matrix are not the same as
those of the element’s generalized eigenvalue problem, (Ke − λj Me ) ϕj = 0,
16 Model Updating 919

j = 1, 2, 3, . . . . Each eigenvalue can be thought of as a spring coefficient for a

deflection defined by its eigenvector.
A new matrix of eigenvectors may be introduced:

ˆ = S
 (32)

so that a new stiffness matrix may be written as

 
00 T
K̂e = κ T ; κ=S S , (33)
0P

and from Eqs. (31) and (32), it is apparent that S is an orthogonal matrix.
It seems at first that the number of updating parameters might be quite large, but
the number can be reduced very considerably by applying engineering judgment.
For example, the first two stiffness eigenvalues might be very sensitive parameters
for many of the lower vibration modes of the complete structure. In this case only the
first two diagonal terms, p1 , p2 , would be chosen. Otherwise modifying the terms
of κ will lead to changes in both the stiffness eigenvalues and eigenvectors.

6.4.1 Example: Eigenvalue Decomposition of a Beam Element

A beam element stiffness matrix with the properties, EI = 1 and l = 1, can be written
as

⎡ ⎤
12 6 −12 6
⎢ 6 4 −6 2 ⎥
Ke = ⎢
⎣ − 12
⎥
−6 12 −6 ⎦
6 2 −6 4

and possesses eigenvalues 0,

√ 0, ⎤2, 30 (with units of stiffness) and eigenvectors
⎡
1 1/2 0√ 2/√10
⎢ 0 −1 1/ 2 1/ 10 ⎥
=⎢ ⎣ 1 −1/2
√ ⎥.
0√ −2/√ 10 ⎦
0 −1 −1/ 2 1/ 10
The rigid body modes describe pure translation and pure rotation about the center
$2
of mass of the beam. Since Ke = ψTj+2 pj ψj +2 , it is seen that Ke ∈ 4 × 4 is a
j =1
rank 2 matrix, so that only the strain “modes” contribute to the stiffness of the beam
element. The rigid body modes of (Ke − pj I) are the same as the rigid body modes
of (Ke − λj Me ) so that the mass matrix has no influence on the rigid body modes,
which span the null space of the element stiffness matrix.
920 J. E. Mottershead et al.

6.4.2 Example: Generic Element Parameters for a Pinned-Pinned

Beam
The eigenvectors of the stiffness matrix formed from two uniform beam elements
with pinned ends are shown in Fig. 6. The first and third are symmetric, whereas the
second and fourth modes are antisymmetric.
To select modifications to all the eigenvalues but only the symmetric eigenvec-
tors, the matrix S is chosen so that

⎡ ⎤
s11 s13
⎢ 1 ⎥
S=⎢
⎣
⎥
s31 s33 ⎦
1

⎡ 2 p + s2 p ⎤
s11 1 31 3 0 (s11 s13 p1 + s31 s33 p3 ) 0
⎢ 0 p 0 0 ⎥
and W = ST PS = ⎢
⎣ (s13 s11 p1 + s33 s31 p3 ) 0
2 ⎥.
s13 p1 + s33
2 2 p
3 0 ⎦
0 0 0 p4
The element stiffness matrix can be reconstructed as

Fig. 6 Mode shapes of the pinned-pinned beam. (Reproduced by kind permission of Elsevier)
16 Model Updating 921

⎡ ⎤⎡ T ⎤
κ11 0 κ13 0 ψ1
  ⎢ 0 κ22 0 0 ⎥ ⎢ ψT2 ⎥ ⎢ ⎥
Ke = ψ1 ψ2 ψ3 ψ4 ⎢ ⎣ κ31 0 κ33 0 ⎦ ⎣ ψT
⎥ , and there are five updating
⎦
3
0 0 0 κ44 ψT4
parameters, κ 11 , κ 13 = κ 31 , κ 22 , κ 33 , and κ 44 .

6.4.3 Example: Updating a System of Three Beams with Offset Central

Span
We consider the system of three beams connected in-line but with the axis of the
central beam offset from the axes of the two outer beams as shown in Fig. 7. The
breadth of all three beams is 0.2 m, and the material is steel (E = 210 GN/m2 ,
ρ = 7860 kg/m3 ). The system is represented by a finite element model consisting of
ten beam elements as indicated in the figure with rigidly fixed ends. Each node has
three degrees of freedom, axial and transverse displacements and a rotation about
the third axis. The fourth and seventh elements have offset nodes at the left-hand
and right-hand ends, respectively.
The connection matrix defining the offset node at the fourth element is given by

⎡ ⎤
10 a
⎣0 1 0⎦.
00 1

In the first case, a finite element model with initial parameters E0 = 180 GN/m2
and a0 = 0.015 m is updated using the correct choice of parameters. Updating is
carried out using the standard Moore-Penrose pseudo-inverse based on convergence
of the first five natural frequencies. It is seen from Fig. 8 that exact convergence to
the true solution is obtained in two iterations.
In the second case, the finite element model is corrected using generic element
parameters. This time the finite element mass matrix is correct, but in the stiffness

0.1

0.05
4 5 6 7
1 2 3 8 9 10

0.3 0.4 0.3

Fig. 7 System of three beams with offset central span (dimensions in meters). (Reproduced by
kind permission of Elsevier)
922 J. E. Mottershead et al.

Fig. 8 Parameter convergence – E and a. (Reproduced by kind permission of Elsevier)

matrix, the offset is initially given by 0.8 of its true value. The elastic modulus is
correct. The generic element is formed from the group of elements, 3, 4, 7, 8, at
the junctions between the thick and thin beam sections. The stiffness matrices from
elements 3 and 4 are uncoupled from elements 7 and 8, but each pair of elements
has identical eigenvalues. Therefore the eigenvalues of the generic element stiffness
matrix occur in pairs. It is found that the first natural frequency of the beam is not
very sensitive to the generic element parameters, which represent (shape) stiffnesses
at the junctions – well away from the most strained portion of the beam in the
middle. The second and third natural frequencies are very sensitive, and the fourth
mode is axial and therefore insensitive. A single generic element parameter is not
sufficient to produce good results, but in Figs. 9 and 10 results are shown from two
updating parameters κ 11 , κ 22 . It is seen that the second and third natural frequencies
converge correctly after approximately 20 iterations when the parameters κ 11 , κ 22
are fully converged.
Updating was achieved using the weighted (regularized) updating Eq. (22) using
the first ten natural frequencies. The second and third diagonal terms of Wε were
given by values of 13,000 with the remaining diagonal terms set to unity (off-
diagonal terms set to zero), Wθ = diag 250 50 . It is seen that the generic element
parameters are able to provide an updated model that accurately reproduces the
dynamic behavior of the offset beam (in the frequency range considered) even
though the updated model lacks physically meaning. Nevertheless, such a model
may be meaningful within the updating frequency range.
16 Model Updating 923

Fig. 9 Parameter convergence – κ11 , κ22 . (Reproduced by kind permission of Elsevier)

Fig. 10 Convergence of natural frequencies – generic element parameters. (Reproduced by kind

permission of Elsevier)

7 Stochastic Model Updating

The stochastic model updating problem may be expressed as

ze − ze = Gj θ − θ + εj +1 (34)
j +1

by the assumption of small perturbation about the mean. In Eq. (34) the over-bar
denotes the mean; ze , ze are experimentally measured outputs, typically natural
frequencies and mode-shape terms; θj + 1 is the (j + 1)th estimate of parameter
distribution to be determined, with mean θj +1 . The mean sensitivity matrix is
924 J. E. Mottershead et al.

denoted by Gj = G θj , and εj + 1 represents errors introduced from various

sources including inaccuracy of the model and measurement imprecision.
Model updating of the means is a deterministic problem given by

θj +1 = θj + Tj ze − zaj θj (35)

where zaj θj is the predicted output of the model at the jth iteration. The
transformation matrix Tj is the generalized pseudo-inverse of the sensitivity matrix
Gj :

T −1 T
Tj = Gj Wε Gj + Wθ Gj Wε , (36)

and Wε and Wϑ are weighting matrices, to allow for regularization of ill-posed

sensitivity equations (see Eq. (22) in Sect. 5).
It is seen from Eq. (34) that the matrix of output covariances is given by

T
Cov ze , ze = Gj Cov θj +1 , θj +1 Gj + Cov εj +1 , εj +1 (37)

ze = ze − ze ; θ = θ − θj . (38)

Then, if the error covariances are deemed to be small, an estimate of the

parameter covariances may be obtained by inversion, using Eq. (36) to obtain

T
Cov θj +1 , θj +1 = Tj Cov ze , ze Tj . (39)

Equation (39) allows for the computation of Cov(θj + 1 , θj + 1 ) using only
the transformation matrix, Tj , obtained at the final step of deterministic updating
of the means and the measured output covariance. It avoids expensive forward
propagation of uncertainty through the model required by alternative approaches. It
was shown [61] that Eq. (39) may be developed straightforwardly from expressions
given previously by Haddad Khodaparast et al. [36]:

T
Cov θj +1 , θj +1 = Cov θj , θj − Cov θj , zj Tj
T
− Tj Cov zj , θj + Tj Cov zj , zj Tj (40)
T
+ Tj Cov ze , ze Tj

and Govers and Link [27]

16 Model Updating 925

Cov θj +1 , θj +1 = Cov θj , θj

T T
+ Tj Cov ze , ze Tj − Tj Cov zj , zj Tj .
(41)

7.1 Example: Stochastic Model Updating of a Three

Degree-of-Freedom System

The example considered is the three degree-of-freedom mass-spring system shown

in Fig. 11.
The nominal values of the parameters of the “experimental” system are
mi = 1.0 kg (i = 1, 2, 3), ki = 1.0N/m (i = 1, 2, . . . , 5), and k6 = 3.0N/m.
The erroneous random parameters are assumed to have Gaussian distributions
with mean values, μk1 = μk2 = μk5 = 2.0 N/m, and standard deviations
σk1 = σk2 = σk5 = 0.3 N/m. The true mean values are the nominal values
with standard deviations σk1 = σk2 = σk5 = 0.2 N/m (20% of the true mean
values). Parameters k1 , k2 , and k5 are assumed to be independent.

Case 1 – Consistent Set of Updating Parameters

This example comprises a consistent updating problem where three uncertain
stiffnesses, k1 , k2 , k5 , are deemed to be responsible for observed variability in
the three natural frequencies of the system. Eqs. (39) and (40) were applied, and
the initial cloud of predicted natural frequencies was made to converge upon the
cloud of “measured” natural frequencies as shown in Fig. 12. The measured data
consisted of 30 separate measurement points (30 points in the three-dimensional
space of the natural frequencies), and the predictions were represented by 1000
points, needed for forward propagation by Latin hypercube sampling (LHS) with
imposed correlation from a normal distribution θj ∈ Nn θj , Cov θj , θj , in
order to determine zj from θj .
Figure 12 shows the results produced by the two methods, where it is apparent
that the updated covariance ellipses from the two solutions are almost indistinguish-

Fig. 11 Three degree-of-freedom mass-spring example

926 J. E. Mottershead et al.

Fig. 12 Frequency scatter plots (Case 1). (Reproduced by kind permission of Elsevier)

able from each other or from the covariance ellipse of the “measured” data. Note that
the covariance ellipses on the scatter plots encompass 95% of the data (two-sigma
ellipses).
Typical convergence characteristics are shown in Fig. 13, and the updated
parameter values are given in Table 2. The adopted convergence criterion was that
the deviation of the predicted eigenfrequencies with respect to the reference ones
should be less than a specified tolerance.
The CPU times shown in Table 2 are determined with respect to the solution
from Eq. (39). It is seen that for this particular three degree-of-freedom problem,
calculation of the parameter covariance matrix is approximately 300 times faster by
Eq. (39) than by Eq. (40).

Case 2 – Inconsistent Updating Parameter Set

Case 2 presents an example of an inconsistent updating problem where the updating
parameter set does not include all the uncertain parameters responsible for the
16 Model Updating 927

Fig. 13 Convergence plots (Case 1): (a) Mean values of the estimates and (b) standard deviation
of the estimates (dash-dotted line: reference values). (Reproduced by kind permission of Elsevier)

observed variability in the reference responses. As in the previous cases, the

reference data were produced with randomized k1 , k2 and k5 , while the updating
parameter set is composed of k1 , k2 , and k6 , i.e., the uncertain k5 is not included
in the updating parameter set. In this case regularization was applied with Wθ = I,
Wε = 0.1 × I.
Figures 14 and 15 show the results of the updating process. The scatter plots
of Fig. 14 show that the output means are reconstructed faithfully, but the choice
of an inconsistent set of updating parameters has resulted in large errors in the
reconstructed covariance ellipses. The updating parameters k1 , k2 , k6 are fully
928 J. E. Mottershead et al.

Table 2 Parameters and eigenfrequencies values

Updated (error %)
Reference Eq. (40) Updated (error %)
30 obs. Initial (error %) 1000 obs. Eq. (39)
k1 [N/m] 1.001 2.0 (99.73) 1.001 (−0.03) 1.014 (1.26)
k2 [N/m] 0.992 2.0 (101.55) 0.993 (0.06) 0.966 (−2.68)
k5 [N/m] 1.001 2.0 (99.84) 1.001 (−0.02) 1.008 (0.69)
σk1 [N/m] 0.197 0.3 (52.59) 0.194 (−1.59) 0.194 (−1.36)
σk2 [N/m] 0.208 0.3 (44.58) 0.213 (2.35) 0.211 (1.40)
σk5 [N/m] 0.211 0.3 (41.94) 0.211 (−0.05) 0.211 (−0.11)
f1 [Hz] 0.1586 0.2030 (28.02) 0.1586 (−0.00) 0.1586 (−0.00)
f2 [Hz] 0.3180 0.3960 (24.54) 0.3180 (−0.00) 0.3180 (−0.00)
f3 [Hz] 0.4505 0.4823 (7.06) 0.4505 (−0.00) 0.4505 (0.00)
# Iterations – – 9 6
CPU time ratio – – ∼300 1

Fig. 14 Frequency scatter plots (Case 2 – Eq. (39)). (Reproduced by kind permission of Elsevier)
16 Model Updating 929

Fig. 15 Convergence plots (Case 2 – Eq. (39)): (a) Mean values of the estimates and (b) standard
deviation of the estimates (dash-dotted line: reference values). (Reproduced by kind permission of
Elsevier)

converged after 30 iterations as shown in Fig. 15. This result demonstrates that the
selection of updating parameters on the basis of reconstructing the output means is
not sufficient to ensure that the output covariances will be well reconstructed.
The inconsistent parameter problem was addressed by Silva et al. [61] who
showed how the updated parameters should be selected based on the scaled covari-
ances of the outputs and a scaled sensitivity matrix, with columns gθk corresponding
to candidate parameters θ k . Based on an assumption that the updating parameters are
mutually independent and independent of measurement noise, the output covariance
930 J. E. Mottershead et al.

4 8 12 16 20

3 7 11 15 19
1 5 9 13 17 21

2 6 10 14 18

Fig. 16 Pin-jointed truss. (Reproduced by kind permission of Elsevier)

matrix was shown to be given by a sum of rank 1 matrices with each term associated
exclusively with a single parameter. The cosine distance between the column gθk and
its projection gθ k on the hypersurface defined by the range of the matrix of output
covariances was used. A cosine distance of zero (or close to zero) is an indicator of
a correctly chosen updating parameter.

7.2 Example: Parameter Selection for Stochastic Model Updating

The pin-jointed truss shown in Fig. 16 has overall dimensions 5 m × 1 m and is

composed of 21 elements in total, each with a stiffness matrix given by

 
1 −1
K = ki ; i = 1, 2, . . . , 21.
−1 1

The five diagonal bars of nominal stiffness EA L = 1.485 × 10 N/m are

8

each randomized for updating. The true mean value of each is equal to the
nominal stiffness, and the standard deviations are given by σkj = 0.135μkj ,
j = 3, 7, 11, 15, 19. For the purposes of parameter selection, the initial estimates
of all the mean stiffnesses, ki , i = 1, 2, . . . , 21, are considered to be 70% of the
reference values, and the standard deviations are given by σkj = 0.27μk3 .
Parameter selection results are shown in Figs. 17, 18, 19, and 20. It is seen that
the correct parameters for updating are recognized correctly in each case of different
numbers of outputs.
It can be seen from the figures that the first bar element k1 has zero cosine
distance. This happens because the boundary condition prevents any extension
or compression of k1 , so that all the outputs are insensitive to it. When the
constraints are removed, so the truss is in the free-free condition, the cosine distance
corresponding to parameter k1 becomes finite as shown in Fig. 21.
16 Model Updating 931

Fig. 17 Cosine distance – first ten eigenvalues. (Reproduced by kind permission of Elsevier)

Fig. 18 Cosine distance – first 15 eigenvalues. (Reproduced by kind permission of Elsevier)

8 Validation of Updated Models

Basic requirements for model validation were formulated in Sect. 3. It was

explained that even when comparisons of experimental modal analysis results
with analytical predictions show satisfactory agreement this does not automatically
mean that the updated model is capable of predicting the structural response for
other loading and/or boundary conditions or for other structural configurations.
In the following some findings are reported from a benchmark study defined
within the European COST Action F3 on “Structural Dynamics” and described
932 J. E. Mottershead et al.

Fig. 19 Cosine distance – all eigenvalues. (Reproduced by kind permission of Elsevier)

Fig. 20 Cosine distance – all eigenvalues and eigenvectors. (Reproduced by kind permission of
Elsevier)

in [14]. The study was aimed at investigating the quality of updated models
under real practical conditions where neither the modeling assumptions nor the
assumptions for updating were unique but defined differently by the participants
of the benchmark. The study should show if the expected non-uniqueness of the
results due to different computational methods, different structural idealizations,
and different parameter sets could be tolerated with regard to the intended purpose.
The following requirements for a validated model were defined:

1. The model must be capable of predicting the experimental modal data and/or
the frequency response functions (FRFs) within the active frequency range and
16 Model Updating 933

Fig. 21 (a) Cosine distance – free-free condition – first ten eigenvalues. (b) Cosine distance –
free-free condition – first 20 eigenvalues. (Reproduced by kind permission of Elsevier)

within certain accuracy limits, of course. The term active frequency range was
related to the frequency range used for computational model updating (CMU).
The above criterion represents a minimum requirement which does not yet say
much about the prediction quality of the model. The prediction quality should
therefore be checked using the following additional criteria:
2. Prediction of the eigenfrequencies and modes beyond the active frequency range.
3. Prediction of the modal data and/or FRFs of a modified structure. For the
benchmark structure, two structural modifications were considered consisting of
additional masses fixed at two different locations as shown in Fig. 22.

The participants were allowed to generate any initial FE model that they found
suitable.
934 J. E. Mottershead et al.

Fig. 22 Geometry and accelerometer location

8.1 Benchmark Data

The benchmark structure was a laboratory structure built to simulate the dynamic
behavior of an aeroplane. The structure was initially built for a benchmark study
on experimental modal analysis conducted by the Structures and Materials Action
Group (SM-AG19) of the Group for Aeronautical Research and Technology in
EURope (GARTEUR) [6, 14, 15, 41]. The test bed was designed and manufactured
by ONERA, France. Figure 22 shows the test structure geometry and the location
of the measured degrees of freedom. The overall length of the structure was 1.5 m,
the wing span was 2.0 m, and the overall mass was 44 kg. The material used was
aluminum. In order to increase the damping, a 1.1 × 76.2 × 1700 mm3 viscoelastic
constraining layer was bonded to the wings. The modal test data for up to 14 modes
for the unmodified and also for the two mass-modified structures were provided for
the participants. Further details are described in Link and Friswell [41].
16 Model Updating 935

8.2 Summary of Model Validation Results

The methods applied by the benchmark participants could be classified according

to:

1. The type of test/analysis residuals

2. The type of FE model (beam or shell elements)
3. The type and number of updating parameters

Very different choices were made by the participants. The model validation was
essentially performed in two steps:

(a) Initial model tuning: This step included updating the parameters of the initial
model of the unmodified structure and to check if the model was capable
of predicting the experimental modal data within the active frequency range
(criterion 1). Some participants extended the checks with respect to the passive
frequency range (criterion 2).
(b) Check of prediction capability concerning the modified structures (criterion 3).

Since all the participants used different residuals for their objective function
to be minimized, and different types of structural idealization (beam or plate
elements), it was difficult to make recommendations on what residual and what
parameter choice were the best. It was necessary to carefully select the parameters
describing the connections, particularly when beam models were used. The most
important issue was to find an appropriate parameter set. With this requirement
fulfilled, good prediction results were found, even with the simple eigenfrequency
residual. Looking at the great variety of parameters, it became obvious that even
though the parameters might be called physical or geometric (like Young’s modulus
or a beam offset), they must be interpreted as non-unique equivalent parameters
describing lumped stiffness and mass properties. It was interesting to note that
good prediction capabilities of the updated models were achieved in many different
ways. In principle it was found that the higher the requirement needed to meet the
structure’s intended purpose, the greater are the number of above validation criteria
that must to be satisfied.

9 Industrial Example Problem

The updating and validation of large-scale finite element models is a challenging

task because of the high degree of complexity of today’s mechanical systems and the
number of candidate updating parameters potentially involved. In order to succeed,
a systematic approach should be adopted as shown in Fig. 23.
A bottom-up strategy has proven to be very effective especially when the overall
system is composed of many interconnected components. This requires the step-
by-step approach shown in Fig. 24, starting with single components (component
936 J. E. Mottershead et al.

FE
(re-) modeling

automated update
test planning analytical modal
analysis

Z0, X
N.O.K.
Z0T, XT adjustment of selected
experimental
Comparison inertia and stiffness
modal analysis
parameters
O.K.

end

Fig. 23 Model validation strategy

Fig. 24 Bottom-up strategy

level) over subassemblies (subsystem level) until the complete system is obtained
(complete system level). Then model updating and validation become feasible as
demonstrated, for example, in Schedlinski and Staples [58], Schedlinski et al.
[59, 60], and Schedlinski [56], and in the automotive example described below.
It is usually assumed that all deviations of the analysis from test data are
due entirely to uncertainties in the finite element model. There are however test
uncertainties too (e.g., exact mounting conditions, signal analysis related errors,
16 Model Updating 937

observability and controllability, data analysis, nonlinearities) which must be kept

as small as possible.
Test planning makes use of the finite element model, which not only enables
the design of the test but also considerably simplifies the later correlation with the
analytical results. Test planning should cover the following aspects:

• Boundary conditions (fixed, free, flexible)

• Target modes (frequency range, local and global modes)
• Measurement degrees of freedom (MDOFs)
– Assessment of required measurement information
– Selection of sensor locations based on FE model, accessibility, and visualiza-
tion of modes
– Unique mode shape recognition using the auto-MAC
• Excitation
– Assessment of suitable excitation points
– Selection, e.g., based on analytical Mode Indicator Functions (MIF)
– Frequency resolution

For test planning and computational model update, several commercial software
tools are readily available capable of handling large-scale finite element models.
Typical parameters for model updating are described in Sect. 6.
It has proven to be effective to update inertia and stiffness properties first, e.g.,
based on eigenvalue and eigenvector residuals. After successful updating of these
parameters, damping parameters can be adjusted by minimizing the deviations in
the resonance regions between measured and simulated FRFs.
A common difficulty in computational model updating is the selection of
updating parameters. A necessary condition, but not a sufficient one, is that the
parameters should be sensitive. Subset selection methods based on comparing the
columns of the sensitivity matrix to the vector of residuals can also be helpful.
However, engineering understanding of the physical structure and the finite element
model is almost always the most important factor. In particular, close inspection
of mode shapes to recognize parameters for updating in highly strained regions
is a skill that the practitioner should aim to develop for great advantage in model
updating.

9.1 Automotive Example Problem

In the following the automotive exhaust system shown in Fig. 25 shall be updated
and validated with particular focus on global vibration behavior and damping.
Special attention will be paid to the modeling of interface stiffnesses and damping
using the bottom-up strategy introduced above. The identification of local damping
parameters will be carried out at every validation step. One important aspect of this
procedure is that the properties of the individual joints can be developed separately.
938 J. E. Mottershead et al.

Fig. 25 Complete
automotive exhaust system
with shaker excitation

Fig. 26 Components and subsystems of the exhaust system

For the bottom-up concept, the exhaust system needs to be separated into a
number of components, subsystems, and subassemblies. Figure 26 shows the chosen
subsystems:
At the component level:

• Converter 1
• Isolation element
• Front part including converter 2
• Rear part with mufflers

At the subassembly and system levels:

• Assembly of rear and front part (outlined in green in Fig. 26)

• Complete system
16 Model Updating 939

Each component and each (sub-) assembly are first subjected to an experimental
modal analysis. Frequency response functions are experimentally determined, and
eigenfrequencies and eigenvectors are identified. Model updating is then carried out
step by step starting at the component level and leading eventually to the overall
system. The bottom-up procedure has several advantages:

• Structural modeling deficiencies of components can be located more easily and

resolved.
• The number of candidate parameters at each validation step can be kept low,
which improves convergence and uniqueness of the updating process, and favors
the finding of physically meaningful results.
• For the validation of the (sub-) assemblies, the main focus can be set on the
interfaces (joints) between the constituent components, where generally the
largest modeling uncertainties exist.
• The separate consideration of joints allows the identification of individual joint
parameters (stiffness and damping).

In what follows, the bottom-up process is demonstrated on the rear part of the
exhaust system at the component level. Then the rear and front parts (Assembly I,
encircled in Fig. 26) are considered at the subassembly level.

9.1.1 Component Level

The agreement of finite element eigenfrequencies and eigenvectors with data
determined from experimental modal tests for each component is checked and, if
necessary, improved by adjusting mass distributions as well as local stiffnesses by
remodeling and model updating. Frequency deviations and MAC values serve as
evaluation criteria for the achieved quality of the finite element model.
Figure 27 and Table 3 show the initial correlation, i.e., the MAC matrix as
well as the MAC values and frequency deviations. The first six eigenvectors
exhibit promising MAC values greater than 80%, but there are two large frequency
deviations of more than 4%.
As the first step, the masses of the finite element model and of the physical
part are compared and reviewed. The total masses match sufficiently well. If large
deviations had been found, the mass of the finite element model could be adjusted
either globally or locally via the densities of the materials or by (additional) mass
elements representing single local masses. At the second step, local stiffnesses of the
finite element model are updated, typically shell thicknesses and Young’s moduli.
Preferably model areas are selected for update which either exhibit a certain degree
of modeling uncertainty or are especially sensitive to parameter changes.
The already mapped eigenvectors of the rear part of the exhaust system represent
global bending modes, characterizing the global vibration behavior of the overall
setup. Since the bending modes are primarily influenced by the stiffness of the tube
between the two mufflers, the modulus of elasticity of the pipe is chosen as the
update parameter. Eigenvectors that were not paired are basically local vibration
modes of the mufflers with limited influence on the global dynamic behavior of the
940 J. E. Mottershead et al.

MAC Matrix For Selected MDOF

1
1

2
0.9
3

4 0.8
EMA Mode # (AGA_shapes_global_final_010.unv)

5
0.7
6

7
0.6
8

9 0.5

10
0.4
11

12
0.3
13

14 0.2

15

16 0.1

17
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
FEA Mode # (aga_ht_01_sol103.op2)

Fig. 27 Initial MAC matrix for the rear part of the exhaust system

Table 3 Initial correlation for the rear part of the exhaust system
No. EMA FEA EMA [Hz] FEA [Hz] Dev. [%] MAC [%]
1 1 7 12.67 12.88 1.70 96.98
2 2 8 22.72 23.99 5.59 93.55
3 3 9 33.74 35.29 4.59 95.65
4 4 10 49.52 50.06 1.10 95.99
5 5 11 61.37 61.28 −0.14 83.27
6 6 12 69.94 70.27 0.46 92.92
7 9 14 173.35 169.54 −2.19 57.83
8 12 16 282.53 260.21 −7.90 52.64
30% Upper limit of frequency deviations
50% Lower limit for MAC values
16 Model Updating 941

overall setup. Therefore, proper representation of these local modes is rated as less
important.
The actual model update is carried out using the dedicated model validation
software ICS.sysval [57] that makes special use of the MSC.Nastran structural
optimization capabilities (SOL200). The software allows for a direct use of Nastran
models and thus is capable of handling industrial size finite element models.
The quality of the updated rear part of the exhaust system is presented in Fig. 28
and Table 4. The MAC values are improved by the changes applied to the model,
and the relative frequency deviations are, except for one eigenfrequency, now below
3%. This result can be regarded as good with respect to the goal of validating the
global vibration behavior of the complete exhaust system.
After the successful adaptation of mass and stiffness, the updating process
concentrates on the damping. For this purpose damping must be defined in the finite
element model locally, e.g., as structural damping for individual materials or parts
of the model or as discrete viscous damping, e.g., for joints.

MAC Matrix For Selected MDOF

1
1
2 0.9
3
EMA Mode # (AGA_shapes_global_final_010.unv)

4 0.8
5
0.7
6
7
0.6
8
9 0.5
10
0.4
11
12
0.3
13
14 0.2
15
16 0.1

17
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
FEA Mode # (aga_ht_04c_sol103.op2)

Fig. 28 MAC matrix after model update for the rear part of the exhaust system
942 J. E. Mottershead et al.

Table 4 Correlation after model update for the rear part of the exhaust system
No. EMA FEA EMA [Hz] FEA [Hz] Dev. [%] MAC [%]
1 1 1 12.67 12.45 −1.71 96.97
2 2 2 22.72 23.38 2.87 94.07
3 3 3 33.74 34.20 1.37 96.42
4 4 4 49.52 49.87 0.71 96.18
5 5 5 61.37 60.23 −1.86 88.67
6 6 6 69.94 68.60 −1.91 94.34
7 9 8 173.35 167.25 −3.52 73.13
30% Upper limit of frequency deviations
50% Lower limit for MAC values

Fig. 29 Error sums before (FEA initial, left) and after (FEA updated, right) damping update for
the rear part of the exhaust system

The evaluation of the finite element model in terms of damping is based

on the frequency-domain displacement residual. The non-paired eigenfrequencies of
the local modes are disregarded. By adjusting the damping, the differences between
the analytical and measured FRF peaks are minimized at the paired eigenfrequen-
cies. During the model update process, the convergence should be checked for
different damping seed values manually selected based on user experience. By
cataloging identified damping values for typical cases, a knowledge database can be
obtained for material and joint damping that may be used for future modeling tasks.
For the rear part of the exhaust system, structural damping of materials (muffler
housings, acoustic wool in mufflers, and connecting tube) was selected as update
parameters. Figure 29 shows the individual error sums before (left bars) and after
(right bars) updating the damping, as well the total sum for all paired measured
eigenfrequencies (the pair of bars to the extreme right). Figure 30 shows a measured
FRF next to its counterpart calculated with the finite element model before and after
16 Model Updating 943

Fig. 30 Example of FRFs before (FEA initial) and after (FEA updated) damping update for the
rear part of the exhaust system

damping updating. The reduction of the error sums due to the damping update and
the significantly improved match of the FRFs in the resonance areas are clearly
visible.

9.1.2 Subassembly Level

After successful model updating of all components with respect to mass, stiffness,
and damping, the joints within the assemblies can be updated. The approach is
similar to that taken with the components: first the overall mass is checked and,
if necessary, adjusted; second the stiffnesses of the joints are assessed based on
relative frequency deviations and MAC values and if necessary updated. Then, in
the third step, the joint damping is updated based on the deviations of the calculated
FRFs from their measured counterparts.
In the case of Assembly I of the exhaust system (consisting of the rear and
the front part), the clamped joint is represented in the finite element model by
a completely modeled clamp assembly. For model updating Young’s modulus
and structural damping of the clamp material are chosen as updating parameters.
Alternatively, a spring/viscous damper element representation of the clamp could
have been considered.
Figure 31 and Table 5 show the MAC matrix and the correlation table after
stiffness update of the clamp. For Assembly I 12 global vibration modes can be
paired with high or very high MAC values. The relative frequency deviations are
below 3% for the first eight natural frequencies. Figure 32 shows a FRF from the
test, as well as its FE-generated counterpart before and after damping updating.
Again, the updating of damping significantly reduces the deviations in the resonance
regions between calculation and test.
944 J. E. Mottershead et al.

MAC Matrix For Selected MDOF

1
1
0.9
2
EMA Mode # (AGA_shapes_global_final_070.unv)

3 0.8

4
0.7
5
0.6
6

7 0.5

8
0.4
9
0.3
10

11 0.2

12
0.1
13
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
FEA Mode # (aga_ht_vtkv02_05a_sol103.op2)

Fig. 31 MAC matrix after model update for Assembly I

10 Conclusions

This chapter describes the complete procedure for finite element model updating
by the sensitivity method. It begins with the general formulation of an objective
function and linearization of the output in terms of parameters, thereby permitting
parameter estimation by iteration. After the separate treatment of systematic errors,
typically model structure errors, the procedure consists essentially of defining a
residual, describing the finite element discrepancy with respect to test data, such
as eigenfrequencies, modes shapes, or frequency response functions. The resulting
equations are generally ill-conditioned and require regularization by the application
of side constraints. The corner of the L-curve defines an optimal value of the
regularization parameter that maximizes the condition of the combined system of
updating equations and side constraints. The selection of updating parameters is a
16 Model Updating 945

Table 5 Correlation after model update for Assembly I

No. EMA FEA EMA [Hz] FEA [Hz] Dev. [%] MAC [%]
1 1 1 7.53 7.46 −0.88 97.81
2 2 2 10.55 10.62 0.71 95.68
3 3 3 23.70 23.77 0.28 94.23
4 4 4 31.79 32.01 0.71 91.12
5 5 5 42.28 41.75 −1.25 95.02
6 6 6 53.48 53.77 0.55 93.27
7 7 7 61.71 60.37 −2.17 95.81
8 8 8 69.33 70.14 1.17 94.46
9 9 9 128.42 136.66 6.41 83.50
10 11 11 162.16 169.27 4.38 90.63
11 12 10 173.73 166.22 −4.32 77.50
12 13 12 195.36 184.26 −5.69 72.69
30% Upper limit of frequency deviations
50% Lower limit for MAC values

Fig. 32 Example of FRFs before and after damping updating for Assembly I

crucial step requiring deep understanding of the physical test structure and the finite
element model. Numerous parametrization techniques are described. Stochastic
model updating is capable of determining the statistics (means and covariances)
of updating parameters responsible for observed output variability in nominally
identical test structures. Updated models are usually said to be validated when
demonstrated to be capable of predicting the behavior of the physical system under
different conditions from those used in the updating process. Numerical examples
are used throughout to illustrate the main points.
946 J. E. Mottershead et al.

Finally, an industrial example, that of an automotive exhaust system, rep-

resentative of many multicomponent engineering assemblies, is described. The
bottom-up strategy is adopted, whereby the standard process of mass and stiffness
updating is improved by carrying out additional updating of damping parameters.
This leads to a significantly improved match between calculated and measured
FRFs. The highlighted procedure for damping update is in particular important
for nonmetallic materials such as catalyst ceramics and for joints, where reliable
damping parameters are not readily available. The validated finite element model of
the assembly matches the measured eigenfrequencies, eigenvectors, and FRFs much
better than the initial model, while realistic damping estimates are obtained, thereby
highlighting the potential of the model updating and model validation procedures.

Acknowledgments The content and images presented for the automotive example problem are
published with the kind permission of the Volkswagen AG, Germany.

References
1. Adhikari S, Friswell MI (2010) Distributed parameter model updating using the Karhunen-
Loève expansion. Mech Syst Signal Process 24:326–339
2. Ahmadian H, Gladwell GML, Ismail F (1997) Parameter selection strategies in finite element
model updating. ASME J Vib Acoust 119:37–45
3. Ahmadian H, Mottershead JE, Friswell MI (1998) Regularisation methods for finite element
model updating. Mech Syst Signal Process 12(1):47–64
4. Ahmadian H, Mottershead JE, Friswell MI (2002) Physical realisation of generic element
parameters in model updating. Trans Am Soc Mech Eng J Vib Acoust 124(4):628–632
5. Au SK (2012) Connecting Bayesian and frequentist quantification of parameter uncertainty in
system identification. Mech Syst Signal Process 29:328–342
6. Balmes E (1997) Garteur Group on Ground Vibration Testing – results from the tests of a
single structure by 12 laboratories in Europe. Proceedings of the International Modal Analysis
Conference IMAC XV, Orlando, pp 1346–1352
7. Batou A (2015) Model updating in structural dynamics – uncertainties on the position and
orientation of sensors and actuators. J Sound Vib 354:47–64
8. Beck JL, Katafygiotis L (1998) Updating models and their uncertainties. I: Bayesian statistical
framework. J Eng Mech 124(4):445–461
9. Behmanesh I, Moaveni B, Lombaert G, Papadimitriou C (2015) Hierarchical Bayesian model
updating for structural identification. Mech Syst Signal Process 64–65:360–376
10. Carvalho J, Datta BN, Gupta A, Lagadapati M (2007) A direct method for model updating with
incomplete measured data and without spurious modes. Mech Syst Signal Process 21:2715–
2731
11. Collins JD, Hart GC, Hasselman TK, Kennedy B (1974) Statistical identification of structures.
AIAA J 12(2):185–190
12. D’Ambrogio W, Fregolent A (2000) The use of antiresonances for robust model updating. J
Sound Vib 236:227–243
13. Datta BN, Deng S, Sokolov VO, Sarkissian DR (2009) An optimization technique for damped
model updating with measured data satisfying quadratic orthogonality constraints. Mech Syst
Signal Process 23:1759–1772
14. Degener M (1997) Ground vibration results from the tests of an aircraft model performed as
part of an European round robin exercise. Proceedings of the CEAS international forum on
aeroelasticity and structural dynamics, Rome
16 Model Updating 947

15. Degener M, Hermes H (1996) Ground vibration test and finite element analysis of the
GARTEUR SM-AG19 testbed. Deutsche Forschungsanstalt für Luft- und Raumfahrt e V
Institut für Aeroelastik, IB 232-96 J 08
16. Fang SE, Ren WX, Perera R (2012) A stochastic model updating method for parameter
variability quantification based on response surface models and Monte-Carlo simulation. Mech
Syst Signal Process 33:83–96
17. Fang SE, Zhang QH, Ren WX (2015) An interval model updating strategy using interval
response surface models. Mech Syst Signal Process 60–61:909–927
18. Fox R, Kapoor M (1968) Rate of change of eigenvalues and eigenvectors. AIAA J 6:2426–2429
19. Friswell MI (1989) The adjustment of structural parameters using a minimum variance
estimator. Mech Syst Signal Process 3(2):143–155
20. Friswell MI, Mottershead JE (1995) Finite element model updating in structural dynamics.
Kluwer Academic Publishers, Dordrecht
21. Friswell MI, Mottershead JE, Ahmadian H (1998) Combining subset selection and parameter
constraints in model updating. Trans Am Soc Mech Eng J Vib Acoust 120(4):854–859
22. Friswell MI, Mottershead JE, Ahmadian H (2001) Finite element model updating using
experimental test data: parameterisation and regularisation. R Soc Philos Trans Math Phys
Eng Sci 359:169–186
23. Gladwell GML, Ahmadian H (1995) Generic element matrices suitable for finite element
model updating. Mech Syst Signal Process 9(6):601–614
24. Goller B, Pradlwarter HJ, Schuëller GI (2009) Robust model updating with insufficient data.
Comput Methods Appl Mech Eng 198:3096–3104
25. Goller B, Broggi M, Calvi A, Schuëller GI (2011) A stochastic model updating technique for
complex aerospace structures. Finite Elem Anal Des 47:739–752
26. Goulet J-A, Michel C, Smith IFC (2013) Hybrid probabilities and error domain structural
identification using ambient vibration monitoring. Mech Syst Signal Process 37:199–212
27. Govers Y, Link M (2010) Stochastic model updating – covariance matrix adjustment from
uncertain experimental modal data. Mech Syst Signal Process 24(3):696–706
28. Govers Y, Link M (2012) Using stochastic experimental modal data for identifying stochastic
finite element parameters of the AIRMOD benchmark structure. Proceedings of the inter-
national conference on noise and vibration engineering, USD2012, Leuven, Belgium, pp
4697–4715
29. Hansen PC (1994) Regularisation tools: a MATLAB package for analysis and solution of
discrete ill-posed problems. Numer Algorithms 6:1–35
30. Hua XG, Ni YQ, Chen ZQ, Ko JM (2008) An improved perturbation method for stochastic
finite element model updating. Int J Numer Methods Eng 73:1845–1864
31. Hua XG, Ni YQ, Ko JM (2009) Adaptive regularisation parameter optimization in output-
error-based finite element model updating. Mech Syst Signal Process 23:563–579
32. Hua XG, Wen Q, Ni YQ, Chen ZQ (2017) Assessment of stochastically updated finite element
models using reliability indicator. Mech Syst Signal Process 82:217–229
33. Jacquelin E, Adhikari S, Friswell MI (2012) A second-moment approach for direct probabilis-
tic model updating in structural dynamics. Mech Syst Signal Process 29:262–283
34. Katafygiotis L, Beck JL (1998) Updating models and their uncertainties. II: model identifiabil-
ity. J Eng Mech 124(4):463–467
35. Kenigsbuch R, Halevi Y (1998) Model updating in structural dynamics: a generalised reference
basis approach. Mech Syst Signal Process 12(1):75–90
36. Khodaparast HH, Mottershead JE, Friswell MI (2008) Perturbation methods for the estimation
of parameter variability in stochastic model updating. Mech Syst Signal Process 22(8):1751–
1773
37. Khodaparast HH, Mottershead JE, Badcock KJ (2011) Interval model updating with irreducible
uncertainty using the Kriging predictor. Mech Syst Signal Process 25(4):1204–1226
38. Kuo Y-C, Datta BN (2012) Quadratic model updating with no spillover and incomplete
measured data: existence and computation of solution. Linear Algebra Appl 436:2480–
2493
948 J. E. Mottershead et al.

39. Lallement G, Piranda J (1990) Localisation methods for parameter updating of finite element
models in elastodynamics. IMAC IIX, Orlando, pp 579–585
40. Link M (1999) Updating of analytical models – basic procedures and extensions. In: Silva
JMM, Maia NMM (eds) Modal analysis and testing. Kluwer Academic Publishers, Dordrecht,
pp 281–304
41. Link M, Friswell MI (2003) Working group 1: generation of validated structural dynamic
models- results of a benchmark study Utilising the GARTEUR SM-AG19 test-bed. In: Golinval
J-C, Link M (eds) COST action F3 “structural dynamics” (1997–2001)- an European co-
operation in the field of science and technology. Mechanical Systems & Signal Processing,
vol 17, no (1), pp 9–20
42. Mao X, Dai H (2012) Finite element model updating with positive definiteness and no spillover.
Mech Syst Signal Process 28:387–398
43. Mares C, Mottershead JE, Friswell MI (2006) Stochastic model updating: part 1- theory and
simulated examples. Mech Syst Signal Process 20(7):1674–1695
44. Mottershead JE, Foster CD (1991) On the treatment of ill-conditioning in spatial parameter
estimation from measured vibration data. Mech Syst Signal Process 5(2):139–154
45. Mottershead JE, Friswell MI (1993) Model updating in structural dynamics: a survey. J Sound
Vib 167(2):347–375
46. Mottershead JE, Friswell MI, Ng GHT, Brandon JA (1996) Geometric parameters for finite
element model updating of joints and constraints. Mech Syst Signal Process 10(2):171–182
47. Mottershead JE, Mares C, Friswell MI, James S (2000) Selection and updating of parameters
for an aluminium space-frame model. Mech Syst Signal Process 14(6):923–944
48. Mottershead JE, Mares C, James S, Friswell MI (2006) Stochastic model updating:
part 2- application to a set of physical structures. Mech Syst Signal Process 20(8):2171–2185
49. Mottershead JE, Link M, Friswell MI (2011) The sensitivity method in finite element model
updating: a tutorial. Mech Syst Signal Process 25(7):2275–2296
50. Mthembu L, Marwala T, Friswell MI, Adhikari S (2011) Model selection in finite element
model updating using the Bayesian evidence statistic. Mech Syst Signal Process 25:2399–2412
51. Natke HG (1991) On regularization methods applied to the error localization of mathematical
models. Proc Int Modal Analysis Conf IMAC IX, Florence, pp 70–73
52. Natke HG (2004) Einführung in Theorie und Praxis der Zeitreihen und Modalanalyse. Vieweg
Verlag, Braunschweig/Wiesbaden
53. Natke HG, Lallement G, Cottin N (1995) Properties of various residuals within updating of
mathematical models. Inverse Prob Eng 1:329–348
54. Patelli E, Govers Y, Broggi M, Martins Gomes H, Link M, Mottershead JE (2017) Sensitivity
or Bayesian model updating: a comparison of techniques using the DLR AIRMOD test data.
Arch Appl Mech 87(5):905–925
55. Prells U (1996) A regularisation method for the linear error localisation of models of
elastomechanical systems. Inverse Prob Eng 3:197–217
56. Schedlinski C (2012) Finite element model validation of a large spinning facility. Proceedings,
ISMA 2012, Leuven
57. Schedlinski C (2018) ICS.sysval product description. http://www.ics-engineering.com
58. Schedlinski C, Staples B (2004) Computational model updating of axisymmetric systems.
Proceedings of the noise and vibration engineering conference, ISMA 2004, Leuven
59. Schedlinski C, et al (2004) Experimental modal analysis and computational model updating of
a body in white. Proceedings of the noise and vibration engineering conference, ISMA 2004,
Leuven
60. Schedlinski C, et al (2008) Computational model updating of structural damping and acoustic
absorption for coupled fluid-structure-analyses of passenger cars. Proceedings of the Noise and
Vibration Engineering Conference, ISMA 2008, Leuven
61. Silva T, Maia NMM, Link M, Mottershead JE (2016) Parameter selection and covariance
updating. Mech Syst Signal Process 70–71:269–283
62. Simoen E, De Roeck G, Lombaert G (2015) Dealing with uncertainty in model updating for
damage assessment: a review. Mech Syst Signal Process 56–57:123–149
16 Model Updating 949

63. Smith SW (1998) Iterative matrix approximation for model updating. Mech Syst Signal Process
12(1):187–202
64. Tikhonov AN, Arsenin VY (1977) Solutions of ill-posed problems. Wiley, New York
65. Titurus B, Friswell MI (2008) Regularization in model updating. Int J Numer Methods Eng
75(4):440–478
66. Wang X, Hill TL, Neild SA, Shaw AD, Khodaparast HH, Friswell MI (2018) Model updating
strategy for structures with localized nonlinearities using frequency response measurements.
Mech Syst Signal Process 100:940–961
67. Webber B, Paultre P, Proulx J (2009) Consistent regularisation of nonlinear model updating for
damage identification. Mech Syst Signal Process 23:1965–1985
68. Xie D (2011) A numerical method of structure-preserving model updating problem and its
perturbation theory. Appl Math Comput 217:6364–6371
69. Yuan Y (2009) A symmetric inverse eigenvalue problem in structural dynamic model updating.
Appl Math Comput 213:516–521
70. Yuen K-V (2010) Bayesian methods for structural dynamics and civil engineering. Wiley,
Singapore
71. Yuen K-V (2012) Updating large models for mechanical systems using incomplete modal
measurement. Mech Syst Signal Process 28:297–308
72. Zhang EL, Feissel P, Antoni J (2011) A comprehensive Bayesian approach for model updating
and quantification of modelling errors. Probab Eng Mech 26:550–560
Nonlinear System Analysis Methods
17
Janette J. Meyer, Raymond M. Bond, and Douglas E. Adams

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 952
2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954
3 Methods for Nonlinear Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
3.1 Coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
3.2 Frequency Response Function Distortion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 959
3.3 Higher-Order FRFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 961
3.4 Hilbert Transform in the Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964
3.5 Hilbert Transform in the Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 966
3.6 Restoring Force Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 967
3.7 Vibro-Acoustic Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 970
4 Methods for Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973
4.1 Nonlinear Autoregressive Moving Average with Exogenous
Inputs (NARMAX) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973
4.2 Direct Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 976
4.3 Reverse Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 978
4.4 Nonlinear Identification Through Feedback of the Outputs (NIFO) . . . . . . . . . . . . . 983
5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 988

J. J. Meyer
School of Engineering, Vanderbilt University, Nashville, TN, USA
e-mail: [email protected]
R. M. Bond · D. E. Adams ()
Vanderbilt University, Nashville, TN, USA
e-mail: [email protected]; [email protected]

© The Society for Experimental Mechanics 2022 951

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_22
952 J. J. Meyer et al.

Abstract

Practical techniques for experimentally detecting and characterizing system

nonlinearities are demonstrated through a test bed consisting of a composite
panel, both with and without a disbond, undergoing an electrodynamic shaker
excitation. Techniques for detecting and characterizing system nonlinearities
are applied to force and response data collected from this test bed, and then
models are identified for those nonlinearities. Both time and frequency domain
techniques are utilized, and the underlying theory and experimental requirements
for each technique are discussed. References to the literature are provided
throughout the chapter for more in-depth discussion of the techniques.

Keywords

Nonlinearity · Coherence · FRF distortion · Higher-order FRFs · Hilbert

transform · Restoring force · Vibro-acoustic modulation · NARMAX · Direct
parameter estimation · Reverse path · NIFO

Nomenclature

x(t) Time domain response

f (t) Time domain input
X(ω) Frequency domain response
F (ω) Frequency domain input
ω,  Frequency
Gxf (ω) Average cross-spectral density
Gxx (ω), Gff (ω) Average auto-spectral densities
H (ω) Frequency response function (FRF)
Hn (ω) nth order FRF
m, k, c; [M], [K], [C] Linear mass, stiffness, damping
[B(ω)] inverse of the FRF matrix
fi ({x}, {ẋ} Functional form of the i th nonlinearity
[Ai ] Coefficients of nonlinear model terms

1 Introduction

Nonlinearities abound in structural dynamic systems. While idealized linear stiff-

ness and damping representations are effective for characterizing many structures,
a wide range of materials, joints, or loading scenarios present stiffness or damping
characteristics which change significantly based on the amplitude or direction of the
response. This dependence of the stiffness or damping on the response is a hallmark
of a nonlinear structural dynamic system. Some materials have inherent nonlinear
stiffness and damping characteristics. For instance, most elastomeric isolators
provide a stiffness that varies with the relative displacement that is experienced
17 Nonlinear System Analysis Methods 953

across the isolator. Joints can also exhibit nonlinear characteristics because the
internal forces imposed by the joint on a structural assembly vary with the relative
motion across the joint. Backlash, which is a form of hysteresis, is an example of this
type of nonlinear behavior introduced by bolted joints. Other sources of nonlinear
stiffness include structural members that are preloaded in compression, which
allows for buckling as the structure dynamically responds. Nonlinear damping is
common as well, and it can, in some cases, result in dynamic phenomena that
would be impossible in linear systems. For instance, in cases where fluid-structure
interaction injects a type of nonlinear damping, sustained steady-state oscillations
can be present in the absence of a cyclic forcing function. Even sensors and data
acquisition systems can introduce nonlinear effects in experimental measurements
as in the case of clipping of a measurement or operating the sensor outside its linear
calibration range.
Because there are many possible sources of nonlinearity, it is fortunate for the
experimentalist that there are many tools available for system identification of these
nonlinearities. There are nonlinear analysis methods based in the time domain and
methods based in the frequency domain, methods that use sinusoidal excitations
and methods that use random excitations, and so on. This chapter does not discuss
every possible method for analyzing experimental structural dynamic data but,
rather, presents a set of representative methods applied to an experimental test
bed. Through these examples, we will illustrate how these methods work, how the
methods are similar, and how the methods are different in terms of the data and
signal processing techniques required. These examples are meant to illustrate a set
of considerations that an experimentalist might evaluate when selecting the best
method for nonlinear system identification – comprised of detecting nonlinearity,
characterizing nonlinearity, and estimating the parameters in a nonlinear model of a
nonlinear structural dynamic system – in a particular application.
Despite the wide range of methods that are available for nonlinear system
identification, there are a number of common considerations when implementing
any of these methods. Some of these recommended “best practices” in nonlinear
system identification include the following:

• When the option is available, acquire time histories, and do not perform any
averaging of the data upfront. This practice is recommended because nonlinearity
is often masked through the averaging process. Likewise, the experimentalist
should keep in mind that both correlated and uncorrelated noise that appears on
the excitation and response measurements can change the nonlinear technique
that is most suitable for a particular experiment. This is because noise in the
excitation measurement is filtered by input nonlinearities (e.g., u2 (t) where u(t)
denotes an excitation time history) whereas noise in the response measurement is
filtered by output nonlinearities (e.g., x 3 (t) where x(t) denotes a response time
history).
• In order to fully observe both modal (or linear) response characteristics and
nonlinear characteristics that accompany these modal responses, the sampling
frequency should be set high enough to ensure that the effects of aliasing do
954 J. J. Meyer et al.

not inhibit the nonlinear analysis process. A higher sampling frequency than
needed for a corresponding linear system may be required because nonlinearities
introduce additional response characteristics that are nonlinear functions of the
modal responses (e.g., harmonics, modulation).
• The correct forcing function must be selected depending on the nonlinear system
identification approaches that are chosen. For example, when the experimentalist
seeks to detect bifurcation in the steady-state dynamic response amplitude, a slow
swept sinusoidal excitation is recommended because these excitations provide
experimental data that can be analyzed to detect and track the dynamic response
through a bifurcation. Several theoretical studies for validation of test results are
also facilitated through the use of sinusoidal excitations.
• It is recommended that the measurement degrees of freedom be selected in the
experiment in order to fully observe the parts of the system that are thought
to potentially exhibit localized nonlinearities. For example, it is recommended
that response measurements on both sides of a mechanical joint be used to fully
observe the relative motion across the joint.
• When designing the experiment and selecting the test parameters, the exper-
imentalist should keep in mind that all aspects of the system, such as the
stiffness and damping as well as the steady-state response, can depend on the
initial conditions and the amplitude of the excitation and response. For example,
one commonly employed method for studying these dependencies is to select
different amplitudes of excitation for different tests in order to selectively expose
nonlinear behaviors in the locations of the structural system.

In the remainder of this chapter, an experimental test bed that will be used to
apply a number of nonlinear experimental identification method is first described.
Then, various methods for the detection, characterization, and parameter estimation
of nonlinear systems are summarized and applied to this experimental test bed. For
each of these methods, the underlying theory that connects nonlinear phenomena
to the corresponding nonlinear identification approach is also discussed. Some of
the advantages and drawbacks of these techniques are described along the way
to provide the experimentalist with guidance on selecting the most appropriate
methods for a particular application.

2 Experimental Setup

Figure 1 shows the experimental setup used to demonstrate many of the nonlinear
analysis methods described below. The same experimental setup will be used to
demonstrate the application of the methods presented here in order to allow the
reader to better compare and contrast the methods. The test specimen used were
sandwich composite panels each comprised of two fiberglass face sheets bonded
to an aluminum honeycomb core. Two panels of the same size were used during
17 Nonlinear System Analysis Methods 955

testing, both measuring about 20 cm by 20 cm. The first panel was unmodified after
fabrication, while a portion of the face sheet of the second panel was separated
from the core using a razor blade. The purpose of this modification was to introduce
a significant nonlinearity into the structure. The disbond between the face sheet
and core creates a nonlinear stiffness because the interface between the core and
the face sheet has zero stiffness in tension but significant stiffness in compression.
During testing, the edges of two sides of the panel were clamped to steel blocks
using aluminum beams screwed to the blocks, as shown in Fig. 1. Foam tape was
used between the beam and the panel and between the panel and the steel block
to prevent rattling, which is a common source of nonlinearities in experimental
setups. Screws were tightened using a torque wrench to provide consistent boundary
conditions.
Several different excitation methods and waveforms were used throughout testing
to accommodate the requirements of the different analysis methods. The main
excitation source was an electrodynamic shaker (SmartShaker K2007E01) with a
carbon fiber stinger attached to the back side of the panel. Single frequency sine
waves, sine sweeps, stepped sine, and random inputs were used. In addition to
shaker inputs, an instrumented impact hammer (PCB 086C03) was used to provide
impulsive inputs to the panel. Finally, a piezoelectric stack actuator (PI P-010.10P)
was used for high-frequency, sinusoidal excitation. Input forces from the shaker and
the actuator were measured using force sensors (PCB 288D01 and PCB 208C03).
Figure 2 shows the shaker and piezo-actuator force sensors attached to the test
structure.
Three different measurement systems were used to measure the response of the
panels. The three-dimensional scanning laser vibrometer (Polytec PSV-400-3D)
shown in Fig. 3 was used to measure the surface velocity of the panel. A one-
dimensional laser displacement sensor (Keyence LK-H157) was used to measure the
displacement of the panel. And, a single axis accelerometer (PCB 352C22) was used

Fig. 1 Test specimen and

fixture
956 J. J. Meyer et al.

Fig. 2 Force sensors used to

measure input from the stack
actuator (top left) and shaker
(bottom right)

Fig. 3 Laser vibrometer used

to measure response of the
panels
17 Nonlinear System Analysis Methods 957

Fig. 4 Accelerometer, three

co-located lasers from the
vibrometer (red dot), and
displacement laser (line)

Table 1 Analysis methods Detection and characterization Modeling

Coherence NARMAX
FRF distortion Direct parameter estimation
Higher-order FRFs Reverse path
Hilbert transform (Time) NIFO
Hilbert transform (Freq.)
Restoring force
Vibro-acoustic modulation

to measure the acceleration of the panel. For most tests, only the laser vibrometer
was used to measure the response. In some tests, the response of the panel was
measured simultaneously using all three measurement systems, as shown in Fig. 4.
Table 1 lists the analysis methods presented in the remainder of the chapter.

3 Methods for Nonlinear Characterization

3.1 Coherence

Perhaps the most straight forward to apply but least conclusive test for nonlinear
behavior is examining the coherence of a set of measured input and output data
when performing a modal test to estimate frequency response functions (FRFs) for
a system. Coherence is typically calculated during the course of a test, so it is a con-
venient first check for symptoms of structural nonlinearities. Careful interpretation
must be used when trying to identify nonlinearities through coherence, because low
coherence can be caused by a number of other factors, and a nonlinear system may
958 J. J. Meyer et al.

show high coherence if all input data is consistent and the system is linearizable at
typical input levels.
To review, coherence is a measure of the quality of fit for a linear FRF estimated
using a set of input/output measurements [2]. For a system with response x(t) and
2 (ω), is a function of average cross-spectral
input f (t), the coherence function, γxf
density (Gxf (ω)) and auto-spectral densities (Gxx (ω),Gff (ω)):

|Gxf (ω)|2
2
γxf (ω) = (1)
Gxx (ω)Gff (ω)

where ω is the frequency.

For the theoretical case where the system is linear and there is no error or
noise in the measurements, the coherence is unity for all frequencies. In real-world
measurements of linear systems, the coherence is typically high at frequencies
which are well excited with a high signal to noise ratio (SNR). Coherence is
typically low for a linear system at anti-resonances or other frequencies where the
response amplitude is low, and the SNR is therefore lower.
Low coherence values at frequencies that have poor SNR are to be expected
even for linear systems (e.g., at anti-resonances or outside the excitation bandwidth)
and should not be confused with an indication of nonlinearity. However, if the
coherence is low at frequencies where peaks in the FRF appear and the response
amplitude is reasonably high, nonlinearities are one possible explanation. Several
other explanations for low coherence should be ruled out before drawing any
conclusions about the linearity of the system. For instance, peaks in the FRF with
low coherence could be explained by a narrowband input to the system which is
not being measured, clipping or saturation of the response signal, or low forcing
amplitude at that frequency, causing a peak as a result of a near divide by zero
situation. If, however, low coherence is observed where there is a high response
amplitude without any other evident sources of error, nonlinearity is a possible
cause. Another indication of possible nonlinearities is peaks with low coherence
at integer multiples of a linear modal frequency or an excitation frequency. These
trends are not definitive, however, so if the coherence of the measurements indicates
the possibility of structural nonlinearities, further characterization of these potential
nonlinearities is warranted.

3.1.1 Example: Undamaged Panel Versus Panel with Disbond

The coherence of modal impact measurements for each of the panels described
in Sect. 2 were determined from data acquired during testing. In this example,
measurements from a test that used an impact force and a measured acceleration
were used. However, any type of measured output can be used. The choice of input
can also vary, as long as the force spans the frequency range of interest. Impacts,
sine sweeps, and random excitation are all viable choices.
Each panel was impacted five times at a selected point using an instrumented
impact hammer which measures the input force, and the response of the panel was
17 Nonlinear System Analysis Methods 959

1 1
Coherence

Coherence
|FRF|

|FRF|
0.5 0.5

0 0
0 1000 2000 3000 0 1000 2000 3000
(a) (b)

Fig. 5 Coherence ( ) and corresponding FRF ( ) determined from impact testing. (a)
Undamaged panel. (b) Panel with disbond

measured using an accelerometer. Average cross-spectral and auto-spectral densities

were calculated from the five data sets, which were then substituted into Eq. 1 to
calculate the coherence. Figure 5 shows the results. The value of the coherence of the
undamaged panel (solid line) is nearly one across the frequency range excited by the
impact, with the exception of anti-resonances (e.g., 2 kHz) seen in the FRF (dotted
line). The coherence of the panel with the disbond shows significant deviation from
unity at several frequencies. Several dips (e.g., 600 Hz and 1 kHz) correspond to
anti-resonances. However, other deviations (e.g., 1.9 kHz) correspond to peaks in the
FRF. Because it is known a priori that this panel contains a disbond, these deviations
can be attributed to the nonlinearities associated with the disbond. If the condition
of the panel was unknown ahead of time or if no undamaged panel was available for
comparison, further testing would be needed to confirm the cause of the reduction
in coherence.

3.2 Frequency Response Function Distortion

The principle of superposition is fundamental to many of the methods that are used
in linear system analysis. As a consequence of superposition, amplifying an input
signal by some factor results in the output increasing by the same factor. Nonlinear
systems tend to violate this principle, and the resulting effects can be used to detect
the presence of structural nonlinearities. The frequency response function (FRF) of
a dynamic system describes the amplitude and phase of the output relative to the
input in the steady state. The FRF of a linear system is independent of the input
amplitude. If inputs of different amplitudes are used to estimate the FRF of a linear
system, the FRF will remain the same in the absence of measurement error. Many
nonlinear systems, on the other hand, exhibit amplitude dependent nonlinearities, so
the FRFs estimated from two different amplitude inputs tend to differ. This behavior
can be used to observe and detect nonlinearities by exciting the system at different
amplitudes and comparing the estimated FRFs.
960 J. J. Meyer et al.

Harmonic distortion is a concept related to FRF distortion. Harmonic distortion is

the phenomenon exhibited by many nonlinear systems in which the system responds
at harmonics of an input frequency. In a linear system, the system responds only
at the frequencies excited by the input, but nonlinear systems do not in general
obey this rule. By exciting the system with a single frequency and observing the
output spectrum, the presence of harmonics of the forcing frequency can be a good
indicator of nonlinear behavior.

3.2.1 Example: Undamaged Panel Versus Panel with Disbond

The response of the undamaged panel to a sine sweep generated by the shaker was
measured using the 3D laser vibrometer described in Sect. 2. Twenty repetitions of
the sweep were performed in order to determine the average auto- and cross-spectral
density functions, and the FRF (H (ω)) associated with the input, f (t), and output,
x(t), was estimated using the H1 estimator:

Gf x (ω)
H (ω) = . (2)
Gff (ω)

The sweep was repeated at a higher amplitude and a second FRF was estimated.
Figure 6a shows the two FRFs. There is little variation between the two FRFs
estimated from inputs of different amplitudes, which suggests the system is
behaving as a linear system. The same testing and analysis were performed on the
panel with the disbond. Unlike the undamaged panel, the response of the damaged
panel shows significant amplitude dependence. The high and low FRFs shown in
Fig. 6b have noticeable discrepancies, suggesting that the system is nonlinear.
Although a sine sweep excitation and a velocity measurement were used in
this example, any excitation method and output measurement that can be used to
estimate a system’s FRF is a valid choice to perform FRF distortion analysis.

10−2 10−1

10−2
|FRF|
|FRF|

10−3 10−3

10−4

10−4 10−5
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
Frequency (Hz) Frequency (Hz)
(a) (b)

Fig. 6 FRFs determined from low ( ) and high ( ) amplitude inputs. (a) Undamaged panel.
(b) Panel with disbond
17 Nonlinear System Analysis Methods 961

3.3 Higher-Order FRFs

Higher-order frequency response functions (FRFs) are a generalization of first-order

FRFs which are capable of characterizing nonlinearity in a system’s response. In
their most general form, higher-order FRFs express a system’s response to multi-
frequency excitations. The order of the FRF indicates the number of simultaneous
frequency components used in the input signal. The complexity of data acquisition,
visualization, and analysis increases dramatically with increasing order. To fully
determine the higher-order FRF matrices, all possible combinations of excitation
frequencies must be applied. Data visualization is challenging beyond second-order
FRFs because more than three dimensions are required (first-order FRFs require two
dimensions, second-order FRFs require three dimensions, etc.). For these reasons,
a simplification is often made to make the method more practical. As developed in
[13], a single frequency excitation can be used to determine the diagonal compo-
nents of the higher-order FRFs, thus simplifying data acquisition and interpretation.
The FRFs are called diagonal because they are evaluated along an n-dimensional
line of equal frequencies. Some information about a system’s response to certain
combinations of excitation frequencies is lost with this simplification, but the trade-
off in terms of time efficiency is typically justified. Visualization of the diagonal
components of each higher-order FRF requires only two dimensions, where the
system’s response is plotted against the nth multiple of the excitation frequency.
By determining the system’s response at multiples of the excitation frequency, the
type of nonlinearity present in the system can be identified.
Higher-order FRFs are derived from the Volterra model. For a nonlinear system
excited by a single frequency excitation (f (t) = F ej ωt ), the relationship between
the input, f (t), and output, x(t), is expressed as

 ∞  ∞
x(t)=F ej ωt h1 (τ1 )e−ωτ1 dτ1 +F 2 ej 2ωt h2 (τ2 , τ2 )e−j 2ωτ2 dτ2 dτ2 +. . .
−∞ −∞
 ∞  ∞
+F e n j nωt
... hn (τn , . . . , τn )e−j nωτn dτn . . . dτn . (3)
−∞ −∞

where h(t) is the impulse response of the system. Rewriting in terms of Fourier
transforms gives

x(t) = H1 (j ω)F ej ωt + H2 (j ω, j ω)F 2 ej 2ωt + · · · + Hn (j ω, . . . , ω)F n ej nωt .

(4)

where Hn (j ω) is the nth order FRF. It is assumed that the nonlinearity can
be expressed in polynomial form; thus this method is not applicable for non-
polynomial nonlinearities such as hysteretic damping, etc. In addition, the derivation
assumes a complex input of the form f (t) = ej ωt . This assumption is necessary for
the real quantity x(t) on the left-hand side of Eq. 4 to equal the sum of the complex
962 J. J. Meyer et al.

values on the right-hand side. In practice, this input is unrealizable, and a purely real
sinusoidal excitation is typically used instead. Changing the form of the input can
have consequences, however. For a nonlinear system, when a purely real excitation
is used, extra terms are produced which contaminate the theoretical FRF. The source
of these extra terms can be understood by recalling the complex form of a sinusoidal
input

F j ωt
F cos(ωt) = (e + e−j ωt ). (5)
2

In a linear system, the two exponential components in Eq. 5 can be considered

separately, and superposition can be applied. The theoretical FRF can be determined
based on applying Eq. 4 for each of the two inputs. For a nonlinear system,
superposition cannot be assumed, and the entire right-hand side of Eq. 5 must be
substituted for F ej ωt in Eq. 4. This substitution reveals the source of the extra terms.
For example, when formulating the second-order term, the right-hand side of Eq. 5 is
squared, resulting in a ( F2 )2 term in addition to the exponential terms. The additional
terms are said to “contaminate” the FRF. This contamination is unavoidable when
testing a system experimentally.
The term higher-order transfer function may be used to indicate that experimental
results will differ from the theoretical higher-order FRF. Equations for the higher-
order transfer functions are similar to the equations for the theoretical higher-order
FRFs except that they contain the contaminating terms mentioned above. Written
in terms of the Fourier transforms of the input (F (j ω)) and output (X(j ω)), the
theoretical higher-order FRFs are

X(j ω)
H1 (j ω) =
F (j ω)
2X(j 2ω)
H2 (j ω, j ω) =
F (j ω)2
(6)
..
.
2n−1 X(j nω)
Hn (j ω, . . . , j ω) = .
F (j ω)n

When estimating higher-order FRFs experimentally, the above equations are

typically used, and the extra terms are not identified. For example, when estimating
the first-order FRF, H1 ∗ , where ∗ indicates an estimation based on experimental
data, to a first approximation, the standard equation for the linear FRF is used:

H1 ∗ (j ω) = H1 (j ω) + O(Y (j ω))2 . (7)

This equation acknowledges the presence of the extra terms by including

O(Y (j ω))2 , and the terms themselves are ignored in the calculations.
17 Nonlinear System Analysis Methods 963

The experimental procedure for acquiring the data necessary to estimate a

system’s higher-order FRFs is based on a stepped sine excitation. First, the
frequency range of interest should be determined by using a sine sweep or impact
test to identify the relevant natural frequencies of the system. Then, the range of
the stepped sine excitation is identified by selecting the lowest frequency to be
no more than one third the lowest natural frequency of interest and the highest
frequency to be greater than the highest frequency of interest. The one third factor
was used here because FRFs up to third order are typically estimated. In general,
if an nth-order FRF is to be estimated, the lowest frequency used in the stepped
sine input should be chosen to be 1/n times the lowest natural frequency of interest.
The frequency increment used in the stepped sine depends on the desired frequency
resolution of the estimated higher-order FRF and, more practically, on the amount
of time available for the test. Multiple data sets at each frequency of the stepped
sine input should be acquired to allow for averaging. Alternatively, a single time
history can be acquired that is long enough to be subdivided for averaging in the
data post-processing phase. The measurement should be designed such that only the
steady-state response, rather than the transient response, of the system is measured
and an integer number of cycles is acquired to reduce leakage and ensure that
the resulting frequency vector contains the excitation frequency. Note that these
requirements can be enforced either when acquiring data or in the data processing
after acquisition.
Input and output measurements are acquired at each of the frequencies of the
stepped sine. A force transducer should be used to measure the input force of the
actuator or shaker. Response measurements (displacement, velocity, or acceleration)
can be made with any appropriate sensor. The first step in the data processing
procedure is to calculate the force and response spectra via the discrete Fourier
transform (DFT). Next, at each frequency step, i , in the stepped sine input, the
amplitude and phase of the response spectrum at the first n harmonics of the
excitation frequency(i , 2i , 3i ,. . . , ni ) are determined. Only the amplitude
and phase of the force spectrum at the excitation frequency (i ) are recorded. These
values are then used in Eq. 6 to estimate the higher-order FRFs of the system.

3.3.1 Example: Panel with Disbond

A stepped sine input was applied to the panel with the disbond, and the response was
measured with an accelerometer near the location of the disbond. The measured
acceleration and force data were processed as described above, and the diagonal
components of the first-, second-, and third-order FRFs were estimated using Eq. 6.
Figure 7 shows the amplitudes of the estimated FRFs. The first-order FRF matches
well with an FRF estimated from a sine sweep (not shown). The peak near 410 Hz
corresponds to a natural frequency of the panel, and based on the width of the
peak, this is a highly damped mode. The second-order FRF can be used to detect
non-symmetric nonlinearities, i.e., nonlinearities that cause a response at two times
the excitation frequency. If a non-symmetric nonlinearity is present, a peak at
half the natural frequency identified in the first-order FRF should be present. This
peak would be present because when the structure is excited at half the natural
964 J. J. Meyer et al.

0
1010

|H1 |
100
0
200 300 400 500 600
|H2 | 1010

−5
1010
100 200 300 400 500 600
10−2
10
|H3 |

−4
1010
100 200 300 400 500 600
Frequency (Hz)

Fig. 7 First (H1 ), second (H2 ), and third (H3 ) order FRFs estimated for the panel with disbond

frequency, the nonlinearity causes a response at twice the excitation frequency,

which corresponds to the natural frequency and results in a resonant-like response.
In Fig. 7, no clear peak is present, although a wide area of increased response
can be observed between 180 and 300 Hz. The third-order FRF can be used to
identify symmetric (e.g., cubic) nonlinearities that cause a response at three times
the excitation frequency. Peaks at one third the excitation frequency are a result
of the nonlinear response coinciding with the natural frequency. In Fig. 7, a peak
is observed near 136 Hz, which is about one third of the 410 Hz natural frequency
identified in the first-order FRF. Other peaks are also present in the third-order FRF
which do not correspond to natural frequencies of the panel. Because the panel
is highly damped, the peaks in the first-order FRF that correspond to the panel’s
natural frequencies are not very distinct. That makes this method less effective at
characterizing the nonlinearities in the panel. Lightly damped system are better
candidates for the application of this method.

3.4 Hilbert Transform in the Time Domain

One common characteristic of nonlinear system responses is an amplitude depen-

dent period of oscillation. The swinging pendulum is the most famous example
of this behavior. For a linear system, the instantaneous frequency will remain
constant and independent of time, but for some nonlinear systems, the instantaneous
frequency is a function of the response amplitude, and that relationship can be
identified using the Hilbert transform along with associated signal processing
techniques which are required to limit the signal to the response due to a single mode
of vibration. By tracking the changes in frequency with amplitude, the nonlinearity
in a system can often be detected, classified, and located. The plot of amplitude
17 Nonlinear System Analysis Methods 965

versus frequency is called the system backbone and is the main by-product of Hilbert
transform time domain techniques. The Hilbert transform can be used with free
response data to estimate the instantaneous frequency and instantaneous amplitude,
which can be used to show the system backbone and characterize nonlinear behavior
in a system. The Hilbert transform is used to shift all frequency components of a
signal by ±π/2 rad, and the combination of the original signal and the phase shifted
signal can be used to estimate the instantaneous frequency and amplitude of the
signal. For a linear system, the instantaneous frequency will remain constant and
independent of time, but for some nonlinear systems, the instantaneous frequency
is a function of the response amplitude, and that relationship can be identified using
the Hilbert transform along with associated signal processing techniques which are
required to limit the signal to the response due to a single mode of vibration.
To apply time domain-based Hilbert transform analysis to experimental data,
the free response (i.e., impulse response) of the system should first be estimated.
The free response can be estimated from the frequency response function (FRF)
of the system determined using an impulsive force, such as the force from the
impact of a modal impact hammer, by taking the inverse Fourier transform of the
FRF. Alternatively, the free response of the system can be measured directly if it
is possible to give the system non-zero initial conditions (e.g., a pendulum being
released from a horizontal position). Once the free response of the system, x(t), is
determined, an “analytical” signal, X(t), is formed according to

X(t) = x(t) − j x̃(t) (8)

where x̃(t) is the Hilbert transform of x(t). An analytic signal is a signal whose
imaginary part is equal to the Hilbert transform of the real part. (See [9] for a
description of the theory of analytical signals in the context of Hilbert transforms
for vibration analysis.) Equation 8 can be written in polar form as

X(t) = A(t)ej ψ(t) (9)

where

A(t) = x 2 (t) + x̃ 2 (t)
  (10)
x̃(t)
ψ(t) = tan−1 ,
ix(t)

and A(t) is called the instantaneous amplitude (or envelope) and ψ(t) is called
the instantaneous phase. The instantaneous frequency, ω(t), can be found by
numerically differentiating the instantaneous phase. The backbone curve can be
generated by plotting the instantaneous amplitude (A) versus the instantaneous
frequency (ω). Curvature in the backbone curve indicates nonlinearity in the system.
This time domain analysis based on the Hilbert transform is most effective on
lightly damped, single degree of freedom systems. When applying this method to
966 J. J. Meyer et al.

multi-degree of freedom systems, filtering must be applied in order to remove all but
one of the modes from the data. Results from systems with closely spaced and/or
highly coupled modes can be inconclusive. In addition to filtering the measured
data, filtering is necessary to smooth the instantaneous frequency.
Hilbert transform-based analysis can also be used to perform parameter esti-
mation to determine the stiffness and damping characteristics of a system. Two
methods, one using the free response of a system [7] and one using the forced
response of a system [8], were developed that generate a force-displacement curve
that can be used to identify the nonlinear stiffness in the system and a force-velocity
curve that can be used to identity the nonlinear damping in the system. As with the
analysis presented above, lightly damped single degree of freedom systems are the
best candidates for this analysis.

3.5 Hilbert Transform in the Frequency Domain

The Hilbert transform can also be used to identify nonlinear systems in the
frequency domain by checking whether the frequency response functions (FRFs)
of the system are invariant under the Hilbert transform. For linear systems, the FRF
is invariant under the Hilbert transform, but nonlinear systems do not necessarily
share that property. As presented in [14], linear systems are causal, meaning their
impulse response functions are always zero for t < 0. The Hilbert transform acts
as an identity transformation (i.e., the output equals the input) on causal functions.
Therefore, the following relationships hold for a linear system:

H{Im(H (ω))} = Re(H (ω))

(11)
H{Re(H (ω))} = −Im(H (ω))

where H{·} is the Hilbert transform and H is an FRF.

Nonlinear systems (that are also minimum phase systems) are also causal in the
sense that the response never precedes the input. However, due to the data processing
procedure used to calculate the Hilbert transform, nonlinear systems can exhibit
“artificial” non-causality where the impulse response function is not always zero for
t < 0. The non-causality is a result of imposing evenness and oddness conditions
while converting one-sided FRFs to double-sided FRFs and then taking the inverse
Fourier transform. In terms of Hilbert transform analysis, the result of this non-
causality is that the relationships in Eq. 11 do not hold. Therefore, the violation of
these equalities can be used to identify the presence of nonlinearity in a system.
One potential drawback to this method is that in order to detect nonlinearity in
a system, a linear version of the system is required for comparison. The linearized
version of the system can be achieved either by using a separate specimen that is
known to be identical to the specimen being tested (e.g., the unmodified panel used
in the previous examples) or, if possible, by acquiring data from the specimen at a
low enough amplitude such that the nonlinearities are not excited. In either case, the
17 Nonlinear System Analysis Methods 967

only data required to apply this method are FRFs from the linearized system and the
(potentially) nonlinear system. Care should be taken to ensure that the suspected
nonlinearities in the system are sufficiently excited during the data acquisition
process. Once the FRFs have been estimated, the equalities in Eq. 11 are checked by
plotting the real and imaginary parts of each FRF against the transformed imaginary
and real parts. If significant deviation from the linearized system is observed, then
the system can be identified as nonlinear.
Complications can arise when determining the Hilbert transform of the real and
imaginary components of the FRF, however, because the Hilbert transform is an
integration over all frequencies, and the experimentally estimated FRFs are naturally
band limited. Several correction approaches are outlined in [14]; however many are
most effective for lightly damped systems. Therefore, this technique is not generally
appropriate for heavily damped systems.

3.6 Restoring Force Method

The restoring force method is a non-parametric approach for characterizing nonlin-

earities in a system. The derivation of the restoring force is based on a straight-
forward rearrangement of Newton’s second law of motion. By simultaneously
measuring a structure’s displacement, velocity, and acceleration, the so-called
restoring force can be determined and plotted as a function of displacement and
velocity. The nonlinearities in the system can be visualized and characterized
through this restoring force surface. The equation of motion for a single degree
of freedom, linear mass, spring, and damper system is

mẍ + cẋ + kx = f (t) (12)

where x is the displacement of mass, m, c is the damping coefficient, k is the

stiffness coefficient, f (t) is an applied force, and dots indicate time derivatives. To
extend this equation to a system with arbitrary nonlinearities, Eq. 12 can be rewritten
as

mẍ + fRF (x, ẋ) = f (t), (13)

where fRF (x, ẋ) is called the restoring force and represents the internal force
that counters the system’s applied and inertial forces. The restoring force can be
determined by measuring the system’s acceleration, mass, and the applied force
using

fRF (x, ẋ) = f (t) − mẍ. (14)

To determine how this force relates to the system’s displacement and velocity,
the displacement and velocity must be determined either by direct measurement
or by integration of the measured acceleration. The restoring force is plotted as a
968 J. J. Meyer et al.

Fig. 8 Ideal restoring force surfaces. (a) Linear. (b) Quadratic stiffness. (c) Cubic stiffness.
(d) Bilinear damping

surface over the displacement-velocity plane, and the shape of the surface is used to
determine the nature of the system’s nonlinearities. Figure 8 shows ideal restoring
force surfaces for a linear system, a system with a quadratic stiffness nonlinearity,
a system with a cubic stiffness nonlinearity, and a system with bi-linear damping.
The nonlinearity can be characterized by observing the trends of the surface in the
force-displacement plane for stiffness nonlinearities and the force-velocity plane
for damping nonlinearities. Once the type/types of nonlinearity is/are determined,
the fRS (x, ẋ) term in Eq. 13 can be replaced with an algebraic expression for the
nonlinear force (e.g., μx 2 for quadratic stiffness) and a least-squares curve-fitting
can be applied to the data to determine the nonlinear parameter(s).
The design of experiment is particularly important when acquiring data for the
restoring force method. The first consideration is whether a single degree of freedom
system can be assumed. The restoring force method can be extended to systems
with multiple degrees of freedom systems, but, to do so requires the use of modal
vectors, which can change with amplitude in a nonlinear system. In some cases, a
multi-degree of freedom system can be tested in a fixture which isolates the mode of
interest or restricts the motion of the structure such that the assumption of a single
degree of freedom system is a valid simplification. Alternatively, if the mode of
interest is well-separated from other modes of the multi-degree of freedom system,
the input excitation can be designed to excite only the mode of interest.
The second consideration when designing the experiment to acquire data for
the restoring force method is the type of excitation to use. The excitation force
must sufficiently excite the nonlinearity and, importantly, must result in sufficient
coverage of the displacement-velocity plane. The former requirement informs the
selection of the input amplitude, while the latter informs the choice of signal type.
For example, the restoring force in response to a steady-state, periodic excitation
is typically a single closed trajectory in the displacement-velocity plane. Therefore,
this type of excitation does not provide sufficient coverage of the displacement-
velocity plane to generate a complete restoring force surface. It is possible to
improve the coverage of the plane by using periodic inputs at multiple amplitudes,
but, in most cases, a better choice of inputs is a random signal. A discussion of the
effectiveness of other types of inputs for generating restoring force surfaces can be
found in [15].
17 Nonlinear System Analysis Methods 969

The final consideration is what physical states to measure. The restoring force
method requires time histories for the displacement, velocity, and acceleration
of the system. In the example below, all three of these quantities are measured
simultaneously to avoid the need to integrate or differentiate measurements to
estimate unknown states. Direct measurement avoids the many numerical issues
involved in estimating unknown states, but in many cases, simultaneously measuring
displacement, velocity, and acceleration is not an option. In those cases, numerical
integration and/or differentiation is required to estimate states that are not directly
measured. In all cases, a force transducer should be used to measure the input force.

3.6.1 Example: Panel with Disbond

The restoring force method was applied to the damaged panel at the location of
the disbond. The panel was excited with a band-limited random force using the
electromagnetic shaker. The input signal was generated by filtering white noise to
limit the bandwidth of excitation to frequencies in the vicinity of one mode of the
panel in order to better fit the single degree of freedom assumption. The acceleration,
velocity, and displacement of a chosen point were simultaneously measured using an
accelerometer, a laser velocimeter, and a laser displacement transducer, as shown in
Fig. 4. The ability to measure all three of these quantities simultaneously eliminated
the need for numerical integration or differentiation. The input force and response
time histories were measured for 300 s in this case, but comparable results could
have been obtained using shorter time histories.
The restoring force surface was generated by plotting the calculated restoring
force against displacement and velocity divided into finite ranges. The restoring
force was calculated using Eq. 14. In general, if a single degree of freedom system
is assumed, the total mass of the system begin tested can be used. The ranges
of displacement and velocity measurements were then determined and divided
into 50 increments of equal length. These increments were used to divide the
displacement-velocity plane into a grid of equally sized squares. Restoring forces
for measurements corresponding to (x, ẋ) pairs in each square are averaged to form
the restoring force surface over the 50 × 50 grid. Figure 9a shows the restoring force
surface for this experiment using a 3D bar plot.
The random input data used in this experiment provided good coverage of the
displacement-velocity plane. In the event that there are holes in the surface due
to lack of data within a given square of the plane, a longer acquisition time or
a change in excitation signal may address the problem. Alternatively, the surface
could be discretized using larger displacement and velocity intervals to construct
the surface. To determine whether nonlinearities are present in the system (at the
chosen force amplitude), the contours of the surface are analyzed. In this case, the
slope of the surface tends to change at the lower and higher displacement ranges.
A representative profile view of the force-displacement plane is shown in Fig. 9b.
This nonlinear trend in the displacement-force plane indicates that a stiffness
nonlinearity is present. Figure 9c shows a representative view of the force-velocity
plane. The trend in this plane appears more linear. It cannot, however, be concluded
that nonlinear damping is not present in the system. These results could indicate
970 J. J. Meyer et al.

Restoring Force
Restoring Force

0
Displacement
(b)

Restoring Force
Velocity Displacement

0
Velocity
(a) (c)

Fig. 9 (a) Restoring force surface for the panel with face sheet disbond. Representative profiles of
the restoring force surface (b) in the force-displacement plane and (c) in the force-velocity plane

that the nonlinearity was not being excited. Overall, these results clearly show the
presence of nonlinear stiffness in the panel, which is consistent with expectations
based on the asymmetric stiffness at the location of the disbond.

3.7 Vibro-Acoustic Modulation

The vibro-acoustic modulation (VAM) method uses two or more simultaneous

excitation forces to detect nonlinearity in a system. Because the principle of
superposition does not hold for nonlinear systems, interactions between these
multiple inputs are a useful indication of nonlinearity. The principle of superposition
states that the response of a linear system to a multi-frequency excitation is equal to
the sum of the responses of the system excited at each individual frequency. When
excited by a multi-frequency signal, a nonlinear system may respond at additional
frequencies, which are not present in the input, including integer multiples of the
excitation frequencies (as described in Sect. 3.2) and frequencies which are linear
combinations of the excitation frequencies. The amplitude and frequency of these
response components can be used to characterize the nonlinearities in the system [6].
To apply VAM, a structure is excited by a signal of the form

f (t) = cos(1 t) + cos(2 t) (15)

where 1 is chosen to be at or near a low-frequency mode of the structure. The

purpose of this signal, called the pumping signal, is to excite the nonlinearity in the
system. 2 is chosen to be much greater than 1 and, if possible, in a frequency
17 Nonlinear System Analysis Methods 971

range in which the structure’s response is not strongly influenced by its linear
modes. Nonlinearities cause the high-frequency signal, called the probing signal, to
be modulated by the low-frequency response. For example, consider a single degree
of freedom system with a quadratic nonlinearity. The equation of motion is

mẍ + cẋ + kx + ksq x 2 = f (t). (16)

When excited by the force given in Eq. 15, the steady-state response of the
structure will contain terms of the form
    
1
cos(1 t) cos(2 t) = cos (2 + 1 )t + cos (2 − 1 )t (17)
2

due to the quadratic term. These terms represent the modulated part of the response,
where the system responds at the probing frequency (2 ) plus and minus the
pumping frequency (1 ). For a cubic nonlinearity, the response will contain terms
of the form cos((2 + 21 )t) and cos((2 − 21 )t). In general, nonlinearities
that can be expressed as polynomial functions of the system response result in
modulation at frequencies that are linear combinations of the pumping and probing
frequencies.
To identify the presence of nonlinearity in a structure, its response to a multi-
frequency excitation is measured and its response spectrum is determined. The
response is analyzed (visually) to determine if peaks are present at the probing
frequency plus and minus some multiple of the pumping frequency. These peaks
are called sidebands. The frequencies at which the sidebands occur may indicate the
order of the nonlinearity (i.e., quadratic, cubic, etc.) and the amplitude may indicate
the strength of the nonlinearity, where a larger amplitude indicates a stronger
nonlinearity. However, underlying linear effects, such as a strong modal response
at or near the modulation frequencies, can affect sideband characteristics [10].
Care must be taken when drawing conclusions about the nature of the nonlinearity
indicated by the sidebands.
When applying VAM, the forces used are typically single frequency sinusoidal
inputs, as described above. However, the low-frequency sinusoidal input can be
replaced with an impulsive force, such as from a modal impact hammer. When
this substitution is made, the method is referred to as Impact Modulation (IM).
In IM, rather than the probing signal being modulated by only one low-frequency
component, it is modulated by every natural frequency excited by the impact
that excites the system’s nonlinearities. In some cases, IM is more convenient to
implement because a modal impact hammer requires much less instrumentation and
setup than a low-frequency shaker.

3.7.1 Example: Undamaged Panel Versus Panel with Disbond

VAM was applied to both the undamaged panel and the damaged panel. During
each test, a panel was excited simultaneously by a low-frequency sinusoidal force
generated by the modal shaker and a high-frequency sinusoidal force generated by
972 J. J. Meyer et al.

the smaller piezoelectric actuator, as shown in Fig. 2. The frequency of the pumping
signal was chosen to be 453 Hz, which is near a natural frequency of the panels
(determined by a sine sweep test). The probing signal was chosen to be 15 kHz,
which is well outside the range where the panel has strong modal responses. The
response of each panel was measured using the laser vibrometer; however, any type
of response measurement could have been used. Only the response data was used in
the analysis. Time data was acquired for 1.28 s at a sampling rate of 51.2 kHz. The
sampling rate was chosen to be greater than twice the probing frequency plus five
times the pumping frequency to ensure that all sidebands of interest would be in the
usable frequency range. Time histories were transformed into the frequency domain
via the DFT. Twenty spectra were acquired and averaged. In some cases, averaging
may not be appropriate because it may make the presence of the nonlinearity less
evident. Figure 10 shows the results for both the undamaged (solid black) and
modified (dotted red) panel. The large peak at 15 kHz corresponds to the probing
frequency. For the panel modified with a simulated disbond, sidebands at the
probing frequency plus and minus one, two, and three times the pumping frequency
can be clearly seen. A fourth sideband on the left and a fifth sideband on the right of
the probing frequency can also be seen. The presence of these sidebands indicates
nonlinearity in the system. The spectrum of the undamaged panel also shows side-
bands, although their amplitudes are significantly smaller than those of the modified
panel.

100−2

10−4
Response ( ms )

10−6

10−8

10−10
1 1.2 1.4 1.6 1.8 2
Frequency (Hz) ×104

Fig. 10 Response spectra for the undamaged panel ( ) and panel with disbond ( ) deter-
mined from vibro-acoustic modulation testing
17 Nonlinear System Analysis Methods 973

4 Methods for Parameter Estimation

4.1 Nonlinear Autoregressive Moving Average with Exogenous

Inputs (NARMAX)

NARMAX, meaning Nonlinear Autoregressive Moving Average with eXogenous

inputs, refers to both a type of discrete time model and an associated parameter
estimation approach [5]. The NARMAX approach is attractive because it is both
very general and widely applicable to nonlinear parameter estimation. Unlike some
other methods, this technique does not require the form of the nonlinearity to
be assumed, although an understanding of the mechanism of nonlinearity can be
useful in formulating the model. To understand the NARMAX model, it is useful to
start by building from related discrete time linear models. For example, consider a
single degree of freedom mass-spring-damper system, with the equation of motion
mẍ(t) + cẋ(t) + kx(t) = f (t), where x(t) is the displacement of the mass,
f (t) is the applied force, and m, c, and k are the mass, damping, and stiffness,
respectively. Approximating the first derivative of displacement using the backward
difference approximation and the second derivative with the second-order central
difference approximation, the equation of motion can be transformed into a discrete
time equivalent. The discrete time equation of motion at time t = (i − 1)t , where
t = 1/fs , fs is the sampling frequency, and xi = x(it ) is as follows:

mfs2 (xi − 2xi−1 + xi−2 ) + cfs (xi−1 − xi−2 ) + kxi−1 = fi−1 (18)

Grouping each lagged output, the response at time t = it can be expressed as
a weighted sum of the response and force at earlier times:

xi = b1 fi−1 + a1 xi−1 + a2 xi−2 (19)

where the ai and bi are the coefficients to be estimated.

This model is called an ARX model, meaning that it is an autoregressive model,
i.e., the response is a function of itself at earlier times, with exogenous inputs. If
the system is nonlinear, this model needs to be extended to include nonlinear terms.
Adding a cubic damping term to the mass spring damper system from above is a
good example of how the number of coefficients needed to model a nonlinear system
drastically increases relative to the linear equivalent. Starting from the equation of
motion mẍ(t) + cẋ(t) + c3 ẋ(t)3 + kx(t) = f (t), and following the same approach
as above, the discrete time equivalent would be

mfs2 (xi −2xi−1 +xi−2 )+cfs (xi−1 −xi−2 )+c3 fs3 (xi−1 − xi−2 )3 +kxi−1 = fi−1
(20)
974 J. J. Meyer et al.

After expanding the cubic term, Eq. 20 becomes

mfs2 (xi − 2xi−1 + xi−2 ) + cfs (xi−1 − xi−2 )

+ c3 fs3 (xi−1
3
− 3xi−1
2
xi−2 + 3xi−1 xi−2
2
− xi−2
3
) + kxi−1 = fi−1 (21)

After grouping terms and defining coefficients as before, the response at time
t = it can be written as

xi = b1 fi−1 + a1 xi−1 + a2 xi−2 + a3 xi−1

3
+ a4 xi−1
2
xi−2 + a5 xi−1 xi−2
2
+ a6 xi−2
3

(22)

Equation 22 shows that by adding a single cubic damping nonlinearity, this

NARX model (Nonlinear ARX model) has four additional coefficients that must be
estimated. The large number of coefficients that can be needed to model the system
is one of the most significant practical differences of nonlinear discrete time models
as compared to linear counterparts. When the type of nonlinearity is not known,
coefficients for all combinations of lagged input and responses are possible.
The number of unknown terms increases even more when noise is considered.
The effects of noise on nonlinear systems are more complex than they are for linear
systems. Because superposition does not hold for nonlinear systems, even simple
additive noise on the output measurement can result in cross product terms in the
model where noise is multiplied by the response or input. In the nonlinear damping
example above, if the measured response (x  ) is x  = x + e, where x is the true
response, e is a random noise term, the finite equation of motion can be written
by substituting xi = xi − ei . Substituting this expression into just one of the four
2 x
nonlinear terms in (22), a4 xi−1 i−2 , shows how the additive noise on the response
can cause numerous cross product terms to arise:

2  
2
a4 xi−1 xi−2 = a4 (xi−1 − 2xi−1 ei−1 + ei−1
2
)(xi−2 − ei−2 )
2     2 
= a4 (xi−1 xi−2 −2xi−1 xi−2 ei−1 +xi−2 2
ei−1 −xi−1 ei−2 −2xi−1 ei−1 ei−2 +ei−1
2
ei−2 ).

As illustrated in this example, even additive noise on the response can cause
complex noise terms mixed with response terms. For this reason, noise terms need
to be included in the model formulation from the beginning in order to produce an
unbiased estimate of model parameters. The inclusion of these error terms is the
distinction between a NARX model and a NARMAX model. In general terms, the
recurrence relation for a NARMAX model is as follows:

xi = F (xi−1 , . . . , xi−no ; fi−1 , . . . , fi−ni ; ei−1 , . . . , ei−ne ) (23)

where F (. . . ) is some nonlinear function and no , ni , and ne are the number of time
delays included for the output, input, and error terms, respectively. Traditionally,
F (. . . ) is a polynomial function of up to order np , denoted as F (np ) (. . . ), although
17 Nonlinear System Analysis Methods 975

other forms of the nonlinear recurrence relation are possible and have been shown
to be effective. Although the recurrence relation is nonlinear, it remains linear in the
parameters, so coefficients can be estimated using linear least-squares techniques.
The rudimentary solution approach would be to select a model order and number
of time delays and then estimate all of the model coefficients at once to produce
a model. This approach might accurately represent the data that were used in the
estimate, but the model is unlikely to represent the system well in general or
accurately predict the output given a different input. Furthermore, the number of
coefficients which would need to be estimated would be much larger than for a linear
model. For instance, even if only one time delay was used on the input and output,
three time delays were used for the error term, and a cubic polynomial recurrence
relation were used, there would be 56 coefficients in total to estimate because all
combinations of the input, output, and noise terms of different powers would have
an associated coefficient.
Rather than trying to estimate every coefficient at once, a better approach is to use
only as many terms as is necessary to reach an acceptable fit. This is accomplished
by using an orthogonal least squares approach to evaluate the significance of each
term, estimating that term and then iteratively continuing until an acceptable quality
of fit is achieved. Once the parameters have been estimated, the model can be
evaluated for validity, used to predict the output of the system, or be used for further
analysis of the underlying dynamics of the system.

4.1.1 Example: Panel with Disbond

The damaged panel was tested in order to identify a NARMAX model which
captures the nonlinear dynamics of the structure. Displacement data was acquired
via the laser displacement transducer for a point near the disbond, while the panel
was being excited by the modal shaker applying band-limited random force. The
magnitude of the RMS (root mean square) amplitude of the force was chosen to be
large enough to excite the nonlinearity in the panel. The applied force was measured
using the force sensor mounted between the panel and the shaker’s stinger, as shown
in Fig. 2. The length of the time histories acquired was arbitrarily chosen to be
3 s. The number of samples acquired during data acquisition determines how over-
determined the parameter estimation problem is. The measurement duration should
be long enough to ensure that the nonlinear response is captured. When in doubt,
acquiring long time histories which can be truncated in post processing is a good
practice.
The NARMAX model was formulated by first choosing the maximum number of
lags to allow for the input, output, and error terms. In this case, ni = no = ne = 2
was chosen. The choice of two lags was reasonable because displacement was
measured and the equations of motions for the physical system are second order.
If quantities other than displacement are measured or if the model fit is poor
with the initial choice of ni , no , and ne , additional time delays could be used
to improve the model fit. The other key parameter that needs to be specified in
order to formulate the NARMAX model is the maximum order of the nonlinearity,
976 J. J. Meyer et al.

np , for polynomial nonlinear functions. As demonstrated above, increasing the

polynomial order increases the number of unknown coefficients significantly. Prior
knowledge about the nature of the nonlinearity (perhaps gained as a result of
applying one of the characterization methods described previously) can be used to
inform the order choice. If the model form is unknown, the polynomial order can
be chosen heuristically. For this example, the order was chosen to be 3, meaning
that all first-, second-, and third-order polynomials formed by the factors xi−1 , xi−2 ,
fi−1 ,fi−2 , ei−1 , and ei−2 are potential terms in the model. Note that non-polynomial
nonlinearities can also be assumed. Many software packages (e.g., MATLAB) have
built-in functions that support the use of wavelets, neural networks, etc.
With the model form established, there are several possible choices for estimating
the model coefficients. One option is to use a brute-force approach to solve for
every possible coefficient, regardless of the significance to the total response. Again,
commercial software packages such as MATLAB support this type of approach.
Alternatively, an orthogonal least-squares approach can be used to identify and
estimate terms which are significant. For this application, the orthogonal least-
squares approach, described in [4], was used. In the orthogonal least-squares
approach, noise-related terms are estimated by assuming e is a zero-mean, random
(uncorrelated) time series. The result of the orthogonal least-squares solution
approach is a subset of terms from the assumed model that sufficiently model the
system’s response. Note that failure to include noise in the model can lead to biased
parameter estimates. This bias often results in the model being able to predict the
system’s response to the type of input used in the estimation process well, but
not being able to accurately predict the system’s response to inputs of different
amplitudes and/or types.
Figure 11 shows the one-step ahead model predictions for both a random input
(using a different data set than the one used for parameter estimation) and a
sine sweep. There is good agreement between the measured response (solid black
line) and the predicted response (dotted red line) in both cases. One-step ahead
predictions, in which the system output is predicted using the estimated parameters
and measured data, are the least stringent of the model validation techniques.
More stringent techniques include model predicted output, in which the output is
generated using only the measured inputs, and correlation tests, as described in [14].
Model validation is important in order to ensure that the model has captured the
system’s dynamics, and not simply fit an erroneous model to a single set of data. If
the latter is true, the model will be useless to predict the system’s response to inputs
other than the one used for the parameter estimation.

4.2 Direct Parameter Estimation

Direct parameter estimation is a time domain method for estimating mass, damping,
and stiffness parameters for both linear and nonlinear systems based on experi-
mental measurements [11]. The equations of motion must be defined based on a
lumped parameter model of the system, so for nonlinear systems, the form of the
17 Nonlinear System Analysis Methods 977

0.2 0.05

0.1
0
Response

Response
0

-0.05
-0.1

-0.2 -0.1
0 0.05 0.1 0.15 0.2 0.055 0.06 0.065
Time (s) (a) Time (s)

0.4 0.4
0.3 0.3
0.2 0.2
Response

Response

0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
0 0.05 0.1 0.15 0.2 0.055 0.06 0.065
Time (s) Time (s)
(b)

Fig. 11 Measured response ( ) and one-step ahead predictions ( ) from NARMAX model
using (a) a random input and (b) a sinusoidal input

nonlinearity must be explicitly assumed. An overdetermined set of equations based

on the equation of motion is generated using the measured force, displacement,
velocity, and acceleration time histories. A least-squares method can be used to
estimate the parameter matrices. The estimated parameters can then be used with
the underlying equation of motion to simulate the system’s response for validation
purposes.
In practice, the first step in applying direct parameter estimation is to formulate
the equations of motion of the system, including any nonlinear terms. In general,
the equations take the form

[M]{ẍ} + [C]{ẋ} + [K]{x} + fi ({x}, {ẋ}) = {f (t)} (24)

i
978 J. J. Meyer et al.

where fi is the functional form of the i th nonlinearity included in the model. An

important requirement in model formulation is to explicitly express the form of the
nonlinearity (or nonlinearities). For example, when testing a structure with bolted
interfaces, quadratic (i.e., non-symmetric) stiffness elements of the form ksq x 2
may be included to account for micro-impacts that occur between the overlapping
surfaces. It may also be appropriate to include nonlinear damping terms. There is
no general rule for determining the type or types of nonlinearities to include, but
techniques such as restoring force surfaces can be useful in selecting an appropriate
form for the nonlinearity. Ultimately, the validity of the assumed model can be
checked using a procedure described below.
After the equations of motion have been established, the system parameters
are estimated by solving an overdetermined system of equations. This set of
equations is generated by repeating Eq. 24 for a select number of points in time.
The equations are populated using measured time histories of the displacements,
velocities, accelerations, and applied force. Ideally, each of these quantities should
be measured simultaneously. Alternatively, one state can be measured and the others
estimated using an appropriate numerical integration or differentiation scheme.
Once formulated, the system of equations is solved using a least-squares algorithm.
To check the validity of the chosen model, the parameters determined from the
least-squares calculation are used with Eq. 24 to simulate the system response to the
force f (t). The simulated response is compared with the measured response (i.e.,
displacement, velocity, and/or acceleration). A close match between the measured
and simulated response indicates that the chosen model captures the dynamics of
the system. Conversely, large discrepancies indicate that the assumed form of the
model is inadequate, and the equations of motion need to be revised.

4.3 Reverse Path

The reverse path method is a frequency domain method that can be used to
decompose a system’s response into an underlying linear system and an uncorrelated
nonlinear system. This decomposition allows unbiased estimates of the linear
parameters of the system along with estimates for the nonlinear parameters. Bias
in linear parameters will usually occur if the parameters are estimated directly
from nonlinear dynamic response measurements without taking into account the
nonlinearity. The bias occurs because the frequency response functions (FRFs)
contain information about both the linear and nonlinear parts of the system. The
goal of the reverse path method is to separate the linear and nonlinear components
of the system response so that the linear parameters can be estimated without
“contamination” from the nonlinearities in the system.
The reverse path method was initially proposed in [3] for single degree of
freedom systems and was extended to multi-degree of freedom systems in [12]. The
multi-degree of freedom approach is typically referred to as conditioned reverse
path because the approach is based on conditioned spectra and FRFs, which will
17 Nonlinear System Analysis Methods 979

be defined below. Consider the system with N degrees of freedom defined by the
equations of motion

Nn
[M]{ẍ(t)} + [C]{ẋ(t)} + [K]{x(t)} + [Ai ]{xni (t)} = {f (t)} (25)
i=1

where [M],[C], and [K] are the N xN (linear) mass, damping, and stiffness matrices,
x(t) is the Nx1 displacement vector, f (t) is the N x1 vector of input forces, and
xni (t) is a qi x1 vector that contains the nonlinear function(s) that describes the
i th type of nonlinearity. There are Nn types of nonlinearities (e.g., cubic stiffness,
quadratic damping, etc.). The i th type of nonlinearity occurs at qi different locations
within the system. [Ai ] is an Nxqi matrix containing the coefficients of the
corresponding nonlinear function vector ({xni (t)}). Note that each of the xn vectors
can be different sizes. Accordingly, each of the [Ai ] matrices can be different sizes
to match the length of {xni }.
For example, consider a system that has one cubic and one quadratic spring
between two degrees of freedom and a second cubic spring between two other
degrees of freedom. In this example, Nn = 2 because there are two types of
nonlinearities (cubic stiffness and quadratic stiffness). Arbitrarily taking the cubic
stiffness to be the first nonlinearity (i = 1), q1 = 2 because there are two separate
cubic stiffness nonlinearities. Because the quadratic stiffness occurs at only one
location, q2 is equal to 1.
To formulate the reverse path form of Eq. 25, the Fourier transform of the
equation is taken, and the terms are rearranged such that the force appears on the
left-hand side of the equation:

Nn
{F (ω)} = [B(ω)]{X(ω)} + [Ai ]{Xni (ω)} (26)
i=1

where [B] = −ω2 [M] + j ω[C] + [K]. Note that [B] is the reciprocal of the
frequency response matrix, [H ]. By placing the force on the left-hand side and
the internal forces of the system on the right-hand side, the standard “path” of the
system has been reversed: applied forces are now outputs, and response vectors {X}
and {Xni } are now inputs. As will be shown below, this formulation allows for the
system to be recast in terms of uncorrelated inputs, which allows the underlying
linear system to be distinguished from the nonlinear portion of the system.
Extracting the linear portion of the response is a recursive procedure. First, the
portion of the total response that is correlated to the first nonlinear response is
removed. Then, the portion of the remaining response that is correlated to the second
nonlinearity is removed. This process continues until the portion of the remaining
response that is correlated with the last nonlinearity is removed. The remaining
response is uncorrelated with the system’s nonlinearities and is thus assumed
to be the linear portion of the total response. The portion of the force that is
980 J. J. Meyer et al.

Xn1 F(+1) F
L1F ∑

Xn2(-1) F(+2) F(-1)

L2F ∑

Xn3(-1:2) F(+3) F(-1:2)

L3F ∑

F(-1:3)
••
Xni(-1:i-1) F(+i) •
LiF ∑ F(-1:i-1)
F(-1:i)
••
•
XnNn(-1:Nn-1) F(+Nn)
LNnF ∑ F(-1:Nn-1)

X(-1:Nn) F(+X) F(-1:Nn)

B

Fig. 12 Conditioned reverse path system model with uncorrelated inputs

uncorrelated with the nonlinearities in the system can be similarly identified.

Figure 12 shows the reverse path model with uncorrelated inputs and outputs.
Subscripts in parentheses are used to denote whether the quantity is correlated (+) or
uncorrelated (−) with the response indicated. For quantities uncorrelated with more
than one response, colon notation is used to identify all the uncorrelated responses.
For example, Xn3(−1:2) indicates the portion of the third nonlinear response vector
that is uncorrelated with the first and second nonlinear response vectors. For each
uncorrelated input, a frequency response matrix [LiF ] relates the input to the
portion of the force that is correlated with that nonlinearity. After the effect of
each nonlinearity in the system is removed, the remaining response (X(−1:Nn ) ) and
force (F(+X) = F(−1:Nn ) ) can be used to estimate the linear frequency response
matrix, and linear parameters can be estimated using any linear parameter estimation
technique. The nonlinear parameters can be estimated recursively as will be shown
below.
In practice, instead of working with the response and force spectra directly, power
spectral density (PSD) matrices are used in a recursive algorithm in order to extract
the linear system. The outputs of the algorithm are the conditioned PSD matrices
necessary to estimate the linear frequency response matrix. Typically, either the
H1 or H2 estimator is used, and therefore the required quantities are GF F (−1:Nn )
and GF X(−1:Nn ) for the H1 estimator and GXF (−1:Nn ) and GXX(−1:Nn ) for the H2
estimator. Because the measured response(s) and force(s) are processed to remove
nonlinear components (i.e., conditioned) before being used to estimate the FRF, the
result is called the “conditioned” estimate. Calculating the conditioned estimates
begins by finding the “unconditioned” PSD matrices, which are calculated directly
from the measured data. The unconditioned PSD matrices are calculated by the
standard formula
17 Nonlinear System Analysis Methods 981

2
GP Q = E[P ∗ QT ] (27)
T

where P and Q can each be any one of the quantities {X},{F }, or {Xni }, T is
the total acquisition time, E[·] is the expected value (indicating that the mean of
multiple averages should be used in the calculation), (·)∗ is the complex conjugate,
and (·)T is the transpose. The algorithm starts by calculating the unconditioned PSD
matrices GXX , GXF , GF F , and for i, j = 1, 2, . . . , Nn , GXni Xni , GXni Xnj , GXni X ,
and GXni F .
Next, the conditioned PSD matrices are calculated recursively. In each step of
the recursion, the portions of the spectra that are correlated with the rth nonlinearity
are removed. Starting with r = 1, the conditioned PSD matrices, GXni Xnj (−1:r) , are
calculated using

[GXni Xnj (−1:r) ] = [GXni Xnj (−1:r−1) ]

− [GXni Xnr (−1:r−1) ][GXnr Xnr (−1:r−1) ]−1 [GXnr Xnj (−1:r−1) ] (28)

where r < i, j and [Gij (−1:0) ] refers to the unconditioned PSD matrix. To find
conditioned PSD matrices involving X or F (GXX(−1:r) , GXF (−1:r) , GXni X(−1:r) ,
etc.), the subscripts i and/or j are replaced with X or F in Eq. 28. The recursion
to calculate the conditioned PSD matrices continues until r = Nn . Then the linear
frequency response matrix is estimated using either of the following:

H1 estimator: H T = [G−1
F F (−1:Nn ) ][GF X(−1:Nn ) ]
(29)
H2 estimator: H T = [G−1
XF (−1:Nn ) ][GXX(−1:Nn ) ].

The final step in the conditioned reverse path analysis is to estimate the nonlinear
parameters of the system, which are contained in the [Ai ] matrices. A formula for
the [Ai ] matrices based on the conditioned PSD matrices can be derived from Eq. 26
∗
by transposing the equation, pre-multiplying by {Xni(−1:i−1) } and taking (2/T )E[·],
which yields

Nn
[GXni F (−1:i−1) ] = [GXni X(−1:i−1) ][B]T + [GXni Xnj (−1:i−1) ][Aj ]T . (30)
j =1

Note that in the summation, [GXni Xnj (−1:i−1) ] = 0 when j < i because the
i thnonlinear function vector is uncorrelated with the nonlinear function vectors
1 through i − 1. By pre-multiplying Eq. 30 by G−1 Xni Xni (−1:i−1) , post-multiplying
by [H ]T , and noting that the first term in the summation becomes [ATi ], the
following equation from which the nonlinear parameters can be determined is
derived:
982 J. J. Meyer et al.


[Ai ]T [H ]T = [G−1
Xni Xni (−1:i−1) [GXni F (−1:i−1) ][H ] − [GXni X(−1:i−1) ]
] T

Nn 
− [GXni Xnj (−1:i−1) ][Aj ]T [H ]T . (31)
j =i+1

Because [Ai ] is unknown, the left-hand side should be multiplied out symboli-
cally. The estimation process begins by identifying [ANn ] and continues backwards
to [A1 ]. The values of the parameters in each [Ai ] matrix are frequency dependent,
and, assuming the parameters are constants, their true values can be estimated by
taking the spectral mean.

4.3.1 Example: Panel with Disbond

Conditioned reverse path analysis was applied to the damaged panel. Time histories
of the response to a band limited random excitation were acquired using a laser
displacement transducer. The input force was applied using the modal shaker and
was measured using a force transducer. Twenty data sets, each 3 seconds long, were
acquired as the panel was excited by band limited random excitation. The input force
was generated by filtering Gaussian white noise with a low-pass filter having a cutoff
frequency of 5 kHz, which was selected based on the frequency range in which the
modes of interest occurred. Measurements were transformed from the time domain
into the frequency domain via the DFT. No averaging of the data is performed
prior to applying the reverse path analysis. The choice to measure displacement
was made to avoid having to perform numerical differentiation in order to calculate
the nonlinear function vectors.
A single, cubic stiffness nonlinearity that depends only on the displacement at a
point measured near the disbond was assumed. Therefore, N = 1, {X} = X,Nn = 1,
[A1 ] = μ, where μ is the coefficient of the cubic stiffness, {Xn1 } = F[x(t)3 ]
(where F is the Fourier transform), and {F } = F . Data was assembled into a single
3 × Nf × Nave matrix, where Nf is the number of frequency points in the spectra
and Nave is the number of data sets acquired. In general, the first N columns will
be the response spectra, the next Nu columns will be the force spectra (where Nu
is the number of inputs used), and the last Nn columns will be the spectra of the
nonlinear function vectors. This assembly simplifies the coding required to calculate
the unconditioned and conditioned spectra. Iterative loops were used to calculate the
unconditioned PSD matrices according to Eq. 27. Expected values were found by
averaging the PSD matrices from each of the twenty data sets. Next, iterative loops
were used to find the conditioned PSD matrices according to Eq. 28.
Once the conditioned PSD matrices were determined, an estimate for the linear
FRF was made using the H2 estimator in Eq. 29. Figure 13 shows the estimate
(dashed blue line) along with the FRF of the nonlinear system (solid black line)
found by estimating the FRF directly from the measured data (i.e., with no
conditioning). For reference, the FRF determined from a low amplitude sine sweep
is shown (red dotted line), which is meant to represent the linear FRF of the panel.
17 Nonlinear System Analysis Methods 983

0.06
0.06
0.05
0.055

|FRF|
|FRF|

0.04
0.05
0.03
0.045
0.02

0.01 0.04
300 350 400 450 500 550 600 400 405 410 415 420 425 430
Frequency (Hz) Frequency (Hz)
(a) (b)

Fig. 13 Magnitudes of FRFs of the panel with disbond: FRF of the nonlinear system ( ), FRF
of the underlying linear system estimated using the conditioned reverse path method ( ), and
FRF of the panel excited by a low amplitude sine sweep ( ). Figure (b) zooms in on the peak
near 420 Hz

Fig. 14 The nonlinear

parameter μ as a function of
frequency from conditioned 105
reverse path analysis
|µ|

100

200 300 400 500 600 700

Frequency (Hz)

The linear FRF estimated by the conditioned reverse path analysis shows good
agreement with the FRF determined in the selected frequency range from the low
amplitude sine sweep.
The last step in the analysis was to estimate the nonlinear coefficient, μ(ω),
which is the only element in the [A1 ] matrix, using Eq. 31. Figure 14 shows the
μ as a function of frequency. Because the parameter is assumed to be a constant, the
estimated value is found by taking the mean over the frequency range of interest.

4.4 Nonlinear Identification Through Feedback of the Outputs

(NIFO)

The nonlinear identification through feedback of the outputs (NIFO) technique is a

frequency domain nonlinear system identification technique which utilizes spatial
information to estimate nonlinear parameters in a simple solution approach [1]. If
the location and form of the nonlinearity is understood, this technique can readily
estimate the nonlinear parameters and separate the linear and nonlinear dynamics.
984 J. J. Meyer et al.

F(ω) Linear X(ω)

+
- System

Nonlinear
System

Fig. 15 Nonlinear feedback system diagram

The formulation of the nonlinear system in this approach is an underlying linear

system with a nonlinear feedback loop. This idea is represented in Fig. 15.
The advantage to this formulation is that the overall response can be considered
as a multiple-input system, with the inputs to the system being a combination of true
forcing functions and nonlinear functions of the outputs. Furthermore, the portion of
the response due to the nonlinearity is a function of the underlying linear frequency
response function. In this form, the coefficient of the nonlinearity can be identified
in a single step along with the frequency response of the underlying system. For
a system with Ni measured inputs, No measured outputs, Navg data sets, and Nn
nonlinearities, the response of the system, {X(ω)}, can be expressed in terms of
the linear frequency response matrix, [HL (ω)], the coefficients of nonlinearities,
μn (ω), the vectors that describes the degree of freedom at which the i th nonlinear
force is acting, {Bni }, the vector of forcing functions, {F (ω)}, and the nonlinear
displacement associated with the i th nonlinearity, {Xni (ω)} as

[X(ω)]No x1 = [HL (ω)] [HL (ω)]μ1 (ω){Bn1 } ... [HL (ω)]μNn (ω){BnNn } (No x(Ni +Nn ))

{F (ω)}Ni x1
× . (32)
−{Xn (ω)}Nn x1 ((N +N )x1)
i n

Equation 32 is the same form as a multi-input, multi-output (MIMO) frequency

response function (FRF) relationship, where there are No output measurements and
Ni + Nn input measurements, with the only difference being that Nn of these inputs
are a function of the system response. To simplify notation, the FRF matrix above
can be renamed as

[HN L ] = [[HL ] [HL ]μ1 {Bn1 } ... [HL (ω)]μNn (ω){BnNn }],

and the rightmost vector {FN L } = [{F }, −{Xn }]T . Making these substitutions,
Eq. 32 becomes

[X]No x1 = HN L (No x(Ni +Nn ))

FN L ((Ni +Nn )x1)
(33)
17 Nonlinear System Analysis Methods 985

Treating {FN L } as a standard forcing vector and [HN L ] as a standard FRF matrix,
[HN L ] can be estimated using standard MIMO FRF estimation techniques. In this
case, the H2 FRF estimator considering several data sets is used. Once the matrix
[HN L ] is estimated, the underlying linear frequency response function matrix and
the coefficient of nonlinearity can be found in a single step. The first Ni columns
of [HN L ] are [HL ], and the last Nn columns are equal to the product [HL ]μi {Bni },
i = 1..Nn .
The strength of this technique is the simplicity with which it decouples the
system’s linear and nonlinear response characteristics if the location and form
of the nonlinearity is well understood. Unlike methods such as the conditioned
reverse path method, estimation of the linear system and the nonlinear parameters
are accomplished simultaneously. The simple solution approach makes NIFO an
attractive method for multi-degree of freedom systems with known nonlinearities.

4.4.1 Example: Panel with Disbond

NIFO analysis was applied to the damaged panel using measured displacement data
at a point that coincided with the location of the disbond and force measurements
from the impedance head on the shaker used to excite the panel. Twenty data sets,
each 3 seconds long, were acquired as the panel was excited by band limited random
excitation. The input force was generated by filtering Gaussian white noise with a
low-pass filter with a cutoff frequency of 5 kHz, which was selected based on the
frequency range in which the modes of interest occurred. In this example, only one
input and one output location were used. However, the NIFO formulation easily
accommodates the addition of more input and/or output locations.
The first step in formulating the NIFO analysis is to identify the location(s)
and type(s) of nonlinearity (ies). In this example, a cubic stiffness nonlinearity
that depends only on the displacement measured at the location of the disbond is
assumed. Therefore, Nn = 1, {Xn (ω)} = Xn1 (ω) = F[x(t)3 ] (where F is the
Fourier transform), μ1 is the coefficient of Xn1 , and {Bn1 } = 1. If multiple output
(i.e., measurement) locations had been used, zeros would be added to the B vector
in all the rows except the row corresponding to the location of the nonlinearity. In
the case that a nonlinearity is assumed that acts between two output locations, the
B vector contains a one in the row corresponding to one of the locations, a negative
one in the row corresponding to the second location, and zeros in all other rows.
After determining the form and location of the nonlinearity, X(ω) and F (ω)
were calculated by transforming the measured time histories into frequency spectra
via the DFT. Then, Xn (ω) was calculated by transforming x(t)3 into the frequency
domain via the DFT. These quantities were used in Eq. 33 in order to solve for
[HN L (ω)]. Figure 16 shows a comparison of FRF estimates of the system. The solid
black line shows the FRF estimated from the measured input/output data. This is
typically referred to as the FRF of the nonlinear system. The dashed blue line shows
HL (ω), the estimate of the FRF of the underlying linear system. For comparison,
the red dotted line is an FRF determined from a sine sweep performed at a low
amplitude, which estimates the linear FRF. The linear FRF (HL (ω)) estimated from
986 J. J. Meyer et al.

0.07 0.06
0.06
0.055
0.05
0.04 0.05

| FRF|
|FRF|

0.03 0.045
0.02
0.04
0.01
0 0.035
300 350 400 450 500 550 400 405 410 415 420 425 430
Frequency (Hz) Frequency (Hz)
(a) (b)

Fig. 16 Magnitudes of FRFs of the panel with disbond: FRF of the nonlinear system ( ), FRF
of the underlying linear system estimated using the NIFO method ( ), and FRF of the panel
excited by a low amplitude sine sweep ( ). Figure (b) zooms in on the peak near 420 Hz

Fig. 17 The nonlinear

parameter μ as a function of
frequency from NIFO 105
analysis
|µ|

100

200 300 400 500 600 700

Frequency (Hz)

the NIFO analysis shows good agreement with the FRF determined from the low
amplitude sine sweep.
Finally, μ1 (ω) was determined from the second column of [HN L ]. Figure 17
shows the parameter as a function of frequency. The parameter is assumed to be a
constant, so the estimate would be determined by taking the mean over the frequency
range of interest.

5 Summary

The nonlinear analysis methods presented in this chapter are a representative subset
of the methods available to the experimentalist for detecting, characterizing, and
modeling a system’s nonlinear dynamics. Table 2 summarizes the requirements to
implement each method. The principles demonstrated here provide a fundamental
basis upon which most nonlinear analysis approaches, including methods not
presented here, are derived. These principles and guidelines provide a powerful tool-
box for moving beyond linear systems analysis and experimentally characterizing
nonlinear dynamic structures.
17 Nonlinear System Analysis Methods 987

Table 2 Summary of nonlinear analysis methods

Methods for detection and characterization
Method Output data required Excitation Comments
required
Coherence FRF using any type of data Random, sine
sweep, impact,
etc.
FRF distortion FRF using any type of data Random, sine Must acquire data at multiple
sweep, impact, input levels
etc.
Higher-order Any output type (response Stepped sine For nth order FRF, 1 ≤ 1/n
FRFs spectra required) from 1 to 2 lowest modal frequency of
interest; 2 > highest modal
frequency of interest
Hilbert Any output type (acquire Impact Other inputs can be used, but
transform (time time histories) a forced response
domain) necessitates an estimate of
the system’s mass
Hilbert FRF using any type of data Random, sine
transform sweep, impact,
(frequency etc.
domain)
Restoring Displacement, velocity, and Random, Non-measured outputs can be
forces acceleration (acquire time multi- estimated by numerically
histories) amplitude integrating and/or
sinusoids differentiating measured
outputs input must generate
outputs that densely cover the
displacement-velocity plane.
Vibro-acoustic Any output type (response Low-frequency Low-frequency signal should
modulation spectra required) sinusoid and be chosen to match a natural
high-frequency frequency low-frequency
sinusoidal force can be replaced by an
impact.
Methods for modeling
NARMAX Any output type (acquire Random Other input types may also
time histories) be acceptable.
Direct Displacement, velocity, and Random Non-measured outputs can be
parameter acceleration (acquire time estimated by numerically
estimation histories) integrating and/or
differentiating measured
outputs other input types may
also be acceptable.
Conditioned Displacement. Velocity Random, sine Form of nonlinearity must be
reverse path only if damping sweep assumed a priori
nonlinearity is assumed
(response spectra required).
NIFO Displacement. Velocity Random, sine Form of nonlinearity must be
only if damping sweep assumed a priori
nonlinearity is assumed
(response spectra required).
988 J. J. Meyer et al.

References
1. Adams D, Allemang R (2000) A frequency domain method for estimating the parameters of a
non-linear structural dynamic model through feedback. Mech Syst Signal Process 14(4):637–
656
2. Bendat JS, Piersol AG (1980) Engineering applications of correlation and spectral analysis.
Wiley-Interscience, New York, p 315
3. Bendat J, Palo P, Coppolino R (1992) A general identification technique for nonlinear
differential equations of motion. Probab Eng Mech 7(1):43–61
4. Billings S, Tsang K (1989) Spectral analysis for non-linear systems, part I: parametric non-
linear spectral analysis. Mech Syst Signal Process 3(4):319–339
5. Chen S, Billings SA (1989) Representations of non-linear systems: the narmax model. Int J
Control 49(3):1013–1032
6. Donskoy D, Sutin A, Ekimov A (2001) Nonlinear acoustic interaction on contact interfaces
and its use for nondestructive testing. NDT & E Int 34(4):231–238
7. Feldman M (1994) Non-linear system vibration analysis using hilbert transform–I. Free
vibration analysis method ’Freevib’. Mech Syst Signal Process 8(2):119–127
8. Feldman M (1994) Non-linear system vibration analysis using hilbert transform–II. Forced
vibration analysis method ’Forcevib’. Mech Syst Signal Process 8(3):309–318
9. Feldman M (2011) Hilbert transform in vibration analysis. Mech Syst Signal Process
25(3):735–802
10. Meyer JJ, Adams DE (2015) Theoretical and experimental evidence for using impact
modulation to assess bolted joints. Nonlinear Dyn 81(1–2):103–117
11. Mohammad K, Worden K, Tomlinson G (1992) Direct parameter estimation for linear and
non-linear structures. J Sound Vib 152(3):471–499
12. Richards C, Singh R (1998) Identification of multi-degree-of-freedom non-linear systems
under random excitations by the reverse path spectral method. J Sound Vib 213(4):673–708
13. Storer D, Tomlinson G (1993) Recent developments in the measurement and interpretation of
higher order transfer functions from non-linear structures. Mech Syst Signal Process 7(2):173–
189
14. Tomlinson G, Worden K (2000) Nonlinearity in structural dynamics: detection, identification
and modelling. Boca Rotan, CRC Press. https://doi.org/10.1201/9780429138331
15. Worden K (1990) Data processing and experiment design for the restoring force surface
method, part II: choice of excitation signal. Mech Syst Signal Process 4(4):321–344
Structural Health Monitoring and Damage
Identification 18
R. Fuentes, E. J. Cross, P. A. Gardner, L. A. Bull, T. J. Rogers,
R. J. Barthorpe, H. Shi, N. Dervilis, C. R. Farrar, and K. Worden

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 991
1.1 Motivation and Definition of SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 991
1.2 Statistical Pattern Recognition Approach to SHM . . . . . . . . . . . . . . . . . . . . . . . . . . 992
1.3 Fundamental Axioms of SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 995
1.4 Historical Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 997
2 SHM Data and Damage-Sensitive Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
2.1 Vibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
2.2 Acoustic Emissions (AE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1011
2.3 Guided Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
2.4 Performance, Operational, and Environmental Parameters . . . . . . . . . . . . . . . . . . . 1015
3 Advanced Topics in Signal Processing and Feature Extraction . . . . . . . . . . . . . . . . . . . . 1016
3.1 Inference via the Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1020
3.2 Recursive Estimation in the Nonlinear Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1023
4 Statistical Pattern Recognition for Damage Identification . . . . . . . . . . . . . . . . . . . . . . . . 1026
4.1 Pattern Recognition for Feature Discrimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026
4.2 Data-Driven Models in SHM: Learning and Prediction . . . . . . . . . . . . . . . . . . . . . . 1027
4.3 Outlier Analysis for Damage Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1028
4.4 The Problem of Inclusive Outliers: Robust Outlier Analysis . . . . . . . . . . . . . . . . . . 1031
4.5 Probabilistic Classification through Supervised Learning . . . . . . . . . . . . . . . . . . . . 1032
4.6 The Problem of Feature Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035
4.7 Outstanding Challenges in Data-Driven SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036
5 SHM in Changing Environmental and Operational Conditions . . . . . . . . . . . . . . . . . . . . 1036
5.1 Removing Confounding Influences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1038

R. Fuentes · E. J. Cross · P. A. Gardner · L. A. Bull · T. J. Rogers · R. J. Barthorpe · H. Shi

N. Dervilis · K. Worden ()
Dynamics Research Group, Department of Mechanical Engineering, University of Sheffield,
Sheffield, UK
e-mail: [email protected]; [email protected]
C. R. Farrar
Engineering Institute, Los Alamos National Laboratory, Los Alamos, NM, USA

© The Society for Experimental Mechanics 2022 989

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_23
990 R. Fuentes et al.

5.2 Linear and Nonlinear Cointegration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1041

5.3 Nonlinear Cointegration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1046
6 Physics-Based Models in SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
6.1 Inverse Model-Driven SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
6.2 Forward Model-Driven SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1049
7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
7.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1052

Abstract

Structural dynamics is fundamentally concerned with the design, operation, and

understanding of physical structures. A significant concern in the management
of these, often very high-value, assets, is their state of health. When a structure
sustains damage, this can have an extremely negative effect on its availability,
and this will have serious implications for profitability and also the safety
of any human operators or occupants. It is therefore important to implement
some means of monitoring structural health so that incipient damage can be
detected and remedial actions can be taken before negative consequences occur.
The pertinent damage identification methodology for engineering structures
is Structural Health Monitoring (SHM). This chapter presents an overview
of SHM, with particular reference to implementations based on monitoring
structural vibrations and waves. The main philosophy under discussion here is
data-based SHM, where diagnostics are based on the interpretation of measured
data directly, without recourse to physics-based models. The main technologies
for carrying out data-based SHM are statistical pattern recognition and machine
learning, and the chapter gives some background on these methods and provides
some case studies to illustrate their use. One of the main approaches to damage
detection is novelty detection, where one develops a statistical model of measured
features from the undamaged structure of interest, and monitors subsequent data
to see if there are deviations from the model, indicative of damage. A serious
problem with this approach is that it is prone to false alarms if there are benign
changes to the data, like operational or environmental variations. Such benign
changes – referred to here as confounding influences, need to be compensated for,
if the SHM system is to be reliable and error-free (as far as possible). The chapter
considers how confounding influences arise, and how they can be removed in the
data-driven context by data normalization. Finally, the chapter concludes with
some discussion of how physics-based models can still have a potentially useful
role in data-driven SHM.

Keywords

Structural Health Monitoring (SHM) · Data-based and model-based SHM ·

Statistical pattern recognition · Machine learning · Confounding influences ·
Data normalization
18 Structural Health Monitoring and Damage Identification 991

1 Introduction

1.1 Motivation and Definition of SHM

In brief, Structural Health Monitoring (SHM) is any automated monitoring practice

that seeks to assess the condition or health of a structure. In terms of a more
formal definition, one might adopt: “SHM is the process of implementing a
damage identification strategy for aerospace, civil and mechanical engineering
infrastructure, where damage is defined as changes to the material and/or geometric
properties of these systems, including changes to the boundary conditions and
system connectivity, which adversely affect the systems performance” [1, 2].
The origins of SHM in the context of engineering can arguably be traced back
as far as the time when tap-testing for fault detection became common, although
the field did not really become established in research communities until the 1980s,
when considerable interest was generated in the structural assessment of oil rigs,
and later in aerospace structures and their health [3]. SHM has now established itself
as a popular and still growing research field, which is more and more becoming a
focus in terms of structural asset management within the engineering community.
As Wenzel has observed [4], bridges (which are arguably the paradigm for critical
infrastructure) have always been monitored, with this activity varying between
visual inspection and continuous monitoring with dense sensor networks.
At its heart, SHM has significant aspirations that, once achieved, will hugely
benefit society; it strives toward an ideal where one is able to monitor (in an
automated fashion) a structure in such a way that any damage introduced, or
any growth of inherent faults, would be immediately detectable, or detectable in
time to plan and apply remedial actions which have minimal impact on operation.
Furthermore, after detection, it is desirable that any fault could be located and its
severity inferred so that decisions on remedial actions are optimal. These global
objectives for SHM are formalized in Rytter’s hierarchy [5], which classifies these
aims into “levels” of increasing difficulty as follows:

Level 1 (DETECTION): The method gives a qualitative indication that damage

might be present in the structure.
Level 2 (LOCALIZATION): The method gives information about the probable
position of the damage.
Level 3 (CLASSIFICATION): The method indicates what type of damage is
present.
Level 3 (ASSESSMENT): The method gives an estimate of the extent of the
damage.
Level 4 (PREDICTION): The method offers information about the safety of the
structure, for example, estimates a residual life.

At this point, it is perhaps important to stress what SHM is not; for disambigua-
tion, it is important to discuss the differences between SHM and related fields
associated with damage identification (a more detailed discussion can be found
992 R. Fuentes et al.

in [6]). Although tap-testing was mentioned earlier in terms of the origins of

SHM, the example more properly belongs to the field of Non-Destructive Test-
ing/Evaluation (NDT/NDE). NDE/NDT concerns the assessment of a structure
or component’s health through (offline) non-damaging procedures; examples of
tools commonly used for NDE are: x-ray, electron microscopy, measurement of
acoustic emissions, and full scale vibration tests. In fact, all of these techniques
supplement SHM technology; however, NDE is distinguished by its application as
one-off planned events, often applied to preestablished small areas of a structure
where the presence of damage is suspected. This is fundamentally different to SHM
where monitoring is usually intended to be continuous and global, and is usually
understood to require permanently installed sensors. In the future, it is likely that
NDE inspection will form the basis for distinguishing between structural health
and performance anomalies where this cannot be accomplished automatically. It
is therefore true to say that NDE may be incorporated as part of an SHM system,
but not vice versa. Another field related to SHM is Condition Monitoring (CM). CM
usually concerns the health of rotating machinery; it has seen been widely applied
and in some areas, accepted as part of everyday industrial practice. Its comparative
success (in terms of industrial uptake) relative to SHM can be attributed to a number
of simplifying factors: machinery often operates in a controlled environment, and it
is usually easy to access and is typically on a small scale. Importantly, rotating
machinery often have well-defined dynamic responses for particular fault categories,
which makes fault detection and identification a more easily attainable goal than it
perhaps is for SHM [7].

1.2 Statistical Pattern Recognition Approach to SHM

There are potentially many ways to organize a discussion of SHM. The approach
here follows [8], in defining the SHM process in terms of a four-step statistical
pattern recognition paradigm. This following four-step process includes [2]:

1. Operational evaluation
2. Data acquisition, normalization, and cleansing
3. Feature selection and information condensation
4. Statistical model development for feature discrimination

All studies in the fields of SHM and CM address some parts of this paradigm,
but the number of studies that address all portions of the paradigm is very limited.

1.2.1 Operational Evaluation

Operational Evaluation (OE) attempts to answer four questions regarding the
implementation of an SHM capability.

1. What are the life-safety and/or economic justifications for performing SHM?
2. How is damage defined for the system being investigated and, for multiple
damage possibilities, which cases are of the most concern?
18 Structural Health Monitoring and Damage Identification 993

3. What are the conditions, both operational and environmental, under which the
system to be monitored functions?
4. What are the limitations on acquiring data in the operational environment?

OE thus attempts to set the limitations on what will be monitored and how the
monitoring will be accomplished; it attempts to tailor the SHM process to features
that are unique to the system being monitored and tries to exploit unique features of
the damage that is to be detected.

1.2.2 Data Acquisition, Normalization, and Cleansing

The data acquisition part of the SHM process involves selecting the excita-
tion methods, the sensor types, number and locations, and the data acquisi-
tion/storage/transmittal hardware. Again, this process will be application specific.
Cost (time/effort) will play a major role in making these decisions. The frequency
with which the data should be collected is another consideration that must be
addressed.
As data can be measured under varying conditions, the ability to normalize the
data becomes very important in the SHM process. Data normalization is the process
of separating changes in sensor readings caused by damage from those caused by
varying operational and environmental conditions, which are benign and must not
raise an alarm. Where environmental or operational variations are an issue, one may
need to normalize the data in some temporal fashion to facilitate the comparison
of data measured at different points in an environmental or operational cycle.
Sources of variability in the data acquisition process and with the system being
monitored need to be identified and minimized. In general, not all of such sources
can be eliminated; therefore, it is necessary to make the appropriate measurements
such that these sources can be statistically quantified. Variability can arise from
changing environmental and test conditions, changes in the data reduction process
and unit-to-unit inconsistencies [2]. Section 5 of this chapter provides a more in-
depth discussion of the various methods available for dealing with environmental
and operational changes.
Data cleansing is the process of selectively choosing data to pass on to, or
reject from, the feature selection process. The data cleansing process is usually
based on the knowledge gained by individuals directly involved with the data
acquisition. For example, an inspection of the test setup may reveal that a sensor was
loosely mounted, and, hence, based on the judgment of the individuals performing
the measurement, this set of data or the data from that particular sensor may be
selectively deleted from the feature selection process. Signal processing techniques
such as filtering and resampling can also be thought of as data cleansing procedures.

1.2.3 Feature Extraction and Information Condensation

The part of the SHM process that usually receives most attention in the technical lit-
erature is the identification of data features that allow one to distinguish between the
undamaged and damaged structure. As such, numerous articles have been devoted
to the feature extraction portion of SHM [9, 10, 11, 12]. Inherent in this feature
994 R. Fuentes et al.

selection process is the condensation of the data. The best features for damage
identification are usually application specific. One of the most common feature
extraction methods is based on correlating measured system response quantities,
such as vibration amplitude or frequency, with the first-hand observations of the
degrading system. Another method of developing features for damage identification
is to apply engineered flaws – proxies – similar to ones expected in actual operating
conditions, to systems and develop an initial understanding of the parameters that
are sensitive to the expected damage [13, 14].
The flawed system can also be used to validate that the diagnostic measurements
are sensitive enough to distinguish between features identified from the undamaged
and damaged system. The use of analytical tools such as experimentally validated
finite element models can be a great asset in this process. In many cases, the
analytical tools are used to perform numerical experiments where the flaws are
introduced through computer simulation. Damage accumulation testing, during
which significant structural components of the system under study are degraded
by subjecting them to realistic loading conditions, can also be used to identify
appropriate features. This process may involve induced damage testing, fatigue
testing, corrosion growth, or temperature cycling to accumulate certain types of
damage in an accelerated fashion. Insight into the appropriate features can be gained
from several types of analytical and experimental studies as described above, and is
usually the result of information obtained from some combination of these studies.
The operational implementation and diagnostic measurement technologies
needed to perform SHM produce more data than traditional uses of structural
dynamics information. A condensation of the data is advantageous and necessary
when comparisons of many feature sets obtained over the lifetime of the structure
are envisioned. Also, because data will be acquired from a structure over an
extended period of time and in an operational environment [2], robust data reduction
techniques must be developed to retain feature sensitivity to the structural changes
of interest in the presence of environmental and operational variability. To further
aid in the extraction and recording of the high-quality data needed to perform SHM,
the statistical significance of the features should be characterized and used in the
condensation process. A discussion of some of the most popular damage-sensitive
features is provided in Sect. 2 of this chapter, while Sect. 3 discusses some of the
more advanced techniques that can be applied to problems such as data compression
and recursive and online estimation of features.

1.2.4 Statistical Model Development

The final part of the SHM process is the development of statistical models for
discrimination between features from the undamaged and damaged structures.
Statistical model development is concerned with implementing algorithms that
operate on the extracted features to qualify/quantify the damage state of the
structure. The algorithms usually fall into three categories. When data are available
from both the undamaged and damaged structure, the algorithms fall into the general
class referred to as supervised learning. Classification and regression analysis
are categories of supervised learning algorithms. Unsupervised learning refers to
18 Structural Health Monitoring and Damage Identification 995

algorithms that are applied to data only containing examples from the undamaged
structure; outlier or novelty detectors are the primary forms of algorithms applied
in unsupervised learning applications. All of the algorithms implicitly or explicitly
analyze probability distributions of the measured or derived features to enhance the
damage identification process.
The damage identification of a system can be described as a five-step pro-
cess along the lines of Rytter’s hierarchy as discussed above. The answers to
the questions posed represent increasing knowledge of the damage state. When
applied in an unsupervised learning mode, statistical models are typically used to
answer questions regarding the existence and location of damage. When applied
in a supervised learning mode and coupled with analytical models, the statistical
procedures can be used to better determine the type of damage, the extent of damage
and remaining useful life of the structure (Although prognosis is distinguished from
the other SHM problems, in that it cannot be accomplished in a data-based fashion
alone, physics-based models of the relevant damage mechanisms and processes
are required.). The statistical models are also used to minimize false indications
of damage. False indications of damage fall into two categories: (i) false-positive
damage indication (indication of damage when none is present) and (ii) false-
negative damage indication (no indication of damage when damage is present).
Errors of the first type are undesirable, as they will cause unnecessary downtime
and consequent loss of revenue, as well as loss of confidence in the monitoring
system. More importantly, there are clear safety issues if misclassifications of the
second type occur. Many pattern recognition algorithms allow one to weigh one
type of error above the other; this weighting may be one of the factors decided at
the operational evaluation [2] stage. Reviews that focus on the statistical modeling
portion of the SHM process include [15, 16, 17, 18].

1.3 Fundamental Axioms of SHM

In terms of practical progress toward the aims and objectives of SHM, a great deal of
experience and knowledge has already been gained. Reflecting on work in the field
and the lessons learned over the years, a number of fundamental “axioms” have been
formulated [19]:

Axiom I: All materials have inherent flaws or defects.

Axiom II: The assessment of damage requires a comparison between two system
states.
Axiom III: Identifying the existence and location of damage can be done in an
unsupervised learning mode, but identifying the type of damage present and the
damage severity can generally only be done in a supervised learning mode.
Axiom IVa: Sensors cannot measure damage. Feature extraction through signal
processing and statistical classification are necessary to convert sensor data into
damage information.
996 R. Fuentes et al.

Axiom IVb: Without intelligent feature extraction, the more sensitive a mea-
surement is to damage, the more sensitive it is to changing operational and
environmental conditions.
Axiom V: The length and time scales associated with damage initiation and
evolution dictate the required properties of the SHM sensing system.
Axiom VI: There is a trade-off between the sensitivity to damage of an algorithm
and its noise rejection capability.
Axiom VII: The size of damage that can be detected from changes in system
dynamics is inversely proportional to the frequency range of excitation.

The axioms are intended to help to specify an SHM system in practice. They were
originally proposed with the intention of stimulating discussion; however, despite
the fact that the original paper has been cited many times, there has been very little
dissent. In 2010, an Axiom VIII was proposed: “Damage increases the complexity
of the structure” [20]. The new axiom appears to be a little more nuanced than
the others and has generated interesting research on the nature of complexity as an
indicator of health.
Where there has been some critical discussion, concerns Axiom II, with a number
of authors arguing that “baseline-free” SHM is possible. The counterargument here
is that the “baseline-free” methods are always carried out with respect to some
idealization, whether this is a model or a conceptualization, and the “comparison”
in the Axiom is thus implicitly present.
The axioms have greater or lesser importance, depending on the specific SHM
context, for example, Axiom III provides some interesting points for discussion in
the context of civil infrastructure SHM. Because only the lowest global modes are
generally excited in practice for say a bridge, the modes will usually involve large-
scale coherent motions. If the structure has a dense sensor network, a breakdown in
correlations between neighboring sensors may give information about the location
of damage. This fact means that localization of damage could be possible despite
the fact that data from damage states are not available. Severity of damage could
also be inferred from the extent of the breakdown of correlation. Another possibility
is that within the highly heterogeneous structure that is a bridge, changes in local
modes, for example, those of cables, may also allow some localization of damage.
Although one could discuss each axiom in turn, the section will conclude
here with some comments on Axiom V. Axiom V is the reason why operational
evaluation is a critical stage in the development of any SHM system. The design
of the sensor system must be tailored to the time and length scales on which one
expects damage to accumulate. In the context of certain aerospace alloys, a critical
crack size smaller than a millimeter means that confident detection of the crack
almost immediately after initiation is critical. In the context of civil infrastructure,
it will usually be the case that damage will accumulate on longer time scales and
will have a greater spatial extent before catastrophic failure would occur. Even so,
these anticipated time and length scales must be considered very carefully in the
specification for the monitoring system; because civil structures are so much larger,
for confident detection sensor densities may well need to be relatively high.
18 Structural Health Monitoring and Damage Identification 997

1.4 Historical Overview

A detailed historical overview definitely falls outside the scope of this chapter, but it
is worth pointing out some of the key developments over the last 40 years. Specific
references are not cited, instead the reader is referred to [3, 16, 7] and [2] for more
detailed summaries of this subject.
To date, the most successful application of SHM technology has been for CM of
rotating machinery [7], as this is particularly well suited to the data-driven paradigm.
Often this pattern recognition is performed only in a qualitative manner based on a
visual comparison of features such as frequency spectra obtained from the system at
different times. For rotating machinery systems, the approximate damage location is
generally known, making a single-channel fast Fourier transform analyzer sufficient
for most periodic monitoring activities. Typical damage that can be identified
includes loose or damaged bearings, misaligned shafts, and chipped gear teeth.
The success of CM is due in part to: (i) minimal operational and environmental
variability associated with this type of monitoring, (ii) well-defined damage types
that occur at known locations, (iii) large databases that include data from damaged
systems, (iv) well-established correlation between damage and features extracted
from the measured data, and (v) clear and quantifiable economic benefits that
this technology can provide. These factors have allowed this application of SHM
to make the transition from a research topic to industry practice several decades
ago, resulting in comprehensive condition management systems such as Health and
Usage Monitoring Systems (HUMS), which are new standard in modern rotary-
wing aircraft.
During the 1970s and 1980s, the oil industry made considerable efforts to develop
vibration-based damage identification methods for offshore platforms. This damage
identification problem is fundamentally different from that of rotating machinery
because the damage location is unknown and the majority of the structure is not
readily accessible for measurement. A common methodology adopted by this indus-
try was to simulate candidate damage scenarios with numerical models, examine the
changes in resonance frequencies that were produced by these simulated changes,
and correlate these changes with those measured on a platform. A number of
very practical problems were encountered which prevented the adaptation of this
technology, and efforts at further developing it for offshore platforms were largely
abandoned in the early 1980s. The aerospace community began to study the use
of vibration-based damage identification during the late 1970s and early 1980s in
conjunction with the development of the space shuttle. This work has continued
with current applications being investigated for the National Aeronautics and
Space Administration’s (NASA) space station and future reusable launch vehicle
designs.
Since the mid-1990s, studies of damage identification for composite materials
have been motivated by the development of a composite fuel tank for a reusable
launch vehicle. The failure mechanisms, such as delamination caused by debris
impacts, and corresponding material response for composite fuel tanks are sig-
nificantly different to those associated with metallic structures. Moreover, the
998 R. Fuentes et al.

composite fuel tank problem presents challenges because the sensing systems must
not provide a spark source. This challenge has led to the development of SHM based
on fiber-optic sensing systems. Reference [21] provides a more detailed discussion
of SHM applied to aerospace structures.
The civil engineering community has studied vibration-based damage assess-
ment of bridge structures and buildings since the early 1980s. Modal properties and
quantities derived from them, such as mode shape curvatures and dynamic flexibility
matrix indices, have been the primary features used to identify damage in bridge
structures. Environmental and operating condition variability presents significant
challenges to the bridge monitoring application. The physical size of the structure
also presents many practical challenges for vibration-based damage assessment.
Regulatory requirements in Asian countries, which mandate that the companies
that construct the bridges periodically certify their structural health, are driving
current research and commercial development of bridge SHM systems. The reviews
of the technical literature presented by [3] and [22] show an increasing number
of research studies related to damage identification. These studies identify many
technical challenges to the adaptation of SHM that are common to all applications
of this technology. These challenges include the development of methods to
optimally define the number and location of the sensors; identification of the
features sensitive to small damage levels; the ability to discriminate changes in these
features caused by damage from those caused by changing environmental and/or
test conditions; the development of statistical methods to discriminate features from
undamaged and damaged structures; and performance of comparative studies of
different damage identification methods applied to common datasets. These topics
are currently the focus of various research efforts by many industries including:
defense, civil infrastructure, automotive, and semiconductor manufacturing where
multidisciplinary approaches are being used to advance the current capabilities of
SHM and CM.

2 SHM Data and Damage-Sensitive Features

This section will provide a general overview of the types of dynamic response data
typically used in SHM: vibration, acoustic emission (AE) and ultrasound, together
with the basic signal processing required to extract damage-sensitive features often
used in these contexts. Note that in the interest of focusing on the conceptual above
the mathematical, a lot of detail will be skipped; in particular, with respect to the
mathematics behind the signal processing context such as sampling and windowing.
For these details, the reader is encouraged to consult other excellent dedicated signal
processing texts such as the classic [23] or sections elsewhere in this book. The
focus of this section is to give an overview of the breadth of damage-sensitive
features and how they relate to the damage process across different scales and
domains.
18 Structural Health Monitoring and Damage Identification 999

2.1 Vibration

Vibration-based SHM is now a very popular technique for assessing the state of
an engineering system, with a significant part of SHM research and industrial
applications being devoted to using vibration measurements to infer the presence
and location of damage; this development dates back to the 1990s [3].
In SHM, vibration monitoring is based on the premise that an adverse change
to the structure will cause a change in the dynamic response, which should be
quantifiable using vibration measurements [19]. This change in the dynamics often
manifests itself in changes to measurable parameters such as mode shapes, natural
frequencies, or damping. Damage alters the structural response in ways that can be
captured through a range of damage-sensitive features. Vibration data can cover a
wide range of frequencies, and this is application dependent. Structures that vibrate
at very low frequencies include most civil infrastructure, such as bridges, buildings,
and stadia, as well as offshore oil rigs and wind turbines. The excitation source
for most civil structures tends to be its own environment. Wind, traffic loading,
and earth movement are common excitation sources in civil infrastructure. In wind
turbines, the low-frequency loading comes predominantly from the aerodynamic
loading caused by the rotation of the blades, and the effects that blades have on each
other. This loading tends to be transmitted through the main driven shaft into the
gearbox and subsequently into the tower.
In the aerospace industry, loading and natural frequencies tend to lie at higher
values. In rotary wing aircraft, the main excitation source, coming from blade
rotation, tends to be low-to-medium frequency, in the 2–50 Hz range, where the
loading is almost purely harmonic. Fixed-wing aircraft tend to be excited by a
number of sources, but the excitation tends to be broadband, and is typically
approximated by Gaussian white noise, or other colored noise for analysis purposes.
The excitation sources in fixed wing aircraft come predominantly from aerodynamic
loads, such as buffeting, friction, and gusts. The frequency ranges involved in this
can range from medium frequency narrowband buffeting in the 50–1000 Hz range,
to much higher-frequency broadband noise involving noise up to several kiloHertz.
There is a strong overlap between SHM and machine CM. As discussed in Sect.
1.2, CM is now a mature field given the fact that the dynamics of rotating machinery
are fairly well understood and such systems normally operate in controlled environ-
ments and loading conditions. For example, a wide variety of bearing and gearbox
faults in gear and bearing components can be often easily classified as they manifest
themselves in specific frequency bands. There are many similarities between the
types of features that are sensitive to damage in both domains. Hence, much of the
discussions that will follow regarding damage-sensitive features also apply to the
field of CM.

2.1.1 Ways to Measure Vibration

The piezoelectric accelerometer is arguably the most popular and practical instru-
ment for measuring vibration. The shapes and sizes vary according to application
1000 R. Fuentes et al.

and cost, but they all generally rely on the same principle – that of converting a
mechanical acceleration into an electrical signal through the piezoelectric effect. An
alternative method for measuring vibration is a laser vibrometer. Laser vibrometry
relies on the Doppler effect on the reflection of a laser beam on a moving surface,
to measure velocity normal to a surface. This technique solves some of the main
drawbacks of accelerometers. Being noncontact, it does not add mass to the system
and does not require a mounting point. Furthermore, large areas of a structure can be
scanned with minimal setup; this can save large amounts of test time compared with
setting up accelerometers, if the channel count is high. The main drawbacks of laser
vibrometry are the fact that the surface has to be reflective and in the line of sight,
and the laser setup is more suitable for laboratory than operational environments.
When measuring displacement, the most popular tool is the strain gauge. The
basic working principle of a strain gauge is that as a material strains, its electrical
resistance will change accordingly and one can measure this change by passing an
electric current through it. There are many classes of strain gauges, but most of them
constitute a wire or a metallic filament that is bonded to the surface of the structure.

2.1.2 Damage-Sensitive Features from Vibration

• To what degree is the excitation source known and can it be measured?
• What is the nature of the dynamic response (stationary, periodic, stochastic,
transient)?
• How many dimensions (DOFs) are being measured?
• What type of damage is one interested in finding? Is there any prior knowledge
regarding? Where this might manifest itself?

By far the most important factor that dictates which features might be useful for
damage detection purposes and what further processing must be carried downstream
is whether or not the excitation can be measured. This point is of fundamental
importance, given that the input excitation has a direct impact on the output dynamic
response. If the premise for detecting damage is based on spotting a change in the
dynamic response, the best way to distinguish whether a change is the result of
damage or a change in excitation is to measure the excitation. In the discussion that
follows, the various different damage-sensitive features will be analyzed in terms of
the points above.
As such, the raw dynamic response data constitute a damage-sensitive feature,
although it is not always a particularly good one. One reason for this is that in
most cases, a time-history is highly redundant in terms of its information content.
This fact results in higher dimensions in the feature vectors, which is in fact
problematic for the machine learning techniques further downstream. It is useful
to use transformations of the raw data into domains that are low dimensional, so
they summarize the data well, are fast to compute, and of course, also sensitive
to damage. It is also often the case that one will choose a feature that has a clear
physical interpretation, so that besides being useful for the purposes of automatically
detecting damage using a machine learning algorithm, one can also understand the
damage mechanism. To this end, Fourier analysis, which transforms signals from the
18 Structural Health Monitoring and Damage Identification 1001

time to the frequency domain, is one of the most useful and widespread techniques,
so a fair amount of attention will be devoted here to features derived from the Fourier
transform.
Throughout this section, examples will be drawn from an investigation originally
carried out in [24].

Features Based on Signal Statistics

Basic signal statistics constitute the simplest forms of damage sensitive features.
This includes the mean, mean-square, variance, skewness, kurtosis, and root-mean-
squared (RMS). Damage has been long known to manifest itself in changes in these
values. Damage introduces changes to the dynamical system that governs the output
measurements. For example, plasticity induced by a growing crack could alter the
damping of a structure, which would lead to a change in the total energy transfer
between input excitation and observed strain/acceleration. This energy change could
be directly observed in all the statistics that measure overall spread such as peak
amplitudes, RMS, and variance. Damage causes changes in the natural frequencies
of the system, and if the system is being excited at a constant frequency, damage
will lead to a change in spread and energy-measuring statistics too.
The higher-order moments, skewness, and kurtosis have also found uses in
damage detection. Kurtosis effectively provides as measure of how non-Gaussian
a signal is. A structural crack may result in an additional impact-like response as
it opens and closes during oscillation cycles, and this may become evident in the
kurtosis. In a similar way, the effect of a crack could be described by a bilinear
stiffness model, where the material stiffens up as the crack closes and loses stiffness
as it opens. These effects generate two different dynamic behaviors in the up and
down strokes of oscillation cycles which will be evident in a skewed probability
distribution of the raw signals.
Basic statistics offer a simple way of deriving damage-sensitive quantities with
relatively minimal computational complexity. On the other hand, one must bear in
mind that changes in these features can easily be caused by changing operational
and environmental effects. Changing excitation amplitudes and frequencies will
easily cause similar changes to these features as will damage. The same is true for
changing material properties from changes in temperature and humidity.

Auto and Cross-Correlations

The correlation function is a similarity measure over time, which is a particularly
useful concept in the analysis of signals. An autocorrelation function, δ xx (τ), of a
signal x(t) is a self-similarity measure, defined by,

δxx (τ ) = E [x(t + τ ) x(t)] (1)

It is essentially a measure of how much a signal looks like itself when shifted
in time by an amount τ . It is a useful damage-sensitive feature as it provides a
normalization of the time axis (to the time-shift), so that one could effectively
compare several autocorrelation feature vectors like-for-like. One could not do this
1002 R. Fuentes et al.

with a raw time history unless it was perfectly synchronized to a specific event,
such as an impulse or the phase of a periodic excitation source. The autocorrelation
function also highlights periodicity in waveforms, which will be evident as strong
peaks in δ xx (τ ). If one has two different stationary random processes x(t) and y(t)
(usually the input and output of a system), one can define the cross-correlation
functions,

δyx (τ ) = E [y(t + τ ) x(t)] (2)

δxy (τ ) = E [x(t + τ ) y(t)] (3)

In a sense δ yx (τ ) detects causal relationships between signals; often these are

the input and output to some system under investigation or test. In SHM, the
auto- and cross-correlation functions are useful in a variety of contexts, often as
damage-sensitive features themselves, but also serving as essential steps in the
computation of other features, such as power spectra and coherence functions. In
the context of vibration monitoring, the cross-correlation has been used extensively
as a damage-sensitive feature [25, 26, 27]; its use in SHM extends to applications in
Acoustic Emissions and ultrasound monitoring [28, 29, 30], the uses of which are
discussed below in the respective sections. An important point to bear in mind when
computing a cross-correlation function is its normalization. This is not discussed
here but the reader is referred to any standard signal processing book for this.

Frequency Spectra
Transforming vibration signals into the frequency domain is undoubtedly a powerful
approach to understanding vibration waveforms. The frequency spectrum of a
discrete signal, derived using a discrete Fourier transform (DFT) is defined as (The
nomenclature of spectra is quite precise, in that account should really be taken as
to whether a spectrum represents power or power density, etc. The distinction is
represented in the units of the object of interest, and should take account of the
effects of windowing, etc. In this chapter, spectra are only discussed in terms of their
usage as features for SHM, and their absolute scales are not relevant. For this reason,
the discussion may be a little cavalier in places, concerning units and scales.),


N −1
2π n
Xn = xi e−i N r (4)
r=0

where n defines the spectral line index, corresponding to a specific frequency

bin, and r represents the time index of waveform x. Performing this sum has
a computational complexity of O(N2 ), which is fairly prohibitive for practical
computation. This high computational cost led to the development of the fast
Fourier transform (FFT) algorithm, which reduced the computational complexity
of computing Xn to O(N log N). The FFT has been singled out as one of the
most important algorithmic developments of the twentieth century [31], and it has
18 Structural Health Monitoring and Damage Identification 1003

enabled widespread application of Fourier analysis across a multitude of disciplines,

including structural dynamics and SHM. Moving to the frequency domain is a
particularly useful way of interpreting the output of a system as the amplitude and
phase information is broken down into distinct bands associated with each response
frequency. The output of a DFT/FFT can be readily interpreted as the amplitude
contribution of each individual sine tone to the total signal. It is very effective when
used in outputs of harmonically excited structures, but less so when the excitation is
broadband and stochastic. A common function derived from the Fourier transform
that is used to examine the output of a stochastic excitation in the frequency domain
is the power spectral density (PSD). An example of a PSD is given in Figure 1c for
undamaged and damaged class datasets. The PSD defines the relative distribution of
power over a given frequency bandwidth. Any change in natural frequency, damping
ratio, and input excitation will lead to a change in the shape of the PSD, thus making
it a useful damage-sensitive feature.
The definition for the power spectrum follows conveniently from that of the
autocorrelation of Eq. (2) (not shown here in detail). Taking the Fourier transform
of τ (xy) yields the convenient form,

F [τ (xy)] = X (ω) X(ω)∗

(5)
Sxx = X(ω)2

Fig. 1 Damaged and undamaged features from the five-story LANL test structure: (a) raw
acceleration, (b) autocorrelation function, (c) PSD, (d) FRF, and (e) coherence.
1004 R. Fuentes et al.

which defines the power spectrum as the square-modulus of the Fourier transforms
of x(t). Similarly, the Fourier transform of the cross-correlation function leads to the
definition of the cross-spectral power Sxy , which becomes handy in the computation
of other features. Note that a power spectrum is an output-only measure and is
therefore also sensitive to changes in the input excitation, so one would have to take
this into account. If a measurement of the input excitation is available, the Frequency
Response Function (FRF) naturally takes this into account, as by definition an FRF
is a ratio between inputs and outputs in a system. The FRF has been identified
as a useful damage-sensitive feature since the early days of SHM [32, 33, 34, 35,
36]. An FRF can be built as a ratio of acceleration, displacement, velocity, and
combinations thereof. An example of an FRF on undamaged and damaged states is
given in Fig. 1d. Mathematically, the FRF is defined as a ratio of inputs,

Syx
H (ω) = (6)
Sxx

Note that the cross-spectral density is used in the above computation; this is
required in order to recover the phase information of the FRF. An amplitude-only
estimate can be obtained from,

Syy
|H (ω)|2 = (7)
Sxx

While the FRF is invariant under changes in input excitation, it is not invariant
under the changing physics imposed by changing operational and environmental
variables, such as that in mass and stiffness estimates arising from different
temperatures and usage cases. More often than not, an FRF is a damage-sensitive
feature found in laboratory environments, where one has the luxury of measuring
and potentially controlling input loads. While in some industries it is possible to
measure an input load in operation, there are many instances where measuring
an input load is simply not possible, especially when wave or wind loading are
involved, and one must resort to using other damage-sensitive features such as the
PSD, and let the machine learning model the changes in loading conditions. Some
of this will be discussed in more detail in Sect. 4.

2.1.3 Coherence
The next in line of useful features derived in the frequency domain is the coherence
function. Coherence is a measure of how linearly related the output of a system
is related to the input of a system; it thus provides an indication of whether
nonlinearities have developed in a system, and nonlinear behavior can often be an
indicator of the presence of damage. The coherence function takes values between
zero and one, with unity indicating perfect linear correlation between input and
output. The coherence function, γ 2 , can be explained as a fraction of the output
power (of signal y) that can linearly correlate to the input power (of signal x) and
can be written as,
18 Structural Health Monitoring and Damage Identification 1005

Smm (ω)
γ 2 (ω) = 1 − (8)
Syy (ω)

While the coherence function can be a useful damage-sensitive feature, care

must be taken so as to capture effects associated to the breakdown of linearity,
and not artifacts of the testing strategy or the signal processing. As an example, a
white-masking effect could be introduced in the input-output relationship for lightly
damped systems when multiple datasets are averaged, and some of the results in the
current average contain transient responses that have not died out, hence appearing
to be uncorrelated. The coherence function is shown in Figure 1e, for damaged and
undamaged class datasets.

Time-Frequency Analysis
While frequency domain analyses carried out through Fourier transform methods
are well suited for signals of a purely periodic nature, they largely fail to correctly
characterize transient signals. This is due to the nature of the Fourier transform
itself; it works by fitting infinitely long sinusoids to finite signals, and this introduces
issues. One of the most popular ways of dealing with this problem is by pre-
multiplying the vibration signal by a window function that reduces the effects
of the signal edges. The obvious downside is that any information contained in
those edges would be attenuated. The Short Time Fourier Transform (STFT) deals
with this problem by splitting the signal into overlapping windows, and applying a
windowed FFT on each signal segment. The resulting vector of frequency spectra
for each window can then be used as features in a statistical inference algorithm.
The particular case where the windowing function is Gaussian leads to the Gabor
transform, which was used extensively before the generalization to the STFT.
The STFT can be particularly useful for extracting features from a system that
is continuously changing, either from inherent complexities of the system such
as nonlinearities, or changes in its environment and operation. An example of an
STFT applied to a synthetic nonlinear oscillator is illustrated in Fig. 2. The features
from this example will be used to illustrate dimensionality reduction and novelty
detection in Sect. 4.
The discrete Fourier transform of a signal over a finite time window of size
N will yield a complex vector of size N, of which only N/2 vectors are typically
independent, given that the DFT is defined over negative and positive frequencies.
There is an inherent relationship between the frequency resolution and the time
resolution. Large STFT windows lead to higher frequency resolution. However,
because the DFT/FFT applied within each window gives fundamentally a stationary
distribution, large windows imply a loss of time resolution. Therefore, there is an
inherent trade-off between frequency and time resolution in the STFT. This trade-
off is illustrated in Fig. 3a. If one desires high frequency resolution, a large window
must be used, but this will result in a loss of time resolution and vice versa. This
is a problem, given that typically one requires higher frequency resolutions and
less time resolution to analyze slow moving (low-frequency) processes, and high
1006 R. Fuentes et al.

s =1 s =4 s =8
80

70

60
Frequency (Hz)

50

40

30

20

10

20 40 60 80 100 120 140 160 180 200 220

Feature Vector Number

Fig. 2 Feature vectors comprising of amplitude spectra, derived using an STFT from the accelera-
tion response of a simulated nonlinear 3-DOF system excited with white noise at increasing loads.
Note the natural frequency increases slightly at different levels, and harmonics are evident at the
highest loading

Fig. 3 Illustration of time-frequency resolution trade-off performed by (a) short time Fourier
transform (STFT) and (b) discrete wavelet transform. (Figure from [38])
18 Structural Health Monitoring and Damage Identification 1007

time resolution with less frequency resolution for fast-evolving processes (usually
at higher frequencies). The STFT does not cater for both of these needs.

Time-Scale Analysis
The wavelet transform fixes some of the shortcomings of the STFT when applied to
transient signals. Wavelet analysis is broadly divided into two classes: continuous
and discrete. The continuous wavelet transform (CWT) of a signal x(t) is defined as
its convolution with a wavelet function ψ(a, t),
 ∞  
1 t −b
c (a, b) = x(t)ψ dt (9)
|a|1/2 −∞ a

where a and b are scale and translation parameters, respectively. As an example of

a simple analyzing mother wavelet ψ, take the complex Morlet wavelet, defined as
a wave windowed by a Gaussian,

ψ(t) = (π C)−0.5 e−t

2 /B
e2π iCt (10)

where B and C are bandwidth and center-frequency parameters, respectively.

The idea of the CWT is to perform this convolution between x(t) and ψ at
varying scales and translations of the mother wavelet. In principle, it is a similar
procedure to the STFT, except that the basis function is transient, which can better
represent transient signals. The scale of the mother wavelet will be related to a
physical frequency; low scales correspond to high frequencies and high scales to
low frequencies.
Some care should be exercised when using a CWT as an SHM feature, as
the transform naturally leads to redundant information. Using a CWT, one can
analyze the energy of a signal at any arbitrary scale (frequency) and discrete time
point. This arbitrary resolution makes it an excellent tool for the visualization of
transient signals. However, for analysis purposes, the CWT can result in a very
inefficient transform if care is not taken to select appropriate scales that provide
useful information about the signal. The CWT has found some use in the field of
machine CM, given its ability to tune into very specific frequencies [37]. In this
context, it is often used in conjunction with a Hilbert transform.
An alternative to the CWT is the discrete wavelet transform (DWT). The names
of these two transforms may be misleading as both operate on discrete data, but
the CWT operates on a continuous space of scales, while the DWT operates on a
discrete grid. The DWT is designed to overcome the redundancy issues of the CWT;
it does so by specifying a mother wavelet that produces a half-band filter. In effect,
the DWT splits the signal into approximation (cA) and detail coefficients (cD). The
approximations contain the low-frequency part of the signal, and the details contain
the high frequencies. There is also a decimation step, so that the number of points
in x(t) equals the sum of the number of points in cA and cD. The idea is to represent
a signal in two bands, with equal numbers of points.
1008 R. Fuentes et al.

Fig. 4 Illustration of a
multilevel wavelet x(t)
decomposition down to three
levels. The wavelet
coefficients inside the dashed
area fully define x(t) cA1 cD1

cA2 cD2

cA3 cD3

The time frequency trade-off occurs when one applies the DWT recursively,
in a tree. The multilevel wavelet decomposition is a DWT designed to split the
signal x(t) into sets of cA and cD that represent all the frequencies in the signal,
but with the right time resolution. The tree is illustrated in Fig. 4. The DWT is
applied recursively to all the approximation coefficients, cAi . Each approximation
coefficient thus represents half of the frequency spectrum of the coefficients one
level above, but with half the number of points. Thus, this decomposition results
in each coefficient representing a frequency band at the appropriate sample rate for
that band. This is significant, as it allows low-frequency data to be represented at a
low rate, and the opposite for high frequency data. The area enclosed by the dashed
region in Fig. 4, represents the entire spectrum of the signal. The multiresolution
approach of the multilevel DWT is a natural trade-off between time and frequency
resolution, compared with the STFT. This point is illustrated in Fig. 3. It is possible
to apply an inverse DWT that perfectly reconstructs the original signal from these
coefficients.
The value of this multiresolution analysis to SHM is great; in particular, for
applications where the data are of a transient nature, or contain dynamics at different
scales, and one requires to extract features that represent each scale succinctly. The
use of wavelets in the SHM literature is quite popular now. Both the DWT and CWT
require the use of a basis function to perform the wavelet decomposition. A lot of
studies, in the SHM literature, have focused on finding suitable bases for particular
SHM problems.

Time Series Models

Time series models provide a means for extracting features, and forecasting time
series, purely in the time domain. Arguably the most popular time series model is
the linear autoregressive (AR) formulation, which views a signal y(t) as a linear
function of its previous p values,
18 Structural Health Monitoring and Damage Identification 1009


p
yt = ai yt−i + η (11)
i=1

where θAR = {a1 , . . . ak } are autoregressive coefficients, which effectively encode

a spectral representation of y(t), and η is a noise term. This form is essentially a
linear regression problem on lagged versions of the signal, so the AR coefficients
can be estimated with ordinary least squares (OLS) regression. Just as in the case of
a frequency domain representation, these coefficients can highlight a change in the
dynamics, and so have been extensively studied as damage-sensitive features [39,
3, 40, 41, 42]. Figure 5 illustrates the AR coefficients of a three-DOF mass-spring-
damper system with and without damage.
Time series models distinguish themselves from spectral models since they can
be used as predictors, and this can be readily exploited by SHM algorithms. A
prediction is a one (or multiple) step-ahead forecast of where the signal will be in
the future, so one could use the model/prediction residual, , as a damage-sensitive
feature. The premise is that a change in the underlying dynamics would cause a
change in the “true” AR coefficients of the system; any predictions on this, using a
baseline undamaged AR model, will result in an increased model discrepancy, .
Most of the early studies into the use of autoregression focus purely on the
damage detection and localization problem with no external influences [3, 42].
However, attention has shifted toward SHM under changing environmental and
operational conditions. An example of an AR feature vector resulting from a

Fig. 5 Illustration of AR coefficients from an undamaged and damaged system, where damage is
represented by a stiffness reduction on a three-DOF mass-spring-damper system
1010 R. Fuentes et al.

synthetic undamaged and damaged three-DOF mass-spring-damper system is given

in Fig. 5. In this case, the damage was introduced through a 20% stiffness reduction
in one of the springs.
Extensions of the linear AR model exist to account for Moving Average (MA)
terms. For example, a linear mass-spring-damper system excited by unknown
Gaussian noise can be captured by an ARMA model. These forms can be extended
to account for external or eXogenous inputs (ARMAX) [43]. However, a linear
AR, ARMA, or ARMAX model would fail to model nonlinear dynamics, and
thus some further extension of this model is required if the system in question
behaves nonlinearly. If the nonlinear parametric form of the underlying system is
known, then this should be used. One of the most general models is the Nonlinear
Autoregressive Model with eXogenous inputs (NARMAX) [44, 45]. The exogenous
or MA terms can be dropped if it is appropriate for the application. The functional
mapping between a signal y(t) and its lagged values y(t – p) for a NARMAX
model can really be anything from nonlinear multinomials to more complex Neural
Network [46] or other nonparametric forms such as radial basis function networks
or Gaussian Process regression [47, 48].
The essence of time series models from an SHM perspective is to capture
the dynamics of the process and use these to make predictions about the model.
Whether the model is linear or nonlinear, parametric or nonparametric, is a modeling
choice and it is the topic of system identification. It is no surprise that there is
an abundance of these methods in the SHM literature, given that the time-domain
system identification is a rather mature subject. In the use of the NARX model, for
example, all that is required is a suitable nonlinear regression method between xt
and the p lagged values. Regression methods can be categorized into parametric
and nonparametric. This results in either of two distinct models in their approach
to modeling the dynamics, often known as white and black-box models. White-box
models are often parametric, and they are distinguished by the relative ease by which
one can recover the underlying parameters of the equations of motion that generated
the data. This freedom gives not only the ability to predict the signal output but
also yields some insight into the physics. On the other hand, black-box models are
associated with nonparametric regression methods such as artificial neural networks
(ANN), support vector machines (SVM), radial basis function (RBF) networks,
relevance vector machines (RVM), and Gaussian processes (GP). These methods
can all be used in the context of NARX models.

Modal Analysis-Based Features

The topic of damage-sensitive features derived for SHM applications extends far
beyond the scope of this small section, so the focus will be on providing the reader
with a very general overview of the topic. More thorough discussions can be found
in [1]. Modal analysis falls under the general category of system identification
techniques, and there are various ways of performing modal testing to extract
the modes. Experimental modal analysis (EMA) is now a mature field, used in
industry primarily to understand and evaluate structural dynamics with the purpose
18 Structural Health Monitoring and Damage Identification 1011

of avoiding certain resonance frequencies, managing structural damping to control

fatigue life, and to validate and update finite element models (FEM) [49].
The usefulness of modal analysis in damage detection and localization problems
has been identified a long time ago [50, 51, 52]. Early studies have focused
on examining the link between modal parameters and the structural degradation
process. Natural frequencies extracted through EMA have been identified as a
primary feature, as they will tend to decrease as the structure is degraded [50]. The
curvature of mode shapes has also been identified as a feature that is useful for
localizing damage [51]. Furthermore, FEM model updating techniques have also
been explored as a means of detecting and locating damage [53].
One of the key aspects of EMA is that input excitations to the system are available
for modeling and analysis. This is of relevance as extracting modal parameters
involves finding a set of parameters for the system equations of motion that agree
in some way with the data being measured. This is termed system identification,
and it is (now) relatively simple to do this using frequency response functions
(FRFs), which measure the input-output relationship between the forcing and the
acceleration response. Modal parameter estimation is often done by finding a set of
analytical FRFs that provide a good fit to the measured FRFs.
In order to compute an FRF, a measurement of the input force is normally
required. However, this input loading may be very difficult to measure in most
practical engineering applications, which is the main reason why EMA is usually
confined to laboratory settings. Methods have been developed to estimate modal
parameters without the need for the measurement of input loads, and they are
classed as operational modal analysis (OMA). One of the most popular techniques
for doing this is Stochastic Subspace Identification (SSI) [54], which fits a state-
space model to the vibration response of the system. One of the key assumptions
of this method (and most OMA methods) is that the loading can be approximated
by white Gaussian noise. Section 3 will discuss the state-space methodology often
used in OMA in more detail.

2.2 Acoustic Emissions (AE)

Acoustic emissions (AE) are high-frequency stress waves that are released from
a material when the internal structure undergoes a change. These waves are often
recorded as bursts, and can be generated by processes such as friction, corrosion,
stress, and growing cracks. Because the application of stress leads to the generation
of AE, this is a popular technique in NDT for assessing the loading history of a
structure, thanks to the Kaiser effect [55]. Kaiser discovered, in the 1950s [55], that
a structure will emit AE if loaded up to a stress that it has not been loaded to before.
Any subsequent application of stress will result in much reduced AE levels.
It is worth establishing some contrast between what is classed as structural
vibration, and AE. Physically, they are similar phenomena, but the excitation source
1012 R. Fuentes et al.

and frequency are much different. In theory, structural vibration can go up to any
arbitrarily high frequency, but practical constraints limit vibration analysis to the
tens of kilohertz range. Unless a structure is unreasonably stiff and lightweight,
its first few natural frequencies will lie within the 0–10 kHz range. Mechanical
excitation within this frequency range tends to be associated with environmental
loads, shock, and rotational motion. This is thus classed as structural vibration.
Going into the tens of kiloHertz range, one finds that natural mechanical
excitation arise from completely different sources. To generate a wave at this
frequency, an impulse would have to be much shorter than the average impulse used,
for example, in modal hammer testing. Micro-cracks tend to generate very short
impulses, sending mechanical stress waves across the material, and this is what is
referred to as AE. When generated from material dislocation, AE can be observed
as discrete bursts, or as is referred to in the AE literature as hits. Figure 6 illustrates
a series of AE hits generated from a yielding steel sample (taken from [56]).
A lot of research followed on from Kaiser’s original thesis, and it is now a
well-understood fact that stress causes AE. This fact is particularly useful given
that a crack introduces a discontinuity in a material, and therefore cause stress
concentration. On the other hand, crack-growth estimation methods rely heavily on
stress concentration factors to estimate residual life. There is a strong link between
AE and crack growth; this has been made a long time ago [57]. The strain energy
release rate from a crack can manifest itself as stress waves. This quantity is thus
well correlated to the count of discrete AE hits and their energy.
There are two key advantages of AE over vibration; the first is that damage
can often be identified at a much earlier stage, owing to the fact that micro-cracks
will generate AE before an appreciable change in the vibratory response appears.
A further advantage of AE is that the source of damage can be located accurately
with relative ease. If multiple sensors are used, the time-of-flight difference between
different sensors can be used to find the spatial location of the acoustic source. If
the geometry of the material is simple and the material properties isotropic, then
all that is required is the wave propagation speed and a triangulation scheme. If
the geometry is complex, a look-up table approach has been suggested [58] called
Delta-t mapping. Another, better method based on Gaussian Process regression has

Fig. 6 Illustration of AE bursts, generated from a yielding steel sample [56]

18 Structural Health Monitoring and Damage Identification 1013

Fig. 7 Typical damage-sensitive features used in the analysis of AE signals [1]

been suggested in [59]. Both of these methods require an example dataset of time of
arrivals with known source locations.
AE is often analyzed in terms of features derived from the transient bursts that
are characteristic of energy release in solids. Features that are important involve
those that quantify the level of energy in the burst, how fast it rises and decays, and
its peak amplitude. These basic features are illustrated in Fig. 7. Note that these
AE bursts are often defined by their exceedance above a threshold value which is
often formed using an order statistic, or a simple maxima over the background noise
floor. However, the choice of threshold has a big effect on the resulting features, and
a simplistic threshold may under- or overestimate important quantities such as rise-
time or accurate wave time-of-arrival differences for localization purposes. In this
case, more advanced onset estimation techniques such as [60] are recommended.

2.3 Guided Waves

While AE is a passive technique, where one listens to the response of the material
to stress and other material effects, it is also possible to excite the material with a
pulse and listen to the response. This is the principle of ultrasound testing, which is
1014 R. Fuentes et al.

predominantly an NDT technique, where a pulse is sent across the thickness of the
material, generating bulk waves, and listening to the reflection of that pulse. This
principle of A, B, and C scans is built around mapping of ultrasonic time of flight
across and between surfaces. The use is widespread in nondestructive evaluation of
engineering structures, but it is not strictly an SHM technique as there are practical
impediments to its implementation in that context.
Guided waves present a solution to this problem. These waves are guided
by the boundaries of the medium in which they propagate. Unlike compression
and shear waves, analytical solutions for wave propagation in these cases involve
consideration of the boundary conditions of the medium. If one considers the
problem of only one boundary condition (a surface), this leads to Rayleigh
waves [61]. One of the key aspects of Raleigh waves comes from the fact that
they propagate along√ a surface. Their geometrical amplitude attenuation is thus
proportional to 1/ R. In contrast, the amplitude of bulk waves decreases with 1/R,
as the wavefront propagates as a sphere. More interesting wave modes arise if there
are two bounding surfaces (a plate). The analytical solution of wave propagation in
these cases leads to the popular Lamb waves [62]. These are wave modes where
both of the bounding surfaces of the plate interact with each other to create either
symmetric or antisymmetric modes. These are illustrated in Fig. 8. The interesting
thing about Lamb waves is that they are dispersive; their propagation speed depends
on the frequency-thickness product. Many more Lamb wave modes exist if one
keeps increasing the frequency. Figure 9 shows the dispersion curves for symmetric
and antisymmetric modes in an aluminum plate.
Beyond being inherently interesting, there are some very practical implications
to the (rather condensed) discussion above. The first is that a discontinuity such as a
crack, or delamination of a composite medium will scatter and reflect Lamb waves.
It is thus possible to monitor the integrity of a structure by generating Lamb waves
and monitoring the response at various points in the structure. This fits well with
the machine learning approach to structural damage identification; if the response
changes, scattering or reflection is likely to have been the cause.
The second practical point is the frequency-thickness product. To efficiently use
Lamb wave propagation for SHM, one needs to be able to both generate a pulse at
the right frequency and acquire the resulting waveform. Working with one particular

a) Symmetric b) Antisymmetric

Fig. 8 Illustration of lamb waves showing (a) symmetric and (b) antisymmetric modes
18 Structural Health Monitoring and Damage Identification 1015

Fig. 9 Dispersion curves for symmetric and antisymmetric lamb wave modes in aluminum [63]

mode is thus desirable. The problem is that at high ωd, the propagation speeds of
different modes become very similar and are thus hard to separate from each other.
This task is usually straightforward at low ωd where the modes are very clearly
separated. A typical piezoelectric transducer will have resonances in the range of
10 kHz to 10 MHz, which restricts the usable plate thickness to no more than a few
centimeters, in most common engineering materials.
One of the key disadvantages in implementing guided-wave techniques in prac-
tice lies with the sensors. In transmission and reception, a piezoelectric transducer
is normally used to convert a mechanical waveform into an electrical signal, and
vice versa. The problem is that piezoelectric sensors are also very sensitive to
temperature changes, so it is often difficult to establish a baseline condition under
temperature fluctuations. Some studies suggest that the problem renders guided
waves an impractical technique [64], and that the solution is to increase the number
of sensors. However, this may not be a practical engineering solution in problems
where a large sensor array would imply extra power (these sensors can use pulses
of 100 V), and space may not be available. Aerospace applications are one such
case. The signal processing solution has been suggested by Cross [65], where the
use of cointegration as a means of removing environmental trends in SHM data is
demonstrated with Lamb wave data under changing temperature. An example of
this is given in Sect. 5.

2.4 Performance, Operational, and Environmental Parameters

Performance metrics, in this context, refer to data that do not relate directly
to the dynamic response, but could help explain certain characteristics of the
response itself. It is generally good practice to collect any associated operational and
environmental parameters that affect the structural dynamics, such as temperatures,
pressures, humidity, and other performance indicators. In some cases, these may be
enough to identify certain types of damage and faults. For example, in the case of
1016 R. Fuentes et al.

wind turbines, various types of faults have been identified by modeling the power
output, and more specifically, how this varies with wind speed [66].

3 Advanced Topics in Signal Processing and Feature

Extraction

There are certain instances where one may wish to adopt more advanced techniques
for the computation and estimation of damage-sensitive features. This section will
discuss two such instances; the first is when one is faced with the problem of having
to deal with large quantities of SHM data to a point where considering the problem
of data compression becomes necessary. The second example is the case where
one may need to perform SHM in an online manner, and so the features must be
computed recursively as data is gathered. This section will discuss some solutions
to these problems, focusing on the most recently developed methods.

Dealing with Large Quantities of Data: Compressive Sensing

Currently, one of the gold standards in data compression is the discrete wavelet
transform, discussed in Sect. 2 and already used extensively in data compression
tasks for SHM [1, 67, 68]. However, these traditional approaches to data compres-
sion can be computationally costly, and could be restrictive in SHM applications
where computational power might be limited, such as with wireless sensor nodes.
Recently, the field of compressive sensing (CS) [69] has challenged the traditional
Nyquist-Shannon sampling theorem, and promises to allow one to infer information
from much higher frequencies, using only a limited number of time domain samples.
These ideas have the potential to bring huge advantages to the field of SHM where
they have already begun to see some applications [70, 71].
The general goal of compressive sensing is to acquire a signal using a much
smaller number of measurements than that required by the Nyquist-Shannon
sampling theorem. It is now a well-established principle that a signal could be
efficiently compressed, or coded, using a basis such as discrete cosine transform
(DCT), Fourier transform, or wavelets (if the signal contains strong transients).
This statement assumes that one has some knowledge of which coefficients in the
representative basis play an important role in the signal.
However, there is a drawback to this approach, since the entire signal must be
acquired for it to be then transformed into a sparse domain. This presents two issues:
data storage and processing. Storing an ultrasound dataset before it is compressed
can be a challenge given the very high sample rates, and the potentially large
scanning areas required by a high-resolution C-scan. Subsequently transforming
these large quantities of data into a compressed domain such as the wavelet domain
can also be computationally time consuming.
Compressive sensing solves this problem by using a combination of ideas,
with two basic underlying assumptions. The first is that the signal has a sparse
representation in some domain. The second is that some knowledge of what
18 Structural Health Monitoring and Damage Identification 1017

this domain may be is available. At the center of CS is the concept of using

l1 -regularised regression to find a coefficient set from a base dictionary that is sparse.
The following sections describe in more detail the essence of sparsity in the context
of ultrasonic NDT, the use of l1 -regularized regression, and the use of dictionaries;
these are all standard techniques in the field of CS.
A signal, x, with a high number of (time-domain) measurements n, is sparse
in a transform domain if a very small number of coefficients, m  n, in the
domain are sufficient to accurately represent the signal. Such a transform could be
represented as,

x= β (12)

where is a basis function set, or dictionary, and β is the coefficient vector that
represents the signal in the transform domain (note that in this chapter all vectors
are assumed to be column vectors unless otherwise specified). A good example of a
sparse representation would be a sinusoid at a fixed frequency, which may contain
a high number of points in the time domain, but may be fully represented by one
complex coefficient in the Fourier domain.
There is particular interest in the problem of dimensionality reduction, for the
purposes of algorithm design, in SHM and many other areas; this is also central
to the idea of CS and so it is worth a brief discussion. A way of “compress-
ing” a dataset is to project the n-dimensional measurement vector x to a lower,
m-dimensional space using a linear or nonlinear transformation. One popular
approach is to use transformations, such as principal component analysis (PCA),
independent component analysis (ICA), or factor analysis. Such a linear transfor-
mation could be written down as,

z= x (13)

where z is now a low-dimensional representation of x. An interesting projection

results if the rotation matrix, , is set to be a random matrix. Johnson and
Lindenstrauss [72] have shown that if is distributed according to a Gaussian, or
Bernoulli distribution, this linear dimensionality reduction preserves, with low error,
certain features of x, such as pairwise distances. This random matrix transform is a
key ingredient in the formulation of the CS problem.
In order for the compressed version of x, through the random transformation of
(13), to be of immediate practical use, there needs to be an algorithm that is able to
recover the original measurement. This is where the Least absolute shrinkage and
selection operator (Lasso) comes into play. The Lasso solves the classical linear
regression problem of Xβ = y, where X is a matrix with column-wise vectors of
inputs, y is an output, and β holds the regression coefficients. The Lasso encourages
sparse solutions for β through a penalty term based on an l1 norm [73]. The
optimization problem can be formulated as,
1018 R. Fuentes et al.

 
1
minimize : y − Xβ22 + λβ1 (14)
2N

The l1 penalty is regularized by the term λ. A general lq penalty could be


computed using the sum β = N j =1 |β| , and the Lasso is the special case when
q

q = 1. This constraint ensures that the optimization problem remains convex [73].
If q = 0, the resulting constraint would yield a subset selection problem that is non-
convex and combinatorially hard, thus not computationally efficient. The Lasso is
thus an attractive method for recovering sparse solutions in high dimensions while
maintaining efficient computation.
The regularization parameter, λ, dictates the degree of sparsity in the solution.
A high value of λ encourages a low number of nonzero coefficients, and vice
versa. Therefore, an appropriate value of λ needs to be chosen for each problem
in particular. The authors of the Lasso suggest using cross-validation [73] in order
to estimate the best (Cross-validation involves iteratively holding out subsets of the
available training data as test sets, in order to assess generalization performance and
avoid model over-fitting.).
This brings the discussion of CS to the last step [69], which is concerned with
finding a sparse set of coefficients β that best describe the random matrix projection
x (the compressed signal representation). This is where the power of the Lasso
is unleashed. What is available to the regression problem is not the full signal,
but rather a projection of it through . The coefficient set can be inferred if the
basis dictionary is also projected through the sensing matrix to yield the following
regression problem,

β= x (15)

where, as before, is a random matrix projection, is a basis function set, and

x is the (uncompressed n-dimensional) signal of interest. The Lasso minimization
of (14) can now be used in order to obtain a sparse solution for β.
Other reconstruction algorithms can also be used. In [74], a probabilistic
reconstruction method is used to reconstructed randomly compressed ultrasound
sequences. In this case, the reconstruction is also able to provide confidence
bounds on the accuracy of the reconstructed signal. An example of an original
and reconstructed ultrasound pulse, sampling significantly below the Nyquist rate
is illustrated in Fig. 10.

Fig. 10 Illustration of an
original and reconstructed
ultrasound pulse using
compressive sensing, and the
probabilistic reconstruction
scheme in [74]
18 Structural Health Monitoring and Damage Identification 1019

Recursive Estimation in the Linear Case: Kalman Filtering

Recursive estimation is an important topic in SHM, as it allows for the online
estimation of damage-sensitive features as well as online estimation of unknown
quantities of interest that cannot be measured directly, such as loading conditions.
This section will review the basic principles of recursive estimation in the simplest
case, when the dynamics of the system are linear, or when one is solving for features
derived from a linear operator.
This section will be followed by a discussion of the recursive estimation problem
in the more difficult nonlinear case.
When monitoring a dynamical system, one is often restricted to making mea-
surements that do not necessarily measure the state of the system directly, but
are related to the underlying process driving those measurements through some
function. A state-space model provides a solution to this type of problem; it models
the observations yt as some function of the underlying dynamics of xt . If the
relationship between x and y is linear, then this can be represented by the standard
linear state-space formulation,

xt = Axt−1 + wt wt ∼ N (0, Q)
(16)
yt = Cxt + vt vt ∼ N (0, R)

where C represents the linear function linking observations to underlying dynamics,

and A represents the linear dynamics – the time evolution of x. The observation
model and the dynamics are both modeled with white Gaussian noises, vt and wt ,
with zero mean, and covariance matrices R and Q, respectively (Standard linear
state-space models also include additive terms to account for control inputs. These
are omitted here for simplicity.). State-space models have applications spanning
various areas of science and engineering, and are now implemented in other fields
such as financial-time series modeling; they are useful whenever one makes use
of measurements that can be somehow related to the underlying state of a system,
and those measurements are corrupted by noise. It is easy to relate this picture to a
structural dynamics context. An MDOF linear vibrating system can be described by
a linear superposition of SDOF systems; this is the foundation of modal analysis.
The underlying driving functions are the SDOF oscillators, each of which has a
characteristic natural frequency, and they are related to a physical location on the
structure via a mode shape. Any good structural dynamics textbook will contain
methods for representing MDOF systems in state-space form [75]. There may be
more than one valid state-space representation of a system and one must adopt the
one that is most suitable to the problem at hand. In SHM, there could be two different
contexts where one might seek a state-space representation:

1. The state vector x represents the parameters of a model, and one is interested in
estimating those parameters, as they evolve through time.
2. The state vector x represents some hidden variables that better model the
underlying linear dynamics of a set of measurements y.
1020 R. Fuentes et al.

In the first case, the state-space modeling approach is cast as a parameter

identification problem. In this case, a state-space model could be seen as an
alternative view to the popular recursive least squares (RLS) algorithm, which
essentially solves the standard ordinary least squares problem, point by point. In this
interpretation of state-space models, the state vector will usually have some direct
physical meaning, and it will often be easy to interpret the results. In the second
interpretation, the meaning of x can be more subtle; in fact, it does not necessarily
need to have physical meaning at all. In this interpretation, if one possesses a
physical model of the structure being considered (say, via a finite element model or
an analytical model), then C and A can be derived, and one is interested in inferring
the state x and in monitoring the residuals vt and wt . An estimate of A and C that
correctly describes the measured data is therefore required, and can be obtained via
the popular stochastic subspace (SSI) methods or via the expectation maximization
(EM) algorithms.
Even though in both cases, xt represents something fundamentally different, both
have useful SHM applications, and most importantly, both require a method for
providing an estimate for xt . If linear Gaussian assumptions are made, the Kalman
filter algorithm provides an efficient and intuitive solution for the inference problem,
and so it is applicable to both the parameter estimation, and the latent variable
approach to implementing state-space models for feature extraction or direct novelty
detection in SHM.

3.1 Inference via the Kalman Filter

The key point of the state-space model is that the observations are modeled as
independent of each other, and dependent only on the state vector at time t, while
the state vector is dependent on all previous values on the Markov chain. There are
two important conditional probability relationships at play: the probability of the
observation vector given the state vector p(yt |xt ), and the probability of the state
vector given the same state vector at a previous point in time p(xt |xt – 1 ). If these
two densities are assumed to be linear Gaussian, they can be written down as,

p ( xt | xt−1 ) = N ( xt | Axt−1 , Q) (17)

p ( yt | xt ) = N ( yt | Cxt , R) (18)

The main interest is in estimating the unknown state vector, xt . Because xt is

not observed, the true state of the system will never be known, but a probability
density can be estimated, and if Gaussianity is assumed on the residuals of the
dynamics of the state, the distribution is fully specified by the mean and variance
of the state vector at every point in time. There are three probability distributions
for xt of interest, and they are commonly referred to (in the statistical time series
communities) as prediction, filtering, and smoothing, where each of them compute
the following conditional probabilities:
18 Structural Health Monitoring and Damage Identification 1021

1. Prediction: Probability of state vector, xt at time t given observations up to time

t – 1, p(xt |y1 , . . . , yt − 1 ).
2. Filtering: Probability of state vector, xt given observations up to time
t p(xt |y1 , . . . , yt ).
3. Smoothing: Probability of state vector, xt given all the observations available
p(xt |y1 , . . . , yT ).

For the state-space model, the filtering distribution can be shown to be,

p ( xt | y1:t ) ∝ p ( yt | xt ) p ( xt | xt−1 ) p ( xt−1 | y1:t−1 ) (19)
xt−1

So far, this does not assume linearity or Gaussianity; one is free to model
the conditional probabilities inside equation (19) with any arbitrarily complex
distribution. However, the assumption of linearity and Gaussianity simplifies things
significantly because of Gaussian identities; the product of two Gaussians is itself
a Gaussian, and the integral of a Gaussian is Gaussian too. These properties are
what make the representation of the conditional densities as Gaussians (equations
(17) and (18)) so efficient. The Kalman filter algorithm effectively solves equation
(19) for the linear Gaussian case. The Kalman algorithm involves two steps, a time
update and a measurement update. In the first case, the Gaussian mean and variance
of the state vector is propagated forward using the physical model to generate a
state prediction. The measurement-update step then takes in new measurements and
computes the filtered state vector, xtt in light of the measurements gathered at time t.
This procedure effectively uses Bayes’ rule to shrink the uncertainty of the state,
given the measurements. More details about the algorithm can be found in [76].

3.1.1 Inference for Parameter Identification

The Kalman filter is similar in nature to the recursive least squares (RLS) algorithm.
RLS has been extensively investigated in the structural dynamics community as a
method for identifying system parameters in real-time [77, 78, 49, 79]. For example,
a sequential or recursive form for a linear AR parameter estimate (such as the one
described by Eq. (11)), mapping yt (note it is one-dimensional) to its lagged version,
can be described by letting the state-space vector xt be the AR parameter vector a,

xt = xt−1 − Kt (yt xt − yt ) (20)

where Kt is a function of the filtered variance, υt , at time index (t – 1), which

represents the confidence in the state vector (in this case the parameters),

−1
Kt = υt−1 yt 1 + υt−1 yt2 (21)

and the update to the variance is,

1022 R. Fuentes et al.

−1
υt = υt−1 − υt−1
2
xt2 1 + υt−1 yt2 (22)

These recursive parameter estimates can be shown to give the same solution to
the AR parameter estimation problem that OLS provides, once all the time steps
have been processed. The equations above show the RLS parameter updates for a
single variable case, but the generalization to a multivariate form is essentially a
matrix version of equations (20), (21), and (22). This version is in fact provided by
the Kalman filter recursions. Using the Kalman filter recursions, RLS can effectively
be achieved for a variety of problems. The state vector xt represents the parameters,
and the observation matrix C would represent the observations. The state transition
matrix A can have different forms, but a popular choice is to set it to the identity
matrix; this assumes that the parameters follow a random walk, the volatility of
which is provided by the state-transition noise model, wt ∼ N (0, Q). Different
choices of the state transition matrix could put different constraints on the parameter
updates. One could solve for a variety of problems through a careful choice of C
and A. An application of interest in SHM is to estimate the AR coefficient vector
recursively. This could be modeled by setting the observation matrix Ct in every
Kalman filter recursion to be the lags of the signal of interest y. In other words, the
observation matrix varies with time, and is,

Ct = yt−1 , . . . , yt−p (23)

for an AR model with p lags. Because A is defined as an identity matrix, the

state transition reduces to specifying that the AR parameters should be close to
the previous ones in time, to within a specified variance,

xt = xt−1 + w (24)

As an illustration of this approach, Fig. 11 shows the fit of an AR model with

40 lags to the response of the second mass of a three-DOF nonlinear system, while
it undergoes a step change in the response due to a cubic nonlinearity. Note that
the estimates for both the AR parameters (xt ) and their variances, w change as the
system changes its dynamics. It is useful to be able to infer parameters in real-time,
and as will be pointed out in Sect. 4, these could be used as features in novelty
detection based on static data.
It should be noted that the filtering distribution can be used directly on the raw
measurement vectors from a system in order to carry out novelty detection, but this
use falls outside the scope of this chapter; more details can be found in [18].
18 Structural Health Monitoring and Damage Identification 1023

Fig. 11 Sequential estimation of AR parameters using a Kalman filter (a) shows the response of
a nonlinear three-DOF system to white noise, with a step change in response, (b) shows the state
vector xt , containing the AR parameters, and (c) shows the variance of each dimension of xt

3.2 Recursive Estimation in the Nonlinear Case

The linear Bayesian filtering equations shown in the previous section cover a wide
class of models which can be very expressive, which includes all linear multi
degree-of-freedom systems encountered in dynamics. However, many real-world
systems are nonlinear and using the linear formulations may yield suboptimal
performance and, in the worst case, fail to capture the dynamics of the system
altogether. Nonlinear recursive Bayesian estimation is not straightforward. The
complication arises from the solution to the integral of equation (19), which often
does not admit a closed-form solution in instances when the dynamics (p(xt |xt – 1 ))
are nonlinear and/or non-Gaussian. The Kalman filter presents a solution for these
two integrals in the special case when the dynamics are linear and the noise is
assumed to be Gaussian, but in the absence of closed-form solutions one must resort
to approximations which offer solutions to these equations.
1024 R. Fuentes et al.

This section will give a brief overview of the different approximations available
that are relevant to SHM, but it is not intended as an in-depth review of the topic. The
first approach to address prediction in nonlinear systems is often linearization. The
Extended Kalman Filter (EKF) is the most popular linearization of nonlinear state-
space models [80, 76, 81]. The EKF linearizes the system using a first-order Taylor
series approximation, through a Jacobian of the nonlinear functions and forming
transition and observation functions based on this. The EKF has been adopted
widely within the SHM community as a tool to deal with nonlinear systems.
By far the most common application of this model is when the EKF is used
to estimate parameters of a model, via a state augmentation approach. In this
case, based on the estimates of the parameters found, inference is made about the
condition of the structure when considering changes in the values of the parameters.
Examples of the state-augmentation approach to parameter estimation can be seen
in [82, 83, 84, 85, 86].
The EKF is known to be ineffective when the nonlinearity in the model is not
weak, that is, it cannot be well approximated (locally) by a linear model with minor
correction. Most importantly, the EKF is known to provide a bad approximation to
the propagation of arbitrary complex distributions through the nonlinear transition
functions. Even though the EKF is easy to implement, it often results in poor
performance. The Unscented Kalman Filter (UKF) [87, 88, 76] was designed to
remove the shortcomings of the EKF; it makes use of the unscented transform
which is a method for Gaussianizing the nonlinear transformation of a probability
distribution by means of a set of sigma points, and these deterministic points from
the previous distribution are propagated through the nonlinear transition and used to
estimate the mean and covariance of a Gaussian that best describes their distribution
following the transformation. The UKF has been reported to outperform the EKF
in various instances [89, 90, 91, 92]. These results are supported by the work in
[93] which focuses on application of the UKF to a structural dynamics identification
problem. The UKF has been used in many similar ways to the EKF within structural
dynamics [94, 95, 96, 97]. A key advantage of the UKF over an EKF approach is
that one does not have to derive and compute the Jacobian of the system.
While the UKF provides an improvement over the EKF in terms of the accuracy
with which it will propagate the state densities, and is computationally very efficient,
it is based on a heuristic and cannot guarantee even an approximate solution.
Ultimately, the best approximations can often be achieved using sequential Monte
Carlo (SMC) methods [98, 76]. These algorithms are a subset of the Monte Carlo
approach to approximating probability distributions. In an SMC setting, an explicit
evolving relationship between the distributions it exploited, that is, a sequence of
probability distributions through time is modeled. The benefit of using SMC as
opposed to either the EKF or UKF is that it is flexible enough to handle fully
nonlinear, non-Gaussian models without resorting to linearization at any point. The
trade-off is an increase in computational time.
The SMC algorithm consists of three basic steps: propagation, weighting, and
resampling. SMC is useful for estimating a sequence of probability distributions,
and one such sequence is shown in Fig. 12. SMC is most applicable when modeling
18 Structural Health Monitoring and Damage Identification 1025

Fig. 12 A sequence of
probability distributions to be
estimated using an SMC
scheme, for example, the
latent states of an SSM
representing the displacement
and velocity of a nonlinear
oscillator

Fig. 13 (a) The propagation and weighting of particles from the first two steps of the sequence
shown in 12, (b) Re-sampling of weighted particles following their propagation and weighting

nonlinear systems where the noise (process or observation) is non-Gaussian.

Importantly, there exist results to prove convergence of the SMC estimation of the
distributions toward the true distributions [99, 100]. There exist several variations on
the SMC formulation, the simplest one being the Bootstrap Particle Filter introduced
by [101]; here, this will be the focus of the explanation, but it should be noted that
it is often possible to construct more efficient filters, for example, the Auxiliary
Particle Filter [102].
The first step of the implementing the particle filter is to propagate a set of
particles, each with a weight 1/N (for N particles), through fθ (xt |xt – 1 , ut – 1 ). This
step defines both the functional form of the transition – for example, the equations
that propagate a Duffing oscillator through time from t to t + 1 – and the process
noise model. In this way, values for the particles at t + 1 can be sampled given the
positions of the particles at time t and any control inputs applied ut – 1 .
In Fig. 13a, the weights of the particles are represented by the sizes of the dots
on their locations. On the left, the set of equally weighted particles is seen, that is,
all the dots are the same size. The propagation of these particles is shown using the
lines which indicate their position (on the right) after passing through the transition
density. The next step is to calculate the weighting of the particles in their new
positions at time t + 1. This must be done via a weighting function which, in the
bootstrap filter, is defined as equal to the likelihood given the observation model.
This procedure enables the estimation of the probability density over xt , via
an importance sampling approach. However, as can be seen in Fig. 13a, this can
1026 R. Fuentes et al.

result in a large number of particles ending up having very low weights. These
particles contribute little information to understanding the distribution of interest.
In an ideal case, each of the particles should contribute equally to understanding
this distribution, this occurs when all the particles have an equal weight. In order to
ensure that, before every propagation step, the particles are equally weighted. The
particles are resampled according to their importance weights. The simplest manner
in which to do this is to resample the particles based on a multinomial resampling
procedure where the probabilities in the multinomial distribution are given by the
normalized importance weights. The basic premise of the resampling procedure is
to remove particles with low weights and to replace them with copies of particles
with higher weights. Alternatively, it is possible to employ residual, stratified, or
systematic resampling methods [103, 104] which can be more efficient, although
this will not be discussed here specifically.
Within SHM, the particle filter is being used for the modeling of nonlinear
dynamical systems [94, 105, 106, 107, 108, 109]. Again, the most common
application of these models is in the attempted recursive estimation of system
parameters. In addition to this point, the flexibility of an SMC scheme allows this
to occur even when the dynamics of the system are nonlinear. The efficacy of the
SMC approach when compared to an EKF is discussed in [110], with reference to
its application in structural dynamics and SHM. In conclusion, it should be noted
that, firstly, the SMC approach to handling nonlinearity in an SSM setting should be
considered the gold-standard when compared to either the EKF or UKF. Secondly,
care should be taken when using these models for recursive parameter estimation
to ensure that the transition model for the parameters does not bias their solutions.
It should also be noted that there is no guarantee that these recursive estimation
methods will converge to the true posteriors over the parameters; therefore, it
is worth considering that the use of a particle MCMC scheme may be a more
appropriate tool for parameter estimation [111].

4 Statistical Pattern Recognition for Damage Identification

The preceding sections have discussed various methods for extracting damage-
sensitive features from data that are measured from a structural or mechanical
system. As discussed, the analysis of these features should, in theory, allow for
the discrimination between the observations that relate to the normal operating
condition(s) and those that relate to damage. This section is concerned with
statistical analysis of these features, in order to inform damage identification.

4.1 Pattern Recognition for Feature Discrimination

It is generally accepted that pattern recognition theory offers a natural framework to

address the feature discrimination problem [17]. Following this approach, machine
learning algorithms have the potential to predict which groups of measured data
18 Structural Health Monitoring and Damage Identification 1027

relate to different operating conditions; for example, is the system operating

normally, under extreme temperatures, or, most critically, is the system damaged?
Feature discrimination can be formalized using Rytter’s hierarchy [5], as dis-
cussed in Sect. 1.2, specifically levels I–III. The first level, damage detection,
is typically addressed using outlier analysis, which utilizes the idea of novelty
detection algorithms. The concept of novelty detection can be considered as a form
of unsupervised learning; an introduction is provided in Sect. 4.3.
Damage assessment and classification (levels II and III) are more difficult, as
the algorithms require more information in order to make informed predictions,
which is often unavailable. Generally speaking, supervised learning is applied in an
attempt to classify and assess any features that might relate to damage; an example
of supervised learning is provided in Sect. 4.5.

4.2 Data-Driven Models in SHM: Learning and Prediction

When categorizing the measured data, x, from a system or structure, algorithms

(or “machines”) can be used to learn which diagnostic labels, y, are associated
with certain patterns within the input data (Note that for the remainder of this
section, the notation x will be used to refer to a multivariate measurement vector
of observations.). Therefore, a dataset must be available (in some form) in order to
train the algorithm. The process of learning from a subset of training data can be
defined in various ways. In the context of SHM, a visual introduction is provided.

4.2.1 Acoustic Emission Dataset

An acoustic emission (AE) dataset – collected by Pullin et al. at Cardiff University
[112] – is used to demonstrate statistical pattern recognition in the context of SHM.
These data were recorded during experiments in which a box-girder of a bridge was
exposed to cyclic loading from 0.1 to 85 kN [1]; details of the test procedure can
be found in [112, 113, 114]. Briefly, the AE burst signals were extracted from the
background noise of the measured data by setting a threshold based on the mean
and six standard deviations; an example of a burst signal is shown in Fig. 7. A total
of 91 AE burst signals were identified from the measured data.
The object of this dataset is to distinguish between different AE sources,
particularly those relating to crack growth, as this information might help to
inform damage detection, classification and prognosis. There are various ways to
implement machine learning in order to analyze the observed data. Time series
analysis could be applied to the burst signals directly, to learn a function in the time
domain and monitor the behavior of the signals [113]. In this example, however, and
in agreement with existing work [112], features are extracted from the burst signals
in the hope that they are sensitive to damage. The features used in this case are those
discussed in Sect. 2 for AE data; specifically, these are rise time, peak amplitude,
duration, and ring-down count [1].
Each AE burst, which represents an observation of the monitored system, is
represented by four features. In order to visualize these data, and to aid discussion,
1028 R. Fuentes et al.

Fig. 14 PCA feature space

for the 91 AE burst signals
projected onto two
dimensions such that the
variance is maximized

dimension reduction is applied to represent each observation in two dimensions,

such that x ∈ R2 ; linear PCA is applied [115]. PCA can be interpreted such that
the resulting features (or principal components) are a linear sum of the features
that define the original data. Furthermore, the projection of the data is such that
the variation has been maximized. The first two principal component scores are
plotted in Fig. 14, which allows for visualization of the 91 AE burst signals, as they
are represented in two dimensions. The goal of pattern recognition algorithms is to
classify these observations according to the condition of the structure that generated
each AE signal.

4.3 Outlier Analysis for Damage Identification

Novelty detection algorithms that utilize outlier analysis have been used extensively
for damage detection in practical applications of structural health monitoring (SHM)
[116, 117, 114]. The problem is to identify, from the measured data, if a machine or
structure has deviated from the normal condition, that is, if the data are novel [116].
In an engineering context, outliers can be suitably defined for novelty detection
as:

Data that deviate so much from other observations, as to arouse suspicions that they were
generated by some different mechanism. [118]
18 Structural Health Monitoring and Damage Identification 1029

Specifically, outlying data should indicate a significant change in the underlying

physics of the system, rather than benign fluctuations in measurement noise.
Although this description is conceptually simple, detecting informative outliers
from noisy engineering data is a nontrivial task.

4.3.1 Statistical Outlier Analysis

Parametric statistical approaches assume that the measured data can be represented
by some d-dimensional random vector, X, where each feature, Xi , can be considered
a random variable X ∈ Rd ∴ X = {X1 , . . . , Xd }. The random vector X is assumed to
be defined by some probability distribution function (pdf) f, such that X ∼ f. Using
these assumptions, the parameters of f can be estimated from the available data, and
a discordancy test can be used as a measure of novelty [119, 116].
Typically, the normal condition data are assumed to be multivariate Gaussian-
distributed, defined by the mean vector μ and covariance matrix  (i.e., location
and scatter), such that f = N (μ, ) [1]. The parameters can be then estimated
through the sample mean and covariance, μ̂,  ˆ , to provide a maximum-likelihood
estimate of the distribution f. To illustrate this concept, a subset of the AE data is
assumed to define a class which is considered as the normal data. In this case, these
are the AE bursts due to frictional processes away from the crack [113], shown by
the red ו markers in Fig. 15. In theory, any observations that are generated by an
alternative mechanism (in this case, crack-related events [113]) can be considered
as outliers.
The maximum likelihood estimate of f for the normal data is illustrated in
Fig. 15a; this is defined by the sample mean and covariance μ̂,  ˆ from the
training data (×). Visually, it is anticipated that any novel data, that is, crack-
related signals, should appear significantly different to the training data. With this in
mind, an appropriate measure of discordancy is the Mahalanobis squared-distance
(MSD) [114, 116]. The MSD can be interpreted as a covariance-weighted squared-
Euclidean-distance from the sample mean μ̂ of the normal data – if the covariance
is equal to the identity they become synonymous [115],
 
MSD i = xi − μ̂ ˆ −1 xi − μ̂ ,
 (25)

where xi is the potential outlying observation, such that xi ∈ X.

A critical MSD value, or threshold, must be defined in order to classify data as
normal or novel. The definition of this threshold proves to be nontrivial, thus, it is
approximated in various ways. When assuming Gaussian-distributed features, the
outliers can be seen as Chi-squared distributed, and the relevant critical values can
be used to define a threshold [119, 120, 121]. An issue with this technique, however,
is that it assumes a specific form of the data distribution; additionally, it is based on
the asymptotic distribution of distance measures [120]. Alternatively, extreme value
statistical theory [122] can be used to define the threshold; this approach has less
strict assumptions regarding the distribution of the data, and as a result, the method
should generalize across various datasets. In this example, however, a simple Monte
1030 R. Fuentes et al.

Fig. 15 Standard MSD outlier analysis with the AE dataset. The training and validation sets are
shown by × and • markers, respectively: (a) Observations in the feature space. The maximum
likelihood estimate of f is shown by the sample mean (+) and covariance (dotted lines represent
two and three standard deviations), (b) the MSD for each of the 91 burst signals

Carlo (MC) sampling technique is used to define a threshold [116]. This method
assumes Gaussian-distributed features, and represents a 99% confidence bound for
a dataset with the same dimensions as the training set. For further details of the MC
sampling regime, the reader is referred to [116].
The corresponding MSD for each observation in the AE data is shown in
Fig. 15b. As this is a maximum-likelihood estimate, the model risks overtraining;
therefore, it is critical to ensure that the model will generalize when applied to new
data through validation. A detailed discussion of validation techniques can be found
in [115]; however, in this simple example, a distinct validation set is used to ensure
generalization, shown by the red • markers in Fig. 15. As expected, the normal
condition data, relating to frictional processes, have a low discordancy for the
training and validation sets; this supports the assumption that these data are sampled
from the same underlying distribution, f. On the other hand, the crack-related AE
signals show higher measures of discordancy, suggesting they were generated by
some alternative and novel mechanism.

4.3.2 Outlier Analysis as One-Class Classifiers

Traditionally in SHM, outlier analysis is referred to as unsupervised learning [1].
This is true in the sense that the data labels, y, are not used explicitly with the
18 Structural Health Monitoring and Damage Identification 1031

input data, x, to learn a predictive model; instead, the model is built using the input
data alone. However, in the exclusive setting (which has been demonstrated in the
AE example), the training data can be assumed to represent the normal condition
only [116], that is, one class of data. Considering this implicit labeling, it can be
useful to consider outlier analysis and novelty detection tools techniques as one-
class classifiers [123]; in other words, the measurements are either from the normal
class (inlying) or they are not (outlying).

4.4 The Problem of Inclusive Outliers: Robust Outlier Analysis

As discussed, the previous example considered exclusive outlier analysis, such that
a set of data can be assumed to be representative of the normal condition. While
this is often valid for practical applications of SHM (e.g., the initial measurements
from a system should represent the undamaged structure) there are scenarios where
outliers can be hidden within the training data [116, 117]. Typically, these outlying
data are referred to as inclusive outliers, and their presence can contaminate the
parameter estimates μ̂,  ˆ . In the context of SHM, the inclusive problem can occur
when data are recorded over a range of operational and damage conditions without
descriptive labeling (unsupervised learning). For example, when considering the AE
data, this occurs when the normal condition data, shown by red markers in Fig. 15a,
are undefined.
If inclusive outliers are not considered when defining a model for novelty
detection, they can significantly influence the parameter estimates of f, leading to
masking or swamping effects [124]. Masking is caused by inclusive outliers that lead
to an increased covariance, ; ˆ thus, these outliers can mask the detection of future
anomalies (leading to false negatives). Alternatively, outliers can shift the model
location μ̂, leading to swamping, causing normal data to appear as outlying (false
positives) [124]. An example of masking with the AE data is shown in Fig. 16; in
this example, almost all of the crack-related data appear as inlying, which is clearly
uninformative.

4.4.1 The Minimum Covariance Determinant

Outlier tools that utilize robust statistics [125, 119, 120, 121] look to account for,
and expose, inclusive outliers that are hidden within the dataset. Robust statistical
methods were introduced into the field of engineering/SHM in [117]. Roughly
speaking, these algorithms accurately estimate f by finding which h-subset of
observations, H, (from the available data) have been generated by the normal
condition, such that the size of the set (cardinality) is |H| = h. The optimal h-subset
can then be used to determine robust estimates of μ̂ and . ˆ
In the existing literature, two typical ways to define H consider the minimum
volume enclosing ellipsoid (MVEE) [121] or minimum covariance determinant
(MCD) [120]. The MVEE approach defines H by searching for the smallest-
volume ellipsoid that encapsulates h observations in the feature space. Alternatively,
1032 R. Fuentes et al.

Fig. 16 The problem of inclusive outliers when using standard MSD outlier analysis: (a)
Observations in the feature space. The maximum likelihood estimate of f is shown by the sample
mean (+) and covariance (dotted lines represent two and three-sigma), (b) the MSD for each of
the 91 burst signals

MCD methods define H as the subset whose covariance matrix has the minimum
determinant. Both definitions can be interpreted as a way to describe the (majority)
h-subset that is the most concentrated in the feature space [121]. Intuitively, this
group is assumed to be generated by the same underlying mechanism, f.
The fast-MCD algorithm is applied to the AE data for demonstration; for further
details relating to the search method used to approximate the optimal h-subset, refer
to [125]. Figure 17 illustrates how the AE burst signals corresponding to crack-
related events are successfully flagged as outliers, despite these inclusive outlying
data being hidden within the training set. The robust approximation of the normal
condition distribution, f, is shown in Fig. 17a, and the associated MSD for each
observation is shown in Fig. 17b.

4.5 Probabilistic Classification through Supervised Learning

Often it is necessary to classify data into multiple groups, rather than simply inlying
or outlying (one-class classifiers). For multi-class classification, each observation,
xi , is associated with a label, yi , which categorizes the observations into K groups,
18 Structural Health Monitoring and Damage Identification 1033

Fig. 17 Outlier analysis using the robust MCD algorithm: (a) Observations in the feature space.
The maximum likelihood estimate of f is shown by robust estimates of the mean (+) and
covariance (dotted lines represent two- and three-sigma), (b) the robust MSD for each of the 91
burst signals

such that yi ∈ {1, . . . , K}. Predictive class labels for unseen observations can then
be used to inform the SHM strategy.
In the context of the AE data, the 91 observations can be approximately split into
three classes, such that a group of observations are associated with [113]:

• Class 1 – frictional processes away from the crack (clamping in the experimental
setup)
• Class 2 – crack-related events (crack extension and crack-face rubbing)
• Class 3 – crack-related events at a distance from the sensor

This labeled dataset is illustrated in Fig. 18. Continuing with the parametric,
statistical approach, it is useful to assume that the measured data can be represented
by some random variable, X; however, for the multiclass problem, it makes sense
to consider X as a mixture of Gaussians, or a Gaussian mixture model (GMM).
Following this approach, each of the K classes can be defined by its own Gaussian
distribution, fy ,


x| y ∼ N μy ,  y , (26)
1034 R. Fuentes et al.

Fig. 18 Multiclass classification of the AE data: (a) Observations in the feature space, illustrating
the training set (× markers) and the test set (• markers), (b) model predictions; the maximum a
posteriori (MAP) estimate of f is shown by the MAP of the mean (+) and covariance (dotted lines
represent two- and three-sigma)


∴ , p ( x| Y = y) = N μy ,  y = fy (27)

A distinct pair of parameters (μy ,  y ) are used to define the distribution of X for
each of the K classes; note, for the AE data, K = 3.
Furthermore, it is useful to consider that the label space can be defined by some
random variable, Y. In this case, as the labels are discrete, Y is assumed to be
categorically distributed; for more details behind this intuition, the reader is referred
to [126],

y ∼ Cat (λ), (28)

where λ = {λ1 , . . . , λK } are the mixing proportions for each class y ∈ Y, such that,

p(y) = λy = P (Y = y) ∀y ∈ Y. (29)

A maximum likelihood approach can be used to approximate the parameters {μy ,

Σ y , λy } of the distributions p(x|y) and p(y) (as in the outlier analysis example). How-
ever, in order to improve model generalization, and to avoid validation procedures,
18 Structural Health Monitoring and Damage Identification 1035

a Bayesian approach is adopted to estimate the parameters of the mixture model.

This involves considering the parameters to be random variables themselves, and
incorporating a prior belief over the parameter estimates to prevent overtraining,
particularly with small datasets. For details explaining the Bayesian approach to
statistical modeling, the reader is referred to [126].
Having approximated the parameters of the mixture model, a generative Bayes’
classifier can be defined to predict the probability of a label y ∈ Y for unseen
observation x̂ [126],
 
 p x̂ y p(y)
p y| x̂ =  , (30)
p x̂

Note: this generative model is a Baye’s classifier; however, it is not a fully

Bayesian method. For this to be the case, the full posterior distribution over
each label estimate p(y|x) would have to be defined analytically or approxi-
mated. For more information on fully Bayesian models, the reader is directed
to [127, 126].
 
As p x̂ is a normalizing constant (to ensure the integral of the pdf p y| x̂
is one), the predicted label, ŷ, can be defined as the most likely label given the
observation x̂,
   
ŷ = argmax p x̂ y p(y) (31)

The resulting multiclass classification for the AE data is shown in Fig. 18. A
random sample of 30% of the total data is used to train the algorithm (× markers),
while the model predicts the label for the remaining 70% (• markers).

4.6 The Problem of Feature Dimensionality

In SHM, the measured data are often high-dimensional (e.g., vibration observa-
tions). As a result, even large volumes of data records can be sparse in their feature
space, leading to a poor representation of the distribution of data, and insufficient
information to build a reliable model. This phenomenon is referred to as the curse
of dimensionality [128, 115].
In the context of outlier analysis, the distance measures (used to define outliers)
can lose their meaning for sparse data in high dimensions [128]; specifically, it has
been shown that the magnitude of the distances between any pair of observations
can become similar [128]. In this case, any observation might be considered as a
potential outlier.
To combat issues of dimensionality, feature selection tools look to identify a low-
dimensional subset of variables from the measured data that are sensitive to damage
[1]. These low-dimensional data can then be used to build pattern recognition
models. Sensitivity analysis [1] of variables over the input data can help identify
1036 R. Fuentes et al.

representative features objectively [129]. Alternatively, the use of genetic algorithms

(GA) has been shown to provide promising results when applied to vibration data
[130].
Dimension reduction techniques offer another method for data compression,
while retaining as much information as possible from the full feature space. Linear
PCA is typically used, which has been demonstrated with the AE data example.
Alternatively, nonlinear variations include kernel-PCA [126] and auto-encoder
networks [131]. When applying these techniques, the resulting features can be
interpreted such that variation within the available data is maximized. While these
methods are highly effective, a feature projection which maximizes variance is not
necessarily indicative of damage, as any major changes in the data might relate
to environmental conditions or other permissible variations during operation. With
this is mind, engineering judgment, along with prior knowledge of the type of data
and the application, is required during signal processing to ensure sensible and
informative features are extracted from the measured data.

4.7 Outstanding Challenges in Data-Driven SHM

One of the most significant challenges for data-based SHM is a lack of data [132].
For example, in order to define a complete labeled dataset for an engineering struc-
ture, the system must be measured across all operational and damage conditions,
while the structure is regularly inspected by an engineer to annotate the measured
data. Additionally, the dataset recorded from one structure is not necessarily relevant
to another (nominally) identical one. Therefore, traditional supervised learning of
expensive systems (such as aerospace or civil structures) is clearly impractical and
infeasible. Currently, this fact forces a dependence on unsupervised techniques in
many practical applications, specifically, novelty detection. An alternative approach,
however, is to utilize novel learning strategies for SHM, which can make use
of datasets with limited information or annotation. These tools are emerging
technologies in the SHM literature, including semi-supervised and active learning
[132, 133].

5 SHM in Changing Environmental and Operational

Conditions

This section discusses how the influence of changing environmental and operational
conditions can be problematic when attempting to infer structural condition from
monitoring data.
As previously alluded to, the effect of changing environmental and operational
conditions on a structure is an important issue in SHM, and has been identified
as a key concern to the research community [134]. This interest arises from the
inconvenient fact that measured responses from a structure that demonstrate sensi-
tivity to damage or structural degradation, will, in general, also exhibit sensitivity
18 Structural Health Monitoring and Damage Identification 1037

to any change in operational and environmental conditions [19]. In these cases,

the confounding effects of the environmental and operational variation must be
accounted for in some way before a reliable measure of structural condition can be
inferred. The problem is often referred to as the data normalization problem within
the SHM community [16].
In the SHM literature, undoubtedly the most commonly occurring discussion of
confounding influences on damage-sensitive features that arises from the sensitivity
of structural response to temperature. For civil infrastructure, for example, tem-
perature is generally considered to be the dominant environmental factor affecting
the normal dynamic response, due to its effect on the stiffness of structural
components, and also its potential effect on the boundary conditions of a structure
(for instance from the freezing of foundations etc.). Historically, many studies have
found fluctuations in modal frequencies to be correlated with ambient temperature,
although different mechanisms have been used to explain this, see for example
[135, 136, 137, 138]. Cornwell et al. [136] suggested that the thermal gradient
across the deck of the Alamosa Canyon Bridge drives the observed fluctuations in
modal frequency. In colder climates, significant shifts in frequency between above
and below freezing temperatures have been attributed to an increase in stiffness
explained by the Young’s modulus of the asphalt on the deck at colder temperatures
[137]. In this case, the modal frequencies of the Z24 bridge deck (Switzerland) were
observed to have a bilinear relationship with temperature. A similar behavior has
also been observed in a steel truss footbridge in the USA [139]. More recently, Xu
et al. [140] showed that the displacement of the Tsing Ma Bridge deck was linearly
related to the changes of ambient temperature, after conducting a 6-year monitoring
campaign. While investigating the thermal effects on the Zhenjiang Bay Bridge
(a cable-stayed bridge with a 480 m main span) in China, Cao et al. discovered that
the effect of temperature differed across the structure and that the temperature in the
concrete lagged 5–6 h behind the measured ambient air temperature [141]. Similarly
the monitoring campaign of the Tamar Suspension Bridge in the UK revealed that
the structural temperature lagged 10–60 min behind air temperature depending on
which part of the bridge was measured [142].
Besides temperature, the importance of other environmental and operational
conditions has also been considered. Wind-induced vibration is a critical factor
for the design and maintenance of tall and slender structures like wind turbines,
skyscrapers, and long-span bridges. A suspension bridge in Japan was tested by
Mahmoud et al. [143]; they found that the natural frequencies, mode shapes, and
damping ratios were all related to the variations in the ambient wind. Cross et al.
[144] observed that the modal frequencies of the Tamar Bridge in the UK were
not significantly correlated to wind speed unless the wind speed exceeded 25 mph,
when a modal frequency dependency on the amplitude of deck vibration could
be observed. The effect of humidity alongside temperature has also been studied.
In [145], the effect of humidity and temperature on the modal parameters of a
reinforced concrete slab was investigated; it was found that increased humidity
effectively adds mass to a structure and, therefore, has a strong negative correlation
with modal frequency.
1038 R. Fuentes et al.

model estimate
1.5
SSI frequency

1
Normalised natural frequency

0.5

−0.5

−1

−1.5

−2

06/11 07/11 08/11 09/11 10/11 11/11

Date of measurement

Fig. 19 Linear regression model of first deck modal frequency of the Tamar Bridge, with traffic
loading input only [144]

Operational conditions in SHM can be considered to be a broad terminology

that may include factors like traffic loading, flight conditions, operating speed,
and payloads, which may vary with time, resulting in significant influences on
the response and dynamic properties of structures of interest. The effect of traffic
loading is of particular interest for bridges, where the modal properties of bridge
decks have been found to be more or less influenced depending on the mass of
the traffic relative to that of the structure itself. For example, the effect of traffic
loading has been addressed in [146], where, for long-span bridges, the influence
of traffic loading on the structure’s modal frequencies was considered negligible
due to the fact that the mass of a single vehicle is very small in comparison to the
mass of the “superstructure.” For the Tamar Suspension Bridge, the fundamental
modal frequency of the deck has been shown to vary linearly with the estimated
instantaneous traffic load. Figure 19 shows how a simple linear regression model
with traffic load as an input can reproduce the variation of the fundamental modal
frequency of the deck [144].

5.1 Removing Confounding Influences

A review of the relevant literature reveals a number of potential options explored

for dealing with the problem of operational or environmentally induced variations
in structural response. Perhaps the most common approach has been to attempt
18 Structural Health Monitoring and Damage Identification 1039

to model the monitored parameters or damage sensitive features in question with

respect to those environmental/operational factors considered to be driving its/their
variation [147, 148, 149, 139, 150, 151, 152, 137, 153]. If a model can predict
the value of a damage-sensitive feature given the conditions affecting it, the error
of the model could be suitable as a robust indicator of structural condition. Often
these approaches have employed a simple regression of the damage-sensitive feature
(normally modal frequencies) onto measured structural temperature [148, 149, 139,
137, 153]. More complex approaches for regression have also been explored [150,
151, 152], where modal parameters of the Ting Kau bridge, Hong Kong, have
been regressed onto measured temperature using support vector machines, principal
component analysis, and neural networks. In a very similar vein, tracking the
correlation between the measured strain of a harbor wall and temperature has been
explored in [147].
For complex relationships between measurands, where it might be beneficial to
model their functional relationship in distinct regimes, an alternative approach has
been suggested that combines Classification and Regression Trees (CART) with a
nonlinear regression model (particularly Gaussian process (CP) regression in this
case) [154]. A treed-GP partitions the input domain into different regimes (leaves)
using a Bayesian CART algorithm and employs a different GP in each regime. The
GP can also be extended to the heteroscedastic context (the variance of the noise is
time dependent); see for example [155].
In the approaches described thus far, the main limiting factor is that the changing
environmental and operational conditions have to be identified and accurately
measured. While this may be feasible where only one or two environmental or
operational factors are important, such as temperature, where multiple factors
affect the features of interest, a substantial monitoring campaign will then become
necessary.
Other approaches, which do not rely on measurements of the environmen-
tal/operational conditions being available, have also been explored. A simple
potential solution to the problem is to use a long span of response data to define the
normal condition of a system, an idea explored in [156]. This could be, for example,
data collected over a whole year, where all ranges of environmental/operational
conditions have occurred. New measurements may then be compared in some way
with the defined normal condition. Evidently this approach requires storage of a
large amount of data, and a further drawback is that using a large normal condition
set may reduce feature sensitivity to damage [153]. This issue was explored in the
context of damage detection in composite panels in [157].
A number of other studies employ what may be described as projection methods
or latent variable models [158], which, without measurement of the changing envi-
ronment, attempt to capture the variation in the feature data caused by it. PCA has
been used in a number of studies to re-express multivariate SHM feature data with a
new set of orthogonal coordinate axes [159, 160]. The assumption employed in these
studies is that the high-variance signatures of changes induced by environmental
and operational conditions in SHM features will be trapped in the higher variance
principal components. In [159], this assumption is exploited by discarding the
1040 R. Fuentes et al.

higher-variance principal components and projecting temperature-dependent data

onto the minor components which constitute a temperature-independent feature set.
In [160], only the higher-variance principal components are retained and used as
a model to predict/reconstruct the feature data. This idea of linear projection and
trapping of environmental variation was also independently proposed through the
use of factor analysis (FA), which is a very similar algorithm to PCA [161]. In the
next section the use of cointegration, an alternative projection method introduced
by a subset of the authors of this chapter, will be introduced and demonstrated. The
idea behind cointegration is to exploit the fact that some monitored variables will
share common trends induced by the changing environment; where variables bear
a linear relationship, these common trends can be removed by via a simple linear
combination of the variables themselves.
Although such approaches have proven very useful in some applications, it has
been shown that nonlinearity can hamper the effectiveness of employing the projec-
tion methods discussed above [162]. In [163], a remedy to problems introduced by
nonlinearity was to cluster feature data into several (linear) regions and then employ
PCA separately to each region. An auto-associative neural network, which may be
said to be equivalent to nonlinear PCA, is used in [164] for data normalization of
features extracted from an autoregressive type model. An auto-associative neural
network (nonlinearly) maps its inputs onto themselves. The premise of using them
for data normalization is that, if the network is trained on data from an undamaged
structural condition, it will learn the effect of latent variation on the features input
to the network. It is then expected that the network error will increase if damage
occurs. Dervilis et al. [165] proposed to improve the robustness and efficiency of
AANNs by adopting an auto-associator using a radial basis function (RBF). The
main difference between a traditional multilayer perceptron (MLP) network and
RBF network is that, instead of representing the nonlinear transfer function with
the scalar products of the input vector and the weight vector, the RBF network
adopts a nonlinear function of the distance (Euclidean distance in this case) between
them, which is hugely advantageous because an RBF network does not require a
cumbersome nonlinear optimization for the model parameters. The RBF network
was examined with an experimental case study to detect damage in a wind turbine
and was proved to be a fast online damage detection algorithm. Shi et al. introduce
nonlinear cointegration for SHM in [166, 167], also discussed in the next section.
Along similar lines, a new approach for data normalization has recently emerged,
where, for multiple sensor arrays, Gaussian Process regression is used to predict
the measurement of each single sensor from the measurements of all other sensors
in the network [161]. Given suitable training data from different environmental
and operational conditions, the GP should be able to accurately predict structural
response at each sensor if the structure continues to operate in a similar way as in
the period where the training data were recorded. In a similar way to the regression
techniques, the GP regression model error is used as an indicator of abnormal
structural response.
A final alternative, rather than treating the management of EOVs as a regression
task, is to manage these changes as a classification task. That is, to separate the
18 Structural Health Monitoring and Damage Identification 1041

behavior of the system into discrete regimes based upon the current environmental
or operational state. This approach has been explored in the literature; in [18] in
the context of mixture modeling and in [168], where an online clustering approach
detects, and subsequently is used to classify, emerging behavioral regimes.

5.2 Linear and Nonlinear Cointegration

Where measurements of changing environmental and operational conditions are not

readily available, projection methods, as discussed above, have proven to be useful
for removing their influences from damage-sensitive features. This section briefly
introduces and demonstrates the idea of cointegration for removal of trends from
SHM data, as one of the most promising of these methods (the bias of the authors
should not go unnoted here). The full details of the theory can be found in [169].
Cointegration originates from the field of econometrics and is, in fact, a property
of nonstationary time series. Informally speaking, two or more nonstationary time
series are cointegrated if some linear combination of them is stationary. Economists
traditionally test for cointegration between various econometric time series as a
means of determining whether there is a statistically significant relationship between
them. Although engineers may well be interested in problems of a similar nature, it
is the stationary linear combination created during the cointegration process that
is of practical interest here. If a number of variables from some process under
investigation are cointegrated, the stationary linear combination of them found
during the cointegration process will be purged of all common trends in the original
datasets, leaving a residual equivalent to the long-run dynamic equilibrium of the
process. For measurements from structural monitoring campaigns, it is likely that
the nonstationarity in each variable will originate from the same cause, such as
a temperature changes. This means that cointegration can be used to remove the
common trends in the measured variables originating from confounding influences.
As an example, Fig. 20 shows 2 months’ worth of measurements of deck and
tower displacement of the Tamar Bridge (with any gaps in the record removed).
These measurements were taken by the Vibration Engineering Section at the
University of Exeter using a total position station (TPS); more details can be found
in [170]. In the figure, one can see an oscillatory trend and a large mean shift
occurring over a longer time period. The oscillatory trend is a daily trend induced
by temperature variation, the mean shift is also temperature induced and occurs as
the season changes and the environment gets colder [171]. Figure 21 shows these
variables under a linear combination established using just 3 days’ worth of data.
The combination remains stationary for the duration of the period and one can
clearly see that the seasonal trend visible in Fig. 20 has been purged.
For use in SHM, if the dataset used to establish this combination is representative
of the normal condition of the structure, the linear combination can be used as
a diagnostic tool by projecting all new data onto it. In the event that the residual
becomes nonstationary, the structure can be said to have departed from its normal
condition, or be damaged – if damage sensitive features make up the linear
1042 R. Fuentes et al.

Hanger 44
4 Hanger 62
Hanger 80
Hanger 98
2
Hanger 112
Hanger 123
Normalised easterly displacements

0 Top of Plymouth side tower

Top of south Plymouth tower
−2 Top of north Plymouth tower

−4

−6

−8

−10

−12

−14
0 500 1000 1500 2000 2500 3000
data point reference

Fig. 20 TPS measurements of the Tamar Bridge deck and tower displacements

30

20

10

−10

−20

−30
0 500 1000 1500 2000 2500 3000
Sample point number

Fig. 21 Linear combination of variables shown in Fig. 20, along with error bars at plus and minus
three standard deviations of the combination, added to act as a statistical process control chart
18 Structural Health Monitoring and Damage Identification 1043

combination. In the following, a formal definition of cointegration is provided along

with a brief overview of how the most stationary combination of variables of interest
may be found.
Definition A set of nonstationary variables, say yi , are cointegrated if some linear
combination zi of them is stationary,

zi = β T yi . (32)

If some β can be found such that zi is stationary, then β is called the cointegrating
vector. If yi includes a total of n variables, there may be as many as n – 1 linearly
independent cointegrating vectors.
Some constraints fall on the nonstationary variables in yi if they are to be
cointegrated; they must share common trends and they must also be “integrated”
to the same order. A nonstationary time series, y, is integrated order d, denoted
y ∼ I(d) if, after differencing the series d times, it becomes stationary. In essence,
each time series must have the same degree of nonstationarity. For the purposes
of SHM, the intent would be to use monitored variables that are cointegrated and
find a cointegrating vector that will create a stationary residual sequence suitable
to be used as a damage-sensitive feature. While it is likely that variables measured
from the same structure will share common trends, this cannot be said of the order of
integration of each monitored variable; this should be ascertained before any attempt
is made to find the cointegrating vector. This is commonly achieved by using the
Augmented Dickey Fuller test, which is a unit root test [172, 173].
Once it has been ascertained to what order all process variables of interest are
integrated to, it remains to find the cointegrating vector that will result in the most
stationary combination of them. There are two common approaches for this in
econometrics; the first is the Engle-Granger two step estimation procedure [174],
often employed when there are only two process variables included in the analysis;
the second is the Johansen procedure [175], a more complex maximum-likelihood
multivariate estimation procedure. The Johansen procedure is typically chosen when
attempting to combine SHM variables. A quick outline of the theory behind finding
the cointegrating vector will be briefly described below, but readers are referred to
[169] for more detail.

Finding the Cointegration Vector

The Johansen procedure uses a maximum-likelihood approach to estimate the
parameters of a Vector Error-Correction Model (VECM) of the variables under
consideration. A VECM takes the form,


p−1
yi = yi−1 + Bj yi−j + φD(t) + εi , (33)
j =1

where yi denotes an n-vector including all n variables to be analyzed, with the

subscript i relating to time, i = 1, . . . N, p represents the model order, or the number
1044 R. Fuentes et al.

of lags to be included in the model, and εi is a normally distributed noise process;

εi ∼ N (0, ). A term to describe a deterministic trend D(t) can also been included.
If the variables yi are cointegrated, then with the correct parameter choice, yi
will be stationary. It is shown in [175] that (roughly) a sufficient condition for a
cointegrating vector to exist is that the matrix  should be rank-deficient. Suppose,
that the n × n matrix [] has rank n – r, then a basic theorem of linear algebra
asserts that it will admit a decomposition  = αβ T where α and β are both n × r
in dimension. It transpires that the columns of β are (up to a linear transformation)
the desired cointegrating vectors. In principle then, construction of the cointegrating
vectors looks fairly straightforward; one constructs a VECM, then once the matrix
 is found, its decomposition leads directly to β. Unfortunately, things are not quite
so straightforward; because  is rank-deficient, standard least-squares regression
procedures cannot be applied to the parameter estimation problem. A more complex
reduced-rank regression approach is required; this can be summarized in terms of
the Johansen procedure, detailed in references [169, 175].

5.2.1 Using Cointegration for SHM

Figures 20 and 21 showed how a temperature trend could be purged from measure-
ments of bridge deck and tower displacement using cointegration. As the bridge
is in a healthy state and performs well, another example is used here to illustrate
how cointegration can be helpful for damage detection. In [157], a dataset from
a Lamb wave inspection of a composite plate in a damaged and undamaged state
is used. In the experiment, which was carried out under the Brite-Euram project
DAMASCOS (BE97 4213), the panel was subjected to a fluctuating temperature in
an environmental chamber; details of the experimental procedure can be found in
[176]. For the example shown here, the features extracted for damage detection were
the amplitude of 20 spectral lines from the area around the peak of the frequency
spectrum. A time history of these spectral line amplitudes is plotted in Fig. 22. Here
one can see that the features fluctuate throughout the test. Separated by the thick
vertical lines, one can also see three regimes, the first is when the temperature in the
environmental chamber is held steady, the second when the temperature is fluctuated
periodically, the third is when damage was introduced with the temperature change
continuing. The challenge is to be able to automatically detect the introduction of
the damage.
In this example, a training data period used to establish the most stationary
cointegrating vector was selected as data points 1000–2000 in Fig. 22; this includes
355 data points from the steady temperature regime and data from the fluctuating
temperature regime. The selection of the training data for cointegration essentially
establishes the “normal condition” for the structure. All remaining data are projected
onto the cointegrating vector found from the training dataset and the stationarity of
the resulting combination/residual assessed. If the residual becomes nonstationary,
this indicates that the relationship between the variables combined has changed.
Figure 23 shows all of the feature data projected onto the most stationary linear
combination of the training data found via the Johansen procedure. Here, one can
see that the residual has successfully been purged of the temperature trends visible in
18 Structural Health Monitoring and Damage Identification 1045

0.9

0.8

0.7
Feature amplitude

0.6

0.5

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000
Sample point number

Fig. 22 Time history of 20 dimensional feature from a Lamb wave inspection of a composite plate

20

15

10

−5

−10
0 500 1000 1500 2000 2500 3000
Sample Point Number

Fig. 23 Cointegrated signal (linear combination of 20 spectral lines)

1046 R. Fuentes et al.

the original data and that, on the introduction of damage, a large mean shift is visible
and detected by control chart limits (which were set at three standard deviations of
the residual in the training period). For more details and for a comparison between
cointegration and a PCA-based projection method, interested readers are referred
to [157].

5.3 Nonlinear Cointegration

In the previous section, the power of the cointegration method for removing
environmental trends from damage-sensitive features was demonstrated. This has
been shown to work well in a number of situations; it is limited, however, by
its linear nature. The method is only valid where the variables of interest are
linearly related during their normal condition. Alternative approaches must be
sought where variables are nonlinearly related in an undamaged condition. An
example of this comes from the famous Z24 benchmark dataset; here, the extracted
modal frequencies have a nonlinear relationship which can be attributed to the
change in the structure when sub-zero temperatures were experienced.
In this case a nonlinear combination of response variables is needed to remove
unwanted environmental and operational trends. A number of possibilities for
nonlinear cointegration for SHM are explored in [162]. Figure 24 is an example
of just one of these approaches. Here the modal frequencies of the Z24 bridge have
been nonlinearly combined using GP Regression. Here the second modal frequency

Fig. 24 Nonlinear cointegration of the Z24 deck modal frequencies; the vertical dotted line
indicates the time at which damage was introduced into the structure
18 Structural Health Monitoring and Damage Identification 1047

has been chosen as a target for the regression, where the model inputs are the
remaining modal frequencies. One can see that the GP regression, established using
datapoints from the undamaged condition, is able to predict the large trends induced
by cold temperatures but remains unchanged after the introduction of damage. The
model error would now be a good candidate for a damage-sensitive feature that has
been purged of its temperature dependency.

6 Physics-Based Models in SHM

This section provides an overview of methodologies that utilize physics-based

models in order to perform damage state inferences. Traditionally this category has
been defined as model-driven (or physics-based) as these approaches use some form
of physical law-based model, unlike data-driven techniques [19, 177, 178, 179]. In a
historical context, this category referred solely to approaches that combined physics-
based models with inverse techniques. These methods then posed the problem of
SHM as inferring (or “updating”) a set of model parameters using operational data,
leading to decisions about structural integrity, based on the interpretation of these
“updated” model parameter values [180, 10]. Herein, these methods are classed
as inverse model-driven. However, another subdivision exists within the general
model-driven category, known as forward model-driven. These approaches seek
to integrate the physics and data-based techniques, in which physics-based models
provide some level of training data or labels to traditional data-driven approaches
[181, 182]. The focus of this section is on outlining the two philosophies, providing
a review of methodologies, and highlighting outstanding challenges.

6.1 Inverse Model-Driven SHM

Inverse model-driven methods make decisions about structural integrity using

“updated” model parameters inferred from in-service data. The application of these
methods often involves the construction of a high-fidelity model of the structure,
for which structural condition decisions are to be made, typically in the form of an
FEA model. The assumption is that the model is a satisfactory representation of the
structure, such that changes in the data due to damage will correspond to changes in
a defined model parameter set.
There are two main procedures for updating a model in order to make health
decisions; the first is a two-step process, where initially the model is calibrated
so that it more accurately represents the structure in question. This is generally
performed by model updating, based on in-service data of the undamaged condition.
In the second stage, new in-service monitoring data from an unknown structural
condition is used to update the model again. After the update, changes in the
inferred model parameters from the baseline calibration are used to perform damage
identification [183]. The second stage is to match response deviations between the
undamaged state and the unknown state of the structure. These two procedures
1048 R. Fuentes et al.

can be used to perform damage identification at levels 1–4 of Rytters hierarchy.

Prognosis may also be achievable because an updated physics-based model is
generated through the inverse model-driven procedure [184].
SHM via an inverse model-driven approach relies on model updating or parame-
ter estimation methods. Model updating refers to techniques where certain model
parameters are adjusted such that the residual between observational data and
model predictions is minimized [185]. This task is broadly attempted in two
general approaches: direct methods, where structural matrices are updated to
reproduce measured data, and sensitivity methods, where errors between predictions
and observations, are minimized via changing a set of defined parameters [185,
186]. Commonly in SHM, sensitivity-based techniques are selected over direct
approaches; this is because attempting to update full structural matrices within a
direct approach often leads to a lack of control over the updated matrix values,
leading to inferred parameters with little physical meaning.
Initial development of model updating methodologies approached the problem
from a deterministic view, for example, the well-established iterative sensitivity-
based method [187]. Such techniques approached the problem of model updating
using optimization technologies, whereby a cost function is developed, typically
in a least-squares formulation, and parameter steps made via sensitivity matrices
[188, 187]. However, these approaches require regularization due to the problem of
model updating being ill-posed [10]. These deterministic methods also have diffi-
culties in handling variability and uncertainties that are present, for example, from
environmental conditions, parametric variability, and model-form uncertainties. For
these reasons, alternative frameworks for approaching model updating have been
developed.
Two popular philosophical approaches for handling uncertainties within model
updating are fuzzy and Bayesian methods [189]. Fuzzy techniques are non-
probabilistic approaches that transform uncertainties into fuzzy inputs, that is, as a
fuzzy number – a quantity that is characterized by a membership function – and then
perform multiple optimization problems [190]. Fuzzy model updating technologies
assume that the fuzzy input variables are independent and equally likely, which will
result in the worst-case range of parameters being inferred. Bayesian methods, per
contra, take a probabilistic view of parameter estimation, using Bayes’ theorem to
update model parameters and their uncertainties. In certain scenarios, these methods
contain inherent model regularization contained within the marginal likelihood,
sometimes referred to as the Bayesian Occam’s razor [191]. Beck and Katafygiotis
provide a review of Bayesian model updating [192, 193]. Nonetheless, most of the
current model updating methodologies fail to account for uncertainties associated
with model form errors, known as model discrepancy. Failure to consider this form
of uncertainty will often lead to bias in the estimated parameters, and therefore
incorrect health statements in an inverse model-driven context.
Inverse model-driven technologies suffer from several challenges when imple-
mented as part of an SHM strategy. Firstly, the type and number of parameters to
18 Structural Health Monitoring and Damage Identification 1049

use must be selected [194, 195]. In scenarios where damage is unknown, as is often
the case, this can lead to an especially large number of parameters. Parametrization
becomes increasingly challenging as model fidelity increases, where there are a
large number of potential parameter sets. Another difficulty is that of interpreting the
updated parameters to make a decision about the structure’s health. This problem
can be especially difficult when parameters affect structural stiffness, as multiple
phenomena influence changes in stiffness. An accurate understanding of the physics
must inform whether updated parameters are no longer physically meaningful rather
than altered by the presence of damage, and constraints placed on the updating
process when this is the case. As mentioned, variability and uncertainties within
the “target” data must be handled as part of the updating process. Moreover, these
issues are confounded by the problem that a solution, or a unique stable solution,
for the inverse approach cannot always be achieved due to ill-conditioning. These
non-identifiability issues become of increasing concern when the parameter values
are being used for health diagnostics, as repeats of the update may lead to different
conclusions.

6.2 Forward Model-Driven SHM

Forward model-driven techniques utilize physics-based models in a forward manner,

whereby their predictions form training data, in a supervised-learning context, or
a labeling method, in a semi-supervised procedure [181, 182]. Health decisions
are subsequently made based on the inferred classifier using standard data-driven
techniques. These classifiers can be trained using models alone, or a combination
of model outputs and data. The main motivation for this class of methods is in
improving the lack of available damage state data problem, by generating it from
physics-based models, and removing issues associated with interpreting health
states from updated model parameters.
Forward model-driven methods are comprised of two main components: gener-
ating representative damage state features from computer models, and using those
predictions to train machine learning or pattern recognition approaches. The second
component, well studied within the data-driven category of SHM, has been demon-
strated to be effective when labeled damage state data are available. Within a for-
ward model-driven approach, these techniques generally remain mathematically and
algorithmically the same, with the only difference arising from the source of training
data, that is, computer model-based predictions. Consequently, the major challenges
in establishing a forward model-driven strategy are in developing methodologies
and technologies that achieve the objective of the first component, namely the
generation of representative damage-state features from computer models.
Few examples of forward model-driven approaches exist within the literature.
FEA models have been used to generate features for ANNs in performing damage
identification in bridges [196, 197]. Satpal et al. implemented a combined model
1050 R. Fuentes et al.

updating and SVM approach, where model predictions trained the classifier [198],
with Hariri-Ardebili and Pourkamali-Anaraki applying a similar methodology to
concrete dams [199]. Most of these approaches utilize deterministic FEA model
outputs, with a few adding arbitrary noise terms to replicate variability, while others
propagate “known” parameter uncertainties through Monte Carlo realizations.
None of these methods consider model form errors, and either do not attempt to
validate their models or implement full-system damage state data in the validation
process.
Generating representative predictions from computer models means addressing
several key challenges and will provide improvements to the existing forward
model-driven literature. Firstly, there must be a method for determining whether
computer model predictions of health states are representative of those obtained
operationally. This requires the definition of what a valid computer model prediction
is within the forward model-driven context. In order to develop this definition an
understanding of how these predictions are used within classification methods must
be established. In a data-driven framework, features extracted from operational data
are often employed in training decision bounds that capture the expected behavior
of the particular damage feature under each damage scenario in the training set.
This requirement means that health state data generated from computer model
predictions must capture the inherent variability and progression of the health state
in question. A computer model will therefore be valid if its predictions generate
statistical distributions of health states that are statistically similar to those from
observations.
Secondly, generating statistically representative predictions will involve some
level of calibration and a validation procedure. Unfortunately, both these processes
require data from the real-world structure, leaving the conundrum of how to cali-
brate and validate the computer model given that structural condition data is neither
feasible to obtain nor cost-effective in the majority of applications. If this question
is not addressed, forward model-driven approaches simply become an expensive
and demanding way to perform substandard data-driven SHM, introducing further
approximations and modeling challenges. One solution to this problem is the
division of the structure in question, and hence the computer model, into a set of
components – sub-assembly, etc. – for which obtaining health state data is feasible
and economically viable. In this scenario a full system, such as an aeroplane, is
divided into various subsystems, for example, wing panels, riveted joints, landing
gear assemblies, and coupons, where each subsystem can be tested with damage
types that are expected to be likely causes of failure in the full-system. Small-
scale test strategies can then be developed, or existing certification tests used to
collect datasets that can be implemented in calibrating and validating the set of
computer models. The usefulness of forward model-driven technologies rest on the
ability to utilize and integrate these subsystem datasets into calibrated and validated
subsystem level computer models, which, when propagated through to the full
system, via an algebra of computer models and uncertainty management, produce
valid, that is, statistically representative predictions, which have required no full-
system health state data. Obviously this is an incredibly ambitious goal; nonetheless
18 Structural Health Monitoring and Damage Identification 1051

methods such as multilevel uncertainty integration strategies offer techniques for

undertaking such a challenge [200, 201, 202].
Thirdly, procedures for calibrating computer models naturally have to involve
mechanisms for handling multiple sources of uncertainty, especially those from
model form errors, known as model discrepancy [203, 204, 205]. Statistically rep-
resentative predictions will not often be achievable without capturing observational
variability, along with parameter uncertainties and accounting for any functional
model discrepancy – the differences between computer model outputs and observa-
tional data. At the point of writing, forward-model driven SHM is still a developing
field, but one which, by addressing these challenges could become a solution to
some of the key difficulties facing SHM, such as the scarcity of training data.

7 Summary

This chapter has presented a general overview of data-driven SHM, placing a

focus on signal processing and statistical learning techniques. The process of SHM,
which largely follows Rytters’ hierarchy of detection, localization, classification,
assessment, and prediction, has been outlined in terms of its various components.
The chapter has placed a strong emphasis on the detection stage; even this first
stage presents some difficult challenges and requires an appropriate combination
of operational evaluation, feature engineering, and machine learning algorithms to
succeed. Today, one of the key challenges in SHM is dealing with the confounding
influence of operational and environmental trends in SHM data. The chapter has
presented an overview of the techniques available to deal with this challenge,
focusing on removal of the confounding influences, as well as their identification
through techniques such as robust statistics. Another problem of interest in SHM is
the use of physical models to aid in the damage identification process. Physical
and phenomenological models often provide a level of insight into the system
dynamics that cannot be achieved with purely data-driven models. However,
physics-based models often fail to accurately predict the output of complex systems,
or handle operational and environmental changes. The last section of this chapter has
discussed the progress that has been made in using physical models in SHM.

7.1 Applications

For a long time, SHM has been confined to laboratory experiments and demon-
strations. However, the field is beginning to make a transition toward real-world
applications, within both the private and public sectors, with a plethora of applica-
tions materializing the usage of such monitoring systems. A significant number of
structures undergo routine inspections and maintenance in order to ensure structural
stability of the system. The costs of these routine inspections could be significantly
reduced if these inspections are shown to be unnecessary when a structure continues
to be healthy, and this could be indicated automatically by implementing an SHM
1052 R. Fuentes et al.

system. SHM could offer robust and online monitoring and necessary maintenance
or repairs could be addressed based on this technology. A very bold and strong
branch of industry in which SHM takes flesh is the offshore wind industry. A
good very recent example research initiative is provided by the “New Partnership
in Offshore Wind” – a UK Engineering and Physical Research Council (EPSRC)
Prosperity Partnership. This proposal brings together two major industrial players:
Siemens-Gamesa Renewable Energy and Ørsted, with world-leading academic
researchers in a £7.64 M, 5-year program in order to address fundamental research
challenges that will help to reduce the levelized cost of electricity (LCoE) from
offshore wind and to support UK supply chain growth. SHM is one of the core
interests in this major project.
Another (historically) major player in SHM applications is the offshore oil and
gas industry, where a platform may undergo routine maintenance or emergency
component replacement, which, in turn, would be an economic and environmental
drawback. Industries such as energy and aerospace have always been keen on
life extension of critical components beyond the originally designed fatigue life.
Applications of SHM systems are arising in additive manufacturing, biological
systems, telescopy, and even in monitoring of advanced infrastructure such as
accelerators. A good example of where this could be critical was highlighted in
the failure that occurred in the £5bn CERN Large Hadron Collider in September
2008 [206]. That problem delayed the restart of the experiment at a critical point.
SHM technology has the potential to benefit all sectors of industry concerned
with monitoring key infrastructure (consider the most recent events concerning the
Morandi Bridge in Genova, Italy). SHM provides the potential to move from time-
based inspection and maintenance into condition-based maintenance approaches.
The basic philosophy behind the condition-based maintenance is that a holistic
and robust sensor network will monitor the system and via smart measurement
processing will alert the operator about potential system abnormalities.

References
1. Farrar C, Worden K (2012) Structural health monitoring: a machine learning perspective.
Wiley, Hoboken
2. Farrar CR, Worden K (2007) An introduction to structural health monitoring. R Soc Philos
Trans Math Phys Eng Sci 365:303–315
3. Doebling S, Farrar C, Prime M, Shevitz D (1996) Damage identification and health moni-
toring of structural and mechanical systems from changes in their vibration characteristics: a
literature review. Technical report. Los Alamos National Laboratories, Los Alamos
4. Wenzel H (2009) The character of SHM in civil engineering. In: Boiler C, Chang F-K, Fujino
Y (eds) Encyclopedia of structural health monitoring. Wiley, Chichester, pp 2031–2037
5. Rytter A (1993) Vibrational based inspection of civil engineering structures. PhD thesis,
Department of Building Technology and Structural Engineering, Aalborg University
6. Worden K, Dulieu-Barton J (2004) Damage identification in systems and structures. Int J
Struct Health Monit 3:85–98
7. Randall R (2011) Vibration-based condition monitoring: industrial, aerospace and automotive
applications. Wiley-Blackwell, Chichester
18 Structural Health Monitoring and Damage Identification 1053

8. Farrar C, Doebling S, Nix D (2001) Vibration-based structural damage identification. Philos

Trans R Soc A Math Phys Eng Sci 359(1778):131–149
9. Fassois S, Sakellariou J (2007) Time-series methods for fault detection and identification in
vibrating structures. Philos Trans R Soc A Math Phys Eng Sci 365(1851):411–448
10. Friswell M (2007) Damage identification using inverse methods. Philos Trans R Soc A Math
Phys Eng Sci 365:393–410
11. Mal A, Banerjee S, Ricci F (2007) An automated damage identification technique based on
vibration and wave propagation data. Philos Trans R Soc A Math Phys Eng Sci 365:479–491
12. Staszewski W, Robertson A (2007) Time-frequency and time-scale analyses for structural
health monitoring. Philos Trans R Soc A Math Phys Eng Sci 365:449–477
13. Papatheou E, Manson G, Barthorpe R, Worden K (2010) The use of pseudo-faults for novelty
detection in SHM. J Sound Vib 329:2349–2366
14. Papatheou E, Manson G, Barthorpe RJ, Worden K (2014) The use of pseudo-faults for damage
location in SHM: an experimental investigation on a Piper Tomahawk aircraft wing. J Sound
Vib 333:971–990
15. Hayton P, Utete S, King D, King S, Anuzis P, Tarassenko L (2007) Static and dynamic novelty
detection methods for jet engine health monitoring. Philos Trans R Soc A Math Phys Eng Sci
365(1851):493–514
16. Sohn H (2007) Effects of environmental and operational variability on structural health
monitoring. Philos Trans R Soc A Math Phys Eng Sci 365:539
17. Worden K, Manson G (2006) The application of machine learning to structural health
monitoring. Philos Trans R Soc A Math Phys Eng Sci 365(1851):515–537
18. Fuentes R (2017) On Bayesian networks for structural health and condition monitoring. PhD
thesis, University of Sheffield
19. Worden K, Manson G (2007) The application of machine learning to structural health
monitoring. Philos Trans R Soc A Math Phys Eng Sci 365(1851):515–537
20. Farrar C, Worden K, Park G (2010) Complexity: a new axiom for structural health monitor-
ing? In Proceedings of 5th European workshop on structural health monitoring, Sorrento, pp
882–888
21. Boiler C, Buderath M (feb 2007) Fatigue in aerostructures - where structural health
monitoring can contribute to a complex subject. Philos Trans R Soc A Math Phys Eng Sci
365:561–587
22. Sohn H, Farrar CR, Hemez F, Czarnecki J (1996) A review of structural health monitoring
literature 1996–2001, Technical report. Los Alamos National Laboratory, Los Alamos
23. Oppenheim AV, Schaffer RW (1975) Digital signal processing. Prentice-Hall, Englewood
Cliffe
24. Figueiredo E, Park G, Figueiras J, Farrar CR, Worden K (2009) Structural health monitoring
algorithm comparisons using standard data sets: LANL report LA-14393, Technical report.
Los Alamos National Laboratory, Los Alamos
25. Sohn H, Farrar C, Hunter N, Worden K (2001) Structural health monitoring using statistical
pattern recognition techniques. ASME J Dyn Meas Syst Control 123:706–711
26. Yang Z, Yu Z, Sun H (oct 2007) On the cross correlation function amplitude vector and its
application to structural damage detection. Mech Syst Signal Process 21:2918–2932
27. Wang L, Yang Z, Waters T (nov 2010) Structural damage detection using cross correlation
functions of vibration response. J Sound Vib 329:5070–5086
28. Ng CT, Veidt M (2009) A Lamb-wave-based technique for damage detection in composite
laminates. Smart Mater Struct 18:074006
29. Michaels JE (2008) Detection, localization and characterization of damage in plates with an
in situ array of spatially distributed ultrasonic sensors. Smart Mater Struct 17:035035
30. Michaels J, Michaels T (2005) Detection of structural damage from the local temporal
coherence of diffuse ultrasonic signals. IEEE Trans Ultrason Ferroelectr Freq Control
52:1769–1782
31. Dongarra J, Sullivan F (2000) Guest editors introduction to the top 10 algorithms. Comput
Sci Eng 2:22–23
1054 R. Fuentes et al.

32. Salawu O (1997) Detection of structural damage through changes in frequency: a review. Eng
Struct 19:718–723
33. Sampaio R, Maia N, Silva J (1999) Damage detection using the frequency-response-function
curvature method. J Sound Vib 226:1029–1042
34. Kessler SS, Spearing S, Atalla MJ, Cesnik CE, Soutis C (2002) Damage detection in
composite materials using frequency response methods. Compos Part B 33:87–95
35. Owolabi G, Swamidas A, Seshadri R (2003) Crack detection in beams using changes in
frequencies and amplitudes of frequency response functions. J Sound Vib 265:1–22
36. Maia N, Silva J, ALMAS E, Sampaio R (2003) Damage detection in structures: from mode
shape to frequency response function methods. Mech Syst Signal Process 17:489–498
37. Peng Z, Chu F (2004) Application of the wavelet transform in machine condition monitoring
and fault diagnostics: a review with bibliography. Mech Syst Signal Process 18:199–221
38. Taha MMR (2006) Wavelet transform for structural health monitoring: a compendium of uses
and features. Struct Health Monit 5(3):267–295
39. Nair KK, Kiremidjian AS, Law KH (2006) Time series-based damage detection and
localization algorithm with application to the ASCE benchmark structure. J Sound Vib 291:
349–368
40. Krishnan Nair K, Kiremidjian AS (2007) Time series based structural damage detection
algorithm using Gaussian mixtures modeling. J Dyn Syst Meas Control 129(3):285
41. Yao R, Pakzad SN (2012) Autoregressive statistical pattern recognition algorithms for damage
detection in civil structures. Mech Syst Signal Process 31:355–368
42. Lu Y, Gao F (2005) A novel time-domain auto-regressive model for structural damage
diagnosis. J Sound Vib 283(3–5):1031–1049
43. Shumway RH, Stoffer DS (2011) Time series analysis and its applications. Springer,
New York
44. Chen S, Billings S a, Cowan CFN, Grant PM (1990) Practical identification of NARMAX
models using radial basis functions. Int J Control 52(769892610):1327–1350
45. Billings S (2013) Nonlinear system identification NARMAX methods in the time, frequency,
and spatio-temporal domains. Wiley, Chichester
46. Bishop C (2006) Pattern recognition and machine learning. Springer, New York
47. Rasmussen C, Williams C (2006) Gaussian processes for machine learning, vol 38. The MIT
Press, Cambridge, MA
48. Worden K, Manson G, Cross EJ (2014) On Gaussian process NARX models and their higher-
order frequency response functions. In: Solving computationally expensive engineering
problems. Springer, Cham, pp 315–335
49. Mottershead J, Friswell M (1993) Model updating in structural dynamics: a survey. J Sound
Vib 167(2):347–375
50. Cawley P, Adams RD (1979) The location of defects in structures from measurements of
natural frequencies. J Strain Anal Eng Des 14:49–57
51. Pandey A, Biswas M, Samman M (1991) Damage detection from changes in curvature mode
shapes. J Sound Vib 145(2):321–332
52. Hearn G, Testa RB (1991) Modal analysis for damage detection in structures. J Struct Eng
117(10):3042–3063
53. Jaishi B, Ren W-X (2006) Damage detection by finite element model updating using modal
flexibility residual. J Sound Vib 290(1–2):369–387
54. Peeters B, De Roeck G (1999) Reference-based stochastic subspace identification for output-
only modal analysis. Mech Syst Signal Process 13(6):855–878
55. Kaiser J (1950) Untersuchungen über das auftreten von geräuschen heim Zugversuch. PhD
thesis, Technical University of Munich (TUM)
56. Fuentes R, Howard TP, Marshall MB, Cross EJ, Dwyer-Joyce RS (2016) Observations on
acoustic emissions from a line contact compressed into the plastic region. Proc Inst Mech
Eng J Eng Tribol 230(11):1371–1376
57. Baram J, Rosen M (1979) Fatigue life prediction by distribution analysis of acoustic emission
signals. Mater Sci Eng 41(1):25–30
18 Structural Health Monitoring and Damage Identification 1055

58. Baxter MG, Pullin R, Holford KM, Evans SL (2007) Delta T source location for acoustic
emission. Mech Syst Signal Process 21(3):1512–1520
59. Hensman J, Mills R, Pierce SG, Worden K, Eaton M (2010) Locating acoustic emis-
sion sources in complex structures using Gaussian processes. Mech Syst Signal Process
24(1):211–223
60. Kurz JH, Grosse CU, Reinhardt HW (2005) Strategies for reliable automatic onset time pick-
ing of acoustic emissions and of ultrasound signals in concrete. Ultrasonics 43(7):538–546
61. Rayleigh L (1884) On waves propagating along the plane surface of an elastic solid. Proc
Lond Math Soc 17:4–11
62. Lamb H (1917) On waves in an elastic plate. Proc R Soc Series A 93:93–114
63. Worden K (2001) Rayleigh and Lamb waves – basic principles. Strain 37:167–172
64. Croxford A, Wilcox P, Drinkwater B, Konstantinidis G (2007) Strategies for guided-wave
structural health monitoring. Proc R Soc A Math Phys Eng Sci 463(2087):2961–2981
65. Cross E, Worden K (2011) Approaches to nonlinear cointegration with a view towards
applications in SHM. In: Proceedings of 8th international conference on damage assessment –
DAMAS 2011, Oxford, UK
66. Papatheou E, Dervilis N, Maguire A, Antoniadou I, Worden K (2015) A performance
monitoring approach for the novel Lillgrund offshore wind farm. IEEE Trans Ind Electron
46(c):1–1
67. Zhang Y, Li J (2006) Wavelet-based vibration sensor data compression technique for civil
infrastructure condition monitoring. J Comput Civ Eng 20:390–399
68. Staszewski W, Tomlinson G, Boiler C (2004) Health monitoring of aerospace structures,
Smart Sensor Technologies and Signal Processing. Wiley, Hoboken
69. Candès E (2006) Compressive sampling. In: Proceedings of the international congress of
mathematicians Madrid, August 22–30, 2006. European Mathematical Society Publishing
House, Zurich, pp 1433–1452
70. Yuequan Bao Y, Beck JL, Li HH (2011) Compressive sampling for accelerometer signals in
structural health monitoring. Struct Health Monit Int J 10:235–246
71. Mascareñas D, Cattaneo A, Theiler J, Farrar C (2013) Compressed sensing techniques for
detecting damage in structures. Struct Health Monit Int J 12:325–338
72. Johnson WB, Lindenstrauss J (1984) Extensions of Lipschitz mappings into a Hilbert space.
Contemp Math 26:189–206
73. Tibshirani R (1996) Regression selection and shrinkage via the lasso. J R Stat Soc B
58(1):267–288
74. Fuentes R, Mineo C, Pierce SG, Worden K, Cross EJ (2019) A probabilistic compressive
sensing framework with applications to ultrasound signal processing. Mech Syst Signal
Process 117:383–402
75. Rao SS (2010) Mechanical vibrations, 5th edn. Prentice Hall, Upper Saddle River
76. Särkkä S (2013) Bayesian filtering and smoothing. Cambridge University Press,
Cambridge
77. Caravani P, Watson ML, Thomson WT (1977) Recursive least-squares time domain identifi-
cation of structural parameters. J Appl Mech 44:135
78. Paleologu C, Benesty J, Ciochia S (2008) A robust variable forgetting factor recursive least-
squares algorithm for system identification. IEEE Signal Proc Lett 15(3):597–600
79. Cooper J, Worden K (2000) On-line physical parameter estimation with adaptive forgetting
factors. Mech Syst Signal Proc 14:705–730
80. Jazwinski AH (1970) Stochastic processes and filtering theory, vol 64. Academic, New York
81. Grewal MS, Andrews AP (2001) Kalman filtering: theory and practice. Wiley, Hoboken
82. Yang J, Lin S, Huang H, Zhou L (2006) An adaptive extended Kalman filter for structural
damage identification. Struct Control Health Monit 13:849–867
83. Pan S, Yang J, Huang H (2007) An adaptive extended Kalman filter for structural damage
identification II: unknown inputs. Struct Control Health Monit 14:497–521
84. Zhou L, Wu S, Yang JN (2008) Experimental study of an adaptive extended Kalman filter for
structural damage identification. J Infrastruct Syst 14(1):42–51
1056 R. Fuentes et al.

85. Corigliano A, Mariani S (2004) Parameter identification in explicit structural dynamics:

performance of the extended Kalman filter. Comput Methods Appl Mech Eng 193(36–38):
3807–3835
86. Mariani S, Corigliano A (2005) Impact induced composite delamination: state and parameter
identification via joint and dual extended Kalman filters. Comput Methods Appl Mech Eng
194(50–52):5242–5272
87. Julier SJ, Uhlmann JK, Durrant-Whyte HF (1995) A new approach for filtering nonlinear
systems. In: American control conference, proceedings of the 1995, vol 3. IEEE, pp 1628–
1632
88. Julier SJ, Uhlmann JK (1997) New extension of the Kalman filter to nonlinear systems. In:
Signal processing, sensor fusion, and target recognition VI, vol 3068. International Society
for Optics and Photonics, pp 182–194
89. Wan EA, Van Der Merwe R (2000) The unscented Kalman filter for nonlinear estimation.
In: Adaptive systems for signal processing, communications, and control symposium 2000.
AS-SPCC. The IEEE 2000. IEEE, pp 153–158
90. LaViola JJ (2003) A comparison of unscented and extended Kalman filtering for estimating
quaternion motion. In: American control conference, 2003. Proceedings of the 2003, vol 3.
IEEE, pp 2435–2440
91. St-Pierre M, Gingras D (2004) Comparison between the unscented Kalman filter and the
extended Kalman filter for the position estimation module of an integrated navigation
information system. In: IEEE intelligent vehicles symposium. CiteSeerX, pp 831–835
92. Gustafsson F, Hendeby G (2012) Some relations between extended and unscented Kalman
filters. IEEE Trans Signal Process 60(2):545–555
93. Wu M, Smyth AW (2007) Application of the unscented Kalman filter for real-time nonlinear
structural system identification. Struct Control Health Monitor 14(7):971–990
94. Chatzi EN, Smyth AW (2009) The unscented Kalman filter and particle filter methods for
nonlinear structural system identification with non-collocated heterogeneous sensing. Struct
Control Health Monitor 16(1):99–123
95. Eftekhar Azam S, Ghisi A, Mariani S (2012) Parallelized sigma-point Kalman filtering for
structural dynamics. Comput Struct 92-93:193–205
96. Mariani S, Ghisi A (2007) Unscented Kalman filtering for nonlinear structural dynamics.
Nonlinear Dyn 49(1–2):131–150
97. Xie Z, Feng J (2012) Real-time nonlinear structural system identification via iterated
unscented Kalman filter. Mech Syst Signal Process 28:309–322
98. Doucet A, Johansen AM (2009) A tutorial on particle filtering and smoothing: fifteen years
later. Handb Nonlin Filter 12(656–704):3
99. Crisan D, Doucet A (2002) A survey of convergence results on particle filtering methods for
practitioners. IEEE Trans Signal Process 50(3):736–746
100. Hu X-L, Schön T, Ljung L (2008) A basic convergence result for particle filtering. IEEE Trans
Signal Process 56(4):1337–1348
101. Gordon NJ, Salmond DJ, Smith AF (1993) Novel approach to nonlinear/non-Gaussian
Bayesian state estimation. IEE Proc F (Radar Signal Proc) 140(2):107–113
102. Pitt MK, Shephard N (1999) Filtering via simulation: auxiliary particle filters. J Am Stat
Assoc 94(446):590–599
103. Douc R, Cappé O (2005) Comparison of resampling schemes for particle filtering. In: Image
and signal processing and analysis, 2005. ISPA 2005. Proceedings of the 4th international
symposium on. IEEE, pp 64–69
104. Hoi JD, Schon TB, Gustafsson F (2006) On resampling algorithms for particle filters. In:
Nonlinear statistical signal processing workshop, 2006 IEEE. IEEE, pp 79–82
105. Chatzi E, Smyth A (2013) Particle filter scheme with mutation for the estimation of time-
invariant parameters in structural health monitoring applications. Struct Control Health Monit
20:1081–1095
106. Corbetta M, Sbarufatti C, Manes A, Giglio M (2013) Stochastic definition of state-space
equation for particle filtering algorithms. Chem Eng Trans:1–6
18 Structural Health Monitoring and Damage Identification 1057

107. Corbetta M, Sbarufatti C, Giglio M, Todd MD (2018) Optimization of nonlinear, non-

Gaussian Bayesian filtering for diagnosis and prognosis of monotonic degradation processes.
Mech Syst Signal Process 104:305–322
108. Sbarufatti C, Corbetta M, Giglio M, Cadini F (2017) Adaptive prognosis of lithium-ion
batteries based on the combination of particle filters and radial basis function neural networks.
J Power Sources 344:128–140
109. Xue S, Tang H, Xie Q (2009) Structural damage detection using auxiliary particle filtering
method. Struct Health Monit 8(2):101–112
110. Ching J, Beck JL, Porter KA (2006) Bayesian state and parameter estimation of uncertain
dynamical systems. Probabilist Eng Mech 21(1):81–96
111. Andrieu C, Doucet A, Holenstein R (2010) Particle Markov chain Monte Carlo methods. J R
Stat Soc B Stat Methodol 72(3):269–342
112. Rippengill S, Worden K, Holford KM, Pullin R (2003) Automatic classification of acoustic
emission patterns. Strain 39:31–41
113. Manson G, Worden K, Holford K, Pullin R (2001) Visualisation and dimension reduction of
acoustic emission data for damage detection. J Intell Mater Syst Struct 12:529–536
114. Farrar CR, Worden K (2012) Structural health monitoring: a machine learning perspective.
Wiley, Hoboken
115. Bishop CM (2006) Pattern recognition and machine learning. Springer, Berlin
116. Worden K, Manson G, Fieller NR (2000) Damage detection using outlier analysis. J Sound
Vib 229(3):647–667
117. Dervilis N, Cross E, Barthorpe R, Worden K (2014) Robust methods of inclusive outlier
analysis for structural health monitoring. J Sound Vib 333(20):5181–5195
118. Hawkins DM (1980) Identification of outliers. Springer, Dordrecht
119. Rousseeuw PJ, Leroy AM (2005) Robust regression and outlier detection. Wiley, New York
120. Hubert M, Debruyne M (2010) Minimum covariance determinant. Wiley Interdiscip Rev
Comput Stat 2(1):36–43
121. Van Aelst S, Rousseeuw P (2009) Minimum volume ellipsoid. Wiley Interdiscip Rev Comput
Stat 1(1):71–82
122. Papatheou E, Dervilis N, Maguire AE, Campos C, Antoniadou I, Worden K (2017)
Performance monitoring of a wind turbine using extreme function theory. Renew Energy
113:1490–1502
123. Moya MM, Koch MW, Hostetler LD (1993) One-class classifier networks for target recogni-
tion applications. Proc World Congr Neural Netw:359–367
124. Zimek A, Gaudet M, Campello RJ, Sander J (2013) Subsampling for efficient and effective
unsupervised outlier detection ensembles. In: Proceedings of the 19th ACM SIGKDD
international conference on knowledge discovery and data mining. ACM, pp 428–436
125. Rousseeuw PJ, Driessen KV (1999) A fast algorithm for the minimum covariance determinant
estimator. Technometrics 41(3):212–223
126. Murphy KP (2012) Machine learning: a probabilistic perspective. MIT Press, Cambridge, MA
127. Gelman A, Carlin J, Rubin D, Vehtari A, Dunson D, Stern H (2013) Bayesian data analysis,
3rd edn. CRC Press, New York
128. Beyer K, Goldstein J, Ramakrishnan R, Shaft U (1999) When is “nearest
neighbor” meaningful? In: Database theory – ICDT’99. Springer, Berlin/Heidelberg,
pp 217–235
129. Manson G, Worden K, Allman D (2003) Experimental validation of a structural health
monitoring methodology: part III. Damage location on an aircraft wing. J Sound Vib
259(2):365–385
130. Worden K, Manson G, Hilson G, Pierce S (2008) Genetic optimisation of a neural damage
locator. J Sound Vib 309(3):529–544
131. Dervilis N (2013) A machine learning approach to structural health monitoring with a view
towards wind turbines. PhD thesis, University of Sheffield
132. Bull L, Worden K, Manson G, Dervilis N (2018) Active learning for semi-supervised
structural health monitoring. J Sound Vib 437:373–388
1058 R. Fuentes et al.

133. Rogers TJ, Worden K, Fuentes R, Dervilis N, Tygesen UT, Cross EJ (2019) A Bayesian non-
parametric clustering approach for semi-supervised structural health monitoring. Mech Syst
Signal Process 119:100–119
134. Farrar C, Sohn H, Worden K (2001) Data normalization: a key for structural health
monitoring, technical report. Los Alamos National Laboratory, Los Alamos
135. Alampalli S (2000) Effects of testing, analysis, damage, and environment on modal parame-
ters. Mech Syst Signal Process 14(1):63–74
136. Cornwell P, Farrar CR, Doebling SW, Sohn H (1999) Environmental variability of modal
properties. Exp Tech 23(6):45–48
137. Peeters B, Roeck GD (2001) One-year monitoring of the Z24-bridge: environmental effects
versus damage events. Earthq Eng Struct Dyn 30:149–171
138. Sohn H, Dzwonczyk M, Straser EG, Kiremidjian AS, Law KH, Meng T (1999) An
experimental study of temperature effect on modal parameters of the Alamosa canyon bridge.
Earthq Eng Struct Dyn 28(8):879–897
139. Moser P, Moaveni B (2011) Environmental effects on the identified natural frequencies of the
Dowling hall footbridge. Mech Syst Signal Process 25(7):2336–2357
140. Xu Y, Chen B, Ng C, Wong K, Chan W (2010) Monitoring temperature effect on a long
suspension bridge. Struct Control Health Monit 17(6):632–653
141. Cao Y, Yim J, Zhao Y, Wang ML (2011) Temperature effects on cable stayed bridge using
health monitoring system: a case study. Struct Health Monit 10(5):523–537
142. de Battista N, Brownjohn JM, Tan HP, Koo K-Y (2015) Measuring and modelling the
thermal performance of the Tamar suspension bridge using a wireless sensor network. Struct
Infrastruct Eng 11(2):176–193
143. Mahmoud M, Abe M, Fujino Y (2001) Analysis of suspension bridge by ambient vibration
measurement using time domain method and its application to health monitoring. In: Proceed-
ings of the international modal analysis conference-IMAC, vol 1. Society for Experimental
Mechanics, pp 504–510
144. Cross E, Koo K-Y, Brownjohn J, Worden K (2013) Long-term monitoring and data analysis
of the Tamar bridge. Mech Syst Signal Process 35:16–34
145. Xia Y, Hao H, Zanardo G, Deeks A (2006) Long term vibration monitoring of an rc slab:
temperature and humidity effect. Eng Struct 28(3):441–452
146. Kim C-Y, Jung D-S, Kim N-S, Kwon S-D, Feng MQ (2003) Effect of vehicle weight on
natural frequencies of bridges measured from traffic-induced vibration. Earthq Eng Eng Vib
2(1):109–115
147. Lanata F, Schoefs F (2011) Multi-algorithm approach for identification of structural behav-
ior of complex structures under cyclic environmental loading. Struct Health Monit 11:
51–67
148. Sohn H, Dzwonczyk M, Straser E, Law K, Meng T, Kiremidjian A (1998) Adaptive modeling
of environmental effects in modal parameters for damage detection in civil structures. In:
Smart systems for bridges, structures, and highways, pp 127–138
149. Moaveni B, He X, Conte J, Fraser M, Elgamal A (2009) Uncertainty analysis of Voigt bridge
modal parameters due to changing environmental condition. In: Proceedings of international
conference on modal analysis (IMAC-XXVII)
150. Ni Y, Hua X, Fan K, Ko J (2005) Correlating modal properties with temperature using long-
term monitoring data and support vector machine technique. Eng Struct 27(12):1762–1773
151. Hua X, Ni Y, Ko J, Wong K (2007) Modeling of temperature–frequency correlation using
combined principal component analysis and support vector regression technique. J Comput
Civ Eng 21:122
152. Ni Y, Zhou H, Ko J (2009) Generalization capability of neural network models for
temperature-frequency correlation using monitoring data. J Struct Eng 135(10):1290–1300
153. Worden K, Sohn H, Farrar C (2002) Novelty detection in a changing environment: regression
and interpolation approaches. J Sound Vib 258(4):741–761
154. Worden K, Cross E (2018) On switching response surface models, with applications to the
structural health monitoring of bridges. Mech Syst Signal Process 98:139–156
18 Structural Health Monitoring and Damage Identification 1059

155. Dervilis N, Shi H, Worden K, Cross E (2016) Exploring environmental and operational
variations in SHM data using heteroscedastic Gaussian processes. In: Dynamics of civil
structures, vol 2. Springer, New York, pp 145–153
156. Surace C, Worden K (1997) Some aspects of novelty detection methods. In: Proceedings of
the third international conference on modern practice in stress and vibration analysis, Dublin
157. Cross E, Manson G, Worden K, Pierce S (2012) Features for damage detection with
insensitivity to environmental and operational variations. Proc R Soci A rspa20120031
158. Kullaa J (2011) Distinguishing between sensor fault, structural damage, and environmental
or operational effects in structural health monitoring. Mech Syst Signal Process 25(8):1–8
159. Manson G (2002) Identifying damage sensitive, environment insensitive features for damage
detection. In: Proceedings of the IES conference
160. Yan A, Kerschen G, De Boe P, Golinval J (2005) Structural damage diagnosis under varying
environmental conditions part I: a linear analysis. Mech Syst Signal Process 19(4):847–864
161. Kullaa J (2004) Structural health monitoring under variable environmental or operational
conditions. In: Proceedings of the 2nd European workshop on structural health monitoring
162. Shi H (2018) On nonlinear cointegration methods for structural health monitoring. PhD thesis,
University of Sheffield
163. Yan A-M, Kerschen G, Boe PD, Golinval J-C (2005) Structural damage diagnosis under
varying environmental conditions part II: local PCA for non-linear cases. Mech Syst Signal
Process 19:865–880
164. Sohn H, Worden K, Farrar C (2002) Statistical damage classification under changing
environmental and operational conditions. J Intell Mater Syst Struct 13(9):561
165. Dervilis N, Choi M, Taylor S, Barthorpe R, Park G, Farrar C, Worden K (2014) On damage
diagnosis for a wind turbine blade using pattern recognition. J Sound Vib 333(6):1833–1850
166. Shi H, Worden K, Cross E (2016) A nonlinear cointegration approach with applications
to structural health monitoring. In: Journal of Physics: Conference Series, vol 744, IOP
Publishing, p 012025
167. Shi H, Worden K, Cross E (2018) A regime-switching cointegration approach for removing
environmental and operational variations in structural health monitoring. Mech Syst Signal
Process 103:381–397
168. Rogers T, Worden K, Fuentes R, Dervilis N, Tygesen U, Cross E (2019) A Bayesian non-
parametric clustering approach for semi-supervised structural health monitoring. Mech Syst
Signal Process 119:100–119
169. Cross EJ, Worden K, Chen Q (2011) Cointegration: a novel approach for the removal of
environmental trends in structural health monitoring data. Proc R Soc Lond A Math Phys Eng
Sci 467:2712–2732
170. Koo K-Y, Brownjohn J, List D, Cole R (2013) Structural health monitoring of the Tamar
suspension bridge. Struct Control Health Monit 20(4):609–625
171. Cross E (2012) On structural health monitoring in changing environmental and operational
conditions. PhD thesis, University of Sheffield
172. Dickey D, Fuller W (1979) Distribution of the estimators for autoregressive time series with
a unit root. J Am Stat Assoc 74:427–431
173. Dickey D, Fuller W (1981) Likelihood ratio statistics for autoregressive time series with a
unit root. Econometrica 49(4):1057–1072
174. Engle R, Granger C Cointegration and error correction: representation, estimation and testing.
Econometrica 55(2):251–276
175. Johansen S (1988) Statistical analysis of cointegration vectors. J Econ Dyn Control 12(2–3):
231–254
176. Pierce SG, Dong F, Atherton K, Culshaw B, Worden K, Manson G, Monnier T, Guy P, Baboux
J-C, Assad J et al (2001) Damage assessment in smart composite structures: the Damascos
program. In: Smart structures and materials 2001: smart structures and integrated systems,
vol 4327. International Society for Optics and Photonics, Bellingham, pp 223–234
177. Fan W, Qiao P (2011) Vibration-based damage identification methods: a review and compar-
ative study. Struct Health Monit 10(1):83–111
1060 R. Fuentes et al.

178. Farrar C, Worden K (2013) Structural health monitoring: a machine learning perspective.
Wiley, Chichester
179. Vagnoli M, Remenyte-Prescott R, Andrews J (2018) Railway bridge structural health
monitoring and fault detection: state-of-the-art methods and future challenges. Struct Health
Monit 17(4):971–1007
180. Friswell MI, Mottershead JE (2001) Inverse methods in structural health monitoring. Key Eng
Mater 204–205:201–210
181. Barthorpe RJ (2011) On model- and data-based approaches to structural health monitoring.
PhD thesis, University of Sheffield
182. Gardner P, Lord C, Barthorpe RJ (2018) A probabilistic framework for forward model-driven
SHM. In: European Workshop on Structural Health Monitoring (EWHSM), pp 1–13
183. Gorl E, Link M (2003) Damage identification using changes of Eigen frequencies and mode
shapes. Mech Syst Signal Process 17(1):103–110
184. Farrar CR, Lieven NA (2007) Damage prognosis: The future of structural health monitoring.
Philos Trans R Soc A Math Phys Eng Sci 365:623–632
185. Friswell M, Mottershead J (1995) Finite element model updating in structural dynamics.
Springer, New York
186. Friswell MI, Mottershead JE (2001) Physical understanding of structures by model updating.
In: Proceedings of international conference on structural system identification, pp 81–96
187. Mottershead JE, Link M, Friswell MI (2011) The sensitivity method in finite element model
updating: a tutorial. Mech Syst Signal Process 25(7):2275–2296
188. Jaishi B, Ren W-X (2005) Structural finite element model updating using ambient vibration
test results. J Struct Eng 131(4):617–628
189. Simoen E, De Roeck G, Lombaert G (2015) Dealing with uncertainty in model updating for
damage assessment: a review. Mech Syst Signal Process 56:123–149
190. Haag T, Carvajal González S, Hanss M (2012) Model validation and selection based on
inverse fuzzy arithmetic. Mech Syst Signal Process 32:116–134
191. Rasmussen C, Ghahramani Z (2001) Occam’s razor. In: Advances in neural information
processing systems 13: proceedings of the 2000 conference, p 294
192. Beck JL, Katafygiotis LS (1998) Updating models and their uncertainties. I: Bayesian
statistical framework. J Eng Mech 124(4):455–461
193. Katafygiotis LS, Beck JL (1998) Updating models and their uncertainties. II: model identifi-
ability. J Eng Mech 124(4):463–467
194. Friswell MI, Penny JET, Garvey SD (1997) Parameter subset selection in damage location.
Inverse Probl Eng 5(3):189–215
195. Friswell MI, Mottershead JE, Ahmadian H (1998) Combining subset selection and parameter
constraints in model updating. J Vib Acoust Trans ASME 120:854–859
196. Ko JM, Sun ZG, Ni YQ (2002) Multi-stage identification scheme for detecting damage in
cable-stayed Kap Shui Mun bridge. Eng Struct 24(7):857–868
197. Lee JJ, Lee JW, Yi JH, Yun CB, Jung HY (2005) Neural networks-based damage detection
for bridges considering errors in baseline finite element models. J Sound Vib 280(3–5):
555–578
198. Satpal SB, Guha A, Banerjee S (2016) Damage identification in aluminum beams using
support vector machine: numerical and experimental studies. Struct Control Health Monit
23:446–457
199. Hariri-Ardebili MA, Pourkamali-Anaraki F (2018) Support vector machine based reliability
analysis of concrete dams. Soil Dyn Earthq Eng 104:276–295
200. Nagel JB, Sudret B (2016) A unified framework for multilevel uncertainty quantification in
Bayesian inverse problems. Probabilistic Engineering Mechanics 43:68–84
201. Li C, Mahadevan S (2016) Role of calibration, validation, and relevance in multilevel
uncertainty integration. Reliab Eng Syst Saf 148:32–43
202. Gardner P, Lord C, Barthorpe RJ (2018) A multi-level uncertainty integration strategy for
forward model-driven SHM. In: Proceedings of the international conference on noise and
vibration engineering, pp 3681–3692
18 Structural Health Monitoring and Damage Identification 1061

203. Kennedy MC, O’Hagan A (2001) Bayesian calibration of computer models. J R Stat Soc B
Stat Methodol 63(3):425–464
204. Gardner P, Lord C, Barthorpe RJ (2017) Bayesian calibration and bias correction for forward
model-driven SHM. In: Proceedings of the 11th international workshop on structural health
monitoring, pp 2019–2027
205. Gardner P, Lord C, Barthorpe RJ (2018) Bayesian history matching for forward model-driven
structural health monitoring. In: Proceedings of IMAC XXXVI
206. Fault shuts large hadron collider for two months. The Guardian 2017
Experimental Dynamic Substructures
19
Randall L. Mayes and Matthew S. Allen

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
2 Experimental Substructure Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065
2.1 Connecting Substructures with Compatibility and/or Equilibrium Equations . . . . 1066
2.2 Connecting Substructures in the Physical Domain . . . . . . . . . . . . . . . . . . . . . . . . . . 1067
2.3 Connecting Substructures in the Modal Domain: Component
Mode Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1070
2.4 Connecting Substructures in the Frequency Domain:
Frequency-Based Substructuring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1075
3 Dealing with Experimental Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077
3.1 Common Experimental Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077
3.2 The Transmission Simulator Approach to Mitigate Traditional
Experimental Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1081
3.3 An Example Using the Transmission Simulator Approach . . . . . . . . . . . . . . . . . . . 1082
3.4 Transmission Simulator Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1084
3.5 Practical Guidance Using the Transmission Simulator Approach . . . . . . . . . . . . . . 1089
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091

Notice: This manuscript has been authored by Sandia Corporation under Contract No. DE-AC04-
94AL85000 with the US Department of Energy. The US government retains and the publisher,
by accepting the article for publication, acknowledges that the US government retains a non-
exclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of
this manuscript or allow others to do so, for US government purposes.

R. Mayes ()
Sandia National Laboratories, Albuquerque, NM, USA
e-mail: [email protected]
M. Allen
Department of Engineering Physics, University of Wisconsin, Madison, WI, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 1063

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_24
1064 R. Mayes and M. Allen

Abstract

This chapter deals with experimental dynamic substructures which are reduced
order models that can be coupled with each other or with finite element derived
substructures to estimate the system response of the coupled substructures. A
unifying theoretical framework in the physical, modal or frequency domain is
reviewed with examples. The major issues that have hindered experimental based
substructures are addressed. An example is demonstrated with the transmission
simulator method that overcomes the major historical difficulties. Guidelines for
the transmission simulator design are presented.

Keywords

Experimental dynamic substructuring · Component mode synthesis ·

Frequency based substructures · Transmission simulator · Experimental
models · Reduced order models

Nomenclature

FRF Frequency response function

RMS Root mean square
DOF Degree of freedom
g Connection force at a single connection DOF
u Physical displacement of a single DOF
ω Frequency in radians per second
B Matrix associated with compatibility of connected DOF
L Matrix associated with equilibrium forces of connected DOF
Y Frequency response function matrix
 Mode shape matrix
u Vector of displacements
g Vector of connection forces
q Vector of non-redundant connection DOF (physical or generalized coordinates)
ξ Vector of generalized modal DOF
† Superscript indicating the Moore-Penrose pseudo-inverse of a matrix

1 Introduction

An experimental dynamic substructure is an experimentally derived model of some

convenient portion of a structure. The substructure can only capture the response of
the subsystem in a finite frequency band. The model input and response is generally
defined only for locations where sensors obtain information on the subsystem.
All the substructures of a system can be combined to simulate the full system
response over a finite frequency band. Figure 1 shows an example using the SEM
substructuring testbed, in which a substructure from a finite element model is
combined with an experimental substructure to obtain the full system response.
19 Experimental Dynamic Substructures 1065

Fig. 1 Substructures from FE and experiment combined to obtain full system response of SEM
testbed

Analytical substructures have been used for decades. Usually they are reduced
order models of the subsystems derived from finite element models of the subsys-
tems. Because they are reduced order models, when they are combined to simulate
the response of the full system, they often reduce the computational burden by many
orders of magnitude compared to the full system finite element model.
In addition to their computational advantage, substructures can be developed and
shared by different organizations to obtain full system response. Important local
dynamics can often be observed in the substructure alone that might be masked in
the full system model response.
Experimental substructures share some of the advantages above and in some
cases may provide a cheaper or more accurate solution than can be obtained from
first principles models. If hardware is available, an experimental substructure may
be extracted from an appropriate experiment.
While the experimental dynamic substructuring concept has been around for
decades, early attempts revealed that accurate experimental substructures can be
difficult to obtain because of several challenges that will be addressed in this chapter.
Since about 2005, new approaches have been developed which are overcoming
many of these challenges. Most of the work has been with linear substructures.

2 Experimental Substructure Technology

Over the years, many different substructure technologies have developed. Some
approaches utilize physical degrees of freedom (DOF), generalized DOF, frequency
response or impedance functions, and even state space formulations. These different
1066 R. Mayes and M. Allen

representations used to be considered almost different technologies, but they are

actually related. Here we utilize a general framework [1] since this terminology
shows the relationship between different approaches. The general framework
matrix definitions will be clearly illustrated with simple models to show how two
substructures can be combined to represent the full system in the physical, modal,
and frequency domains.

2.1 Connecting Substructures with Compatibility and/or

Equilibrium Equations

Whether substructuring is analytical or experimental, it is required that compatibil-

ity and force equilibrium be enforced between the boundaries of the substructures
to simulate full system response. If we consider two substructures formulated with
their physical DOF, compatibility can be expressed for any two connected DOF as

u1 − u2 = 0 (1)

where u represents the scalar displacement of the connection DOF and the super-
script denotes the substructure. This forces these two DOF to have the same
displacement. In the general framework, all the compatibility equations can be
written as a matrix equation as

Bu = 0 (2)

where the B matrix typically contains either −1, 0, or 1 in all its elements.
Force equilibrium for any two connected DOF can be expressed as

g1 + g2 = 0 (3)

where the g values represent the amplitude of the equal and opposite connection
forces on each substructure. In the general framework, another matrix equation can
describe all the equilibrium equations for all connection forces as

LT g = 0 (4)

where L is another Boolean matrix with either 0 or 1 in all its elements and
superscript T denotes the matrix transpose.
A special property relates B and L which helps establish relationship between
substructuring approaches that emphasize either compatibility or equilibrium. This
can be derived by using the so-called primal form of substructuring to eliminate
redundant DOF. Mathematically this is stated
19 Experimental Dynamic Substructures 1067

u = Lq (5)

where q is a subset of u that eliminates redundant attached DOF. Combining Eqs.

(2) and (5) gives

BLq = 0. (6)

Since the vector q can have an infinite number of solutions depending on the
excitation forces applied to the system, it is required that

BL = 0 (7)

enforcing the rows of B to be orthogonal to the columns of L. This property is

sometimes stated that L is in the null space of B. Many linear algebra tools have a
programmed function that finds the null space of another matrix. (For example, in
MATLAB L = null(B) can be used to find L.) Hence, one knows L if B has been
established, or vice versa.
The principles of displacement compatibility and force equilibrium between the
boundary DOF of connected substructures are foundational to all substructuring
technologies. The B and L matrices are found in every substructuring approach in
the general framework, showing how the different approaches are related. In the
next sections, we will demonstrate these principles for substructures in the physical,
modal, and frequency domains to connect two simple substructures.

2.2 Connecting Substructures in the Physical Domain

Consider the two substructures in Fig. 2, which consist of two masses and a spring
with the values given in Table 1. The goal is to connect DOF u2 to u3 to get response
of the full system.
Compatibility to combine these two substructures would be written with the
boundary DOF as

u2 − u3 = 0 (8)

k1 k2

m1 WWW m 2
m3 WWW m4

u1 u2 u3 u4

Fig. 2 Two simple substructures – constrain displacement of m2 to displacement of m3 to get full

system response
1068 R. Mayes and M. Allen

Table 1 Mass and stiffness m1 = 1 k1 = 1.317e5

for two substructures m2 = 0.7 k2 = 5.1681e4
m3 = 0.1
m4 = 1

which can be put into the form of Eq. (2) as

⎧ ⎫
⎪
⎪ u1 ⎪
⎨ ⎪ ⎬
u2
Bu = [0 1 − 1 0] =0 (9)
⎪
⎪ u ⎪
⎩ 3⎪ ⎭
u4

which is sometimes called the constraint equation. Here there is only one constraint,
so B has only one row, but normally there are several constraints. Construct a
Boolean L matrix to eliminate u3 . This would conform to Eq. (5) as

u = Lq
⎧ ⎫ ⎡ ⎤
⎪ u1 ⎪
⎪ 1 0 0 ⎧ ⎫
⎨ ⎪ ⎬ ⎢ ⎨ u1 ⎬
u2 0 1 0⎥ (10)
=⎢ ⎥ u2
⎪ ⎣0
⎪ u3 ⎪
⎪ 1 0 ⎦⎩ ⎭
⎩ ⎭ u4
u4 0 0 1

so that the reduced q vector now contains only u1 , u2 , and u4 (since u3 = u4 ). By

Eq. (4), L should provide force equilibrium as

⎤⎧g ⎫ ⎧ ⎫
⎡
1000 ⎪⎪ 1⎪
⎨ ⎪
⎬ ⎨0⎬
⎢ ⎥ g2
L g=⎣0110⎦
T
= 0 (11)
⎪
⎪ g3 ⎪
⎪ ⎩ ⎭
0001 ⎩ ⎭ 0
g4

which specifies that the constraint forces on mass 1 and 4, g1 and g4 , = 0 and the
constraint forces on mass 2 and 3 must be equal and opposite, g2 + g3 = 0. One
can also check to see that L is in the null space of B according to Eq. (7) so that
the constraint, Eq. (6), is still satisfied. The equations of motion can be written in a
block diagonal form as
  1
     1
  
M1 0 ü K1 0 u1 f g1
2 + = 2 + 2 (12)
0 M2 ü 0 K2 u2 f g

with the superscripts representing the values associated with the two different
substructures, and the f vectors represent the external forces. Damping could
easily be included, but we neglect these terms to save space when all elements
19 Experimental Dynamic Substructures 1069

of the matrices are written in the examples below. This is in the dual form
which includes all DOF for both substructures. To obtain the primal form which
eliminates redundant (connected) DOF, substitute for the u vector from Eq. (10)
and premultiply by LT to get

   1   1  1
M1 0 K 0 T f T g
LT Lq̈ + LT
Lq = L 2 +L (13)
0 M2 0 K2 f g2

which can be expanded with our values for L, q, and the two substructures as

⎡ ⎤⎡ ⎤
⎡ ⎤ m 0 0 0 1 0 0 ⎧ ⎫
1000 ⎢ 1 ⎥⎢ ⎥⎪⎨ ü1 ⎪
⎬
⎢ ⎥ ⎢ 0 m2 0 0 ⎥⎢ 0 1 0⎥
⎣ 0 1 1 0 ⎦⎢ ⎥⎢ ⎥ ü2 +
⎣ 0 0 m3 0 ⎦⎣ 0 1 0 ⎦⎪⎩ ü ⎪⎭
0001 4
0 0 0 m4 0 0 1
⎡ ⎤⎡ ⎤ ⎧ ⎫
⎡ ⎤ k −k1 1 0 0 ⎧ ⎫ ⎡ ⎤⎪f ⎪ ⎡ ⎤
1000 ⎪ 1⎪
0 0
1000 ⎢ 1 ⎥⎢ ⎥⎪⎨ u1 ⎪
⎬ ⎨ ⎪
⎪ ⎬ 0
⎢ ⎥ ⎢ − k1 k1 0 0 ⎥⎢ 0 1 0 ⎥ ⎢ ⎥ f2 ⎢ ⎥
⎣ 0 1 1 0 ⎦⎢ ⎥⎢ ⎥ u2 = ⎣ 0 1 1 0 ⎦ + ⎣0⎦
⎣ 0 0 k2 −k2 ⎦ ⎣ 0 1 0 ⎦ ⎪⎩u ⎪ ⎭ ⎪ f3 ⎪
⎪ ⎪
0001 4 0001 ⎪ ⎩f ⎭ ⎪ 0
0 0 −k2 k2 0 0 1 4
(14)

which now represents the primal form of the full system with the substructures
attached. Note the far right-hand term goes to zero in accordance with Eq.
(11), because the internal connection forces are in equilibrium. Completing the
multiplication gives

⎡ ⎤⎧ ⎫ ⎡ ⎤⎧ ⎫ ⎧ ⎫
m1 0 0 ⎨ ü1 ⎬ k1 −k1 0 ⎨ u1 ⎬ ⎨ f1 ⎬
⎣ 0 (m2 + m3 ) 0 ⎦ ü2 + ⎣ − k1 (k1 + k2 ) −k2 ⎦ u2 = f2 + f3
⎩ ⎭ ⎩ ⎭ ⎩ ⎭
0 0 m4 ü4 0 −k2 k2 u4 f4
(15)

which provides the coupled system equations of motion in the physical domain,
without the redundant DOF u3 . The matrices that come from compatibility and
equilibrium, B and LT , are used in various ways in the physical, modal, and
frequency domains for various methods of substructuring. Because of the various
formulations and methods to minimize the experimental errors in experimental
substructuring, the form of B and L may be modified, but there is always some
enforcement of compatibility and equilibrium. Solving the eigenvalue problem
from Eq. (15) gives natural frequencies and mass-normalized mode shapes of the
assembled system that are listed in Table 2.
1070 R. Mayes and M. Allen

Table 2 Mode shapes for u1 −0.5976 −0.5699 −0.5639

coupled physical system u2 = u3 −0.5976 −0.2695 0.9057
u4 −0.5976 0.7855 −0.1606
Frequency (Hz) 0 41.93 93.23

2.3 Connecting Substructures in the Modal Domain: Component

Mode Synthesis

There are multiple linear substructure formulations in the modal domain. Each
formulation approximates the motion with a (usually) reduced number of DOF
pertaining to shape functions that can be superimposed to approximate the actual.
Typical types of shapes associated with the DOF might include vibration mode
shapes with free boundary conditions, mode shapes with constraints at the inter-
face, or static deflection shapes with constraints. Component Mode Synthesis
(CMS) is the general class of this type of substructuring. It is used in analytical
substructures, but is a very common approach for experimental substructures
since modal DOF and mode shapes are extracted from a modal test. Because
mode shapes are orthogonal with respect to the mass and stiffness matrices, the
modal mass and stiffness matrices have the nice property of being diagonal or
uncoupled.
The example from the previous section using two substructures will be utilized
again to illustrate the concepts, but now we shall assume that substructure 1 is
derived from a FE model and substructure 2 is derived from a modal test with free
boundary conditions. Both substructures will be described in terms of their free
modal DOF and free mode shapes. Analytical substructure 1 is derived from an
eigenvalue analysis, and the experimental substructure 2 is obtained from a modal
test.
One decision that is made early in the CMS process is what is the frequency
band for the analytical eigenvalue analysis for the first substructure and for the
modal test of the second substructure. Here we assume that 100 Hz is chosen as
the highest frequency for which modes will be calculated or extracted for both
substructures. CMS always provides an approximate solution, which assumes the
effects of modes above the cutoff frequency (100 Hz for this example) are not of
great importance. The error associated with neglecting higher-frequency modes is
called modal truncation error. Performing the eigenvalue analysis on substructure
1 and a modal test on softly suspended substructure 2 with accelerometers on
m3 and m4 , we obtain the frequency and mode shape information in Tables 3
T
and 4. The mode shapes are mass-normalized, i.e., r Mr = 1 for every
mode r derived from the analysis. Note that the modal test did not extract the
elastic mode, because it was above the chosen bandwidth of 100 Hz. We will
consider the effect of using only the rigid body mode for substructure 2 in this
example.
19 Experimental Dynamic Substructures 1071

Table 3 Frequencies of substructures

Mode # Analytical substructure 1 frequency (Hz) Experimental substructure 2 frequency (Hz)
1 0 0
2 90.00 Above 100 Hz

Table 4 Mass-normalized mode shapes for two substructures

Analytical mode # 1 2 Experimental mode # 1
DOF u1 0.767 −0.642 u3 0.954
DOF u2 0.767 0.917 u4 0.954

To derive the B and L compatibility and equilibrium matrices in the modal

domain, start with the physical DOF of Eq. (9) which is repeated here as
⎧ ⎫
⎪ u1 ⎪
⎪ ⎪
  ⎨ u2 ⎬
Bu = 0 1 −1 0 =0 (16)
⎪ u3 ⎪
⎪ ⎪
⎩ ⎭
u4

but the modal estimate of these DOF is

⎧ ⎫
⎪
⎪ u1 ⎪
 1
⎨ ⎪ ⎬  1
u2  0 ξ
= (17)
⎪
⎪ u ⎪ 0  2 2
⎩ 3⎪ ⎭ ξ
u4

where the superscripts denote the substructure,  are the free mode shapes at the
analytical or measured DOF, and ξ are the generalized modal DOF. Insert the mode
shapes from Table 4 and Eq. (17) into Eq. (16) to get
⎡ ⎤
0.767 −0.642 0 ⎧ 1⎫
  ⎢ 0.767 0.917 ⎥ ⎨ ξ1 ⎬
0 ⎥
Bu = 0 1 −1 0 ⎢⎣ 0 ξ1 = 0 (18)
0 0.954 ⎦ ⎩ 22 ⎭
ξ1
0 0 0.954

Combining the first and second matrices gives a new constraint matrix denoted
Bm that is not Boolean yielding
⎧ 1⎫
  ⎨ ξ11 ⎬
Bm ξ = 0.767 0.917 −0.954 ξ =0 (19)
⎩ 22 ⎭
ξ1
1072 R. Mayes and M. Allen

which is the modal constraint equation. One finds a modified version of the Lm
matrix as the null of Bm from a matrix algebra program as

⎡ ⎤
−0.5996 0.6237
⎢ ⎥
Lm = ⎣ 0.7606 0.2491 ⎦ (20)
0.2491 0.7409

which is no longer Boolean. Analogous to the physical domain Eq. (10), one has

⎧ ⎫
⎪
⎨ ξ1 ⎪
1
⎬
ξ21 = Lm q (21)
⎪
⎩ 2⎪ ⎭
ξ1

where the q vector has only two DOF. In the modal domain, the equations of motion
for the two substructures are

⎡ ⎤
 ⎧ 1⎫ 1  1  T 1  T 
⎨ ξ̈ ⎬ \ ω2
I10 ⎢ \ 0 ⎥ ξ 1 f 1 g1
+ ⎣   ⎦ = + (22)
0 I2 ⎩ ξ̈ ⎭
2 2 2 T 2 T
0 \ ω2
\
ξ 2 f 2 g2

where the physical DOF u are replaced by Eq. (17) and the first substructure (first
row of Eq. (12)) is premultiplied by 1T . The second substructure is characterized
from the frequencies and mode shapes of the test. Here mode shapes are mass
normalized to give the identity mass matrix, and the stiffness matrix is the circular
frequency squared for each mode. Analogous to the primal method in the physical
domain, one premultiplies by LTm and substitutes Eq. (21) in for the modal ξ DOF
to obtain

⎡ 1 ⎤
    \ ω2  
I10 q̈ ⎢ \ 0 ⎥ q
LTm Lm 1 + LTm ⎣  2 ⎦ Lm 1
0 I2 q̈2 0 \ ω2 q2
\
(23)
 1
  
1T
 f 1T
 g1
= LTm T 2 + LTm T
2 f 2 g 2
19 Experimental Dynamic Substructures 1073

and the far right-hand term goes to zero analogous to the physical Eq. (11). Putting
in the numbers obtained from Eq. (20) and the circular frequencies squared derived
from Table 3 gives

     
10 q̈1 1.8501 −0.6056 q1
+ 1.0e5
01 q̈2 − 0.6056 0.1983 q2
⎧ ⎫
⎪
⎪ f1 ⎪ (24)
 ⎨ ⎪ ⎬
−0.9479 0.2374 0.2374 0.2374 f2
=
−0.3185 −0.7065 −0.7065 −0.7065 ⎪
⎪ f ⎪
⎩ 3⎪ ⎭
f4

Solving the eigenvalue problem from the left-hand side of Eq. (24) gives
estimates for the natural frequencies of the assembly, which are 0 and 72.03 Hz.
This is over 70% error for the first elastic modal frequency compared to the full
physical system model. Comparing the CMS elastic mode shape to the true shape
of the full physical system model in Table 5 also shows some significant error. DOF
2, 3, and 4 are all the same since only a rigid shape for substructure 2 was utilized
in the modal model (since no elastic modes were available below the chosen cutoff
frequency).
This provides a gross illustration of a typical modal truncation problem encoun-
tered when utilizing free mode shapes as the basis vectors for experimental CMS.
The lower-frequency free modes in the testable frequency range of the experimental
substructure may not strain the experimental substructure like the fully coupled
system does.
Now consider a modified experiment where experimental substructure 2 is
modified by attaching a fixture (mass of 0.2) to DOF u3 that will cause the
experimental substructure first elastic mode to drop down to 75.3 Hz which is below
the cutoff frequency. Now two modes are included in the experimental substructure.
However, the assembly now has the additional mass of the fixture, which is not part
of the assembly of interest. To address this, one can simply add a third substructure
and subtract its effect. The equations that set up this modified fixture test are given
below, with substructure 3 representing the fixture mass.

Table 5 First elastic mode shape comparison

Mode 2 full physical system Mode 2 truncated CMS
u1 −0.5699 −0.8018
u2 = u3 −0.2695 0.4454
u4 0.7855 0.4454
1074 R. Mayes and M. Allen

⎡ 1 ⎤
⎧ 1⎫ \ ω2 ⎧ 1⎫
⎡ ⎤ ⎪ ⎪ ⎢ 0 0
I1 0 0 ⎨ ξ̈ ⎪
⎪ ⎬ ⎢ \
 2 ⎥ ⎪ξ ⎪
⎥⎨ 2⎬
⎣ 0 I2 0 ⎦ ξ̈2 + ⎢ 0 \ ω2 ⎥
⎪ 3⎪ ⎢ \ 0 ⎥ ξ ⎪
0 0 −I3 ⎪⎩ ⎪ ⎭ ⎣  3 ⎦ ⎪
⎩ 3⎭
ξ̈ 0 0 \ − ω2 ξ
\ (25)
⎧ T 1⎫ ⎧ T ⎫
⎪
⎨ f ⎪
1
⎬ ⎪
⎨ g ⎪
1 1
⎬
2 T
= 2 f + 2 g2
T
⎪
⎩ 3T 3 ⎪ ⎭ ⎪
⎩ 3T 3 ⎪
⎭
 f  g

After inserting the mode shapes, the constraint equation becomes

Bu = Bξ
⎧ ⎫ ⎧ ⎫
⎪
⎪ η11 ⎪
⎪ ⎪
⎪ 0⎪⎪
 ⎪
⎪
⎪ 1
⎪
⎪
⎪
η2 ⎬ ⎨ 0 ⎪
⎪ ⎪
⎪
⎪
⎪
−0.7670 0.9167 0.8771 1.6013 0.0000000 ⎨ ⎬
= Bm ξ = η1 = 0
2
0.0000 0.0000 −.8771 −1.6013 −2.2361 ⎪ ⎪ ⎪0⎪
⎪ ⎪
⎪
⎪ η2 ⎪⎪ ⎪ ⎪
⎪ ⎪
⎪
⎪
⎪
⎩ 23 ⎪ ⎩ ⎪
⎭ ⎪ ⎭
η1 0
(26)

and one can then find the null space of Bm to obtain

⎧ ⎫
⎤⎪
1 ⎪
⎡ ⎪ 1 ⎪
⎪ gm1
⎪ ⎧ ⎫
0.4310 0.0386 0.8469 −0.2796 −0.1320 ⎪ ⎨ gm2 ⎪
⎪ ⎪
⎬ ⎨0⎬
⎢ ⎥
LTm gm = ⎣ 0.7868 0.0705 −0.2796 0.4896 −0.2410 ⎦ gm1 2 = 0 (27)
⎪ 2 ⎪ ⎩ ⎭
0.1536 0.8651 −0.1777 −0.3244 0.3020 ⎪ ⎪
⎪ m1 ⎪
⎪
g ⎪
⎪
⎪ 0
⎩ g3 ⎭
m1

When one does this, the answers are exactly the same as the full physical system.
Other advantages of adding a fixture will be discussed later. Note several things
here. In this case, Lm was determined from a matrix algebra program as the null
space of Bm . The gm vector is the far right-hand term of Eq. (25). Five DOF are
contained in the equation of motion, two modes for substructures 1 and 2 and one
rigid body mode with 0 frequency for the fixture, substructure 3. Two constraints
are applied to couple substructures 1 and 2 and to couple substructures 2 and 3.
When one substitutes Lm q into Eq. (25) and premultiplies by LTm , the five-DOF
problem is reduced by the two constraints to a three-DOF problem. The far right-
hand term of Eq. (25) also goes to zero. Note the subtraction of the stiffness and
mass of substructure 3 in the equation of motion. To summarize the use of the
primal formulation with either physical or generalized DOF, one needs the equation
of motion, some form of the constraint B from which L is derived, and then the
equation of motion can be modified to eliminate redundant DOF using the L matrix.
19 Experimental Dynamic Substructures 1075

2.4 Connecting Substructures in the Frequency Domain:

Frequency-Based Substructuring

Frequency-based substructuring (FBS) is generally developed with the dual formu-

lation, where all DOF on each substructure are retained in the solution. The equation
of motion is formulated with a focus on the equilibrium forces where

g = −BT λ (28)

which automatically satisfies equilibrium Eq. (11) since LT BT = 0 because BT is

in the null space of LT . λ is a vector of Lagrange multipliers corresponding to the
interface force magnitudes. The equation of motion is rearranged slightly with each
substructure block diagonal with the stiffness and mass matrices as

Mü + Ku + BT λ = f. (29)

The mass, stiffness (and damping) matrices correspond to the impedance.

Generally measurements are gathered in terms of frequency response functions,
which are the inverse of impedance. Using Y as the frequency response function
(FRF), receptance matrix yields

Y−1 u + BT λ = f. (30)

Repeating the constraint equation

Bu = 0 (9)

and using it with Eq. (30), one can eliminate λ and obtain a long result as
  −1 
u = Y − YB BYB
T T
BY f (31)

which still retains all DOF from each substructure. This provides response of
every possible output from every possible input. Many times only a few rows are
of interest. Two convenient subsets of this large matrix will be given here and
demonstrated with the example problem. Suppose one desires the response at a point
on substructure 2 due to a local force on the substructure 1. This can be extracted
from Eq. (31) as
 −1
Yri = Y2rc Y2cc + Y1cc Y1ci (32)

where subscript r is a response DOF, subscript i is the input force DOF, and subscript
c implies all the connection DOF, and the superscripts denote FRFs for the particular
substructure by itself. If, for example, there were 9 connection DOF, the size of the
1076 R. Mayes and M. Allen

Ycc matrix for each substructure would be 9x9 for every frequency line. Of course,
the connection DOF must be in the same order in each substructure matrix.
Now consider the example problem from Fig. 2. Choosing f1 as the input and u4
as the output for this case and noting that the connection DOF on substructure 1 is
2 while the connection DOF on substructure 2 is 3 gives

Y41 = Y43 (Y33 + Y22 )−1 Y21 (33)

The other common problem is to obtain a response on the same substructure as

the force, which can be written generally as
 −1
Yri = Y1ri − Y1rc Y2cc + Y1cc Y1ci (34)

where this assumes the force and input are both on substructure 1. For the example
problem, choosing f1 and u1 as input and response yields
 −1
Y11 = Y11
1
− Y12
1 2
Y33 + Y22
1 1
Y21 . (35)

Equations (32) and (34) can be applied to many problems.

For an example, Eq. (33) is selected. We could obtain the analytical substructure
1 FRFs with either a direct FRF calculation or superposition of modal FRF solutions
from the FE model. Damping is usually included. We could obtain the experimental
substructure 2 FRFs by measuring response of both u3 and u4 due to input on
the connection DOF u3 in a “free” boundary condition modal test. Y43 could be
obtained by reciprocity with Y34 . Deriving the acceleration/force FRFs from all the
free modes of vibration of each substructure yields the FRF magnitudes given in
Fig. 3. Note that the resonant frequency for substructure 2 is outside the 100 Hz
bandwidth, but the effect of the elastic resonance is captured in the “tails” of the
120 Hz mode visible in the 100 Hz band.

2 FRFs Substructure 1 2 FRFs Substructure 2

10 10
H43
H33

1 1
10 10

0 0
10 10

-1 H21 -1
10 10
H22

-2 -2
10 10
0 20 40 60 80 100 0 20 40 60 80 100
Frequency - Hz Frequency - Hz

Fig. 3 FRFs for substructures 1 and 2 (not connected)

19 Experimental Dynamic Substructures 1077

2 Y41 FRF Combined Substructures 1 and 2

10

X: 42
Y: 55.35

1
10
X: 93
Y: 4.47

0
10

-1
10
0 20 40 60 80 100
Frequency - Hz

Fig. 4 FRF response at u4 due to force at u1

In Fig. 4 is given the response of u4 due to the force on u1 with the substructures
coupled using Eq. (33). The new frequencies near 42 and 93 Hz match the full sys-
tem response that one would compute with the full assembly equations of motion.
The clear advantage of the FBS is that even if the resonances of higher modes are
not captured, their information is still captured in the residual effects or “tails” of
the resonances. Given this advantage, why not use FBS all the time? Some of the
difficulties of experimental substructuring will be addressed in the next section.

3 Dealing with Experimental Difficulties

The two substructure example used in Sect. 2 is instructive, but it only illustrated one
of the several difficulties encountered in experimental substructures: that difficulty
was the effect of modal truncation shown in the CMS method. The lumped mass
system with only one connection DOF per substructure is not representative of most
substructures of interest.

3.1 Common Experimental Difficulties

The general framework [1] enumerates several of the experimental difficulties to

which we add a few others including

1. Measuring rigid body modes

2. Modal fitting of nonlinear response
1078 R. Mayes and M. Allen

3. Mass errors introduced by sensors

4. Modal truncation errors
5. Measuring rotational DOF motion and forces
6. Continuity of the attachment interface
7. Difficulty in mounting sensors at connection points
8. Dynamic effects in the joints
9. Experimental errors

Mitigation strategies for each of these difficulties are provided below.

3.1.1 Measuring Rigid Body Modes

For free substructures, the most important modes to extract properly are often not
the elastic modes, but the rigid body modes. One can have all the elastic modes
represented perfectly and still obtain erroneous results. One way to understand the
importance of rigid body modes is to consider the real part of acceleration/force
FRFs. The natural frequency of a SDOF system is determined by the frequency
at which the FRF real part goes to zero. When many modes of two substructures
are active, the way the FRF real parts sum to zero determines where the natural
frequencies of the connected system will be. The effect of the rigid body modes is
entirely on the real part in the FRFs. If one eliminates or has poor estimates of the
rigid body modes, this drastically changes the estimates of the frequencies at which
the real parts will sum to zero. Usually the rigid body mode shapes can be estimated
from mass properties of the structure. These may be either measured or obtained
from a solid model. A good check on rigid body mode shapes is to synthesize FRFs
at all the measurement DOF using only the rigid body mode shapes. Then compare
the real part of the measured FRFs at low frequency to the synthesized FRFs. If
there are large differences, something is probably wrong, and the substructuring
effort will fail.
Consider again the two substructure example in Fig. 2. The mass line of the
accelerance FRFs will be (m3 + m4)−1 = 0.909. The magnitude of the mass lines
and the true FRFs are plotted in Fig. 5 and show agreement at low frequency. If
there were error in the rigid body mode shapes, this plot would indicate a mismatch
at low frequency.

3.1.2 Modal Fitting of Nonlinear Response

If one has to fit the modal frequency, damping, and shapes as in the experimental
CMS approach, inaccuracies in the fitting can create large errors in the model. As an
example, some algorithms extract the frequency and damping and then use those in
a separate step to fit the mode shapes. If the damping estimate is off by a factor of 2,
then the residue calculation (which leads to the mode shape) will compensate by a
factor of 2 for the damping error in order to fit the FRFs. In many mode shape fitting
algorithms, the drive point mode shape is used to determine the modal mass scaling
for all the shapes. If the drive point response is very weak for a certain mode, large
errors in the scaling can result. Large errors in mode shape create large errors in the
CMS substructure model. A nonlinear response will often make it difficult to obtain
19 Experimental Dynamic Substructures 1079

2
10
H33
H43
1 Mass Line H33
10
Mass Line H43

0
10

-1
10

-2
10
0 20 40 60 80 100
Frequency (Hz)

Fig. 5 FRFs for substructure 2 compared with mass line for rigid body mode shapes

accurate modal parameters and shapes. Therefore, fitting at low excitation levels is
usually the best method to obtain FRFs fit the linear model form. There are some
exceptions to this general rule.

3.1.3 Mass Errors Introduced by Sensors

When performing an experiment with a shaker, the force sensor is mounted to the
substructure. Significant errors can be introduced if the load cell mass is significant
with respect to the test article. Even in the designed direction of force measurement
of the sensor, roughly half of the mass of the load cell is not accounted for in the
FRF measurement. The mass of the sensor in the lateral (or shearing) directions is
not corrected at all in the FRF measurement. Examples have been observed in which
the load cell causes a 10% downward shift in the modal frequency (which is roughly
20% increase in modal mass!). For light or small structures, this makes an impact
hammer a better approach. Impact hammers can also excite a structure to higher
frequencies than typical modal shakers. Shakers are desirable to produce linearized
FRFs using random input, but beware of mass loading effects of load cells. A good
check is to remove the load cell and impact the structure with a hammer to see if
the impact FRFs overlay well with the shaker and load cell FRFs. Mass loading can
also be experienced from accelerometers, but most testers are aware of this. Mass
loading from the load cell is more often overlooked.
The next six difficulties to be addressed were largely responsible for the failures
that were experienced in the early decades of experimental dynamic substructuring.
They can all be, at least partially, mitigated by proper testing with a fixture (later
called a transmission simulator) attached at the interface. This approach will be
described later.
1080 R. Mayes and M. Allen

3.1.4 Modal Truncation Errors

All modal methods of model reduction are only accurate in a limited frequency
band. Even then, some mode shape bases are better than others, giving accurate
results over a wider bandwidth for a similar number of modal DOF. For example,
the Craig-Bampton shapes (fixed boundary mode shapes plus static constraint
shapes) usually provide a better basis than free mode shapes given a certain
frequency band. However, measuring fixed base shapes and static constraint shapes
is usually unfeasible because 1. no fixed base exists in practice and 2. measuring
static constraint shapes is unfeasible in most laboratories. Generally free modes
are measured since this is easily accomplished with soft supports such as bungee
cords. However, free tests put no strain at the boundary locations which is very
different from the connected boundary condition that is of interest. When free modes
are used, the resulting modal truncation errors can theoretically be corrected by
adding the residual flexibility as in the Rubin and Craig-Chang methods. Residual
flexibilities can theoretically be extracted from the FRFs of a free modal test.
This becomes very demanding if there are many connection DOF since one must
theoretically have a drive point at every connection DOF, including the rotations. It
is beyond the state of the art to apply moments and measure rotations on systems
with a significant number of connection DOF. As was shown in the CMS example
of Sect. 2.3, one way to deal with out-of-band modes is to bring them down in
frequency into the band by attaching a rigid fixture. The fixture approach to be
described later enhances the result, but modal truncation error will increase with
increase in frequency.

3.1.5 Measuring Rotational DOF Motion and Forces

Most load cells and motion sensors measure translational motion. Measuring
moments and rotations usually requires much larger, more expensive sensors that
are more difficult to attach to a test structure and may cause huge mass loading as
discussed previously. Some progress is being made in sensing rotational motion, but
little progress has been made on applying and measuring moments except for very
specific applications. In most substructuring applications, the effects of moments
and rotations cannot be neglected without causing significant errors. Some efforts
have been attempted to obtain rotations by differencing translation measurements
and dividing by the distance between them. This can be effective for a small number
of connection DOF. For more general cases, the transmission simulator approach
with the CMS method to be discussed later addresses this need.
One great difficulty of the pure FBS method using raw measured FRFs is
measurement of all the drive point FRFs if there are many connection DOF. This
is certainly impractical if there are rotational connection DOF. Since there are
frequently many connection DOF, it can often prove impractical even if only the
translation connection DOF need to be measured. Therefore, large numbers of
connection DOF or rotation connection DOF make the raw FRF approach quite
difficult.
19 Experimental Dynamic Substructures 1081

3.1.6 Continuity of the Attachment Interface

As shown in the compatibility theory, substructuring compatibility equations focus
on very discrete attachment locations. Many attachments are not discrete but are
continuous, such as a bolted flange which has large areas in contact. Usually such
attachments are over-simplified to discrete locations with little knowledge of the
error this may induce. The transmission simulator approach to be discussed can
address this challenge as well.

3.1.7 Difficulty in Mounting Sensors at Connection Locations

Many times one cannot even attach a sensor at the location of the connection DOF.
For example, a small shoulder or a large threaded joint or a bolt hole can provide
difficult mounting conditions. Use of the transmission simulator approach to be
described later does not require sensors to be mounted at the precise connection
location.

3.1.8 Dynamic Effects in the Joints

One difficulty with traditional substructuring is that it perfectly enforces compat-
ibility at the attachment locations, when in the real hardware, there is actually
compliance in the joint connection. The joint is also a significant source of energy
dissipation. Testing with the transmission simulator, to be addressed later, captures
both joint compliance and damping.

3.1.9 Experimental Errors

The examples given previously show that substructuring theory works very well
with perfect data. However, experimental measurements always have both random
(noise) and bias (sensitivity) errors. This tends to provide difficulty, especially in
multiple connection applications. Typical accelerometers are considered calibrated
if their sensitivity is within 3–5% error at every frequency in the usable range. If
there is a 5% error in displacement at one location and a negative 5% error in
another location, the internal forces to correct for the false incompatibility can cause
huge errors. This incompatibility creates very ill-conditioned substructure matrices
when substructures are assembled at the physical connection DOF. Ill-conditioning
in either stiffness/mass or FRF matrices often manifests itself with spurious non-
physical modes. Traditionally, one way to overcome such problems is to instrument
more than necessary to provide least squares fitting through experimental data.
Another mitigating strategy is to weaken the constraint. The transmission simulator
approach in the next section takes advantage of both of these strategies.

3.2 The Transmission Simulator Approach to Mitigate

Traditional Experimental Difficulties

The difficulties 4–9 listed above can be mitigated by the transmission simulator
approach. The transmission simulator approach does require some additional
1082 R. Mayes and M. Allen

upfront preparation, but it eliminates or mitigates the most grievous difficulties

that can cause experimental substructures to fail. The mitigation of difficulties
will be listed along with the transmission simulator features in the description
that follows. A transmission simulator consists of two entities, an instrumented
physical fixture and a mathematical model of the fixture (usually a FE model). The
fixture is fabricated to mate to the experimental substructure exactly like the next
assembly will fit. The fixture should emulate the impedance of the nearest portion
of the next assembly. Some “features” may be included on the fixture for attaching
sensors or imparting various input forces. The FE model of the fixture is utilized
to determine where sensors should be placed and loads applied in the experimental
substructure test. A chosen number of free modes, usually going slightly beyond
the frequency bandwidth of interest, are chosen to represent the fixture motion,
so it is possible to represent both rigid and elastic motion with the approach. The
fixture is then instrumented with enough sensors to capture the mode shapes so
that all shapes are linearly independent on the sensor set. Usually the number of
sensors is 1.5 to 2 times the number of free modes chosen to take advantage of
least squares fitting of the experimental data (difficulty 9). The sensors are placed
at convenient locations on the fixture. They do not have to be placed at the actual
connection locations. The FE model of the fixture is used to convert all the measured
motions to the modal DOF of the fixture. The modal DOF inherently capture the
rotational motion and translational motion at connection DOF that are difficult to
measure directly (difficulties 5 and 7). In addition the modal DOF of the fixture
approximate the motion at the connection for continuous interfaces (difficulty 6).
Because the fixture is mounted to the substructure in the test with the same type
of joint that will be in the assembled system, the compliance and damping in the
joint is captured in the substructure model (difficulty 8). The mass of the fixture
also helps bring modes down into the testable bandwidth as demonstrated in the
example in Sect. 2.3 (difficulty 4). If the chosen mode shapes span the space of
the actual connection motion, the transmission simulator approach will be accurate.
To the extent the chosen mode shapes do not span the space of the connection
motion, inaccuracy (a form of modal truncation error) will be introduced in the
connected substructures model. However, the mode shapes of the transmission
simulator generally improve the connection shape bases a great deal over free modes
without the fixture (difficulty 4).

3.3 An Example Using the Transmission Simulator Approach

Consider the hardware shown in Fig. 6. A FE substructure of the cylinder will be

coupled with an experimental model of the plate/beam at the eight bolted locations
to produce full system response. The transmission simulator, shown in Fig. 7, is a
ring with tabs on it to allow for exciting the experiment in the tangential direction.
In this problem, the model of the transmission simulator was “welded” into the
FE model of the cylinder. Then the experimental substructure with transmission
simulator was coupled to the cylinder, and two transmission simulators were
19 Experimental Dynamic Substructures 1083

Fig. 6 Full system assembled for cylinder with plate/beam – experimental substructure at right

Fig. 7 Transmission simulator – triaxial accelerometer locations shown at right

subtracted to produce full system response. In hindsight, the authors believe that
even better results could be obtained if the tabs on the transmission simulator were
shorter, so the tab modes would have been above the testable frequency band.
Measurements were made on the end of the beam (Fig. 6) so that drive point
response for an axial input on the end of the beam in the assembled system was
developed with the transmission simulator method using FBS coupling. The FRFs
for the experimental substructure were developed from analytical rigid body mode
shapes and measured modes of the experiment to 4,000 Hz so they do not include
residual effects of modes above 4,000 Hz. FRFs for the cylinder were calculated
from 100 modes of the FE model of the cylinder. The resulting coupled FRF out to
2,000 Hz is shown in Fig. 8.
Since the plate/beam substructure was fairly simple, a FE model was generated
for it. The FE model matched test data frequencies for the plate/beam alone quite
well. A virtual test was then performed on the FE model of the plate and beam,
and a virtual experimental model without the transmission simulator was generated
from 21 modes (all the free modes to 4,000 Hz). The CMS free mode models were
coupled at all 48 connection DOF (24 rotations and 24 translations) to the 100 mode
1084 R. Mayes and M. Allen

1000y driving point

4
10

2
10

Substructure Model
Test Data
0
10
2 3
10 10
Frequency - Hz

Fig. 8 FBS coupled analytical and experimental substructures versus test data

free mode model of the cylinder. The result is shown in Fig. 9. The FRF errors are
pretty drastic, similar to the example CMS problem without the fixture given in
Sect. 2.3. It is difficult to even establish correspondence of the resonances between
the test data and substructure result. Many more free modes above 4,000 Hz (which
might be quite difficult to extract) would need to be included in the experimental
substructure to improve the CMS result. The result using the transmission simulator
fixture in Fig. 8 is clearly more useable, and no rotational measurements were
required.

3.4 Transmission Simulator Theory

Transmission simulator (TS) theory is developed with its own version of the B
and L matrices and at least one additional substructure, the instrumented fixture,
or transmission simulator [2]. We will consider the theory in light of the goal of
coupling an experimental substructure to a FE substructure.

3.4.1 Preparation to Implement Transmission Simulator Theory

Before the measurement process for the experimental substructure is undertaken,
the fixture is fabricated, and a model of the fixture is generated, usually with
19 Experimental Dynamic Substructures 1085

Drive Point 1000Y FRF

5
10

4
10

3
10

2
10

1
10
Test Data
No Fixture 21 Mode Result to 4000 Hz
0
10
3
10

Fig. 9 FBS coupled model with virtual experimental model computed for 24 translations and 24
rotations with all free modes out to 4,000 Hz and no transmission simulator – a demonstration of
free modes modal truncation error with no fixture

FE software. The FE model of the fixture, or transmission simulator, is used to

calculate modes out significantly beyond the frequency of interest at many candidate
measurement DOF, e.g., a factor of 10 more DOF than the required number of
sensors. Then a number of retained modes will be selected that are deemed to be
able to span the space of the connection motion. Usually modes past the frequency
of interest are retained. The candidate measurement DOF are all chosen at locations
where sensors can easily be installed. Then an algorithm is used to reduce the
number of sensor DOF down from the candidate set to about 1.5 to 2 times
the number of retained modes. Various approaches can suffice for this, such as
the effective independence method [3], min-mac [4], or optimizing to obtain a
condition number of the mode shape matrix that is no more than 3.5–5 [5]. This is
important since the TS mode shape matrix must be inverted for this method, and ill-
conditioning of the mode shape matrix can ruin the results. Once the sensor locations
are established, one usually identifies force input locations that will excite all free
modes of the TS well. One must have a drive point response sensor co-located with
each force input. If all of the measurement DOF are physically located so they are
in contact with DOF on the FE model, only one TS is required. If some DOF are
not physically in contact with the FE substructure as nodes A10X on the tabs in
Fig. 6, then one needs to attach the TS model to the FE model at all DOF where
the TS and FE model occupy the same space. See Fig. 10 which has the TS bonded
1086 R. Mayes and M. Allen

Fig. 10 Substructuring example using transmission simulator

in with the FE model. For development of the theory, assume this is the case. This
means that two TS will need to be subtracted (one from each substructure). With
this information in hand, one can now develop the compatibility equations and apply
them to the appropriate equations of motion.

3.4.2 Transmission Simulator Method Using CMS

The TS method enforces the motion of the TS to be the same on the FE substructure
and the experimental structure at the measurement DOF on the TS, but the constraint
is weakened using the free modes of the TS as a basis. Consequently, the measured
DOF, as opposed to the actual connection DOF, will be utilized in establishing
compatibility. This allows one to use easy to instrument locations on the TS as
opposed to requiring the measurement at all connection DOF which provides a great
advantage. It also allows us to approximate connections even if the connection is not
discrete but continuous. Since we have a FE substructure with a TS attached to it,
an experimental substructure with TS, and a TS substructure, one can write for the
physical DOF

exp
uFmeas
E
− umeas = 0
exp
(36)
umeas − uTmeas
S
=0

where subscript meas denotes the measurement DOF on the TS. Building the B
matrix for these DOF gives

  ⎧ uF E ⎫
⎨ meas ⎬
I −I 0 exp
Bumeas = umeas = 0 (37)
0 I −I ⎩ T S ⎭
umeas

but in the CMS approach the displacements are expressed as a function of the mode
shapes as follows.

  ⎡ F E ⎤ ⎧ FE ⎫
0 0 ⎪
⎨ξ ⎪ ⎬
I −I 0 ⎣ meas
exp ⎦ exp
Bumeas = 0 meas 0 ξ =0 (38)
0 I −I ⎪
⎩ TS ⎪ ⎭
0 0 Tmeas
S
ξ
19 Experimental Dynamic Substructures 1087

The mathematical key to the TS is to premultiply each set of constraints by the

pseudo-inverse of the retained TS mode shapes as

 ⎡ F E ⎤ ⎧ FE ⎫
  0 0 ⎪
⎨ξ ⎪ ⎬
Tmeas
S† 0 I −I 0 ⎣ meas
exp ⎦ exp
BmT S ξmeas = 0  meas 0 ξ
T
0 meas S†
0 I −I ⎪
⎩ TS ⎪ ⎭
0 0 Tmeas
S
ξ
⎧ FE ⎫
 T S† F E exp  ⎨ξ ⎪
⎪ ⎬
meas meas −Tmeas
S† 
meas 0 exp
= exp ξ =0
Tmeas meas − I ⎪ ⎩ TS ⎪
0 S† 
⎭
ξ
(39)

which projects the constraint onto the mode shapes of the TS. To the extent that the
mode shapes of the TS span the true motion, this will provide a good approximation.
Since there are fewer retained modes in the TS than the number of measured DOF,
this reduces the number of constraints which weakens the constraint. The constraint
can only be satisfied in a least squares sense instead of an absolute sense, i.e.,
Eqs. (37) and (38) will not be perfectly satisfied, and the physical displacements
at the physical connection DOF will not match precisely. The advantage of this
is that it does not force the substructuring problem to enforce compatibility for
the experimental errors in the measured data (e.g., the 3–5% sensitivity errors
of accelerometers). This improves the conditioning of the matrices tremendously
and eliminates most of the spurious modes that would be obtained if one utilized
compatibility Eq. (38) which would enforce the measured motion with its errors to
satisfy compatibility. The related LmTS matrix is just the null matrix of BmTS . The
block diagonal modal equations of motion start with
⎧ ⎫
⎡ ⎤ ⎪ F E ⎪ ⎡ 2F E ⎤ ⎧ FE ⎫
IF E
0 0 ⎨ ξ̈ exp ⎪
⎪ ⎬ ω 0 0 ⎪
⎨ξ ⎪ ⎬
⎣ 0 I exp 0 ⎦ ξ̈ +⎣ 0 ω 2 exp 0 ⎦ ξF E
⎪ TS ⎪ ⎪ ⎪ ⎪
0 0 −2IT S ⎪⎩ ⎭ 0 0 −2ω2T S ⎩ ξF E ⎭
ξ̈
⎧ T FE
⎫ ⎧ ⎫ (40)
⎪ f ⎪ ⎬ ⎪ g ⎪
F E F E T FE
⎨ ⎨ ⎬
T exp T
= exp f + exp gexp
⎪
⎩ T ST T S ⎪ ⎭ ⎪ ⎩ T ST T S ⎪⎭
 f  g

where one can see that two TS models are being subtracted in the mass and stiffness
(and also damping if it is included). Analogous to Eq. (21)
⎧ FE ⎫
⎪
⎨ξ ⎪ ⎬
exp
ξ = LmT S q (41)
⎪
⎩ TS ⎪ ⎭
ξ
1088 R. Mayes and M. Allen

which one substitutes into Eq. (40) and premultiplies by transpose of LmTS to obtain
a result that has the same form as was seen in examples previously,

⎡ ⎤
IF E 0 0
LTmT S ⎣ 0 Iexp 0 ⎦ LmT S q̈
0 0 −2IT S
⎡ ⎤ ⎧ ⎫
⎪ F E T fF E ⎪
ω2F E 0 0 ⎨ ⎬
+ LTmT S ⎣ 0 ω2 exp 0 ⎦ LmT S q = LTmT S expT fexp
⎪
⎩ T ST T S ⎪ ⎭
0 0 −2ω2T S  f
(42)

where the far right-hand term of Eq. (40) goes to zero because of the equilibrium
condition, similar to Eqs. (11) and (14). New frequencies and eigenvectors are
obtained by solving the eigenvalue problem of the left side. If those eigenvectors
are given in matrix , the new motion will be

u = ξ = LmT S q = LmT S p (43)

so the new mode shapes are formed as LmTS . The number of constraint equations
in Eq. (39) is 2* NTS , where NTS is the number of retained modes for the TS. The
total number of modes obtained by solving the eigenvalue problem of Eq. (42) is
NFE + Nexp + NTS − Nconstraints = NFE + Nexp − NTS .

3.4.3 Transmission Simulator Method Using FBS

For FBS, consider just attaching two substructures, the FE and the experimental
substructures. (For brevity, subtracting the TS is not shown in the example below;
subtraction could be done in a second step.) Compatibility begins with physical
DOF equation as
 E 
  uFmeas
Bumeas = I −I exp =0 (44)
umeas

and premultiply by the pseudo-inverse of the TS mode shape matrix at the measured
DOF to get a new B matrix as
 E   E 
  uFmeas  T S+  uFmeas
BF BS,T S umeas = Tmeas
S+
I −I exp =  meas − T S+
meas exp = 0.
umeas umeas
(45)

BFBS,TS can be substituted directly into FBS Eq. (31) for the traditional B matrix.
Eq. (31) is repeated here as
19 Experimental Dynamic Substructures 1089

  −1 
u = Y − YB BYB
T T
BY f. (31)

To show how this would work in the subset of equations that cover many
situations, first consider Eq. (32) where the force is applied to the FE substructure
and the response is on the experimental substructure. The classical FBS equation is
 −1
exp exp
Yri = Yrc Ycc + YFccE YFciE (46)

where the exp and FE superscripts have replaced the substructure numbers. With the
TS approach using the FE and EXP substructures, with BFBS_TS the form is

T
 T

T −1
exp exp
Yri = Yrc Tc S+ Tc S+ Ycc Tc S+ + Tc S+ YFccE Tc S+ Tc S+ YFciE (47)

T
where all the c DOF are actually the measured DOF (i.e., Tc S+ = Tmeas S+ ). If

one looks at the kernel for the connection DOF in the parentheses which must be
inverted, the matrix has been reduced in size from the number of measurement DOF
on the TS to the number of modes of the TS. This has the effect weakening the
constraint, which reduces the ill-conditioning of the kernel for inversion. If the force
is applied to the FE substructure and measured on the FE substructure, the traditional
Eq. (34) is repeated here as
 −1
exp
Yri = YFriE − YFrcE Ycc + YFccE YFciE (48)

where the number superscripts have been replaced by the appropriate substructure
name. The TS version of Eq. (34) for our example is

T
 T

T −1 T S+ F E
exp
Yri = YFriE − YFrcE Tc S+ Tc S+ Ycc Tc S+ + Tc S+ YFccE Tc S+ c Yci
(49)

where the c DOF are the measured DOF.

3.5 Practical Guidance Using the Transmission Simulator

Approach

There are several practical guidelines to using the TS approach which deserve
mention here.

1. The TS fixture should be designed to be as simple as possible to fabricate and

model. Furthermore, it is best if the fixture can be fabricated from one piece
1090 R. Mayes and M. Allen

of material to eliminate the uncertainty and nonlinearity of joints on the fixture

dynamics.
2. Fabricate the fixture of the same material and with the same interface geometry
as the assembly it is replacing to capture the joint compliance and dissipation as
realistically as possible.
3. The fixture should capture the local impedance of the assembly it is replacing.
This allows the joint to be excited realistically. The mass of the fixture will help
bring modes down in frequency into the testable bandwidth. It is best to err on
the side of making the fixture too stiff, rather than too flexible.
4. Cautiously add features to help obtain difficult measurements. In the example,
tabs were added to allow accelerometer and force input locations that would
be difficult to obtain on a pure ring. However, in the example, the tabs were
too long, so that they had their own modes without much motion of the rest of
the fixture. This actually degrades the results with the method. (In hindsight,
shorter tabs would have improved results in the example.) Avoid features that
result in repeated roots for the fixture modes. Add features that cause each natural
frequency to be distinct. Add features that provide force input locations that will
improve the excitation of poorly excited modes of the TS.

In addition, several guidelines are worth mentioning that are not specifically
related to the TS approach.

1. Reduce the frequency bandwidth to the minimum required for the problem
requirements. All substructuring methods suffer from modal truncation errors
at some frequency. As one increases the bandwidth and adds more shapes to the
bases, the possibility that the bases are insufficient to represent high-frequency
motion increases. Many substructuring results have failed in an attempt to obtain
a result for a much larger bandwidth than was truly required.
2. Where forces will be applied, a drive point motion sensor should be located
perfectly co-linear with the input force. Either apply the force on the surface
directly opposite of the sensor, or build a cap to put over the sensor, so that the
force input may be co-linear with the motion sensor. (This is required for proper
mode shape scaling in the CMS method or connection drive point measurements
in FBS substructuring.)
3. Use low-level force inputs to minimize nonlinear response and improve the
capability for fitting the modal parameters. (In most cases, low-level excitation
will not excite nonlinear response as much as high-level excitation. There are
some exceptions.)
4. Utilize hammer force inputs to a. minimize mass distortion due to shaker load
cell mass and b. increase the frequency bandwidth of excitation over what can
typically be obtained with shakers. If the substructure is massive, the mass of
the load cell and frequency capability of the shaker may not be hindrances, so
shakers may be a very good option.
5. When using CMS, make sure FRFs synthesized from the modal parameters
reproduce the test FRFs adequately. For free substructures, the low-frequency
19 Experimental Dynamic Substructures 1091

real part of the FRFs (mass line) should match to show that the analytical rigid
body mode shapes are accurate. The elastic resonance amplitudes should match
to show that the frequency/damping/mode shape/modal mass was extracted
properly. Especially look to see that responses match for FRFs from references
that were NOT used to extract the shapes. If they do not match, the drive point
mode shape from the original reference may not have been extracted accurately.
6. To minimize the number of input references, FBS may be performed with FRFs
accurately synthesized from modal parameters. The advantage of this is that one
does not have to have a drive point FRF at every connection DOF. One only needs
to use enough drive points to extract all the modes in the desired bandwidth.
This approach helps eliminate experimental noise from the measurements. The
disadvantage is that it cannot capture the residual effects of out-of-band modes.
7. Be aware of accuracy limitations of a coupled substructure model. The frequen-
cies of the coupled model are usually most accurate in terms of percentage error.
Usually, the mode shapes are less accurate, and the damping is the least accurate.

References
1. De Klerk D, Rixen DJ, Voormeeren SN (2008) General framework for dynamic substructuring:
History, review, and classification of techniques. AIAA J 46(5):1169–1181
2. Allen MS, Mayes RL, Bergman EJ (2010) Experimental modal substructuring to couple
and uncouple substructures with flexible fixtures and multi-point connections. J Sound Vib
329:4891–4906
3. Kammer DC (1991) Sensor placement for on-orbit modal identification and correlation of large
space structures. J Guid Control Dyn 14:251–259
4. Carne TG, Dohrmann CR (1995) A modal test design strategy for model correlation. In: 13th
international modal analysis conference, Nashville, pp 927–933
5. Bergman EJ, Allen MS, Kammer DC, Mayes RL (2010) Probabilistic investigation of sensitivi-
ties of advanced test-analysis model correlation methods. J Sound Vib 329:2516–2531
Structural Dynamics Modification
and Modal Modeling 20
Mark Richardson and David Formenti

Contents
1 Modal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
2 Design Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
3 Eigenvalue Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
4 Measurement Chain to Obtain an EMA Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . 1098
4.1 Critical Issues in the Measurement Chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1098
4.2 Calculating FRFs from Experimental Vibration Data . . . . . . . . . . . . . . . . . . . . . 1098
4.3 Sensing Force and Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1099
4.4 Sensitivity Flatness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1099
4.5 Transverse Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1100
4.6 Sensor Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1100
4.7 Sensor Mounting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
4.8 Leakage Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
4.9 Finite Length Sampling Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
4.10 Leakage-Free Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
4.11 Leakage-Free Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1102
4.12 Reduced Leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1102
4.13 Linear Versus Nonlinear Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1102
4.14 Random Excitation and Spectrum Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
4.15 Curve Fitting FRFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
4.16 Modal Models and SDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
5 Structural Dynamic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
5.1 Structural Resonances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
5.2 Truncated Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105
5.3 Sub-structuring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105

M. Richardson ()
Vibrant Technology, Inc., Centennial, CA, USA
e-mail: [email protected]
D. Formenti
BlackHawk Technology, Santa Cruz, CA, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 1093

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_25
1094 M. Richardson and D. Formenti

5.4 Rotational DOFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105

6 Time Domain Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106
6.1 Finite Element Analysis (FEA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106
6.2 FEA Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
7 Frequency Domain Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
8 Parametric Models Used for Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
8.1 Rational Fraction Polynomial Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1108
8.2 Partial Fraction Expansion Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1108
8.3 Experimental FRFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1109
9 FRF-Based Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1110
9.1 Modal Frequency and Damping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
9.2 Modal Residue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
10 Transformed Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1112
11 Dynamic Model in Modal Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1112
11.1 Damping Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114
11.2 Lightly Damped Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115
12 Scaling Mode Shapes to Unit Modal Masses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115
12.1 Modal Mass Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1116
12.2 Modal Stiffness Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117
12.3 Modal Damping Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117
12.4 Unit Modal Masses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1118
13 SDM Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1118
14 SDM Equations Using UMM Mode Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1119
15 Scaling Residues to UMM Mode Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1120
15.1 Driving Point FRF Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1121
15.2 UMM Mode Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1122
15.3 Triangular FRF Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1122
16 Integrating Residues to Displacement Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1122
17 Effective Mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124
18 Diagonal Mass Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124
18.1 Checking the Engineering Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1125
19 Effective Mass Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1125
20 SDM Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1127
20.1 Cap Screw Stiffnesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1127
21 EMA Mode Shapes of the Plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1128
22 FEA Mode Shapes of the Plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1130
22.1 Mode Shape Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1130
22.2 Modal Frequency Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1131
22.3 Hybrid Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1131
22.4 RIB FEA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1132
22.5 RIB Impact Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
22.6 Hybrid Modal Model of the RIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
23 Substructure Modal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
23.1 Block Diagonal Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
24 Calculating New Modes with SDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136
25 SDM Versus FEA Modes: Plate and RIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136
25.1 SDM Mode Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1137
26 SDM Versus EMA Modes: Plate and RIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1139
27 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1140
28 Modeling a Tuned Vibration Absorber with SDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1141
29 Adding a Tuned Absorber to the Plate and RIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1142
30 Modal Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1144
30.1 EMA Modes of the Plate and RIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1144
30.2 Using SDM to Explore Joint Stiffnesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1146
20 Structural Dynamics Modification and Modal Modeling 1095

30.3 Current Versus Target Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1146

30.4 Solution Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1148
31 FEA Modal Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1149
32 Difference Between Modal Sensitivity and FEA Model Updating . . . . . . . . . . . . . . . . 1152
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1152

Abstract

The structural dynamics modification (SDM) method, also called eigenvalue

modification, was first commercialized in the earlier 1980s by a software
company founded by the authors. It has proved to be a useful engineering tool for
providing a quick look at the influence of physical modifications to a mechanical
structure on its modes of vibration.
SDM can be used with experimental mode shapes as well as analytical mode
shapes. It provides meaningful results even when used with a few mode shapes
and combined experimental and analytical mode shape data. This chapter also
provides background information on the many details to be considered when
acquiring experimental mode shapes. Finally, the example provided shows that
SDM gives very accurate results when both translational and rotational joint
stiffnesses are used to model the attachment of a rib stiffener to a plate.

Keywords

Frequency response function (FRF) · Experimental modal analysis mode shape

(EMA mode shape) · Finite element analysis mode shape (FEA mode shape) ·
Operational modal analysis mode shape (OMA mode shape) · Unit modal mass
mode shape (UMM mode shape) · Hybrid mode shape (combined EMA &
FEA mode shape data)

1 Modal Models

SDM is unique in that it works directly with a modal model of the structure, either
an Experimental Modal Analysis (EMA) modal model, a Finite Element Analysis
(FEA) modal model, or a Hybrid modal model consisting of both EMA and FEA
modal parameters. EMA mode shapes are obtained from experimental data, and
FEA mode shapes are obtained from an analytical finite element computer model.
A modal model consists of a set of scaled mode shapes. In this chapter the mode
shapes used in a modal model are scaled to Unit Modal Masses, called UMM mode
shapes. FEA mode shapes are commonly scaled to UMM mode shapes using the
mass matrix of the FEA model. In this chapter, it will be shown how EMA mode
shapes can also be scaled to UMM mode shapes without using a mass matrix.
A modal model preserves the mass, damping, and stiffness properties of a
mechanical structure and is used by SDM to represent the dynamic properties of
the unmodified structure.
1096 M. Richardson and D. Formenti

Fig. 1 SDM input-output diagram

2 Design Modifications

Once the dynamic properties of an unmodified structure are defined in the form of its
modal model, SDM can be used to predict the dynamic effects of mechanical design
modifications to the structure. These modifications can be as simple as additions to
or removals of point masses, linear springs, or linear dampers, or more complex
modifications can be modeled using FEA elements such as rod and beam elements,
plate elements (membranes), and solid elements such as prisms, tetrahedrons, and
brick elements.
SDM is computationally very efficient because it solves an eigenvalue problem
in modal space. In contrast, FEA mode shapes are obtained by solving an eigenvalue
problem in physical space.
Another advantage of SDM is that the modal model of the unmodified structure
must only contain data for the DOFs (points and directions) where the modification
elements are attached to a geometric model of the structure. SDM then provides a
new modal model of the modified structure, as depicted in Fig. 1.

3 Eigenvalue Modification

A variety of numerical methods have been developed over the years which only
require a modal model to represent the dynamics of an unmodified structure. Among
the more traditional methods for performing these calculations are modal synthesis,
the Lagrange multiplier method, and diakoptics. However, the local eigenvalue
modification technique, developed primarily through the work of Weissenburger,
Pomazal, Hallquist, and Snyder [1], is the technique commonly used by the SDM
method today.
All of the early development work on SDM was done primarily with analytical
FEA mode shapes. The primary objective was to provide a faster means of
investigating physical changes to a structure without having to solve a much larger
eigenvalue problem. FEA mode shapes are obtained by solving the problem in
20 Structural Dynamics Modification and Modal Modeling 1097

physical coordinates, whereas SDM solves a much smaller eigenvalue problem in

modal coordinates.
In 1979, Structural Measurement Systems (SMS) began using the local eigen-
value modification method together with an EMA modal model derived from a
modal test [2–5]. The computational efficiency of this method made it very attractive
for use in a laboratory on a desktop calculator or computer. More importantly, it gave
reasonably accurate results using only a small number of EMA mode shapes in the
modal model of the unmodified structure.
A modal model with only a few mode shapes in it is called a truncated modal
model. Regardless of whether EMA or FEA mode shapes are used, truncated modal
models have been shown to adequately characterize the dynamics of a structure. The
effects of using truncated modal models were investigated in [2, 3].
The fundamental calculation of SDM is the solution of an eigenvalue problem.
The solution is computationally efficient because a small dimensional eigenvalue
problem is solved. Computational speed is virtually independent of the number of
DOFs in the modal model. Hence, large modifications involving many DOFs are
handled as efficiently as smaller modifications.
The SDM computational process is straightforward. All physical modifications
are converted into appropriate changes to the mass, stiffness, and damping matrices
of the equations of motion, in the same manner as an FEA model is constructed.
These modification matrices are then transformed to modal coordinates using
the mode shapes of the modal model of the unmodified structure. The resulting
transformed modifications are then added to the modal properties of the unmodified
structure, and these new equations are solved for the new modes of the modified
structure.
To illustrate this process, if there were 1000 DOFs in an FEA model, solving
for its FEA mode shapes requires the solution of an eigenvalue problem with mass
and stiffness matrices of the size 1000 by 1000 [6–8]. By contrast, if the dynamics
of an unmodified structure is represented with a modal model consisting of ten
mode shapes, new mode shapes resulting from a structural modification are found
by solving an eigenvalue problem with transformed mass and stiffness matrices of
the size 10 by 10.
The size of the eigenvalue problem in modal space is independent of the number
of structural modifications made to the structure. Many modification elements can
be attached to a 3D geometric model of the structure, and the SDM solution time
does not significantly increase.
SDM requires two inputs:

1. A modal model that adequately represents the dynamics of the unmodified

structure
2. Finite elements attached to a geometric model of the structure that characterize
the structural modifications

With these inputs, SDM calculates a new modal model that represents the
dynamics of the modified structure. It will also be shown in later examples that
1098 M. Richardson and D. Formenti

SDM obtains results that are very comparable to those obtained from an FEA eigen-
solution.

4 Measurement Chain to Obtain an EMA Modal Model

If a modal model containing EMA mode shapes is used with SDM, the accuracy
of the mode shapes will directly influence the accuracy of the results calculated
with the SDM method. To understand the potential errors that can occur in an EMA
modal model, it is important to review the steps in the measurement chain required
to obtain EMA mode shapes.
Three major steps are commonly used to obtain an EMA modal model:

1. Acquire experimental vibration data from the test article.

2. Calculate a set of frequency response functions (FRFs) from the vibration data.
3. Curve fit the FRFs to estimate the EMA mode shapes of the test article.

4.1 Critical Issues in the Measurement Chain

Following is a list of issues to consider in implementing a measurement chain:

1. Nonlinearity of the test structure dynamics

2. Boundary conditions of the test structure
3. Excitation technique
4. Force and response sensors
5. Sensor mounting
6. Sensor calibration
7. Sensor cabling
8. Signal acquisition and conditioning
9. Spectrum analysis
10. FRF calculation
11. FRF curve fitting
12. Creating an EMA modal model

All of these issues involve assumptions that can impact the accuracy of the EMA
modal model and ultimately the accuracy of the SDM results. Only a few of these
critical issues will be addressed here, namely, sensors, sensor mounting, sensor
calibration, FRF calculation, and FRF curve fitting.

4.2 Calculating FRFs from Experimental Vibration Data

To create an EMA modal model, a set of calibrated inertial FRF measurements is

required. These frequency domain measurements are unique in that they involve
20 Structural Dynamics Modification and Modal Modeling 1099

subjecting the test structure to a known measurable force while simultaneously

measuring the structural response(s) due to the force. The structural response is
measured either as acceleration, velocity, or displacement using sensors that are
either mounted on the surface or are non-contacting but still measure the surface
vibration.
An FRF is a special case of a Transfer Function. A Transfer Function is a
frequency domain relationship between any type of input signal and any type of
output signal. An FRF defines the dynamic relationship between the excitation force
applied to a structure at a specific location in a specific direction and the resulting
response motion at another specific location in a specific direction. The force input
point and direction and the response point and direction are referred to as the degrees
of freedom (or DOFs) of the FRF.
An FRF is also called a cross-channel measurement. It requires the simultaneous
acquisition of both the excitation force and one of its resultant responses. This
means that at least a two-channel data acquisition system or spectrum analyzer is
required to measure the signals required to calculate an FRF. The force (input) and
the response (output) signals must also be simultaneously acquired, meaning that
both channels of data are amplified, filtered, and sampled without introducing any
artificial phase difference between the two signals.

4.3 Sensing Force and Motion

The excitation force is typically measured with a load cell. The analog signal from
the load cell is fed into one of the channels of the data acquisition system. The
response is measured either with an accelerometer, laser vibrometer, displacement
probe, or another sensor that can measure the surface vibration. Accelerometers are
most often used today because of their availability, relatively low cost, and variety
of sizes and sensitivities. The important characteristics of both the load cell and
accelerometer are:

1. Sensitivity
2. Usable amplitude range
3. Usable frequency range
4. Transverse sensitivity
5. Mounting method

4.4 Sensitivity Flatness

The most common type of sensor today is referred to as an IEPE/CCLD/ICP/Delta

tron/Isotron style of sensor. This type of sensor requires a 2–10 milliamp current
supply, typically supplied by the data acquisition system, and has a built-in charge
amplifier and other signal conditioning. It also has a fixed sensitivity. Typical
sensitivities are 10mv/lb. or 100mv/g.
1100 M. Richardson and D. Formenti

The ideal frequency spectrum for any sensor is a “flat magnitude” over its usable
frequency range. The documented sensitivity of most sensors is typically given at a
fixed frequency (such as 100 Hz, 159.2 Hz, or 250 Hz) and is referred to as its 0 dB
level.
The sensitivity of an accelerometer is specified in units of mv/g or mv/(m/sˆ2)
with a typical accuracy of +/−5% at a specific frequency. The frequency spectrum
of all sensors is not perfectly flat, meaning that its sensitivity varies somewhat
over its usable frequency range. The response amplitude of an ICP accelerometer
typically rolls off at low frequencies and rises at the high end of its usable frequency
range. This specification is the flatness of the sensor, with a typical variance of
+/−10% to +/−15%.
All of this equates to a possible error in the sensitivity of the force or response
sensor over its usable frequency range. This means that the amplitude of an FRF
might be in error by the amount that the sensitivity changes over its measured
frequency range.

4.5 Transverse Sensitivity

Adding to its flatness error is the transverse sensitivity of a sensor. Both force and
vibration have a direction associated with them. That is, a force or motion is defined
at a point in a specific direction.
A uniaxial (single axis) transducer should only output a signal due to force or
motion in the direction of its sensitive axis. Ideally, any force or motion that is not
along its sensitive axis should not yield an output signal, but this is not the case with
most sensors.
All sensors have a documented specification called transverse sensitivity or cross
axis sensitivity. Transverse sensitivity specifies how much of the sensor output is due
to a force or motion that is sensed from a direction other than the measurement axis
of the sensor. Transverse sensitivity is typically less than 5% of the sensitivity of the
measurement axis. For example, if an accelerometer has a sensitivity of 100mv/g,
its transverse sensitivity might be 5%, or about 5mv/g. Therefore, 1 g of motion in a
direction other than the sensitive axis of an accelerometer might add 5mv (or 0.05 g)
to its output signal.

4.6 Sensor Linearity

Another area affecting the accuracy of an FRF is the linearity of each sensor output
signal relative to the actual force or vibration. If a sensor output signal were plotted
as a function of its input force or vibration, all its output values should lie on a
straight line. Any values that do not lie on a straight line are an indication of the
nonlinearity of the sensor. The nonlinearity specification is typically less than 1%
over the usable frequency range of a sensor.
20 Structural Dynamics Modification and Modal Modeling 1101

As the amplitude of the measured signal becomes larger than the specified input
amplitude range of the sensor, the signal will ultimately cause an overload in the
internal amplifier of the sensor. This overload results in a clipped output signal from
the sensor. A clipped output signal is the reason why it is very important to measure
amplitudes that are within the specified amplitude range of a sensor.

4.7 Sensor Mounting

Attaching a sensor to the surface of the test article is also of critical importance.
The function of a sensor is to “transduce” a physical quantity, for example, the
acceleration of the surface at a point in a direction. Therefore, it is important to
attach the sensor to a surface so that it will accurately transduce the surface motion
over the frequency range of interest.
Mounting materials and techniques also have a usable frequency range just like
the sensor itself. It is very important to choose an appropriate mounting technique
so that the surface motion over the desired frequency range is not affected by the
mounting material of method. The use of magnets, tape, putty, glue, or contact
cement is convenient for attaching sensors to surfaces. But attaching a sensor using
a threaded stud is the most reliable method, with the widest frequency range.

4.8 Leakage Error

Another error associated with the FRF calculation is a result of the FFT algorithm
itself. The FFT algorithm is used to calculate the Digital Fourier Transform (DFT)
of the force and response signals. These DFTs are then used to calculate an FRF.

4.9 Finite Length Sampling Window

The FFT algorithm assumes that the time domain window of acquired digital data
(called the sampling window) completely contains the acquired signal. If an acquired
signal is not fully captured within its sampling window, the DFT of the signal will
contain leakage error.

4.10 Leakage-Free Spectrum

The spectrum of an acquired signal will be leakage-free if one of the following

conditions is satisfied.

1. If a signal is periodic (like a sine wave), then it must make one or more complete
cycles within the sampled window.
1102 M. Richardson and D. Formenti

2. If a signal is not periodic, then it must be completely contained within the

sampled window.

If an acquired signal does not meet one of the above conditions, there will be
errors in its DFT and errors in the FRF that is calculated using the DFT. Leakage
error causes both amplitude and frequency errors in a DFT and in a FRF that uses
the DFT.

4.11 Leakage-Free Signals

Leakage is eliminated by using testing signals that meet one of the two conditions
stated above. During impact testing, if the impulsive force and the impulse response
signals are both completely contained within their sampling windows, leakage-free
FRFs will be calculated using those signals.
During shaker testing, if a Burst Random or a Burst Chirp (fast swept sine) shaker
signal is used to excite the structure, leakage-free FRFs can be calculated using
those signals. A Burst Random or Burst Chirp signal is terminated prior to the end
of its sampling window so that both the force and structural response signals are
completely contained within their sampling windows.

4.12 Reduced Leakage

If one of the two leakage-free conditions cannot be met by the acquired force and
response signals, then leakage errors can be minimized in their spectra by applying
an appropriate time domain window to the sampled signal before it is transformed
using the FFT. A Hanning window is typically applied to pure (continuous) random
signals. Pure random signals are never completely contained within their sampling
windows. Using a Hanning window prior to transforming them with the FFT will
minimize leakage in their frequency spectrum.

4.13 Linear Versus Nonlinear Dynamics

Both EMA and FEA modal models are defined as solutions to a set of
linear differential equations. Using a modal model assumes that the linear
dynamic behavior of the test article can be adequately described using these
equations. However, many real-world structures may not exhibit linear dynamic
motion.
Real-world structures can have dynamic behavior ranging from linear to slightly
nonlinear to severely nonlinear. If the test article is in fact undergoing nonlinear
motion, significant errors will occur when attempting to extract modal parameters
from a set of FRFs which are based on a linear dynamic model.
20 Structural Dynamics Modification and Modal Modeling 1103

4.14 Random Excitation and Spectrum Averaging

To reduce the effects of nonlinear behavior, random excitation combined with signal
post-processing must be applied to the acquired data. The goal is to yield a set of
linear FRF estimates to represent the dynamics of the structure subject to a certain
force level.
This common method for testing a nonlinear structure is to excite it with one or
more shakers using random excitation signals. If these signals continually vary over
time, the random excitation will excite the nonlinear behavior of the structure in a
random fashion.
Each time a nonlinear signal is transformed using the FFT, the nonlinear
components of the signal will appear as random noise spread over the frequency
range of the DFT. If multiple DFTs of the response of a randomly excited structure
are averaged together, the nonlinear components (random noise) will be “averaged
out” of the average DFT, leaving only the linear resonant response peaks.

4.15 Curve Fitting FRFs

The first step of an FRF-based EMA is to calculate a set of FRFs that accurately
represent the linear dynamics of the test article over a frequency range of interest.
The second step is to curve fit the FRFs using a linear parametric model of an
FRF. The unknown parameters of the FRF model are the modal parameters of the
structure. The goal of these two steps is to obtain an accurate EMA modal model.
If the test article has a high modal density including either closely coupled modes
(two modes represented by one resonance peak) or repeated roots (two modes with
the same frequency but different mode shapes), extracting an accurate EMA modal
model from the FRFs can be challenging.
The linear parametric curve fitting model is a summation of contributions from
all modes at each frequency sample of the FRFs. This model is commonly curve fit
to the FRF data using a least-squared-error method. This broadband curve fitting
approach also assumes that all resonances of interest have been adequately excited
over the frequency span of the FRFs.
A wide variety of FRF-based curve fitting methods are commercially available
today. All FRF-based curve fitting methods assume that the FRFs adequately
represent the linear dynamics of the test article and are leakage-free.

4.16 Modal Models and SDM

SDM will give accurate results when an accurate modal model of the unmodified
structure is used. The modal model can contain EMA mode shapes, FEA mode
shapes, Hybrid mode shapes consisting of both EMA and FEA modal parameters,
or a mixture of all three types of mode shapes.
1104 M. Richardson and D. Formenti

The advantage of SDM is that with a reasonably accurate modal model of the
unmodified structure, numerous structural modifications can be quickly explored.
This could include exploring alternate boundary conditions which are difficult to
model with an FEA model.
Later in this chapter, a Hybrid modal model containing both translational and
rotational DOFs will be used with SDM to model the attachment of a RIB stiffener to
an aluminum plate. The new mode shapes calculated by SDM will then be compared
with both FEA and EMA mode shapes of the plate with the RIB attached to it.

5 Structural Dynamic Models

The dynamic behavior of a mechanical structure can be modeled either with a set
of differential equations in the time domain or with an equivalent set of algebraic
equations in the frequency domain. Once the equations of motion have been created,
they can be used to calculate mode shapes and to calculate structural responses to
static loads or dynamic forces.
The dynamic response of most structures usually includes resonance-assisted
vibration. Dynamic resonance-assisted response levels can far exceed the defor-
mation levels due to static loads. Resonance-assisted vibration is often the cause
of noisy operation, uncontrollable behavior, premature wear out of parts such as
bearings, and unexpected material failure due to cyclic fatigue.

5.1 Structural Resonances

Two or more spatial deformations assembled into a vector format are called an
Operating Deflection Shape (or ODS).

A mode of vibration is a mathematical representation of a structural resonance. An

ODS is a summation of mode shapes.

Each mode is represented by its natural frequency (its modal frequency), a

damping decay constant (the decay rate of a resonance when forces are removed
from the structure), and its spatially distributed amplitude levels (its mode shape).
These three modal properties (frequency, damping, and mode shape) provide a
complete mathematical representation of each structural resonance. A mode shape
is the contribution of a resonance to the overall deformation (the ODS) on the
surface of a structure at each location in each direction.
It is shown later that both the time and frequency domain equations of motion
can be represented solely in terms of modal parameters. This powerful conclusion
means that a set of modal parameters can be used to completely represent the linear
dynamics of a structure.
When properly scaled, a set of mode shapes is called a modal model. The
complete dynamic properties of the structure are represented by its modal model.
20 Structural Dynamics Modification and Modal Modeling 1105

SDM uses the modal model of the unmodified structure together with the FEA
elements that represent structural modifications as inputs and calculates a new modal
model for the modified structure.

5.2 Truncated Modal Model

All EMA and FEA modal models contain mode shapes for a finite number of modes.
An EMA modal model contains a finite number of mode shapes that were obtained
by curve fitting a set of FRFs that span a limited frequency range. An FEA modal
model also contains a finite number of mode shapes that are defined for a limited
range of frequencies. Therefore, both EMA and FEA modal models represent a
truncated (approximate) dynamic model of a structure.
Except for so-called lumped parameter systems (like a mass on a spring), all
real-world structures have an infinite number of resonances. But SDM still provides
usable results because of the following property.

The dynamic response of most structures is dominated by the excitation of their

low frequency modes.

When using the SDM method, all the low frequency modes should be included in
the modal model. In order to account for the higher frequency modes that have been
left out of the truncated modal model, it is also important to include several modes
above the highest frequency mode of interest in the modal model.

5.3 Sub-structuring

To solve a sub-structuring problem, where one structure is mounted on or attached

to another using FEA elements, the free-body dynamics (the six rigid-body modes)
of the structure to be mounted on the other must also be included in its modal model.
This will be illustrated by the example later in this chapter.

5.4 Rotational DOFs

Another potential source of error in using SDM is that certain modifications

require mode shapes with both translational and rotational DOFs. Normally only
translational motions are acquired experimentally, and therefore the resulting FRFs
and mode shapes only have translational DOFs. If a modal model does not contain
rotational DOFs, accurate modifications that involve torsional stiffnesses and/or
rotary inertia effects cannot be accurately modeled.
FEA mode shapes derived from rod, beam, and plate (membrane) elements
have rotational DOFs included in them. When rotational stiffness and inertia are
important, FEA mode shapes with rotational DOFs in them can be used in a Hybrid
modal model as input to SDM. Later in this chapter, SDM will be used to model
1106 M. Richardson and D. Formenti

the attachment of a RIB stiffener to a plate structure. Mode shapes with rotational
DOFs and spring elements with rotational stiffness will be used to correctly model
the joint stiffness between the RIB and the plate.

6 Time Domain Dynamic Model

Modes of vibration are defined by assuming that the dynamic behavior of a

mechanical structure or system can be adequately described by a set of time domain
differential equations. These equations are a statement of Newton’s second law
(F = Ma). They represent a force balance between the internal inertial (mass),
dissipative (damping), and restoring (stiffness) forces and the external forces acting
on the structure. This force balance is written as a set of linear differential equations:

[M] {ẍ (t)} + [C] {ẋ (t)} + [K] {x (t)} = {f (t)} (1)

where

[M] ← Mass matrix (n by n)

[C] ← Damping matrix (n by n)
[K] ← Stiffness matrix (n by n)
{ẍ (t)} ← Accelerations (n-vector)
{ẋ (t)} ← Velocities (n-vector)
{x(t)} ← Displacements (n-vector)
{f(t)} ← Externally applied forces (n-vector)

These differential equations describe the dynamics between n-discrete points and
directions or n-degrees of freedom (DOFs) of a structure. To adequately describe its
dynamic behavior, enough equations can be created involving as many DOFs as
necessary. Even though equations could be created between an infinite number of
DOFs, in a practical sense, only a finite number of DOFs is ever used, but they could
still number in the hundreds of thousands.
Notice that the damping force is proportional to velocity. This is a model for
viscous damping. Different damping models are addressed later in this chapter.

6.1 Finite Element Analysis (FEA)

Finite element analysis (FEA) is used to generate the coefficient matrices of

the time domain differential equations written above. The mass and stiffness
matrices are generated from the physical and material properties of the structure.
Material properties include the modulus of elasticity, inertia, and Poisson’s ratio (or
“squeezability”).
Damping properties are not easily modeled for real-world structures. Hence the
damping force term is usually left out of an FEA model. Even without damping,
20 Structural Dynamics Modification and Modal Modeling 1107

the mass and stiffness terms are enough to model resonant vibration; hence, the
equations of motion can be solved for modal parameters.

6.2 FEA Modes

The homogeneous form of the differential equations, where the external forces on
the right-hand side are zero, can be solved for mode shapes and their corresponding
natural frequencies. This is called an eigen-solution. Each natural frequency is an
eigenvalue, and each mode shape is an eigenvector. The analytical mode shapes are
referred to as FEA mode shapes. The transformation of the equations of motion (1)
into modal coordinates is covered later in this chapter.

7 Frequency Domain Dynamic Model

In the frequency domain, the dynamics of a mechanical structure or system are

represented by a set of linear algebraic equations, in a form called a Transfer
Function model or MIMO (Multiple Input Multiple Output) model. This model
contains Transfer Functions between all combinations of input and response DOF
pairs:

{X (s)} = [H (s)] { F (s)} (n − vector) (2)

where

s ← Laplace variable (complex frequency)

[H(s)] ← Transfer Function matrix (n by n)
{X(s)} ← Laplace transform of displacements (n-vector)
{F(s)} ← Laplace transform of externally applied forces (n-vector)

This model is also a complete description of the dynamics between n-DOFs of

a structure. Equations can be created between as many DOF pairs of the structure
as necessary to adequately describe its dynamic behavior over a frequency range
of interest. Like the time domain differential Eq. (1), these Eq. (2) are finite
dimensional.

8 Parametric Models Used for Curve Fitting

Curve fitting is a numerical process by which an analytical FRF model is matched

to experimental FRF data in a manner that minimizes the squared error between
the experimental data and the analytical curve fitting model. The purpose of curve
fitting is to estimate the unknown modal parameters of the curve fitting model. More
1108 M. Richardson and D. Formenti

precisely, the modal frequency, damping, and mode shape of each resonance in the
frequency range of the FRFs are estimated by curve fitting an analytical model to a
set of FRFs.

8.1 Rational Fraction Polynomial Model

The Transfer Function matrix in Eq. (2) can also be expressed analytically as a
ratio of two polynomials. This is called a rational fraction polynomial matrix form
of the Transfer Function matrix. Expressed in terms of m-modes, the denominator
polynomial has (2 m + 1) terms, and each numerator polynomial has (2 m) terms:

 
[b0 ] s2m−1 + [b1 ] s2m−2 + [b2 ] s2m−3 + · · · + b2m−1
[H (s)] = (n by n) (3)
a0 s2m + a1 s2m−1 + a2 s2m−2 + · · · + a2m

where

m= number of modes in the analytical curve fitting model

a0 s2m + a1 s2m – 1 + a2 s2m – 2 + . . . + a2m ← the characteristic polynomial
a0 , a1 , a2 , . . . , a2m ← real-valued coefficients
[b0 ]s2m – 1 + [b1 ]s2m − 2 + [b2 ]s2m − 3 + . . . + [b2m − 1 ] ← numerator polynomial
matrix (n by n)
[b0 ], [b1 ], [b2 ], . . . , [b2m – 1 ] ← real-valued coefficient matrices (n by n)

Each Transfer Function in the (n by n) matrix has a unique numerator polynomial

(n by n) matrix and the same denominator polynomial, called the characteristic
polynomial.

8.2 Partial Fraction Expansion Model

The Transfer Function matrix in Eq. (2) can also be expressed in partial fraction
expansion form. When expressed as shown in Eqs. (4) and (5), any Transfer Function
value at any frequency is a summation of terms, each term called the resonance
curve of a mode of vibration:
 ∗

m
[rk ] r
[H (s)] = −  k ∗ (4)
2j (s − pk ) 2j s − pk
k=1

or
   t

m
Ak {uk } {uk }t A∗k u∗k u∗k
[H (s)] = −   (5)
k=1
2j (s − pk ) 2j s − p∗k
20 Structural Dynamics Modification and Modal Modeling 1109

where
m= number of modes of vibration
[rk ] ← Residue matrix for the kth mode (n by n)
pk = − σk + jωk ← Pole location for the kth mode
σk ← Damping decay of the kth mode
ωk ← Damped natural frequency of the kth mode
{uk } ← Mode shape for the kth mode (n-vector)
Ak ← Scaling constant for the kth mode
t – denotes the transposed vector

Figure 2 shows a Transfer Function for a single resonance, plotted over half of
the s-plane.

8.3 Experimental FRFs

An FRF is defined as the values of a Transfer Function along the jω-axis in the
s-plane.

Fig. 2 Transfer Function and FRF of a single resonance

1110 M. Richardson and D. Formenti

Fig. 3 Log magnitude of an experimental FRF

An experimental FRF can be calculated from acquired experimental data if each

excitation force and all responses caused by that force are simultaneously acquired.
Figure 3 shows the magnitude and phase of a typical experimental FRF.

9 FRF-Based Curve Fitting

Curve fitting is commonly done using a least-squared-error algorithm which

minimizes the difference between an analytical FRF model and the experimental
data. The outcome of FRF-based curve fitting is a pole estimate (frequency and
damping) and a mode shape (a row or column of residue estimates in the residue
matrix) for each resonance that is represented in the experimental FRF data.
All forms of the curve fitting model, Eqs. (3), (4), and (5), are used by different
curve fitting algorithms. If the rational fraction polynomial model (3) is used, its
numerator and denominator polynomial coefficients are determined during curve
fitting. These polynomial coefficients are further processed numerically to extract
the frequency, damping, and mode shape of each resonance represented in the FRFs.
20 Structural Dynamics Modification and Modal Modeling 1111

9.1 Modal Frequency and Damping

Modal frequency and damping are calculated as the roots of the characteristic
polynomial. The denominators of all three curve fitting models (3), (4), and (5)
contain the same characteristic polynomial. Therefore, global estimates of modal
frequency and damping are normally obtained by curve fitting an entire set of FRFs.
Another property resulting from the common denominator of the FRFs is that
the resonance peak for each mode will occur at the same frequency in each
FRF. Mass loading effects can occur when the response sensors add a significant
amount of mass relative to the mass of the test structure. If the sensors are moved
from one point to another during a test, some resonance peaks will occur at a
different frequency in certain FRFs. When mass loading of this type occurs, a local
polynomial curve fitter, which estimates frequency, damping, and residue for each
mode in each FRF, will provide better results.

9.2 Modal Residue

The modal residue, or FRF numerator, is unique for each mode and each FRF.

A modal residue is the magnitude (or strength) of a mode in an FRF. A row or

column of residues in the residue matrix defines the mode shape of a mode.

The relationship between residues and mode shapes is shown in the numerators
of the two curve fitting models (4) and (5).
Figure 4 shows an analytical curve fitting function overlaid on the log magnitude
of an experimental FRF.
If the partial fraction expansion model (5) is used, the pole (frequency and
damping) and residues for each mode are explicitly determined during the curve

Fig. 4 Curve of an experimental FRF

1112 M. Richardson and D. Formenti

fitting process. To achieve more numerical stability, curve fitting can be divided into
two curve fitting steps.

1. Estimate frequency and damping (global or local estimates).

2. Estimate residues using the frequency and damping estimates.

10 Transformed Equations of Motion

Since the differential equations of motion (1) are linear, they can be transformed to
the frequency domain using the Laplace transform without loss of any information.
In the Laplace (complex frequency) domain, the equations have the form:

s2 [M] {X (s)} + s [C] {X (s)} + [K] {X (s)} = {F (s)} + {ICs} (7)

where

{ICs} ← vector of initial conditions (n-vector)

{X(s)} ← Laplace transforms of displacements (n-vector)
{F(s)} ← Laplace transforms of applied forces (n-vector)

All physical properties of the structure are preserved in the left-hand side of the
equations, while the applied forces and initial conditions {ICs} are contained on the
right-hand side. The initial conditions can be treated as a special form of the applied
forces and hence will be dropped from consideration without loss of generality in
the following development.
The equations of motion can be further simplified:

[B (s)] {X (s)} = {F (s)} (n − vector) (8)

where

[B (s)] = s2 [M] + s [C] + [K] ← system matrix (n by n) (9)

Equation (8) shows that any linear dynamic system has three basic parts:
applied forces (inputs), responses to those forces (outputs), and the dynamic system
represented by its system matrix [B(s)].

11 Dynamic Model in Modal Coordinates

The modal parameters of a structure are the solutions to the homogeneous equations
of motion. That is, when {F(s)} = {0}, the solutions to Eq. (8) are complex
valued eigenvalues and eigenvectors. The eigenvalues occur in complex conjugate
20 Structural Dynamics Modification and Modal Modeling 1113

 
pairs pk , pk∗ . The eigenvalues are the solutions (or roots) of the characteristic
polynomial, which is derived from the following determinant equation:

det [B (s)] = 0 (10)

The eigenvalues (or poles) of the system are:

pk = −σk + j ωk , k = 1, . . . m

pk∗ = −σk − j ωk , k = 1, . . . m

m = number of modes
pk = −σ k + jωk ← pole for the kth mode
pk∗ = −σ k − jωk ← conjugate pole for the kth mode
σ k ← damping of the kth mode
ωk ← damped natural frequency of the kth mode, k = 1, . . . m.

Each eigenvalue has a corresponding eigenvector,

  and hence the eigenvectors
also occur in complex conjugate pairs, {uk } , u∗k .

Each complex eigenvalue (also called a pole) contains the modal frequency and
damping. Each corresponding complex eigenvector is the mode shape.

Each eigenvector pair is a solution to the algebraic equations:

[B (pk )] {uk } = {0} k = 1, . . . m (n − vector) (11)

  ∗   ∗ 
B pk uk = {0} k = 1, . . . m (n − vector) (12)

The eigenvectors (or mode shapes) can be assembled into a matrix:

      
[U] = {u1 } , {u2 } , . . . , {um } , u∗1 , u∗2 , . . . , u∗m
(13)
← mode shape matrix (n by 2 m)

This transformation of the equations of motion means that all vibration can be
represented in terms of modal parameters.

Fundamental Law of Modal Analysis: All vibration is a summation of mode

shapes.

Using the (n by 2 m) mode shape matrix [U], the time domain response of a
structure {x(t)} is related to its response in modal coordinates {z(t)} by:
1114 M. Richardson and D. Formenti

{x (t)} = [U] {z (t)} (n − vector) (14)

Applying the Laplace transform to Eq. (14) stated gives:

{X (s)} = [U] {Z (s)}

where

{Z (s)} ← Laplace transform of displacements in modal coordinates (2m − vector)

Applying this transformation to Eq. (8) gives:

s2 [M] [U] + s [C] [U] + [K] [U] {Z (s)} = {F (s)} (n − vector) (15)

Pre-multiplying Eq. (15) by the transposed conjugate of the mode shape matrix
([U]t ) gives:

s2 [U]t [M] [U]+s[U]t [C] [U] + [U]t [K] [U] {Z (s)} = [U]t {F (s)} (2 m by 2 m)
(16)

Three new matrices can now be defined:

[m] = [U]t [M] [U] ← modal mass matrix (2 m by 2 m) (17)

[c] = [U]t [C] [U] ← modal damping matrix (2 m by 2 m) (18)

[k] = [U]t [K] [U] ← modal stiffness matrix (2 m by 2 m) (19)

The equations of motion transformed into modal coordinates now become:

s2 [m] + s [c] + [k] {Z (s)} = [U]t {F (s)} (2 m by 2 m) (20)

11.1 Damping Models

In Eq. (1), the damping of the structure was modeled with a linear viscous force
term which is proportional to surface velocity (1). This is called a non-proportional
damping model. Non-proportional damping is the most commonly used damping
model, unless there is a known physical reason for using a different damping model.
If the structure model has no damping ([C] = 0), then it can be shown that
the modal mass and stiffness matrices are diagonal matrices and the equations of
motion in modal coordinates (20) are uncoupled.
20 Structural Dynamics Modification and Modal Modeling 1115

Table 1 Damping models

Damping Mode shapes Modal matrices
None Normal Diagonal (m by m)
Non-proportional Complex Non-diagonal (2 m by 2 m)
Proportional Normal Diagonal (m by m)
Light Almost normal Almost diagonal (m by m)

If damping is modeled with a proportional damping matrix ([C] = α[M] + β[K]),

where α & β are proportionality constants, this is called a proportional damping
model. With proportional damping, the modal mass, damping, and stiffness matrices
are diagonal matrices, and the equations of motion in modal coordinates (20) are
again uncoupled.

11.2 Lightly Damped Structures

When they vibrate, all real-world structures have several damping mechanisms
which dissipate their vibration energy. On earth, the surrounding air always provides
one damping mechanism. After all excitation forces are removed, all structural
vibration will be damped out by the damping mechanisms.
A structure is assumed to be lightly damped if its damping forces are signifi-
cantly less than its internal mass (inertial) and stiffness (restoring) forces.
If a structure exhibits troublesome resonance-assisted vibration, it is usually
because it is lightly damped. A common way to define a lightly damped structure is
as follows:

A structure is called lightly damped if its modes have less than 10% of critical
damping.

If a structure is lightly damping, then it can be shown that its modal mass,
damping, and stiffness matrices in Eq. (20) are approximately diagonal matrices.
Furthermore, its mode shapes can be shown to be approximately normal (or real-
valued). In this case, the 2m-equation (20) are redundant and can be replaced to
m-equations, one corresponding to each mode.
The damping cases are summarized as follows (Table 1).

12 Scaling Mode Shapes to Unit Modal Masses

Mode shapes are called “shapes” because they are unique in shape, but not in value.
In other words, the mode shape vector {uk } for each mode (k) does not have unique
values. The “shape” of {uk } is unique, but its shape values are arbitrary.
1116 M. Richardson and D. Formenti

Another way of saying this is that the ratio of any two mode shape components
is unique. A mode shape is also called an eigenvector, meaning that its “shape” is
unique, but its values are arbitrary. Therefore, a mode shape can be arbitrarily scaled
using any scale factor.

Curve fitting a set of un-calibrated FRFs will yield un-scaled mode shapes; hence,
they are not a modal model and cannot be used with SDM.

12.1 Modal Mass Matrix

SDM requires a modal model to describe the dynamics of the unmodified structure.
In order to accurately model the structural dynamics, the mode shapes of the modal
model must be scaled to preserve the mass, stiffness, and damping properties of the
structure. SDM requires mode shapes which are scaled so that the modal masses are
one or unity. These are called UMM mode shapes.
When the mass matrix is post-multiplied by the mode shape matrix and pre-
multiplied by its transpose, the result is the diagonal matrix shown in Eq. (21). This
is a definition of modal mass:
⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
[U]t [M] [U] = ⎢
⎣ m ⎥=⎢
⎦ ⎣
1
Aω
⎥
⎦ (21)
.. ..
. .

where

[M] ← mass matrix (n by n)

[U] = [{u1 }, {u2 }, . . . , {um }] ← mode shape matrix (n by m)
⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
⎢ m ⎥=⎢ 1 ⎥ ← modal mass matrix (m by m)
⎣ ⎦ ⎣ Aω ⎦
.. ..
. .

The modal mass of each mode (k) is a diagonal element of the modal mass
matrix:

1
mk = ← modal mass k = 1, . . . , m (22)
Ak ωk

pk = −σk + jωk ← pole location for the kth mode

ωk ← damped natural frequency of the kth mode
Ak ← scaling constant for the kth mode
20 Structural Dynamics Modification and Modal Modeling 1117

12.2 Modal Stiffness Matrix

When the stiffness matrix is post-multiplied by the mode shape matrix and pre-
multiplied by its transpose, the result is a diagonal matrix, shown in Eq. (23). This
is a definition of modal stiffness:
⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
[U]t [K] [U] = ⎢
⎣ k ⎥=⎢
⎦ ⎣
σ2 +ω2
Aω
⎥
⎦ (23)
.. ..
. .

where

[K] ← stiffness matrix. (n by n)

⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
⎢ k ⎥=⎢ σ2 +ω2 ⎥ = modal stiffness matrix (m by m)
⎣ ⎦ ⎣ Aω ⎦
.. ..
. .

The modal stiffness of each mode (k) is a diagonal element of the modal stiffness
matrix:

σ2k + ω2k
kk = ← modal stiffness k = 1, . . . , m (24)
Ak ωk

where

σk ← modal damping of the kth mode

12.3 Modal Damping Matrix

When the damping matrix is post-multiplied by the mode shape matrix and pre-
multiplied by its transpose, the result is a diagonal matrix, shown in Eq. (25). This
is a definition of modal damping:
⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
[U]t [C] [U] = ⎢
⎣ c ⎥=⎢
⎦ ⎣
2σ
Aω
⎥
⎦ (25)
.. ..
. .

where

[C] ← damping matrix (n by n)

1118 M. Richardson and D. Formenti

⎡ ⎤ ⎡ ⎤
.. ..
. .
⎢ ⎥ ⎢ ⎥
⎢ c ⎥=⎢ 2σ ⎥ = modal damping matrix (m by m)
⎣ ⎦ ⎣ Aω ⎦
.. ..
. .

The modal damping of each mode (k) is a diagonal element of the modal
damping matrix:

2σk
ck = ← modal damping k = 1, . . . , m (26)
Ak ωk

12.4 Unit Modal Masses

Each of the modal mass, stiffness, and damping matrix diagonal elements (22),
(24), and (26) includes a scaling constant (Ak ). This constant is necessary
because the mode shapes are not unique in value and therefore can be arbitrarily
scaled.
One of the common ways to scale mode shapes is to scale them so that
the modal masses are “one” or “unity.” Normally, if a mass matrix [M] were
available, the mode vectors would simply be scaled such that when the triple product
[U]t [M][U] was formed, the resulting modal mass matrix would be an identity
matrix.

13 SDM Dynamic Model

The local eigenvalue modification process used by SDM requires a modal

model of the unmodified structure. The modal model consists of the modal
frequency, modal damping (optional), and mode shape of each mode in the
model.
The dynamic model for the unmodified structure was given in Eq. (1). Similarly,
the dynamic model for the modified structure is written:

[M + M] {ẍ (t)} + [C + C] {ẋ (t)} + [K + K] {x (t)} = {f (t)} (27)

where

[M] ← matrix of mass modifications (n by n)

[C] ← matrix of damping modifications (n by n)
[K] ← matrix of stiffness modifications (n by n)
20 Structural Dynamics Modification and Modal Modeling 1119

14 SDM Equations Using UMM Mode Shapes

Unit Modal Mass (UMM ) scaling is normally done on FEA mode shapes because
the mass matrix is available for scaling them. However, when EMA mode shapes
are extracted from experimental FRFs, no mass matrix is available for scaling the
mode shapes to yield Unit Modal Masses.
The mode shapes are eigenvectors and hence have no unique values, but if they
are scaled so that the modal mass matrix is an identity matrix, the equations of
motion in modal coordinates (20) become:

s2 [I] + s [2σ] + 2 {Z (s)} = [U]t {F (s)} (m − vector) (28)

where

[I] ← identity modal mass matrix (m by m)

[2σ] ← diagonal modal damping matrix (m by m)
[2 ] ← diagonal modal frequency matrix (m by m)
[2 ] = [σ2 + ω2 ]

In Eq. (28), the complete dynamics of the unmodified structure is represented by

modal frequencies, modal damping, and mode shapes that are scaled to unit modal
masses. All mass, stiffness, and damping properties of the unmodified structure are
preserved in the modal model that consists of UMM mode shapes.
Using the UMM mode shapes, the equations of motion (27) for the modified
structure can be transformed to modal coordinates:

s2 [m] + s [c] + [k] {Z (s)} = [U]t {F (s)} (m − vector) (29)

where

[m] = [I] + [U]t [M] [U] (m by m) (30)

[c] = [2σ] + [U]t [C] [U] (m by m) (31)

[k] = 2 + [U]t [K] [U] (m by m) (32)

For a lightly damped structure, the mode shapes are almost real-valued so the
mode shape matrix has dimension (n by m).

The homogeneous form of Eq. (29) is solved by the SDM method to find the modal
properties of the modified structure.
1120 M. Richardson and D. Formenti

Using the approach of Hallquist et al. [2], an additional transformation of the

modification matrices [M],[C], and [K] is made which results in a reformu-
lation of the eigenvalue problem in modification space. For a single modification,
this problem becomes a scalar eigenvalue problem, which can be solved quickly
and efficiently. The drawback to making one modification at a time, however, is that
if many modifications are required, computation time can become significant and
errors will accumulate.
A more practical approach is to solve the homogeneous form of Eq. (29) directly.
This is still a relatively small (m by m) eigenvalue problem which can include as
many structural modifications as desired, but only needs to be solved once.
Equations (30) to (32) also indicate another advantage of SDM:

Only the mode shape components where the modification elements are attached to
the structure model are required.

This means that only mode shape data for those DOFs where the modification
elements are attached to the structure is necessary for SDM.

15 Scaling Residues to UMM Mode Shapes

Without a mass matrix, EMA mode shapes can be scaled to Unit Modal Masses by
using the relationship between residues and mode shapes.

Residues are related to mode shapes by equating the numerators of curve fitting
models (4) and (5):

[r (k)] = Ak {uk } {uk }t k = 1, . . . , m (33)

where

[r(k)] ← residue matrix for the mode (k) (n by n)

Residues are the numerators of the Transfer Function matrix in Eq. (4) when it is
written in partial fraction form. For convenience, Eq. (4) is rewritten here:


m
[r (k)] [r (k)]∗
[H (s)] = −   (34)
2j (s − pk ) 2j s − p∗k
k=1

* – denotes the complex conjugate

Residues have engineering units associated with them and hence have unique
values. FRFs have units of (motion/force), and the FRF denominators have units
20 Structural Dynamics Modification and Modal Modeling 1121

of Hz or (radians/second). Therefore, the residues in the numerator have units of

(motion/force-seconds).
Equation (34) can be written for the jth column (or row) of the residue matrix
and for mode (k) as:

⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ r1j (k) ⎪
⎪ ⎪
⎪ u1k ujk ⎪
⎪ ⎪
⎪ u1k ⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎪
⎪ r2j (k) ⎪
⎪
⎪
⎪
⎪
⎪ u2k ujk ⎪
⎪
⎪
⎪
⎪
⎪ u2k ⎪
⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎨ · ⎬ ⎪ ⎪
⎨ · ⎪
⎬ ⎪
⎨ · ⎪ ⎬
· = Ak · = Ak ujk · k = 1, . . . , m
⎪
⎪ ⎪
⎪ ⎪
⎪  2 ⎪ ⎪ ⎪
⎪ ⎪
⎪ (35)
⎪
⎪ r (k) ⎪
⎪ ⎪
⎪ jk ⎪
u ⎪ ⎪
⎪ u ⎪
⎪
⎪
⎪
jj ⎪ ⎪ ⎪ ⎪ jk ⎪
⎪
⎪
⎪ · ⎪ ⎪
⎪
⎪
⎪
⎪
⎪
⎪ · ⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ · ⎪⎪
⎪
⎪
⎩ ⎭ ⎩ ⎭ ⎩ ⎭
rnj (k) unk unk unk
Unique Variable Variable

This relationship states that residues have unique values and preserve the physical
properties of the structure, but mode shapes do not have unique values and
therefore can be scaled in any manner.

The scaling constant Ak must be chosen so that Eq. (35) remains valid. Either
the value of Ak can be chosen first and the mode shapes scaled accordingly, or the
mode shapes can be scaled first and Ak calculated so that Eq. (35) is still satisfied.
To obtain UMM mode shapes, simply set the modal mass equal to one and solve
Eq. (22) for Ak :

1
Ak = k = 1, . . . , m (36)
ωk

15.1 Driving Point FRF Measurement

Mode shapes scaled to Unit Modal Mass (UMM mode shapes) are obtained from
the jth column (or row) of the residue matrix by substituting Eq. (36) into Eq. (35):
⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ u1k ⎪
⎪ ⎪
⎪ r1j (k) ⎪
⎪ ⎪
⎪ r1j (k) ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ u ⎪
⎪ ⎪
⎪ r (k) ⎪
⎪ ⎪
⎪ r (k) ⎪
⎪
⎪
⎨
2k ⎪
⎬ ⎪
⎨
2j ⎪
⎬  ω ⎪
⎨
2j ⎪
⎬
· 1 · k ·
= = k = 1, . . . , m (37)
⎪
⎪ · ⎪
⎪ A ujk ⎪
⎪ · ⎪
⎪ r (k) ⎪
⎪ · ⎪ ⎪
⎪
⎪ ⎪
⎪
k
⎪
⎪ ⎪
⎪
jj
⎪
⎪ ⎪
⎪
⎪
⎪ · ⎪
⎪ ⎪ · ⎪
⎪ ⎪ ⎪
⎪ · ⎪ ⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
unk rnj (k) rnj (k)
1122 M. Richardson and D. Formenti

15.2 UMM Mode Shape

The driving point residue rjj (k) (where row index j equals column index j) plays an
important role in this scaling process. The driving point residue for each mode (k)
is required in Eq. (37) for scaling mode shapes to UMM. Driving point residues are
obtained by curve fitting a driving point FRF.

A drive point FRF is any measurement where the excitation force DOF is the same
as the response DOF.

15.3 Triangular FRF Measurements

In some cases, it is difficult or even impossible to make a good driving point FRF
measurement. In those cases, an alternative set of measurements can be made from
which to create UMM mode shapes. From Eq. (37) the following equation can be
written:

rjp (k) rjq (k)
ujk = k = 1, . . . , m (38)
Ak rpq (k)

Equation (38) can be substituted for ujk in Eq. (37) to calculate UMM mode
shapes. To calculate a starting component ujk , three FRFs are required (FRFjp ,
FRFjq , FRFpq ). DOF(j) is the (fixed) reference for the jth column (or row) of
FRF measurements, so the two measurements FRFjp and FRFjp would normally be
made. In addition, one extra measurement FRFpq is also required in order to obtain
the three residues required by Eq. (38). Since the FRFs (FRFjp , FRFjq , FRFpq )
form a triangle of off-diagonal FRFs in the FRF matrix, they are called a triangular
FRF measurement. Eq. (38) leads to the following conclusion:

A set of triangular FRF measurements which do not include driving point FRFs
can be curve fit and their residues used to create UMM mode shapes

16 Integrating Residues to Displacement Units

Vibration measurements are commonly made using either accelerometers that

measure acceleration responses or vibrometers that measure velocity responses.
Excitation forces are typically measured with a load cell. Therefore, FRFs cal-
culated for experimental data will have units of either (acceleration/force) or
(velocity/force)
20 Structural Dynamics Modification and Modal Modeling 1123

Modal residues always have the units of the FRF multiplied by (radians/second).

• Residues extracted from FRFs with units of (acceleration/force) will have units
of (acceleration/force-seconds).
• Residues extracted from FRFs with units of (velocity/force) will have units of
(velocity/force-seconds).
• Residues extracted from FRFs with units of (displacement/force) will have units
of (displacement/force-seconds).

Since the modal mass, stiffness, and damping Eqs. (21), (23), and (25) assume
units of (displacement/force), residues with units of (acceleration/force-seconds)
or (velocity/force-seconds) must be “integrated” to units of (displacement/force-
seconds) before scaling them to UMM mode shapes.
Integration of a time domain function has an equivalent operation in the
frequency domain. Integration of a Transfer Function is done by dividing it by the
Laplace variable(s):

[Hv (s)] [Ha (s)]

[Hd (s)] = = (39)
s s2

where

[Hd (s)]= transfer matrix in (displacement/force) units

[Hv (s)]= transfer matrix in (velocity/force) units
[Ha (s)]= transfer matrix in (acceleration/force) units

Since residues are the result of the partial fraction expansion form of an FRF,
residues can be “integrated” directly using the formula:

[rv (k)] [ra (k)]

[rd (k)] = = k = 1, . . . , m (40)
pk (pk )2

where

[rd (k)]← residue matrix in (displacement/force) units

[rv (k)] ← residue matrix in (velocity/force) units
[ra (k)] ← residue matrix in (acceleration/force) units
pk = −σk + jωk ← pole location for the kth mode

If light damping is assumed and the mode shapes are real-valued, Eq. (40) can
be simplified to

[rd (k)] = Fk [rv (k)] = (Fk )2 [ra (k)] (41)

1124 M. Richardson and D. Formenti

Table 2 Residue scale factors

To change Transfer Function units Multiply residues by
From To
ACCELERATION DISPLACEMENT
FORCE FORCE F2
VELOCITY DISPLACEMENT
FORCE FORCE F

where
ωk
Fk ∼
= 2  k = 1, . . . , m (42)
σk + ω2k

Equations (41) and (42) are summarized in the following Table 2 where F ∼
=
ω
(seconds)
(σ +ω )
2 2

17 Effective Mass

A useful question to ask is:

“At one of its DOFs, what is the effective mass of a structure at one of its resonant
frequencies?”

Another way to ask the question is:

“At one of its DOFs, if a structure were treated like an SDOF mass-spring-damper what is
its effective mass , effective stiffness & effective damping?”

The answer to those questions follows from a further use of the modal mass,
stiffness, and damping Eqs. (21), (23), and (25) and the definition UMM mode
shapes.
It has already been shown that residues with units of (displacement/force-
seconds) can be scaled to UMM mode shapes. One further assumption is necessary
to define the effective mass at a DOF.

18 Diagonal Mass Matrix

If the mass matrix [M] is assumed to be a diagonal matrix, then pre-multiplying and
post-multiplying it by UMM mode shapes changes Eq. (21) to:


n
 2
massj ujk = 1 k = 1, . . . , m (43)
j=1
20 Structural Dynamics Modification and Modal Modeling 1125

where

massj ← jth diagonal element of the mass matrix

ujk ← jth component of the UMM mode shape (k)

If the structure is viewed as a mass-spring-damper at DOF(j), the effective mass

at the frequency of mode (k) at DOF(j) is determined from Eq. (43) as:

1
effective massj =  2 j = 1, . . . , n (44)
ujk

If each DOF(j) is treated a driving point, Eq. (37) can be used to write the mode
shape component ujk in terms of the modal frequency ωk and driving point residue
rjj (k):

ujk = ωk rjj (k) j = 1, . . . , n (45)

Substituting Eq. (45) into Eq. (44) gives another expression for the effective mass
at the frequency of mode (k):

1
effective massj = j = 1, . . . , n (46)
ωk rjj (k)

18.1 Checking the Engineering Units

Assuming that the driving point residue rjj (k) has units of (displacement/force-
seconds) as discussed earlier, and the modal frequency ωk has units of
(radians/second), then the effective mass would have units of (force-sec2 /
displacement), which are units of mass.
Using the effective mass, the effective stiffness and damping of the structure
can be calculated using Eqs. (29) and (31).

19 Effective Mass Example

Suppose that we have the following data for a single mode of vibration:

Frequency = 10.0 Hz.

Damping = 1.0%
1126 M. Richardson and D. Formenti

⎧ ⎫
⎨ −0.1 ⎬
Residue vector = + 2.0
⎩ ⎭
+ 0.5

Also, suppose that the FRF measurements that were curve fit to obtain this data
have units of (Gs/Lbf). Also assume that the driving point is at the second DOF of
the residue vector, and therefore driving point residue is equal to 2.0.
Converting the frequency and damping into units of radians/second:

Frequency = 62.83 Rad/Sec

Damping = 0.628 Rad/Sec

The residues always carry the units of the FRF measurement multiplied by
(radians/second). For this example, the units of the residues are:
 
Residue units → Gs/ (Lbf − Sec) → 386.4 Inches/ Lbf − Sec3

In these units, the residues become:

⎧ ⎫
⎨ −38.64 ⎬  
Residue vector = + 772.8 Inches/ Lbf − Sec3
⎩ ⎭
+ 193.2

Since the modal mass, stiffness, and damping Eqs. (21), (23), and (25) assume
units of (displacement/force), the above residues with units of (acceleration/force)
must be converted to (displacement/force) units. This is done by using the
appropriate scale factor from Table 2. For this example:
 2  
1
F ∼
2
= = 0.000253 Seconds2
62.83

Multiplying the residues by F2 gives:

⎧ ⎫
⎨ −0.00977 ⎬
Residue Vector = + 0.1955 Inches/ (Lbf − Sec)
⎩ ⎭
+ 0.0488

Using Eq. (37) the residue mode shape must be multiplied by the following
scale factor to obtain a UMM mode shape:
 
ω 62.83
SF = = = 17.927
rjj +0.1955
20 Structural Dynamics Modification and Modal Modeling 1127

Therefore,
⎧ ⎫
⎨ −0.175 ⎬
UMM Mode Shape = + 3.505 Inches/ (Lbf − Sec)
⎩ ⎭
+ 0.875

Using Eq. (44), the effective mass at the driving point is:

1 1
effective mass = 2
= = 0.0814Lbf − sec2 /in.
(u2 ) (3.505)2

Or, using Eq. (46), the effective mass at the driving point is:

1 1
effective mass = = = 0.0814Lbf − sec2 /in.
ω r22 (62.83) (0.1955)

20 SDM Example

In this example, SDM will be used to model the attachment of a RIB stiffener to an
aluminum plate. The new mode shapes obtained from SDM will be compared with
the FEA mode shapes of the plate with the RIB attached and with the EMA mode
shapes obtained from an impact test of the actual plate with the RIB attached. Mode
shapes will be compared in three cases.

1. EMA versus FEA mode shapes of the plate without the RIB
2. SDM versus FEA mode shapes of the plate with the RIB attached
3. SDM versus EMA mode shapes of the plate with the RIB attached

The plate and RIB are shown in Fig. 5. The dimensions of the plate are 20 in
(508 mm) by 25 in (635 mm) by 3/8 in (9.525 mm) thick. The dimensions of the
RIB are 3 in (76.2 mm) by 25 in (635 mm) by 3/8 inches (9.525 mm) thick.
Two roving impact modal tests were conducted on the plate, one before and one
after the RIB stiffener was attached to the plate. FRFs were calculated from the
impact force and the acceleration response only in the vertical (Z-axis) direction.

20.1 Cap Screw Stiffnesses

The RIB stiffener was attached to the plate with five cap screws, shown in Fig. 5b.
When the RIB is attached to the plate, translational and torsional forces are applied
between the two substructures along the length of the plate centerline where they are
attached together. Both translational and torsional stiffness forces must be modeled
in order to represent the real-world plate with the RIB stiffener attached.
1128 M. Richardson and D. Formenti

Fig. 5 (a) Aluminum plate. (b) RIB and cap screws. (c) Plate and RIB attached

The joint stiffness was modeled using six-DOF springs located at the five cap
screw locations, as shown in Fig. 6. Each six-DOF FEA spring model contains three
translational DOFs and three rotational DOFs. The six-DOF FEA springs were
given large stiffness values to model a tight fastening of RIB to the plate using the
cap screws.

• Translational stiffness: 1 × E6 lbs/in (1.75E+05 N/mm)

• Torsional stiffness: 1 × E6 in-lbs/degree (1.75E+05 mm-N/degree)

21 EMA Mode Shapes of the Plate

FRFs were calculated from data acquired while impacting the plate in the vertical
direction, at each of the 30 points shown in Fig. 7. The plate was supported on bubble
wrap laying on top of a table as shown in Fig. 5. A fixed reference accelerometer
was attached to the plate. (The location of the reference accelerometer is arbitrary.)
The EMA modal parameters were estimated by curve fitting the 30 FRFs
calculated from the roving impact test data. EMA mode shapes for 14 modes were
obtained by curve fitting the FRFs, each mode shape having 30 DOFs (1Z through
30Z). A curve fit on one of the FRFs is shown in Fig. 4.
20 Structural Dynamics Modification and Modal Modeling 1129

Fig. 6 FEA springs modeling the cap screws

Fig. 7 Impact test points

1130 M. Richardson and D. Formenti

22 FEA Mode Shapes of the Plate

An FEA model of the plate was constructed using 80 FEA plate (membrane)
elements. The following properties of the aluminum material in the plate were used:

1. Young’s modulus of elasticity: 1E7 lbf/inˆ2 (6.895E4 N/mmˆ2)

2. Density: 0.101 lbm/inˆ3 (2.796E-6 kg/mmˆ3)
3. Poisson’s ratio: 0.33
4. Plate thickness: 0.375 in (9.525 mm)

The FEA model shown in Fig. 9 has 99 points (or nodes). The eigen-solution
included the first 20 FEA modes, 6 rigid-body mode shapes and 14 flexible-body
mode shapes. Each FEA mode shape has 594 DOFs (3 translational and 3 rotational
DOFs at each point). The FEA mode shapes were scaled to UMM mode shapes;
hence, they constitute a modal model of the plate (Fig. 8).

22.1 Mode Shape Comparison

The Modal Assurance Criterion (MAC) values between each EMA mode shape and
each flexible-body FEA mode shape are displayed in the bar chart in Fig. 9.
MAC is a measure of the co-linearity of two mode shapes. The following rule of
thumb is commonly used with MAC

• MAC = 1.00 → two shapes are the same (they lie on the same straight line)
• MAC > = 0.90 → two shapes are similar
• MAC < = 0.90 → two shapes are different

The diagonal MAC bars in Fig. 9 indicate that each flexible-body EMA mode
shape closely matched one-for-one with the each flexible-body FEA mode shape.
The worst-case pair of matching mode shapes is the first pair with MAC = 0.97.

Fig. 8 Curve of an experimental FRF

20 Structural Dynamics Modification and Modal Modeling 1131

Fig. 9 FEA model with FEA Quads

22.2 Modal Frequency Comparison

The modal frequencies of the matching FEA and EMA mode pairs are listed
in Table 3. Each EMA modal frequency is higher than the frequency of its
corresponding FEA mode. The pair with the highest difference is different by
100 Hz.
The frequency differences indicate that the stiffness of the actual aluminum plate
is greater than the stiffness of the FEA model. These frequency differences could
be reduced by increasing the modulus of elasticity or increasing the thickness of the
FEA plate elements. However, since the EMA and FEA modes shapes are closely
matched, the EMA frequency and damping can be combined with the FEA mode
shapes to provide a more accurate modal model of the plate.

22.3 Hybrid Modal Model

In most cases, EMA mode shapes will not have as many DOFs in them as FEA
mode shapes. But in most all cases, EMA mode shapes will have more accurate
modal frequencies than FEA mode shapes. Also, EMA mode shapes always have
non-zero modal damping, whereas FEA mode shapes typically have no damping.
1132 M. Richardson and D. Formenti

Table 3 FEA versus EMA modes – plate without RIB

FEA frequency EMA
Shape pair (Hz) frequency (Hz) EMA damping (Hz) MAC
1 91.4 102 0.031 0.968
2 115 129 0.250 0.991
3 190 208 0.458 0.990
4 217 242 0.107 0.993
5 251 284 0.106 0.984
6 332 367 0.642 0.985
7 412 469 0.159 0.975
8 424 477 0.339 0.985
9 496 567 3.130 0.991
10 564 643 0.936 0.991
11 626 714 3.680 0.984
12 654 742 0.923 0.987
13 689 802 0.443 0.983
14 757 859 3.090 0.984

If a pair of EMA and FEA mode shapes is highly correlated (their MAC value is
close to 1.0), a Hybrid mode shape can be created by replacing the frequency and
damping of each FEA mode shape with the frequency and damping of its closely
matching EMA mode shape.
In a Hybrid mode shape, the frequency and damping of each FEA mode shape
is replaced with the frequency and damping of its closely matching EMA mode
shape.

In Fig. 10 and Table 3, each FEA mode shape has a high MAC value with a
corresponding EMA mode shape. Therefore, a Hybrid modal model of the plate can
be created by replacing the modal frequency and damping of each FEA mode shape
with the modal frequency and damping of its closely matching EMA mode shape.
A Hybrid modal model has several advantages for modeling the dynamics of an
unmodified structure with SDM

• Its modal frequencies and damping are more realistic.

• It can have DOFs at locations where EMA data was not acquired.
• Its mode shapes can include rotational DOFs which are not typically included in
EMA mode shapes.
• FEA mode shapes are typically scaled to UMM mode shapes.

22.4 RIB FEA Model

An FEA model of the RIB in a free-free condition (no fixed boundaries) was created
using 30 FEA quad plate elements. The FEA RIB model is shown in Fig. 11.
20 Structural Dynamics Modification and Modal Modeling 1133

Fig. 10 MAC between FEA

and EMA mode shapes

The frequencies of the first 16 FEA modes of the RIB are listed in Table 4.
Because it has free-free boundary conditions, the first six modes of the FEA model
are rigid-body mode shapes with zero “0” frequency. These FEA mode shapes are
UMM mode shapes, so they constitute a modal model of the RIB.

22.5 RIB Impact Test

The RIB was impact tested to obtain its EMA modal frequencies and damping, but
not its mode shapes. The RIB was only impacted once, and the resulting FRF was
curve fit to obtain its EMA modal frequencies and damping. The curve fit of the FRF
measurement is shown in Fig. 12, and the resulting EMA frequencies and damping
are listed in Table 4.

22.6 Hybrid Modal Model of the RIB

We have already seen that pairs of the EMA and FEA mode shapes of the plate
are strongly correlated based upon their high MAC values. The only significant
difference between the EMA and FEA mode shapes was their modal frequencies,
and each EMA mode also has modal damping, while the FEA mode shapes do not.
1134 M. Richardson and D. Formenti

Fig. 11 FEA RIB model

Table 4 FEA versus EMA FEA EMA EMA

RIB frequencies frequency frequency damping
Shape pair (Hz) (Hz) (Hz)
1 0.0
2 0.0
3 0.0
4 0.0
5 0.0
6 0.0
7 117.0 121.0 0.78
8 315.0 330.0 0.72
9 521.0 582.0 0.89
10 607.0 646.0 2.49
11 987.0 1.07E+03 3.86
12 1.07E+03 1.18E+03 1.24
13 1.45E+03 1.60E+03 8.72
14 1.67E+03 1.79E+03 2.55
15 1.99E+03 2.24E+03 3.92
16 2.32E+03 2.44E+03 2.97
20 Structural Dynamics Modification and Modal Modeling 1135

Fig. 12 Curve fit of a RIB FRF

Before it is attached to the plate, the RIB is a free body in space. It is

essential that the rigid-body modes of the RIB be included in its modal model
to correctly model its free-body dynamics. Rigid-body modes are typically not
measured experimentally, but they are included in the FEA mode shapes.
A RIB hybrid modal model was created by replacing the frequency and
damping of each FEA mode shape with the EMA modal frequency and damping
from its closely matching EMA mode shape. The six rigid-body FEA mode shapes
were also retained in the hybrid modal model to define the free-body dynamics of
the RIB.

23 Substructure Modal Model

In order to model the RIB attached to the plate using SDM, the Hybrid modal
model of the RIB was added to the Hybrid modal model of the plate to create a
modal model for the entire unmodified structure. This is called a substructure modal
model.
Figure 13 shows how the points on the RIB are numbered differently than the
points on the plate. This ensures that the DOFs of the RIB modes are uniquely
numbered compared to the DOFs of the plate modes.

23.1 Block Diagonal Format

When the modal model of the RIB is added to the modal model of the plate, the
unique numbering of the points on the plate and RIB creates a modal model in block
diagonal format. In block diagonal format, the DOFs of the RIB mode shapes are
zero valued for DOFs on the plate, and likewise the DOFs of the plate mode shapes
are zero valued for the DOFs of the RIB.
1136 M. Richardson and D. Formenti

Fig. 13 Point numbers of the plate and RIB

The plate modal model contains 14 modes with 594 DOFs (297 translational and
297 rotational DOFs) in each mode shape. The RIB modal model contains 16 modes
with 264 DOFs (132 translational and 132 rotational DOFs) in each mode shape.
Therefore, the substructure modal model contains 30 modes and 858 DOFs (429
translational and 429 rotational DOFs) in each mode shape.

24 Calculating New Modes with SDM

The five FEA springs shown in Fig. 13 were used by SDM to model the five cap
screws that attach the RIB to the plate. These springs were used together with the
substructure modal model for the unmodified structure as inputs to SDM.
Even though the mode shapes in the substructure modal model have 858 DOFs
in them, only the mode shape DOFs at the attachment points of the FEA springs
are used by SDM to calculate the new frequencies and damping of the plate with
the RIB attached. Following that, all 858 DOFs of the unmodified mode shapes are
used to calculate the new mode shapes of the modified structure.

25 SDM Versus FEA Modes: Plate and RIB

An FEA model consisting of the 80 quad plate elements of the plate, 30 quad plate
elements of the RIB, and the 5 springs was also solved using an FEA eigen-solver.
The SDM and FEA results are compared in Table 5.
20 Structural Dynamics Modification and Modal Modeling 1137

Table 5 SDM modes versus FEA modes – plate with RIB

FEA frequency SDM frequency SDM damping
Shape pair (Hz) (Hz) (Hz) MAC
1 96.0 108.2 0.035 1.00
2 170.5 187.6 0.369 0.99
3 222.6 253.3 0.118 0.98
4 232.7 311.5 0.293 0.92
5 245.1 351.7 0.104 0.98
6 415.0 479.2 0.171 0.98
7 423.0 521.3 0.713 0.91
8 459.1 537.4 2.770 0.95
9 530.7 619.1 0.863 0.91
10 596.0 1412.0 3.185 0.63

The first nine pairs of mode shapes in Table 5 have MAC values greater than
0.90, indicating a strong correlation between those SDM and FEA mode shapes.
The FEA modal frequencies are lower than the SDM frequencies for those first nine
mode shape pairs.
It will be shown later in FEA Model Updating how SDM can be used to find
more realistic material properties for the FEA model so that its modal frequencies
more closely match the EMA frequencies.

25.1 SDM Mode Shapes

Figure 14 is a display of the first ten SDM mode shapes. Five of the ten mode shapes
clearly reflect the torsional coupling between the RIB and the plate. All ten mode
shapes show the intended effect of the RIB stiffener on the plate.

All bending of the plate along its centerline has been eliminated by attaching the
RIB to it.

Both the RIB and plate are flexing together in unison, both being influenced
by the torsional stiffness created by the 6-DOF springs that modeled the cap
screws.
Attaching the RIB to the plate has created new modes with mode shapes
that did not exist before the modification. This confirms the law stated
earlier.

Fundamental Law of Modal Analysis: All vibration is a summation of mode

shapes
1138 M. Richardson and D. Formenti

SDM Mode Shape 1 SDM Mode Shape 2

SDM Mode Shape 3 SDM Mode Shape 4

SDM Mode Shape 5 SDM Mode Shape 6

SDM Mode Shape 7 SDM Mode Shape 8

Fig. 14 SDM mode shapes

20 Structural Dynamics Modification and Modal Modeling 1139

SDM Mode Shape 9 SDM Mode Shape 10

Fig. 14 continued

Fig. 15 Impact points on plate with RIB

26 SDM Versus EMA Modes: Plate and RIB

To compare the SDM mode shapes with EMA mode shapes, the plate with the RIB
attached was impact tested using a roving impact hammer. The plate was impacted
at 24 points on the plate in the (vertical) Z-direction, as shown in Fig. 15. This
provided enough EMA mode shape data for comparison with the SDM mode shapes.

A driving point FRF measurement is not required since the EMA mode shapes do
not require UMM scaling to compare them with SDM mode shapes using MAC.
1140 M. Richardson and D. Formenti

Fig. 16 Curve fit of an FRF from the plate with RIB

The curve fit of a typical FRF from the impact test is shown in Fig. 16. The 24
FRFs were curve fit to extract the EMA mode shapes for the modified plate.

27 Conclusions

In Table 6, the modal frequencies of the first three SDM modes agree closely with
the frequencies of the first three EMA modes. In Table 6, the first eight SDM mode
shapes agree closely with the first eight EMA mode shapes, all having MAC values
close to 1.0.
The close agreement between the first eight mode shapes from all three cases,
SDM, FEA, and EMA, verifies that the joint stiffness provided by the five cap
screws was correctly modeled in SDM using 6-DOF springs and mode shapes with
rotational DOFs in them. This example has demonstrated that even with the use
of a truncated modal model containing relatively few mode shapes, SDM provides
realistic and useful results.
Several options could be explored to obtain closer agreement between the SDM,
FEA, and EMA mode shapes:

1. Add more FEA springs between the RIB and the plate to model the stiffness
forces between the two substructures.
20 Structural Dynamics Modification and Modal Modeling 1141

Table 6 EMA versus SDM modes for the plate and RIB
EMA EMA damping SDM SDM
Shape pair frequency (Hz) (Hz) frequency (Hz) damping (Hz) MAC
1 103.8 0.142 108.2 0.034 0.99
2 188.5 0.377 187.6 0.369 0.99
3 242.5 0.254 253.3 0.118 0.99
4 277.8 0.941 311.5 0.293 0.98
5 259.8 0.254 351.7 0.104 0.97
6 468.6 0.710 479.2 0.171 0.98
7 504.1 6.202 521.3 0.713 0.97
8 572.5 1.877 537.4 2.770 0.97
9 620.3 0.818 619.1 0.865 0.85
10 803.3 6.070 801.1 0.544 0.86

2. Use more FEA quad plate elements for the plate and RIB. Increasing the mesh of
nodes for the plate elements usually provides more accurate FEA mode shapes.
3. Include more modes in the modal model of the unmodified plate and RIB. Extra
modes will provide a more complete dynamic model of the two substructures as
input to SDM.

28 Modeling a Tuned Vibration Absorber with SDM

Another use of SDM is to model the addition of tuned mass-spring-damper vibration

absorbers to a structure. A tuned vibration absorber is designed to absorb some of
the vibration energy in the structure so that one of its modes of vibration will absorb
less energy and hence the structure will vibrate with less overall amplitude.

A tuned absorber is used to suppress resonant vibration in a structure. The primary

effect of adding a tuned absorber is to replace one of its resonances with two lower
amplitude resonances.

The mass and stiffness of the tuned absorber are chosen so that its natural
frequency is “close to” the resonant frequency of a structural resonance to be
suppressed. Ideally, the absorber should be attached to the structure at a point and
in a direction where the magnitude of the resonance is large, near an anti-node of
its mode shape. The absorber will have no effect if attached at a node of the mode
shape, where its magnitude is zero.
SDM models the attachment of a tuned absorber to a structure by solving a sub-
structuring problem like the one in the previous plate and RIB example. A tuned
absorber is modeled by attaching an FEA mass to the structure using an FEA spring
and FEA damper. SDM solves for the new modes of the structure with the tuned
vibration absorber attached.
1142 M. Richardson and D. Formenti

To begin the design, a mass must be chosen for the tuned absorber. The following
rule should be used in choosing an absorber mass.

Rule of Thumb: The mass of a tuned absorber should not exceed 10% of the mass
of the structure.

After the mass has been chosen, the frequency of the structural mode to be
suppressed together with the mass of the absorber will determine the stiffness of
the spring required to attach the absorber to the structure. These three values are
related to one another by the formula:

k = m ω2 (47)
where

m← tuned absorber mass

ω ← frequency of the structural mode to be suppressed
k ← tuned absorber stiffness

Adding a damper is optional. If a damper is added between the absorber mass

and the structure, its damping value must also be chosen. A realistic damping value
of a few percent of critical damping is calculated using the following formulas.
 
k = m ω2 + σ 2 (48)

where
σ= √ω ← damping decay constant
1−%2
% ← percent of critical damping
The mode shape of the unattached tuned absorber is simply the UMM rigid-body
mode shape of the mass substructure in free space. In order to use SDM to model a
tuned absorber, two more steps are necessary:

1. The free-free mode shape of the tuned absorber must be added in block diagonal
format to the mode shapes of the unmodified structure. The block diagonal
format was explained in the previous plate and RIB example.
2. The attachment DOF (point and direction) of the tuned absorber must be defined.
A geometric model of the structure is usually required for this.

29 Adding a Tuned Absorber to the Plate and RIB

SDM will be used to model the attachment of a tuned vibration absorber to one
corner of the flat aluminum plate used in the previous example. The absorber will
be designed to suppress the amplitude of the high-Q resonance at 108 Hz, shown in
the blue FRF magnitude plot in Fig. 17.
20 Structural Dynamics Modification and Modal Modeling 1143

Fig. 17 Synthesized FRFs (1Z:1Z) before and after absorber

The plate and RIB weighs about 21.3 lbm (9.7 kg). For this example, the absorber
weight is chosen as 0.5 lbm (0.23 kg). In order to absorb energy from the plate and
RIB at 108 Hz, the attachment spring stiffness must be chosen so that the absorber
will resonate at 108 Hz.
The absorber parameters are:

Mass: 0.5 lbm (0.23 kg)

Stiffness: 586.6 lbf/in (104.8 N/mm)
Damping: 0.5%

Only the modal model data of the unmodified plate and RIB at DOF 1Z is
required. Since the mass will be attached to the plate and RIB as a substructure, the
mode shape of the free-body mass is added to the mode shapes of the unmodified
plate and RIB in block diagonal format, explained in the previous example.
To model the tuned absorber, the modal model for the unmodified plate and RIB
substructure together with a modal model and the spring and damper of the absorber
are used as inputs to SDM. SDM then solves for the new modes of the plate and RIB
with the absorber mass attached by the spring and damper to one corner of the plate
(DOF 1Z).
Figure 17 shows the log magnitudes of two overlaid driving point FRFs of the
plate and RIB at DOF 1Z, before (blue) and after (red) the tuned absorber was
attached to the plate. These overlaid FRFs clearly show that the resonant frequency
at 108 Hz has been removed from the plate and RIB and replaced with two new
resonances, one at 84 Hz and the other at 128 Hz. The two new modes also have
lower Q’s (less amplitude) than the Q of the mode they replaced.
1144 M. Richardson and D. Formenti

Table 7 Modes before and after tuned absorber attached at 1Z

Before TA Before TA After TA After TA
Shape pair frequency (Hz) damping (Hz) frequency (Hz) damping (Hz) MAC
1 108.2 0.0345 84.3 0.149 0.96
2 108.2 0.0345 127.6 0.333 0.93
3 187.6 0.369 190.4 0.422 0.99
4 253.3 0.118 258.7 0.250 0.98
5 311.5 0.293 317.9 0.488 0.98
6 351.7 0.104 354.4 0.217 0.99
7 479.2 0.171 480.7 0.239 1.00
8 521.3 0.713 524.9 0.924 0.97
9 537.4 2.77 538.7 2.808 0.98
10 619.1 0.863 622.1 1.055 1.00
11 801.1 0.544 801.1 0.544 1.00

The MAC values in Table 7 show that the two new mode shapes are essentially
the same as the mode shape of the original 108 Hz mode. Notice also that the tuned
absorber had very little effect on the other resonances of the structure.
Figure 18 shows how the tuned absorber mass moves with respect to the plate. In
Fig. 18a. the tuned absorber is moving in-phase with the plate below it. In Fig.18b.
it is moving out-of-phase with the plate below it. (An animated picture shows this
relative motion more clearly.)

30 Modal Sensitivity Analysis

It is well-known that the modal properties of a structure are very sensitive to

changes in its physical properties.

Because of its computational speed, SDM can be used to quickly solve for
the modal parameters of thousands of potential modifications to a structure. The
calculation and ordering of multiple SDM solutions from best to worst is called
Modal Sensitivity Analysis.

30.1 EMA Modes of the Plate and RIB

In a previous example, SDM was used to model the attachment of a RIB stiffener
to the aluminum plate shown in Fig. 5a–c. To validate the SDM mode shapes using
experimental data, the plate with the RIB attached was tested with a roving impact
hammer test.
20 Structural Dynamics Modification and Modal Modeling 1145

Fig. 18 (a) Absorber in-phase with the 84 Hz mode. (b) Absorber out-of-phase with the 128 Hz
mode
1146 M. Richardson and D. Formenti

Table 8 EMA versus SDM modes for the plate with RIB
SDM
EMA EMA damping SDM damping
Shape pair frequency (Hz) (Hz) frequency (Hz) (Hz) MAC
1 103.8 0.144 108.2 0.0345 1.00
2 188.5 0.360 187.6 0.369 0.99
3 242.5 0.262 253.3 0.118 0.99
4 259.7 0.378 311.5 0.293 0.98
5 277.4 1.164 351.7 0.104 0.97
6 468.6 0.760 479.2 0.171 0.98
7 503.6 6.035 521.3 0.713 0.97
8 572.6 4.953 537.4 2.77 0.98
9 618.8 1.828 619.1 0.863 0.87
10 657.5 6.541 801.1 0.544 0,95

The plate was impacted at 24 points on the plate in the (vertical) Z-direction to
gather enough data to uniquely define the EMA mode shapes for comparison with
the SDM mode shapes. In Table 8, the first eight EMA and SDM mode shape pairs
have MAC values close to 1.0, indicating that they are closely matched. But the
EMA and SDM modal frequencies are all different from one another.

30.2 Using SDM to Explore Joint Stiffnesses

The first mode of the plate and RIB involves twisting of both the plate and RIB,
as shown in Fig. 14. The mode shape is influenced by both the translational and
rotational stiffness of the spring stiffeners used to attach the RIB to the plate.
Using the Hybrid modal model containing the mode shapes of the plate without
the RIB attached, SDM can be used to quickly calculate the modes of the plate and
RIB using many different translational and rotational stiffnesses of the springs used
to attach the RIB to the plate. These solutions are then ordered from best to worst.
Modal Sensitivity Analysis can be performed by calculating and ordering multiple
SDM solutions.

30.3 Current Versus Target Frequency

A Modal Sensitivity window is set up in Fig. 19a to perform sensitivity calculations

on the plate and RIB. The window contains two spreadsheets. The frequencies of the
30 modes of the unmodified plate and RIB substructures are listed in the Current
Frequency column of the upper spreadsheet. This modal model contains 14 mode
shapes of the plate without the RIB attached and 16 free-body mode shapes of the
RIB. The mode shapes are sorted according to frequency, beginning with the rigid-
body mode shapes of the RIB.
20 Structural Dynamics Modification and Modal Modeling 1147

Fig. 19 (a) Modal sensitivity setup for 2500 solutions

The EMA modal frequencies of the plate and RIB are listed in the Target
Frequency column in the upper spreadsheet. These frequencies are used for ranking
the SDM solutions from best to worst.
Ten Shape Pairs have been selected in the upper spreadsheet. The Selected Pairs
are used to order the solutions from best to worst. The best solution is the one
which minimizes the difference between each Solution Frequency and each Target
Frequency.
1148 M. Richardson and D. Formenti

Fig. 19 (b) Best solution with eight shape pairs selected

30.4 Solution Space

In Table 8, the first EMA mode shape (at 103.8 Hz) has a lower frequency than the
first SDM mode shape (at 108.2 Hz). Therefore, the best Modal Sensitivity solution
should require less stiffnesses than the stiffnesses (1,000,000) used to attach the RIB
to the plate.
The lower spreadsheet defines ranges of stiffness values for the translational and
rotational stiffnesses of the five FEA springs. Each stiffness has a range of 50 Steps
(or values) in its solution space. Each SDM solution will use a stiffness value from
20 Structural Dynamics Modification and Modal Modeling 1149

the Minimum Property (1000) to the Maximum Property (2,000,000) of each

stiffener. The solution space has 50 steps × 50 steps = 2500 stiffness values in it.
SDM will solve for new modes using all combinations of stiffness values in the
solution space of the two stiffeners.
Figure 19b shows the Modal Sensitivity window after 2500 solutions have been
calculated and ordered from best to worst. The modal frequencies of the best
solution are displayed in the Solution Frequency column of the upper spreadsheet.
The damping values are displayed in the Solution Damping column.
The stiffness values used to calculate the best solution are displayed in the
lower spreadsheet. The translational stiffness used to calculate the best solution
is 4.18 E04 lbf/in. The rotational stiffness used to calculate the best solution is
2.05 E05 (lbf-in)/deg. Much less translational and rotational stiffness of the five
spring stiffeners was required to closely match the frequencies of the first eight
EMA modes of the plate and RIB.

31 FEA Modal Updating

Because of its computational speed, SDM can be used to quickly evaluate thousands
of modifications to the physical properties of an FEA model. In Table 3 their MAC
values indicate that each FEA mode shape of the plate closely matches with an EMA
mode shape, but each FEA modal frequency is less than the EMA frequency of its
matching mode shape.
The physical properties used for the plate elements in the FEA model of the
aluminum plate were:

1. Young’s modulus of elasticity: 1E07 lbf/inˆ2 (6.895E4 N/mmˆ2)

2. Density: 0.101 lbm/inˆ3 (2.796E-6 kg/mmˆ3)
3. Poisson’s ratio: 0.33
4. Plate thickness: 0.375 in (9.525 mm)

The plate is made from 6061-T651 aluminum. A more accurate handbook value
for the density of this alloy of aluminum is 0.0975 lbm/inˆ3 (2.966E-6 kg/mmˆ3).
In addition, the quad plate elements were given a nominal thickness of 0.375 in
(9.525 mm). Plate stiffness is very sensitive to its thickness!
Error in the density or thickness of the elements in the FEA plate model could be
the reason why the frequency of each FEA mode shape was less than the frequency
of its corresponding EMA mode shape.
An FEA Modal Updating window is set up in Fig. 20a to perform SDM
calculations using multiple density and thickness values. The FEA Frequency of
each of the 14 FEA mode shapes of the plate is listed in the upper spreadsheet. Each
EMA frequency is listed as a Target Frequency.
All 14 mode Shape Pairs are selected, meaning that each Solution Frequency
will be compared with each Target Frequency to order the solutions from best to
1150 M. Richardson and D. Formenti

Fig. 20 (a) Setup for 100 FEA model updating solutions

worst. The best solution is the one which minimizes the difference between each
Solution Frequency and each Target Frequency.
The Solution Space is defined in the lower spreadsheet. The plate thickness
and density are selected, and each has 10 Property Steps (or values) between its
Property Minimum and Property Maximum. Solutions will be calculated over
a solution space of 100 property values, using all combinations of 10 different
thicknesses and 10 different densities.

The properties of the original FEA model are required in order to update those
properties.
20 Structural Dynamics Modification and Modal Modeling 1151

Fig. 20 (b) Best solution for updating density and thickness

To perform FEA Model Updating, the properties of the unmodified model must
be removed from the mass and stiffness matrices before the new properties can be
added.
Figure 20b shows the Model Updating window after 100 solutions have been
calculated and ordered from best to worst. For all 14 Shape Pairs, each Solution
Frequency closely matches each Target Frequency. The Solution MAC between
each Shape Pair also indicates that the mode shapes of all 14 mode shapes were not
changed by updating the density and thickness.
The updated density (0.0967) more closely matches the handbook density for
6061-T651 aluminum. The updated thickness (0.417 in.) is more than the thickness
originally used, but it resulted in new modal frequencies that more closely matched
the experimental frequencies.
1152 M. Richardson and D. Formenti

32 Difference Between Modal Sensitivity and FEA Model

Updating

In order to calculate the new modes of a modified structure, SDM only requires a
modal model of the unmodified structure together with FEA elements. For Modal
Sensitivity Analysis, the properties of modification elements are used. For FEA
Modal Updating, the properties of FEA elements are used.

In Modal Sensitivity Analysis, multiple SDM solutions are calculated over a

solution space of modification element properties, and the solutions are ordered
from best to worst based on how closely the Solution frequency and damping of
each selected Shape Pair match Target modal frequency and damping.
In FEA Model Updating, multiple SDM solutions are calculated over a solution
space of FEA model properties, and the solutions are ordered from best to worst
based on how closely the Solution frequency and mode shape of each selected
Shape Pair match Target modal frequency and mode shape values.

Whether SDM is used for Modal Sensitivity or FEA Model Updating studies,
thousands of potential property changes can be quickly evaluated and sorted from
best to worst based on how close a Solution is to Target values. In these applications,
SDM is very useful for “closing the gap” between analytical and experimental
results.

References
1. Hallquist J (1974) Modification and synthesis of large dynamic structural systems. PhD
dissertation, Michigan Technological University
2. Formenti D, Welaratna S (1980) Structural dynamics modification – an extension to modal
analysis. SAE paper no. 811043
3. Structural Measurement Systems, Inc. An introduction to the structural dynamics modification
system. Technical note no.1, February 1980
4. Ramsey K, Firmin A (1982) Experimental modal analysis structural modifications and FEM
analysis – combining forces on a desktop computer. First IMAC proceedings, Orlando, 8–10
5. Wallack P, Skoog P, Richardson MH (1988) Simultaneous structural dynamics modification
S2 DM). Proceedings of 6th IMAC, Kissimmee
6. Bathe KJ, Wilson EL (1976) Numerical methods in finite element analysis. Prentice Hall, Inc,
Englewood Cliffs
7. Yang TY (1986) Finite element structural analysis. Prentice Hall, Inc, Englewood Cliffs
8. McGuire W, Gallagher RH (1979) Matrix structural analysis. Wiley, New York
Toward Robust Response Models:
Theoretical and Experimental Issues 21
Nuno Maia, António Urgueira, Raquel Almeida, and Tiago Silva

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154
1.1 Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154
1.2 Classification of the Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1155
2 Coupling/Uncoupling Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1158
2.1 Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1158
2.2 Uncoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1162
3 Measurement of Rotational Degrees of Freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164
3.2 Experimental Methods for Measuring Rdofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1166
3.3 Experimental Methods for Exciting Rdofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1184
4 Condensation (Reduction) Versus Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1186
4.1 Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1187
4.2 Expansion of Measured Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1195
5 Transmissibility as a Means to Estimate the Dynamic Response . . . . . . . . . . . . . . . . . . . 1198
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1198
5.2 Theoretical Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1199
5.3 Other Possible Applications of Transmissibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 1202
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1202

N. Maia ()
Department of Mechanical Engineering, LAETA/IDMEC/IST (Instituto Superior Tecnico),
University of Lisbon, Lisbon, Portugal
e-mail: [email protected]
A. Urgueira · R. Almeida · T. Silva
UNIDEMI, Department of Mechanical and Industrial Engineering, FCT NOVA – Faculty of
Science and Technology, Universidade NOVA de Lisboa, Caparica, Portugal

© The Society for Experimental Mechanics 2022 1153

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_26
1154 N. Maia et al.

Abstract

The purpose of building a reliable response model requires awareness and

understanding of the theoretical and practical problems and pitfalls that may be
encountered along the process. Those issues have to deal with many different
aspects of structural dynamics that will be tackled in this chapter. First of all,
it is necessary to frame the problem in the context of structural dynamics and
modal analysis; then, it is crucial to discuss the substructuring/coupling problem,
especially when facing a very complex structure, which needs to be studied
by parts. This leads us inevitably to the difficult task of measuring angular
responses, that is, to measure the response at rotational degrees of freedom. A
comprehensive discussion on this matter is given here.
Not only due to the lack of information regarding rotational measurements, a
constant issue is the incompleteness of the experimental data and its expansion to
the numerical model size or the condensation/reduction of the numerical model
to the experimental one. A significant discussion on this subject is also given in
this chapter.
Finally, the more recently developed transmissibility theory for N degrees of
freedom allows for the estimation of unmeasured frequency response functions,
contributing to the building of robust response models.

Keywords

Response model · Coupling · Rotational measurements ·

Expansion/Reduction techniques · Transmissibility · FRF estimation

1 Introduction

1.1 Foreword

The dynamic analysis of real structures is, in general, a complex matter that can
be approached from various points of view, depending essentially on the purpose
of the study and ultimately on the computational resources available. In any case,
it may be also constrained by the complexity of the structure itself, which may
prevent an easy numerical modeling, if that is the case, or an easy access when
it comes to experimental testing. Sometimes, one may “simply” wish to evaluate
the condition of the structure to detect any malfunctioning or damage progression,
through the direct measurement of the output response, other times one wishes to
know the response at one coordinate due to an applied force at another coordinate.
There are cases, however, where it is important to have a model of the structure that
can reproduce, as accurately as possible, its effective behavior in practice. Such a
model may have just a few degrees-of-freedom or many thousands, depending once
more on the objectives of the study. The main advantage of having a good model
representing the dynamics of the structure is that one can use it to predict what
will happen if a modification is made – like adding mass, stiffness, or damping at
21 Toward Robust Response Models: Theoretical and Experimental Issues 1155

some coordinates – or what will be the consequence of coupling a new substructure,

without having to go through a series of experimental testing, which tends to be
costly and time-consuming.
There are essentially two options when it comes to build a model: the exper-
imental approach and the theoretical approach. Most researchers working on the
experimental side tend to select a model that incorporates the information directly
available from tests. On the other hand, the theoretical models are constructed in
order to incorporate the spatial distributed properties.
In this chapter, devoted to the response (or impedance) modeling, one shall
address its relation with respect to other modeling approaches, its advantages,
disadvantages or limitations, referring also to some topics that are directly con-
nected, like associated coupling techniques, the measurement of rotational degrees
of freedom, incompleteness of the models, leading to reduction or expansion issues,
and estimation of unmeasured responses.

1.2 Classification of the Models

Essential to the communication among different organizations and research groups

is the terminology and format used to describe the dynamic properties of a system
or subsystem model, be it theoretically or experimentally derived. In what follows,
the three different model formats are described.

1.2.1 The Spatial Model

Any structure can be modeled numerically, in a more or less accurately way; this
is often accomplished using the finite element method, where one ends up with a
description of the model in terms of mass, stiffness, and damping matrices. It is
known as the spatial model. The dynamic equilibrium equation is given by:

M ẍ + C ẋ + Kx = f (1)

where M, K, and C are matrices of order N (the number of degrees of freedom)

representing the mass, stiffness, and viscous damping, respectively. In such a
numerical model the damping is often neglected, as it requires experimental
information to be taken into consideration. This can be a reasonable option to start
with when the damping is really low. As a common alternative, one may assume
a Rayleigh (proportional) type of damping, where the damping matrix is directly
proportional to the stiffness and/or mass matrices, although some constants still have
to be evaluated experimentally. In any case, it will be necessary to “tune” the model
so that it represents the reality as close as possible, undergoing a refinement process
involving some kind of comparison with real experimental data; such a process is
known as model updating.
1156 N. Maia et al.

1.2.2 The Response Model

Sometimes, due to the complexity of the structure, it may not be so easy to build a
numerical or analytical model and it is preferable to excite the structure with forces
at some locations and measure the responses at those or other locations. By doing
so, a series of the so-called frequency response functions (FRFs) can be grouped
together to build a matrix, where each element is an FRF relating the response at
coordinate i to the force at coordinate j, expressed as:

Xi (ω)
Hij (ω) = (2)
Fj (ω)

provided that all other forces are zero.

One has, therefore, the FRF matrix H representing the response model.
By taking into account all the degrees of freedom, one may write:

X = HF (3)

Ideally, if all the dofs can be measured and excited, one can write the inverse
relation

F = ZX (4)

where Z = H−1 is the dynamic stiffness matrix. In practice, this is not the case, as
one can only measure a reduced number of dofs, one may end up with a reduced or
condensed Z matrix. This will be discussed in Sect. 4.
Apparently, one might think that it would be indifferent to measured Z or H.
However, that is not the case, as each element Zij represents the force applied at
dof i when dof j responds with a unit displacement and all the remaining dofs other
than j do not move, that is, numerous dofs would have to be constrained to have
no motion, which is not possible, or at least impractical. This is not the case with
the FRF matrix H, as each element Hij represents the response at dof i due to a
unit force at dof j, provided that all other forces are zero. Moreover, a measured
FRF matrix at one stage of the design can be later expanded by simply performing
new measurements at additional coordinates to form new rows and columns without
modifying any existing elements. This is not the case whenever a dynamic stiffness
matrix needs to be expanded, since all the existing elements would have to be re-
measured after the new constraints had been imposed to the structure (assuming this
could be somehow accomplished in practice).
Concluding, in practice, the response model is build based upon the measurement
of FRFs, which can be in terms of displacement, velocity, or acceleration. Due to
the wide use of accelerometers to acquire responses, the FRFs are often measured
in terms of accelerations, that is, accelerance FRFs.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1157

1.2.3 The Modal Model

It may also happen that some parts of the structure be modeled using the spatial
model and other parts using the response model. In the “middle” stands the
modal model, based on the so-called modal properties, constituted by the natural
frequencies ωr , mode shapes φ r , and damping ratios ξ r . In some applications it may
be convenient to have the structure described in such terms. In contrast with the other
two kinds of models, the modal model is not directly obtained from the structure;
it is deduced from either the spatial model, solving a generalized eigenproblem, or
the response model, using a modal identification technique.

1.2.4 Relation Among the Models

The three models are, of course, interrelated. From (1), assuming harmonic excita-
tion, it follows that
 
K − ω2 M + iωC X = F (5)

Comparing Eqs. (3), (4), and (5), one concludes that the spatial and response
models are related in the following way:
 −1
H (ω) = Z −1 (ω) = K − ω2 M + iωC (6)

To relate the response and modal models one has to use the orthogonality con-
ditions. Let φ be the mass-normalized mode shape matrix. Assuming proportional
damping, as it is usual to do when modeling with the finite element method, the
orthogonality conditions are:

φ T Mφ = I
 
φ T Kφ = diag ωr2 (7)

φ T Cφ = diag (2ξr ωr )

from which one can deduce the three matrices:

M = φ −T φ −1
 
K = φ −T diag ωr2 φ −1 (8)

C = φ −T diag (2ξr ωr ) φ −1

Substituting (8) in (6), inverting and pre- and post-multiplying by φ T and φ,

respectively, yields:
 
φ T H −1 φ = diag ωn2 − ω2 + i2ξr ωωr (9)
1158 N. Maia et al.

Fig. 1 Interrelation among the three dynamic models for an undamped model

and thus,
  −1
H (ω) = φ diag ωn2 − ω2 + i2ξr ωωr φT (10)

Provided that the models are complete, that is, that one has the information
from all the degrees of freedom, one can easily go from one model to the
other. From the spatial model to the modal model one has to solve a generalized
complex eigenproblem to obtain the complex natural frequencies (which include
the information about damping) and the mode shapes. The inverse process simply
implies the application of Eq. (8). Going from the response model to the modal
model requires a modal identification process, whereas the opposite operation is
much easier, just involving the use of Eq. (10). Figure 1 illustrates the interrelation
among the three models, for the undamped case (for clarity of exposition). The
incompleteness of the models will be addressed in Sect. 4.

2 Coupling/Uncoupling Techniques

2.1 Coupling

The basic principle inherent to the coupling philosophy is the assumption that the
whole (complex) structure is formed by different substructures (or components),
each of them being first analyzed individually and independently from the others.
This is the idea underlying the nowadays well-known “substructuring,” or “cou-
pling” approaches for solving static and dynamic problems.
The order of the matrices used to formulate the equations of motion of the
assembled structure depends either on the number of coordinates (connection and
interior ones) or on the number of kept modes pertaining to each component model.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1159

One can say that two groups of coupling techniques emerge from the large variety
of methods that have been used in different fields of research and industry, classified
as:

• Impedance coupling techniques, which benefit from the reduction performed on

the subsystem models in terms of coordinates
• Modal coupling techniques, which are suitable for the use of reduced models in
terms of modes

The former group deals primarily with the coupling of subsystems whose models
are described either by their spatial or response properties. The first of these types of
models is used extensively in the finite element method, but is rarely used in cases
involving experimental modeling. Although response models can be obtained by
theoretical analysis, they mostly constitute the raw data available from modal tests.
The techniques forming the latter group are applied in those situations where the
component models are described by their modal properties – modal models. This
type of model is easily generated from an eigensolution, if a theoretical tool such
as the finite element method is used, or they can be derived from an identification
process carried out on measured FRF data.
Depending on whether spatial or response component models are used directly as
input data into a coupling process, the assembling techniques are here designated as
spatial or FRF coupling techniques, respectively. The former is ideal for the use of
FE methods, whereas the latter comprises both theoretical and experimental fields
of work. The FRF coupling technique is generally referred to as the “Impedance
Coupling” technique since, at the system level, it assembles mathematically the
generalized impedance properties.
The application of the conceptually simple impedance coupling technique is
straightforward when the components are amenable to theoretical modeling, but
practical complex systems have demanded subsystem impedances to be derived
from measured data. Although the models obtained via this latter approach have
the advantage of reflecting in a closer way the “true” dynamic characteristics of a
structure, they are contaminated by errors arising during the acquisition and analysis
of measured data. Consequently, the accuracy of the dynamic properties of the
assembled structure results will be somehow affected.
The experimental approach to the impedance coupling problem, herein desig-
nated as FRF coupling, was one of the main reasons that motivated a breakthrough
to the development of suitable techniques and equipment to measure, assess, and
analyze data. The main difficulties encountered in those applications were mainly
related to the mathematical inconsistency of the measured models and to the
inadequacy of experimental means to measure some terms in the FRF matrices
of certain components. Mostly, those FRFs were related to rotational response
measurements. Rotational responses are necessary to formulate proper constraints
between connected components. Success in the prediction of results for a coupled
structure is dependent on how the connection coordinates are measured and included
in a coupling process. For instance, if a single connection point is assumed between
1160 N. Maia et al.

two components that respond to excitations in all three planes, it is vital to include
the three rotations in addition to the three translations, in order to properly formulate
the constraint conditions. In terms of response models, the FRFs related to rotational
response/excitations represent 75% (or 60% if symmetry properties are assumed) of
the total elements of the corresponding FRF matrix. Section 3 addresses the topic of
measuring FRFs involving rotational dofs.
The FRF coupling method makes use of subsystem models derived directly
from FRF data (commonly available from experimental studies but seldom from
theoretical modeling). The dynamic properties of those models are synthesized
in terms of the FRF matrix and generally denoted as H(ω) (such as receptance,
mobility, or accelerance matrices). The coordinates involved in the connection
between components A and B should be identified and represented by the index
c (and similarly i and j for the remaining ones), leading to the following partitioned
FRF matrices:

   
HA A
ii H ic
HB B
jj H j c
HA = A A and H B = (11)
H ci H cc H cj H B
B
cc

By invoking the constraint equations that impose that at the connection coordi-
nates the forces obey to the action/reaction principle and that the displacements of
each component A and B are the same, the FRF matrix of the coupled structure [77]
turns out to be

 −1  −1 −1
HC = HA ⊕ HB (12)

where the operation sign “⊕” means “coupled to.” Denoting by Zkl the element k,l
in the generalized impedance matrix Z = H−1 , the operation sign “⊕” implies that
the connectivity between the coordinates of both components must be respected,
leading to the following assembling of matrices:

⎡ ⎤−1
ZA
cc ZAic 0
⎢ B ⎥
= ⎣ ZAci Z cc + Z cc Z cj ⎦
A B
HC (13)
0 ZB
jc ZB
jj

The FRF matrices of each substructure (available over a frequency range of

interest, which is the same for both structures) are “added” together frequency by
frequency until the whole FRF matrix is completely calculated.
A refined version for the calculation of the FRF matrix of the coupled structure
reduces the number of inversions to be carried out [59]:
21 Toward Robust Response Models: Theoretical and Experimental Issues 1161

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
HA A HA A T
ii H ic 0 ic  −1 H ic
⎢ ⎥ ⎢ A⎥ ⎢ A⎥
HC = ⎣HA ci H cc 0 ⎦ − ⎣ H cc ⎦ H cc + H cc
A A B
⎣ H cc ⎦ (14)
0 0 H jj B B
H jc HBjc

Herein only one inversion is required and, additionally, it is applied only to the
sum of the sub-matrices whose order depends solely on the number of connection
coordinates. More interior coordinates can then be included in the analysis without
affecting significantly the required computational time. The result of this will be
a quicker calculation of the required FRF matrix and, as in the latter approach, it
will be able to deal with redundancies on the interior coordinates whenever they are
present in each component.
There are several fields where one can encounter applications of FRF coupling
techniques. For instance, the mechanical behavior of the human body has long
been characterized using biodynamic measurements on various human body parts
in several positions and postures. In the case of the interaction between a human
body and a vibrating structure, the dynamics involved in the structure alone is as
important as the dynamics of the human body. Thus, the use of FRF coupling to
combine biodynamic measurements with the dynamic behavior of the structure is
essential to understand the vibration transmission phenomena in such a complex
assembly. The research work [101] presents the advantages and also the challenges
of using FRF coupling techniques between a mechanical structure and biodynamic
measurements.
Experimental/analytical substructuring methods have long been explored to
expedite testing and analysis of built-up systems in various fields. However, many
of these efforts have failed because the substructuring calculations can be very
sensitive to experimental uncertainty and truncation of the subcomponent models.
The work by Allen et al. [3] presents a review of the literature regarding uncertainty
in experimental/analytical substructuring, highlighting the phenomena that have
been observed such as inherent ill-conditioning, cross-axis sensitivity, uncertainty
modeling and propagation, as well as the fact that experimental measurements may
exhibit phenomena that are not physically realizable.
The other coupling method one shall be concerned with is the modal coupling
technique, also referred to in the related literature as time domain or component
mode synthesis methods. The basic philosophy is the same as the previous one, that
is, it permits the use of reduced component models in order to achieve a reduced
order in the final equation of motion matrices of the assembled structure. However,
unlike the impedance-based methods, which take advantage on the reduction of the
number of coordinates, those methods use a reduction performed on the number of
modes used to describe each component model, while still accounting for all the
physical dofs. By using a Ritz-type transformation, the reduced number of principal
coordinates is related to the number of modes that is taken into account for the modal
estimation; generally, the information relating to the higher natural frequency modes
is discarded. Essentially, there are two modal coupling approaches that differ from
each other according to the dynamic displacement shapes used to form the truncated
1162 N. Maia et al.

set of the natural modes. In the first approach, the elastic modes pertaining to a
fixed-interface component are retained, whereas in the second one the modes are
obtained by assuming the component to be vibrating in a freely supported condition
at its attachment points. This being the most readily simulated condition during
an experimental test, it constitutes an attractive technique for the use of combined
experimental/theoretical analysis of dynamic structural systems.
The use of a set of truncated modes [140] to establish the compatibility equations
sometimes leads to unacceptable errors in the prediction of the assembled system
responses. Thus, for rigidly connected subsystems a more accurate definition is
necessary, either by including more modes or, if these represent an unreasonable
number, by providing some information about the effects of the neglected modes.
Two possible ways may be used to improve the structural definition of each
component; the first one, by using additional masses attached to the connection
points in an attempt to generate a more realistic condition for the component when
it is vibrating together with the remaining parts (the localized flexibility properties
near the connection area are better represented, since more modes are brought to the
frequency range of interest); the second one, by seeking to compensate for the lack
of flexibility due to the truncation of the set of natural modes by using additional and
important information concerning the flexibility effects of the out-of-range modes
[141]. Component mode synthesis is also addressed in Sect. 4, as a model reduction
technique.
A different approach to the coupling procedure has been proposed by Rixen
and van der Valk [114, 115], where substructures are no longer characterized by
their matrices, but rather by the impulse response on their interface. The impulse-
based substructuring (IBS) method proposed in Rixen and van der Valk [114]
provides a systematic way to couple components through their interface using
impulse response functions obtained analytically, experimentally, or numerically.
The method expresses the interface problem in a dual form, meaning that the
interface forces are computed in order to enforce the interface compatibility
conditions. The time response is then computed by applying a convolution product
(Duhamel’s integral) in each substructure.

2.2 Uncoupling

Generally, the interest of knowing the properties of one of the components leads
us to the formulation of the uncoupling techniques. This situation is regularly
encountered in practice in structural monitoring and vibration control techniques,
where monitoring and controlling of individual (critical) components in an assembly
can be very valuable. This problem arises when substructures cannot be measured
separately, but only when coupled to neighboring substructures.
Reliable solutions of the uncoupling problem could lead to promising
developments, both in the field of diagnostics and in the field of vibration control.
Uncoupling procedures may either start with impedance-like matrices (impedance-
based approach) or with mobility-like matrices (mobility-based approach).
21 Toward Robust Response Models: Theoretical and Experimental Issues 1163

The work [31] highlights pitfalls due essentially to lack of measured DoFs,
producing ill-conditioning in the neighborhood of certain frequencies.
The uncoupling problem, that is, the identification of the dynamic behavior of
a structural subsystem, starting from the known dynamic behavior of the complete
system, and from information about a second component subsystem, is revisited
in the general framework of frequency-based substructuring [28, 33]. Several
approaches have been proposed in the literature to tackle the uncoupling problem.
However, all of them present some pitfalls that have been also highlighted, such as
modal truncation, lack of information on coupling DoFs (rotational DoFs), and ill-
conditioning in the neighborhood of particular frequencies. It has been shown that
the last problem can be circumvented by including internal DoFs in the measured
dataset, of course together with coupling (or interface) DoFs. In previous papers,
two frequency-based approaches were considered: an impedance-based approach
and a mobility-based approach. In both approaches, the FRF matrix of the coupled
system is assumed to be known at the coupling DoFs, and possibly at some internal
DoFs of one subsystem. In this chapter, an approach derived through the dual
formulation, within the general framework of frequency based substructuring, is
developed and discussed. Possible difficulties in the use of additional internal DoFs
are envisaged and investigated.
For instance, our focus can be the identification of the dynamic properties of a
subsystem that works as a joint. In this case there are three main formulations [9];
let a structure C be composed of two substructures, A and B; A is divided into two
parts, connected by B, which represents the joint: one of the formulations involves
only the information about the connection coordinates (c), the second one requires
only the information of the interior coordinates of A (i), and the third one uses a
mixture of connection and interior coordinates.
Considering the first formulation, the following equation is written:
  −1
HB
cc = cc H cc − H cc
HA A C
− I cc H A
cc (15)

If one cannot access the connection coordinates of the global structure C after all
components being coupled, a second formulation can be used, as follows:

   −1
HB
cc = H A
ci H A
ic H A
ci H A
ii − H C
ii H A
ic ci H ic − H cc
HA A A
(16)

In the third formulation, all types of coordinates can be used

  −1
cc = H cc H ci H ic − H ic
HB ci H ic − H cc
A A A C
HA A A
(17)

In both cases two and three, the number of interior coordinates should be greater
than the number of connection coordinates.
1164 N. Maia et al.

An additional formulation can be written whereby the inversion process is

applied to a matrix whose size depends on all the coordinates,

   H A H A   H C H C  −1 
HA

HB
cc = H ci H cc
A A ii
A − HC HC
ic ii ic ic − H A
cc (18)
HAci H cc ci cc HA
cc

An uncoupling technique presumes that a fictitious subsystem that is the negative

of the residual subsystem is added to the coupled system, and appropriate com-
patibility and equilibrium conditions are enforced at interface DoFs. Compatibility
and equilibrium can be required either at coupling DoFs only (standard interface),
or at additional internal DoFs of the residual subsystem (extended interface), or
at some coupling DoFs and/or some internal DoFs of the residual subsystem
(mixed interface). Using a mixed interface, rotational coupling DoFs could be
eliminated and substituted by internal translational DoFs. This would avoid difficult
measurements of rotational FRFs. This possibility is verified in the paper [34] using
simulated experimental data. A sensitivity study was performed on the uncoupling
techniques taking into account the uncertainties in the dynamic properties of the
components [32].

3 Measurement of Rotational Degrees of Freedom

3.1 Introduction

The study of the dynamic behavior of structures has taken on, in recent decades,
a key role in engineering, even in areas as diverse as: conditioned maintenance,
aerodynamics, structural modification, damage detection, fatigue life study, seismic,
comfort and safety. When it is required the knowledge of the dynamic behavior
of some structure, two techniques are commonly used: a numerical one, called
finite element method (FEM), which requires the knowledge of the mass, stiffness,
and damping properties of the elements that make up the structure under study
and an experimental one, named experimental modal analysis (EMA). However,
the numerical models developed using the FEM lack, in most cases, experimental
validation.
One of the mathematical models generally used to describe the dynamic behavior
of a structure is the response model, which relates the response to an excitation, at
two given locations. The response may either be a linear (translation) or an angular
(rotation) displacement. The excitation could be a force or a moment. Thus, one can
consider relations between translation and force, translation and moment, rotation
and translation, and rotation and moment. However, the experimental determination
of any other relations, except the relation between translation and force, continues
to present a major challenge for the scientific community. The biggest obstacle for
the experimental determination of these quantities is the excitation of the structures
with a pure moment, being the second one the corresponding measurement of
21 Toward Robust Response Models: Theoretical and Experimental Issues 1165

the rotational motion. It is clear that the lack of such information will lead
to obtaining an incomplete experimental model; the missing information could
easily represent more than 50% of the complete model. In a schematic way, and
considering the possibility of 6 dofs for each measurement point, the global mobility
matrix describing the FRFs can hence be subdivided by grouping together all the
translational dofs (Tdofs) and all the rotational dofs (Rdofs), resulting:

T ⇒ translational response (Tx , Ty , Tz )

 
H TF H TM R ⇒ rotational response (Rx , Ry , Rz )
H = with : (19)
H RF H RM F ⇒ force excitation (Fx , Fy , Fz )
M ⇒ moment excitation (Mx , My , Mz )

The restriction of the structural FRF measurement to Tdofs results in the fact that
only 25% of the complete mobility matrix is effectively known.
The effects of spatial incompleteness in structural coupling predictions are firstly
identified, in the ends of 1960s, by Smith [129] and later by Ewins and Sainsbury
[47]. The importance of assessment of Rdofs in FRF coupling has been studied
by several researchers [40, 45, 52, 73, 127, 140]. The studies were unanimous
in concluding that the absence of Rdofs in coupling applications could cause an
underestimation of correct predictions. Urgueira and Ewins [141] and Duarte and
Ewins [40] also verify that coordinate incompleteness is not the only source of error
in the prediction of the dynamic behavior of a coupled system, residuals can also be
a problem. This is particularly important for Rdofs at high frequencies, for which
the residual terms are normally much larger than those related to Tdofs.
Also, in the fields of structural dynamic modification (SDM) and finite element
Updating, it was soon understood the importance of the inclusion of Rdofs in
the analysis, especially for beam-like and plate-like structures [5, 20, 30, 91, 96,
127, 128]. Avitabile and Piergentili [7] investigated the effects of truncation on the
synthesized impedance used for hybrid modeling; the role of the Tdofs and Rdofs is
compared, and they concluded that there is a much greater effect of the truncation
in the antiresonances of the FRF of the Rdofs than of the Tdofs.
The consideration of the Rdofs seems to have a strong importance whenever one
wants to validate or correct a numerical or analytical model, using data acquired
experimentally on a prototype. Once the theoretical model is validated, it can be
used to evaluate further project modifications. Moreover, if the modification to be
implemented intends to include rotational inertias [87] or rotational stiffnesses, the
rotational receptances of the unmodified system have to be obtained. Cafeo et al.
[18] also verified that the combined use of translational and rotational vibration
measurements may reduce the number of locations that are necessary to represent
mode shapes with an accuracy similar to the one obtained with purely translational
measurements.
The knowledge of Rdofs proved to be also essential in investigation fields
such as:
1166 N. Maia et al.

• Identification of the dynamic characteristic of joints [79, 153]

• Identification of unknown forces from measured responses [133]
• Vibration isolation [120]
• Analysis and control of vibrational power transmissions to machinery supporting
structures [67, 103]
• Determination of rigid body properties and damage detection [134, 142]
• Vehicle collision studies

As previously referred, the determination of Rdofs is a complex task with

two important problems to solve: (i) the experimental measurement of rotational
motion and (ii) the application of a pure moment to correctly excite the structure,
at some point. The solution to these problems has been the subject of study by
many researchers; since the late 1960s, great advances have been achieved and
multiple experimental techniques have been developed. The principal techniques for
measuring Rdofs and to apply an excitation moment to a structure are next referred,
as well as their main advantages and disadvantages.

3.2 Experimental Methods for Measuring Rdofs

The experimental techniques developed to obtain Rdofs can be divided into two
main groups: the indirect and direct techniques. The first group, as the name sug-
gests, implies that the rotational responses are indirectly obtained from translational
measured ones; the known techniques are (i) the block excitation technique, (ii) the
mass additive technique, (iii) the finite difference technique, (iv) the use of the laser
Doppler vibrometer, (v) the use of PZTs and strain gauges, and (vi) sensors using
piezoelectric accelerometers. Within the second group, corresponding to the direct
techniques, the measurements are obtained directly by the use of special transducers
located at the measurement points. Two types are known: (i) the angular transducers
and more recently the (ii) micro electro mechanical systems (MEMS).
It should be noted that there is another group of techniques, called estimation
techniques based on expansion methods that will not be addressed here, because
they are not considered as purely experimental techniques, since they require a finite
element model of the structure. A brief review of the experimental techniques for
obtaining rotational responses is given in the next section.

3.2.1 Indirect Techniques for Measuring Rdofs

Block Excitation
The use of a block excitation technique to measure the rotational receptances
requires the application of a moment, at a point of the structure, to enable the
measurement of the “rotation due to moment” FRFs. In most situations, the other
FRFs relating translation to moment can be obtained using directly the Maxwell’s
“Rule of Reciprocity,” implying that translation/moment FRFs are identical to the
rotation/force FRFs. Smith [129] proposes a first technique to obtain the complete
21 Toward Robust Response Models: Theoretical and Experimental Issues 1167

v1

d
z
z y
d
v2

z
Accelererometer
Drive
Vibration Wire
Exciters

x
y

Fixture

Test Structure

Fig. 2 Exciter block (adapted from [129])

structural mobility matrix, using a solid block attached at the measurement point
and two shakers connected to that block, allowing for force and moment excitations
(Fig. 2).
In 1972 the work of Ewins and Sainsbury [47] presented one of the most
tested experimental techniques to measure rotational receptances, using a T-block.
This block is rigidly attached to the structure at the location of interest and,
by measuring two translations at two points conveniently chosen, they can be
posteriorly converted into a rotation and a translation, by using simple geometric
relationships. Considering that the rigid T-block is attached at point P (see Fig. 3),
the kinematic relationship between the set of measured translational responses and
the set of estimated translational and rotational responses at point P is as follows:

 ẍA +ẍB
ẍA = ẍ + s θ̈ ẍ = ẍp =
⇒ 2
ẍA −ẍB (20)
ẍB = ẍ − s θ̈ θ̈ = θ̈p = 2s

However, measuring rotations only solves half of the problem; moment excitation
may also be of interest and one can extend the above technique to relate the force fx
and the moment mθ to the responses ẍp and θ̈p . So, performing two run tests with
forces f1 and f2 (within the same frequency range), respectively, applied at the right
and left end of the block, the following relationships can be written [77]:
1168 N. Maia et al.

e1 e2
s s s s
f1 f2
&x&A 1
&x&B 1
&x&A 2
&x&B 2

fx fx
mT mT

P P

Fig. 3 T-block technique to measure the rotational response at point P (adapted from [77])

     
fx = f1 − m ẍp 1 fx = f2 − m ẍp 2
test 1 ⇒   and test 2 ⇒  
mθ = e1 f1 − Ip θ̈p 1 mθ = −e2 f2 − Ip θ̈p 2
(21)

where m and Ip are, respectively, the mass and moment of inertia of the block
(including accelerometers), about point P. The acceleration values ẍp 1 , ẍp 2 ,
   
θ̈p 1 , and θ̈p 2 at point P may be simply related to the measured values through
Eq. (20). The receptance equation relating these quantities is
         
xp 1 Hxx Hxθ fx f1 + ω2 m xp 1
test 1 ⇒   = = H (ω)  
θp 1 Hθx Hθθ mθ e1 f1 + ω2 Ip θp 1
(22)
         
x Hxx Hxθ fx f2 + ω2 m xp 2
test 2 ⇒  p 2 = = H (ω)  
θp 2 Hθx Hθθ mθ − e2 f2 + ω2 Ip θp 2
(23)

and therefore, dividing (22) by f1 and (23) by f2 , one obtains

      
x /f 1 + ω2 m xp 1 /f1
test 1 ⇒  p 1 1 = H (ω)    (24)
θp 1 /f1 e1 + ω2 Ip θp 1 /f1
       
x /f 1 + ω2 m xp 2 /f2
test 2 ⇒  p  2 2 = H (ω)    (25)
θp 2 /f2 − e2 + ω2 Ip θp 2 /f2

By suitable combination of (24) and (25) and taking into account the relation-
ships (20) for (xp )1 and (θ p )1 , and equivalent ones for (xp )2 and (θ p )2 , the receptance
matrix is found to be:
21 Toward Robust Response Models: Theoretical and Experimental Issues 1169

1
H (ω) = − T G( − MT G)−1 (26)
ω2

where
     
0.5 0.5 1 1 (ẍa /f )1 (ẍa /f )2
T = , = , G= (27)
(2s)−1 −(2s)−1 e1 −e2 (ẍb /f )1 (ẍb /f )2

and
 
m 0
M= (28)
0 Ip

The required receptance matrix is obtained in terms of T, Π and M, which

are constant for a given test configuration, together with G, which incorporates
data measured at the same angular frequency ω in the two test runs. The resulting
receptance data have thus been fully corrected to account for the mass and inertia of
the added block and transducers.
A large number of researchers have attempted to apply this technique trying
to obtain the Rdofs, not only to test its performance but also for posterior
application in other studies involving: structural coupling [45, 47, 48, 123, 140],
structural modification [127, 128], complex damped structures [116], and others
like Gleeson [54].
New techniques to obtain Rdofs using the T-block continued to be developed,
like in the works by Yoshimura and Hosoya [157, 158], Maia et al. [78], Montalvão
et al. [88], Silva et al. [125], Mottershead et al. [90, 91], and Mottershead et al.
[91]. In Yoshimura and Hosoya [157] a T-shaped rigid block is attached to the
structure at the measurement point and a set of hammer impacts are applied to
several points on the T-block, resulting in a set of forces and moments applied to
the fixture point, and a set of responses at the accelerometers placed on the T-block.
After that, a numerical procedure makes use of the measured data to fully estimate
the force and moment applied at the fixture point of the T-block, as well as the
translational and rotational responses at the same point. Finally, the FRFs of the
structure at the fixture point are estimated (including Rdofs) without the effect of
the mass and the moment of inertia of the T-block. The same researchers [158]
presented, 2 years later, a new T-block sensor, named T-square-block (T2 -block) to
measure six dofs (three dofs in translation and three dofs in rotation) at some points
of a car body. Maia et al. [78] propose a method based on impedance coupling
techniques, in which rotational/moment receptances are estimated without having
to measure them. It is also shown that neither a moment exciter nor eccentric forces
applied to a rigid fixture (T-block) are needed to estimate rotational receptances. The
structure is excited with a conventional shaker on a single location (a translation
dof) and several sets of response measurements are taken, both in translation and
rotation. This technique works perfectly when exact theoretical data is used, but
fails when directly applied to real cases or with theoretical data polluted with
1170 N. Maia et al.

noise. To help improving the results, modified versions were proposed [88, 125]
based upon coupling techniques, introducing a structural modification by rotating a
T-block, with the force applied at its centroid. Mottershead et al. [90, 91] pro-
posed a technique in which the T-block is treated as a structural modification at
the connection point, deriving the corresponding full force and response vec-
tors. The T-block mass and elastic properties are included in a finite element model.
The receptances are estimated without any mass, inertia, or stiffness effects of the
T-block, which are removed by the T-block finite element model, thus avoiding ill-
conditioning problems.
The application of this technique has, as its main advantage, the possibility of
moment excitation at a given point on a structure and the determination of the
corresponding rotation responses. However, it is an indirect technique, that is, the
rotational responses are obtained from translational ones, which makes it somehow
vulnerable to ill-conditioning. Some of these particular features have been identified
[75, 127, 147] and are listed below:

• The sensor (T-block and accelerometers) represents additional masses attached to

the structure under test. Thus the measurement of angular accelerations on light
structures should be considered carefully so that the mass loading effects caused
by the sensor do not alter significantly the dynamic properties of the structure.
• The sensitivity of the sensor is directly proportional to the length of the upper
arm (Eq. 20); in principle the larger this length, the more sensitive will be the
sensor. However, the sensor has its own dynamic characteristics, and depending
on the test frequency range, the sensor dynamics can affect the dynamics of the
structure.
• Another important issue in the application of this technique is the correct
attachment of the transducer to the structure, such that the behavior of the
structure be transmitted to the transducer faithfully, the same being true for the
transmission of the applied excitations.
• The accelerometers used to instrument the T-block should be matched in terms
of sensitivities; two identical accelerometers should be used reducing the chance
of inaccurate results caused by significant differences in calibration.
• The cross-axis sensitivities of the accelerometers can be very important when
subtracting two FRF signals of similar size, particularly when the motion in
one of the transverse directions is large. In such cases, the difference between
the measured FRFs can be of the same order of magnitude as the cross-
axis sensitivity component. This is the situation that may arise in determining
the rotation/translation, translation/rotation, and rotation/rotation FRFs, where
subtraction of measured translational FRFs is required. To minimize these effects
it is necessary to align the maximum cross-axis sensitivity direction of each
accelerometer with the direction of minimum transverse motion at the attachment
point, something that may not be a simple task.

Another similar indirect technique for measuring the rotational mobility has
been developed by Cheng and Qu [24], Qu and Cheng [105], and Qu et al. [107]
21 Toward Robust Response Models: Theoretical and Experimental Issues 1171

and is called “tip excitation technique (TET)”; instead of a T-block, it uses an

L-shaped beam tip fixed at one point of the original structure, a flexible plane, where
the rotational mobility is to be evaluated. The tip acts as a mechanical converter
transforming an exciting force applied at the tip into an excitation moment applied
to the original structure and reciprocally, to convert a rotational response into a
translational one. The rotational compliance is evaluated using simple analytical
relations that partially eliminate the effect of tip inertia on the structure. In Cheng
and Qu [24] a numerical study is presented, and in Qu et al. [107] the experimental
measurements were performed on a rectangular plate and the results were compared
with the ones resulting from numerical simulations. It is concluded that this
technique presents better results in the low to middle frequency ranges for flexible
plane structures.

Mass Additive Technique

The mass additive technique to estimate rotational dofs at a measurement point was
developed by Yasuda et al. [154]. This technique also requires the addition of a rigid
block to the structure at a point of interest, but this time, unlike the previous one,
the excitation is not directly applied to the measurement block (the added mass), to
produce the moment and/or force excitation. As a consequence there is no possible
estimation of direct FRFs.
The rigid body motion of the added mass is measured using six or more
translational transducers (Fig. 4). After this a least-square procedure is used to
compute the rigid body motions, from which the rotational and translational dofs,
at the attachment point, are estimated. Next, the modal parameters are estimated
using a complex exponential algorithm. Finally, a mass modification procedure is
implemented to compensate for the added mass, that is, the extra mass is analytically
subtracted from the process like in the block excitation technique. Yasuda et al.
[154] applied this method to two structures, a cantilever beam and an automotive
frame; the results obtained for the beam looked promising, but some errors have
been detected for the more complex automotive frame.
This technique has been applied in various and different areas like: structural
dynamic modifications [56, 64], correction of modal scaling factors [150], to
determine the moments and rotations in a structure due to impact excitation [71,
133], and model updating [56].
The need for measuring multidirectional responses is becoming increasingly
important in different applications; several researchers tried to develop transducers
in order to measure the response of the six dofs at a point, using various linear
piezoelectric accelerometers mounted on a light fixture [57, 148] or using as an
acceleration sensor a piezoelectric quartz element [82].

Finite Difference Technique

In the two abovementioned methods (block exciter and mass additive technique)
special fixtures with accelerometers have to be attached to the structure and their
dynamic influence has to be corrected.
1172 N. Maia et al.

Accelererometers

Additive Measured FRFs

Mass

Test Structure
Vibration
Exciter

Application of Mass Additive

Method

Point where rotational dofs

are estimated

Test Structure

Fig. 4 Procedure for the mass additive technique (adapted from [154])

At the beginning of the 1980s, Sattinger [122] developed a method to gen-

erate experimental rotational mobility functions using conventional measurement
techniques without requiring the use of special fixtures. The approach permits to
represent the mobilities involving rotational responses and/or moment excitation
as spatial derivatives of conventional translational mobilities; the derivatives are
approximated as finite difference sums of sets of those translational values. The
method is demonstrated in a set of experiments conducted on a free-free beam;
good agreement was obtained between experimentally and theoretically generated
versions for rotational velocity/force mobilities, although exhibiting a large amount
of scatter in some frequency bands. The author believes that this scatter is due to
the subtraction of nearly equal translational quantities and suggests eliminating this
scatter by the application of a smoothed process, as curve fitting, to the translational
data, before the subtracting operation. The derivation of rotational data could be
accomplished using either first-order [23] or second-order [122] finite difference
formulas. In Sattinger [122], the backward, central, and forward finite difference
formulas are given, the choice between them depends on whether the location where
rotation is desired is an inside, middle, or outside location, or whether it is on the
positive or negative end of the global axis. The formulation of this method can also
be observed in Urgueira [140], Duarte [39], or Lofrano [75].
21 Toward Robust Response Models: Theoretical and Experimental Issues 1173

Fig. 5 Finite difference technique for Rdofs measurements (adapted from [140])

To illustrate the application of this method consider the situation presented in

Fig. 5, where three (or two) accelerometers (according to the finite difference
formula employed) are placed by each other, in a constant spacing s. One of
the accelerometers is located at the point where rotation needs to be deduced,
represented as point P, and the position of the other accelerometers are dependent
on the finite difference formula employed and the direction of the global axis.
The first- and second-order finite difference transformation matrices needed for
the estimation of the necessary rotational FRFs will be introduced next.

Finite Difference Transformation Matrices for First-Order Approximation

If a first-order approximation is considered, only two location sites have to be taken
for the estimation of the rotational parameters, for instance, points B and C in Fig. 5.
If P = B and the coordinate system is the one shown in the same figure, or if P = C
and the coordinate system is in the opposite direction, the first-order forward (f )-
difference transformation matrix is expressed as:
 
0 1
T 1f = (29)
1/s −1/s

If P = C and the coordinate system is the one assumed in Fig. 5, or if P = B

and the coordinate system is in opposite direction to the assumed one, the first-order
backward (b)-difference transformation matrix is expressed as:
 
0 1
T 1b = (30)
− 1/s 1/s

The accuracy of the matrices in (29) and (30) is equivalent.

1174 N. Maia et al.

Finite Difference Transformation Matrices for Second-Order Approximation

If three location sites are considered, three different formulas can be employed
depending on the accelerometers and coordinate system positions, as shown in
Fig. 5. The second-order forward (f )-, central (c)- and backward (b)-difference
transformation matrices [23] are, respectively:

 
1 0 0 2s
T 2f = (31)
2s − 1 4 −3

if P = A with the coordinate system assumed in Fig. 5 (or if P = C with the

coordinate system is in the opposite direction), or

 
1 0 2s 0
T 2c = (32)
2s −1 0 1

if P = B; in this case the assumed coordinate system has no importance. Finally,

 
1 0 0 2s
T 2b = (33)
2s 1 −4 3

if P = C with the coordinate system assumed in Fig. 5 (or if P = A and the coordinate
system is in the opposite direction). All the transformation matrices for the second-
order approximation are of equivalent accuracy.
The set of desired FRFs related to translational and rotational coordinates, for a
specific point, are determined solving the following equation:

 
H yy (ω) H θy T (ω)
H est. (ω) = = (T ? (ω)) (H meas. (ω)) (T ? (ω))T (34)
H θy (ω) H θθ (ω)

where Hmeas. (ω) contains the values of translational FRFs measured for each
frequency. The subscript “?” is explained after eq. (36). The size of this matrix is
related to the approximation employed and symmetry should always be considered.
If a first-order approximation is chosen, two measuring points are needed, in this
case points B and C, assuming that the measured matrix is symmetric, and thus,

 
HBB HBC
H meas. = (35)
sym HCC

Using the same assumptions for the second-order formulation, three measure-
ment points (A, B, and C) are required, so at least six FRFs have to be measured to
obtain the Hmeas. matrix.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1175

⎡ ⎤
HAA HAB HAC
H meas. = ⎣ sym HBB HBC ⎦ (36)
sym sym HCC

The subscript ? used in matrix T? (ω) aims to emphasize that any one of the
matrices in (29), (30) or (31), (32), (33) could be used depending on the order
approximation used, first or second, and the point where the rotation needs to be
derived. The formulation presented in Eq. (34) is generally referred as FRF-based
approach and was used in Urgueira [140], to evaluate the use of different transducer
spacing to estimate matrix in Eq. (34) in a short and a long beam. This evaluation
was made simultaneously with other ones, either using the T-block approach or the
data from a continuous measurement with a laser transducer.
Substituting Eqs. (35) and (29) or (30) in Eq. (34), the first-order forward (f )-
and backward (b)-difference approximations can be obtained, resulting respectively
in:
 
HCC − 1s (HCC − HCB )
(H est. )1f = (37)
sym. s12 (HCC − 2HCB + HBB )

and
 
s (HCC − HCB )
1
HCC
(H est. )1b = (38)
CC − 2HCB + HBB )
1
sym. s2
(H

A similar result occurs for the second-order forward (f ), central (c), and
backward (b) approximations. The development of Eq. (34) using matrix (36) and
transformation matrix (34), (32) or (33) leads to:
 
HCC − 2s
1
(HCA − 4HCB + 3HCC )
(H est. )2f =
sym. 1
4s 2
(HAA − 8HBA + 6HCA + 16HBB − 24HCB + 9HCC )
(39)
 
2s (HCB − HBA )
1
HBB
(H est. )2c = (40)
sym. 1
4s 2
(HAA − 2HCA + HCC )
 
2s (HCA − 4HCB + 3HCC )
1
HCC
(H est. )2b =
sym. 1
4s 2
(HAA − 8HBA + 6HCA + 16HBB − 24HCB + 9HCC )
(41)

Martinez et al. [85] suggest a similar process to the one proposed by Sattinger
[122]; however, instead of using experimental translational FRFs directly (FRF-
based approach), they use the modal model referred to the measured coordinates,
generally derived from a row (or column) of the measured FRF matrix over a
1176 N. Maia et al.

selected frequency range encompassing k modes. The information related to the

effects of the out-range modes has to be synthesized in a residual matrix HResid. .
Therefore, the measured response and modal models are related by
 −1
H meas. (ω) = φ kmeas ‘
λk 2 − ω 2 ‘ φ Tkmeas + H Resid.meas (42)

In order to use the finite difference approach to the translational modal parame-
ters and to avoid modal truncation problems, a similar procedure has to be employed
to the residual matrix as well. This procedure, named Modal-based approach can
be summarized as follows:
  −1
H est. (ω) = T ? φ kmeas ‘ λk 2 − ω2 ‘ φ Tkmeas + H Resid.meas T ? T
  −1  
= T ? φ kmeas ‘ λk 2 − ω2 ‘ φ Tkmeas T ? T + T ? H Resid.meas T ? T
(43)

The main advantage of the Modal-based approach in relation to FRF-based

approach proposed initially by Sattinger [122] is that the noise is removed from
measured data and therefore smoother curves are obtained. This research approach
is also used and reported by Maleci and Young [83] and Urgueira [140], where the
modal-based approach produces better predictions than the FRF-based one.
Comparing the finite-difference technique (both approaches) with the block
exciting technique, two main advantages can be easily identified:

• No special apparatus is required, only the transducers normally used in standard

modal testing.
• The mass and stiffness properties of the original structure are not altered by the
use of an excitation block (T-block).

However, some problems, summarized below, were identified when it is

applied:

• The quality of the predictions is dependent on the spacing and order used in
the approximation, and these two points are inter-related. According to Urgueira
[140], increasing the approximation order for the rotation/force FRF requires
a smaller spacing between the transducers, but increasing the order for the
rotation/moment FRF requires a larger spacing. So, the definition of the ideal
space between accelerometers is not an easy task, with a linear answer.
• The first-order approximation produces better results for the rotation/force
FRFs, while the second-order approximation produces better results for the
rotation/moment FRFs [41].
• The direct use of measured translational FRFs in the FRF-based approach
normally yields unacceptably noisy rotational FRF data, because relatively small
21 Toward Robust Response Models: Theoretical and Experimental Issues 1177

errors in the measured data may result in large errors in the estimated responses,
due to the fact that very similar quantities are subtracted. The use of regenerated
FRF curves including the effects of the out-range modes (residual compensation)
is recommended to solve this problem [41, 140].
• It is advisable to use similar accelerometers (mass and sensitivity) with small
dimensions and mass.

Elliott et al. [43, 44] study the effects of random and bias errors when central dif-
ference finite approximation technique is used to determine the moment excitation
and moment mobilities in plate and beam structures.
Liu et al. [74] present an enhanced stiffness identification method for the spindle-
tool holder joint, in which the Rdofs are included. The FRFs associated to the
Rdofs are determined using a finite-difference technique, to depress the influence
of “modal truncation” and measurement noise; residual compensation theory is
introduced to regenerate the Rdofs. An obvious improvement in the results was
observed when residual compensation was used.

Laser Doppler Vibrometer

The use of the laser Doppler vibrometer (LDV) [37, 156] brought to experimental
structural vibrations analysis new possibilities of measurements, impossible to
obtain with the traditional vibration transducers (such as accelerometers, strain
gauges), in particular measurements: on extremely lightweight objects when the
local mass loading of the sensor may distort the results, or when fluid flows
are studied and one wants to avoid disturbing the flow itself by the introduction
of contacting transducers on structures, on hot temperatures, or even in rotating
components, etc.
The laser Doppler vibrometer (LDV) is an optical velocity-measuring device
based on the measurement of the Doppler shift of the frequency of laser light
scattered by a moving object [84].
The advantages in the use of laser vibrometers are related to the fact that:

• The noncontact nature of the instrument makes it particularly attractive for

the use in light-weight structures where the measurement interaction must be
minimized.
• No loading effects at all have to be taken into account.
• The measurements can be done far from the vibrating structure, on hot or
cool surfaces, in strong magnetic fields, without all the problems that a normal
accelerometer would face.
• It is possible to measure a quasi-continuous line or area on the structure, thus
reducing the degree of incompleteness in terms of coordinates, which sometimes
is a disadvantage of experimentally derived models when compared with the
theoretical ones.
• The measurements are relatively fast and easy.
• The resolution and the accuracy of the measurements are high.
1178 N. Maia et al.

• The cost of the equipment is decreasing, and the possibility to measure many
points in a very short time is a very much appreciated quality.

A drawback could be the fact that the output of the laser is normally velocity and
not acceleration.
The use of laser vibrometers in the measurement of vibrations begins in the
1980s, but their limited sensitivity and low signal-to-noise ratio allowed measure-
ments only on very diffusive surfaces or by applying a retro-diffusive tape at the
point of measurement; despite the referred difficulties, Oliver [98] refers to its
application as a promising technique for the determination of dynamic Tdofs and
Rdofs of a structure. In a work by Urgueira [140] a continuous measurement with
a laser vibrometer was made on a segment of a long and a short beam, where three
accelerometers and a T-block were also attached, in order to carry out an assessment
of the various techniques to predict the rotational information associated with the
receptance matrix.
In the early 1990s, hardware and software developments increased the perfor-
mance of the instrumentation and many researches started using LDV to obtain the
Rdofs.
Cafeo et al. [17] developed a non-contacting measurement approach, capable
of simultaneously measuring one dynamic translation and two dynamic angular
rotations (see Fig. 6). The transducer system is based on the positional measurement
of two collimated light beams reflected from a planar reflective target. The light
beams are created by lasers, and the locations of their respective reflections are
measured with two separate two-dimensional photo-detectors. Using the geometric

Fig. 6 Three degrees of

freedom laser vibrometer
(adapted from [17])
21 Toward Robust Response Models: Theoretical and Experimental Issues 1179

orientation between the origins of the light beams and the corresponding reflections
onto the photo-detectors, vertical translation (z) as well as roll (θ x ) and pitch
(θ z ) angular deflections of the planar target can be determined. The time-varying
x–z coordinates’ signals from the photo-detectors are digitized and then processed
to estimate the three variables. The algorithm used to extract the variables from
the geometric layout and photo-detector signals is explained in Trethewey et al.
[139]. An evaluation of the transducer dynamic performance in a modal test
environment of two structures, a cantilever beam and a printed circuit board,
demonstrated the ability of the system to produce high-quality time and spectral
data. The researchers used the obtained structural rotation data to refine the spatial
description of mode shape, and verified that a significantly improved definition of
the mode shape spatial characteristics was observed by integrating the translational
and rotational modal deflections at each response measurement site [18]. Using the
same technique to measure rotational data, Cafeo et al. [19] presented an SDM
study with a beam element, where the modification involves the fixation of the free
end of one cantilever beam; the measured rotational data are used as a database
for SDM. The experimental results obtained with a real beam with dual fixed-
fixed boundary conditions were compared with the SDM predictions using the
experimental cantilever database, exhibiting a good correlation.
A novel non-contacting device to measure the same three dofs of vibration
(translation, pitch, and roll) was proposed by Sommer et al. [130], utilizing a
single laser beam reflected by a planar reflective surface target onto two transparent
two-dimensional position-sensing photo-detectors (see Fig. 7). The advantage of
this system over the original dual-laser beam setup [17] is the need for less hardware
and smaller overall size.
Bokelberg et al. [13] presented a new laser system to simultaneously measure
three translational (x, y, z) and three rotational (θ x , θ y , θ z ) displacements of a
vibration object. The system is based on sensing the positions of the three collimated
light beams reflected from a tetrahedral target attached to the vibration object.
The dynamic position of the target is determined by using the system geometry,
coordinate transformations, and kinematic closure procedures. The operational
principle and experimental evaluation of this measurement technique are presented

Fig. 7 Three degrees of

freedom laser device (adapted
from [130])
1180 N. Maia et al.

Fig. 8 Laser system to measure 6 dofs (adapted from [13])

in Bokelberg et al. [14, 15]. Figure 8 shows the conceptual design for this six
degrees-of-freedom vibration measurement system. The disadvantages observed
with the use of this system are: the laser system set-up must be moved to each of
the desired locations for measurement, which is very time consuming when a large
number of data points have to be gathered, and as a consequence of its dimensions
the system is hardly portable.
The estimation of rotations from translational data using fitting functions was
the approach chosen by various researchers. Two groups of fitting functions have
been particularly studied: polynomials [140, 151] and splines [93, 94]. These
techniques have shown considerable potential. The polynomials proved to be ideal
for representing uncomplicated relationships. The main disadvantage, however, is
the inability of the polynomial fits to produce acceptable results for the higher
frequency modes; the results from tests show that the accuracy of the estimated
Rdofs depends on the density of the original displacement measurements and also
that the estimation improves with an increase in the number of measurement points
between nodes. The splines, being composed of sets of polynomial pieces, turned
out to be more flexible than single polynomials and capable of representing very
complex relationships. The accuracy of the estimates of rotations from translational
data, when spline approximations of the mode shape functions are used, depends on
several factors, namely the combination of the smoothing factor and the number of
knots in the fit function, and density and distribution of the measurement points on
the structure.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1181

Ratcliffe and Lieven [108] present a technique that uses the simple expedient
of fitting a plane, in a least-square sense, to experimental translation data, to
calculate the two out-of-plane rotations. The accuracy of the method increases when
the distance between measurement points decreases, so the use of laser Doppler
techniques proved to be ideal since they have the capacity of measuring a structure
using a finer grid because of their speed of operation.
In 1994 two new techniques for measuring Rdofs were presented [131], using
a Scanning Laser Doppler Vibrometer (SLDV); in these techniques a traditional
point-by-point scanning procedure was substituted by a continuous scan along a
line or around a circle on a harmonically vibrating surface, giving a modulated
output which can be used to analyze structural vibration in more than one dof.
Using continuous linear scanning in one dimension it is possible to extract the
translation; and one angular vibration component is extracted while using circular
scanning; two components of angular vibration can be derived directly from the
frequency spectrum sidebands, together with the translational dofs. Interesting
researches conducted by Ziaei-Rad et al. [159] and Martarelli [84] compare the
experimental results obtained with the application of three techniques to obtain
Rdofs on a coupled structure. Those techniques are the discrete point measurements,
the continuous linear scanning, and the continuous circular scanning. The most
important conclusions drawn from those works are summarized in Ziaei-Rad et al.
[159]. It was found that:

• The linear and circular scan techniques offer advantages over discrete scanning
in terms of time saving.
• Better results are obtained in the linear scan when short lengths are measured.
• The effects of changing radius, in the tried range, are not very significant.
• The FRF resonance peaks drop slightly when the frequency of scanning is
increased, because the speckle noise always affects laser Doppler measurements.

Another technique, using Continuously Scanning Laser Doppler Vibrometer

(CSLDV), for six dofs was presented by Stanbridge and Ewins [132] and named
conical-scanning LDV. This technique requires performing a small radius circular
scan with an SLDV on a focal lens that focuses the scanning laser beam to a
point and leads to a conical scan. Two measurements have to be performed: (i) one
focusing on the conical scan at a measurement point and (ii) another one intercepting
that same point with a very small circle. Five of the six dofs are easily obtained:
the three translations and two out-of-plane rotations. However, the lens makes the
measurement of the in-plane rotation very difficult and impractical.
Giuliani et al. [53] developed a new type of scanning device, allowing the
measurement of all six dofs at a given location with a single setup and a single point
LDV, by combining variable small circle scans. The main aim of the research was to
combine the two separated measurement approaches, performed on the previously
referred technique, into a single system, based on a single point LDV in combination
with a newly developed scanning head, and add the measurement of a missing in-
plane rotation around a point, to provide a measurement technique that can resolve
1182 N. Maia et al.

the dynamic response of all six dofs at a point on the structure. Experimental results
have shown the ability of the new scanner to distinguish between the different dofs
[53] and to measure the out-of-plane deflection shapes of a specimen, bending along
the three axial directions [82].

PZTs and Strain Gauge Transducers

A group of researchers developed a new technique using low-cost PZT transducers
to estimate rotational quantities [10]; special reference is made to bimorph patches
that are able to measure the local curvature of a structure. They provide curvature
measurements from which, by interpolation techniques (single polynomial, third
degree polynomials, or using the analytical modes), rotations can be determined.
Despite the use of these sensors being quite simple, inexpensive, with a minimum
intrusiveness due to their low weight and size, it has proven the necessity for
solving a series of problems related to their amplification, bonding, and assembly
to the structure, calibration, and cross-sensitivity. If all of these items are carefully
considered, the results obtained using a supported beam proved to be very promising
when compared to those obtained by a rotational accelerometer and by finite
models.
These kinds of transducers, bimorph PZTs, present an important advantage
over strain gauges in estimating curvature [66], as no geometrical and physical
properties of the cross section are needed. Differently from strain gauges, bimorph
PZT transducers are directly sensitive to curvature: the curvature estimation is
independent of the structure to which they are attached, so that no information about
the structure under test is needed.
Cicogna et al. [25] also studied the use of PZT patches (bimorph) to obtain the
local curvatures on a cantilever beam and a free-free plate, and from it the rotational
quantities, by interpolation techniques. A methodology based on genetic algorithms
was also presented, considering the possibility of including more than one patch to
find the optimal size and position of the bimorph onto beam-like structures.

3.2.2 Direct Techniques

Angular Transducers
Systron-Donner Corporation under contract to the Air Force Weapons Laboratory
developed an angular instrumentation system [89], using a proper angular displace-
ment sensor. This system uses inertial techniques to measure three axes of angular
displacement with extreme accuracy over a very wide frequency range (1–2000 Hz),
with no need for temperature.
Two other techniques, used on flight control and navigation, for measuring Rdofs
data could be referred: the first one is based on the use of magneto-hydrodynamic
sensors [136] and the second one based on the use of gyroscope sensors [1];
however, the large dimensions of these sensors prevent their use in most vibration
applications.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1183

Micro-Electro-Mechanical-Sensor (MEMS)
MEMS inertial sensors, comprising acceleration and angular rate sensors [4, 92]
(i.e., micro-accelerometers and micro-gyroscopes (based on Coriolis angular rate)),
are used in a wide range of areas [76, 134, 142] such as in automotive industry appli-
cations (for crash air-bag deployment systems), consumer electronics applications
(such as Apple iPhones, Nintendo Wii, computers), and medical applications. They
are able to measure the responses associated to all six dofs, in one go, as they can
easily measure, simultaneously, three rotations and three translational motions.
A MEMS sensor is produced by micromachining techniques to form minute
springs, seismic masses, and motion or force sensing elements from a silicon wafer.
When the body of the accelerometer is moved by an externally applied force, the
motion of the seismic mass is detected by differential capacitive, piezo-resistive, or
other types of sensing elements. The signal produced is amplified, conditioned, and
filtered by circuit components mounted inside the same IC package [155].
Certain types of silicones may show characteristics that reveal that they can be
stronger than steel, though with only a third of the weight, allowing in recent years
the building of sensors with dimensions of just a few millimeters. This is considered
a great advantage, because the small size and very low weight give the possibility to
insert transducers in spaces considered inaccessible for other kinds of instruments,
and also to install them on smaller objects, because of the lower insertion effect, as
it happens with the smartphones. It should also be stressed that the MEMS devices
are cheaper when compared to other available sensors, so all these explain why the
interest on MEMS technology is so high.
In the beginning, their use was limited to measurements considered of “low-
quality,” but during the last years MEMS technology has quickly improved,
leading to sensors with higher qualities that supposedly can be used in rigorous
measurement applications. Their utilization has been introduced in measurement
fields, like modal analysis [11, 26, 68] or structural health monitoring [27]. Cigada
et al. [26] draw attention to the fact that the metrological limits of this type of
sensors are not yet well defined, and the reason is mainly due to testing costs: these
sensors are so cheap that any kind of test other than gravity sensitivity check is more
expensive than the sensor itself, so a true calibration for this kind of sensors does
not exist so far. In order that their use becomes widespread, in those research fields,
it is important to investigate and clarify the following:

• The sensor performances in measuring very low accelerations (low frequency

band), where the signal to noise ratio is a critical feature
• Transverse vibration sensitivity
• Intrinsic noise
• Temperature influence

Direct Piezoelectric Rotational Accelerometers

In the last decade the first piezoelectric rotational accelerometers appeared in the
market, capable of measuring directly angular accelerations [42]. However, their use
in the field of experimental structural dynamics is rare when compared to the classic
1184 N. Maia et al.

translational accelerometers, mainly due to their high cost and mass modification
effect on lightweight structures.
Drozg et al. [38] evaluate the performance of two direct piezoelectric rotational
accelerometers to show the possibilities of their usage in structural dynamic
applications. They concluded that the force-excited rotational responses measured
with those two accelerometers are more accurate and reliable, compared to the
indirect rotational response obtained via the T-element.

3.3 Experimental Methods for Exciting Rdofs

The excitation of structures with a pure, sufficiently powerful and well-controlled

moment is a task very difficult to achieve in practice. The development of techniques
that make possible the application of moments in structures has been taking
place simultaneously with the development of techniques for the measurement of
rotational dofs. As it happens to the techniques for measuring rotational dofs, the
ones for excitation can also be divided into indirect and direct techniques. The
indirect ones deduce the moment mobilities from measured force mobilities, while
the direct techniques use a mechanical system to create a force couple to excite the
structure under test. Further reviews of these techniques can be found in Bregant
and Sanderson [16], Elliott et al. [44], and Mottershead et al. [91]

3.3.1 Indirect Techniques to Apply a Moment Excitation

The earliest attempts to measure moment receptances were indirect approaches,
which made use of supports with different forms [45, 47, 107, 122]. Such supports
are small rigid bodies, of accurately defined geometry, which are attached to the
structure under study; the support is excited by one or various forces in different
directions and/or positions. The moment excitation is determined based on the
geometry of the block. Theoretically, the indirect approach poses no problem, but
in practice none of these techniques succeed in applying easily a single moment
component, and some of them require a complex post-analysis to extract the
different excitation components.

3.3.2 Direct Techniques to Apply a Moment Excitation

The limitations observed in the application of indirect techniques led the scientific
community to develop new procedures that allow the application of a pure moment.
As referred in Sect. 3.2.1, Smith [129] proposed a first known technique to obtain
the complete structural mobility matrix applying a crude moment exciter using two
electromagnetic shakers in a configuration capable of applying a couple by using
two equal and opposite forces applied to a special fixture (Fig. 2), but its acting
frequency range was quite limited. Several other techniques have been proposed to
apply a pure moment:

• Twin-shaker arrangements [119, 121]

• Synchronized hammers [22]
21 Toward Robust Response Models: Theoretical and Experimental Issues 1185

• Magneto-strictive exciters [60, 61, 102]

• Using centrifugal forces [137]

Let us make a brief summary of these methods:

Sanderson [119] constructed a moment exciter with twin shakers for the measure-
ment of beam moment mobility. To apply the moment, two two-exciter methods are
implemented using two different fixtures in the form of T-block and I-block. Their
works include a theoretical study [121] and an experimental one [119] on potential
bias errors, where two main types are identified. The first bias error is an error
in the measurement of the moment acting on the structure, due to the rotational
inertia, and it was verified that this kind of error increases with the frequency and
is most prominent at resonance frequencies; the correction of this error could be
easily applied if rotational inertia cancellation is applied. The second error is a bias
error in the rotational velocity that can be seen to be composed of two separate
parts: the sensitivity of the measurement object to false excitation and the quality
of excitation. To correct these errors several proposals have been presented. A
special prototype moment exciter based on magneto-strictive rods was developed by
Petersson and has been described in Petersson [102]. Jianxin and Gibbs [60] studied
the effectiveness of the two different twin-shaker configurations and the Petersson
magneto-strictive actuator in the application of a pure moment in beam structures.
They concluded that using a twin-shaker system it is not possible to achieve a pure
moment excitation, the reason being that the magnitude of the forces exerted by
electromagnetic shakers depends upon the structural response of the system being
excited. Where cross mobility is non-zero, results reveal themselves unreliable,
and it is difficult to estimate the discrepancy as the shakers are non-matching. The
Petersson magneto-strictive rods actuator seems to avoid the problems of matching
and setting up in the twin-shaker systems. The same conclusions were drawn by
Jianxin and Mak [61] on their experimental research using a magneto-strictive rods
moment exciter, where they proposed methods to improve the moment exciter to
generate a pure and sufficiently powerful moment.
The method of moment excitation by synchronized hammers Champoux et al.
[22] uses two impact hammers to apply a pure moment by simultaneously exciting
at points at either side of a point of interest but in opposite directions. The main
advantage of this method is that no additional fixtures are needed, since the two
forces are applied on each side of the point of interest; however, this is only true if
both sides of the test structures can be easily accessed, which often is not the case.
If only one side of the structure is available, an additional T-block or I-block has to
be used to transmit the couple.
Trethewey and Sommer III [137, 138] presented a device that allows the
generation of a pure moment by the centrifugal forces generated by eccentric masses
attached to two symmetrically connected rotating wheels (see Fig. 9). However,
this system also generated an undesirable moment in an orthogonal direction; to
cancel it, a second identical counter-rotating system of masses was added on top
of the previous one. The major drawback of the device is its considerable weight,
although other restrictions have been identified, such as the narrow frequency range,
1186 N. Maia et al.

Fig. 9 Device for the

generation of a pure moment
by the centrifugal forces
generated by eccentric masses
(adapted from [138])

the limitation to harmonic excitations, and problems related to the fixture to the
structure to be studied.
The error analysis previously performed, in the two-force configuration exci-
tation arrangements, was mainly focused on the initial load of the exciters, the
matching of the two forces, and the system resonance. Tao and Mak [135] also
investigated the distance requirement of two-forces configuration for moment
excitation.
In addition to the specific limitations of each one of the above-described methods,
a main drawback is common to all of them: the large space occupied by the
arrangements and the weight, which makes them unsuitable for use in many real
applications.
For everything that has been said in this subsection, it is concluded that despite
the fact that a lot of research work has already been accomplished, much more is
required in order to create an effective system for applying a truly pure moment.

4 Condensation (Reduction) Versus Expansion

As the experimental data is usually collected from few coordinates and in a limited
frequency range, when the comparison of numerical and experimental data is
needed, the problem of model incompleteness arises. Hence, it is inevitable to deal
with model incompleteness. Several methods have been proposed [77] and are being
used in order to reduce the theoretical model to the test points or to expand the
experimental data over the numerical model dofs.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1187

4.1 Model Reduction

The general equilibrium equation for a given structure is

M ẍ + C ẋ + Kx = f (44)

If one neglects damping and defines primary p and secondary s coordinates, Eq.
(44) can be partitioned as
       
M pp M ps ẍ p K pp K ps xp fp
+ = (45)
M sp M ss ẍ s K sp K ss xs 0

Note that forces can only act on the primary coordinates and that this subset
is composed of a given number of dofs to be retained after model reduction. The
selection of the retained dofs can be challenging and must be in such a way that the
mode shapes can be described as well as possible.
The following subsections present some of the most referred reduction methods,
with special emphasis on the transformation matrix proposed by each one. Detailed
reviews on model reduction techniques have been published by Qu [104], Koutso-
vasilis [69], Besselink et al. [12].

4.1.1 Guyan Reduction

Introduced by Guyan [55], the Guyan or static reduction technique can be found in
several commercial computational applications, although its accuracy is limited to
the lower order modal properties, as it neglects the inertial terms of Eq. (45). Hence,
one can state that

K sp x p + K ss x s = 0 (46)

and the transformation matrix of the Guyan reduction TS is given by

 
I
TS = (47)
− K −1
ss K sp

Using the transformation matrix TS , the spatial model coordinates can be related
to the primary ones by using the following transformation
 
xp
= T S xp (48)
xs

Consequently, the reduced model matrices are given by

M R = T TS M T S and KR = T TS K T S (49)
1188 N. Maia et al.

4.1.2 Dynamic Reduction

As the inertial terms were neglected in the formulation of the transformation matrix
TS , the dynamic response of the reduced model is obviously exact only in static
conditions. Moreover, as the excitation frequency increases, the accuracy of the
reduced model responses tends to degrade. Thus, the Guyan reduction technique
can be extended in order to reproduce the exact response of a structure at any
given frequency ω [100]. This extension assumes harmonic excitation and it results
on the modification of Eq. (47) in order to include the inertial effects at the
selected frequency ω, so that the transformation matrix of the dynamic reduction
technique is
 
I
TD =  −1   (50)
− K ss − ω M ss
2 K sp − ω2 M sp

and, consequently, the reduced model matrices are obtained similarly to Eq. (49).
Qu and Fu [106] present a dynamic reduction technique based on the subspace
iteration method in the eigenproblem. The transformation matrix of this iterative
technique is derived from the ith approximation of the first m mass-normalized
(i)
eigenvectors m , such that
   
(i)
pm I
(i)
= =   (i)
m (i) (i) (i) −1 pm (51)
sm sm pm

(i+1) (i+1)
If one considers the subspace approximation X m of m , defined as

(i+1)
Xm = K −1 M m
(i)
(52)

the reduced system matrices are given by

 T  T
M R = Xm
(i+1) (i+1)
M Xm and K R = Xm
(i+1) (i+1)
K Xm (53)

4.1.3 Improved Reduction System (IRS)

The improved reduction system (IRS) technique, proposed by O’Callahan [95],
develops Guyan’s technique to include the contribution of the secondary coordi-
nates’ inertias as pseudo-static forces, which generates the transformation matrix
TI ,

T I = T S + S M T S M −1
R KR (54)

where TS is the transformation matrix of the Guyan reduction technique and MR

and KR are statically reduced mass and stiffness matrices, respectively, and
21 Toward Robust Response Models: Theoretical and Experimental Issues 1189

 
0 0
S= (55)
0 K −1
ss

Hence, the reduced model matrices are generated as in Eq. (49).

The transformation of Eq. (54) is a perturbation on Guyan’s reduction technique
and therefore some of its limitations remain, namely being exact only in static
conditions. Friswell et al. [50] defined the IRS technique based on a perturbation
on the dynamic reduction, where the transformation is exact for a selected non-
zero frequency. This reduction technique is the dynamic IRS (DIRS) and its
transformation matrix is:

T DI = T D + S D M T D M -R1 K R (56)

where
 
0 0
SD =  −1 (57)
0 K ss − ω2 M ss

The IRS technique was also modified by Friswell et al. [50] in order to
improve its accuracy and is known as the iterated IRS (IIRS) technique, where the
transformation matrix TI is replaced by
 
(k) (k−1) (k−1) −1 (k−1)
T II = T S + S M T II MR KR (58)

(0)
where the superscript (k) denotes the kth iteration, for k > 1. Note that T I I = T S ,
(1)
T I I = T I is given in Eq. (54) and the reduced model matrices are given by
   
(k) (k) T (k) (k) (k) T (k)
MR = T II M T II and KR = T II K T II (59)

Further details on this technique are given in Friswell et al. [50, 51].
Xia and Lin [152] proposed a modification to the IIRS technique that leads to the
following transformation matrix:
 
(k) (k−1) (k) −1
T III = T S + S M T III MD KS (60)

with

K S = T TS KT S (61)

and
(k) (k−1)
M D = T TS MT I I I (62)
1190 N. Maia et al.

(0) (1)
Note that T I I I = T S and T I I I = T I , as for the IIRS, and the reduced
model matrices are obtained similarly to Eq. (59). This technique proved to be
computationally more efficient than the base IIRS [152].

4.1.4 System Equivalent Reduction Expansion Process (SEREP)

The system equivalent reduction expansion process (SEREP) was introduced by
O’Callahan et al. [97], although a similar technique was proposed by Kammer [62].
This technique requires the solution of the generalized eigenproblem

 
K − ω2 M X = 0 (63)

in order to obtain the modal model defined by the natural frequencies and mode
shapes. The computed analytical eigenvectors must then be partitioned so that
 A = [ Ap  As ]T , where the subscript A refers to the analytical model. Fur-
thermore, the Moore-Penrose pseudo inverse of  Ap is used to map the complete
analytical eigenvectors onto the secondary dofs by means of the transformation
matrix TSE ,

 
 Ap  T −1
T SE =  Ap  Ap  TAp =  A  +
Ap (64)
 As

and the reduced model matrices are obtained as in Eq. (49). Note that if the
eigenvectors are mass-normalized, the reduced model matrices can be efficiently
computed, regarding the orthogonality conditions, as

 T  T
M R = +
Ap +
Ap and K R = +
Ap  +
Ap (65)

where  is the mass-normalized modal matrix and  is the diagonal matrix of the
squared eigenvalues.
Due to ill-conditioning issues, SEREP can only be applied if the number of
primary dofs is greater or equal to the number of eigenvectors. However, as it can
be derived from the SEREP formulation, the reduced model has exactly the same
eigenfrequencies and eigenmodes as the full model for the modes used to form TSE .
Note that the IIRS technique proposed by Friswell et al. [50] converges exactly to
the same solution of SEREP, avoiding the extraction of modal properties.
Kammer [63] proposed a hybrid reduction technique that combines both Guyan
and SEREP formulations. The transformation matrix of this hybrid reduction
technique is

 
T H = T SE + T S I - Ap  +
Ap (66)
21 Toward Robust Response Models: Theoretical and Experimental Issues 1191

4.1.5 Modal Truncation

Considering the modal matrix , one can derive the system matrices in modal
coordinates, such that

m =  T M and k =  T K (67)

respectively, the modal mass and stiffness matrices. Note that if the mass-normalized
modal matrix is used, Eq. (67) is recast as

T M = I and T K =  (68)

In the modal domain, the equation of motion is completely decoupled, and

therefore, it constitutes a set of linearly independent equilibrium equations related
to each eigenvalue. Hence, one can obtain a reduced model just by truncation
of the modal system matrices at a given mode of interest. Thus, for a truncated
modal system defined by R and R , one can obtain the following reduced model
matrices:

 −1  −1
M R = TR -1
R and K R = TR -1
R R (69)

Note that the reduced mass and stiffness matrices can be directly obtained from
the experimental modal data, after normalization. If damping is not neglected, a
second-order modal truncation must be considered, as in Koutsovasilis [69].
The use of modal truncation is the base of an efficient approach to the assembly
of substructures or super-elements, the modal-based assembly (MBA), incorporated
in the commercial software FEMTools (FEMtools 3.7.0, 2013).

4.1.6 Component Mode Synthesis

Developed as a modal coupling technique, one of the most accurate, although
expensive, model reduction techniques is given by the component mode synthesis
(CMS) methods. These methods attain a reduced order model for the substructures
of a given assembled structure by reducing the number of mode shapes used to
describe the dynamics of each substructure, while preserving all the physical dofs.
A general CMS algorithm can be framed into three main categories: the fixed-
interface, the free-interface, and the residual-flexible free interface [29, 69, 77,
141]. The choice for a specific CMS is problem dependent, although in structural
mechanics the fixed-interface method is usually recommended, as it accurately
retains the lower to mid frequency eigenvalues. Among several CMS hypothesis,
the Craig-Bampton (CB) method [8] is perhaps the most used and therefore it is
presented here.
Regarding a fixed-interface CMS method, Eq. (45) can be recast in terms of
connection c and interior i dofs, as
1192 N. Maia et al.

       
M cc M ci ẍ c K cc K ci xc fc
+ = (70)
M ic M ii ẍ i K ic K ii xi fi

where xc and fc are the displacement and force vectors associated to the connection
dofs between substructures, respectively; and xi and fi are the ones related to the
interior dofs. Note that the connection dofs are actually the primary ones, in the
sense of what was previously described.
The fixed-interface CMS algorithm is based on the assumption that the secondary
dofs can be expressed in terms of both elastic and static mode shapes, considering
that the connection dofs are fixed (xc = 0) and no forces are acting at the interior dofs
(fi = 0). At a first stage, one may approximate the vector of interior displacements
xi by a linear combination of the im elastic fixed-interface modes or CB modes,
given by:

x i = im pm (71)

where pm is related to the number of calculated modes (at initial stage m = i).
However, the underlying philosophy of this reducing technique is to retain less
modes, say k modes leading to a reduced modal matrix ik while preserving all
the physical dofs, to describe the dynamics of each substructure.
On the other hand, the vector of interior displacements xi can also be approx-
imated by using a linear combination of the constraint or static modes. In the CB
method, these static modes are in fact given by the Guyan reduction, as it can be
inferred from the manipulation of Eq. (70). Hence, one can relate the displacements
at the interior coordinates to the connection ones, through

x i = −K −1
ii K ic x c = T S x c (72)

Thus, connection and interior dofs can be related by the transformation matrix of
the CB method TCB , derived from Eqs. (71) and (72), as follows:

      
xi T S ik xc xc
= = T CB (73)
xc I 0 pk pk

Consequently, the reduced model matrices, for the k modes, are obtained from
T TCB M T CB and T TCB K T CB .
Note that the CB method can be improved by considering a more accurate
approximation of the static modes. Koutsovasilis [69] proposed to replace the Guyan
reduction by the IRS technique or, as it was proposed by Urgueira and Ewins [141],
using the residual information of the neglected modes to improve the displacement
approximation. However, one can state that the Guyan reduction can be replaced by
any other more accurate technique.
21 Toward Robust Response Models: Theoretical and Experimental Issues 1193

4.1.7 Sum of Weighted Accelerations Technique (SWAT)

Carne et al. [21] presented the sum of weighted accelerations technique (SWAT), a
force identification technique based on the concept of a modal filter. This technique
uses the rigid body mode shapes in order to derive a weighting vector that separates
the rigid body accelerations from the elastic response within a frequency band of
interest. Hence, if one considers the transformation of generalized coordinates into
physical ones, given by

x = q (74)

the acceleration vector can be approximated by a sum of modal contributions,

defined as

ẍ =  q̈ (75)

By the SWAT technique, a weighting matrix W is derived such that a vector of

rigid body accelerations ẍ RB , also known as SWAT dofs, can be extracted from the
measured one:

ẍ RB = W T ẍ = W T q̈ (76)

here with mass-normalized modes. Rewriting Eq. (75) in terms of the contribution
of the rigid body modes RB and the flexible ones , one has
 
ẍ RB
ẍ = [RB ] (77)
ẍ

Using Eqs. (76) and (77), it is possible to define the weighting matrix W that
extracts only the rigid body accelerations from the measured ones, as described by
Allen and Carne [2]:

 +  I 
W = [RB ] T
(78)
0

Note that this technique can be used as a reduction technique in the sense that the
model can be replaced by its local time-dependent response [86].
On the other hand, if one truncates the elastic mode shapes to a given number of
modes of interest (k), defined as the number of primary dofs to retain, as addressed
by the CB method, one can relate the spatial model coordinates to the primary k
ones based on Eq. (77), as
     
xp x RB x RB
x= = [RB k ] = T SW (79)
xs xk xk
1194 N. Maia et al.

Thus, the reduced system matrices are given in a similar way to Eq. (49) and their
dimension is given by the number of rigid body modes plus the number of modes of
interest (k).

4.1.8 Reduction of Damped Models

Jeong et al. [58] derive a transformation matrix equivalent to the IIRS technique in
the state-space formulation to be used for damped systems. The equation of motion
for non-proportional damping must be reformulated in the state-space form to derive
the quadratic eigenvalue problem:

Aẏ + By = b (80)

with
       
C M K 0 x f
A= , B= , y= and b= (81)
M 0 0 −M ẋ 0

From Eq. (80), one can write the generalized eigenvalue problem,

˜ = A
B ˜˜ (82)

whose solutions are real or exist in complex conjugate pairs, given by

   
˜ =  ∗ ˜ =  0
 and  (83)
  ∗ ∗ 0 ∗

where ˜ and  ˜ are matrices of complex conjugate pairs of eigenvectors and

eigenvalues.
Rewriting Eq. (82) in a partitioned form, one has
     
B pp B ps ˜p
 App Aps ˜p

= ˜
B sp B ss ˜s
 Asp Ass ˜s 

(84)

from which one derives a transformation matrix (TSS ) equivalent to the IIRS
transformation given in Eq. (58):
     
(k) I 0 0 (k−1) (k−1) −1 (k−1)
T SS = + B T SS BR AR (85)
− B −1
ss B sp
−1
0 B ss

(0)
where the superscript (k) denotes the kth iteration, for k > 1. Note that T SS =
 
I
, equivalent to the static reduction and the reduced model matrices
− B −1
ss B sp
are given by
21 Toward Robust Response Models: Theoretical and Experimental Issues 1195

   
(k) (k) T (k) (k) (k) T (k)
AR = T SS A T SS and B R = T SS B T SS (86)

Das and Dutt [35] proposed a modified SEREP formulated in state-space,

including the gyroscopic effect and both internal and external damping. This
technique is thought to reduce finite element models of rotor systems, although it
can be applied to any linear system of equations.

4.2 Expansion of Measured Data

As the referred reduction methods point out, the majority of the available techniques
have been developed either in the spatial or modal domains, and they can be directly
used to expand the model data over the secondary coordinates. Such approach can
be viewed as an inverted reduction, where the mode shape expansion is achieved
by using the transformation matrices of the above referred reduction techniques.
Hence, one can obtain the complete rth mode shape  r considering the expansion of
 pr , using one of the several transformation matrices developed for model reduction
purposes, such as,
 
p
r = = T pr (87)
s r

If one considers not just one particular mode shape but a set of mode shapes, Eq.
(87) can be recast as
 
p
= = T p (88)
s

where  is a rectangular modal matrix constituted by a set of measured mode shapes

 p expanded over the set of secondary coordinates.

4.2.1 Kidder’s Method

Kidder [65] creates a spatial transformation using the analytical mass and stiffness
matrices and the generalized eigenproblem
     
K pp K ps M pp M ps p
− ωr2 =0 (89)
K sp K ss M sp M ss s r

From Eq. (89), the expanded mode shapes are given by

 
I
r =  −1   pr = T K pr (90)
− K ss − ωr2 M ss K sp − ωr2 M sp
1196 N. Maia et al.

where TK is the transformation matrix of the Kidder’s method.

Note that the expansion process is achieved mode by mode and that experimental
mode shapes can be expanded over the entire model coordinates.

4.2.2 Expansion Using Analytical Modes

Perhaps the simplest approach to expand measured data is the one where the data
at unmeasured dofs is directly replaced by its analytical counterpart, although
numerical discontinuities may arise from this kind of procedure [77]. However, in
this context, a feasible technique was introduced by Lipkins and Vandeurzen [72].
This expansion technique relies on the assumption that each mode can be obtained
by a linear combination of the analytical modes. Considering the analytical modal
matrix partitioned in terms of primary and secondary coordinates and selected and
discarded mode shapes, one has
 
 11  12
A = (91)
 21  22

where the submatrix [ 11  21 ]T is related to the selected mode shapes. Hence,

the expanded experimental mode shapes are generated from the following linear
combination:
   
 X1  11
X = = υ (92)
 X2  21

where υ is a vector of unknown coefficients obtained in a least squares sense, as

υ = ( 11 )+  X1 (93)

and therefore the mode shapes at the secondary coordinates are computed by

 X2 =  21 υ (94)

4.2.3 Expansion of Frequency Response Functions (FRF)

Rather than expand the mode shapes, the need to expand a set of measured FRFs
has been addressed by several authors [16].
Silva et al. [126] used a modified version of the Kidder’s expansion method
in order to expand a set of measured FRFs Hpj (ω) over the secondary model
coordinates, considering that the experimental responses are due to a set of q
excitations applied individually at j locations. Instead of being based on the solution
of the eigenproblem, this technique is based on the notion of dynamic stiffness and
it is derived from the equilibrium equation for a particular frequency, as
       
K pp K ps M pp M ps H pj (ω) Ij
−ω 2
= (95)
K sp K ss M sp M ss H sj (ω) 0
21 Toward Robust Response Models: Theoretical and Experimental Issues 1197

where Ij is a vector of zeros except for the element corresponding to the excitation
location j. From Eq. (95), the expanded FRF vector is given by

 
I
H j (ω) =  −1   H pj (ω) = T M H pj (ω)
− K ss − ω2 M ss K sp − ω2 M sp
(96)

which can also be extended to be used for damped systems, if one considers the
general equilibrium equation (44).
For the sake of simplicity, from now on the explicit reference to the frequency
dependence of the FRF matrix and sub-matrices will be omitted.
Avitabile and O’Callahan [6] proposed an expansion technique, in the FRF partial
fraction form, written to include the effect of all the modes of the system. This
method can be implemented in one of two ways: (i) one needs to synthesize the
translation FRFs from the measured modal quantities through their partial fraction
form and to expand them to rotational dofs using SEREP; additionally, one includes
the residual effect of the unmeasured modes and the expansion of them using the
dynamic expansion method; and (ii) where the synthesized FRFs are expanded to
the rotational dofs using SEREP and the residual effect is related to the difference
of the SEREP expansion and the dynamic one at a particular frequency. However,
avoiding modal identification, one can address the straightforward expansion of
the experimental rectangular FRF matrix using the transformation matrix of the
modified Kidder’s method (TM ), as proposed by Silva and Maia [124] using a
proper classification for each dof in the system.
As it is obvious, the expanded FRF of Eq. (96), or the ones obtained by the
expansion techniques of Avitabile and O’Callahan [6], can only complete a sub-
matrix of H, related to the coordinates where the force excitation is applied.
However, the rotational FRFs are required by several methods, from coupling to
model updating. The work of Avitabile and O’Callahan [6] proposed a strategy to
expand the set of measured FRFs in order to obtain a complete FRF matrix, as given
in the work of Drozg et al. [38]. To implement such a strategy one must write the
FRF matrix in the following partitioned form:

 
H pq H pu
H = (97)
H sq H su

where the subscript u is related to the non-excited coordinates, neither by forces

nor by moments. Note that one has both translational and rotational dofs among the
secondary coordinates and that the non-excited set of coordinates also encompasses
translational and rotational dofs, respectively, identified by the subscripts st , sθ , uf ,
and uτ . Thus, Eq. (97) can be rewritten as
1198 N. Maia et al.

⎡ ⎤
H pq H puf H puτ
⎢ ⎥
H = ⎣ H st q H st uf H st uτ ⎦ (98)
H sθ q H sθ uf H sθ uτ

and the complete FRF matrix can be obtained from Hpq using the following steps:

1. Expand Hpq to the secondary coordinates:

 
H pq
= T H pq (99)
H sq

2. Transpose the expanded FRFs:

H pu = H Tsq (100)

3. Expand Hpu to the secondary coordinates:

 
H pu
= T H pu (101)
H su

5 Transmissibility as a Means to Estimate the Dynamic

Response

5.1 Introduction

The dynamic characteristics of a structure are often derived from a set of measured
frequency response functions. However, it may happen that due to the cost related
to data acquisition, or to practical reasons such as the inaccessibility of locations
for measurement, certain FRFs cannot be experimentally measured. In these
circumstances, it is useful to have some tools that can provide the prediction of
such dynamic information. The transmissibility concept seems to play an important
role to circumvent these situations.
The notion of transmissibility is presented in every classic textbook on vibrations,
associated to the single degree-of-freedom system, when its basis is moving
harmonically; it is defined as the ratio between the modulus of the response
amplitude and the modulus of the imposed amplitude of motion.
The first attempts to extend the idea of transmissibility to a system with N
dofs, that is, how to relate a set of unknown responses to another set of known
responses, for a given set of applied forces, were given by Vakakis et al. [99, 143–
145], although their generalization was still limited to a very particular type of N
degree-of-freedom systems, one where a set constituted by a mass, stiffness, and
damper is repeated several times. The works of Ewins and Liu [46] and Varoto and
McConnell [146] also extend the initial concept to N degrees-of-freedom systems,
21 Toward Robust Response Models: Theoretical and Experimental Issues 1199

but again in a limited way, the former using a definition that makes the calculations
dependent on the path taken between the considered coordinates involved, the
latter by restricting the set of coordinates where the displacements are known
to be coincident to the set of applied forces. An important contribution to the
generalization of the transmissibility concept in frequency domain for multiple dof
systems was presented by Ribeiro [109]. In this generalization, the transmissibility
matrix between two sets of response functions is built from any of the mobility
matrices of the structure. In most practical cases, the known (or measured) responses
shall constitute one of the sets, while the other set includes the responses at any of
the other coordinates. The generalization of the transmissibility concept suggests
that if the transmissibility matrix can be evaluated in the laboratory or numerically
beforehand, then by measuring in service some responses, one would be able to
estimate the responses at the inaccessible coordinates.
The next section presents the original and an alternative formulation of the
transmissibility presented by Ribeiro et al. [111]; some important properties will
also be referred.

5.2 Theoretical Description

5.2.1 Fundamental Formulation

The approach proposed in Ribeiro [109] and Ribeiro et al. [111] is based on
harmonically applied forces (easy to generalize to periodic ones). If one has a vector
FA of magnitudes of the excitation forces (and/or moments) at coordinates A, a
vector XU of unknown response amplitudes at coordinates U, and a vector XK of
known response amplitudes at coordinates K, then one can write:

XU = H U A F A (102)

XK = H KA F A (103)

where HUA and HKA are the receptance frequency response matrices relating
coordinates U and A, and K and A, respectively. Assuming that the number of known
coordinates is equal or higher than the number of applied forces, that is, #K ≥ # A,
one can eliminate FA between Eqs. (102) and (103), leading to

XU = H U A H +
KA X K (104)

or
(A)
XU = T U K XK (105)

where H + KA is the pseudo-inverse of HKA . Thus, the transmissibility matrix is

defined as
1200 N. Maia et al.

T UK = H UA H +
(A)
KA (106)

Note that the set of coordinates where the forces are applied (A) need not coincide
with the set of known responses (K). The only restriction is that – for the pseudo-
inverse to exist – the number of K coordinates must be greater or equal than the
number of A coordinates.
An important property of the transmissibility matrix is that it does not depend on
the magnitude of the forces; one simply has to know or to choose the coordinates
where the forces are going to be applied (or not, as one can even choose more
coordinates A if one is not sure whether or not there will be some forces there
and later on state that those forces are zero) and measure the necessary frequency
response functions.
As previously referred, one of the requirements of this approach is that the
number and locations of the dynamic loads be known. However, in real structures,
it may be difficult to point out a priori the coordinates where the dynamic loads are
applied; examples of these situations are the case of structures loaded by traffic or
by wind. To apply the transmissibility concept to structures subjected to distributed
loads, the same authors in Ribeiro et al. [110] developed a method for response
prediction from a reduced set of known responses, using the transmissibility matrix.
For situations of non-harmonic excitations, as in the case of structures submitted
to environmental loads, the transmissibility concept has been extended too, using
the spectral densities of the response measurements [112].

5.2.2 Alternative Formulation

An alternative approach developed by Ribeiro et al. [113] evaluates the transmissi-
bility matrix from the dynamic stiffness matrices, where the spatial properties (mass,
stiffness, etc.) are explicitly included.
The dynamic behavior of an MDOF system can be described in the frequency
domain by the following equation (assuming harmonic loading):

Z X=F (107)

where Z represents the dynamic stiffness matrix, X is the vector of the amplitudes of
the dynamic responses, and F represents the vector of the amplitudes of the dynamic
loads.
From the set of dynamic responses, as defined before, it is possible to distinguish
between two distinct subsets of coordinates K and U; from the set of dynamic loads
it is also possible to distinguish between two subsets, A and B, where A is the subset
where the dynamic loads may be applied and B is the set formed of the remaining
coordinates, where the dynamic loads are never applied. One can write X and F as

   
XK FA
X= ,F = (108)
XU FB
21 Toward Robust Response Models: Theoretical and Experimental Issues 1201

With these subsets, Eq. (107) can be partitioned accordingly:

     
Z AK Z AU XK FA
= (109)
Z BK Z BU XU FB

Taking into account that coordinates B represent the ones where the dynamic
loads are never applied, and considering that the number of these coordinates is
greater or equal to the number of coordinates U, from Eq. (109) it is possible to
obtain the unknown response vector:

F B = 0, #B ≥ #U
⇓ (110)
XU = −Z +BU Z BK X K

where Z +BU is the pseudo-inverse of ZBU . Therefore, this means that the transmissi-
bility matrix can also be defined as

T U K = −Z +
(A)
BU Z BK (111)

Equation (111) is an alternative definition of transmissibility, based on the

dynamic stiffness matrices of the structure. Therefore,

T UK = H UA H + +
(A)
KA = −Z BU Z BK (112)

Considering that the dynamic stiffness matrix for an undamped system is

described in terms of the stiffness and mass matrices, Z = K − ω2 M, one can
now relate the transmissibility functions to the spatial properties of the system. To
make this possible, one must bear in mind that it is mandatory that both conditions
regarding the number of coordinates be valid, that is,

#B ≥ #U and #K ≥ #A (113)

5.2.3 Transmissibility Properties

Observing Eqs. (104) and (106), it is possible to conclude that the transmissibility
matrix is independent from the force vector FA . (Note that FA is eliminated between
Eqs. (102) and (103).) This means that any change verified in one of the force
values, acting along with coordinates of set A, will not affect the transmissibility
(A)
matrix T U K . This change can be due, for instance, to the alteration of mass values
associated to coordinates A or stiffness values of springs interconnecting those
(A)
coordinates. Additionally, to highlight that characteristic of matrix T U K , it can be
verified from Eq. (111) that there are no coordinates A involved in the matrices Z
(i.e., neither ZAK nor ZAU ). As it was presented by Maia et al. [81], this statement
reinforces the previous conclusion extracted from Eq. (106) and will lead to the
formulation of two properties, as follows:
1202 N. Maia et al.

Property 1. The transmissibility matrix does not change if some modification is

made on the mass values of the system where the loads can be applied – subset A.
Property 2. The transmissibility matrix does not change if some modification is
made on the stiffness values of springs interconnecting coordinates of subset A –
(where the loads can be applied).

However, any changes in the mass values associated to coordinates A and/or any
changes in the stiffness values of springs interconnecting coordinates A will affect
the inertia forces and elastic forces, respectively, acting along those coordinates and
thus belonging to FA .
According to properties 1 and 2, if a modification is operated on the original
system, the transmissibility matrix remains constant, and therefore, it is possible to
estimate the FRFs associated to the unknown coordinates without the necessity of
measuring the responses on those coordinates, that is,

+   +
T (A)
U K = H U A H KA = H U A H KA (114)

where HUA , HKA and H U A , H KA are the receptance frequency response matrices
relating coordinates U and A, and K and A, for the original and the modified system.
Thus, if the receptance matrix relating coordinates K and A of the modified system
(H KA ) is known, the receptance matrix relating coordinates U and A (H U A ) can be
estimated:

H U A = T U K H KA
(A)
(115)

5.3 Other Possible Applications of Transmissibility

The potentialities of application of the transmissibility are not limited to the

determination of unmeasured FRFs, a number of other possible applications are
reported in Maia et al. [81], like in structural coupling [36, 49, 70], damage detection
[80, 117, 118, 160], sensor placement [149], etc.

References
1. Algrain MC, Saniie J (1991) Estimation of 3D angular motion using gyroscopes and linear
accelerometers. IEEE Trans Aerosp Electron Syst 27:910–920
2. Allen M, Carne T (2006) Comparison of Inverse Structural Filter (ISF) and Sum of
Weighted Accelerations (SWAT) time domain force identification methods. In: 47th
AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics, and materials conference
14th AIAA/ASME/AHS adaptive structures conference 7th. Structures, structural dynamics,
and materials and co-located conferences. Reston: American Institute of Aeronautics and
Astronautics. https://doi.org/10.2514/6.2006-1885. Accessed 24 Mar 2014
3. Allen MS, Kammer DC, Mayes RL (2010) Uncertainty in experimental/analytical substruc-
turing predictions: a review with illustrative examples. In: Proceedings of ISMA 2010, Leuven
21 Toward Robust Response Models: Theoretical and Experimental Issues 1203

4. Antonello R, Oboe R (2011) MEMS gyroscopes for consumers and industrial applications.
In: Minin O (ed) Microsensors. InTech
5. Avitabile P, O’Callahan J (1987) Expansion of rotational degrees of freedom for struc-
tural dynamic modifications. Presented at the 5th international modal analysis conference
(IMAC V), London, pp 950–955
6. Avitabile P, O’Callahan J (2003) Frequency response function expansion for unmeasured
translation and rotation dofs for impedance modelling applications. Mech Syst Signal Process
17(4):723–745. http://www.sciencedirect.com/science/article/pii/S0888327001914578
7. Avitabile P, Piergentili F (1997) Consideration for effects of rotational degrees of freedom for
hybrid modeling applications. Presented at the 15th international modal analysis conference
(IMAC XV), Orlando, pp 792–798
8. Bampton MCC, Craig RR Jr (1968) Coupling of substructures for dynamic analyses. AIAA J
6(7):1313–1319. https://doi.org/10.2514/3.4741. Accessed 25 Mar 2014
9. Batista FC, Maia NMM (2011) Uncoupling techniques for the dynamic characteriza-
tion of substructures. In: Linking models and experiments, vol 2. Springer, New York,
pp 383–392
10. Bello M, Sestieri A, D’Ambrogio W, La Gala F (2003) Development of a rotation transducer
based on bimorph PZTs. Mech Syst Signal Process 17:1069–1081
11. Beskhyroun S, Ma QT (2012) Low-cost accelerometers for experimental modal analysis. In:
Proceedings of the 15th world conference on earthquake engineering
12. Besselink B et al (2013) A comparison of model reduction techniques from structural
dynamics, numerical mathematics and systems and control. J Sound Vib 332(19):4403–4422.
http://linkinghub.elsevier.com/retrieve/pii/S0022460X1300285X. Accessed 4 Mar 2014
13. Bokelberg EH, Sommer III HJ, Trethewey MW (1993) Simultaneous measurement of six
coordinate vibration: three translations and three rotations. Presented at the 11th international
modal analysis conference (IMAC XI), Kissimmee, pp 522–527
14. Bokelberg EH, Sommer HJ III, Trethewey MW (1994) A six-degree-of-freedom laser
vibrometer, part I: theoretical development. J Sound Vib 178:643–654
15. Bokelberg EH, Sommer HJ III, Trethewey MW (1994) A six-degree-of-freedom laser
vibrometer, part II: experimental validation. J Sound Vib 178:655–667
16. Bregant L, Sanderson M (2000) Rotational degrees of freedom an historical overview on
techniques and methods. Presented at the Proceedings of the international seminar on modal
analysis (ISMA 25), Leuven
17. Cafeo JA, Rieker JR, Trethewey MW, Sommer III HJ (1991) Application of a three degree of
freedom laser vibrometer for experimental modal analysis. Presented at the 9th international
modal analysis conference (IMAC IX), Florence, pp 1161–1167
18. Cafeo JA, Trethewey MW, Sommer HJ III (1992) Measurement and application of experi-
mental rotational degrees of freedom for mode shape refinement. Int J Anal Exp Modal Anal
7:255–269
19. Cafeo JA, Trethewey MW, Sommer III HJ (1993) On the use of measured rotational degrees-
of-freedom in structural dynamic modification. Presented at the 11th international modal
analysis conference (IMAC XI), pp 96–101
20. Cafeo JA, Trethewey MW, Sommer HJ III (1995) Beam element structural dynamics
modification using experimental modal rotational data. J Vib Acoust 117:265–271
21. Carne TG, Bateman VI, Mayes RL (1992) Force reconstruction using a sum of weighted
accelerations technique. In: 10th international modal analysis conference (IMAC X), San
Diego
22. Champoux Y, Cotoni V, Paillard B, Beslin O (2003) Moment excitation of structures using
two synchronized impact hammers. J Sound Vib 263:515–533
23. Chen W-H, Cherng J-S (1985) Modal synthesis via combined experimental and finite element
techniques with consideration of rotational effects. J Sound Vib 103:1–11
24. Cheng L, Qu YC (1997) Rotational compliance measurements of a flexible plane structure
using an attached beam-like tip, part 1: analysis and numerical simulation. J Vib Acoust
119:596–602
1204 N. Maia et al.

25. Cicogna TR, Varoto PS, Trindade MA (2008) Measurement of rotational FRFs in isotropic
beams and plates using piezoceramic bimorphs and genetic algorithm optimization. Presented
at the 26th IMAC: conference and exposition on structural dynamics 2008 (IMAC XXVI),
Orlando
26. Cigada A, Lurati M, Redealli M, Vanali M (2007) Mechanical performance and metrological
characterization of MEMS accelerometers and application in modal analysis. Presented at the
conference & exposition on structural dynamics (IMAC XXV), Orlando
27. Cigada A, Lurati M, Redealli M, Vanali M (2008) Evaluation of the metrological per-
formances of a MEMS accelerometers to the aim of their application in structural health
monitoring. Presented at the conference & exposition on structural dynamics – technologies
for civil structures (IMAC XXVI), Orlando
28. Cloutier D, Avitabile P (2010) Investigation on the use of various decoupling approaches. In:
Proceedings of the IMAC-XXVIII, Jacksonville
29. Craig RR Jr (2000) Coupling of substructures for dynamic analyses: an overview. In: AIAA
dynamics specialists conference, Atlanta, p 1573
30. Crowley SM, Mehzad J, Brown DL (1986) An investigation of structural modification using
an H-frame structure. Presented at the 4th international modal analysis conference (IMAC
IV), Los Angeles, pp 1268–1278
31. D’Ambrogio W, Fregolent A (2008) Promises and pitfalls of decoupling procedures. In:
Proceeding of 26th IMAC, Orlando
32. D’Ambrogio W, Fregolent A (2008) Sensitivity of decoupling techniques to uncertainties in
the properties. In: Proceedings of ISMA 2008, Leuven, pp 3737–3749
33. D’Ambrogio W., Fregolent A., 2009. Decoupling procedures in the general framework of
frequency based substructuring. In: Proceedings of the IMAC-XXVII, Orlando
34. D’Ambrogio W, Fregolent A (2014) Are rotational DoFs essential in substructure decoupling?
Dynamics of coupled structures, vol 1, Proceedings of IMAC 32
35. Das AS, Dutt JK (2008) Reduced model of a rotor-shaft system using modified
SEREP. Mech Res Commun 35(6):398–407. http://www.sciencedirect.com/science/article/
pii/S0093641308000451. Accessed 4 Mar 2014
36. Devriendt CMG, Fontul M (2005) Transmissibility in coupled structures: identification of the
sufficient set of coupling forces. In: Proceedings of the 1st international operational modal
analysis conference. Presented at the IOMAC 1, Copenhagen
37. Drain LE (1980) The laser Doppler technique. Wiley, Chichester/New York
38. Drozg A, Rogelj J, Čepon G, Boltežar M (2018) On the performance of direct piezoelectric
rotational accelerometers in experimental structural dynamics. Measurement 127:292–298
39. Duarte ML (1996) Experimentally-derived structural models for use in further dynamic
analysis. PhD thesis. Imperial College of Science, Technology and Medicine London, London
40. Duarte ML, Ewins DJ (1995) Some insights into the importance of rotational degrees-of-
freedom and residual terms in coupled structure analysis. Presented at the 13th international
modal analysis conference (IMAC XIII), Nashville, pp 164–170
41. Duarte ML, Ewins DJ (2000) Rotational degrees of freedom for structural coupling analysis
via finite-difference technique with residual compensation. Mech Syst Signal Process 14:
205–227
42. Dumont M, Kinsley N (2016) Using accelerometers for measuring rotational degrees-of-
freedom. Sound & Vibration
43. Elliott AS, Moorhouse AT, Pavić G (2008) Measurement of force and moment mobilities
using a finite difference technique. Presented at the Acoustics’ 08, Paris, pp 1808–1812
44. Elliott AS, Moorhouse AT, Pavić G (2012) Moment excitation and the measurement of
moment mobilities. J Sound Vib 331:2499–2519
45. Ewins DJ, Gleeson PT (1975) Experimental determination of multidirectional mobility data
for beam. Shock Vib Bull 45:153–173
46. Ewins DJ, Liu W (1998) Transmissibility properties of MDOF systems. Presented at the 16th
international modal analysis conference (IMAC XVI), Santa Barbara, pp 847–854
21 Toward Robust Response Models: Theoretical and Experimental Issues 1205

47. Ewins DJ, Sainsbury MG (1972) Mobility measurements for the vibration analysis of
connected structures. Shock Vib Bull 42:105–122
48. Ewins DJ, Silva JMM, Maleci G (1980) Vibration analysis of a helicopter plus an externally-
attached structure. Shock Vib Bull 50:155–171
49. Fontul M, Ribeiro AMR, Silva JMM, Maia NMM (2004) Transmissibility in structural
coupling. In: Proceedings of international conference on noise and vibration engineering.
Presented at the ISMA 2004, Leuven
50. Friswell MI, Garvey SD, Penny JET (1995) Model reduction using dynamic and iterated
IRS techniques. J Sound Vib 186(2):311–323. http://www.sciencedirect.com/science/article/
pii/S0022460X85704518. Accessed 4 Mar 2014
51. Friswell MI, Garvey SD, Penny JET (1998) The convergence of the iterated IRS
method. J Sound Vib 211(1):123–132. http://www.sciencedirect.com/science/article/pii/
S0022460X97913687. Accessed 5 Mar 2014
52. Gialamas T, Tsahalis D, Bregant L, Otte D, Van der Auweraer H (1996) Substructuring by
means of FRFs: some investigations on the significance of rotational Dofs. Presented at the
14th international modal analysis conference (IMAC XIV), Dearborn, pp 619–625
53. Giuliani P, Di Maio D, Schwingshackl CW, Martarelli M, Ewins DJ (2013) Six degrees of
freedom measurement with continuous scanning laser Doppler vibrometer. Mech Syst Signal
Process 38:367–383
54. Gleeson PT (1979) Identification of spatial models for the vibration analysis of lightly damped
structures. PhD thesis. Imperial College London, University of London, London
55. Guyan RJ (1965) Reduction of stiffness and mass matrices. Am Inst Aeronaut Astronaut J
3(2):961–962
56. Helderweirt S, Van der Auweraer H, Mas P, Bregant L, Casagrande D (2001) Application
of accelerometer-based rotational degree of freedom measurements for engine subframe
modeling. In: 19th international modal analysis conference, IMAC XIX, pp 1298–1304
57. Ivarsson L, Sanderson MA, Troshin AG (2000) Design, theory and validation of a low mass
6-d.o.f. transducer. J Sound Vib 230:661–688
58. Jeong J, Baek S, Cho M (2012) Dynamic condensation in a damped system through
rational selection of primary degrees of freedom. J Sound Vib 331(7):1655–1668. http://www.
sciencedirect.com/science/article/pii/S0022460X11008959. Accessed 7 Mar 2014
59. Jetmundsen B, Bielawa RL, Flannelly WG (1988) Generalised frequency domain. J Am
Helicopter Soc 33:55–64
60. Jianxin S, Gibbs BM (1998) Measurement of point moment mobility in the presence of non-
zero cross mobility. Appl Acoust 54:9–26
61. Jianxin S, Mak CM (2002) Direct measurement of moment mobility and a moment excitation
system. Appl Acoust 63:139–151
62. Kammer DC (1987) Test-analysis model development using an exact modal reduction. Int J
Anal Exp Modal Anal 2(3):174–179
63. Kammer DC (1991) A hybrid approach to test analysis model development for large space
structures. J Vib Acoust 113(3):325–332
64. Kanda H, Wei ML, Allemang RJ, Brown DL (1986) Structural dynamic modifications using
mass additive technique. Presented at the 4th international modal analysis conference (IMAC
IV), Los Angeles, pp 691–699
65. Kidder RL (1973) Reduction of structural frequency equations. AIAA J 11(6):892
66. Kim YY, Kang JH (1999) Strain-based rotational mode measurement in a beam. J Sound Vib
222:225–237
67. Koh YK, White RG (1996) Analysis and control of vibrational power transmission to
machinery supporting structures subjected to a multi-excitation system, part III: vibrational
power cancellation and control experiments. J Sound Vib 196:509–522
68. Kok R, Furlong C, Pryputniewicz J (2003) Application of MEMS accelerometers for modal
analysis. Presented at the SEM annual conference & exposition on experimental and applied
mechanics
1206 N. Maia et al.

69. Koutsovasilis P (2009) Model order reduction in structural mechanics: coupling

the rigid and elastic multi body dynamics. PhD thesis, Technische Universität
Dresden
70. Law SS, Li J, Ding Y (2011) Structural response reconstruction with transmissibility concept
in frequency domain. Mech Syst Signal Process 25:952–968. https://doi.org/10.1016/j.ymssp.
2010.10.001
71. Li S, Fladung W, Brown D, Bono R (1995) Indirect measurement of rotation/moment
impedance functions (SAE Technical Paper No. 951271). SAE International, Warrendale
72. Lipkins J, Vandeurzen U (1987) The use of smoothing techniques for structural modification
applications. In 12th international seminar on modal analysis, pp S1–S3
73. Liu W, Ewins DJ (1999) The importance assessment of RDOF in FRF coupling analysis.
Presented at the 17th international modal analysis conference (IMAC XVII), Kissimmee, pp
1481–1487
74. Liu Z, Song X, Zhao Y, Cai L, Guo H, Ma J (2015) Stiffness identification of spindle-
toolholder joint based on finite difference technique and residual compensation theory. Adv
Mech Eng. https://doi.org/10.1155/2013/753631
75. Lofrano M (2003) Techniques to estimate the rotational frequency response functions in
experimental modal analysis with applications to beam structures. PhD thesis. Escola de
Engenharia de São Paulo, Universidade de São Paulo, São Paulo
76. Maenaka K (2008) MEMS inertial sensors and their applications. Presented at the 5th
international conference on networked sensing systems, INSS 2008, pp 71–73
77. Maia NMM, Montalvão e Silva JM, He J, Lieven NAJ, Lin RM, Skingle GW, Todd W-
M, Urgueira APV (1997) Theoretical and experimental modal analysis. Research Studies
Press/Wiley, Taunton, Somerset/New York
78. Maia NMM, Silva JMM, Ribeiro AMR (1997) Some applications of coupling/uncoupling
techniques in structural dynamics. Part 3: estimation of rotational FRFs using MUM.
Presented at the 15th international modal analysis conference (IMAC XV), Orlando, pp 1453–
1462
79. Maia NMM, Silva JMM, Ribeiro AMR, Silva PLCGC (1998) On the dynamic characteriza-
tion of joints using uncoupling techniques. Presented at the 16th international modal analysis
conference (IMAC XVI), Santa Barbara, pp 1132–1138
80. Maia NMM, Ribeiro AMR, Fontul M, Montalvão D, Sampaio RPC (2007) Using the
detection and relative damage quantification indicator (DRQ) with transmissibility. KEM
347:455–460. ISSN: 1662–9795
81. Maia NMM, Urgueira APV, Almeida RAB (2011) Whys and wherefores of transmissibility.
In: Beltran-Carbajal F (ed) Vibration analysis and control – new trends and developments.
InTech
82. Maio DD, Copertaro E (2013) Experimental validation of a newly designed 6 degrees
of freedom scanning laser head: application to three-dimensional beam structure. Rev Sci
Instrum 84:121708
83. Maleci G, Young JW (1985) The effect of rotational degrees of freedom in system analysis
(SA) via building block approach (BBA). Presented at the 3rd international modal analysis
conference (IMAC III), Orlando, pp 1040–1045
84. Martarelli M (2001) Exploiting the laser Scanning facility for vibration measurements. PhD
thesis. Imperial College of Science, Technology and Medicine London, London
85. Martinez DR, Miller AK, Came TG (1985) Combined experimental analytical modeling of
shell/payload structures. Presented at the joint ASCE/ASME mechanics conference, AMD,
167, pp 167–194
86. Minnicino II M, Hopkins D (2004) Overview of reduction methods and their implementation
into finite element local-to-global techniques. Technical report, U.S. Army Research Labora-
tory, Aberdeen Proving Ground, ARL-TR-3340
87. Montalvão D (2003) Determination of rotational terms of the dynamic response by means of
modal analysis techniques de termos rotacionais da resposta dinâmica por técnicas de análise
modal. Master thesis. Instituto Superior Técnico, UTL, Lisboa
21 Toward Robust Response Models: Theoretical and Experimental Issues 1207

88. Montalvão D, Ribeiro AMR, Maia NMM, Silva JMM (2004) Estimation of the rotational
terms of the dynamic response matrix. Shock Vib 11:333–350
89. Morris HD, Peters RB (1980) A precision inertial angular vibration measuring system. Shock
Vib Bull 50:1–10
90. Mottershead JE, Kyprianou A, Ouyang H (2003) Estimation of rotational frequency
responses. Presented at the 5th international conference on damage assessment of structures
(DAMAS 2003), Southampton, pp 157–165
91. Mottershead JE, Kyprianou A, Ouyang H (2005) Structural modification. Part 1: rotational
receptances. J Sound Vib 284:249–265
92. Nasiri S (2013) A critical review of MEMS gyroscopes technology and commercialization
status. InvenSense, Santa Clara, CA, USA
93. Ng’andu NA, Fox CHJ, Williams EJ (1993) Estimation of rotational degrees of freedom
using curve and surface fitting. Presented at the 11th international modal analysis conference
(IMAC XI), Kissimmee, pp 620–626
94. Ng’andu NA, Fox CHJ, Williams EJ (1995) On the estimation of rotational degrees of
freedom using spline functions. Presented at the 13th international modal analysis conference
(IMAC XIII), Nashville, pp 791–797
95. O’Callahan J (1989) A procedure for an improved reduced system (IRS) model. In:
Proceedings of 7th international modal analysis conference, pp 17–21
96. O’Callahan JC, Chaur-Ming C (1984) Study of a structural modification procedure with three
dimensional beam elements using a local eigenvalue modification procedure. Presented at the
2nd international modal analysis conference (IMAC II), Orlando, pp 945–952
97. O’Callahan J, Avitabile P, Riemer R (1989) System equivalent reduction expansion process
(SEREP). In: Proceedings of 7th international modal analysis conference, pp 29–37
98. Oliver D (1988) Non-contact vibration imager for wide range of component sizes and
displacement amplitudes. Presented at the 6th international modal analysis conference (IMAC
VI), Kissimmee, pp 629–638
99. Paipetis SA, Vakakis AF (1985) A method of analysis for unidirectional vibration isolators
with many degrees of freedom. J Sound Vib 98(1):13–23. https://doi.org/10.1016/0022-
460X(85)90398-0. ISSN: 0022–460X
100. Paz M (1984) Dynamic condensation method. Am Inst Aeronaut Astronau J 22(5):724–727
101. Perrier S, Champoux Y, Drouet J-M (2013) Integrating biodynamic measurements in
frequency-based substructuring to study human-structure interaction. In: Topics in experi-
mental dynamic substructuring, vol 2, Proceedings of IMAC 31, a conference and exposition
on structural dynamics
102. Petersson B (1987) On the use of giant magnetostrictive devices for moment excitation. J
Sound Vib 116:191–194
103. Petersson BAT, Heckl M (1996) Concentrated excitation of structures. J Sound Vib 196:
295–321
104. Qu Z-Q (2004) Model reduction techniques: with applications in finite element analysis.
Springer, London
105. Qu YC, Cheng L (1995) Rotational compliance measurement of a flexible plane structure.
Presented at the 13th international modal analysis conference (IMAC XIII), Nashville, pp
948–954
106. Qu Z-Q, Fu Z-F (2000) An iterative method for dynamic condensation of structural
matrices. Mech Syst Signal Process 14(4):667–678. http://linkinghub.elsevier.com/retrieve/
pii/S0888327098913024. Accessed 11 Mar 2014
107. Qu YC, Cheng L, Rancourt D (1997) Rotational compliance measurements of a flexible plane
structure using an attached beam-like tip, part 2: experimental study. J Vib Acoust 119:603
108. Ratcliffe MJ, Lieven NAJ (1996) Measuring rotations using a laser Doppler vibrometer.
Presented at the 14th international modal analysis conference (IMAC XIV), Dearborn, pp
1002–1008
109. Ribeiro AMR (1998) On the generalization of the transmissibility concept. In: Proceedings
of the NATO/ASI conference on modal analysis and testing, Sesimbra
1208 N. Maia et al.

110. Ribeiro AMR, Maia NMM, Silva JMM (2000) Response prediction from a reduced set of
known responses using the transmissibility. In: Proceedings of IMAC XVIII. Presented at the
18th international modal analysis conference, San Antonio, pp 425–427. ISBN: 0-912053-
67-4
111. Ribeiro AMR, Silva JMM, Maia NMM (2000) On the generalisation of the transmissibility
concept. Mech Syst Signal Process 14:29–35
112. Ribeiro AMR, Fontul M, Silva JMM, Maia NMM (2002) Transmissibility matrix in harmonic
and random processes. In: Proceedings of the international conference on structural dynamics
modelling (SDM). Presented at the SDM, Funchal, Madeira
113. Ribeiro AMR, Fontul M, Maia NMM, Silva JMM (2005) Further developments on the
transmissibility concept for multiple degree of freedom systems. Sci Bull “Politehnica” Univ
Timisoara, Romania, Trans Mech 50(64):1224–6077
114. Rixen DJ, van der Valk PLC (2013) An impulse based substructuring approach for impact
analysis and load case simulations. J Sound Vib 232:7174–7190
115. Rixen DJ, van der Valk PLC (2014) Impulse based substructuring as paradigm for coupled
analysis of dynamic components. In: 6th European conference on computational fluid
dynamics, Barcelona
116. Sainsbury MG (1976) (Experimental and theoretical techniques for the) Vibration analysis of
damped complex structures. PhD. Imperial College London, University of London
117. Sampaio RPC, Maia NMM, Ribeiro AMR, Silva JM (1999) Damage detection using the
transmissibility concept. In: Proceedings of ICSV 6. Presented at the sixth international
congress on sound and vibration, Copenhagen
118. Sampaio RPC, Maia NMM, Ribeiro AMR, Silva JM (2001) Transmissibility techniques for
damage detection. In: Proceedings of IMAC XIX. Presented at the 19th international modal
analysis conference, Kissimmee, pp 1524–1527
119. Sanderson MA (1995) Direct measurement of moment mobility: part ii: an experimental
study. J Sound Vib 179:685–696
120. Sanderson MA (1996) Vibration isolation :moments and rotations included. J Sound Vib
198:171–191
121. Sanderson MA, Fredö CR (1995) Direct measurement of moment mobility: part I: a
theoretical study. J Sound Vib 179:669–684
122. Sattinger SS (1980) A method for experimentally determining rotational mobilities of
structures. Shock Vib Bull 50:17–28
123. Silva JMM (1978) Measurement and applications of structural mobility data for the vibration
analysis of complex structures. PhD thesis. Imperial College London, University of London,
London
124. Silva TAN, Maia NMM (2015) Estimation of rotational frequency response functions.
Presented at the 33rd international modal analysis conference (IMAC XXXIII) (ed: Allemang
RJ), vol 6. Springer International Publishing, pp 35–48
125. Silva JMM, Maia NMM, Ribeiro AMR (2001) Indirect estimation of rotational frequency-
response-functions. Presented at the 19th international modal analysis conference (IMAC
XIX), Kissimmee, pp 1535–1542
126. Silva TAN, Maia NMM, Barbosa JI (2012) A model updating technique based on FRFs for
damped structures. In: Sas P (ed) ISMA2012 international conference on noise and vibration
engineering. Katholieke Universiteit Leuven, Leuven
127. Skingle GW (1989) Structural dynamic modification using experimental data. PhD thesis.
Imperial College of Science, Technology and Medicine. University of London, London
128. Smiley RG, Brinkman BA (1984) Rotational degrees of freedom in structural modifica-
tion. Presented at the 2nd international modal analysis conference (IMAC II), Orlando,
pp 937–939
129. Smith JE (1969) Measurement of the total structural mobility matrix. Shock Vib Bull 40:
51–84
21 Toward Robust Response Models: Theoretical and Experimental Issues 1209

130. Sommer III HJ, Erickson MJ, Trethewey MW (1994) Single-beam laser vibrometer for
simultaneous measurement of translation, pitch and roll with neural network calibration.
Presented at the 12th international modal analysis conference (IMAC XII), Honolulu,
pp 1196–1201
131. Stanbridge AB, Ewins DJ (1994) Measurement of translational and angular vibration using a
scanning laser Doppler vibrometer. Presented at the 1st international conference on vibration
measurements by laser techniques: advances and applications, Ancona, pp 37–47
132. Stanbridge AB, Ewins DJ (1996) Measurement of total vibration at a point using a conical-
scanning LDV. Presented at the second international conference on vibration measurements
by laser techniques: advances and applications, Ancona, pp 126–136
133. Stebbins A, Blough JR, Shelley SR, Brown DL (1996) Measuring and including the effects of
moments and rotations for the accurate modeling of transmitted forces. Presented at the 14th
international modal analysis conference (IMAC XIV), pp 429–436
134. Silva TAN, Almeida RAB, Urgueira APV (2018) On the use of measured rotational responses
in structural dynamic analysis. J Phys: Conf Series 1106, conference 1
135. Tao J, Mak CM (2007) Error due to two-force excitation in moment mobility measurement.
Appl Acoust 68:1494–1501
136. Titterton DH, Weston JL (2004) Strapdown inertial navigation technology (chapter 4.3.2).
Institution of Electrical Engineers, Stevenage
137. Trethewey MW, Sommer III HJ (2002) Application of a pure moment exciter for measurement
of moment-rotational dof frequency response functions. Presented at the 20th international
modal analysis conference (IMAC XX), Los Angeles, pp 1153–1158
138. Trethewey MW, Sommer III HJ (2002) Measurement of rotational dof frequency response
functions with pure moment excitation. Presented at the international conference on structural
dynamics modelling, Funchal, Madeira, pp 81–88
139. Trethewey MW, Sommer HJ III, Cafeo JA (1993) A dual beam laser vibrometer for
measurement of dynamic structural rotations and displacements. J Sound Vib 164:67–84
140. Urgueira APV (1989) Dynamic analysis of coupled structures using experimental data. PhD
thesis. Imperial College of Science, Technology and Medicine. University of London, London
141. Urgueira APV, Ewins DJ (1989) A refined modal coupling technique for including residual
effects of out-of-range modes. Presented at the 7th international modal analysis conference
(IMAC VII), Las Vegas
142. Urgueira A, Venâncio N, Riscado P, Almeida RAB, Silva T (2018) Experimental determi-
nation of rigid body properties: an evaluation on the use of piezoelectric or MEMS tri-axial
accelerometers. MATEC web conference, 148
143. Vakakis AF (1985) Dynamic analysis of a unidirectional periodic isolator, consisting of
identical masses and intermediate distributed resilient blocks. J Sound Vib 103(1):25–33.
https://doi.org/10.1016/0022-460X(85)90245-7. ISSN: 0022–460X
144. Vakakis AF, Paipetis SA (1985) Transient response of unidirectional vibration isolators
with many degrees of freedom. J Sound Vib 99(4):557–562. https://doi.org/10.1016/0022-
460X(85)90540-1. ISSN: 0022–460X
145. Vakakis AF, Paipetis SA (1986) The effect of a viscously damped dynamic absorber on a
linear multi-degree-of-freedom system. J Sound Vib 105(1):49–60. https://doi.org/10.1016/
0022-460X(86)90219-1. ISSN: 0022–460X
146. Varoto P, McConnell KG (1998) Single point vs multi point acceleration transmissibility
concepts in vibration testing. Presented at the 12th international modal analysis conference
(IMAC XII), Santa Barbara, pp 83–90
147. Varoto PS, Lofrano M, Cicogna TR, Oliveira LPR, McConnell KG (2006) Moment mobility
FRF measurement techniques. Presented at the IMAC-XXIV conference and exposition on
structural dynamics, St Louis, pp 1298–1304
148. Wang DH, Yuan G (2011) A six-degree-of-freedom acceleration sensing method based on six
coplanar single-axis accelerometers. IEEE Trans Instrum Meas 60:1433–1442
1210 N. Maia et al.

149. Wang J, Law SS, Yang QS (2014) Sensor placement method for dynamic response recon-
struction. J Sound Vib 333:2469–2482. https://doi.org/10.1016/j.jsv.2013.12.014
150. Wei ML, Zhang Q (1989) Correction of modal scaling factors using mass perturbation
technique. Presented at the 7th international modal analysis conference (IMAC VII), Las
Vegas, pp 830–835
151. Williams EJ, Green JS (1990) A spatial curve fitting technique for estimation of rotational
degrees of freedom. Presented at the 8th international modal analysis conference (IMAC
VIII), Kissimmee, pp 376–381
152. Xia Y, Lin R-M (2004) Improvement on the iterated IRS method for structural eigen-
solutions. J Sound Vib 270(4,5):713–727. http://www.sciencedirect.com/science/article/pii/
S0022460X03001883
153. Yang K-T, Park Y (1993) Joint structural parameter identification using a subset of frequency
response function measurements. Mech Syst Signal Process 7:509–530
154. Yasuda C, Riehle PJ, Brown DL, Allemang RJ (1984) An estimation method for rotational
degrees of freedom using a mass additive technique. Presented at the 2nd international modal
analysis conference (IMAC II), Orlando, pp 877–886
155. Yazdi N, Ayazi F, Najafi K (1998) Micromachined inertial sensors. Proc IEEE 86:1640–1659
156. Yeh Y, Cummins HZ (1964) Localized fluid flow measurements with AN He-Ne laser
spectrometer. Appl Phys Lett 4:176–178
157. Yoshimura T, Hosoya N (2000) FRF estimation on rotational dofs by rigid block attachment
approach. Presented at the international conference on noise and vibration engineering (ISMA
25), Leuven
158. Yoshimura T, Hosoya N (2002) FRF measurement of rotational dofs by T-block approach
(Improvement of accuracy for random noise and application to a Car Body). Presented at the
international conference on noise and vibration engineering (ISMA 2002), Leuven
159. Ziaei-Rad S, Martarelli M, Ewins DJ (2000) Measuring Rdofs using laser Doppler vibrometer.
Presented at the international conference on noise and vibration engineering (ISMA 25),
Leuven
160. Maia NMM, Neves MM, Silva, TAN (2018) Transmissibility versus damage detection.
Proceedings of international conference on noise and vibration engineering (ISMA 2018),
Leuven, Belgium
Linear Modal Substructuring
with Nonlinear Connections 22
Peter Avitabile

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1212
2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1215
2.1 Equations of Motion and Modal Space Representation . . . . . . . . . . . . . . . . . . . . . . 1215
2.2 Model Reduction and Model Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1215
2.3 Structural Dynamic Modification and System Modeling . . . . . . . . . . . . . . . . . . . . . 1216
2.4 Mode Contribution Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1217
2.5 Response of Linear Components Interconnected with Nonlinear
Connection Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218
2.6 Expansion of Transient Time Response from Reduced Order
System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218
3 Test Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1220
3.1 Nonlinear Response Prediction (Thibault and Marinone) . . . . . . . . . . . . . . . . . . . . 1220
3.2 Nonlinear Response Prediction with Expansion for Full Field
Dynamic Strain Prediction (Harvie) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1227
3.3 Nonlinear Response Prediction with Embedded Subcomponent
Models (Obando) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1232
4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1240
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1242

Abstract

Large finite element models predominate the structural dynamic community,

and inclusion of discrete nonlinear connections further complicates the models.
The need for highly reduced order, computationally efficient nonlinear models
is the focus of this work. System models generated from highly reduced

P. Avitabile ()
Structural Dynamics and Acoustic Systems Laboratory, University of Massachusetts Lowell,
Lowell, MA, USA
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 1211

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_34
1212 P. Avitabile

order components interconnected with nonlinear connection elements to predict

accurate system response with efficient approaches is identified. In addition,
expansion from the reduced order model to the full space of the original finite
element model is shown with no degradation in the predicted full-field response
for both displacement and strain for nonlinear dynamic environments.

Keywords

Nonlinear response · Nonlinear connection · Nonlinear system models ·

Reduced order model · Full-field expansion

Nomenclature

Matrix
[M] Analytical mass matrix
[K] Analytical stiffness matrix
[U] Analytical modal matrix
M Diagonal modal mass matrix
 
K Diagonal modal stiffness matrix
[T] Transformation matrix
[E] Experimental modal vectors
[I] Identity matrix
2 Diagonal matrix of ω2 values

Vector
{X} Displacement

Subscript
n Full set of finite element DOF
a Tested set of experimental DOF
d Deleted (omitted) set of DOF
G Guyan
U SEREP

Superscript
T Transpose
g Generalized inverse
−1 Standard inverse
A Component A
B Component B

1 Introduction

The structural dynamics community routinely develops extremely large finite

element models to address complicated structural systems that are used for dynamic
response predictions. Many of these models are sufficiently described by linear
models, but there is a large category of models that contain a variety of different
22 Linear Modal Substructuring with Nonlinear Connections 1213

nonlinear characteristics. A sub-class of these models contains components that

are connected with nonlinear coupling elements of different types (bilinear, slip-
stick, gap, etc.). While the components themselves are relatively linear compared
to the connection elements, the overall model becomes nonlinear, and the solution
becomes more involved and requires substantial computational resources. In addi-
tion, the connection elements may have a variety of different states depending on
the pressure, preload, temperature, etc. which then further complicates the analysis
because there may be a variety of different cases that need to be studied and
evaluated in order to determine the worst case scenario for response and strain in
the system for the various dynamic events of concern.
The use of highly accurate, reduced order component models interconnected
with nonlinear connection elements is very desirable in order to identify system
response characteristics in an efficient manner. Further the components may also
have subcomponents that are dynamically active and need to be included in the
overall system response. This may be due to the need to include the subcomponent
because there is a significant contribution to the system response or due to the need
to have an accurate description of the subcomponent loads for detailed analysis
on the subcomponent. In any event, the need for efficient reduced order models
is clear. Several different approaches and techniques have been developed over
the course of several years that are described herein and are useful for these
applications.
The first steps in this work were done to develop linear reduced order models to
form system models. Van Zandt [1, 2] developed linear reduced order component
and system models to show the accuracy with extremely small order models; these
were more traditional reduced order models. Butland [3–5] extended this work
specifically for a Craig-Bampton component synthesis model with both analytical
and experimental results. At the same time these reduced order time domain
approaches were developed, there were also companion frequency domain tech-
niques by Wirkkala [6] and Nicgorski [7–10] that complimented the time domain
approaches and included both analytical and experimental approaches. In addition to
the reduced order models and prediction of highly accurate time response, Chipman
[11–13] extended this work in an important regard which addressed the expansion
of real-time operating data collected at very small sets of measured points to a much
larger space through the use of orthogonal projection functions. Initially, Chipman’s
work was just intended to address better visualization of real-time operating data,
but then this was quickly adapted by Pingle [14–17] in order to expand the highly
reduced order model to predict full field dynamic response and dynamic strain for
the large set of finite element degrees of freedom. There were several important
papers published by Baqersad [18–21] which clearly tied all the previous work
together with linear model representations for industrial wind turbine blades where
many different response scenarios were considered ultimately including a wind
turbine in a rotating condition where a very small set of measurement sensors were
able to adequately capture the dynamic response and dynamic strain and validated
with traditional foil strain gauge measurements. Baqersad’s work clearly showed the
1214 P. Avitabile

feasibility and robustness of the approach of using a highly reduced order model for
integration with measured data that is eventually used to predict full field dynamic
response and dynamic strain. This starting body of work addressed linear response
models for both component and system model representations and included both
analytical and experimental models for these predictions.
Nonlinear models extended the concept using linear modal component and
system model work as the basis for predicting system models where the connections
between components are made up from nonlinear connection elements. Two
approaches were taken. One utilized a physical model approach and the other
utilized a modal model approach, but both provided very reduced models which
makes all of the nonlinear response model computation very efficient. Thibault [22,
23] worked with highly reduced physical models that were used to generate system
models with nonlinear connections. Marinone [24, 25] utilized a modal superpo-
sition approach coupled with a structural dynamic modification/system modeling
approach to couple components together. The approach of structural modification
relies on the ability to identify an appropriate set of modes to adequately span
the space of the solution. Utilizing concepts from structural dynamic modification
[26], both Thibault and Marinone developed an approach to identify the necessary
modes for the analysis. Both Thibault and Marinone achieved the same accuracy
in the different approaches and with significant computational savings. Thibault’s
approach is better suited for analytical modeling scenarios, whereas Marinone’s
approach works equally well with analytical and experimental data. The work
performed was mainly done at the reduced order system model level. Following this
work, Nonis [27–29] presented an expansion methodology that utilized uncoupled
component modes to determine full field mode shapes for assembled systems. This
approach [30] was adopted for the nonlinear response expansion to the finite element
degrees of freedom.
Harvie [31–33] then further extended this approach to include not only full
field nonlinear dynamic displacement but also full field dynamic strain to complete
the picture. That work addressed all the previous work and extended the work to
show that highly reduced order component models interconnected with nonlinear
connections could predict both full field dynamic displacement and dynamic
strain with a high degree of accuracy with extremely efficient computational
models.
Obando [34–37] then directed the focus to subcomponents embedded in com-
ponents used to form nonlinear system models. This work clearly showed that
important subcomponent items that have significant dynamic response can be
embedded into the complete system model representation without the need for
specific degrees of freedom to be included and that the expansion process retained
the necessary information to expand proper dynamic response information for the
subcomponent in the full space.
The theory from all this work will be summarized followed by a few of
the models that were developed to illustrate the techniques deployed and results
obtained. The referenced papers contain many more cases evaluated as part of the
work and provide further substantiation of the results shown here.
22 Linear Modal Substructuring with Nonlinear Connections 1215

2 Theory

The theory is well documented in all the referenced papers, and only basic summary
equations are presented to summarize all the work done. These mainly deal with
model reduction and model expansion along with structural dynamic modification
and system modeling concepts that are used throughout all this work.

2.1 Equations of Motion and Modal Space Representation

The general equation of motion for a multiple degree of freedom system in matrix
form is

[M1 ] {ẍ} + [C1 ] {ẋ} + [K1 ] {x} = {F(t)} (1)

The eigensolution can be performed to obtain the frequencies and the modal
matrix. The physical system can be transformed to modal space using the modal
matrix as

[U1 ]T [M1 ] [U1 ] {p̈1 } +[U1 ]T [C1 ] [U1 ] {ṗ1 } +[U1 ]T [K1 ] [U1 ] {p1 } =[U1 ]T {F(t)}
(2)

which can be written as diagonal matrices of modal mass, modal damping, and
modal stiffness as typically done and shown as
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
.. .. ..
. . .
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ M1 ⎥ {p̈1 } + ⎢ C1 ⎥ {ṗ1 } + ⎢ K1 ⎥ {p1 } = F(t) (3)
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
.. .. ..
. . .

2.2 Model Reduction and Model Expansion

The main concept behind model reduction and expansion involves the mapping
between a large model containing ‘n’ DOF and a reduced model containing a limited
‘a’ set of DOF. The general transformation between the full model and reduced DOF
is given as

xa
{xn } = = [T] {xa } (4)
xd

where the full ‘n’ space model is made up of ‘a’ master DOF and ‘d’ deleted DOF.
The transformation matrix, [T], contains the appropriate mapping information to
relate the full and reduced models. By employing energy conservation principles,
1216 P. Avitabile

the mass and stiffness of the reduced ‘a’ DOF can be calculated using the
transformation matrix with

[Ma ] = [T]T [Mn ] [T] and [Ka ] = [T]T [Kn ] [T] (5)

These reduced mass and stiffness matrices must preserve the full space model
characteristics in order for them to be useful. Many model reduction schemes
such as Guyan [38], dynamic [39], and IRS [40] all provide good reduced models
but typically contain some distortion. Other model reduction schemes such as the
system equivalent reduction expansion process (SEREP) [41], hybrid [42], and
KM_AMI [43] all provide an exact representation of the full space model for the
modes included in the reduced model approximation. For the work presented here,
SEREP was used for the development of the reduced order model. The SEREP
transformation matrix is written as

[TU ] = [Un ] [Ua ]g (6)

The SEREP reduction is generally used because of its unique characteristic in

that the full space mass and stiffness matrices (which are generally very large) are
not needed in order to form the reduced order mass and stiffness matrices; this is not
true of any other reduction scheme and provides significant benefits. The reduced
matrices are given as

 T  g 
[Ma ] = [TU ]T [Mn ] [TU ] = Ua g Ua
     (7)
T
[Ka ] = [TU ]T [Kn ] [TU ] = Ua g 2 Ua g

The reduced mass and stiffness matrices are of size ‘a,’ but rank ‘m.’ Therefore
when the number of retained modes is less than the number of retained DOF (m < a),
the reduced matrices will be rank deficient. Care must be taken when dealing with
the rank-deficient matrices. These are instances when KM_AMI would be used to
form full ranked matrices. For the models here, full rank was maintained using
a = m for all cases studied.

2.3 Structural Dynamic Modification and System Modeling

For the modeling approaches used in this work, the identification of modes needed
to address the structural dynamic response is needed. As the system response
progresses, the different nonlinear states are tracked, and the modes needed to
identify the response will change depending on the particular linear state of the
system. The structural dynamic modification and system modeling procedures are
used to change from one linear state to the next linear state. The modes of each
component are used in linear combinations to form each of the modified states. The
22 Linear Modal Substructuring with Nonlinear Connections 1217

physical system equations describing the coupling of one component to another

component can be written as

  A    A
Mn ẍn
 B  + [M12 ]
Mn ẍB
n
  A    (8)
Kn xn A fn A
+  B  + [K12 ] =
Kn xn B fn B

These components at the full space of the finite element model are reduced to
form more efficient representation of the system components which is the work
addressed by Thibault. These are then written in reduced space as

  A    A
Ma ẍa
  + [M12 ]
Ma B ẍB
a
  A    (9)
Ka xa A fa A
+  B  + [K12 ] =
Ka xa B fa B

Marinone developed essentially the same set of equations but from a modal space
perspective and utilizes structural dynamic modification for the identification of the
change of state for the system as different nonlinear characteristics occur. The basic
equation describing the change in state is

⎡⎡ ⎤ ⎤ ⎡⎡ ⎤ ⎤
.. ..
. .
⎢⎢ ⎥  ⎥ ⎢⎢ ⎥  ⎥
⎢⎢ M1 ⎥ + M12 ⎥ {p̈1 } + ⎢⎢ K1 ⎥ + K12 ⎥ {p1 } = {0}
⎣⎣ ⎦ ⎦ ⎣⎣ ⎦ ⎦
.. ..
. .
(10)

where
 
M12 = [U1 ]T [M12 ] [U1 ]
  (11)
K12 = [U1 ]T [K12 ] [U1 ]

2.4 Mode Contribution Matrix

The number of modes necessary for the dynamic response is determined from the
mode contribution matrix developed by Thibault given as

[U12 ] = [U1 ]T [M2 ] [U2 ] (12)

1218 P. Avitabile

The mode contribution matrix is used to determine which original mode shapes
are necessary for the accurate reconstruction of each of the desired final mode
shapes. If a dynamic response involves multiple system states, then the [U12 ] matrix
must be computed for each configuration to determine the number of original system
modes to appropriately span the space of the solution and avoid truncation. If a
component mode has a high value in the contribution matrix for a certain desired
system mode and that component is not included in the reduction or modes retained,
then the system mode will be in error and is said to be truncated. Conversely,
component modes with low contribution values for the desired system modes do
not participate significantly in the system modes and therefore are not necessary
to include in the solution. The mode contribution matrix is important because it
can identify the minimal set of component modes that are necessary to include in
a system model; retaining fewer modes in a reduced model can result in higher
computational savings.

2.5 Response of Linear Components Interconnected

with Nonlinear Connection Elements

The basic schematic describing Thibault’s equivalent reduced model technique

(ERMT) and Marinone’s modal modification response technique (MMRT) to
calculate the nonlinear response of the system is shown in Fig. 1; the schematic
overviews both techniques. In both techniques there needs to be a clear identification
of the modes needed to describe the system response in all the possible different
linear states (as is typically done for any structural dynamic response model). The
direct integration of equation of motion is then performed using Newmark, for
instance. However, the computation time is dramatically reduced using either the
reduced order models (ERMT) or mode superposition approach (MMRT). At each
time step, the state of the system is checked, and if there is any change in the linear
state of the system, then the model is updated to reflect the change and then the
integration process is continued; if there is no change, then the integration process
continues.

2.6 Expansion of Transient Time Response from Reduced Order

System Models

The expansion from the reduced model to the full space model (using SEREP) is

xa  g
{xn } = = [Un ] Ua {xa } (13)
xd

This equation is used for the expansion and generally is used for vector
expansion. However, Chipman and Pingle have used this for the expansion of real-
22 Linear Modal Substructuring with Nonlinear Connections 1219

ERMT MMRT

FORMULATE PHYSICAL IDENTIFY MODE FORMULATE MODAL

DATA BASE CONTRIBUTION MATRIX DATA BASE

EXCITATION
F
T F NONLINEAR

GENERATE FORCING X
CONNECTION
ELEMENT

FUNCTION & DETERMINE

EXCITATION
INITIAL CONDITIONS
F
T F NONLINEAR
CONNECTION
X
ELEMENT

COMPUTE MODAL FORCE

& MODAL INITIAL
CONDITIONS

USE M & K MATRICES PERFORM NUMERICAL UPDATE MODAL

THAT REFLECT THE INTEGRATION FOR NEXT DATABASE AND INITIAL
CURRENT SYSTEM STATE ΔT & CALCULATE DYNAMIC CONDITIONS
RESPONSE

COMPUTE PHYSICAL PERFORM STRUCTURAL

RESPONSE FROM MODE CHANGES - SDM OR
SUPERPOSITION SYSTEM MODELING

CHECK FOR
NONLINEARITY
NO NO
YES ANY CHANGE IN THE ANY CHANGE IN THE
YES
CURRENT LINEAR CURRENT LINEAR MODAL
PHYSICAL SYSTEM STATE?
STATE?

Fig. 1 Overview schematic of nonlinear response using ERMT and MMRT

time operating data for individual components. The premise being that the modes
that are the primary contributors for the modal response are the same modes to
use in the expansion of real-time operating data (made up from those same modal
contributions). Nonis further showed that any system model that is formed from
component modal representations can be expanded from reduced space to full
space using that same set of uncoupled modes used to form the system model.
Harvie further showed that this could be extended from linear system models to
nonlinear system model response under the same premise that the modes used to
form the nonlinear response are made up from various linear states. Therefore, the
expansion also works for nonlinear system response when the system is comprised
of linear components interconnected with nonlinear connection elements (using the
uncoupled component modes from the expansion process that spans the space for
all of the different nonlinear states that exist). Harvie also extended the full field
expansion to identify the full field dynamic strain in conjunction with the finite
element model using the full field displacements obtained at each time step.
Obando then extended the reduction and expansion process to include subcompo-
nent (or ancillary) information embedded into the individual components; this might
be a very dynamically active subcomponent (such as a circuit board, for instance)
which may need significant finite element fidelity to characterize the subcomponent
but then makes the component model extremely large in terms of overall model
description but is critical to the overall system response.
1220 P. Avitabile

There are several example models that have been developed for the illustration of
the techniques described. In order to evaluate the results, there are two correlation
tools used in the evaluations: the MAC and TRAC. The modal assurance criteria
(MAC) is used to compare two deformation shapes to identify the degree of
correlation; MAC values approaching 1.0 indicate very good correlation, whereas
values approaching 0.0 show no degree of correlation. The test response assurance
criteria (TRAC) is analogous to the MAC and is used to compare two time response
traces at a particular point. The MAC and TRAC are given as

 2
{xn1 }T {xn2 }
MACRA =  
{xn1 }T {xn1 } {xn2 }T {xn2 }
and (14)
 2
{xn2 }T {xn1 }
TRAC =   
{xn2 }T {xn2 } {xn1 }T {xn1 }

Note that the MACRA is used to calculate the correlation between real-valued
vectors and no complex notation is needed. In this work, the MAC is calculated
between the deformation of the full space reference solution and estimated solution
obtained from the reduced order model at each time step. Similarly the TRAC
is used to compare the time response from the reduced order model to the time
response from the full space finite element solution at each degree of freedom.

3 Test Cases

There are several test cases that were developed to illustrate the techniques
identified. The earlier linear reduced order model cases are not presented here;
essentially the nonlinear models presented here can all be reduced to simpler linear
models. The first cases illustrate the results obtained from the approaches first
developed by Thibault and Marinone. These are followed by Harvie’s work which
addresses nonlinear response and expansion to full space for dynamic response and
dynamic strain at full field. The last cases are intended to show Obando’s embedded
subcomponent work.

3.1 Nonlinear Response Prediction (Thibault and Marinone)

Due to the complexity of the nonlinearities introduced into actual structural systems,
a representative structure was fabricated to demonstrate the proposed technique.
The laboratory structure is comprised of two cantilevered aluminum beams that are
mounted to a common frame and base plate, as shown in Fig. 2. This structure
was designed to be a simplified representation of a typical structural system that
 22 Linear Modal Substructuring with Nonlinear Connections

Fig. 2 Laboratory structure representation of contact nonlinearities (middle) with beam model characteristics (left) for Beam A and Beam B and component
frequencies (right)
1221
1222 P. Avitabile

contains various contact and component interconnection features, such that the
forced nonlinear response of the system would have similarities to real-world
structures of interest.
Beam A is shown in red and Beam B is shown in blue to distinguish the
two components apart from one another. Two planar beam element models were
generated using MATLAB. Each beam model is clamped for 3 in of their length
using translational and rotational springs to replicate the cantilevered boundary
condition that is applied in the actual test fixture. Figure 2 also shows the
individual beam model information (left) – dimensions, material properties, and
modeling characteristics of the beam component models. Also shown is the resulting
component frequencies (right) for Beam A and Beam B uncoupled from each other
for the first ten bending modes of each beam component with bending about the
weaker axis; torsion was not considered for this model.
The cantilevered beam component models developed were assembled into
common matrix space as a linear system, which was discussed previously and is
shown in Fig. 3. Note that the FEM nodes for 3 in of each beam are clamped for
the cantilevered boundary condition that was applied using 10E6 lb/in translational
springs and 10E6 lb-in/rad rotational springs. This was performed to simulate the
boundary condition that exists on the actual test fixture, where the beams are
clamped between a top plate and a base block. Translational and rotational springs
were used to allow for fine-tuning of the models in order to account for possible
flexibility of the test fixture if this was a concern for the cases where test data is
introduced.
The full space linear component models were then reduced down to ‘a’ space
using SEREP where the ADOF were selected to correspond with transducer
locations on the actual tested structure. SEREP reduction was performed for the
case where the number of ADOF and modes retained in the reduced order models
are equal. Therefore, the selection of ADOF does not affect the reduced model
characteristics. In addition, the DOF at potential contact locations was retained
in the reduction process in order to formulate system models for possible contact
configurations. The full space and reduced space component models are shown in
Fig. 4, where the red points indicate the ADOF retained in the reduced order models
and the black arrow denotes the force pulse input location, which is at DOF 105
of Beam A. Figure 4 also shows the ADOF and modes retained in the reduced

Full ‘n’ Space

Beam A

Beam B

Fig. 3 Schematic of full space linear beam models

22 Linear Modal Substructuring with Nonlinear Connections 1223

F
Full ‘n’ Space
Beam A

Beam B

F
‘a’ Space
Beam A

Beam B

Reduced
# of DOF Retained Modes Retained DOF
Model
Beam A 33,41,49,57,65,73,81,89,97,101,
17 1-17
'a' Space 105,113,121,129,137,141,143
Beam B 149,151,155,163,171,179,187,
14 1-14
'a' Space 191,195,203,211,219,227,235

Fig. 4 Full space and reduced models with force pulse input location

component models; note that only translational DOF were retained in the reduced
component models.
An analytical force pulse was designed to be frequency band-limited, exciting
modes up to 1000 Hz while minimally exciting higher-order modes. The force
pulse was applied at DOF 105 of Beam A for all analytical cases studied. The force
pulse input DOF was selected to avoid being located at the node of a mode, where
minimal response of a particular mode could result. When computing the system
response to the input force pulse using the Newmark direct integration method,
the initial conditions defined at time t = 0 were zero initial displacement and zero
initial velocity. The time step (t) used was 0.0001 s. An additional parameter that
was used in the direct integration process was damping. For the analytical models
developed in this work, the modal damping ζ, for all component modes as well as
for all system modes, was assumed to be 1% of critical damping.
Although many cases were studied, the case described here consists of Beam A
coming into contact with Beam B in three different configurations at two possible
contact locations once a specified gap distance is closed between Beams A and
B, which is 0.05 in for this case. Each system is a potential configuration of
the two components depending on the relative displacements of the two beams
where no contact is also a possible configuration. The spring element used at the
contact locations was a 1000 lb/in translational spring and can be considered to be
a relatively hard contact for the system studied in this work. The spring stiffness
is applied to the full space physical model for each system configuration, and an
1224 P. Avitabile

Fig. 5 Mode contribution and frequencies of each configuration possibility

eigensolution is computed using the modified system mass and stiffness matrices.
The contact for configuration 1 occurs between DOF 141 of Beam A and DOF 191
of Beam B. The contact for configuration 2 occurs between DOF 101 of Beam A
and DOF 151 of Beam B. The contact for configuration 3 occurs when both contacts
for configurations 1 and 2 are closed simultaneously. The natural frequencies of
the modified system configurations as well as for the unmodified components are
listed in Fig. 5 along with the mode contribution matrices that are used to identify
the unmodified component modes that contribute in the modified system modes
for system configurations 1, 2, and 3, respectively. The mode contributions are
computed using full space models such that modal truncation is not of concern.
The various box colors indicate the amount that each unmodified component mode
contributes in a particular modified system mode; the actual contribution ranges for
each color are shown.
Although each mode contribution matrix is different, the matrices show that all
potential system modes are formed from the original modal database, although the
contribution from each mode changes depending on the configuration. As a result,
22 Linear Modal Substructuring with Nonlinear Connections 1225

Full ‘n’ Space Reduced‘a’ Space

Beam A Beam A

Beam B Beam B

SEREP REDUCTION
Beam A Beam A

Beam B Beam B

Beam A Beam A

Beam B Beam B

Beam A Beam A

Beam B Beam B

Fig. 6 Full space ‘n’ and reduced ‘a’ space models for two beams with multiple contacts

the transformation matrix that contains the original component mode shapes has the
necessary information to expand any of the possible states of the system. In order to
expand time response data, however, the transformation matrix used must contain an
appropriate selection of modes that are needed for accurately forming the potential
system modes.
For this case, three potential modified system configurations exist, which results
in three separate mode contribution matrices. The ‘a’ space reduced model devel-
oped for this case retains 17 DOF and modes 1 to 17 of Beam A and 14 DOF and
modes 1 to 14 of Beam B. The models developed for this case are shown in Fig. 6
for the full space and reduced ‘a’ space models. To confirm that the 31 DOF ‘a’
space models are sufficient for this case, the system response at DOF 141 is plotted
in the time and frequency domain for the reduced model with comparison to the full
space solution in Fig. 7.
The results obtained when using the 31 DOF ‘a’ space model can be observed
to correlate very well with the full space solution in Fig. 7. Two different time
correlation tools MAC and TRAC are used to quantify the similarity of the reduced
model results with the full space solution. Both plots are shown in Fig. 7 where the
y-axis scale for the MAC and TRAC plots is 0.999–1.0 so that the slight differences
between the full space and reduced space solutions could be observed. The solution
time for each model is listed with the average MAC and TRAC values in Table 1 to
show the significant decrease in computation time and highly accurate results when
the reduced models are used.
The solution obtained using the ‘a’ space model was then expanded to ‘n’ space
using the linear transformation matrix developed in Fig. 6. Figure 8 shows the MAC
and TRAC plots over the first 0.2 s (note that the TRAC is now calculated for all 276
DOF). The y-axis scale for the MAC and TRAC plots is 0.9–1.0 so that the slight
differences between the full space and reduced space solutions could be observed.
The 276 DOF ‘a’ space expanded model for the two-beam system with multiple
hard contacts produces very good results as well.
The solution times for all models are listed with the average MAC and TRAC
values in Table 1 to show the significant decrease in computation time, and highly
accurate results are obtained when the reduced models are used. The favorable MAC
and TRAC results in Table 1 show that the 31 DOF ‘a’ space model is sufficient
1226 P. Avitabile

Fig. 7 Reduced model time response (DOF 141) compared to reference solution (left) with MAC
of reduced order model deformed shape compared to reference deformed shape for 31 ADOF
(middle) and TRAC correlation for 31 ADOF (0–0.2 s) – two beams with multiple hard contacts

Table 1 Solution times and average MAC/TRAC for two beams with multiple hard contacts –
transient portion of time response (0–0.2 s)
Solution time
Model # of DOF (sec) Average MAC Average TRAC
‘n’ space 276 88.5 0.99996 0.99958
‘a’ space 31 4.9
‘n’ space 276 88.5 0.99996 0.99973
‘a’ space expanded 276 5.1

for accurately computing the initial transient portion of the time response for this
particular case. In addition, the solution time for the full space model is 88.5 s in
contrast to the reduced ‘a’ space model, which is only 4.9 s.
Table 1 alsolists the average MAC and TRAC for each solution along with the
solution time (the solution time for the expanded ‘a’ space is the time required to
solve the ‘a’ space model and then expand to ‘n’ space). Even with the additional
calculation due to multiplying the time solution by the expansion matrix, the time
required is significantly reduced in comparison to the full ‘n’ space solution, going
22 Linear Modal Substructuring with Nonlinear Connections 1227

Fig. 8 MAC (left) and TRAC (right) comparisons between 276 DOF ‘n’ and 276 DOF ‘a’
expanded space models for transient portion of time response (0–0.2 s) – two beams with multiple
hard contacts

from 88.5 to 5.1 s. The average MAC and TRAC have negligible improvements
because the remaining error is due to truncation which is not improved by expansion
of the time responses.
Table 1 shows that expanding the ‘a’ space time response to ‘n’ space yielded an
accurate solution while significantly reducing the time required. Rather than solving
the full space solution directly, reducing the model and then expanding the solution
yielded nearly identical results at a fraction of the time.
The other key point of this case study is that the nonlinear time response was
achieved using the modes of the originally uncoupled component beam models.
Even though the linear mode shapes used in expanding the time responses were not
identical to the mode shapes used when the configuration was in a different linear
state, a separate expansion matrix for each nonlinear perturbation was not required.
Because a sufficient number of modes have been used to span the space of the
analytical model and all perturbed nonlinear variations of the model, the expansion
does not distort the data and allows for very accurate expansion for all perturbed
configurations of the two-beam system. As a result, the highly accurate expanded
time solution was obtained very efficiently. In addition, an actual measured test
configuration was also studied and produced results comparable in efficiency and
accuracy to the analytical results presented; these experimental results are included
in the references.

3.2 Nonlinear Response Prediction with Expansion for Full Field

Dynamic Strain Prediction (Harvie)

A finite element model was generated using Abaqus/CAE [38]. The two-beam
system, as illustrated in Fig. 8, was generated to imitate a large, complicated model
to accurately demonstrate the principles at hand while maintaining a feasible model
size on which reference calculations are performed; note that the diagram shown in
1228 P. Avitabile

Property Beam A Beam B

Width 50.8 mm (2 in) 50.8 mm (2 in)

Geometric
Height 25.4 mm (1 in) 25.4 mm (1 in)
Length 1.27 m (50 in) 3.56 m (140 in)
# elements 250 700

Modeling
# nodes 251 701
# DOF 502 1402
Node
0.2 0.2
spacing
1.27 m
2.03 m
(50 in) Material Aluminum Aluminum
F
F (80 in)

1.75 GN/m 7.03 kg/m3 7.03 kg/m3

Material
Beam A 1.75 GN/m
Density 3 3
(10e6 lb/in) (10e6 lb/in) (2.54e-4 lb/in ) (2.54 e-4 lb/in )
1.75 MN/m Beam B 1.75 MN/m
(10e3 lb/in) (10e3 lb/in) Young's 68.9 GPa 68.9 GPa
3.56 m
(140 in) Modulus (10e6 psi) (10e6 psi)

Fig. 9 Physical representation of two-beam system with geometric, modeling, and material
properties of two-beam system

Fig. 9 is not to scale. The main beam, Beam B, is 140 in in length and joined to
the smaller Beam A, 50 in in length, using 10,000,000 lb/in translation springs; the
main beam is grounded using 10,000 lb/in translation springs. The full space model
contains nodes with 0.2 in spacing on each beam; therefore there are 251 nodes on
Beam A and 701 nodes on Beam B. Each node contains a shear DOF and a rotational
DOF to capture planar beam bending only. Details on the properties of the structure
can be found in Fig. 9; note that the modeling properties are only applicable for the
large N-space model.
To introduce discrete nonlinearities to the system, a gap-spring interface was used
to simulate a contact; the stiffness of the spring contact is either set to a predefined
stiffness value when the specified gap distance is closed or set to zero when the
specified gap distance is open. The nonlinear cases have two contact locations
between the beams, as shown in Fig. 10; these contact locations were chosen so
that both contact springs could engage during the response. The initial gap distance
was set to 0.003 in for the nonlinear cases; once again this value was chosen merely
so that both contact springs would engage during the response. The three possible
configurations that the beams can encounter with the springs engaged are also shown
in Fig. 10 along with the frequencies for each configuration.
For all cases, a frequency band-limited analytical force pulse was utilized to
excite a frequency range of roughly 400 Hz; the frequency range excited by the
force pulse includes roughly 13 modes in all configurations. The use of this force
pulse allows for minimal excitation of higher modes and controls the number of
modes that substantially participate in the system response due to the impulse.
The force was applied at the left-most node of Beam B, and the force was applied
perpendicular to the beam to excite modes along the weak axis; this approximation
22 Linear Modal Substructuring with Nonlinear Connections 1229

Frequency (Hz)
Mode
Original Config 1 Config 2 Config 3

1 3.88 4.14 4.07 4.14

2 14.61 15.01 15.05 15.08

3 26.91 27.18 26.97 27.34

4 38.37 43.42 43.11 43.49

Original State
5 44.03 67.08 66.73 67.10
12 in 20 in 18 in
6 67.10 76.99 92.65 98.80

7 96.85 99.19 113.84 144.79

8 141.42 145.76 144.79 187.79

Configuration 1
9 153.93 190.90 189.58 208.79

10 190.93 227.85 229.99 234.73

Configuration 2 11 255.93 262.04 264.22 269.05

12 322.30 325.28 328.99 329.20

13 335.25 383.76 338.36 391.88

Configuration 3
14 404.54 404.55 406.64 407.48

15 496.67 499.47 497.10 499.56

16 584.91 585.47 590.07 591.58

17 590.09 591.95 615.35 615.52

18 707.96 707.97 708.67 708.68

19 814.95 815.99 816.16 816.98

20 901.54 907.93 909.59 915.61

21 949.60 950.01 950.13 950.71

22 1078.50 1078.53 1078.71 1078.73

Fig. 10 Contact locations and configurations for nonlinear case with natural frequencies to
1000 Hz of full space system with and without hard contacts

was made to demonstrate the principles at hand and could be extended to different
forcing functions and locations.
The large N-space model was reduced from 1904 modes and DOF to an n-space
model with 194 modes and DOF by retaining every tenth node from the full space
model. A comparison of model sizes for this reduction is shown in Fig. 11. All nodes
in both models contain both shear and rotary DOF. This reduction was performed to
produce a more reasonable-sized model.
The n-space model was reduced further for use with ERMT. Two model sizes, a
space and aa-space, were generated for the various analyses performed. The a-space
model contains 24 modes and DOF, while the aa-space model contains only 13
modes and DOF, as outlined in Fig. 11; also shown are the specific DOF and modes
used in the models. The a-space model includes modes of the original system up
1230 P. Avitabile

# Nodes # DOF # Modes Active DOF

N-space
952 (shear & rotary
1904 1904 1-1904
with 0.2” spacing)

n-space
97 (shear & rotary 1-2, 21-22, 41-42,
194 194
with 0.2” spacing) …1903-1904

1,101,201,301,401,501,
a-space 601,701,801,861,921,
24 (shear, variable
24 24 1021,1121,1301,1341,
spacing, max 10”)
1401,1403,1463,1523,
1623,1723,1803,1903
aa-space
1,201,401,601,801,
13 (shear, variable
13 13 921,1121,1301,1401,
spacing, max 20”)
1403,1523,1723,1903

Fig. 11 Description of full space model and reduced order models used

to approximately 1300 Hz, while the aa-space model includes modes up to only
400 Hz.
The mode contribution matrices were calculated at full space for the modified
system in all configurations, and the matrices are not shown for brevity. The a-space
model with 24 modes contains enough modal information to span the space of the
problem for all configurations encountered.
Direct integration of the equations of motion was performed using Newmark
time integration to compute the time response for all cases. Newmark integration
was utilized for similarity to the solver used in Abaqus, where the Hilber-Hughes-
Taylor (HHT) variation of the Newmark method is used. The damping of the system
was approximated using one percent of critical damping for all modes. Proportional
damping was assumed to keep a straightforward solution procedure, but a state space
solution could be used to solve systems with non-proportional damping.
The full space displacements were determined using both a full space solution
and the proposed efficient technique. Comparisons of the computation time and
accuracy for this case are shown in Table 2. The solution time was reduced
significantly by utilizing such a smaller model to solve for the response. For
this case, very high accuracy was obtained using an extremely reduced model.
The displacement results for the full and predicted displacement are shown in
Fig. 12 along with the strain. The frequency range of the original system modes
included in the reduction extends over 1300 Hz, so the energy distribution of the
response is accurately captured in the frequency domain. Both the displacement
and strain are predicted accurately using the reduced order model implemented in
this case. The model used in the efficient technique is able to better predict the
higher-order curvature in the dynamic displacement and strain response because
more modes were included in the reduction. While the efficient calculation of the
dynamic displacement matches nearly perfectly to the reference solution, some
high-frequency content is present in the reference strain solution that is not captured
22 Linear Modal Substructuring with Nonlinear Connections 1231

Table 2 Comparison of reduced and full solution for hard contact case with more modes
Model # of DOF Solution time (sec) Average MAC Average TRAC
Full space 1904 740.18 0.9998 0.9999
Reduced 24 0.28

Displacement on Beam A with Hard Contacts Strain on Beam A with Hard Contacts
0.004 Reference
Expanded Efficient
15 Calculation
(24 modes)
0.002 10

Strain (Microstrain)
Dis plac ement (in)

5
0.000

-0.002
-5

-10
-0.004

-15
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time (s) Time (s)

Fig. 12 Displacement and corresponding strain on Beam A for hard contact case

using the efficient calculation. The slight differences in the strain calculations can
be further reduced by including even more modes in the reduced order model. The
strain and displacement results calculated efficiently compare very accurately to the
full space model for this case, with a solution time that is dramatically reduced
compared to the full space solution.
The results obtained from this case compared very well with the full space
solution because more modes were retained in the reduced solution to accurately
represent the system dynamics for all possible configurations. The full solution took
over 12 min to compute, but accurate strain and displacement data could be obtained
in less than a second without compromising accuracy. Although the addition of hard
contact springs in the system causes the necessity for additional modes to be retained
in a reduced order model, the full space model can still be substantially reduced to
retain only the dynamic characteristics that are necessary to the response.
1232 P. Avitabile

3.3 Nonlinear Response Prediction with Embedded

Subcomponent Models (Obando)

Further extending the models and results presented in the previous cases, this case
will introduce a modally active embedded component to one of the components and
will be included in the system response with no active degrees of freedom on the
subcomponent with nonlinear response predicted with highly reduced component
models.
Multi-component structural systems were addressed in the context of retaining
embedded structural information of ancillary subcomponents for the calculation of
nonlinear reduced order model time response. For linear forced response, expansion
was shown to return a precise approximation of the ancillary subcomponent even in
cases where the reduction does not include active DOF at that subcomponent level.
The model was developed as a natural extension of the nonlinear work discussed
in the last cases. A full space finite element model consisting of two systems,
one of which contains a dynamically active ancillary subcomponent, were reduced
to a smaller set of degrees of freedom and used for the prediction of the forced
time response of the system as seen in Fig. 13. Gap-spring contact elements were
introduced to generate nonlinear response between the two systems. The reduced
order model (with embedded ancillary subcomponent information) was then used
to calculate the response at ADOF and then to expand back to the full space
finite element model and extract the predicted forced response of the ancillary
subcomponent. The results for the linear and the nonlinear hard and soft contact
cases will be discussed here, but further information can be found in the references.
In the linear case, the SEREP model will be discussed. A reduced model was
computed such that no DOF from the ancillary subcomponent (gray beam at the

System/Assembly 2
NDOF Subcomponents

Assemble
Unreduced
Systems
System/Assembly 1

Connect System 1 & 2 and

perform Reduction to ADOF

Extract
Subcomponent
characteristics Calculate Forced
Response and
Expand back to
NDOF

Fig. 13 Sequence for the development of reduced system response models

22 Linear Modal Substructuring with Nonlinear Connections 1233

Reference System SEREP Reduced %

Freq. (Hz) System Freq. (Hz) Difference

Mode 1 16.8 16.8 0%

Mode 2 37.5 37.5 0%

Mode 3 68.1 68.2 0%

Mode 4 84.6 84.6 0%

Mode 5 102.1 102.1 0%

Mode 6 129.1 130.2 1%

Mode 7 210.0 211.5 1%

Mode 8 282.0 283.2 0%

Mode 9 343.0 351.2 2%

Mode 10 396.2 396.6 0%

System 2: 7 ADOF – 23, 31, 33, 41, 43, 55, and 63 447%
Mode 11 477.3 2611.8
System 1: 5 ADOF – 65, 117, 169, 199 and 205 Mode 12 645.3 3023.1 369%

Fig. 14 Comparison of SEREP reduced order model (12 DOF) frequencies with respect to (206
DOF) the reference solution. “x” indicates relative location of nodes in FEM

top of the red beam) was preserved in the reduction process. The model is shown
in Fig. 14 and a comparison of the response as well as correlation results of the
expansion process in Fig. 15.
The SEREP reduction and expansion process resulted in high correlation with
full space model using only a very small fraction of the DOFs. Furthermore, the
omission of the connecting DOF for the ancillary subcomponent did not yield any
additional error. Addition of modes beyond the 12 modes indicated in the [U12 ]
matrix in the SEREP reduced model showed large improvement from the resulting
expanded model response as illustrated in Fig. 16 where expansion is used on a
17 mode and DOF reduced model. When the reduction process is successful (as
it was with the SEREP and KM_AMI models), the modes selected span the space
of the system response. Modes beyond the 12 indicated by the [U12] smooth the
approximation of the system response and further addition of modes results in better
results until the reduced ‘a’ space model approaches the full ‘N’ space solution and
hence spans the whole space of the full assembled system response.
The following models expand the reduced order modeling of the three-beam
system to include nonlinear response. Gap-spring contact elements are introduced
to generate nonlinear response between the two systems as done in the first example
shown here. However, the possibility of nonlinear contacts is introduced at the
subcomponent level as shown in Fig. 17. For the first nonlinear case, four different
configurations are possible (initial configuration, one configuration for each spring
coming in contact, and one for both springs in contact) as seen in Fig. 17. Using
the calculated [U12 ] for all configurations, the necessary number of component
modes were determined in order to properly characterize the system; using the
force described above, the system modes must be able to characterize the response
over a 250 Hz frequency span. While this is the initial frequency range of interest
for any structural dynamic study, there also needs to be a consideration for the
nonlinear contacts which occur that may require a frequency range beyond that
initially determined from the applied forcing function.
1234 P. Avitabile

-4
x 10 Time Domain Response of Ancillary Beam at Node 1
6
Guyan Red. Model
Reference Model
4
Dis plac ement at Node 1

-2

-4
Guyan Red. Model
Reference Model

-6
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Time (sec) 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time (sec)

Fig. 15 Left: Comparison of time response at node 1 of the ancillary subcomponent from the
expansion of a 12DOF SEREP reduced model versus the 206 DOF full space reference solution (in
magenta). Zoomed in region shows the response for the first 0.1 s. Right: MAC (blue) and TRAC
(red) bar plots showing the correlation of the expanded SEREP reduced model to the reference
model

In these nonlinear cases, the location of the nonlinearity can have a significant
effect on the accuracy of the prediction and on the number of modes (and DOF)
required in the reduction/expansion process. SEREP reduction was used to reduce
the active DOF of the system to an ‘a’ set not including DOFs on the ancillary beam.
The forced responses of the reduced ADOF systems were computed. The dynamic
characteristics of the ancillary subcomponent were then extracted using the system
information available from the reduction process. A SEREP reduced model using
16 modes was created. Figure 18 shows the comparison of the predicted response
of the expanded model and the reference model.
22 Linear Modal Substructuring with Nonlinear Connections 1235

MAC (0.95-1.00)
1

0.995

0.99

0.985

0.98

0.975

0.97

0.965

0.96

0.955

0.95
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time [s]

TRAC
1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
NDOF

Fig. 16 MAC and TRAC plots showing correlation of the expanded 17 DOF SEREP reduced
model to reference model. MAC y-axis is showing values from 0.95 to 1.0
1236 P. Avitabile

Case A Case B
gap
gap

Original Configuration 1 Configuration 2 Configuration 3

State

Nonlinear Contacts

Fig. 17 Nonlinear gap-spring contact cases analyzed. The nonlinear configurations for Case A are
shown. Each case indicates the closing or opening of the gap springs (in green when closed)

-3
x 10 Time Domain Response at Ancillary Node 1
2
Ref. Model
1.5 Reduced Model

0.5
Dis plac ement Node 1

-0.5

-1

-1.5

-2

-2.5

-3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Time (sec)

Reduced
Model

16 DOF
Beam A – ADOF 23, 31, 35, 39, 43, 47, 55 and 63.
Beam B – ADOF 65, 99, 135, 169, 179, 189, 199, and 205.

Fig. 18 Predicted response at ancillary node 1 (in yellow) from 16 DOF reduced model. MAC
from 0.95–1.0 (top right) and TRAC (bottom right) correlation of models

Using 16 modes allows for the accurate prediction of up to 8 modes (which spans
approximately 300 Hz) of the three-beam assembled system. Figure 19 shows the
comparison of the FFT of the displacement at the left spring contact location. The
FFT of the displacement at the left contact location in Fig. 19 shows no significant
effect of mode truncation of the higher-order modes, and the response is an accurate
22 Linear Modal Substructuring with Nonlinear Connections 1237

x 10-3 Time Response at Contact 1 FFT of Time Response at Contact 1

3 -120
Ref. Model Ref. Model
Red. Model Red. Model
-140
2

-160
1

dB Displacement
Displacement

-180
0
-200

-1
-220

-2
-240

-3 -260
0 0.05 0.1 0.15 0 100 200 300 400 500
Time (sec) Frequency (Hz)

Fig. 19 Comparison of response (left) and FFT (right) of reference model and 16 mode reduced
order model at spring contact location 1

approximation of the full NDOF solution. Note that no ADOF has been placed
at the ancillary subcomponent or at the DOFs that connect the ancillary beam to
the top beam. Nevertheless, the embedded information in the reduction process has
successfully allowed the prediction of the time response at all NDOF.
The effect of higher-order modes can be exacerbated if the type of contact is a
hard impact, thus producing a narrow time pulse that translates into a high-order
pulse in the frequency domain. The next case will explore a hard contact spring
acting at two locations of the structure. A comparison of the two types of contacts,
soft and hard, can be seen in Fig. 20.
Nonlinear contacts of 10,000 lb/in were implemented instead of the soft contacts
of 100 lb/in previously used. Because the hard contact excites frequencies in
the range of 700 Hz, the 16 mode model previously used cannot give the best
approximation of the response of the system. The selection of modes preserved in
the reduction must form a linearly independent set of vectors spanning the space of
the full response of the system. In other words, the selected projection vectors in the
transformation matrix must be able to approximate any other vector in the space as
linear combinations of the mode shape vectors preserved in the reduced space. A 21
mode and DOF model was created to better approximate the hard contact nonlinear
response of the system. Figure 21 shows the correlation and response for the 21
DOF model and the reference model.
The modes used in the reduced model resulted in an accurate approximation of
the response of the NDOF system. However, the effects of mode truncation could
not be completely mitigated. Caution must be exercised when simply adding modes
to the reduced order model since linear independence of the vectors formed in the
expansion process is very important. The reduced space matrices can become ill-
conditioned for certain choice and number of DOF, and the predicted response is
1238 P. Avitabile

-3
x 10 Time Response and Force Pulse for Soft Spring at Contact Location 1
5 0.1
Dis plac ement

Force (lbf)
0 0

-5 -0.1
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Time (s)
-3
x 10 Time Response and Force Pulse for Hard Spring at Contact Location 1
2 4

1 3

0 2
Dis plac ement

Force (lbf)

-1 1

-2 0

-3 -1
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Time (s)

Fig. 20 Comparison of displacement and force magnitude for soft (left) and hard (right) types of
nonlinear contacts at DOF 37 of the three-beam system at contact location 1 (left gap spring)
22 Linear Modal Substructuring with Nonlinear Connections 1239

-3 MAC
x 10 Time Domain Response at Ancillary Node 1 1
2
Ref. Model 0.9
1.5 Reduced Model
0.8

1 0.7

0.6
0.5
Dis plac ement Node 1

0.5

0 0.4

-0.5 0.3

0.2
-1
0.1

-1.5 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time [s]
-2
TRAC
1
-2.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.9
Time (sec)
0.8
Reduced
0.7
Model
0.6

0.5

0.4

0.3
21 DOF
Beam A – ADOF 31, 33, 39, 43, 47, 55, 57 and 61. 0.2

Beam B – ADOF 65, 79, 89, 99, 109, 119, 149, 159, 169, 0.1
179, 189, 199 and 205.
0
0 20 40 60 80 100 120 140 160 180 200
NDOF

Fig. 21 Predicted response at ancillary node 1 from expansion of 21 DOF reduced order model
with hard contacts. MAC (top right) and TRAC (bottom right) correlation of models

then subjected to high levels of numerical error. Figure 22 shows that the FFT and
response at the contact location 1 is reasonably approximated by the reduced model
of 21 modes.
The cases considered have shown that full field results can be obtained from
reduced order models with subcomponent interactions from the embedded informa-
tion preserved in the reduction process. Using the necessary number of modes in
the reduced model to span the space of all modes of interest allows the response at
the ancillary and any other DOF to be predicted accurately. The [U12 ] contribution
matrix and the effect of the higher-order modes from nonlinear interactions must be
taken into account as well as the linear independence and well-determined behavior
of the reduced matrices in order to obtain a good approximation of the dynamic
characteristics of the system. Moreover, when proper precautions are taken to reduce
the N space system to ‘a’ space, no ADOF are needed in the dynamically active
ancillary subcomponent.
Additional cases have been studied to investigate the nonlinear connection
between the subcomponent and the attached component. Generally these cases
produced similar results and are not presented here due to space limitations.
1240 P. Avitabile

-3
x 10 Time Response at Contact 1 FFT of Time Response at Contact 1
3 -100
Ref. Model Ref. Model
Red. Model Red. Model
2
-150

dB Displacement (in)
1
Displacement (in)

-200

-250
-1

-300
-2

-3 -350
0 0.05 0.1 0.15 0 200 400 600 800
Time (sec) Frequency (Hz)

Fig. 22 Comparison of response (left) and FFT (right) of reference (in red) and expansion of 21
DOF reduced order model at contact location 1 (left gap spring)

4 Conclusions

Understanding of the linear reduction/expansion methodology showed the advan-

tage of using reduced models for full field prediction of systems that may undergo
localized nonlinear response due to component contact or coupling interactions. For
these types of problems, the presence of highly nonlinear coupling elements such as
hard contacts, isolation mounts, gap springs, bilinear springs, etc. is predicted with
a piecewise linear approximation providing the dynamic response of the structure
as a superposition of all possible configurations of the system.
For all cases considered, linear and nonlinear, high correlation values were
obtained between the reference model and the predicted expanded model response
from the reduced models. For such models, the computation time was greatly
reduced as well as the number of DOF in the model. There was no degradation
in the predicted full field dynamic response or the full field dynamic strain. The
models clearly showed that very good and very efficiently calculated results could
be obtained for linear components interconnected with highly nonlinear connection
elements.

Author Contributions
This work has spanned several years and several student theses that are acknowl-
edged in this section.
22 Linear Modal Substructuring with Nonlinear Connections 1241

Van Zandt, T., Development of Efficient Reduced Models for Multi-Body Dynamics
Simulations of Helicopter Wing Missile Configurations, MS, University of
Massachusetts Lowell, Lowell, Massachusetts, 2004
Wirkkala, N., Efficient Method of Model Updating Using Target Definition from
Impedance Based Modeling for Multi-Body Dynamics Simulations, MS, Uni-
versity of Massachusetts Lowell, Lowell, Massachusetts, 2004
Butland, A., A Hybrid Component Mode Synthesis Approach for System Modeling
Applications, MS, University of Massachusetts Lowell, Lowell, Massachusetts,
2006
Nicgorski, D., Investigation on Experimental Issues Related to Frequency Response
Function Measurements for Frequency Based Substructuring, MS, University of
Massachusetts Lowell, Lowell, Massachusetts, 2006
Chipman, C., Real Time Operating Data Expansion for Improved Visualization, MS,
University of Massachusetts Lowell, Lowell, Massachusetts, 2006
Pingle, P., Prediction of Full Field Dynamic Stress/Strain from Limited Sets of Mea-
sured Data, PhD, University of Massachusetts Lowell, Lowell, Massachusetts,
2010
Thibault, L., Development of Equivalent Reduced Order Model Technique for
Linear Modal Components Interconnected with Nonlinear Connection Elements,
MS, University of Massachusetts Lowell, Lowell, Massachusetts, 2012
Marinone, T., Efficient Computational Nonlinear Dynamic Analysis using Modal
Modification Response Technique, MS, University of Massachusetts Lowell,
Lowell, Massachusetts, 2012
Harvie, J., Computationally Efficient Reduced Order Models for Full Field Non-
linear Dynamic Strain Predictions, MS, University of Massachusetts Lowell,
Lowell, Massachusetts, 2013
Baqersad, J., A Non-contacting Approach for Full Field Dynamic Strain Monitor-
ing of Rotating Structures using Photogrammetry, Finite Element and Modal
Expansion Techniques, PhD, University of Massachusetts Lowell, Lowell, Mas-
sachusetts, 2015
Obando, S., Use of Expansion of Highly Reduced Order Models for the Accurate
Prediction of Full Field Dynamic Characteristics in the Forced Response of Lin-
ear and nonlinear Systems and Components, PhD, University of Massachusetts
Lowell, Lowell, Massachusetts, 2016

Acknowledgments Some of the work presented herein was partially funded by Redstone
Arsenal Award No. W31P4Q-07-P-0606 entitled “System Modeling Approaches of Heli-
copter/Wing/Missile Configurations,” by NSF Civil, Mechanical and Manufacturing Innovation
(CMMI) Grant No. 0900543 entitled “Dynamic Stress-Strain Prediction of Vibrating Structures in
Operation,” and by Air Force Research Laboratory Award No. FA8651-10-1-0009 “Development
of Dynamic Response Modeling Techniques for Linear Modal Components.” Any opinions,
findings, and conclusions or recommendations expressed in this material are those of the authors
and do not necessarily reflect the views of the particular funding agency. The authors are grateful
for the support obtained.
1242 P. Avitabile

References
1. Van Zandt T et al (2006) Development of efficient reduced models for rigid body dynamics
simulation for helicopter missile wing combinations. In: Proceedings of the twenty-fourth
international modal analysis conference, Feb 2006, St Louis
2. Avitabile P et al (2007) Development of efficient reduced models for flexible body dynamic
simulation. In: Proceedings of the twenty-fifth international modal analysis conference, Feb
2007, Orlando
3. Butland A, Nicgorski D, Avitabile P (2008) A hybrid craig-bampton approach for system
modeling applications. In: Proceedings of the twenty-sixth international modal analysis
conference, Feb 2008, Orlando
4. Butland A, Avitabile P (2009) Component mode synthesis using reduced order test verified
components. In: Proceedings of the twenty-seventh international modal analysis conference,
Feb 2009, Orlando
5. Butland A, Avitabile P (2010) A reduced order, test verified component mode synthesis
approach for system modeling applications. Mech Syst Signal Process 24(4):904–921
6. Wirkkala N, Avitabile P, Van Zandt T (2006) Modification of target definition for model
updating using impedance based modeling for helicopter missile wing combinations. In:
Proceedings of the 24th international modal analysis conference, Feb 2006, St. Louis
7. Nicgorski D, Butland A, Avitabile P (2008) Experimental issues related to frequency response
function measurements for frequency based substructuring. In: Proceedings of the 26th
international modal analysis conference, Feb 2008, Orlando
8. Nicgorski D, Avitabile P (2009) Conditioning of FRF measurements for use with frequency
based substructuring. In: Proceedings of the twenty- seventh international modal analysis
conference, Feb 2009, Orlando
9. Nicgorski D, Avitabile P (2010) Experimental issues related to frequency response func-
tion measurements for frequency-based substructuring. Mech Syst Signal Process 24(5):
1324–1337
10. Nicgorski D, Avitabile P (2010) Conditioning of FRF measurements for use with frequency
based substructuring. Mech Syst Signal Process 24(2):340–351
11. Chipman CC, Avitabile P (2008) Expansion of real time operating deformations for improved
visualization. In: Proceedings of the twenty-sixth international modal analysis conference, Feb
2008, Orlando. ProQuest
12. Chipman C, Avitabile P (2009) Expansion of transient operating data. In: Proceedings of the
twenty- seventh international modal analysis conference, Feb 2009, Orlando
13. Avitabile P, Chipman C (2012) Expansion of transient operating data. Mech Syst Signal Process
31:1–12
14. Pingle P, Avitabile P (2010) Prediction of full field dynamic stress/strain from limited sets of
measured data. In: Proceedings of the twenty-eighth international modal analysis conference,
2010, Jacksonville. Springer
15. Pingle P, Avitabile P (2011) Limited experimental displacement data used for obtaining full-
field dynamic stress strain. In: Proceedings of the twenty-ninth international modal analysis
conference, 2011, Jacksonville. Springer
16. Pingle P, Avitabile P (2011) Full field dynamic stress/strain estimation from limited sets of
measured data. In: Proceedings of the twenty-ninth international modal analysis conference,
Feb 2011, Jacksonville. Springer
17. Avitabile P, Pingle P (2012) Prediction of full field dynamic strain from limited sets of
measured data. Shock Vib 19(5):765–785
18. Baqersad J, Niezrecki C, Avitabile P (2015) Predicting full-field strain on a wind turbine for
arbitrary excitation using displacements of optical targets measured with photogrammetry
technique. In: Proceedings of the thirty-third international modal analysis conference, Feb
2015, Orlando. Springer
22 Linear Modal Substructuring with Nonlinear Connections 1243

19. Baqersad J et al (2015) Extracting full-field dynamic strain response of a rotating wind turbine
using photogrammetry. In: SPIE/NDE, March 2015, San Diego. International Society for
Optics and Photonics
20. Baqersad J, Niezrecki C, Avitabile P (2015) Full-field dynamic strain prediction on a wind
turbine using displacements of optical targets measured by stereophotogrammetry. Mech Syst
Signal Process 62:284–295
21. Baqersad J, Niezrecki C, Avitabile P (2015) Extracting full-field dynamic strain on a wind
turbine rotor subjected to arbitrary excitations using 3D point tracking and a modal expansion
technique. J Sound Vib 352:16–29
22. Thibault L et al (2012) Equivalent reduced model technique for nonlinear dynamic response.
In: Proceedings of the thirtieth international modal analysis conference, Feb 2012, Jacksonville
23. Thibault L et al (2013) Equivalent reduced model technique development for nonlinear system
dynamic response. Mech Syst Signal Process 36(2):422–455
24. Marinone T et al (2012) Efficient computational nonlinear dynamic analysis using modal
modification response technique. In: Proceedings of the thirtieth international modal analysis
conference, Feb 2012, Jacksonville
25. Marinone T et al (2012) Efficient computational nonlinear dynamic analysis using modal
modification response technique. Mech Syst Signal Process 31:67–93
26. Avitabile P, O’Callahan J (1991) Understanding structural dynamic modification and the effects
of truncation. Proc Int J Anal Exp Modal Anal, invited contribution 6(4):215
27. Nonis C et al (2013) Development of full space system model modes from expansion of
reduced order component modal information. In: Proceedings of the thirty-first international
modal analysis conference, Feb 2013, Garden Grove
28. Nonis C, Avitabile P, Obando SE (2013) System model modes developed from expansion of
uncoupled component dynamic data. In: Recent advances in structural dynamics, July 2013,
Pisa
29. Avitabile P, Nonis C, Obando SE (2014) System model modes developed from expansion of
uncoupled component dynamic data. J Mech Eng 60(5):287–297
30. Marinone T, Thibault L, Avitabile P (2013) Expansion of nonlinear system response using
linear transformation matrices from reduced component model representations. In: Proceedings
of the thirty-first international modal analysis conference, Feb 2013, Garden Grove. Springer
31. Harvie J, Obando SE, Avitabile P (2013) Reduced order system model nonlinear response and
expansion for full field results. In: Recent advances in structural dynamics, July 2013, Pisa
32. Harvie J, Avitabile P (2014) Computationally efficient nonlinear dynamic analysis for
stress/strain applications. In: Proceedings of the thirty-second international modal analysis
conference, Feb 2014, Orlando. Springer
33. Avitabile P, Harvie J, Obando SE (2014) Efficient reduced order nonlinear response with
expansion for full field results. In: EURODYN 2011, Porto
34. Obando SE, Avitabile P (2014) Prediction of forced response on ancillary subsystem com-
ponents attached to reduced linear systems. In: Proceedings of the thirty-second international
modal analysis conference, Feb 2014, Orlando. Springer
35. Obando SE, Avitabile P (2015) Prediction of forced response using expansion of perturbed
reduced order models with inexact representation of system modes. In: Proceedings of the
thirty-third international modal analysis conference, Feb 2015, Orlando. Springer
36. Obando SE, Avitabile P (2015) Prediction of nonlinear forced response on ancillary subsystem
components attached to reduced linear systems. In: Proceedings of the thirty-third international
modal analysis conference, Feb 2015, Orlando
37. Obando SE, Avitabile P (2016) Developments in the prediction of full field dynamics for the
nonlinear forced response of reduced order system models. In: Proceedings of the thirty-fourth
international modal analysis conference, Feb 2016, Orlando
38. Guyan RJ (1965) Reduction of stiffness and mass matrices. AIAA J 3(2):380–380
39. Paz M (1984) Dynamic condensation. AIAA J 22(5):724–727
1244 P. Avitabile

40. O’Callahan JC (1989) A procedure for an improved reduced system (IRS) model. In:
Proceedings of the 7th international modal analysis conference. Union College, Schenectady
41. O’Callahan J, Avitabile P, Riemer R (1989) System equivalent reduction expansion process
(SEREP). In: Proceedings of the 7th international modal analysis conference. Union College,
Schnectady
42. Kammer DC (1991) A hybrid approach to test-analysis-model development for large space
structures. J Vib Acoust 113(3):325–332
43. Marinone T, Butland A, Avitabile P (2012) A reduced model approximation approach
using model updating methodologies. In: Topics in modal analysis II, vol 6. Springer, New
York/London, pp 625–636
 Part IV
Applications and Miscellaneous Topics
Civil Structural Testing
23
Ruben Boroschek and Joao Pedro Santos

Contents
1 Why Structural Health Monitoring (SHM) for Civil Structures . . . . . . . . . . . . . . . . . . . 1250
2 Basic Considerations on Loading and Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1251
3 Data Acquisition and Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1251
3.1 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1251
3.2 Selection of Data Acquisition Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1257
4 Basic Procedures for SHM in Civil Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1258
4.1 Initial Conditions Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1259
4.2 Force Vibration with Shaker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1260
4.3 Ambient Vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1261
5 Identification Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1262
5.1 The Fourier Transform and Time-Frequency Analysis . . . . . . . . . . . . . . . . . . . . 1264
5.2 Extended Logarithmic Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1265
5.3 Eigensystem Realization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1267
5.4 The Periodogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1270
5.5 Frequency Domain Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1274
5.6 Natural Excitation Technique ERA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1276
5.7 Stochastic Subspace Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1278
5.8 Multivariable Output-Error State sPace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1282
5.9 Iterative Modal Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1285
6 Stability and Cluster Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1287
7 Modal Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1290
7.1 Distance Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1290
7.2 Spurious Pole Elimination Using Hard Stability Criteria . . . . . . . . . . . . . . . . . . 1293
7.3 Modal Characterization Based on Parametric Procedures . . . . . . . . . . . . . . . . . . 1294
7.4 Modal Characterization Using Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1294

R. Boroschek ()
University of Chile, Santiago, Chile
e-mail: [email protected]
J. P. Santos
Structural Monitoring Unit, National Laboratory for Civil Engineering, Lisboa, Portugal
e-mail: [email protected]

© The Society for Experimental Mechanics 2022 1247

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_29
1248 R. Boroschek and J. P. Santos

7.5 Assessing the Number of Modes Using Cluster Validity . . . . . . . . . . . . . . . . . . . 1299

7.6 Cluster Representatives as Mode Features’ Estimates . . . . . . . . . . . . . . . . . . . . . 1302
8 Modal Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1303
8.1 Problem Statement and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1303
8.2 Tracking Based on Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1306
8.3 Cluster-Based Modal Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1308
9 An Important Note to Increase the Speed of Identification . . . . . . . . . . . . . . . . . . . . . . . 1311
10 Application Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1311
10.1 Building Excited by Ambient Vibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1311
10.2 Buildings Earthquake Vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1319
10.3 Base-Isolated Buildings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1322
10.4 Bridges Under Operational and Ambient Excitation . . . . . . . . . . . . . . . . . . . . . . 1323
11 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1335
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1335

Abstract

Civil engineering structures are especially complex due to their size, geometric
and physical uniqueness, intrinsic nonlinearity, and also due to the variations
of their properties as a function of environmental or operational loadings
and support conditions. While overloading can produce severe but recoverable
changes in modal properties, in the order of 30%, environmental conditions like
temperature and humidity can produce state changes, particularly in supporting
soils and boundary conditions, that have been observed to impose changes
of 50% on modal properties without the presence of damage. Some of these
variations are abrupt while others are slow, yet regardless of their speed they
impose challenges on the experimental setups and identifications techniques.
The basic civil structural instrumentation is mainly based on accelerometers
and displacements sensors, being their characteristics described in the present
chapter. The environmental excitations are in general nonstationary and input
excitations are usually not measured. So, response is generally analyzed using
Operational Modal Analysis (OMA) techniques, even if Experimental Modal
Analysis (EMA) have also been extensively used. The most commonly applied
technique for identification is the Stochastic Identification algorithm in their
covariance and data-driven versions. In order to observe the complexities of
civil infrastructure, examples are given for buildings under environmental and
earthquake loads and bridges under environmental and traffic loads. Identifi-
cation and automatic tracking algorithm are presented theoretically and with
examples.

Keywords

Civil structures · Earthquake action · Ambient vibration · Operational modal

analysis · Modal identification methods · Modal estimation · Modal tracking ·
Cluster analysis · Stabilization diagrams · Bridge use case · Building
use case
23 Civil Structural Testing 1249

Nomenclature

[A] System state matrix

agi Base acceleration
aP Acceleration at position p
(k)
ai Average distance between the i-th object of a cluster and the remaining object
of the same cluster
ADC Analog to digital
ARX Auto-regressive model with exogenous input
(k)
bi Average distance between the i-th object of a cluster and the objects assigned
to other clusters
B(PK ) Overall between-cluster distance measured from a partition P of k clusters
[C] Observation matrix
COV Reference to covariance-driven in SSI
Ck k-th cluster of partition PK
d(xi ,xj ) Distance between observations i and j measured for quantity x
DAS Data acquisition systems
EFDD Enhanced frequency domain decomposition
EMD Empirical mode decomposition
ERA Eigensystem realization algorithm
FDD Frequency domain decomposition
FFT Fast Fourier transform
[H] Hankel matrix
ITD Ibrahim time domain decomposition
K Number of clusters in cluster partition
[L] Matrix obtained from LQ decomposition
LVDT Linear variable differential transformer
MAC Modal assurance criterion
MIMO Muitiple input multiple output
MOESP Multivariate output-error state space
MP Mean phase
MPD Mean phase deviation
MPC Mean phase collinearity
Next-ERA Natural excitation technique eigensystem realization algorithm
[O] Observability matrix
Pk Cluster partition containing k clusters
PP Peak picking
[Q] Matrix obtained from LQ decomposition
RXX , RXY Cross-correlation
SIL(Pt ) Silhouette width of partition P containing t clusters
S/N Signal-to-noise ratio
SHM Structural Health Monitoring
(k)
si Silhouette width of the i-th object assigned to a cluster
sk Silhouette width of cluster k
SSI Stochastic subspace identification
STFT Short-time Fourier transform
SVD Singular value decomposition
SXX , SXY Periodogram
T Overall distance measured from a cluster partition
[T] Toeplitz matrix
[U] Hankel matrix of observations
UPS Uninterruptable power supply
v(t) Decay response
1250 R. Boroschek and J. P. Santos

W(PK ) Overall within-cluster distance measured from a partition P of k clusters

{y} Response vector
ẏ, ÿ First and second time derivatives of y
[Y] Hankel matrix of response
{z} State space vector
βt Damping of mode t
γ2 Coherence
λ Eigenvalue
ω, f Natural frequency of mode t
[] Singular values matrix
[φ] Modal shape matrix
[ψ] Eigenvector matrix

1 Why Structural Health Monitoring (SHM) for Civil

Structures

Civil engineering structures are in general large and unique. When dynamic testing
is performed to a complete structure not located in laboratory, but in the field,
the environmental conditions are uncontrolled and the structure is usually under
operation. The sensitivity of the system to various ambient, boundary, and loading
conditions, along with the restrictions for testing due to size, usage, and safety
of the occupants and of the structure itself, generally requires special attention
and modifications of the techniques typically applied to mechanical systems and
described in other chapters. Due to their exposure to the environment and to
usage, civil engineering structures exhibit slow or abrupt deterioration, damage, and
collapse. The consequence of damage to these structures is not only related to loss
of functionality and costly repairs. The deterioration could easily put the users and
surrounding population and structures at risk.
The main challenges for dynamic testing are basically:

– Nonlinear response. Civil structures are inherently nonlinear due to the

material used in their construction and due to their support conditions.
Constituent materials are in general nonlinear even at low deformations levels.
At larger deformations and extreme loads, hysteretic response is common and
acceptable. Also, its boundary conditions (foundations and soils) are sensitive
to environmental variations and loading history and intensity.
– Uniqueness. In general, each structure is unique in terms of its design, bound-
ary conditions, construction procedures, and maintenance. Knowledge of
similar structures helps understand the overall behavior, but cannot necessarily
predict its particular properties and response.
– Limited Modelling Capacity. Numerical models are approximations that can
only predict roughly the actual behavior of the structure. Nonlinear models
can be developed but in general they lack basic information on the material,
elements, and system hysteretic response. The typical viscoelastic models
are good approximations, but they rarely represent the reality of an existing
system.
23 Civil Structural Testing 1251

Despite these characteristics, it has been possible to design, construct, and

maintain these types of structures for centuries. This ability is not based on
numerical models, but rather a rough understating of their actual response under
varying conditions of existing similar structures.
This chapter presents the instrumentation, the identification methodologies, and
on-site examples of civil testing in real structures. Laboratory and component testing
are neither presented nor discussed.

2 Basic Considerations on Loading and Response

Civil structures are expected to be used for extended periods of time. A typical
design considers a life span of more than 50 years for buildings and 100 years for
bridges while other critical structures like dams and reservoirs can be in operation
for several hundred years. Due to the long exposition to the environment and
use, change of state and deterioration are common, and regular maintenance and
repair and reinforcement are generally necessary. To detect these changes, a clear
understanding of the structures’ characteristics and responses to the environment
and loading condition has to be developed. These state changes have to be
distinguished between progressive and abrupt as well as recoverable and non-
recoverable, to allow defining the appropriate instrumentation and procedures of
detection.
Typical progressive damages consist of material deterioration (carbonation,
corrosion, abrasion, etc.) or foundation modification (settlement, scouring, etc.),
while abrupt changes can be related with extreme loads like overloading, freezing,
blasting, extreme winds, and earthquakes, or when the structure and components
reach a limit state due progressive loading.
For example, in bridge piers (Fig. 1a, b) and dams (Fig. 1c, d), it can be observed
that frequency values of modes are significantly affected when the water levels rise
and submerge large lengths/section of this type of structural systems.
A recent study described in Pereira, Magalhães, Gomes, Cunha, and Lemos [67],
where the modal quantities of an arch dam were monitored and estimated during the
first filling of the reservoir, shows that, in fact, the effect of the water level can have
a significant influence on the properties of a large number of natural modes, with
frequency values decreasing proportionally with the increase in water level.

3 Data Acquisition and Sensors

3.1 Sensors

Civil structure modal testing is commonly conducted using accelerometers (Fig. 2).
Table 1 shows the typical accelerometer sensitivity, range, and bandwidth required
for different testing conditions and excitations. In general, the DC component on the
acceleration bandwidth is not required for modal testing but it is needed to detect
permanent displacements due to change of state, residual displacements, or damage.
1252 R. Boroschek and J. P. Santos

Fig. 1 Dynamic effects imposed by the water level: (a) changes observed in piers of a curved
bridge [87], (b) changes observed in natural frequencies of a dam model [57]

In civil engineering, rigid systems are considered those with natural frequencies
greater than 33 Hz, so higher frequencies are generally needed for component
testing.
Data from spatially distributed sensors installed in civil structures must be
precisely synchronized. Time lag between records of different sensors must be
less than 1 ms. The lag between recordings can cause apparent complex modes on
the identification [44] and in most cases can be identified from the phase between
records.
23 Civil Structural Testing 1253

Table 1 Accelerometer typical characteristics

Bandwidth
Excitation (Hz) Range (G, +/–) Sensitivity (G)
Impact 0.05–300 50 1e-3
Force 0.05–40 2 1e-3
vibration
Ambient 0.05–40 0.25 1e-5
Earthquake 0.05–40 4 1e-3
Traffic 0.05–200 4 1e-3
Blasting 0.05–200 10 (in adjacent structure) 1e-3
>100 (structure close to source)
Use 0.05–20 0.25 1e-4

Fig. 2 Accelerometer types used in civil dynamic testing: (a) force balance uniaxial,
(b) force balance triaxial, (c) microelectromechanical systems (MEMS), (d) piezoelectrical, and
(e) integrated station with GPS

Due to the large dimensions of civil structures, data is usually stored locally
and afterwards transmitted by cable or using cellular networks to a central storage
system. The transmission of analog signals is generally avoided due to electrical
contaminations and data loss. The ADC (analog-to-digital converter) and the signal
conditioning units are usually installed next to the sensor, and the data transmission
is made digitally and, in general, using fiber-optic medium, so as to fulfill the
recommended minimum signal-to-noise ratio, which would be higher than 3.
Other sensors used in some civil dynamic testing and monitoring consist of strain
gages, displacement sensors, velocity meters, tilt-meters, and image-based optical
sensors. Thorough examples of sensor use can be found in  Chap. 2, “Sensors and
Their Signal Conditioning for Dynamic Acceleration, Force, Pressure, and Sound
Applications”. In this section, only a brief reference exemplifying those used in
civil testing is made.
1254 R. Boroschek and J. P. Santos

Velocity sensors and tilt-meters are capable of measuring the structural vibration
speed and can be used as alternatives to accelerometers. They are based on laser
or inductance, can achieve very high sensitivities, and can therefore be suitable for
testing where there is need to measure vibrations with low amplitudes. However,
these sensors are generally limited in capturing lower frequencies and therefore
are not appropriate for slender flexible structural systems and for continuous
monitoring. As a result, this type of sensors seems to be most suited for seismic
movements [72].
Strain gages based on several types of transducing principles can be found in the
market and in civil monitoring applications, such as resistance-based, optic-based
(most common are Fiber Bragg Gratings, FBG), Carlsson or vibrating wire, with
numerous solutions for mounting and fixing to civil structures, such as glue, fixed,
embedded, or welded. For the purpose of modal civil testing, only the first two types,
resistance (Fig. 3a, b) and optic (Fig. 3c), are capable of providing accurate dynamic
data output. Regarding their mounting on civil structures, these two types of strain
gages are widely available for gluing (Fig. 3a) and welding (Fig. 3b) to metallic
components and for embedding in concrete (Fig. 3c).
Given the adequate signal conditioning and ADC features, these sensors are
capable of providing resolutions that can be much smaller than 1×10–6 m/m,
especially when assembling them in Wheatstone bridges. A typical use-case of
Wheatstone Bridge assemblies consists of measuring flexural strain in the top and
bottom flanges of Steel I-section. With the appropriate assembly, the flexural strain

Fig. 3 Strain gages used in civil testing: (a) resistance-based glued, (b) resistance-based welded,
(c) embedding resistance strain gage inside concrete element
23 Civil Structural Testing 1255

can be multiplied by a factor of 2 or 4. Even if these resolutions result in smaller

sensitivities to detect natural modes, when compared to the use of accelerometers,
the fact that strain gages, appropriately arranged, are self-compensated for tem-
perature effects along with their capacity to measure vibration waves with any
(large) period makes them appropriate for detecting low order natural modes in
large flexible civil structures, especially when these are composed of large-section
elements.
Relative displacement sensors with the capacity to output accurate data at
dynamic rates are also widely used in civil structural testing and generally fall into
the class of LVDT (Fig. 4c), precision potentiometer, magnetostrictive (Fig. 4a, b),
and optical. These types of sensors are generally fixed, or glued, to both structural
nodes for which the relative displacements are to be obtained, with due care to
keep rotations free at both ends. They are generally deployed between both ends of
joints, bearing supports, base isolation or control devices to measure the vibrations
imposed to these and also to capture the structural natural modes associated with
the displacements imposed to them. Their capacity to measure vibration waves with
long periods makes them particularly suitable for monitoring longitudinal mode
shapes in high pier bridges such as cable-stayed and suspended ones.
Load cells generally consist of mechanical assemblies of strain gages, either
resistance-based or optical, with resistant housing for high durability and force
resistance. Their shapes can vary significantly so as to adapt to specific applications,
such as, for example, those installed in downward end of the hangers of a bow-string

Fig. 4 Relative displacement sensors used in civil testing: (a, b) magnetostrictive sensor, (c)
LVDT
1256 R. Boroschek and J. P. Santos

Fig. 5 Load cells used in civil testing: (a) standard and (b) specifically fabricated

bridge (Fig. 5a). When load cell solutions are not available in the market, appropriate
strain gage assemblies can be defined and deployed in specific structural elements,
such as the bridge bearing device shown in Fig. 5b which, after calibration, is now
capable of measuring the forces applied to it. All remarks made above for strain
gages apply to load cells regarding sensitivity, applicability, and ability to detect
(mostly) lower order natural modes, even if characterized by very low values of
frequency.
Environmental sensors are used for dynamic civil testing with the objective of
measuring the effects of any demands that may influence the structural dynamic
response or which may influence the structural response by changing the materials’
properties or the boundary conditions.
Regarding the demands, wind is generally measured with anemometers such as
the one shown in Fig. 6a or with weather stations combining anemometers with
additional weather measurements, such as the one shown in Fig. 6b, c. Anemometers
can be based on cups or on ultrasonic measurements (Fig. 6a–c), where the latter are
more durable, precise, and require less maintenance. Water level can be measured
using buoy systems, even if the newer more robust and precise ones are based on
radar measurements such as the one shown in Fig. 6d and installed on the downward
faces of box girders or beams.
Concerning additional actions that may change the dynamic structural response,
scour is one of the most difficult conditions to measure in real bridges and can
be measured using sonar-based sensors such as the one shown in Fig. 6f, which
can be fixed to the upstream side of a bridge’s pile header or submerged beams to
detect the variation of pressure or deformation depending on the scouring level.
Structural element temperature is generally measured using glued of embedded
resistance thermometers such as those shown in Fig. 6e, or using thermistors, which
are generally less precise and less durable than the former. Rain, hail, and humidity
can generally be measured with high precision using weather stations such as the
one shown in Fig. 6b.
23 Civil Structural Testing 1257

Fig. 6 Ambient sensors: (a, b) ultrasonic anemometer from GILL Instruments, (c) cup-based
anemometer from Vaisala, (d) radar-based water level sensor from OTT (e, fit) weather stations
from VAISALA, (f) embedded resistance thermometers of the type PT100, (h) sonar sensor
UDM200 for scour monitoring

3.2 Selection of Data Acquisition Systems

Data acquisition systems (DAS) must be selected based on the number of sensors
(scalability), type of sensors (generality), sensitivity, stability, robustness for the
environmental actions, and usage conditions. DAS could be concentrated or dis-
tributed depending on the size and level of noise present on the system. For highly
electrical environmental noise, digital and optical data transmission is preferred.
Typical sensitivity varies depending on the sensors, but 24 bits ADC converters are
the most common. Synchronization of the different channel data is essential, and
nowadays synchronization is in the order of the nanosecond or lower.
Sampling frequency is highly dependable on the parameters to be measured so
flexibility on DAS per channels is higher convenient. Meteorological stations can be
sampled in terms of minutes or hours and accelerations in terms of microseconds.
All connection to the DAS should consider the environmental conditions and
the level of vibrations presented. Many structures require military type or bolted
connectors for long-term monitoring. Connections based on friction are in general
not reliable.
Voltage protection and stability is a must for long-term monitoring. Typically,
UPS systems are used to stabilize, clean, and guarantee the voltage supply to the
1258 R. Boroschek and J. P. Santos

system. See  Chaps. 2, “Sensors and Their Signal Conditioning for Dynamic
Acceleration, Force, Pressure, and Sound Applications”,  4, “Applied Digital
Signal Processing” and  5, “Introduction to Spectral and Correlation Analysis:
Basic Measurements and Methods” for more detail on DAS. Vandalism is common
so additional protection to the sensors and DAS should be considered.

4 Basic Procedures for SHM in Civil Engineering

Several different methods are used to identify the modal and dynamic response
properties of civil engineering structures. They have been developed for more than
a century, but most documented tests on real structures started to be reported in
the 1950s with the development of more flexible and transportable sensing and
recording systems [22, 41].
All civil structures present different levels of nonlinearity associated with
constituent material, foundation conditions, and interaction between components,
even if no damage is present. The nonlinearity and the need to evaluate the different
states that could be present in a structure require a clear understanding of the
possible excitation conditions and transitions between different response levels.
Hence, key decisions for the selection of testing and analysis procedures should
include:

1. Expected levels of excitation

2. Appropriate sensing and recording devices
3. Sensor locations
4. Applicable identification procedures
5. Desired identification properties
6. Representative model of the structure at each testing amplitude level

In reinforced concrete structures, the strain–stress relation is nonlinear. This

nonlinearity could be elastic (recoverable) or inelastic depending on the strain levels
[106] or boundary conditions. The opening and closing of micro-cracks causes
an important stiffening effect and corresponding modal parameter change. If this
micro-cracking is not excessive, once the loading is removed the system returns to
its original stiffness. One very early example of this condition was documented by
Keightley et al. when testing a dam reservoir under different periodic shaking loads
[51]. The boundary condition, typically soil, is also a highly nonlinear material and
highly sensitive to humidity and temperature variations [34, 36, 37]. Intrinsic energy
dissipation mechanisms are also highly sensitive to amplitude [30]. These conditions
require identification procedures that can capture the desired parameters and which
are stable under these conditions. In general, results are valid for a specific response
amplitude and environmental conditions.
Another important consideration is the differences between models’ assumptions
and reality. Typically, civil structures are modeled as linear viscoelastic. In practice
the energy dissipation properties are mixture of different mechanisms where friction
between components and radiation at foundation level also play an important role
23 Civil Structural Testing 1259

[47, 94]. Even at very low amplitudes of vibrations, the variations in damping with
amplitude are important, and at large amplitudes the response can also include
hysteretic behavior among many others [4, 24]. Some numerical models have
been proposed to capture the variation of damping with amplitude in buildings
[45, 52, 89], but they are in general not used for the identification procedures.
This chapter presents a summary of the most common testing procedures of full
structures under dynamic loads at field conditions. It presents the basic instrumen-
tation, identification methods for system identification, response monitoring, and
change of state detection.
The basic excitations considered for the identification depend on the structure
size, complexity, foundation characteristics, usage, location, time for testing, and
budget. The most common testing techniques are:

1. Initial conditions testing

2. Vibratory shakers
3. Ambient vibrations
4. Operational conditions
5. Extreme loading events: earthquake, blasting, wind

4.1 Initial Conditions Testing

Initial conditions, displacement and velocities, are excited by pulling [27, 56, 59],
pushing [101], or impacting [60] the structure. The excitation has to be selected and
applied to excite the local or the global response of the structure. Local responses are
obtained when the perturbation is applied to the elements under study, like beams,
columns, slabs, or walls. They are typically exited by hand systems that generate
pulling or pushing, heel dropping, jumping, hammer impacting, blast impacting, and
others, depending on the element characteristics and response levels desired. Global
responses can be obtained by pulling, pushing, and blast impacting the structure.
Several pulling mechanisms have been developed based on jacks, accompanied
with quick release mechanisms (collapse mechanism [6], breakable fuse [10], etc.).
Additionally, truck or heavy equipment can be used to pull the structural systems.
This type of testing is quite limited in existing structures because the level and
location of loading should be carefully selected and controlled so that the response
of the system is under the desired response levels. In general, the loading should
be large enough to excite the structure to the desired level without damaging
it. To accomplish this, a numerical model of the structure should be developed
and calibrated under real conditions (typically derived from ambient test). The
models are only approximations to reality, so several variations of the models are
recommended to quantify, in a reliable way, the possible states that the structure or
its components can reach.
To globally test a structure, it should generally not be in operation. Preparation
for the test including modeling, development of loading devices, and setup of the
sensors and recording system can take from weeks to months, depending on the
level of response required.
1260 R. Boroschek and J. P. Santos

The procedure applied to define the initial displacement or velocity controls the
modal parameter excited. For low frequency modes, simple displacement patterns
are recommended. When applying initial velocities, the impact durations should be
related to the frequency of interest.
Response levels can reach high values, so response monitoring sensors should
be anchored. Typical recommended acceleration sensor sensitivity is 0.001 g with
a range of +/–10 g and a bandwidth between 0.05 to 20 Hz. DC components for
this excitation are in general not required for acceleration. The measurement of
the initial displacements or the applied impact force is generally recommended, so
as to be able to obtain more information like static stiffness or inertial mass. For
modal characterization, the most common procedures are Ibrahim Time Domain
Decomposition (ITD), Eigensystem Realisation Algorithm (ERA), Peak Picking
(PP), Short-Time Fourier Transform (STFT), and Logarithmic Decay, among others.

4.2 Force Vibration with Shaker

This type of testing is based on an excitation produced by a rotatory [43, 93, 105]
or linear shaker [26]. The shaker generally produces the force by an eccentric or
inertial mass. Inertial forces are in the range of +/–50 kN for practical movable
systems. Due to its dimensions and at the high end of loading, the system should be
anchored to the structure. The frequency band of excitation is typically between 0.2
to 15 Hz and in general has a single-frequency nature. Response levels can reach
quite high values so the response monitoring sensors should be anchored. Typical
recommended sensor sensitivity is 0.001 g and a bandwidth between 0.1 to 20 Hz.
DC component for this vibration is not generally required.
Due to the medium and high levels of vibration that can be obtained by the shaker
at resonance conditions, the structure should be nonoperative or with restrictive
operations (especially in buildings). Also, the engineer responsible of the test should
have previously analyzed the possible damage states that could be induced in the
structure. For low levels of vibration excitations, no special care is normally needed,
but for large testing loads, a set of numerical models is recommended to estimate
the response.
Due to the preparatory requirements, namely the anchoring of the shaker and
sensors, these procedures have very limited application. Nevertheless, the possibility
of controlling the amplitude, direction, and location of the applied loading and
response levels allows the observation of the structures’ nonlinearities, as observed
by [41, 51].
Typically, testing time could be extremely long at large amplitudes, due to the
nonlinearity of the system when searching for resonance frequencies. Data analysis
is relatively simple but requires extensive observation and control by the testing
operators. The experimental results typically consist of acceleration, velocity, or
displacement amplification response as a function of excitation frequency. Typical
analysis procedures are peak picking (PP), enhanced frequency domain decomposi-
tion (EFDD), and others.
23 Civil Structural Testing 1261

4.3 Ambient Vibrations

Ambient vibrations are produced mainly by wind, traffic, micro-tremors, and usage.
In general, the excitations are not broad banded and their main energy is below
20 Hz. This bandwidth is good enough to excite the low-frequency modes of
the system. Typical wind-induced drift in buildings is observed in the order of
2/1000 [95] and acceleration response amplitude in the order of 0.001 g or even
lower, so sensitive accelerometers are needed. Other sources of ambient excitation
could have larger amplitude so typical range is +/–0.25 g. Recommended sensor
sensitivity for wind-induced vibration is 10–5 g and a bandwidth between 0.1 to
30 Hz. The DC frequency component for this vibration is in general not required.
Due to the low level of excitation, anchorage is not required and normal friction is
sufficient to keep the sensor in position, unless a permanent monitoring system is
desired.
Due to the quick deployment of sensors and the extreme low risk of damaging
the structure, these testing procedures are quite useful and convenient to identify the
modal properties of a structure. Nevertheless, due to the extreme low-level response,
the properties observed are restricted to a linear range and in general are not good
predictors of higher response, especially when damage occurs.
Depending on the objective of the monitoring the density of the recording
locations must be adequately chosen. If only frequency is desired, only few degrees
of freedom are needed, and care should be taken not to put the sensor in a nodal
point of the mode of interest. If the shape of the mode is required, then a much
densely recording network is needed. The simplicity of the procedures allows the
deployment of several sensors that should be properly synchronized.
For short-term recording it is not required to have a large number of recording
points. In this case, one or several sensors are defined as fixed references and another
set is moved around the structure to capture the response. Later the mode shape is
integrated with the help of the reference point. Typical recording times for each
setup are from 20 to 60 min, to obtain reliable properties.
Permanent or long-term ambient monitoring is also quite popular to detect
slow and abrupt variations of modal properties. For this case the sensors should
be permanently anchored and protected from the environment and the users of
the structure. Long time stability of the sensors and recording system should be
carefully studied for the selection of the components and synchronization between
sensors. Due to the extended monitoring time, data have to be carefully analyzed
before any attempt for identification or classification is done. This analysis should
include at least the verification of the bandwidth of the response, the interference or
contamination by periodic excitation, like motors, the occurrence of impacts to the
structures, or impacts due to traffic or users. The portions of contaminated data have
to be removed.
For long-term monitoring, data can be obtained in packages of shorter durations.
Typically, 5–15 min are sufficient. For each data package system identification
has to be performed, and later the modal properties identified have to be linked
to previous obtained values. This link, in the absence of abrupt changes, can be
1262 R. Boroschek and J. P. Santos

done using a distance metric, as the one presented in Eq. (81), which considers
a normalized frequency and shape similarity (MAC) between consecutive modes
and with consideration of the number of sensors. Damping is not generally used in
the tracking due to its uncertainty. Example of tracking procedures from ambient
vibrations can be found in [17, 49, 68].
Because of the sensitivity of civil structures to the environment, specially
temperature, humidity and level of excitation, and the modal parameters vary in
time [20, 32, 38, 92, 107]. Typically, modal frequency can have a daily change
due to temperature of 1–2% in buildings, if freezing of soil is not present. In case
of freezing, the variation on frequency can be as strong as 50%. Due to rain or
high humidity the frequency change is on the order of 6%. The effects due to rain,
extreme low temperatures, vary daily (temperature and radiation) and seasonal.
So, to understand the magnitude and effect of the change, consistent observation
periods must be considered. The relation between variations and modal properties
is multivariate so the largest the observation periods the best.
The modal changes due to environmental fluctuations are quite large so if the
system is required to detect deterioration, these variations can be easily masked
by them. In order to increase the identification capacity of change of state, several
approaches have been implemented and generally named as data normalization
procedures. These consist of removing the effects of the actions, from the responses,
using Principal Component Analysis [102], multilinear or nonlinear regression,
neural networks [81, 107], and ARX models, among others [65].
In general, the excitation is not measured so the typical identification procedures
are based on Operational Modal Analyses. Among the most used identification
techniques and some references for examples of applications are PP, ERA [15],
Extended Frequency Domain Decomposition [11], and Stochastics Subspace Iden-
tification [63, 62, 64].

5 Identification Methods

The identification methods used can be classified according to their basic properties:

(a) Solution space: time, frequency, state

(b) Parametric or non-parametric
(c) Direct or iterative
(d) Type of excitation

Table 2 presents the most common methods used in civil structures. There are
also other methods used such as linear [78] and nonlinear time filters [76], and for
others see for example Peeters et al. [66].
23 Civil Structural Testing 1263

Table 2 System identification techniques

Name Parametric Domain Excitation Comments Reference
Peak Picking No Frequency Ambient, or Easily  Chap.
(PP) colored or understood and 10,
white noise used “Experi-
force excitation mental
Modal
Analysis
Meth-
ods”
Short-Time No Frequency All type The initial step [21]
Fourier in most
Transform identification
(STFT) process Best to
evaluate
nonlinearity of
the system
Hilbert–Huang No Time and All Includes the [42]
Transform frequency possibility of
nonlinear
systems
Frequency No Frequency and Ambient, or More robust [13, 12]
Domain time colored or tan PP
Decomposition white noise
(FDD)/Extended force excitation
FDD (EFDD)
Eigensystem Yes Time Impulse [48]
Realisation
Algorithm
(ERA)
Natural Yes Time Ambient [46]
Excitation
Technique
ERA
(NeXT-ERA)
Stochastic Yes Time Ambient Assumes [62, 99]
Subspace properties of
Identification the unknown
(SSI COV-SSI input
DATA) excitations
Multivariable Yes Time Input-output [100]
Output-Error
State sPace
(MOESP)
N4SID Yes Time Ambient Includes SSI [98]
-Input/output and MOESP
Iterative Modal Yes Time Earthquake Iterative [3]
Response
1264 R. Boroschek and J. P. Santos

5.1 The Fourier Transform and Time-Frequency Analysis

Among the initial steps in the identification analysis is the evaluation of its fre-
quency content by way of the Fourier Transform (FT) and the STFT or spectrogram.
The visualization of the component and its variations with time allows the user to
identify the bandwidth of the signal, the presence of permanent or stationary signals,
and the possible nonlinear characteristics or variation of apparent predominant
signals.
The FT presents a summary of the energy for each frequency component in the
whole signal. The STFT uses the FT in a relative short duration moving window, as
shown in Fig. 7. The results are plotted in a 3D plot or color map, Fig. 8. The STFT
uses a constant time window so careful selection of its duration shall be performed
to capture the predominant frequencies and its variation with time. As indicated by
Cohen [21], this transformation is redundant so it does not have an inverse. This
can be overcome if the S-Transform is used instead [91]. Another alternative for
some nonlinear system is the Hilbert–Huang Transform that uses the empirical
mode decomposition (EMD) and the Hilbert transform to capture the variation
frequency due to nonlinear responses. An example of application can be found in Shi
et al. [86].

Fig. 7 Spectrogram construction. (a) Time series. Two windows sections to be FFT. (b) FFT of
two windows. (c) 3D plot of FFT centered on the mid time of selected window. No treatment is
use in this example to select the time window to reduce leakage
23 Civil Structural Testing 1265

Fig. 8 Spectrogram of sine function that presents a linear varying frequency (0–15 Hz) and a
linear increase of amplitude (0–3). (a) Time series, (b) spectrogram, and (c) Fourier transform of
the whole signal

5.2 Extended Logarithmic Decay

5.2.1 Linear Adjustment of the Decay

The analysis on a single predominant frequency decaying response can be per-
formed with and extended version of the logarithmic decay presented in Chapter
12. In this case assuming a viscoelastic one degree-of-freedom system, the decay
follows the rule

v(t) = e−βωt [ρ cos (ωD t − θ )] (1)

Figure 9 presents the acceleration response of a wharf structure subjected to a

quick release [10]. Because modes are clearly separated, the signal is band pass
filtered to select the second mode. From the statistics of the zero crossing and
evaluation of the linearity of the frequency, the amplitude can be evaluated (red
points in Fig. 10). If all maximum absolute values of each cycle are identified
and numbered, a plot can be made of ln(|v(t)|) versus its sequential number. If the
resulting plot is linear, we can conclude that a viscous energy dissipation mechanism
is predominant. If it presents other shape, friction or other mechanisms can be
1266 R. Boroschek and J. P. Santos

Fig. 9 Acceleration response of wharf structure. Two modes are present

detected. Due to noise, the arrangement of the data is not always linear, so a
least square regression can be performed to obtain the slope, which relates to an
equivalent viscous damping of the system given by Fig. 11:

slope
β= (2)
π

This procedure is not appropriate when multiple predominant decaying frequen-

cies are present. In this case other methods like ITD and ERA are more appropriate.

5.2.2 Nonlinear Adjustment of the Decay

The decay of the response for a single observation point from a system with several
degrees of freedom can be expressed as

v(t) = e−βi ωi t ρi cos (ωDi t − θi ) (3)
i

where the parameters are optimized through the minimization of an error function is
not recommended due to the large number of local minima, ERA and ITD are more
23 Civil Structural Testing 1267

Fig. 10 Zero crossing and maximum absolute values of the second mode after extracting from
original signal

appropriate methods. Nevertheless, once a close solution is found, several studies

on the decay can be performed. For the original response of the initial condition test
shown in Fig. 9, the nonlinear fit is shown in Fig. 12. Figure 12 presents the fitted
parameters for the whole window. Figures 13 and 14 present the fit to two smaller
size windows. Results of the fits are different, indicating the nonlinearity present
in the response. Additionally, the error function of the fit can be used to evaluate
the sensitivity of the different parameters to the error. For example Fig. 15 presents
the case observed when fixing all parameters except the damping for the second
mode indicating that values between 5% and 7% can give similar adjustment to the
response.

5.3 Eigensystem Realization Algorithm

The method developed by JUANG and PAPPA [48] allows the parameter determina-
tion from a response decaying signal. This decaying signal can be obtained by direct
experimental test (initial conditions testing) or derived from ambient vibrations
(NeXT ERA, Random Decrement). See additional details of ERA in  Chap. 10,
“Experimental Modal Analysis Methods”.
1268 R. Boroschek and J. P. Santos

Fig. 11 Estimation of damping by least square fit to maximum absolute values

The algorithm uses the decay response of the system organized on a Hankel
matrix:
⎡ ⎤
[Yk ] [Yk+1 ] · · · Yk+j
⎢  ⎥
⎢ p×q ⎥
⎢ ⎥
⎢ [Yk+1 ] [Yk+2 ] · · · Yk+j +1 ⎥i
[Hk−1 ] = ⎢
⎢ .. .. ..
⎥
⎥ (4)
(prxqs) ⎣ . . ··· . ⎦
[Yk+i ] [Yk+i+1 ] · · · Yk+i+j
↔j

where p is the number of sensors, q the number of independent excitations, and i

and j are the selected number of observations. There are several recommendations
for defining the number of observations and the size of the problem to solve,
but the number of columns should be at least twice the number of expected
modal components present on the signal and the number of rows should allow
the representation of the decay without including the noise portion. The stability
diagram, presented in  Chap. 11, “Experimental Modal Parameter Evaluation
Methods” and later described below, is the most common procedure to select the
representative modes.
23 Civil Structural Testing 1269

Fig. 12 Nonlinear adjustment to multimode decaying signal. Whole window

The singular value decomposition of [H]

[H ] = [U ] [] [V ]T (5)
(prxqs) (prxpr)(prxps)(psxqs)

selecting a reduced number of non-null singular values.

[H0 ] = [U2N ] [2N ] [V2N ]T (6)

(prxqs) (prx2N )(2N x2N ) (2N xqs)

and considering
 
T
Ep = [I ] [0] ··· [0] (7)
(pxpr) (pxp) (pxp) (pxp)

The system state and observation matrix can be found as

[A] = [2N ]−1/2 [U2N ]T [H1 ] [V2N ] [2N ]−1/2 (8)

T
[Cd ] = Ep [U2N ] [2N ]1/2 (9)
1270 R. Boroschek and J. P. Santos

Fig. 13 Nonlinear adjustment to multimode decaying signal. Initial 4 s window

The eigenvalues and vectors of [A] are

[A] { ui } = λui { ui } (10)

from which the modal shape at the observed and locations can be found as
 
φyi = [Cd ] { ui } (11)
(px1) (px2N )(2N x1)

along with the natural frequency and damping of the mode

|ln (λui )| Re al (ln (λui ))

ωi = ; βi = − (12)
t |ln (λui )|

5.4 The Periodogram

Several identification techniques are based on the power spectral estimation or peri-
odogram.  Chaps. 4, “Applied Digital Signal Processing” and  5, “Introduction
to Spectral and Correlation Analysis: Basic Measurements and Methods” presents
23 Civil Structural Testing 1271

Fig. 14 Nonlinear adjustment to multimode decaying signal. Window from 4 s to end

the basic concepts of frequency domain analysis, power spectra, and periodogram.
Here they are used to identify the modal properties of a system using PP, FDD, and
EFDD.
Two estimators of the FRF are usually used.

Y (f )X∗ (f ) SY X
H1 (f ) = ∗
= (13)
X(f )X (f ) SXX

where Y(f ) and X(f ) are the Fourier transform of the windowed signal, or

Y (f )Y ∗ (f ) SY Y
H2 (f ) = = (14)
X(f )Y ∗ (f ) SXY

For the evaluation of Ĥ (f ) we use the estimation of the periodogram

1  1 
Na Na
ŜXX (f ) = (SXX )n , ŜY Y (f ) = (SY Y )n , (15)
Na Na
n=1 n=1
1272 R. Boroschek and J. P. Santos

Fig. 15 Error as a function of second mode damping and optimal value for whole window

1  1 
Na Na
ŜXY (f ) = (SXY )n , ŜY X (f ) = (SY X )n (16)
Na Na
n=1 n=1

The coherence in the frequency space is (see also  Chaps. 5, “Introduction to

Spectral and Correlation Analysis: Basic Measurements and Methods” and  6,
“Frequency Response Function Estimation”):

 2  
 
γ 2 = ŜXY (f ) / ŜXX (f )ŜY Y (f ) (17)

The coherence has values between 0 and 1. An unitary coherence reflects a clear
agreement and consistent modal properties while a value lower than 1, for a given
mode, can be obtained depending on:

1. Noise
2. Nonlinearity of system
3. Leakage
4. Out of phase effects
23 Civil Structural Testing 1273

The signal-to-noise ratio can be estimated as

γY2X
S/N = (18)
1 − γY2X

In Fig. 16, the acceleration time series of channel 6 of the office building
described in Sect. 10.1 is shown and the corresponding record shows important
bursts of response. In case the signal is long enough, the high amplitude section
can be removed if the remaining segments can be used to identify the structure
properties. The record should be initially tested by estimating the periodogram.
If clear peaks are present, in the absence of noise, they can be associated with
structural properties. As an example, the power spectral density estimated by the
Welch method, for all channel above ground for the TC Building, is presented in
Sect. 10.1. In order to have a reliable spectrum several averages should be taken.
When nonoverlapping time segments are used the uncertainty of the estimation is
related by the inverse of the number of averages. The largest the number of averages,
the smaller the uncertainty. This requires long records. In order to avoid leakage,
the selected time segments shall be windowed. Typical window used for ambient
vibrations are Hanning and Hamming. The length of the window affects the shape

Fig. 16 Ambient acceleration response of 8 floors in an office building

1274 R. Boroschek and J. P. Santos

Fig. 17 Periodogram of all aboveground channels, 1–9, of an office building calculated by the
Welch method

of the spectra and the identified damping. The largest the window, the smallest the
effect on the damping estimation.
For typical structures with frequencies varying from 1 to 10 Hz, a window of
length 20–30 s gives good results when averaged at least 30 times. This means
records of more than 15 min duration must be taken. For lower frequencies, larger
duration and windows are recommended.
From an initial visual inspection of Fig. 16 and of the predominant operational
frequencies in Fig. 17, several peaks can be observed and, in this case, at least 8
are clearly detected. Nevertheless, not all channel presents all the peaks, because
the operational response at these frequencies is small. The first three operational
frequencies are 1.76, 2.17, and 2.62 Hz. Care should be taken in identifying a peak
as a single modal frequency. In order to distinguish repeated frequency, the FDD
methods consist of a simple procedure.

5.5 Frequency Domain Decomposition

The methods make use of this property and perform a singular value decomposition
of the observed responses periodogram [12, 13]
23 Civil Structural Testing 1275

     H
+
svd Syy (j ωi ) = U (ωi ) \ S(ωi )\ U (ωi ) (19)

If only one mode dominates the response at frequency ωr the modal shape can
be obtained as
 
φ̂r = {u1 (ωr )} (20)

This method allows the identification of close modes. Figure 18 presents the
singular value decomposition of the estimation shown in Fig. 17. Some repeated
root can be observed for frequencies close to 9.5 Hz. For damping the Enhanced
Frequency Domain Decomposition,
  [12, 13] recommend selecting a band around
ωr , where the modal vector φ̂r is similar (MAC higher than 0.8), as shown in
Fig. 19. This segment is transformed to the time domain and identification is done
on the decay by different techniques, shown in Fig. 20. The damping identified by
this method, in case of close modes, is not as reliable as when time domain methods
are used. In the periodogram-based methods the number of points that define the

Fig. 18 Singular value decomposition of cross-spectral density matrix of 6 channels above ground
1276 R. Boroschek and J. P. Santos

Fig. 19 Selection of frequency band for each mode to perform a damping estimation using a
response function

periodogram should be large so a good resolution of the frequency of the peak is

obtained.

5.6 Natural Excitation Technique ERA

The method was developed by James III et al. [46] and uses the cross-correlation of
the response of structures excited by white noise with respect to a reference channel.
It can be shown that cross-correlation has the same characteristics of a decaying
signal of a dynamic system under initial excitations.
The cross-correlation is calculated numerically in the time or frequency domain,
being in the former defined as:

N −k
1 
R̈ÿi ÿj (k t) = ÿi (l)ÿj (k + l) (21)
N −k
l=0

For the case of one degree of freedom, the cross-correlation has the form of
23 Civil Structural Testing 1277

Fig. 20 Time response function for first operational mode. Envelope calculated with Hilbert
transform is also shown

 
σ2 −βωτ β
Ryy (τ ) = 3β
e  sin (ωD τ ) + cos (ωD τ ) (22)
4ωD 1 − β2

Hence, the methodology takes the ambient response, shown in Fig. 16, and
consists of obtaining its cross-correlation using a reference channel, as shown in
Fig. 21. In this case we present the decaying function with respect to Channel 6. Due
to the long duration of ambient records, the cross-correlation quickly diminishes to
very low values associated with noise. The portion that represents the decay function
should be selected and used for the analysis with ERA. The longer the period or the
smaller the damping, the longer the decay signal to be selected. Several different
reference sensors shall be analyzed to reduce the effect of a defective reference,
which may be generated by low amplitudes at certain frequencies or by the presence
of noise.
The sampling rate of the signal shall be related to expected predominant
frequencies and should allow for a good representation of the decay. Extreme high
sampling rates can limit the ability for modal identification and therefore down
sampling the signal is generally recommended to a sampling frequency of no more
than four times the maximum expected modal frequency present in the signal.
1278 R. Boroschek and J. P. Santos

Fig. 21 IFR from correlation of response acceleration with respect to channel 6

As indicated in  Chap. 11, “Experimental Modal Parameter Evaluation Meth-

ods”, different model sizes should be tried to identify the most probable modal
parameters. The results of the different models can be represented in a stability or
consistency diagram, as shown in Fig. 22. In the stability diagram the poles comply
with a set of rigid and soft criterion in order to distinguish between operational and
spurious modes. Several poles could comply with these criteria. One way to select
the representative pole is to construct a cluster plot of poles that comply with the
criteria. In Fig. 23a, frequency damping clusters are used where it can be seen that
the dispersion is mainly due to damping. Several procedures have been used to select
the representative value from each cluster, but one of the best criteria is to select the
pole in the densest region of the cluster. In the figure this is indicated by the red
circle. The selected modal parameters are summarized in Table 3. The spread of
possible results should always be kept in case needed for further analysis.

5.7 Stochastic Subspace Identification

Two SSI methods are commonly used [63]: the covariance driven (COV) and the
data driven (DATA). Chapter 14 presents the description of the methods and the
proper selection of the record length and number of rows in the observation matrix
23 Civil Structural Testing 1279

Fig. 22 Stability diagram showing poles that comply with hard and soft criteria. Identified fre-
quency and damping is marked in red circle and shown in table. First singular value decomposition
from FDD is shown in dark gray

for best results. As previously referred, care should be taken to select the appropriate
length of records according to an expected number and magnitudes of the operation
frequencies and damping present on the signal. The most common procedure to
select the appropriate model is the construction of a stability diagram and to impose
hard and soft discriminating criteria. The method of selecting a gap or jump from the
singular value decomposition matrix is not practical due to the noise level present
in the signals.
Theoretically, the excitation shall be white noise even if, as will be shown
later, the identification can be performed adequately with color noise and even for
transient event like earthquake responses.
Here SSI-COV is summarized for the key elements that are used. The methods
make use of the Toeplitz matrix of the Hankel matrix of observations:
⎛    ⎞
ref ref
Ri ref Ri−1 · · ·R1
⎜ ⎟
   T ⎜      ⎟
⎜ ref ref
···
ref ⎟
=⎜ Ri+1 Ri R2 ⎟
ref ref
T1/i = Yf Yp ⎜ ⎟ (23)
⎜ ⎟
⎝  ···   ···  ···  ··· ⎠
ref ref ref
R2i−1 R2i−2 · · · Ri
1280 R. Boroschek and J. P. Santos

Fig. 23 Cluster of selected frequency and damping. Identified frequency and damping is marked
in red circle and shown in table

Table 3 Frequency and NExT- SSI-

damping derived from EFDDFreq/Damp ERAFreq/Damp COVFreq/Damp
different methods for TC Mode (Hz)/(%) (Hz)/(%) (Hz)/(%)
Building
1 1.76/2.26 1.78/2.00 1.77/1.69
2 2.17/2.01 2.16/2.47 2.17/2.65
3 2.62/1.76 2.61/1.49 2.60/1.91
4 5.17/1.03 5.18/1.37 5.17/1.56
5 6.32/1.57 6.34/2.38 6.32/2.20

This matrix is decomposed using singularly values to estimate the system matrix
 
ref
T1/i == [U1 ] [S1 ] [V1 ]T (24)

or
   
ref ref
T1/i = [Oi ] i (25)
23 Civil Structural Testing 1281

with

[Oi ] = [U1 ] [S1 ]1/2 [T ]

  (26)
= [T ]−1 [S1 ]1/2 [V1 ]T
ref
i

It can be shown that the discrete state matrix can be obtained by a shifted block
Toeplitz matrix [62]

   
ref ref
T2/i+1 = [Oi ] [A] i
    
ref †
= [S1 ]−1/2 [U1 ]T T2/i+1 [V1 ] [S1 ]−1/2
ref ref
[A] = [Oi ]† T2/i+1 i
(27)

After finding [A], the modal properties are obtained in a similar way as described
in the ERA method section.
For the case of the TC Building, if the record length for the analysis is short, then
the response could not have all modes of interest and consequently the resulting
estimation may be poor. Very long record increases the computational cost and field
work but does not necessarily give additional information. So careful selection of
the appropriate length shall be considered. Because time domain methods require
much shorter data length than frequency domain methods, the records can be split in
shorter length windows and the results of each selected window can be statistically
analyzed given a more stable results for the number of windows selected.
The results of the analysis of a single 15-min recording, considering 30 s and
1-, 5-, and 10-min windows, are presented in Table 4. For the 30-s duration a total
of 30 windows are considered and for the 10-min window only one is considered.
For the 30-s and 1-min windows not all modes are detected in all windows. In this
case the average frequency and damping values are similar, but a large dispersion
of the values in each window is found for the shorter windows. Table 5 shows
the minimum, medium, and maximum results of the analysis of 5-min windows.
Selected values are similar to those obtained by the NExT-ERA and EFDD methods,
Table 3.

Table 4 Mean values Mode 30 s 1 min 5 min 10 min

damping estimate for
1 1.54 1.34 1.38 1.72
different window length (%)
2 2.15 2.23 2.23 2.62
3 1.47 1.38 1.60 1.95
4 1.70 1.63 1.62 1.54
5 1.73 1.92 2.00 2.03
6 1.82 2.00 2.07 2.21
1282 R. Boroschek and J. P. Santos

Table 5 Minimum, Mean, and Maximum Values from three 5-min independent windows
Mode Mean frequency Min–Mean–Max Mean damping Min–Mean–Max
1 1.76–1.77–1.78 1.17–1.38–1.69
2 2.17–2.17–2.18 1.78–2.23–2.68
3 2.60–2.61–2.61 0.98–1.60–2.23
4 5.17–5.18–5.20 1.51–1.61–1.78
5 6.32–6.35–6.37 1.71–2.00–2.38
6 8.03–8.06–8.08 1.83–2.07–2.20

5.8 Multivariable Output-Error State sPace

Multivariable Output-Error State sPace (MOESP) [100] is commonly used in civil

structures when the input excitation is known. The methods transform the state space
equation with known input into a decay response, to latter apply ERA.
Starting from the state space and observation equations

{zk+1 } = [A] {zk } + [B] {uk } (28)

{yk } = [Cd ] {zk } + [D] {uk } (29)

Considering the initial state {z0 } in a recurrent formulation:

{z1 } = [A] {z0 } + [B] {u0 } (30)

We get


k−1
{zk } = [A]k {z0 } + [A]k−i−1 [B] {ui } (31)
i=0

In a similar way for {yk }:

k−1 

{yk } = [Cd ] [A] {z0 } + [Cd ]
k
[A]k−i−1
[B] {ui } + [D] {uk } (32)
i=0
23 Civil Structural Testing 1283

then
⎧ ⎫ ⎡ ⎤
⎪
⎪ {y0 } ⎪
⎪ [Cd ]
⎪
⎨ {y1 } ⎪
⎬ ⎢ ⎥
⎢ [Cd ] [A] ⎥
.. =⎢ .. ⎥ {z0 }
⎪
⎪ . ⎪
⎪ ⎣ . ⎦
⎪
⎩ ⎪
⎭
{ys−1 } [Cd ] [A]s−1
⎡ ⎤⎧ ⎫
[D] 0 ··· ⎪
⎪ {u0 } ⎪
0 ⎪ 0
⎢ [Cd ] [B] [D] ···
⎪
⎥ ⎨ {u1 } ⎪
0 ⎬ 0
⎢ ⎥
+⎢ .. .. .. ⎥ . .. ..
⎣ . . . ⎦⎪⎪ .. . ⎪
⎪ .
⎪
⎩ ⎪
⎭
[Cd ] [A]s−2
[B] [Cd ] [A] s−1
[B] · · · [Cd ] [B] [D] {us−1 }
(33)
⎡ ⎤
[Cd ]
⎢ [Cd ] [A] ⎥
⎢ ⎥
if [Os ] = ⎢ .. ⎥ is the extended observability matrix, and using the
⎣ . ⎦
[Cd ] [A]s−1
Toeplitz matrix, then
⎧ ⎫ ⎧ ⎫
⎪
⎪ {yk } ⎪
⎪ {uk } ⎪
⎪
⎪ ⎪
⎪
⎨ ⎪
⎬ ⎪
⎨ ⎪
{yk+1 } {uk+1 } ⎬
.. = [Os ] {zk } + [Ts ] .. (34)
⎪
⎪ . ⎪
⎪ ⎪
⎪ . ⎪
⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭
{yk+s−1 } {uk+s−1 }

If we include the Hankel matrix of input and observations

⎡ ⎤
{yi } {yi+1 } · · · {yi+N −1 }
⎢ {yi+1 } {yi+2 } · · · {yi+N } ⎥
⎢ ⎥
Yi,s,N = ⎢ .. .. . . .. ⎥ (35)
⎣ . . . . ⎦
{yi+s−1 } {yi+s } · · · {yi+N +s−2 }
⎡ ⎤
{ui } {ui+1 } · · · {ui+N −1 }
⎢ {ui+1 } {ui+2 } · · · {ui+N } ⎥
⎢ ⎥
Ui,s,N = ⎢ .. .. . . .. ⎥ (36)
⎣ . . . . ⎦
{ui+s−1 } {ui+s } · · · {ui+N +s−2 }

we reduce the equation to

Y0,s,N = [Os ] Z0,N + [Ts ] U0,s,N (37)

1284 R. Boroschek and J. P. Santos

Constructing the Hankel matrix of input and output and performing LQ decom-
positions, we get
  * +* +
U0,s,N [L11 ]ms x ms [0]ms x ls [Q1 ]T ms x N
[H ] = ms x N = (38)
Y0,s,N ls x N
[L21 ]ls x ms [L22 ]ls x ls [Q2 ]T ls x N
* +
U0,s,N = [L11 ] [Q1 ]T
[H ] = (39)
Y0,s,N = [L21 ] [Q1 ]T + [L22 ] [Q2 ]T

with
,
[I ] if i = j
[Qi ] T
Qj = (40)
[0] if i = j

then

Y0,s,N = [Os ] Z0,N + [Ts ] [L11 ] [Q1 ]T = [L21 ] [Q1 ]T + [L22 ] [Q2 ]T (41)

Post multiplying Equation (41) by [Q2 ] and considering the orthogonality

relations, we get

Y0,s,N [Q2 ] = [Os ] Z0,N [Q2 ] = [L22 ] (42)

The application of singular value decomposition to [L22 ] results in

 - * +
1 nxn [0]
- n x (ls−n) [N1 ]T n x ls
[L22 ] = [M1 ]ls x n [M2 ]ls x (ls−n)
[0](ls−n)x n 2 (ls−n)x (ls−n) [N2 ]T (ls−n)x ls
 
≈ [M1 ] [N1 ]T
1
(43)
   
= diag (σ1 · · · σn ) ; = diag (σn+1 · · · σls ) (44)
1 2

By selecting n number of singular values, we get the Observability Matrix as

 1/2
[Os ] = [M1 ] (45)
1

and [A] as

[Cd ] = [Os (1 : l, :)] (46)

[Os (1 : (s − 1) l, :)] [A] = [Os ((l + 1) : sl, :)] ⇒ [A]

(47)
= [Os (1 : (s − 1) l, :)]+ [Os ((l + 1) : sl, :)]
23 Civil Structural Testing 1285

Fig. 24 (a and b) Top story record TC Building 2009 low amplitude earthquake. Record is
analyzed by a sliding nonoverlapping window of 5 s. (c) Evolution of first predominant frequency
with time

After [A] is found the procedure described in the ERA section is used to obtain
the modal parameters.
In the TC Building more than 1700 earthquakes of different amplitudes were
recorded, some of which produced damage. Figure 24 presents one of the recorded
acceleration responses. The upper stories (6 channels) are analyzed considering the
full record length and dividing them in 5-s length nonoverlapping windows to track
an approximation to the variations of the predominant frequency. For this small
amplitude response, the building is nearly elastic so no apparent change is detected.
In contrary the Magnitude 8 earthquake of 2010, shown in Fig. 25, produced damage
in the building, and in this case a clear reduction of the frequency is observed. This
is a permanent change that later is corroborated by ambient vibrations.

5.9 Iterative Modal Identification

A modal minimization approach of an error function is an alternative to the direct

MOESP method. This was initially presented in civil engineering by Beck [3] where
1286 R. Boroschek and J. P. Santos

Fig. 25 (a,b) Records acquired from the TC Building 2010 damaging earthquake, analyzed by a
sliding nonoverlapping window of 7s. (c) Evolution of first predominant frequencies over the same
time period

a sequential search of modal parameters and an error function is used. This is

quite a slow process; nevertheless it can provide more flexibility by controlling
the modes that participate in the response, given different weighting factors to
different record locations and to evaluate the shape of the error surface as a function
of all independent variables. This gives important insight on the sensitivity of the
parameters.
The initial dynamic modal equations used to develop the identification algorithm
are


k
ÿj (t) + 2ωj ξj ẏj (t) + ωj2 yj (t) = Lj,i agi (t) (48)
i=1


N
ap (t) = φj,p · ÿj (t) (49)
j =1

where ωj , ξ j , Lj, i , and φ j, p are modal parameters of mode j, ωj being the modal
angular frequency, ξ j the modal damping ratio, Lj, i modal participation ratio for
23 Civil Structural Testing 1287

base acceleration agi (t) and φ j, p the modal shape vector component corresponding
to position p. Additionally, yj (t) and ap (t) represent modal response of mode j and
estimated acceleration at position p, respectively.
Several target function can be established for the optimization problem for
example [19].
.
/-
/ αp · - 1a0,p (t) − ap (t)22
/
/p t
E=/ - -1 22 (50)
0 α · a (t)
p 0,p
p t

Weight coefficients, α p , allow to control each mode in the identification process.

The optimization is performed mode by mode, retrieving the estimated response
from the target acceleration in a cyclic procedure until convergence is found.

6 Stability and Cluster Diagrams

The identification methods listed in Table 2, from which several are described in
Sect. 5, are based on solutions obtained either in the time domain, in the frequency
domain, or in both. Those that are reported to be more suited for modal estimation in
civil structures are based on solutions obtained in the time domain [14, 48, 64], such
as the SSI, the ERA, the Next-ERA, or the MOESP models (see Sect. 5). The main
reason for this choice concerns the fact that while frequency and time-frequency
methods estimate the existence of modes in continuous functions, time-based
methods allow observing them as single-valued data objects (poles), for each of the
n orders assumed for these models. The computational and theoretical simplicity
associated with analyzing this type of data, when compared to the complexity
associated with the analysis of continuous functions, has made researchers and
practitioners choose this class of methods for civil dynamic testing. Additionally,
time methods generally use less data and provide more reliable values for damping.
Even if time-based algorithms are preferable when compared to those which are
frequency-based, they may still face the challenge of depending on the structure of
the Hankel matrix which, in turn, depends on the model order n. In theory, the model
order may be chosen as twice the number of modes needed to describe the structural
system. However, this theoretical value is difficult to use in civil structural testing
due to the number of spurious modes, associated with noise and other nonstructural
effects, that is generally measured on site. One solution to this problem can simply
consist of estimating the parameters associated with different model orders, bearing
in mind that orders lower than twice the number of modes will underestimate their
number while higher values will tend to generate spurious modes.
The identification of which mode estimates (or poles) can be associated with
natural or spurious modes is generally conducted by choosing multiple model
orders, n, from a value which clearly underestimates the system’s dynamics, to one
1288 R. Boroschek and J. P. Santos

Fig. 26 (a) Stability diagram obtained from 1 h of data acquired on the deck of a suspended bridge
and corresponding (b) cluster diagram (frequency vs. damping), (c) cluster diagram (frequency vs.
MP), (d) cluster diagram (frequency vs. MPD), (e) cluster diagram (frequency vs. MPC)

which is known to produce both operational modes and noise-associated estimates.

This strategy is described in Chapter 12 and known as stability diagram (already
shown in Sect. 5), due to the fact that, when estimates are observed in a “frequency
Vs. order” scatter plot, those associated with operational modes repeat themselves
over the majority of model orders. An example of a stability diagram obtained from
1 h of monitoring data acquired on a suspended bridge is shown in Fig. 26a.
When observing the same estimates in a “frequency vs. damping” scatter plot,
shown in Fig. 26b, each vertical alignment observed in the stability plot (Fig. 26a)
can be observed as a tight clusters in Fig. 26b. Cluster diagrams are not necessarily
drawn from “frequency vs. damping” scatter plots as the one shown in Fig. 26b.
Numerous quantities associated with the coordinates of mode shapes can also be
used to construct cluster diagrams, namely:
23 Civil Structural Testing 1289

– The mean phase, MP, which can be computed as the angle of the best straight line
fit of the mode shape in the complex plane. Bearing in mind the singular value
decomposition of the mode shape matrix , USVT = [ Re (j) Im (j) ], where
U and V are orthonormal columns and that V12 and V22 denote the elements
(1,2) and (2,2) of the V matrix, then the MP can be obtained as [71]:

3 4
1 2 −V12
MP φj = arctan (51)
V22

– The mean phase deviation, MPD, which measures how much each mode shape
coordinate deviates from MP and can be computed simply through the scalar
product between the mode shape matrix, [ Re (j) Im (j) ]T , and [ V22 − V12 ]T .
As a result, the MPD can be obtained as [71]:

 
-n    

φj i  arcos  ( 5j i ) 22 ( j i ) 12 
Re φ V −Im φ V
1 2
i=1  V12 22 | j i |
2 +V 2 φ 
MP D φj = -n   (52)
 
i=1 φj i

where φ ji is the deviation of the phase from the MP and n is the number of modal
coordinates.

– The mean phase collinearity, MPC, which measures how collinear, in the
complex plane, mode shapes are. It lies between 0 (not collinear at all) and 1
(perfectly collinear) and can be obtained as [71]:

6  62    1 1 2 2
6 6
1 2 6Re φ̃j 62 + C +1 sin (θMP C ) −1
1 2 2
ξMP C Re φ̃jT Im φ̃j 2 ξMP
MP C φj = 6  62 6  62
6 6 6 6
6Re φ̃j 6 +6Im φ̃j 6
2 2
(53)

where
6  62 6  62
6 6 6 6
6Im φ̃j 6 − 6Re φ̃j 6
ξMP C =     2
2
(54)
2Re φ̃jT Im φ̃jT
3 5 4
φMP C = arctan |ξMP C | + sign (ξMP C ) 1 + ξMP
2
C (55)
1290 R. Boroschek and J. P. Santos

and where φ̃j is the vector whose components consist of:

-n
i=1 φj i
φ̃j i = φj i − (56)
n
The cluster plots shown in Fig. 26c–e reveal the cluster sets observed in the f
vs. MP, f vs. MPD, and f vs. MPC bi-dimensional spaces. The first one, Fig. 26c,
reveals a clear trend exhibited by the modes shapes of this case study, to align around
phase angles of 0, π /2, and –π /2, while the second, Fig. 26d, is similar, in pattern,
to the one showing f vs. damping (Fig. 26b). The cluster diagram shown in Fig. 26e
reveals the damping nature of the observed mode shapes of this particular case
study, which are either well-aligned (MPC ∼ = 1) and associated with proportional
damping, or not collinear at all (MPC ∼ = 0) and therefore associated with highly
nonproportional damping. Regardless of the additional results and conclusions that
can be obtained from these modal features, the important information underlying
their density in these bi-dimensional spaces can be of importance for modal
identification and experimental dynamics of civil structures under uncharacterized
operational loading.

7 Modal Estimation

Civil structural dynamic testing was traditionally conducted using forced vibrations
and small data acquisition periods, mostly due to the technical limitations associated
with acquiring large data sets and with measuring small magnitude vibrations.
Nowadays, sensors and data loggers are capable of measuring changes of one
hundredth of the micro-g, where mainstream hard disk drives can store several
terabytes of data and where computer processing units (CPU) can process the data
in real time for dynamic feature identification. As a consequence, civil dynamic
testing can be conducted continuously and autonomously for the purpose of system
identification and structural health monitoring [55, 68, 82].

7.1 Distance Matrices

Diagrams such as those shown in Sect. 6 can be obtained from pairs of modal-
based quantities to reveal patterns that can be directly matched to different modes.
This type of patterns can, however, be obtained from each modal quantity alone
by computing distances between pairs of poles. These distances between each pair
of poles can be organized, in a simple manner, using distance matrices whose
visualization, for example, in the form of pixel maps (as those shown in Fig. 27),
allows visualizing how the modal information is organized as well the ability of
each quantity to provide reliable modal estimations.
The quantification of the distances can be conducted using several metrics, from
which the Minkowski distance is the simplest and well-known. It can be defined, in
its relative form, as
23 Civil Structural Testing 1291

Fig. 27 Pixel plots of the distance matrices obtained from (a) frequencies, (b) damping ratios, (c)
1-MAC, (d) MP, (e) MPD, and (f) MPC
1292 R. Boroschek and J. P. Santos

1-n  p 2 1
1 2   p
i=1 xi − xj
d xi , xj = 1  2 (57)
max |xi | , xj 

where the parameter p must be taken larger than the 1. When p=1, the distance is
known as Manhanttan distance and assumes the form of the absolute value of the
difference between modal features. When p=2, the Minkowski distance assumes the
form of the well-known Euclidean distance, or square root of sum of squares. For
all matrices shown in Fig. 27, this distance with p=1 was used, with the exception
of Fig. 27c, which was defined as d=1−MAC, where the MAC stands for the Modal
Assurance Criterion (MAC), [1], as follows:

-n 
 ∗ 2
i=1 φiA φiB
MAC (φiA , φiB ) = -n ∗ ∗ (58)
i=1 φiA φiA φiB φiB

where φ iA and φ iB represent modal coordinates of mode shapes A and B, respec-

tively.
Distance matrices and their corresponding pixel plots allow observing the
existence of modes through the existence of homogeneous groups estimations
with distances close to zero along the main diagonal of the matrices. This is the
case of the pixel plot representing the distance defined as d=1-MAC, shown in
Fig. 27c, where for the structural system considered herein, a clear main diagonal
comprising approximately 20 green colored groups can be observed. Along with
these groups, the existence of several located far from the main diagonal can also be
observed in civil testing, revealing the similarity between mode shapes with different
characteristic frequencies.
When considering civil structures, there are applications where the modes are
clearly well-separated in the frequency domain. In these cases, the frequency-based
distance matrix can have a similar appearance as the one obtained from the MAC.
There are, however, numerous civil structures where the density of modes in the
frequency domain renders the frequency matrix less clearly defined along the main
diagonal, as is the case of the structural system under analysis (Fig. 27a). The use
of each modal feature for modal identification is therefore case dependent and must
be considered for each application.
Regardless of the structural system under analysis, it must be referred that a
common result obtained from civil testing of structures under operation consists
in observing that the damping, the MP, the MPC, and the MPD distance matrices,
exemplified in Fig. 27b, d–f, do not generally exhibit compact regions only along
the main diagonal, since all these quantities repeat themselves across several distinct
modes, especially when large frequency domains and numbers of modes are to be
identified.
It must be referred, however, that these four modal quantities can be used
for the analysis of narrow frequency domains, for distinguishing between modes
with similar frequency values and whose shapes appear similar for the sensor set
23 Civil Structural Testing 1293

installed. In these cases, the phase and collinearity can also contribute to the correct
identification.

7.2 Spurious Pole Elimination Using Hard Stability Criteria

Modal estimates obtained from the time-domain identification methods can produce
poles which can be associated with physical modes or with other effects influencing
the dynamic test under analysis. The latter are named spurious poles (see Chapter
12) and can be characterized by not being observed across the model orders. As
a result, a common procedure for distinguishing between physical and spurious
poles consists of defining limits (named as hard criteria) to the values of the modal
features, namely those defined in section 6 (f, d, MAC, MP, MPD, or MPC). These
limits are generally in accordance with the precision requirements needed for the
experimental dynamic property at hand.
Hard criteria can be seamlessly obtained from the distance values between pairs
of modal estimates (shown in Fig. 27). One possible procedure for hard criteria
implementation consists of analyzing the distances between poles obtained at each
model order, n, with those obtained for the preceding model order, n−1. The
algorithm associated with this type of stability criteria is based on searching if there
is at least one pole j, in model order n−1, that meets all or a set of the following
conditions when compared to pole j, of model order n,
1 2
df (i, j ) ≤ lim df (59)

dd (i, j ) ≤ lim (dd ) (60)

dMAC (i, j ) ≤ lim (dMAC ) (61)

dMP (i, j ) ≤ lim (dMP ) (62)

dMP D (i, j ) ≤ lim (dMP D ) (63)

dMP C (i, j ) ≤ lim (dMP C ) (64)

where the ranges used are highly case dependent and a set of possible ranges used
for the frequency, damping and MAC limits can be chosen as,
1 2
0.005 ≤ lim df ≤ 0.05 (65)

0.01 ≤ lim (dd ) ≤ 0.1 (66)

0.999 ≥ lim (dMAC ) ≥ 0.95 (67)

1294 R. Boroschek and J. P. Santos

The application of this procedure using the limits lim(df )=0.005, lim(dd )=0.01,
and lim(dMAC )=0.999 resulted in the new stability and cluster plots shown in
Fig. 28a, b, where it can be observed that numerous poles located in non-dense
regions of the cluster plot were considered spurious and therefore eliminated.

Another possible procedure based on hard stability criteria consists of checking,

for each pole j of any order, that all or a set of the conditions shown above are met
between this and all the other i poles present in the pole set, regardless of their order.
Under this procedure, a pole is considered stable if the conditions are met between
it and an arbitrary number of s different poles, which can be defined as a fraction of
the maximum model order. The results obtained using these procedures are shown
in Fig. 28c, d for an s value of 5 poles.

7.3 Modal Characterization Based on Parametric Procedures

Modal characterization of civil structures resorts to one of the methods shown in

Sect. 5, or others, to estimate modal quantities, after which one or more steps can
be taken to ensure that these estimates are free of noise or of other spurious effects
[70]. Finally, each of these features can then be analyzed to characterize each of the
structural modes.
The characterization of the modes may follow the manual identification of
either (i) maximum values in spectra generated by frequency or time-and-frequency
methods or (ii) vertical alignments in stabilization diagrams/clusters in cluster
diagrams generated by time-based estimation methods.
Semi-empirical strategies may comprise the automatic or manual identification
of local maxima in spectra or in histograms computed from stability diagrams [70,
73]. As alternatives to these comprise strategies, those based on curve fitting using
splines, polynomial, or auto-regressive models can also be used [54, 72].

7.4 Modal Characterization Using Clustering

An alternative nonparametric class of nonparametric methods also used for civil

structural testing relies in clustering algorithms, which are pattern recognition
that are capable of, by themselves, labeling data simply by studying the intrinsic
structure of the pole set [8, 54, 71, 84, 85]. The aim of a clustering method can
be defined as the division of a data set into groups, which must be as compact
and separated as possible according to a certain predefined metric. To fulfill
this objective, allocation rules must be defined so that pair-wise dissimilarities
between objects assigned to the same cluster tend to be smaller than those allocated
in different clusters [28]. Considering a given partition containing K clusters,
Pk = {C1 , . . . , Ck }. The within-cluster distance W(Pk ) can be defined as [39]:
 23 Civil Structural Testing

Fig. 28 Spurious pole elimination: (a) stability and (b) cluster diagrams using by-order pole elimination, (c) stability, and (d) cluster diagrams using entire-set
pole elimination
1295
1296 R. Boroschek and J. P. Santos

1  
K
W (Pk ) = dij (68)
2
k=1 C(i)=k C(j )=k

where C(i) is an allocation rule which assigns element i to cluster k, based on a

distance dij . The total variation of data can be defined as in Eq.(69), where N is
the total number of objects considered in the cluster analysis. Finally, the between-
cluster distance B(Pk ) can be simply obtained by subtracting the other two defined
distances, B(Pk ) = T − W(Pk ).

1 
N N
T = dij (69)
2
i=1 j =!

Each cluster can be described by a prototype, which generally consists of an

object of the same type as the ones being clustered. The location of each prototype
can be obtained using centroids, medoids, or other representatives.
Numerous families of clustering methods can be found in the literature but the
most used are the iterative and hierarchical methods, from which the k-means is the
most well-known. It requires the definition of an initial set of clusters’ prototypes,
which is known as the initialization phase, after which each iteration starts by
allocating the objects to the clusters according to an allocation rule, c(i). This
step is represented in Fig. 29b, where the dashed lines show the allocation of
each object into the closest cluster prototype, according to the distance defined.
The second step of each iteration is named representation and consists of finding
the set of K prototypes that best represents the clusters defined in the allocation
phase. The k-means is based on representing the clusters by their centroids (new
prototypes shown in Fig. 29c). These two steps, allocation and representation, are
afterwards repeated (Fig. 29d, e) until an objective function, which depends on
cluster compactness and separation, reaches its global minimum value. The k-means
takes the squared within-cluster dissimilarity measured across the K clusters as
objective function [39].
The hierarchical agglomerative methods provide a merging hierarchy so that
all partitions are defined, regardless of their number of elements [39], and starts
by considering single-object clusters and, at each level of the hierarchy, merge a
selected pair of clusters into a new one. This merging produces a new level in
the hierarchy (which contains one less cluster). This hierarchy can be displayed
in a dendrogram plot, which allows for a clear visualization of the structure of
high dimensional data, such as modal features acquired from dynamic tests, in a
single and unambiguous plot such as the one shown in Fig. 29g, where clusters are
represented by vertical lines and hierarchy levels by horizontal lines.
Cluster merging is made, at each level, based on merging rules, among which
the two simplest are the “single link” and “complete link.” The first states that the
two clusters containing the closest objects in a data set should be merged, the latter
chooses the ones with the farthest objects. The most widely used merging rules are
23 Civil Structural Testing 1297

Fig. 29 K-means clustering algorithm: (a) initialization (three cluster prototypes to describe
eleven objects); (b) iteration 1, allocation; (c) iteration 1, representation; (d) iteration 2, allocation;
(e) iteration 2, representation, (f) final set of clusters’ prototypes. Hierarchical agglomerative
clustering (g)

the “average link” and the “ward,” or “minimum variance” rule. The first defines
that the pair of clusters to be merged, at each level, is the one with smaller average
distance, while the “ward” rule defines that, at each level, the pair of merged clusters
must generate a new partition with the smallest variance possible. From the defined
agglomerative hierarchy, cluster partitions containing any number of clusters (from
1 to N) can be obtained by cutting the dendrogram plots (horizontally) between two
hierarchy levels, as show in Fig. 29g.
Clustering methods, regardless of their type or nature, use distance matrices
obtained between pairs of all elements being clustered as input. For modal analysis
in the realm of experimental dynamics, distance metrics can consist of those
obtained from the quantities presented in Sect. 6 or others considered appropriate
for the application at hand Fig. 27.
To exemplify the automatic modal characterization using clustering, each of the
six matrices shown in Fig. 27 was used as input in an independent hierarchical
agglomerative procedure based on the minimum variance rule, without any elimina-
tion of spurious poles, and assuming that 24 modes are present in the domain [0;2]
Hz. The results are shown in Fig. 30, in the form of stability diagrams overlapped
1298 R. Boroschek and J. P. Santos

Fig. 30 Application of the hierarchical agglomerative clustering to distances metrics obtained

from applying the Manhattan distance to: (a) frequency; (b) damping; (c) 1-MAC; (d) MP; (e)
MPD; and (f) MPC

with the spectra of the first singular value (extracting using the method described in
Sect. 5.5).
A possible criterion to analyze whether a clustering process has identified the
modes adequately consists of analyzing whether it classifies each entire vertical
alignment of poles as a single cluster. When conducting this analysis to the results
shown in Fig. 30, it can be observed that, for the present case, the clustering
processes which used frequencies and MAC as input generated more trustworthy
mode classifications, when compared to the remaining ones. This conclusion is in
line with the one obtained from the analysis of the distance matrices themselves,
shown in Fig. 27. However, it can also be observed that even if the matrices obtained
from frequencies and MAC are significantly different (Fig. 27a, c, respectively), the
resulting clustering obtained from them generates similar mode configurations.
23 Civil Structural Testing 1299

Fig. 31 Application of the hierarchical agglomerative clustering to a distance metric obtained

from applying the Manhattan distance to frequency and 1-MAC of the entire pole set

Fig. 32 Application of the hierarchical agglomerative clustering to a distance metric obtained

from applying the Manhattan distance to frequency and 1-MAC of the pole set after spurious pole
elimination

As an alternative to clustering based on single-feature distances, procedures

using matrices obtained from several features can also be used and may prove
advantageous by combining the information underlying each feature. The mode
identification based on the combination of frequencies and mode shapes is exem-
plified in Fig. 30 for the entire pole set and in Fig. 31 for the non-spurious poles,
after the elimination described in Sect. 7.2. As it can be observed, both generate
trustworthy mode classifications, thus suggesting that clustering can be used both
for spurious pole elimination as well as mode characterization (Fig. 32).

7.5 Assessing the Number of Modes Using Cluster Validity

Clustering algorithms are able to assign poles to any amount of predefined modes
(each represented by one cluster), irrespective of whether this number of modes
really exists in the structural system’s dynamic response. Even though the analysis
of pole allocations, such as the one shown in Fig. 31, is able to provide hints on the
number of modes observed, the results may become subjective. To exemplify the
problem of generating any number of clusters without knowing the real number
of mode shapes in advance, Fig. 33 shows the application of clustering to the
pole set analyzed so far using 5–55 clusters to represent the modes. Its analysis
1300 R. Boroschek and J. P. Santos

Fig. 33 Application of the hierarchical agglomerative clustering using: (a) 5; (b) 15; (c) 25; (d)
35; (e) 45; and (f) 55

allows observing that 5 or 15 modes (Fig. 33a, b) appear to be too few to represent
the dynamic response of the structural system, since several different vertical
alignments are classified into the same modes. However, it can also be observed
that 45 and 55 clusters (Fig. 33e, f) divide several vertical alignments into different
modes, and therefore might not be representing the dynamic response well.
To obtain a quantitative and objective estimation of the number of modes
observed in an experimental data set, an evaluation known as cluster validity [96]
may be conducted on any pole assignment. Cluster validity consists of computing
indices for all possible pole classifications [58, 69] containing from k=2 modes to
an arbitrary number which can be as large as the application at hand allows, and can
be much larger than the number of degrees of freedom measured on site.
Each validity index provides a measure of each poles’ cluster compactness
and separation, regardless of the number of modes present in the evaluated
23 Civil Structural Testing 1301

classification. The most truthful number of modes is then assessed by comparing the
indices obtained from each of the evaluated partitions. Depending on each index’s
theoretical background, its maximum or minimum value will point out the most
truthful number of clusters.
The fact that cluster validity indices are based on different theoretical back-
grounds can make their ability to analyze data and point out different structural
behaviors case-dependent [5, 58, 96]. For the purpose of identifying the number
of modes, the Global Silhouette (SIL) index can be used [74]. To define its
mathematical formulation, a pole set of N objects and K mode partitions of this set
into t modes, Pt = (C1 , . . . , Ct ), is considered
 as well as its
 k-th mode, consisting
(k) (k)
of Mk poles 1 ≤ Mk ≤ N, defined as Ck = T1 , . . . , TMk .
The silhouette width of the i-th object belonging to Ck is defined as

(k) bi(k) − ai(k)

si =   ∈ [−1, 1] (70)
max ai(k) , bi(k)

(k)
where ai is the average distance between the i-th pole of Ck and the remaining j
poles assigned to the same cluster is given by


Mk −1
(k) 1
ai = dij , 1 ≤ i ≤ Mk (71)
Mk − 1
j =1
i = j

(k)
Similarly, the minimum average distance bi between object i and all the objects
clustered in the remaining clusters is given by
⎛ ⎞
1 
Mr
(k)
bi = min ⎝ dij ⎠ , 1 ≤ i ≤ Mk (72)
r = 1, . . . , K Mr
j =1
r = k

where r is any cluster of the partition Pt with Mr poles.

Finally, the silhouette index of cluster Ck , sk , and the global silhouette index of
the entire partition containing t modes, SIL(Pt ) , are respectively given by

1 
Mk
(k)
sk = si (73)
Mk
i=1

1 
K
SI L (Pt ) = sk , t = 2, . . . , K (74)
K
k=1
1302 R. Boroschek and J. P. Santos

Fig. 34 SIL index obtained for cluster partitions containing K=2–75

where the higher Silhouette (SIL) corresponds to the partition exhibiting more
compact and separated clusters, and thus the most truthful classification of poles
into the most correct number of monitored modes. For the pole set analyzed herein,
and shown in the previous figures, the maximum SIL value is obtained for K=36
modes (Fig. 34). However, it can be argued that cluster validity cannot be applied
blindly without considering the context and physical problem at hand. Even if the
higher value of the SIL index (or of other validity indices) might indicate the best
clustering partition, its value may be too high to consider for the problem at hand.
An alternative criterion consists of choosing the first local maximum, as long as
it is located within a domain which is known possible for the structural system
under analysis and the number of sensors deployed on site. For the example shown
so far, this criterion yields K=27 modes (Fig. 34), which appear to be more in
accordance with the dynamics of the structural system at hand since it leads to less
splits divisions of vertical alignments in the stability diagram shown in Fig. 35b
when compared to the one obtained for K=36 clusters, shown in Fig. 35a.
The maximum number of modes considered for cluster validity can also be
defined as a divisor of the maximum order chosen for the modal estimation
algorithm, and shown in the stabilization diagram. In Fig. 34, clustering partitions
were analyzed up to the value of the maximum order, K=75.

7.6 Cluster Representatives as Mode Features’ Estimates

Clustering methods and validity indices allow assessing which poles are allocated to
each of the modes to characterize it according to the chosen modal features. These
allocated pole subsets exhibit, however, variability and each pole by itself may not
be the best representative of each mode shape. Thus, to find the best modal features
characterizing each mode, accurate clusters’ representatives can be obtained.
Several types of clusters representatives can be obtained, and their definition
can greatly influence the result of the clustering process and its results. Both the
23 Civil Structural Testing 1303

Fig. 35 Application of the hierarchical agglomerative clustering with the Manhattan distance
obtained from frequencies and 1-MAC for K=27 and 36 clusters

k-means and the hierarchical agglomerative use the clusters’ centroids (average
coordinate values), computed in the p-dimensional space, where p is the number
of modal features used to defined the distance matrices above described [80, 82].
Some methods such as the k-medoids use the medoids (median coordinate values)
to obtain such estimates [79] and others use density-based representatives [8].
To exemplify the definition of the cluster representatives for the example shown
so far, the partition comprising K=27 modes was considered and from it the cluster
centroids were computed as representatives of each mode. These are shown as dotted
lines in the stabilization diagram represented in Fig. 36a and crosses in the cluster
plot represented in Fig. 36b. The corresponding mode shapes are also obtained in
the same manner, from those associated with each pole. The first six shapes can be
found in Fig. 37.

8 Modal Tracking

8.1 Problem Statement and Motivation

The need for modal tracking arises from structural health monitoring of civil
structures, whose requirements nowadays include permanent and continuous iden-
tification of damage in an early stage, not only for ensuring safety but also for
1304 R. Boroschek and J. P. Santos

Fig. 36 Stabilization (a) and cluster (b) diagrams of the modes comprising K=27 clusters

infrastructure management [23, 79, 83]. Continuous health monitoring demands that
civil dynamic testing be conducted also continuously under operation, and therefore
in an automated manner [16, 49, 85].
In practice, continuous dynamic testing can be seen as a set of singular tests
conducted at predefined successive time steps, taken with a time difference equal
to the needs of precision and detection readiness. Each of these singular successive
tests can generate a set of modes, or mode representatives, such as those shown in
Fig. 36 and which, over time:

– May be associated with true modes of the structural system or with spurious
estimates
– May be excited in some of the singular tests but not in others, depending on the
operational and environmental loading
– May exhibit different values according to the operational and environmental
loading or to changes in the structural system

For this reason, the label associated with each mode, at a certain singular
test conducted at a specific time-instant, may differ when obtained in a different
time-instant, on the same structural system and using the same sensor set. The
task of establishing the relation between the same mode characterized in different
periods of time in the same structure can be named as modal tracking. Modal
tracking strategies can assume several formats and strategies, from which boundary-
based tracking, where the modes’ representatives are controlled by comparison
against pre-known modal features, and cluster-based tracking, where natural modes’
 23 Civil Structural Testing

Fig. 37 Mode shapes in the complex plane

1305
1306 R. Boroschek and J. P. Santos

Fig. 38 Time series of modal representatives acquired hourly on a suspended steel bridge

representatives are allocated using cluster analysis and without the need for pre-
known features, are popular in civil structural testing. Each of these and other
strategies found in the literature has, in turn, different variants depending on the
authors and the specificities of each structural system being tested [54, 85, 108]
To have examples of strategies for modal tracking, the example shown in Sect.
7 with modal estimation conducted from February to May 2018 is considered.
The mode representatives were obtained hourly during this period by applying the
SSI method (see Sect. 5.7) followed by cluster-based strategy described in Sect.
7. The resulting time series of modal representatives is shown in Fig. 38 for the
frequency domain [0.9;1.3] Hz, where four to six near-horizontal alignments span
over time, indicating that modes’ frequencies are repeating themselves throughout
the monitoring period under analysis. In the same plot there are, however, frequency
bands with higher variability than others, indicating either important sensitivity to
operational and environmental loading, or the existence of close (maybe overlapped)
modes, in the frequency domain.

8.2 Tracking Based on Boundaries

The definition of boundaries for modal tracking can be used for civil dynamic
testing, for the continuous safety control of operating civil structures. This type of
modal tracking requires that the modal features (frequencies, damping ratios, and
mode shape coordinates) of modes be known in advance and is usually based on the
definition of upper and lower boundaries for each feature (f, d, MAC, MP, MPD, and
MPC) and for each of the modes, identified herein with index j. These boundaries
can generally be defined as,
1 2 1 2
fi ∈ 1 − αf,inf fj ; 1 + αf,sup fj (75)

1 2 1 2
di ∈ 1 − αd,inf dj ; 1 + αd,sup dj (76)
23 Civil Structural Testing 1307

1 2 1 2
MACij ∈ 1 − αMAC,inf ; 1 + αMAC,sup (77)

1 2 1 2
MPi ∈ 1 − αMP,inf MPj ; 1 + αMP,sup MPj (78)

1 2 1 2
MPDi ∈ 1 − αMPD,inf MPDj ; 1 + αMPD,sup MPDj (79)

1 2 1 2
MPCi ∈ 1 − αMPC,inf MPCj ; 1 + αMPC,sup MPCj (80)

where α QUANTITY, sup and α QUANTITY, sup express the ratio between the upper or lower
boundaries and the corresponding representative value of each known-in-advance
mode. For simplicity, the same α value can be used for each quantity across all
modes and even across features or entirely variable across modes and features and
even across the time periods under analysis. Depending on the structural system and
on the influence of environmental and operational actions on the mode estimation, it
might be necessary to define boundaries, α QUANTITY , and reference feature values, fj ,
dj , MACj , MPj , MPDj , and MPCj , as functions of the values assumed by the actions.
Each newly arrived cluster representative, identified with the index i, can therefore
be allocated into the pre-known mode j if it is within a chosen set of the boundaries
represented in Eq. (75) to Eq. (80).
To exemplify the application of boundary-based modal tracking, the time series
shown in Fig. 38 is considered with the objective of tracking the mode with a
frequency value of approximately 1.15 Hz using as features the frequencies, the
damping ratios, and the MAC values and a value of α=0.1 across for all features.
The results of the tracking are shown in Fig. 39, where the mode representatives that
do verify are represented in blue color whereas the remaining ones are represented
in gray color.
The boundary-based tracking defined for frequency alone can be observed in
Fig. 39a, where the upper and lower boundaries can be directly identified in the
plot. Conversely, and in line with the results obtained for the modal estimation, the
boundaries applied to damping do not produce as clear divisions in the frequency
domain (Fig. 39b) since a large number of modes observed in a structural system
exhibit identical damping values. Unlike for damping, the unique shape of each
mode seems to produce, for the structural system under analysis, clear divisions
between the representatives of a mode and those representing the remaining ones,
as it can be observed by the clear time-series produced by applying the boundaries
on the MAC values (Fig. 39c). This is a common result for civil structural testing,
where for some structures the MAC is a useful feature for separating modes with
identical frequency value. Unsurprisingly, when applying boundaries to sets of these
features, for the present case study, those which include damping are less performant
and those which are based on the MAC outperform the remaining ones (Fig. 39c–g).
1308 R. Boroschek and J. P. Santos

Fig. 39 Modal tracking of mode with f =1.15 Hz using α=0.1 and the following features (a)
frequency, (b) damping, (c) MAC, (d) frequency + damping, (e) damping + MAC, (f) frequency
+ MAC, and (g) frequency + damping + MAC

8.3 Cluster-Based Modal Tracking

Modal tracking is based on clustering resorts to the same principles as cluster-based

modal estimation, described in Sect. 7.4. As in that section, a clustering method
is applied to a certain data set to allow for the best division between modes [54,
80, 85], followed by the computation of a cluster validity index [80] to ascertain
23 Civil Structural Testing 1309

the most likely number of modes present. The difference (and additional challenge)
consists of the fact that, for modal tracking, the data is dispersed through time and
therefore might suffer slight variations in its inner structure, which may demand for
additional procedures, depending on the structural system and on its actions’ effects,
such as moving-windows processes, baseline creation for posterior direct allocation,
among others.
To provide an example of cluster-based tracking, the hierarchical agglomerative
clustering was applied a single time to the time series shown in Fig. 38, with a cut
at k=6 clusters (modes). Its results allow observing (Fig. 40) that the clustering
method is capable of dividing the time series in different subsets according to the
distance metrics used. When considering frequencies, the division relies mostly on
the density of the representatives in the frequency domain, as it can be seen in
Fig. 40a, while clustering based solely on damping generated a mixture of the modes
representatives over time (Fig. 40b), as happened with boundary-based tracking.
Unlike when using the frequencies and damping, the MAC-based cluster tracking
(Fig. 40c) allowed observing modes with dissimilar shapes and identical frequency
values, such as those observed in the frequency domain between 0.95 Hz and 1.05
Hz, in Fig. 40c, where two mode shapes were identified by the clustering methods
in a dense region of representatives. These overlapped modes can be observed in
large symmetric civil structures, where the MAC-based tracking assumes particular
importance. The modes time series obtained by applying clustering to distances
obtained from frequencies + MAC are similar to those obtained for the MAC alone,
yet the addition of the frequencies as input seems to produce less miss allocations.
The application exemplified herein (Fig. 40) consists of a single-time cluster
process applied to a large time-series, which can be used in practice. However, as
referred in the previous paragraph, clustering can also be applied in smaller time-
windows which can slide along time to consist of a baseline-free modal tracking
procedure. An alternative may consist of conducting the procedure described herein
for defining a baseline without pre-known knowledge of the structural system,
after which the mode representatives obtained over time are compared against and
allocated. Combinations of both and of other methods can also be used. Neural
networks, decision trees, and other supervised classification algorithms can be used
to learn the baselines defined either manually or by a clustering process such as the
one described in order to improve the results of posterior allocation. When using
these methods, environmental and operational actions can seamlessly be included
into the process, with the possibility of large improvements in the results over time.
The assessment of the number of clusters assumes particular importance for
cluster-based modal tracking and can make the difference in delivering sensitive and
robust Structural Health Monitoring strategies. When too few modes are considered,
different high variability of some time series can be obtained, thus leading to large
confidence boundaries for safety control. This is the case of the partition containing
K=4 clusters obtained from the time series shown so far (Fig. 41a) and where it
can be observed that the blue cluster is allocated to representatives located in the
frequency band around 1.0 Hz and in a small band characterized by f =1.3 Hz.
When too many modes are considered, the appropriate clustering method may tend
 1310

Fig. 40 Modal tracking of modes with frequencies between f =0.9 Hz and f =1.3 Hz using distance metrics obtained from (a) frequency, (b) damping,
(c) MAC, and (d) frequency + MAC
R. Boroschek and J. P. Santos
23 Civil Structural Testing 1311

to allocate the additional clusters to noise/spurious modes, thus allowing for their
seamless identification and cleansing. This is clearly identifiable in the cyan cluster
of Fig. 41d, which appears to be associated with a spurious mode, and in the light
green cluster of the optimal partition, with the highest SIL value (Fig. 41c).

9 An Important Note to Increase the Speed of Identification

Most of the techniques use the stability diagram to select physical modes. This
procedure requires the formulation of the response matrix that is later decomposed
by the QR transform for different dimensions which could be quite slow. Döhler
and Mevel [29] developed a methodology to compute only one QR transform and
to derive the smaller dimension matrixes in a simple way speeding the process
considerably.
An additional recommendation to increase the speed of the identification, when
several sensors are present in a structure, is to use a small number of channels to first
identify the frequency and damping and later use a larger set of sensors to identify
the mode shape more densely. In order to do this, you need to apply the identification
technique for the shape given the frequency and damping of the mode is known like
the one presented by Caicedo and Marulanda [18]
A comparison of the methods described above can be found in Peeters and
Ventura [64] and Giraldo et al. [33].

10 Application Examples

10.1 Building Excited by Ambient Vibration

There are two common testing applications, namely those based on temporary and
permanent sensor arrangement.

10.1.1 Temporary Testing

The temporary arrangement can be performed by wired or wireless sensors. For
identification of mode shapes, it is critical to pay attention to the synchronization
between sensors. For wire systems this is typically attained by using an arrangement
of analog sensors connected to a single digitization station that provide the common
timing. For wireless distribution sensors, in some limited cases the sensors can be
synchronized by radio frequency or others, but in general when moving between
floor this is quite limited. So, the most reliable procedure for wireless-based testing
is to have a very precise clock in each sensor and this clock should be synchronized
at the beginning and end of the measurement and in some long duration monitoring
at intermediate steps. To obtain reliable results the maximum time drift between
sensor times should be smaller than 1 ms.
When a detailed mode shape is required, then two options are available, either
sensors are installed in all desired location at each set up or a set of reference sensors
 1312

Fig. 41 Modal tracking of modes with frequencies between f =0.9 Hz and f =1.3 Hz using distance metrics obtained from frequency + MAC and a number of
R. Boroschek and J. P. Santos

clusters equal to (a) K=4, (b) K=5, (c) K=6, and (d) K=7
23 Civil Structural Testing 1313

(fixed sensors) is defined along with a set of movable or rowing sensors. In this last
case a much smaller number of sensors is required and the modal shape can be
reconstructed later using the reference sensors amplitude and phase. For the case
of rowing sensors for each set up, the system identification procedure will provide
the frequency and damping of each mode and a partial description of the shape.
For several set ups there will be redundant data for frequency and damping and
complimentary data for the shape. The reference and wireless rowing sensors are
preferred for the procedure for long distance or highly detailed mode shapes not
only because cables installation is difficult, due to physical limitations, but also due
to the possibility of noise inclusion in the data.
The sensor sensitivity shall consider the level of response of the structure; never-
theless a sensitivity of 1e−5 g is generally sufficient in most cases. The bandwidth
is related with the range of frequencies of interest, but for low rise structures
a bandwidth between 0.5 to 15 Hz is generally sufficient. For tall buildings, a
bandwidth increase to 0.01–15 Hz may be required, generally demanding for more
expensive sensor sets.
Due to the low level of excitations (typically bellow 0.0005 g) and the bandwidth
of interest, the sensors can be deployed directly on the surface without any
permanent or temporary attachment.
Sampling frequency has to be related with modes of interest but typically 50 Hz
is sufficient and 200 Hz desirable.
Duration of monitoring depends on the objective of the monitoring. When a
particular environmental response is desired the duration must consider the period
of the desired vibrations. Typically, measurement durations are 20–60 min. The
longer time recordings allow for the evaluation of variations during the monitoring
period due to possible sensitivity to the amplitude and frequency response of the
structure.
The identification methods also impose requirements on the duration of the
recording. Typically, the frequency domain methods require extended periods of
time in order to obtain a more robust estimate of the response function. Time domain
methods generally require considerable shorter recording times.
Recording building vibrations present challenges related to usage due to
increased level of noise, namely people moving, causing localized slab vibration,
machinery operations such as elevators, air conditioning, copy machines, printers,
and electrical noise produced by illumination and communication system, among
many others. Several measures can be taken to reduce their impact and one of
the simplest and most effective consists of recording long periods of time and
latter select sections of the recorded data that have a smaller influence of noise.
Additionally, the application of procedures to reduce the noise effect by filtering
and averaging is commonly used.
IT should be pointed out that buildings’ modal characteristics are sensitive to
amplitude of the response and environmental conditions (temperature, time of day,
rain (even previous rain cycles), and weather conditions). Hence, it is very important
to document and record the specific occupations levels and activity, as well as
environmental conditions. The most relevant conditions are temperature, humidity
1314 R. Boroschek and J. P. Santos

of the ground, freezing conditions and mass additions due to snow and usage.
For short duration monitoring, the environmental conditions can be considered
constant. Nevertheless, if the monitoring extends by hours and a sudden change
of loading amplitude or weather conditions occur, this may affect the structural
properties.
The best practices are to report at least the following:

(a) Building location

(b) Building structural characteristics and soil conditions
(c) Weather conditions
(d) Monitoring Time
(e) Sensor characteristics (bandwidth and sensitivity among others)
(f) Sensor locations and direction
(g) Number of recording setups
(h) Duration of each recording
(i) Sampling frequency
(j) Particular situation that occurred

Results should include

(k) Figures showing the recording signal in time and frequency space
(l) Procedures and consequence of data cleansing
(m) Filtering and decimation
(n) System identification used and their controlling parameters.
(o) Modal results including the confidence of the analyst on the results

10.1.2 Permanent Ambient Vibration Monitoring

All the comments presented for temporary ambient monitoring apply in this
case; nevertheless the long-term monitoring requires several modifications on the
equipment and the analysis procedures.
For this case, sensor location is fixed, so the required sensor density to describe
modal shape defines the number of permanent sensors and their recording directions.
In this case the sensor must be anchored to the structure to avoid motion due the
normal long-term operation of the building. Cabling between sensors and central
recording units or for synchronization between digital units is convenient unless
other procedures are used for synchronization between especially distant sensors
(now day PTP is being used). Energy must be provided permanently and a UPS
system must be considered if required for continuity of operations. Special attention
should be given to the selection of the recording system for long-term operations
(i.e., temperature conditions and an automatic restarting).
Because for permanent continuous monitoring the data acquired can be very
large, different strategies for data cleansing, conditioning, identification, tracking,
storage, and referencing should be considered. All these processes must be done in
real time or near-real-time. It is not reasonable nowadays to record data continuously
and analyze it later.
23 Civil Structural Testing 1315

The cleansing procedures must be able to detect at least:

(a) Periods of malfunction of the system

(b) Number of operative sensors
(c) Noise level present in each channel
(d) Procedures to discard noisy channels to avoid alterations of the identification
and the tracking activities

Data conditioning must be able to process the data based on preestablished

parameters or adaptive parameters.
System identification must be conducted in real-time, or use selected periods
of monitoring time to perform a window system identification. Results should
include description of the analysis window, the conditions of the structure, and the
environment at the time of the analysis.
To handle the large amounts of data several strategies have been applied. Without
been exhaustive, the following are among the most common:

(a) Obtain parameters of interest at predetermined time periods or response

characteristics. Tag and storage of the derived results. Store only a subset
of the original time data for post processing. Processed data is informed via
electronics means (SMS, email, periodic electronic reports (daily, weekly or
when a change of state is detected)).
(b) Similar to (a) but raw data is stored locally and periodically. If not required
in a predefined period of time (for example a month), data is removed from
the system.
(c) When allowed by the communication bandwidth, raw data is transmitted
to a processing server where data is processed, made available, and stored
for future analysis. In this case the cost of store raw data can be very high
depending on its duration, sampling frequency, and number of channels. This
procedure is recommended at the beginning of a project. After robustness
of the whole process is obtained, some of the storing requirements can be
relaxed but maintaining raw data allows the revaluation and the testing of
new procedures.

The data is typically analyzed by using moving windows. The window length will
depend on the system identification used and the robustness required for the results.
For example, if methods in the time domain are used, a representative value can
be obtained for an hour of measurements by performing 12 consecutive analyses
in corresponding 5 min of non-overlapping data. Each 5-min data will provide
information on the desired parameters and the 12 results could represent hour values
with appropriate information on the dispersion of the results.
Because excitation levels are different during recordings, each analysis window
could contain different modes and for a given mode its value can vary. So, a tracking
strategy must be developed to link consecutive modes that are identified as the same
in different time windows. This is generally solved using distance metrics strategies
1316 R. Boroschek and J. P. Santos

as mentioned earlier, for example, in Gonzalez and Boroschek [35] the distance
metric used is shown in Eq. (81), as follows:

 
fref − fi  1 1 22
dref,i = (1 − α) α 1 − MAC φref , φi (81)
max (fi )
j

The distance is a combination of frequency and mode shape difference weighted

by a constant that is related to the number of sensors present at the time of the
analysis. For the case that no extreme events occurred between analysis windows,
two modes are consider the same if the distance is less than a threshold value.
After the classification of each mode is done, using the tracking algorithm,
the responses and characteristics of each identified mode can be expressed as
a function of time. This collection of information will present the effect of the
environment and noise. The most common effect is variation of the predominant
frequency with temperature and humidity in the structure and in the foundation’s
surrounding soil. The variations observed in typical building can achieve up to
10% without any presence of damage and higher in the presence of soil freezing.
As described previously, the continuous ambient vibrations monitoring is the most
reliable procedure to detect changes of state of health on the structure so this
sensitivity to environmental conditions must be considered if small changes in
modal and response parameters are desired.
The most common strategy in buildings is to perform a normalization of the data
by procedures previously described. When the information on the environmental
variables is nonexistence or limited, removal of the induced variation can be done
by a Principal Component Analysis (PCA) [102]. Alternatively, a modal parameter
dependency model based on recorded environmental variables can be used. The
model must consider the time lag between the response and the environmental
variables. For this case ARX, MLP and others have been used [65, 81].
After data normalization, the effect of environmental variables is diminished so
it is possible to detect more subtle state changes. Several procedures can be used
and control charts are commonly applied [103].
One of the case studies we have been presenting is an office building. It is
located in a semiarid region in Chile so it is excited by operations, wind, and
earthquakes under varying environmental conditions, as shown in Fig. 42. Sensors
are permanently located in the structure so evaluation of state changes can be
identified and tracked automatically.
Sensors are located in three levels of the structure, shown in Fig. 43. This allows
a global identification of the structure response and a limited representation of the
mode shape. The sensors are of the capacitive type with a sensitivity of 1 × 10−5
g and a range of 4 g. This allows the recording of the ambient and earthquake
vibrations with the same system. Sampling rate is set to 200 Hz to record the
amplitudes due to earthquake motions. For ambient vibration a sampling between
50 to 100 Hz will be appropriate.
23 Civil Structural Testing 1317

Fig. 42 Temperature variations recorded in TC Building

Fig. 43 TC Building

Sensors are anchored to the floor slab and they are located close to vertical
structural elements to limit the effect of the slab vibrations and deformations on the
DC component and its predominant frequency on the acceleration signal. Sensors
are protected by a cover to avoid impact due to human activities and are located
1318 R. Boroschek and J. P. Santos

away from doors, elevator shafts, and local permanent equipment (air conditioners
and others). All sensors are wired and connect to a signal conditioning system
that performs antialiasing filtering, digitizing and transferring to a local computer
system. In the computer system, the data is analyzed and stored.
Data is stored using two different schemes. For ambient vibrations, 15 min
records are saved to a hard drive and for earthquake events a trigger base criterion is
used. For ambient vibrations the data is analyzed according to the following steps:

1. Data Cleansing. Data is automatically inspected for noise contamination,

aliasing, recording errors, impacts that occur due to the use of the building and
earthquakes. The earthquake records in general are not used for the ambient
evaluations. The analysis uses the frequency content of the signal along with
the RMS and PGA acceleration and also the value and energy evolution of
the signal to identify noisy signals. An information vector is formed for the
specific record, indicating which sensor data of the set is valid or not.
2. Data decimation. Data is decimated to 20 Hz for identification. Larger sam-
pling frequencies do not contribute to the identification of modal properties
and increase the computing and storage demands, as shown in Fig. 16.
3. System Identification. All aboveground channels are used for the identifica-
tion. The SSI-COV data is used considering order between 20 and 100, as in
Fig. 22.
4. Selection of Physical Modes. Several strategies have been developed to
identify physical from spurious modes, as mentioned above. In this case a
density-based criterion is used, Fig. 23.
5. Modal Tracking. When the target application is a continuous monitoring
system, the mode identified in one time period should be linked to the next
observations. In the absence of damage this is generally done using a distance
criteria based on frequency and shape. Damping is not used because its values
are highly uncertain and many modes can have similar values. A typical
distance equation should consider a normalized distance and shape similarity,
typically using MAC. It should also consider the number of sensors that define
the mode, as shown in Fig. 44.
6. Mode Normalization. Due to the variability of environmental conditions, the
properties of each mode changes with temperature, humidity, and response
amplitude among others. If change of state detection is required, the influence
of the this variation should be considered or removed. One possibility is
to obtain the principal components of the data and remove the effects of
environment, as is shown in Fig. 45 to visualize the linear relation between
mode 1 and 2. Figure 46 shows the normalized frequencies. A much smaller
variation is observed, allowing for more clear visualization of slow or abrupt
changes. Alternatively, a model base on the environmental variables can be
developed using linear multivariate regression or ARX.
7. State Change. From the normalized data a state change can be identified using
control charts or other methods.
23 Civil Structural Testing 1319

Fig. 44 First three identified and tracked frequencies

10.2 Buildings Earthquake Vibrations

Earthquake loading is characterized by transient signals with high energy between

0.05 to 30 Hz. The response in buildings is typically measured by accelerometers
to capture the global behavior and, in some cases, they are complemented by strain
gages and displacement meters to detect local deformation information. Tilt-meters
are also used to determine residual deformations.
Similar requirements as those described for permanent ambient vibration mon-
itoring relating to anchoring, cabling, common timing and location are needed.
In addition, due to loss of power during an earthquake safe UPS and recording
system are needed with a typical autonomy of 2–4 days. For the case of earthquakes,
damage can be expected, so sensor location should consider all possible expected
scenarios. Some guidelines are described in the following points:

1. Input recording. At the lowest basement level and at the ground surface level
a minimum of three horizontal sensors are recommended to capture horizontal
motions and possible wave passage causing plane rotations. Vertical sensors
should be located at the end of walls and at the center of the foundation to capture
vertical motions and local and global rocking.
1320 R. Boroschek and J. P. Santos

Fig. 45 Mode 1 vs. Mode 2

2. Floor displacements and rotations. To capture the motions of the floor slabs,
three horizontal sensors should be installed. To capture the in-slab deformation
additional horizontal sensors are needed. To capture vertical amplification and
rocking, sensors should be located at opposite sides of the buildings and at
borders of structural walls.
3. Storey drifts. To capture drifts the maximum number of possible floors should be
instrumented. Preference should be given to the first two floors (where damage
is typically concentrated) and others distributed in height.
4. Derived inertial forces. Inertial forces can be estimated based on floor accelera-
tions. These forces can be used to estimate floor and base shear and overturning
moments.
5. Vertical amplifications. In very tall buildings vertical amplification of motions is
possible, so vertical sensors should be distributed in height.
6. Irregularity and damage scenarios. If predefined damage scenarios are the
objective of the instrumentation, local sensor should be considered. A mix of
strain and displacements sensor are common to track local responses.
7. Tilting and permanent displacement. To capture permanent displacements due to
damage, acceleration sensors should have a bandwidth that contains the DC com-
ponent. Integration of the acceleration signal should not filter the DC components
23 Civil Structural Testing 1321

Fig. 46 Normalized frequencies

[9]. In this case accelerometers should have a very low noise, typically below
1 × 10−5 g. Tilt-meters are able to capture permanent inclinations of the
structure, and they should be considered to obtain the tilt in two orthogonal
directions and in more than one location in height: at ground level to detect
permanent foundation rotations and upper floors to detect tilting due to structural
damage.

There is a strong dependency of the modal parameters due to amplitude [40, 97],
while damage can severely change the systems’ responses, as shown in Fig. 25.
Foundation conditions play an important role in the response, extending natural
periods or increasing or decreasing the energy dissipation due to soil structure
interaction [25].
Comparison of pre- and post-earthquake ambient vibration responses has been
used to identify permanent modal properties changes. If the ambient vibrations are
obtained immediately before and after the earthquake, the change due to environ-
mental variables should be small (with the exception of the possible evacuation of
the building). The change in periods can be an indication of the severity of the
state change. Changes in predominant period below 20% (without the presence of
localized damage) are related with nonstructural or light damage. Period changes
between 20% and 50% can be associated with moderate damage.
1322 R. Boroschek and J. P. Santos

The possibility of damage can be identified by the magnitude of the permanent

drifts, the presence of tilting, and the level of period extension. To track the possible
nonlinear response the STFT, S-Transform or regressive models [77] or windowed
MIMO techniques [88] can be used.
To estimate possible damage and to correlate with design variables the acceler-
ation and displacement at non-instrumented floors is sometimes necessary. Several
schemes have been proposed to obtain drifts along the building height, even at non-
instrumented floors. Typically linear or spline interpolation between floors or the
use of the knowledge of the mode shape to interpolate for each modal response is
used [50].
For the analysis of possible damage, displacements information are critical,
particularly the relative displacement between floors or drifts. The displacements
can be obtained by double integrating acceleration records. This has to be done with
care in order to avoid the introduction of spurious low noise components [7, 9].

10.3 Base-Isolated Buildings

Base-isolated buildings are one of the main solutions to avoid damage in highly
seismic regions. In this case the building is sitting on top of the isolators which
present a high vertical stiffness and extreme low stiffness in the horizontal direction.
These conditions produce a first modal shape that is similar to the seismic influence
vector (constant in height). The influence vector represents the acceleration at
all the degrees-of-freedom of the system assuming a rigid body motion of the
structure with respect to the points of earthquake input. When the first mode
has the same shape as the influence vector, the modal participation factor for
all other modes is close to zero so actually the isolation layer is like an energy
barrier.
The isolator can be of several types but the most common is natural rubber with or
without a lead plug to increase damping capacity or steel frictional pendulum. Both
isolators are highly nonlinear and do not present a viscous damping mechanism [75]
so typical system identification techniques have very limited applications.
In order to record the response in this type of buildings, sensors have to
have DC components. Permanent displacements are common after strong shak-
ing. Sensors described in the earthquake response sections are the same but a
concentration of acceleration and displacement sensors are needed at the interface
level.
Nonlinearity is always present in the response so nonparametric identification
tools like PP, spectrogram, S-Transform are appropriate [31, 59, 90]. Also MIMO
techniques used in sequential windows give information, with large uncertainties,
on the possible model parameters [2, 53, 61, 104].
23 Civil Structural Testing 1323

10.4 Bridges Under Operational and Ambient Excitation

10.4.1 The Suspended 25 de Abril Bridge in Lisbon

The suspended 25 de Abril bridge (Fig. 47a), located in Lisbon, Portugal, has a total
length of 2177 m, with a 1013 m long central suspended span, two 483 m lateral
suspended spans, and three non-suspended spans, each approximately 100 m long
(Fig. 47b). The bridge’s suspension system comprises four cables, where the two
main cables were installed during its construction before 1966 and the secondary
ones in 1999, when the railway was installed on the bridge. Hanging from the four
cables are 1344 hangers with rope-like section, divided in groups of 8 at each 168
suspended sections. Each of the two 190 m high pylons (Fig. 47d) consists of a
steel truss which supports the deck at an height of approximately 70 m, while piers
P2, P5, and P6 also support the deck at this height. Piers P1 and P7 connect the
bridge structure to the south tunnel and to the north viaduct and are composed of
concrete. These structures also include the anchorage of the main cables, while the
anchorage of the secondary ones are made with concrete structures built next to each
of these two piers. The bridge deck consists of a steel truss carrying 6 roadway lanes
and two railway lines (Fig. 1b), comprising four main beams which span between
the 200 transverse stiffening suspended frames, each of which is 11.20 m apart
from the adjacent ones and also consisting of steel trusses, as it can be observed in
Fig. 47c. Along with the main beams, diagonal elements span between each pair
of transverse suspended frames. The roadway system consists of seven stringers
supporting numerous transverse beams as well as the pavement, while the railway
is supported by four stringers, one per rail.

10.4.2 Structural Monitoring System and Data

The structural monitoring system installed on the 25 de Abril Bridge acquires data
from 214 measurements obtained from 171 standalone physical sensors at a rate of
500 samples/s. The quantities being measured consist (Fig. 48) of:

– Accelerations on five sections of the deck, on two sections of the pylons and
on the ground at the foundations of pier P1 and the pylons, using force-balance
accelerometers (Fig. 49b, c)
– Relative displacements between the piers P1, P2, P5, and P7 and the deck, using
magnetostrictive displacement transducers (Fig. 49e)
– Rotations of the nodes where the cables change their angle, located on the top of
piers P2 and P5, and on the pylons, using servo-inclinometers
– Stress in the mid-span of the deck’s suspended spans and in three sections of
each pylon, using Wheatstone bridges of resistance strain gages (Fig. 49d, f)
– Railway traffic on top of pier P1 in both railways, using rubber pads placed in
between the rail and the sleepers, and instrumented with a mesh of fiber optic
sensors
1324 R. Boroschek and J. P. Santos

Fig. 47 The 25 de Abril bridge: (a) overall view, (b) side view, (c) stiffening truss section, and
(d) tower side view

– Wind action on the mid-span and south quarter-span of the main central
suspended span, using three-dimensional ultrasonic anemometers (Fig. 49g)
– Temperature in the deck and pylon sections where stress is being measured, using
thermistors sharing the strain gage housing (Fig. 49d, f)

Data is being acquired from the sensors at a rate of 500 samples per second,
so as to take advantage of the hardware’s analog low pass filters, thus leading to
over 9000 million values per day. Data is then digitally filtered to a value of 20 Hz
and decimated to 50 samples per second per sensor. These samples are then kept
23 Civil Structural Testing 1325

Fig. 48 Structural monitoring system of the 25 de Abril bridge

as digital inventory of the bridge’s use and structural response and for Structural
Health Monitoring.
For the present example, the accelerometers installed on the five cross-sections
of the bridge deck and on the four sections of the pylons are considered. The
instrumented cross-sections consist of those shown in Fig. 48 and named as 0, 22S,
66S, 22N, 66N (“S” stands for South and “N” for North, taken from the mid central
span). Their locations consist of the center and quarters of the main span and as the
center of the lateral suspended spans. The accelerometers installed on the pylons are
located on their top and mid-height.
The linear character of the elements constituting long-span bridge structures,
with one dimension significantly larger than the other two, motivates the conversion
of the linear acceleration acquired by each sensor into section-wise accelerations
associated with the vertical, longitudinal, lateral (or transversal), and rotational
components of this physical quantity. This conversion can be conducted in bridge
dynamic testing and has been conducted in the present example, as it can be
observed in Fig. 50, which shows typical sections of the deck and pylons, where
the accelerations acquired from the sensors (in blue in Fig. 50) were converted
according to Eqs. (82)–(86), into the section-wise components (in green in the same
figure).

ah = at = a1t (82)
1326 R. Boroschek and J. P. Santos

Fig. 49 Components of the structural monitoring system of the 25 de Abril bridge: (a) enclo-
sure with acquisition, conditioning, and transmission equipment, (b) accelerometer housing,
(c) accelerometer, (d) strain gage and thermistor housing, (e) magnetostrictive displacement
transducer including housing, (f) Wheatstone bridge with thermistor, and (g) anemometer

av = (a2v + a3v) /2 (83)

al = (a2l + a3l) /2 (84)

ar = (a2v − a3v) /2 for the deck (85)

ar = (a2l − a3l) /2 for the pylons (86)

The singular value spectra of the vertical/longitudinal, transversal, and rotational

acceleration components acquired on the structure are shown in Fig. 51. These were
obtained by decomposing the periodogram matrix in singular values and vectors, as
described in Chapter 12 and in Sect. 5.5, and allow observing that a large number of
operational modes can be identified in the first and second singular values from 0 to
5 Hz. From 5 to 10 Hz, the noise generated by the road traffic increases significantly
and no clear peaks are generally observed in the first two singular values. In this
frequency subdomain, the last two singular values appear as those which still allow
observing some modes, even if not as clearly as in the subdomain 0–5 Hz. In fact,
23 Civil Structural Testing 1327

Fig. 50 Conversion of sensor acceleration into section-wise measurements: (a) on the deck and
(b) on the pylons

the operational action of the traffic generates not only the baseline noise vibration
shown in the SV spectra represented in Fig. 51, but it also changes the structural
system being measured by adding a significant amount of mass to it, as it can be
observed in Fig. 52, where the values of the RMS of the acceleration on the deck
and the frequencies of the first three vertical modes are shown. By observing this
plot, it can be concluded that the value of the RMS follows the amount of vehicles
crossing the bridge daily, and exhibits maxima around the morning and afternoon
rush hours, at the same time instants when the frequencies seem to exhibit minima.
These results hint about the importance of the mass traveling on the deck on the
identification of these lower order modes.
Other actions such as the railway traffic and wind are also known to be directly
influencing the structural system, not only through the excitation but also by
changing the properties of the materials and of the boundary conditions, as is
the case of the temperature, whose influence on the latter can be of particular
importance. Under these conditions, the estimation of the dynamic properties
through testing can be particularly challenging and the corresponding results can
exhibit very important variations observed over time.

10.4.3 Modal Estimation and Tracking

The need to conduct permanent identification in civil structures with sufficient
precision, for structural health monitoring purposes, demands that the modal
identification be conducted also continuously and therefore automatically. Under
these constraints, the identification methods as well as the strategies for post-
processing their output must be robust and precise, yet efficient. For the present
case study, where a high number of modes is observed over a small frequency
range, the coexistence of robustness and precision with computational efficiency can
be challenging and demands for computationally efficient variants of the strategies
described in Sects. 7 and 8.
1328 R. Boroschek and J. P. Santos

Fig. 51 Singular value spectra

When using a method such as the SSI-COV, described in Chapter 14 and in Sect.
5.7, very large values of model order, n, have to be used to detect high numbers
of modes, such as the number observed in Fig. 51, for a certain frequency domain
under analysis. This can be observed when trying to perform estimation using a
model order of 50 in 1 h of data acquisition on the 25 de Abril Bridge, which is
shown in the stability diagram of Fig. 53, where only ten modes can be observed in
that frequency domain (0–10 Hz). The model order needed to appropriately capture
all physical modes in this frequency domain, along with the spurious ones that
will necessarily be observed, requires prohibitive amounts of time and memory. An
efficient alternative which can be used consists of dividing the frequency domain
under analysis into subdomains which can be as small as needed to capture all
relevant physical modes. This strategy was used in the present example, where
23 Civil Structural Testing 1329

Fig. 52 RMS of acceleration on the deck and variation of the frequency values of the first three
vertical modes over one day. The three frequencies are not to scale and the changes shown in the
plot correspond to relative values of approximately 4% for the first two and 1.5% for the third

subdomains with a 1 Hz range were separately analyzed using the SSI-COV with a
maximum model order of 50. Each of the 1 Hz frequency slices was obtained using
the SSI-COV only after the application of a band-pass filter followed by decimation,
so as to ensure that the process is as efficient as possible. The corresponding stability
diagram is shown in Fig. 54, where a large number of modes can be identified in the
vertical alignments represented in the first 6 Hz. In accordance with the conclusion
obtained from the singular value spectra shown in Fig. 51, this stability diagram
does not show stable vertical alignments from 6 Hz to 10 Hz, thus suggesting that
poles obtained along this frequency domain are majorly related to noise, known to
be generated by road traffic.
Following the calculation of the poles for the frequency domain under analysis,
choice was made to perform spurious pole elimination based on hard criteria defined
on the values of f, d, and MAC, chosen as 0.01, 0.05, and 0.99, respectively. The
algorithm chosen for the stability verification consisted in checking whether, as each
level, the poles verify all three conditions in the level immediately below. The results
obtained are shown in Fig. 55 for the frequency range [0;5] Hz, whereas poles
with higher frequencies were not considered given their low stability. The major
motivation for performing the hard criteria verification prior to applying cluster
analysis consisted in computational simplicity. Lower number of poles used as input
in clustering proved important in reducing computation times, not only due to the
smaller data sets but also due to easier convergence.
Using the poles considered as stable, these were considered for cluster analysis
and validation. The SIL index was obtained for cluster partitions comprising from
two to half of the maximum SSI order, taken as the theoretical number of modes that
can be detected. Choice was made to select the cluster partition with the higher SIL
value instead of the first relative maximum, so as to increase the detection incidence
of each mode. The application of the hierarchical agglomerative clustering using the
Ward criterion along with the SIL index allowed obtaining, for the data set acquired
during 1 h and analyzed herein, the cluster representatives, taken as de centroids,
shown in Fig. 56, where a large density of modes identified in the frequency range
[0,5] Hz can be observed.
1330 R. Boroschek and J. P. Santos

Fig. 53 Stability diagram obtained from applying the SSI to the frequency range [0;10] Hz

The time-series of cluster representatives obtained over a period of 8 months in

2018 is shown in Fig. 57, whose observation confirms the high density of modes
in the range [0;5] Hz. During the period under analysis, a large number of clearly
defined near-horizontal alignments over time can be observed, thus revealing modes
that are repeatedly identified over time with small variability. However, large dense
bands with high variability can also be observed, in which modes are not as easily
distinguished.
No significant changes can be observed in the frequency values between the
colder months (February and March) and the warmer ones (July and August),
yet important changes are observed daily in the majority of the near-horizontal
alignments, as exemplified in Fig. 52, and there are even several modes which
are only detected when certain operational conditions are met, such as those with
frequencies of approximately 1.5 and 3.0 Hz, which are only detected during week
23 Civil Structural Testing 1331

Fig. 54 Stability diagram obtained from applying the SSI to 10 frequency slices of 1 Hz

days, as it can be observed in Fig. 57. This apparent inexistence of correlation with
temperature along with a strong correlation with traffic loading is not the general
case in experimental bridge testing and is due to the large span and slender character
of the 25 de Abril Suspended Bridge. In general, changes of approximately 5% can
be observed in modal frequencies due to temperature in smaller spans, in which
the influence of road traffic is also smaller. Railway bridges can be considered as
an exception, regarding traffic loading, since trains crossing the bridge change the
vibration regime. In these cases, one possibility consists of removing the time-series
obtained during the train crossing, merging the remaining subsets of the signal using
appropriate windowing functions, and in analyzing them using methods such as
the SSI-COV. An alternative strategy consists of analyzing the decaying response
obtained immediately after the train crossing using the ERA methods (Chapter 12
and Sect. 5.3).
1332 R. Boroschek and J. P. Santos

Fig. 55 Stable poles obtained using hard criteria applied to f, d, and MAC

Fig. 56 Mode representatives obtained using hierarchical clustering and the SIL index and shown
in: (a) stability diagram and (b) f vs. d cluster diagram
23 Civil Structural Testing 1333

Fig. 57 Mode representatives obtained over 8 months of 2018

The modal tracking strategy used to relate the representatives of each mode
obtained in different periods of monitoring time of the 25 de Abril Bridge relied
also on clustering methods and on the SIL index to define a baseline reference
in an unsupervised way, that is, without human intervention regarding the number
of modes and their feature values. A 2-week period was selected for defining the
baseline, to which the strategy described in Sect. 8.3 was applied separately for
each band of 0.5 Hz. The distance matrix chosen as input was the one obtained
from the frequencies and the MAC only. A total of 76 distinct modes were obtained
(characterized through their clusters’ centroids) and used as baseline reference
to which each of the following mode representatives, obtained at posterior time
instants, was compared. The allocation of each new mode into the baseline ones
was conducted by calculating the distance matrix (with the same metric, based on
1334 R. Boroschek and J. P. Santos

Fig. 58 Mode representatives obtained over 8 months of 2018 and colored according to the mode
identification

f and MAC) between the baseline and the newly arrived modes. The newly arrived
modes were then allocated into the baseline ones based on the smaller distance value
found in this non-symmetric distance matrix.
The time-series of frequencies tracked for each of the 76 modes can be found
in Fig. 58, where it can be readily observed that several modes exhibit smaller
variability and/or higher detection incidence, thus rendering them appropriate
for structural health monitoring. It can also be observed, especially for higher
frequencies, that some modes are associated with noise and other effects, due to their
visible variability. This disadvantage is a consequence of choosing the maximum
SIL value as criterion and of choosing the maximum number of clusters equal to half
the maximum SSI order. These choices resulted, however, in higher mode detection
incidences over time.
23 Civil Structural Testing 1335

11 Concluding Remarks

Testing of civil engineering structures is an old field but still exhibits limited
applications. The benefits of testing, and in particular of structural health monitoring
(SHM), are evident for the safety of the user and for a rational maintenance of the
structural systems. A state of development has been reached where the hardware
is cost effective, the software is available, and broadband communications are
available with a wide coverage, robustness, and low costs.
The importance of monitoring and particularly of SHM for the recovery of
infrastructure after extreme events like hurricanes and earthquakes, where the
damage systems could easily outnumber the number of practitioners by a factor of
10, has already been seen in recent cases. Monitoring hardly accessible structures
due to its remote location, temperature of operation, or just the difficulty and
complexity of the inspection are clear benefits of the presented techniques.
To increment its use, several steps should be performed and, among these, the
education of undergraduate and graduate civil engineering students on the basis
of experimental techniques, and its applications to laboratory and real structures,
naturally assumes particular importance. The development of academic and business
public benchmark cases for the validation and promotion of the systems should also
be undertaken so as to promote normalization of techniques and procedures. Finally,
the development of guidelines must also be considered, not only for the users to
clarify what can be gained, what is possible at this stage of developments, and what
are the typical cost and reliable systems to be applied but also for the practitioners
to standardize products, reduce cost, and validate results.

Acknowledgments The authors would like to acknowledge all the support given by the Depart-
ment of Civil Engineering, of the University of Chile, and by the Structures Department of the
Portuguese National Laboratory for Civil Engineering (LNEC).
In addition, the authors would also like to acknowledge the support given by Conicyt and
Fondecyt Chile, projects 1950629, 1000912, 1070319, and the Portuguese Foundation for Science
and Technology (FCT) through the SAFESUSPENSE project – POCI-01-0145-FEDER-031054
(funded by COMPETE2020, POR Lisboa and FCT).

References
1. Allemang R (2003) The modal assurance criterion – twenty years of use and abuse. Sound
and Vibration 37(8):14–21. https://doi.org/10.1016/j.chemgeo.2006.02.014
2. Astroza R, Gutiérrez G, Repenning C, Hernández F (2018) Time-variant modal parameters
and response behavior of a base-isolated building tested on a shake table. Earthquake Spectra
34(1):121–143. https://doi.org/10.1193/032817EQS054M
3. Beck JL (1978) Determining models of structures from earthquake records. EERL report
78-01
4. Bernal D, Döhler M, Kojidi SM, Kwan K, Liu Y (2015) First mode damping ratios for
buildings. Earthquake Spectra 31(1):367–381. https://doi.org/10.1193/101812EQS311M
5. Bishop CM (2006) Pattern recognition and machine learning. pattern recognition. Springer,
Berkeley
1336 R. Boroschek and J. P. Santos

6. Bixio AR, Dolce M, Nigro D, Ponzo FC, Braga F, Nicoletti M (2001) Repeatable dynamic
release tests on a base-isolated building. Journal of Earthquake Engineering 5(3):369–393.
https://doi.org/10.1080/13632460109350398
7. Boore DM, Sisi AA, Akkar S (2012) Using pad-stripped acausally filtered strong-motion
data. Bulletin of the Seismological Society of America 102(2):751–760. https://doi.org/10.
1785/0120110222
8. Boroschek RL, Bilbao JA (2019) Interpretation of stabilization diagrams using density-
based clustering algorithm. Engineering Structures 178:245–257. https://doi.org/10.1016/J.
ENGSTRUCT.2018.09.091
9. Boroschek RL, Legrand D (2006) Tilt motion effects on the double-time integration of linear
accelerometers: an experimental approach. Bulletin of the Seismological Society of America
96(6). https://doi.org/10.1785/0120050167
10. Boroschek RL, Baesler H, Vega C (2011) Experimental evaluation of the dynamic properties
of a wharf structure. Engineering Structures 33(2). https://doi.org/10.1016/j.engstruct.2010.
10.014
11. Brincker R, Ventura CE (2015) Introduction to operational modal analysis. Wiley, Chichester
12. Brincker R, Ventura C, Andersen P (2001a) Damping estimation by frequency domain
decomposition. In: Proceedings of IMAC 19: a conference on structural dynamics,
pp 1–6
13. Brincker R, Zhang L, Andersen P (2001b) Modal identification of output-only systems using
frequency domain decomposition. Smart Materials and Structures 10(3):441–445. https://doi.
org/10.1088/0964-1726/10/3/303
14. Brownjohn JMW (2003) Ambient vibration studies for system identification of tall buildings.
Earthquake Engineering and Structural Dynamics 32(1):71–95. https://doi.org/10.1002/eqe.
215
15. Brownjohn JMW, Magalhaes F, Caetano E, Cunha A (2010) Ambient vibration re-testing and
operational modal analysis of the Humber Bridge. Engineering Structures 32(8):2003–2018.
https://doi.org/10.1016/j.engstruct.2010.02.034
16. Cabboi A, Magalhães F, Gentile C, Cunha A (2011) Automated modal identification and
tracking: application to an iron arch bridge. Structural Control and Health Monitoring
19(1):88–106. https://doi.org/10.1002/stc
17. Cabboi A, Magalhães F, Gentile C, Cunha Á (2017) Automated modal identification and
tracking: application to an iron arch bridge. Structural Control and Health Monitoring
24(1):e1854. https://doi.org/10.1002/stc.1854
18. Caicedo JM, Marulanda J (2011) Fast mode identification technique for online monitoring.
Structural Control and Health Monitoring 18(4):416–429. https://doi.org/10.1002/stc.381
19. Carreño RP, Boroschek RL (2011) Modal parameter variations due to earthquakes of different
intensities. In: Conference proceedings of the Society for Experimental Mechanics Series,
vol 4
20. Clinton JF, Bradford SC, Heaton TH, Favela J (2006) The observed wander of the natural
frequencies in a structure. Bulletin of the Seismological Society of America 96(1):237–257.
https://doi.org/10.1785/0120050052
21. Cohen L (1995) Time-frequency analysis. Prentice Hall PTR
22. Crawford R, Ward HS (1964) Determination of the natural periods of buildings. Bulletin of
the Seismological Society of America 54(6A):1743–1756
23. Cremona C, Santos JP (2018) Structural health monitoring as a big data problem. Structural
Engineering Internacio:1–18. https://doi.org/10.1080/10168664.2018.1461536
24. Cruz C, Miranda E (2017a) Evaluation of damping ratios for the seismic analysis of
tall buildings. Journal of Structural Engineering 143(1):04016144. https://doi.org/10.1061/
(ASCE)ST.1943-541X.0001628
25. Cruz C, Miranda E (2017b) Evaluation of soil-structure interaction effects on the damping
ratios of buildings subjected to earthquakes. Soil Dynamics and Earthquake Engineering
100(May):183–195. https://doi.org/10.1016/j.soildyn.2017.05.034
26. Cunha A, Caetano E (2006) Experimental modal analysis of civil engineering structures.
Sound and Vibration June:12–20. https://doi.org/10.1007/978-1-4020-6239-1_299
23 Civil Structural Testing 1337

27. Cunha A, Caetano E, Delgado R (2001) Dynamic tests on large cable-stayed bridge. Journal
of Bridge Engineering 6(1):54–62. https://doi.org/10.1061/(ASCE)1084-0702(2001)6:1(54)
28. Cury A, Crémona C, Diday E (2010) Application of symbolic data analysis for structural
modification assessment. Engineering Structures 32(3):762–775
29. Döhler M, Mevel L (2012) Fast multi-order computation of system matrices in subspace-
based system identification. Control Engineering Practice 20(9):882–894. https://doi.org/10.
1016/j.conengprac.2012.05.005
30. Eyre R, Tilly GP (1977) Damping measurements on steel and composite bridges – abstract.
In: Symposium on dynamic behaviour of bridges (abstract), pp 22–39
31. Furukawa T, Ito M, Izawa K, Noori MN (2005) System identification of base-isolated building
using seismic response data. Journal of Engineering Mechanics 131(3):268–275. https://doi.
org/10.1061/(ASCE)0733-9399(2005)131:3(268)
32. Gao W, Zhou Y (2003) Eigenvalue-based algorithms for testing positive realness of SISO
systems. IEEE Transactions on Automatic Control 48(11):2051–2055. https://doi.org/10.
1109/TAC.2003.819302
33. Giraldo DF, Song W, Dyke SJ, Caicedo JM (2009) Modal identification through ambient
vibration: comparative study. Journal Of Engineering Mechanics-Asce 135(8):759–770.
https://doi.org/10.1061/(asce)0733-9399(2009)135:8(759)
34. Gonzalez I (2013) Stochastic model updating of ballast stiffness in cold conditions. In: The
6th International conference on structural health monitoring of intelligent infrastructure
35. Gonzalez W, Boroschek RL (2019) Modal tracking on a building with a reduced number of
sensors system. In: International modal analysis conference
36. González L, Boroschek R, Verdugo G (2009) Experimental study to correlate change in
the structural frequency and foundation soil water content. In: Proceedings of the 17th
international conference on soil mechanics and geotechnical engineering: the academia and
practice of geotechnical engineering, vol 3. https://doi.org/10.3233/978-1-60750-031-5-2040
37. Guéguen P, Langlais M, Garambois S, Voisin C, Douste-Bacqué I (2017) How sensitive are
site effects and building response to extreme cold temperature? The case of the Grenoble’s
(France) City Hall building. Bulletin of Earthquake Engineering 15(3):889–906. https://doi.
org/10.1007/s10518-016-9995-3
38. Guillier B, Chatelain JL, Perfettini H, Oubaiche EH, Voisin C, Bensalem R et al (2016)
Building frequency fluctuations from continuous monitoring of ambient vibrations and their
relationship to temperature variations. Bulletin of Earthquake Engineering 14(8):2213–2227.
https://doi.org/10.1007/s10518-016-9901-z
39. Hastie T (2011) The elements of statistical learning, data mining, inference and prediction,
2nd edn. Springer, New York
40. Haviland RW (1976) A study of the uncertainties in the fundamental translational periods and
damping values for real buildings. Res. Rep. No. 5, Pub. No. R76-12, Department of Civil
Engineering, MIT, Cambridge, p 115
41. Hisada T, Nakagawa K (1956) Vibration tests of various types of building structures up to
failure. In: 1st World conference of earthquake engineering (WCEE)
42. Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q et al (1998) The empirical mode
decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis.
Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences
454(1971):903–995. https://doi.org/10.1098/rspa.1998.0193
43. Hudson DE (1964) Resonance testing of full-scale structures. Journal of the Engineering
Mechanics Division 90:em3
44. Imregun M, Ewins DJ (1995) Complex modes – origins and limits. In: 13th International
modal analysis conference, pp 496–506
45. Iwan WD, Cifuentes A (1986) A model for system identification of degrading structures.
Eathquake Engineering and Structural Dynamics 14(December 1985):877–890
46. James III, G. H., Carne, T. G., Lauffer, P., J., Lauffer, J. P. (1995). The Natural Excitation
Technique (NExT) for modal parameter extraction from operating wind turbines. The
International Journal of Analytical and Experimental Modal Analysis, 10(4), 260–277.
SAND92-1666
1338 R. Boroschek and J. P. Santos

47. Jeary AP (1986) Damping in tall buildings – a mechanism and a predictor. Earthquake
Engineering & Structural Dynamics 14(5):733–750. https://doi.org/10.1002/eqe.4290140505
48. Juang J-N, Pappa RS (1985) An eigensystem realization algorithm for modal parameter
identification and model reduction. Journal of Guidance, Control, and Dynamics 8(5):620–
627. https://doi.org/10.2514/3.20031
49. Kaya Y, Safak E (2015) Real-time analysis and interpretation of continuous data from
structural health monitoring (SHM) systems. Bulletin of Earthquake Engineering 13(3):917–
934. https://doi.org/10.1007/s10518-014-9642-9
50. Kaya Y, Kocakaplan S, Şafak E (2015) System identification and model calibration of multi-
story buildings through estimation of vibration time histories at non-instrumented floors.
Bulletin of Earthquake Engineering 13(11):3301–3323. https://doi.org/10.1007/s10518-015-
9774-6
51. Keightley WO, Housner GW, Hudson DE (1961) Vibration tests of the Encino Dam intake
tower. Technical report: Caltech EERL.1961.003. California Institute of Technology
52. Li QS, Yang K, Wong CK, Jeary AP (2003) The effect of amplitude-dependent damping on
wind-induced vibrations of a super tall building. Journal of Wind Engineering and Industrial
Aerodynamics 91(9):1175–1198. https://doi.org/10.1016/S0167-6105(03)00080-1
53. Loh C-H, Weng J-H, Chen C-H, Lu K-C (2013) System identification of mid-story isolation
building using both ambient and earthquake response data. Structural Control and Health
Monitoring 20(2):139–155. https://doi.org/10.1002/stc.479
54. Magalhães F (2010) Operational modal analysis for testing and monitoring of bridges and
special structures. University of Porto, Faculty of Engineering
55. Magalhães F, Cunha Á, Caetano E (2009) Online automatic identification of the modal
parameters of a long span arch bridge. Mechanical Systems and Signal Processing 23(2):316–
329. https://doi.org/10.1016/j.ymssp.2008.05.003
56. Magalhães F, Cunha Á, Caetano E, Brincker R (2010) Damping estimation using free decays
and ambient vibration tests. Mechanical Systems and Signal Processing 24(5):1274–1290.
https://doi.org/10.1016/j.ymssp.2009.02.011
57. Mendes P (2010) Observação e análise do comportamento dinâmico de barragens de betão.
Faculdade de Engenharia da Universidade do Porto
58. Milligan G, Cooper M (1985) An Examination of Procedures for Determining the Number of
Clusters in a Data Set. Psychometrika 50(2):159–179
59. Moroni MO, Sarrazin M, Boroschek R (1998) Experiments on a base-isolated building in San-
tiago, Chile. Engineering Structures 20(8). https://doi.org/10.1016/S0141-0296(97)00086-2
60. Murray TM, Allen DE, Ungar EE (2003) Floor vibrations due to human activity. Steel Design
Guide Series 11, D811(10M797)
61. Nagarajaiah S, Xiaohong S (2000) Response of base-isolated USC hospital building in
Northridge earthquake. Journal of Structural Engineering 126(10):1177–1186. https://doi.org/
10.1061/(ASCE)0733-9445(2000)126:10(1177)
62. Peeters B, De Roeck G (1999) Reference-based stochastic subspace identification for output-
only modal analysis. Mechanical Systems and Signal Processing 13(6):855–878. https://doi.
org/10.1006/mssp.1999.1249
63. Peeters B, De Roeck G (2001) Stochastic system identification for operational modal analysis:
a review. Journal of Dynamic Systems, Measurement, and Control 123(4):659. https://doi.org/
10.1115/1.1410370
64. Peeters B, Ventura CE (2003) Comparative study of modal analysis techniques for bridge
dynamic characteristics. Mechanical Systems and Signal Processing 17(5):965–988. https://
doi.org/10.1006/mssp.2002.1568
65. Peeters B, Maeck J, De Roeck G (2000) Dynamic monitoring of the Z24-Bridge: separating
temperature effects from damage. In: Proceeding of the European COST F3 conference on
system identification structural health monitoring, November 2015, pp 377–386
66. Peeters B, Van der Auweraer H, Guillaume P, Leuridan J (2004) The PolyMAX frequency-
domain method: a new standard for modal parameter estimation? Shock and Vibration
11(3–4):395–409. https://doi.org/10.1155/2004/523692
23 Civil Structural Testing 1339

67. Pereira S, Magalhães F, Gomes JP, Cunha Á, Lemos JV (2018) Dynamic monitoring
of a concrete arch dam during the first filling of the reservoir. Engineering Structures
174(July):548–560. https://doi.org/10.1016/j.engstruct.2018.07.076
68. Rainieri C, Fabbrocino G (2014) Operational modal analysis of civil engineering structures.
Springer, New York. https://doi.org/10.1007/978-1-4939-0767-0
69. Rendón E, Abundez I, Arizmendi A, Quiroz EM (2011) Internal versus external clus-
ter validation indexes. International Journal of Computers and Communications 5(1):
158–163
70. Reynders E (2012) System identification methods for (operational) modal analysis: review
and comparison. Archives of Computational Methods in Engineering 19(1):51–124. https://
doi.org/10.1007/s11831-012-9069-x
71. Reynders E, Houbrechts J, De Roeck G (2012) Fully automated (operational) modal analysis.
Mechanical Systems and Signal Processing 29:228–250
72. Rodrigues J (2004) Identificação Modal Estocástica Métodos de análise e aplicações em
estruturas de engenharia civil. University of Porto, Faculty of Engineering
73. Rodrigues J, Brincker R, Andersen P (2004) Improvement of frequency domain output-only
modal identification from the application of the random decrement technique. In: Proceedings
of the 23rd international modal analysis conference
74. Rousseeuw P (1987) Silhouettes: a graphical aid to the interpretation and validation of cluster
analysis. Journal of Computational and Applied Mathematics 20(1):53–65
75. Ryan KL, Polanco J (2008) Problems with Rayleigh damping in base-isolated buildings.
Journal of Structural Engineering 134(11):1780–1784. https://doi.org/10.1061/(ASCE)0733-
9445(2008)134:11(1780)
76. Safak E (1989) Adaptive modeling, identification, and control of dynamic structural systems.
II: applications. Journal of Engineering Mechanics 115(11):2406–2426. https://doi.org/10.
1061/(ASCE)0733-9399(1989)115:11(2406)
77. Şafak E (1989) Adaptive modeling, identification, and control of dynamic structural systems.
I: theory. Journal of Engineering Mechanics 115(11):2386–2405. https://doi.org/10.1061/
(ASCE)0733-9399(1989)115:11(2386)
78. Şafak E (1991) Identification of linear structures using discrete-time filters. Journal of Struc-
tural Engineering 117(10):3064–3085. https://doi.org/10.1061/(ASCE)0733-9445(1991)117:
10(3064)
79. Santos JP (2014) Smart structural health monitoring techniques for novelty identification
in civil engineering structures. PhD thesis. Instituto Superior Técnico – University of
Lisbon
80. Santos J, Crémona C, Orcesi AD, Silveira P (2013a) Multivariate statistical analysis for early
damage detection. Engineering Structures 56:273–285
81. Santos J, Crémona C, Orcesi A, Silveira P (2013b) Baseline-free real-time novelty detection
using vibration-based symbolic features. In: EVACES’13, Ouro Preto, pp 1–8
82. Santos J, Orcesi AD, Crémona C, Silveira P (2015) Baseline-free real-time assessment of
structural changes. Structure and Infrastructure Engineering: Maintenance, Management,
Life-Cycle Design and Performance 11(2):145–161. https://doi.org/10.1080/15732479.2013.
858169
83. Santos JP, Cremona C, Silveira P, Calado L (2016a) Real-time damage detection based on
pattern recognition. Structural Concrete. https://doi.org/10.1002/suco.201500092
84. Santos J, Mata J, Cremona C, Silveira P (2016b) Detection of structural changes using
symbolic representations of stochastic subspace identification estimates. In: 8th Workshop
on structural health monitoring, pp 1–10
85. Santos JP, Crémona C, Silveira P (2018) Automatic modal operational analysis of a long-span
suspended bridge using pattern recognition. In: 40th IABSE symposium, 19–21 September
2018, Nantes, France. Tomorrow’s megastructures. IABSE, p 8
86. Shi W, Shan J, Lu X (2012) Modal identification of Shanghai World Financial Center both
from free and ambient vibration response. Engineering Structures 36:14–26. https://doi.org/
10.1016/j.engstruct.2011.11.025
1340 R. Boroschek and J. P. Santos

87. Silva R, Rodrigues J (2011) Influência da altura de água nas frequências de vibração de pontes
com pilares submersos. Lisbon
88. Siringoringo DM, Fujino Y (2018) Seismic response of a suspension bridge: insights from
long-term full-scale seismic monitoring system. Structural Control and Health Monitoring
June:e2252. https://doi.org/10.1002/stc.2252
89. Spence SMJ, Kareem A (2014) Tall buildings and damping: a concept-based data-
driven model. Journal of Structural Engineering 140(5):04014005. https://doi.org/10.1061/
(ASCE)ST.1943-541X.0000890
90. Stewart JP, Conte JP, Aiken ID (1999) Observed behavior of seismically isolated buildings.
Journal of Structural Engineering 125(9):955–964. https://doi.org/10.1061/(ASCE)0733-
9445(1999)125:9(955)
91. Stockwell RG, Mansinha L, Lowe RP (1996) Localization of the complex spectrum: the S
transform. IEEE Transactions on Signal Processing 44(4):998–1001. https://doi.org/10.1109/
78.492555
92. Su J, Xia Y, Ni Y, Zhou L, Su C (2017) Field monitoring and numerical simulation
of the thermal actions of a supertall structure. Structural Control and Health Monitoring
24(4):e1900. https://doi.org/10.1002/stc.1900
93. Takeuchi M (1960) Vibrational characteristics of buildings, part I – vibrational characteristics
of actual buildings determined by actual tests. In: Proceedings of the 2nd world conference
on earthquake engineering, Tokyo
94. Tamura Y, Yoshida A (2008) Amplitude dependency of damping in buildings. In: Structures
congress 2008. American Society of Civil Engineers, Reston, pp 1–10. https://doi.org/10.
1061/41000(315)39
95. Tamura Y, Suda K, Sasaki A (2000) Damping in buildings for wind resistant design. In:
Proceedings of the international symposium on wind and structures for the 21st century, pp
115–130
96. Theodoridis S, Koutroumbas K (2009) Pattern Recognition, 4th edn. Elsevier, London
97. Uebayashi H, Nagano M, Hida T, Tanuma T, Yasui M, Sakai S (2016) Evaluation of the struc-
tural damage of high-rise reinforced concrete buildings using ambient vibrations recorded
before and after damage. Earthquake Engineering & Structural Dynamics 45(2):213–228.
https://doi.org/10.1002/eqe.2624
98. Van Overschee P, De Moor B, Sista E (1994) N4SID: subspace algorithms for the identifica-
tion of combined deterministic-stochastic systems. Automatica 30:75–93. https://doi.org/10.
1016/0005-1098(94)90230-5
99. Van Overschee P, De Moor B, Van Overschee P, De Moor B (1996) Subspace identification
for linear system: theory – implementation – applications. In: Conference proceedings of the
international conference of IEEE Engineering in Medicine and Biology Society, 2008, pp
4427–4430. https://doi.org/10.1109/IEMBS.2008.4650193
100. Verhaegen M (1994) Identification of the deterministic part of MIMO state space models
given in innovations form from input-output data. Automatica 30(1):61–74. https://doi.org/
10.1016/0005-1098(94)90229-1
101. Ward HS, Crawford R (1966) Wind-induced vibrations and building modes. Bulletin of the
Seismological Society of America 56(4):793–813
102. Yan A, Kerschen G, De Boe P, Golinval J-C (2005a) Structural damage diagnosis under
varying environmental conditions – part II: local PCA for non-linear cases. Mechanical
Systems and Signal Processing 19(4):865–880
103. Yan A, Kerschen G, De Boe P, Golinval J-C (2005b) Structural damage diagnosis under
varying environmental conditions – part I: a linear analysis. Mechanical Systems and Signal
Processing 19(4):865–880
104. Yoshimoto R, Mita A, Okada K (2005) Damage detection of base-isolated buildings using
multi-input multi-output subspace identification. Earthquake Engineering and Structural
Dynamics 34(3):307–324. https://doi.org/10.1002/eqe.435
105. Yousefianmoghadam S, Behmanesh I, Stavridis A, Moaveni B, Nozari A, Sacco A (2018)
System identification and modeling of a dynamically tested and gradually damaged 10-story
23 Civil Structural Testing 1341

reinforced concrete building. Earthquake Engineering and Structural Dynamics 47(1):25–47.

https://doi.org/10.1002/eqe.2935
106. Zembaty Z, Kowalski M, Pospisil S (2006) Dynamic identification of a reinforced concrete
frame in progressive states of damage. Engineering Structures 28(5):668–681. https://doi.org/
10.1016/j.engstruct.2005.09.025
107. Zhou HF, Ni YQ, Ko JM (2011) Eliminating temperature effect in vibration-based structural
damage detection. Journal of Engineering Mechanics 137(12):785–796. https://doi.org/10.
1061/(ASCE)EM.1943-7889.0000273
108. Zonno G, Aguilar R, Castañ B, Lourenç PB (2018) Preliminary validation of an automatic
modal identification methodology for structural health monitoring of historical buildings.
International Journal of Structural and Civil Engineering Research 7(2):144–150. https://doi.
org/10.18178/ijscer.7.2.144-150
Aerospace Perspective for Modeling
and Validation 24
Robert Coppolino

Contents
1 Theoretical Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1344
1.1 Categories of Dynamic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1344
1.2 Variational Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1346
1.3 The Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1346
2 Structural Dynamic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1348
2.1 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1348
2.2 Dynamic Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1350
2.3 Effective Modeling Guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1351
2.4 Further Thoughts on Structural Dynamic Modeling . . . . . . . . . . . . . . . . . . . . . . . . 1352
3 Matrix Structural Dynamic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1353
3.1 Guyan Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1353
3.2 The Hurty-Craig-Bampton Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1354
3.3 The Benfield-Hruda Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1355
3.4 The MacNeal-Rubin Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1356
3.5 Application of Hurty-Craig-Bampton and MacNeal-Rubin Methodology . . . . . . . 1359
3.6 Detailed Structural Dynamic Loads and the Mode Acceleration Method . . . . . . . 1361
4 Verifiction and Validation of Structural Dynamic Models . . . . . . . . . . . . . . . . . . . . . . . . 1362
4.1 System Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1363
4.2 Modal Test Planning and the Test Analysis Model (TAM) . . . . . . . . . . . . . . . . . . . 1364
4.3 Target Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1366
4.4 Automated Response DOF Selection for Mapping
of Experimental Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1366
4.5 Measured Data Acquisition, Data Analysis, Experimental Modal Analysis . . . . . 1368
4.6 Modal Test-Analysis Correlation and U.S. Government Standards . . . . . . . . . . . . 1369
4.7 Overview of Efficient Structural Dynamic Sensitivity Analysis . . . . . . . . . . . . . . . 1371
4.8 Residual Mode Augmentation (RMA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1372
5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1376
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1377

R. Coppolino ()
Measurement Analysis Corporation, Torrance, CA, USA

© The Society for Experimental Mechanics 2022 1343

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_30
1344 R. Coppolino

Abstract

The present chapter on the Aerospace Perspective for Modeling and Validation
offers an overview of generally accepted practices in the U.S. aerospace com-
munity, with emphasis on space launch systems and their spacecraft payloads.
Inclusion of practices in the important U.S. aircraft community requires a more
extensive document. In keeping the present chapter within the confines of
“accepted practices,” valuable subject matter related to scientific and technical
innovations that are not yet generally accepted or peer “vetted” is considered
beyond the scope of the present exposition. This chapter includes two distinct
areas of subject matter, namely (1) modeling and analysis practices employed
in prediction of structural dynamic behavior, and (2) integrated test-analysis
practices employed in model verification, validation, and updating.

Keywords

US · Aerospace · Structural · Dynamic · Model · Validation

1 Theoretical Foundations

Review of theoretical foundations of structural dynamics is an essential preliminary

step for subsequent discussions on Structural Dynamic Models (Sect. 2), and
Matrix Structural Dynamic Analysis (Sect. 3). Both of these topics are products
of successive (a) theoretical underpinnings from categorization (basic assumptions
and approximations), (b) variational principles, and (c) the finite element method.

1.1 Categories of Dynamic Systems

Harry Himmelblau, a long-time colleague in the Southern California aerospace

community, once stated that “The universe is composed of bananas and everything
that’s not a banana. Linear systems are the bananas, nonlinear systems are every-
thing else.” The single degree of freedom system (SDOF) model, shown in Fig. 1,
serves to describe the fundamental classifications of “bananas” and “everything
else.”
Dynamic equilibrium of the SFOF system is described by

   
MÜ = FL U, U̇, pL + FN U, U̇, pN + FE (t), (1)

where the “linear” restoring force, FL , is typically described in terms of two linear
system parameters, pL = (K, B), as

 
FL = − BU̇ + KU , (2)
24 Aerospace Perspective for Modeling and Validation 1345

Fig. 1 SDOF system

schematic

and the “nonlinear” restoring force is described in one of two general forms, namely,
(a) an algebraic nonlinear function of displacement, velocity, and fixed parameters,
pN , or more generally, (b) a hysteretic nonlinear function of displacement, velocity,
and evolving parameters, pN . A typical algebraic nonlinear restoring force (cubic
stiffness and “fluid” drag type dissipation) is described by,

   
FN = − B2 U̇ U̇ + K3 U3 . (3)

The more general hysteretic nonlinearity is described by the evolving functional

relationship,

 
(FN (t + t), {pN (t + t)}) = −N U (t), U̇ (t), {pN (t)} . (4)

The above functional relationship describes the behavior of materials undergoing

elastic-plastic deformation and structural joints that experience stick-slip frictional
deformation, as well as all things not falling into the “linear” and “algebraic
nonlinear” categories.
Based on the above discussion, it appears that Himmelblau’s viewpoint of the
universe should be modified as follows:

(a) “Bananas” are the “linear” systems described by Eq. 2.

(b) “Plantains” are the “algebraic” nonlinear systems described by Eq. 3.
(c) “All Else” are the “hysteretic,” evolving nonlinear systems described by Eq. 4.

Thus Himmelblau’s modified maxim is “The universe is composed of bananas,

plantains and everything else.” Fortunately, the majority of subject matter discussed
in this chapter on the aerospace perspective for modeling and validation relates to
“bananas.” That being said, several very important applications falling into the other
two categories will be cited herein.
1346 R. Coppolino

1.2 Variational Principles

The entirety of the aerospace perspective for modeling and validation, regardless of
“banana, plantain, or all else” categorization of the subject dynamic system, depends
on variational principles. Specifically the foundational variational principles [1], in
hierarchical sequence include:

1. Hamilton’s Principle [2]

2. Lagrange’s Equations [3]
3. The Ritz Method [4]
4. Toupin’s Variational Principle [5].

Innovations due to Trefftz [6] and Galerkin [7], while generally important
in twentieth-century developments in the field of structural mechanics, are of
secondary significance to the above four variational viewpoints. In particular,
Hamilton’s Principle, Lagrange’s Equations, and The Ritz Method form the overall
foundation of partial differential equations of mathematical physics, finite element
analysis, and matrix structural analysis. Toupin’s Variational Principle provides a
systematic foundation for treatment of hydroelastic dynamic systems, especially
within the NASTRAN software environment [8].

1.3 The Finite Element Method

The finite element method [9], as it is widely employed in the aerospace community,
is a building block approach that models structural components, represented in
Fig. 2 as a classical “potato shaped” form, in terms of distributed displacement shape
functions.
In matrix form, the “interior” and “boundary” displacements, Ui and Ub ,
respectively, relate functionally to the displacement field of the entire element, U, as

Np
 
Nh
u (x, y, z, t) = p,n (x, y, z) · qp,n (t) + h,n (x, y, z) · uh,n (t), (5)
n=1 n=1

Fig. 2 Finite element

24 Aerospace Perspective for Modeling and Validation 1347

where the “ h ” shape functions are associated with discrete “boundary” displace-
ments, and the “ p ” shape functions and associated generalized displacements,
“qp ”, relate to interior displacements (with boundary displacements set to “zero”).
In matrix form, the element displacement field relates to the shape functions and
displacement DOF as,

 qp
{u} = p h (6)
uh

Employing Hamilton’s Principle and the Ritz Method (details in Reference [10]),
the resulting matrix equations describing linear structural dynamic behavior of the
finite element are of the form,

Mpp Mph q̈p Bpp Bph q̇p Kpp Kph qp Qp

+ + = , (7)
Mhp Mhh üh Bhp Bhh u̇h Khp Khh uh Fh

where the symmetric, positive semi-definite coefficient matrices relate to the

element’s mass, viscous damping, and stiffness characteristics. It should be noted
that (1) the damping matrix is typically not defined on an element level and (2)
element externally applied loads are typically limited to the element boundary.
Finite elements fall into three general classes [9], namely:

1. “H” elements, defined exclusively in terms of “ h ” shape functions, the most

common type of elements used in aerospace structural dynamics practice.
2. “P” elements, defined exclusively in terms of “ p ” shape functions (which
encompass the boundary).
3. The most general hybrid “H-P” elements.

The collection of all contributing finite elements for a subject structural system
is described by the matrix set of equations (including additional discrete nonlinear
forces),

[M] {ü} + [B] {u̇} + [K] {u} = [N ] {FN } + [E ] {FE } (8)

where, [ N ] is the nonlinear force allocation matrix for discrete nonlinear

forces,{FN }, and [ E ] is the externally applied load allocation matrix associated
with discrete external forces, {FE }. The discrete, generalized displacements and
velocities associated with the nonlinear forces are
   
{uN } = N T
{u} and {u̇N } = NT
{u̇}. (9)

Finally, the nonlinear forces, {FN }, are calculated employing a generalization of

Eq. 4, which symbolically is represented by

{FN (t + t), {pN (t + t)}} = − {N ({uN (t)}, {u̇N (t)}, {pN (t)})}. (10)
1348 R. Coppolino

In order to clarify the logic associated with the above two dynamic system
equations, let us assume that,

{N} = − [BN ] {u̇N } − [KN ] {uN } (11)

Such a case would result in a special form of Eq. 8 wherein

 
[M] {ü} + [B] {u̇} + [K] {u} = − N BN N
T {u̇} −  K  T {u̇} + [ ] {F } ,
N N N E E
or,
 
[M] {ü} + B + N BN N T {u̇} + K +  K  T {u} = [ ] {F }
N N N E E
(12)

which consistently reduces to the appropriate linear system with additional discrete
forces {FN }.
Equations 8–10 represent a quite general finite element (matrix) set of structural
dynamic equations that are employed in the aerospace industry. Two examples
of widely used (locally) nonlinear applications include (a) the nonlinear Space
Shuttle Orbiter payload bay attachments and (b) general launch vehicle-to-launch
pad interface loads during lift-off dynamics. Specialized matrix methods widely
employed in the aerospace industry are the subject of the next subsection. In a sense,
the above two applications span the categories of “bananas,” “plantains,” and “all
else.”

2 Structural Dynamic Models

This section offers an overview of theoretical modal analysis of linear systems,

which is the predominant basis of the Aerospace Perspective for Modeling and
Validation. In spite of the complicating effects of nonlinearity, modal analysis per-
sists in playing a role in their presence. Knowledge of the bandwidth of a system’s
dynamic environment is essential for determination of guidelines for construction
of relevant structural dynamic models, which must adhere to “granularity” or grid
spacing requirements that follow rigorous frequency-wavelength relationships.

2.1 Modal Analysis

Eigenvalues and eigenvectors associated with a linear structural dynamic system

(Eq. 8) are solutions of either (a) the undamped algebraic equation.

[K] [] = [M] [] [λ] , λn = ωn 2 , which results from [M] {ü} + [K] {u} = {0},
{u} = [] · eiωn t ,
(13)
24 Aerospace Perspective for Modeling and Validation 1349

or (b) the damped algebraic equation

−M−1 B −M−1 K φV φV
= [λ] , λn = σn + iωn , which results from
I 0 φu φu
(14)

[M] {ü} + [B] {u̇} + [K] {u} = {0} , {u̇} = {v} , {u} = [φu ] · eλn t , {v} = [φv ] · eλn t

Nearly all traditional structural dynamics methodology employed in the U.S.

aerospace industry employs undamped (real) modal analysis as a staple. However,
emerging methodology is in the process of exploring potential benefits of damped
(complex) modal analysis, which will not be discussed in the present chapter.
The well-known orthogonality properties of undamped normal modes (for unit
generalized mass scaling) are,
     
T [M] [] = [I], T [K] [] = [λ] = ωn2 (15)

In addition, the distribution of kinetic energy for each mode is the term-by-term
product,

{KE }n = {[M] {n }} ⊗ {n }, (16)

which has an accumulated total value of 1.0 (based on unit generalized mass mode
scaling).
A truncated set of normal modes (typically selected on the basis of a dynamic
frequency cut-off to be discussed in the next subsection) forms the basis of the
modal coordinate transformation,

{u} = [] {q}, (17)

resulting in the modal form of Eqs. 9–11,

         
T M {q̈} + T B {q̇} + T K {q} = T N {N } + T E {FE }
(18)
 
{q̈} + [β] {q̇} + [λ] {q} = T N {N} + T E {FE } ([β] = [2ζn ωn ] is typically
assumed diagonal),
   
{uN } = N
T
 {q}, {u̇N } = N
T
 {q̇} (19)

{FN (t + t), {pN (t + t)}} = − {N ({uN (t)}, {u̇N (t)}, {pN (t)})} (20)
1350 R. Coppolino

In the case of a linear structural dynamic system, the transient response to applied
external loading involves a summation of responses of the uncoupled SDOF modal
equations,
 
q̈n + 2ζn ωn q̇n + ω2n qn = Tn E {FE } (21)

2.2 Dynamic Bandwidth

Applied external loads, {FE (t)}, imposed on integrated launch vehicle and spacecraft
systems during launch and ascent typically occur within the 50–70 Hz base
frequency band. The resulting dynamic stresses are generally termed “primary”
structural loads. Additional “secondary” structural loads (not addressed in this
chapter) that occur in a much wider frequency band are driven by acoustic and
extremely brief shock environments. While U.S. government standards [11, 12]
specify the primary structure frequency band for spacecraft and launch vehicle
systems, the shock response spectrum (SRS) function [13] provides a rigorous basis
for estimation of dynamic bandwidth.
Within the present context, the SRS function associated with a specific force time
history, F(t), is the solution map for linear SDOF systems,

q̈n + 2ζn ωn q̇n + ω2n qn = F (t) , (22)

with natural frequency, ωn = 2πfn , for 0 ≤ fn ≤ fmax , with a selected reference value
for critical damping ratio, ζn . It is highly recommended that the solutions to Eq. 22
be subject to “static equilibrium” initial conditions,

qn (0) = F(0)/ωn2 , q̇n (0) = 0, (23)

in order to suppress unwanted initial transients. The normalized displacement SRS

function is defined as the following solution map,
  
 2πf2 qn (t)
n
SRS (fn ; ζn ) = max
(24)
|F (t)|max

The characteristics of the normalized SRS for a typical transient force history are
depicted in Fig. 3.
While the SRS peak at 10 Hz clearly indicates the dominant frequency content
in the force history, the asymptotic trend for f ≥ 30 Hz (f*) indicates the onset
of quasi-static SDOF response for all modes with frequency above f*. When f* is
determined as the overall maximum value for all components of {FE (t)}, the dynamic
bandwidth for response of any linear structural dynamic system subjected to such
loading is rigorously established. Note that 0 ≥ f ≥ 50 to 70 Hz is the generally
assumed dynamic bandwidth for spacecraft and launch vehicle systems in the U.S,
24 Aerospace Perspective for Modeling and Validation 1351

Fig. 3 Typical applied force Applied Force History

history and associated SRS 1
(ζn = 0.05) 0.5

F(t)
0
-0.5

-1
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Normalized SRS
101
Normalized SRS
100

10-1 SRS
Cut-off f*
10-2
100 101 102
fD (Hz)

which should be reviewed and potentially revised for new systems. It must also be
noted that estimation of f* is somewhat less certain when significant nonlinearities
are present in a subject dynamic system.

2.3 Effective Modeling Guidelines

In order to develop a relevant dynamic model, general requirements should be

addressed based on

1. Frequency band width 0 < f < f*, and intensity of anticipated dynamic environ-
ments, {FE (t)}.
2. General characteristics of structural or mechanical components.

Dynamic environments are generally (a) harmonic, (b) transient, (c) impulsive,
or (d) random. For all categories, the cut-off frequency (f*) is reliably determined
by SRS analysis, as described in the previous subsection. The overall intensity
level of a dynamic environment is described by the peak amplitude for harmonic,
transient, and impulsive events, or by the statistical amplitude (e.g., mean plus a
multiple of the standard deviation) for a long duration random environment. With
the cut-off frequency (f*) established, the shortest relevant wavelength (L) of forced
vibration for components in a structural assembly may be calculated. For finite
element modeling, the quarter wavelength (L/4) is of particular interest, since it
is a rough estimate of the grid spacing needed to characterize system dynamics at
the cut-off frequency. Note that the actual grid spacing requirement is a function of
the specific elements being used. Grid spacing (quarter wavelength) guidelines, for
typical structural components, are summarized in Table 1.
1352 R. Coppolino

Table 1 Grid spacing guidelines for typical structural components

Component Mode type Quarter wavelength (l/4) Additional data
√ 
String Lateral T/ρA / (4f∗ ) T = tension, A = area, ρ = mass
density
√ 
Rod Axial E/ρ / (4f∗ ) E = elastic modulus
√ 
Rod Torsion G/ρ / (4f∗ ) G = shear modulus
√
Beam Bending (π/2) (EI/ρA)1/4 / 2πf∗ EI = flexural stiffness
√ 
Membrane Lateral N/ρh / (4f∗ ) N = stress resultant
√
Plate Bending (π/2) (D/ρh)1/4 / 2πf∗ D = flexural stiffness, h = thickness
√ 
3-D Elastic Dilational E/ρ / (4f∗ )
√ 
3-D Elastic Shear G/ρ / (4f∗ )
√ 
Acoustic Dilational B/ρ / (4f∗ ) B = Bulk Modulus

Accurate modeling of stress concentrations, for a unified dynamic-stress model,

may be addressed by employment of adaptive mesh refinement [9], or by utilization
of documented stress concentration formulae [14]. If mesh refinement is utilized,
the number of model degrees of freedom may increase substantially, imposing
undue computational resource penalties on a model that had been appropriately
designed for study of dynamic response and loads. If localized stress concentra-
tions do not affect a model’s dynamics fidelity, it is prudent to recover detailed
stresses by employing (a) stress concentration formulae and/or (b) separate, detailed
local or global “stress” finite element models (designed using adaptive mesh
refinement).

2.4 Further Thoughts on Structural Dynamic Modeling

Additional topics of significance with regard to structural dynamic modeling which

are not explicitly included in this chapter include:

(a) Damping and Joint Characterization (Misconceptions and Realities) [15]

(b) Behavior and Modeling of Non-Standard Materials [15]
(c) Fluid-Structure Interaction (Propellant Tanks) [8].

Treatment of the first two topics in the U.S. aerospace industry typically includes
attempts to “fit” actual physical behavior of damping, joints, and non-standard
materials (1) within the context of linear “M,B,K” modeling practices and (2)
if deemed appropriate by specialized localized nonlinear models. Modeling of
propellant tank fluid-structure interaction is extensively treated in the NASTRAN
environment employing methodologies documented in reference [8] and further
refinements/developments as well as proprietary techniques developed by U.S.
aerospace corporations .
24 Aerospace Perspective for Modeling and Validation 1353

3 Matrix Structural Dynamic Analysis

The U.S. aerospace industry, specifically the launch vehicle and spacecraft com-
munity, over the past 60 years has developed and/or adopted a series of matrix
structural dynamic analysis methods to address its unique modeling and validation
goals. This section provides an overview of the broadly accepted methodologies,
which fall into the categories of (a) model order reduction, (b) modal substructure
definition, (c) component mode synthesis, (d) system dynamic response analysis
including treatment of local nonlinearities, and (d) accurate estimation of local
structural member loads and stresses.

3.1 Guyan Reduction

The underlying idea that defines Guyan Reduction (GR) [16] is static
condensation . . . and Bob Guyan’s monumental formulation was published as a one-
half page technical note! The degrees of freedom describing a structural dynamic
system are first separated into “analysis” and “omitted” subsets, which lead to the
partitioned matrix equation (ignoring damping)

Moo Moa Üo Koo Koa Uo Fo

+ = (25)
Mao Maa Üa Kao Kaa Ua Fa

If only the [Maa ] partition of the mass matrix were non-zero, and external
forces were only applied to “analysis” degrees of freedom, the relationship between
“analysis” and “omitted” degrees of freedom (the GR transformation) would be

Uo −K−1
oo Koa {U }
= a (26)
Ua Iaa

In that situation, the above reduction transformation would be exact. However,

when the mass matrix partitions, [Moo ], [Moa ], and [Mao ], are non-zero, the
reduction transformation is approximate (its columns are Ritz shape functions).
Application of the reduction transformation, in a symmetric manner following the
Ritz method yields, respectively, the (statically exact) reduced stiffness matrix and
(dynamically approximate) mass matrix

 −K−1
T
−K−1  
Kaa = oo Koa ·
Koo Koa
· oo Koa = [Kaa ] − Kao K−1
oo Koa
Iaa Kao Kaa Iaa
(27)

T
 −K−1
oo Koa Moo Moa −K−1
oo Koa
Maa = · · (28)
Iaa Mao Maa Iaa
1354 R. Coppolino

The reduced (approximate) matrix equation set is therefore

   Fo
Maa {üa } + Kaa {ua } = −Kao K−1
oo Iaa . (29)
Fa

GR was initially developed to permit efficient, approximate modal analysis of

structural dynamic models that were considered of excessively large order for
computers in the late 1960s. Advancements in computer hardware and numerical
methods have eliminated that original need. That being said, GR is still employed
as a tool in modal test planning to be discussed in Sect. 4.2.

3.2 The Hurty-Craig-Bampton Method

The Hurty-Craig-Bampton (HCB) method [17, 18] continues to be the most widely
applied method for definition of component substructures (also called superele-
ments). The matrix equations of a single HCB component are a logical extension
of Guyan Reduction. In a simple, direct manner, the HCB component is defined
on the basis of Eq. 25, wherein the boundary degrees of freedom (which form the
interface with an adjacent substructure) are the “a” subset, and the interior degrees
of freedom are the “o” subset. The interior displacements, {uo } are expressed in
terms of selected “low frequency” normal modes of the “interface-fixed” eigenvalue
problem,
 
[Koo ] [on ] = [Moo ] [on ] ωn2 . (30)

Therefore, the approximate relationship between physical and HCB degrees of

freedom (the HCB transformation) is defined as,

Uo on −K−1
oo Koa qn
= (31)
Ua 0an Iaa Ua

In practice, the number of “interior” component modes included in [on ] should

be consistent with the established dynamic cut-off frequency, f*, discussed in Sect.
1.2. Application of the HCB transformation to Eq. 25, in a symmetric manner
following the Ritz method yields,

Inn Pna q̈n ω2n 0na qn Ton 0na Fo

+ = −1 . (32)
PTna Maa üa 0an Kaa ua − Kao Koo Iaa Fa

The above result, compared with Eq. 29, illustrates that HCB is a mathematical
extension of GR. That being noted, the two methods are typically employed for
differing purposes.
24 Aerospace Perspective for Modeling and Validation 1355

Fig. 4 Hypothetical HCB

launch vehicle stack

In a sense, the HCB component can be thought of as an H-P finite element (see
Sect. 1.3) with “qn ” representing the “P” degrees of freedom and “ua ” representing
the “H” degrees of freedom. Considering the hypothetical launch vehicle “stack”
dynamic system in Fig. 4, the assembled HCB system will adopt a “collective”
matrix form conforming to Eq. 32.
On the assumption that Eq. 32 (also) represents the assembled dynamic system,
the system modes, [sys ], and modal frequencies, [ωsys ], (eigenvalue solution for
Eq. 32) relate to the physical system DOF (Eq. 31) as follows:

Uo on −K−1
oo Koa sys,qn sys,qn 
= , where = sys . (33)
Ua 0an Iaa sys,ua sys,ua

3.3 The Benfield-Hruda Method

While the Hurty-Craig-Bampton method is the most widely used approach (in the
U.S. aerospace industry) for definition of integrated launch vehicle/spacecraft “com-
ponent” or “superelement” dynamic models, the Benfield-Hruda (BH) method [19]
endures as the most widely used method for assembly of launch vehicle/spacecraft
dynamic models. The nuance of the BH method is that a “main body” substructure
is defined by the matrix equation set,

 
Maa,m {üa } + Kaa,m {ua } = {0}. (34)

Assuming that the collection of HCB components (Eq. 32) “maps” onto the main
body physical DOF set, the assembled dynamic system equations are,
1356 R. Coppolino

Inn Pna q̈n ω2 0na qn Ton 0na Fo

+ n = −1 .
Pna Maa +Maa,m
T üa 0an Kaa + Kaa,m ua − Kao Koo Iaa Fa
(35)

The partition related to main body degrees of freedom, {ua }, represents a main
body structure “augmented” with “rigid body” substructure branches. In actuality,
the branches are “rigid bodies” with locally flexible attachment DOF partitions.
A signature feature of the HB method is intermediate computation of augmented
main body modes, specifically,

   
Kaa + Kaa,m [a ] = Maa + Maa,m [a ] ωa2 . (36)

The intermediate stage transformation and resulting matrix dynamic equations,

respectively, are

qn In 0 qn
= , and (37)
ua 0 a qa

In Pna a q̈n ωn2 0na qn Ton 0na Fo

+ = .
Ta PTna Iaa qa 0an ωa2 ua − a Kao K−1
T
oo Ta Fa
(38)

The system modes, [sys ], and modal frequencies, [ωsys ], (eigenvalue solution
for Eq. 38) relate to the physical system DOF (Eqs. 31 and 35) as follows:

Uo on −K−1
oo Koa In 0 sys,qn sys,qn 
= , where = sys .
Ua 0an Iaa 0 a sys,ua sys,ua
(39)

The three key steps in the BH method are depicted in the hypothetical launch
vehicle “stack” dynamic system below in Fig. 5, where, “I” represents the collection
of HCB components (branch) and main body, “II” represents the assembled main
body and “rigid” branch, and “III” represents the assembled dynamic system
(Eq. 35) .

3.4 The MacNeal-Rubin Method

While the Hurty-Craig-Bampton and Benfield-Hruda methods are the most widely
employed strategies for assembly of launch vehicle/spacecraft dynamic models,
another approach, the MacNeal-Rubin [20, 21] (MR) method is noteworthy for
its utilization in challenging situations involving localized interface nonlinearities.
While there are a multitude of MR-type formulations, the present discussion focuses
24 Aerospace Perspective for Modeling and Validation 1357

Fig. 5 Hypothetical BH launch vehicle stack

on foundational equations, which apply to the unconstrained structure (damping

ignored),

[M] {ü} + [K] {u} = [] {F} , (40)

MacNeal noted that “static” equilibrium of the above system associated with
attachment loads is described by
 
[K] {uS } = [] {F} or {uS } = K−1 [] {F} (41)

In actuality, the unconstrained stiffness matrix is singular and its inverse must be
conditioned in one of a variety of mechanics-mathematical techniques. One general
approach, which encompasses unconstrained structures that have greater than 6
DOF singularities (rigid body motions) due to mechanisms, employs a small “shift”
operator, described by

[K + λS M] {uS } = [] {F} or {uS } = [K + λS M]−1 [] {F} (42)

Expressing the “static” displacements in terms of a truncated set of low frequency

modes (including rigid body and mechanism modes), i.e.,

{us } = [ ] {q }, (43)

the “shifted static” displacements are

1358 R. Coppolino

 −1  
{us } = [ ] ω2 + λs T  {F}. (44)

At the heart of MacNeal’s formulation [20] is the definition of residual “quasi-

static” response,

   −1  
usρ = {us } − {us } = [K + λs M]−1 [] − [ ] ω2 + λs T 
(45)

{F} = ρ {F}.

Rubin’s
 “improved”
 formulation [21] added a residual coefficient term associ-
ated with F̈ based on a MacLauren series expansion. Both MacNeal’s and Rubin’s
formulations resulted in unorthodox, non-standard mixed displacement and force
DOF dynamic equations that posed a challenge to conventional matrix structural
dynamics practices.
An alternative (conventional) formulation was introduced by Coppolino [22] that
employed the residual matrix, [ ρ ], residual “Ritz” vectors, which augment the
lower frequency modes, [ ]. The augmented DOF transformation is therefore,

 q
{u} =  ρ (46)
qρ

It is important to note that the residual vectors are a linear combination of the
higher frequency modes of the dynamic system (and the higher frequency modes
need not be known in order to form the residual vectors). Due to this property, the
lower frequency modes and residual vectors are mutually orthogonal to each other,
and the residual vectors are typically converted to pseudo-modes, [ρ ], by solution
of the following matrix equations:
       
Mρ = ρT [M] ρ , Kρ = ρT [K] ρ , (47)

        
Kρ φρ = Mρ φρ ωρ2 , ρ = ρ φρ .

Therefore, the augmented unconstrained modal transformation is

 q
{u} =  ρ , (48)
qρ

and the uncoupled MacNeal-Rubin equations [22] become

I 0 q̈ ω2 0 q T 

+ = {F} . (49)
0 Iρ q̈ρ 0 ωρ2 qρ Tρ 
24 Aerospace Perspective for Modeling and Validation 1359

The above form of the MacNeal-Rubin method is identical to uncoupled

(truncated set) modal equations for an unconstrained system. However, exact static
behavior is guaranteed by Eq. 49, in contrast to an approximation employing a
truncated set of system modes, which typically requires many more high frequency
normal modes than “residual” modes to achieve satisfactory convergence.

3.5 Application of Hurty-Craig-Bampton and MacNeal-Rubin

Methodology

A prominent past application of the Hurty-Craig-Bampton and MacNeal-Rubin

methods is found in space shuttle payload dynamic loads analyses. Attachment of
cargo manifests in the space shuttle payload bay, as illustrated in Fig. 6, involved
fittings that exhibited nonlinear stick-slip mechanical behavior.
The typical HCB cargo element (contractor) model was supplied in the following
matrix form,

Inn Pna q̈n 2ζn ωn 0na q̇n ωn2 0na qn 0o

+ + = , (50)
PTna Maa üa 0an Baa u̇a 0an Kaa ua Fa

Fig. 6 Space shuttle cargo bay payload interface

1360 R. Coppolino

where the generalized damping matrix terms were defined on the basis of cargo
element engineering judgment and experimental data, if available. The nonlinear
interface forces are represented by {Fa }, and the applied interior DOF forces, {Fo },
(see Eq. 32) are assumed null due to the fact that the cargo is totally enclosed within
the orbiter’s payload bay (Note: Payload bay acoustic loads were generally treated
in a separate analysis).
The space shuttle dynamic model (often related to liftoff and abort landing
configurations) was provided in a standard format, which ultimately was described
by MR equations of the form,

I 0 q̈l 2ζ ω 0 q ω2 0 q

+ +
0 Iρ q̈ρ 0 2ζρ ωρ qρ 0 ω2ρ qρ

T e T a
= T {Fe } + {−Fa } (51)
ρ  e Tρ a

where the generalized damping matrix terms were defined on the basis of cargo
element engineering judgment and experimental data. In addition, the distribution,
[ e ], and time histories of standardized external applied load transients, {Fe },
respectively were provided by the space shuttle contractor. Physical interface
displacements, on the space shuttle payload bay side, were recovered via the modal
transformations,

         
{ua }s = aT  {q } + aT ρ qρ , {u̇a }s = aT  {q̇ } + aT ρ q̇ρ
(52)

The relative displacements and velocities associated with the nonlinear stick-slip
interface DOF are described by,

{ua } = {ua } − {ua }s , {u̇a } = {u̇a } − {u̇a }s . (53)

Finally, the stick-slip interface forces were appropriately computed with hys-
teretic equations of the type,

{Fa (t + t), {pa (t + t)}} = − {N ({ua (t)}, {u̇a (t)}, {pa (t)})}. (54)

Integrated system dynamic responses associated with Eqs. 50–54 are computed
by one among a variety of stable numerical integration methods.
24 Aerospace Perspective for Modeling and Validation 1361

3.6 Detailed Structural Dynamic Loads and the Mode

Acceleration Method

The relationship between system dynamic response, expressed in terms of system

modal DOF displacements, and detailed structural member loads and stresses
(internal loads) is typically described by the matrix equation,

{σ } = [Kσ ] {u} + [Bσ ] {u̇} = [Kσ ] {q} + [Bσ ] {q̇}, (55)

where the velocity dependent term is generally ignored. It should be noted that the
modal matrix, [], is associated with a truncated (base-band) set of system modes.
It is a well-established fact that when the truncated mode set does not include quasi-
static residual modes (see Eq. 51), serious inaccuracies in internal loads may occur.
In contrast, inclusion of quasi-static residual modes automatically accounts for the
quasi-static response of all higher frequency band modes, eliminating the modal
truncation deficiency.
A second, more prevalent approach to recovery of internal loads, namely the
mode acceleration method, was introduced in 1945 by D. Williams [23]. From a
matrix viewpoint, the method is derived from the assembled physical (non-reduced
order) dynamic system equations with externally applied, inertial and dissipative
forces on the right-hand side. Consider a linear structural dynamic system,

[K] {u} = [e ] {Fe } − [M] {ü} − [B] {u̇}, (56)

with incorporation of right-hand side modal substitutions,

[K] {u} = {Ftotal } = [e ] {Fe } − [Mn ] {q̈n } − [Bn ] {q̇n }. (57)

It should be noted that interpretation of [ e ]{Fe } as inclusive of the nonlinear

forces in Eq. 8 renders the present derivation applicable to a system with local
nonlinear forces.
Since the system stiffness matrix is most often singular due to a free-free
condition (and sometimes mechanisms), the full set displacement DOF are separated
into “relative” (l) and “reference rigid body” DOF (r) partitions resulting in,

Kll Klr ul Ftotal,l ul Ill K−1

ll Klr u l
= → = → (58)
Krl Krr ur Ftotal,r ur 0rl Irr ur

Kll 0lr u l Ill 0lr Ftotal,l

→ = ,
0rl Krr − Krl K−1
ll Klr ur − Krl K−1
ll Irr Ftotal,r

where [Kll ] is the positive-definite relative stiffness matrix. Focusing on the upper
partition, the mode acceleration relative displacements (akin to strains) are,
1362 R. Coppolino

    −1         
u l = Kll Ftotal,l = K−1
ll [ e ] l {Fe } − K−1
ll [M n ] l {q̈n } − K−1
ll

[Bn ]l {q̇n } (59)

Introducing the internal “stress” (or member load) relationship,

 
{σ} = [Kσ ] u l , (60)

the system-level mode acceleration equations become,

 
{σ} = [LTMF ] {Fe } − LTMq̈ {q̈n } − LTMq̇ {q̇n }, (61)

    
[LTMF ] = [Kσ ] K−1
ll [ e ]l , LTM q̈ = [Kσ ] K−1
ll [Mn ]l ,
  
LTMq̇ = [Kσ ] K−1
ll [Bn ]l .

The key advantage of the mode acceleration method is that it automatically

accounts for the quasi-static response of all higher frequency modes, without
the need to explicitly compute those modes. The second, vital advantage of the
mode acceleration method relates to  the fact that [LTMF ]{Fe } is the quasi-static
contribution to internal loads, and LTMq {q̈n } is the system mode acceleration
contribution to internal loads. The modal velocity contribution to internal loads is
often neglected.
In closing the present discussion on the mode acceleration method, it should
be noted that expressions for load transformation matrices are often developed by
employment of Hurty-Craig-Bampton (HCB) component models, which employ
expansion transformations from HCB to system modal accelerations and velocity
DOF. While such an approach is employed out of necessity (HCB models based
on proprietary FEMs are provided by the payload contractors to payload integration
contractors), the resulting load transformation matrices should be equivalent to those
described by Eqs. 56– 61.

4 Verifiction and Validation of Structural Dynamic Models

Over the past 60 years, the U.S. aerospace community has developed, refined,
and standardized an integrated approach to structural dynamic model verification
and validation. One name for this overall approach is the Integrated Test Analysis
Process (ITAP) for structural dynamic systems, which is summarized in Fig. 7 and
discussed in the immediately following subsections.
24 Aerospace Perspective for Modeling and Validation 1363

Fig. 7 Overview of the

integrated test analysis
process

4.1 System Dynamic Model

The system dynamic model to be verified, validated, and updated in the ITAP
process should be developed employing (a) adherence to the guidelines discussed in
Sects. 2.2, 2.3 and 2.4 and consistency with design documentation and engineering
drawings. It is inexcusable today that “consistency” is often not practiced in many
aerospace organizations, in spite of the availability of modern CAE tools that
facilitate automated drawing-to-FEM consistency.
Beyond the above recommended modeling practices, a vital preliminary step
in the ITAP process is initial model verification. Over the past 60 years, rigid
body checks have been a standard part of the model verification process. Using
a geometrically defined rigid body displacement matrix, [ RB ], the rigid body
mass properties of a structure (both the complete assembly and subassemblies in
unconstrained form) defined by,

[MRB ] = [RB ]T [M] [RB ], (62)

are correlated with reference engineering properties (total mass, center of gravity,
mass moments of inertia) of the designed system. It should be noted that aircraft
and space system development organizations generally include weight engineering
departments which manage system mass budgets throughout the development
process. The Society of Allied Weights Engineers (SAWE) was established in the
U.S. in 1939 to support the weights engineering community.
1364 R. Coppolino

The companion initial model verification calculation

[KRB ] = [RB ]T [K] [RB ], (63)

is employed on the structure (both the complete assembly and subassemblies in

unconstrained form) to verify integrity of the modeling process. [KRB ] must be
a null matrix for an unconstrained structure. It should be noted that in special
situations involving structures with “mechanisms” additional unconstrained veri-
fication calculations are required on the assembled system model (i.e., more than
six stiffness free “modes”).

4.2 Modal Test Planning and the Test Analysis Model (TAM)

As a general rule, the number of system model degrees-of-freedom is substantially

greater than a practical number of acceleration sensor channels. Before 1965,
allocation of a modal test accelerometer array was a task guided primarily by general
understanding of the character of anticipated system modes. Moreover, the assembly
of a system mass allocation matrix (corresponding to the test accelerometer array)
was conducted on the basis of engineering experience and “expertise.” The mass
allocation matrix was deemed necessary for validation of test mode orthogonality,
especially in aeronautical applications.
Introduction of Guyan reduction (described in Sect. 3.1) in 1965 offered a
systematic strategy for definition of an appropriate test accelerometer array and
system mass allocation matrix. In the context of modal test planning, the “analysis
set” DOFs (Eq. 26) correspond to instrumented (accelerometer) response channels
on the test article and it is assumed (based on engineering intuition) that the “omit”
degrees of freedom (approximately) follow a static relationship with respect to
the “analysis” set degrees of freedom. The reduced stiffness and mass matrices
(corresponding to Eqs. 27 and 28) are called the test-analysis model (TAM) stiffness,
[KTAM ], and mass, [MTAM ], matrices, respectively, within the context of the modal
test plan; it should be noted that the TAM mass matrix is the more important quantity
employed in the modal test plan.
Appropriateness of a selected “analysis” set may be verified by comparison of
approximate modes, []TAM , and natural frequencies associated with the reduced
eigenvalue problem (Eqs. 26–28) with the results for the exact eigenvalue problem
(Eq. 25). The exact and approximate modal vectors (which use Eq. 26 as an
expansion transformation) are,

o −K−1
oo Koa [ ]
[]exact = , []TAM = a exact (64)
a exact
Iaa

The choice of a specific set of “analysis” degrees of freedom is acceptable if all

“TAM” approximate modes and natural frequencies are “satisfactory” agreement
24 Aerospace Perspective for Modeling and Validation 1365

with the “exact” system modes and natural frequencies. A very clear example
of an inadequate selection of “analysis” set degrees of freedom is discussed in
Reference [24], which describes difficulties encountered in modal testing of a trim
body automobile test article. In that experience, the front seats were not modeled or
designated by “analysis” degrees of freedom and local seat flexing modes which
coupled with overall body vertical bending motions were obscured. In response
to this and other experiences, a variety of automated “analysis” DOF selection
algorithms were developed. Reference [25] describes the currently accepted strategy
from which other DOF selection methodologies have been developed, to be
discussed in Sect. 4.4.
Utilization of Guyan reduction to designate (1) the “analysis” set degrees of
freedom as the modal test accelerometer array and (2) the reduced mass matrix,
[MTAM ], as the test-analysis model (TAM) mass (allocation) matrix is the most
prevalent, contemporary test planning strategy in the U.S. aerospace industry.
Allocation of force excitation resources on the basis of modal gains, {n }T [ e ],
complements the allocated accelerometer array. Additional strain gage sensor
arrays have been successfully employed by NASA/JPL [25] and others since
the 1980s as part of modal survey planning and execution. Such information
provides a basis for correlation and validation of load transformation matrices
(see Eq. 61).
Finally, it should be pointed out that allocation of an accelerometer array
may require special considerations to avoid the mapping of unwanted modes,
specifically breathing modes of plate and shell subassemblies (that are not
included in the target mode set). In such situations, accelerometers directed
normal to a “breathing” surface should not be allocated. Moreover, if possible,
tangentially oriented accelerometers should be allocated only at stiffening rings
and spars to further suppress measurement of breathing modes, while focusing
on “body” modes (e.g., axial, bending, and torsion modes). The above cited
practice, while popular, must be employed with caution and selectively as (a)
slight imperfections and manufacturing details cause mixing of “body” and
“breathing” modes, and (b) suppression of sometimes important shell “bulge”
modes.
In summary, the following data sets form the basis of modal test planning in the
U.S. aerospace industry:

(a) TAM mass matrix, [MTAM ], derived from appropriate application of the Guyan
reduction process.
(b) TAM predicted modes and natural frequencies based on solution of the exact
(unreduced) eigenvalue problem. The accelerometer array modal partition
represents the predicted test modes. It should be noted that some organizations
use the Guyan reduction solution as the predicted mode set.
(c) Applied force allocation, based on predicted mode gains to excite all target
modes.
(d) Additional designated strain gages at critical locations based on pre-test flight
loads analyses (the “NASA/JPL” strategy [25]).
1366 R. Coppolino

4.3 Target Modes

Selection of a subset of significant “target” modes from the complete set of predicted
modes of a test article is a practice employed by many aerospace organizations.
A common, yet often inadequate approach to target mode selection involves a
(significant) modal effective mass criterion (see Eq. 35). Additional target modes,
however,
 may be selected based on modal excitations imposed on non-boundary
DOF, Ton {Fo }, also included in Eq. 35. In any case, the practice of focus
exclusively on target modes, rather than all modes within the structural dynamic
frequency band (0 < f < f*) is risky due to mixing with “non-target” modes as a
result of imperfections and manufacturing details.

4.4 Automated Response DOF Selection for Mapping

of Experimental Modes

The difference between exact and approximate analytical eigenvectors, described in

Eq. 64, is the residual displacement error matrix,

o + K−1
oo Koa a .
[R ] = (65)
0a

Note that the residual error associated with the “measured” or “analysis” DOF
partition is null. The modal kinetic energy distribution for the complete system is,

[E ] = [M] ⊗ [] (66)

where the column sum for each individual mode is unity (if the modes are
normalized to unit modal mass). The residual kinetic energy matrix is now defined
in a similar manner as,

[ER ] = [MR] ⊗ [R] (67)

Like the residual displacement error, [R], the residual kinetic energy matrix
is exactly “zero” at the rows corresponding to the measured DOF. The expected
characteristic that residual energy is pronounced at “omitted” yet dynamically
significant DOF in any particular mode is demonstrated in the illustrative example
structure along with its first six predicted modes, pictured in Fig. 8 (from Reference
[26]) which follows.
Six DOFs (1,3,5,7,10,19), which are insufficient for mapping modal behavior
on the “upper portion” of the bar, are first selected as response “measurement”
DOFs. The residual kinetic energy associated with this selected response DOF set
is illustrated in Fig. 9 as a surface with the horizontal axes corresponding to FEM
DOF number and mode number, respectively. The surface clearly indicates that the
24 Aerospace Perspective for Modeling and Validation 1367

Fig. 8 Simple free-free bar structure and normal modes

Fig. 9 Residual kinetic

energy surface for the 6 DOF
Res Energy (%)

measured set
40
20
0
6
5
4
3 15
2 10
Mode 5
1 DOF

“measurement” DOF set is adequate for the first three modes. Pronounced residual
kinetic energy terms in modes 4–6 suggest that at least two to three more DOFs on
the “upper portion” of the bar structure must be instrumented to adequately map
modes 4–6.
It is of interest to note the orthogonality of the approximate “measured” modes,
shown below, for the present DOF selection (1,3,5,7,10,19)
   
[ORTAM ] = TTAM [M] [TAM ] = Ta [MTAM ] [a ]
⎡ ⎤
100 10 −1 −55 −58 −46
⎢ 10 100 −5 44 51 45 ⎥
⎢ ⎥
⎢ ⎥
⎢ − 1 −5 100 −18 3 29 ⎥
=⎢ ⎥ (68)
⎢ − 55 44 −18 100 23 60 ⎥
⎢ ⎥
⎣ − 58 51 3 23 100 33 ⎦
− 46 45 29 60 33 100
1368 R. Coppolino

Fig. 10 Residual kinetic

energy surface for the

Res Energy (%)

improved measured DOF set
10
5
0
6
5
4
3
15
2 10
Mode 5
1 DOF

The “acceptable” orthogonality among modes 1–3 confirms adequacy of the

selected DOF set for these modes, while the extremely poor orthogonality of the
remaining modes indicates a need for additional measurement DOF.
Following the recommendation for additional “measured” DOF implied by the
“6 DOF based” residual kinetic energy, shown in Fig. 9, the refined residual kinetic
energy (augmented with DOFs 13, 15, and 17) is
The residual kinetic energy reduces substantially as illustrated in Fig. 10. The
improvement in orthogonality of “approximate” measured modes, shown below,
indicates substantial improvement of the refined response measurement DOF set.
⎡ ⎤
100 0 −2 0 5 0
⎢ 0 100 −6 ⎥
⎢ 0 2 0 ⎥
⎢ ⎥
⎢ −2 0 100 0 7 0 ⎥
[ORTAM ] = ⎢ ⎥ (69)
⎢ 0 2 0 100 0 20 ⎥
⎢ ⎥
⎣ 5 0 7 0 100 0 ⎦
0 −6 0 20 0 100

Following its introduction in Reference [26], the automated response DOF

selection technique, today known as the RKE method, was applied in planning of
two International Space Station (ISS) modal tests. RKE planning results for the ISS
P5 modal test [27] are illustrated in Fig. 11. One noteworthy feature of the ISS
P5 modal test plan was inclusion of measured DOF on the NASA/MSFC shuttle
payload simulation test fixture, which had a few modes below 50 Hz.

4.5 Measured Data Acquisition, Data Analysis, Experimental

Modal Analysis

Laboratory-centric tasks addressing measured data acquisition, measured data

analysis, and experimental modal analysis, have undergone a high degree of and
24 Aerospace Perspective for Modeling and Validation 1369

Fig. 11 ISS P5 modal test plan predicted mode orthogonality

refinement during the past 50+ years. That being said, there are no specific
standards representing aerospace industry practice in the U.S. In keeping with the
intent of the present chapter, this subject matter is not addressed herein. Information
on the laboratory-centric tasks is covered in other chapters in this handbook.

4.6 Modal Test-Analysis Correlation and U.S. Government

Standards

The relationship between test modes, [TEST ], and their analytical counterparts,
[TAM ], is described by the transformation,

[TEST ] = [TAM ] [COR] + [R] (70)

where [COR] is the cross-orthogonality matrix and [R] is the residual error matrix.
Employing [MTAM ] as a weighting matrix, the least squares solution for cross-
orthogonality is,

 
[COR] = [ORTAM ]−1 TTAM MTAM TEST , (71)

where the TAM mode orthogonality matrix (consistent with Eq. 68) is,

 
[ORTAM ] = TTAM MTAM TAM ≈ [I] (72)

It should be noted that [ORTAM ] is exactly an identity matrix only if the

“analysis” modes are the unit mass normalized modes of the Guyan reduced
structural dynamic model.
1370 R. Coppolino

An additional useful modal coherence matrix, [COH], is defined as follows,

   
COHt = [I] − [ORTEST ]−1/2 RT MR [ORTEST ]−1/2 (73)

= [I] − [ORTEST ]−1/2 CORt ORTAM COR [ORTEST ]−1/2 .

It should be noted that (a) TAM mode orthogonality, (b) test mode orthogonality,
(c) test-TAM mode cross-orthogonality matrices, along with (d) corresponding
test-analysis natural frequency differences, represent the accepted test-analysis
correlation metrics in the U.S. aerospace industry. Modal coherence, [COH], is
employed by some investigators to judge the degree to which test and analysis modal
clusters are linear combinations of one another; however, this metric is not widely
recognized.
The following standards are imposed on space systems contractors by U.S.
government agencies:

Air Force Space Command [28]

For all spacecraft modes up to 70 Hz, test mode orthogonality (unit mass
normalized)

| ORTEST,ij |≤ 0.10 (10%) for i = j

Test-Analysis mode cross-orthogonality (both sets unit mass normalized)

| CORij |≥ 0.95 (95%) for i = j

| CORij |≤ 0.10 (10%) for i = j

Difference between corresponding test and analysis modal frequencies must be

less than 3%.

NASA [29]

For all significant modes, test mode orthogonality (unit mass normalized)

| ORTEST,ij |≤ 0.10 (10%) for i = j

Test-Analysis mode cross-orthogonality (both sets unit mass normalized)

| CORij |≥ 0.90 (90%) for i = j

| CORij |≤ 0.10 (10%) for i = j

24 Aerospace Perspective for Modeling and Validation 1371

Difference between corresponding test and analysis modal frequencies must be

less than 5%.
It should be noted that the stricter standards imposed by the U.S. Air Force Space
Command, particularly for spacecraft, are suited for situations in which test data is
employed directly to construct a hybrid test-analysis dynamic model [30], which
circumvents an often lengthy model correlation and reconciliation process.

4.7 Overview of Efficient Structural Dynamic Sensitivity Analysis

The final goal in the ITAP process involves reconciliation of the test article FEM
and modal test data. While many aerospace and software organizations have custom
methodologies that address the reconciliation task, there is no universally accepted
approach to such an endeavor. That being said, a numerically efficient technique,
namely Residual Mode Augmentation (RMA) [31] for calculation of FEM sen-
sitivity to large parametric variation has gained acceptance in U.S. aerospace
government and industry organizations.
Efficient computation of structural dynamic modal frequency and mode shape
sensitivities associated with variation of physical stiffness and mass parameters is
essential for (1) practical design sensitivity and uncertainty studies and (2) reconcil-
iation of finite element models with modal test data. Sensitivity analysis procedures
fall in two distinct categories, namely (a) modal derivatives for small parametric
variation and (b) altered system modes associated with “large” parametric variation.
The latter category is generally applicable to modal testing, which often requires
significant local parameter changes at joints to effect FEM-test reconciliation.
However, many investigators and commercial software packages employ estimated
modal derivatives in optimization strategies, which address FEM-test reconciliation
objectives.
Since the 1960s, methods for computation of modal frequency and mode
shape derivatives have been developed. Fox and Kapoor [32] introduced an exact
derivative formulation that required knowledge of all modes of the original system;
application of the procedure when a truncated set of modes was employed produced
compromised derivatives. In response to this difficulty, Nelson [33] derived an
exact formulation for computation of mode shape derivatives for truncated mode
sets. Efforts to refine and extend application of mode shape derivatives for finite
parameter change sensitivity computations have been pursued by many investiga-
tors. However, the need for modal frequency and mode shape sensitivities that map
over very large ranges for multiple parameters suggests application of alternative
Ritz strategies.
The Ritz method [4] is one of the most significant developments in analytical
mechanics of the past century. This method provides a logical energy formulation
for consistent reduction of mass and stiffness matrices employing a set of trial
vectors as a reduction transformation. Effectiveness and accuracy of the reduction
process depend on selection of an appropriate trial vector set. When a truncated set
of baseline system mode shapes is used as the trial vector set (popularly known as
1372 R. Coppolino

Structural Dynamic Modification (SDM) [34], the Ritz method often produces poor
estimates for the altered system. Augmentation of the truncated baseline system
mode shapes with appropriately defined additional vectors, however, has been found
to produce extremely accurate altered system modal frequencies and mode shapes.
Quasi-static residual vectors [20], appended to a truncated set of mode shapes,
were found to produce extremely accurate modes for offshore oil platform models
subjected to localized alterations [35]. Residual Mode Augmentation (RMA),
introduced in 2002 [27] and thoroughly discussed in Reference [31], is a procedure
that defines augmented trial vectors, which are appropriate for structures subjected
to highly distributed, as well as localized, alterations.

4.8 Residual Mode Augmentation (RMA)

The matrix equations describing exact free vibration of baseline and altered
structures, respectively, are

[KO ] [O ] − [MO ] [O ] [λO ] = [0], and

[KO + p · K] [] − [MO + p · M] [] [λ] = [0]. (74)

It is implicitly assumed that the stiffness and mass changes scale linearly with
respect to the parameter, p. Therefore, changes in “beam” depth may not be directly
applied, since the axial stiffness (AE) scales linearly with depth and the flexural
stiffness (EI) scales as the cube of depth. The appropriate formulation for Eq. 74
permits linear sensitivity of “AE” and “EI” separately.
The relationship between mode shapes of the baseline and altered structures is
expressed as the cross-orthogonality of orthonormal mode shape sets,
 
[COR] = TO [MO ] [], (75)

where the baseline and altered structure modes are unit mass normalized.
The most fundamental Ritz approximation, used in SDM [34], employs a
truncated set of low frequency eigenvalues as the reduction transformation described
by

[] = [OL ] [φ], (76)

where the reduced baseline structure stiffness and mass matrices, respectively, are
   
[kO ] = TOL KO OL = [λOL ], [mO ] = TOL MO OL = [IOL ], (77)

the reduced stiffness and mass sensitivity matrices, respectively, are

24 Aerospace Perspective for Modeling and Validation 1373

   
[k] = TOL KOL , [m] = TOL MOL , (78)

and the reduced altered structure free vibration equation is

 
λOL + p · k [φ] − IOL + p · m [φ] [λ] = [0]. (79)

A well-known result of this type of trial vector reduction strategy is that the
approximate altered structure eigenvalues are generally higher than results for the
exact solution, and the approximate mode shapes do not closely follow the exact
shapes when parametric alterations are large.
When structural alterations are well-dispersed, parametric structural changes
may affect many physical degrees of freedom and require a description in terms of
several independent scaling parameters, “pi ”. The expressions for altered stiffness
and mass matrices in such a situation are


N 
N
[K] = [KO ] + pi [Ki ], [M] = [MO ] + pi [Mi ], (80)
i=1 i=1

The altered system free vibration matrix equation for this situation is
   

N 
N
KO + pi [Ki ] [] − MO + pi [Mi ] [] [λ] = [0], (81)
i=1 i=1

Definition of residual vectors associated with dispersed, independent alterations

of a baseline structure is accomplished by first computing the lowest frequency
mode shapes of the baseline structure (Eq. 74) as well as the lowest mode shapes
associated with each independent alteration of the structure.
 
KO + pi Ki [iL ] − MO + pi Mi [iL ] [λiL ] = [0] (for i = 1, . . . , N),
(82)

The selected value of each independent scaling parameter is sufficiently large to

produce a substantial change in mode shapes (with respect to the baseline structure).
An initial set of trial vectors that adequately (and perhaps redundantly) encompass
all potential (low frequency) altered system mode shapes is

[] = 1L 2L . . . NL (83)

This set of trial vectors is expressible as the sum of (a) a linear combination of
baseline system mode shapes and (b) trial vectors (that are linearly independent of
the baseline system mode shapes).

[] = [OL ] [COR] +   (84)
1374 R. Coppolino

The cross-orthogonality coefficient matrix is determined based on the following

least-squares solution.
     
TOL MO  = TOL MO OL [COR] + TOL MO   = [IOL ] [COR] + [0],
(85)

where,
    
[COR] = TOL [MO ] [] and   = IOL − OL TOL MO [] (86)

The “purified” trial vector set is linearly independent of the baseline system mode
shapes in a manner similar to MacNeal’s residual vectors, as follows:
    
  MO OL =  T IOL − MO OL TOL
T

    (87)
[MO OL ] =  T MO OL − MO OL TOL MO OL ≡ [0]

    
  KO OL =  T IOL − MO OL TOL [KO OL ]
T

 
=  T [KO OL − MO OL λOL ] ≡ [0].

While the “purified” trial vector set has the above property, it includes an
unnecessarily large number of vectors. An appropriate, substantially smaller set of
residual vectors is identified by singular value decomposition of the generalized
mass matrix.
 
[A] =   MO   ,
T
(88)

The singular value decomposition process involves solution of the eigenvalue

problem,
  
[A] φρ = φρ λρ λρ1 ≥ λρ2 ≥ λρ3 ≥ .. . . . (89)

The cut-off criterion, noted below employed to define suitable reduced trial
vector set, is

λρN
≤ tol = 10−N (where N ∼ 4 to 6 is usually adequate).
λρ1

The augmented trial vector set (replacing the reduction transformation of

Eq. 76) is
24 Aerospace Perspective for Modeling and Validation 1375

 
OL = OL   φρ . (90)

The form of the resulting Ritz, multi-parameter sensitivity model (associated

with selected values of the scaling parameters) is,
   

N 
N
kO + pi [ki ] [φ] − mO + pi [mi ] [φ] [λ] = [0], where (91)
i=1 i=1

     
T T T
[kO ] = OL KO OL , [mO ] = OL MO OL , [ki ] = OL Ki OL ,
 
T
[mi ] = OL Mi OL .

Recovery of mode shapes in terms of physical degrees-of-freedom is accom-

plished with,

[] = OL [φ]. (92)

Since its introduction in 2002 [27], RMA has exhibited the capability to
accurately follow modal sensitivity trends over an extremely wide range of para-
metric variation. The simple cantilevered (planar) beam example, provided in
Fig. 12, demonstrates typical RMA performance (“100%” is baseline). Actual cross-
orthogonality checks are also excellent.
Additional recent advances and evaluations of the RMA method are found in
Reference [31].

Support EI1 EI2 EI3

Baseline: EI is uniform

50% EI2 Loss 99.9% EI2 Loss

Mode Baseline Exact Approx Exact Approx
1 28.78 28.47 28.47 5.71 5.71
2 180.39 172.40 172.40 79.86 79.86
3 467.65 467.65 467.65 320.70 320.70
4 505.11 504.82 504.82 467.65 467.65
5 989.87 950.46 950.46 733.08 733.08
6 1400.30 1400.30 1400.30 870.15 870.15
7 1636.50 1633.60 1633.60 1400.30 1400.30
8 2325.20 2325.20 2325.20 2042.60 2042.60
9 2445.00 2357.10 2357.10 2170.70 2170.70
10 3237.10 3237.10 3237.10 2325.20 2325.20
11 3416.00 3403.70 3403.70 3237.10 3237.10
12 4130.80 4130.80 4130.80 4017.60 4017.60
13 4550.00 4403.90 4403.90 4130.80 4130.80
14 5001.50 5001.50 5001.50 4201.20 4201.20
15 5844.20 5813.10 5813.10 5001.50 5001.50
16 5848.10 5844.20 5844.20 5844.20 5844.20
17 6654.20 6654.20 6654.20 6654.20 6654.20
18 7311.70 7104.70 7104.70 6668.30 6668.30
19 7427.00 7427.00 7427.00 6935.70 6935.70
20 8158.30 8158.30 8158.30 7427.00 7427.00

Fig. 12 RMA sensitivity performance for a cantilevered beam example

1376 R. Coppolino

5 Concluding Remarks

The aerospace perspective for modeling and validation, particularly in the U.S., is
primarily the product of (1) developments and applications in predictive structural
dynamics and (2) evolution of an integrated test-analysis process that focuses on
the goal of correlation and reconciliation of predicted and measured normal modes
of structures. Practices in predictive structural dynamics employ finite element
analysis (FEM) and matrix structural analysis methodologies that focus on the
assumption of linear behavior of launch vehicles and their spacecraft payloads. The
important area of aircraft systems, which is not covered in this chapter, incorporates
unique issues associated with the coupling of structural dynamics and aerodynamics
(aeroelasticity). The present state of the practice has been heavily influenced by
developments in digital computer hardware and computer aided analysis (CAE) over
the past 60 years.
Currently accepted best practices in the U.S. aerospace industry include:

1. Definition of linear finite element models of major components (substruc-

tures also called “superelements”) that faithfully address the frequency band
(0 ≤ f ≤ f*) and intensity of dynamic environments.
2. Description of substructures in terms of component mode models, primarily
employing the Hurty-Craig-Bampton (HCB) method. The MacNeal-Rubin (MR)
method is rarely employed.
3. Assembly of a launch vehicle/payload system dynamic model and computation
of assembled system normal modes based on the HCB or the Benfield-Hruda
(BH) component mode synthesis method. The MacNeal-Rubin (MR) method is
rarely employed.
4. Incorporation of local nonlinear model features in the launch vehicle/payload
system dynamic model, when deemed necessary. Transient dynamic response
strategies developed during the Space Shuttle program are particularly relevant;
the MR method was employed in these situations.
5. Estimation of detailed structural dynamic stresses and member loads by the
mode acceleration method, which (a) accounts for the quasi-static response of all
modes above f* and (b) provides a systematic account of “static” and “modal”
contributions to stresses and loads. Employment of MacNeal-Rubin residual
“modes” is an alternative to the mode acceleration method that satisfies “a” but
does not address “b.”
6. Verification and validation of component structural dynamic models employing
a systematic integrated test-analysis process (ITAP) that include modal test
planning, measured data analysis, experimental modal analysis, and test-analysis
correlation and model reconciliation.
7. Application of the residual kinetic energy (RKE) method and extensions of
RKE for allocation of accelerometer resources for modal testing and test-
analysis model (TAM) mass matrix definition. The TAM mass matrix is essential
for conventional test mode orthogonality and test-analysis cross-orthogonality
evaluations.
24 Aerospace Perspective for Modeling and Validation 1377

8. Adherence to NASA STD-5002 or SMC-C-004, which specifies minimum stan-

dards for test mode orthogonality, test-analysis cross-orthogonality, and modal
frequency correspondence criteria to assure acceptable verified and validated
post-test component structural dynamic models.
9. Utilization of residual mode augmentation (RMA) to enhance accuracy and
reliability of structural dynamics modification (SDM) for (large multiple param-
eter variation) sensitivity analyses to effect component model correlation and
reconciliation.

Many other excellent analytical and experimental tools are employed in structural
dynamic model verification and validation, which are not uniformly accepted and
standardized in U.S. aerospace community practice. It is hoped that accepted prac-
tices and standards will become more comprehensive with continuing experiences
and unfolding innovations in this community.

References
1. Coppolino R (2014) Variational foundations of modern structural dynamics. IMAC XXXII
2. Hamilton WR (1835) On a general method in dynamics. Phil Trans R Soc
3. Lagrange JL (1788) Mechanique Analitique, Paris
4. Ritz W (1909) Über eine neue Methode zur Lösung gewisser Variationsprobleme der mathe-
matischen Physik. Journal für die Reine und Angewandte Mathematik
5. Toupin RA (1952) A variational principle for the mesh-type analysis of a mechanical system.
J Appl Mech 74
6. Trefftz E (1926) Ein Gegenstuck zum Ritzschen Verfahren. In: Proceedings of the 2nd
International Congress of Applied Mechanics
7. Galerkin BG (1915) . . . Some questions of elastic equilibrium of rods and plates. Vestnik
inzhenerov i tekhnikov 19
8. Coppolino RN (1975) A numerically efficient finite element hydroelastic analysis, vol 1: theory
and results, NASA CR-2662, April
9. Zienkiewicz OC, Taylor RL, Zhu JZ (2005) The finite element method, its basis and
fundamentals, 6th edn. Elsevier
10. Coppolino R (2010) Chapter 23: finite element methods of analysis. In: Piersol A, Paez T (eds)
Harris’ shock and vibration handbook, 6th edn. McGraw-Hill
11. Independent structural loads analysis, Air Force Space Command, SMC-S-004 (2008)
12. Load analyses of spacecraft and payloads, NASA-STD-5002 (1996)
13. Paez T (2010) Chapter 8: transient response to step and pulse functions. In: Piersol A, Paez T
(eds) Harris’ shock and vibration handbook, 6th edn. McGraw-Hill
14. Pilkey W, Pilkey D (eds) (2008) Peterson’s stress concentration factors, 3rd edn. Wiley
15. Coppolino R (2016) Structural dynamics modeling – tales of sin and redemption. Sound and
Vibration Magazine, January
16. Guyan R (1965) Reduction of stiffness and mass matrices. AIAA J 3
17. Hurty W (1965) Dynamic analysis of structural systems using component modes. AIAA J 3(4)
18. Craig R, Bampton M (1968) Coupling of substructures for dynamic analysis. AIAA J 6(7)
19. Benfield W, Hruda R (1971) Vibration analysis of structures by component mode substitution.
AIAA J 9(7)
20. MacNeal RH (1971) A hybrid method for component mode synthesis. Comput Struct 1
21. Rubin S (1975) Improved component-mode representation for structural dynamic analysis.
AIAA J 13(8)
1378 R. Coppolino

22. Coppolino R (1978) Employment of residual mode effects in vehicle/payload dynamic loads
analysis, Government/Industry Workshop on Payload Loads Technology, NASA CP-2075
23. Williams D (1945) Dynamic loads on aeroplanes under given impulsive load with particular
reference to landing and gust loads on a large flying boat, Great Britain Royal Aircraft
Establishment Reports SME 3309 and 3316
24. Coppolino RN, Borowski VJ (1997) Integrated modal testing and analysis of a body-on-frame
automobile. In: Proceedings of the 15th IMAC
25. Chen JC, Garba JC (1985) Structural analysis model validation using modal test data. ASME,
AMD 67:109–137
26. Coppolino RN (1998) Automated response DOF selection for mapping of experimental normal
modes. In: Proceedings of the 16th IMAC
27. Coppolino RN (2002) International Space Station P5 Modal Survey: test planning through
FEM reconciliation. In: Proceedings of the 20th IMAC
28. Load analyses of spacecraft and payloads, NASA-STD-5002 (1996)
29. Independent structural loads analysis, air force space command, SMC-S-004 (2008)
30. Coppolino RN (1985) Hybrid experimental/analytical dynamic models of aerospace structures.
ASME, AMD 67:79–108
31. Coppolino RN (2018) Methodologies for verification and validation of Space Launch System
(SLS) structural dynamic models, NASA CR-2018-219800, vol 1
32. Fox RI, Kapoor MP (1968) Rates of change of eigenvalues and eigenvectors. AIAA J 6
33. Nelson RB (1976) Simplified calculation of eigenvector derivatives. AIAA J 14
34. Avitabile P (2002) Twenty years of structural dynamic modification-a review. In: 20th IMAC
35. Coppolino R (1981) Structural mode sensitivity to local modification, SAE Paper 811044
Applied Math for Experimental Structural
Dynamics 25
Chuck Van Karsen and Andrew Barnard

Contents
1 Domains and Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1380
1.1 Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1383
2 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1385
2.1 Basic Concepts and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1385
2.2 Transposition Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1389
2.3 Special Matrix Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1389
2.4 Symmetric Matrix Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1390
2.5 Matrix Measures (Determinant and Trace) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1391
2.6 Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1392
2.7 Vector Space Applied to Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1392
2.8 Spectral Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1392
2.9 Singular Value Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1393
2.10 Eigen Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1394
2.11 Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1394
2.12 LU Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1395
2.13 Moore-Penrose Generalized Inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1399
3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1400
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1401

Abstract

Explanation of the methods, techniques, and theoretical aspects of experi-

mental structural dynamics requires a fundamental understanding of several
mathematical concepts. Data acquisition and analysis and their relationship to
structural dynamics theory rely on an understanding of domains and transforms.

C. Van Karsen () · A. Barnard

Mechanical Engineering-Engineering Mechanics Department, Michigan Technological
University, Houghton, MI, USA
e-mail: [email protected]; [email protected]

© The Society for Experimental Mechanics 2022 1379

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0_32
1380 C. Van Karsen and A. Barnard

Theoretical and computational methods in structural dynamics (modeling, cor-

relation, parameter estimation, etc.) make use of linear algebra concepts and
techniques. This chapter presents the fundamentals of these concepts and
methods.

Keywords

Frequency response function · Transfer function · Fourier transform · Laplace

transform · Impulse response function · Linear algebra · Matrix mathematics ·
Matrix · Singular value decomposition · Eigen problem · Eigen solution ·
Eigenvalue · Eigenvector · Matrix inverse

Acronyms

ω Frequency, radians/second
f Frequency, Hertz (cycles/second)
t time, seconds
s Laplace
√ variable, radians/sec
j −1
an , bn Fourier Series coefficients
cn Fourier Series coefficients (complex form)
θ Angle, radians
[A ]Matrix
{x} Vector
ẋ First time derivative of x
ẍ Second time derivative of x
[A nxm ]Matrix consisting of n rows and m columns
σ Eigenvalue

1 Domains and Transforms

Experimental structural dynamics uses three primary domains to communicate

information, system properties, and characteristics of results (predicted or mea-
sured). These are the time , frequency, and Laplace domains. These domains can
be thought of as coordinate systems where the choice of coordinate system depends
on the characteristics of the quantity that is being described or communicated. As
an example consider the following:
Suppose it is necessary to communicate the geometry of a simple cylinder by
describing points in 3D space that lie on the surface of the cylinder, Fig. 1. The
easiest way to accomplish this would be to use a cylindrical coordinate system.
The same points could also be described using a Cartesian coordinate system
as well, although it may be harder to visualize. If the point coordinates exist in
cylindrical coordinates, they can be changed to Cartesian by using a coordinate
transformation, a set of equations which relate (R, θ, Z) to (X, Y, Z). Changing
back to cylindrical coordinates is accomplished by using the inverse coordinate
25 Applied Math for Experimental Structural Dynamics 1381

Fig. 1 Coordinate system example

Fig. 2 Domains used in experimental structural dynamics

transformation. Information is not gained or lost with this process; it is only

presented in a form that is best suited for the application.
The primary domains (coordinate systems) that are used in experimental struc-
tural dynamics are shown in Fig. 2. Experimental data is acquired in the time
domain and is observed in terms of events. The fundamental laws of physics which
govern structural dynamics are expressed as differential equations with time as the
independent variable. Many simulation tools produce results in the time domain as
well. Time (t) is typically expressed in units of seconds.
1382 C. Van Karsen and A. Barnard

In many cases it is necessary to determine the periodic nature (frequency content)

of data. This is accomplished by transforming from the time to the frequency domain
using the Fourier transform. The frequency domain is complex valued having both
real (cosine) and imaginary (sine) components. Information can be transformed
back and forth between the time and frequency domains without any gain or loss of
information. Frequency (ω) is expressed in units of radians per second. Alternatively
frequency (f ) can be expressed in units of Hertz (cycles per second) ω = 2π f.
The Laplace domain is used to observe a system or structure’s characteristics in
terms of poles and zeros or poles and residues. This is accomplished by transforming
from the time to the Laplace domain using the Laplace transform. The Laplace
domain is complex valued, having both real and imaginary components. Information
can be transformed back and forth between the time and Laplace domains without
any gain or loss of information. The Laplace variable (s = σ + jω) is expressed
in units of radians per second and is complex valued having a real part σ , and an
imaginary part ω.
An important relationship in structural dynamics is the ratio between output of
a system or structure to an input to the system or structure. This relationship is
expressed as an impulse response function in the time domain, a frequency response
function in the frequency domain, and as a transfer function in the Laplace domain.
An example of this relationship for a single degree of freedom system is shown in
Fig. 3.

Fig. 3 SDOF input-output relationships in each domain

25 Applied Math for Experimental Structural Dynamics 1383

The mathematical descriptions of these functions in pole residue form for a single
degree of freedom system are:

Impulse response function:

∗t
h(t) = Aeλt + A∗ eλ (1)

Frequency response function:

A A∗
H (ω) = + (2)
j ω − λ j ω − λ∗

Transfer function:

A A∗
H (s) = + (3)
s − λ s − λ∗

The parameters which are characteristics of the system’s input-output relation-

ship, A, λ, are seen in each domain’s mathematical description. It should be noted
that the Fourier transform of the impulse response function is the frequency response
function and the Laplace transform of the impulse response function results in the
transfer function.

1.1 Frequency Domain

Transformation between the time and frequency domain can be accomplished in one
of three ways:

1. Fourier series
2. Integral Fourier transform
3. Discrete Fourier transform (DFT)

The choice of method depends on the nature and form of the time domain
information (data). The Fourier series and integral Fourier transform are used when
continuous time histories are available. The DFT is used for discrete (sampled) time
histories.
Any time history that is periodic with respect to time, Fig. 4, can be expressed by
the following trigonometric function which is known as a Fourier series (In addition
the time history must have a finite number of discontinuities within any period, it
must have a finite number of maxima and minima within any period, and it must be
absolutely integrable over any period. These are known as the Dirichlet conditions.):
∞
a0 
X(t) = + an cos (2π nf t) + bn sin (2π nf t) (4)
2
n=1
1384 C. Van Karsen and A. Barnard

1.5
X (t )
1

0.5
Amplitude

–0.5

–1
t+T
t
–1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time - seconds
X (t ) = X (t + T )

Fig. 4 Periodic time history

where:

1
f =
T
 t+T
2
an = X(t) cos (2π nf t) dt
T t
 t+T
2
bn = X(t) sin (2π nf t) dt
T t
The an and bn coefficients provide a description of the information (data) in the
frequency domain. The an ’s represent frequency content that is co-sinusoidal with
respect to time t, and the bn ’s represent frequency content that is sinusoidal with
respect to time t.
Using Euler’s identity, e±iθ = cos θ ± i sin θ , the trigonometric form of the
Fourier series can be written in complex exponential form:

∞

X(t) = cn ei2π nf t (5)
n=−∞

where
 t+T
1
cn = X(t)e−i2π nf t dt for n = ± (1, 2, 3, . . . )
T t

c0 = a0

 2 1
a + bn 2
2
cn = n for n = ± (1, 2, 3, ..)
2

The cn ’s are complex valued with the real part representing frequency content
that is co-sinusoidal with respect to time t, and the imaginary part representing
frequency content is sinusoidal with respect to time t.
25 Applied Math for Experimental Structural Dynamics 1385

In either formulation the Fourier series is a simple concise method for expressing
the periodic properties of time information in the frequency domain and considering
experimental data or the results of a simulation all that is required is knowledge of
the period (T).

1.1.1 Integral Fourier Transform

There are many situations where the time domain data is not periodic or has a finite
length with no period. In order to transform these types of data to the frequency
domain, it is necessary to use the integral Fourier transform.
 ∞
X(f ) = X(t)e−i2πf t dt time to frequency
−∞
 ∞
(6)
X(t) = X(f )e i2πf t
df frequency to time
−∞

For a given set of time data X(t), the integral Fourier transform will determine
the amount of frequency content at any frequency f. The only issue is the limits on
the integration. If the integration time is truncated, the frequency domain result will
be incorrect. An example of this is shown in Fig. 5.
For other properties of the integral Fourier transform, see references.
Implementation and limitations of the integral Fourier transform on sampled data
is presented in  Chap. 4, “Applied Digital Signal Processing.”

2 Linear Algebra

2.1 Basic Concepts and Definitions

The analysis of information in both the test and analytical aspects of structural
dynamics requires the efficient manipulation and interrogation of large data sets,
and therefore the use of matrix and vector algebra is a mandatory requirement. In
this context there are fundamentally two types of problems to be solved:

Systems described by coupled algebraic expressions:

[A] {x} = {b} (7)

And systems described by coupled ordinary differential equations:

[M] {ẍ} + [C] {ẋ} + [K] {x} = {F (t)} (8)

1386 C. Van Karsen and A. Barnard

a. without truncation
1
0.5
0
–0.5
–1
t
time

0.8

0.6

0.4

0.2

frequency
–1 1
t t

b. with truncation

0.5

–0.5

–1

time

1
0.8
0.6
0.4
0.2
0
–0.2
frequency

Fig. 5 Integral Fourier transform ((a) without truncation, (b) with truncation)
25 Applied Math for Experimental Structural Dynamics 1387

For systems described by matrix expressions, there are three cases which can
occur:

[A]nxm {x}mx1 = {b}nx1 (9)

1. Underdetermined: n < m applications include optimization, finite element model

updating, and projection of data onto subspaces.
2. Determined: n = m applications include analytical structural dynamics using
finite element models.
3. Overdetermined: n > m applications include time and frequency domain param-
eter estimation, least squares applications, and sensor placement algorithms.

Two basic entities are used in the language of linear algebra, a matrix and a
vector. A matrix is an array of numbers or expressions. The elements of the matrix
are referred to by their row/column location in the matrix.
⎡ ⎤
a11 a12 a13
[A] = ⎣ a21 a22 a23 ⎦ (10)
a31 a32 a33

A vector is a special case of a matrix which has only one row or column.
⎧ ⎫
⎪
⎪ b1 ⎪
⎨ ⎪ ⎬
b2
{b} = or c = c1 c2 c3 c4  (11)
⎪
⎪ b ⎪
⎩ 3⎪ ⎭
b4

A diagonal matrix has nonzero terms only on the diagonal:

⎡ ⎤
d11
⎢ d22 ⎥
[D] = ⎢
⎣
⎥ dnm = 0 for n = m
⎦ (12)
d33
d44

There are several elementary rules for matrix operations.

Matrix addition:

[C]ixj = [A]nxm + [B]pxq i = n = p j =m=q cij = anm + bpq

         
12 56 12 56 6 8
[A] = [B] = [C] = [A] + [B] = + =
34 78 34 78 10 12
(13)
1388 C. Van Karsen and A. Barnard

Multiplication by a scalar:
⎡ ⎤
ka 11 ka 12 ka 13
k [A] = ⎣ ka 21 ka 22 ka 23 ⎦ (14)
ka 31 ka 32 ka 33

Multiplication of two matrices:

[A] [B] = [B] [A] NOT commutative (15)

([A] [B]) [C] = [A] ([B] [C]) Associative (16)

([A] + [B]) ([C] + [D]) = [A] [C] + [A] [D] + [B] [C] + [B] [D] (17)

Matrix cancelation:

[A] [B] = [0] Implies one of the following;

[A] = [0]
(18)
[B] = [0]
[A] and [B] are singular

Matrix multiplication:

[A]nxm [B]pxq = [C]nxq requirement m = p

⎡ ⎤⎡ ⎤ ⎡ ⎤
a11 a12 a13 b11 b12 c11 c12
⎣ a21 a22 a23 ⎦ ⎣ b21 b22 ⎦ = ⎣ c21 c22 ⎦
a31 a32 a33 b31 b32 c31 c32
(19)
c21 = a21 b11 + a22 b21 + a23 b31

m
cij = aik bkj
k=1

Two special matrices are often used:

Identity matrix
⎡ ⎤
1000
⎢0 1 0 0⎥
[I ] = ⎢
⎣0 0 1 0⎦
⎥
(20)
0001

1 s on the diagonal, zeros elsewhere
25 Applied Math for Experimental Structural Dynamics 1389

Zero matrix
⎡ ⎤
000
[0] = ⎣ 0 0 0 ⎦ Zeros in all locations (21)
000

The transpose of a matrix is formed by interchanging rows and columns.

⎡ ⎤ ⎡ ⎤
123 147
[A] = ⎣ 4 5 6 ⎦ [A]T = ⎣ 2 5 8 ⎦ (22)
789 369

If complex numbers are entities of the matrix, a Hermitian transpose (or

Hermitian) is formed by interchanging rows and columns and then taking the
complex conjugate.
⎡ ⎤ ⎡ ⎤
1 − 2i 4 2 + 4i 1 + 2i 3 + 9i 3
[A] = ⎣ 3 − 9i 9 11 − i ⎦ [A]H =⎣ 4 9 1−i ⎦ (23)
3 1 + i −4 − 8i 2 − 4i 11 + i −4 + 8i

2.2 Transposition Rules

The following rules apply to matrix transposition :

([A] + [B])T = [A]T + [B]T

 T T
[A] = [A]
(24)
([A] [B])T = [B]T [A]T
([A] [B] [C])T = [C]T [B]T [A]T

2.3 Special Matrix Forms

Symmetric [A] = [A]T

Skew Symmetric [A] = − [A]T
⎡ ⎤ ⎡ ⎤
123 0 2 3
[A] = ⎣ 2 4 5 ⎦ [A] = ⎣ − 2 0 −5 ⎦ (25)
356 −3 5 0

Hermitian [A] = [A]H

Skew Hermitian [A] = − [A]H
1390 C. Van Karsen and A. Barnard

⎡ ⎤ ⎡ ⎤
1 6 + 5i 4i i 2 + 2i 8
[A] = ⎣ 6 − 5i 6 7 − 8i ⎦ [A] = ⎣ − (2 − 2i) 8i − (6 − 3i) ⎦
− 4i 7 + 8i 0 8 6 + 3i 0
Diagonal terms are real Diagonal terms are imaginary or zero
(26)

2.4 Symmetric Matrix Rules

The following rules apply to symmetric matrices:

[A] = [A]T ; [B] = [B]T ; [A] [B] = ([A] [B])T

(27)
[A] = [A]T ; [C] = [B]T [A] [B] ; [C] = [C]T

Orthogonal [A] [A]T = [A]T [A] = [I ]

Unitary [A] [A]H = [A]H [A] = [I ]

(28)
Idempotent [A] = [A] any positive integer m
m

Nilpotent [A]k = [0] any positive integer k

For a Toeplitz matrix, all elements on any diagonal are equal. This matrix is often
used in time domain identification techniques.
⎡ ⎤
t1 t5 t6 t7
⎢ t2 t1 t5 t6 ⎥
[T ] = ⎢
⎣ t3
⎥ (29)
t2 t1 t5 ⎦
t4 t3 t2 t1

A Toeplitz matrix is useful in computing the convolution of discrete data.

 k
y(k) = H (k)u (k − i)
⎡ i=0 ⎤⎧ ⎫ ⎧ ⎫
H0 H1 H2 H3 H4 ⎪ ⎪ ⎪
⎪ u4 ⎪ ⎪ y4 ⎪
⎪
⎢ 0 H H H H ⎥⎪ ⎪u ⎪ ⎪ ⎪
⎪y ⎪ ⎪
⎢ 0 1 2 3⎥⎪ ⎨ 3⎪ ⎬ ⎪
⎨ 3⎪ ⎬ (30)
⎢ ⎥
⎢ 0 0 H0 H1 H2 ⎥ u2 = y2
⎢ ⎥⎪ ⎪ ⎪ ⎪
⎣ 0 0 0 H0 H1 ⎦ ⎪ ⎪
⎪ u1 ⎪
⎪
⎪
⎪
⎪
⎪ y1 ⎪
⎪
⎪
⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
0 0 0 0 H0 u0 y0 ⎭

The Hankel matrix has all elements on any diagonal perpendicular to the main
diagonal equal to each other. This matrix is used in the Eigensystem realization
algorithm (ERA) modal parameter estimation method.
25 Applied Math for Experimental Structural Dynamics 1391

⎡ ⎤
h1 h2 h3 h4
⎢ h2 h3 h4 h5 ⎥
[H ] = ⎢
⎣ h3
⎥ (31)
h4 h5 h6 ⎦
h4 h5 h6 h7

In a Vandermonde matrix, the first column is 1, and successive columns are the
second column raised to increasing integer powers.

⎡ ⎤
1 v1 v12 v13
⎢1 v22 v23 ⎥
⎢ v2 ⎥
⎢ ⎥
[V ] = ⎢ 1 v3 v32 v33 ⎥ (32)
⎢ ⎥
⎣1 v4 v42 v43 ⎦
1 v5 v52 v53

This matrix occurs in curve fitting and some frequency domain parameter
estimation algorithms.

2.5 Matrix Measures (Determinant and Trace)

There are two common matrix measures: determinant and trace. A determinant is
only defined for square matrices and is equal to zero if the matrix is singular.

 
 a a 
|[A]| =  11 12  = a11 a22 − a12 a21
a21 a22
⎛ ⎞
 a11 a12 a13       
   a22 a23   a21 a23   a21 a22 
 
|[A]| = ⎝ a21 a22 a23 ⎠ = a11   
− a12   
+ a13  
 a a a  a32 a33  a31 a33  a31 a32 
31 32 33
     
  
|[A] [B]| = |[A]| |[B]| |[A]| = [A]T  [A]∗  = [A]H 
 
|k [A]| = k n [A] for [A]nxn
where ∗ indicates the complex conjugate
(33)

The trace is simply the sum of the diagonal elements of a square matrix.


n
tr ([A]) = aii (34)
i=1
1392 C. Van Karsen and A. Barnard

2.6 Vector Space

The set of all n dimensional vectors make up an n-dimensional vector space, Rn . n

vectors {e}i in Rn are considered as linearly independent if the equation:

0 = a1 {e}1 + a2 {e}2 + a3 {e}3 . . . an {e}n (35)

only has a solution when all constants ai are zero. In this case the n vectors {e}i
span the vector space. Any n-dimensional vector {x} can be expressed uniquely as a
linear combination of the n linearly independent vectors {e}i :

{x} = b1 {e}1 + b2 {e}2 + b3 {e}3 . . . bn {e}n (36)

2.7 Vector Space Applied to Matrices

Consider this matrix equation, [A]nxm {x}mx1 = {b}nx1 . The rank of a matrix is the
number of linearly independent columns or rows rank([A]) = r.

rank ([A]) = m [A] is full column rank .

rank ([A]) = r < m, n [A] is rank deficient.If [A] is square, then it is singular.
(37)

Column space of [A] is the vector space spanned by its columns.

Row space of [A] is the vector space spanned by its rows.
The null space of [A] is the set of vectors {x} that satisfies [A]{x} = {0}.

2.8 Spectral Decomposition

In most cases a square matrix [A] can be decomposed into a product of three
matrices:

[A] = [] [] []−1 (38)

where [] is a diagonal matrix of eigenvalues λi and the matrix [] is made up of
column vectors {φ}i called eigenvectors. This means that [A] can be diagonalized
[]−1 [A][] = []. Eigenvalues and eigenvectors satisfy the eigenvalue problem
([A] − λi [I]){φ}i , and eigenvalues satisfy the characteristic equation |([A] −
λi [I])| = 0.
A matrix [A] is diagonalizable if it has n linearly independent eigenvectors. If [A]
has distinct eigenvalues, it is diagonalizable.
25 Applied Math for Experimental Structural Dynamics 1393

 −1     
φ11 φ12 a11 a12 φ11 φ12 λ1 0
= (39)
φ21 φ22 a21 a22 φ21 φ22 0 λ2

Real symmetric matrices are always diagonalizable and have real eigenvalues
and eigenvectors. The eigenvector matrix is orthogonal:

[]T = []−1 ∴ []T [A] [] = [] (40)

2.9 Singular Value Decomposition

Any n x m matrix [A] of rank r can be decomposed into a product of three matrices:

[U ]nxn
[A] = [U ] [S] [V ]H [S]nxm
(41)
[V ]mxm
[U ] and [V ] are unitary [U ]H [U ] = [V ]H [V ] = [I ]

These matrices can be partitioned such that:

  
  [ ] [0] [V1 ]H
[A] = [U1 ] [U2 ] [U1 ]nxr [U2 ]nx(n−r)
[0] [0] [V2 ]H
(42)
[ ] is diagonal [V1 ]mxr [V2 ]mx(m−r)

As a result:


r
[A] = [U1 ] [ ] [V1 ]H ⇒ [A] = {u1 }i σi {v1 }H
i (43)
i=1

Singular values are real and satisfy the following relationship:

⎡ ⎤
σ1 0 0 ···
⎢0 σ2 0 · · · ⎥
⎢ ⎥
[ ]rxr =⎢
⎢0 .. ⎥ σ1 ≥ σ2 ≥ · · · ≥ σr > 0 (44)
⎣ 0 . ···⎥
⎦
.. .. ..
. . . σr

The number of nonzero singular values is the rank, r, of [A].

1394 C. Van Karsen and A. Barnard

σi2 , {u1 }i are the eigenvalues and eigenvectors of [A] [A]T

{u2 }i are the eigenvectors of [A] [A]T with zero eigenvalues
σi2 , {v1 }i are the eigenvalues and eigenvectors of [A]T [A]
{v2 }i are the eigenvectors of [A]T [A] with zero eigenvalues
(45)
[U1 ] spans the column space of [A]
[U2 ] spans the column null space of [A]
[V1 ]T spans the row space of [A]
[V2 ]T spans the row null space of [A]

2.10 Eigen Solutions

Two, square, nonsingular, symmetric matrices can be decomposed into their associ-
ated eigenvalues and eigenvectors.

[[A] − λ [B]]nxn {X} = {0}

λ1 , λ2 , · · · λn are the eigenvalues, (46)
and {x}1 , {x}2 , · · · {x}n are the associated eigenvectors

2.11 Inverse Problems

2.11.1 Solution of Determined Equations

When a system of equations has as many independent equations as unknowns, a
unique solution will always exist.

[A]nxn {x}nx1 = {b}nx1 [A] is square, full rank and invertible

{x} = [A]−1 {b} {b} is alwys in the range space of [A] (47)
[A]−1 is the matrix inverse of [A]

The inverse of a nonsingular matrix [A] is a matrix [A]−1 that when multiplied
by [A] is the identity matrix.

[A] [A]−1 = [A]−1 [A] = [I ] (48)

25 Applied Math for Experimental Structural Dynamics 1395

Matrix inverses have the following properties.

([A] [B])−1 = [B]−1 [A]−1

!−1 !T
[A]T = [A]−1
!−1 !H
[A]H = [A]−1 (49)
1
(k [A])−1 = [A]−1
k
 
 −1  1
[A]  =
|[A]|

There are many methods for inverting a matrix. One that is particularly relevant
in the field of structural dynamics is the adjoint method. The adjoint of [A] is a
matrix where the elements are the cofactors of [A] transposed. The cofactors of [A]
are the signed minor determinants.
 
−1 Adjoint ([A])
[A] = (50)
|[A]|
 
Cofactor([A])ij = cij = (−1)i+j  Mij 
 
Mij is a submatrix of [A] obtained by deleting the ith row and j th column
(51)
⎡      ⎤
 e f  b c b c 
     
⎢ +h i  −h i  +e f ⎥
⎢ 
⎢      ⎥⎥
⎢  d f   a c   a c  ⎥
⎢ − + − ⎥
⎢ g i   g i   d f  ⎥
⎢  ⎥
⎡ ⎤ ⎣  d e   a b   a b  ⎦
+ − +
ab c
adjoint ([A]) g h g h d e 
[A] = ⎣ d e f ⎦ ⇒ [A]−1 = =
|[A]| |[A]|
gh i
(52)

2.12 LU Decomposition

LU decomposition can be used to calculate determinants and matrix inverses, and

therefore it is useful in the solution of determined equations. Any nonsingular square
matrix [A] can be factored into the product of two matrices:

[A] = [L] [U ] (53)

1396 C. Van Karsen and A. Barnard

[L] is a lower triangular matrix, and [U] is an upper triangular matrix.

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
6 −2 −4 4 2 0 0 0 3 −1 −2 2
⎢ 3 −3 −6 1 ⎥ ⎢ 1 −1 0 0⎥ ⎢0 2 4 1 ⎥
[A] = ⎢
⎣ − 12
⎥ ⇒ [L] = ⎢
⎦ ⎣
⎥ [U ] = ⎢ ⎥
8 21 −8 −4 2 1 0⎦ ⎣0 0 5 −2 ⎦
−6 0 −10 7 − 2 −1 −2 2 0 0 0 4
(54)

In terms of LU decomposition, the matrix inverse is:

[A]−1 = ([L] [U ])−1 = [U ]−1 [L]−1 (55)

The determinant becomes the product of the determinants of the decomposition

matrices.

|[A]| = |[L]| |[U ]| (56)

The solution to a set of determined equations can be found algebraically without

the calculation of an inverse.

[A] {x} = {b} ⇒ [L] [U ] {x} = {b}

[L] {z} = {b} solved using forward substitution (57)
[U ] {x} = {z} solved using backward substitution

2.12.1 Solution of Underdetermined Equations

An underdetermined condition exists when the number of independent equations is
less than the number of unknowns.

[A]nxm {x}mx1 = {b}nx1 n < m (58)

[A] is rectangular with more columns than rows and is assumed to be full-row
rank. A solution will always exist, but there are an infinite number of solutions. {b}
is always in the column space of [A]. If {x}1 is a solution and {x}N is any vector
in the null space of [A], then the linear combination, {x}s = {x}1 + {x}N , is also a
solution.
A solution is obtained in the following manner. Begin by defining a new vector
{x} = [A]T {z}. The original matrix equation becomes:

[A] [A]T {z} = {b} (59)

25 Applied Math for Experimental Structural Dynamics 1397

Since [A] is full-row rank, [A][A]T is nonsingular and can be inverted.

Solving for{x}:

!−1
{z} = [A] [A]T {b}
!−1
{x}+ = [A]T {z} = [A]T [A] [A]T {b} (60)
!−1
[A]+ = [A]T [A] [A]T Right Generalized Inverse of [A]

Since there are an infinite number of solutions, which solution is {x}+ , for any
arbitrary solution {x}s :

{b} = [A] {x}s

!−1
{x}+ = [A]T [A] [A]T [A] {x}s = [P ] {x}s
!−1 (61)
[P ] = [A]T [A] [A]T [A]

{x}+ = [P ] {x}s

{x}+ is the orthogonal projection of the general solution {x}s onto the row space
[A] or the column space of [A]T . {x}+ is the minimum norm (length) solution, and
[P] is an orthogonal projector.

2.12.2 Solution of Overdetermined Equations

An overdetermined condition exists when the number of independent equations is
greater than the number of unknowns.

[A]nxm {x}mx1 = {b}nx1 n > m (62)

[A] is rectangular with more rows than columns and is assumed to be full-column
rank. {b} may or may not be in the column space of [A]. If {b} is in the row space
of [A], a unique solution exists. If {b} is not in the row space of [A], no exact
solution exists, but an approximate solution can be found. A solution is obtained
in the following manner. Begin by pre-multiplying by the transpose of [A]:

!
[A]T [A] {x} = [A]T {b} (63)
1398 C. Van Karsen and A. Barnard

Since [A] is full-column rank, [A][A]T is nonsingular and can be inverted.

Solving for {x}:

 −1
{x} = [A]T [A] [A]T {b}
 −1 (64)
[A]+ = [A]T [A] [A]T Left Generalized Inverse of [A]

Combining this solution with the original equation:

!−1 " # $ %
[A] {X} = [A] [A]T [A] [A]T {b} = P̂ {b} = b̂ (65)

" # $ %
P̂ is an orthogonal projector onto the row space of [A]. b̂ is the orthogonal
projection of {b} onto the row space of [A].
If {b} is in the row space of [A], {b} ∈ R[A]

 T −1 T
{x}
" #= [A] [A] [A] {b}
P̂ {b} = {b} (66)
{x} is a unique solution to [A] {x} = {b}

If {b} ∈ R[A]

& '  T −1 T

"x̂ #= [A]$ [A]
% [A] {b}
P̂ {b} = b̂ (67)
& ' $ %
x̂ is a unique solution to [A] {x} = b̂

& '
x̂ minimizes the Euclidean norm(length) of error vector {e}.

$ %
{e} = {b} − b̂ (68)

This solution technique is normally referred to as least squares (error) solution.

Direct inversion of the normal form ([A]T [A]), however, is normally not recom-
mended to obtain the least squares solution because it is costly (computationally)
and inaccurate. LU decomposition of ([A]T [A]) and backward/forward substitution
is faster but can also be inaccurate.
Another technique for the least squares solution which is both fast and accurate
is a method which uses QR decomposition (also referred to as QR factorization).
Beginning with (62), [A] is uniquely factored into the following form:
25 Applied Math for Experimental Structural Dynamics 1399

[A]nxm = [Q]nxn [R]nxm

[Q]T [Q] = [I ]
(69)
 
[U ]
[R] = upper triangular
[0]

Pre-multiplication by [Q]T gives:

[Q] [R] {x} = {b}

(70)
[Q]T [Q] [R] {x} = [R] {x} = [Q]T {b}

Partitioning and backward substitution yields the final solution{x}:

   
[U ] [Q1 ]T
{x} = {b}
[0] [Q2 ]T (71)
[U ] {x} = [Q1 ] {b}
T

2.13 Moore-Penrose Generalized Inverse

Every matrix [A]nxm of rank r has an mxn [A]+ matrix called a generalized inverse
if it satisfies the following relationships:

[A] [A]+ [A] = [A]

(72)
[A]+ [A] [A]+ = [A]+

There are an infinite number of matrices [A]+ which can satisfy these conditions.
If in addition [A]+ also satisfies:
 H
[A]+ [A] = [A]+ [A]
 H (73)
[A] [A]+ = [A] [A]+

Then the inverse is unique and is called the Moore-Penrose generalized inverse
[A]† . The M-P inverse is computed using singular value decomposition. Assuming
[A] has rank r.
  
  [ ]rxr [0] [V1 ]H
[A]nxm {x}mx1 = {b}nx1 ⇒ SVD ⇒ [A] = [U1 ] [U2 ]
[0] [0] [V2 ]H
[A] = [U1 ] [ ] [V1 ]H
(74)
1400 C. Van Karsen and A. Barnard

The Moore-Penrose inverse is:

⎡ ⎤
σ1−1 0 0 · · ·
⎢ 0 σ −1 0 · · · ⎥
⎢ 2 ⎥
[A]† [V1 ] [ ]−1 [U1 ]H [ ]−1 =⎢
⎢ 0 0 ... · · · ⎥
⎥ (75)
⎣ ⎦
.. .. .. −1
. . . σr

SVD is computationally costly, but it is accurate and stable. The nature of [A]
may make it difficult to determine the value of r.
The Moore-Penrose generalized inverse can solve matrix equations for any case
regardless of the size and rank of [A].
& '
[A]nxm {x}mx1 = {b}nx1 ⇒ x̂ = [A]† {b} (76)

Case 1: Minimum norm: n < m and rank([A]) = n.

!−1
[A]† = [A]T [A] [A]T (77)

Case 2: Least squares: n > m and rank([A]) = m.

!−1
[A]† = [A]T [A] [A]T (78)

Case 3: General or rank deficient case: r = rank ([A]) < min (n, m).

[A]† = [V1 ] [ ]−1 [U1 ]H (79)

3 Summary

This chapter presented the details of mathematical methods discussed in other

chapters of this book. When approaching structural dynamics problems, it is
imperative to understand concepts of transforms and linear algebra to best approach
a solution to a problem. Structural dynamics relies heavily upon mastery of the
concepts presented in this chapter, some of which are subtle. Mathematics is
the language and tool of the structural dynamics engineer. Using this chapter as
a reference for the other chapters in this book will enable the reader to fully
understand how structural dynamics methods are implemented and why certain
operations are used.
25 Applied Math for Experimental Structural Dynamics 1401

References
1. Bendat JS, Piersol AG (2000) Random data analysis and measurement procedures, 3rd edn.
ISBN: 0471-31733-0
2. Golub GH, Van Loan CF (2012) Matrix computations, 4th edn. Johns Hopkins University Press.
ISBN-13: 978-1421407944
3. Hayek SI (2001) Advanced mathematical methods in science and engineering, 1st edn. ISBN:
0-8247-0466-5
4. Kisak PF (ed) (2016) Fourier analysis: “decomposing a waveform into a harmonic series”, 1st
edn. CreateSpace Independent Publishing Platform. ISBN-13: 978-1523323913
5. Kreyzig E (2004) Advanced engineering mathematics, 8th edn. ISBN: 9971-51-283-1
6. Oppenheim AV, Schafer RW (1999) Discrete-time signal processing, 2nd edn. Prentice Hall.
ISBN: 0-13-754920-2
7. Spiegel MF (1974) Schaum’s outline of Fourier analysis with applications to boundary value
problems, 1st edn. McGraw-Hill Education. ISBN-13: 978-0070602199
8. Strang G (2016) Introduction to linear algebra, 5th edn. Wellesley-Cambridge Press. ISBN-13:
978-0980232776
Index

A Angular phase difference, 745

Absorption coefficient, 736, 737 Angular wavenumber, 721, 727
Accelerance, 747 Anti-alias filter ring, 100
Acceleration transducer, 61–65 Anti-aliasing filters, 87, 334
Accelerometer(s), 499, 957, 963, 969, 1254 Antiresonance, 271
base strain sensitivity, 99 Antisymmetric modes, 275
characteristics, 1251 Applied digital signal processing
types used in civil dynamic testing, 1253 blind deconvolution, 175
Accessible emission limits (AEL), 129 blind extraction, 173–175
Acoustic emissions, 1011–1013 blind separation, of structural modes,
Acoustic wavelength, 722, 723, 727 182–185
Acoustically screened, 725 BSS, 177–182
Actuator, 496 cepstrum, 207–226
Adjoint, of matrix, 1395 condition monitoring, for
Aerodynamics, 375 rotating/reciprocating machines,
Aeroelastic damping, 455 167–171
Air damping, 735 cyclic modulation spectrum, 195–197
Air-borne-sound radiation, 686 deterministic signals and traditional Fourier
Aleatory uncertainty, 901 analysis, 152–155
Aliasing, 87, 396 discrete-random separation, 175–177
Allocation rule, 1296 envelope spectrum, 194–196
Ambient sensors, 1257 experimental signals and stochastic signal
Ambient vibrations, 1261–1262, 1285, 1316 modelling, 155–163
Amplitude correction factor (ACF), 243 filtration, 172–173
Amplitude decay, 712, 744 gears, 203–205
Amplitude-frequency characteristic, 744 modal analysis, 167
Analog filters, 83–84 order analysis, 190–191
Analog technology, 10 reciprocating machines and engines,
Analog to digital converter (A/D converter), 205–207
6, 7, 10, 11, 86, 153 reference-based filtering, 172
fundamental requirements, 86 rolling element bearings, 202–203
types, 87–90 scaling and dimensions, Fourier transform,
Analytical methods, 684 164–166
Anemometer, 1256, 1257 second order cyclostationary indicators,
Angle of lag, 745 191–194
Angle–time cyclostationary (AT-CS), 161 signal extraction, 171–172
Angular displacement sensor, 1182 spectral correlation density, 197–200
Angular frequency, 708, 709, 752, 765, 1286 time-frequency representations, 185–190
of Ω, 722 Arctangent phase demodulation, 118

© The Society for Experimental Mechanics 2022 1403

R. Allemang, P. Avitabile (eds.), Handbook of Experimental Structural Dynamics,
https://doi.org/10.1007/978-1-4614-4547-0
1404 Index

Artificial neural network (ANN), 207 Blind deconvolution, 175

Aspect ratio, 729–731, 735 Blind extraction, 173
Assemblies, 712, 740 cyclostationary signals, 174–175
Assembly of substructures, 1191 impulsive signals, 173–174
Assumptions, 537 principle of, 174
Asymptotically elastic behavior, 691 Blind source separation (BSS), 177
Asynchronous averaging, 308 intermittent sources, 181
with Hanning window, 309 maximum kurtosis of sources, 181
Asynchronous signal averaging, 298 mutual decorrelation of cyclostationary
Auto moment functions, 632–635 sources, 181
Auto power spectrum, 269 mutual decorrelation of sources, 181
Auto spectra, 514 mutual independence of sources, 181
Auto-Regressive, Moving Average (ARMA) notion of source, 178
terminology, 555 optimization algorithm, 182
Autocorrelation, 256 problem statement, 178
Automotive body structure, 524 separation operator, 180
Autonomous modal parameter estimation, sparse sources, 181
672–674 types of mixture, 179–180
Autopower spectrum, 246, 249 Block diagonal format, 1135
Autoregressive moving-average (ARMA) Block excitation technique, 1166–1171, 1176
approach, 25–26 Block-based spectrum, 244–246
AutoRegressive with eXogenous inputs Blocksize, 238
(ARX(m,n)) model, 26, 554 Bode diagram, 747
AutoRegressive-Moving-Average Bonded metal foil strain gages, 49
(ARMA(m,n)) model, 554 Bonded strain gage technology, 8
Autospectral density, 376 Boundary-based modal tracking, 1307, 1309
Autospectrum, 316 Boundary conditions, 343
Auxiliary particle filter, 1025 Boundary DOF, 1067
Average link rule, 1297 Boundary element method (BEM), 754–757
Averaged cross spectra, 278 Boundary integral equation, 754, 756
A-weighting, 72 Boxcar, 301
Bragg cell, 112, 116, 118
Bridge, 498, 1328
B Bridge, suspended 25 de Abril, 1323
Back-scatter, 106, 107, 114, 121 anemometer, 1324, 1326
Backscattered-modulation, 114 band-pass filter, 1329
Balanced truncation, 852 deck, 1323, 1327
Base isolated buildings, 1322 detection incidences, 1334
Baseline, for modal tracking strategy, 1333 frames, 1323
Baseline noise, 1327 hangers, 1323
Bayesian evidence, 900 inclinometers, 1323
Bayesian model, 1048 magnetostrictive displacement transducer,
Bearing characteristic frequency (BCF), 169 1323, 1326
Beat signal, 110, 124 modal estimation and tracking, 1327–1334
Bedding damping, 686 model order, 1328
Bending stiffness, 714, 726 pylons, 1325, 1327
Bending vibrations, 726, 729, 734 railway, 1323
Benfield truss, 849 strain gage, 1323, 1326
Benfield-Hruda (BH) method, 1355–1356 structural monitoring system and data,
Between-cluster distance, 1296 1323–1329
Bias error, 249, 310 suspended span, 1324
Bimodal calculation, 747 suspension system, 1323
Bimodal transformation, 753 temperature, 1331
Bimorph PZT transducers, 1182 thermistor housing, 1326
Index 1405

thermistors, 1324 force vibration with shaker, 1260

time-series, 1334 initial conditions testing, 1259
traffic loading, 1331 loading and response, considerations on,
transmission equipment, 1326 1251
Wheatstone bridge, 1323, 1326 modal characterization, 1294–1299
Brite-Euram project, 1044 modal identification, 1290
Broadband excitation, 731 and operational loading, 1290
Buildings, 1317 system identification methods, 1262–1287
Bulk modulus, 700, 749 Civil structural testing, 1306, 1307
Bulk-optics, 144 Classical theory, 396
Burst length, 320, 326 Classification and regression trees (CART),
Burst random excitation, 443 1039
Burst random signal, 325 Classification of laser products, 129
Clipping, 514
C Closed hysteresis, 708
Cables, 92 Closed loop, 136, 142
Cable-stayed bridges, 1255 Cluster(s), 581, 1301, 1309
Calibration, 117, 127, 416–418 compactness, 1300
Candidate sensor set, 518 diagrams, 1288, 1290, 1294
Cap screw stiffness, 1127 number of, 1309
Capacitance technology, 9 validity, 1299–1302
Capacitive accelerometers, 59 Cluster-based modal tracking, 1308–1311
Capture blocks, 319 Clustering, 1298, 1299
Carrier frequency, 113 algorithms, 1294–1299
Carrier signal, 119, 120, 125, 129, 130 methods, 1309, 1329
Cat-eyes reflectors, 122 partitions, 1302
Cavitation, 740 Co-Quad, 11
Centroids, 1296 Coefficient condensation, 602
Cepstral prewhitening, 210 Coherence, 105, 106, 110, 121, 514, 957–959,
Cepstrum, 209 1004
comb notch lifter, 210–212 distance, 121, 128
complex, 207 functions, 278
definition, 207 Coherent averaging, 608
exponential lifter, 210–213 Coherent light, 107, 108, 122
longpass lifter, 214 Coincidence frequency, 727
for machine diagnostics, 214–223 Column space, 1392
for modal analysis, 220–226 Common Statistical Subspace Autonomous
modal suppression, 211–214 Mode Identification (CSSAMI), 674
power, 208 Companion matrix, 556–558
properties, 209 Compatibility, 1066
Ceramic adhesives, 53 Compensated, exponentially decaying
Characteristic phase lag theory, 16 sinusoids (CEDS), 391
Characteristic space, 536, 541, 542, 619, 625 Complex amplitude, 696, 701, 723, 745, 771,
Charge amplifier, 76 772
Chemical composition, 758 Complex bulk modulus, 701
Chirp excitation, 442 Complex cepstrum, 207
Cholesky decomposition, 269, 278, 284 Complex correlation coefficient (CCF), 671
Circle Fit method, 573 Complex Fourier coefficients, 233
Circular convolution, 397 Complex Fourier series, 233
Civil dynamic testing, 1304 Complex mode, 647
modal tracking, 1306–1311 Complex mode indication function (CMIF),
Civil engineering structures, 1250, 1258 634–637, 639, 650
ambient vibrations, 1261–1262 Complex modulus, 696, 697, 753, 773, 787
damage, 1250 Complex shear modulus, 767
1406 Index

Complex Z mapping function, 600–601 functions, 230

Component mode synthesis (CMS), 838, 1070, See also Finite element modelfinite element
1161 model, 859
methods, 1191–1192 Correspondence principle, 749
Compressive sensing, 1016–1018 Couette flow, 761, 762
Computational efficiency, 1327 Coulomb friction elements, 705
Computational simplicity, 1329 Coupling effects, 747
Computed order tracking, 191 Coupling technique, 1158–1162
Concrete structures, 1258 Craig-Bampton (CB) method, 1080, 1191,
Condenser microphone(s), 55, 69–70 1192
construction, 55–57 Creep, 694–696, 702, 758, 759
electret/prepolarized design, 58–59 compliance, 691, 692
externally polarized microphones, 58 function, 695, 701, 703, 705
primary components, 56 modulus, 787
standards, 72 test, 691, 759
Condition monitoring (CM), 992 Cross axis sensitivity, see Transverse
Condition monitoring, for sensitivity, of sensor
rotating/reciprocating machines Cross-channel measurement, 1099
bearing faults, 169–170 Cross-correlation, 1276
gear faults, 168–169 functions, 499
internal combustion engines, 170–171 Cross linking, 758
typical rotor problems, 167–168 Cross modal assurance criterion (Cross MAC),
Condition number, 596 664
Conditioned coherences, 282 Cross orthogonality, 525–526, 880, 882
Confidence limits, 274 check, 880
Confounding influences, 1037–1041 Cross power spectra (CPS), 566
Conical-scanning LDV, 1181 Cross power spectrum, 269
Connection coordinates, 1163 Cross-spectral density, 368
Connectors, 93 matrix, 1275
Consistency, 658–672 Cumulative coherences, 285
Consistency diagrams, 581, 638–641 Curie temperature, 48
trail, 640, 643 Current feedback, 326
Constant relative bandwidth, 237 Curve fitting, 424–425, 687
Constrained support, 512 FRF-based, 1110–1112
Contact surface, 715, 720 parametric models, 1107–1110
damping, 686 Curvilinear method, 708, 711
Contiguous blocks, 319 Cut-off frequency, 727–729, 731, 732, 734
Continuous circular scanning, 1181 Cyanoacrylate adhesives, 52
Continuous linear scanning, 1181 Cyclic averaging, 299, 306
Continuous-scan laser doppler vibrometry with Hanning window, 310
(CSLDV), 138–140 Cyclic convolution, 242
Continuous wavelet transform (CWT), 823, Cyclic modulation spectrum, 195–197
1007 Cyclic signal averaging, 298
Continuum mechanics, 693, 705 Cyclo-non-stationary signals, 157, 160–163
Control system, 487 Cyclostationary signals, 157–159, 174–175
Convolution, 242 CS1 model, 160
Convolutive mixture, 179 CS2 model, 160–161
Coordinate incompleteness, 1165 Cylindrical coordinates, 731
Coordinate modal assurance criterion
(COMAC), 669, 883 D
Coordinate orthogonality check (CORTHOG), Damage, 487
671, 886 and earthquake vibrations, 1319–1322
Coordinate transformation, 1380 identification of, 1303
Correlation, 272, 859 scenarios, 1320
Index 1407

Damped complex exponential response Determinant, 1391

methods, 21–24 Determined equations, solution of, 1394–1396
Damped model reduction, 1194–1195 Deterministic excitation methods, 329
Damping, 685 Deterministic signals, 322
of assemblies, 712–741 Deviator, 698–701
contact-surface, 686 Diagonal mass matrix, 1124
couplings, 687, 752 Diakoptics, 1096
damping in guides, 686 Dielectric losses, 686
electromechanical, 686 Difference method formulations, 576
external, 686, 757 Differential amplifier, 77
high, 771–772 Differential operator, 693, 699, 700, 749–750,
material, 686 779, 781
models, 1114 Differential vibrometers, 131–132
non-conservative, 685 Diffusion processes, 686
parameters, 687, 767, 778 Digital decoding, 117, 118, 125, 128
ratio, 687, 743, 752, 764, 767, 775, 776, Digital demodulation, 117, 118
1307 Digital filters, 84
soil, 778–779 Digital image correlation (DIC), 9, 10,
of solids, 687–712 413–414, 443, 539, 621
and stiffness, 551 air turbulence, 447–448
structural, 686 blurring, 437–438, 445
work, 688, 689, 697, 705, 706, 708 burst random excitation, 443
Damping of solids calibration, 416
linear models, 689–704 camera angle, 446
nonlinear models, 704–712 camera setup and calibration, 444–445
physical phenomena, 687–690 curve fitting considerations, 423–425
Dams, modal quantities of, 1251 data processing parameters, 448–449
Dark speckles, 125, 126 3D SLDV and 3D stereo-DIC, 455
Data acquisition, 10, 1290 Fourier transformation
analog technology, 10 for high rate testing, 463
classification, 14 frequency response function measurement
Data acquisition systems (DAS), selection of, considerations, 420
1257 hardware, 415
Data-based SHM, 1036 heating effects, 439
Data cleansing, 993, 1318 image correlation parameters, 448
Data condensation, 536 impact testing, 440
Data decimation, 1318 and leakage considerations, 422
Data domains, 541, 624 lens adjustment, 436–437
Data loggers, 1290 lighting, 439
Data normalisation, 993, 1040 long sampling requirements, 430–431
problem, 1037 patterning, 416
Data sieving/filtering/decimation, 601 periodic random excitation, 442–443
Decay coefficient, 743 phase stepping, 435
Deconvolution, 396 projected speckle patterns, 451
Deflectometry, 452 proper sampling parameters, 430
Deformation energy, 714 pseudo random excitation, 442
Degrees of freedom (DOFs), 424, 500, 538, pure random excitation, 442
621, 625, 645–647, 652, 665, 776, sampling theory relationships, 429–430
1300 sine excitation, 441
number of input, 543 single point measurement, aliasing with,
number of output, 543 432
DelrinTM , 787–788 software, 416
Delta-sigma converter, 90 speckle pattern and target shape, 446
Depth-of-field, 124 strain mode shapes, 449–451
1408 Index

Digital image correlation (DIC) (cont.) Dynamic compliance (receptance), 746

stroboscope lights and high-speed Dynamic condensation, 865
measurements, 433 Dynamic environments, 1351
structural dynamic testing/modal analysis, Dynamic IRS (DIRS) technique, 1189
427 Dynamic mass, 747
structural motion and frequency, 427–428 Dynamic model, in modal coordinates,
swept sine/chirp excitation, 441–442 1112–1115
temporal vs. frequency domains, 428 Dynamic reduction technique, 1188
temporal aliasing for image processing, 432 Dynamic response, of structural system, 1300
traditional modal analysis sensing, 458–463 Dynamic stiffness, 746, 771
Digital signal processing (DSP), 21 matrix, 1156
Digital-to-analog (DAC) converter(s), 315, 328 Dynamic testing, challenges for, 1250
Digitization bit noise, 514 Dynamic transmission behavior, 719
Digitization noise, 119 Dynamic viscosity, 738, 760
Dilatation, 699, 700
Dirac delta function, 754
Direct parameter estimation, 976, 978 E
Direct piezoelectric rotational accelerometers, Earthquake, 1285, 1286
1184 and base isolated buildings, 1322
Direct system parameter identification (DSPI), Earthquake vibrations
26 and building, 1319–1322
Discrete attachment, 1081 Effective independence, 517, 519
Discrete Fourier transform (DFT), 154, 238, distribution, 519
396, 422, 1002 Effective mass, 1124–1127
Discrete orthogonal polynomials, 599 Effective modal frequencies, 603
Discrete relaxation times, 709 Effective modal mass, 565, 631
Discrete-time Fourier transform (DTFT), 154 Eigen frequencies, 731
Discrete-time signals Eigen solutions, 1107, 1394
and digital implementation, 153–155 Eigenfrequency, 735
Displacement field, 756, 786 Eigensystem realization algorithm (ERA), 23,
Displacement sensors, 1255 535, 1263, 1267–1270, 1285, 1331
Displacement vector, 748, 755 Eigenvalue, 1107, 1112
Dissipation, 686, 715, 741, 745 Eigenvalue decomposition (ED), 270, 279, 604
Distance matrices, 1290–1291 Eigenvalue-eigenvector methods, 548
Distance metrics, 1315 Eigenvalue-eigenvector solution, 538
Distortion dependent stiffnesses, 708 Eigenvectors, 687, 751–753, 775, 1107, 1113,
Diversity combining, 126 1116
Domains, 1380 Elastic mode, 506
Doppler effect, 106, 143 Elastoplastic deformation, 715
Double Least Squares (DLS), 536 Electrets, 58–59
Double mirror configuration, 137, 138 Electrical cables, 92
Double-sided versus single-sided spectra, 229, Electrical energy, 686
236 Electromagnetic, 685, 686
Dove prism, 140 noise, 91
Down mixing, 116, 119 Electromechanical damping, 686
Drive point, 1090 Electromechanical processes, 685
FRF measurement, 1122 Electrostatic noise, 91
reciprocity, 514 Elemental stiffness matrix, 748
Drop-out, 118, 121, 123, 125, 126 Elliptical shape, 708
3D Scanning Laser Doppler Vibrometry (3D EMA modal model
SLDV), 455 FRF calculation, 1099
3D scanning vibrometer, 142–143 linear vs. non-linear dynamics, 1102
Dual Craig-Bampton method, 846–848 measurement chain, 1098
Dynamic bandwidth, 1350–1351 sensing force and motion, 1099
Index 1409

sensitivity flatness, 1099 Experimental dynamic substructures

sensor linearity, 1100 experimental difficulties, 1077, 1086–1088
sensor mounting, 1101 finite frequency band, 1064
EMA mode shapes, 1146 SEM substructuring testbed, 1064
Embedded sub-component models, 1231–1240 substructure technologies, 1065–1077
Energy dissipation, 685, 686, 688, 712, 740, transmission simulator (TS) theory,
747 1084–1089
Energy spectral density (ESD), 166 transmission simulator approach, 1081,
function, 236, 251 1082
Engle-Granger two step estimation procedure, TS approach, 1089
1043 Experimental errors, 1081
Enhanced coordinate modal assurance criterion Experimental modal analysis (EMA), 12–14,
(ECOMAC), 670 16, 18, 20, 21, 23–26, 29, 225, 275,
Enhanced Frequency Domain Decomposition 313, 322, 535, 617–619, 674, 675,
(EFDD), 1275, 1281 687, 775, 1095
Enhanced frequency response function (eFRF) measurements, 232
eFRF r (ω), 651 Experimental modal parameter evaluation
FRF SVD development, 652–655 methods, 621–624
historical development, 652 autonomous modal parameter estimation,
theoretical definition, 650 672–674
Ensemble/average, 320 characteristic space, 625
Environmental fluctuations, 1262 data dimensionality, 626
Environmental loading, 1304, 1306 data domain, 624
Environmental sensors, 1256 equation normalization, 629–630
Epistemic uncertainty, 900 general (two-stage) solution procedure,
Epoxy adhesives, 53 628–629
Equation condensation methods, 607 generalized frequency, 626
Equation normalization, 545–546, 629–630 kernel equations, 626–628
Equation of motion, 726, 748, 750, 751, 774, modal frequency evaluation/validation
786, 818, 1215 tools, 632–644
transformed, 1112 modal vector evaluation/validation tools,
Equilibrium, 1066 645–672
modulus, 692 modal vectors, modal scaling, residues,
Equivalent linearization, 719 630–631
Equivalent noise bandwidth, 249 overdetermined linear models, 628
Equivalent Reduced Model Technique Experimental structural dynamics, 12
(ERMT), 1218, 1219 applications in, 229, 231
Erbium-fiber lasers, 131 classification methods, 13
Ergodic, 321 conferences, 28–30
Error covariance matrix, 518 damped complex exponential methods,
Estimate modal characteristics, 316 21–24
Euclidean distance, 1292 data acquisition classification, 14
Euclidean norm, 1398 definition, 6
Euler Bernoulli beam theory, 773 frequency response function method, 18–21
Evolution equations, 707, 708, 711 government reports, 31
Excitation, 313 handbooks, 30
methods, 319 mathematical input-output model method,
signal, 320 24
Expansion technique, of measured data pioneers/contributers, 33–39
frequency response functions, 1196–1198 sinusoidal input-output method, 15–18
Kidder’s method, 1196 sound and vibration magazine, 30
using analytical modes, 1196 textbooks, 31
Expected value function, 367 Exponential averaging, 295
Experimental difficulties, 1077 Extended Kalman filter (EKF), 1024
1410 Index

External damping, 686, 741, 757, 759 Forced response decomposition method, 17
External variable, 693 Forced vibrations, 744–745, 1290
Externally polarized microphone, 58 Force identification technique, see Sum of
Extreme loads, 1251 weighted acceleration technique
(SWAT)
Force transducer, 61–65, 68–69
F Force vibration, 1260
Fading memory, 782 Force windows, 331
Fast Fourier transform (FFT), 6, 7, 11, 12, 15, Forward model-driven SHM, 1049–1051
16, 21, 238, 265, 357, 381 Fourier coefficients, 232, 302
FEA modal updating, 1149–1152 Fourier series, 152, 153, 164, 1383
FEA mode shapes, 1107 complex, 233
FEA RIB model, 1132–1134 theory, 232
Feature extraction and information Fourier spectra, 774
condensation, SHM, 993–994 Fourier spectrum, 491
Ferroelectric polycrystalline ceramic materials, Fourier transform, 153, 231, 297, 301, 357,
47 775, 979, 1264–1265, 1382
Field equations, 754 deterministic transient signals, ESD, 166
Field-of-view (FOV), 427 periodic and quasi-periodic signals,
Field point, 755 164–165
Filled polymers, 709, 710 stationary random signals, PSD, 165–166
Filter(s), 81 Fourier transformation, 754
analog filters, 83–84 and leakage considerations, 422
digital, 84 Fractional coherence, 285
implementations, 83, 85 Fractional derivatives, 699, 788
ripple, 84 Grünwald definition of, 779–781
Finite difference technique, 1166, 1171–1177 numerical calculation of, 781–784
Finite element analysis (FEA), 1095, 1106 Fractional-order derivative, 694
Finite element method (FEM), 687, 715, 719, Free boundary conditions, 505
720, 779–790, 1164, 1346–1348 Free decay, 316
Finite element model, 515, 1227 Free-field microphone, 70
model reduction, 862–863 Free-free boundary, 344
Finite Fourier series, 807 Free response, 316
Finite impulse response (FIR) filters, 85 Free vibrations, 742–744, 764–766
Finite time elements, 750 Frequency(ies), 1307
First-order approximation, 1176 analysis, 237
finite difference transformation matrices matrix, 1292
for, 1173 modulation, 108
Fisher information matrix (FIM), 518 normalization, 597
Fixed base, 1080 range, 485
Fixed-interface CMS algorithm, 1192 ratio, 728, 744
Fixed-interface component, 1162 resolution, 301
Fixture, 1082 response functions, 1004
Flash A/D converter, 87–88 shaping, 319
Flattop window, 241 shift, 106, 109, 110, 112–114, 132
Flow effects, 685 spectrum, 1002–1004
Flow viscosimeter, 760 Frequency based substructuring (FBS), 1075
Flutter predictions, 487 Frequency demodulation (FDM), 116
Foam suspension, 512 Frequency domain, 555, 1380, 1382–1385
Focal length, 124 characteristic, 302
Force appropriation method, 673 dynamic model, 1107
Force balance, 1106, 1253 estimation, 592
Forced normal mode excitation method, Frequency domain assurance criterion (FDAC),
16–18 671
Index 1411

Frequency domain decomposition (FDD), Grounding, 90

1263, 1274–1277 Guided waves, 1013–1015
Frequency-energy plot (FEP), 809–812, Guidelines, 1089
820–824, 826 Guyan reduced structural dynamic model,
Frequency response assurance criterion 1369
(FRAC), 670, 884 Guyan reduction, 863, 1353–1354
Frequency-response curves, 687 technique, 1187
Frequency response function (FRF), 222, Guyan-Irons method, 835–837
223, 226, 231, 260, 262, 266, 276,
297, 313, 345, 380, 396, 425, 426,
535, 541, 544, 617, 624, 626–629, H
805, 957–959, 965, 966, 978, 980, Half power bandwidth method, 568, 570
982–984, 986, 1011, 1156, 1383 Half-power methods, 567
auto moment functions, 632–635 Halfwidth value, 767
CMIF, 635, 636 H1 algorithm, 269
coupling method, 1159, 1160 H2 algorithm, 269
data, 553 Hv algorithm, 269
distortion, 960 Hamilton’s dynamical equations, 547, 548
expansion process of, 1196–1198 Hamilton’s principle, 1346
experimental, 1110 Hankel matrix, 24, 1279, 1283, 1287, 1390
higher order FRFs, 961–964 Hanning window, 241, 244, 306, 308, 311,
matrix, 264, 546 314, 1102
measurement considerations, 420 characteristics, 305
model, 18–21 Hard criteria, 1329
FRF-based approach, 1175, 1176 Hard stability criteria, 1293–1294
FRF-based curve fitting, 1103, 1110–1112 Harmonic(s), 256
Friction hysteresis, 716 deformation, 765
Frictional stress, 711 velocity, 722
Fringe distance, 134 Harmonic-balance method, 719
Full field expansion, 1219 Harmonic transfer function (HTF), 140
Full space linear beam models, 1222 Health monitoring, 1304
Fundamental solution, 754–756 Heavily damped systems, 332
Fuzzy method, 1048 Helium-neon gas laser, 130
Helium-neon laser, 106, 127, 130
Hermitian matrix, 1389
G Heterodyne configuration, 110, 112
Gage factor, 49 Heterodyne technology, 10
Galvanometric motors, 136 Heterodyning efficiency, 110, 113, 121
Gamma function, 694, 781 H-frame structure, 334
Gas pumping, 686, 715, 735 Hierarchical agglomerative clustering method,
Gauss quadrature, 756 1309
Gaussian mixture model (GMM), 1033 Hierarchical agglomerative methods,
Gaussian probability distribution, 323 1296–1299
General framework, 1066 High rate testing, DIC
General input-output model, 13, 14 camera calibration, 470
Generalized complex eigenproblem, 1158 camera protection, 468–469
Generalized frequency, 544, 596 camera selection, 464–466
Generalized mass, 551 2D vs. 3D stereo-DIC, 466
Generalized modal DOF, 1071 environmental concerns, 466
Geometric damping, 778, 779 extended noise floor measurements, 469
Glass transition temperature, 702, 703 IRIG, 472
Global stiffness matrix, 749 motion blur, 471–472
Grünwaldian fractional derivatives, 786 painting techniques, 472–473
Ground loops, 92 polarization, 471
1412 Index

High speed gear mesh (HSGM), 216 Independent component analysis (ICA), 181
High speed shaft (HSS), 216 Inductance technology, 9
Hilbert Huang Transform, 1264 Inelastic, 687
Hilbert transform, 1277 Inertance, 747
in frequency domain, 966–967 Inertia restraint, 562
in time domain, 964–966 Inertial forces, 1320
Ho-Kalman procedure, 24 Infinite impulse response (IIR) filters, 86
Holograms, 144 Initial conditions, 706, 743, 754, 755, 785
Homodyne, 110, 129, 144 Initial modulus, 692
Homogeneous isotropic continuum, 749 Inner product, 522
Homogeneous strips, 773 In-phase vibration, 725
Hurty-Craig-Bampton (HCB) method, In-plane vibrometers, 134–135
839–842, 1354–1355 Input auto power matrix, 284
Hybrid, 869–870 Input autospectra, 289
excitation method, 329 Input cross spectrum matrix, 289
modal model, 1095, 1132 Input-input cross-spectra matrix, 553
random signal, 328 Input locations, 526–529
reduction technique, 1190 Input-output relationship, 268
Hydrostatic, 698, 778 Input recording, 1319
pressure, 779 Instantaneous amplitude, 965
Hysteresis, 688, 698, 707, 716–719, 742, 779 Instantaneous frequency, 965
curve, 689, 697, 705, 706, 708–710, 717, Instantaneous phase, 965
719 Instationary phase, 706
frictional, 716 Instrumentation amplifier, 77
point-symmetrical, 705–708 Integral(s), 781
static, 705 distortion laws, 709
Fourier Transform, 1385
transformations, 750
I Integrated circuit piezotronic (ICP® ), 8
Ibrahim time domain (ITD) approach, Integrated electronics piezo-electric
22, 535 (IEPE), 8
Identification of damage, 1303 accelerometer, 96, 99
IEC classification scheme, 129 Integrated electronics piezoelectric (IEPE)
IMAC Advisory Board, 38 sensor
Impact excitation, 331 advantages, 77
Impact hammer, 489 cable length check, 81
Impact modulation (IM), 971 coupling capacitor time constant, 80
Impact signal, 331 electrical considerations, 79
Impact testing, 331 linearity, 79
Impedance coupling techniques, 1159 mechanical considerations, 79
Impedance head, 69 precautions, 80, 81
Impedance matrix, 263, 1160 transducer time constant, 79, 80
Improved accuracy, 275 Integrated test analysis process (ITAP), 1362
Improved reduction system (IRS) technique, Integrating A/D converter, 89
864, 1188–1190 Integro-differential equation, 750
Impulse based substructuring (IBS) method, Inter-Range Instrumentation Group (IRIG)
1162 Timing, 472
Impulse response functions (IRFs), 22, 230, Interface reduction approach, 852–853
353, 535, 541, 617, 622, 625, 627, Interference, 106, 109, 111, 113, 114, 116,
629, 1162, 1383 117, 120–122, 129–132, 134
Impulsive, 316 Interior coordinates, 1163
signals, 173–174 Intermediate shaft gear mesh, 216
Inclusive outliers, 1031–1032 Internal combustion (IC) engines, 170–171
Incomplete equivalent model, 777 Internal variable, 693, 707, 711
Index 1413

International Modal Analysis Conference Large deformation, 704

(IMAC), 29 Laser class, 129
International Seminar on Modal Analysis Laser coherence, 121
(ISMA), 28 Laser diode, 114, 116
Intersection volume, 134 Laser Doppler vibrometry (LDV), 1166,
Intuition, 522 1177–1182
Invariant manifold approach, 805, 819 back-scatter issues, 121–122
Inverse discrete Fourier transform (IDFT), 154, demodulation of Doppler signals, 116–118
155, 252 differential vibrometers, 131–132
Inverse Fourier transform, 757 diversity combining, 126
Inverse modal assurance criterion (IMAC), 670 in-plane vibrometers, 134–135
Inverse model-driven SHM, 1047–1049 laser safety and standards, 128–129
Inverse problems, 1394–1400 lasers sources and Doppler effect, 105–108
IRS technique, see Improved reduction system Mach-Zehnder interferometer, 112–114
(IRS) technique Michelson interferometer, 110–111
Isolated resonance frequency, 776 multi beam, 143–144
Isolated resonance point, 776, 777 noise and resolution, 118–121
Isoparametric concept, 756 optical homodyne and heterodyne, 108–110
Isothermal, 702 optical-fiber-vibrometers, 131–132
Isotropic materials, 699 rotational vibrometers, 132–134
Iterated IRS (IIRS) technique, 1189 self-mixing, 114–116
Iterative modal identification, 1285 single point vibrometers, 129–131
speckle noise, 122–123
tracking filter, 125
J uncertainty and calibration, 126–128
Jenkin-Element, 706 See also Scanning vibrometersscanning
Johansen procedure, 1043 vibrometers
Joint(s), 685, 686, 712, 715, 716, 719, 720, 736 Laser-optical sensors, 687
compliance, 1081 Laser velocimeter, 969
sliding, 496 Laser vibrometer, 501, 956, 1000
Journal bearings, 737–739 Latent variable models, 1039
Layout and documentation, 502–503
K Leakage, 240, 272
Kalman filter, 172 bias error, 310
Kelvin-Voigt model, 692, 697, 698, 751–753, error, 273, 281, 297, 310
760 reduction method, 335
Kernel equation, 544, 584 Leakage error, FRF calculation, 1101
frequency domain, 626 Leakage-free signals, 1102
time domain, 627 Leakage-free spectrum, 1101
Kidder’s method, 1196 Least absolute shrinkage and selection operator
Kinetic energy, 742 (Lasso), 1017
Kirchhoff-Helmholtz integral, 720 Least squares (LS), 266, 536, 591
Kirchhoff’s theory of a plate, 726 error algorithm, 1110
k-means, 1296 frequency response functions, 282
k-medoids, 1303 methods, 266
solution, 577, 1398
Least squares complex exponential (LSCE),
L 22, 535
Lagrange multiplier method, 899, 1096 Least squares, global SDOF formulations, 579
Lagrangian approach, 142 Least squares, local SDOF formulations, 578
Laminated components, 712, 715 Least-squared-error method, 1103
Laminated strips, 773 Left-hand eigenvectors, 753
Laplace domain, 1380, 1382 Leibniz’s rule, 358
Laplace transform, 1112 Lifting approach, 139–141
1414 Index

Light amplification by stimulated emission of linear adjustment, 1265–1266

radiation (LASER), 105 nonlinear adjustment, 1266–1267
Lightly damped structures, 1115 Logarithmic decrements, 764
Lightly damped systems, 332 Logarithmic radiation efficiency, 721, 723
Linear algebra Long dimension (NL ), 543, 626
concepts and definitions, 1385–1389 Longitudinal modes, 105, 106, 130
determined equations, solution of, Longitudinal oscillations, 774
1394–1396 Loss factor, 688, 696, 702, 713, 715, 722, 726,
eigen solutions, 1394 735, 736, 741, 745, 753, 754, 767,
inverse problems, 1394–1400 773
matrix forms, 1389 Loss modulus, 696, 697, 763
singular value decomposition, 1393 Low order (First) frequency domain methods,
symmetric matrix rules, 1390–1391 589
transposition rules, 1389 Lower triangular matrix, 1396
Linear averaging, 293 LQ decompositions, 1284
Linear correlation, 283 LU decomposition, 1395
Linear FRF estimates, 1103 Lumped parameter mass, 262
Linear independence, 519
Linear inelastic, 689
Linear instantaneous mixture, 179 M
Linear models MAC, see Modal Assurance Criterion (MAC)
analysis, 820–821 Mach-Zehnder (MZ) interferometer, 112–114
creep and relaxation, 694–696 Machine learning, 1000, 1026, 1027, 1049
harmonic stress and strain function, MacNeal method, 845–846
696–697 MacNeal-Rubi (MR) method, 1356–1359
N -parameter model, 692, 693 Magnetostrictive, 685, 686
operator notation, 693 Magnitude averaging, 294
temperature dependence of viscoelastic Manhanttan distance, 1292
material properties, 702 Manufacturing processes, 758
thermo-rheological simple materials, Masing model, 705, 707
703–704 Mass, 686
three dimensional stress state, 697–702 additive technique, 1166, 1171
three parameter model, 689–692 errors, 1079
Linear normal modes (LNMs), 802, 807 matrix, 748
invariance, 800 normalized, 1070
modal superposition, 800 weighting, 521
Linear parametric curve fitting model, 1103 Mass-normalized modal matrix, 1190
Linear spectrum, 246 Mass-normalized mode shape matrix, 1157
Linear systems Mass-normalized mode shapes, 139
properties of, 231 Material damping, 685, 686, 715, 735, 763,
simplest, 231 778, 779
Linear time invariant (LTI), 152 Material deterioration, 1251
Linear variable differential transformer Material equation, 693, 694, 696, 700, 701
(LVDT), 1255 Material loss factor, 688, 702, 703
Linear viscoelastic, 697, 698 Material memory, 695
Linearity, 264, 487 Mathematical input-output model method, 24
Lissajous patterns, 10 Matrix coefficient polynomial model, 552
Load cell solutions, 1256 Matrix inverses, 1395
Load point, 754–756 Matrix of transfer functions, 774, 775
Local modelling, 716 Matrix structural dynamic models
Local operator, 781 application of Hurty-Craig-Bampton
Locus diagram, 747 and MacNeal-Rubin methods,
Log decrement method, 568, 572 1359–1360
Logarithmic decay Benfield-Hruda method, 1355–1356
Index 1415

Guyan Reduction, 1353–1354 COMAC, 669

Hurty-Craig-Bampton method, 1354–1355 cross MAC, 664
MacNeal-Rubin method, 1356–1359 definition, 660
and mode acceleration method, 1361–1362 ECOMAC, 670
Maximum likelihood approach, 1034 FMAC, 669
Maxwell and Betti’s theorem, 756 IMAC, 670
Mean phase (MP), 1289, 1292, 1293 MACRV, 669
Mean phase collinearity (MPC), 1289, 1292, MACSR, 668
1293, 1306 matrix, 522
Mean phase correlation (MPCr ), 648 misuse/abuse of, 672
Mean phase deviation (MPD), 647, 1289, PMAC, 668
1292, 1293, 1306 pole weighted/state vector MAC,
Mean stress, 698, 699, 701 665–667
Measured damping, 511 SMAC, 668
ratio, 508 uses of, 671–672
Measurement inaccuracy, 904 WMAC, 668
Measurement situation, 267 Modal assurance criterion square root
Mechanical energy, 685 (MACSR), 668
Mechanical impedance, 747 Modal assurance criterion using reciprocal
Mechanical shock, 351, 355, 372, 406 vectors (MACRV), 669
Mechanical vibration energy, 686 Modal assurance criterion with frequency
Medoids, 1296 scales, 669
Melts, 762 Modal-based approach, 1176
Memory integral formulation, 750 Modal-based assembly (MBA), 1191
Memory integrals, 694, 700 Modal characteristics, 775
MEMS sensor, see Micro-Electro-Mechanical- Modal confidence factor, 22
Sensor (MEMS) Modal coordinates, dynamic model in,
Metal strain gages, 49 1112–1115
construction, 50–53 Modal correlation coefficient, 670
history, 50 Modal coupling techniques, 1159, 1161
Method of weighted residuals, 756 Modal damping, 332, 510, 751, 775
Metrics, 1294 matrix, 1117, 1118
Metrological characteristics, 127 ratios, 685
Michelson interferometer, 110–111 Modal features, 1302
Micro electro mechanical systems (MEMS), Modal forces, 751
1166, 1183, 1253 Modal frequency evaluation/validation tools
transducers, 54 auto moment functions, 632–635
Micro slip, 715 consistency diagrams, 638–641
Microplastic deformations, 686 modal parameter clustering, 643–644
Min-Mac, 522, 524, 525 mode indication functions, 635–638
Minicomputer FFT systems, 12 model order relationships, 632
“Minimum variance” rule, 1297 pole surface consistency plots, 641–643
Minkowski distance, 1290 Modal identification, 1197, 1290
Mobility (admittance), 746 Modal kinetic energy, 516–518
Modal A, 631 Modal mass, 631, 751, 775
Modal analysis, 419, 420, 425, 427, 440, matrix, 1116
449, 450, 452, 453, 455, 458, 751, Modal model, 1157
1348–1349 Modal Modification Response Technique
test, 343 (MMRT), 1218, 1219
Modal analysis-based features, 1011 Modal parameter(s), 317, 537
Modal assurance criterion (MAC), 641, 658, clustering, 643–644
659, 661–664, 687, 879, 1130, 1220, Modal parameter estimation (MPE), 13, 15,
1225, 1292, 1293, 1298, 1306, 1307, 19, 23, 28, 32
1309, 1329 algorithms, 540, 544
1416 Index

Modal parameter estimation methods, Modal tracking, 1304, 1306–1311, 1318

621–624 Modal truncation, 1191
algorithms, 619 augmentation, 852
autonomous modal parameter estimation, Model updating, 1048
672–674 Modal vector(s), 564
characteristic space, 625 Modal vector complexity plot (MVCP),
data dimensionality, 626 648–649
data domain, 624 Modal vector conditioning, 645
equation normalization, 629–630 central axis rotation, 646
general (two-stage) solution procedure, modal vector complexity plots,
628–629 648–649
generalized frequency, 626 real normalization, 646
kernel equations, 626–628 vector complexity, 647
modal frequency evaluation/validation vector normalization, 645
tools, 632–644 Modal vector consistency, 658–672
modal vector evaluation/validation tools, Modal vector evaluation/validation tools
645–672 eFRF, 650–655
modal vectors, modal scaling, residues, modal vector conditioning, 645–649
630–631 modal vector consistency, 658–672
overdetermined linear models, 628 weighted modal vector orthogonality,
Modal participation factor, 659 656–658
Modal properties, 1157 Modal vector normalization, 620
Modal results, 1314 Mode acceleration method,
Modal scale factor (MSF), 658, 659 1361–1362
Modal scaling, 550 Mode contribution matrix, 1218
Modal sensitivity, 1146, 1147 Mode indication functions (MIF), 635
analysis, 1144, 1146, 1152 CMIF, 635–637
Modal space, 1215 MvMIF, 637–638
Modal stiffness, 751, 775 Model expansion, 870–877
matrix, 1117 Model incompleteness, 1186
Modal synthesis, 1096 Modelling errors, 904–906
Modal TAM, 868–869 Model order, 1293
Modal test Model reduction, 859
acquisition methods, 500–501 component mode synthesis method,
constrained support, 512 1191–1192
constraints, 484 damped model reduction, 1194–1195
controlled inputs, 496–497 dynamic condensation, 865
excitation level, 485–487 dynamic reduction technique, 1188
free boundary conditions, 505–511 and expansion concept, 1215–1216
frequency range, 485 Guyan reduction technique, 863, 1187
impulsive inputs, 489–496 hybrid, 869
linearity, 487 improved reduction system technique, 864,
location measurements, 502 1188–1190
model validation, 515–529 measured data expansion, 1195–1198
multiple inputs, 497 modal TAM, 868
natural/operational inputs, 498–499 modal truncation, 1190–1191
operating environment, 513 model expansion, 870
operational loads, 488 primary coordinates, 1187
purpose of, 483 sum of weighted accelerations technique,
quality criteria, 513–514 1193–1194
structural damping, 488 system equivalent reduction expansion
suspension system design, 511 process, 866–868, 1190
test display model, 503–504 test data consideration, 876–879
transducers, 499 vector correlation, 879
Index 1417

Model updating, 521, 687 Multiple input, multiple theory, 276

industrial example problem, 935–945 Multiple input relationship, 19
modelling errors and measurement Multiple-degree-of-freedom structure,
inaccuracy, 904 361–365
parameter estimation, 902 Multivariable Output-Error State sPace
parameterization, 902–914 (MOESP), 1263, 1282–1286
process, 1155 Multivariate mode indication function
regularization, 909 (MvMIF), 635, 637–638
sensitivity analysis, 906–909 Mutual correspondence criterion, 670
stochastic model updating, 923
validation, 931 N
Model validation, 483, 500, 505, 515 Narrowband noise, 95
Mode normalization, 1318 Natural angular frequency, 743, 775
Mode representatives, 1306 Natural Excitation Technique ERA (NeXT
Mode shape(s), 128, 138–141, 1115–1118, ERA), 1263, 1276–1281
1292, 1299, 1302, 1303, 1305, 1316 Natural frequencies, 687, 743, 753, 767, 773,
expansion, 1195 776, 777
Modern multiple degree of freedom (MDOF) Natural inputs, 488, 498
methods, 580 Neural networks, 1262
Modes, 1309 New England Photoelasticity Conference, 8
shapes, 1290 Newmark time integration, 1230
Modified force vector, 787 Newtonian liquids, 760, 762
Modified stiffness matrix, 787 Nodal displacements, 748, 784, 786
Modified version of the Kidder’s expansion Nodal forces, 749
method, 1196 Noise and Vibration Conference (NVC), 29
Moment excitation, 1167, 1172, 1184–1186 Noise contamination, 272
Moment mobility, 1185 Non-conservative, 685
Moment of inertia, 763 Non-contact vibration, 141
Mono-axial, 700, 709 Non-destructive testing/evaluation, 992
Monte-Carlo (MC) sampling technique, 1030 Non-isothermal, 702
Moore-Penrose Generalized Inverse, Nonlinear Autoregressive Model with
1399–1400 eXogenous inputs (NARMAX),
Moore-Penrose pseudo inverse, 1190 1010
MOOZ random signal, 328 Nonlinear autoregressive moving average with
m-th order state vectors, 558 exogenous inputs (NARMAX),
Multi beam, 143–144 973–977
Multi-input, multi-output (MIMO), Nonlinear characterization
984, 1322 coherence, 957–959
Multi-mode fibers, 131 FRF distortion, 960
Multi-reference FRFs, 497 higher order FRFs, 961–964
Multichannel vibrometer, 144 Hilbert transform, in frequency domain,
Multiple coherence, 281, 288 966–967
Multiple coherence function, 335 Hilbert transform, in time domain, 964–966
Multiple degree of freedom (MDOF), 622, 623 restoring force method, 967–970
measurement, 539 VAM, 970–972
methods, 777 Nonlinear cointegration, 1046
Multiple input estimation, 274 Nonlinear connection elements, 1213, 1214,
Multiple input excitation, 275 1219
Multiple input frequency response function, Nonlinear identification through feedback of
276 the outputs (NIFO), 983–986
Multiple input, multiple output (MIMO), 7, 12, Nonlinear model(s), 704–705
19, 618, 620, 646, 659 analysis, 821–826
algorithms, 537 hysteresis curve, general shape of, 708
model, 1107 mathematical model, 711
1418 Index

Nonlinear model(s) (cont.) Octave analysis, 237

nonlinear viscoelasticity, models for, Off-diagonal terms, 522
708–709 Operating deflection shape (ODS), 566, 1104
point-symmetrical hysteresis, without measurements, 231, 248
reversal points, 705–708 Operating vector (Peak-Pick) estimation, 566
rheological model, 710–711 Operational conditions, 758
static hysteresis and viscoelasticity, models Operational evaluation, 992
for, 709 Operational inputs, 488
static hysteresis, models for, 705 Operational loading, 1290, 1304
Nonlinear normal modes (NNMs), 800 Operational loads, 488
analytical methodologies, 820 Operational modal analysis (OMA), 182,
energy-based formulation, 815–819 220–226, 230, 498, 535, 617
frequency-energy dependence, 805–810 measurements, 232
invariant manifold approach, 805, 819 Optical access, 124
limitations, 801 Optical beating, 108
linear modal analysis, 820–821 Optical derotator, 140–141
modal interactions, 810–813 Optical-feedback interferometry, 114
mode bifurcations and stability, 813–815 Optical-fiber-vibrometers, 131–132
nonlinear modal analysis, 821–826 Optical technology, 9–10
numerical techniques, 819 Optical windows, 124, 128
reduced-order modeling, 826–827 Optimal set, 515
Rosenberg’s definition, 803–804 Order-frequency spectral correlation (OFSC),
Nonlinear response, 1090 163
Nonlinear response prediction, 1220–1227 Ordinary coherence, 280, 287
full field dynamic strain prediction, function, 272, 289, 308
expansion for, 1227–1231 Orthogonal polynomials (OP), 536, 597–599
with embedded sub-component models, Orthogonal projector, 1397
1231–1239 Orthogonality, 550, 861, 868, 880, 882
Nonlinear system analysis methods conditions, 1157, 1190
experimental setup, 954–956 Outlier analysis, 1028–1031
nonlinear characterization, 957–972 Over-determined equations, solution of,
parameter estimation, 973–986 1397–1399
Nonlinear system models, 1214, 1219 Over-determination factor, 544, 601, 628
Nonlinear viscoelasticity, 708–709 Overlap, 248
Nonlinearity(ies), 272, 281, 322, 345, 487, 735 processing, 311
Non-local operators, 781
Non-parametric model, 13 P
Non-parametric spectrum estimation, 237 Parameter estimation
Non-proportional damping, 1114, 1194, 1290 direct parameter estimation, 976, 978
Normalized equivalent bandwidth, 244 NARMAX, 973–976
N-parameter model, 692, 693, 748–749 NIFO, 983–986
Null space, 1392 reverse path method, 978–983
Number of references, 543, 626 2-Parameter model, 751–753
Number of responses, 626 3-Parameter model, 689–698
Numerical integration, 750, 788 Parameter selection, 930
Numerical methods, 684 Parameterization
Nyquist curve, 767, 768 generic elements, 918–922
Nyquist frequency, 239, 432 mass, damping and stiffness matrix
Nyquist-Shannon sampling theorem, 121, 1016 multipliers, 915–916
material properties, thicknesses and
O sectional properties, 916–917
Objective function, 1296 offset nodes, 917–918
Observability, 519 Parametric model, 13
matrix, 1284 Parseval’s theorem, 302
Index 1419

Partial coherence, 282, 283 FEA mode shapes, 1130

function, 289 modal model, 1136
Partial fraction Point tracking, 413
expansion model, 1108 Poisson’s ratio, 700
residue model, 559 Polarization, 107, 110, 111, 124, 131
Partial modal assurance criterion (PMAC), 668 Polarizing beam splitter, 111, 113
Partitioned eigenvectors, 1190 Poles, 1290, 1293, 1298, 1299, 1301, 1302,
Peak hold, 296 1329
Penalty function methods, 899, 900 surface consistency plots, 641–643
Performance metrics, 1015 weighted vector, 641
Periodic chirp signal, 330 Poly-reference frequency domain (PFD), 23
Periodic random excitation, 442 Polyreference Least Squares Complex
Periodic random signal, 324 Frequency (LSCF), 535
Periodic signals, 152, 164, 176–177 Poly-reference time domain (PTD) approach,
Periodogram, 253, 1270–1274, 1326 22–23
correlation estimate, 255–256 Polyester adhesives, 53
Periodogram-based spectrum and correlation Polyimide adhesives, 53
estimation, 253 Polynomial coefficient estimation, 595
Permanent Ambient Vibration Monitoring, Polynomial function, 138
1314–1318 Polyreference frequency domain (PFD), 536
Permanent displacement, due to damage, 1320 Polyreference time domain (PTD), 535, 639
Personal computer (PC), 7 Poppet valves, 132
Perturbed boundary condition, 344 Position sensor, 136, 142
Phase demodulation (PDM), 116 Potential energy, 713, 742
Phase frequency response, 747, 769 Power cepstrum, 208
Phase lock loops (PLLs), 116, 117 Power reflectivity, 115
Phase-modulation, 108 Power spectra, 268, 1271
Phase resonance, 13, 621 Power spectral averages, 320
methods, 539 Power spectral density (PSD), 165–166, 806,
Phase separation, 13, 621 980–982, 1003
methods, 539 Precision, 1327
Phase-shift, 742 Prepolarized microphones, 58–59
Phase shifting, 111, 112 Pressure microphone, 71
Phase spectrum, 248 Pressure transducer, 61–65, 73–74
Phase stepping, 435 Pressure variations, 720–722
Photodetector, 108–111, 113, 119, 121, 125, Primary domains, 1380, 1381
134, 135 Principal component analysis (PCA), 283, 290,
Photogrammetry, 411 604, 1039, 1262, 1316
point tracking, 413 Principal response functions, 606
target-less approaches, 415 Progressive damages, 1251
See also Digital image correlation (DIC) Projection methods, 1039
Physical modes, 1318 Propagation time delay, 274
Piezoelectric accelerometers, 65–68, 687, 999 Proportional damping, 620, 656, 1157
Piezoelectric ceramics, 49 matrix, 1115
Piezoelectric crystal technology, 8 model, 1115
Piezoelectric effect, 686 Prototype, 1296
Piezoelectric energy, 685 Pseudo cross orthogonality, 657
Piezoelectric force gauge, 68 Pseudo orthogonality, 880, 882, 883, 888
Piezoelectricity, 47–49 Pseudo orthogonality check (POC), 656, 669,
Piezoelectric strain gages, 54 880
Piezoelectric transducer, 62, 65 Pseudo-random, 399
Piezoelectric transduction element, 48 excitation, 442
Plate signal, 323
EMA mode shapes, 1128 Pumping losses, 715
1420 Index

Pure random signal, 323 Reduced order component models, 1213

PZT transducers, 1182 Reduced order model, 859, 860, 862–865
Reduced order system models, 1218–1220
Reduced shear speed, 762
Q Reduced structural matrix approach, 26
QR transform, 1311 Reduced viscosity, 762
Quadratic eigenvalue problem, 1194 Reduction methods, 834
Quadrature mixing, 118 Reference-based filtering, 172
Quantification of margin and uncertainty Reference energy, 688
(QMU), 610 Reference sensors, 1311
Quantization interval, 119 Reference signal, 116, 118, 137, 143
Quarter-wave plate, 111, 113 Refractive index, 115, 123, 124
Quartz, 47 Regularization, 909–914
Quasi-periodic signals, 164 Relative displacement sensors, in civil testing,
Quasi-static methods, 758 1255
Relative history, 291
Relative motions, 686
R Relative spectrum, 291
Radial basis function (RBF), 1040 Relativistic approach, 106
Radiation efficiency, 721–725, 728, 729, 731, Relaxation, 694–696
733 function, 691, 692, 695, 696, 703, 750, 758
Radiation loss factor, 721–722, 726, 734, 735 test, 691, 702, 758, 759
Railway bridges, 1331 time, 694, 700, 709
Random, 497 Residual, 756
errors, 249, 268, 297, 310 flexibility, 562
excitation methods, 315 from FRFs, 561
excitation signals, 314 inertia, 562
incidence microphone, 71–72 kinetic energy surface, 1367
noise, 95, 1103 terms, 561
signal, 321 Residual mode augmentation (RMA),
Random decrement, 1267 1371–1375
analysis, 312 Residue (modal vector) estimation, 590–595
averaging, 22, 312 Residues, 558
Random vibration, 369, 406 from IRFs, 560
control, 381 from multiple reference FRFs, 559
Random vibration testing from single reference FRFs, 558
autospectral density, 376–380 Resistance technology, 7
random test control loop, 365–369, 380 Resolution, 116–121, 140
test conduct, 384–386 Resonance-assisted vibration, 1104
test setup and instrumentation, 382–384 Resonance cavity, 105
Rank of matrix, 1392 Resonance curve of a mode of vibration, 1108
Rapidly changing speckles, 126 Resonance frequencies, 767, 772, 777
Rate-independent friction forces, 717 Resonance test, 767, 772, 773
Rational fraction polynomial (RFP), 535, 1108 Response (exponential) windows, 332
Rayleigh damping, 537, 620, 751, 753 Response model, 1156, 1164
Rayleigh’s criterion, 291 Response-only modal analysis, 535, 617
Rayleigh wave dispersion measurement, 779 Response Vector Assurance Criteria (RVAC),
Rayleigh-Ritz approach, 837 884
Real normalization, 646 Restoring force method, 967–970
Real symmetric matrices, 1393 Retro-reflective tape, 122
Real time analyzers (RTA), 11 Reversal points, 705, 708
Reciprocity, 501 Reverse path method, 978–983
Rectangular window, 241 Reynold’s Equation, 739
Recursive estimation, 1019–1020 RF interference, 97
Index 1421

RF telemetry systems, 98 finite difference transformation matrices

Rheological N -parameter model, 692 for, 1174–1176
Rheological model, 689, 692, 693, 697, 700, Second order modal truncation, 1191
705, 710–711, 754, 785 Second stage of modal parameter estimation
RIB hybrid modal model, 1135 (MPE-2), 646
RIB impact test, 1133 Seismic, 779
Right-hand eigenvectors, 753 Selective excitation, 346
Rigid body correction, 454 Self-mixing (SM), 114–116
Rigid body mode, 344, 506 SEM Testbed, 1065
Ritz method, 1353, 1371 Semiconductor strain gage, 53
RMS spectrum, 246 Sensing technologies, 47
Road traffic, 1329 capacitive accelerometers, 59
Robustness, 1327 piezoelectricity, 47–49
Root mean square (RMS) averaging, 295, 1329 See also Condenser microphonecondenser
Rotating optical measurements microphones, 57
camera setup, 453 Sensitivity analysis
measurement, duration and shutter time, frequency-domain displacement response
frequency of, 452–453 residual, 908–909
mode extraction challenges and effects of procedures, 1371
harmonics, 454–455 undamped eigenvalue residuals, 906
rigid body correction, 454 undamped mode-shape residuals, 907
Rotational compliance, 1171 Sensitivity
Rotational degrees of freedom, 1080, 1105, civil structures, 1316
1164–1166, 1197 errors, 1087
direct techniques for, 1182–1184 matrix, 903
indirect technique for, 1166–1182 method, 899
in investigational fields, 1165, 1166 Sensor(s), 1251–1256, 1290
moment excitation, 1184–1186 bonded strain gage technology, 8
Rotational mobility, 1170 digital technology, 11
Rotational modal deflections, 1179 DOF, 1085
Rotational receptances, 1165, 1166 dynamic models, 60–65
Rotational responses, 1159, 1170 IEPE/CCLD/ICP/Deltatron/Isotron style
Rotational vibrometers, 132–134 of, 1099
Roving accelerometer, 501 inductance and capacitance technology, 9
Roving hammer, 501 installation related design considerations,
Row space, 1392 90–93
Rubin and Craig-Chang methods, 1080 linearity, 1100
Rubin method, 844 mounting, 1101
Rule of Reciprocity, 1166 optical technology, 9–10
piezoelectric crystal technology, 8
resistance technology, 7
S selection and use, 65–74
Safety, 128, 129, 144 transverse sensitivity of, 1100
Sample-hold, 125 Sensor system, 74
Scaled modal assurance criterion (SMAC), 668 analog to digital converter, 75
Scan frequency, 138, 139 filtering, 75
Scanning Laser Doppler Vibrometer (SLDV), impedance buffer, 75
1181 transducer, 75
Scanning vibrometers voltage gain, 75
CSLDV, 138–140 Shaker excitation, 314
optical derotator, 140–141 Shaker testing, 440–443
step-scan vibrometers, 135–137 Shannon’s sampling theorem, 291
TLDV, 141–142 Shape bases, 1082
Second-order approximation, 1176 Shape functions, 786
1422 Index

Shear modulus, 700, 701, 749, 759–760 Single degree of freedom system (SDOF)
Shear stress, 708, 712, 759 model, 1344
Shift function, 703, 704 Single input relationship, 19
Shock and vibration symposium (SVS), 29 Single-input/single-output (SISO), 370
Shock response spectrum (SRS), 351, 372, 388 frequency, 267
function, 1350–1351 Single-mode, 131
Shock testing Single point vibrometers, 129–131, 136
characteristics and limitations, 387 Singular value, 289, 290
classical theory, 396–397 spectra, 1326
classical waveforms, 386 Singular value decomposition (SVD), 285,
decaying sinusoids, 389–393 604, 605, 608, 635, 653, 654, 1275,
electrodynamic and electrohydraulic 1284, 1393
shakers, 386 methods, 536
oscillatory waveforms, 389 Sinusoidal input-output method, 15–18
peak-velocity capability, 387 Skew Hermitian matrix, 1389
shock response spectrum, 388 Skew Symmetric matrix, 1389
shock-response spectrum, 387 Slip effects, 686
synthesis, 394 Slow random signal, 326
system frequency response function, Slow swept sine signal, 330
398–401 Smoothed periodogram, 254
transient waveform, 397 Society for Experimental Mechanics (SEM), 8
waveform classification, 388 Society for Experimental Stress Analysis
Short circuiting, 725 (SESA), 8
Short dimension, 543, 626 Society of Automotive Engineering (SAE), 29
Short time Fourier transform (STFT), 185–186, Soil damping, 778–779
1005 Soil freezing, 1316
Shot-noise, 119–121 Sommerfeld number, 737
Signal classes, 232 Sonar based sensors, 1256
periodic signals, 232 Sound and Vibration Magazine, 30
random signals, 233–235 Sound field, 722, 729
transient signals, 236 Sound power, 722–725, 733
Signal clipping, 94 Space shuttle dynamic model, 1360
Signal conditioning, 76 Sparsity pattern of stiffness matrix, 850
charge amplifier, 76 Spatial incompleteness, 1165
differential amplifier, 77 Spatial model, 1155
voltage amplifier, 76 Speckle, 122–123, 125, 126, 138, 140
Signal-strength, 125 Spectral analysis, 230
Signal to noise, 331 Spectral correlation (SC), 163
Signal to noise ratio (SNR), 116, 120, 130, density, 197–200
135, 294, 326, 958 Spectral decomposition, 1392
Signal type, 319 Spectral density, 368
Silhouette (SIL) index, 1301, 1302 function, 368
Silhoutte index, 1311 Spectrogram, 1264, 1265
Silicon strain gauges, 74 Spectrum analysis, principle for, 237–238
Silicone oils, 762 Spike, 125, 126
Simple Bearing Theory, 737 Spring-pots, 785
Sine excitation, 441 Spurious poles, 1293–1294, 1297
Single input, single output measurements, 280 Square root of the modal assurance criterion,
Single cluster, 1298 668
Single degree of freedom (SDOF), 20, 352, Squeeze-film dampers, 737, 739–741
622–623, 652, 653 SSI-COV, 1331
algorithms, 566 Stability, see Hard stability criteria
methods, 583 Stability diagram, 582, 1288, 1297, 1302,
structure, 352–360 1330, 1331
Index 1423

Stability/stabilization diagrams, see Strain frequency response function (SFRF)

Consistency diagrams matrix, 450
Stabilization diagram, 1294, 1302, 1303 Strain gages
Stable averaging, 296 in civil testing, 1254
Stand-off distance, 121, 135 frequency response, 60
State change, 1318 metal, 49–53
State estimation problem, 518 piezoelectric, 54
State-space formulation, 1194 semiconductor, 53
State vector, 550, 641 Strain gauge(s), 499
modal assurance criterion, 665–667 transducers, 1182
Static bearing, 738, 739 Strain tensor, 699
Static hysteresis, 705, 709–711, 719 Stress field, 756
Static reduction technique, 516, 1187 Stress tensor, 697
Stationary methods, 779 Structural damping, 488, 686
Stationary random signals, 323 Structural dynamic(s), 412, 415, 420, 427, 431,
Stationary signals, 157–158 451, 459, 463
Statistical averaging methods, 720 in frequency domain, 356–360
Statistical model development, mechanical shock, 372–375
994–995 random processes, 365–369
Statistical pattern recognition, SHM random vibration, 369–371
cleansing, 993 single-degree-of-freedom, 352–356
data acquisition, normalisation and in time domain, 361–365
data-driven models, 1027 Structural dynamic models, 509, 1104–1106
feature dimensionality, 1035–1036 automated response DOF selection,
feature discrimination, 1026–1027 1366–1369
feature extraction and information dynamic bandwidth, 1350–1351
condensation, 993–994 measured data acquisition, measured data
inclusive outliers, 1031–1032 analysis and experimental modal
operational evaluation, 992 analysis, 1368
outlier analysis for damage identification, modal analysis, 1348
1028–1031 modal test planning and test-analysis
statistical model development, model, 1364–1365
994–995 modelling guidelines, 1351–1352
Steady state, 706, 712, 754, 767 system dynamic model, 1363–1364
random excitation, 399 target modes, 1366
Step relaxation, 317, 485, 491 thoughts on, 1352
signal, 334 verification and validation, 1362–1375
Step signal, 316 Structural dynamic modification (SDM), 1165,
Step-scan vibrometers, 135–137 1216, 1372
Stepped-sine excitation, 497 advantage, 1104
Stereophotogrammetry, 412 calculating new modes, 1136
Stiffness matrices, 686, 748, 749, 774 design modifications, 1096
Stochastic model updating, 901, 923–930 EMA modal model, 1098–1104
Stochastic signal modelling and experimental vs. EMA modes, 1137–1141
signals vs. FEA modes, 1136–1137
cyclo-non-stationarity, 160–163 example, 1127–1128
cyclostationary signal model, 158–161 inputs, 1097
stationary signal model, 157–158 modal models, 1095
time-varying distributions, 155–157 mode shapes, 1115–1118
transient signal model, 163 tuned vibration absorber modeling,
Stochastic Subspace Identification (SSI), 1263, 1141–1144
1278–1282, 1306 UMM mode shapes, 1118–1124
Storage modulus, 696, 702, 703, 760, 763 unmodified structure, modal model of,
Storey drifts, 1320 1118
1424 Index

Structural health monitoring (SHM), 1309 System equivalent reduction expansion process
acoustic emissions, 1011–1013 (SEREP), 866–868, 1190, 1216,
challenges in data-driven, 1036 1222, 1232–1234
in changing environmental and operational System identification, 1318
conditions, 1036–1047 methods, 1262–1287
compressive sensing, 1016–1018 System impedance matrix, 546
condition monitoring, 992 System-level characteristic constraint (S-CC)
definition, 991 modes, 853–854
forward model-driven methods, 1049–1051 System matrix, 263, 546
fundamental axioms, 995–996 System modelling procedures, 1216
guided waves, 1013–1015 System properties, 261
historical overview, 997–998 System response characteristics, 1213
inverse model-driven method, 1047–1049 System transfer function matrix, 262
Kalman filter, 1020–1022
linear and nonlinear cointegration,
1041–1043 T
non-destructive testing/evaluation, 992 Target mode partitions, 521
nonlinear cointegration, 1046 T-block technique, 1167, 1168
objectives for, 991 Technical rubber materials, 709
origins of, 991 Temperature dependence, 702
performance metrics, 1015 Temporary sensor testing, 1311–1314
physics-based models in, 1047–1051 Test-analysis correlation, 518
recursive estimation, 1023–1026 Test-analysis model (TAM), 517, 525,
Structure-borne-sound radiation, 686 1364–1365
Suboptimal, 520 Test design, 484, 490, 495, 505, 529
Subranging A/D converter, 88, 89 Test plan, 483, 497, 503, 506, 513, 525–529
Subspace, 606 Test Response Assurance Criteria (TRAC),
iteration method, 1188 885, 1220, 1225
Substructure modal model, 1135 Testing configurations, 261
Substructure reduction method, 837 Thermal drift, 100
Dual Craig-Bampton method, 846–848 Thermal energy, 685
free interface modes, 842–844 balances, 769
Hurty/Craig-Bampton method, 839–842 Thermo mechanical coupling, 686
interface reduction approach, 852–853 Thermometers, 1256
MacNeal method, 845–846 Threshold, 125
numerical examples, 848, 851 Tilt-meters, 1319
numerical techniques, 839–848 Time averaging, 294
Rubin method, 844 Time dependent, 696, 743, 754, 783
system-level characteristic constraint Time domain, 712, 755, 757, 775, 787, 1380
(S-CC) modes, 853–854 dynamic model, 1106–1107
Substructuring approach, 1158 methods, 1313
Substructuring methods, 1161 model, 554
Sum of weighted accelerations technique signals, 94–100
(SWAT), 1193–1194 Time invariance, 264
Superposition principle, 704 Time series models, 1008–1009
Supervised learning, 994 Time shift function, 704
Support conditions, 504–513 Time synchronous averaging (TSA), 168, 203,
Suspended Bridge, see Bridge, suspended 25 205
de Abril Time window, 240
Symmetric locations, 291 Time-frequency analysis, 1005–1006
Symmetric matrix, 1389 Tip Excitation Technique (TET), 1171
rules, 1389–1391 Toeplitz matrix, 1283, 1390
Symmetric modes, 275 Total least squares (TLS), 266, 536
Synchronized hammers, 1184, 1185 approach, 591
Index 1425

Total loss factor, 740 Twin-shaker configurations, 1185

Tourmaline, 47 Two-degree-of-freedom (2DOF), 801
Trace, 1391
Tracking CSLDV, 142 U
Tracking filter, 125 Unbounded space, 754
Tracking laser Doppler vibrometry (TLDV), Uncertainty, 126, 128, 135
141–142 quantification, 900
Traditional multiple coherence function, 283 Uncoupling techniques, 1162–1164
Trail, 640 Undamped eigenvalue residuals, 906
Transducer(s), 499 Underdetermined equations, solution of,
connectors, 93 1396–1397
Transfer function, 1383 Unified matrix polynomial approach (UMPA),
matrix, 263, 546 536, 555, 618, 624
model, 1107 Uniform window, 301, 306
Transfer function analyzer (TFA), 10, 11 characteristics, 304
Transfer path analysis (TPA), 172 Unit modal mass, 1118
Transformation methods, 608 mode shapes, 1118–1124
Transformed equations of motion, 1112 Unscented Kalman filter (UKF), 1024
Transient control, 396, 404 Unsupervised learning, 994
Transient deterministic signal, 331 Upper triangular matrix, 1396
Transient random signals, 157, 163 UPS, 1257
Transient signals, 152, 153 U.S. aerospace industry
Translational modal deflections, 1179 structural dynamic models, 1348–1352
Translation FRFs, 1197 verification and validation, structural
Transmissibility, 1198 dynamic models, 1362–1375
applications, 1202
functions, 1201
fundamental formulation, 1199–1200 V
generalization of, 1199 Validation, 931–935
matrix, 1199, 1200 Van der Monde form, 587
properties, 1201–1202 Van der Monde matrix form, 596
Transmission simulator, 1079, 1082 Vandermonde matrix, 1391
approach, 1081 Variational principles, 1346
Transmission simulator (TS) theory Vector complexity, 647
CMS, 1086–1088 Vector correlation, 879–889
FBS, 1088 Vector error-correction model (VECM), 1043
implementation, 1084 Vector expansion, 870–876
TS approach, 1089–1091 process, 859
Transposition rules, 1389 Vector normalization, 645
Transverse mode, 105 Vector space, 1392
Transverse sensitivity, of sensor, 1100 Velocity-dependent, 710
Trial functions, 754, 757 Velocity-independent, 705, 710
Trial/shape functions, 748 Velocity superposition, 106
Triangular FRF measurement, 1122 Vibration amplitudes, 375
Triaxial sensors, 521 Vibration-based SHM, 999
Triboelectric effects, 93 autocorrelation function, 1001
Trigger condition, 298 coherence, 1004
Trigger function, 312 damage-sensitive feature, 1000–1004
Triggering issues, 297 frequency spectrum, 1002–1004
Truncated modal model, 1097, 1105, 1140 modal analysis-based features, 1011
Truncated modal system, 1191 time series models, 1008–1009
Truncated modes, 1162 time-frequency analysis, 1005–1006
T-square-block, 1169 time-scale analysis, 1006, 1007
Twin-shaker arrangements, 1184 ways for measuring, 999–1000
1426 Index

Vibration measurements, 105 Weighted modal assurance criterion (WMAC),

Vibro-acoustic modulation (VAM), 970–972 668
Virtual coherence, 285 Weighted modal vector orthogonality, 656–658
Virtual DOFs, 603 Weighting function, 754, 782
Virtual forces, 285 Welch method, 1273
Virtual FRFs, 606 Welch’s correlation function estimates,
Viscoelasticity, 749 251–252
Viscoplasticity, 716 Welch’s method, 249
Viscous damping model, 1106 Wheatstone Bridge, 1254
Visualization set, 523 White noise, 312, 316
Voltage amplifier, 76 Wide-sense stationarity (WSS), 158
Voltage controlled oscillator (VCO), 116 Wigner-Ville spectrum (WVS), 187–188
Voltage feedback, 326 Wigner–Ville distribution (WVD), 186–188
Volterra model, 961 Wind turbines, 498
Window function, 320
W Window scaling factors, 243
Ward rule, 1297 Wireless distribute sensors, 1311
Wavefronts, 109, 112 Within-cluster distance, 1294
Wave propagation, 684, 685 WLF function, 704
Wave velocity, 735
Wavelets, 188 Y
constant percentage bandwidth, 189–190 Young’s modulus, 709, 735
packets, 190
Wavenumber, 108
WAVSYN pulse, 389 Z
Weather stations, 1256, 1257 Zero padding, 243, 252
Weighted least squares (WLS), 591 Zoom Fourier transform, 299