Structural Dynamics and Linear Systems

Compute, Test, Calibrate and Validate

Thomas Abrahamsson
Copyright p 2019 Thomas Abrahamsson
P UBLISHED BY C HALMERS U NIVERSITY OF T ECHNOLOGY
ARCHIVE . ORG
Licensed under the Creative Commons Attribution - NonCommercial 3.0 Unported License (the
“License”). You may not use this file except in compliance with the License. You may obtain a
copy of the License at http://creativecommons.org/licenses/by-nc/3.0. See the License
for the specific language governing permissions and limitations under the License.

First printing: August 12, 2019

 Preface

Working with practical and theoretical vibrational problems I have come to realize the importance
of linear system theory in vibrational engineering. In traditional vibration engineering education
material, much of the modern linear system theory is left unnoticed, since the focus is usually on
modeling and analysis of linear/non-linear elements/components/structures to which the second
order differential equations provided by Newton’s second law provides a theoretical base. That
base in itself has been so strong that the need to take inspiration from the first-order description
commonly used in linear system theory field has not been strong. A book by Meirovitch [30] is an
exception from the rule. One intention of this book is to link vibration theory and more general
system theory even further.
In this treatise I have not striven to be complete, i.e. to include and compare all available
methods and techniques, neither in a historical sense or in current practice. I am restricting the
presentation to those methods and tools that I have found to be efficient and sufficiently accurate
for real life vibrational problems in my own work. Those interested in comparative studies must
seek such in other sources or build personal experience from their own work. The book starts with
a short chapter on structural mechanics but is basically aimed for those already familiar with basic
solid/structural mechanics and vibrational theory. I have tried to strike a good balance, and neither
over-explaining nor under-explaining matters. My strive has been to produce a material which I
myself would have appreciated as a student eager to learn more on vibrations and related matters.
The familiarity with matrices and matrix operations is crucial.
Besides modelling, analysis and computation, much of practical vibrational engineering is
related to dynamic testing. A testing that can have a value on its own or be used in conjunction with
modelling and analysis in the validation and substantiation of computational models. Here the exper-
imental modal analysis and system identification play important roles. To understand the underlain
techniques and principles on which these relies, a dose of linear system theory is very helpful.
The well-developed theories for linear system identification and the linear mathematical/numerical
models that the identification provides gives a solid base for comparison between the real-world
testing and desktop modelling and analysis. These comparison can be restricted to comparisons
with diverse correlation metric or brought further with the parameterization and calibration och the
4

computational models. Another intended purpose of the book is to present methods for which such
validation and calibration can be done efficiently also for very large computational models.
Theories by themselves may be interesting, but when implemented in practical useful tools
they may become valuable. The practical use of linear system theories has been strongly related
to available computational resources. During the last decades the implementation of those has
been very much simplified by the introduction of high level computer languages such as M ATLAB.
The reader of this book is strongly advised to test the methods presented here in a M ATLAB or
M ATLAB-like environment to increase insight without too much programming effort.

Thomas Abrahamsson, Göteborg, August 12, 2019

 Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

I Computational Structural Dynamics

1 Introduction with Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.1 Motivation 13
1.2 Notation 16
1.2.1 General notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Fundaments of Linear Structural Dynamics . . . . . . . . . . . . . . . . . . . . . 19

2.1 Basic concepts and principles 19
2.1.1 The particle and its degrees-of-freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.2 Linear kinematics. Vector transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.3 Particle joining. Compatible and constrained motion . . . . . . . . . . . . . . . . . . . 23
2.1.4 Dynamic equilibrium. Newton’s second law and Euler’s rotation equation . . . 23
2.1.5 Linear constitutive relations for materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 The stiffness method 26
2.2.1 System assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Degree-of-freedom reordering and partitioning . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.3 Prescribed displacement. Fixed boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.4 Static condensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 The finite element method 31
2.3.1 FE modelling of stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.2 FE modelling of mass inertia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.3 One-dimensional FE elements for trusses and frames . . . . . . . . . . . . . . . . . . . . 33
2.3.4 Two-dimensional FE elements for plane stress and plane strain . . . . . . . . . . . . . 38
2.3.5 Three-dimensional FE element analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Problems 41

3 Linear State-Space Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3.1 The state-space formulation 45
3.1.1 Structural dynamics in state-space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 State-space realization forms 48
3.2.1 State-space realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3 Problems 53

4 Decoupling of System States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.1 Modal decomposition 57
4.1.1 Decomposition of the undamped 2nd order structural dynamics equation . . 57
4.1.2 Rayleigh quotient. Rayleigh’s and Courant’s eigenvalue theorems . . . . . . . . . 59
4.1.3 Decomposition of the damped 2nd order structural dynamics equation . . . . 63
4.1.4 Spectral decomposition of stiffness and mass relations . . . . . . . . . . . . . . . . . . . 66
4.1.5 State-space modal decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Eigenvalue enclosure methods 69
4.2.1 Gerschgorin’s disks for eigenvalue enclosure . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.2.2 Givens’ method with Sturm sequence checking . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.3 Wittrick-Williams eigenvalue counting algorithm . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 Matrix iteration 76
4.3.1 Inverse iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.2 Lanczos’ method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4 Matrix decomposition and transformation 83
4.4.1 LDL0 decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4.2 Householder transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Problems 86

5 Time Domain Solution Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

5.1 Continuous time solution for structural dynamics equation 93
5.1.1 Viscously damped single-degree-of-freedom systems . . . . . . . . . . . . . . . . . . . 93
5.1.2 Viscously damped multi-degree-of-freedom systems . . . . . . . . . . . . . . . . . . . . 98
5.2 Continuous time solution for the state-space system 98
5.2.1 State transition matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2.2 State observability, controllability and Grammians . . . . . . . . . . . . . . . . . . . . . 100
5.2.3 Checking observability and controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2.4 Continuous-time Markov parameters and the Hankel matrix . . . . . . . . . . . . . 106
5.3 Response bounds 106
5.3.1 Worst-case forcing function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.4 Numerical discrete time solutions 109
5.4.1 The second order mass, stiffness and damping system . . . . . . . . . . . . . . . . . . 109
5.4.2 The first order state-space system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.3 Stability of zero-order-hold time integration . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4.4 Computation of the discrete-time transition matrix . . . . . . . . . . . . . . . . . . . . . 115
5.4.5 Discrete time observability and controllability . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4.6 Hankel matrix and Markov parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.5 Problems 119

6 Frequency Domain Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

6.1 Frequency response 125
6.1.1 Transfer functions from the structural dynamics equation . . . . . . . . . . . . . . . . 126
6.1.2 The realness of loading and response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.1.3 Modal summation of frequency response functions . . . . . . . . . . . . . . . . . . . . 128
6.1.4 State-space models in frequency domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.2 Exact dynamic condensation 134
6.2.1 Exact condensation of discrete system dofs . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.2.2 Continuous systems. Exact finite elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.3 Parseval’s theorem 141
6.4 Problems 143

7 Model Reduction and Substructuring . . . . . . . . . . . . . . . . . . . . . . . . . 147

7.1 State transformation 147
7.2 Modal reduction methods 148
7.2.1 Mode displacement method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.2.2 Mode acceleration method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.3 Substructuring methods 152
7.3.1 The Guyan method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7.3.2 The Craig-Bampton method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.3.3 Surrogate parametric model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.4 State-space reduction methods 158
7.4.1 State-space reduction based on transfer strength . . . . . . . . . . . . . . . . . . . . . 158
7.4.2 State reduction by use of Grammians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.5 Problems 162

II Testing and Test Data Driven Modeling

8 Modal Analysis and System Identification . . . . . . . . . . . . . . . . . . . . . 169

8.1 Experimental modal analysis 169
8.1.1 Theoretical foundation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.1.2 Mobility circle fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.2 Introduction to State-Space System Identification 173
8.3 State-space subspace identification 174
8.4 Problems 179

9 Correlation and Comparison Metric . . . . . . . . . . . . . . . . . . . . . . . . . . 181

9.1 Vector correlation metric 181
9.1.1 Modal assurance criterion - MAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.1.2 Coordinate Modal Assurance Criterion - COMAC . . . . . . . . . . . . . . . . . . . . . 183
9.1.3 Modal observability correlation - MOC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
9.2 Data correlation metric 185
9.2.1 Frequency response metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.2.2 Mode indicator functions - MMIF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
9.3 Experimental mode expansion 189
9.4 Vector and matrix norms 191
9.5 Problems 193

10 Data Driven Substructuring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

10.1 State-space model on coupling form 195
10.1.1 Physically motivated modeling constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.2 System coupling 199

11 Vibration Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

11.1 Planning sensor and actuator placement 204
11.2 Testpiece excitation and response data processing 205
11.2.1 Periodic excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11.2.2 Aperiodic excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.2.3 Signal processing of periodic signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.2.4 Data processing caveats and remedies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
11.2.5 Testpiece support system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
11.3 Vibration testing hardware 219
11.3.1 Accelerometers for vibration testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
11.3.2 Force transducers for vibration testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
11.3.3 Laser Doppler vibrometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
11.3.4 Hardware Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
11.3.5 Some Recent Hardware Trends in Vibration Testing . . . . . . . . . . . . . . . . . . . . 231
11.3.6 Some future perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
11.4 Problems 234

III Model Calibration and Validation

12 Validation and Calibration Concepts . . . . . . . . . . . . . . . . . . . . . . . . . 239

12.1 Models and model structures 241
12.1.1 Model validation and falsification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
12.1.2 Model calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
12.1.3 Model verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
12.1.4 Experimental support for validation and calibration . . . . . . . . . . . . . . . . . . . . 246
12.2 Problems 248

13 Property Variability and Model Uncertainty . . . . . . . . . . . . . . . . . . . . 251

13.1 Real world variability 252
13.1.1 Individual Spread - A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
13.1.2 Random outcomes, probability and basic statistics . . . . . . . . . . . . . . . . . . . . 260
13.1.3 Random sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
13.2 Parameter estimation statistics 267
13.2.1 Parameter identifiability and model distinguishability . . . . . . . . . . . . . . . . . . . 270
13.3 Model structure selection 272

14 Model Calibration Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

14.1 Minimizing a single-variable-function 275
14.2 Minimizing a quadratic functional 277
14.2.1 Newton’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
14.2.2 Gauss-Newton’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
14.2.3 Levenberg-Marquardt’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
14.3 Computational aspects 280
14.3.1 Gradients from finite differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
14.3.2 Parameter settings and randomized starts . . . . . . . . . . . . . . . . . . . . . . . . . . . 281

15 Validation and Cross-Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

15.1 Classical validation 283
15.1.1 Pragmatic validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
15.1.2 Validation guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
15.2 Cross-validation 287
15.2.1 K-Fold cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
15.2.2 Bootstrapping cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
15.2.3 Monte-Carlo cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
15.2.4 Cross-validation estimates of parameter statistics . . . . . . . . . . . . . . . . . . . . . . 289
15.2.5 Statistical validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
I
Computational Structural
Dynamics

1 Introduction with Notation . . . . . . . . . . . . 13

1.1 Motivation
1.2 Notation

2 Fundaments of Linear Structural Dynamics

19
2.1 Basic concepts and principles
2.2 The stiffness method
2.3 The finite element method
2.4 Problems

3 Linear State-Space Models . . . . . . . . . . . 45

3.1 The state-space formulation
3.2 State-space realization forms
3.3 Problems

4 Decoupling of System States . . . . . . . . . 57

4.1 Modal decomposition
4.2 Eigenvalue enclosure methods
4.3 Matrix iteration
4.4 Matrix decomposition and transformation
4.5 Problems

5 Time Domain Solution Procedures . . . . 93

5.1 Continuous time solution for structural dynamics
equation
5.2 Continuous time solution for the state-space system
5.3 Response bounds
5.4 Numerical discrete time solutions
5.5 Problems

6 Frequency Domain Solutions . . . . . . . . 125

6.1 Frequency response
6.2 Exact dynamic condensation
6.3 Parseval’s theorem
6.4 Problems

7 Model Reduction and Substructuring 147

7.1 State transformation
7.2 Modal reduction methods
7.3 Substructuring methods
7.4 State-space reduction methods
7.5 Problems
 1. Introduction with Notation

1.1 Motivation
It is safe to say that all physical systems are non-linear. Stressed by a sufficiently strong stimuli, all
systems react in a way that the superposition principle of linear systems is violated. For smaller
stimuli levels however, the system’s behavior is often such that a linear mathematical model captures
its essential characteristics to sufficiently accuracy for an application in mind. Such systems with
sufficiently mild stimuli will be named linear systems and are those treated in this book. The
onset to nonlinear behavior is sometimes sudden and drastic and should not be ignored if relevant.
However, a good understanding of linear system characteristics is a good foundation for further
studies of non-linear phenomena. Non-linear behavior, such as chaotic motion, mode saturation
and sub- or super-harmonic resonance are thus not treated in this text and information about these
found elsewhere.
A further sub-classification of linear systems, into time-invariant and time-varying linear
such, is usually made. This classification is justified, since all the solution methods applicable to
time-invariant systems do not apply for time-varying systems. Physical systems are often slowly
time-varying such that a short time-scale governs their vibration characteristics and a longer time-
scale governs their time-varying properties. Over a sufficiently short duration of time such systems
can be approximated as time-invariant. The experience from vibration tests is that results on the
same test object can rarely be reproduced from one day to another, except under very controlled
environmental conditions. The reason is usually that environmental conditions, such as temperature
and humidity, vary slowly with time and that affects the mechanical properties of the system under
study which thus becomes time-varying. The slowness of these environmental processes in relation
to the rather short time span of the simulations for the dynamic problems most often at hand
motivates the focus in this text which is on linear time-invariant systems.
Mathematically, a system is said to be linear if it satisfies the homogeneity and additivity
properties such that

• If r1 is the response to stimulus s1 and s2 is the response to stimulus u2 then the response to
s1 (t) + s2 (t) is r1 (t) + r2 (t). From this it follows that:
14 Chapter 1. Introduction with Notation

• If the response to stimulus s(t) is r(t), then the response to αs(t) is αr(t), where α is an
arbitrary constant.

It is these properties, also known as superposition properties, that may be exploited so effectively
for linear systems. Also, there is a vast amount of results from numerical linear algebra that supports
the development of methods for those. These results have contributed much to the development in
vibrational engineering, control engineering, signal processing and system identification that are
major players in the field of structural dynamics.
In the era before the 1990’s there was a saying in industry that “Nobody trust the result of
analysis - but the analyst himself. Everybody believes in test results - except for the experimentalist
himself”. In modern industry, that saying is not longer fully relevant, and a slow shift of paradigm
has been made since then. Where, in the old days, product development was mainly based on
testing of product prototypes, present day’s product development is mostly based on analysis and
simulation. This is the process known as virtual prototyping. While in the old days, too much
confidence was often put in the outcome of test, in present days too much confidence is often put in
the result of modeling and analysis. All analysts know that the modeling is usually made under lack-
of-information conditions, and more or less justified assumptions have to be made in the modeling.
In the field of structural dynamics this regards material properties, product geometry, properties
of structural joints, boundary conditions and, not to the least extent, the loading conditions. To
make the results of analysis and simulation more credible, the models on which they rely need to
be properly verified and validated [40] .
In the present era in which simulation has a very important role, the computational solid me-
chanics is increasingly important in the design and performance assessment of engineering systems.
Automobiles, aircraft, high-rise buildings and weapon systems are examples of engineered systems
that have become more and more reliant on computational models and simulation results to predict
their performance, reliability or safety. Although important decisions are based on computational
solid mechanics, the credibility of these models and simulations is often not questioned by the
general public, the technologists who design and build the systems, or the decision makers who
commission their manufacture and govern their use. What is the basis if this trust? Both the public
and the decision makers tend to trust graphical and numerical presentations of computational results
that are plausible and that make sense to them. This trust is also founded on faith in the knowledge
and abilities of the engineers and scientists who develop, exercise, and interpret the models. Those
responsible for the computational models and simulations on which society depends so heavily are,
therefore, keepers of the public trust with an abiding responsibility for ensuring the veracity of their
simulation results. They always need to ensure that their approximations of reality are appropriate
for answering specific questions about their engineered system. Primarily, an analyst should strive
to establish the accuracy of a computational model that is appropriate for its intended use. These
accuracy requirements vary from problem to problem and can be influenced by people perception
and economic considerations, as well as by engineering judgement. The truth of a scientific theory,
or of a prediction made from that theory, cannot be proven in the sense of deduction logic. However,
scientific theories and subsequent predictions can and should be tested for trustworthiness by the
accumulation of evidence from reality.
For models to be truly credible, they need to be validated against experimental data obtained
by testing of the real world products they are representing. They, together with the computer
program that runs them, also need to be properly verified to be free from bugs, discretization
errors and other errors. This has been obvious for a long time for safety critical products as
aircraft and spacecraft. For aircraft, the certifying authorities Federal Aviation Administration and
Joint Aviation Authorities, require that models are substantiated by test results to be used in the
certification process. Otherwise the aircraft designed with the help of simulation is not allowed
to fly in commercial traffic. Another product, that is also designed against a small safety factor
1.1 Motivation 15

Figure 1.1: Schematic illustration of the stimuli-to-response relation of a system. In the calibration
setting in this book, the system is assumed to have a not fully known parameter setting, but the load
is assumed to be known.
........................................................................................

to represent unknowns, is the wind turbine. For wind turbine blades, the certifying bodies also
require that validation testing is made as part of commissioning before the wind turbine is set into
production.
Finite element models of mechanical systems are most often developed from first principles.
Often it is based on Hooke’s law for the constitutive relation that governs stiffness and on Newton’s
law for the relation between forces and acceleration. Numerous times it has been found that the
modeling of solid parts that are cut out from a solid piece of material or molded in one piece is
almost trivially modeled to a very high accuracy with the finite element method. That is provided
that the material data and geometry of the parts are available to high precision. However, the
modeling task becomes more involved when parts are put together into an assembly by joining
techniques such as riveting, bolting, gluing or welding. It is often at the modeling of such joints
that the modeler’s experience comes to test. In a complete assembly, a vast part of the contribution
to system damping is often attributed to the physical processes in the joints during the motion of
the structure. It normally takes a very good insight into physics and precise information about the
details of the system to get the damping models right at first time without complementary testing.
The assumption of linearity simplifies matters in the respect that many methods and techniques
become applicable. Such are Fourier decomposition methods that give the duality between time
domain and frequency domain. Another method that is meaningful only for linear models is
the modal decomposition method. Also the powerful state-space sub-space system identification
methods become available as giving a data processing link between the pure time domain test data
and data that can be computed more easily from the finite element model.
Delimitation: This book concerns the validation of structural dynamics models that are linear
and time-invariant. That does not mean that the calibration criteria are linear in parameters. On the
contrary, very few (if any) model parameter affect the calibration criterion function linearly. The
book also just treats the calibration and validation of the loaded system, see Fig. 1.1. It does not do
any attempt to treat the validation of the applied load. Load validation is another important topic
that also needs to be covered as the system response so heavily depends on the applied loading.
 16 Chapter 1. Introduction with Notation

1.2 Notation
1.2.1 General notation

Table 1.1: General notation

Entity Notation Example(s)
Scalar Lower case italic latin a
Complex conjugate Superscript * a∗
Extreme value Superscript ? a?
Approximation of a Tilde ã ≈ a
Data vector Bold faced lower case letter c
Matrix Bold faced upper case letter D
Time domain quantity Plain latin letter e, G
Frequency domain quantity Latin letter with caret H
ĝ, Ĥ
Modal domain quantity Greek letter β,Ψ
Vectors in 2D/3D Arrow over symbol ~
AB
1.2 Notation 17

Table 1.2: Data vector related notation

Entity Explanation
a Plain vector a
ãa Approximation of a
āa Transformation of a
a∗ Complex conjugate of a
{aa} Curly brackets embrace row/column vector data
[aa, b ] Side-by-side concatenation of column vectors a and b of consistent size
{aa; b } Stacked concatenation of column vectors a and b of arbitrary size
aT Transpose of a
aH Conjugate transpose of a (Hermitian transpose)
ai The vector element i of a , i.e. a scalar element
ak Indexed vector a , i.e. k:th vector a
a k0 Gradient of a with respect to k:th variable

Table 1.3: Matrix related notation

Entity Explanation
A Plain matrix A
A
à Approximation of A
A
Ā Transformation of A
A∗ Complex conjugate of A
[AA] Square brackets embrace matrix data
A, B ]
[A Side-by-side concatenation of matrices A and B of consistent row dimension
A; B ]
[A Stacked concatenation of matrices A and B of consistent column dimension
Ai j The element on row i and column j in A, i.e. a scalar element
A: j All elements of column j, i.e. a column vector
A i: All elements of row i, i.e. a row vector
A iJ All elements of row i in column set J
AI j All elements of column j in row set I
A −1 Inverse of A
AT Transpose of A
AH Hermitian transpose (conjugate transpose) of A, i.e. AH = [A A∗ ]T
A −T Inverse of transpose of A , i.e. A −T , [A A−1 ]T
AT ]−1 = [A
A−H −H
Inverse of Hermitian transpose of A, i.e. A , [A A−1 ]H
AH ]−1 = [A
A† Pseudo-inverse of rectangular A (Penrose inverse)
Ak Indexed matrix A , i.e. k:th matrix A
Ai jk The element on row i and column j of k:th matrix A
A k0 Gradient of A with respect to k:th variable
I n×m The n-by-m identity matrix with ones along the main diagonal from upper left
 2. Fundaments of Linear Structural Dynamics

This book is about linear structural dynamics and assumes that the reader has knowledge about
basic solid mechanics and structural mechanics. In linear structural dynamics the fundamental
concept of a structure is that it is an aggregation of parts. These parts are often named components,
substructures, elements or segments in books like this. The linear part of the subject is attributed to
the linearization that is made to form the resulting equation that governs the motion of the structure
when it is subjected to stimulus. When such stimuli act together to create a response that is the
sum of the responses that are created by the stimuli when acting separately, independently of the
magnitude and number of stimuli, the system is said to be linear. The dynamics part of the subject
relate to the time variation and is in opposition to statics. For structural dynamics phenomena like
vibration, wave propagation, shaking, squeek, shock, quake, impact, rattle and the like are relevant.
Observations of these phenomena has resulted in various mathematical representations that could
be said to represent first principles from a platform of mathematical physics. The first principles
that relate to linear structural dynamics are basically five. Two principles relate to Newton’s laws
relating acceleration of mass to force and the relation of forces between bodies in mechanical
interaction. One first principle relate to the compatibility of the motion of parts that are rigidly
joined together. Another relate to the deformation of materia under stress stimulus. A material that
reacts with deformation that is linear in low/moderate levels of stress variation fits well within the
field of linear structural dynamics for which Hooke’s (generalized) law gives a good mathematical
description. The second law of thermodynamics with the principle of conservation of energy is the
fifth. These first principles will be used, explicitly or implicitly, in the following.

2.1 Basic concepts and principles

2.1.1 The particle and its degrees-of-freedom
A particle is a figure-of-thought and as such an aggregate of constituents of ordinary matter that are
rigidly connected to form a unity that occupy some volume in space. Another figure-of-thought is
that there are material points that form those constituents. The minimal set of variables that can
describe the motion in 3D space of that particle is called the particle’s degrees-of-freedom (dof). If
there is no restriction imposed on the motion of the particle, it has been proved that the minimum
20 Chapter 2. Fundaments of Linear Structural Dynamics

number of such dofs of the particle is six, three translational dofs that describe the position of the
particle in a global coordinate system and three rotational dofs that describe its angular orientation
in space. To a particle we can associate a local coordinate system xyz with origin at position O in a
global (inertial) coordinate system XY Z. With the transformation R that transforms from local to
global coordinates, the global location P of a material point within the particle, located at a point
given by the local position vector p , can thus be found at P = O + R p in the global coordinate
system.
The transformation matrix R can also be used for transformation of any vector quantity given
in a local coordinate system in its transformation to a global coordinate system. In the context of
this book it relate to position vectors, displacement vectors, force and moment vectors. Although
not a vector per se, the rotations that specify a particle’s orientation are treated as vectors in linear
structural mechanics. Since rotations are special in this respect it is worthwhile discuss those
further.
Rotational dofs. Various sets of variables that describe a particle’s orientation are in use. A
set often used in aerospace applications is called the (set of) Bryant angles. A particle is by them
associated with three angles that are defined as a well-defined sequence of rotations about three
particle-fixed local axes xyz in an exact turn. The first rotation is about the local x-axis and is called
a rotation by a roll angle α. The second turn is the rotation about the particle-fixed y-axis by the
pitch angle β . The final rotation is about the particle-fixed z-axis and is a jaw angle rotation γ. By
successive rotations over the three angles in turn, the particle can be proven to be able to reach
any orientation in space. A particularly nasty feature of the rotation angles is that the angular set
[α, β , γ] is not a 3D vector quantity and rotations therefore cannot be added as vectors and the order
of the rotation sequence is not arbitrary. An illustration of that is shown by two different rotation
sequences of a matchbox in Fig. 2.1. Using Bryant angles, the transformation matrix R can be
written
 
cosβ cosγ −cosβ sinγ sinβ
R = cosαsinγ + sinαsinβ cosγ cosαcosγ − sinαsinβ sinγ −sinαcosβ  (2.1)
sinαsinγ − cosαsinβ cosγ sinαcosγ + cosαsinβ sinγ cosαcosβ
which is seen to be highly nonlinear in the rotation angles.
After two successive rotations with first angular rotations (α1 , β1 , γ1 ) with R1 and thereafter
rotations (α2 , β2 , γ2 ) with R 2 the global location of a point at local p can be found at P = O +R R2 R 1 p .
It can be easily verified that the successive transformations do not commute, i.e. R 1 R 2 6= R 2 R 1 and
thus the order of the rotation sequence is essential and rotation angles of the sequence cannot simply
be added, i.e. R 1 (α1 , β1 , γ1 )RR2 (α2 , β2 , γ2 ) 6= R 2 R 1 6= R (α1 + α2 , β1 + β2 , γ1 + γ2 ). The rotations
thereby does not fulfil an essential criterion for being a vector, i.e. that vector components could be
superimposed in an arbitrary sequence.
A theorem by Euler lay the foundation for an alternative set of rotation parameters. It states
that:
Theorem 2.1.1 — Eulers’s theorem on rotations. A general reorientation of a body from any
angular orientation to any other angular orientation can be made by a single rotation about one
fixed (space-fixed or body-fixed are the same) axis.

Let the fixed axis over which the rotation occurs be given by the unit direction vector v =
{v1 , v2 , v3 } with normalization v21 + v22 + v23 = 1 and let the positive angular rotation about that axis
be θ . This forms a set of four rotation parameters. However, the normalization reduces the set to
three independent parameters, just as many as the three Bryant angles. It can be shown [41] that the
transformation matrix R expressed in these four parameters is
θ
R = I +V V V sin2
V sinθ + 2V (2.2)
2
2.1 Basic concepts and principles 21

Figure 2.1: Illustration over two rotational sequences a and b of two matchboxes lying with
same orientations (0a and 0b). Lower row from right to left (0a through 3a) are rotations 90° in
a sequence over local x − y − z axes in turn. Upper row from right to left (0b through 3b) are
rotations 90° over local y − x − z axes in turn. It is seen that from same orientations, the orientation
configuration in the end of the two sequences are completely different and thus the sequence is
important. Black arrow indicates fixed axis over which matchbox from orientation 0a can be turned
by 180° to reach same final orientation 3a.
........................................................................................

with the components of v arranged in the skew-symmetric V as

 
0 −v3 v2
V =  v3 0 −v1  (2.3)
−v2 v1 0

An important property of the rotation matrix R is that it is orthonormal, i.e. R −1 = R T and the
transformation from local-to-global P = R p after rotation can be reversed into a global-to-local
transformation p = R−1 P = RT P. This can be understood by a counter-rotation by −θ about the
same axis which leads to
−θ −θ
R −1 = I +V V V sin2
V sin − θ + 2V = I −V V V sin2
V sinθ + 2V = RT (2.4)
2 2
Example 2.1 Rotations in 2D
A long and slender particle 12 has its length axis x0 in a global xy plane originally coinciding with

the global x axis. Its local z0 axis coincides with the global z axis. It is rotated by the yaw angle γ to
a new orientation and no other rotations occur. Since then α = β = 0, the transformation matrix is
 
cosγ −sinγ 0
R =  sinγ cosγ 0 (2.5)
0 0 1

 22 Chapter 2. Fundaments of Linear Structural Dynamics

Small angle rotation. The transformation given by Eq. (2.2) makes it trivial to obtain
approximations for small angle rotations δ θ . For small δ θ we have that sinδ θ ≈ δ θ when
truncating after first order terms. The small angle transformation matrix thus becomes
R ≈ I +V
Vδθ (2.6)
While the finite rotation is not a vector quantity, this result can be used to prove that the
infinitesimal rotation is indeed a vector quantity. To see that, consider two infinitesimal rotations
dθ1 and dθ2 that are made in sequence. Let dθ1 be a rotation about the unit vector v 1 and dθ2 be a
rotation about the unit vector v 2 . The transformation matrix associated with the first rotation is thus
V 1 dθ1
R 1 = I +V (2.7)
and the transformation matrix for the second rotation is
V 2 dθ2
R 2 = I +V (2.8)
One can then write the transformation sequence
R 1 R 2 = [II +V V 2 dθ2 ] ≈ I +V
V 1 dθ1 ][II +V V 2 dθ2 ≈ R 2 R 1
V 1 dθ1 +V (2.9)
This shows that two successive infinitesimal rotations about two different axes can be added
and thus infinitesimal rotations can indeed be considered as vector quantities. In linear structural
dynamics this is relaxed to also hold for small angle rotations. They are thus treated as vectors
which allowed particle rotations to be superimposed in combined load cases.

2.1.2 Linear kinematics. Vector transformation

The above shows that a general rotation with (infinitesimally) small angular rotations over perpen-
dicular axes can be added without considering the order of the rotational sequence and can thereby
be considered a vector quantity. Let that small rotation with rotations over local x, y and z axes be
the rotational vector {δ α, δ β , δ γ} expressed in the orthogonal set of Bryant angles. For a particle
with a given orientation {α, β , γ} these small local rotations supplements to the particle local
displacements {ūx , ūy , ūz } to form the six particle dofs. For simplicity of notation, let us instead
denote the local translational displacement vector {ū1 , ū2 , ū3 } , {ūx , ūy , ūz } and the local small
angle rotational displacement vector {ū4 , ū5 , ū6 } , {δ α, δ β , δ γ} . With the transformation matrix
R (α, β , γ), the corresponding (generalized) displacement vector in a global coordinate system u is
thus
   

 u1 
 
 ū1 




 u2







 ū2 


   
u3 R 0 ū3
  
u, = , T ūu (2.10)

 u4 
 0 R   ū4 

u  ū 
 
   
 
 5  5

 

 
u6 ū6

and since R −1 = R T the transformation global-to-local is instead

 T 
R 0
ūu = u = TTu (2.11)
0 RT
For a particle that is part of a structure undergoing general motion due to applied loading,
the particle orientation may vary with time. In linear structural mechanics however, the vector
transformations are always made from the orientation state the particles have in the structures virgin
state. The transformation matrix T is thus considered being fixed in linear structural dynamics
analysis.
 2.1 Basic concepts and principles 23

Figure 2.2: Illustration over two components, I and II, that are perfectly bonded via rigidly
connected interfacing particles. Local coordinate systems of particles also shown.
........................................................................................

2.1.3 Particle joining. Compatible and constrained motion

Compatible motion. Structural components that are bonded together meet at a common interface.
It is common practice to consider this bonding to be perfectly rigid (perfect bonding) and thus
the interfacing particles of the two components share the same pattern of motion, see Fig. 2.2.
To simplify analysis, the translational and rotational motion of the particles are expressed in
local coordinate systems with same axes orientations. Most often this is also the orientation of
the global coordinate system. In global coordinates, the 6-dof motion of particles I and II are
u I = {u1I ; u2I ; u3I ; u4I ; u5I ; u6I } and u II = {u1II ; u2II ; u3II ; u4II ; u5II ; u6II }. If now also the origins of
the local coordinate systems are made to coincide then, since the particles are rigidly connected, it
hold that

u I = u II (2.12)

This is the compatibility relation for the rigidly constrained motion between the perfectly
bonded particles.
Kinematic constraints. The motion of the particle can be constrained to move according to
some pattern which would reduce its number of dofs. A linear scalar such constraint, that reduces
the dofs by one, can be expressed c T u = 0. Let u = {uuI ; u II } and the compatibility relation u I = u II
can be written C T u = 0 with

C T = −II 6×6 I 6×6

 
(2.13)

which reduces the number of system dofs by six.

2.1.4 Dynamic equilibrium. Newton’s second law and Euler’s rotation equation
Consider a rigid particle with mass M and principle mass moments of inertia Jx , Jy and Jz subjected
to forces f¯x , f¯y and f¯z and force couples m̄1 , m̄2 and m̄3 acting in and about axes of a coordinate
system xyz that is fixed to the particle and let the axes be in the directions of the principle axes of
inertia. Let also the angular rotation velocity be expressed in Bryant angles as ω = {α̇, β̇ , γ̇} and
let the translational displacement vector be ūu = {ūx , ūy , ūz }. Newton’s second law than gives that

1 0 0 ū¨x   f¯x 
    

M 0 1 0 ū¨y = f¯y (2.14)

¨   ¯
0 0 1 ūz fz
 24 Chapter 2. Fundaments of Linear Structural Dynamics

Let J being the inertia matrix related to a coordinate system with axes parallel to the principle
axes as
 
Jx 0 0
J =  0 Jy 0  (2.15)
0 0 Jz

and then Euler’s rotation equation, based on Newton’s second equation, gives
 
m̄x 
ω + ω × [JJ ω ] = m̄y
J ω̇ (2.16)
m̄z
 

or

Jx α̈ + (Jz − Jy )β̇ γ̇ = m̄x

Jy β̈ + (Jx − Jz )α̇ γ̇ = m̄y (2.17)
Jz γ̈ + (Jy − Jx )α̇ β̇ = m̄z

Under the assumption that the angular velocities are small, let them be denoted δ α, δ α and
δ α for that reason, the second order terms are neglected to give the linearized equation

Jx δ¨α = m̄x
Jy δ¨β = m̄y (2.18)
Jz δ¨γ = m̄z

Eqs. (2.14) and (2.18) leads to the linearized equations of translational and rotational motion of
the particle
 ¨   ¯ 
ū

M 0 0 0 0 0   x   fx 
¨y  f¯y 
 
 0 M 0 0 0 0 
 ū 



 

 0 0 M 0 0 0  ū¨z ¯z 
    
f
  
M ūu¨ , 

 δ¨α  = m̄x  , f̄f
 (2.19)
 0 0 0 Jx 0 0  
    
 0 0 0 0 Jy 0   δ¨β  m̄y 
 
 

   
 ¨   
  
0 0 0 0 0 Jz δγ m̄ z

which forms one fundamental rule of linearized structural dynamics.

2.1.5 Linear constitutive relations for materials

The theory of linear elasticity is based on the observation the deformation of most materials is
directly proportional to the load applied. For small load the deviation between the true deformation
state and the linear approximation is negligible for these materials. For many materials the linear
approximation holds sufficiently well up to moderate load levels so that linearization becomes useful
in engineering practice. Such is, for instance, the case for steel for which the linear approximation
holds well up to a significant portion of the rupture strength at a load level at which the material
collapses. The linear dependency was formulated by Hooke from observations of uniaxial (x)
tension/compression testing into Hooke’s law σxx = Eεxx where E is the proportionality factor
(a.k.a. the Young’s modulus) that relate the strain response εxx to the stress loading σxx . In the
uniaxial testing of an isotropic material it was also found that the material also shrunk in the
directions perpendicular to the loading (y and z) in a linear manner which were formulated as
σxx = −νEεyy and σxx = −νEεzz where the proportionality factor −νE includes a second elasticity
parameter, ν, known as the Poisson ratio.
2.1 Basic concepts and principles 25

Figure 2.3: Stress components acting on surfaces on infinitesimal cube of volume dxdydz. Magni-
tude and directions indicated by arrows. Stresses acting on opposite and hidden sides of cube are
of same magnitude but opposite direction. Shear stresses σyz = σzy , σxz = σzx and σxy = σyx and
therefore there are just six independent stress components.
........................................................................................

For material batches that have been formed by processes that create nonisotropic internal micro-
structures, a more complex linear stress-strain relation has been motivated by physical observations.
That holds for cold-rolled sheet metal and, not to the least, for composite material. Let Fig. 2.3
define the six independent stress components of the stress state σ that act on an infinitesimal cube
from within the material. Let further u (x, y, z) , {u(x, y, z); v(x, y, z); w(x, y, z)} be the deformation
state of the material in the point (x, y, z)n let the deformation gradient define the six independent
strain components

∂u ∂v ∂w
εxx = εyy = εzz = (2.20)
∂x ∂y ∂z
∂u ∂v ∂u ∂w ∂v ∂w
εxy = + εxz = + εyz = + (2.21)
∂y ∂x ∂z ∂x ∂z ∂y

The generalized Hooke’s law formulates a linear relation between the stress state σ , {σxx ; σyy ;
σzz ; σyz ; σxz ; σxy } and the strain state ε , {εxx ; εyy ; εzz ; εyz ; εxz ; εxy } as σ = E ε where E is a
symmetric [16] matrix, often called the material stiffness matrix, of material constants. The explicit
form of the generalized Hooke’s law is thus

    

σxx   e11 e12 e13 e14 e15 e16  εxx 





σyy  
  e22 e23 e24 e25 e26  
 
ε yy



  
 
e33 e34 e35 e36  εzz
  
σzz 
= (2.22)

 σ yz 
 
 e44 e45 e46 
 εyz 

e55 e56 
   

 σ
  xz



   ε
 xz


  
σxy sym e66 εxy

Note that, for symmetry reason, the material stiffness matrix holds 21 independent elasticity
constants e jk . These 21 constants are required to characterize a fully anisotropic material and thus
require substantial material testing to obtain the material property data. However, for isotropic
material, the number of independent coefficients is just 2 and the material stiffness matrix can be
 26 Chapter 2. Fundaments of Linear Structural Dynamics

written
 
2e11 2e12 2e12 0 0 0

 2e11 2e12 0 0 0  
1 2e11 0 0 0 
E=   (2.23)
2
 e11 − e12 0 0  
 e11 − e12 0 
sym e11 − e12

where
1−ν ν
e11 = E e12 = E e11 − e12 , 2G (2.24)
(1 − 2ν)(1 + ν) (1 − 2ν)(1 + ν)

in which e11 − e12 has been used to define the isotropic material’s shear modulus, G. Since there
are only 2 independent coefficients in E there is a relationship between the elasticity parameters
E, ν and G that reads

E
G= (2.25)
2(1 + ν)

2D stress and strain states. In a so-called plane stress state (in xy) it holds that σzz = σxz =
σxy = 0 and Hooke’s law for the isotropic material can be reduced. With the reduced stress
σ = {σxx ; σyy ; σxy } and reduced strain ε = {εxx ; εyy ; εxy } it again holds that σ = E ε and the
material stiffness, and its inverse the material compliance matrix E −1 , become
   
1 ν 0 1/E −ν/E 0
E 
E= 1 0  E −1 =  1/E 0  (2.26)
1 − ν2
sym (1 − ν)/2 sym 1/G

For the special case of plane strain state (in xy) it instead holds that εzz = εxz = εxy = 0 and
Hooke’s law for the isotropic material can be reduced to give the material stiffness and compliance
as
   
1−ν ν 0 1 − ν −ν 0
E 1+ν 
E=  1−ν 0  E −1 = 1 − ν 0 (2.27)
(1 + ν)(1 − 2ν) E
sym 1 − 2ν sym 1

2.2 The stiffness method

The stiffness method (aka the displacement method) is a general method to set up a complete
system’s governing equations. It makes use of the compatibility relation of rigidly joined particles
of interfacing components. It also makes us of Newton’s third law for the interaction forces.
Together with a description of the loading and imposed boundary conditions it gives a full set of
equations that can be solved for the joined particle’s motion.

2.2.1 System assembly

To illustrate the joining process (aka the assembly process) let us consider two components A and
B, see Fig. 2.4 with global displacements uA and uB respectively. Let f A and f B be the global loads
that are vectorially associated with u A and u B . Let further u a be the partition of u A that describe the
motion of the particles of A that are rigidly connected to the particles of B, and further u b be the
2.2 The stiffness method 27

Figure 2.4: Two components, A and B, about to be rigidly joined by fixing rigid particles a and b
(solid black) of A and B together. The association to interface and internal dofs of the different sets
of displacements are indicated.
........................................................................................

partition of u B that is associated to B’s joining particles to A. The displacement partitions u c and
u d are the disjunct sets of displacements to component particles that do not join. That is so that
   
ua ub
uA = and u B = (2.28)
uc ud

Let further the loads be arranged accordingly so that

   
fa fb
fA = and f B = (2.29)
fc fd
Since component A has nothing to do with the displacements u d its force-displacement relation
can be written on partitioned form as
    
K aa K ac 0 u a   f a 
K Tac K cc 0  u c = f c (2.30)
0 0 0 ud 0
   

and similarly for component B that has nothing to do with displacements u c as

    
K bb 0 K bd u b   f b 
 0 0 0  uc = 0 (2.31)
K Tbd 0 K dd ud fd
   

At this time it is convenient to invoke the compatibility relation u a = u b , ūu¯ in which the
displacements of the joined particles ūu¯ is introduced. Let also the load f a that act on the joining
particles of A be split into two parts f aX and f aI so that f a = f aX + f aI with f aI being the interface
loading acting from the joining particles of B on A, and f aX are other external forces that act on the
same particles. Similarly, let f b = f bX + f bI with f bI being the interface loading that by virtue of
Newton’s third law on action and interaction is f bI = − f aI . Let these two relations be introduced
to Eqs. (2.30) and (2.31) to give

K aa K ac 0  ūu¯   f aX   f aI 
      
K Tac K cc 0  u c = f + 0 (2.32)
   c  
0 0 0 ud 0 0
 28 Chapter 2. Fundaments of Linear Structural Dynamics

and
K bb 0 K bd  ūu¯   f bX   f aI 
      
 0 0 0  uc = 0 − 0 (2.33)
T
K bd 0 K dd ud fd 0
     

The assembly process constitute of the summation of these two equations together to yield

K aa + K bb ] K ac K bd  ūu¯   f aX + f bX   ¯f̄f 
      
[K
Ku ,  K Tac K cc 0  uc = fc , f ,f (2.34)
T   c
K bd 0 K dd u f
  
d d fd

where ¯f̄f has been introduce to denote the total external forces that act on the joined particles of A
and B.
Stiffness matrix checking. Since the forces of f are the complete set of forces that act on
the combined system, the static equilibrium condition ∑ f j = 0 provides a convenient procedure
to check the correctness of the assembled stiffness matrix K . This can be made by column-wise
checking by observing that f = K {0; . . . ; 0; uk ; 0; . . . ; 0} = K :k uk are the loads required to produce
a single displacement uk with the remaining displacements in the displacement vector u fixed to
zero. For load equilibrium it is thus required that (∑ j K jk )uk = 0, or for uk 6= 0 that ∑ j K jk = 0, for
all columns of K . This is a requirement that can be easily checked for a system which possesses
only translational dofs and forces as a necessary but not sufficient criterion for the correctness of K .
For systems involving also rotational dofs and moments a corresponding test is not so convenient.

2.2.2 Degree-of-freedom reordering and partitioning

Vector reordering and partitioning of the rearranged system equations are often used to demonstrate
various concepts in structural statics and structural dynamics. The reordering of a vector u can
be achieved mathematically a transformation operation to obtain a reordered vector ūu as ūu = T u
with T being an orthonormal Boolean matrix and thus also the originally ordered vector u can be
obtained as ūu = T −1 u = T T u . The reordering of an equation system is illustrated in the following
example.
 Example 2.2 Reordering and partitioning
Let the stiffness equation of a 5-dof system be described on matrix form as
    
K11 K12 K13 K14 K15  u1   
 f1 
K21 K22 K23 K24 K25   
 u 2





 f 2



 
K31 K32 K33 K34 K35  u3 = f3 (2.35)
 
K41 K42 K43 K44 K45   u4 
  
  f4 
 


     
K51 K52 K53 K54 K55 u5 f5

Let u a = {u2 ; u4 } be one partition with two selected displacement elements and u b = {u1 ; u3 ; u5 }
the other partition with the remaining displacement dofs. Let then a re-ordered displacement vector
be ūu = {uua ; u b } and the similarly reordered force vector be f̄f = { f a ; f b }. Re-partition the stiffness
matrix accordingly using first a Boolean operations solution and then a hands-on solution!

Boolean solution. The Boolean transformation to obtain the modified order 2-4-1-3-5 is
 
0 1 0 0 0 1st row ← 2nd
0 0
 0 1 0 2nd row ← 4th

T = 1 0 0 0 0  3rd row ← 1st (2.36)
0 0 1 0 0 4th row ← 3rd
0 0 0 0 1 5th row ← 5th
2.2 The stiffness method 29

Since the Boolean matrix T is orthonormal we have that T T T = I and therefore the equation system
K u = f can be written T T T K T T T u = f or [T
T K T T ]{T
T u} = {TT f } and we note that
      
0 1 0 0 0   u1 
 
 u2 
 
 f2 
0 0 0 1 0  u
 
u

 f4 
 
  2  4
     

{TT u} = 
1 0 0 0 0  u3 = u1
 and similarly {TT f } = f 1 (2.37)
0 0 1 0 0  
 u4





 u3





 f 3




      
 
0 0 0 0 1
 
u5 u5 f5

so that the rearrangements of u and f are now in order. The associated stiffness matrix T K T T
is after a first post-multiplication of T T
    
K11 K12 K13 K14 K15 0 0 1 0 0 K12 K14 K11 K13 K15
K21 K22 K23 K24 K25  1 0 0 0 0 K22 K24 K21 K23 K25 
T
    
KT ] = T 
T [K K31 K32 K33 K34 K35  0 0 0 1

 = T K32
0  K34 K31 K33 K35 

K41 K42 K43 K44 K45  0 1 0 0 0 K42 K44 K41 K43 K45 
K51 K52 K53 K54 K55 0 0 0 0 1 K52 K54 K51 K53 K55
(2.38)

which we note has resulted in a proper column-wise rearrangement of the stiffness coefficients.
After the second and final step we have after pre-multiplication with T that
      
0 1 0 0 0 K12 K14 K11 K13 K15 K22 K24 K21 K23 K25
0 0 0 1 0 K22 K24 K21 K23 K25   K42 K44
 K41 K43 K45 

T
   
T K T = 1 0 0 0 0 K32 K34 K31 K33 K35  =  K12 K14
     K11 K13 K15  
0 0 1 0 0 K42 K44 K41 K43 K45  K32 K34  K31 K33 K35 
0 0 0 0 1 K52 K54 K51 K53 K55 K52 K54 K51 K53 K55
(2.39)

and now the matrix rows have also been properly re-arranged.

Hands-on solution. It is obvious that this matrix form can be obtained in a two-step opera-
tion in which the first step consists of swapping the column order of the coefficient matrix to
correspond with the correct partitioning of the displacement vector ūu. The second step is then to
swap the row order of that column-swapped matrix to correspond with the correct partitioning of
the load vector f̄f . To see this clearly, extract the five equations of the matrix system in the order
2-4—1-3-5 to get the forces in the specified partition order f̄f = { f2 ; f4 ; f1 ; f3 ; f5 }. The equations
are then in order (note the order of the LHS terms)

K22 u2 + K24 u4 + K21 u1 + K23 u3 + K25 u5 = f2

K42 u2 + K44 u4 + K41 u1 + K43 u3 + K45 u5 = f4
K12 u2 + K14 u4 + K11 u1 + K13 u3 + K15 u5 = f1
K32 u2 + K34 u4 + K31 u1 + K33 u3 + K35 u5 = f3
K52 u2 + K54 u4 + K51 u1 + K53 u3 + K55 u5 = f5

or on matrix form
      
K22 K24 K21 K23 K25   u2   
 f2  
 K42 K44
 K41 K43 K45  u4   f4 
 
 
       
 
K K u f

aa ab a
 a
 K12 K14 K11 K13 K15  u 1 , = , f 1
K ba K bb u b fb
     
K32 K34  K31 K33 K35   u 
 
f 

 3   3 
  
 
   
K52 K54 K51 K53 K55 u5 f5
 30 Chapter 2. Fundaments of Linear Structural Dynamics

which is the re-partitioned form.



One should note that in practical computer implementations, the rearrangement of the equation
systems does not take place. It would cause too much computational overhead. Methods that are
based on partitions of the problem are instead implemented with other methods of bookkeeping.

2.2.3 Prescribed displacement. Fixed boundary

For many systems, parts of the system boundary may be considered as prescribed by its surrounding.
A light and flexible building built on solid rock ground is a good example for which the building
parts that is attached to ground cannot move because the rigidity of the rock. The motion of
the system boundary is then prescribed to zero. In an earthquake, the motion of the boundary is
governed, and thus prescribed, by the motion of the heavy rock foundation. At least this holds in
idealized situations.
Let the displacements of the system be partitioned according to u = {uua ; u b }, where u b is the
displacement partition that is prescribed by an outside agent. Let the forces required to produce
that prescribed motion be f b and the other external forces that act on the system associated to u a be
f a . The partitioned stiffness equation then becomes
    
K aa K ab u a fa
= (2.40)
K ba K bb u b fb
Since u b is known, the upper matrix row gives that
K aa u a = f a − K ab u b (2.41)
which gives a reduced set of equations to solve. The outside agent forces f b required to enforce
that boundary deformation can thereafter be obtained from the second matrix row as
f b = K ba u a + K bb u b (2.42)
A special situation is when the boundary is fixed so that u b = 0 that leads to the simplified
relations
K aa u a = f a and f b = K ba u a (2.43)
A 3D mechanical system that is not connected to anything, such as a satellite in orbit, possess
at least six so-called rigid body modes. That is that it has six independent motion patterns it could
undergo without causing any flexible deformation. However, a unrestricted system is not limited
to have only six such possible rigid body modes. A gearbox is a good example of a system that
possesses more that six rigid body modes. Besides the overall translational and rotational motion
the full gearbox can undergo, also its shafts may rotate freely without elastic deformation in relation
to the gearbox housing. The number of possible rigid motions is determined from the rank of the
system matrix K . The number of possible rigid body modes are given by dim(K K ) − rank(KK ).
The prescribed displacement boundary conditions imposes a rank reduction of the stiffness
matrix so that dim(K K aa ) − rank(K K ) − rank(K
K aa ) < dim(K K ). The boundary conditions can be
categorized by the number of possible rigid body motion the system possesses after the boundary
conditions have been applied. A boundary-fixed system for which still after applied boundary
constraint possesses one or more possible rigid body modes is said to be hypostatic for which then
dim(K K aa ) − rank(K
K aa ) > 0. Systems that do not possess such possible rigid body motion are either
said to be isostatic or hyperstatic. The isostatic structures are such that the number of constraints
imposed by the boundary is just about enough to make dim(K K aa ) − rank(K
K aa ) = 0 and systems
with more independent constraints than that are hyperstatic.
 2.3 The finite element method 31

2.2.4 Static condensation

Static condensation can be an efficient means of reducing the number of unknown displacements in
an equation system K u = f . It is most often associated to the condensation of dofs to which no
external forces act. Let u a be the partition on which the non-zero external forces f a act and let u b
be the partition on which there are no external forces, i.e. f b = 0 . We thus have the partitioned
form
    
K aa K ab u a fa
= (2.44)
K ba K bb u b 0

and the lower matrix row gives u b = −K K −1

bb K ba u a that can be used to condense the displacement
set u b on u a using the upper matrix row to give

K aa u a + K ab u b = K aa − K ab K −1
 
bb K ba u a = f a (2.45)

in a condensed equation system that involves only the displacement partition u a .

2.3 The finite element method

The finite element method (FEM) has become the most versatile method to analyse the mechanical
behaviour of structures. As it decomposes the structure into finite volume elements of volume Ve of
arbitrary size it can well represent the geometry of the structure in question and also its elasticity
and mass distribution. By being such a versatile method, the literature on the finite element method
is vast, see for instance [5, 8, 52]. Some brief comments on stiffness and mass modelling are made
here and some 1D and 2D element formulations are given for illustration purpose.

2.3.1 FE modelling of stiffness

Textbooks on the finite element method, such as [5, 8, 52], give the element stiffness matrix of the
linearly elastic material as
Z
Ke = BT E BdVe (2.46)
Ve

where E is the material stiffness matrix that relate the stress to the strain as σ = E ε and B is the
strain-displacement matrix1 that relate the strain to the element displacement vector u e as ε = B u e .
The element stiffness matrices are assembled together using the stiffness method outlined above to
form the complete system. In that process, discrete material particles on the elements’ interfacing
surfaces (the elements’ nodes) are rigidly connected to form a unity.
One problem in the modeling of the material of monolithic parts of a structure is to use
appropriate data for the constitutive model that gives E . The most common modeling practice is
to use an isotropic material assumption and to use tabular material data of the Young´s modulus
and Poisson ratio for it. However, for many structures more involved material descriptions need to
be used. This is obvious for structures that consists of composite materials, for which orthotropic
or anisotropic material descriptions are often used. It is less obvious for sheet metal structures,
such as cars, refrigerators, washing machines, aircraft, etc. For sheet metal, the forming process in
which the raw metal is rolled into a sheet often introduce anisotropic material stiffness that can be
significatively different from isotropy. For very accurate models, this is worth considering.
In the finite element modeling, one requirement is that the geometry of the structure is accurately
represented. This is to get convergence, through refined levels of discretization, to the exact solution
1 The use of B in this isolated context should not be confused with the use of B in the state-space description that will

be treated later.
 32 Chapter 2. Fundaments of Linear Structural Dynamics

to the underlying partial differential equations that are set up to model the real world behavior.
However, it is common practice that small features, such as small holes and small fillets, are
disregarded from the model and thus a step from the precise geometry modeling is taken. Such
model deviations from the true geometry are perfectly justified if they are supported by model
verification. However, a big problem in stiffness modelling is the modeling of interfaces and joints
between monolithic parts. These interfaces may be bolted, riveted, welded, glued or held together
by other means of establishing and maintaining contact between parts. Such interfaces have always
been considered as strong candidates to stiffness modelling errors. It is strongly advised that the
modeler gains as much insight as possible into the physical characteristics of joints, as its usually
there the most significant modeling errors are situated.

2.3.2 FE modelling of mass inertia

In the finite element modeling, the mass distribution seems to be the easiest to model correctly
from start and does not require any advanced first principles. As long as the geometry is correctly
described by the finite element mesh and the density ρ of the constituents is known, the modeling
consists of assembling the global mass matrix M from the element mass matrices which are (see
e.g. Ref. [8])
Z
Me = N T N dVe
ρN (2.47)
Ve

Here N is the shape function matrix of the element. The mass matrix evaluated by Eq. (2.47) is
said to be consistent with the stiffness matrix formulation (2.3.2) if the strain-displacement matrix B
is established from the same shape functions. The consistent mass matrix may be densely populated
with non-zero entries. A simpler and historically earlier formulation is the lumped mass matrix,
which is obtained by placing particle masses mk at the nodes k of an element, such that ∑ m j is
the total element mass. Particle “lumps” have no rotary inertia unless assigned, as is sometimes
done for the rotational degrees-of-freedom of beams and plate elements. A lumped mass matrix
is diagonal. Many analysis schemes used in structural dynamics can be made more efficient if
the mass matrix is lumped, creating a diagonal global mass matrix M . However, the consistent
mass formulation has an interesting property that the lumped mass formulation does not have. It
can be shown that the eigenvalues calculated from the consistent finite element model are always
equal or higher than the exact eigenvalues of the underlying partial differential equation model.
This property can be exploited in the model verification process for which we then know that the
eigenvalues should converge from above when the mesh is refined. However, this property is often
lost in the modeling procedure since the stiffness matrix integral () is only evaluated approximately
by use of quadrature rules.
Mass lumping can be made by various algorithms, out of the Hinton-Rock-Zienkiewicz (HRZ)
lumping scheme described below is just one. For all good lumping schemes it is important that
the total mass of each element is accurately represented. The HRZ lumping scheme is an effective
method for producing a diagonal mass matrix. It can be recommended for arbitrary elements. The
idea is to use only the diagonal terms of the consistent mass matrix, but to scale them in such a way
that the total mass of the element is preserved. Specifically, the HRZ procedural steps are as follows

i) Compute and use only the diagonal coefficients of the consistent mass matrix
ii) Compute the total mass me of the element from its volume and density
iii) Compute a sum s by adding the diagonal coefficients mkk associated with translational dofs (but
not rotational dofs, if any) that are in mutually parallel direction
iv) Scale all the diagonal coefficients by multiplying them by the ratio me /s, thus representing
exactly the total mass of the element.
 2.3 The finite element method 33

It is the author’s experience that a mass lumping scheme gives a sufficiently good representation
of the mass distribution in most practical situations. That observation is made based on that present
days finite element models use so dense meshes such that the results created by consistent and
lumped mass modeling schemes are practically equal for the frequency range of interest in structural
dynamics. It is also the author’s experience that commercial finite element code for mass lumping
is better verified and it has been found that the consistent mass modeling schemes might create
spurious eigensolutions if more rarely exercised elements are used for which the stiffness and mass
matrices have not been implemented to be truly consistent.
In model validation, the mass modeling is also easily checked against the true weight of the test
article. It is strongly advised to check the weight of the structure in the situation of a vibration test.
A first calibration step is then normally to adjust the density of the model parts, from the nominal
density given by material data sheets, to density values that are in line with the scale reading. As
the weight of the model is linear in the material density, the total weight can also easily be kept
constant throughout the calibration process by imposing a weight constraint. Say that the total
weight is Wtot and that the weight of parts that are not parameterized in the calibration process is
W0 . The total model weight is then
m
Wtot = W0 + ∑ ak pk (2.48)
k=1

where m is the number of free density parameters, pk is a density parameter of the k:th part and ak
is its associated linear contribution coefficient that can easily be determined by a side calculation.
To keep the weight constant, while adjusting the parameters pk to better fit, we may select one of
the parameters, say the one related to the heaviest part j, to be slaved to the others such that
m
a j p j = Wtot − ∑ ak pk (2.49)
k6= j

That allows us to use unconstrained parameter minimization techniques in the calibration to

minimum deviation to test results.

2.3.3 One-dimensional FE elements for trusses and frames

When appropriate, so-called 1D elements are often engaged to represent structural parts that has
simple enough geometry such as cylinders and prisms with cross-sectional dimensions that are much
smaller than its longitudinal dimension, i.e. its length. They are engaged to save computational
effort since they engage fewer degrees-of-freedom than elements with alternative formulations
such as 2D or 3D elements. When such elements are subjected to pure tension/compression with
forces on its ends that act only in a longitudinal sense it is normally called a rod element with only
two dofs with end particles free to move only in a direction corresponding to its longitudinal axis
(say, the local x-axis), see Fig. 2.5a. Provided that the cross-sectional area A is uniform over its
length L and its material properties are also uniform its element stiffness matrix and consistent
mass matrices evaluated with Eqs. (2.46) and (2.47) become
   
EA 1 −1 ρAL 2 1
Ke = and M e = (2.50)
L −1 1 6 1 2

where E is Young’s modulus of the material in a 1D stress state and ρ is the material’s density.
When the same geometrical entity is loaded at its ends with pure twisting couples, the normal
nomenclature is to call it a shaft element, see Fig. 2.5b. With G being the shear modulus of the
material and Kv being the cross-sectional factor of torsion for the cross-sectional geometry in
34 Chapter 2. Fundaments of Linear Structural Dynamics

Figure 2.5: a) Rod element with end particles (1 and 2) loaded by longitudinal forces f1 and f2
respectively responding with longitudinal displacements u1 and u2 . b) Shaft element with end
particles loaded by couples acting around x-axis m1 and m2 respectively to which it responds with
angular displacements φ1 and φ2 .
........................................................................................

question its stiffness and consistent mass matrices become

ρALrp2 2 1
   
GKv 1 −1
Ke = and Me = (2.51)
L −1 1 6 1 2

with rp being the polar radius of inertia of the cross-section rp2 = A1 A (y2 + z2 )dA.
R

In the situation the element is loaded through its end particles in bending only (neglecting
shearing), the element is normally called an (Euler-Bernoulli) beam element. This bending can
be about both local cross-sectional axes and a complicated deformation pattern may result also
for simple load conditions. For elements with symmetric geometric and material properties over
its cross-section, however, the modelling is simplified and the bending in the two perpendicular
cross-sectional planes are decoupled. Let a local symmetry axis be denoted the z-axis and study
the bending in a local xy-plane, see Fig. 2.6 (left). The stiffness and mass matrices related to the
element’s motion in the y-direction in that plane are

   
12 6L −12 6L 156 22L 54 −13L
EIz  4L2 −6L 2L2 
 and M e = ρAL 
 4L2 13L −3L2 
Ke = 3   (2.52)
L  12 −6L  420  156 −22L
sym 4L2 sym 4L2

If, on the other hand, the element is loaded such that its deformation takes place in only the
local xz-plane, see Fig. 2.6 (right), the associated stiffness and consistent mass matrices become
   
12 −6L −12 −6L 156 −22L 54 13L
EIy  4L2 6L 2L2 
 M e = ρAL 
 4L2 −13L −3L2 
Ke = 3   (2.53)
L  12 6L  420  156 22L
sym 4L2 sym 4L2

For an element for which tensional, torsional and bending effects all need to be considered its
end particles undergo arbitrary translation and rotation in 3D space. Both ends are thus associated
with six dofs and the stiffness matrix thus involves in total 12 dofs.
The full stiffness matrix can thus be combined from tension/twist/bending Eqs. (2.50-2.53)
into
2.3 The finite element method 35

Figure 2.6: Beam elements with end particles (1 and 2) loaded by lateral forces f1 and f3 and
bending couples f2 and f4 . They respond with lateral displacements u1 and u3 and rotations u2 and
u4 . Left figure shows bending in the xy-plane and right figure shows bending in the xz-plane. Finite
element shape functions also shown. NB! positive rotations about z and y axes are different in these
planes.
........................................................................................

 EA
− EA

L . . . . . L . . . . .
12EIz 6EIz

 L3
. . . L2
. − 12EI
L3
z
. . . 6EIz
L2


12EIy 6EI 12EI 6EI
. − L2 y . . . − L3 y . − L2 y . 
 
L3

GKv
− GK
 v


L . . . . . L . . 
4EIy 6EIy 2EIy
 

 L . . . L2
. L .  
4EIz
. − 6EI z
. . . 2EIz 

Ke =  L L2 L 

EA 
 L . . . . . 
12EIz
− 6EI
 
z

 L3
. . . L2 
12EIy 6EIy

 L3
. L2
.  
GKv
. . 
 
 L
 4EIy 
 L . 
4EIz
sym L
(2.54)

where dots have replaced zeros for reading convenience. The associated consistent mass matrix
36 Chapter 2. Fundaments of Linear Structural Dynamics

is likewise
 
140 . . . . . 70 . . . . .

 156 . . . 22L . 54 . . . −13L
 156 . −22L . . . 54 . 13L . 
140rp2 70rp2
 
 . . . . . . . 
4L2 −3L2
 

 . . . −13L . . 
ρAL  4L2 . 13L . . . −3L2 
Me =  
420 
 140 . . . . . 

 156 . . . −22L

 156 . 22L . 

 140rp2 . . 
 4L2 . 
sym 4L2
(2.55)

Using the HRZ mass lumping scheme, the corresponding diagonal lumped mass matrix is
instead
 
39 . . . . . . . . . . .

 39 . . . . . . . . . .

 39 . . . . . . . . .
 39r 2 . . . . . . . . 
 p 

 L2 . . . . . . .
ρAL  L2 . . . . . .

Me = (2.56)

78  39 . . . . .
 
39 . . . .
 


. . .

 39
39rp2 .
 
 .
L2 . 
 

sym L2

 Example 2.3 Static reduction and condensation of a beam element

Consider the planar motion (in the xy plane) of an 1D element in subjected to two boundary
conditions a) and b) in the figure above. The 2 × 2 reduced stiffness matrix for case a) with free
displacements u2 and u6 can be extracted from Eq. (2.54) and is

 
EIz 12 6L
Ke = (2.57)
L3 6L 4L2
and for the case b) with free displacements u2 , u6 and u12 the associated 3 × 3 element stiffness is
   
12 6L 6L  
EIz  2 2 K aa K ab
K e = 3 6L 4L  2L   , (2.58)
L 2 2 K ba K bb
6L 2L 4L
2.3 The finite element method 37

Using the static condensation given by Eq. (2.45) gives the condensed stiffness matrix K e =
K aa − K ab K −1
bb K ba associated to u2 and u6 as
 
EIz 3 3L
Ke = (2.59)
L3 3L 3L2

The load-deformation pattern in these two cases can be summarized in the elementary cases
given in Fig. 2.7. 

. .......................................................................................

Figure 2.7: Elementary cases for beam element for two sets of boundary conditions at its ends.
Forces and couples required to give specified end displacements/rotations given.
 38 Chapter 2. Fundaments of Linear Structural Dynamics

Figure 2.8: A 3-noded triangular (CST) element with specified node coordinates of connection
particles and a 4-noded rectangular element with element sides parallel with the global x and y axes.
........................................................................................

2.3.4 Two-dimensional FE elements for plane stress and plane strain

For illustration purpose only two specific 2D elements for plane stress or plane strain analysis are
considered in this book. These are the basic rectangular and triangular elements that in combination
can be made to represent any geometry of a 2D domain to an arbitrary accuracy. Their motion is
described by the translations of their corner nodes in a global coordinate system.

Triangular elements. The three-noded triangular element, the so-called constant strain trian-
gle element (CST element), has its corner nodes 1-2-3 at coordinates (x1 , y1 ), (x2 , y2 ) and (x3 , y3 ) in
a global coordinate system XY , see Fig. 2.8, giving it the triangle area Ae = 12 (x1 y2 − x1 y3 + x2 y3 −
x2 y1 + x3 y1 − x3 y2 ). Its strain-displacement matrix B is constant (thereby giving the element its
name) and is
 
−∆y1 0 ∆y2 0 −∆y3 0
1 
B= 0 ∆x1 0 −∆x2 0 ∆x3  (2.60)
2Ae
∆x1 −∆y1 −∆x2 ∆y2 ∆x3 −∆y3

with help-variables

∆x1 = x3 − x2 ∆x2 = x3 − x1 ∆x3 = x2 − x1

∆y1 = y3 − y2 ∆y2 = y3 − y1 ∆y3 = y2 − y1

For an element with constant thickness t its stiffness matrix associated to the element displace-
ment vector ue , {u1 ; v1 ; u2; v2; u3; v3} evaluates to
Z
Ke = B T E B dV = tAe B T E B (2.61)
Ve

If also a situation of plane stress apply (see Eq. 2.26 for the material stiffness E in plane stress) the
6 × 6 stiffness matrix can then be shown to be (with ν + = 1 + ν and ν − = 1 − ν)

2∆y21 + ν − ∆x12 −ν + ∆x1 ∆y1 −2∆y1 ∆y2 − ν − ∆x1 ∆x2 . . .


..
2∆x12 + ν − ∆y21 2ν∆x1 ∆y2 + ν − ∆x2 ∆y1 . . .

tE  .
Ke = + −

..
8Ae ν ν  
. 2∆y22 + ν − ∆x22 ...
sym. ...
2.3 The finite element method 39

. . . 2ν∆x2 ∆y1 + ν − ∆x1 ∆y2 2∆y1 ∆y3 + ν − ∆x1 ∆x3 −2ν∆x3 ∆y1 − ν − ∆x1 ∆y3

. . . −2∆x1 ∆x2 − ν − ∆y1 ∆y2 −2ν∆x1 ∆y3 − ν − ∆x3 ∆y1 2∆x1 ∆x3 + ν − ∆y1 ∆y3 
+ − 2ν∆x3 ∆y2 + ν − ∆x2 ∆y3 

... −ν ∆x2 ∆y2 −2∆y2 ∆y3 − ν ∆x2 ∆x3  (2.62)
... 2∆x22 + ν − ∆y22 2ν∆x2 ∆y3 + ν − ∆x3 ∆y2 −2∆x2 ∆x3 − ν − ∆y2 ∆y3 
... 2∆y23 + ν − ∆x32 −ν + ∆x3 ∆y3 
... 2 −
2∆x3 + ν ∆y3 2

and, provided that the density ρ is uniform over the element, the consistent mass matrix M e =
T
N dV related to the element’s linear shape functions Ni (X,Y ), i = 1, 2, 3 can be shown [26]
R
Ve N
ρN
to be
 
2 0 1 0 1 0

 2 0 1 0 1

ρtAe  2 0 1 0
Me =   (2.63)
12  2 0 1

 2 0
sym. 2

or alternatively the lumped mass matrix obtained by the HRZ lumping scheme

ρtAe
Me = I 6×6 (2.64)
3

Rectangular elements. For a constant thickness rectangular element with sides parallel to the
global X and Y axes with area Ae = ∆x∆y the element stiffness matrix element associated with its
displacement vector u e , {u1 ; v1 ; u2; v2; u3; v3; u4; v4} can be shown to be

tE
Ke = ×
96A ν + ν −
 − 2e
4ν ∆x + 8∆y2 3ν + ∆x∆y 2ν − ∆x2 − 8∆y2 −3(1 − 3ν)∆x∆y ...
.
.. 8∆x2 + 4ν − ∆x∆y 3(1 − 3ν)∆x∆y 4∆x2 − 4ν − ∆y2


 ...
..
4ν − ∆x2 + 8∆y2 −3ν + ∆x∆y

 . ...
..

8∆x2 + 4ν − ∆x∆y

 . ...
sym ...

. . . −2ν − ∆x2 − 4∆y2 −3ν + ∆x∆y −4ν − ∆x2 + 4∆y2 3(1 − 3ν)∆x∆y

... −ν + ∆x∆y −4∆x2 − 2ν − ∆y2 −3(1 − 3ν)∆x∆y −8∆x2 + 2ν − ∆y2 
. . . −4ν ∆x + 4∆y −3(1 − 3ν)∆x∆y −2ν − ∆x2 − 4∆y2
− 2 2 3ν + ∆x∆y


− + −
2 2 2 2

. . . 3(1 − 3ν)∆x∆y −8∆x + 2ν ∆y 3ν ∆x∆y −4∆x − 2ν ∆y   (2.65)
. . . 4ν − ∆x2 + 8∆y2 3ν + ∆x∆y 2ν − ∆x2 − 8∆y2 −3(1 − 3ν)∆x∆y  
... 8∆x2 + 4ν − ∆y2 3(1 − 3ν)∆x∆y 4∆x2 − 4ν − ∆y2  
... 4ν − ∆x2 + 8∆y2 −3ν + ∆x∆y 
... 8∆x2 + 4ν − ∆y2
R T
and the consistent mass matrix M e = N
Ve ρN N dV of the element with uniform density distribution
 40 Chapter 2. Fundaments of Linear Structural Dynamics

that relate to the element’s bilinear shape functions [26] is

 
4 0 2 0 1 0 2 0

 4 0 2 0 1 0 2

 4 0 2 0 1 0
ρtAe  4 0 2 0 1
Me =   (2.66)
36  4 0 2 0

 4 0 2
 4 0
sym 4

or alternatively, using the HRZ lumping scheme

ρtAe
Me = I 8×8 (2.67)
4

Quadratic elements. For quadratic elements, the stiffness matrix simplifies into
Et
Ke = +ν −
×
 96ν
4(3 − ν) 3ν + −2(3 + ν) −3(1 − 3ν) ...
 .
..

 4(1 − 3ν) 3(1 − 3ν) 4ν ...
..
−3ν +

 . 4(3 − ν) ...
..


 . 4(3 − ν) ...
sym ...

. . . −2(3 − ν) −3ν + 3(1 − 3ν)


4ν
... −3ν + −2(3 − ν) −3(1 − 3ν) −2(3 + ν) 
3ν +

... 4ν −3(1 − 3ν) −2(3 − ν) 
+

. . . 3(1 − 3ν) −2(3 + ν) 3ν −2(3 − ν) 
 (2.68)
. . . 4(3 − ν) 3ν + −2(3 + ν) −3(1 − 3ν)

... 4(3 − ν) 3(1 − 3ν) 4ν 

... 4(3 − ν) −3ν + 
... 4(3 − ν)

which is seen to be independent of the length of the element’s edges.

2.3.5 Three-dimensional FE element analysis

Only 3D problems with 1D elements are treated in this book. For 3D problems involving 2D shell
and plate elements or 3D solid elements the reader should consult a specialized book on the finite
element method such as Refs. [5, 8] or [52].
 2.4 Problems 41

2.4 Problems
Problem 2.1 Stiffness and mass matrices of a 3-dof system
Consider the built-up system in the figure.

a) Assemble a stiffness matrix of a 3-dof system combined from the three particle-spring-particle
components by rigidly joining the particles a-to-e, b-to-c and d-to-f. Particles at ends of springs
are restricted to move in 1D only by sliding joints.
b) Make row sum checks of the stiffness matrix to verify that the columns sum to zero.
c) For the same system, establish the related mass matrix when the particle masses are ma = mc =
me = mf = m and mb = md = 2m.

. .......................................................................................
Problem 2.2 Stiffness and mass matrices of two simple truss systems
Consider the planar truss structures A and B in the figure.

a) Assemble the stiffness matrix K of truss system A with two rod elements.
b) Before imposing displacement boundary conditions on A, check that the row-sums of K add to
zero.
c) Impose boundary conditions on A and calculate the joint displacement for the load case shown.
d) Assemble the consistent mass matrix M of truss system A.
e) Check the translational mass of the plane frame in both directions from the mass matrix
elements. Compare with exact mass.
f) Assemble the stiffness matrix K of truss system B.
g) Before imposing displacement boundary conditions on B, check that the row-sums of K add to
zero.
h) Impose boundary conditions on B and calculate the joint displacement for the load case shown.

. .......................................................................................
42 Chapter 2. Fundaments of Linear Structural Dynamics

Problem 2.3 Stiffness and mass matrices of two simple planar frames
Consider the plane frames A and B. NB! There is one hinge joint in each of the frames. The others
are rigid joints.

a) Assemble the stiffness matrix K of planar frame A with two beam&rod elements.
b) Impose boundary conditions and give the resulting stiffness matrix.
c) Assemble the stiffness matrix K of planar frame B.
d) Assume that the tensional deformation is negligible and reduce the problem accordingly (i.e.
reduce its dofs). Give the relevant stiffness matrix K of planar frames A and B.

. .......................................................................................
Problem 2.4 Ground motion affecting an N-storey building
An N-storey buiding with N equal concrete slab floors and similar shear walls between all storeys
is shown in the figure. The building is subjected to ground motion with horizontal acceleration
component ü0 with the other acceleration components being negligible.

a) Express the equations of motion on matrix form.

b) Determine the static deformation matrix S in the relation u = S u0 where u = {u1 ; u2 ; . . . ; uN }T
due to static motion u0 . NB! The matrix can be established by simple inspection in this case.
c) Establish the mass and stiffness matrices for which the ground motion u0 has been condensed.
d) Express the reaction force in shear wall between 1st floor and ground.

GoR 2.12 X1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Problems 43

Problem 2.5 Eigenfrequencies and modes of a simplistic beam model

Consider a simplistic cantilever beam model with two rotational dofs according to the figure. It
is modeled with torsional springs k = 2EI/L and rigid parts between with midpoint translational
masses M = mL/2 with m being the beam’s mass per unit length.
a) Set up the system equations and calculate by hand the eigenfrequencies and modes of the 2-dof
system.
b) Verify that the eigenmodes satisfy the mass and stiffness orthogonality properties.

GoR 2.14 X? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 2.6 Plane stress 2D problem with triangular and rectangular elements
A plate with uniform thickness t is subjected to a localized force f . Data: f =100N, t=1mm,
a=10mm, E=210GPa and ν=0.3.
a) Assemble the stiffness matrix K of the structure using CST and quadratic finite elements with
the displacement boundary condition indicated by the figure.
b) Calculate the vertical displacement of the node on which the force act.
c) Compare with the results of Euler-Bernoulli beam theory for a beam with length 5a and
cross-section 2a × t.
d) Assemble the lumped mass matrix M of the structure using the HRZ mass lumping scheme.

. .......................................................................................
Problem 2.7 Model verification. Mesh refinement and mass lumping
Use a finite element code of your choice that has both the consistent mass and lumped mass
formulations. Do eigenvalue analysis to calculate the first 10 eigenfrequencies with consistent and
lumped mass formulations. Use a model of a 600x800x5mm steel plate. Could you say that the
models with courser meshes are verified if you require that eigenvalues are predicted with a, say,
2% precision?

a) Use quadrilateral elements with 4 cm edges in a first run and then refine the mesh to 2 cm and
then 1 cm elements.
b) Use the finest mesh of a) with consistent and lumped mass matrices.

. ...................................................................................
 3. Linear State-Space Models

3.1 The state-space formulation

A state-space model representation gives a unified approach to the modeling of finite dimensional
dynamical systems. It is often used in control theory and for system identification purpose but
less so in classical structural dynamics theory that more rely on second-order system descriptions.
However, it also suits well in structural dynamics and in particular gives a good bridging link to the
system identification that estimates mathematical models from experimental data.
The state-space system formulation provides a general first-order differential form description
that relate the system response r (t) ∈ Rnr to known system stimulus s (t) ∈ Rns and stimuli from
other little known sources, the so-called process noise w ∈ Rnw . The stimuli in s (t) are assumed
to be known from measurement or from modelling and should be the dominant system stimuli
for an analysis being meaningful. The less dominant stimulus w (t) is not measured or modelled
for practical reasons or pure ignorance. In modern testing, the system responses are measured
with sensors using an electronic data acquisition system (DAQ) that unavoidably mix the true
system response with electric noise v (t) ∈ Rnv . Since unknown stimuli w (t) and sensor noise v (t)
are often best seen as being stochastic, a so-called stochastic state-space description that embeds
deterministic and stochastic parts can be expressed by the state transition differential equation, also
called the dynamic equation together with an output equation for r (t) as

ẋx = A x + B s + E w (3.1a)
r = C x + Ds + v (3.1b)

Here x (t) ∈ RN is the state vector. The system order is thereby N. The matrices A , B , C , D and
E are state-space coefficient matrices out of which A ∈ RN×N is the system matrix, B ∈ RN×ns is
the input matrix, C ∈ Rnr ×N is the output matrix and D ∈ Rnr ×ns is the direct throughput matrix.
In general, for a nonlinear and time-varying system, the system coefficients are state and time
dependent. However, for linear time-invariant systems the matrices are constant coefficient.
In vibration testing in laboratory, the system of interest is often put on soft supports to isolate it
from ambient vibrations and thereby minimize the effect of ambient unknown stimuli. The system
is then excited by some measured stimuli in s (t) that is of sufficient magnitude to totally dominate
 46 Chapter 3. Linear State-Space Models

over w (t) and its responses r (r) are picked up as sensor signals which are filtered, digitized and
processed in the DAQ. In reality, the stimuli s registered by the DAQ is not the true system stimuli
since stimulus sensors are not ideal. Also the support system cannot be made fully ideal and
therefore it does not fully isolate the test object from ambient vibrations which leads to an unknown
vibration source w (t) transmitted to the system. In field testing, the process noise may sometimes
be a major contributor to the noise1 . However, in many cases the stochastic terms may rightfully be
deemed negligible and the signal noise so low that it can be ignored. These circumstances lead to
the deterministic state-space description

ẋx(t) = A x (t) + B s (t) (3.2a)

r (t) = C x (t) + D s (t) (3.2b)

This deterministic state-space model also fit well in structural dynamics simulation which will
be described next. The time domain solution to the initial value problem associated to Eq. (3.2a) is
treated in Ch. 5 and its frequency domain counterpart in Ch. 6.

3.1.1 Structural dynamics in state-space

It is interesting to relate the state-space model structure to a general linear second-order mechanical
system that is so often used in finite element (FE) based structural dynamic modeling. In a very
general setting, the structural dynamics equation reads

M üu + [V
V + G ]u̇u + [K
K + λ K geo + L ]uu = f (t) (3.3)

with the displacement and load vectors u , f ∈ ℜn . The coefficient matrices M , V , G , K , K geo , L
are the mass, viscous damping, gyroscopic, stiffness, geometric stiffness and circulatory matrices
respectively. The λ is the buckling load parameter that can be increased to a level λ = λcr that
gives static buckling. The matrices are often derived by analytical methods using energy principles,
see e.g [30]. The number of model dofs is n and thereby the matrices are n × n.
Let f (t) relate to the nonzero independent stimuli s (t) with the linear relation f = P s s with
P s ∈ RN×ns being a selection matrix. The selection matrix is most often a Boolean matrix used
to in effect select the non-zero loads of f , equating that partition to the non-zero stimuli s and by
P s redistribute the non-zeros back into f . The mass matrix is assumed to be symmetric positive
definite and therefore Eq. (3.3) can be rewritten as

üu + M −1 [V
V + G ]u̇u + M −1 [K
K + λ K geo + L ]uu = M −1 P s s (t) (3.4)

Combined with the trivial equation I u̇u − I u̇u = 0 the two equations can be written as the
state-space dynamic equation on matrix form as
   
0 I 0
ẋx = x+ s ≡ Ax + Bs (3.5)
− M −1 [K
K + λ K geo + L ] − M −1 [V
V + G] M −1 P s

where the state vector x has been introduced so that

   
u u̇u
x= and thus ẋx = (3.6)
u̇u üu

The quantities of interest in mechanical analyses is often a selected set of displacements,

velocities, accelerations or quantities that can be derived from these, e.g. stresses and strains. Let
the quantities of interest be called the responses (a.k.a. outputs) r (t) of the system. In finite element
1 This is often the case in field testing of large civil structures such as bridges with ongoing traffic. Field testing of

offshore platforms with wave loading or wind turbines loaded by wind gust are other examples.
3.1 The state-space formulation 47

vibrational analysis, the responses of the model are often obtained by post-processing the analysis
results. The analysis results in this case are the FE node displacements u and velocities u̇u of the
model. The analyst then specifies what quantities are of interest and lets the post-processor calculate
these using analysis results together with complementary model data. These quantities are often
linearly related to the displacements and velocities given by the analysis, and are therefore natural
ingredients of in a linear state-space model. Displacement and velocity output elements may easily
be extracted. This can be made by letting a selection matrix operate on the state vector x containing
all nodal displacement and velocities of the FE model. Let P d be a Boolean selection matrix that
points to the displacements of interest r d from the displacement partition of the state vector and
P v be the one that points to the velocities of interest r v from the velocity part. Then the output
equation becomes
     
r P 0 C
r (t) ≡ d = d x ≡ d x(t) (3.7)
rv 0 Pv Cv
The time-derivative of the state vector ẋx holds the acceleration data. Let therefore P a be the
selection matrix that points to the accelerations of interest r a from the acceleration partition of the
state vector’s time-derivative ẋx. Then the output equation for accelerations becomes
   
r (t) , r a = 0 P a ẋx = 0 P a [A Ax + B s ] , C a x + D s (t) (3.8)
In summary, for a combined output
 
r d 
r ≡ rv (3.9)
ra
 

the corresponding state-space model quadruple {AA, B , C , D } can be identified to be

   
0 I 0
A= B= (3.10a)
−MM −1 [K M −1 [V
K + λ K geo + L ] −M V + G] M −1 P s
       
Cd Pd 0 Dd 0
C = C v  =  0 Pv  D = D v  =  0  (3.10b)
Ca [00 P a ]A
A Da [00 P a ]B
B
Here it may be noted that only acceleration output has static contribution through the direct
throughput term D a s (t), i.e. direct contribution from the excitation besides that of the dynamics of
the system via the dynamic equation ẋx = A x + B s . This is consistent with Newton’s second law
that directly relates acceleration contribution to force. The velocities and displacements, on the
other hand, need to be obtained as the integral solution to the initial value problem of the dynamic
equation ẋx = A x + B s , x (0) = x 0 .
The velocity output is also related to displacement output by r v = drr d /dt and therefore, as an
alternative to the formulation r = C v x , it also holds that
r v = C d {A
Ax + B s } = C d A x + C d B s = C d A x (3.11)
Pd , 0 ][00; M −1 P s ] = 0 and thus no direct throughput
since it can be verified from above that C d B = [P
to velocity response from s exists in this formulation either. From that it can be concluded that
C v = C dA.
It may also be noted that the selection matrices P s , P d , P v , P a need not be Boolean but can be
general constant coefficient matrices. For instance, for the displacement response r d = C d x the
selection matrix C d may accomodate a coordinate transformation such that the displacements in the
response vector r d can be assocated with another coordinate system than the displacements in x.
For other output quantities which relate linearly to the above quantities, e.g. strains and stresses,
the corresponding state-space description is straight-forward.
 48 Chapter 3. Linear State-Space Models

3.2 State-space realization forms

A state-space model, on the form given by Eq. (3.2a), attempting to mimic the behavior of a
real system may be called a realization within a model structure M of a system. For a multi-
input multi-output system the coefficient matrix quadruple M = {A A , B , C , D} generally holds
2
N + Nnr + Nns + nr ns elements in total. However, for a given input/output relation there is no
unique state-space description. Specific realization forms of technical interest can be obtained
after adding model form constraints. Four such realizations; the balanced realization, the modal
realization, the Jordan form realization, and the coupling realization are of particular interest in this
book. The modal form is applicable for most systems and is a realization in which its states are
fully decoupled. Some systems with esoteric behavior cannot be transformed into a system with
fully decoupled states. For these, the Jordan normal form gives the realization with the minimal
number of couplings between states. The balanced realization is described in Ch. 8.1. The coupling
realization is particularly useful for coupling of state-space models for which force equilibrium and
displacement compatibility need to be enforced at subsystem interfaces and is described in Ch. 10.
The non-uniqueness of the state-space models are given by the following theorem.

Theorem 3.2.1 — State-Space Similarity Theorem. Let a state-space model M = {A C, D}

A, B ,C
be a realization associated to the state vector x . Then a change of variables from x to another set
z (transformation) x = T z with a non-singular square transformation matrix T gives another real-
ization M̄ = {Ā
A, B̄
B, C̄ D} = {T
C , D̄ T −1 A T , T −1 B ,C
C T , D } that has the same input-to-output relation
as M and the models M and M̄ are in this respect similar.

Proof. Setting out from the state-space model M given by Eq. (3.2a) and using the transformation
x = T z give
T żz = A T z + B s and r = C T z + D s
Since T is non-singular this leads to
T −1 T żz = żz = T −1 A T z + T −1 B s ≡ Ā Bs and r = C T z + D s ≡ C̄
Az + B̄ Cz + Ds
which conludes the proof. 

A similarity transformation x = T z that preserves the model’s input/output relation is also

known to preserve the eigenvalues of the system and can thus also be considered similar in that
respect.

3.2.1 State-space realizations

State-space realization on diagonal form. The diagonal form, a.k.a. the modal decomposition
form, is a particular state-space realization that is strongly linked to the free decaying system state,
i.e. s (t) = 0 , from an initial state x (0) = ρ 6= 0 for which it holds that

ẋx(t) = A x (t) , r (t) = C x (t) and x (0) = ρ (3.12)

Under the assumption that the free decay is governed by the solution x(t) = ρ eσt one has

A ρ = σ ρ and r (t) = C ρ eσt (3.13)

for which there are non-trivial solution pairs (σk , φ k ) provided that σk is a root of the characteristic
polynomial det[AA − σk I ] = 0. Such roots are also called system poles and the associated solution
vectors k are the eigenvectors of A. Assuming that A and C stem from a physically realizable
ρ
system (in a physically realizable the response r to a real-valued stimulus s is real, and thus r is
real also when the stimulus is zero), a real form of A ∈ RN×N and C ∈ Rnr ×N is possible. Thus the
3.2 State-space realization forms 49

characteristic polynomial has real-valued polynomial coefficients and the poles are thus either real
or appear in complex-conjugate pairs.
The eigenvalue problem (3.13) for all eigensolutions combined is

AP = PΣ (3.14)

where the modal matrix P has all system eigenvectors ρ k , k = 1, . . . , N as columns and the eigenvalue
matrix Σ has the associated system poles σk , normally sorted in increasing magnitude order, as
elements along its diagonal. It has been shown, see e.g. [20], that Σ is fully diagonal provided that
all system poles are unique, see further Ch. 10. That state-space realization thus has fully decoupled
states. For systems with system poles that are not all unique, but for which some or all appear in
clusters of coalescing poles, it has been shown that a minimal-form Σ has a 2 × 2 block-diagonal
form for the associated poles and is otherwise diagonal. Such systems, with a so-called deficient
system matrix A , are treated in the next section that treats the Jordan normal form while this section
is devoted to systems that can be brought to a fully diagonal form for which the system matrix
A is non-deficient. That includes systems that has repeated eigenvalues but for which A is still
non-deficient.
Using that P −1 A P = P −1 P Σ = Σ together with the state transformation x = P z , the realization
(3.12) becomes

żz = Σ z + P−1 B s ≡ Σ z + B̄
Bs (3.15)
r = C P z + D s ≡ C̄
Cz + Ds

and since Σ is diagonal the first-order differential equation system (3.15) is thus fully decoupled. In
free vibration in which one mode only is active, i.e. zk (t) 6= 0, zm = 0 ∀m 6= k, one notes that the
response is

C :k zk (t)
r (t) = C̄ (3.16)

C (denoted C̄
and the k:th column of C̄ C :k ) is thus the k:th eigenvector of A as seen by the sensors
through the projection of C .
Since the eigenvalues may either be real with real-valued eigenvectors, or appear in complex-
conjugate pairs with associated complex-conjugate eigenvectors, a block-diagonal real form of the
generally complex-valued realization {Λ, B̄ C , D } is possible. For each complex-conjugate pair of
B, C̄
eigenvalues σk = Re{σk } ± iIm{σk } the corresponding 2 × 2 block of the system matrix becomes
 
Re{σk } −Im{σk }
(3.17)
Im{σk } Re{σk }

ρ k } and Im{ρ
and the associated two columns of P become Re{ρ ρ k }.
 Example 3.1 A two-degree-of-freedom problem
Let the parameters of the depicted system be α = 0, β = 1, k = 100 N/m, v = 10 Ns/m and m = 1
kg. Let further the output of the system be the displacement of the right-most mass and the input be
the force applied on the other mass.

........................................................................................
50 Chapter 3. Linear State-Space Models

15
= =1

4
67
10

=1
40
60

=
3

02
53
5

.0
.5

=0
=1
1
0

2
-5

4
-10

-15
-30 -25 -20 -15 -10 -5 0 5

Figure 3.1: Root locus of four system poles for discrete step variation of α with fixed β = 1
(red) and of varying β with fixed α = 1 (black). Arrows indicate increasing parameters α and β .
Asterisks (with values of α and β ) indicate where poles coalesce which result in deficient systems.
........................................................................................

A numerical state-space representation, see Eqs. (3.10a) and (3.11), is then

   
0 0 1 0 0
 0 0 0 1 0  
A=
−200
 B=  C= 0 1 0 0
100 −1 1 1
100 −100 1 −1 0

for which the eigenvalues to three significant digits are σ1,2 = −.0527 ± 6.18i and σ3,4 = −.947 ±
16.14i [rad/s] and the transformed system on diagonal form is
 
−.0527 − 6.180i 0 0 0
 0 −.0527 + 6.180i 0 0 
A=
Ā 
 0 0 −.947 − 16.10i 0 
0 0 0 −.9470 + 16.10i
 
+.269 − .00555i
+.269 + .00555i
B=
B̄ −.427 + .00779i


−.427 − .00779i
 
C = −.00116 + .136i
C̄ −.00116 − .136i .000435 + .0326i .000435 − .0326i

The corresponding real-valued 2 × 2 block-diagonal realization is

 
−.0527 −6.180 0 0
 6.180 −.0527 0 0 
A=
Ā 
 0 0 −.947 −16.10
0 0 16.10 −.947
3.2 State-space realization forms 51

 
.533
 .0111   
B=
B̄ 
 −.853  C = −.00116 .136
C̄ .000435 .0326
−.0158

A root locus plot of pole positions for various combinations of parameters α1 and α2 can be
seen in Fig. 3.1. From that it can be noted that some parameter combinations render coalescing
poles that are the subject of the next numerical example.


State-space realization on Jordan normal form. For some systems with repeated eigenvalues,
i.e. σk = σk+1 = . . . = σk+mk , with multiplicity mk + 1 it is impossible to form the same number of
eigenvectors to diagonalize the system. Such systems have a deficient system matrix A with lesser
than n unique eigenvectors and its principal vectors (sometimes called generalized eigenvectors)
need to be found as the missing columns of P to form the minimal system, called the Jordan normal
form Σ . That system is minimal in the sense that it gives a minimal number of couplings between
its states and at the maximum couples states two-by-two. Systems with rigid body modes or critical
viscous damping are examples of such, as are illustrated by three examples below. The structure of
the minimal coupling form is given by the following theorem.

Theorem 3.2.2 — Jordan Normal Form Theorem. If A ∈ RN×N , then there exists a full rank
P ∈ CN×N such that P −1 A P = diag(JJ 1 , . . . , J t ) is block diagonal with Jordan blocks J k related
to eigenvalues σk with multiplicity mk and ∑tk=1 mk = N. The k:th such block is
 
σk 1 0

 σk 1 

Jk = 
 . .. . .. 

 
 σk 1
0 σk m ×m
k k

Proof. See [21]. 

 Example 3.2 Three variants of the two-degree-of-freedom example

Let the parameters of the system in√Fig. 3.1 be such that; (a) α = 1 and β = 0, (b) α = 0 and
β = 1.553, and (c) α = 1 and β = 2. In all three cases k = 100 N/m, v = 10 Ns/m and m = 1
kg. This will render; (a) an undamped system with one rigid-body mode, (b) a system with one
critically damped mode, and (c) a system with both one rigid-body mode and one critically damped
mode. Principal value calculation will then give the following transformation matrices and resulting
system matrices on Jordan normal form:
(a)
   
.5 0 +.0177i −.0177i 0 1 0 0
.5 0 −.0177i +.0177i 0 0 0 0 
P=  A
Ā =  
0 0 +.25 +.25  0 0 −14.1421i 0 
0 0 −.25 −.25 0 0 0 14.1421i
52 Chapter 3. Linear State-Space Models

(b)
 
−.0378 −.0015 +.0222 − 0.1153i +.0222 + .1153i
+.0545 +.0043 −.0111 − 0.1500i −.0111 + .1500i
P= 
+.5685 −.0158 +.7537 + 0.2039i +.7537 − .2039i
−.8200 −.0106 1 1

 
−15.0372 1 0 0
 0 −15.0372 0 0 
A=
Ā 
 0 0 −.4931 + 6.6319i 0 
0 0 0 −.4931 − 6.6319i

(c)
 
+.0499 +.0035 −.7071 0
−.0499 −.0035 −.7071 0 
P=
−.7053

+.0002 0 −.7071
−.0002 0 −.7071
 
−14.1421 1 0 0
 0 −14.1421 0 0
A=
Ā 
 0 0 0 1
0 0 0 0

 3.3 Problems 53

3.3 Problems
Problem 3.1 SS model of 3-dof mechanical system
Specify the state-space matrices A , B , C and D of the undamped system in the figure. The response
vector r is here composed of the acceleration ü2 (t) of the second carriage and the support reaction
R(t), i.e. the response vector is r = [ü2 R]T .

a) Set up the state-space model for the case that all masses are non-zero.
b) Set up the state-space model for the case when m2 = 0 while the other two masses are non-zero
and thus M is singular.

. .......................................................................................

Problem 3.2 SS model of 2-dof mechanical system

Consider a 2-dof system according to the figure.

a) Set up its state-space matrix quadruple {A C , D } for the single force input s1 and where the
A, B ,C
two outputs r1 and r2 are the compressive forces in the springs.
b) Set up its state-space matrix quadruple {A C , D } for three other outputs r1 , r2 and r3 where
A, B ,C
r1 is the displacement, r2 is the velocity and r2 is the acceleration that are all three co-linear
with the force s1 .

P3.13a-b X9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problem 3.3 SS model with acceleration sensors
For the 3-dof system shown, there are two sensors (black squares) that sense accelerations. Cal-
culate the system eigenvectors as they are observed by the sensors. Use a state-space representation.
Data: k = 1 × 106 N/m, m = 1kg and c = 5 × 103 Ns/m.

P3.13c X9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54 Chapter 3. Linear State-Space Models

Problem 3.4 State-space modelling with various output

For the 2dof system depicted, the force stimulus s(t) is generated by an electromagnetic shaker.
The system responses are measured by one accelerometer mounted on the heavier mass and one
extensiometer measuring the relative displacement between the masses. Express the elements of the
A4×4 , B 4×1 , C 2×4 and D 2×1 ) of a state space model in the quantities k and m.
state-space matrices, (A
Here the output vector r is the measured acceleration ü2 over the relative displacement u2 − u1 .

12/18/2009-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 3.5 State-space modelling with current input and velocity output
For the 2dof system depicted, the force f (t) is generated by an electromagnetical shaker that is fed
by an alternating current i(t) such that f (t) = c0 i(t). The system response is measured by a velocity
meter mounted on the heavier mass. Express the state-space matrices A 4×4 , B 4×1 , C 1×4 , D 1×1 of a
state space model with parameters k, m and c0 . Here the stimulus s is the alternating current i(t)
and the single output r is the measured velocity u̇2 .

10/20/2009-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
 4. Decoupling of System States

4.1 Modal decomposition

Solutions for eigenvalues and eigenmodes, the eigensolution in short, and their use to simplify
structural dynamics analysis has been a very effective enabler for the analysis of large scale systems.
Such analysis provide not only the resonance frequencies of the system that could mean serious
damage to the system if a harmonic excitation should be present for some time at (or close nearby)
those frequencies. Very important in this context is also the eigensolution’s usefulness to break
down, or decompose, the coupled equations of motion into a set on much more manageable
uncoupled scalar differential equations in time. This can be made to viscously damped structural
dynamics equations (under some restrictions) or the state-space equations alike. While the structural
dynamics equations are decomposed into a set of real-valued second-order ordinary differential
equations, the state-space model is decomposed into a set of complex-valued first-order ordinary
differential equations. The procedures will be described in the following.

4.1.1 Decomposition of the undamped 2nd order structural dynamics equation

The decomposition of the governing equations of motion takes its starting point in the free vibration
of the undamped structural dynamics system that is

M üu(t) + K u(t) = 0 (4.1)

Using the assumption that in harmonic stationarity the displacement is governed by the time
function u (t) = ûueiωt this leads to [K
K − ω 2 M ]ûueiωt = 0 and by that the relation

K − ω 2 M ]ûu = 0
[K (4.2)

This system of equations can have non-trivial solutions ûu 6= 0 provided that the determinant
K − ω 2 M | = 0 which is known as the characteristic equation (in variable ω 2 ) of the system
|K
K − ω 2 M | being the characteristic polynomial in ω 2 . As the order of the system is n, the
with |K
characteristic equation provides n distinct roots ωk2q
, k = 1, . . . , n which are the eigenvalues of the
system. The associated angular frequencies ωk = + ωk2 are the system’s natural frequencies, also
58 Chapter 4. Decoupling of System States

called the system’s resonance frequencies. The eigenvalues are distinct and numerable, but they
may not be unique. Some eigenvalues may coalesce and groups of such coalescing eigenvalues
are the so-called repeated eigenvalues that are often present for systems with various degrees of
symmetry. Mathematically it can be proved that for system with symmetric positive definite mass
matrix M > 0 and symmetric positive-semidefinite stiffness matrix K ≥ 0, the roots ω 2 are all
positive and real. It is the non-trivial vectors ûuk , φ k associated to the eigenvalues ωk2 , i.e. the
solutions to

[−ωk2 M + K ]φφ k = 0 (4.3)

that have the remarkable property that they collectively can be used to decouple the system equations.
This property, the so-called orthogonality property, is stated in the following theorem:

Theorem 4.1.1 — Modal decomposition theorem. Let K and M be symmetric with M > 0
and K ≥ 0. Let further ωk2 and φ k be a positive real eigenvalue and its associated real-valued
eigenvector of the partial eigenvalue problem K φ k = ωk2 M φ k and consider the case when all
eigenvalues ωk2 are unique. Since M > 0 then φ Tk M φ k , µk > 0 and φ Tk K φ k , γk ≥ 0. Then any
two different modes φ i and φ j are both M-orthogonal φ Ti M φ j = 0 and K-orthogonal φ Ti K φ j = 0
for all i 6= j. Let further K Φ = M Φ Ω 2 be the full eigenvalue problem with the modal matrix Φ
holding the eigenmodes φ k , k = 1, . . . , n as columns and Ω 2 , diag(ωk2 ). Then it follows from
the above that Φ T M Φ = diag(µk ) and Φ T K Φ = diag(γk ) are decoupling transformations that
simultaneously diagonalizes M and K .

Proof. Let K φ i = ωi2 M φ i and K φ j = ω 2j M φ j be two partial eigenvalue problems for two unique
eigensolutions i and j. Pre-multiply the first of these with the mode of the second and vice
versa which leads to the two equations φ Tj K φ i = ωi2 φ Tj M φ i and φ Ti K φ j = ω 2j φ Ti M φ j . Since
M and K are both symmetric it holds that the second of these equations can be transposed into
φ Tj K φ i = ω 2j φ Tj M φ i . Subtracting the second transposed equation from the first equation gives
0 = (ωi2 − ω 2j )φφ Tj M φ i and thus φ i is M-orthogonal to φ j with φ Tj M φ i = 0 since ωi2 6= ω 2j . If, on
the other hand, the first equation is multiplied by ω 2j φ Tj and the the second by ωi2 φ Ti , using the same
symmetry argument, it leads to that ω 2j φ Tj K φ i = ω 2j ωi2 φ Tj M φ i and ωi2 φ Tj K φ i = ωi2 ω 2j φ Tj M φ i .
Subtracting the second of these from the first leads to (ωi2 − ω 2j )φφ Tj K φ i = 0. Again, since ωi2 6=
ω 2j , it follows that φ Tj K φ i = 0 and the conclusion that the modes are also K-orthogonal to one
another. 

It can be proved that the M and K orthogonality properties still holds for repeated eigenvalues,
see [48] for details.

Decoupling of the undamped equations. The modal matrix Φ = [φφ 1 φ 2 . . . φ n ] can be used
to decouple the forced undamped structural dynamic equation

M üu(t) + K u(t) = f (t) (4.4)

Because the modal matrix is non-singular (its columns are orthogonal and therefore linearly
independent) it can be used in a unique linear forward transformation of variables u = Φ η and
backwards as η = Φ −1 u . Since Φ is time invariant, and therefore üu = Φ η̈
η , the equations of motion
takes the following form after a pre-multiplication with Φ T

Φ T M Φ η̈
η (t) + Φ T K Φ η (t) = diag(µk )η̈ η (t) = Φ T f (t) , Ψ (t)
η (t) + diag(γk )η (4.5)
 4.1 Modal decomposition 59

The rows of the transformed system loading Ψ , i.e. the scalar time function ψk (t) = φ Tk f (t)
are known as the modal loads of the system, the rows of the generalized displacement vector η
are known as the modal displacements ηk while the positive scalars µk and γk are known as the
modal masses and modal stiffnesses respectively. The coupled equation system in Eq. (4.3) is by
the transformation converted into a decouped system of n equations

µ1 η̈1 (t) + γ1 η1 (t) = ψ1 (t) = φ T1 f (t) (4.6a)

µ2 η̈2 (t) + γ2 η2 (t) = ψ2 (t) = φ T2 f (t) (4.6b)
..
.
µn η̈n (t) + γn ηn (t) = ψn (t) = φ Tn f (t) (4.6c)

that can be solved independently as ordinary second order equations in time when combined with
the initial conditions for modal displacements η 0 = Φ −1 u 0 and modal velocities η̇ η 0 = Φ −1 u˙0 given
by the backward transformation from the physical displacements u 0 and velocities u˙0 .
One special case is the free vibration with f (t) = 0 . Using the harmonic assumption for the k:th
mode, i.e. ηk (t) = η̂k eiωt , we have for the non-trivial solution at ω = ωk that (γk − ωk2 µk )η̂k eiωt = 0
and thus γk − ωk2 µk = 0. This gives to the relation between the modal stiffness and the modal mass
as

γk = µk ωk2 ∀k = 1, 2, . . . , n (4.7)

The analytical and numerical time domain solutions for the differential equations in Eq. (4.6a)
will follow in Ch. 5 and the algebraic frequency domain solution for these equations in Ch. 6.

4.1.2 Rayleigh quotient. Rayleigh’s and Courant’s eigenvalue theorems

The Rayleigh quotient is a function that maps an arbitrary vector φ to a scalar ω 2 via two scaling
matrices K and M such that
φ T Kφ
ω2 = (4.8)
φ T Mφ
It can easily be verified that the Rayleigh quotient is scale-invariant so that αφφ gives the same
Rayleigh quotient as φ provided that α is an arbitrary scalar number that is not zero. The Rayleigh
quotient has interesting properties that makes it a useful tool in the theory of free vibration.
For the undamped free vibration problem of the k:th mode it holds that

K φ k = ωk2 M φ k or φ Tk K φ k = ωk2 φ Tk M φ k (4.9)

This leads to a Rayleigh quotient relation for the k:th eigenvalue that reads

φ Tk K φ k
ωk2 = (4.10)
φ Tk M φ k

The Rayleigh quotient for any vector φ̃φ ≈ φ k that is an approximation to the k:th eigenvector
gives an approximation to the k:th eigenvalue ωk2 . To see this, let φ̃φ = φ k +εθθ where θ = ∑ j6=k α j φ j ,
α j are arbitrary scalars, ||φφ k || = ||φφ k || ∀ j, and ε is a small number. Because of the mass and
stiffness orthogonality of the eigenmodes it can be verified that ωk2 φ Tk M θ = φ Tk K θ . Without loss
T
of generality, let the vectors be mass orthonormalized so that φ Tk M φ k = φ̃φ M φ̃φ = 1, a relation
which leads to (since M is symmetric) that

2φφ Tk M θ = −εθθ T M θ (4.11)

60 Chapter 4. Decoupling of System States

Figure 4.1: Rayleigh quotient ω 2 for the 2-dof and 3-dof examples shown for modes normalized
so that ||φφ || = 1. For the 2-dof example it can be verified that: φ 1 = {−0.52; −0.85} and φ 2 =
{0.85; −0.52}. For the 3-dof example: φ 1 = {−0.33; −0.59; −0.73}, φ 2 = {0.73; 0.33; −0.59}
and φ 3 = {−0.59; 0.74; −0.33} (red crosses). NB! only half of space shown (e.g. φ1 > φ2 for
2-dof) since other half gives mirror image and correspond to Rayleigh quotient for −φφ .
. .......................................................................................

The Rayleigh quotient with arbitrary mass-orthonormalized φ̃φ and symmetric K is then
T
φ̃φ K φ̃φ
2
ω = T = φ Tk K φ k + 2εφφ Tk K θ + ε 2 θ T K θ = ωk2 + 2εωk2 φ Tk M θ + ε 2 θ T K θ (4.12)
φ̃φ M φ̃φ
= ωk2 − ε 2 ωk2 θ T M θ + ε 2 θ T K θ = ωk2 + ε 2 θ T [K
K − ωk2 M ]θθ (4.13)
The eigenvalue approximation error ∆ωk2 , ω 2 − ωk2 is thus
∆ωk2 = ε 2 θ T [K
K − ωk2 M ]θθ (4.14)
and the error is of second order in the small factor ε. Since d∆ωk2 /dε|ε=0 = 0, the Rayleigh quotient
is stationary in the vicinity of an eigenvector. In conclusion: For an eigenvector approximation
that is in error by the order of ε, the Rayleigh quotient provides an eigenvalue approximation that is
in error to the order of ε 2 .
If we let the eigenvalues be in ordered sequence so that ω12 ≤ ω22 ≤ . . . ≤ ωn−1 2 ≤ ωn2 then
the smallest value the Rayleigh quotient can take is thus ω1 and the largest value is ωn2 . These
2

two eigenvalues are thus the absolute minimum and maximum the quotient can take for any φ̃φ .
The minimum is then at ω12 for φ̃φ = φ 1 and the maximum is at ωn2 for φ̃φ = φ n . Whether the
Rayleigh quotient is at maximum or at minimum at that stationary point depends on the sign of
θ Tk [K
K − ωk2 M ]θθ k which is indeterminate since K − ωk2 M is indefinite for all ωk2 > 0.
Some illustrations of the Rayleigh quotient are provided i Fig. 4.1.

Courant’s minimax principle. Courant’s minimax principle, in its use for vibration systems,
relate to eigenvalue re-positioning of a system that is subjected to linear constraints on its displace-
ments u . A scalar such linear constraint can be expressed as c T u = 0 (see example below) with
4.1 Modal decomposition 61

c being a constant coefficient vector of same size as u and with r linearly independent and linear
constraints we have that C T u = 0 with C ∈ ℜn×r . The principle is formulated in the following
theorem:
Theorem 4.1.2 — Courant’s minimax theorem. The (r + 1):st eigenvalue of an unconstrained
vibrating system is the maximum value that the minimum of the Rayleigh quotient can take when
the system is subjected to r added linear independent constraints.

Proof. See [9]. 

Since the minimum of the Rayleigh quotient is the minimum eigenvalue of the modified system,
Courant’s minimax principle gives an upper bound on how much the system’s first eigenvalue could
be increased by structural modifications with extra imposed displacement constraints. A typical
example is the addition of extra bearings to support a vibrating axle to increase its fundamental
eigenfrequency or extra stay supports to stiffen vibrating masts.

 Example 4.1 — Constraints imposed on a 3-dof system. Consider the 3-dof system in its
original and in two constrained configurations shown in the figure below. It the original configuration
the stiffness and mass matrices are
   
2 −1 −1 1 0 0
K = k −1 2 −1 and M = m 0 1 0
−1 −1 2 0 0 1

and its tree eigenvalues are ωk2 = (0, 3, 3) × k/m. In the second configuration (middle in figure) a
constraint u3 = 0 is imposed so that c T u = 0 with c T = {0, 0, 1}. The system is reduced to a 2-dof
system with stiffness and mass matrices
   
2 −1 1 0
K =k and M =m
−1 2 0 1

and its two eigenvalues become ω̄12 = k/m and ω̄22 = 3k/m which are seen to agree with Rayleigh’s
theorem on scalar constraints. The first eigenvalue increases and the second stays where it was
before the constraint was added.
62 Chapter 4. Decoupling of System States

In the third configuration (bottom in figure) another constraint u2 = 2u1 is imposed by a rigid
link arrangement so that the resulting constraint equation becomes C T u = 0 with
 
T 0 0 1
C =
2 −1 0

The system is reduced to a 1-dof system with stiffness K = 5k and mass M = 3m and its eigenvalue
becomes ω̄k2 = 5k/3m which is seen to agree with Courant’s minimax theorem for multiple scalar
constraints. It also agrees with Raigleigh’s theorem on one added scalar constraint as the first
eigenvalue has increased from k/m into 5k/3m which is not beyond the second eigenvalue 3k/m of
the system before the second constraint was added. 

Rayleigh’s theorem on scalar constraint. As the Rayleigh quotient is at its absolute minimum
ω12 only when u ≡ φ 1 , a scalar constraint c T u = 0 on u may imply that u cannot be made identical
to φ 1 . Thus the minimum of the Rayleigh quotient u T K u /uuT M u under the condition that c T u = 0
can only be larger than, or equal to, ω12 . Let the smallest eigenvalue of the constrained system be
denoted ω̄12 and the above reasoning leads to that ω12 ≤ ω̄12 .
On the other hand the Rayleigh quotient of the unconstrained system is at its maximum value
2
ωn when φ n . Imposing a scalar constraint on the system in effect reduces its dofs by one, and its
2 . With the same argument as above, an imposed constraint on u may
largest eigenvalue is thus ω̄n−1
imply that u cannot be identical to φ n and thus the maximum value the Rayleigh quotient can take
2
for the modified system is ω̄n−1 ≤ ωn2 .
It can be shown that the bounding holds for any eigenvalue so that ωk2 ≤ ω̄k2 ∀k. This is
formulated in a theorem that states that:
Theorem 4.1.3 — Rayleigh’s theorem on scalar constraints. A system with the eigenvalue
sequence ω12 ≤ ω22 ≤ . . . ≤ ωn−1
2 ≤ ωn2 that is subjected to a scalar constraint ct u = 0 on its
displacements u has an eigenvalue sequence ω̄12 ≤ ω̄22 ≤ . . . ≤ ω̄n−1
2 for which it holds that
2 2 2 2 2 2
ω1 ≤ ω̄1 ≤ ω2 ≤ ω̄2 ≤ . . . ≤ ω̄n−1 ≤ ωn .

An added constraint to a system thus leads to that none of its eigenvalues will be smaller and if any
of its eigenvalues is increased, the increase is maximally up to the next-in-order of the unconstrained
system.

Rayleigh’s theorems on added stiffness or mass. Let ∆K K > 0 be an added stiffness imposed on
an original system {K
K , M } and a theorem by Rayleigh addresses the impact this has on the system
eigenvalues:

Theorem 4.1.4 — Rayleigh’s theorem on added stiffness. For a system {K

K , M } subjected
K > 0 its eigenvalues either stay or increase.
to added stiffness ∆K

Proof. The theorem follows from the observation that the Rayleigh quotient is

u T [K
K + ∆KK ]uu u T K u u T ∆K
Ku uT K u
ω̄ 2 = = + ≥ = ω2 (4.15)
uT M u uT M u uT M u uT M u
K is positive semi-definite.
since ∆K 

M > 0 be an added mass imposed on the system {K

Instead let ∆M K , M } and another theorem by
Rayleigh states:
 4.1 Modal decomposition 63

Theorem 4.1.5 — Rayleigh’s theorem on added mass. For a system {K

K , M } subjected to
M > 0 its eigenvalues either stay or decrease.
added mass ∆M

Proof. The Rayleigh quotient gives

uT K u uT K u uT K u
ω̄ 2 = = ≤ = ω2 (4.16)
u T [M
M + ∆MM ]uu u T M u + u T ∆M
M u uT M u
M is positive semi-definite.
since ∆M 

4.1.3 Decomposition of the damped 2nd order structural dynamics equation

While the real-valued eigenmodes of the undamped eigenvalue problem can be used to fully
decouple the equations of the undamped system, they cannot generally be used to do so for the
viscously damped system. To see this we again use the transformation u = Φ η , and thus also
u̇u = Φ η̇
η and üu = Φ η̈
η , in the damped structural dynamic equation

M üu(t) +V
V u̇u(t) + K u (t) = f (t) (4.17)

After a pre-multiplication with the modal matrix transposed we have

Φ T M Φ η̈
η (t) + Φ T V Φ η̇
η (t) + Φ T K Φ η (t) =
diag(µk )η̈ V η̇
η (t) + V̄ η (t) = Φ T f (t) , Ψ (t)
η (t) + diag(γk )η (4.18)

for which we may note that the transformed viscous damping matrix V̄V = ΦT V Φ is only decoupled
by the eigenmodes under certain circumstances that will be treated in the following.
In modelling, the physical properties and physical behaviour of the system that makes up
the stiffness and mass matrices are usually well known. These matrices are normally based on
sound first principles. The physical phenomena that attributes to damping are usually much more
involved. That is one reason why the modelling of damping is more sketchy. It would often take too
much effort to model it very precisely. The second reason is that for small damping, the response
behaviour is very much dominated by the structure’s mass and stiffness properties and not its
damping. For that reason a very precisely modelling of damping would often give little gain. An
approximation made is then to disregard the off-diagonal elements of the transformed viscous
damping matrix as

V ≈ Ṽ
V̄ V = diag[Ṽ
V ] , diag(2ζk µk ωk ) (4.19)

which defines the relative modal damping ζk and leads to the decoupled equations

diag(µk )η̈
η (t) + diag(2ζk µk ωk )η̇ η (t) = Φ T f (t) , Ψ (t)
η (t) + diag(γk )η (4.20)

or, since also γk = µk ωk2 ,

η̈1 (t) + 2ζ1 ω1 η̇1 (t) + ω12 η1 (t) = ψ1 (t)/µ1 (4.21a)

η̈2 (t) + 2ζ2 ω2 η̇2 (t) + ω22 η2 (t) = ψ2 (t)/µ2 (4.21b)
..
.
η̈n (t) + 2ζn ωn η̇n (t) + ωn2 ηn (t) = ψn (t)/µn (4.21c)

The diagonalization approximation of the viscous damping matrix is particularly good, see
Ref. [17], when the damping is small and the eigenvalues are well separated.
64 Chapter 4. Decoupling of System States

Caughey damping. The severe difficulties in modelling damping accurately from physical
principles has motivated another route to account for the various dissipation phenomena. That route
is based on estimations of relative modal damping in the system modes. Such estimations may stem
from physical testing or experience. With the modal damping ζk given by estimation, one seek
to construct a viscous damping matrix that correspond to that damping and which preserves the
real-valued eigenvectors of the undamped system.
Assume that we have the modal matrix Φ given by the undamped full eigenvalue problem
K Φ = M Φ Ω 2 . Since the transformed viscous damping matrix is V̄ V = Φ T V Φ the physical viscous
damping matrix is given by

V = Φ−T V̄
V Φ−1 (4.22)

which involves the inverse of the modal matrix. A relation for its inverse can be established using
the orthogonality property of the eigenmodes. Using that diag(µk ) = Φ T M Φ we note that

I = diag(1/µk ) diag(µk ) = diag(1/µk ) Φ T M Φ and also I = Φ −1 Φ (4.23)

and therefore

Φ −1 = diag(1/µk )Φ
ΦT M (4.24)

Combined with Eq. (4.22) this gives

ΦT M ]T V̄
V = [diag(1/µk )Φ ΦT M ] = M T Φ diag(1/µk )V̄
V [diag(1/µk )Φ ΦT M (4.25)
V diag(1/µk )Φ

V =
which can be further simplified using that the target modal damping matrix is diagonal V̄
diag(2ζk µk ωk ) and M is symmetric, and therefore
n
2ζk ωk
ΦT M =
V = M Φ diag(2ζk ωk /µk )Φ ∑ {M M φ k }T
M φ k }{M (4.26)
k=1 µ k

A viscous damping V so constructed is called a Caughey damping and is the most general form
that can be diagonalized by the real-valued eigenvectors of the undamped eigenproblem.

Rayleigh damping. While the Caughey damping is the most flexible form that can be de-
coupled by real modes, it has the disadvantage that the modal damping ζk needs to be specified
for all n eigenmodes. In a practical situation there may not be sufficient information available to
substantiate a specification to that level of detail. The Rayleigh damping assumption might then
provide a good balance between available information and modelling accuracy. In the Rayleigh
damping assumption it is assumed that the damping is distributed as a combination of the mass and
the stiffness. Following the assumption, the viscous damping matrix can be written as

K +βM
V = αK (4.27)

with α and β being positive scalar factors which provides a two-factor flexibility in the modelling
of damping. Since the damping is linearly proportional to the stiffness K and mass M this damping
model is also known as proportional damping model. The decoupled equations of motion for that
damping model are

Φ T M Φ η̈
η (t) + Φ T V Φ η̇
η (t) + Φ T K Φ η (t) =
diag(µk )η̈
η (t) + diag(β µk + αγk )η̇ η (t) = Φ T f (t)
η (t) + diag(γk )η (4.28)
4.1 Modal decomposition 65

Figure 4.2: Modal damping ζk versus eigenfrequency ωk for combined stiffness and mass propor-
tional Rayleigh damping (solid curve).
........................................................................................

and thus the equivalent relative modal damping ζk can be obtained from the relations 2ζk µk ωk =
β µk + αγk and γk = µk ωk2 as

ζk = (αωk2 + β )/2ωk (4.29)

In practical situations the Rayleigh damping factors α and γ are often determined from estimated
relative modal damping ζi and ζ j of two eigenmodes with eigenfrequencies ωi and ω j that are in a
frequency range of interest. For those two modes it holds that

ζi = (αωi2 + β )/2ωi ζ j = (αω 2j + β )/2ω j (4.30)

which leads to the relations for the factors

2(ζ j ω j − ζi ωi ) 2ωi ω j (ζi ω j − ζ j ωi )
α= β= (4.31)
ω 2j − ωi2 ω 2j − ωi2

from which it follows that the modal damping for the k:th mode is given by

ωk (ζ j ω j − ζi ωi ) ωi ω j (ζi ω j − ζ j ωi )
ζk = + (4.32)
ω 2j − ωi2 ωk (ω 2j − ωi2 )

This two-factor relation for the relative modal damping is illustrated in Fig. 4.2. It is seen that the
damping tend to infinity for eigenmodes with eigenfrequencies that tend to zero (i.e. rigid-body
modes) unless β = 0 and it also grows with increasing eigenfrequency ωk unless α = 0.

Augmented damping. Since the Caughey and Rayleigh damping models both can be used in
the decoupling of the system equations but both have their own drawbacks, a combination of the two
can be a middle ground with best use of their individual strengths. This is the idea of the augmented
damping. The advantage of the Caughey damping model is that the relative modal damping can
be specified for each individual mode but, on the other hand, estimates of the modal damping for
all system modes are seldom at hand. For modes of which there are no modal estimates given,
the damping can be set to some arbitrary default value (including zero). The Rayleigh damping
model on the other hand imply modal damping on all modes without discrimination. It gives
increasingly higher damping the higher the eigenfrequency of the mode provided it holds a stiffness
proportional term. Such artificial high damping on high-frequency modes can be advantageous
in the numerical integration of the system equations since high-frequency oscillations outside a
frequency range of interest will then die out fast without seriously affecting the simulation accuracy
for the low-frequency dynamics. The augmented damping model assign the appropriate damping to
a set of modes for which the modal damping is known by use of a modified Caughey model while
 66 Chapter 4. Decoupling of System States

the modal damping of the remaining modes are given by the Rayleigh damping model. Let the set
of modes for which the damping is known1 be denoted K and thus the modes numbered k ∈ K are
the modes in the set. The diagonalizable viscous damping matrix

2ζk ωk − αωk2 − β
K +βM +
V = αK ∑ {M M φ k }T
M φ k }{M (4.33)
k∈K µk

then lead to the appropriate damping ζk for the modes in the set while the remaining modes k ∈
/K
2
are assigned the Rayleigh damping values ζk = (αωk + β )/2ωk .

General viscous damping. When the damping is modelled from first principles to a form that
is outside the range of the Caughey damping then the real modes of the undamped eigenproblem
do not suffice to provide a decoupled second-order form. To decouple the system equations one
then have to resort to the first-order description in the state-space formulation, see Sect. 3.1.1.

4.1.4 Spectral decomposition of stiffness and mass relations

Besides the decoupling properties of the modal solution, some other useful spectral properties can
be deduced from the mass and stiffness orthogonality of the undamped system’s eigenmodes. To
see that, let K and M be the coefficient matrices of a system with eigenmodes φ j , modal masses
µ j = φ Tj M φ j and modal stiffnesses γ j = φ Tj K φ j = µ j ω 2j . Let us further assume that both are of
full rank (this implies that K > 0 and M > 0) and therefore the system do not possess rigid body
motion and non of the system eigenvalues are zero. Since the eigenmodes are linearly independent
it is possible to express any vector x as a linear combination of eigenmodes as
n
x= ∑ α jφ j (4.34)
j=1

The combination coefficients α j can be obtained using the mass and stiffness orthogonality
property of the eigenmodes. Let us pre-multiply Eq. (4.34) with the i:th mode and M to get
n
φ Ti M x = ∑ α j φ Ti M φ j = αi µi (4.35)
j=1

Also we can use the stiffness orthogonality of the modes to get

n
φ Ti K x = ∑ α j φ Ti K φ j = αi µi ωi2 (4.36)
j=1

The two relations leads to two expressions for the linear combination coefficients

1 T 1
αi = φ i Mx = φ T Kx (4.37)
µi µi ωi2 i

This leads to two equivalent possible modal expansions of x

n n
1 1
x= ∑ µ j φ Tj M x φ j and x = ∑ µ j ω 2 φ Tj K x φ j (4.38)
j=1 j=1 j

1 In
aerospace the modes for which much is known after ground vibration testing, including modal damping, might
be; wing bending and torsional modes, fuselage bending modes, stabilizer bending and torsional modes, control surface
modes, etc.
 4.1 Modal decomposition 67

or since φ Tj M x and φ Tj K x are both scalars this is equivalent to

n n
1 1
x = [∑ φ j φ Tj ]M
M x or x = [ ∑ 2
φ j φ Tj ]K
Kx (4.39)
µ
j=1 j j=1 µ j ω j

From these relations one can conclude that spectral expansions of the unit matrix are
n n
1 1
I = [∑ φ j φ Tj ]M
M and I = [ ∑ φ φ T ]K
2 j j
K (4.40)
j=1 µ j j=1 µ j ω j

from which one obtains spectral relations for the inverses of the mass and stiffness matrices as
n n
1 1
M −1 = ∑ µ j φ j φ Tj and K −1 = ∑ µ j ω 2 φ j φ Tj (4.41)
j=1 j=1 j

A specific form of mass and stiffness relation for the vector K −1 M φ j appear in some useful
structural dynamics methods such as the mode acceleration method that will be described later.
Using the above relations one notes that this takes a surprisingly simple form as
n n
1 1
K −1 M φ j = ∑ µ j ω 2 φ j φ Tj M φ j = ∑ ω 2 φ j (4.42)
j=1 j j=1 j

4.1.5 State-space modal decomposition

The modal decomposition to decouple the structural dynamics system equations also apply to the
state-space models. Since the system matrix A is not necessarily symmetric, some eigensolution
properties that holds for the undamped symmetric {KK , M } system do not hold here. This require a
more general approach to the state-space eigenproblem, but the basics are the same.
Let us assume that the solution to the homogeneous state-space equation ẋx = A x is on the form

x(t) = eσt ρ (4.43)

where σ and ρ are a scalar constant and a vector, respectively. Introduce the solution (4.43) in
ẋx = A x and divide through by eσt , to obtain

Aρ = σ ρ (4.44)

Eq. (4.44) represents a set of homogeneous algebraic equations and is the basic algebraic
eigenvalue problem. The eigenvalue problem is to determine the values of σ for which Eq. (4.44)
has non-trivial solutions. These are given by the characteristic polynomial |A
A − σ I |. Recalling that
A is an N × N matrix, the eigenvalue problem can be satisfied in N different ways, namely

A ρ k = σk ρ k k = 1, 2, . . . , N (4.45)

where σk and ρ k are the eigenvalues and eigenvectors of A . Since A is not necessarily symmetric, the
characteristic roots may be complex-valued. However, since A is real, the characteristic polynomial
has real coefficients and therefore the roots are either real or appear in complex conjugate pairs.
Another eigenvalue problem of interest is associated with the matrix AT and is known as the
adjoint eigenvalue problem defined by

A T λ k = σk λ k k = 1, 2, . . . , N (4.46)
68 Chapter 4. Decoupling of System States

Because det(A AT − σ I ), the eigenvalues of A and A T are the same. The eigen-
A − σ I ) = det(A
vectors λ j are known as the adjoint of the set of eigenvectors ρ j . Eq. (4.46) can also be written in
the form

λ Tj A = σ j λ Tj j = 1, 2, . . . , N (4.47)

Because of their position relative to A , ρ k are called right eigenvectors of A and λ j are known
as the left eigenvectors of A .
If we pre-multiply Eq. (4.45) with λ Tj and post-multiply Eq. (4.47) with ρ k we have

λ Tj A ρ k = σk λ Tj ρ k (4.48a)
λ Tj A ρ k = σ j λ Tj ρ k (4.48b)

or by subtracting the two

λ Tj ρ k = 0
(σk − σ j )λ (4.49)

If the eigenvalues are distinct, i.e. σk 6= σ j , we note that the right and left eigenvectors
corresponding to different eigenvalues are bi-orthogonal, i.e. λ Tj ρ k = 0. We also note that they are
bi-orthogonal with respect to A since λ Tj A ρ k = λ Tj ρ k = 0 for k 6= j.
It is normal practice to normalize the eigenvectors so as to satisfy λ Tk ρ k = 1 and ||λ
λ k || = ||ρ
ρ k ||,
in which case the eigenvectors are bi-orthonormal, as expressed by

λ Tj ρ k = δ jk (4.50a)
λ Tj A ρ k = σ j δ jk (4.50b)

where δ jk is the Kronecker delta. Next let us introduce the N × N modal matrices of right and left
eigenvectors2

P = [ρ
ρ1 ρ2 ... ρN] Λ = [λ
λ 1 λ 2 ... λ N] (4.51)

as well as the matrix of eigenvalues

Σ = diag(σ
σ k) k = 1, 2, . . . , N (4.52)

and then Eq. (4.51) can be written on compact matrix form

ΛT P = I (4.53a)
T
Λ AP = Σ (4.53b)

Eq. (4.53b) is seen to constitute a similarity transformation, i.e. A and Λ T A P are similar
matrices, meaning they share the same eigenspectrum Σ.
In a general case, in which the eigenvalues of A are not distinct but appear in clusters of multiple
eigenvalues, the diagonal form may not be obtainable. For such, so-called deficient matrices, a
similar matrix of minimum number of non-zero elements is in block-diagonal form. Such matrices
are known as the Jordan normal form, see Sect. 3.2.1. Since the cases with deficient system
matrices A are very rare in practice, let us here only consider the non-deficient case for which
diagonalization of A is possible.
2 Note that P and Λ are the Greek capital letter counterparts to ρ and λ .
 4.2 Eigenvalue enclosure methods 69

We note that for the non-deficient problems the right and left modal matrices effectively
decouples the state equations. With the similarity transformation x = P z and in accordance with
Eqs. (3.2a), we have
Λ T P żz = Λ T A P z + Λ T B s (4.54a)
y = C Pz + Ds (4.54b)
Using Eq. (4.53) this simplifies to
żz = Σ z + Λ T B s (4.55a)
y = C Pz + Ds (4.55b)
We thus have an N-dimensional set of uncoupled first order differential equations
żk = σk zk + λ Tk B s , σk zk + b k s k = 1, 2, . . . , N (4.56a)
N N
y (t) = ∑ C ρ k zk + D s (t) , ∑ c k zk + D s (t) (4.56b)
k=1 k=1
which, in general, are complex-valued. However, since A is a real-valued matrix the roots of the
A − σi I ) = 0 are either real or appear as complex conjugate pairs. We
characteristic equation det(A
find, using rules from complex algebra, that if one eigensolution {σk , λ k , ρ k } has been found for
the two adjoint eigenproblems
A ρ k = σk ρ k and λ Tk A = σk λ Tk (4.57)
then also the complex conjugate solution {σ j , λ j , ρ j } = {σk∗ , λ ∗k , ρ ∗k } is an eigensolution. For
real-valued roots σk also the right and left eigenvectors are real, and the corresponding differential
equation in (4.55) is thus real-valued. If the roots are indeed complex, the corresponding pair of
conjugate differential equations are
żk = σk zk + b k s (t) (4.58a)
ż j = σ j z j + b j s = σk∗ z j + b ∗k s (t) (4.58b)
and thus give complex conjugate solution pairs, i.e. z j = z∗k . The contribution to the output of those
two complex conjugate states are
∆yy(t) = c k zk (t) + c j z j (t) = c k zk (t) + c ∗k z∗k (t) = 2Re{cck zk (t)} (4.59)
since
c j = C ρ j = C ρ ∗k = c ∗k (4.60)
It thus suffice to calculate either of the two complex conjugate solutions for obtaining the
contribution to the output, with potential of halving a computational effort. Analytical time domain
solution procedures for solving these equations are given in Ch. 5.1 and numerical procedures in
Ch. 5.4.

4.2 Eigenvalue enclosure methods

Eigenvalue calculation methods are basically of two types. However, since there are no non-
iterative method to solve for the roots of high-order characteristic polynomials all eigensolution
methods are iterative. These methods can be classified as being either enclosure methods, that will
precisely bracket in (enclose) the position of eigenvalues, or matrix iteration methods. Since the
matrix iteration methods are more efficient but does not guarantee to find all eigenvalues they are
often combined with enclosure methods to give robust eigensolution procedures. The eigenvalue
enclosure methods are described first in this text and then the matrix iteration methods, but in
practical use the sequence is usually the opposite.
 70 Chapter 4. Decoupling of System States

Figure 4.3: A Gerschgorin disk in the complex σ plane. A possible location of an eigenvalue σk is
indicated by a bullet.

4.2.1 Gerschgorin’s disks for eigenvalue enclosure

Under certain circumstances, a theorem of Gerschgorin, provides easily obtainable estimations on
the eigenvalues of the general N × N matrix A . For heavily banded and diagonally dominant3 , and
ultimately tridiagonal4 , matrices A it often gives close bounds at a very modest computational cost.
Let us consider the eigenvalue problem A ρ = σ ρ in index notation
N
∑ ai j ρ j = σ ρi i = 1, 2, . . . , N (4.61)
j=1

Then assuming that ρk is the component of the vector ρ with the largest modulus, i.e. |ρk | =
max|ρ j | ∀ j = 1, 2, . . . , N, we let i = k in Eq. (4.61) and write

N
(σ − akk )ρk = −akk ρk + ∑ ak j ρ j (4.62)
j=1

Using the triangle inequality this leads to

N N
|σ − akk ||ρk | ≤ −|akk ||ρk | + ∑ |ak j ||ρ j | ≤ |ρk |(−|akk | + ∑ |ak j |) (4.63)
j=1 j=1

that after dividing through by |ρk | lead us to

N
|σ − akk | ≤ −|akk | + ∑ |ak j | , rk (4.64)
j=1

First, we observe that |σ − akk | represents the distance from the point akk in the complex plane
to the eigenvalue σ , so that the inequality (4.64) may be interpreted geometrically as a circle in
the complex plane with center at the diagonal element akk and radius rk (see Fig. 4.3). Then, as
Eq. (4.61) admits N solutions, we let k = 1, 2, . . . , N and express the inequality in the form of
Gerschgorin’s theorem:

Theorem 4.2.1 — Gerschgorin’s theorem. Every eigenvalue of a matrix A ∈ CN×N lies within
at least one of the circular disks with centers at akk and radii rk = ∑Nj=1 |ak j | − |akk |.

3A
diagonally dominant matrix A is such that |aii |  ∑ j6=i |ai j | ∀i.
4A
tridiagonal matrix T has non-zero elements only on its diagonal and its first sub- and super-diagonals, i.e. its
elements ti j = 0 for i > j + 1 and j > i + 1.
 4.2 Eigenvalue enclosure methods 71

Figure 4.4: Two Gerschgorin disks associated to a real symmetric positive definite matrix A that
have collapsed into segments along the σ axis. Bullets indicate possible locations of σ1 and σ2 .
........................................................................................

These disks are often referred to as the Gerschgorin disks. It may be noted that for real
symmetric matrices A , the disks in the complex σ plane collapse to segments along the real σ axis,
see Fig. 4.4. Also, we note for the tridiagonal matrix A = T , that the sum in the inequality (4.64)
involves only two terms and that in this case
|σ − tkk | ≤ |tk,k−1 | + |tk,k+1 | (4.65)
Therefore, for diagonally dominant T , i.e. |tkk |  |tk,k−1 | + |tk,k+1 |, we may get a very tight
enclosure for the eigenvalue. Such are often obtained as the result of tridiagonalization of real
symmetric matrices using Householder transformation, see Ch. 4.4.2.

4.2.2 Givens’ method with Sturm sequence checking

As will be demonstrated, the Givens’ method is a very efficient technique for enclosing the
eigenvalues of a real symmetric tridiagonal positive definite matrix T n×n . It do so by counting the
roots σ of the characteristic polynomial |TT − σ I | without actually requiring explicit evaluation
of the polynomial. As modern methods for computing eigenvalues in a narrow spectrum of
large eigenvalue problems are often based on tridiagonalization such as the Lanczos’ method
(that will be explained later), such methods can be used to the advantage in conjunction with
Givens’ method. The method is versatile in that respect it admits a precise enclosing of individual
eigenvalues and operates most effective when their approximate locations are known in advance.
These approximations could, for example, be the centers of Gerschgorin’s discs. The characteristic
determinant associated to the tridiagonal symmetric matrix T has the form
 
T11 − σ T12 0 ... 0 0
 T12
 T22 − σ T23 ... 0 0  
 0 T23 T33 − σ . . . 0 0 
|T
T − σ I | = det( . .. ) (4.66)
 
.. .. .. ..
 .. . . . . . 
 
 0 0 0 . . . Tn−1,n−1 − σ Tn−1,n 
0 0 0 ... Tn−1,n Tn,n − σ
We note that if any Ti j = 0 (i 6= j) the problem of obtaining the determinant can be broken down
to smaller sub-problems, each having their own characteristic polynomials with roots common with
the original problem. Without loss of generality we may therefore assume that all Ti j 6= 0 (i 6= j).
Denoting by pi (σ ) the principle minor determinant of i:th order of the matrix T − σ I , and with the
definition of p0 ≡ 1, it can be shown by induction5 that
2
p1 (σ ) = T11 − σ and pi+1 (σ ) = (Ti+1,i+1 − σ )pi − Ti,i+1 pi−1 i = 2, 3, . . . , n − 1 (4.67)
5 This capitalizes on Laplace’s formula for determinants; Let A n×n have minors ai j that are the sub-determinants
72 Chapter 4. Decoupling of System States

Some important properties of the polynomials pi (σ ), out of which of course pn (σ ) is the

characteristic polynomial, may be established without actual knowledge of their explicit expressions.
Let us consider an interval a < σ < b on the real σ axis where a and b are not roots of any
polynomial pi (σ ). The first step of Givens’ method is the determination of the number of roots
of the characteristic polynomial lying in the interval [a, b]. To get this we exploit the following
statements that relate to the principal determinant polynomials:

I. If pi−1 (σ ) = 0 then pi (σ ) and pi−2 (σ ) are non-zero and of opposite signs.

II. As σ passes through a zero of pn (σ ), the quotient pn (σ )/pn−1 (σ ) changes sign from positive
to negative.

To demonstrate statement I, we assume that pi−1 (σ ) = 0 and obtain from Eq. (4.67) that
2
pi (µ) = −ti−1,i pi−2 (µ) (4.68)

If we further assume that pi (σ ) = 0, then according to Eq. (4.67) pi−2 (σ ) must also be zero,
so that three consecutive polynomials in the sequence are zero. Then we must conclude from Eq.
(4.67) that pi−3 (σ ) = pi−4 (σ ) = . . . = p0 (σ ) = 0 , which is a contradiction since p0 (σ ) = 1 by
definition. Hence, if pi−1 (σ ) = 0, then pi (σ ) 6= 0 and pi−2 (σ ) 6= 0, so that we may conclude from
Eq. (4.68) that pi (σ ) and pi−2 (σ ) must have opposite signs.
Further, to prove statement II, we use a Rayleigh theorem that states that eigenvalues of T never
decrease when the system giving T is subjected to a constraint and that the eigenvalues’ upper
bounds are set by the sequence of eigenvalues of the constrained system matrix. First, we denote
the roots of the characteristic polynomial pn (σ ) by σ1 , σ2 , . . . , σn and assume that they are ordered
so as to satisfy σ1 < σ2 < . . . < σn . Moreover, the polynomial pn−1 (σ ) represents the determinant
of the matrix obtained by constraining T − σ I by striking out its last row and column. We denote
the roots of pn−1 (σ ) by the ordered sequence σ10 < σ20 < . . . < σn0 . Then, according to Rayleigh’s
theorem, the two sets of eigenvalues satisfies the inequalities

σ1 ≤ σ10 < σ2 ≤ σ20 < . . . < σn−1 ≤ σn−1

0
< σn (4.69)

A typical plot of pn (σ ) and pn−1 (σ ) is shown in Fig. 4.5, in which the vertical lines through
σ1 , σ10 , σ2 , . . . separate regions in which the ratios pn (σ )/pn−1 (σ ) are of opposite sign. Note that,
of A with the i:th row and j:th column removed. The cofactors ci j of A are the positive or negative minors defined as
A| = ∑ni=1 ci j Ai j
ci j = (−1)i+ j ai j . The Laplace formula gives a cofactor column expansion of the determinant of A as: |A
for any j.

........................................................................................

Figure 4.5: Regions with signs of the ratio pn (σ )/pn−1 (σ ). It is seen that the sign changes from
positive to negative when σ passes an eigenvalue σi .
4.2 Eigenvalue enclosure methods 73

Figure 4.6: Plots of the polynomials p0 (µ), . . . , p4 (µ) indicating the sign changes from top p0 to
bottom p4 at σ = µ.
........................................................................................

since the matrix T is positive definite, both pn (0) and pn−1 (0) are positive. It is clear from the figure,
as σ passes through the roots σ1 , σ, . . . , σn , the sign of pn (σ )/pn−1 (σ ) changes from positive to
negative. It follows that the sequence of polynomials p1 (σ ), p2 (σ ), . . . , pn (σ ) fulfills statements I
and II. A sequence of polynomials possessing these characteristics is known as a Sturm sequence.
By this we are now in the position to consider Sturm’s theorem which states that:

Theorem 4.2.2 — Sturm’s first theorem. If the polynomials p0 , p1 (σ ), . . . , pn (σ ) represent a

Sturm sequence on the interval [a, b] and if s(σ ) denotes the number of sign changes in the
consequtive sequence of numbers p0 (σ ), p1 (σ ), . . . , pn (σ ), then the number of roots of the
polynomial pn (σ ) in [a, b] is equal to s(σ = b) − s(σ = a).

Proof. Sturm’s theorem can be proved by induction. To do so, let us assume that the number of
sign changes s(µ) in the sequence of numbers p0 (µ), p1 (µ), . . . , pn (µ) is equal to the number of
roots of pn (σ ) corresponding to σ < µ. As an example, we show the sequence of five polynomials
p0 (σ ), . . . , p4 (σ ) in Fig. 4.6. For the particular value of µ shown, there are two sign changes
in the sequence of numbers p0 (µ), . . . , p4 (µ) and there are exactly two roots, σ1 and σ2 , of the
characteristic polynomial p4 (µ) for σ < µ. As µ increases, the number s(µ) remains the same
until µ crosses the root σ3 , at which point s(µ) increases by one. This can be explained by the
fact that, according to the second property of the Sturm sequence, the number of sign changes
remains the same as µ crosses any root of pi−1 (σ ), i = 1, 2, . . . , n. At the same time, according to
statement II, there is an additional sign change as µ crosses a root of pn (σ ). Hence, the number
s(µ) increases by one every time µ crosses a root of pn (σ ). 

At this time it should be obvious that the characteristic polynomials of the principal minors
are not needed in explicit form. Only their signs at µ1 = a and µ2 = b are needed. These can be
calculated recursively using Eq. (4.67). When searching for a specific eigenvalue, the enclosure
[a, b] containing it is usually being narrowed down by the use of the bisection method.
74 Chapter 4. Decoupling of System States

An alternative formulation utilizes the Sturm sequence property of the L D L T factorization (see
Ch. 4.4.1) of the symmetric system matrix Z (σ ) = K − σ M = L D LT . It has the advantage that
no tridiagonalization operations are required to establish the Sturm sequence. Using the structure
of the L matrix (it has ones on the diagonal and zeros above) one has that det(L L) = det(L L)T = 1
and thus det(ZZ (σ )) = det(L L(σ )D D(σ )L LT (σ )) = det(D D(σ ). The principle minor determinant of i:th
T
order is thus pi = det(L L1:i,1:i D 1:i,1:i L 1:i,1:i ) = ∏ij=1 d j (σ ). With p0 ≡ 1 one notes that the number
of sign shifts of the series p0 , p1 , . . . , pn is equal to the number of negative diagonal elements of D .
This leads to an alternative form of the Sturm theorem:
Theorem 4.2.3 — Sturm’s second theorem. If the polynomials p0 (σ ), p1 (σ ), . . . , pn (σ ) with
pi = ∏ij=1 d j (σ ) represents a Sturm sequence of the characteristic polynomial K − σ M on the
enclosure [a, b] and if s(σ ) denotes the number of negative signs of the numbers d j (σ ) , j =
1, 2, . . . , n of D in L D L 0 = K − σ M then the number of roots of the polynomial pn (σ ) in [a, b] is
equal to s(σ = b) − s(σ = a).

The following example will illustrate the process for both formulations in parallel.
 Example 4.2 — Sturm sequence checking for 3dof system. The 3-dof system in the figure

has the three eigenvalues ω12 = 0.069k/m, ω22 = 0.584k/m and ω32 = 3.097k/m. We pretend for a
while that we do not know them but instead want to calculate how many eigenfrequencies that are
below a certain number, say ω 2 = 2k/m so that σ = 2, by Sturm sequence checking (that number
is obviously two). This will be done using Sturm’s first and second theorems in a) and b) below.

a) To bring the system to symmetric standard form we factorize the mass matrix so that M =
M M 1/2 which leads to the eigenvalue problem K φ = ω 2 M 1/2 M 1/2 φ or with φ = M −1/2 φ̄φ to
1/2

the symmetric standard form M −1/2 K M −1/2 φ̄φ , T φ̄φ = ω 2 φ̄φ . In this case it also gives a tridiagonal
matrix T since the mass matrix is diagonal and the stiffness matrix is tridiagonal with

   
√ 1 0 0 3 −1 0
M 1/2 = m 0 2 0 and K = k −1 2 −1 (4.70)
0 0 2 0 −1 1
it can be verified that
   
3 −0.50 0 t11 t12 0
T = (k/m) −0.50 +0.50 −0.25 , (k/m) t12 t22 t23  (4.71)
0 −0.25 +0.25 0 t23 t33
The Sturm sequence is thus
p0 , 1 (4.72)
p1 = t11 − σ = 3 − 2 = 1.0
p2 = (t22 − σ )p1 − t12 p0 = (0.5 − 2) · 1.0 − 0.52 · 1 = −1.75
2

2
p3 = (t33 − σ )p2 − t23 p1 = (0.25 − 2)(−1.75) − 0.252 · 1.0 = 3.0
Since there are two sign swaps, between p1 & p2 and p2 & p3 , the Sturm sequence checking
tells that there are two eigenvalues below ω 2 = 2k/m (as it should!).
 4.2 Eigenvalue enclosure methods 75

b) On the other hand a LDL0 decomposition of K − ω 2 M (see Ch. 4.4.1) gives

 
3−2·1 −1 0
K − ω 2 M = k  −1 2−2·4 −1  = kL LD L T = (4.73)
0 −1 1−2·4
   
1 0 0 1 0 0 1 −1 0
k −1 1 0 0 −7 0  0 1 0.143
0 0.143 1 0 0 −6.86 0 0 1

which is seen to hold two negative diagonal elements in D . According to this, the system thus has
two eigenvalues below ω 2 = 2k/m (as it should!). 

4.2.3 Wittrick-Williams eigenvalue counting algorithm

The Sturm sequence check for discretized systems enables us to find out exactly how many
eigenfrequencies that are enclosed within in a specified frequency domain. Using the Sturm
sequence checking after any eigensolution extraction method thus leave us with the possibility
to ensure that no mode has been missed by the solution algorithm. Under certain conditions, of
which the exact representation of distributed parameter (continuous) systems are the most notable,
however, the Sturm sequence checking breaks down. This has to do with the fact that not all of the
generalized degrees-of-freedom are present in the vector of unknowns u of the eigenvalue problem

Z (ω )u = 0 (4.74)

where Z (ω) is the dynamic stiffness matrix. Since not all the system’s dofs are included in u, also
such vibratory solutions exist such that the trivial solution u = 0 and |Z
Z | = 0 is important. As an
example, let us consider the exact dynamic reduction of the system
       
M aa M ab üua K aa K ab u a fa
+ = (4.75)
M ba M bb üub K ba K bb u b fb =0

When the system is vibrating in stationary harmonic motion with amplitude ûu at angular
frequency ω we have

K aa − ω 2 M aa K ab − ω 2 M ab ûua
       
Z aa Z ab ûua f
, = a (4.76)
K ba − ω 2 M ba K bb − ω 2 M bb ûub Z ba Z bb ûub 0

We may eliminate the degrees-of-freedom ûub using the second row equation Z ba ûua + Z bb ûub = 0
to receive

Z aa + Z ab Z −1
[Z ua , Z̄
bb Z ba ]û Z aa ûua = f̂f a (4.77)

In free harmonic vibration, i.e. when f̂f a = 0 , situations may occur in which ûua = 0 and
|Z̄
Z aa | = 0 and ûub 6= 0 , see Fig. 4.7.
Taking the Sturm sequence check as a basis, Wittrick and Williams [50] , devised a method for
the exact computation of the number of natural frequencies in specified frequency ranges of such
systems. They noted the similarity of computing the number of sign changes of the determinants
of the principle minors of matrix Z and the number of negative diagonal elements of D given by
LDL0 -factorization of Z , i.e. Z = L D L T . Define s(σ ) as the negative sign count of the dynamic
stiffness matrix Z (σ ) established at the trial frequency µ, i.e. the number of negative diagonal
elements of D (µ). Also, define as J0 (σ ) the number of eigenfrequencies of the system with the
dofs u a fixed to zero. That is the basis of the Wittrick-Williams theorem that is stated as:
 76 Chapter 4. Decoupling of System States

Figure 4.7: A 3-dof system. When the dofs u1 and u3 are condensed, the system posses an eigenvalue
with the displacement vector equal to zero, here the single element u2 = 0. The corresponding
mode of the condensed dofs is indicated by dashed rectangles.
........................................................................................

Theorem 4.2.4 — Wittrick-Williams theorem. The number of eigenfrequencies J of a system

below a certain trial frequency ω = µ is equal to the sum of the sign count s(µ) of the dynamic
stiffness matrix added with the number of eigenfrequencies J0 (µ) of the system with all elements
of the condensed displacement vector fixed, i.e. J(µ) = s(µ) + J0 (µ).

Proof. See reference [50]. 

4.3 Matrix iteration

One obvious route to take to solve the eigenproblem would be to first solve for the roots ωk2 of
the characteristic equation |KK − ω 2 M | = 0 and from them obtain the pertaining eigenvectors from
the eigenvalue equation K φ k = ωk2 M φ k . These roots could be found by enclosure methods by, for
instance, bi-sectioning to an arbitrary accuracy. However, such procedures have been found to be
less efficient than matrix iteration methods. Matrix iteration methods get their efficiency from fast
matrix-vector multiplications in a computer and the good convergence behaviour of the problem’s
eigenvectors. However, the matrix iteration methods do not warrant that all eigensolutions that
could be of interest are solved for and are therefore often combined with enclosure methods to
ensure that all eigenvalues (and its vectors) in a given range (or complex domain) are solved for.
The inner working and the convergence properties of two well-known matrix iteration methods is
described next.

4.3.1 Inverse iteration

It is easy to verify that the original undamped eigenproblem K φ = ω 2 M φ can be brought to
its shifted form (KK + σ M )φφ = (ω 2 + σ )M M φ through trivial left-hand side and right-hand side
additions of σ M φ . Introducing the shifted stiffness matrix K σ , K + σ M and the shifted eigenvalue
ωσ2 , ω 2 + σ then lead to the shifted eigenproblem that reads

K σ φ = ωσ2 M φ (4.78)

At this point, such trivial shifting operation may seem pointless. However, one notes that if the
system possess at least one rigid body mode, and thus have a positive semi-definite stiffness matrix
K ≥ 0, the system is brought to a form with a positive definite shifted matrix K σ > 0 provided
σ > 0 since the mass matrix M is positive definite. That is useful since such K then cannot be
inverted because its singularity while K σ can be inverted since it is non-singular.
The inverse iteration scheme iterates on Eq. (4.78) to give progressively better approximations
of a system eigenvector setting out from a starting guess ψ guess of that vector. The starting guess
does not have to be particularly good, but must not be orthogonal to the eigenvector searched for.
An update from the guess is then obtained by solving for a better eigenvector approximation φ̃φ as

K σ {ωσ−2 φ̃φ } , K σ ψ update = M ψ guess (4.79)

4.3 Matrix iteration 77

Since eigenvectors are invariant to scaling, the resulting scale factor ωσ−2 6= 0 embedded in
ψ update is immaterial. The update should be a better approximation, and can be used as a better
guess in another round for an even better update, and so on in an iterative manner. This thinking
leads to the inverse iteration algorithm:

BASIC I NVERSE I TERATION WITH S HIFT

1: K , M , σ , kits , ψ 0 , φ ) . Iterate kits times, ψ 0 is guess

procedure BASIC I NVERSE I TERATION(K
2: Kσ = K + σM
3: for k = 1, . . . , kits do
4: ψ k = K −1σ M ψ k−1
5: end for
6: return φ = ψ k . This is the eigenvector approximation
7: end procedure

It is natural to ask if such procedure converges to the true eigensolution as the number of
iterations tends to infinity. A convergence analysis can give the answer. As premises for that
analysis, let the eigensolution of (4.78) be ordered such that ω1σ 2 ≤ ω 2 ≤ . . . ≤ ω 2 and let the
2σ nσ
n
starting guess ψ guess , ψ 0 be a linear combination ψ 0 = ∑ j=1 α j φ j of the true system eigenmodes
φ j that is required to be non-orthogonal to φ 1 , i.e. α1 6= 0. We then have for the steps of the
iteration sequence:
n n n
ψ 1 = K −1 −1 −1 −2
σ M ψ 0 = K σ ∑ αi M φ i = ∑ αi K σ M φ i = ∑ αi ωiσ φ i =
i=1 i=1 i=1
n n
−2 −2 −2 ω1σ 2
α1 ω1σ φ 1 + ∑ αi ωiσ φ i = ω1σ (αi φ i + ∑ αi ( ) φ i) (4.80a)
i=2 i=2 ωiσ

n n
ω1σ 4
K −1
ψ 2 =K −4 −4
σ M ψ 1 = ∑ αi ωiσ φ i = ω1σ (αi φ i + ∑ αi ( ) φ i) (4.80b)
i=1 i=2 ωiσ
..
.
n n
ω1σ 2k
K −1
ψ k =K −2k −2k
σ M ψ k−1 = ∑ αi ωiσ φ i = ω1σ (αi φ i + ∑ αi ( ) φ i) (4.80c)
i=1 i=2 ωiσ
−2k 2 < ω 2 , ω 2 , . . . , ω 2 and α 6= 0. The k:th
and thus ψ k → ω1σ α1 φ 1 as k → ∞ provided ω1σ 2σ 3σ nσ 1
approximation ψ k thus hold insignificant components of other eigenmodes than φ 1 when k is large.
By virtue of the convergence analysis, three properties of the inverse iteration algorithm are
obvious. A first observation is that the updated vectors may grow without bounds as k → ∞ if
ω1σ2 < 1. A proper normalization, e.g. by ψ ← ψ /||ψ ψ k ||, after each computation of ψ k can
k k
cure this problem without affecting the convergence at a small extra computational cost. The
second observation is that a good approximation of the true eigensolution can only be guaranteed if
kits is large. A practical iteration stop criterion for the algorithm is thus needed to strike a good
balance between computational accuracy and speed. The third observation is that the algorithm
converges to the eigensolution with the minimal eigenvalue ω1σ provided α1 6= 0 and therefore to
the corresponding eigenvalue ωr2 (and its eigenvector φ r ) that give the smallest ω1σ 2 = ω 2 + σ . By
r
2 2
selecting a proper shift σ such that σ ≈ −ωr and thus the corresponding ω1σ is small, the algorithm
can be taylored to converge to the r:th eigensolution of the original eigenproblem K φ = ωr2 M φ .
A termination criterion for the eigensolution convergence can be obtained by use of the Rayleigh
quotient. Let the approximation of ω1σ 2 at the k:th iteration be denoted ω 2 with the Rayleigh
1k
78 Chapter 4. Decoupling of System States

quotient

2
ω1k = ψ Tk K σ ψ k /ψ
ψ Tk M ψ k (4.81)

2
The quotient of two successive approximations ω1(k−1) 2 can be monitored during the
and ω1k
iteration process and when the quotient approaches unity the iterations can come to a stop. The
approximation to the r:th eigenvalue is then ωr2 = ω1k
2 − σ . This is the basis for the following

algorithm:

I NVERSE I TERATION WITH S HIFT, N ORMALIZATION AND A DAPTIVE S TOP

1: procedure I NVERSE I TERATION(K K , M , σ , ε, ψ 0 , ωr2 , φ r ) . Find solution close to shift

2: Kσ = K + σM . Shift K
3: 2 = ψ T K ψ /ψ
ω10 ψ T Mψ . Rayleigh quotient eigenvalue approximation
0 σ 0 0 0
4: 2 = ω2 + 1
k = 1, ω11 10
while |ω1k2 /ω 2
5: 1,k−1 − 1| > ε do
6: ψ k = K −1
σ M ψ k−1 . Update eigenvector approximation
7: ψ k ← ψ k /||ψ ψ k || . Normalize
8: 2 T
ω1k = ψ k K σ ψ k /ψ T
ψk Mψ k . Rayleigh quotient of update
9: k ← k+1
10: end while
11: return ωr2 = x k . This is eigenvalue approximation
12: return φ r = x k . This is eigenvector approximation
13: end procedure

Another observation from the convergence analysis is that the sequence of eigenvector approxi-
mations ψ 0 , ψ 1 , ψ 2 , . . . become increasingly rich on the true eigenvectors closest to the solution
2 , φ } because of the factors (ω /ω )2 < 1 that shrinks more rapidly for eigensolutions with
{ω1σ 1 1σ iσ
2 further away from ω . The sequence of eigenvector approximations, here as columns of
ωiσ 1σ
matrix T such that T = [ψ ψ 0 , ψ 1 , ψ 2 , . . .], is thus a sequence of vectors with strong domination of
eigenvectors associated with the eigenspectrum close to the smallest eigenvalue ω1σ 2 . This sequence
−1 −2
{MM ψ 0 , K M ψ 0 , K M ψ 0 , . . .}, known as the Krylov sequence, is utilized in the construction of
the Lanczos method that is particularly good for finding a subset of eigensolutions. That method is
described in Sect. 4.3.2.

Repeated eigenvalues. For structures with various forms of symmetry it is not uncommon that
their eigenvalues appear in tight clusters or even coalesce. That may happen also for structures
without such symmetries and is thus a more generic issue. Systems with two or more rigid-body
modes are perfect examples of such. For these the corresponding eigenvalues at zero repeat with a
multiplicity that corresponds to the number of linearly independent rigid-body modes. For each
group of coalescing eigenvalues, the convergence analysis reveals that convergence is to one specific
linear combination of the modes. To see this, let the start guess be ψ 0 = ∑ni=1 αi φ i where we assume
that the factors αi of the m repeated eigenvalues closest to ω1σ2 are not all zeros, i.e. m α φ 6= 0 .
∑i=1 i i
2 2 2
Using that ω1σ = ω2σ = . . . = ωmσ and repeating the convergence analysis under that extra premise
4.3 Matrix iteration 79

we have that
m n m n
ω1σ 2
K −1
ψ 1 =K −2
σ M ψ 0 = ∑ αi ωiσ φ i + ∑ −2
αi ωiσ −2
φ i = ω1σ ( ∑ αi φ i + ∑ αi ( ) φ i)
i=1 i=m+1 i=1 i=m+1 ωiσ
(4.82a)
..
.
m n
ω1σ 2k
K −1
ψ k =K −2k
σ M ψ k−1 = ω1σ ( ∑ αi φ i + ∑ αi ( ) φ i) (4.82b)
i=1 i=m+1 ωiσ

and we note that the convergence is towards the contribution of the repeated eigensolutions that is
present already in the starting guess, i.e. the contribution ∑m
i=1 αi φ i . To converge to other, linearly
independent, eigenmodes the inverse iteration could be restarted with other randomized guesses to
form a set of modes from which an orthogonal set could be constructed. However, this strategy is
not without foreseeable problems and the orthogonalization process described next can be a remedy.

Gram-Schmidt orthogonalization. For repeated eigenvalues for which the eigenvectors

are not unique any linear combination from a specific set of orthogonal eigenvectors is also an
eigenvector. The inverse iteration as describe above, however, converges to the eigenvector given by
the content of the starting guess ψ 0 and does not give the full orthogonal set of eigenmodes. This
problem is solved by an orthogonalization process, known as the Gram-Schmidt orthogonalization
process, that successively deflates from the guess already found eigenvectors of the orthogonal set.
It takes support from the following theorem.

Theorem 4.3.1 — Gram-Schmidt theorem. Let the columns of matrix X = [xx1 · · · x m ] be

vectors of a mass orthogonal set of vectors for which x Ti M x j = 0 ∀i 6= j. Let further µ j be
a mass normalization constant such that µ j = x Tj M x j . Given a vector x̃xk of size consistent
with M that cannot be linearly combined from X , i.e. x̃xk − X α 6= 0 ∀αα , a nonzero vector x k
that is mass orthogonal to the vector set of X can be constructed by the deflation operation
x k = (II − ∑mj=1 µ −1 T xk
j x j x j M )x̃

Proof. An orthogonality check between two vectors x i and x k with i 6= k gives: x Ti M x k = x Ti M (II −
∑mj=1 µ −1 T xk = x Ti M x̃xk − ∑mj=1 µ −1
j x j x j M )x̃
T T x = x T M x̃
j x i M x j x j M x̃ k i xk − µ −1 T T x = x T M x̃
j x i M x i x i M x̃ k i xk −
−1 T
µ j µ j x i M x̃xk = 0. The vectors x k and x i , i = 1, . . . , m are thus M-orthogonal. 

The orthogonalization technique lay ground for a more general inverse iteration procedure that
gives a set of mass and stiffness orthogonal eigenvectors with eigenvalues that are closest to a given
shift. When computed, the eigenvalue approximations can be obtained by use of the Rayleigh
quotient. The procedure is as follows:

I NVERSE I TERATION

1: procedure I NVERSE I TERATION(K K , M , m, σ , ε, x guess , Φ ) . Solve for m eigenvectors

2: Kσ = K + σM . Shift K
3: X = empty . X empty at start
4: for n = 1 to N do
5: x old = 0 . Temporarily
6: X T M )xxguess
x new = (II − Xp . Deflate
7: x new ← x new / x Tnew M x new . Mass normalize
8: while ε < ||xxnew − x old || < 2 − ε do . Loop until x new ≈ ±xxold
 80 Chapter 4. Decoupling of System States

9: x old ← x new . Update old

10: xnew ← K −1 µ M x old . Update new
11: x new ← (II − Xp X T M )xxnew . Deflate
12: x new ← x new / x Tnew M x new . Mass normalize
13: end while
14: X ← [X X , x new ] . Concatenate to columns of X
15: end for
16: return Φ = X . Eigenvector approximations in columns of Φ
17: end procedure

4.3.2 Lanczos’ method

While the inverse iteration computational scheme uses the Krylov sequence and ortogonalization to
successively deflate found eigensolutions closest to the shift, the Lanczos method takes another
stance to the use of the Krylov sequence. In the inverse iteration, the iteration vectors are made
orthogonal to the previously found (approximations of) system eigenvectors and iterates from the
same starting vector a number of times before another eigenvector approximation is found. In that
iteration process many, successively better, eigenvector approximations are computed, but only
the final and best is kept before iterations to the next sought eigensolution can commence. The
Lanczos method, however, does not waste Krylov vectors in that manner and can therefore be
more computationally efficient. The Lanczos method uses the Krylov sequence for reduction of the
eigenproblem and leaves the iterative solutions for the reduced system to a final stage.
The convergence analysis of the inverse iteration with shift (see Eqs. (4.80a) - (4.80b))
revealed that the Krylov sequence {M M ψ 0 , K −1 −2
σ M ψ 0 , K σ M ψ 0 , . . .} is increasingly dominated by
the eigenvectors closest to the shift σ . That property of the Krylov sequence is taken advantage of
by the Lanczos method. It is constructed by a sequence of Krylov vectors that are made orthogonal
and used for system reduction. Since the vectors are dominated by the eigenvectors closest to the
shift, the reduced system give good approximations of the eigensolutions in the spectrum closest to
the shift.
The Lanczos method is based on the standard eigenvalue problem formulation H φ = ω 2 φ with
an Hermitian matrix H . Therefore, to use it, the eigenvalue problem K φ = ω 2 M φ needs to be
brought to this form. The following theorem states that this can be made.

Theorem 4.3.2 — Eigenproblem from general symmetric to standard Hermitian form.

Let K and M be real-valued symmetric with M > 0 and K ≥ 0. Let further σ be a scalar shift
such that K σ , K + σ M has no eigenvalues at zero and is therefore nonsingular and invertible.
Then the general symmetric eigenvalue problem K σ φ = ω 2 M φ can be brought to the standard
form H ψ = ω 2 ψ that preserves the eigenvalues and with H being symmetric.

Proof. Since a symmetric and real matrix M > 0 can be decomposed into M = M 1/2 M 1/2 , the
matrix M 1/2 > 0 is symmetric and real. It can therefore be inverted into a real and symmetric
M −1/2 . Using the transformation φ = M −1/2 ψ , the symmetric eigenproblem K φ = ω 2 M φ becomes
K M −1/2 ψ = ω 2 M M −1/2 ψ and therefore also M −1/2 K M −1/2 ψ = ω 2 M −1/2 M 1/2 M 1/2 M −1/2 ψ =
ω 2 ψ with preserved eigenvalues ω 2 holds. This is the form of the standard eigenvalue problem
H ψ = ω 2 ψ with H = M −1/2 K M −1/2 that is symmetric since H T = [M M −1/2 ]T K T [M
M −1/2 ]T =
−1/2 −1/2
M KM = H. 

The Lanczos transformation sequence. Let ψ 0 be a start vector, assumed to be a linear com-
bination of eigenmodes i.e. ψ 0 = ∑nj=1 α j φ j . In practice it is often generated as a vector with ran-
domized elements. Let that vector be normalized such that its norm is unity, i.e. let ψ 0 ← ψ 0 /||ψψ 0 ||.
4.3 Matrix iteration 81

ψ 0 , H ψ 0 , H 2 ψ 0 , . . .}. Let
Let that unitary vector constitute the first entry of the Krylov sequence {ψ
us form the next entry of the Krylov sequence as

ψ1 = Hψ0 (4.83)

and make it orthogonal to the previous vector ψ 0 by the linear operation

ψ 1 ← (II − ψ 0 ψ T0 )ψ
ψ 1 =(II − ψ 0 ψ T0 )H
Hψ 0 =
ψ T0 H ψ 0 ) = H ψ 0 − (ψ
H ψ 0 − ψ 0 (ψ ψ T0 H ψ 0 )ψ
ψ 0 , H ψ 0 − α00 ψ 0 (4.84)

that defines the coefficient α00 . Further, make it unitary with a normalizing operation

ψ 1 ← ψ 1 /γ1 with γ1 = ||ψ

ψ 1 || (4.85)

Let this be the next entry of an un-orthodox and orthonormal Krylov sequence {zz0 , z 1 , . . .}. This
process can be continued recursively for additional entries of the orthonormal Krylov sequence as

ψ p+1 = H ψ p (4.86)

which are made orthogonal to the previous p vectors by the deflating operation
p
ψ p+1 ← ψ p+1 − ∑ (zzTj H z p )ψ
ψj (4.87)
j=0

and further unit normalized by

ψ p+1 ← ψ p+1 /γ p+1 with γ p+1 = ||ψ

ψ p+1 || (4.88)

The vector sequence {ψ ψ 0 , ψ 1 , . . .} so generated are thus not identical to a true Krylov sequence
of vectors, but since the orthonormalization process just involves linear operations on the Krylov
sequence both the un-orthodox and true sequences span the same subspace.
Combining Eqs. (4.86)-(4.88) we notice that the sequence of vectors can be explicitly written

H ψ 0 = γ1 ψ 1 + α00 ψ 0 (4.89a)
H ψ 1 = γ2 ψ 2 + α01 ψ 0 + α11 ψ 1 (4.89b)
H ψ 2 = γ3 ψ 3 + α02 ψ 0 + α12 ψ 1 + α22 ψ 2 (4.89c)
..
.
p
H ψ p = γ p+1 ψ p+1 + ∑ α j p ψ i (4.89d)
j=0

with α j p = ψ Tj H ψ p . On matrix form these relations combine into

H Ψ , H [ψ ψ 0 ψ 1 . . . ψ p] =
 
α00 α01 α02 α03 ...
 γ1 α11
 α12 α13 ... 

 0
 γ2 α22 α23 ... 

[ψ
 0
ψ 0 ψ 1 . . . ψ p]  0 γ3 α33 ...  + [0 . . . 0 γ p+1 ψ p+1 ]

 .. .. .. .. .. 
 .
 . . . . 

 0 0 0 ... γ p−1 α p−1,p−1 α p−1,p 
0 0 0 ... 0 γp α pp
, ΨT + R (4.90)
82 Chapter 4. Decoupling of System States

Pre-multiplying this relation with Ψ T leads to

ΨT H Ψ = ΨT ΨT + ΨT R = T + ΨT R (4.91)

since the columns of Ψ are mutually orthonormal.

Using the structures of Ψ and R and the imposed mutual orthogonality between the vectors ψ j
we note that
0 . . . 0 ψ T0 ψ p+1
 T  
ψ0
ψ T  0 . . . 0 ψ T ψ 
T  1 1 p+1 
Ψ R =  ..  [0 . . . 0 γ p+1 ψ p+1 ] = γ p+1  . =0 (4.92)

 .   .. 
ψ Tp 0 . . . 0 ψ Tp ψ p+1

This lead to the relation Ψ T H Ψ = T . Because H is symmetric, and thus Ψ T H Ψ is symmetric,

it leads to the conclusion that T , as defined in Eq. (4.90), is symmetric. Since all elements of T
below the first sub-diagonal are all zero it leads to the conclusion that T is tri-diagonal6 and has the
following symmetrical form
 
α00 γ1 0 0 ... 0 0
 γ1 α11 γ2
 0 ... 0 0  
 0
 γ2 α22 γ3 ... 0 0  
T = 0
 0 γ3 α33 . . . 0 0   (4.93)
 .. .
. .
. .
. . . .
. .
. 
 .
 . . . . . .  
 0 0 0 . . . γ p−1 α p−1,p−1 γ p 
0 0 0 ... 0 γp α pp

The Lanczos sequence of Krylov vectors in Ψ leads to a reduction basis of the full n × n
eigenproblem H φ = ω 2 φ . Let Ψ ∈ Rn×p be used for transformation φ = Ψ θ and the eigenproblem
reads

H Ψ θ = ω̃ 2 Ψ θ (4.94)

Let further Ψ T be used to reduce that problem into

Ψ T H Ψ θ = ω̃ 2 Ψ T Ψ θ (4.95)

Since ΨT H Ψ = T and ΨT Ψ = I this leads into the p × p eigenvalue problem

T θ = ω̃ 2 θ (4.96)

which by virtue of the reduction gives an approximating subset of the eigensolutions of the full
eigenproblem. The problem has been reduced into a significatively smaller problem than the
original problem if p  n and can therefore be solved by robust but relatively slow eigensolvers
such as the iterative QZ algorithm7 without much speed penalty. Once solved for eigenvalues
ω 2j ≈ ω̃ 2j and eigenvectors θ j , the eigenvectors of the smaller problem can be projected on the full
size eigenvectors φ j again using the transformation φ j ≈ Ψ θ j .

L ANCZOS ’ EIGENSOLUTION FOR PARTIAL EIGENSOLUTION AT SHIFT

6 For the reason it is tri-diagonal it has been denoted T

7 The QZ algorithm is not described in this book. See e.g reference [20] for details.
 4.4 Matrix decomposition and transformation 83

1: procedure L ANCZOS(K K , M , σ , p, ψ 0 , Ω 2 , Φ ) . Gives p eigensolutions around σ using ψ 0

1/2
2: M = decomp[M M] . Decompose mass matrix
−1/2
3: H =M K + σ M ]M
[K M −1/2 . Create Hermitian matrix
4: for k = 0, . . . , p − 1 do
5: ψ p+1 = H ψ p
6: for j = 0, . . . , k do
7: α j p = ψ Tj H ψ p
8: ψ p+1 ← ψ p+1 − ∑ pj=0 α j p ψ j
9: end for
10: γ p+1 = ||ψ ψ p+1 ||
11: ψ p+1 ← ψ p+1 /γ p+1
12: end for
13: T = T (γ j , α jk ) . Set up T as in Eq. (4.93)
2
14: [Θ
Θ, Ω̃
Ω ] = qz[T T] . Solve smaller eigenvalue problem with QZ
15: Ψ = [ψ ψ 0 . . . ψ p] . Set up transformation matrix Ψ
2
16: return Ω 2 = Ω̃
Ω . This is approximation to eigenvalues
17: return Φ = Ψ Θ . This is approximate eigenmode matrix
18: end procedure

4.4 Matrix decomposition and transformation

Some well-known matrix decomposition and transformations are workhorses in numerical proce-
dures for finding the system’s eigensolutions. A few of them have been mentioned and are briefly
described below.

4.4.1 LDL0 decomposition

The LDL0 decomposition applies to real symmetric matrices A but also to complex Hermitian
matrices. A lower triangular matrix L and a diagonal matrix D are formed such that

A = LDLT (4.97)

where the structure of L is such that its diagonal elements are all one (1) and all its elements above
the diagonal are zero. Because of its structure it can easily be verified for the determinants that
|L LT | = 1 and thus
L| = |L

|A LD L T | = |L
A| = |L L||D LT | = |D
D||L D| (4.98)

The remarkable property that |A

A| = |D
D| is exploited in some eigensolution procedures. The
LDL0 decomposition can be seen as a by-product of the Gauss elimination step in the solution of
linear equations

Ax = LDLT x = b (4.99)

With a given LDL0 decomposition, a set of equations can be solved in two steps out of which the
Gauss elimination step is

Ly = b (4.100)

The first phase of the back-substitution step then proceeds as

Dz = y (4.101)
84 Chapter 4. Decoupling of System States

and in the second, and final, phase of the back-substitution the solution is found as

LT x = z (4.102)

In particular, the solution process can be used to invert the symmetric matrix A . Let b = I and
obviously x = A −1 since then A x = I . This process is illustrated in a step-by-step example below.
 Example 4.3 The inversion of the real symmetric 3×3 matrix A below will serve as an illustration
of the use of the LDL0 process. The matrix inverse and its determinant are also given in the outset
for verification purpose.
   
3 9 −6 815/6 −79/2 17/2
A= 9 26 −23 A −1 = −79/2 23/2 −5/2 |A
A| = −6
−6 −23 −11 17/2 −5/2 1/2

As a first step, the Gauss elimination process L y = I is made that obviously end up in the
inverse of L . The row operations to eliminate sub-diagonal elements of the first two columns of A
are illustrated in Tab. 4.1 which gives
   
3 9 −6 1 0 0
D L T = 0 −1 −5 and L −1 = −3 1 0 (4.103)
0 0 2 17 −5 1

The extraction of the diagonal elements of D from D L T is given by simple row operations and
gives
  
3 0 0 1 3 −2
D L T = 0 −1 0 0 1 5 
0 0 2 0 0 1

which specifies D and the upper triangular L T . The inverse of D is then

 
1/3 0 0
D −1 =  0 −1 0 
0 0 1/2

Table 4.1: The Gauss elimination step. Row operations shown for LH and RH sides of equation.
Operations LHS RHS Comment
   
3 9 −6 1 0 0 Starting point
 9 26 −23  0 1 0
−6 −23 −11 0 0 1
   
3 9 −6 1 0 0
R2 ← R2 − 3 · R1  0 −1 −5  −3 1 0
R3 ← R3 + 2 · R1 0 −5 −23 2 0 1 1st column sub-diagonals eliminated
   
3 9 −6 1 0 0
 0 −1 −5 −3 1 0
R3 ← R3 − 5 · R2 0 0 2 17 −5 1 2nd column sub-diagonals eliminated
 4.4 Matrix decomposition and transformation 85

and it can be verified that L −T D −1 L −1 is indeed the inverse of A since

     
1 −3 17 1/3 0 0 1 0 0 815/6 −79/2 17/2
L −T D −1 L −1 = 0 1 −5  0 −1 0  −3 1 0 = −79/2 23/2 −5/2
0 0 1 0 0 1/2 17 −5 1 17/2 −5/2 1/2


4.4.2 Householder transformation

 86 Chapter 4. Decoupling of System States

4.5 Problems
Problem 4.1 Normalized eigenmode of a 2-dof system
The first eigenvalue of the system in the figure is ω1 = 0.5412Ω. Use that information to calculate
a corresponding eigenmode that is normalized such that its algebraically largest element is +1.

. .......................................................................................

Problem 4.2 Anti-resonance frequency of a 2-dof system

Study the 2-degree-of-freedom system in problem 4.1 for the case the system is excited with a
single load f2 6= 0, i.e. when f1 = 0.

a) Compute the corresponding anti-resonance frequency, i.e. when u2 = 0

b) Give u1 as frequency function of f2 .

. .......................................................................................

Problem 4.3 Sturm eigenvalue bracketing followed by inverse iteration.

a) Use the Sturm sequence check to find out how many eigenfrequencies of the depicted system
that are lower than 2.7Ω and 3.2Ω and thus how many eigenfrequencies that are in the range
between the two. Hint. Use LDLT factorization of the dynamic stiffness matrix at the two
given frequencies.
b) Use inverse iteration technique to compute the lowest eigenfrequency of the system of. Express
the solution in Ω.
c) Use inverse iteration with shifting to compute the second eigenfrequency of the system. Use
the result from a) for an appropriate shift. Again, express the solution in Ω.

. ...................................................................................
4.5 Problems 87

Problem 4.4 Exact condensation. Wittrick-Williams eigenvalue search.

For the system of problem 4.3, compute the number of eigenfrequencies below 2.7Ω and 3.2Ω.
This time use the Wittrick-Williams algorithm but first do an exact condensation of u1 such that
only u2 and u3 remain.
. .......................................................................................

Problem 4.5 Inverse iteration for 3-dof system

Do two inverse iteration steps and give approximations to both the eigenvalue and the corresponding
eigenvector of the system depicted. Use as start vector u0 = {1; 1; 1} as first guess in both tasks a)
and b) below.
a) Calculate approximations to the first eigenvalue and its mode.
p
b) Iterate to the second eigenvalue and its mode. One has the approximation ω2 ≈ k/m, use that
approximation as an inverse iteration shift.

12/18/2009-3 and 12/01/2012-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.6 Inverse iteration for 2nd mode

For the 3-dof system, the first eigenvector and approximations to the first two eigenvalues are
known. Use inverse iteration with shift to calculate an approximation to the second eigenmode and
a better approximation to the corresponding second eigenvalue. Do two iterations and start with a
vector that is linearly independent to the first mode.

12/20/2008-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.7 Gerschgorin eigenvalue bounds of a 4-dof system

Use Gerschgorin disks to compute bounds for the four eigenvalues of the depicted system.

. .......................................................................................
88 Chapter 4. Decoupling of System States

Problem 4.8 Eigenvalue of rod system with inverse iteration

A step-wise area-varying rod undergoing pure longitudinal motion is depicted. Model the rod
system with two suitable discrete finite elements with a lumped mass formulation. Use inverse
iteration with shift to calculate its second eigenfrequency. Use the second eigenfrequency of a
corresponding uniform rod with cross-sectional area 2A as the shift (see Fig. 2.5 in Ch. 6.2.2) and
1;-1 as the start vector. Do two iterations and give the Rayleigh quotient approximation for the
eigenvalue after the last iteration.

01/14/2020-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.9 Trace inverse iteration outcome using Rayleigh quotient

Do two inverse iterations with shift for the depicted 3-dof system. Use the Rayleigh quotient
approximation for the iteration start vector ψ 0 = {0.5 ; 1.0 ; -0.8} as shift. After each iteration
step; normalize the vector so that its absolutely largest element becomes +1. Note the eigenvector
approximation sequence and calculate the eigenvalue approximation sequence by the use of the
Rayleigh quotient.

09/11/2007-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.10 Inverse iteration for 1st eigenvalue of beam system

For the 2-segment beam system in the figure do determine the first eigenfrequency for bending
vibration within the plane of the figure. Use inverse iteration with start vector {1 ; 1} for the
mid-point node and two finite beam elements. Stop after two iterations. Use lumped mass matrix
according to the HRZ mass lumping scheme. Data: ρA1 = 40.0kg/m, EI1 = 840kNm2 , L1 =2.00m,
ρA2 = 160.0kg/m, EI2 = 14.1MNm2 , L2 =2.00m.

05/27/2003-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Problems 89

Problem 4.11 Inverse iteration to largest eigenvalue

Do one inverse iteration step to obtain approximation to largest _largest eigenvalue of the depicted
3-dof system. Use ψ 0 = {−1; 1; −1} as start vector. Give approximations to both the eigenvalue
and the corresponding eigenvector.

04/10/2010-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.12 Eigenfrequencies in a range for a 4-dof system p

Howp many eigenfrequencies does the 4-dof system have in the range between 0.5 k/m and
4.0 k/m?

. .......................................................................................

Problem 4.13 Fundamental eigenfrequency of continuous rod with end mass

A rod has a small rigid mass attached to its right end. The rod is uniform with tensional stiffness
AE, mass per unit length ρA and length L. The weight of the end mass is ρAL/10. Determine
eigenfrequencies to two significant numbers of the system in its longitudinal vibration using the
Wittrick-Williams algorithm with a bisection search. Use Rayleigh’s theorem on constraints to
bracket the eigenvalues to obtain a start guesses for iterations.p
The sequence of eigenvalues of the
rod with one end fixed and
p the other free are ω k = (k − 1/2)π E/ρL2 k = 1, 2, . . . and with both
ends fixed are ωk = kπ E/ρL2 .

a) The rod has its left end fixed. Determine its 1st eigenfrequency.
b) The rod’s left end is free. Determine the eigenfrequency of the system’s 1st elastic mode.
c) The rod’s left is connected to a discrete 2-dof spring-mass system. Do an exact condensation of
it to the end displacement of the rod so that only one unknown remains (the end displacement
of the rod). Determine the system’s 2nd eigenfrequency.
p Hint! Use that the eigenfrequency of
the 2-dof system with fixed interface dof is 10E/ρL . 2

P4.10-12 X7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
90 Chapter 4. Decoupling of System States

Problem 4.14 Eigenvalue bounding

A 4-dof system is considered, see figure. Its data are: m = 1kg, k = 1kN/m. Its eigensolu-
tion is given here for reference.
p The given modes are mass orthonormalized. Eigenvalues:
[10.15, 12.33, 14.46, 17.58] × k/m. Coefficient matrices and mode matrix:
     
103 0 −2 −1 1 0 0 0 .9999 −.0106 .0005 .0097
 209 −4 −5   1 0 0  Φ = .0006 .0447 −.9982 .0397 
 
K =  M =
 309 −3   1 0  .0098 .0106 −.0392 −.9991
sym. 609 sym. 4 .0052 .4994 .0225 .0044
For the system you are asked to:

a) Factorise the mass matrix to M = U T U using Cholesky factorisation or via L D L T factorization

of M .
b) Use the factorisation matrix U to bring the eigenvalue problem to symmetric standard form
Ax = x.
c) Bound its eigenvalues by use of the Gerschgorin discs.
d) Bound the eigenvalues further by doing Sturm sequence checks at the centres of the discs using
results of L D L T factorization.
e) Do inverse iteration with shift to iterate on the second eigenvalue (and vector). Do 3 iterations.
Use shifts at centre of bound given by d). Use z0 = {1; 1; 1}T .
f) Do Rayleigh quotient iteration to iterate on second eigenvalue (and vector). Do 3 iterations.

P6.10 X4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Problems 91

Problem 4.15 Augmented damping

For the 4-dof system in the figure, assign 1% and 2% relative viscous damping to the two first
eigenmodes respectively. Assign stiffness-proportional viscous damping to the other two modes
such that no mode has less than 3% damping.

a) Establish the viscous damping matrix and make a diagonalization of it using the undamped
system eigenmodes.
b) Check the relative viscous damping of each mode. Does it correspond to the assigned values?

. .......................................................................................

Problem 4.16 Inverse iteration for planar beam problem

Calculate the first eigenfrequency for in-plane motion of the two-beam system shown using inverse
iteration. Use transverse displacement and midpoint rotation as the only two free dofs and start
guess {1; 1} for the mode and. Perform (at least) two iterations. Use Hinton-Rock-Zienkiewicz
mass lumping for the two beam elements.
Data: ρA1 = 40.0kg/m, ρA2 = 160.0kg/m, EI1 = 840kNm2 , EI2 = 14.1MNm2 and L1 = L2 =
2.0m.

5/27/2003-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.17 Sturm sequence check for 3-dof problem

For the 3-dof system shown use a Sturm sequence analysis to determine
p how many eigenfrequencies
that are in the frequency range from 0.8ω0 to 1.3ω0 with ω0 = k/m.

5/28/2002-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92 Chapter 4. Decoupling of System States

Problem 4.18 Rayleigh’s theorem on constraints

Use Rayleigh’s theorem on constraints to best bracket the two lowest eigenfrequencies of the
coupled two-rod system A+B. Four eigenvalue series are given for the individual rods A and B for
two different boundary condition sets.

9/13/2001-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 4.19 Inverse iteration with shift

Do one inverse
p iteration step to the second eigenvalue ω22 for which an approximation is given
as ω̃22 = k/m. Use {1; 1; −1} as start vector for the iteration step. Give the Rayleigh quotient
approximation to the eigenvalue and the corresponding eigenvector.

1/12/2010-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
 5. Time Domain Solution Procedures

Time domain solutions are presented for the structural dynamics equation and the state-space
equation. The first part of the chapter is devoted to exact analytical solutions to the initial value
problem, a so-called continuous-time problem. These solutions build the foundation for time-
stepping solution procedures that can be implemented for computational use for which solutions
are only obtained at specific discrete times. These procedures are presented in the subsequent part.

5.1 Continuous time solution for structural dynamics equation

5.1.1 Viscously damped single-degree-of-freedom systems
To obtain an analytical solution to the second-order structural dynamics equation M üu +V V u̇u + K u =
f (t) a corresponding single-degree-of freedom system may first serve as a prototype, see Fig. 5.1.
Its solution is uniquely given by the solution to its ordinary differential equations (ODE) together
with its initial conditions, i.e. its initial displacement state u0 and velocity state u̇0 at time t = 0.
The problem constitutes a mathematical initial-value problem stated as

Solve ODE: mü + vu̇ + ku = f (t) (5.1a)

Under condition: u(t = 0) = u0 , u̇(t = 0) = u̇0 (5.1b)

. .......................................................................................

Figure 5.1: Single-degree-of-freedom system with mass m, stiffness k and viscous damping v
loaded by force f (t) and responding with displacement u(t).
94 Chapter 5. Time Domain Solution Procedures

The homogeneous solution. The solution to the ODE (5.1a) is given by the superposition
of its homogeneous solution (for which f (t) ≡ 0 with u0 and u̇0 not both zero) and its particular
solution (for which f (t) 6= 0 and u0 = u̇0 = 0). Let us write the homogeneous solution as the linear
combination of the contribution from non-zero initial displacement u0 and non-zero initial velocity
u̇0 as

u(t) = u0 hd (t) + u̇0 hv (t) u̇(t) = u0 ḣd (t) + u̇0 ḣv (t) (5.2)

Topsimplify notation, now let ω0 be a frequency parameter related to stiffness and mass such that
ω0 = k/m. This parameter is also known as the system’s natural frequency. Also let a damping
parameter ζ and another√ frequency parameter
p ωd be parameters related to damping, stiffness and
mass such that ζ = v/2 km and ωd = ω0 |1 − ζ 2 |. The frequency parameter ωd is also known as
the system’s damped natural frequency. The mathematical textbook solution for the homogeneous
problem is split by three levels of damping. For a system with weak damping, ζ < 1, the system
is known as being undercritically damped and the functional form of the factors hd and hv of the
homogeneous solutions for the displacement are

ζ 1
hd = (cosωd t + p sinωd t) e−ζ ω0t hv = sinωd t e−ζ ω0t (5.3)
|1 − ζ 2 | ωd

and for velocity

ω0 ζ
ḣd = − p sinωd t e−ζ ω0t ḣv =(cosωd t − p sinωd t) e−ζ ω0t (5.4)
|1 − ζ 2 | |1 − ζ 2 |

This also gives the solution for cases with negative damping ζ < 0 for which an unstable
solution that is seen to grow without bounds as time passes.
For the rare, but still possible, situation the damping parameter of the system is exactly ζ = 1
the system is known as being critically damped and its factors are for the displacement

hd = (1 + ω0t) e−ω0t hv = t e−ω0t (5.5)

and for the velocity

ḣd = −ω02t e−ω0t ḣv = (1 − ω0t) e−ω0t (5.6)

Lastly, for a heavily damped system with ζ > 1, known as a overcritically damped system, the
factors for the displacement are

ζ 1
hd = (coshωd t + p sinhωd t) e−ζ ω0t hv = sinhωd t e−ζ ω0t (5.7)
ζ2 −1 ωd

and for the velocity

ω0 ζ
ḣd = − p sinhωd t e−ζ ω0t ḣv = (coshωd t + p sinhωd t) e−ζ ω0t (5.8)
2
ζ −1 2
ζ −1

The particular solution. Mathematical textbooks give expressions for the particular solutions
of the ODE (5.1a) for various specific forcing functions f (t) such as for it being sinusoidal in time,
exponential in time, constant in time, etc. However, there is no analytical textbook solution for a
general f (t). To bring the analytical path to a successful stop one is helped by the fact that any
5.1 Continuous time solution for structural dynamics equation 95

given forcing function can be seen as a sequence of impulses, and that there is a solution to the
impulse excitation problem as follows.
Let δ (t) be the infinite Dirac impulse function such that δ (t 6= 0) = 0, δ (0) 6= 0 and δτ ,
R 0+
0− δ (t)dt = 1. It can be observed that the impulse δτ carries the units of force × time. Let such
an impulse forcing function act on a system initially at rest, i.e. u(t = 0− ) = 0 and u̇(t = 0− ) = 0.
From time t = 0− to t = 0+ no elastic forces ku and damping forces vu̇ evolve and the initial value
problem is then

Solve ODE: mü = δ (t) (5.9)

− −
Under condition: u(t = 0 ) = 0 , u̇(t = 0 ) = 0

This differential integration is integrable from t = 0− to t = 0+ and has the velocity and displacement
solutions at t = 0+
Z 0+ Z 0+ Z 0+
1 1
u̇(0+ ) = δ (t)dt = δτ /m u(0+ ) = ( δ (t)dt)dt = 0 (5.10)
m t=0− m t=0− t=0−

For times larger than t = 0+ the motion of the one-degree-of-freedom system initially at rest is
thus governed by the homogeneous ODE in the problem

Solve ODE: mü + vu̇ + ku = 0 (5.11a)

Under condition: u(t = 0) = 0 , u̇(t = 0) = δτ /m (5.11b)

The solution for this homogeneous initial value problem is given for the various levels of
damping in Eqs. (5.2) - (5.7) above. With u0 = 0 and u̇0 = δτ /m we thus have the response to a
unitary impulse as u(t) = (δτ /m)hv (t). Let u(t)/δτ be called the (unit) impulse response function
h(t) of the system, illustrated in Fig. 5.2, and we thus have for the three levels of damping that
1
h(t) = sinωd t e−ζ ω0t for undercritical damping ζ < 1 (5.12a)
mωd
t
h(t) = e−ω0t for critical damping ζ = 1 (5.12b)
m
1
h(t) = sinhωd t e−ζ ω0t for overcritical damping ζ > 1 (5.12c)
mωd
The superposition principle for linear systems can now be used to the advantage. Viewing the
general loading f (t) as a sequence of impulses of strength f (τ)dτ, for any given time t = τ, each
such giving a contribution to the system response. The total displacement response at any given
time t can be obtained by a superimposing integration as
Z t
u(t) = h(t − τ) f (τ)dτ (5.13)
τ=0

This integral is known as the convolution integral or the Duhamel integral.

The total solution. The total solution is now the superposition of the contribution from the
loading and from non-zero initial conditions. The particular solution and the homogeneous solution
are superimposed into
Z t
u(t) = h(t − τ) f (τ)dτ + u0 hd (t) + u̇0 hv (t) (5.14)
τ=0

An illustration of the use, for an undamped single-degree-of-freedom system, is provided by

the following example.
96 Chapter 5. Time Domain Solution Procedures

Figure 5.2: Impulse responses for various classes of damping: negative damping ζ < 0, zero
damping ζ = 0, positive undercritical damping 0 < ζ < 1, critical damping ζ = 1 and overcritical
damping ζ > 1.
........................................................................................

 Example 5.1 Solve for transient response to harmonic loading. Let the undamped p 1dof
system be initially at rest, i.e. u0 = u̇0 = 0. The system’s natural frequency is ω0 = k/m. The
system is abruptly given the stimulus f (t) = f0 cosω0t starting at t = 0. The excitation frequency
is thus at the resonance frequency of the system and violent vibrations are expected.

For the undamped system ζ = 0 and thus e−ζ ω0t = 1 and its damped natural frequency ωd
equals its natural frequency, i.e. ωd = ω0 . The system’s displacement response is then given by the
Duhamel integral (5.14) with the proper impulse response function given by Eq. (5.12a) as
Z t
f0
u(t) = sinω0 (t − τ) cos(ω0 τ)dτ (5.15)
mω0 τ=0

A mathematical handbook solution, see e.g. [1], for this integral gives
f0t
u(t) = sinω0t (5.16)
2mω0
The solution is thus seen to be 90° out-of-phase with the excitation force f0 cosω0t and growing
linearly without bounds as time increases.


Response to a load step. A important load case, that will be utilized later for developing an
efficient numerical time-stepping procedure, is when a system initially at rest is subjected to a step
load of constant magnitude. In that case the forcing function is zero up to time t = 0 and then
constant in time, say f (t) ≡ f0 . The Duhamel equation Eq. (5.13) then reads
Z t
u(t) = f0 h(t − τ)dτ (5.17)
τ=0
5.1 Continuous time solution for structural dynamics equation 97

Figure 5.3: Step responses for various levels of damping: positive undercritical damping 0 < ζ < 1,
critical damping ζ = 1 and overcritical damping ζ > 1.
........................................................................................

The integral solution for this case for the various classes of damping are

f0 ζ
u(t) = 2
(1 − (cosωd t + p sinωd t)e−ζ ω0t ) ζ < 1 (5.18a)
mω0 1−ζ 2

f0
u(t) = (1 − (1 + ω0t)e−ω0t ) ζ = 1 (5.18b)
mω02
f0 ζ ζ √
−2ωd t −(ζ − ζ 2 −1)ω0 t
u(t) = (2 − (1 + + (1 − )e )e ) ζ > 1 (5.18c)
2mω02
p p
ζ2 −1 ζ2 −1

The velocity u̇(t) at an arbitrary time can be obtained by time differentiation of these displace-
ments u(t) to give
f0
u̇(t) = sinωd t e−ζ ω0t ζ <1 (5.19a)
mωd
f0
u̇(t) = te−ω0t ζ =1 (5.19b)
m
f0 √ 2 √ 2
u̇(t) = (e−(ζ − ζ −1)ω0t − e−(ζ + ζ −1)ω0t ) ζ >1 (5.19c)
2mωd

The displacement and velocity responses to a step load for some values of ζ are seen in Fig.
5.3.

Stability. At this point it is appropriate to briefly discuss the important concept of system
stability. It is discussed in the view of the following two definitions:
Definition 5.1.1 — Stability. An equilibrium state of a mechanical system is said to be stable
if for some initial state around that equilibrium state, the motion is such that the system stays
close to this equilibrium state. Such a system is called stable.

Definition 5.1.2 — Asymptotic stability. An equilibrium state of a mechanical system is said

to be asymptotically stable if for some initial state close to, but not at, that equilibrium state the
motion will converge to this equilibrium state. Such a system is called asymptotically stable.

From these definitions one thus notes that an asymptotically stable system thus belongs to a
subclass of the stable systems. All other systems thus belongs to the complement set of systems:
the unstable systems.
 98 Chapter 5. Time Domain Solution Procedures

The theory of unstable systems is the theory of instability. Since this is a very important topic
on its own the literature on the subject is vast, for mechanical systems see e.g. [51] and [7]. In
structural dynamics two of the most important phenomena are related to instability due to flutter
and instability in mass transport by conveyors. Instability analysis can lead to a a prediction of
the maximum speed an aircraft can fly without risking failure by excessive flutter vibration and
the maximum transport velocity that is achievable by a conveyor belt before dynamic stability is
lost. Both these have catastrophic consequences which must be avoided. For linear systems the
instability analysis focus on the system damping versus some other system parameter, such as
transport velocity, and may give ranges for these parameters for which stability is guaranteed. Such
analysis require appropriate modelling, such as aeroelastic modelling, that is not covered in this
book.

5.1.2 Viscously damped multi-degree-of-freedom systems

To be written! A numerical time-stepping procedure for the second-order differential equation will
be presented in Sect. 5.4.1.

5.2 Continuous time solution for the state-space system

By splitting the excitation into a sequence of impulse excitations, each giving contribution that
can be superimposed to the system response. Chopping the excitation into impulses is trivial,
but obtaining a high-order system’s response to these impulses is more involved. By modal
decomposition methods, see Sect. 4.1.1, the coupled system equations of high order, may be
reduced to a system of scalar first order decoupled equations. One route forward would then be to
first decouple the equations and then solve for these to finally superimpose the results to obtain
the system’s response. The unit impulse response function g(t) of a linear such first-order system
ẋ(t) + ax(t) = s(t) is well-known from calculus and is

g(t) = e−at , t > 0 g(t) = 0, t < 0 (5.20)

Let the excitation impulse over an infinitesimal time increment dτ be s(t)dτ, then the response
contribution to this impulse is

dx(t, τ) = g(t − τ)s(τ)dτ (5.21)

Assuming that the initial conditions are homogeneous, the integrated response is
Z ∞
x(t) = g(t − τ)s(τ)dτ (5.22)
0

But since g(t − τ) is zero for t − τ < 0, or equally for τ > t, the upper bound of the integral may
be replaced by t. The superposition integral, or convolution integral, formulation for the response
thus is
Z t
x(t) = g(t − τ)s(τ)dτ (5.23)
0

Although this prototype solution could be included in a possible route forward, it is rarely made
since the couples state-space solution has a general exponential form that can be more easily
incorporated in a numerical integration scheme. That will be described next.
 5.2 Continuous time solution for the state-space system 99

5.2.1 State transition matrix

Similarly to the scalar case treated above, we now seek a matrix integral solution to the state-space
initial-value problem

ẋx(t) = A x + B s , x (0) = x 0 (5.24)

which can be shown to be

Z t
x(t) = G(t)xx0 + G(t − τ)B
B s(τ)dτ (5.25)
0

Here G (t) is the state-transition matrix, which is G (t) = eAt . To proof this, we first consider
the general linear scalar initial-value problem. În i general linear system the coefficients are
time-varying and the initial value problem is

ẋ(t) = a(t)x(t) + b(t)s(t) with x(0) = x0 (5.26)

Its homogeneous solution, when s(t) ≡ 0, is known from calculus, see e.g. [22], and is
Rt
x(t) = x0 e 0 a(τ)dτ with x(0) = x0 (5.27)

Above, we have given the time domain solution (5.25), to the time-invariant multivariate problem.
For the multivariate problem, a closed-form convolution integral solution cannot be found for the
case the coefficient matrices A and B are time-varying. Since this is possible for the scalar case it
may come as a surprise and it may be interesting to know why.
As in the scalar case, it is natural to try a (matrix) integral solution on a similar form for the
time-variant homogeneous matrix initial-value problem ẋx(t) = A (t)xx(t), x (0) = x 0 , as
Rt
x (t) = x 0 e 0 A (τ)dτ (5.28)

where the exponential is defined by the series ẋx(t) = A (t)xx(t), x (0) = x 0 , as

Z t Z t Z t
Rt
0 A (τ)dτ
1
e =I+ A (τ)dτ + A (τ)dτ A (τ)dτ + . . . (5.29)
0 2 0 0

However, one notes that since

Z t Z t
d R0t A (τ)dτ 1
e A(t)
= A (t) + [A A (τ)dτ + A (τ)dτ A (t)] + . . . (5.30)
dt 2 0 0

is not generally equal to

Z t
1
A (t)(II + A (τ)dτ + . . .) (5.31)
2 0

unless the matrices A(t) and 0t A(τ)dτ commute, i.e. that A(t) 0t A(τ)dτ = 0t A(τ)dτ A(t). In
R R R

special cases, such as for time-invariant systems with constant matrices A or uncoupled systems
with diagonal A , the commutative property is seen to hold. We have thus found a distinct property
which distinguish the time-variant multivariate matrix differential equation from the time-invariant.
This distinction rules out closed form solutions along the suggested route for the time-invariant
case but for time-invariant systems the above suggest that the route is open.
We continue with the time-invariant case and utilize the commutative property when appropriate.
We use the transition matrix G (t) as defined by
Rt
0 A dτ
t2 2
G (t) = e = eAt = I + tA
A+ A +... (5.32)
2!
 100 Chapter 5. Time Domain Solution Procedures

which we require to obey the matrix differential equation

−1
A G−1 (t) ,
G (t) = −A
Ġ G−1 (0) = I (5.33)
The solution to the matrix differential equation is
t2 2
G −1 (t) = e−AAt G −1 (0) = e−AAt ≡ I − tA
A+ A +... (5.34)
2!
We have already observed that the constant matrix A commute with G −1 , such that A G −1 =
G −1 A and therefore Eq. (5.33) may also be written as
−1
G
Ġ G−1 A
= −G (5.35)
Now let us pre-multiply the state-space equation with G −1 to obtain
G −1 ẋx = G −1 A x + G −1 B s (5.36)
d G −1 −1
or by using that dt (G x )
G x + G −1 ẋx we have
= Ġ
d −1 −1
(G G x = G −1 A x + G −1 B s
G ẋx) − Ġ (5.37)
dt
But in view of Eq. (5.35) this reduces to
d −1
G ẋx) = G−1 B s
(G (5.38)
dt
which can be integrated readily. Taking into consideration that G −1 (0) = I we have
Z t Z t
G −1 (t)xx(t) = G −1 (0)xx0 + G −1 (τ)B
Bs (τ)dτ = x 0 + G −1 (τ)B
Bs (τ)dτ (5.39)
0 0
Pre-multiply this equation with the state-space transition matrix G , we obtain the desired
response
Z t Z t
x (t) = G (t)xx0 + G (t) G −1 (τ)B
Bs (τ)dτ = eAt x 0 + eA (t−τ) B s (τ)dτ (5.40)
0 0

where the state transition matrix property G = eAt has been utilized. We have thus showed that the
solution to the problem is indeed the solution given by Eq. (5.25). This solution will be used in
numerical time-stepping integration schemes that are presented in Sect. 5.4.2.

5.2.2 State observability, controllability and Grammians

State observability and state controllability plays important roles in experimental vibration engi-
neering. Loosely speaking, the state observability condition tells whether the states of the system
can be uniquely determined from recorded response r (t). Similarly, the state controllability tells
whether the model’s states may be excited independently from each other by the stimuli s (t). In
vibration testing, often targeting eigenmodes of a structure, the observability and controllability of
the modes given by the chosen sensor-actuator configuration are critical for success. In the sensor
placement, the observability criterion should therefore be fulfilled such that no critical states are
non-observable. Likewise, state controllability of critical states should be maintained by proper
actuator placement such that these states are excited and therefore contribute to the measured
response.

State observability. The concept of state observability is linked to the response and states of
the system. Given a state-space model and the set of responses r and stimuli s , the question of
observability is whether the model states x (t) are deducible from the set. This is more rigorously
formulated in the following definition:
5.2 Continuous time solution for the state-space system 101
Definition 5.2.1 — State observability. A linear system is said to be observable at time t0 if
the state x (t0 ) can be uniquely determined from the response r (t) when t ≥ t0 . If the system is
observable for all times, then the system is said to be completely observable.

A, B , C , D } can be investigated provided the model

The observability of a state-space model {A
and its inputs and outputs are known. When the output is given at times t ≥ t0 , also the time
derivatives of the output at t may be determined. Using that r = C x + D s, and therefore the
sequence

ṙr = C ẋx + D ṡs = C A x + C B s + D ṡs (5.41)

2
r̈r = C Aẋx + C Bṡs + Ds̈s = C A x + C A B s + C Bṡs + Ds̈s
..
.

leads to maximum N independent equations (but not more according to the Cayley-Hamilton
theorem, see below) for x . On matrix form these equations are

r̄r¯ = Oxx + G s̄s¯ (5.42)

with r̄r¯ being the vectorial concatenation of r , ṙr , r̈r , . . . and s̄s¯ being the vectorial concatenation of
s , ṡs, s̈s, . . .. The observability matrix of the system O is
 
C
 CA 
 2 

O =  CA  (5.43)

 .. 
 . 
C A N−1

and matrix G is irrelevant for observability. If the nr N × N observability matrix has a rank of less
than N, some linear combination of the N columns add to zero, and therefore there are (transformed)
states that do not contribute to r̄r¯ . It is thus necessary for observability of all system states that the
observability matrix is of full rank. Is the full rank condition also sufficient for observability? To
examine this, we start with multiplying Eq. (5.42) with the transpose of O to obtain

OT r̄r¯ = OT Oxx + OT G s̄s¯ (5.44)

If O is of full rank then OT O is non-singular and thus the state vector x can be determined as
its unique solution, which is

x = (OT O)−1 OT (r̄r¯ − G S ) (5.45)

Is this also the solution to Eq. (5.42)? If it is not, for the two different solutions x 1 and x 2 we
would have O(xx1 − x 2 ) = 0 which means that some linear combination of the columns of O is zero,
which contradicts the assumption that O is of full column rank. In conclusion we may thus state
that: A realization is uniquely observable if and only if the observability matrix O has full rank N.
It is important to notice that observability of states or lack of observability of states is not a given
system property but can be actively affected by the selection of sensor configuration. In the planning
of a vibration test, the observability issue can be addressed and various sensor configurations can
be compared with respect to observability. Since the selection of sensor configuration affect the
C matrix, different configurations can be evaluated for best observability in a pretest planning
phase using FE analyses if an FE model is available. In practice, because of numerical issues, the
observability matrix is often obtained from the time-discrete state-space models, see Sect. 5.4.5.
102 Chapter 5. Time Domain Solution Procedures

State controllability. The concept of controllability relate to the input and the states of a
system. The state-space first order differential equation can be used to examine the concept. The
controllability is defined by the following:
Definition 5.2.2 — State controllability. The system ẋx = A x + B s is said to be state control-
lable at time t = t0 if there exists a piece-wise continuous input s (t) that will drive the initial
state x (t0 ) to any final state x (t f ) within a finite time interval t f − t0 . If this is true for all initial
times and all initial states, the system is said to be completely state controllable.

If a time-invariant system is state controllable it is thus also completely state controllable. For
such systems a quantitative test of controllability can be derived. To this end let the initial time be
t0 = 0, for which the solution to the state-space equation is
Z tf
x (t f ) = e At f
x(0) + eA (t f −τ) B s (τ)dτ (5.46)
0

which, for non-singular eAt f , can be reduced to

Z tf
e−AAt f x(t f ) − x (0) , ∆xx(t f ) = e−AAτ B s (τ)dτ (5.47)
0

The Cayley-Hamilton theorem giving A k = ∑N−1 k

J=0 α jk A (for any k, i.e. also for k ≥ N, trivially
for k < N) together with the definition of the matrix exponential eAt = I + tA A + t 2 A 2 /2! + . . . give
N−1
e−AAτ = ∑ α j (τ)AA j (5.48)
j=0

with
∞
α j (τ) = ∑ (−1)k α jk τ k /k! (5.49)
k=0

Hence, Eqs. (5.47) through (5.49) give

N−1 Z tf N−1
j
∆xx(t f ) = ∑A B 0
α j (τ)ss(τ)dτ , ∑ A j B c j (t f ) (5.50)
j=0 j=0

with
Z tf
c j (t f ) = α j (τ)ss(τ)dτ (5.51)
0

Eq. (5.50) can then be written in matrix form as

B A B A 2 B . . . A n−1 B ][ccT0 c T1 . . . c TN−1 ]T

∆xx(t f ) = [B (5.52)

and represents a set of N equations and Nnu unknowns. The matrix equation has a solution for any
∆xx(t f ) provided that the N × Nns matrix

B A B A 2 B . . . A N−1 B ]
C = [B (5.53)

has N independent columns. The matrix C is known as the controllability matrix and thus the
system is completely state controllable if C has rank N.
5.2 Continuous time solution for the state-space system 103

Controllability and observability Grammians. Suppose a realization is given by ẋx = A x +

B s , x (t0 ) = x 0 and the quest is to bring the state to zero by a finite amount of energy input u in a
fixed time t f . Is that possible? To answer that question, the initial-value solution for a given s
Z tf
x (t f ) = eA (t f −t0 ) x (0) + eA (t f −τ) B s (τ)dτ (5.54)
t0

can be used. For a non-singular eA (t f −t0 ) , and to bring x (t f ) = 0 , we thus have

Z tf Z tf
−xx0 = e−AA(t f −t0 ) eA (t f −τ) B s (τ)dτ = eA (t0 −τ) B s (τ)dτ (5.55)
t0 t0

This is an integral equation for the sought stimuli s . Its solution can be shown to be
T
(t0 −τ)
BT eA
s (τ) = −B G−1 x0
c (t0 ,t f )x (5.56)

where
Z tf
T
Gc (t0 ,t f ) = eA (t0 −τ) B B T eA (t0 −τ)
dτ (5.57)
t0

is the controllability Grammian. To see that Eq. (5.57) really gives the solution, let it enter into the
integral of Eq. (5.55) which renders
Z tf
T
− eA (t0 −τ) B B T eA (t0 −τ)
G−1 x0 dτ =
c (t0 ,t f )x (5.58)
t0
Z tf
T
−[ eA (t0 −τ) B B T eA (t0 −τ)
dτ]G−1 x0 = −Gc (t0 ,t f )G−1
c (t0 ,t f )x x0 = −xx0
c (t0 ,t f )x
t0

since Gc (t0 ,t f ) is constant. The process requires that the controllability Grammian is non-singular
and therefore invertible, which is therefore the Grammian-related condition for controllability.
Similarly, for the realization ẋx = A x + B s , r = C x , the corresponding observability Grammian
Z tf
T
(t0 −τ) T
Go (t0 ,t f ) = eA C C eA(t0 −τ) dτ (5.59)
t0

must be non-singular [24] for the realization to be observable from the output during the time range
t0 ≤ t ≤ t f .
The Grammian singularity test R t1
is known to be a test for linear dependence of functions l j and lk
for which the Grammian G jk = t0 l j lk dτ should be non-zero for linear independence. In the case
here, the functions for independence test are the system’s state sequence eAt B to impulse excitation.
Let the start time for control stimuli vary, i.e. let t0 = t, and observe the following property of
the controllability matrix
d d t f A (t−τ) T A T (t−τ)
Z
Gc (t,t f ) = e BB e dτ =
dt dt t
Z tf
d A (t−τ) T A T (t−τ)
[e BB e AT − B B T
]dτ − e0 B B T e0 = A Gc (t,t f ) + Gc (t,t f )A (5.60)
t dt

For large time ranges t f − t and damped (asymptotically stable) systems the state impulses
eventually die out asymptotically and do not contribute more to the Grammian. Thus, for large
control times the Grammian derivative with respect to initial time variation is zero. The infinetely
long time Grammian G∞ c is thus governed by the Lyapunov equation

A G∞ ∞
AT − B B T = 0
c (t,t f ) + Gc (t,t f )A (5.61)
104 Chapter 5. Time Domain Solution Procedures

On the other hand, the corresponding long time observability Grammian G∞

o can be shown
(see [24]) to be the solution of another Lyapunov equation
A T G∞ ∞
CT C = 0
A −C
o (t,t f ) + Go (t,t f )A (5.62)
Numerical methods for solving the Lyapunov equations exist but are computationally expensive.
For large order systems (say N > 10.000) the solution may take hours to obtain even with a fast
computer (as of year 2019). Recall that the Grammians obtained by the Lyapunov equation are for
infinitely long control and observation times. The general controllability Grammian, as defined by
Eq. (5.57), is seen to be dependent on control time and is thus not unique. Also, observe that the
Grammians are not invariant to similarity transformations. Therefore, a similar realization as in
Eq. (3.12) normally gives other Grammians Gc and Go . This is used to the advantage for model
reduction using the balancing realization, one important realization form discussed below.

State-space realization on balanced form. Moore [31] has showed that it is possible to find a
similarity transformation x = T b z such that the controllability Grammian of Eq. 5.61 G∞
c and the
∞
observability Grammian Go of Eq. 5.62 simultaneously become diagonal and equal (balanced), i.e.
G∞ ∞
c = Gc . The corresponding state-space model

żz = T −1 −1
b AT bz + T b Bs (5.63)
r = C T bz + Ds
is called a balanced realization, with Grammians balanced over the control and observation range
[0, ∞].
The procedure to obtain the balancing transformation involves the solution of two Lyapunov
equations and goes as the following.

Balancing procedure:
1. Solve two Lyapunov equations for matrices P and Q
A T P A − P +C
CT C = 0 AQAT − Q + BBT = 0 (5.64)

2. Make a Cholesky decomposition of the symmetric and positive definite Q

RT R = Q (5.65)

3. Compute a singular value decomposition to obtain the singular value matrix Σ for singular
values in decreasing order in the relation
U T = RPRT
U ΣU (5.66)

4. Then the balancing transformation T b is given by

T b = R T U Σ−1/2 (5.67)

It should be noted that, although the controllability and observability Grammians are both
diagonalized and equal, the realization M̄ = {T T −1 −1
b A T b , T b B , C T b , D } related to Eq. (5.63) is
generally not taken to diagonal form by the transformation z = T b x .
For large-scale models the solution to the Lyapunov equations is very computationally de-
manding. However, the balanced form is very well suited for model reduction as states which
contribute little to the input/output relation can be singled out by the associated small elements of
the diagonalized Grammians and by that reduced from the model. State reduction schemes based
on this are presented in Sect. 7.4.1.
 5.2 Continuous time solution for the state-space system 105

5.2.3 Checking observability and controllability

The rank of the controllability and observability matrices may be determined by rank computation,
more often than not based on QR factorization, see e.g [20]. However, for high-order systems these
rectangular matrices may be of very high dimensionality and therefore require heavy computation.
Other, sometimes cheaper, tests have thus been developed. Such are the tests related to the theorems
by Popov and Belovitch with applications suggested by Hautus. These tests are called the PBH
tests, see e.g. [24], and are especially useful for theoretical analysis and also in numerical problems
whenever determination of matrix eigenvalues and eigenvectors is computationally feasible.

The PBH eigenvector test. The PBH eigenvector tests for controllability and obsevability are:

i) A state-space realization pair {A

A , B} will be controllable if and only if there, to the adjoint
T T
eigenvalue problem λ A = σ λ , exist no left eigenvector λ of A that is orthogonal to all the
columns of B . This should be checked.
ii) A state-space realization pair {A
A , C } will be observable if and only if there, to the eigenvalue
problem ρ A = σ ρ , exist no right eigenvector ρ of A that is orthogonal to all rows of C . This
should be checked.

For an engineer that knows vibration theory the controllability theorem should come as no big
surprise. It simply means that if the load distribution is orthogonal to any mode of the system, that
mode will not be driven by the loading and is therefore not controllable. As an alternative we have
the PBH rank test as:
PBH rank test. The PBH rank tests for controllability and obsevability are:

i) A state-space realization pair {A

A, B } will be controllable if and only if the matrix [[iωII − A ] B ]
has rank N for all frequencies ω.
ii) A state-space realization pair {A A]T C T ]
A, C } will be observable if and only if the matrix [[iωII −A
has rank N for all frequencies ω.

These conditions will clearly be met for all frequencies ω that do not match eigenvalues of A ,
because |iωII − A | =
6 0 for such ω. The point of the theorem is that the rank must be N even when
iω is indeed an eigenvalue of A . The case of multiple eigenvalues (of multiplicity m) deserves a
further treatment. It may be found during the PBH testing that

λ Tk B 6= 0 ∀k ∈ [1, m] (5.68)

and thus the test for controllability may, although here possibly falsely, be considered to be
fulfilled. However, for multiple eigenvalues we know that also an arbitrary linear combination
of the eigenvectors λ 0 = [λ
λ 1 λ 2 . . . λ m ]α
α , Λ α is also an left eigenvector. We thus require that
T T
B λ = B Λ α 6= 0 . We know that a solution to

BT Λ]ns ×m α m×1 = 0
[B (5.69)

can always be found if m > ns . In the case m = ns a non-trivial solution may also be found provided
B T Λ is singular. However, since B is not rank deficient, if designed properly, and Λ is never rank
deficient, the then quadratic B T Λ cannot be singular. The conclusion is thus that controllability is
lost if the number of inputs ns are fewer than the highest multiplicity of any of the eigenvalues of
the system. A similar analysis of the observability reveals that observability is lost if the highest
multiplicity of any eigenvalue is higher than the number of outputs nr .
 106 Chapter 5. Time Domain Solution Procedures

5.2.4 Continuous-time Markov parameters and the Hankel matrix

Markov parameters and the Hankel matrix are used in system identification theory. They set out
from the matrix integral solution of the initial value problem

ẋx = A x + B s with x (0) = x 0 (5.70)

for which the exact solution is given by the convolution

Z t
x (t) = G (t)xx0 + G (t − τ)B
Bs (τ)dτ (5.71)
0

with the state transition matrix G given by Eq. 5.32.

An important special case is for the impact stimuli s (t) = δ (t){1 1 . . . 1}T of a system at rest
with x 0 = 0 and δ (t) being the Dirac delta function. For such an ideal impact (can also be called a
multi-input simultaneous hit), the convolution integral in (5.71) evaluates to

A0 B + tA
x(t) = [A A B/1! + t 2 A2 B/2! + . . .][1 1 . . . 1]T (5.72)

and the system response r , r hit to such a hit is thus given by

C A 0 B + tC
Dδ (t) +C
r hit (t) = [D C A B /1! + t 2C A 2 B /2! + . . .]{1 1 . . . 1}T (5.73)

The terms of the system’s impact response series define what is often called the system’s
Markov parameters as

h0 = D , h k = C A k−1 B k = 1, 2, . . . (5.74)

from which it can be noted that the impact response is

∞
r hit (t) = [hh0 δ (t) + ∑ h k+1t k /k!]{1 1 . . . 1}T (5.75)
k=0

The infinite-size Hankel matrix H of the system is constructed from its Markov parameters as
follows
   
h1 h2 h3 h4 . . . CB C AB C A2B C A3B . . .
h 2 h 3 h 4 h 5 ...   C A B C A 2 B C A 3 B C A 4 B ... 
   
H = . = (5.76)
   
h3 h4 h5 h6 ..  C A 2 B C A 3 B C A 4 B C A 5 B ... 

.. .. .. .. . . .. .. .. ..
   
..
. . . . . . . . . .

Although the Hankel matrix is of infinite size, its rank is bounded to be rank(H ) ≤ N. This
follows from the Cayley-Hamilton theorem stating that A raised to any power can be expressed as a
linear combination with a finite number of terms as A k = ∑N−1 k
j=0 α jk A . This in turn leads to that
the number of linearly independent columns (and rows, and thus the rank) of the Hankel matrix
is limited. The number of positive singular values of the Hankel matrix is thus also bounded. Its
largest singular value σ1 (H ) is associated to the largest impulse response of the system.

5.3 Response bounds

Consider the response of a linear time-invariant system excited from the start of time. Using the
matrix of impulse response h (t) we have for the causal system’s response
Z t Z t
r (t) = h (t − τ)ss(τ)dτ = C G (t − τ)B
Bs (τ)dτ (5.77)
−∞ −∞
5.3 Response bounds 107

We want to bound the response vector norm ||rr (t)|| under certain restrictions put on the stimuli
vector norm ||ss(t)||. Before proceeding, we define what is meant by the norm of a vector-valued
function of time. Starting with the r-dimensional vector function r (t), we define its time-varying
q-norm to be
r
||rr (t)||q = ( ∑ |ri (t)|q )1/q (5.78)
i=1

Normally, only the Euclidean norm (q = 2) or the extreme norm (q = ∞) are of interest.
We then extend the definition of a vector norm to consider also the norm across time as well.
We define the (p, q)-norm of the vector-valued time-dependent function r(t) to be
Z ∞
||rr (t)|| p,q = ( ||rr (t)||qp dt)1/p (5.79)
−∞

Notice that this norm is independent of time. When p is ∞, the (p, q)-norm is the peak value
of the q-norm of the vector function y (t) over all times. Let us consider the two, probably most
interesting, cases when q is either 2 or ∞. First, when (p, q) is (∞, 2), the norm becomes the
maximum value of all times of the Euclidean norm of the vector r (t)

||rr (t)||∞,2 = sup ||rr (t)||2 (5.80)

−∞<t<∞

Second, when (p, q) is (∞, ∞) the norm is the maximum value over all times of the maximum
value of any component of the vector r (t)

||rr (t)||∞,∞ = sup ||rr (t)||∞ = sup max |ri (t)| (5.81)
−∞<t<∞ −∞<t<∞ i∈[1,r]

From now on, we restrict our attention to all stimuli with a given (2, 2)-norm equal to s? , i.e.
||ss(t)||2,2 = s? . This is the time domain root-mean-square value of the stimuli. The following
theorem then give norm bounds on the response r from a given level of stimuli.

hT (t)dt.
R∞
Theorem 5.3.1 — Response bound theorem. Let the matrix S be such that S = −∞ h (t)h
The (∞, 2)-norm of the vector of responses ||rr (t)||∞,2 to any stimuli vector s (t) from the class of
vectors having a (2, 2)-norm equal to s? will be bounded to
p
||rr (t)||∞,2 ≤ s? 2 max eig(SS )

and the (∞, ∞)-norm of the vector of responses to the same class of stimuli is bounded to
p
||rr (t)||∞,∞ ≤ s? 2 max diag(SS )

Proof. See Wilson [49] . 

This theorem let us bound the transient response of a structure without either knowing the actual
loading history or solving the initial value problem for a given load history. Whatever the actual
peak response to any given load, it will be less than or equal to the bounds given. In the following
section we will give an expression for a load history of norm s? that exactly matches the worst-case
bound for all loads within its class. It should be mentioned here, and always considered in practice,
that the scaling of the response vector elements is critical. Thus for mixed response vectors, e.g.
vectors holding both stresses and strains, the elements should be properly normalized before the
bounds are actually computed. A normalization will affect the state-space matrix C .
 108 Chapter 5. Time Domain Solution Procedures

We note that the matrix S is

Z ∞ Z ∞
S= h (t)hhT (t)dt = B B T G T (t)C
C G (t)B C T dt (5.82)
−∞ −∞
Z ∞
= C( B B T G T (t)dt)C
G (t)B C T ≡ C GcC T
−∞

with Gc , the controllability Grammian, given by Eq. (5.57).

5.3.1 Worst-case forcing function

As an analytical method, the preceding bound on the system response provides a means for analyzing
its worst-case transient response. It only requires that the class of loading can be characterized in
terms of their given (2,2)-norm. A system response bound analysis would reveal the most critical
stress components, the most critical displacement, etc., using only model data without simulation.
It is often necessary, however, to qualify a system by experimentally applying a loading that is
a worst-case simulation of the true operating environment. In this case, it might be necessary to
know at least one loading history that excites the system so that its response is at its limit. This load
history could then be applied knowing that all other load histories with the same (2,2)-norm would
excite the system less. One such transient loading, considering the bound of the (∞, 2)-norm of the
response is presented here. The response convolution expression (5.77) indicates that we look for a
loading on the form

s (t) = h (−t)T a (5.83)

in which a is a vector of yet to be determined constants. The response at an arbitrary time is then
Z ∞
r (t) = h (t − τ)hh(−τ)T a dτ (5.84)
−∞

and, specifically, the response at t = 0 is

Z Z ∞
r (0) = −∞∞ h (−τ)hh(−τ)T a dτ = [ h (t − τ)hh(−τ)T dτ)]aa ≡ S a (5.85)
∞ −∞

We are looking for a loading that produces the largest (∞, 2)-norm, i.e. the largest Euclidean
norm of the response at a certain time of all times. Let that time be t = 0. We thus want to maximize
||rr (0)||2 . On the other hand, the (2,2)-norm of the input is
Z ∞ Z ∞
||ss(t)||2,2 = [ s (t)T s (t)dt]1/2 = [ a T h (−t)hh(−t)T a dt]1/2 = [aaT S a ]1/2 ≡ s? (5.86)
−∞ −∞

The matrix S is positive semi-definite and symmetric, and so it has an eigen-decomposition

given by

S = E ΛE T (5.87)

which is orthonormal, i.e. E E T = I. We denote the ordered sequence of eigenvalues σk in Σ with

0 ≤ σ1 ≤ σ2 ≤ . . . ≤ σr and the corresponding eigenvectors in E with e k . The response quantity
||rr (0)||2 is maximized, under the condition that ||ss(t)||2,2 = s? , if we align the vector a with the
eigenvector of S that corresponds to its largest eigenvalue σ ? as thus

a = αr e r (5.88)
 5.4 Numerical discrete time solutions 109
√
where the scalar constant αr must be αr = s? / σ ? in order to satisfy Eq. (5.86). To show that
||rr (0)||2 is really maximized under these conditions we note that the quotient of response to load,
i.e.

||rr (0)||22 aT S T S a aT E Λ E T E Λ E a
= = (5.89)
||ss(t)||22,2 aT Sa aT E ΛE a

should be maximized. Since E is a square r × r matrix of full rank we may express any r × 1 vector
a as a linear combination of its columns e k . We then have
r
a= ∑ αk e k (5.90)
k=1

with αk being arbitrary constants. The matrix product E T a is then

 T
α1 e T1 e 1
   
e1 α1
e T  r α2 e T e 2  α2 
 2 2 
E T a =  .  ∑ αk e k =  .  =  .  , α (5.91)
  
.
.  .   .. 
.
k=1
e Tr αr e Tr e r αr

since E is orthonormal. The quotient in Eq. (5.89) is then

||rr (0)||22 α T Σ2α ∑rk=1 σk2 αk2

= = (5.92)
||ss(t)||22,2 α T Σα ∑rk=1 σk αk2

Since all eigenvalues are positive, this quotient is bounded by

||rr (0)||22 ∑rk=1 σk2 αk2

≤ ∀k ∈ [1, r] (5.93)
||ss(t)||22,2 σ ? αk2

with equality if αk 6= 0 and α j = 0 ∀ j 6= k. At equality, the quotient is at its maximum at σ ? .

Therefore, we should align a with the corresponding eigenvector. The worst-case loading is then
given by

s? s?
s (t) = √ ? h (−t)T e r = √ ? B T Φ (−t)T C T e r (5.94)
σ σ

5.4 Numerical discrete time solutions

5.4.1 The second order mass, stiffness and damping system
An exponential integrator. The analytical solutions to the homogeneous differential equation of
the mass, damping and stiffness system Eqs. (5.3) to (5.7) lend themselves to a basis for numerical
integration of the system’s differential equation. Together with the Duhamel solution for a step load,
presented in Eqs. (5.18a) to (5.19a), it can be used to form an efficient time-stepping integration
algorithm. For the time stepping with a constant time step ∆t , T , the algorithm produces solutions
at the times t = 0, T, 2T, . . . , kT, . . . , nT , where tk , kT is an arbitrary time and nT is the final time
for evaluation.
Let uk , u(kT ), uk+1 , u(kT + T ), u̇k , u̇(tk ) and u̇k+1 , u̇(tK + T ) be the displacement and
velocity at two adjacent arbitrary discrete times kT and kT + T . Also let the loading fk and fk+1
be the loads that act on the system at these two adjacent discrete times. If the load function is
sufficiently smooth in time, the approximation f¯k = ( fk + fk+1 )/2 is an approximation to the loading
110 Chapter 5. Time Domain Solution Procedures

Figure 5.4: Illustration of discretization process for 2 full cycles of a continuous forcing function
f (t), here f (t) = 2 − cos(4πt), from t = 0 to t = 1. Vertical stem lines indicate discrete times and
black bullets illustrate the mean values of the function for two neighbouring discrete times. Two
specific times t = tk and t = tk+1 and the associated mean value of the force f¯k+1 are highlighted
........................................................................................

function f (t) that is constant between times t = kT and t = kT + T and that is increasingly accurate
for decreasing time step T , see Fig. 5.4.
Starting at time t = 0 for which k = 0 the approximate solution for the displacement and velocity
at time t = T is given by the matrix relation
      
u1 ā11 (T ) ā12 (T ) u0 b̄1 (T ) ¯
= + f (5.95)
u̇1 ā21 (T ) ā22 (T ) u̇0 b̄2 (T ) 1

where the coefficients are given by Tab. 5.1.

As the end condition u1 , u̇1 after the first time step is the initial condition for the succeeding
second step, and so on, it could be understood that for any discrete time instant tk+1 we have that
      
uk+1 ā11 (T ) ā12 (T ) uk b̄1 (T ) ¯
= + f (5.96)
u̇k+1 ā21 (T ) ā22 (T ) u̇k b̄2 (T ) k+1

which is used to form a time-stepping algorithm that starts with computing the solution at time
t = T and progresses step-by-step k = 1, 2, 3, . . . up to final time t = nT .
It may be noted that the sole approximation is for the loading term for which the load is
approximated over the duration of the step to be constant in time. Such an approximation is often
called a zero-order-hold approximation. Since the coefficients are based on exponential functions as
given by Tab. 5.1, the time stepping algorithm is therefore known as an exponential zero-order-hold
algorithm. The numerical stability of the method is perfect, i.e. numerical stability is guaranteed for
arbitrary long time steps T as long as the system is stable. Its accuracy depends on the quality of the
approximation of the forcing function. For slowly varying forcing functions, long time steps can be
taken without significant loss of accuracy. For rapidly varying forcing functions the accuracy may
be brought to arbitrary accuracy (within numerical precision of the computer) by using sufficiently
small time steps.

The Newmark time integration method. The exponential integrator presented above has
good stability and is highly effective and therefore has much of the features desired for an algorithm.
However it is restricted to the class of systems for which the equations may be decoupled by the
5.4 Numerical discrete time solutions 111

Table 5.1: Zero-order-hold exponential integrator coefficients for various classes of damping. NB!
All coefficients b̄1 should be divided by mω02 and b̄2 should be divided by mω0 .
Damping Coefficient Expression
ζ <1 ā11 (cosωd T + √ ζ sinωd T ) e−ζ ω0 T
1−ζ 2
1 −ζ ω0 T
ā12 ωd sinωd T e
ā21 − √ 0 2 sinωd T e−ζ ω0 T
ω
1−ζ
ā22 (cosωd T − √ ζ 2 sinωd T ) e−ζ ω0 T
1−ζ
b̄1 1 − (cosωd T + √ ζ 2 sinωd T )e−ζ ω0 T
1−ζ
b̄2 √ 1 2 sinωd T e−ζ ω0 T
1−ζ

ζ =1 ā11 (1 + ω0 T ) e−ω0 T
ā12 T e−ω0 T
ā21 −ω02 T e−ω0 T
ā22 (1 − ω0 T ) e−ω0 T
b̄1 1 − (1 + ω0 T )e−ω0 T
b̄2 ω0 Te−ω0 T
ζ >1 ā11 (coshωd T + √ ζ2 sinhωd T ) e−ζ ω0 T
ζ −1
1 −ζ ω0 T
ā12 ωd sinhωd T e
ā21 − √ω20 sinhωd T e−ζ ω0 T
ζ −1
ā22 (coshωd T − √ ζ2 sinhωd T ) e−ζ ω0 T
ζ −1 √
b̄1 1− 1
(1 + √ ζ
+ (1 − √ ζ
)e−2ωd T )e−(ζ − ζ 2 −1)ω0 T
2 ζ 2 −1 2
√ 2 ζ −1 √ 2
b̄2 1
√ 2 (e −(ζ − ζ −1)ω 0 T − e−(ζ + ζ −1)ω0 T )
ζ −1

corresponding undamped system’s eigenmodes. It also needs an, at least partial, eigensolution to
be computed prior to the actual time integration. In the late 1950’s Newmark [32] developed a time
integration method that has been in extensive use since then. Its popularity is also due to the fact
that it can easily be adapted to nonlinear systems (not treated in this text). The development of the
method makes use of the Taylor series expansion. For an arbitrary function with s:th order time
derivatives given at instant tk , the Taylor expansion at time t = tk + T ≡ tk+1 is

2 s  ds f 
 
T T
fk+1 = fk + T f˙k + f¨k + . . . +  s  + Rs (5.97)
2 s! dt t=tk

where Rs is the residual of the series after the s:th term expressed as

Z tk+1 (s+1)
1 d f (τ)
Rs = (tk + T − τ)s dτ (5.98)
s! t=tk dt (s+1)

For the displacement and velocity at time tk+1 the Taylor series truncated to the level that the
112 Chapter 5. Time Domain Solution Procedures

displacement and velocity residuals involves the acceleration are

Z tk+1
u k+1 =uuk + T u̇uk + (tk+1 − τ)üu(τ)dτ (5.99)
t=tk
Z tk+1
u̇uk+1 =u̇uk + üu(τ)dτ (5.100)
t=tk

To this point no approximation has been made but it is also an endpoint for the exact analysis
since the time history of the acceleration üu(τ) needs to be known to solve the residual integrals. A
zero-order hold approximation of the acceleration gives a possible route forward. Significantly for
the Newmark method is that this zero-order-hold approximation is different for the displacement
and velocity residuals. Assuming, for the moment, that the acceleration at time tk+1 is known to
be üuk+1 it can be assumed that for the velocity approximation the acceleration over the duration
from tk to tk+1 can be represented by some intermediate value üu(τ) ≈ (1 − γ)üuk + γ üuk+1 with some
parameter γ. That γ is one Newmark parameter that can be tuned to adjust the behaviour of the
numerical integration. Another zero-order-hold approximation üu(τ) ≈ (1 − 2β )üuk + 2β üuk+1 is used
for the displacement with β being another tunable Newmark parameter. These displacement related
approximation into Eq. (5.99) and the velocity related approximation into Eq. (5.100) lead to

u k+1 =uuk + T u̇uk + (1/2 − β )T 2 üuk + β T 2 üuk+1 + r dis (5.101)

u̇uk+1 =u̇uk + (1 − γ)T üuk + γT üuk+1 + r vel (5.102)

where the residuals r dis and r vel can be shown to be [32]

... ....
r dis =(β − 1/6)T 3 u (τ̄) + O(T 4 u ) (5.103)
... ....
r vel =(γ − 1/2)T 2 u (τ̄¯ ) + O(T 3 u ) (5.104)

for some τ̄ and τ̄¯ in the range [tk ,tk+1 ].

Neglecting the residuals, the displacement and velocity relations Eqs. (5.101)-(5.102) into the
structural dynamics equation established at time tk+1 , i.e. M üuk+1 +V
V u̇uk+1 + K u k+1 = f k+1 , leads
to the relation for the acceleration at time step k + 1 being

V +β T 2 K ]üuk+1 = f k+1 −V
M +γTV
[M K {uuk +T u̇uk +(1/2−β )T 2 üuk } (5.105)
V {u̇uk +(1−γ)T üuk }−K

This relation, together with the residual-free relations for displacements and velocities in Eqs.
(5.101)-(5.102) constitutes the Newmark time integration algorithm. Together with given initial
conditions for displacements and velocities it can be used in a recursive time-marching fashion
to simulate the displacement, velocity and acceleration at discrete times with spacing T . The
initial acceleration state üu0 is needed at the start of the algorithm. This state, consistent with the
displacement and velocity state, can be obtained from the structural dynamics equation as

üu0 = M −1 { f 0 − K u 0 −V
V u̇u0 } (5.106)

The Newmark algorithm is summarized in the following:

N EWMARK ’ S T IME I NTEGRATION A LGORITHM

1: procedure N EWMARK(K K ,V
V , M , F , u 0 , u̇u0 , β , γ,U
U , U̇
U , Ü
U)
2: nt = column dimension of F . Number of time steps
3: f 0 = F :,1 . Force at t = 0 from first column of F
4: üu0 = M −1 { f 0 − K u 0 −VV u̇u0 } . Initial acceleration state
 5.4 Numerical discrete time solutions 113

5: U :,1 = u̇u0 ; Ü
U :,1 = u 0 ; U̇ U :,1 = üu0 . Data into output matrices
−1 2 −1
6: M = [M
M̄ M + γTV V + βT K] . Invert modified mass matrix once and for all
7: for k = 0, . . . , nt − 1 do
−1
8: M { f k+1 −V
üuk+1 = M̄ V {u̇uk + (1 − γ)T üuk } − K {uuk + T u̇uk + (1/2 − β )T 2 üuk }}
9: u k+1 = u k + T u̇uk + (1/2 − β )T 2 üuk + β T 2 üuk+1
10: u̇uk+1 = u̇uk + (1 − γ)T üuk + γT üuk+1
11: U :,k+2 = u k+1 ; U̇U :,k+2 = u̇uk+1 ; Ü
U :,k+2 = üuk+1 . More data into output matrices
12: end for
13: return U , U̇U , Ü
U . Return displacement, velocity and acceleration in data matrices
14: end procedure

The behaviour of the Newmark algorithm depends on the choice of the algorithm parameters β
and γ. The displacement and velocity displacements residuals (5.103) and (5.104) suggest that a
good choice would be β = 1/6 and γ = 1/2 to minimize the errors. However, a stability analysis
(see e.g. Ref. [17]) reveal that such a choice would render the algorithm only conditionally stable
and the time steps T needs to be sufficiently small. Another choice that lead to unconditionally
stable integration, but with larger error, is obtained for the choice β = 1/4 and γ = 1/2. For that
reason it has evolved as a most recommended Newmark parameter setting. Since the choice of
parameters is arbitrary, much research has been made to find the optimal setting. Since the setting
affect both accuracy and stability, the final balancing decision is left to the user. The accuracy of the
algorithm is often evaluated in the artificial damping the algorithm introduces and the periodicity
error. Both are evaluated in free decay, where the artificial damping is the difference between
the model’s damping and the apparent damping in simulation results. The periodicity error is the
difference between the free decay period time of the modal oscillations and the apparent ditto in
simulation output. The accuracy and stability properties for some popular Newmark parameter
choices are summarized in Table 5.2.

Table 5.2: Properties of some named members of the Newmark algorithm family.
Algorithm name γ β Stability Artificial Periodicity
limit damping error
Central difference 1/2 0 2 0 −ω 2 T 2 /24
Fox&Goodwin 1/2 1/12 2.45 0 O(ω 4 T 4 )
Linear acceleration 1/2 1/6 3.46 0 ω 2 T 2 /24
Average constant acceleration 1/2 1/4 ∞ 0 ω 2 T 2 /12
Artificially damped α + 1/2 (α + 1)2 /4 ∞ αωT /2 (1/12 + α 2 /4)ω 2 T 2

5.4.2 The first order state-space system

The time domain transient solutions of the scalar and matrix differential equations of the state-space
system have been given in convolution integral form. For practical purpose, the use of these is
limited, unless implemented in a computational scheme. The process of deriving a computational
scheme is the process of time-discretization, necessitated by present day’s computer architectures.
The convolution solution conveniently lends itself to discretization. Consider the solutions at two
consequent discrete times kT and kT + T (where fs = 1/T is often referred to as the sampling
frequency) given by
Z kT
xk = eA kT x0 + eA (kT −τ) B s(τ)dτ (5.107)
0
 114 Chapter 5. Time Domain Solution Procedures

and
Z kT +T
x k+1 =e A (kT +T )
x0 + eA (kT +T −τ) B s (τ)dτ = (5.108)
0
Z kT Z kT +T
eA T [eA kT x 0 + eA (kT −τ) B s (τ)dτ] + eA (kT +T −τ) B s (τ)dτ =
0 kT
Z kT +T
eA T x k + eA (kT +T −τ) B s (τ)dτ
kT

It is interesting to note that, for a sample interval in which no excitation take place, the state x k+1
can be computed without approximation, using the discrete time transition matrix Ā A = exp(A AT )
and the previous state x k independently of the sampling step T .
Using the zero-order-hold assumption, i.e. the excitation is assumed to be constant over the
entire sampling period, we have for the excitation term
Z kT +T Z kT +T
A (kT +T −τ)
e B s (τ)dτ ≈ [ eA (kT +T −τ) dτ]B
Bs k =
kT kT
−1 A(kT +T −τ) kT +T
A e
[−A ]kT B s k A−1 [e0 − e
= −A AT
Bs k = A −1 (eAT − I )B
]B Bs k (5.109)

We thus have the explicit time stepping algorithm

Ax k + B̄
x k+1 = Ā Bs k yk = C xk + Dsk (5.110)

with

A = eA T
Ā B = A −1 (Ā
B̄ A − I )B
B (5.111)

We note that we are required to provide the matrix ĀA = eA T which has the series expansion
of Eq. (5.32). In chapter 5.4.4 approximate methods for computing this are presented. It can be
shown that for stable systems with poles of A that have negative real parts that ||Â
A|| < 1 which is a
requirement for the time-stepping numerical integration scheme (5.110) to be stable as will be seen
below.

5.4.3 Stability of zero-order-hold time integration

Before addressing the stability aspects of state-space time-stepping algorithms, we consider the
diagonalization of the continuous-time realization

ẋx = A x + B s (5.112)

Provided that the matrix A is not deficient, see e.g. Refs. [20, 48], the system may be fully
decoupled by a suitable similarity transformation x = T z , and thus

żz = T −1 A T z + T −1 B s , diag(σc )zz + B̄

Bs (5.113)

The discrete-time counterpart is

z̆zk+1 = Ă
Az k + B̆
Bs k (5.114)

where
1
A = ediag(σ T ) = I + diag(σ T ) +
Ă diag2 (σ T ) + . . . (5.115)
2!
1
= diag(1 + σ T + (σ T )2 + . . .) = diag(eσ T ) , diag(σ̆ )
2!
 5.4 Numerical discrete time solutions 115

Figure 5.5: Projection of the stable continuous-time system’s eigenvalues, all on the marked left
half-plane. The unit circle embraces the eigenvalues of the discrete-time eigenvalues.
........................................................................................

The eigenvalues σ̆ of the discrete-time state-transition matrix ĂA are thus eσ T . For a stable
continuous-time system the real part of the eigenvalues σ are negative or zero. Therefore, the stable
system’s discrete time eigenvalues are bounded by the unit circle in the complex eigenvalue plane
(see Fig. 5.5). To see this we note that

|eλ T | = |eRe(σ T )+iIm(σ T ) | = |eRe(σ T ) eiIm(σ T ) | ≤ |eRe(σ T ) ||eiIm(σ T ) | = |eRe(σ T ) | ≤ 1 (5.116)

The argument of the complex eigenvalue may be any angle, since Im(λ T ) may be any real
number, large or small, positive or negative. We note in particular that for an undamped system,
i.e. for which Re(σ T ) = 0, the eigenvalues of the discrete-time system are all located on the unit
circle since then |eRe(σ T ) | ≡ 1.
One should note that although the transformation of eigenvalues from continuous-time to
discrete-time is unique, the reverse transformation is not so. One may also note that, for any integer
number n, the continuous-time eigenvalues σ = σ0 + 2inπ/T all correspond to the same discrete-
time eigenvalue eσ 0T . Thus no unique continuous-time eigenvalue correspond to the discrete-time
eigenvalue eσ0 T . However, for small time steps T such that Re(λ T ) < π for all eigenvalues the
reverse transformation is indeed unique.

5.4.4 Computation of the discrete-time transition matrix

One notes that for stable systems the algorithm (5.114) is unconditionally stable, since the magni-
tude of the discrete-time eigenvalues is less than or equal to one. For a homogenous state-space
stable system with non-homogenous initial condition, i.e. x 0 = 0 and s (t) ≡ 0 , the solution is
bounded. The discrete-time homogenous solution is also exact, provided that the state-transition
matrix is computed without truncation. The algorithm is thus free from algorithmic damping. For
calculation, an obvious approach is to follow the route given in the previous chapter, e.g. first
diagonalize the system by a similarity transformation and then compute diagonal elements of the
A˜ by complex exponential functions. This can be made with high accuracy,
state-transition matrix Ā
but will be costly. Another method is to use the series expansion of the state-transition matrix and
truncate the series at an appropriate number of terms. The series expansion may then be calculated
by a number of matrix multiplication. Using the computationally efficient Horner’s rule we have
for the exponential series truncated to order m

1 1 1
A˜ = eT A ≈ I + T A (II + T A (II + . . . +
Ā T A (II + T A ) . . .)) (5.117)
2 m−1 m
A question then naturally arises. How many terms are sufficient for an accurate representation
of the transition matrix? We will not answer that question here but just address the stability aspect
116 Chapter 5. Time Domain Solution Procedures

Figure 5.6: (a) Stability boundaries for increasing order n of the expansion of the state transition
A. (b) Magnitude of discrete time eigenvalue λ̆ versus imaginary part of normalized
matrix Ă
continuous time eigenvalue.

of the time-stepping algorithm with a truncated transition matrix. Conditions for algorithm stability
for an undamped system will be established. For a first-order truncated series we have

A˜ ≈ I + T A
Ā (5.118)

We now seek the stability bound for the time-stepping algorithm based upon this truncation, i.e.
A˜ . We assume, for simplicity but with some loss of generality1 , that we
we seek the eigenvalues of Ā
have a diagonal continuous-time realization matrix A = diag(σ ) and thus

A˜ ≈ I + T A = diag(1 + T σ ) , diag(σ̃(1) )
Ā (5.119)

The discrete-time eigenvalues are thus σ̆(1) = 1 + T σ . We know that the stability boundary is
at the unit circle, i.e. when |λ̆(1) | = 1. Separating the real and imaginary part of the discrete-time
eigenvalues we then have for stable eigenvalues A = diag(σ ) and thus

|1 + Re σ̃(1) + iIm σ̃(1) | ≤ 1 (5.120)

or equivalently

(1 + Re σ̃(1) )2 + Im2 σ̃(1) ≤ 1 (5.121)

This defines a stability disc in the complex eigenvalue plane, the disc having unitary radius and
center at [Re σ̃(1) , Im σ̃(1) ] = [−1, 0]. This is illustrated in Fig. 5.6. It is interesting to see that for an
undamped system, i.e. when Re σ̃(1) = 0, the algorithm is unstable for all T > 0 since |Im σ T | > 0.
This is usually not acceptable, particularly not for long simulation times. We continue the stability
investigation, now for an algorithm based upon a second order series truncation. We then have
1 1
A ≈ I + T A (II + T A ) = diag(1 + T σ + (T σ )2 ) , diag(λ̆(2) )
Ă (5.122)
2 2
1 Deficient matrices cannot be put on this form by similarity transformation
 5.4 Numerical discrete time solutions 117

with discrete-time eigenvalues σ̃(2) . Again, their magnitude should be less than or equal to unity. It
is straightforward to express this condition into a condition involving the real and imaginary part of
the continuous-time eigenvalues as

8Re σ̃(2) + 8Re2 σ̃(2) + 4Re3 σ̃(2) + 4Re σ̃(2) Im2 σ̃(2) + Re4 σ̃(2)
+ 2Re2 σ̃(2) Im2 σ̃(2) + Im4 σ̃(2) ≤ 0 (5.123)

The stability region associated with (5.123) is shown in Fig. 5.6. Again it can be seen that
the truncated series based algorithm is unstable for undamped systems. At this moment we may
wonder how much further the series has to be expanded before we receive an algorithm which is
also stable for undamped systems. Of course, we wish to use the shortest possible expansion that
give sufficient stability and accuracy. We know that an algorithm based on the full series expansion
is unconditionally stable and exact, such that no algorithmic damping is present. This knowledge
encourage us to proceed the expansion further, this time using a third order expansion. We then
have
1 1 1 1
A˜ ≈ I + T A(II + T A(II + T A)) = diag(1 + T σ + (T σ )2 + (T σ )3 ) , diag(σ̃(3) ) (5.124)
Ā
2 3 2 6
The stability analysis becomes increasingly more involved as the expansion proceeds, but the
basic principle remains. We do not carry out the analysis here but show the resulting stability
bounds in Fig. 5.6. It is seen from the figure that the stability region now contains part of the
imaginary axis. Thus the √ algorithm is also stable for undamped systems. The stability condition
is that |Im σ̃(3) = 0| ≤ 3. A similar analysis for a fourth order series expansion algorithm give
√
the stability condition |Im σ̃(4) = 0| ≤ 8. Interestingly, for a further expansion to fifth and sixth
order we again loose the stability in the undamped case (the interested reader with enough time can
prove this as outlined above). A fourth order expansion seems to be a good compromise between
speed, accuracy and stability. In that case we have

1 1 1
A˜ ≈ I + T A (II + T A (II + T A (II + T A )))
Ā (5.125)
2 3 4
The corresponding discrete-time zero-order-hold matrix is

1 1 1
B˜ = A −1 (eT A − I )B
B̄ A˜ − I )B
B = A −1 (Ā B ≈ T (II + T A (II + T A (II + T A )))B
B (5.126)
2 3 4
Since the stability border is not the entire left half-plane in the complex λ domain, we should take
care not to use too long time steps. Generally, for a third order expansion, it is suggested that T
should be selected as
√
T = 3κ/|Im max(σ )| (5.127)

where max(σ ) is the continuous-time eigenvalue of largest imaginary part and κ is the Couchy
number which is recommended to be in the range of [0.95,0.98]. However, if the load varies
significantly over the sample period, the zero-order hold assumption cease to be valid and shorter
time steps are required. It may also happen that one require a higher time resolution for other
purposes.

5.4.5 Discrete time observability and controllability

It is interesting to note that, for high order models, the time-continuous observability matrix Eq.
(5.43) involve large entries when N is large and ||A A|| > 1. Its numerical evaluation may then be
 118 Chapter 5. Time Domain Solution Procedures

troublesome since that may involve high powers of A . The observability matrix of the stable
discrete-time model
2 N−1 T
C T (C
OT = [C A)T (C
C Ā A )T . . . (C
C Ā C Ā
A ) ] (5.128)

with ||Ā
A|| ≤ 1 does not suffer from this problem and is thus better suited for numerical evaluation.
Similar to the observability matrix of continuous-time systems, the controllability matrix has
bad numerical properties when the system order gets large and ||A A|| > 1. The observability matrix
of the associated discrete-time model
2 n−1
B Â
C = [B̂ AB̂
B Â
A B̂
B . . . Â
A B]
B̂ (5.129)

of the stable system is then better suited. This matrix related to discrete-time entities is
sometimes called the reachability matrix.

5.4.6 Hankel matrix and Markov parameters

After studying the discrete-time realization it is natural to consider the discrete-time Markov
parameters. They will appear later in the context of system identification, where they play an
important role. The Markov parameters are defined as the impulse response matrices of the system.
Let us see how they relate to the state-space realization triple {Â
A, B̂
B,CC }. To this end we study the
initial-value impulse response problem

Ax k + B̂
x k+1 = Â Bs k yk = C xk x0 = 0 sk = δ k (5.130)

Here δ k is a vector of ones for k = 0 and a vector of zeros for all other k > 0, i.e. a load
sequence of initial unitary impulses. For k = 0, 1, 2 we have

Bδ 0
k = 0 : x 1 = B̂ Bδ 0
y 1 = C x 1 = C B̂ (5.131a)
1
Ax 1 = Â
k = 1 : x 2 = Â AB̂Bδ 0 A B̂
y 2 = C x 2 = C Â Bδ 0 (5.131b)
2 2
Ax 2 = Â
k = 2 : x 3 = Â A B̂
Bδ 0 A B̂
y 3 = C x 3 = C Â Bδ 0 (5.131c)
k−1 k−1
A B̂
or generally y k = C Â Bδ 0 . This defines the Markov parameter matrices h k as h k = C Â
A B̂ B for
k > 0. An important matrix connected to the Markov parameters is the Hankel matrix formed from
2 j + 1 Markov parameters
 
hk h k+1 . . . h k+ j
h k+1 h k+2 . . . h k+ j+1 
H = . (5.132)
 
. .. . . .. 
 . . . . 
hk+ j hk+ j+1 . . . hk+2 j

A significant feature of the Hankel matrix is that it is constant along the anti-diagonals. For the
single-input single-output case, in which hk are scalars, H is a standard Hankel matrix. In the case
of multi-input multi-output systems, in which the Markov parameter matrices form blocks constant
along the anti-diagonal, the matrix H is a so-called block Hankel matrix.
 5.5 Problems 119

5.5 Problems
Problem 5.1 Duhamel solution
For the 2-dof system shown first set up the governing equations and solve the associated eigenvalue
problem for eigenmodes and eigenvalues ω1 and ω2 . Also solve for the frequency ωa and mode
shape that is associated to anti-resonance for excitation at dof # 1. Then:

a) Solve for u(t) when f1 (t) = fˆ1 sinωt and f2 ≡ 0 for three full cycles of the fundamental mode,
i.e. 0 < t < 3 × 2π/ω1 . The system is at rest at t = 0. Use the mode displacement method and
the Duhamel integral solution. Consider four cases: I) ω = ω1 /2, II) ω = ω1 , III) ω = ω2 and
IV) ω = ωa . Plot u1 (t) and u2 (t).
b) Solve the same problem as in a) (case II only) but with a mathematical handbook solution.
Again, plot u1 (t) and u2 (t).
c) Solve for u (t) using the mode displacement method with the Duhamel integral for excitation
by support motion, i.e. when acceleration ü0 is prescribed. Let ü0 = asinω1t and study the
response for three full cycles. The system is at rest at t = 0.

p2.22a-f,i X2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 5.2 Mode displacement response at one instant of time in free vibration
For the undamped 5dof system, the right-most mass has been set in motion with an impulse at t = 0,
such that u5 (t = 0+ )=1m/s. The others are in their neutral position and are then at rest. Use the
mode displacement method to obtain the velocities of the masses at t=0.1s. Use the first two modes
given in the mode matrix Φ , at eigenfrequencies 9.04 and 26.31rad/s, in the synthesis.

10/20/2009-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
120 Chapter 5. Time Domain Solution Procedures

Problem 5.3 Modal solution for internal load

At one particular instant of time, the external loading of the 2-dof system is f1 = 0.745N and
f2 = 0. A modal solution at that time is given as η1 = 1.11mm and η1 = 0.16mm while the
modal accelerations are η̈1 = 0.768m/s2 and η̈1 = −0.432m/s2 . Data: m = 1kg, k = 1kN/m,
ω1 = 17.9rad/s ω2 = 68.4rad/s and Φ = [-0.5544 0.4389 ; -0.6207 -0.7840]. The mode matrix Φ
contains mass normalized modes.
Calculate the simultaneous force in spring A by use of:
a) The mode displacement method.
b) The mode acceleration method.

3/6/2011-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 5.4 Initial transient with zero-order-hold and Newmark

The undamped single-dof system below is released from its initial displacement state q(0) = q0
and velocity state q̇(0) = 0. The response is well-known and is a periodic oscillatory motion of
constant displacement amplitude q0 and period T = 2π/ω with ω 2 = k/m. Calculate the time
discrete solutions. Use step lengths ∆t = T and ∆t = T /2 and note differences. Compare the results
for a)-f). Repeat for:
a) Exact solution as reference.
b) Using a state-space model and zoh with exact evaluation of eA∆t .
c) Using zoh with approximation eA ∆t ≈ I + ∆tA
A.
d) A
Using zoh with approximation e ≈ I + ∆tA
∆t A + 2!1 (∆tA
A)2 .
A 1 A)2 + 3!1 (∆tA
A)3 .
e) Using zoh with approximation e ≈ I + ∆tA
∆t A + 2! (∆tA
f) Using Newmark’s method with ∆t = T /2, T /4 and T /8 and with parameters γ = 0.5 and
β = 0.25.

P7.5b-c X9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 5.5 Show non-singularity of state transition matrix

A is non-singular (has only non-zero eigenvalues)
Show that the discrete-time state transition matrix Ă
for most practical systems, i.e. systems with finite damping. Hint: relate on the one hand the
eigenvalues of the continuous-time realization to the eigenvalues of the discrete-time realization
and on the other hand damping to the real part of the continuous-time realization’s eigenvalues.
. .......................................................................................
5.5 Problems 121

Problem 5.6 Numerical integration stability

Show that for a stable state-space model {A A, B} (i.e. all its eigenvalues have negative real parts) the
exponential integrator produces a stable numerical solution (i.e. it does not grow without bounds in
free vibration). In other words, show that the time-stepping procedure x k+1 = eA T x k is stable for all
time steps T .
P7.5 X9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 5.7 Numerical integration stability

In a state-space algorithm (xxk+1 = A x k + B s k , y k = C x k ) for discrete time one has been given the
coefficient matrices below. Can one expect a numerically stable time integration?
     
+0.96 −0.30 0 0 −0.02 −1.440
+0.30 +0.96 0 0  +0.50 T +0.076
A=
  B̂ B= +0.49 Ĉ
 C = 
 0 0 +0.90 −0.06  +0.093
0 0 +0.06 +0.90 −0.10 +0.550
. .......................................................................................

Problem 5.8 3-dof system subjected to sine pulse load

Use numerical simulation to calculate the output vector r due to problem 3.1. Use a half-sine pulse
of unit maximum magnitude as the input s(t). Let the duration of the pulse be half of the period of
the system’s highest natural frequency. Let m1 = m2 = m3 = 1 kg, k1 = 1 kN/m and k2 = k3 = 10
N/m. Simulate for a four times the duration of the pulse of with increasing number of terms (2,
3, 4, 5 and 6) in the expansion eT A . Let T be one quarter of the duration of the pulse. Note the
algorithmic damping, is it positive, negative or varying for increasing number of terms?
. .......................................................................................

Problem 5.9 Response extremes

For the system in problem 3.1, let the single output be the support reaction R(t). Let the (2,2)-norm
of the input be ||s(t)||2,2 = σ = 1Ns1/2 . Let m1 = m2 = m3 = 1kg, k1 = 1kN/m, k2 = k3 = 10N/m,
c1 = 1Ns/m, c2 = 0.2Ns/m and c3 = 0.1Ns/m.
a) Calculate the (∞, ∞)-norm of the output.
b) Create the worst-case loading function and simulate the system’s response. Does the maximum
response obtained by simulation match the calculation?
c) Make simulations of the responses to half-sine pulses of various duration (with (2,2)-norm
1Ns1/2 ) and plot the dynamic amplification factor (as peak over reference load 1N) versus a
duration time that is normalized with respect to the period of the most high-frequency natural
mode.
. .......................................................................................

Problem 5.10 Mode acceleration solution in forced vibration

Calculate the displacements u1 and u2 for the 2-dof system at the specific time instant when t = 3ms.
The loading is saw-tand pulses according to the figure and the modal accelerations are known to
be η̈1 = 0.019m/s2 and η̈2 = −1.201m/s2 at t = 3ms. The resonance frequencies of the system
are f1 = 224Hz and f1 = 1600Hz. Other data: k = 1MN/m, m = 5kg, f2 ≡ 0, modal matrix:
Φ = [0.010 1.000; 1.000 − 0.020].
3/13/2004-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 5.11 Mode displacement solution in free vibration

An undamped 3-dof system is in free vibrational motion. At t = 0s, the modal displacements are
122 Chapter 5. Time Domain Solution Procedures

η1 = 0.10mm, η2 = −0.05mm and η3 = 0.01mm and the modal velocities are all zero. Calculate
the physical displacements u1 , u2 and u3 at t = 30s. The system’s eigenfrequencies are ω1 = 90rad/s,
ω2 = 120rad/s and ω3 = 200rad/s.
The modal matrix is: Φ = [0.54 − 0.15 0.23; 0.25 0.45 − 0.05; 0.11 0.25 0.27]
5/27/2003-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
 6. Frequency Domain Solutions

Frequency domain solutions play an important role in the understanding of linear vibration. Fre-
quency domain analysis procedures are based upon Fourier’s assumption that the vibration solution
can be factorized into one spatial and one temporal part. In some cases this factorization drastically
improves the possibility to achieve a solution. The benefit of frequency domain solution techniques,
for the linear systems to which they apply, is that they rely on algebraic solutions to the problem as
opposed to time domain methods that rely on more involved solution techniques for differential
equations. The duality between time and frequency domain and the transformation between the
two provided by direct and inverse Fourier transformation sometimes gives a simple analysis route
forward when other routes are harder to take.
Many technically important systems are excited by stationary harmonic loading with one or
more strong frequency components. Electrical motors at constant speed with an out-of-balance rotor
is a good example for a mono-frequency excitation mechanism. Combustion engines with moving
pistons, cranks and crank-shafts are other good examples that provide stationary multi-frequency
excitation when at stationary operating conditions.
Although more abstract, the frequency domain concepts often provide additional insight into a
vibration problem at hand. Phenomena like system resonance and anti-resonance and a system’s
frequency response function together with characteristic spectra for the loading can provide useful
information in design and problem-solving for vibrating systems. Such phenomena, and other
frequency related entities, will be described in the following.

6.1 Frequency response

The frequency response analysis is based on the assumption that, given that the system excitation is
stationary harmonic, the response is also stationary harmonic. A prerequisite is therefore stationarity
which, in theory, means that the excitation behaviour has been constant from beginning of time
and will stay so forever. In practice, it often means that that the excitation behaviour has been
approximately constant for a while and will remain so over the duration of our interest. That while
is long enough for an initial transient vibration (the homogeneous part of the differential equations)
to fade out to be negligible in comparison to the response caused by the harmonic loading.
 126 Chapter 6. Frequency Domain Solutions

In the following, the harmonic assumption will be applied to the second-order time differential
equation given by the structural dynamics equation and later to the first-order time differential
equation of the state-space description.

6.1.1 Transfer functions from the structural dynamics equation

Let a mathematical representation of a stationary harmonic forcing function at angular frequency ω
be

f (t) = f̂f eiωt (6.1)

where f̂f is a data vector with complex-valued and time-invariant force amplitudes as elements.
Following Fourier’s harmonic assumption then the response share the same harmonic behaviour
and thus the displacement response can be written

u (t) = ûueiωt (6.2)

with ûu being a data vector with complex-valued and time-invariant displacement amplitudes. The
velocity and acceleration consistent with Fourier’s assumption follows from time differentiation of
the displacement which leads to

u̇u(t) = iω ûueiωt , v̂veiωt üu(t) = −ω 2 ûueiωt , âaeiωt (6.3)

Here v̂v and âa are data vectors with complex-valued velocity and acceleration amplitude elements.
With the stationary harmonic force and its displacement, velocity and acceleration responses into
the structural dynamics equation M üu +V V u̇u + K u = f (t) leads to

[−ω 2 M + iωV
V + K ]ûueiωt , Z (ω)ûueiωt = f̂f eiωt (6.4)

which, after elimination of the common time function multiplier eiωt , gives the displacement
amplitudes as the result of an algebraic inversion of the dynamic stiffness matrix Z (ω) = K +
V − ω 2 M as
iωV

ûu = Z −1 (ω) f̂f , H d (ω) f̂f (6.5)

The frequency dependent matrix function H d (ω) is usually called the system’s frequency
response function. For this specific case with displacement response, the frequency response
function is also called the receptance of the system. From the definition of the velocity and
acceleration amplitudes in Eq. (6.3) one notes that

V + K ]−1 f̂f , H v (ω) f̂f

v̂v = iω[−ω 2 M + iωV (6.6)

and

V + K ]−1 f̂f , H a (ω) f̂f

âa = −ω 2 [−ω 2 M + iωV (6.7)

The frequency response functions for the system’s velocity and acceleration, H v and H a , are
known as the system’s mobility and accelerance matrices respectively.

6.1.2 The realness of loading and response

We note that the mathematical load description given by Eq. (6.1) is a complex-valued quantity and
as such it cannot be created in the real world. However, a related and real-valued forcing that can
be created is
∗
2 f (t) = f̂f eiωt + conj( f̂f eiωt ) = f̂f eiωt + f̂f e−iωt = 2Re{ f̂f eiωt } (6.8)
6.1 Frequency response 127

where a second term has been added that can be deduced from the first using the complex conjugate
of the force amplitude vector and negative frequency ω for the harmonic function.
To be physically sound, we must demand from the model that this real forcing function produces
a real-valued response. The displacement response can be obtained by superposition using the
receptance H d as
∗
2uu(t) = H d (ω) f̂f eiωt + H d (−ω) f̂f e−iωt = H d (ω) f̂f eiωt + H d (−ω) conj( f̂f eiωt ) (6.9)
For the superimposed u (t) to be real for any combination of ω and f̂f this demands that
H d (−ω) = H ∗d (ω) for which we then have
2uu(t) = H d (ω) f̂f eiωt + H ∗d (ω) conj( f̂f eiωt ) = 2Re{H
H d (ω) f̂f eiωt } = 2Re{ûueiωt } (6.10)
This result leads to the conclusion that there is no need to calculate a complex conjugate
solution for the complex conjugate loading provided that the system’s transfer function fulfills the
mentioned criterion. The real-valued loading Re{ f̂f eiωt } thus gives the real-valued displacement
response Re{ûueiωt } with ûu given by the algebraic relation in Eq. (6.5).
Since it holds for an arbitrary square matrix A that conj(A A−1 ) = [conj(AA)]−1 it also demands for
a structural dynamics model to be physically realizable that the dynamic stiffness matrix Z holds
the property Z (−ω) = Z ∗ (ω). Since the stiffness, viscous damping and mass matrices are all real-
valued it is easy to verify that this holds for the dynamic stiffness matrix Z (ω) = −ω 2 M + iωV V +K
and thus also for the receptance. It can be verified that this also holds for the system’s mobility H v
and accelerance H a given by Eqs. (6.6) and (6.7).

 Example 6.1 Harmonic excitation from eccentric rotor

Task. A rotor with a small mass eccentricity spins with a stationary spinning speed ω around
an axle fitted to a large mass M that is restricted to move in 2D translation only, see the figure. The
task is to calculate the horizontal and vertical displacements u1 and u2 for a given spinning speed
ω = Ω.
........................................................................................

Figure 6.1: A viscously damped 2-dof system (dofs: u1 and u2 ) subjected to an imbalance load from
a very small mass with eccentricity e spinning around A with angular frequency ω. Its coefficient
matrices given. Spinning mass creates imbalance forces f1 (t) = meω 2 cosωt = meω 2 Re{eiωt }
and f2 (t) = meω 2 sin ωt = meω 2 cos(ωt + π/2) = meω 2 Re{eiπ/2 eiωt }
. .......................................................................................

Complex-valued loading. The imbalance force is meω 2 and its horizontal and vertical load
components f1 and f2 are given in the figure text. The load vector is then
eiωt
     
f̂f 1 iωt 1
f (t) = Re{ e } = meω Re i(ωt+π/2) = meω Re{ iπ/2 eiωt }
2 2
f̂f 2 e e
 128 Chapter 6. Frequency Domain Solutions

Figure 6.2: Two full cycles of stationary harmonic loads f1 (t) and f2 (t) (dashed lines) and resulting
stationary harmonic displacements u1 (t) and u2 (t) (solid lines). Forces and displacements are
plotted with different scales.
........................................................................................

and the complex valued load vector amplitudes are thus the elements of the load vector

meω 2
   
2 1
f̂f = = meω
meω 2 eiπ/2 i

Solution. Solve for the complex-valued displacement vector as given by Eq. (6.5), i.e. ûu =
Z −1 f̂f . Start with setting up the dynamic stiffness matrix Z which is
     
2 2 30 −2 0 0 2 1 0
Z = K + iωV V − ω Z = MΩ + iMΩω − Mω
−2 50 0 1 0 1

when, as specified, ω = Ω the dynamic stiffness matrix and the corresponding receptance H d are
   
2 29 −2 −1 1 49 + i −2
Z = MΩ Hd = Z =
−2 49 + i (1417 + 29i)MΩ2 −2 29

Now, when ω = Ω the displacement amplitude vector can be established as

    
−1 me 49 + i −2 1 me 49 + 3i
u
û = Z f
f̂ = =
(1417 + 29i)M −2 29 i (1417 + 29i)M 2 + 29i

The time history of the displacements are then

 
me 1 49 + 3i iΩt
u (t) = Re{ûueiωt } = {with ω = Ω} = Re{ e }
M 1417 + 29i 2 + 29i

The stationary harmonic load history and the resulting displacement history for a time range
involving two full harmonic cycles are shown in Fig. 6.2. It can be observed that the system
damping results in that loads and displacements are not truly in-phase or completely anti-phase to
one another. 

6.1.3 Modal summation of frequency response functions

The mass and stiffness orthogonality of the natural modes of the system can be used to set up simple
relations between displacements, velocities and accelerations caused by the loading. Assuming
6.1 Frequency response 129

that the viscous damping is of Caughy type, and therefore the undamped modes can be used for
complete decoupling, the decoupled modal equations are (see Sect. 4.1.3)

diag(µk )η̈
η + diag(2ζk µk ωk )η̇ η = Φ T f (t)
η + diag(µk ωk2 )η (6.11)

and from its solution the displacements, velocities and accelerations can be obtained as

u = Φη u̇u = Φ η̇
η üu = Φ η̈
η (6.12)

Using again the Fourier assumption, a stationary harmonic loading f (t) = f̂f eiωt leads to har-
η eiωt with η̂
monic response η (t) = η̂ η being a data vector with complex-valued modal displacement
η eiωt and η̈
η (t) = iω η̂
amplitudes. This implies that η̇ η (t) = −ω 2 η̂
η eiωt . Inserted into Eq. (6.11) this
leads to

diag(−µk ω 2 )η̂
η eiωt + diag(2iζk µk ωk ω)η̂ η eiωt = Φ T f̂f eiωt
η eiωt + diag(µk ωk2 )η̂ (6.13)

which simplifies into

η = Φ T f̂f
diag(µk (ωk2 + 2iζk ωk ω − ω 2 ))η̂ (6.14)

The benefit from the modal decoupling is the simple inverse relation
1
η = diag−1 (µk (ωk2 + 2iζk ωk ω − ω 2 ))Φ
η̂ ΦT f̂f = diag( ΦT f̂f
)Φ (6.15)
µk (ωk2 + 2iζk ωk ω − ω 2)
Using the transformation from modal displacements to physical displacements as given by Eq.
(6.12), i.e. ûu = Φ η̂
η , we finally have
1
ûu = Φ diag( ΦT f̂f
)Φ (6.16)
µk (ωk2 + 2iζk ωk ω − ω 2)
and the receptance H d , i.e the transfer function that relates displacement to forces, is thus
1
H d (ω) = Φ diag( ΦT
)Φ (6.17)
µk (ωk2 + 2iζk ωk ω − ω 2)
By solving for the two roots, denoted σk and σ̄k , of the denominator polynomial equation
ωk2 + 2iζk ωk ω − ω 2 = 0 one has that ωk2 + 2iζk ωk ω − ω 2 = (iω − σk )(iω − σ̄k ). The roots, most
often called the system poles, are thus (σk , σ̄k ) = (ζk ± i(1 − ζk2 )1/2 )ωk which is seen to appear
in complex conjugate pairs or are both real. For under-critically damped modes with 0 < ζk ≤ 1
the poles are complex-valued scalars, for undamped modes with ζk = 0 they are purely imaginary
with (σk , σ̄k ) = ±iωk and for critically damped and over-critically damped modes with ζk ≥ 1 they
are purely real with (σk , σ̄k ) = (ζk ± (ζk2 − 1)1/2 )ωk . Using the pole representation of the transfer
function, the receptance is thus
1
H d = Φ diag( ΦT
)Φ (6.18)
µk (iω − σk )(iω − σ̄k )
H d and accelerence H a = −ω 2 H d are therefore
and the mobility H v = iωH
iω −ω 2
H v = Φ diag( ΦT
)Φ H a = Φ diag( ΦT (6.19)
)Φ
µk (iω − σk )(iω − σ̄k ) µk (iω − σk )(iω − σ̄k )
Using that the eigenvectors φ k are the columns of the modal matrix Φ the receptance can be
expressed on summation form as
n
φ φT
Hd = ∑ µk (iω − σkk )(iω
k
− σ̄k )
(6.20)
k=1
130 Chapter 6. Frequency Domain Solutions

The usefulness of the modal expression for the transfer functions is strongly related to the
modal expansion on series for for individual elements of the transfer function. Let H i jd be one
such element of H d . It is observed, see Eq. (6.20) that this element can be obtained from individual
components φik and φ jk of the eigenvectors as
n
φik φ jk /µk
H i jd = ∑ (iω − σk )(iω − σ̄k ) (6.21)
k=1

One obvious advantage of this modal summation form, for the situation that a modal solution is
at hand, is that it lend itself to evaluation without the need for matrix inversion as in Eq. (6.5). The
downside is that the eigensolution has to be computed first which requires effort.
The summation form in Eq. (6.21) is called the pole-residue representation of a transfer function
where the residues are the numerators φik φ jk /µk . Another popular representation is the pole-zero
representation which can be deduced from Eq. (6.21) as
n
φik φ jk /µk αi j ∏2n−2 (ω − zi jk )
H i jd = ∑ (iω − σk )(iω − σ̄k ) ∏n (iωk=1− λ )(iω − λ̄ )
= (6.22)
k=1 k=1 k k

where the 2n − 2 zeros z i jk are the so-called transmission zeros of the transfer function and the
constant αi j is called the transmission gain. For undamped systems the transmission zeros are
real-valued and the frequencies ω = zi jk for where they occur are called the system’s anti-resonance
frequencies with respect to the input-output pair (i, j). The anti-resonance frequencies are thus
unique to the pair (i, j) as opposed to the system poles (σk , σ̄k ) which are system-wide quantities.

Some properties of direct transfer functions. One special case of interest is the direct transfer
functions of the undamped system, i.e. the case when the stimulus and response concern the same
dof and therefore i = j. The system poles for an undamped system are σk = iωk and σ̄k = −iωk
and thus the direct receptance for that system is
n n 2
φ jk φ jk /µk φ jk
H j jd = ∑ (iω − iωk )(iω + iωk ) = ∑ µk (ω 2 − ω 2 ) (6.23)
k=1 k=1 k

which is seen to be a function that tends to ±∞ as ω → ωk from above or below. The gradient of
the direct receptance with respect to frequency is therefore
n 2ω
2φ jk
H j jd /dω =
dH ∑ µk (ω 2 − ω 2 )2 (6.24)
k=1 k

For all positive frequencies ω this gradient is strictly positive since all terms are positive as
a consequence of that the modal masses µk are all positive quantities and that the eigenvector
elements φ jk and eigenvalues ωk are real-valued.
The gradient of the direct accelerance with respect to frequency is

d n −ω 2 φ jk
2 n 2 ω 2ω 2
2φ jk k
H j ja /dω =
dH
dω ∑ µk (ω 2 − ω 2 ) = − ∑ µk (ω 2 − ω 2 )2 (6.25)
k=1 k k=1 k

which is seen to be a negative function of frequency.

Both the receptance and accelerance are thus frequency functions that varies with constant-sign
gradients between the eigenfrequencies at which the functions tend to ±∞. For positive frequencies,
the functions tend to ±∞ as many times as the system order n when (ωk2 − ω 2 ) → 0 from below or
above. The constant-sign gradients then implies that between all eigenfrequencies the direct transfer
 6.1 Frequency response 131

Figure 6.3: Transfer functions from loading f1 in linear scale. Upper row: mobilities, lower row:
receptances. Leftmost column: direct transfer function, middle and rightmost columns: cross
transfer functions.√ Embedded
p figure: 3-dof system for which undamped eigenfrequencies are:
ω1 , ω2 , ω3 = 0, 1, 3 × k/m
........................................................................................

functions must cross zero. The behaviour is illustrated for a 3-dof system with a rigid-body mode
in Fig. 6.3 where it is also shown that for cross-receptances, i.e. when i 6= j, this constant-sign
gradient behaviour does not hold.
For illustration purpose, the transfer functions are often represented by their absolute value
(magnitude) in a logarithmic scale as part of a so-called Bode plot. Such illustrations are shown
in Fig. 6.4 in which the constant-sign property of the direct transfer functions manifest itself as
a clear peak-valley structure with very deep valleys at the zero-crossings. Since there is one zero
crossing between each eigenvalue, there needs to be one distinct steep valley between all function
peaks that occur at the eigenfrequencies. For damped system, this distinct peak-valley behaviour of
the direct transfer functions is less pronounced and lesser so for increasing damping. For small
damping, however, this basic pattern is still clear. An example is shown in Fig. 6.5.
The behaviour of the direct transfer functions is of technical importance in the quality assessment
of test results. Test that are set up with accelerometer and force sensor pairs that are meant to
be co-located and collinear should ideally show the typical direct transfer function behaviour
mentioned above. If test results does not show this behaviour it is a strong indication that something
went wrong in in the test set-up, in the testing itself or in data processing of test results.

6.1.4 State-space models in frequency domain

The frequency domain counterpart for the steady-state dynamics of the stochastic time domain
model given by Eqs. (3.1a) and (3.1b) is based on the frequency descriptions of the known harmonic
stimuli with amplitude ŝs and frequency ω that is s (t) = Re{ŝseiωt } and the spurious process noise
w (t) = Re{ŵweiωt }. The harmonic assumption is that this lead to a response r (t) = Re{r̂r eiωt } that
is mixed with signal noise v (t) = Re{v̂veiωt } and that the state time variation also is harmonic
x (t) = Re{x̂xeiωt }. The assumption leads to
d
(x̂xeiωt ) = iωII x̂xeiωt = {A w}eiωt and r̂r eiωt = {C
Ax̂x + B ŝs + E ŵ C x̂x + D ŝs + v̂}eiωt (6.26)
dt
a relation illustrated as a block diagram in Fig. 6.6.
132 Chapter 6. Frequency Domain Solutions

Figure 6.4: Magnitude of transfer functions from loading f1 in logarithmic scale. Upper row:
receptances, lower row: accelerances. Leftmost column: direct transfer function, middle and
rightmost columns: cross transfer functions.
........................................................................................

Figure 6.5: Same as Fig. 6.4 but for damped system with small damping v. Dashed curves are for
corresponding undamped system.
........................................................................................
6.1 Frequency response 133

For the deterministic case, in which w and v are both zero, this leads to the relation between
input and output as

C (iωII − A )−1 B + D ]ŝseiωt , H (ω)ŝseiωt

r̂r eiωt = [C (6.27)

which, in turn, defines the system’s transfer function H , also known as its frequency response
function, as

H (ω) = C (iωII − A )−1 B + D (6.28)

One thus notes that the differential equation system in time domain has led to an algebraic
system of equation in the frequency domain for which the response amplitudes can be obtained as

r̂r (ω) = H (ω)ŝs(ω) (6.29)

In particular one notes, since the transfer function H (ω) is complex-valued, that the response
amplitudes in r̂r are generally complex-valued and thus contain both magnitude and phase informa-
tion. That is illustrated in the example below.
For the modal form realization (3.15) the frequency domain transfer function can be expressed
as

C (iωII − Σ)−1 B̄
H (ω) = C̄ B+D (6.30)

Since the state matrix Σ is diagonal, with the k:th eigenvalue σk on its k:th row and column
diagonal, one notes in particular that a transfer function element hi j can be expressed as

1
hi j (ω) = c̄ci: diag( )b̄b: j + di j (6.31)
iω − σk

or using the so-called pole-residue form as

N
c̄ik b̄k j
hi j (ω) = ∑ iω − σk + di j (6.32)
k=1

where the scalars c̄ik b̄k j are the N residues of the transfer function hi j . Since the eigenvectors of A
are generally complex-valued, the residue may be a complex-valued scalar.

. .......................................................................................

Figure 6.6: Block diagram representation of Eq. (6.26). Process noise ŵ w will be colored, i.e. its
spectrum will be distorted before it enters the output, by the system’s dynamics while output noise
v̂v is directly transmitted to the output. The process noise may represent unknown environmental
excitation of the system. The output noise may be introduced by the experimental system measuring
the measured response r̂r .
 134 Chapter 6. Frequency Domain Solutions

6.2 Exact dynamic condensation

The exact dynamic condensation plays its most important role in the analysis of continuous systems.
For such systems, as will be shown later, the infinite number of dofs that the continuous system
possesses is subdivided into a finite set of dofs of particular interest and an infinite set of dofs
which constitutes the system’s remaining dofs. The influence that these remaining dofs have on
the dynamics of the dofs of interest is obtained through a process known as dynamic condensation.
The dofs of interest are often called the active dofs while the rest are called the condensed dofs.
Although most important for continuous systems, the concept is easily demonstrated on discrete
(finite-dimensional) system that will be treated first.

6.2.1 Exact condensation of discrete system dofs

Let some dofs of a discretized system be collected into one set u a , the active set, that for some
reason are of particular interest. Let the remaining dofs u c be the remaining dofs that is of less
interest and for that reason can be reduced and left out (condensed) as variables of the resulting
equation. The partitioned version (see partitioning example 2.2) of the dynamic equation Eq. (6.4)
is then
      
2 M aa M ac V aa V ac K aa K ac ûu
[−ω + iω + ] a
M ca M cc V ca V cc K ca K cc ûuc
    
Z (ω) Z ac (ω) ûua f̂f a
, aa = (6.33)
Z ca (ω) Z cc (ω) ûuc f̂f c

from which it can be extracted from the second matrix row that

ûuc = Z −1
cc { f̂f c − Z ca û
ua } (6.34)

which can be used to eliminate ûuc from the first matrix row as then

Z aa − Z ac Z −1
[Z cc Z ca ]û Z aa ûua = f̂f a − Z ac Z −1
ua , Z̄ cc f̂f c (6.35)

For the case the loading f̂f c on the condensed dofs is zero, the resulting equation is thus

Z aa ûua = f̂f a
Z̄ (6.36)

The active dofs that are kept are normally those forced by external forces or other imposed
boundary condition. They are therefore sometimes called interface dofs since it is those dofs that
interfaces with the outside world as seen by the system. For the same reason the condensed dofs are
sometimes called internal dofs or internal variables. The interface dofs are often much fewer than
the number of internal dofs and therefore the resulting equation system Eq.(6.36) is significantly
smaller than the original equation system Eq. (6.33). To obtain the form (6.36), however, is is
required that the inverse of the then large dynamic stiffness matrix Z cc associated to the internal dofs
need to be inverted and no efficiency gain is achieved by the condensation operation for discrete
systems. For continuous systems the condensation is obtained as the exact solution to the system’s
differential equation which can lead to a very efficient condensation technique. Such systems are
the topic of the next section.
Although it cannot be used as an efficient computational procedure, the condensation form leads
to some useful insight. One such insight is that the condensed system (6.36) can possess solutions
ûua = 0 when |Z̄ Z aa | = 0 and the original system (6.33) does not posses a similar solution for the
full dof set {û a ; ûuc } unless ω → ∞. One notes that the full dynamic stiffness matrix in Eq. (6.33)
u
is populated by second order polynomials of ω that does not tend to infinity causing the trivial
solution {ûua ; ûuc } = 0 unless ω tends to infinity. Since the reduced form involves the inverse of Z cc ,
 6.2 Exact dynamic condensation 135

Figure 6.7: Beam element in vibrating motion due to end loading fˆ1 , fˆ2 , fˆ3 and fˆ4 .
........................................................................................

Z aa tend to infinity for all frequencies for which |Z

the reduced dynamic stiffness matrix Z̄ Z cc (ω)| → 0.
For undamped systems, these frequencies are the eigenfrequencies of the eigenvalue problem

K cc ûuc = ω 2 M cc ûuc (6.37)

which are the roots of the characteristic equation |K K cc − ω 2 M cc | = 0 which can be deduced from the
full set of equations (6.33) by fixing the dofs ûua . For damped systems the corresponding frequencies
are instead the complex-valued roots of the characteristic polynomial |K V cc − ω 2 M cc | = 0.
K cc + iωV
These frequencies are called the general anti-resonance frequencies of the condensed system. It is
at those specific frequencies that the interface does not move at all although there is a non-zero
loading acting on the interface, i.e. ûua = 0 with f̂f a 6= 0 in Eq. (6.36).

6.2.2 Continuous systems. Exact finite elements

Continuous systems by necessity need to be dynamically condensed to be useful in computational
structural dynamics, since they have infinitely many dofs. For some very few but practically
interesting cases this condensation can be made with exact means. Such is the cases for the uniform
beams and rods that will be described in the following.

Continuous uniform beam elements. In Euler-Bernoulli beam theory the transversal deflec-
tion w(x,t) in a local z direction obeys the differential equation

(EIy (x)w00 (x,t))00 + ρAẅ(x,t) = W (x,t) (6.38)

where (·)0 denotes differentiation with respect to the longitudinal length coordinate x and W is
the distributed transversal loading. The Euler-Bernoulli assumptions are that the shear center and
torsional center both are in the principle plane xz. In stationary harmonic vibration with a stationary
loading W (x,t) = Ŵ (x)eiωt it follows from Fourier’s assumption that

w(x,t) = ŵ(x)eiωt (6.39)

where ŵ(x) is the deflection mode shape.

In the special case the bending stiffness EIy and mass distribution ρA are constant and if the
distributed load is zero and thus only end forces/couples act to create vibration (see Fig. 6.7) it
holds that
ρA
ŵ0000 (x) − ŵ(x) = 0 (6.40)
EIy
136 Chapter 6. Frequency Domain Solutions

With the so-called wave number ν defined so that ν 4 = ρAω 2 /EIy this differential equation
has the analytical solution

ŵ(x) = a sinνx + b cosνx + c sinhνx + d coshνx (6.41)

for which the integration constants (a, b, c, d) can be determined from given boundary conditions
on the translatory and rotatory displacement of the beam element’s ends. The end’s displacements
and loading relate to the deflection mode shape so that

û1 = ŵ(0) û3 = ŵ(L) û2 = −ŵ0 (0) û4 = −ŵ0 (L) (6.42)
fˆ1 = −EI ŵ000 (0) fˆ3 = EI ŵ000 (L) fˆ2 = EI ŵ00 (0) fˆ4 = −EI ŵ00 (L)

The end loading for six elementary cases of boundary conditions with unitary end displace-
ments/rotations are given in Fig. 6.8 and thus gives the beam’s dynamic stiffness for various end
conditions. The end loading refers to frequency functions
p κ̂1 thru κ̂17 , here called the Koloušek
functions. With the frequency parameter β = ω/π 2 EI/ρAL4 , related to the wave number ν so
that β π 2 = ν 2 L2 , these Koloušek functions can be written
p p
κ̂1 = π β (sh − s)/(1 − ch · c) κ̂2 = π β (ch · s − sh · c)/(1 − ch · c)
κ̂3 = π 2 β (ch − c)/(1 − ch · c) κ̂4 = π 2 β sh · s/(1 − ch · c)
κ̂5 = π 3 β 3/2 (sh + s)/(1 − ch · c) κ̂6 = π 3 β 3/2 (ch · s + sh · c)/(1 − ch · c)
p
κ̂7 = 2π β sh · s/(ch · s − sh · c) κ̂8 = π 2 β (sh + s)/(ch · s − sh · c)
κ̂9 = π 2 β (ch · s + sh · c)/(ch · s − sh · c) κ̂10 = π 3 β 3/2 (ch + c)/(ch · s − sh · c)
κ̂11 = 2π 3 β 3/2 ch · c/(ch · s − sh · c) κ̂12 = π 3 β 3/2 (1 + ch · c)/(ch · s − sh · c)
κ̂13 = π 3 β 3/2 (ch · c)/(2sh · s) κ̂14 = −π 3 β 3/2 (ch · s − sh · c)/(2sh · s)
p
κ̂15 = −π β (ch · s − sh · c)/(1 + ch · c) κ̂16 = −π 2 β sh · s/(1 + ch · c)
κ̂17 = −π 3 β 3/2 (ch · s + sh · c) · s/(1 + ch · c) (6.43)
p p p p
where s, c, sh and ch abbreviates sin(π β ), cos(π β ), sinh(π β ) and cosh(π β ).
For very low frequencies when β tends to zero the numerators and denominators of the Koloušek
functions κ̂1 thru κ̂14 both tend to zero. Truncated Taylor series expansions for very small β for
these instead give the approximations

κ̂1 = 2 + .007142857π 4 β 2 κ̂2 = 4 − .009523810π 4 β 2 κ̂3 = 6 + .030952381π 4 β 2

κ̂4 = 6 − .052380952π 4 β 2 κ̂5 = 12 + .128571429π 4 β 2 κ̂6 = 12 − .371428571π 4 β 2
κ̂7 = 3 − .019047619π 4 β 2 κ̂8 = 3 + .039285714π 4 β 2 κ̂9 = 3 − .085714286π 4 β 2
κ̂10 = 3 + .139285714π 4 β 2 κ̂11 = 3 − .485714285π 4 β 2 κ̂12 = 3 − .235714286π 4 β 2
κ̂13 = 0 + .166666667π 4 β 2 κ̂14 = 0 − .333333333π 4 β 2 (6.44)
6.2 Exact dynamic condensation 137

Figure 6.8: Elementary cases for the vibrating uniform beam element with various prescribed
displacement/rotation boundary conditions. Forces and couples required to give the specified
prescribed motion are given. The end forces and couples required to produce the specified end
motion are expressed in Koloušek functions κ1 thru κ17 given explicitly in Eqs. (6.43) and (6.44).
138 Chapter 6. Frequency Domain Solutions

p
Figure 6.9: Eigenfrequencies in frequency parameter β = ω/π 2 EI/ρAL4 of the Euler-Bernoulli
beam subjected to various boundary conditions.
6.2 Exact dynamic condensation 139

Continuous uniform rod elements. The rod with longitudinally distributed load X(x,t), see
Fig. 6.10, obeys the differential equation

−(EA(x)u0 (x,t))0 = X(x,t) − ρA(x)ü(x,t) (6.45)

For the case of stationary harmonic vibration in a uniform rod, i.e. the factors EA and ρA are
both constants, we have that X(x,t) = X̂(x)eiωt and u(x,t) = û(x)eiωt which leads to

−EAû00 (x)eiωt = X̂(x)eiωt − ω 2 ρAû(x)eiωt (6.46)

and thereby the second order spatial differential equation with constant coefficients

EAû00 + ω 2 ρAû = X̂ (6.47)

For a rod with length L loaded only by concentrated forces at its ends it gives the homogeneous
differential equation

EAû00 + ω 2 ρAû = 0 (6.48)

p
which has the solution expressed in the frequency parameter β , ω ρL2 /E as

û(x) = asin(β x/L) + bcos(β x/L) (6.49)

for which the integration constants a and b need to be determined from the rod’s boundary condi-
tions at its ends. Three cases of particular interest can be identified which are considered next.

Case I. The fixed-free rod. The rod with one end fixed, see Fig. 6.11(I), and with a concen-
trated force fˆ2 loading the free end, the conditions at the ends are û(0) = 0 and EAû0 (L) = fˆ2 .
These give the integration constants a = fˆ2 L/(EAβ cosβ ) and b = 0 and thereby the displacement
solution
L sin(β x/L) L tanβ ˆ
û(x) = fˆ2 and in particular û2 , û(L) = fˆ2 f2 (6.50)
EA β cosβ EA β

The dynamic stiffness at the free end of the exactly condensed fixed-free rod is thus

fˆ2 EA β
Z22 (ω) , = (6.51)
û2 L tanβ

........................................................................................

Figure 6.10: (Left) A rod element 12 of length L loaded in tension/compression. (Right) magnifica-
tion of part of rod with indicated cross-sectional area A and longitudinal distributed loading acting
on centerline (but indicated with arrows offset center).
140 Chapter 6. Frequency Domain Solutions

Figure 6.11: Dynamic stiffness and eigenfrequencies of uniform rod in three elementary cases.
........................................................................................

On the other hand, the cross-sectional normal force at the fixed end is

cos(β x/L) EA β
fˆ1 = EAû0 (x = 0) = fˆ2 |x=0 = fˆ2 /cosβ = û2 (6.52)
cosβ L sinβ

That force acts in the negative sense on the rod at its fixed end and thus gives the rod’s cross-stiffness
as

fˆ1 EA β
Z12 (ω) , =− (6.53)
û2 L sinβ

In free vibration, the nontrivial solution û2 6= 0 although fˆ2 = 0 is thus given by Z(ω) = 0 that
leads to that β /tanβ = 0 which has the possible solutions
q
π
βk = (2k − 1)π/2 or similarly ωk = (2k − 1) E/ρL2 ∀k = 1, 2, . . . , ∞ (6.54)
2
with the corresponding eigenmodes

φk (x) = asin((2k − 1)πx/2L) (6.55)

Case II. The free-free rod. Loaded by a concentrated force at one end, see Fig. 6.11(II), the
rod’s boundary conditions are û0 (0) = 0 and EAû0 (L) = fˆ2 that gives the integration constants a = 0
and b = fˆ2 L/(EAβ sinβ ) and thus the displacement solution

L cos(β x/L) L 1 ˆ
û(x) = − fˆ2 and in particular û2 , û(L) = − f2 (6.56)
EA β sinβ EA β tanβ

which gives the dynamic stiffness of the rod with the internal dynamics condensed to end 2 as

fˆ2 EA
Z(ω) , =− β tanβ (6.57)
û2 L

that gives the free vibration solution when β tanβ = 0 which happens for
q
βk = kπ or similarly ωk = kπ E/ρL2 ∀k = 0, 1, 2, . . . , ∞ (6.58)

with the corresponding eigenmodes

φk (x) = bcos(kπx/L) (6.59)

Case III. The fixed-fixed rod. The boundary conditions for the fixed-fixed rod, see Fig.
6.11(III), are û(0) = 0 and û(L) = 0. These give the integration constant b = 0 and the relation
 6.3 Parseval’s theorem 141

asinβ = 0. For non-trivial solutions û(x) 6= 0 it is thus required that asin(β x/L) 6= 0 although
asinβ = 0. That leads to the non-trivial solutions for β as
q
βk = kπ or similarly ωk = kπ E/ρL2 ∀k = 1, 2, 3, . . . , ∞ (6.60)

for which the corresponding eigenmodes are

φk (x) = asin(kπx/L) (6.61)

The results for the three elementary cases are summarized in Fig. 6.11.

6.3 Parseval’s theorem

A particularly useful time-frequency transform is that given by Parseval’s theorem. The theorem
states that for real-valued functions r(t) and s(t) we have
Z +∞ Z +∞
r(t)s(t)dt = r̂( f )ŝ∗ ( f )d f (6.62)
t=−∞ f =−∞

+∞ −2πi f t
R
where r̂( f ) = −∞ e r(t)dt and similarly ŝ( f ) are the Fourier transforms of the real-valued
r(t) and s(t) for which it holds that r̂(− f ) = r̂∗ ( f ) and ŝ(− f ) = ŝ∗ ( f ). The special case in
which Rr(t) = s(t) is of particular interest since it evaluates the mean square (m.s.) of the function
+∞ 2
s2ms = −∞ s (t)dt. In signal processing, this is normally named as the power of the signal s(t) for
which Parseval’s theorem states that
Z +∞ Z +∞ Z 0 Z ∞
2 2 2
s (t)dt = |ŝ( f )| d f = |ŝ( f )| d f + |ŝ( f )|2 d f
t=−∞ f =−∞ f =−∞ f =0
Z ∞ Z ∞
2
= |ŝ(− f )| d f + |ŝ( f )|2 d f
f =0 f =0
Z ∞ Z ∞ Z ∞
∗ 2 2
= |ŝ ( f )| d f + |ŝ( f )| d f = 2 |ŝ( f )|2 d f (6.63)
f =0 f =0 f =0

The usefulness of the theorem should be understood in the context of frequency contribution to
the mean square level of the signal. The contribution from frequency ranges from zero up to, say,
f1 and f2 > f1 are
Z f1 Z f2
2
2 |ŝ( f )| d f and 2 |ŝ( f )|2 d f (6.64)
f =0 f =0

and therefore the contribution ∆s2ms from signals in a frequency range from f1 to f2 is
Z f2
∆s2ms = 2 |ŝ( f )|2 d f (6.65)
f = f1

Even more useful, in this era of computation, is the discrete-time discrete-frequency version of
Parseval’s theorem that states that
N N
∑ rk sk = ∑ r̂( fk )ŝ∗ ( fk ) (6.66)
k=1 k=1

where r̄( fk ) and s̄( fk ) are the discrete Fourier transforms of the discrete-time values rk and sk with
N N
∑ s2k = ∑ |ŝ( fk )|2 (6.67)
k=1 k=1
142 Chapter 6. Frequency Domain Solutions

as an important special case.

Eq. (6.67) thus makes it possible to calculate the contribution to the power of the signal from
each discrete frequency fk of the spectrum. As we know that a low-damped system’s response under
broad-band excitation is dominated by the contribution around the system resonances this makes
it possible to give good mean square signal approximations from relative few frequency domain
response calculations. That is provided that the system’s eigenfrequencies are known beforehand.
This is a feature that can be explored in model validation to check whether the model and test data
predict the same level of output power without very extensive computations.
 6.4 Problems 143

6.4 Problems
Problem 6.1 Harmonic response of 2-dof system
Two carriages are subjected to a harmonic load f (t) = f̂eiωt , with f̂ = { fˆ1 fˆ2 }T and fˆ2 = 0. The
driving frequency is half the first natural frequency which is at ω1 = 0.5412Ω.

a) Calculate the displacement, velocity and acceleration magnitudes.

b) The response at a frequency corresponding to a natural frequency, i.e. when ω = ωn , cannot be
calculated by use of the harmonic response assumption. Explain why.

. .......................................................................................

Problem 6.2 Mode displacement solution for 3-dof system

For the 3-dof system given in the figure and table below, compute the stationary harmonic dis-
placement, velocity and acceleration response of the left-most carriage. The driving frequency is
ω = 0.5(ω1 + ω2 ).

a) Calculate the exact solution.

b) Calculate solution by use of the modal displacement method with one mode, with two modes,
and with all three modes. Note the convergence rate (decrease of error).

(n) (n) (n)

ωn x1 x2 x3
Mode 1 0.468 1.0000 -0.2811 -0.9995
Mode 2 1.510 0.9998 -0.2804 1.0000
Mode 3 44.727 0.5614 1.0000 -0.0005

P2.9 X3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
144 Chapter 6. Frequency Domain Solutions

Problem 6.3 Mode acceleration solution for 3-dof system

Calculate the same quantities, as in problem 6.2 by use of the mode acceleration method.
a) Compare convergence rate with that obtained in problem 6.2. Calculate error (in %) compared
to exact solution for three modal solutions with increasing number of modes.
b) Compare convergence rate of the displacement solution as compared with that of the acceleration
solution. Use error metric as in a).
. .......................................................................................

Problem 6.4 Modal superposition for internal forces

For the system in problem 6.2, calculate the load in the left-most spring using displacements
calculated by:
a) The exact solution.
b) The solutions of the mode displacement method with first one and then two modes.
c) The solutions of the mode acceleration method with first one and then two modes.
. .......................................................................................

Problem 6.5 Dynamic condensation of two 2-dof systems

Calculate the harmonic response u2 of the two rigidly coupled subsystems in the figure.
a) Obtain solution by regular use of the displacement method of the resulting 3-dof system as a
reference.
b) Use dynamic condensation of the two identical subsystems’ degrees-of-freedom to their com-
mon degree-of freedom u2 . Calculate u2 of the condensed system.

P6.11 X2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 6.6 Stationary harmonic vibration in 2-dof system

A rigid body (mass M and moment of inertia J) is fixed to the end of a cantilever beam. The mass
of the beam is negligible. A harmonic force f1 (t) is applied to the body at a frequency that is the
eigenfrequency of the system when the moment of inertia is set to J = 0. Determine the endpoint
deflection amplitude u1 .

12/18/2009-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Problems 145

Problem 6.7 Stationary harmonic longitudinal vibration in continuous system

The two-segment system depicted has two rods connected via a rigid mass that is loaded by a
stationary harmonicpload. Determine the displacement amplitude of the mass when the load
frequency is Ω = π3 EA/mL2 with m = ρA being the mass per unit length.

a) Use two ordinary rod finite elements and a rigid mass in the model.
b) Use two exact continuous rod elements and a rigid mass in the model.

P4.15 X3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 6.8 System with tuned mass absorber

A damped system is driven by a harmonic force f2 = fˆ2 cosΩt with Ω = 300rad/s. It could be
represented by the depicted 2-dof system with undamped eigenfrequencies ω1 = 134.2rad/s and
ω2 = 471.1rad/s and thus the load frequency is between those. To remedy a vibration problem with
excessive vibration magnitude û2 at Ω, an undamped tuned mass absorber (TMA) is attached as
shown in Configuration 2 with mass mtma = 0.02m and stiffness ktma = mtma Ω2 . Data: m = 10kg,
k = 2 × 106 N/m and c = 305Ns/m.

a) Plot the magnitude of transfer function f2 → u2 for both configurations. Let the frequency vary
between 0 and 500 rad/s. Compare in particular the transfer function value at Ω.
b) Calculate the displacement vector u (magnitudes and phases) when the loading is f2 (t) =
Re{ fˆ2 eiΩt } = fˆ2 cosΩt with fˆ2 = 100N. Calculate the force in the TMA spring (magnitude and
phase). Does the force in the TMA spring fully counteract the external force?
c) In Configuration 1 the relative modal dampings are ε1 = 0.01 (at ω1 ) and ε2 = 0.35 × 10−3 (at
ω2 ). Make a Rayleigh damping (instead of the nonproportional damping given by dashpot c.
Repeat the calculations specified in b).
d) In Configuration 2 the relative modal damping values are ε2 = 0.01 , ε2 = 36.0 × 10−6 and
ε3 = 0.36 × 10−3 . Construct a viscous damping matrix, that replaces the non-proportional
viscous damping, given those modal damping values. Repeat the calculations specified in b).

X1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
 7. Model Reduction and Substructuring

In order to reduce computational time to solve large-scale structural dynamics problems, it has been
found effective to use approximate models of reduced dimension. An accurate large-scale model
then acts as a baseline from which the reduced and therefore approximate model is deduced. By that
computational efficiency is gained and computer predictions and simulations are made practically
feasible. Ideally, the model reduction scheme does only introduce negligible approximation errors
so that the reduced model can act as a good surrogate leading to results with sufficient confidence
for the application in mind. The development of model reduction methods has followed two major
routes. One, using insight and ingenuity and no prior mathematical rigour, with the second-order
structural dynamic equation as a base. The other, based on linear system theory and with solid
mathematical backup, with the first-order state-space equation as its base. The first route has lead
to well-known methods like the mode displacement method, the mode acceleration method, the
Guyan method and the Craig-Bampton method which will be described hereafter. The second route
has led to Grammian based methods that do model reduction with error control, so that the model’s
accuracy of the input-to-output relation of the model can be best preserved in the reduction process.
For a good treatise of classical model reduction techniques in structural dynamics see Ref. [10].
For state-space reduction methods see e.g. Ref. [3].
In procedures for model calibration, the use of model reduction techniques is of major im-
portance since the calibration involves many iterative updates of the finite element model. The
reduction of computational effort provided by the model reduction methods is then of utmost
importance to get reasonable computational times.

7.1 State transformation

To introduce the reduction methods we first gravitate for a while around the transformation of
the structural dynamics equation of a viscously damped reciprocal system with symmetric n × n
coefficient matrices

M üu +V
V u̇u + K u = f (t) (7.1)
 148 Chapter 7. Model Reduction and Substructuring

This equation can be transformed using a displacement and force transformation so that

u = T ūu and f̄f = T T f (7.2)

For the transformation from u to ūu to be unique, the quadratic transformation matrix T needs to
be non-singular and thus of full column rank n. The transformation matrix is time invariant so that
also u̇u = T ūu˙ and üu = T ūu¨ hold. By pre-multiplying Eq. (7.1) by the transpose of T and letting in
the transformations for displacements, velocities and acceleration we have

T T M T ūu¨ + T T V T ūu˙ + T T K T ūu = T T f (t) , f̄f (7.3)

M , T T M T , V̄
Introducing the transformed mass, damping and stiffness matrices as M̄ V , T TV T
K , T T K T respective we obtain the transformed equation of motion
and K̄

M ūu¨ + V̄
M̄ V ūu˙ + K̄
K ūu = f̄f (t) (7.4)

Since no reduction has yet been made not much has been gained by this transformation operation.
The system still has n degrees-of-freedom to solve for. One exception is the modal transformation
in which the modal matrix Φ is taken for the transformation matrix, i.e. T = Φ . As already noted
in Chapter 4.1.1 this brings a set of decoupled differential equations that normally can be solved
more easily than the coupled equations in (7.1).

7.2 Modal reduction methods

The rationale behind the use of modal reduction methods in structural dynamics is twofold. The
first takes its support in the duality between time-domain and frequency domain given by Parseval’s
theorem (see Sect. 6.3) and the physical properties of the structural loads. The physical processes
giving rise to structural loads are often relatively slow in the meaning the root-mean-square
contribution of high-frequency load components are missing or are negligible. The frequency is
then set in relation to the eigenfrequency spectrum of the structural system under consideration.
If the load root-mean-square contribution of significance is embraced by a number, say at the
maximum a couple of hundred, lowest eigenfrequencies of the structure the loading is said to be
slow. An approximate analysis, whether it is conducted in the frequency or time domain, then can
focus on the response contribution from the modes related to these eigenfrequencies only and still
give good accuracy.
While the first rationale is based on physics, the second rationale is based on computational
efficiency. Precise modeling of, sometimes very complex, mechanical systems often leads to very
large discretized models. It is not uncommon with models that involve millions of degrees-of-
freedom. Since the coefficient matrices of such models are often very sparse, it has led to the
development of very efficient numerical eigenproblem solvers for sparse problems, such as the
inverse iteration method and the Lanczos method that can be made to focus on relatively few
eigenfrequencies in a selected frequency range. They thus only solve a partial eigenproblem in
order to save computational effort. The following two sections cover the two most common modal
reduction methods that work under these premises.

7.2.1 Mode displacement method

As seen in Ch. 4.1.3 the modal superposition for the response is given by

u (t) = Φ η (t) (7.5)

where the solution for the modal displacements η j (t) of vector η are given by the Duhamel
convolution equation, see Sect. 5.1.1. Let us partition the modal matrix Φ into two parts Φ lo and
7.2 Modal reduction methods 149

Φ hi with low-frequency and high-frequency modes so that Φ = [Φ Φlo Φ hi ]. Let Φ lo hold m modes
and thus hi holds n − m modes. Eq. (7.5) can the be written (t) = [Φ
Φ u Φlo Φhi ]ηη (t) or on summation
form as
m n
u (t) = ∑ φ j η j (t) + ∑ φ j η j (t) (7.6)
j=1 j=m+1

The modal method approximation is then to truncate this series to involve only the low frequency
modes as
m
u (t) ≈ ∑ φ j η j (t) (7.7)
j=1

To get insight into the approximation errors introduced by this truncation let us assume that the
system under consideration is undamped and that the system loads can be written as

f (t) = a(t) f 0 (7.8)

where f 0 is a time invariant load distribution vector and a(t) is an arbitrary time dependent scale
factor. For an undamped system initially at rest, the transient response of a mode can be obtained
as (see Sect. 5.1.1)

φ Tj f 0 Z t
η j (t) = a(τ) sinω j (t − τ) dτ (7.9)
µ j ω j τ=0

and the displacement approximation is thus

φ j φ Tj f 0 Z t
m
u (t) ≈ ∑ a(τ) sinω j (t − τ) dτ (7.10)
j=1 µ j ω j τ=0

As a scalar metric for the displacement let us consider the mass-weighted time-varying quadratic
measure q = u T (t)M
M u (t), i.e.
T
m m
φ Ti f 0 φ j f 0 t
Z Z t
T
q(t) = ∑∑ φ i Mφ j a(τ) sinωi (t − τ) dτ a(τ) sinω j (t − τ) dτ
i=1 j=1 µi ωi µ j ω j τ=0 τ=0
m
1 T 1 t
Z
= ∑ (φφ j f 0 )2 ( a(τ) sinω j (t − τ) dτ)2 (7.11)
µ
j=1 j ω j τ=0

To derive Eq. (7.11), the mass-orthogonality property of the eigenmodes has been used, i.e.
φ Ti M φ j = 0 ∀i 6= j and φ Tj M φ j = µ j .
It is seen that each series term of the metric is a product of one spatial factor that depends on
the distribution of the loading f 0 in relation to the mode shape φ j and one integral temporal factor
that depends on the time dependence a(t) of the loading in relation to eigenfrequencies ω j . One
term of the series is thus small if either of these factors is small provided that the other factor is
not simultaneously big. Modes that give insignificant contribution to the series can thus be left
out without causing serious approximation errors. The convergence of the series for successively
smaller and smaller terms can be assessed by studying these factors individually.
The convergence of spatial type is obtained if modal terms are left out for which the load
distribution f 0 is (almost) orthogonal to the associated modes, i.e. for which the factor (φφ Tj p0 )2 /µ j
is small. Good spatial convergence is obtained if modal terms are successively brought in based on
decreasing magnitude of this factor.
 150 Chapter 7. Model Reduction and Substructuring

A convergence of spectral type is obtained providedR modal terms are successively included in
t
a sequence for which the magnitude of the factors ω1j τ=0 a(τ) sinω j (t − τ) dτ are decreasingly
small. It thus depends on the frequency content of the loading and the system’s eigenfrequencies
in combination. For example in the case of a harmonic load f (t) = f 0 cosωt we have (unless
ω = ±ω j ) for the magnitude of the factor that
Z t
1 cosω j t − cosωt 2
| cosωτ sinω j (t − τ) dτ| = | |≤ 2 (7.12)
ωj τ=0 ω 2 − ω 2j |ω − ω 2j |

Say that the loading is a linear combination of cosine loadings with maximum frequency ω = ω̄
and that the k:th eigenfrequency ωk is larger than that upper frequency of the loading. Then we
have for the related factor
Z t
1 2
| sinωk (t − τ) dτ| ≤ (7.13)
ωk τ=0 ωk2 − ω̄ 2

which is a factor that is small provided ωk  ω̄. By including all modes with eigenfrequencies
lower than ωk a good spectral convergence is obtained as ωk increases.
The full convergence for the displacements expressed by Eq. (7.7) requires a combination
of both spatial and spectral convergence. Since all eigenmodes φ j are not known until a full
eigensolution has been obtained a spatial convergence check is not practically feasible. That is
because the efficient solution procedures in use are based on that just a partial set of the full
eigensolution need to be computed. On the other hand, error estimates for the spectral convergence
can be found using bounds such as given for the harmonic loading in Eq. (7.13). These indicate that
accurate mode superposition results can be obtained by including all modes in the eigenfrequency
spectrum up to the first eigenfrequency ωk for which the criterion ωk  ω̄ is fulfilled. A practical
such criterion may be to require that ωk ≥ 10ω̄. Since efficient eigenproblem solvers can be set
up to compute all eigenvalues in a given range this leads to a practically useful model reduction
strategy.
Experience shows that this modal model reduction strategy, known as the mode displacement
method, has to be used with caution. Since the spatial convergence goes unchecked it sometimes
leads to unacceptable loss of accuracy. For this reason a variant of the mode displacement method
that involves the representation of the load distribution with the desirable feature that it is asymp-
totically correct for quasi-static loading, i.e. when the upper frequency of the loading is less than
system’s first elastic eigenfrequency. This method, the mode acceleration method will be described
next.

7.2.2 Mode acceleration method

The mode acceleration method is based on the observation that the exact response to the structural
dynamics problem is available in the form

K u (t) = f (t) − M üu(t) −V

V u̇u(t) (7.14)

This exact solution, said to be statically correct, however require that the acceleration üu and
velocity u̇u are fully known. which they are not until the full dynamic solution is known. However, by
using a modally based truncated approximation of these, an approximate solution can be obtained.
Noting that the acceleration and velocity can be obtained by modal superposition as üu = Φ η̈ η and
u̇u = Φ η̇
η we have for the truncated modal series on summation form
m m
K u (t) = f (t) − ∑ M φ j η̈ j (t) − ∑ V φ j η̇ j (t) (7.15)
j=1 j=1
7.2 Modal reduction methods 151

To illustrate the concept, without giving too much detail, the damping forces are left out in the
following. The displacement solution for neglected damping is thus
m
u (t) = K −1 f (t) − ∑ K −1 M φ j η̈ j (t) (7.16)
j=1

Using the spectral expansion of K −1 M φ j as given in Sect 4.1.4 one has

m
1
u (t) = K −1 f (t) − ∑ φ η̈ (t)
2 j j
(7.17)
j=1 ω j

where η j (t) for all modes j are given by the differential equation

µ j η̈ j (t) + µ j ω 2j η j (t) = φ Tj f (t) (7.18)

for which the analytical solution is given by Eq. (5.14) and a numeric time-integration procedure is
given by the exponential integrator Eq. (5.96).
To assess the benefit of the model acceleration method let us consider the solution for an
undamped system initially at rest for which the solution is (see Sect. 5.1)
Z t
1
η j (t) = φ Tj f (τ)sinω j (t − τ) dτ (7.19)
µ jω j τ=0

which into Eq. (7.18) leads to

φ Tj Z t
η̈ j (t) = { f (t) − ω j f (τ)sinω j (t − τ) dτ} (7.20)
µj τ=0

The modal truncated series of modal accelerations inserted into the displacement equation
(7.18) gives

−1
m φ j φ Tj Z t
K
u (t) =K f (t) − ∑ { f (t) − ω j f (τ)sinω j (t − τ) dτ} (7.21)
j=1 µ j ω 2j τ=0

−1
m φ j φ Tj φ j φ Tj Z t
m
K
=[K −∑ ] f (t) + ∑ f (τ)sinω j (t − τ) dτ (7.22)
j=1 µ j ω 2j j=1 µ j ω j τ=0

By utilizing the inverse of K on spectral form K −1 = ∑nj=1 φ j φ Tj /µ j ω 2j (see Sect.4.1.4) the end
result is obtained as
n φ j φ Tj m φ j φ Tj Z t
u (t) = [ ∑ 2
] f (t) + ∑ f (τ)sinω j (t − τ) dτ (7.23)
j=m+1 µ jω j j=1 µ j ω j τ=0

This result shows that the mode acceleration method complements the mode displacement
solution with the missing terms from the modal expansion of the static response K −1 f . However,
these terms do not have to be computed as the displacements are computed using the formula
given by Eq. (7.17). Experience shows that there are many application examples for which the
mode acceleration method provides significantly more accurate results than the mode displacement
method with little extra computational cost.
 152 Chapter 7. Model Reduction and Substructuring

7.3 Substructuring methods

Substructuring methods are particular forms of model reduction methods in which a particular set of
model degrees-of-freedom is retained while another (transformed) set is left absent in the remaining
reduced model. The set that is retained includes those degrees-of-freedom that are associated with
the substructure’s interface(s) to other substructures of the system. Substructuring methods are
very common in large structural dynamics projects in which teams, with subsystem responsibilities,
work together to form a computational model of the entire system of interest. For instance in
aerospace; separate groups can work on wings, fuselage, engine, nacelle, etc. To form a complete
aircraft model the sub-models of these can be reduced to their interfacing degrees-of-freedom and
the system can be coupled and analysed. A substructuring procedure is then normally split in three
steps, sometimes called substructuring phases, to make up a full analysis chain. In the first phase,
the reduction step, the groups make their substructure models and do the reduction. In the second
phase, the synthesis and analysis step, the substructures are assembled to form the complete system
and a full system analysis is made. In the third step, sometimes called the backsubstitution step, the
interface motion of the individual substructures are extracted from the analysis results and used for
separate substructure analyses by the teams.
The substructuring methods, and in particular the Craig-Bampton method that will be described
later, have been found effective and reliable and therefore been introduced as general reduction
methods in which computational algorithms, and not engineering teams, make the splitting of
the system into domains (substructures) of balanced complexity. This procedure is known as
Automated Component Mode Synthesis (ACMS).

7.3.1 The Guyan method

The archetype of substructure methods is the Guyan1 method that is built on exact static condensa-
tion of degrees-of-freedom. To provide a detailed insight, before fully turning to the Guyan method
for dynamic substructuring, the exact static reduction of dofs is described first.

Exact static condensation. Let the dofs u of the substructure under consideration be devisable
into on set of dofs u a that are associated to an interface to which the substructure is connected to
the rest of the world and another set of dofs u b that are unique to the substructure itself. Let us
make a symbolic partitioning (see partitioning example Ex. 2.2) of the dofs such that
 
ua
u= (7.24)
ub

The displacement of that substructure caused by a static load f can be obtained by the partitioned
equation
    
K aa K ab u a fa
Ku , = ,f (7.25)
K ba K bb ub fb

where f a and f b are the static loads that act on u a and u b respectively. The second matrix row gives
K −1
the relation K ba u a + K bb u b = f b and thus u b = −K −1
bb K ba u a + K bb f b . Introduce this result for u b
into the first matrix row of Eq. (7.25) to get an expression involving only the unknown u a as

K aa − K ab K −1
[K ua = f a − K ab K −1
bb K ba ]u bb f b (7.26)

We note that Eq. (7.26) represent a reduced set of dofs (only u a from u ) and it has been obtained
without approximation. It is thus provided by an exact static reduction.
1 The method is sometimes called the Guyan-Irons method, since Guyan and Irons proposed the method independently

at about the same time.

7.3 Substructuring methods 153

K −1
Further insight may be obtained by using the exact static reduction matrix S , −K bb K ba in a
T
transformation of the equation system (7.25). Let the transformation matrix be
 
I 0
T= (7.27)
S N
where it is obvious that its left side partition [II ; S ] is of full column rank independently on the rank
properties of S and where N is a non-unique full-rank quadratic matrix that render all columns of T
linearly independent and thus T invertible. Using the displacement transformation u = T ūu and the
force transformation f̄f = T T f with that transformation matrix in the system equation (7.25) lead to

K aa + S T K ba + K ab S + S T K bb S K ab N + S T K bb N f a + ST f b
    
ūua
= , f̄f (7.28)
N T K ba + N T K bb S N T K bb N ūub NT f b

Using that S = −K K −1 T −T −1
bb K ba and the symmetry properties of K , and thus K ab = K ba and K bb = K bb ,
simplifies this expression into
K aa − K ab K −1 f a + ST f b
    
bb K ba 0 ūua
= , f̄f (7.29)
0 N T K bb N ūub NT f b
Five important observations can be made at this point:
i) The transformation matrix T of the transformation u = T ūu is designed such that the upper
partition of u and ūu are the same, i.e. ūua ≡ u a .
ii) The lower partition ūub = S u a + N u b has lost its interpretation as a set of physical displace-
ments and is better characterized as a set of generalized displacements. These generalized
displacements are often known as internal variables.
iii) The transformed matrix system Eq. (7.29) is such that there is no stiffness coupling between
the dof sets ūua = ua and ūub . This is in agreement with that ua can be solved independently of
the solution of ūub as specified by Eq. (7.26).
iv) The decoupling has been made without introducing approximations. The reduction of the
dofs ūub from analysis thus still produces exact results for the dof set u a . The influence of the
displacement u b has been "condensed" to the displacement set u b which motivates the naming
exact static condensation.
v) To obtain the condensed system there is no need to obtain the nullspace N of the static reduction
matrix S .
Guyan transformation and reduction. In the Guyan reduction approach, the static conden-
sation transformation matrix T is used also on the structural dynamics equation. To simplify
notation, only the undamped case is studied here for which we have M üu + K u = f (t). Since the
transformation matrix is time-invariant we have for the acceleration transformation that üu = T ūu¨ .
Using a partitioning for the mass matrix consistent with the stiffness partitioning given by Eq.
(7.25) we have
 
M aa M ab
M, (7.30)
M ba M bb
which leads to the transformed equations
M aa + S T M ba + M ab S + S T M bb S M ab N + S T M bb N
  
üua
(7.31)
N T M ba + N T M bb S N T M bb N ūu¨ b
K aa − K ab K −1 f a (t) + S T f b (t)
    
bb K ba 0 ua
+ = , f̄f (t)
0 N T K bb N ūub N T f b (t)
154 Chapter 7. Model Reduction and Substructuring

Figure 7.1: Mass and stiffness matrices after transformation. Grey inertia coupling partitions of
mass matrix are neglected in Guyan reduction. Corresponding partitions of stiffness matrix are
identically zero after the static condensation.
........................................................................................

which is seen to introduce inertia coupling effects between the states in the transformed mass
matrix.
With the motivation that either ūu¨ b or the mass coupling M ab N + S T M bb N are assumingly small,
the Guyan approximation strategy to obtain a representation of u a is to neglect the coupling mass
coefficients altogether, see Fig. 7.1, and thus obtain the approximate uncoupled system equation
M aa + S T M ba + M ab S + S T M bb S
  
0 üua
T (7.32)
0 N M bb N ūu¨ b
K aa − K ab K −1
  
bb K ba 0 ua
+ T = f̄f (t)
0 N K bb N ūub
The advantage of this strategy is that the interface dofs u a can be calculated without knowledge of
the remaining dofs ūub and the computation of ūub can be avoided for that reason. The reduced set of
equation of motion for u a is thus

[M K aa − K ab K −1
M aa + S T M ba + M ab S + S T M bb S ]üua + [K ua = f a (t) + S T f b (t) , f̄f a (t) (7.33)
bb K ba ]u

The hereby described substructure can be seen as a superelement with dofs u a . One may note that
this is the result obtained by a reduction transformation in which the reduction matrix R = [II ; S ] is
the leftmost part of the full transformation matrix T . The superelement’s element mass matrix is
thus
M , R T M R = M aa + S T M ba + M ab S + S T M bb S
M̄ (7.34)

while its element stiffness matrix is

K , R T K R = K aa + K ab S = K aa − K ab K −1
K̄ bb K ba (7.35)

and its load vector is

f̄f a (t) , R T f = f a (t) + S T f b (t) (7.36)

To form those there is thus no need to form the non-singular N and they can be used with stan-
dard assembly procedures to form the coupled equations of a system of which the substructure
(superelement) is a part.
Experience has shown that the mass coupling effect from internal variables can often be large
and the Guyan approximation can have a detrimental effect on the accuracy. To alleviate this
neglected coupling other methods have been developed out of which the Craig-Bampton method is
the most well-known. It will be described next.
 7.3 Substructuring methods 155

7.3.2 The Craig-Bampton method

The foundation of the Craig-Bampton method is the exact static condensation used in Guyan
reduction and further it takes bearing on the missing coupling between the internal variables and
the interface dofs that is assumed in the derivation of the Guyan method. The starting point for the
Craig-Bampton reduction is thus the transformation of variables
    
ua I 0 ūua
= (7.37)
ub S N ūub
where, again, is noted by construction that the transformed variables ūua are identical to the interface
dofs u a , i.e ūua ≡ u a . Since N can be any nonsingular square matrix it leaves room for flexibility in
the design of a reduction method. Craig noted that one possible choice for N of suitable size is the
nonsingular real-valued modal matrix Φbb associated to the undamped eigenvalue problem of the
internal dofs
K bb Φ bb = M bb Φ bb Ω 2bb (7.38)
which leads to the transformation
      
ua I 0 ua ua
= ,T (7.39)
ub S Φ bb ūub ūub
The eigenvalue problem (7.38) can be deduced from the free vibration problem of the substruc-
ture
     
K aa K ab u a M aa M ab üua
+ =0 (7.40)
K ba K bb u b M ba M bb üub
by considering the interface dofs u a fixed, i.e. u a = 0 and üua = 0 . The eigenmodes in Φ bb and
the associated eigenvales ωbb thus represent the internal dynamics of the substructure and the
diagonalization properties can thus efficiently be used to rank the eigenmodes as being slow (low
eigenfrequencies) or fast (high eigenfrequencies).
Using mass-ortonormalized modes in Φ bb we have Φ Tbb M bb Φ bb = I and Φ Tbb K bb Φ bb = diag(ωbb 2 ).

Together with that N = Φ bb , the transformed and still unreduced equations (7.32) are then
M aa + S T M ba + M ab S + S T M bb S M ab Φ bb + S T M bb Φ bb üua
  
(7.41)
Φ Tbb M ba + Φ Tbb M bb S I ūu¨ b
K aa − K ab K −1 f a (t) + S T f b (t)
    
bb K ba 0 ua
+ = , f̄f (t)
0 2 )
diag(ωbb ūub Φ Tbb f b (t)
To facilitate a reduction, the internal variables ūub are now partitioned with a slow partition ūubs and
a fast partition ūubf as ūubb = {ūubbs ; ūubbf }. The associated mode matrices of the slow and fast modes
are Φbbs and Φbbf respectively leading to the transformation of variables
 
     ūua 
ua I 0 0
= ūu (7.42)
ub S Φbbs Φbbf  bs 
ūubf
Consistent with that partitioning the transformed equations becomes
M ab Φ bbs + S T M bb Φ bbs M ab Φ bbf + S T M bb Φ bbf  üua 
  
M aa
M̄
Φ Tbbs M ba + Φ Tbbs M bb S I 0  ūu¨ bs
T T ¨ 
Φ bbf M ba + Φ bbf M bb S 0 I ūubf
   T
 
K aa
K̄ 0 0  u a   f a (t) + S f b (t)
+ 0 diag(ωbbs )
 2 0  ūubs = Φ Tbbs f b (t) (7.43)
0 0 2
diag(ωbbf ) ūubf Φ Tbbf f b (t)
   
156 Chapter 7. Model Reduction and Substructuring

Figure 7.2: Mass and stiffness matrices after Craig-Bampton transformation. Grey inertia coupling
partitions of mass matrix and high eigenfrequency partition of stiffness matrix are omitted in
Craig-Bampton reduction. Diagonal line through matrices indicate that those are diagonal.
........................................................................................

where M̄M aa and K̄K aa have been introduced with M̄ M aa , M aa + S T M ba + M ab S + S T M bb S , K̄

K aa ,
−1
K aa − K ab K bb K ba and the slow eigenvalues ωbs and fast eigenvalues ωbf relates to Φ bbs and Φ bbf .
With support of the convergence analysis for the mode displacement method (see Sect. 7.2.1),
and provided that the excitation f b (t) is at the lower end of the frequency spectrum, one can argue
that the generalized displacements of the high-frequency internal variables ūubf (t) are assumingly
small enough to be neglected. One means of eliminating the influence of those on the dynamics of
the interface dofs is by neglecting their coupling giving the approximate equation of motion of the
substructure
M ab Φ bbs + S T M bb Φ bbs 0  üua 
  
M aa
M̄
Φ Tbbs M ba + Φ Tbbs M bb S I 0  ūu¨ bs
¨ 
0 0 I ūubf
   T
 
K aa
K̄ 0 0  u a   f a (t) + S f b (t)
+  0 diag(ωbbs 2 ) 0  ūubs = Φ Tbbs f b (t) (7.44)
2 T
0 0 diag(ωbbf ) ūubf Φ bbf f b (t)
   

This reduction procedure is illustrated in Fig. 7.2.

The decoupled reduced system can thus be extracted as
M aa + S T M ba + M ab S + S T M bb S M ab Φ bbs + S T M bb Φ bbs
  
üua
Φ Tbbs M ba + Φ Tbbs M bb S I ūu¨ bs
K − K ab K −1 f a (t) + S T f b (t)
    
bb K ba 0 ua
+ aa = (7.45)
0 diag(ωbbs2 ) ūubs Φ Tbbs f b (t)
This is the result obtained by a reduction of variables
      
ua I 0 ūua ūua
= , R (7.46)
ub S Φ bbs ūubs ūubs
in which the reduction matrix R is the left part of the full transformation matrix T .
The superelement’s element mass matrix is thus
M aa + S T M ba + M ab S + S T M bb S M ab Φ bbs + S T M bb Φ bbs
 
M=
M̄ (7.47)
Φ Tbbs M ba + Φ Tbbs M bb S I
and its element stiffness matrix and load vectors are K̄ K , R T K R = K aa − K ab K −1
bb K ba and its load
T T
f
vector is f̄ a (t) , R f f
= a (t) + S f b (t). To form those there is thus no need to form the non-
singular N and they can be used with standard assembly procedures to form the coupled equations
of a system of which the substructure (superelement) is a part.
 7.3 Substructuring methods 157

7.3.3 Surrogate parametric model

For to be practically useful, model calibration schemes require rapid calculations of responses as
the parameter setting changes in the iterative search for minimum deviation to test data. For most
industry size FE models in use, sufficiently rapid such calculations are infeasible without model
reduction at present time. Since, on the other hand, also model reduction may take significant time,
a balance needs to be struck between the time spent for model reduction and the accuracy loss
imposed by the reduction process.
A modal reduction scheme is most often employed to create a surrogate model used by the
calibration procedure. The eigenmodes of the corresponding undamped system at the nominal
parameter configuration, belonging to all eigenvalues in a frequency range that significantly overlaps
the frequency range of interest, are then used for reduction. To save computational effort, that
reduction basis is kept within one full calibration cycle and thus not modified as the parameter
settings vary during iterations.
Let the eigenvalue problem be formulated with parameterized mass M (pp) and stiffness K (pp).
Here p is a parameter vector and, in particular, p = p 0 is the nominal parameter setting. At the
nominal parameter setting p 0 the eigenprolem is

K (pp0 )T T Ω with Ω = diag(ω 2j ) ∀ [ω ≤ ω j ≤ ω̄]

T = M (pp0 )T (7.48)

Then the mass and stiffness matrices at any parameter setting p of a reduced model with a
constant reduction matrix T are

M (pp) = T T M (pp)T
M̄ K (pp) = T T K (pp)T
T and K̄ T (7.49)

and, in particular at the nominal configuration

M 0 , T T M (pp0 )T
M̄ K 0 , T T K (pp0 )T
T and K̄ T (7.50)

Since the reduction matrix T is invariant to parameter changes, the gradients of the reduced
order matrices with respect to the k:th parameter pk are

M̄ M /dpk | p =pp0 = T T (dM

M k0 = dM̄ M /dpk | p =pp0 )T
T (7.51)

and

K̄ K /dpk | p =pp0 = T T (dK

K k0 = dK̄ K /dpk | p =pp0 )T
T (7.52)

The gradients of the full size FEM are normally computed by a finite difference approximation
scheme. With finite difference calculation general parameterization is allowed and the need for
source code access to the FE code for modifications is eliminated.
A surrogate model that is linear in the parameters is taken as the first order expansion of the
Taylor series of the reduced order model as
np np
M̄ M 0 + ∑ (pk − pk0 )M̄
M (pp) = M̄ M k0 and K̄ K 0 + ∑ (pk − pk0 )K̄
K (pp) = K̄ K k0 (7.53)
k=1 k=1

with pk0 being the k:th parameter at the nominal setting.

It is observed that, once the reduced order model and its gradients are established from the full
size FE mode, no further evaluation of the FE model is required. That normally leads to very fast
computations.
 158 Chapter 7. Model Reduction and Substructuring

7.4 State-space reduction methods

Many model order reduction techniques have been developed in order to balance the accuracy and
simplicity of the systems, see Refs. [3, 13, 33]. Eigenvalue based model reduction techniques, such
as dynamic condensation, component mode synthesis and modal truncation, continue to receive
attention due to their low computational cost and applicability for very large systems [37, 45].
These approaches produce reduced-order models under the assumption that the eigenmodes with
eigenfrequencies in the vicinity or within the dominant spectrum of the loading are the system’s
dominant eigenmodes to be kept in a reduced-order model. One of the key features of the system is
its input-output behavior and selection metrics for reduction have been developed to best conserve
this feature of the original system. This is the topic of state-space reduction.

7.4.1 State-space reduction based on transfer strength

To obtain accurate reduced-order modal state-space models, it is necessary to determine the
dominant eigenmode subset which keeps the most important features of the system. To reduce
unimportant states a frequency-limited interval modal dominancy index for continuous-time systems
with both under-damped and over-damped eigenvalues has been developed, see Refs. [38, 43]. The
index quantify the two-norm contribution of each eigenmode to the system output deviation. Thus,
the non-dominant modes with less output contribution can be identified and excluded to obtain the
appropriate reduced order model. The main advantage of this index is that it is on explicit form in
terms of modal contribution and the frequency bound of interest. In addition, the retained low-order
model does not contain any uncontrollable and unobservable modes.
Consider the diagonal (i.e. modal) realization of a linear, time invariant and stable continuous-
time system given by Eq. (3.15). Let the partitioned diagonalized state-space representation
be

      
żz1 Σ1 0 z 1 B
B̄
= + 1 s
żz2 0 Σ2 z 2 B2
B̄
 
  z1
C 1 C̄
r = C̄ C2 + Ds (7.54)
z2

where z 1 contains the nr modal coordinates to be retained in the low-order system and Σ 1 is a
diagonal matrix which involves the nr dominant eigenvalues of the full system M = (Σ B, C̄
Σ, B̄ C , D ).
Thus, the truncated system can be written as Mr = (Λ B 1 , C̄
Λ1 , B̄ C 1 , D ).
The low-order model obtained by modal truncation has some guaranteed properties. First,
the H∞ norm of the difference between the full model and the low-order model has an a priori
upper bound. In diagonalized form, the difference between the transfer functions of the full model
G(M, ω) and the reduced-order model, Gr (Mr , ω), can be written as (see Eq. (6.32))
ns
c̄k b̄k
G(M, ω) − Gr (Mr , ω) = ∑ (7.55)
k=nr +1 iω − σk

Thus, the H∞ norm of the error system is upper bounded by the following expression
ns
σ ? (c̄k b̄k )
||G(M, ω) − Gr (Mr , ω)||∞ = sup [σ ? (G(S, ω) − Gr (Sr , ω))] ≤ − ∑ (7.56)
ω∈ℜ k=nr +1 Re(σk )

where σ ? (.) is the largest singular value of matrix (.). Secondly, the eigenvalues of the low-order
model is a subset of the eigenvalues of the original model and therefore they keep their physical
interpretations [13], e.g. the modal truncation preserves the stability property of the full system.
7.4 State-space reduction methods 159

The upper bound of Eq. (7.56) hints to a reduction strategy to yield the low-order model. It
remains to be determined which set of eigensolutions to be used in the reduced-order model in
order to conserve the input-output behavior of the full system to as large extent as possible. To this
end, let the k:th modal coordinate of the diagonalized state-space equation be described by
żk = λk zk + b̄k s
∆rk = c̄k zk (7.57)
where the stimuli vector s ∈ Rns is unit-impulse at t = 0 and the response vectors ∆rk ∈ Rnr are the
modal contributions to the system output. The following metric (or dominance index) quantify the
contribution of the k:th mode to the full system output
Z ∞
Mk = (∆rkH ∆rk )dt (7.58)
0

According to Parseval’s theorem this dominance index can be transformed into the frequency
domain as
Z +∞
Mk = (∆r̂r k (ω)H ∆r̂r k (ω))dω (7.59)
−∞

Utilizing the Laplace transformation, the input-output relation can be written as

∆r̂r k = c̄k (iω − σk )−1 b̄k ŝs(ω) (7.60)

With a unit-impulse signal (ŝs(ω) = 1 ∀ ω) and a substitution of Eq. (7.60) into Eq. (7.59) then
leads to
π b̄H H
Z +∞
−1 H −1 k c̄k c̄k b̄k
Mk = b̄H
k (iω − λk ) c̄k c̄k (iω − λk ) b̄k dω = − (7.61)
−∞ Re(λk )
However, most loading situations of relevance here are such that the (one-sided) load spectrum
is dominated by the content in a specific frequency region from ω1 thru ω2 . For the reduced-order
model to more accurately represent the behavior in this frequency region, the frequency information
of the input can be taken into the account by assuming that ŝs(ω) = 1 ∀ω = [ω1 , ω2 ] and ŝs(ω) = 0
elsewhere. That leads to the following frequency-weighted metric
Z ω2
M̄k = (∆r̂r k (ω)H ∆r̂r k (ω))dω (7.62)
ω1

With Eq. (7.60) this then leads to

b̄H H
k c̄k c̄k b̄k
M̄k = (atan(ω1 − Im(σk )) − atan(ω2 − Im(σk )) (7.63)
Re(σk )
Based on this metric, a frequency interval dominancy based definition of modal dominancy can
be introduced;
Definition 7.4.1 — Modal dominancy. For a given threshold value ε ≥ 0 and frequency
interval [ω1 , ω2 ], let nr be the number of modal metrics for which M̄k > ε for k = 1, 2, . . . , nr .
The modes associated to these are the dominant modes.
As a consequence, the full model has nr dominant and nx − nr non-dominant eigensolutions.
An important feature of this metric is that the retained dominant eigenvalue subspace is both
controllable and observable since each non-observable or non-controllable mode k whould render
M̄k = 0. A fuller description of the reduction procedure, also treating the case with sets of non-
unique coalescing eigenvalues, together with a numerical example that illustrates its performance is
given in [44].
 160 Chapter 7. Model Reduction and Substructuring

7.4.2 State reduction by use of Grammians

Consider the continuous time state-space realization

ẋx = A x + B s , r = C x + D s (7.64)

which is assumed to stem from an asymptotically stable system that is both controllable and
observable. This means that its controllability and observability Grammians
Z tf Z tf
T T
Gc (t0 ,t f ) = eA (t0 −τ) B B T eA (t0 −τ)
dτ Go (t0 ,t f ) = eA (t0 −τ) T
C C eA (t0 −τ) dτ (7.65)
t0 t0

are both non-singular for any control and observation duration t f − t0 > 0. As stated earlier, the
Grammians are not invariant under similarity transformations of the realization. Moore [31] showed
that there exists a similar system for which the Grammians are equal and diagonal. Such realization
is balanced over the interval [t0 ,t f ].
Let us assume that the realization (7.64) has been brought to balanced form, see Sect. 5.2.2, and
moreover the state variables have been permuted such that the diagonal elements of the diagonal
Grammian matrices diag(gk ) are in decreasing order, i.e. that gk+1 ≤ gk ∀k ∈ [1, N − 1]. Let us
sub-divide the state vector into two partitions x 1 and x 2 with response contributions r 1 and r 2
respectively and write the partitioned balanced realization as
       
ẋx1 A 11 A 12 x 1 B1 0 s1
= + r 1 = C 1x1 r 2 = C 2x2 (7.66)
ẋx2 A 21 A 22 x 2 0 B2 s2

where A11 is k × k. Let s1 and s2 be the minimum norm stimuli functions that drive the state
from the origin to [xxT1 0 ]T and [00 x T2 ]T respectively, in the time interval [t0 ,t f ]. It was shown by
Moore that
R tf
t0 ||ss2 ||2 dτ gk ||xx2 (t f )||2
R tf ≥ (7.67)
t0 ||ss1 ||2 dτ gk+1 ||xx1 (t f )||2

If gk  gk+1 and s 1 and s 2 have the same norms it follows that ||xx2 (t f )||2  ||xx1 (t f )||2 . In other
words, the part x 2 of the state x is much less affected by the input than the part x 1 . Analogously,
let r 1 and r 2 be the homogeneous (i.e. s (τ) = 0 ∀t0 < τ < t f ) system’s responses from the initial
states [xxT1 (t0 ) 0 ]T and [00 x T2 (t0 )]T respectively. Then it can be shown that
Z tf Z tf
||rr 2 (τ)||2 dτ  ||rr 1 (τ)||2 dτ (7.68)
t0 t0

if gk+1  gk and ||xx1 (t0 )|| = ||xx2 (t0 )||. This means that the state x 1 affects the output much more
than the state x 2 .
It seems reasonable to assume that the state x 2 does not affect the input-output behavior of the
system very much if gk  gk+1 . This assumption suggests that the realization triple {A C1}
A11 , B 1 ,C
may be a good lower-order approximation of the system (7.64). Applications have also shown this
to be true also in practice.
It should be noted that, since (7.64) is not on diagonal form, a reduction of the less control-
lable/observable states does not preserve the eigenstructure of the realization. Thus eigenvalues are
not only removed, but also shifted in the reduced realization. This is truly an annoying property of
the reduction procedure but also a property shared by other popular schemes for reduction such as
the Gyuan and Craig-Bampton reductions[7.2]. Although the eigenvalues are shifted, it has been
showed by Pernebo and Silverman [36] that the resulting reduced realization is still asymptotically
stable as detailed below.
7.4 State-space reduction methods 161

Balancing truncation. A reduction scheme that is popular in control engineering is based

on the balanced input-output relation M̄ = {TT −1 −1
b A T b , T b B , C T b , D } , {Ā
A, B̄ C , D}. It has the
B, C̄
drawback of requiring the costly balancing operation but has the beneficial properties given by
the balancing reduction theorem below. Let M̄ be balanced with diagonal controllability and
observability Grammians Gc = Go ≡ G ≡ diag(g). Partition the matrices related to the n̄ separated
larger Grammians in g1 from the n smaller in g2 of G as

     
g1 0 Ā A12
A11 Ā B1
B̄  
G= A=
Ā B=
B̄ C = C̄
C̄ C2
C 1 C̄ (7.69)
0 g2 Ā A22
A21 Ā B2
B̄

Let further the systems S ∈ R(n+nr )×(n+ns ) and Sr ∈ R(n̄+nr )×(n̄+ns ) be

   
A B̄
Ā B A11 B̄
Ā B1
S= Sr = (7.70)
C 0
C̄ C1 0
C̄

The model M̄r = {ĀA11 , B̄ C 1 , D } is then a reduced order system obtained from {Ā
B1 , C̄ A, B̄ C, D}
B, C̄
by balanced truncation. It has some guaranteed properties related to stability, controllability,
observability and impulse response which are given by the following theorem.

Theorem 7.4.1 — Balancing truncation theorem. Given the controllable, observable and sta-
ble system M̄, the reduced-order system M̄r obtained by balanced truncation have the following
properties:
1. M̄r is balanced and has no unstable poles.
A11 ) 6= λq = eig(A
2. If λ p = eig(A A22 ) ∀p, q then M̄r is controllable and observable.
3. Let the ordered singular values of S be σi , i = 1, 2, . . . , q and let further the number of
singular values of Sr be k with k < q. The Hankel H∞ -norm of the difference between the
full-order system S and the reduced-order system Sr is then twice the sum of neglected Hankel
singular values as : ||S − Sr ||H∞ ≤ 2(σk+1 + . . . + σq )

Proof. See [3]. 

The last part of the above theorem implies that if the neglected singular values are small, the
frequency response functions of M and Mr are close. The error bounds of continuous-time and
discrete-time state-space model reduction were established by Glover [19] and Enns [12].
 162 Chapter 7. Model Reduction and Substructuring

7.5 Problems
Problem 7.1 Mode displacement and mode acceleration reduction
For the 2-dof system shown first set up the governing equations and solve the associated eigenvalue
problem for eigenmodes and eigenvalues ω1 and ω2 . Also solve for the frequency ωa and mode
shape that is associated to anti-resonance for excitation at dof # 1. NB! the same problem but
without reduction is treated in 5.1. Then:

a) Solve for u (t) when f1 (t) = fˆ1 sinωt and f2 ≡ 0 for three full cycles of the fundamental mode,
i.e. 0 < t < 3 × 2π/ω1 . The system is at rest at t = 0. Use the mode displacement method
together with the Duhamel integral solution. Reduce the problem by considering the first mode
only. Consider four cases: I) ω = ω1 /2, II) ω = ω1 , III) ω = ω2 and IV) ω = ωa . Plot u1 (t)
and u2 (t).
b) Solve the same problem as in a) but using the reduced mode acceleration method together with
the fundamental mode only. Again, plot u1 (t) and u2 (t).

p2.22g-h X2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 7.2 . . . Guyan reduction of 4-dof system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Use Guyan reduction to reduce the system to the two middle dofs of the system shown (i.e. they
should remain). Establish the reduced order 2 × 2 mass and stiffness matrices. Hint: The static
deflection modes can be figured out without calculation noting the stiffness distribution.

. .......................................................................................
Problem 7.3 Guyan reduction of 5-dof system
For the 5-dof system depicted, do a Guyan reduction to eliminate the dependent dofs u3 and u4 of
the lighter masses. Assemble the resulting 3 × 3 mass and stiffness matrices of the reduced system.

12/01/2012-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5 Problems 163

Problem 7.4 Synthesis of reduced components

Calculate the approximate
p displacement response of the shared dof under the shown loading
condition for which Ω = k/m.

a) Use Gyuan’s method to reduce both subsystems.

b) Use the Craig-Bampton method with one flexible mode of the 3-dof subsystem and exact
condensation of the 2-dof subsystem, i.e one static and one dynamic mode for the Craig-
Bampton reduction. Compare results.

p5.1 + p5.2 X12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 7.5 Synthesis of reduced components

For the 3-dof system shownp reduce the system to the displacement u1 . Do frequency response
calculations for u1 at Ω = k/m and for all methods in a)-c). The load is f1 = 1N. Compare with
the exact result.

a) Use Gyuan’s reduction method.

b) Use Craig-Bampton’s reduction with one static and one normal mode.
c) Use the modal displacement method with two normal modes.

p5.3 X12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 7.6 Craig-Bampton reduction of mass

For the 5-dof system depicted in Prob. 7.3 do a Craig-Bampton reduction to eliminate the dependent
dofs u4 and u5 of the lighter pmasses. NB! the dof numbering. Use one fixed-interface mode, given
at eigenfrequency 1.3946 k/m, in the reduction. The associated Guyan reduction matrix T is
given by the figure. Assemble the resulting 4 × 4 mass matrix of the reduced system.
12/01/2012-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
164 Chapter 7. Model Reduction and Substructuring

Problem 7.7 Craig-Bampton reduction of stiffness

A 5-dof component will be used in a system analysis as a substructure, using the Craig-Bampton
technique for its representation. One has already obtained the Guyan reduction matrix S and wants
to proceed with obtaining a single fixed interface mode of the transformation matrix T . The active
dofs are here u1 and u5 and the others are the dependent dofs to be reduced. p
a) Obtain the fixed interface mode that relates to the eigenfrequency ω1 = 0.7423 k/m.
Normalize the mode such that its largest element becomes unity (+1).
b) Determine the element k̄33 of the reduced stiffness matrix T T K T .

10/20/2009-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II
Testing and Test Data Driven
Modeling

8 Modal Analysis and System Identification

169
8.1 Experimental modal analysis
8.2 Introduction to State-Space System Identification
8.3 State-space subspace identification
8.4 Problems

9 Correlation and Comparison Metric . 181

9.1 Vector correlation metric
9.2 Data correlation metric
9.3 Experimental mode expansion
9.4 Vector and matrix norms
9.5 Problems

10 Data Driven Substructuring . . . . . . . . . . 195

10.1 State-space model on coupling form
10.2 System coupling

11 Vibration Testing . . . . . . . . . . . . . . . . . . . . 203

11.1 Planning sensor and actuator placement
11.2 Testpiece excitation and response data processing
11.3 Vibration testing hardware
11.4 Problems
 8. Modal Analysis and System Identification

8.1 Experimental modal analysis

Experimental modal analysis is the process of experimentally determining the eigenstructure of the
system under test. Linearity of the system, under the loading conditions the system will be subjected
to, is then assumed and attempts to validate that is ideally made as part of the test procedure. The
theoretical basis is given first and then the classical experimental modal analysis based on circle
fitting of mobility is next.

8.1.1 Theoretical foundation

In most cases the system under tesing is supposed to be undamped or viscously damped. The linear
second order differential equations governing its vibratory motion is then

M üu(t) +V
V u̇u(t) + K u (t) = f (t) (8.1)

As an alternative to what was given in Sect 3.1.1 this may be written in a symmetrical first
order form by use of the dummy relation M u̇u(t) − M u̇u(t) = 0 as
       
V M u̇u K 0 u f
+ = (8.2)
M 0 üu 0 −M M u̇u 0

The homogeneous harmonic form of Eq. (8.2), i.e. when f (t) = 0 and u (t) = ûueλt , results in
an eigenvalue problem
     
K 0 V M ûu
φ = −λ φ with φ = (8.3)
0 −M M M 0 λ ûu

It is the purpose of the testing to experimentally obtain the pertinent, generally complex-valued,
eigenvalues λk and the associated eigenmodes φ of the system.
Since the eigenmodes φ are determined only by their shape, and not magnitude, a proper
normalization is of interest. Commonly used such are unit vector norm normalizations. These may
170 Chapter 8. Modal Analysis and System Identification

be either infinite-norms, where the largest element of the vectors are set to one, or the 2-norms of
the eigenvectors set to unity. Together with the modal constants
   
T K 0 T V M
βk = φ k φ and αk = φ k φk (8.4)
0 −M M k M 0

the system characteristics may now be written on first order form. We use of the modal matrix
Φ , [φφ 1 φ 2 . . . φ N ] together with the transformation

 
u
x, = Φη (8.5)
u̇u

which results in
   
T V M T K 0
Φ η (t) + Φ
Φ η̇ Φ η (t) = (8.6)
M 0 0 −M
M
 
T f
diag(αk )η̇
η (t) + diag(βk )η
η (t) = Φ , Ψ (t)
0

Ψeiωt and η (t) = η̂

For stationary harmonic excitation, i.e. when Ψ (t) = Ψ̂ η eiωt , we then have

diag(βk + iαk ω)η̂

η = Ψ̂
Ψ (8.7)

or
   
f f
η = Φ diag(1/iαk (ω − ωk )) Φ T
x̂x = Φ η̂ ,H (8.8)
0 0

where it has been used that βk = −λk αk , −iωk αk . One element Hi j of the system transfer matrix
H can be shown to be
(k) (k)
n φi φ j n Rkij
Hi j = ∑ iαk (ω − ωk ) , ∑ iαk (ω − ωk ) (8.9)
k=1 k=1

It may be seen that the generally complex-valued eigenfrequencies ωk , the eigenvectors φ (k)
together with the modal normalization constants αk (known as modal A:s or modal Foss damp-
ings[9.1]) fully describe the system characteristics.
For proportionally damped systems, i.e. where V = aK K + bMM , a theoretical analysis may be
somewhat simplified. Motivated by this, a commonly made simplification is often to approximate
also non-proportionally damped systems as being proportional. This may be justified if the system’s
damping is light and its eigenfrequencies well separated. Using the eigenvalues ωk2 and eigenmodes
φ k of the corresponding undamped system’s eigenproblem

K φ = ω 2M φ (8.10)

and using the M and K orthogonality properties of the modes, we may transform equation 8.1 into
decoupled second-order differential equations as

Φ T M Φ η̈
η (t) + Φ T V Φ η̇
η (t) + Φ T K Φ η (t) =
diag(µk )η̈
η (t) + diag(aκk + bµk )η̇ η (t) = Φ T f (t)
η (t) + diag(κk )η (8.11)
 8.1 Experimental modal analysis 171

Here it has been used that the modal masses and the modal stiffnesses are µk = φ Tk M φ k and
κk = φ Tk K φ k respectively. In stationary harmonic vibration we have, with the modal damping
νk , aκk + bµk and with η = η̂ η eiωt , that

η = ΦT f (t)
diag(κk + iωνk − ω 2 µk )η (8.12)
√
or, with κk = µk ωk2 and introducing the relative viscous damping ζk , νk /2 κk µk , we have

η = Φ diag(1/µk (ωk2 + 2iζk ωk ω − ω 2 )) Φ T f , H f

η̂ (8.13)

One element of the system transfer function H , in structural dynamics also known as the
dynamic flexibility or the receptance, is
n (k) (k)
φi φ j
Hi j (ω) = ∑ µk (ω 2 + 2iζk ωk ω − ω 2 ) (8.14)
k=1 k

with n being the order of the system matrices. It may be seen that the eigenfrequencies ωk , the
eigenvectors φ k , the modal dampings ζk and the modal masses µk fully describes the system
characteristics. These are the quantities that should be determined by data from testing.
In vibration testing it is uncommon to measure structural displacements. It is much more
common to use accelerometers for sensing and from those and force sensors estimate system
transfer functions in form of accelerances. Since accelerations η̈ η relate to displacements such that
η = −ω 2 η̂
η̈ η eiωt , the associated accelerance transfer function elements Ai j = −ω 2 Hi j become
(k) (k)
n −ω 2 φi φ j
Ai j (ω) = ∑ µk (ω 2 + 2iζk ωk ω − ω 2 ) (8.15)
k=1 k

or analogously, for the mobility transfer functions related to velocity Yi j = −iωHi j

(k) (k)
n −iωφi φ j
Yi j (ω) = ∑ µk (ω 2 + 2iζk ωk ω − ω 2 ) (8.16)
k=1 k

The simple relation between accelerance and mobility Yi j (ω) = Ai j (ω)/iω makes it convenient
to estimate accelerance transfer function from test data and convert to mobility transfer functions
for analysis. Such analysis is described next.

8.1.2 Mobility circle fitting

Most simple modal parameter extraction methods set out from the observation that for systems with
small damping and well separated eigenfrequencies close to a natural frequency, the system transfer
functions are totally dominated by the contribution from that frequency’s associated eigenmode.
Let us consider a mobility transfer function, i.e. the function Yi j relating velocity response at
degree-of-freedom i and applied force at degree-of-freedom j, of the viscoelastic system. Using the
mobility equation (8.16) we may separate the contribution from the r:th mode from the others into
(r) (r) (k) (k)
−iωφi φ j n −iωφi φ j
Yi j (ω) = + ∑ (8.17)
µr (ωr2 + 2iζr ωr ω − ω 2 ) k=1, k6=r µk (ωk2 + 2iζk ωk ω − ω 2 )

For frequencies close to the r:th resonance ωr , the contributions of the other modes may be
approximated as being constant, i.e.
(r) (r)
−iωφi φ j (r)
Yi j (ω) = + Bi j (8.18)
µr (ωr2 + 2iζr ωr ω − ω 2 )
172 Chapter 8. Modal Analysis and System Identification

Figure 8.1: Properties of the modal mobility circle. Experimentally determined discrete mobility
values, between each the frequecy increments are the same, are indicated by dots.
........................................................................................

Here the frequency dependent part has the basic form of the mobility Y of a single-degree-of-
freedom system (of mass mr , stiffness mr ωr2 and relative damping ζr ) which is
iω 2ζr ω 2 /ωr + iω(1 − (ω/ωr )2 )
Y (ω) = = (8.19)
µr ωr2 (1 + 2iζr ω/ωr − (ω/ωr )2 ) µr ωr2 |1 + 2iζr ω/ωr − (ω/ωr )2 |2
(r) (r)
By comparing Eqs (8.18) and (8.19) one notes that the eigenvector element product φi φ j of
the multi-degree-of-freedom system is a scaling of the corresponding single-degree-of-freedom
transfer function.
It may be shown that the transfer function Y form a perfect circle in the complex Y -plane (also
known as the Nyquist plane). It can be verified by inspection that the circle is defined by
Re2 (Y − 1/4µr ζr ωr ) + Im2 (Y − 0) = (1/4µr ζr ωr )2 (8.20)
i.e. the circle centre is at (1/4µr ζr ωr , 0) and its radius is 1/4µr ζr ωr , see Fig. 8.1a. From ge-
ometrical considerations we can deduce from the real and imaginary parts of Eq. (8.19) that

θ 1 − (ω/ωr )2
tan = tanγ = (8.21)
2 2ζr ω/ωr
It may be shown that the maximum sweep rate, i.e. when dθ /dω is at its peak, occurs at ω = ωr .
Another valuable result can be obtained by further inspection of this basic modal circle. Suppose
we have two specific points on the circle, one corresponding to a frequency ωb below the natural
frequency, and the other ωa to one above the natural frequency (see Fig. 8.1b). Then we can write
for the corresponding angles of the circle
θb 1 − (ωb /ωr )2 θa (ωa /ωr )2 − 1
tan = , tan = (8.22)
2 2ζr ωb /ωr 2 2ζr ωa /ωr
and from these two equations we can obtain an expression for the relative viscous damping of the
mode
ωa2 − ωb2
ζr = (8.23)
2ωr (ωa tanθa /2 + ωb tanθb /2)
Armed with the above insight into the properties of the mobility near resonance, it is a straight-
forward matter to devise an algorithm to extract the modal parameters of a particular mode given
experimental discrete-frequency data. The algorithm reads;
 8.2 Introduction to State-Space System Identification 173

i) Fit a circle to the experimentally determined mobility transfer function around the resonance
ωr .
dθ
ii) Locate the natural frequency by finding the maximum sweep rate max( dω ) by numerical
differentiation.
iii) Obtain damping estimates, as of Eq. (8.23), by considering two discrete mobility data on either
sides of the natural frequency.
iv) Determine the eigenvector elements by fixing the modal mass (say: µr = 1) and determine
(r)
φk from the radius of the circle associated to a direct mobility element Ykk . The radius Rkk is
(r)
Rkk = (φk )2 /4µr ζr ωr .
(r)
v) Determine the remaining eigenvector elements φ j ( j 6= k) by considering the radii of the
(r)
mobilities Yk j with fixed modal mass µr and the obtained eigenvector element φk . Here
(r) (r)
Rk j = φ j φk /4µr ζr ωr .

The first step can be performed by any curve-fitting routine which finds a circle that gives a
least-square fit to experimental data. The second step is straightforward for experimental data with
linear frequency increments which are most common. In the third step, there are many choices for
the selection of frequencies ωa and ωb . Different choices should be evaluated and the scatter in
damping estimates determined. If the deviation is less than, say, 5%, reasonable good damping
estimates has been obtained. Steps iv and v are relatively straightforward to process.
It is advisable to synthesize the mobility functions from the extracted modal parameters and to
compare it with the experimental raw mobility data. If the correlation between the two is not good,
proper action should to be taken. See Ref. [14] for more information.

8.2 Introduction to State-Space System Identification

The number of methods for system identification from discrete time series input output data of
linear systems is plentiful. Such are the AR, ARX, ARMA, ARMAX, IV and PEM methods1
(see reference [27]). Most of them are iterative in nature and require that initial estimates of
the model parameters are provided, and increasingly more difficult task as the model complexity
(more states, more stimuli and more responses) grows. Recent methods, more useful in vibrational
engineering, are non-iterative and rely heavily on numerical linear algebra. A major advantage
of these methods is that the user just have to do a few choices, of which the estimation of the
number of significant states i.e. the appropriate model order, is the most important. For this choice,
however, there are criteria and correlation indicators supporting the user. One powerful method, the
N4SID method, will be described here. It is a recent and general method for linear time-invariant
systems, developed in the 1990’s, which is for the system identification from input/output data
of combined deterministic and stochastic systems, such as given by equations (3.1a,b). It was
originally developed for time domain test data, but later adaptation to discrete frequency data has
been made [29]. A major advantage with that adaptation is that non-uniform frequency data is
allowed which gives good flexibility to the user.
Before we go into details about the state-space identification procedures, let us for a moment
assume that not only the inputs and outputs of the realization are known but also the state sequence.
Later we will see that what the state-space methods do is to provide us with a such sequence. By
formulating the discrete time realization with noise as

Ax k + B̄
x k+1 = Ā Bs k + w k and r k = C x k + D s k + v k (8.24)
1 AR - AutoRegressive, ARX - AutoRegressive with eXtended input, ARMA - AutoRegressive with Moving Average,
ARMAX - AutoRegressive with Moving Average and eXtended input, IV - Instrumental Variable method, PEM -
Prediction Error Method
 174 Chapter 8. Modal Analysis and System Identification

on matrix form we have

      
x k+1 A B̆
Ă B xk wk
= + (8.25)
rk C D sk vk

We note that the problem of obtaining the state-space matrices is basically a linear regression
problem. The least squares solution to (8.25) is
   †  
A B̆
Ă B x k+1 xk
= (8.26)
C D rk sk

The residual to the least squares problem define the noise sequence to be
      
wk x k+1 A B̆
Ă B xk
= − (8.27)
vk rk C D sk

The covariance matrix of the noise can now also be determined easily by as the sample sum
of the squared residuals. That will give the covariance and cross-covariance matrices for w and
v . If we were to design a feedback controller for the tested system, these allow us to compute the
Kalman filter gain for state feedback control.
Although, theoretically, we are able to extract the state-space model {Ă
A, B̆
B,CC , D } from a single
least squares solution, in practice this procedure has been found to produce low quality estimates
B and D matrices. The Ă
of the B̆ A and C matrices, however, are usually of high quality. One notes
A and C , the estimation problem based on the transfer function in Eq. (6.28) is linear
that, for fixed Ă
B and D . We may calculate these from the measured input/output data by use of least squares
in B̆
calculation. The procedure is as follows.

C (iω1 I − A )−1 B + D − H x (ω1 )

 
 C (iω2 I − A )−1 B + D − H x (ω2 ) 
{B̆
B, D } := arg min ||   ||2 (8.28)
 
..
{B̆
B,DD}|{Ă
A,CC}  . 
−1 x
C (iωK I − A ) B + D − H (ωK )

Since, for a given set {A

A, B }, the system transfer function is also linear in {C
C , D } an improved
update of these can be obtained by solving the complementary least-squares problem

C (iω1 I − A )−1 B + D − H x (ω1 )

 
 C (iω2 I − A )−1 B + D − H x (ω2 ) 
{C
C , D } := arg min ||   ||2 (8.29)
 
..
{C
C , D }|{Ă
 B}
A,B̆ . 
C (iωK I − A )−1 B + D − H x (ωK )

This procedure has been found to produce high quality estimates of the state-space matrices.
The simulation work may be decreased by transforming the realization to diagonal form. Also, the
procedure may be extended to estimate a possible non-homogeneous initial state besides the system
matrices.

8.3 State-space subspace identification

In the following, the N4SID state-space sub-space method, see Ref. [35], for the deterministic
system is presented.
8.3 State-space subspace identification 175

Central to the two categories is the estimation of the system’s observability matrix
 
C
 CA
    
C O
 
x
O =
 C A2
,

, (8.30)
 ..O C A nx −1
  .
C Anx −1

Since the true model order ns of the tested system is hidden (for a continuous system the order
is infinite), a system identification user needs to specify that model order nx that is thought to best
capture the model as seen through test data. Discrete-time data for the SSSI are from the nt steps
of the stimuli sequences (ssk , k = 1, 2, . . . , nt ) and response sequences (rr k , k = 1, 2, . . . , nt ). The data
projection method used for the estimation of Ox is at the core of the method but is beyond the scope
of this text with good reading provided by [34]. [27] For a given Ox , an estimation of the system’s
output matrix C is given by the upper nr rows of Ox . The shift property O = OA A is exploited to
give an estimation of the system’s dynamic matrix A as

†
A=O O (8.31)

The state-space sub-space methods utilize the shift property of the observability matrix O. We
note that the r × N top partition of O holds the output matrix C . We also note that one single block
shift down (r rows) in the observability matrix corresponds to a post-multiplication of C by the
matrix A . Let us denote the top m − r rows of O with O, and the bottom m − r rows with O and we
have

A
O = OA (8.32)

We may thus obtain the state transition matrix A by use of the pseudo-inverse of O as

†
A=O O (8.33)

Armed with the above, we have the necessary tools to obtain all state-space matrices from
input/ output data, provided the observability matrix is known. The establishing of that matrix is
the fundamental step in the sub-space identification.
Before we state the solution to the identification of the observability matrix, we need a few
definitions and some concepts from linear algebra. We introduce the input and output block Hankel
matrices S and R , respectively, as
 
s0 s1 s2 ... s k−1
 s1 s2 s3 ... sk 

.. .. .. ..
 "past" stimuli in
 .. 
 . . . . .  j block rows
   
Sp  s j−1 sj s j+1 ... s j+k−2 
S= =  (8.34)
Sf  sj
 s j+1 s j+2 ... s j+k−1 

 s j+1 s j+2 s j+3 ... s j+k  "future" stimuli in
 
 . .. .. .. .. j block rows
 ..

. . . . 
s 2 j−1 s 2 j s 2 j+1 . . . s 2 j+k−2
176 Chapter 8. Modal Analysis and System Identification

Figure 8.2: a) Graphical interpretation of the orthogonal projection in the 2-dimensional space.
The projection proj(MM a , M b ) is formed by projecting the row space of M a on the row space of M b .
b) Interpretation of the oblique projection in the 2-dimensional space. The oblique projection is
formed by projecting on the row space of M c , the row space of M a along the row space of M b .

and similarly
 
r0 r1 r2 ... r k−1
 r1 r2 r3 ... rk 

.. .. .. ..
 "past" responses in
 .. 
 . . . . .  j block rows
   
Rp  r j−1 rj r j+1 ... r j+k−2 
R= =  (8.35)
Rf  rj
 r j+1 r j+2 ... r j+k−1 

 r j+1 r j+2 r j+3
 ... r j+k 
 "future" responses in
 . .. .. .. .. j block rows
 ..

. . . . 
r 2 j−1 r 2 j r 2 j+1 . . . r 2 j+k−2

The matrices accommodate the available 2 j + k − 1 data samples from testing. As can be seen,
the upper and lower partitions of the block Hankel matrices are called the past and future block
Hankel matrices. We may note that the notation of past and future partitions of the Hankel matrices
is somewhat ambiguous. The partitions are seen to hold many block elements in common. However,
the notation may be justified by the fact that for individual columns there is a distinct border line
between past and future data in the sense that they are ordered consecutively from top to bottom.
The notation was introduced to support intuitive conceptual discussions about the method.
Similarly to the input and output Hankel matrices, we may introduce the combined input/output
block Hankel matrix. The upper partition of it, the past inputs/outputs, is
 
S
Wp= p (8.36)
Rp

Also, the oblique projection of matrices need to be defined for the following presentation. It is
defined via the orthogonal projection of a matrix. We introduce two matrices M a and M b of the
same column dimension, with M b being of full rank. The orthogonal projection of M a along the
row space of M b is illustrated in figure 8.2a and is defined as

proj(M M b M Tb ]† M b
M a , M b ) , M a M Tb [M (8.37)

M Ta M Tb ]
Now let M c be yet another matrix of the same column dimension as M a and M b with [M
8.3 State-space subspace identification 177

being of full rank. The orthogonal projection of M a on the joint row-space of M b and M c is
    "   T #†  
Mb Mb Mb Mb Mb
M a,
proj(M ) = Mb (8.38)
Mc Mc Mc Mc Mc
 
  Mb
, Nb Nc = N bMb + N cMc
Mc

The oblique projection on the row space of M c of the row space M a of along the row space of
M b is then defined as

M a, M c, M b) , N cM c
proj(M (8.39)

with as defined by equation (8.38). The oblique projection is illustrated in Fig. 8.2b.
By now we have the necessary tools to state the sub-space identification theorem.

R f be the
Theorem 8.3.1 — N4SID observability theorem. Let the projected future outputs R̄
oblique projection, on the row space of the combined past input/output Hankel matrix W p , of the
future output Hankel matrix R f along the row space of the future input Hankel matrix S f , i.e. let
the projection and its singular value decomposition be
  
U 1 Σ1 0  
R f = proj(R
R̄ R f ,W
W p, S f ) = V1 V2 (8.40)
U 2 0 Σ2

Then, for a noise-free system for which the singular values Σ 2 = 0 , the system order nx is
equal to the number of non-zero singular values in Σ 1 . Furthermore, the extended observability
matrix for j > nx is
 
C
 CA 
 2 

1/2
OX =  C A  ≡ U 1 Σ 1 (8.41)

 .. 
 . 
C A j−1

Proof. For proof, see reference [34]. 

All ingredients for a sub-space state-space algorithm are now in place. It can be formulated as:

T HE N4SID A LGORITHM

1. Establish with given input/output data the output, input and combined input/output block
Hankel matrices and compute the projected output Hankel matrix R f ,
2. Compute the SVD of R f . Determine the system order by counting the number of its significant
singular values Σ 1 to which there are associated singular vectors in U 1 ,
1/2
3. Compute the extended observability matrix OX = U 1 Σ 1 ,
4. Extract the C matrix from the first nr rows of OX ,
A by use of the shift structure of OX as in Eq. (8.33),
5. Compute the matrix Ă
B and updates of the C and D
6. Compute, by linear regression to test data, the elements of the B̆
matrices by using Eqs. (8.28) and (8.29).
178 Chapter 8. Modal Analysis and System Identification

There are restrictions on the use of the above state-space sub-space algorithm. First, the number
of block rows j in the past and future input and output Hankel matrices must be greater than the
system order. Here the user must make a qualified guess of the system order and provide the
algorithm with large enough data matrices. Second, the input sequence must be persistently exciting
of order 2 j. That is that the input covariance matrix must be of full rank, see reference [35]. It is
a user responsibility to create such input sequence for the test. Third, the intersection of the row
space of the future input Hankel matrix and the row space of the past states must be empty, see
reference [34]. The last restriction is definitely hard to know beforehand, but practical applications
have showed that this is usually not an issue.
 8.4 Problems 179

8.4 Problems
Problem 8.1 Controllability and observability check
A carriage of mass M has two inverted hinged pendulums influenced by gravity on its back with
lengths l2 and l3 . Both have end-tip bobs of equal mass m (see figure). The external force s act and
cause the linear displacement u1 and angular displacements u2 and u3 about the vertical positions.
The three equations of motion, for small angular motion u2 and u3 , are

M ü1 = −mgu2 − mgu3 + s

m(ü1 + l j ü j ) = mgu j , j = 2, 3

a) Check if controllability exist for the system for all length ratios l2 /l3 .
b) Is the system observable with output r = q2 ?
 9. Correlation and Comparison Metric

In many model comparison situations, not to the least in the validation and calibration of models,
the chosen deviation metric is of the essence. In validation this is used to decide whether the model
fulfills or fails a validation test against the deviation metric based criteria. In model calibration,
the model parameters are adjusted to minimize a metric that measures the difference between test
data and model data. The FE model calibration is most often based on frequency domain or modal
domain data while model validation is most often, at least partially, based on eigenvector correlation
analysis. The metric chosen for validation and calibration could be different and fulfil different
purposes. For validation it is essential that the metric is a measure that is closely related to do the
intended purpose of the model and for calibration it is essential that the metric;

i) requires as few computational operations as possible, since the calibration is an iterative

computational process,
ii) does not destroy parameter informativity that may be present in raw test data, and
iii) renders the parameters to be locally identifiable, and preferable also globally identifiable.

In this chapter we describe two eigenvector correlation metrics that are often used in conjunction
with calibration based on modal parameters. These are the Modal Assurance Criterion (MAC) and
an extension of that which is called the Modal Observability Correlation (MOC). It also introduces
another correlation metric, called the Coordinate MAC (COMAC), which is highly suitable for
post-test screening for possible test outliers. It then concludes with descriptions of time domain,
frequency domain and modal domain calibration and validation metrics. Most metrics are based on
either vector norms or matrix norms, the ones used here are summarized in the end of this chapter.

9.1 Vector correlation metric

9.1.1 Modal assurance criterion - MAC
Ideally, when an experimental modal analysis is done and an FE eigenvalue analysis of the test
object has been made, the results of the two should closely resemble. Comparison of experimentally
found modal properties and calculated ditto may be compared in various manners. The most obvious
182 Chapter 9. Correlation and Comparison Metric

Figure 9.1: Comparison of mode shapes from experimental modal analysis and FE analysis.
Experimental vector elements are plotted versus analytically obtained elements. Only real parts
of eigenvectors are considered. Reasonable correlation but different mode shape scaling can be
observed.
........................................................................................

is probably the comparison of resonance frequencies from the test with the eigenvalues from FE
analysis. If they match in magnitude, it is most likely that the experiment has been performed well
and that the analytical model closely match the test object and its supporting conditions.
Another way of comparing is by means of eigenvector comparison. Since the test is normally not
conducted with sensing of the response of all degrees-of-freedom of the FE model, the associated
eigenvector elements have to be extracted from the FEM’s modal matrix Φ . Let us call the partition
of Φ that correspond to the test sensor locations Φ A and orientations holding the eigenvectors φ kA .
Similarly, let us call the experimentally obtained eigenvectors φ kX , which may be collected as
columns in the experimental mode matrix ΦX . Since the eigenmodes may be determined to shape
but not to magnitude, any eigenvector comparison should be made independent of eigenvector
scaling. A graphical means for comparing eigenvectors is shown in Fig. 9.1. When experimental
eigenvector elements are plotted against their analytical counterparts, they should ideally be lined
up along a straight line in the plot. Another, non-graphical, correlation measure is to calculate
the angle between the experimental and analytical eigenvectors. For co-linear and therefor similar
eigenvectors, a such angle should be zero (or 180°). If instead we take the cosine squared of this
angle, we end up with the Modal Assurance Criterion (MAC) correlation number
(φφ TiX φ jA )2
MAC(i, j) = cos2 ∠(φφ iX , φ jA ) = (9.1)
||φφ iX ||22 ||φφ jA ||22
If the modes and are co-linear, this index is unity. If the modes are orthogonal, and thus fully
uncorrelated, it is zero. For partially correlated eigenvector pairs, the number is between zero and
unity.
The MAC(i, j) indices may be considered as elements of a matrix. If no modes are missing in
the experimental and analytical mode sets, and the eigenmodes are ordered according to increasing
frequency order, the MAC matrix should ideally have ones along its diagonal. Note, however, that
although the eigenmodes should be mutually orthogonal (in some sense) does not mean that the
ideal MAC matrix has zeroes everywhere outside its diagonal. This is so because not all structural
degrees-of-freedom are measured and thus present in the mode matrices.
In practice it is often found that the diagonal elements of the MAC matrix is off from being ones.
One reason for this is that, also for good analytical models, multiple eigenvalues or close-to-multiple
 9.1 Vector correlation metric 183

Figure 9.2: Eigenvalues and two highest MAC numbers versus a model parameter p. Correlation of
vectors of any p is made to the second eigenmode of the system with p = 1.4 which then mimics a
test mode.
........................................................................................

eigenvalues exist. For such models, it is well known that the corresponding eigenvectors may
change drastically for small perturbations in model parameters, see Fig. 9.2.
A rule-of-thumb may be that experimental/analytical mode pairs that give a MAC index greater
than 0.95 should be considered as closely correlated, while those giving an index less than 0.8
should be considered as poorly correlated with a grey-zone between.

9.1.2 Coordinate Modal Assurance Criterion - COMAC

Normally, when the MAC has been computed, it can be observed that a simple re-arrangement
of the modes (in either the experimental or analytical mode sets) would give a MAC matrix with
diagonal elements closer to unity. The analytical model and the test object is thus in some respect
similar, but the order of their eigenvalue magnitudes are not the same. When such re-arrangement
has been made, pairs of well correlated modes can often be identified. Let us denote the associated
mode matrices of such re-arranged modes (in which some of the original eigenmodes that show
little correlation to others may have been left out, but m pair of modes remain) Φ X and Φ A , for
the experimental and analytical modes respectively. An index, known as the CO-ordinate Modal
Assurance Criterion number (COMAC) can then be constructed by considering rows, corresponding
to individual dofs, of the mode matrices. These dofs also corresponds to sensors in testing. Before
such indices are calculated, the eigenvectors have to be normalized to be of unit length because
COMAC is not scale invariant. They also should be rotated, by a multiplication of a complex
number of unit magnitude, such that the norm of the eigenvector pair difference ||φφ iX − φ iA || is
minimized. Defining the COMAC number for the i:th sensor as
1
COMACi = ||Φ
Φi:X − Φ i:A || (9.2)
2m
a set of numbers, each of value between zero and unity, is obtained. For a set of paired modes,
which give acceptably large MAC indices, a COMACi index far from being zero indicated an
experimental error for the i:th dof/sensor. Reasons for this are often loose accelerometers, badly
 184 Chapter 9. Correlation and Comparison Metric

connected sensor wires or erroneous sensor calibration but it may be caused by other sources as
well.

9.1.3 Modal observability correlation - MOC

Starting with a state-space realization of a system, we can bring it to a diagonal balanced form.
We do that in sequence such that we first bring it to the decoupled diagonal form, see Sect. 4.1.5.
After that, all diagonal states are then individually brought to a balanced form which brings us
to the resulting diagonal balanced realization in which the modes have equal observability and
controllability Gramians, see Sect. 5.2.2. By that sequence of operations we note that we fix the
realization and the realization is unique except for the ordering of the states. In particular we
note that the columns of the matrix C̄C of the output equation are fixed. Let T 1 be the similarity
transformation matrix of the diagonalizing operation and T 2 the diagonal similarity transformation
matrix of the balancing operation. The diagonal element of T 2 that relate to the k:th state is T 2k .
Then we have

C = C T 1T 2
C̄ (9.3)

In particular we note that for this matrix, the k:th column being C ρ k T 2 , which we see is
a projection of the k:th eigenvector ρ k of A . By that we also note that the k:th column of the
observability matrix, see Eq. (5.43), becomes
 
C ρ k T 2k
 C ρ T 2k σk 
 k 
2 
O:kbd =  C ρ k T 2k σk  (9.4)

 .. 
 . 
C ρ k T 2k σkN−1

and thus each column of Obd is associated to quantities of one single eigenmode only. We use that
property in the sequel. The notation Obd hints that it is the observability matrix of the balanced and
diagonalized realization.
To make the problem of distinguishing between eigenvectors smaller than it is when using MAC,
the correlation metric based on the observability matrix of the diagonal and balanced state-space
realization can be used. One such correlation metric is the Basic Modal Observability Correlation
(bMOC). For this correlation metric we use the columns of the observability matrix of the balanced
diagonal realization. We define the bMOC correlation metric between the j:th and k:th columns of
the observability matrix to be

|OH
: jbd O:kbd |
2
bMOC jk = (9.5)
(OH H
: jbd O: jbd )(O:kbd O:kbd )

We note that the j:th and k:th columns of Obd relate to the corresponding eigenvectors that are
weighted with their associated eigenvalues to increasing power order, see equation (9.4). Using
two sets of experimental data taken for two individuals from an ensemble of components, we see
in comparison between Fig. 9.3 that the cross-correlation between modes drops significantly for
bMOC in comparison with MAC. This is accomplished by augmenting the eigenvectors with the
corresponding eigenvalues in the bMOC.
However, a further distinction between modes can be obtained if we use the scaling property
of the balanced realization. We note that the controllability and observability balancing of the
states makes the columns of the observability matrix fixed. This means that for a modal state
that contribute little to the input/output relation, the norm of the corresponding column of the
 9.2 Data correlation metric 185

Figure 9.3: MAC correlation of two mode sets indicated with color codes.
........................................................................................

Figure 9.4: MAC correlation of two mode sets indicated with color codes.
........................................................................................

observability matrix is low. On the contrary, for a modal state which contributes much to the
input-output relation, the corresponding column norm is large. We use this property to further
distinguish between states which have close eigenvalues and similar eigenvectors but that have
strong dissimilarities in the input-output contribution to form an enhanced version of the bMOC,
the scaled MOC. We define this to be
|OH
: jbd O:kbd |
2
MOC jk = (9.6)
sup2 ((OH H
: jbd O: jbd ), (O:kbd O:kbd ))

In Fig. 9.4 we see that we further add to the distinguishability between modes by using this
metric.

9.2 Data correlation metric

9.2.1 Frequency response metric
Most calibration problems are solved by gradient based minimization techniques. A calibration
scheme that uses a gradient based minimizer, needs to work with a smooth deviation metric for
high likelihood of success. That is to obtain convergence in the search for the parameter’s optimum
settings from start settings of the parameters. A well calibrated model should give high accuracy in
simulation of test output quantities, and ideally predictions with high credibility of other output
quantities not tested. In a frequency domain context, this often translates to that model which
186 Chapter 9. Correlation and Comparison Metric

accurately captures the structural resonances and possibly also its anti-resonances. A metric that
does not discriminate against deviations at frequencies where the structural response is small is the
quadratic functional

Q = δ Hδ (9.7)

in which the deviation vector δ is

H A (pp) − H X )
δ (pp) = log10 vect(H (9.8)

Here H A and H X are the frequency response functions established by FE analysis and provided
by experiments respectively, see Eq. (3.12). The function vect(.) is the vectorizing operation that
makes all frequency response function elements of the nr × ns transfer function, at all n f discrete
frequencies used for evaluation, into a nr ns n f × 1 column vector. Due to the non-uniqueness of the
logarithm function for complex numbers, a mathematically equivalent formulation better suited for
computer implementation is

δ (pp) = log10 (vect(H

H A (pp)./vect(H
H X )) (9.9)

where the ./ operator denotes the element-by-element division. Since finite element model calibra-
tion tends to be very computationally demanding, calibration criteria that lead to computational
efficiency is strongly of the essence. If operations can be spared and therefore reduce calculation
times, it can mean that the calibration issue can move from being an interesting theoretical concept
to being practically useful. However, all computations need to be optimized to provide as much use-
ful information as possible with as little effort as possible. Such optimization targets the sampling
strategy for the discrete frequencies that are selected for frequency response function evaluation.
The half-band-width ∆ωk of a damped structural resonance at frequency ωk is given by

∆ωk = 2ζk ωk (9.10)

with ζk being the relative modal damping of the k:th mode. One observes that the half-band-width
increases linearly with increasing resonance frequency. It then seems to be a good frequency
sampling strategy to utilize frequency steps that increase linearly with frequency. Such sampling
keeps the number of samples over one half-band-width constant over the range. That is to take
steps such that the logarithm of the frequency steps over the frequency range is constant, i.e.

log(ωi+1 − ωi ) = log(ωi − ωi−1 ) ∀i (9.11)

That sampling strategy seems reasonable, provided that relative damping of all modes in the
range are equal, which rarely happens for experimentally found eigenmodes. However, the damping
can be equalized by a procedure that is described below. The influence of the density of discrete
frequency steps and damping level, using the model of the example treated in Chapter 10.4, can be
seen in Fig 9.5. It can be noted that the calibration criterion function is regularized by increasing
damping and making smaller frequency steps.

Damping equalization. A central issue for FRF based model calibration is that of model
damping. Since in general, damping has been found to be very difficult to model using first
principles, it is most often assigned a simple representation for modeling convenience. Such are the
Rayleigh damping and the modal damping models. These simple representations of all physical
dissipation mechanisms that contribute to the system damping often render a model with prediction
accuracy that is sufficient for its intended purpose. In case of modal damping modelling, the model’s
9.2 Data correlation metric 187

Figure 9.5: Normalized deviation metric versus parameter variation from nominal of two parameters,
k9 and m4 of the system seen in chapter 10.4, for (a) three system damping levels at 0.01, 0.1 and
1%, and (b) various frequency sampling density in number of samples per half-bandwidth (p½bw).
........................................................................................

damping is set using the outcome of experimental modal analysis of a modal test of the structure
under investigation, or using engineering judgement in the mapping of modal damping data from
other similarly built structures. The modal damping found in experiments are normally used for
FE simulation without further attempts to understand their physical background. Physically based
parameterization of damping phenomena such as friction, radiation and dissipation is therefore
uncommon.
The nature of the damping mechanisms is normally such that the modal damping varies from
mode to mode. That makes a mapping of experimentally obtained modal damping into modal
damping of FE modes cumbersome. The difficulty arises since the mapping of modal damping
relies on mode shape pairing, meaning that the same amount of modal damping should be assigned
to modes that are similar in their deformation pattern. Mode pairing of EMA modes and FEM
modes are usually made through correlation analysis using MAC correlation analysis. Such MAC
based pairing is normally not straightforward, especially for systems with high modal density
and with sparsely distributed experimental sensor layout. Eigenmode pairing for the purpose of
damping mapping would be unnecessary if the modal damping was same for all modes.
To overcome the problem of mode pairing, a method of damping equalization is suggested. If
all modes have the same amount of damping, there is no need for mode matching. The damping
equalization is achieved by imposing the same modal damping on all experimentally found system
modes by perturbation of a mathematical model of the experimental data found from system
identification using raw frequency response function data H Xraw . Using contemporary state-of-
the-art system identification methods, such as the state-space sub-space method N4SID, these
experimental data can be used to obtain a mathematical state-space model. The experimentally
 188 Chapter 9. Correlation and Comparison Metric

found system transfer function H Xraw can then be represented by the identified state-space system

ẋx = A x + B s , r = C x + D s (9.12)

such that an approximation to the raw data becomes

H X = C [iωII − A ]−1 B + D (9.13)

The experimental state-space system can be brought to diagonal form by a similarity transfor-
mation as seen in equation (3.26a,b). We thus have that
with diagi= in which i are the complex-valued system poles as given by the experimental data.
The relative modal damping n , obtained from these poles are
In the process of damping equalization, the real parts of the poles are perturbed such that the
damping is made equal for all modes. The modal dampings are then set to a single fixed value 0 ,
i.e.
The effect of such damping equalization is that the oscillatory imaginary part of the poles are
preserved and the real damping part is modified such that the perturbed system poles are now
and the modified state-space realization is
with
This in turn give us a modified transfer function for the experimental model, such that the
transfer function used for calibration with damping equalization is
At this stage it should be obvious that the application of the system identification procedure
on the raw test data HrawX has led us to a mathematical model which we can evaluate for any
frequency . In particular it means that we can use the equal log-frequency increments as given
by equation (9.24) for transfer function evaluation. In addition to that, we are also able to make
fictitious modifications of the system under test. A particularly useful such modification is that we
can adjust the system damping level, leaving stiffness and inertia properties intact, such that all
system modal damping are set equal. The model calibration of the FE model can then be made
towards this fictitious experimental model for calibration of parameters that relate to mass and
stiffness only. For the FE based system representation, the modal damping allows for a simple
representation. For a system with given mass and stiffness matrices M and K we have the viscous
damping matrix V to be[24]
with eigenfrequencies n , modal masses mn , and the modal matrix X given by the undamped
system’s eigenvalue problem
In a calibration procedure we are then able to search for the mass and stiffness related parameters
p of the FE model K(p),M(p) that render the transfer function HA given by equations (3.4), (3.5) and
(3.9) and that let the criterion function of equation (9.20) to be minimal. The discrete frequencies
used to evaluate equation (9.33) does not have to match the discrete frequencies used in testing.

9.2.2 Mode indicator functions - MMIF

Many mechanical systems under test are found to be little damped with relative modal dampings
well below 1%. In that case the structural eigenmodes are close to being real-valued. Such real
modes, called normal modes, may be identified more easily by applying a proper load distribution
onto the structure. It is the purpose of pre-test methods to find such load distributions that maximizes
the likelihood to obtain normal modes and thus minimizes the real part of the displacement response
at the eigenfrequencies as will be shown below. Consider the displacement response ûu(ω) for
which we have

ûu = H (ω) f̂f (ω) (9.14)

 9.3 Experimental mode expansion 189

where the individual elements of the dynamic flexibility (receptance) matrix H (ω) are (see Sect.
6.1.3)
N
φiK φ jK
Hi j (ω) = ∑ 2
(9.15)
K=1 µK (ωK + 2iζK ωK ω − ω 2 )

For systems with well separated eigenfrequencies ωK and small damping ζK in particular,
one notes that the response is strongly dominated by the Kth mode at frequency ω = ωK , i.e. the
receptance elements can there be approximated as the pure imaginary number
φiK φ jK
Hi j (ωK ) ≈ (9.16)
2iµK ζK ωK2

For real-valued excitation vectors f̂f (ω), i.e. for which all elements in f̂f are either in-phase
or 180° out-of-phase, we may separate the terms of Eq. (9.14) into their real and imaginary
components. Henceforth, we drop the angular frequency from the notation and have

Re ûu + Im ûu = Re H f̂f + Im H f̂f (9.17)

Now, for the response to be totally dominated by a specific normal mode at ωK , a load vector f̂f
must be found such that the real part Re ûu of the response vector is as small as possible as compared
to the total response ûu. We define the norm of the total response to be

||ûu|| = ûuH ûu =Re ûuT Re ûu + Im ûuT Im ûu = (9.18)

T T T T
f̂f Re Ĥ
H Re Ĥ H Im Ĥ
H f̂f + f̂f Im Ĥ H f̂f (9.19)
T T T
f̂f [Re Ĥ
H Re Ĥ
H + Im Ĥ
H Im Ĥ
H ] f̂f (9.20)

By minimizing the ratio of the norm of the real part to the norm of the total response, under the
condition that the norm of the loading is constant, we have

H ReHˆ f̂f
T T
||Reûu|| f̂f ReĤ
λ , min = min T T T
(9.21)
||ûu|| f̂f [ReĤ
H ReĤ H + ImĤ
H ImĤH ] f̂f
which is the Rayleigh quotient associated to the smallest eigenvalue λ of the eigenvalue problem

H ReHˆ] f̂f = λ [ReĤ

T T T
[ReĤ H ReĤ
H + ImĤ
H ImĤ
H ] f̂f (9.22)

Plotting the smallest eigenvalues as functions of frequency (Multivariate Mode Indicator

Functions, MMIF) clearly show at which frequencies normal modes exist. The load distribution
that best excite these modes are given by the associated eigenvectors to the eigenvalue problem.
Multiple simultaneous drops of the MMIF functions indicate the multiplicity of the structural
resonances at the particular frequency (see Fig. 9.6).

9.3 Experimental mode expansion

In most situations the modal test need to be limited with respect to the number of used sensors,
usually because of cost or time constraints. The aim of the testing then has to be to give sufficiently
rich information for subsequent analysis. When the spatial resolution of the test is low, sometimes
the structural dynamic characteristics are required also for locations which were not instrumented.
In that case, an analytical model may be of help to interpolate the experimental data to such
locations.
190 Chapter 9. Correlation and Comparison Metric

Figure 9.6: Typical graph of MMIF functions. The significant simultaneous drop of two MMIF
functions indicate eigenfrequency doublett ω2 and ω3 .
........................................................................................

A popular scheme for eigenmode expansion is the System Equivalent Reduction Expansion
Process (SEREP). It sets out from the reduced modal matrix Φ A of the analysis model. If we
partition the modal matrix according to measured and omitted degrees-of-freedom of the test, we
have
 
Φ mA
ΦA = (9.23)
Φ oA n×m

where n is the number of analysis dofs and m is the number of retained modes to be used. The
partition associated to the measurement is Φ mA which is ns × m with ns being the number of
measured responses. The eigenfrequency spectrum of the retained modes should span the frequency
range of interest of the test.
Now, if we have an experimentally determined eigenmode φ mX we can find the best linear
combination of the truncated analytical modes that best approximate the experimental mode in a
least square sense as

ΦTmA Φ mA ]−1 Φ TmA φ mX

φ mX ≈ Φ mA α with α = [Φ (9.24)

The combination factors in α can be used for expansion to the full structural size as
 
ΦmA
φ̃φ X ≈ α (9.25)
Φ oA

and therefore, the interpolated eigenvector of the un-measured dofs is

ΦTmA Φ mA ]−1 Φ TmA φ mX

φ̃φ oX ≈ Φ oA [Φ (9.26)

One also obtains an approximation for the measured dofs, smoothed through the use of the
analytical eigenvectors, as

ΦTmA Φ mA ]−1 Φ TmA φ mX

φ̃φ mX ≈ Φ mA [Φ (9.27)

Another popular scheme for eigenvector expansion is through the use of the Guyan reduction
matrix. In the Guyan reduction method, see Ch. 7, the reduction matrix is given by static
 9.4 Vector and matrix norms 191

condensation as
 
I
R= (9.28)
S

in which the static condensation matrix becomes S = −K K −1

oo K om with the stiffness matrix partition
of the omitted (reduced) dofs being K oo and the cross stiffnesses to the measured (active) dofs
being K om . The full expanded eigenvector is then given by the Guyan transformation as
   
φ mX I
= φ (9.29)
φ oX S mX

and therefore

K −1
φ̃φ mX = S φ mX = −K oo K om φ mX (9.30)

is an approximation to the unmeasured partition of the eigenvector φ oX in a formulation that is

alternative to the SEREP formulation that does not require eigenvector calculations.

9.4 Vector and matrix norms

A norm is a function that maps all the elements of a non-zero vector x or a non-zero matrix A on a
positive real scalar number. It is a metric that measures the "strength" of the vector/matrix. The
elements of the vectors/matrices may be real or complex. Let x and y be two vectors or matrices of
same dimension and let nrm() be the norm function that operates on these. A proper norm function
is defined as a function that satisfies the following three properties;

Strict positiveness: nrm(xx) > 0 ∀xx, x 6= 0 (9.31)

Triangle inequality: nrm(xx + y ) ≤ nrm(xx) + nrm(yy) (9.32)
Positive homogeneity: nrm(αxx) = |α|nrm(xx) ∀α ∈ C (9.33)

Vector norms. For vectors x ∈ Cn the p-norms (also known as Hölder norms) are defined as
(
(∑ni=1 |xi | p )1/p 1 ≤ p < ∞
nrm(xx) , ||xx|| p = (9.34)
max(xi ) p=∞

These vector norms satisfy what is called the Hölder’s inequality

|xxH y | ≤ ||xx|| p ||yy||q ∀ p−1 + q−1 = 1 (9.35)

For p = q = 2 this becomes the well-known Cauchy-Schwartz inequality

|xxH y | ≤ ||xx||2 ||yy||2 (9.36)

with equality holding if and only if y = αxx, α ∈ C. An important property of the 2-norm is that it
is invariant to unitary (orthogonal) transformations. Let U be n × n and U T U = I . It follows that
U x ||22 = x H U T U x = x H x = ||xx||22 . This also holds if U is sub-orthogonal and has size of n × m
||U
where m < n.
Matrix norms. For matrices A ∈ Cn×m , an important class of norms are those that are induced
by the p-vector norms introduced above. More precisely, we have the induced (p,q)-norm for A as

||A
Ax ||q
||A
A|| p,q = sup (9.37)
x 6=0 x || p
||x
192 Chapter 9. Correlation and Comparison Metric

In particular, for a few equi-induced norms, i.e. when p = q = [1, 2, ∞], the following expressions
hold
m
||A
A||1 = max ∑ |Ai j | (9.38)
j
i=1
n
||A
A||∞ = max ∑ |Ai j | (9.39)
i
i=1
q q
||A
A||2 = AA H ) = λmax (A
λmax (A AH A ) (9.40)
where λmax (MM ) is the largest eigenvalue of a square matrix M .
There exist other norms besides the induced matrix norms. An example is the Schatten p-norms.
These non-induced norms are unitarily invariant, i.e. ||U
U A || = ||A
A||. To define them, we utilize the
singular values σi of the n × m matrix A such that for m ≤ n
m
A||S,p = ( ∑ σip (A
||A A))1/p 1 ≤ p < ∞ (9.41)
i=1
It follows that the Schatten norm for p = ∞ is
||A
A||S,∞ = σmax (A
A) (9.42)
which is the same as the 2-induced norm of A . For p = 1 we obtain the sum of singular values of A ,
or what is known as the trace norm of A as
m
||A
A||S,1 = ∑ σi (A
A) (9.43)
i=1
For p = 2 the associated norm is also known as the Frobenius norm or the Schatten 2-norm of
A as
min(m,n)
||A
A||S,2 , ||A
A||F = ( ∑ σi2 (A AH A ))1/2
A))1/2 = (tr(A (9.44)
i=1
M ) denotes the trace (= sum of diagonals) of the square matrix M .
where tr(M
All the matrix norms discussed above satisfy the sub-multiplicativity property
||A
AB || ≤ ||A
A|| ||B
B|| (9.45)
Notice that there exist some matrix norms that do not satisfy this relationship. As an example,
consider the proper matrix norm ||A A|| = max|Ai j |.
One may ask if there is a need for that many vector/matrix norms. One answer is that the
multitude of norms are there for mathematical convenience. Often a mathematical solutions can
more easily be bounded by one vector/matrix norm than others and therefor simplifies mathematical
proofs. However the multitude can also be a source of confusion in communication between people.
In the comparison between norms, the following tabulated relations for real matrices A ∈ Rm×n
may therefore be helpful:
√
||A
A||2 ≤ ||A
A||F ≤ n||A A||2 (9.46)
√
max|Ai j | ≤ ||A
A||2 ≤ mn max|Ai j | (9.47)
p
||A
A||2 ≤ || A ||1 ||AA||∞ (9.48)
1 √
√ ||A A||∞ ≤ ||AA||2 ≤ m||A A||∞ (9.49)
n
1 √
√ ||A A||1 ≤ ||A
A||2 ≤ n||A A||1 (9.50)
m
Comparing the vector norms above with the deviation metric Q given by equation (1.1) we note
that the deviation is the square of the Hölder 2-norm of the deviation vector d.
 9.5 Problems 193

9.5 Problems
Problem 9.1 MAC correlation of eigenvector sets
One has obtained analytical and experimental eigensolutions of a system and wants to make a
correlation analysis. To the analytical modes in modal matrix ΦA , one has obtained the eigenfre-
quencies ωk =6.1, 8.5 and 8.9rad/s. The complex eigenvalues λk = (−0.05 + 5.8i) and (-0.04+9.2i)
are associated with the two experimental modes in Φ X . Make a correlation analysis based on MAC
and determine the analysis/experimental pairs of eigenvalues with the best match of modes. Also
give the corresponding MAC numbers.

12/18/2009-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 9.2 SEREP expansion of eigenvector

For a small system, the modal matrix Φ A has been obtained by analysis. The eigenfrequencies
corresponding to the modes are; 7.4, 23.6, 46.8, 57.2 and 91.8rad/s. From an experiment made
on this system, with the sensors placed as shown in the figure, one has the eigenvector φ 1X
corresponding to the first resonance frequency, found at ω1X = 9rad/s.
a) Use SEREP with two analytical modes as a basis to interpolate the experimental mode to the
dofs that were not measured.
b) Motivate your choice of the modes you used for interpolation.

20/10/2009-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
 10. Data Driven Substructuring

As shown in Ch. 3.1, a state space model ẋx = A x + b s , r = C x + D s submitted to a transformation

of variables x = T z preserves its input-to-output relation. One such transformation, that is suitable
for coupling of components on state-space form, is described next. With that, a transformation can
be made to efficiently reduce the redundant dofs that are common to the coupled components. The
elimination can be done using compatibility at the component intersection and is described after an
outlook to identification procedures that lead to a state-space model deduced from test data.

10.1 State-space model on coupling form

Consider two state-space components (I and II) that are subjected to coupling. The motion of
the interface between the components is represented by the displacement vector r ∈ ℜnc . To the
displacement, the vectorially associated (energy conjugate) stimuli vectors s i ∈ ℜnc that act on the
two components are s I and s II . For displacement output, a proper state-space representation of the
two models i = I, II is

ẋxi = A i x i + B i s i (10.1)
r = C ixi

Displacement compatibility at the common interface leads to that the nc displacement outputs
r are same for both components. For an energy conjugate system, a state-space model structure
Ai , B i ,C
{A C i } without a direct throughput term is proper since such a system lacks direct throughput
in consistency with the following lemma.

Lemma 1. An energy-conjugate state-space system {AA, B , C , D } for which the instantaneous

s s T r
power P provided by the stimuli amounts to P = ṙ , the direct throughput matrix needs to be
D = 0.

Proof. With the response r = C x + D s , the response rate is ṙr = C ẋx + D ṡs = C A x + C B s + D ṡs and
thus the instantaneous power is P = s T (CC A x + C B s ) + s T D ṡs. To be independent on the stimuli
 196 Chapter 10. Data Driven Substructuring

rate ṡs it requires that s T D = 0 for which the only possible solution for arbitrary stimulus s is that
D = 0. 

Let further the two systems be such that they could alternatively be represented by second-order
differential equations. Then the relation between the matrices of the triple {A
A, B , C } is given by the
next lemma.

Lemma 2. For a state-space realization ẋx = A x + B s , r = C x of a second-order system

M üu + V u̇u + K u = Ps s with response r = Pd u being a linear combination of u only and with
non-singular mass matrix M , (a) the relation between B and C is such that C B = 0 , and (b) the
relation between A , B and C is such that C A B 6= 0 .

Proof. For a second-order system, one possible state-space realization pair is, see Eq. (3.10a),
 
0  
B= −1 , C = Pd I 0
M Ps

from which it follows that, also after an arbitrary similarity transformation x = T z , the relation of
T −1 B , C T } is such that
the pair {T
 
−1
  0
(C T ) ( T B ) = P d I 0 =0
M −1 P s

which proofs part (a). With the structure of A given by Eq. (3.10a), it can easily be verified that
C A B = P d M −1 P s . Since M is full rank thus C A B 6= 0 , unless under the trivial condition that
P d = 0 and/or P s = 0 , which concludes the proof of (b). 

10.1.1 Physically motivated modeling constraints

A second order structural dynamics finite element model as that given in Eq. (3.3) is physically
sound. Its basis are well-known first principles and well established discretization techniques. The
corresponding state-space companion can therefore also be considered as physically sound.
On the other hand, state-space models identified from test data can suffer from errors made in
testing and errors and approximations introduced by signal processing of test data as the result of
pure mistakes, non-ideal sensors and others. Such errors may lead to state-space model behaviour
that is not motivated by the physics of the tested real-world system. Typical such errors are
related to stimuli measurements. Vibration testing are usually made by either impact stimuli by
use of hammer-like impactors or by electromagnetic or hydraulic shakers. In hammer testing
the system is hit repeatedly at the same location in the same angle of attack towards the system
surface and averages are made to minimize noise effects on estimated frequency response functions.
Ideally these impacts are exact repetitions which they never are in practice. For shaker induced
vibrations the stimuli force obtained by the motion of the vibrating piston of the shaker, through a
tension/compression rod called a stinger, is measured by a force cell. The stinger should ideally
transmit only the force component in the direction of the stinger and the force cell should be ideal
with sensor output only linearly proportional to that force transmitted by the stinger. These ideal
conditions do not occur in practice, and unmeasured shear forces and bending/twisting couples are
always transmitted from the shaker through the stinger-sensor into the tested system and spuriously
affect the measured system response. These are just a few examples of the many potential error
sources that may lead to identified state-space models that are not truly valid for their intended
purpose. Due to the unavoidable imperfections introduced by the testing procedure the fundamental
10.1 State-space model on coupling form 197

laws of physics can be violated by the identified model. If that is the case that model is thus
physically inconsistent. If proper actions are not taken state-space models describing the relation
between force/couple input and motion output (such as displacement, velocity or acceleration) may
violate various physical laws such as:

Stability law. A linear system is stable if the response due to any excitation is bounded. A
system is said to be asymptotically stable if its free motion converges to a fixed equilibrium state. All
system poles of an asymptotically stable system have negative real parts. An unstable system, on the
other hand, has one or more system poles that are positive real which leads to exponential response
growth from arbitrarily small stimulus. A system is said to be marginally stable if it is neither
asymptotically stable nor unstable. An undamped structural dynamics system M üu + (K K + λ K g )uu =
F with λ < λcr is stable but not asymptotically stable since it will vibrate about an equilibrium
point for eternity after impact. On the other hand, a damped system M üu + V u̇u + K u = F with
a symmetric positive definite viscous damping matrix V is asymptotically stable. An identified
state-space model, stemming from test data of a stable system, that has system poles with spurious
positive part thus needs trimming. A well-established technique is to mirror any unstable model
poles in the imaginary axis and thus bringing those poles being negative real instead of positive real.
This trimming will usually just lead to a small effect on the system’s frequency domain transfer
functions.

Passivity law. System passitivity relates to the flow of energy over a system boundary. The
energy conservation principle stipulates that
Es (t) + Ein (t) − Eout (t) = Es (t0 ) ∀t > t0 (10.2)
with Es being the system’s internal energy and Ein and Eout are the energies provided by outside
energy sources or absorbed by outside energy sinks. These energies are all positive quantities. The
system is isolated by a system boundary that clearly encapsulates the inside (the system) from its
outside. For a system to be long time passive in the mechanical sense (LTP) the outside energy
influx Ein is purely mechanical and the system energy at t = t0 is bounded. Assume that there exists
a stationary periodic solution (period T ), to a stationary periodic system stimuli such that
Es (t + nT ) + Em (t + nT ) − Eout (t + nT ) = (10.3)
Es (t + nT + T ) + Em (t + nT ) + ∆Em (T ) − Eout (t + nT ) − ∆Eout (T )
with Em = Ein being the energy provided by the mechanical input. Over one full cycle T , the
internal energy has thus been changed the amount of ∆Em (T ) − ∆Eout (T ). For the LTP system the
energy is bounded 0 ≤ Es (0) < ∞ and thus it is required for long time stationarity that this amount
of change needs to be positive and thus
∆Em (T ) > ∆Eo (T ) ≥ 0 (10.4)
If not positive, it would require non-positive energy Es (t + nT + T ) < 0 after long time t + nT +
T , which is un-physical. A periodic solution under the condition that ∆Em (T ) = 0 can only persist
until the system’s initial energy Es (t0 ) (such as elastic energy, kinetic energy, potential energy,
chemical energy, electro-magnetic energy and heat) is fully drained. For a system to be LTP it is
thus necessary for the net energy provided by the mechanical stimuli over a full cycle to be positive.
The instantaneous power P(t) supplied by a force stimuli F (t) with energy conjugate velocity
response v (t) is P(t) = F T v (t). For a stationary harmonic (and thus periodic) force F (t) = F̂
F (t)eiωt
and stationary harmonic response v (t) = v̂v(t)eiωt the mechanical energy provided to the system
over a full cycle T = 2π/ω is
Z T
H H
∆Em (T ) = F v̂v} = Re{F̂
P(τ)dτ = Re{F̂ F}
F H F̂ (10.5)
τ=0
198 Chapter 10. Data Driven Substructuring
H
F H F̂
where H is the system transfer function for force-to-velocity. The real part of F̂ F evaluates to

F }T Re{H
∆Em (T ) = Re{F̂ H }Re{F̂ F }T Re{H
F } + Im{F̂ H }Im{F̂
F} (10.6)

which is seen to be positive for all positive definite Re{HH }. For a LTP system, the real part of the
transfer function thus needs to be positive definite. This puts a constraint on any system model that
should truly mimic such a system. Since the system is passive only if the real part of the mobility
frequency response function C (iωII + A )−1 B is positive definite, the passitivity constraint is related
to the positive real (PR) lemma, see [2]. The lemma states that for a system model to be PR for all
frequencies ω there needs to be a symmetric positive definite matrix P for some L such that there is
a solution to the Lyapunov equation

P A + A T P = −L
LT L < 0 (10.7)

under the condition that the following relation holds

BT P = C (10.8)

It has been experienced to be far from trivial to find a solution to (10.6) and (10.7) for the
general MIMO system, and no successful attempts have been made to the authors’ knowledge.
Therefore, attempts have been made to prove PR under less general SISO conditions instead, see e.g.
Ref. citeliljerehn. The general problem is hard to solve since it requires the search for a solution
{P
P, L } if such at all exists (the PR case) or to prove that no such solution exists (the non-PR case).

Reciprocity law. Many mechanical systems obey the principle of reciprocity which manifests
itself in symmetric coefficient matrices in the second order differential equations. One exception is
the spinning mechanical system that generates unsymmetric Coriolis forces and another exception
is the aeroelastic system for which the aeroelastic forces cause unsymmetry. A common practice
to avoid that a state-space model identified from test data violates the law of reciprocity is to
process the data such that reciprocity results. Say that MIMO transfer function data H (ω) are taken
from a system that is judged to be reciprocal with energy conjugate stimuli and responses. The
matrix H (ω) is thus square. A simple procedure to modify test data to be symmetric, and form
reciprocal transfer functions H r (ω), is then to make averaging of non-diagonal elements of the
transfer function matrix as

H (ω) + H T (ω))/2
H r (ω) = (H (10.9)

Other methods rely on weighted averages in which test data quality indicators, such as test data
coherence of the individual elements in H , are used.

Displacement-velocity consistency. Newton’s second law gives a direct relation between

force and acceleration. A state space model for acceleration response can therefore include a
direct throughput term in the output equation, i.e. the output equation for acceleration may read
r a = C a x + D a s with D a 6= 0 . However there is no direct relation from force to displacement or
velocity, these are indirectly related to force and needs to be integrated from the dynamic equation
ẋx = A x + B s . The output equation for displacement output thus must read r d = C d x . Since the
velocity output is related to displacement output as r v = ṙr d = C d (A Ax + B s ) = C d A x + C d B s it
must also hold that d C B = 0 C
. With d B = 0 one also notes that the acceleration output may be
Ax + B s ) = C d A 2 x +C
written r a = ṙr v = C d A (A C d A B s which leads to the conclusion that C d A B is
not necessarily zero.
 10.2 System coupling 199

For accelerance data H x , the procedure to improve the estimates of the pair {B
B, C } described in
Sect. 8.2 can be modified to take account for such physically motivated constraint by iteratively
improving the estimates of first B and then C d in a augmented least-squares sense in the sequence
C d A 2 (iω1 I − A )−1 B +C
C d A B − H x (ω1 )
 
C d A 2 (iω2 I − A )−1 B +C
C d A B − H x (ω2 )
 
B ∗ = arg min||  ..  2
 ||2 (10.10)

.
B |{A
A,C d }  2
C d A (iωk I − A)−1 B +C x

C d A B − H (ωk )
CdB
and
C d A 2 (iω1 I − A )−1 B +C
C d A B − H x (ω1 )
 
C d A 2 (iω2 I − A )−1 B +C
C d A B − H x (ω2 )
 
C ∗d = arg min||  ..  2
 ||2 (10.11)

.
C d |{A
A, B } 
C d A 2 (iωk I − A )−1 B +C

C d A B − H x (ωk )
CdB

Mobility and accelerance constraints. If the mechanical system under study is not free to
undergo rigid-body motion its mobility and accelerance transfer functions in statics (ω = 0) need
to be zero. With the velocity output equation r v = C v x and the acceleration output equation
r a = ṙr v = C v ẋx = C v A x +CC v B s , the mobility matrix at statics thus needs to be C v A −1 B = 0 and the
accelerance matrix at statics needs to be 2C CvB = 0
Augmented least squares solutions that enforce such constraints for accelerance data H x can be
obtained by the iterative solutions of
C v A (iω1 I − A )−1 B +C C v B − H x (ω1 )
 
 C v A (iω2 I − A )−1 B +C C v B − H x (ω2 ) 
..
 
 
∗
B = arg min|| 
 .  2
 ||2 (10.12)
−1 x
B |{A A,C d } C v A (iωk I − A ) B + D +C
 C v B − H (ωk )
C v A −1 B

 
C vB
and
C d A (iω1 I − A )−1 B +C
C v B − H x (ω1 )
 
C d A (iω2 I − A )−1 B +C
C v B − H x (ω2 )
..
 
 
C ∗d = arg min|| 
 .  2
 ||2 (10.13)
−1 x
C d |{A C d A (iωk I − A ) B−1+C
A, B }  C v B − H (ωk ) 
 C vA B 
C vB

10.2 System coupling

Let the interface loads of the components be decomposed of two parts, the external force αi s and
the interaction force between the components βi s̄s. The interface of the two components is then
loaded by the total external stimuli s = αI s + αII s (thus αI + αII = 1). For the interaction forces
βI s̄s and βII s̄s it holds that βI = −βII according to Newton’s law on action and interaction. Further,
assume that the models can be brought to coupling form with a model structure like
   
0 I 0 0
ẋxi = A ivd A ivv A ivb  x i + B ivv  (αi s + βi s̄s) r = I 0 0 x i
 
(10.14)
i i i
A bd A bv A bb 0
200 Chapter 10. Data Driven Substructuring

with state vector

 
r
x i =  ṙr  (10.15)
x ib

and with non-singular B ivv with inverse B −i

vv . The two systems are thus
     
ṙr 0 I 0 0
B −I r  = B −I I −I I −I I  I
s r = I 0 0 x I
 
vv r̈ vv A vd B vv A vv B vv A vb x + αI s + βI s̄ (10.16)

ẋxIb A Ibd A Ibv A Ibb 0
and
     
ṙr 0 I 0 0
B −II r  = B −II II −II II −II II  II
  II
vv r̈
 s
vv A vd B vv A vv B vv A vb x + αII s + βII s̄ r = I 0 0 x(10.17)
II II II II
ẋxb A bd A bv A bb 0
Adding the two systems together results in the coupled system equations with the expanded
state vector x̄xT = [rr ṙr x Ib x II
b ] as
 
 ṙr 
 B−I
(B vv + B −II
 
vv r
)r̈

I (=10.18)

 xb 

x II
 
b
   
0 I 0 0 0
B −I I −II II −I I −II II −I I −II II 
 vv A vd +I B vv A vd B vv A vv +I B vv A vv B vv IA vb B vv A vb  x̄x + (αI + αII )ss + (βII + βII )s̄s
 
 A bd A bv A bb 0   0 
II II II
Abd Abv 0 Abb 0
 
r = I 0 0 0 x̄x
Since αI + αII = 1 and βI + βII = 0 this results in the coupled state space model
0 0 0

I
  
0
I II 
Avd Ā
Ā Avv Ā
Avb Ā
Avb   Bvv 
B̄  
x̄x˙ = 
A I A I A I x̄x +   s r = I 0 0 0 x̄x (10.19)
bd bv bb 0   0 
II II II 0
A bd A bv 0 A bb
−I −II −1 are non-singular and its following
for which it is assumed that B Ivv , B II
vv and B vv + B vv ≡ Γ
matrix partitions can therefore be formed
Avd = Γ[B
Ā B−I I −II II
vv A vd + B vv A vd ] (10.20)
B−I
Avv = Γ[B
Ā I −II II
vv A vv + B vv A vv ] (10.21)
I
Ā B−I
Avb = ΓB I
vv A vb (10.22)
II
Avb = ΓB
Ā B−II II
vv A vb (10.23)
Bvv = Γ
B̄ (10.24)
That a state-space model structure on coupling form is realizable, under certain conditions on A
and C , for the energy-conjugate state-space system follows from the proof of the following lemma.

Lemma 3. (Coupling form lemma) A state-space system ẋx = A x + B s , r = C x (A A ∈ Rn×n , B ∈

C ∈ Rnc ×n ) stemming from a second-order system M üu + V u̇u + K u = Ps s can be transformed
Rn×nc ,C
to a system with state-space triple {Ā
A, B̄ C } on the coupling form according to Eq. (10.14) provided
B, C̄
C A ] is of full row rank 2nc and that nc ≤ n.
C ;C
that [C
10.2 System coupling 201

Proof. Let T be a transformation for {A

A, B , C } with inverse Z be such that
 
C  
T = C A and Z ≡ Z 1 Z 2 Z 3 (10.25)
N

Let further N ∈ Rn−2nc ×n be the projection of the nullspace N B ∈ Rn−nc ×n of B on the nullspace
N C ∈ Rn−2nc ×n of [C
C ;C
C A ] as

N B N TB )−1 N B
N = N C N TB (N (10.26)

with the nullspaces N C and N B defined by

C ;C
[C C A ]N
N C = 0 and N B B = 0 (10.27)

The nullspace N C is of full row rank n − 2nc and thus the combined nullspace N is also of full
row rank n − 2nc because B is not in the nullspace of [C ; C A ] since C A B 6= 0. Since both [C ; C A ]
and N are full row rank and N is a nullspace of [C ; C A ] it implies that T is full rank and thus
non-singular. Using that T Z = I one has
 
CZ 1 CZ 2 CZ 3
C A Z 1 C A Z 2 C A Z 3  = I
NZ1 NZ2 NZ3

which, in particular, imply that C A Z 1 = 0 , C A Z 2 = I and C A Z 3 = 0 . The first block row of the
system matrix on coupling form ĀA = T A T −1 = T A Z is thus
     
C A Z 1 Z 2 Z 3 = C AZ 1 C AZ 2 C AZ 3 = 0 I 0

A. The structure of B̄
which concludes the proof of the coupling form structure of Ā B is given by
       
C CB 0 0
B = T B = CA B =
B̄    C AB  = C A B ≡ B̄
   Bvv 
T T −1
N N BN B ) N BB
N C N B (N 0 0

B. The structure of C̄
since C B = 0 and N B B = 0 which thus proofs the coupling form structure of B̄ C
on the other hand is given from the definition of Z by

C = C T −1 = C Z = C Z 1
   
C̄ CZ 2 CZ 3 = I 0 0


 11. Vibration Testing

A successful outcome of a validation and calibration procedure heavily depends on quality of test
data. The purpose of a setting up the test is to calibrate a number of model parameters that are
to some extent uncertain. The starting point of calibration is then the nominal model that often
gives a good, but not perfect, match to test data. It seems to be a good idea to use that nominal
model in planning the test to increase the likelihood of creating a successful test outcome. Such
planning aims at finding the optimal placement for sensors and actuators in the vibration test. The
optimality should then ideally be put on the optimum sensor/actuator configuration that maximizes
the identifiability of the uncertain model parameters. However, the test planning methods that are
available are more focused on optimality in the identification of system modal states. Since there
is a close link between the model parameters and the modal states of the system, there is some
logic to the thinking that a test planned for optimum observability of system states is also good for
identification of model parameters. A popular method for modal test planning, called the method of
effective independence, is described below.
Even if the test has been rigorously planned and all precautions have been taken to conduct the
test with great accuracy, things can still go wrong. Some problems that occur during the test might
be obvious and caught by a test monitoring procedure. These can usually be compensated for and
therefor do not affect the quality of the final test results that are eventually used for calibration.
However, some problems are more subtle and do not reveal themselves during the test. A post-test
screening procedure to find test outliers is therefore of relevance. In such post-test screening, the
test outcome is compared with the outcome of simulation using the nominal model. Minor and
moderate deviations between test and analysis is expected, but if the deviation is unexpectedly
large, it may indicate that there is a problem with test data. Identifying, and possibly curing, such
problems is essential before test data is taken into the calibration procedure. Methods for post-test
screening is the topic of Sect. 9.1.2.
This chapter also discusses some hardware that is commonly used for vibration testing. It
considers hardware for the test data collection. It also considers hardware for test article support
that provide vibration isolation to minimize the biasing boundary condition effect of added stiffness
and mass and noise effects from uncontrolled vibration sources in the surrounding.
 204 Chapter 11. Vibration Testing

11.1 Planning sensor and actuator placement

In the preparation of an experimental modal analysis, important considerations regard the layout
of sensor locations. The successful outcome of the test very much relies on that the important
vibrational characteristics of the structure is well captured by the test data. The worst positions of
the sensors is naturally at vibrational nodal positions shared by the eigenmodes of interest. Such
locations, if they exist, should thus be avoided to all cost. But what about the best locations? How
should such be found in a test preparation?
From here on, we assume that a good finite element model of the test object is at hand and that
the dynamic characteristics of it is suitable for the planning of sensor layout. Under that condition
the method of Effective Independence (EFI) for sensor placement, as described below, has been
found to work well. Finite element models are usually much more detailed with respect to the
number of degrees-of-freedom being employed as compared to the number of sensors to be used in
the dynamic test. Let the partition of the modal matrix of the finite element model that is associated
with system response quantities given by candidate sensor locations and directions be denoted
Φ c . The measured displacement response r c at the candidate sensors then relate to the generalized
structural motion η (t) as (see Eqs. (4.56) and (3.1b))

r c (t) = Φ c η (t) + v (t) (11.1)

where v (t) is measurement noise and it has been assumed that there is no direct throughput D u
from stimuli s to the output r c .
Under the assumption that the measurement noise is stationary, Gaussian and white, Kammer [25]
advocates that the covariance matrix R = E[(η η − η̃
η )(η η )T ] of the estimation error should be
η − η̃
minimized for a good sensor placement. The true system state is here η and the estimate, that
η . In that case, the inverse of the covariance matrix F = R −1
should be as close to η as possible, is η̃
should be as large as possible. For a dynamic test in which the measurement noise is not correlated
between sensors and for which identical statistical properties of each sensor is possessed, the
measurement noise may be characterized by the scalar variance σ 2 . Kammer showed that the
inverse of the covariance matrix, the Fisher information matrix F that condition is

F = σ −2 Φ Tc Φ c , σ −2 F̄ (11.2)

with F̄ being the scaled Fisher information matrix.

Above, we have loosely discussed the maximization of the Fisher information matrix without
defining a way of obtaining its size. Kammer uses the determinant of the Fisher information matrix
as the scalar measure of its size. Thus, keeping the determinant as large as possible while removing
sensors from the candidate set should be a good strategy for sensor reduction. The quantification of
the contribution to the determinant from the sensor individuals is then a key issue. An index vector
e was devised by Kammer, with elements e j related to the j:th sensor, as

e j = |F̄| − |F¯ j | (11.3)

Here F¯ j is the scaled Fisher information matrix obtained after the j:th sensor has been removed
from the candidate set, i.e. the corresponding row has been eliminated from the modal matrix Φ c .
Since removal of a sensor (removal of a row in Φ c ) does never increase the determinant, it is seen
that the indices are bounded to the interval [0, |F̄|] and they are the fractional contributions to the
determinant. The index numbers e j are called the effective independence of the j:th sensor.
Now, by omitting the sensor giving the least index the minimum decrease of the information
matrix determinant is achieved. By repeating the procedure, eventually a set of sensors remains for
which a large determinant of the Fisher information matrix exist. However, it may be seen from the
definition of the Fisher information matrix, equation (5.2), that the matrix becomes singular when
 11.2 Testpiece excitation and response data processing 205

the number of rows of the modal matrix Φ c is lesser than the its number of columns, i.e. when the
number of sensors are fewer than the number of target modes. Therefore, the minimum number of
sensors is always equal to the number of target eigenmodes since a further sensor reduction would
render a null determinant.
In the pretest planning the EFI iterations are performed until the target number of sensors
remain. Since the procedure is recursive it is efficient. However, the resulting set of sensors cannot
be guaranteed to give optimal identifiability of the modes. The result is sub-optimal but has been
found to give close to optimal results in most situations.

11.2 Testpiece excitation and response data processing

In vibration testing there is a variety of excitation methods in use. The excitation time history can
generally be categorized as being either periodic or aperiodic. These categories can be further
divided into sub-categories. For the periodic excitation types, common excitations are either
mono-frequency harmonic, multi-frequency harmonic, swept-sine or excitation time histories that
are more or less random in nature. The periodic random signals may be either broad-band white
noise signals or have some frequency color obtained through band-pass filtering. Commonly used
aperiodic excitations develop from impulse loading from hammer hit tests or step loading from
snap-back tests in which the test article is instantaneously let loose from a prestressed state. As
other examples of aperiodic testing are the use of fully random excitation obtained by random
signal source sequences sent to a shaker or through the assumed random character of the system
loading under stochastic operating conditions.
This section describes the excitation signal generation and the response signal data collection.
Different types of excitation signal are discussed together with the required signal processing for
each type.

11.2.1 Periodic excitation

The periodic excitation type tests are considered to be more accurate and more suitable for model
calibration than the aperiodic tests. One the other hand, the results of aperiodic excitation tests, such
as impact tests and operational random tests, are often more easy to obtain and need less delicate
hardware. The result of aperiodic test may be useful in the preparation for accurate calibration tests
and for obtaining validation data sets but is not regarded sufficient in producing the high-quality
data needed for calibration. That is because the signal processing of such is more involved than
the processing of periodic data which can produce spurious results. In this section the periodic
excitations are described in more detail. These types of excitation regularly produce good quality
raw test data. They are associated with a characteristic time T , which is the duration of one single
full period.

General periodic excitation. The Fourier series expansion of a periodic input signal s(tk ) = sk
with J harmonic components that is sampled at time instants tk = kTs at a rate Ts is

J
c0 2π jkTs 2π jkTs
sk = + ∑ (c j cos + d j sin ) (11.4)
2 j=1 T T

and the linear system responds to that stimulus with the output response

J
a0 2π jkTs 2π jkTs
rk = + ∑ (a j cos + b j sin ) (11.5)
2 j=1 T T
206 Chapter 11. Vibration Testing

Periodic vibration tests rely on that the system come to a steady-state condition. This does not
happen instantaneously as soon as the excitation is started. The excitation has to be applied in a
number of full periods, also called cycles, until the steady-state condition enters asymptotically.
Undamped systems will never come to a steady-state condition and the initial transient caused by
the start of the excitation will never fade out. However, since all structure of practical use have
some damping associated to it, the initial transients will eventually die off and the response often
approach steady-state periodic with good period-to-period repeatability. If the cyclic repeatability
do not enter, it may be an indication to that the system is not linear and/or that the test apparatus is
not in good condition.

Stepped-sine testing. Stepped-sine testing is the classical testing method of measuring a

frequency response function in which the command signal supplied to the shaker is a discrete
sinusoid with fixed amplitude and frequency. The source frequency of excitation at the i:th frequency
step is ω j with periodicity T j = 2π/ω j . In practice, the excitation force applied to the test article is
not purely mono-frequency but consists of various harmonic components as
I
c0 2πitk 2πitk
sk = + ∑ (ci cos + di sin ) (11.6)
2 i=1 Tj Tj

in which d1 largely dominates over other Fourier coefficients ci and di . In the case the other Fourier
coefficients are negligible, the excitation force is thus approximately the mono-frequency time
history
2πtk 2πtk 2πtk
rk = c1 cos + d1 sin = d¯1 sin( + φ j) (11.7)
Tj Tj Tj
Strictly speaking, sine excitation is simply a particular type of periodic signal, but it has several
unique features, and sine testing uses lend itself to simpler signal processing than other periodic
signals, so it is appropriate to treat it separately.
In order to encompass a frequency range of interest, the source signal frequency is stepped from
one discrete frequency to another in such a way as to provide the required frequency resolution
of the frequency response function searched for. Invariably driving through an attached shaker,
the excitation force and test article responses are measured and discretized by the DAQ for further
processing. In this technique, it is necessary to ensure that steady-state conditions have been
attained before the test data are collected. This entails delaying the start of the measurement process
for a short while after a new frequency has been stepped into, as there will be a transient response
because of the changed conditions. The extent of the unwanted transient response will depend on:

• the proximity of the excitation frequency to a natural frequency of the test article,
• the abruptness of the changeover from the previous sinusoidal signal to the new signal with
another frequency,
• the lightness of the damping of the eigenmodes of the nearby eigenfrequencies, and
• the proximity of the excitation frequency to an anti-resonance frequency (transmission zero
of the transfer function) at the response sensor locations.

The more pronounced each of these features is, the more serious is the transient effect and the
longer must be the delay before measurements are made. In practice, it is only in the immediate
vicinity of a lightly damped resonance and at antiresonance that the necessary delay becomes
significant in comparison to the measurement time.
One of the advantageous features of the stepped-sine test method is the facility of taking
measurements just at the frequencies where and as they are required. For example, the typical
11.2 Testpiece excitation and response data processing 207

FRF function has large frequency regions, away from resonance and antiresonance, of relative
slow change of magnitude. In these regions it is sufficient to take measurements at relative widely
spaced discrete frequencies. By contrast, near the resonance and antiresonance frequencies, the
function exhibits much more rapid changes and it is more appropriate to make measurements at
more closely spaced frequencies. It is also more efficient to use less delay and measurement time
away from these critical frequency regions, partly because there are less problems there but also
because these data are less likely to be required with great density for the system identification
and modal analysis phases later on. Thus, we have the possibility of optimizing the measurement
process when using discrete sinusoidal excitation, especially if the whole measurement is under
computer control. However, the disadvantage of the stepped-sine procedure is still the long testing
time in comparison with other tests with other periodic or aperiodic excitations and modern digital
signal processing.
The distinctly more important advantage of stepped-sine testing is however the preciseness
of signal and noise control as compared with other methods which uses other means of signal
processing. At sinusoidal testing, in which the system has to come to steady-state, it is possible
to do good estimates of the signal-to-noise ratios of the measured signals and of the harmonic
distortion of the signals. Signal distortion may indicate that the test-article behaves non-linearly at
that specific frequency and at that specific excitation magnitude.
The signal processing of discrete-time data from stepped-sine testing differs from the signal
processing of other test types. It relies on linear regression analysis of the fit of the measured signal
to a signal model. Let the signal model approximation be
I
2πitk 2πitk
ỹk = a0 + b0tk + ∑ (ai cos + bi sin ) + vk (11.8)
i=1 Tj Tj

where I denotes the highest order of harmonic signal distortion of significance and vk is the
remaining noise. The first two terms is to model a constant offset (also called a DC offset) and
a slow linear signal drift that is often present in sensor signals. By equating this model ỹk to the
measured signal yk at the nk discrete times for which we have samples we get the matrix relation
 a     
1 t1 cos 2πt 2πt1
 1
sin . . . 0
 y1   v1 
Tj Tj

 
b
    
2πt2 2πt2  0
 
1 t
2 cos Tj sin Tj . . .
 
 
   y2   v2  
a

. .
. . .. .. ..    =  ..  −  .. 
 1 (11.9)
. . . . . b1 
 
 .  
 . 
2πtn 2πtn
1 tnk cos Tj k sin Tj k . . .  ... 

   yn  vn 
k k

This can be solved in a least-squares sense for the signal coefficients ai and bi as
   †

 a0 
 1 t1 cos 2πt Tj
1
sin 2πtTj
1
...  
y1 
b0 
  
2πt2
sin 2πt
  
 
   1 t2 cos Tj Tj
2
. . .
  y2 

a1 =  . . .. .. ..   .. 
 (11.10)
. .


 b1 

 . . . . .   . 
. 2πtnk 2πtnk  
ynk
  
. 1 tnk cos Tj sin Tj ...
 
.

The least squares estimate of the fundamental harmonic coefficients a1 and b1 lay the foundation
for the estimation of the frequency response function of the system at the frequency ω j = 2π/T j .
Let the complex harmonic coefficient of the measured force stimulus s be c1s = a1s − ib1s and of
the measured response r be c1r = a1r − ib1r . Then the estimate of the complex-valued FRF from the
reference stimulus to the response is simply

Hrs (ω j ) = c1r /c1s (11.11)

208 Chapter 11. Vibration Testing

One advantage of the stepped-sine testing procedure is that the FRF estimate may be computed
for a sequence of periodic cycles for statistical evaluation. The mean of the FRF estimates may
then be taken as a good transfer function estimate, but also other statistical quantities, such as
the standard deviation, may be used for assessing the quality of the FRF estimate and be used in
the statistical assessment of the covariance of model parameter estimated in a model calibration
exercise.
The signal distortion is a measure that indicate system non-linearity and problems in the test
setup. We know that a linear system in steady-state respond to a mono-frequency sinusoidal
excitation with an output of the same frequency. Thus, if the output signal from a mono-frequency
sinusoidal test includes other frequency components they are due to non-linear effects or that a
good mono-frequency loading has not been achieved in the first place. Provided a steady-state
periodic condition has entered, a good measure of distortion is then the ratio of mean-square of the
higher-order harmonic signal component to the mean-square of the fundamental harmonics given
by
I
1
dist(ω j ) =
a1 + b21
2 ∑ a2i + b2i (11.12)
i=2

Signal noise may also adversely affect the quality of test data. Let the mean-square of the noise
be vms = ||v1 v2 . . . vnk ||22 and the mean square of the harmonic signals be yms = ∑Ii=1 a2i + b2i . Then
a metric of the signal-to-noise ratio is

snr(ω j ) = yms /vms (11.13)

During the stepped-sine test, the harmonic distortion and the signal-to-noise ratio can be
monitored and if they are found to exceed some threshold values one can take action. One such
possible action to reduce harmonic distortion is to reduce the excitation level that often leads to
that the system non-linearity become less pronounced. However, by reducing the excitation level,
the signal-to-noise ratio of the output signals normally gets poorer. The test engineer needs to
consider the signal-to-noise behavior and adjust the excitation level such that a good balance is
struck between a good signal-to-noise ratio and little distortion.

Other types of periodic testing. With the availability of fast algorithms for computing the
frequency spectrum of a signal to provide simultaneous information on all the frequency components
in a given range, it is a natural extension of the mono-frequency sinusoidal test method to use a
more complex periodic input signal which contains not one but all the frequencies of interest. This
is nothing more complicated than a superposition of several sinusoids simultaneously, with the
Discrete Fourier Transform (DFT) capable of extracting the response to each of these components.
What is lost in the process, however, is the possibility to monitor the distortion as a metric of system
non-linearity.
The method of computing the FRF is quite simple: the DFT is computed of both the force and
response signals and the ratio of these gives the FRF as
air − bir
Hrs (ωi ) = (11.14)
ais − bis
where ais , bis , ais and bis are the Fourier coefficients, see Eqs. (11.5) and (11.5), of the excitation s
and the response r respectively. Since both signals are represented by Fourier series with samples
at discrete frequencies ωi , it follows that the FRF determined in this way is defined only at those
specific frequencies.
Two excitation signals modes are common in periodic testing. These are the full cycle excitation
mode and the burst excitation mode. For the full cycle excitation mode, there is excitation signal
 11.2 Testpiece excitation and response data processing 209

Figure 11.1: (a) Burst mode chirp signal with sweep from 5Hz to 30Hz in 1s, and (b) burst mode
random signal (colored noise) as a realization of low-pass filtered white noise.
........................................................................................

energy during the entire period. In the burst excitation mode, there is signal energy for a short
duration burst and then the signal source is quite for the rest of the period. The burst excitation
thus consists of short sections of an underlying continuous signal - which may be a sine sweep or a
realization of a random signal - followed by a period of zero output, resulting in a response which
shows a transient build-up followed by a decay. The duration of the burst cycle is normally selected
so as to provide that the response signal has just died away to insignificance by the end of one full
period. This is essential if just one single period is used in the data processing to avoid leakage (see
Sect. 11.2.3), but it is not critical if many periods are used and the system is in a periodic steady
state.
The chirp signal testing is a traditional method of frequency response function measurement
that can be used in both burst mode and full cycle model. It utilized a sinusoidal signal with a
frequency that is varied continuously through the frequency range of interest and repeated several
times to attain stationary conditions. By that, one is ascertained that there will be signal energy over
the entire frequency range. The chirp name comes from the sound, a chirp sound, that is emitted
during the test. Examples of chirp signals and random signals in burst mode can be seen in Fig.
11.1.

11.2.2 Aperiodic excitation

The aperiodic excitation type tests are less accurate and less suitable for model calibration. However,
they are used extensively in vibration testing and can sometimes be the only realistic choice. That
is for mostly for civil structures that are to large to shake easily with artificial loading, but instead
loaded by its natural loading such as traffic loading. Aperiodic loading by sharp transients, such as
hammer hits, are also commonly used as to get a quick assessment of the structure´s dynamics with
the quickness having a price of poorer accuracy. Since these types of excitation are not common in
model calibration, they are not further elaborated on here. Recommended reading is in Ref. [14].

11.2.3 Signal processing of periodic signals

Many system identification techniques use transfer function estimates. These are based on the
spectra of the measured test stimuli s (t) and of the system’s measured response r (t) to that stimuli.
The spectral analysis is made from the discretized time signals yk , y(tk ) that may either be the
measured stimuli or system responses. For periodic excitation these are sampled over the duration
of a full period T at N time instants (n = 0, 1, . . . , N − 1) by the DAQ. The sample rate is thus
Ts = N/T with angular sampling frequency ωs = 2πN/T and the fundamental angular frequency
of the period is ω0 = 2π/T = 2π/NTs . A processing of the discretized data are made to create the
210 Chapter 11. Vibration Testing

complex-valued Fourier coefficients ck of the signal sequence

1 N−1 −2πikn/N
ck = ∑ yn e k = 0, 1, . . . , N − 1 (11.15)
N n=0

from which the signal can be reconstructed as

N−1
yn = ∑ ck e2πikn/N n = 0, 1, . . . , N − 1 (11.16)
k=0

Using that ω0 Ts = 2π/N and after some rearrangement of terms this leads to
bN/2c
yn = c0 + ∑ (ck eiknω0 Ts + cN−k e−iknω0 Ts ) + c N einπ (11.17)
2
k=1

where the last term exists only if N is an even integer and bN/2c denotes the largest integer that
is smaller than N/2. Introducing the real-valued coefficients ak and bk so that ck = ak + ibk and
utilizing Euler’s formula exp(iθ ) = cosθ + i sinθ leads to an alternative expression for the real
discretized signal
bN/2c
yn = a0 + ∑ (ak coskω0 nTs − bk sinkω0 nTs ) − a N cosNω0 nTs /2 (11.18)
2
k=1

where, again, the last term exists only if N is even. Eq. (11.18) also leads to the interpolation
formula for any t ∈ [0, T ] as
bN/2c
yn = a0 + ∑ (ak coskω0t − bk sinkω0t) − a N cosNω0t/2 (11.19)
2
k=1

where, once again, the last term only exists for even N.
The Fourier coefficients ck , also called the spectral coefficients, that correspond to the angular
frequencies ωk = kω0 are usually evaluated with extremely fast computations in the circuits built
into the DAQ by use of the Fast Fourier Transform (FFT) algorithm. It is seen from Eq. (11.18) that
the maximum argument of the harmonic functions of the Fourier series is ωmax = Nω0 /2 = π/Ts
that corresponds to half of the angular sampling frequency. This maximum angular frequency is
known as the Nyquist frequency of the sampled signal. The increment between two consecutive
discrete frequencies is ∆ω = 2π/T .
The transfer function estimate, the FRF, of the transfer path from the input si to the response r j
is determined from the estimated Fourier coefficients. With the complex Fourier coefficients cks
from the signal from a force transducer and ckr the complex Fourier coefficients of a signal from a
response transducer, the transfer function Hi j (kω0 ) is therefore given by

H(ω = kω0 ) = ckr /cks (11.20)

There are a number of features of the digital Fourier analysis which, if not properly treated, can
give rise to erroneous results. These are generally the result of the discretization approximation and
of the need to limit the duration of the period. The signal aliasing and frequency leakage are two
very important aspects and will be treated in Sect. 11.2.4.

Averaging. A distinct advantage of using periodic excitation is the possibility of averaging.

This means that the effect of unavoidable signal noise can be minimized. Let us assume that the
 11.2 Testpiece excitation and response data processing 211

Figure 11.2: Aliasing of a high-frequency (HF) signal sampled at a too low rate which make it
similar at the constant-rate samples (red circles) of a low-frequency signal (LF) of same amplitude.
........................................................................................

system has come to a periodically cyclic steady-state. Let us also assume that we re-set the clock t
to zero at the start of each period. Then we have as a periodic model of the measured signal at the
j:th periodic cycle at the k:th time step as
m
2πntk 2πntk
r̃k j = a0 j /2 + b0 j tk + ∑ (an j cos + bn j sin ) + vk j (11.21)
n=1 T T
where vk j is the noise that cannot be explained by the modeled harmonics. If we repeat the cycle
for J periods we obtain the average of the estimate r̃¯k = 1J ∑Jj=1 r̃k j as
m
2πntk 2πntk
r̃¯k = ā0 /2 + ā0tk + ∑ (ān cos + b̄n sin ) + v̄k (11.22)
n=1 T T
where ān , n̄n , n = 0, . . . , m are averages of the model coefficients and v̄n is the period-to-period
average of the noise. If we assume that the noise is Gaussian with zero mean that means that v̄n
tends to zero as more and more periods J are involved in the evaluation. This mean can be evaluated
as more data become available as more periods come to end. The cycling can then be terminated
when the noise average becomes smaller than an accepted threshold. The associated means of
the harmonic coefficients are the good candidates for further processing in the transfer function
estimation.

11.2.4 Data processing caveats and remedies

Aliasing. The problem known as aliasing that is associated with DFT is the result from the
discretization of the underlying continuous time history. With the time discretization process, the
existence of very high-frequency components of the continuous signal may well be misinterpreted
if the sampling rate is too slow. In fact, such high-frequency components will masquerade as
lower-frequency components that will be indistinguishable from the contributions of genuine
low-frequency components. In Fig. 11.2 it can be seen that digitizing a low-frequency signal
produces exactly the same set of discrete-time values as result from the same process applied to a
higher-frequency signal. Thus, a signal of frequency ω and one of ω − Nωs are indistinguishable
when represented as a discretized time history. This fact causes a spurious contribution to the
spectrum obtained from the DFT, even when the spectral components are computed exactly. A
signal which has a frequency content above the Nyquist frequency will appear reflected, or aliased,
onto the range [0 , ωs /2].
212 Chapter 11. Vibration Testing

Figure 11.3: (a) Spectral characteristics of anti-aliasing filter for signals sampled at 1000Hz with
Nyquist frequency at 500Hz and cut-off at 50% of Nyquist frequency, and (b) Broadband random
signal before (black) and after (red) filtering.
........................................................................................

The solution to the aliasing problem is to use an anti-aliasing filter which subjects the signal to
a low-pass analog filter with a characteristic of the principle form shown in Fig. 11.3. This has the
result of submitting a modified signal to the DFT process. Because the filters used are inevitably less
than perfect, and has a finite cut-off rate, it remains necessary to reject the spectral measurements
in a frequency range approaching the Nyquist frequency. As a rule-of-thumb, the reject frequency
threshold at ωs /2.56 has been suggested by filter manufacturers. It is so essential that anti-aliasing
precautions are taken that anti-aliasing filters are usually provided as a non-optional feature of DAQ
analyzers on the market.

Leakage. Leakage is a problem which is a direct consequence of the need to take only a
finite length of time history coupled with the assumption of periodicity that is made in the Fourier
transformation process. The problem can be illustrated by the two examples shown in figure Fig.
11.4 where two sinusoidal signals of same frequencies are subjected to the same analysis process
but with different sampling duration. In the first case (a), the signal is perfectly periodic in the time
window T and the resulting spectrum has only one single spectral component, illustrated by the
single line at the frequency of the sinusoidal. In the second case (b), the periodicity assumption
is not valid and there is a discontinuity implied at each end of the sample set. As a result, the
spectrum produced for this case does not indicate the single frequency which the original time signal
possessed. Indeed, that exact frequency is not actually represented in the spectral components at
all. The signal energy has leaked to spread across a number of spectral components that neighbors
the true frequency. These two examples represent a best case and a bad case scenario although the
leakage problem becomes more acute when the signal frequencies are lower and fewer cycles fill
the sample set time window.
Leakage is a serious problem in many applications of digital signal processing and ways of
avoiding or minimizing its effect are of major importance. Various possibilities to alleviate from
the leakage problem include:
a) changing the duration of the measurement time T to match the underlying periodicity of the
signal so as to capture the exact number of full cycles of the contributing harmonic signal
components. Although such a solution can remove the leakage effect altogether, it can only do
so if the signal being measured is truly periodic and the period of that signal can be determined
precisely,
b) the abruptness of the changeover from the previous sinusoidal signal to the new signal with
11.2 Testpiece excitation and response data processing 213

Figure 11.4: (c) illustrates Fourier spectrum (black dots) of samples of signal in (a) sampled over a
number of full signal periods (black). (d) illustrates leakage of Fourier spectrum of signal in (b)
that is not sampled over a full integer number of periods (black).
........................................................................................

Figure 11.5: Hanning (a), Tukey with 20%+20% cosine taper (b) and exponential (c) time windows
w(t).
........................................................................................

another frequency,
c) increasing the duration of the measurement period T such that the separation between the
spectral components - the frequency resolution - is finer,
d) adding zeroes to the end of the sampled time record, called zero padding, thereby partially
achieving the preceding result but without requiring to sample more data, and
e) modifying the signal sample obtained in such a way as to reduce the severity of the leakage
effect. This process is referred to as windowing and is widely employed in signal processing.
Windowing. In many situations, the most practical method to reduce the leakage problem
involves the use of time windowing and there are a range of different window types for different
classes of problem. Some of the most popular windows are the Hanning window, the Tukey (or
cosine-taper) window and the exponential window, see Fig. 11.5.
Windowing involves the imposition of a weighted profile on the time signal prior to performing
the Fourier transform. The weight function, or window, is generally depicted as a time function
w(t) as in Fig. 11.5. The analysed signal is then y(t) ← w(t)y(t), a replacement for the original
signal y(t). The result of using a window is seen in Fig. 11.6a and, for the case with leakage
214 Chapter 11. Vibration Testing

Figure 11.6: Effect of Tukey window on discrete Fourier transform. (a) is non-periodic signal
(black segment) with Fourier coefficients as in (c), and (b) is same signal made “more periodic”
with Tukey window with less leakage that can be seen in (d).
........................................................................................

previously shown in Fig. 11.4, this produces the improved spectrum shown in Fig. 11.6b. The
Hanning or Tukey windows are typically used for continuous signals, such as are produced by
steady-state or continuous random excitation, while the exponential window is used for transient
vibration application, such as data from tap tests or snap-back tests, where much of the important
information is concentrated in the initial part of the time record and would thus be suppressed by
windows of the Hanning or Tukey types.
In all cases, a re-scaling is required to compensate for the attenuation of the signals by the
application of the windows which otherwise would turn up as spurious test article damping
estimates.
The effect of applying a window to the time sequence signal has been shown and benefit
from such modifications prior to undergoing its Fourier coefficient computation can therefore be
understood. It is possible, also, to witness the effect of applying a window by examining the same
process in the frequency domain and, although this is more complex than the direct multiplication
we have just made in the time domain, it deserves a useful role to make such a parallel study.
It is a simple matter to make a Fourier transformation of the time function w(t) which defines
the window, and to define the corresponding frequency-domain function ŵ(ω). Of course, because
w(t) is a continuous function, also ŵ(ω) will be a continuous function. In seeking to define the
spectrum of a signal after windowing, it must be noted that this cannot be obtained simply by
multiplying the original signal spectrum by the spectrum of the window. Instead, it is necessary
to perform a convolution1 of these two frequency-domain quantities so that the required output
spectrum is expressed in terms of its input spectrum ŷ(ω) and that of the window ŵ(ω) by the

1 A multiplication in the frequency domain corresponds to convolution in the time domain, and vice versa, see

reference [].
11.2 Testpiece excitation and response data processing 215

Figure 11.7: Two-sided frequency spectra of the (a) Hanning, (b) Tukey and (c) exponential
windows seen in Fig. 11.5.
........................................................................................

relationship
Z ∞
ŷ(ω) ← ŵ(ω) ∗ ŷ(ω) , ŵ(γ)ŷ(ω − γ)dγ (11.23)
γ=−∞

where * denotes the convolution process. We shall see, in the next section about filtering, how
the adjoint process is made in which a simple multiplication in the frequency domain demands
convolution in the time domain to define the modified time-history of the signal.
For the specific case of the Hanning window w(t) = 12 (1 − cos 2πt
T ), we obtain the spectrum
shown in Fig. 11.7. Similar spectra can be obtained for the other windows used, see again Fig.
11.7.

Filtering. There is another signal conditioning process which has a direct parallel with win-
dowing, and that is the process of filtering. In fact we have already described one type of filter in
our discussion about aliasing. The anti-aliasing filter is of the low-pass type and other common
filters are high-pass, band-pass, narrow-band and notch filter types, see Fig. 11.8. In practice, all
filters will have a finite frequency range over which they function as designed and will exhibit what
is known as roll-off features near its cut-off and cut-on frequencies. Although said to have cut-off
frequencies, these are not distinct features of the filter, and the distinctness of their filtering effect
in the region of these characteristic frequencies determines their filtering capacity.
In the same way that the time-domain characteristic of a window could be transformed to the
frequency domain, so also can the characteristic of a filter be represented in the time domain. In
order to derive expressions for the time domain descriptions of signals which have been filtered, it
is necessary to use the convolution procedure so that, in the case of filtering we can write

ŷ(ω) ← ŵ(ω)ŷ(ω) (11.24)

and
Z ∞
y(t) ← w(t) ∗ y(t) , w(τ)ŷ(t − τ)dτ (11.25)
τ=−∞

However, it is not usual that we want to perform calculations in this way as the frequency
formula is often more convenient.
Many filters have a state-space representation, which is a useful representation in the description
of the full measurement chain for which the filter is one part. For very precise modeling, the filter
model can then be incorporated in the total state-space model and act in series with the structural
model.
216 Chapter 11. Vibration Testing

Figure 11.8: Frequency characteristics of (a) low-pass filter with cut-off at 20% of Nyquist
frequency fNyq , (b) high-pass filter with cut-on at 10% of fNyq , (c) narrow-band filter with band-
pass between 20-30% of fNyq and (d) notch filter with sharp notch at around 25% of fNyq . All
filters are Butterworth filters of state-order 32.
........................................................................................
 11.2 Testpiece excitation and response data processing 217

11.2.5 Testpiece support system

One important preliminary to the whole process of vibration testing is the preparation of the test
article supporting conditions. This is often not given the attention it deserves and the consequences
which accrue can cause an unnecessary degradation of the whole test. Because one wants to isolate
the system under test to the actual test article, and not include dynamic effects of its surrounding,
one has to take the decision to whether the structure to be tested should be tested under conditions
that mimic free or grounded boundary conditions.

Free Support Conditions. By ideally free support conditions is meant that the test article is
not attached to ground at any location and is, in effect, freely suspended in space. In this condition,
the structure will exhibit rigid body modes which are determined solely by its mass and inertia
properties and in which there is no elastic deformation at all. Theoretically, any structure will
possess at least six rigid body modes2 and each of these has a natural frequency of 0 Hz. By testing
a structure in free conditions, we are able to determine the rigid body modes and thus the mass and
inertia properties which can themselves be very useful data.
In practice, of course, it is not feasible to provide a truly free support unless we perform our
tests in orbit - the structure needs to be held in some way - but it is generally feasible to provide
a suspension system which is a close approximation to the free condition. This can be achieved
by supporting the test article on very flexible spring-like elements, such as might be provided by
light elastic bands known as bungee cords, so that the rigid body modes, while no longer having
zero natural frequencies, have values which are very low in relation to those of the elastic modes.
Very low in this context means that the highest rigid body mode frequency, for which the elasticity
is provided mostly in the elastic support, is less than 10% of that of the lowest elastic mode of
the structure in completely free conditions. One added precaution which can be taken to ensure
minimum interference by the suspension on the lowest elastic modes of the structure is to attach
the suspension as close as possible to nodal points of the modes in question. At the same time,
particular attention should be paid to the possibility of the suspension adding significant damping to
otherwise lightly damped test articles. Besides using bungee cords, other means of soft suspension
methods have been tried: Examples of such is to put light-weight test article on soft polymeric
foam or to put vehicles and aircraft on almost fully deflated tires or special purpose air-bags. A soft
gas-filled cushion called AirRide has been developed by the Modal Shop Inc. and is claimed to
provide mounting eigenfrequencies as low as 2.9Hz when supporting a 295kg test article.
As a parting comment on this type of suspension, it is necessary to note that any tested body
will possess at least six rigid body modes and it is necessary to check that the natural frequencies of
all these are sufficiently low before being satisfied that the suspension system used is sufficiently
soft. To this end, suspension wires, soft cushions, etc. should generally be normal to the primary
direction of vibration rather than in the same direction as these supports. Selection of the optimum
suspension points is one of the features offered by the test planning procedure that is described in
more detail in Sect. 11.1.

Fixed Support Conditions. The opposite type of support is referred to as fixed or grounded
because it attempts to fix selected points on the structure to rigid ground. While this condition is
extremely easy to apply in computational modelling, simply by deleting the appropriate degrees-
of-freedom, it is much more difficult to implement with good accuracy in the practical case. The
reason for this is that it is very difficult to provide a base of foundation on which to attach the test
2A structure which includes internal mechanisms may have more than six rigid body modes. An example is a
generator in which the rotor is free to rotate relative to the stator about one axis. This generator structure therefor has
seven rigid body modes.
218 Chapter 11. Vibration Testing

structure which is sufficiently rigid to provide the necessary grounding. The, to the eye, perfectly
stiff concrete floor or steel test rig is never perfectly rigid and often experience significant motion
when subjected to excitation over a broad frequency spectrum. All structures and materials have a
finite impedance and thus cannot be regarded as truly rigid but whereas we are able to approximate
the free condition by a soft suspension, it is less easy to approximate the grounded condition
without taking extraordinary precautions when designing the support structure. Perhaps the safest
procedure to follow is to measure the mobility FRF of the base structure itself over the frequency
range of the test and to establish that this is a much lower mobility than the corresponding levels for
the test article at the points of attachment. If this condition can be satisfied for all the attachment
points to be grounded then the base structure can reasonably be assumed to be grounded. However,
as a word of caution, it should be noted that the degrees-of-freedom involved will often include
rotations and these are notoriously difficult to measure.
From the above comments, it might be concluded that we should always test structures in
freely supported conditions. Ideally, this is so but there are numerous practical situations where
this approach is simply not feasible and again others where it is not the most appropriate. For
example, very large test articles, such as parts of power generator stations or civil engineering
structures, could not be tested in a freely-supported state. Further, in just the same way that low
frequency properties of a freely supported structure can provide information on its mass and inertia
characteristics, so also can the corresponding low-frequency asymptotes of the mobility curves for
a grounded structure yield information on its static stiffness. Another consideration to be made
when deciding on the format of the test is the environment in which the structure is to operate
and it may happen that the operation support condition more resembles the fixed than the free
condition. The fixed boundary conditions of the test will then give structural eigenmodes that more
resemble the operational eigenmodes. In the real world, where we are dealing with approximations
and less-than-perfect data, there is additional comfort to be gained from a comparison made using
modes which are close to those of the functioning system.

Loaded Boundary Conditions. A compromise procedure can be applied to the advantage in

component testing. In this procedure, the test article is connected at certain interface locations to
another component of known mobility. This may be to a complex test rig with well–known dynamic
characterization. It may also be simple components such as stiff components attached to the test
article in a way that they more-or-less just add dead-weight. This modified or loaded test article is
then studied experimentally and the effects of the added component are added analytically. This
method is used increasingly in place of free supports. When a structure is tested in free conditions,
it is often found that there are fewer modes observed in a given frequency range than there will be
for that same structure when it is installed as a components in a full system assembly. Moreover,
the modes of the structure as an isolated free component are often quite different from to those
experienced in the assembled condition in which the interface is far from free.

Perturbed Boundary Conditions. There is an extended procedure of the above idea of loading
the boundary surfaces of test article. This procedure is known as the perturbed boundary condition
approach. This is an approach well suited for model validation experimentation. In validation one
search for validation data that is not used for model calibration. This approach can provide just
that. The test data base for a given test article can be extended or enriched by the repetition of
the vibration test for different boundary conditions. This, in effect, means testing several different
structures, but each is simply related to the others by the differences in the boundary loads. For
simple boundary loading, these boundary loads can be accounted for very precisely, and so a
multiplicity of test data can be derived from just one test article with very little modification of the
test setup and instrumentation.
 11.3 Vibration testing hardware 219

Figure 11.9: Schematics of a vibration test setup

........................................................................................

Local Stiffening Effects. If it is decided to ground the structure or to add mass loading, care
must be taken to ensure that no local stiffening or other distortion is introduced by the attachment,
other than that which is an integral part of the test article itself. Great care must be paid to the
area of the attachment if a realistic and reliable test configuration is to be obtained. It is advisable
to perform some simple checks to ensure that the whole assembly gives repeatable results when
dismantled and reassembled again. Such attention to detail will be repaid by confidence in the test
outcome.

11.3 Vibration testing hardware

Hardware for use in vibration testing is under constant development. It continues to change rapidly
as the result of innovations and, not to the least, advances in solid-state electronics and computer
hardware. Hence specific hardware capabilities change very quickly and only generic hardware and
test setup is discussed here. A more detailed treatise of the various hardware components follows
in subsequent sections.
A vibration test require several hardware components. A schematic arrangement is shown in
figure 11.9. The basic hardware elements are the sensors (or transducers) that convert the physical
system responses into an electric signals that can be taken to a form suitable for computations.
Some sensors require a signal conditioning device to match the voltage requirements of the input
electronics of a multi-channel data acquisition system (DAQ). The basic functionality of the DAQ
is to sample the signals from the sensors to get sensor data on a digital form for further processing
by computers. That is made by its analog-to-digital (A/D) circuitry. Modern DAQ systems can
sample the sensor signal at a rate of 200kHz or more to a digital resolution of 24 bits. Many
DAQ operate in the ±20V range. With a 24 bit resolution, this translates to a voltage resolution
of 20V/224−1 ≈ 2.4µV, see footnote3 . The DAQ also includes an analog signal filter before the
3A typical sensitivity of an accelerometer is 1mVs2 /m. For such an accelerometer, the resolution of the DAQ means
that the acceleration resolution is about 2.4mm/s2 . In a steady-state sinusoidal vibration at 100Hz this corresponds to
a displacement amplitude of 6nm. This is of the dimension of a few layers of atoms and is close to the threshold of
human’s most sensitive vibration sensing system in the fingertips.
220 Chapter 11. Vibration Testing

A/D converter that is adjusted to fit the system’s sampling rate. High frequency components
from the sensors, well above the frequency range of excitation, is assumed to be spurious noise
that is unwanted from the perspective of the test application, and therefore filtered away by the
high-frequency stop-band characteristics of this filter. Sometimes the DAQ also has a filter with
a stop-band for low frequency signal components, most often called a Digital Current (DC) filter.
Some sensors has a DC drift such that the sensor gives a constant, or very slowly varying, non-zero
signal also when the sensor is not subjected to stimulus. That is effectively taken away by the DC
filter. Since the filters cause magnitude and phase distortion of the signals, it is important for the
test engineer to know about the filter characteristics of the DAQ and how to set up and modify the
filters settings to match the test specification for optimal outcome.
Another feature of most DAQ for vibration testing is its signal source. The source stems
from a programmable signal generator that feeds its digital data through digital-to-analog (D/A)
circuitry to deliver an analog signal source. The signal generator can be programmed to deliver
arbitrary source signal types. The most popular in vibration testing are the mono-frequency or
multi-frequency sinusoidal signal, the chirp signal and the random signal with frequency spectrum
color. A subsequent chapter will treat the characteristics of source signal types. The power supply
for the signal source of the DAQ is normally very small, and the signal has to be boosted through a
power amplifier before it is sent to a shaker to create force magnitudes that are meaningful for test
article shaking.
The actual excitation of the structure is usually made by an electromagnetic shaker that drives
a rod that is attached to the test article via a load transducer. The load transducer is attached to
the test article in such a way that the added transducer mass between the load sensing part of the
transducer and the test article is minimized and thus the signal from the transducer best represent
the force that shake the structure. The purpose of the driving rod, often called the stinger, is to
transmit a force in its longitudinal direction without loading the test article in its transverse direction
and without creating bending or torsional couples. There is a big range of electromagnetic shakers,
from tiny little ones that give a couple of Newtons force at the maximum, to large shakers that
give forces in the kilo-Newton range. The size of the test article and the load levels for which the
article should be validated must guide the selection of shaker size. However, almost all shakers
have stronger power demands than what can be delivered by the source signal of the DAQ, and
thus need to be supplied with power from a power amplifier. The power amplifier has a frequency
characteristic of its own. Most often the power amplifier do a strong signal amplification in a
frequency pass-band only. At some laboratories, high quality Hi-Fi power amplifiers are used to
support small size shakers with power. These typically make a strong signal amplification in the
20-20.000Hz range, while the amplification outside of that frequency band may be substantially
smaller.
Other excitation systems exist, with the most common being the impulse hammer and the
hydraulic shaker. The impulse hammer is equipped with a load transducer at the hammer tip
to sense the contact force in the hit of the test article. Impulse hammer testing is considered to
be quick-and-dirty in the respect that an impulse hammer test can be set up quickly but usually
do not provide data that is accurate enough for careful calibration. Hydraulic shakers are used
by specialized laboratories to excite large systems with large vibration magnitudes to emulate
situations such as earthquakes. These shakers require hydraulic power to created high force levels
that cannot easily be obtained by electromagnetic shakers.
The response sensing in vibration testing is commonly made by he use of acceleration sensing
units called accelerometers. A detailed description of modern accelerometer types is given in the
next section. Strain gauges can also be used as sensors to pick up vibrational response. A strain
gauge is made of metallic or semiconductor material that exhibits a change in electrical resistance
when subjected to strain. The use of strain gauges normally require a Wheatstone bridge circuit to
 11.3 Vibration testing hardware 221

Figure 11.10: A typical accelerometer and its schematics.

........................................................................................

measure the resistance change, but strain gauges that includes signal conditioning hardware have
recently been developed to fit into modern DAQ systems. A model validation procedure could
easily incorporate measured strains.

11.3.1 Accelerometers for vibration testing

Accelerometer types. Accelerometers are transducers that generate an electrical signal output as
a result of the acceleration motion it experiences. The schematics of an accelerometer is shown
in figure 6.2. The most common type of accelerometer operates by the piezoelectric effect4 .
Piezoelectric (PE) accelerometers are generally classified in accordance with their operational mode
– low impedance voltage mode or charge mode. The piezoelectric effect is the phenomenon that
an electrical charge is produced as a result to a strain applied to the material. PE accelerometers
utilize this technology by mounting an inertial mass to a piezoelectric crystal. When the sensing
element experiences acceleration, this inertial mass applies a counter-acting force to the crystal
which creates a proportional strain resulting in the output of an electrical charge. Collected on
an electrode, the high impedance electrical charge signal can be conditioned by either internal or
external electronics for measurement purposes. Accelerometers containing internal electronics
are classified as Integrated Electronic PiezoElectric (IEPE), but commonly referred to as voltage
mode accelerometers. Voltage mode IEPE accelerometers incorporate built-in, signal conditioning
micro-electronics to create an output voltage that is proportional to the experienced acceleration.
Alternatively, PE accelerometers requiring external charge amplifiers for signal conditioning are
called charge mode accelerometers.
All IEPE accelerometers require electrical power delivered to the sensor over simple 2-wire or
coaxial cable schemes as a constant current source. The constant current is most often provided
by an IEPE compatible DAQ to which the accelerometer is hooked up. The built-in electronics of
the accelerometer convert the charge signal generated by the piezoelectric material into a usable
voltage signal as output from the transducer. This output is low impedance and therefor the sensor
signal can be transmitted with long cables and used in noisy electromagnetic environments with
little degradation. That is in contrast to charge mode accelerometers that work with lower voltages
and therefor put higher demands on cable shielding and short cable lengths.
Charge mode PE accelerometers give the high impedance electrical charge signal generated
directly from the piezoelectric sensing element. These transducers require an external charge
amplifier or an in-line charge converter to convert the high impedance charge signal to a low
impedance voltage signal suitable for measurement purposes. Since the output is high impedance,
the charge signal is very sensitive to noise from the surrounding environment and several important
precautionary measures should be taken for proper measurements. Special low noise coaxial
cables should be used between the transducer and the external charge amplifier. These cables are
4 Other, more rarely used, types of accelerometer designs include piezoresistive transduction, based upon strain guage

technology and variable capacitive transduction.

222 Chapter 11. Vibration Testing

Figure 11.11: (a) Shear type accelerometer and (b) flexural mode accelerometer.
........................................................................................

specially treated (e.g. lubricated with graphite) to reduce motion-induced noise effects. Also, it is
critical to maintain high insulation resistance of the transducer, cabling and connectors by keeping
them dry and very clean. Given these precautions compared with the simple operation of voltage
mode accelerometers, charge mode accelerometers are generally only used in high temperature
applications (above 120°C) where the voltage mode accelerometers internal electronics fail.
The last important characteristic of all PE transducers (voltage mode and charge mode alike) is
their AC behavior. A piezoelectric material is unable to hold its charge output due to a static input
because the charge will always leak through a high impedance closed circuit. In other words, it
only senses dynamic events and thus cannot be used to measure DC acceleration. The design of
the charge amplifier electronics (whether integrated internal or external) define the low frequency
filtering effect on the measurement signal. With voltage mode accelerometers the filter characteristic
is fixed. With charge mode accelerometers, the external charge amplifier commonly allows the user
to alter the settings to control the AC filtering effects. The cut-on frequency of PE accelerometers
DC filters ranges from tenths of a Hertz to a few Hertz.
A variety of ceramic materials are used for accelerometers, depending on the requirements of
the particular application. All ceramic materials are forced to become piezoelectric by a polarization
process, known as poling. In poling, the material is exposed to a high-intensity electric field. This
process aligns the electric dipoles, causing the material to become piezoelectric. Unfortunately, this
process tends to reverse itself over time until it exponentially reaches a steady state. If ceramic is
exposed to temperatures exceeding its range or electric fields approaching the poling voltage, the
piezoelectric properties may be drastically altered or destroyed. Normal accelerometers should not
be exposed of temperatures exceeding 120 C to avoid this effect. Accumulation of high levels of
static charge also can have this effect on the piezoelectric output.
A couple of mechanical configurations are available to perform the transduction principles of a
piezoelectric accelerometer. These configurations are defined by the nature in which the inertial
force of an accelerated mass acts upon the piezoelectric material. The shear configuration or the flex-
ural configuration is used in most modern accelerometers, see figure 11.11a. In shear mode designs,
the sensing crystals is bonded (sandwiched) between a center post and seismic mass. A compression
ring applies a preload force required to create a rigid linear structure. Under acceleration, the mass
causes a shear stress to be applied to the sensing crystals. By isolating the sensing crystals from the
base and housing, shear accelerometers are good at rejecting thermal transient and base bending
effects. Also, the shear geometry lends itself to small size, which minimizes mass loading effects on
the test structure. With this combination of ideal characteristics, shear mode accelerometers offer
very good performance to the expense of a more complex and costly design. Accelerometers built
with a flexural mode designs, see figure 11.11b, utilize beam-shaped sensing crystals, which are
supported to create strain on the crystal when accelerated. The crystal may be bonded to a carrier
11.3 Vibration testing hardware 223

beam that increases the amount of strain when accelerated. This design offers a low profile, light
weight and excellent thermal stability with a design that is less complex than the shear mode design.

Accelerometer Mounting Considerations. One of the most important considerations in

dealing with accelerometer mounting is the effect the mounting technique has on the accuracy of
the measured frequency response. There is always some compliance between the surface of the
test article and the accelerometer’s reacting piezoelectric material. This means that the motion
experienced by the test article surface is not identical to the motion of the sensing material, and
that effect is most notable in the high frequency range. A critical frequency is the fundamental
eigenfrequency of the accelerometer setup. That is the first natural frequency of the accelerometer,
mounted to rigid ground by the chosen mounting technique. Accelerometer output for frequency
ranges approaching or exceeding this frequency should never be trusted. The accelerometer
manufacturer determines the accelerometer’s operating frequency by securely stud-mounting the
test sensor directly to a high-quality reference accelerometer. This direct coupling technique, a
stud-mount to a very smooth surface, generally yields the highest fundamental frequency and,
therefore, the broadest usable frequency range. The addition of any mass to the accelerometer,
such as an adhesive or magnetic mounting base, lowers the fundamental frequency of the sensing
system and may affect the accuracy and limits the usable frequency range of the accelerometer.
Also, compliant materials, such as a rubber interface pad, can create a mechanical filtering effect by
isolating and damping high-frequency vibration.
For best measurement results, especially at high frequencies, it is important to do a proper
surface preparation. That can be done by preparing a smooth and flat machined surface where the
accelerometer is to be attached. One should also inspect the sensor interface area to ensure that no
metal burrs or other foreign particles interfere with the contacting surfaces. The application of a
thin layer of silicone grease between the accelerometer base and the mounting surface also assists in
achieving a high degree of intimate surface contact required for best high-frequency characteristics.
When it can be used, stud-mounting is recommended by most manufacturers. In stud-mounting,
one should grind or machine on the test object a smooth, flat area at least the size of the sensor base.
Then, a tapped hole perpendicular to the mounting surface should be made. Install accelerometers
with a mounting stud and make certain that the stud does not bottom in either the mounting surface
or accelerometer base. Any stud bottoming or interfering between the accelerometer base and the
structure inhibits acceleration transmission and affects measurement accuracy.
When installing accelerometers onto thin-walled structures, a cap screw passing through a
hole of sufficient diameter is an acceptable means for securing the accelerometer to the structure
by screw mounting. The screw engagement length should always be checked to ensure that the
screw does not bottom into the accelerometer base. A thin layer of silicone grease at the mounting
interface ensures better high-frequency transmissibility.
Most often, mounting by stud or screw is impractical, because the test article modifications
required. For such cases, adhesive mounting offers an alternative mounting method. The use
of separate adhesive mounting bases is recommended to prevent the adhesive from damaging
the accelerometer base or clogging the mounting threads. The type of adhesive recommended
depends on the particular application. Petro-wax, a synthetic form of beeswax, is generally supplied
with most of accelerometers and offers a very convenient, easily removable approach for room
temperature use. Two-part epoxies offer more stiffness, which maintains high-frequency response
and a permanent mount. Other adhesives, such as dental cement, hot glues and instant glues are
also viable options. Surface flatness, adhesive stiffness, and adhesion strength affect the usable
frequency range of an accelerometer. Almost any mounting method at low acceleration levels
provides the full frequency range of use if the mounting surface is very flat and the sensor is pressed
hard against the surface to wring out all extra adhesive. Generally, as surface irregularities or the
224 Chapter 11. Vibration Testing

thickness of the adhesive increase, the usable frequency range decreases. The less stiff, temporary
adhesives reduce the usable frequency range of an accelerometer much more than the more rigid,
harder adhesives. Generally, temporary adhesives are recommended more for low-frequency (<500
Hz) structural testing at room temperature. When quick installation and removal is required over a
wide frequency range up to 10kHz one should use stiffer adhesives suitable for more permanent
installations.
Magnetic mounting bases offer a very convenient, temporary attachment to magnetic surfaces.
For best results, the magnetic base should be attached to a smooth, flat surface. A thin layer of
silicone grease should be applied between the sensor and magnetic base, as well as between the
magnetic base and the structure.

Accelerometer Caveats. Although everyone would like them to be, accelerometers are not
perfect. Understanding the accelerometer’s errors is just as important as understanding how the
accelerometer works in an application. Accelerometer errors are due to various causes. Before
delivered to the end users, the accelerometers are calibrated by the accelerometer manufacturer.
A good practice is also to calibrate them before each test and do a post-calibration after the test
has been conducted to avoid that bad test data are progressed further down the line of its use. The
calibration result obtained is provided as a calibration constant which is unique for each individual
sensor. To it is associated a calibration error. That is because the accelerometer never have a
constant sensitivity to acceleration over its entire workable frequency range and the calibration
constant is the obtained value at one particular reference frequency only. The accelerometers
also have a non-linear signal-versus-acceleration characteristic. Also, varying temperatures affect
the accelerometer’s sensitivity to vibration. However, these effects are normally rather small
and quantifications of these effects are normally given by the manufacturers data sheet for the
accelerometer in question. A technical data sheet of a typical light-weight accelerometer is found
in table 6.1.
Other possible sources of error originated from alignment error and accelerometer cross-
sensitivity. Associated with the accelerometer is its principle axis. This is the axis for which the
accelerometer is most sensitive, meaning that for a given acceleration magnitude, the accelerometer
output signal is the strongest if it is subjected to that acceleration along its principle direction. Cross
axis sensitivity is the variation in the accelerometer’s output because of accelerations applied in
axes perpendicular to the principle axis of the accelerometer. That sensitivity should ideally be
zero, but in practice it never is. A high-quality accelerometer has cross-sensitivity in the range
1 to 3 percent of its principle axis sensitivity. That means that a strong transverse acceleration
component may give a signal contribution that is bigger than the contribution of a small acceleration
component in the principle direction. This cause crosssensitivity errors in the test data processing.
The associated alignment error is not an accelerometer deficiency but a user-induced error effect.
Alignment errors occur from imprecise attachment of the accelerometer such that its principle
direction is not oriented as was intended.

Weight 0.0005 kg Sensitivity ±15% 1.0 mV/(m/s2 )

Size (H × L ×W ) 3.6 × 11.4 × 6.4 mm Measurement range ±500 g (peak)
Frequency range (± 5%) 1 to 10,000 Hz Resonance frequency > 50 kHz
Frequency range (± 10%) 0.7 to 13,000 Hz Non-Linearity <1%
Temperature Range -54 to +121°C√ Transverse Sensitivity < 5%
Spectral Noise (100 Hz) 600 (µm/s2)/ Hz Shock Limit 10,000 g (peak)

Table 11.1: Specification of accelerometer of model 352C22 from PCB Piezotronics Ltd.

To the above, one should add the mass loading effect coming from the accelerometers and effects
 11.3 Vibration testing hardware 225

caused by accelerometer cables. Present day’s accelerometers are often very light-weight and the
dynamic characteristics of the test article thereby affected very little by adding the accelerometers
to it. For very light-weight structures or light structural parts subjected to testing for which the
relative weight of the accelerometers are of concern, the mass loading must be considered. For
very precise testing, also the effect of accelerometer cables must be minimized. Accelerometer
cables may swing freely and have local resonances that may affect the accuracy of the measurement.
Cables also add mass and damping, and if not properly fitted, they may tap on the structure while it
vibrates causing transient excitation strong enough to affect measurement data. For these reasons,
the cable motion should be noticed during test and proper action should be taken if something
troublesome happens with the cables.
When installing accelerometers onto electrically conductive surfaces, a potential exists for
ground noise pick-up. Noise from other electrical equipment and machines that are grounded to the
structure, such as motors, pumps, and generators, can enter the ground path of the measurement
signal through the base of a standard accelerometer. When the sensor is grounded at a different
electrical potential than the signal conditioning and readout equipment, ground loops can occur.
This phenomenon usually results in current flow at the line power frequency (and harmonics thereof)
and signal drift. Under such conditions, it is advisable to electrically isolate the accelerometer
from the test structure. This can be accomplished in several ways. The use of insulating adhesive
mounting bases, isolation mounting studs, isolation bases, and other insulating materials, such
as paper beneath a magnetic base, are effective ground isolation techniques. Be aware that the
additional ground-isolating hardware can reduce the upper frequency limits of the accelerometer.

11.3.2 Force transducers for vibration testing

Force transducer types. The force transducer is a simple form of piezoelectric transducer which
is in the load path from the shaker5 to the test article. When force is applied to this sensor, the
quartz crystals generate an electrostatic charge proportional to the Calibration and Validation of
Structural Dynamics Models Force transducers for Vibration Testing input force. This output is
collected on the electrodes sandwiched between the crystals. The force transducers are ideally
subjected to uniaxial loading along its preferred axis of operation and no couples. In practice, other
load components act on the unit as well but care should be taken to make such spurious loads
small. The force that it transmits, or a known part of it, is applied directly across the crystal which
thus generates a charge that is proportional to the loading. One important property of the force
transducer is the relative stiffness of the crystal and of the housing. The fraction of the load which
is transmitted through the crystal depends on this ratio. In addition, there exists the undesirable
possibility of cross-sensitivity which is also influenced by the design of the housing. Such cross
sensitivity make the transduced give a bias signal because of spurious shear loads and couples that
act on the unit.
The force indicated by the charge output of the sensor will always be slightly different from the
force applied by the shaker, and also from that transmitted to the test article. This is because of
inertia loading caused by acceleration of the small amount of mass that is distributed in the sensor.
The effect of sensor inertia is minimized if the sensor is attached properly to the structure. The
proper mounting is normally indicated by a marking put on the sensor.
Static loads cannot be measured by a piezoelectric force transducer. That is because the quartz
crystals of a piezoelectric force sensor generate an electrostatic charge only when force is applied
to or removed from them. However, even though the electrical insulation resistance is quite large,
the electrostatic charge will eventually leak to zero through the lowest resistance path. In effect, if
you apply and then hold steady a static force to a piezoelectric force sensor, the electrostatic charge
output initially generated will eventually leak back to zero. The rate at which the charge leaks back
5 Or hammer head in hammer impact testing.
226 Chapter 11. Vibration Testing

Figure 11.12: A picture and a schematic illustration of the cross-section of a typical quartz force
sensor.
........................................................................................

to zero is dependent on the lowest insulation resistance path in the sensor, cable and the electrical
resistance/capacitance of the amplifier used. In a charge mode force sensor, the leakage rate is
usually fixed by the impedance of the sensor cable and external signal conditioning unit used. In
an IEPE force sensor with built-in signal conditioning electronics, the resistance and capacitance
of the built-in electronics normally determines the leakage rate. When a rapid dynamic force is
applied to a piezoelectric force sensor, the electrostatic charge is generated quickly and the charge
leakage is insignificant. However, there is a point at which a slow speed dynamic force becomes
quasi-static and the leakage is faster than the rate of the changing force. That point is determined
by the time constant of the exponential discharging rate of the sensor. When leakage of a charge
occurs in a resistive capacitive circuit, the leakage follows an exponential decay. A piezoelectric
force sensor system behaves similarly in that the leakage of the electrostatic charge through the
lowest resistance also occurs at an exponential rate e−t/τ . The Discharge Time Constant (DTC)
τ is the circuit’s capacitance multiplied by its resistance. The DTC is also the time required for
the sensor to discharge its signal to 37% of the original value under steady loading. The same
physical characteristics holds for any piezoelectric sensor, whether the operation be force, pressure
or acceleration. The DTC of a system directly relates to the low frequency monitoring capabilities
of a system.
Unlike the low frequency characteristics of the sensor, which is determined by the sensors
electrical properties, the high frequency response is determined mechanically from the sensor
components. Each force sensor has an unloaded resonant frequency specification which should
be observed when determining upper linear limits of operation. The calibration constants of force
sensors is generally considered to be accurate up to 20% of this resonant frequency value. A
specification of a typical light-weight force transducer is given in Tab. 11.2.

Force Transducer Installation. Proper installation of sensors is essential for accurate dynamic

. .......................................................................................
Weight 0.031 kg Sensitivity (±15%) 112 mV/kN
Size (D × L) 16.5 × 32 mm Measurement range ±44 N
Frequency range (± 5%) 1 to 15,000 Hz Allowable load ±260 N
Low frequency limit (-5%) 0.1 Hz Non-linearity (full scale) < 1 %
Temperature range -54 to +121°C Temperature sensitivity < 0.054%/°C
Discharge time constant > 50s

Table 11.2: Specification of force transducer 221B01 from PCB Piezotronics Ltd.
 11.3 Vibration testing hardware 227

Figure 11.13: Speckle pattern in the interferometer with amplification and cancellation of direct
light from laser source together with back-scattered light from test article. Intensity pattern varies
with test-piece motion.
........................................................................................

measurements. Many force sensors are designed with quartz compression plates to measure forces
applied in an axial direction, aligning the sensor and contact surfaces to prevent edge loading or
bending moments in the sensor will produce better dynamic measurements. Having parallelism
between the sensor and test structure contact surfaces minimizes bending moments and edge
loading. Flatness of mounting surfaces will also affect the quality of the measurement. Using a thin
layer of lubricant on mounting surfaces during installation, creates better contact between sensor
and mounting surface. One other consideration when mounting force sensors is try to minimize
unnecessary mechanical high frequency shock loading of the sensors. The high frequency content
of direct metal-to-metal impacts can often create short duration overloads in structures and sensors.
This problem can be minimized by using a thin damping layer of a softer material on the interface
surface between the structure and sensor being impacted. It should be considered beforehand
whether the slight damping of the high frequency shock is critical to the force measurement
requirements.

11.3.3 Laser Doppler vibrometry

There is increasing use of laser transducers in vibration testing. The major advantages of these are
that they are unintrusive and do not cause any mass loading to the test piece. There are various
optical techniques available, but the most readily available laser transducers are those based on the
Laser Doppler Velocimeter (LDV) concept. This technique has been exploited into basically two
different concepts, the single-point stationary version and the scanning version (SLDV).

The Laser Doppler Velocimeter. The basic LDV transducer is a device which is capable of
detecting the instantaneous velocity of the surface of a structure. The velocity measurement is
made by directing a laser beam at he target point and measuring the Doppler shifted wavelength of
the reflected light which is returned from the moving surface. The Doppler shift is sensed by an
interferometer and the technique is described more fully below. The measurement made is of the
velocity of the target point along the line of the laser beam. The main requirements of the LDV
is that there is a free line of sight to the target measurement points and that the target surface is
capable of reflecting the laser beam adequately. The reflection should be diffuse, so mirror-like
surfaces and light-absorbing surfaces are bad. The specification of a typical device is: 0-250 kHz
frequency range, 10 µm/s-20 m/s vibration velocity range, 0.2-30 m target distance and 1-1000
mmV/s sensitivity.
The main limitations of the LDV as a general-purpose response transducer are the line of sight
requirement, that only vibration along the line of the laser beam can be measured, and the problems
associated with optical noise. The optical noise is associated with the sensing of the speckle pattern,
228 Chapter 11. Vibration Testing

Figure 11.14: Left; a scanning laser doppler vibrometry (SLVD) system that can be set up to
measure vibrations in 3D at multiple visual points that it roves over in a scanning fashion. Right;
two SLVD units on industrial robots to automate much of the measurement setup.
........................................................................................

see figure 11.13, in the interferometer that compares the light of a reference laser beam to that of
the laser beam that is reflected by the moving target surface. The speckle noise result in occasional
signal drop-outs and signal spikes that have to be rejected from the data acquired using this device.
However, against these disadvantages are distinct advantages for measurements which have to
be made in hostile environments, especially in hot-surface measurements for which conventional
transducers cannot be used.
In conventional LDV on market today, the LDV laser beam can be directed towards its target
point either manually or by remote control of positioning mirrors. For positioning control, two
mirrors are integrated into the LDV instrument to direct the laser beam without moving the
instrument. By remote control from a computer, the two mirrors can be set to positions to vary
the elevation and azimuth angles of the laser beam. Normally, the variation can be made within
a ±20°range. A typical application of this device is measuring at a large number of points of a
surface mesh covering part of the test article.

The Scanning Laser Doppler Velocimeter. A natural extension of the capability described
above for directing the measurement laser beam is to incorporate a dynamic feature in the position-
ing mechanism. In the SLDV, the laser device is equipped with a laser beam mirror system that can
be controlled to move the laser beam to scan along a given path. This means that we can exploit the
ability to locate the laser beam direction on demand by devising a scanning process which moves the
beam from one measurement point to the next in a controlled way. In its simplest form, the SLDV
simply moves the beam to the first measurement point, makes a measurement, and then moves to the
next measurement point and repeats that process. The faster this stepping or scanning can be done,
the shorter will be the total measurement time. However, the speed of such a procedure is limited by
a number of factors such as those concerning the time required to dwell at a measurement point in
order to have sufficient information to characterize its behavior. Other factors are those determined
by the physical limitations inside the instrument, such as the inertia of the mirrors which must be
moved and come to rest in order to bring about the desired change of location of the laser beam. In
effect, the latter factors constitute a major barrier to faster measurements of this type, especially at
high frequencies of vibration. The former factors needs to be considered at low frequency vibration.

Basic Principle of Laser Vibrometry. A mono-frequency wave emitted from a stationary

source that is reflected by a moving object and frequency-detected again at the source experiences a
frequency shift known as stemming from the Doppler effect. The frequency shift can be described
11.3 Vibration testing hardware 229

Figure 11.15: A photo of an LDV laser unit and a schematic illustration of the laser beam paths in a
laser doppler vibrometer.
........................................................................................

as

fD = 2v/λ (11.26)

where λ is the wavelength of the emitted wave. To be able to determine the velocity v of an object,
the Doppler frequency shift has to be measured at a known wavelength. This is done in the LDV by
using a laser interferometer. It exploits the physics of optical interference, requiring two coherent
light beams, with their respective light intensities I1 and I1 , to overlap. The resulting intensity is not
just the sum of the single intensities, but is modulated according to the formula
p
Itot = I1 + I2 + 2 I1 I2 cos 2π(r1 − r2 )/λ (11.27)

with a so-called interference 3rd term. This interference term relates to the path length difference
r1 − r2 between both laser beams. If this difference is an integer multiple of the laser wavelength,
the overall intensity is four times a single intensity. Correspondingly, the overall intensity is zero if
the two beams have a path length difference of half of one full wavelength.
The image in figure 11.15 shows how this physical law is exploited technically in the LDV.
The beam of a helium neon laser is split by a beam-splitter (BS1) into a reference beam and a
measurement beam. After passing through a second beam-splitter (BS2), the measurement beam is
focused onto the test article, which reflects it. This reflected beam is now deflected downwards
by BS2 and merged with the reference beam by the third beam splitter (BS3) and is then directed
onto the interference detector. As the path length r2 of the reference beam is constant over time, a
movement r1 (t) of the object under investigation generates a dark and bright fringe pattern typical
of interferometry on the detector, see figure 11.13. One complete dark-to-bright cycle on the
detector corresponds to an object displacement of exactly half of the wavelength of the light used.
In the case of the helium neon laser this corresponds to a displacement of 316 nm.
Changing the optical path length per unit of time manifests itself as the Doppler frequency shift
of the measurement beam. This means that the modulation frequency of the interferometer pattern
determined is directly proportional to the velocity of the object. As object movement away from
the interferometer generates the same interference pattern and frequency shift as object movement
towards the interferometer, this setup cannot determine the direction the object is moving in. For
this purpose, a modulator known as a Bragg cell is placed in the reference beam, which shifts the
light frequency by 40 MHz (by comparison, the frequency of the helium neon laser light is 474
THz). This generates a modulation frequency of the fringe pattern of 40 MHz when the object is at
rest. If the object then moves towards the interferometer, this modulation frequency is reduced and
if it moves away, the detector receives a frequency higher than 40 MHz. This means that it is now
 230 Chapter 11. Vibration Testing

possible not only to detect the amplitude of movement but also to clearly define the direction of
movement.

11.3.4 Hardware Calibration

As with all measurement processes, it is necessary to calibrate the equipment which is used. In
the case of vibration testing, there are two levels of calibrations that should be made. The first of
these is a periodic calibration of individual transducers to check that their sensitivities remains
the same as those specified by the manufacturer. Any marked deviation could indicate internal
damage which is insufficient in magnitude to cause the device to fail completely, but which might
nevertheless constitute a loss of linearity or repeatability which would not necessary be detected
immediately. The second type of calibration is one which can and should be carried out during
each test, preferably twice - once at the onset and again at the end of the test. This type of
calibration is one which provides the overall sensitivity of the complete instrumentation system
without examining the performance of the individual elements.
The first type of calibration is quite difficult to make accurately. As in all cases, the absolute
calibration of a transducer of a transducer or a complete system requires an independent measure-
ment to be made of the quantity of interest, such as force and acceleration, and this can be quite
difficult to achieve. The use of another transducer of the same type is seldom satisfactory as it is not
strictly an independent measure, except in the case of using a reference transducer which has been
produced to have very stable and reliable characteristics and has previously been calibrated against
an accepted standard under strictly controlled conditions. Other means of making independent
measurements of motions is through displacements, which are generally confined to optical devices
and these are not widely available. Independent methods of measuring force are even more difficult
to obtain.
As a result, absolute calibration of transducers is generally undertaken only under special
conditions and is most often performed using a reference accelerometer, both for accelerometers
and - with the aid of a precisely known mass - of force transducers. If legislation or quality assurance
policies require so, transducer calibration need to be made by an accredited test laboratory or sensor
manufacturer.
One reason why the absolute type of calibration has not been further developed for the ap-
plication of FRF testing is the availability of a different type of calibration procedure which is
particularly convenient. With few exceptions, the parameters measured in modal testing are ratios
between response and excitation levels, such as mobility FRF or accelerance FRF. So what is
required is the ability to calibrate the whole measurement system including response and excitation
sensors. The quantities actually measured are usually two voltages, one from the force transducer
and its associated electronics and the other from the response transducer. These voltages are related
to the physical quantities being measured by the sensitivities of the respective transducers. The
voltage from the force transducer subjected to the force f is Uf = αf f and the voltage from the
accelerometer subjected to the acceleration a is Ua = αa f . As mentioned above, there is some
difficulty in determining values for the calibration coefficients αf and αa individually but we note
that, in practice, we only ever use the measured voltages as a ratio to obtain the frequency response
function

a/ f = (αf /αa )(Ua /Uf ) (11.28)

and so what is required is the ratio of the two sensitivities αf /αa . This overall sensitivity can
be more readily obtained by a calibration process because we can easily make an independent
measurement of the quantity now being measured - the ratio of acceleration response to force. If we
undertake an accelerance measurement on a simple rigid mass-like structure, the result we should
obtain is a constant magnitude over the frequency range at a level which is equal to the reciprocal
 11.3 Vibration testing hardware 231

Figure 11.16: A simultaneous acceleration and force calibration setup using the known weight of a
calibration mass.
........................................................................................

of the mass of the calibration block, a quantity which can be accurately determined by independent
weighing.
Figure 11.16 shows a typical calibration block setup which can be used to record the simultane-
ous output voltages of the accelerometer and force transducer. These can then be used to convert
the measured values of (voltage/voltage) to those of (acceleration/force). The scale factor thus
obtained should be checked against the corresponding value computed using the manufacturers’
stated sensitivities and amplified gains to make sure that no major errors have been introduced and
to see whether either of the transducers has changed its sensitivity markedly from the nominal value.
In practice, this check need only be made occasionally as the approximate scale factor for any given
pair of transducers will become known and so any marked deviations will be spotted quite quickly.
A calibration procedure of this type has the distinct advantage that it is very easy to perform
and can be carried out in situ with all the measurement equipment in just the same state as is used
for the FRF measurements proper. In view of this facility, and the possibility of occasional faults
in various parts of the measurement chain, frequent checks on the overall calibration factors are
strongly recommended. As mentioned at the outset, the ideal situation is that this is made at the
beginning and end of each test.

11.3.5 Some Recent Hardware Trends in Vibration Testing

Integrated Electronics PiezoElectric - I EPE. The requirements of signal conditioning in extra
hardware has been eliminated with the introduction of a technique called I EPE6 . I EPE sensors has
the signal condition built-in and is fed by a constant current of the DAQ to drive its electronics. All
I EPE sensors require electrical power by a constant current in the range of 2-20mA over a 2-wire
cable. As the sensor is subjected to stimulus, it reacts by an alternating apparent impedance that
cause the voltage over the sensor to change. This varying voltage is the signal that the DAQ picks
up. I EPE has been adopted as the standard by the industry’s sensor, analyzer and data acquisition
manufacturers.

Transducer Electronic Data Sheet - T EDS The basic function of T EDS capability is to provide
a standardized means of digital communication in what was formerly the analog-only measurement
channel. A T EDS accelerometer can report standard information like: manufacturer, model number,
serial number and calibration value. It can also report detailed information as its frequency response
transfer function, or user supplied data such as orientation direction, vector component and sensor
node name. The digital T EDS communication is accomplished by a polarity swapping of the
I EPE constant current feed that allows a standard 2-wire I EPE accelerometer to toggle into digital
6 The I EPE technique is also called the ICP technique by PCB Piezotronics Inc., the company that developed it.
232 Chapter 11. Vibration Testing

Figure 11.17: (a) M EMS acceleration sensor, and (b) a cubic shaped T RIAX accelerometer.
........................................................................................

communication mode when an appropriate negative supply current is introduced. With normal
polarity the accelerometer responds for accurate analog measurements, but with polarity swapping
the sensor responds with a digital T EDS data stream of predetermined size and format. Just like
analyzers in the sound and vibration market have evolved to include the constant current supply for
I EPE operation, most modern dynamic DAQ systems now include the minimal extra system design
to handle the T EDS feature.

Micro-Electro-Mechanical Systems - M EMS. Most acceleration transducers are made from the
mechanical bonding of discrete components, which are normally manually assembled one at a
time. However, some manufacturers use techniques borrowed from semiconductor manufacturing
to make mass-produced monolithic sensors from silicon. The result is referred to as M EMS. M EMS
accelerometers can be created using mass removal techniques to get useful shapes exploiting the
crystallography of silicon and chemical etchants. The powerful etching equipment in the semicon-
ductor industry provide precise material removal with microscopic dimensions. This capability,
plus the convenient fact that resistors implanted in silicon are by nature piezoresistive, allows an
entire transducer to be created with dimensions on the order of a millimeter, see figure 11.17a.
M EMS accelerometers can be made to measure accelerations in frequencies down to statics (DC)
and are increasingly present in portable electronic devices such as smartphones and video game
controllers but is not yet common in vibration testing.

Triaxial Accelerometers - T RIAX. By far, the most popular configuration of piezoelectric

accelerometers for high channel-count DAQ system users has in recent years become the integral
triaxial accelerometer (T RIAX), see figure 11.17b. The T RIAX is made to simultaneously sense
the 3D acceleration field it is subjected to by registering the three orthogonal linear acceleration
components. The implementation of a single 4 pin sensor connector (three signals and one common
ground) allows for almost a factor of 3 reduction in cabling. That is a huge benefit to many
users who may use hundreds of channels for standard vibration tests in automotive and aerospace
industries. In addition to the reduction in cabling, the integral cubic T RIAX package provides
practical and versatile mounting to any of five available surfaces while the sixth surface fields the
output connector. This allows for convenient transducer mounting and cable routing regardless of
required location or orientation. Recent efforts to de-facto standardization of T RIAX has also been
the continual drive toward smaller, lighter and more robust designs mainly utilizing titanium in the
housing. At the extremes, there are 6.5 mm T RIAX I EPE accelerometers with 5 mV/g sensitivities
weighing a single gram.
A potential benefit of a T RIAX is that the cross-sensitivity effect can be reduced. Let the signal
from three independent piezoelectric crystals be s1 , s2 and s3 and the three orthogonal accelerations
the sensor experiences be ax , ay and az . The signals are related to the acceleration through the
 11.3 Vibration testing hardware 233

sensitivity matrix S such that

    
s1  Sxx Syx Szx ax 
s2 = Sxy Syy Szy  ay , S a (11.29)
s3 Sxz Syz Szz az
   

By calibration, the sensitivity matrix elements can be determined and compensated for by electronic
filtering using the scaled inverse αSS −1 of the sensitivity matrix S such that signals s01 , s02 and s03 and
and can be achieved that are insensitive to cross-acceleration as
 0    
s1  ax  ax 
s02 = αSS −1 S ay = α ay (11.30)
 0
s3 az az
   

This is a feature that cannot be obtained by uniaxial accelerometers.

Efforts has been made to also measure the angular acceleration at the sensor location with a
single sensor device. Such efforts have not yet found commercial success7 .

11.3.6 Some future perspectives

New technologies are emerging for non-contact vibration measurement. The laser doppler vibrom-
eter has been used in vibrations testing since the 1990’s and now high speed cameras and rapid
image processing are evolving. These are utilizing high speed video stroboscope and multiple
cameras. This permits the simultaneous measurement of all three displacement axes on the entire
visible surface of a specimen. The technique uses white-light speckle correlation and the actual
object movement is measured and the in-surface strain becomes available at every point on the
surface. It can cope with large vibration amplitudes and can be employed up to several kHz. It
can thus be used as a multi-axis vibrometer. The specimen needs to be prepared with a random
pattern, e.g. using spray paint, and therefore no significant mass is added to the object under test.
The accuracy of the measured data, and the robustness of the technology are still question marks.
As all points of interest cannot be covered by non-contacting optical measurements there will
still be a need for accelerometer transducers that can also be mounted at locations to which it is hard
to get access. One hardware manufacturer, the Brüel&Kjaer company, predicts that the future of
sensors will be affected by the development of micro-mechanical sensors, the evolution of batteries
and wireless communication. The author’s prediction is that we can await lightweight, wireless
accelerometers that can register accelerations in the full 0-10,000 Hz range, and which might be
charged by the vibration itself using the vibration energy harvesting technique. The sensors might
also communicate its identity data to the DAQ and also give calibration data, self diagnostics and
maybe even its position in space.

7 Itis noteworthy to recall that the evolution has provided animals and humans with an angular motion sensing device
by the labyrinth, located in our inner ear, that provide rotatory motion sensing through fluid motion in (more-or-less)
orthogonal fluid-filled circular canals.
 234 Chapter 11. Vibration Testing

Figure 11.18: A camera-based deformation measurement system for materials and product testing.
........................................................................................

11.4 Problems
Problem 11.1 Circle fit for damping estimate
In a vibration experiment one has been using 0.05Hz frequency steps in a stepped-sine test. One has
obtained system mobilities, and in particular Y j j (ω). In the phase-plane for the mobility one has
found what appears to be the system’s second resonance frequency at about 25.7Hz. One has found
a circle of radius 0.053m/Ns with center at coordinates (0.053, -0.0048i) m/Ns to fit the mobility
data. The experimentally obtained mobilities at frequencies around the resonance are listed in
the figure. They are associated with experimental frequencies starting at 25.5Hz. Determine the
system’s modal damping at the second eigenfrequency.

12/01/2012-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III
Model Calibration and
Validation

12 Validation and Calibration Concepts 239

12.1 Models and model structures
12.2 Problems

13 Property Variability and Model Uncer-

tainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
13.1 Real world variability
13.2 Parameter estimation statistics
13.3 Model structure selection

14 Model Calibration Procedures . . . . . . . 275

14.1 Minimizing a single-variable-function
14.2 Minimizing a quadratic functional
14.3 Computational aspects

15 Validation and Cross-Validation . . . . . 283

15.1 Classical validation
15.2 Cross-validation

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
 12. Validation and Calibration Concepts

Because of many, far from trivial, modeling issues in computational structural dynamics, the
model validation is most often made in conjunction with a model calibration. Model calibration
of structural dynamics models has been a very active research topic since the 1980’s. Model
validation is also known as model updating1 or system identification2 . In the mid 1990’s a research
program on FEM updating was started, called the COST Action, that involved many European
universities. In an early description of the COST Action it was stated that; For fifty years, the finite
element method has been developed for structural computations. It is now currently used in industry
for the modeling and the calculation of mechanical components. For the same period, vibration
measurement devices and experimental modal analysis techniques have known large improvements
due to hardware and software developments. The correlation between theoretical computations
and experimental data has become a sensitive problem for the industry sector. In most cases, the
interaction between theory and experiment is limited to the simple comparison of computation
and measurement results. Thus correcting the model in order to get computation results closer to
experimental data is usually done by "hand" adjustments using a heuristic approach. During the
last twenty years, attempts have been made to correlate rather than simply compare experimental
and theoretical results. In this context, it becomes important to develop European collaboration
and to intensify and coordinate research in the field of structural testing, dynamic analysis and
model updating. The COST Action ended without no real consensus about what methods that could
be recommended for FEM calibration.
In model validation with model calibration, the validation process is made in successive steps.
The first major step is to prepare the calibration and validation tests using the prediction capacity of
1 The use of model updating is avoided in this book since its meaning is ambiguous in the industrial process of

modeling. The finite element models are normally updated numerous times during the model creation as new and better
data become available from the suppliers of system components and materials. It is probably also updated after test data
have become available, which is the process of model calibration in the meaning used in this book.
2 In system identification, the system model parameters are determined from test data in the same meaning as in

model calibration. However, in system identification a full arsenal of models are often tried to find out which model
that best fits the test data. The models range from fully black-box parameterization, with no meaningful connection to
physical properties, to white-box parameterization bases on first principle models. In this book we use model calibration
and system identification of white-box models as being equivalent.
240 Chapter 12. Validation and Calibration Concepts

Figure 12.1: The schematics of the model calibration and validation process.

initial models. The second step is then to conduct the validation and calibration tests. It should be
said here at the outset that the validation and calibration test data must be kept separate and same
data should never be used both for calibration and validation. The next step is to do a post-test
screening of the test results to find possible test outliers and cure the test data such that bad test
data does not enter the validation and calibration process. The following major step is to do model
validation. If the model is here found valid, no more steps are required and the model can be used
with confidence for its intended purpose. However, if the model is found to be invalid, the next step
is to conduct a model calibration. In this step, uncertain model parameters that affect the system’s
mass, stiffness and damping, are calibrated such that the model predictions best match the test data.
The calibrated model is then tried against the validation data to find out about whether the calibrated
model is now valid for its intended purpose or not. If so, the calibration and validation process
has a happy ending. If not, the model has to be re-worked, more uncertain parameters added for
calibration or the validation criterion needs to be re-visited to allow for stronger model inaccuracies
which then can be compensated for by setting higher safety factors in the design process. The
overall process is illustrated in figure 12.1.
Even though the considered models in this book are linear, many hard problems still awaits
the analyst that want to do a finite element model calibration. One is that most structures are
multi-modal, which means that a high model order is required for the model to fit data. Another is
that the model structure is fixed by the structure of the verified finite element model, which restricts
the flexibility of the model to fit test data. Yet another problem is that adjustment to better fit to
data is restricted to adjusting physically related parameters that is part of the first principles physics
 12.1 Models and model structures 241

on which the finite element model is based. A free mathematical parameterization, with no relation
to physical properties, normally gives a better fit but with the downside that the resulting model
is less useful to the modeler. A third major problem is that a finite element model often holds a
large number of more-or-less unknown parameters. The calibration of all these may be a task that
is too daunting and the combination of test signal noise, limited sensing capability and imprecise
calibration metric will prohibit a meaningful calibration when the calibration parameters become
too many. Such matters, and others, are discussed in the following.

12.1 Models and model structures

A model is the conceptual, mathematical, and possibly numerical representation of the physical
phenomena needed to represent real world conditions and scenarios. Thus, for mechanical struc-
tures modeled by the finite element method, the model includes the geometrical representation,
constitutive models and related material parameters, governing differential and constraint equations,
boundary and initial conditions, together with system loading. A model is a particular realization of
parameter settings within a given model structure. In the finite element modeling context the model
structure can best be described by what is given as the input file to the program doing the finite
element computations. The numerical data of the input file can be seen as values of a particular
realization creating the model that is run by the analysis program. With a given input file we fix the
model structure, but if we see the numerical data as the particular realization of free parameters we
can do an almost infinite number of realizations within that structure. As an example we can take
an excerpt from a NASTRAN input data deck that defines the shell element properties (including
the thickness t=0.0026m) and material parameters (including the Young’s modulus E=7GPa and
density ρ = 1600kg/m3 ) for a linear isotropic material:

PSHELL, 1, 1, 0.0026, 1, 1, 0.
MAT1, 1, 7.E+9, 0.3, 1600., 0., 0.

If we want the plate thickness, Young’s modulus and density to be three free parameters, p1 , p2
and p3 , in a model created as a realizations from this given model structure we can specify these as:

PSHELL, 1, 1, P1, 1, 1, 0.0

MAT1, 1, P2, 0.3, P3, 0., 0.

To re-run the same input file with another realization of thickness and modulus of elasticity, we
can then easily edit the file, replace the symbolic parameters with numerical values and run again.
A specific class of model structures is the so-called consistent model structure. This is the
model structure for which we can have a perfect fit to test data with an exact representation of the
underlying parameters. This is also called the true model structure and sometimes the oracle model
structure and the optimal parameter setting of this defines the true model or the oracle model. Since,
for real problems with test data as observations of reality, the parameters and model structure are
always hidden to us. The consistent model structure thus has only theoretical value in this context.
However, methods for calibration and validation are often tried out first in a simulation environment
in which a known model and its response data are taken as substitute for testing. For that model, the
oracle model structure is known. If one develops methods and criteria for calibration and validation,
that fails when tested against synthetic data with a consistent model structure, it will certainly not
be trustworthy in the calibration and validation context with real test data. Examples of consistent
and inconsistent model structures are seen in figure 12.2.
A model can be used for prediction and simulation. In predictions we use the model to get
an extrapolation in time from given time history data. The predictions are normally valid for
 242 Chapter 12. Validation and Calibration Concepts

short durations of time. In a discrete time analysis we do predictions over a finite number of time
samples. If we use the model to predict the system response for the n future time steps we do a
n-step prediction. On the other hand, in simulation we simulate the response over the compete time
domain of interest setting out from a given initial state. This may be a considerable time involving
numerous time steps. In simulation, it is required that the model and the associated numerical time
integration scheme are stable or the response solution will diverge which normally introduces large
simulation errors. That stability is not required for the prediction. The prediction accuracy can still
be good also for unstable models integrated with unstable integration methods.

12.1.1 Model validation and falsification

Model validation is the process of making credible that the models that are used for simulation
or prediction are sufficiently accurate for decision making. It is the process of doing comparative
studies between results of test and analysis. Model validation assesses the degree to which
the computational model is an accurate representation of the physics being modeled. In model
validation, the deviation of the outcome of a simulation or prediction is compared to the outcome of
one or multiple experiments. The tests are conducted on one or more test articles that are samples
of the system that is modelled. Since test results always spread when the same test procedure is
repeated, we cannot define the test results to be “correct” or “the truth”. We thus cannot define
an error based on test data. We can just assume that the test data well represents the physics of
the problem and to which we can measure the deviation of the outcome from our model. Model
validation is thus based on comparisons between numerical simulations and relevant experimental
data. Validation must asses the predictive capacity of the model in the physical realm of interest,
and it must address uncertainties that arise from both experimental and computational procedures.
Together with the model verification process, see section 1.4, it constitutes the Verification and
Validation (V & V) process which is the processes by which evidence is generated, and credibility
is thereby established, that computational models have adequate accuracy and level of detail for
their intended use. In short, we make the following definition:
Definition 12.1.1 — Model Validation. Model Validation is the process that substantiate that
a computational model, within its domain of applicability, possesses a satisfactory range of
accuracy in consistency with the intended use of the model. The deviation between the model
and the test article(s) should then be in agreement with what is prescribed by a suitable validation
criterion.

In the practical implementation of the V&V process it should ideally begin with a statement
of the intended use of the model. That should be made so that the relevant physics are included
in both the model and the experiments performed to validate the model. Modeling activities and

........................................................................................

Figure 12.2: Illustration of two small finite element beam models, (a) and (b), of same beam system
but with two model structures. In (a) the two beam elements have same Youngs’s modulus and
density but in (b) they have individual properties. Model (b) is consistent with (a) but (a) is not
consistent with (b). This is because (b) can mimic (a) with E1 = E2 = E0 and ρ1 = ρ2 = ρ0 but (a)
cannot mimic (b) if E1 differs from E2 or ρ1 differs from ρ2 .
12.1 Models and model structures 243

experimental activities are guided by the response features of interest and the accuracy requirements
for the intended use. Experimental outcomes for component level, or sub-system level tests should,
wherever possible, be provided to modelers only after the numerical simulations for them have
been performed with a verified model. For a particular application, the verification and validation
process end with acceptable agreement between model predictions and experimental outcomes
after accounting for uncertainties in both, allowing application of the model for the intended use. If
the agreement between model and experiment is not acceptable, the process of V&V is repeated by
updating the model and performing additional experiments.
A detailed specification of the model’s intended use should include a definition of the accuracy
criteria by which the model’s predictive capability will be assessed. The accuracy criteria should be
driven by the application requirements for the intended use of the model. Although accuracy criteria
and other model requirements may have changed before, during, or after validation assessments
of the entire system, it is best to specify validation and accuracy criteria prior to initiating model
development and experimental activities in order to establish a basis for defining “how good is
good enough?” with the model being a good-enough model being a validated model.
A model is normally developed with one purpose, or a set of purposes, in mind. Its validity
should be determined with respect to that. If the purpose of the model is to answer a variety of
questions, the validity of the model needs to be determined with respect to each question. Numerous
sets of experimental conditions are usually required to cover the domain of a model’s intended
use. A model may be valid for one set of experimental conditions and invalid in another. A model
is considered valid for a set of experimental conditions if the model’s deviation from testing is
within its acceptable range, which is the amount of accuracy required for the model’s intended
use. If the variables of interest are considered to be random, then properties and functions of the
random variables such as means and variances are usually what is of primary interest and are what
is used in determining model validity. Several versions of a model are usually developed prior
to obtaining a satisfactory valid model. This is usually made in a calibration process in which
the model parameters are adjusted to data that is not used for validation. The full substantiation,
including both verification and validation, is usually part of the total modeling process.
It is often too costly and time consuming to determine that a model is fully valid over the
complete domain of its intended use. In some cases it would even be dangerous or illegal, since the
intended model use might be to predict what would happen in hazardous or catastrophic situations.
Examples of such conditions are brutal weapon attacks, large magnitude earthquakes and failure of
critical sub-systems. Instead, tests and evaluations are conducted within a safe condition envelope
until sufficient confidence is obtained that a model can be considered valid for its intended use also
by means of extrapolation. If a test reveals that a model does not have sufficient accuracy for any
one of the experimental conditions, then the model is fully falsified. However, determining that
a model has sufficient accuracy for numerous experimental conditions does not guarantee that a
model is valid everywhere in its planned usable domain.
Also, the cost of doing validation needs to be considered. Figure 12.3 shows the relationships
between model confidence and cost of performing model validation and the added model value to
the user. The cost of model validation is usually quite significant, especially when extremely high
model confidence is required. If we take a side-view to another field of physics, the particle physics
the Standard Model is the name given to the current theory of fundamental particles and their
interaction. The Standard Model is good since it has been validated with prediction capability with
amazing precision, decimal place after decimal place. All the particles predicted by this theory have
been found in reality. However, the validation to such precision has been made to an extreme cost
and cannot be expected to happen in structural dynamics applications. However, faster computers
and better test equipment will take us a long way. In present days, areas in which model validation
plays an important role are in aerospace, automotive and power generation.
 244 Chapter 12. Validation and Calibration Concepts

Figure 12.3: Schematic picture of the relation between the extra cost taken to provide an added
value to the user.
........................................................................................

A model which cannot be validated is said to be falsified. To be in a process of falsification

is to be in another state-of-mind than being in a process of validation. If one is in the process
of validation, the positive outcome is a validated model. If one is in the process of falsification,
the positive outcome is a falsified model. If one fails to do a positive falsification, one has on the
other hand in some respect validated the model. Some people believe that to be in the falsification
state-of-mind without success of falsifying may end up in a better validated model than if one tries
to validate the model in the first place. One recently proposed verification and validation technique
(Active Nonlinear Tests - ANT) is explicitly formulated as a series of mathematical tests that are
designed to break the model. A serious but failed attempt to break the model will make the model
that could withstand such attempts more credible.

12.1.2 Model calibration

Model calibration is at the hart of model validation and calibration. It is important since many
models fail to pass a validation test without it. This is since a model is seldom in accordance
with test results when all parameter values set to their nominal values. Model calibration is also
the most challenging task of model validation and calibration. While validation is a straight-
forward evaluation of a validation metric based on deviations between test and model outcome,
model calibration is more difficult. Model calibration, also called model updating, model training,
parameter estimation or system identification, requires the solution of a parameter identification
that is non-linear in the parameters. It often amounts to solving a non-linear least-squares problem
formulated in the vector of model parameters p as

p ? = arg min Q(pp) with Q(pp) = δ T δ (12.1)

p

where δ = δ (pp) is a vector-valued deviation metric that quantifies the deviation between results
from simulation and testing.
From linear regression theory we know that when we have as many parameters as we have
observations, we can formulate an equation system with a unique solution which gives a model
with a perfect fit to all observations. That is provided that the equations we formulate for each
observation are linearly independent. In linear regression we also know that the model we develop
can be subjected to over-fitting. The classical case is in the fit of a polynomial model to noisy data.
With N noisy data samples we can achieve a perfect fit with a N:th order polynomial. However that
polynomial has extremely poor prediction capability for interpolation between samples, and is thus
over-fitted.
 12.1 Models and model structures 245

For a non-linear parameter estimation problem, e.g. the calibration problem, there is no guaran-
tee that there is a zero residual solution giving Q(pp? ) = 0. We can only expect Q(pp) to have a global
minimum which we ultimately search for in the calibration process. In structural dynamics model
calibration, the criterion functions that has been developed up until now are usually plagued by local
minima that makes it hard to find the global optimum. The minimization must rely on non-linear
optimization algorithms and involves finite element computations to evaluate the criterion function
as the search for the optimum progresses. This makes the problem a huge computational task and
we cannot expect reasonable computational times if the number of parameters are large, say more
than a few tens.
A central feature in parameter estimation is the parameter identifiability from test data. If the
criterion function is very insensitive to parameter variation of one or more parameters, and the test
data has variability because of noise, the parameters will be estimated with a large variance when
using various realization of the noise. What good does it do to know that a parameter is estimated
to have a mean value at p̄p with a large standard deviation’s coefficient of variation of, say, 400%?
Not much, and one task in model calibration is therefor to find a model parameterization that is
relevant in the sense that it contains a small set of the most important parameters and that these are
properly identified with small variance from test data.

12.1.3 Model verification

Model verification is about assessing modelling errors. It assesses the numerical accuracy of a
computational model, irrespective of the physics being modeled. Both code3 verification, that
address potential errors in the software, and calculation and algorithm verification, that estimates the
numerical errors due under-resolved space and time discrete representations, of the mathematical
partial differential equation model are addressed. In short we say that:
Definition 12.1.2 — Model Verification. Model Verification is the process of determining that
a computational model, including the code that runs it, are free from errors such that it accurately
represents the underlying mathematical model and its perfect solution.

The verification is done to ensure that the model is programmed correctly, the algorithms
have been implemented properly and that the model does not contain errors, oversights or bugs.
Verification does not ensure that the model solves an important problem, meets a specified set of
model requirements or correctly reflects the workings of a real world process.
In this book we assume that the codes that we run and the models we use have already been
verified. We do this, but we also acknowledge that this is an ideal situation that seldom occurs
in practice. Even the best codes are seldom free from errors. The correction-of-error lists that is
supplied with new code versions and the disclaimer notes of commercial codes are evidence of that.
We know that serious code developers do their best in their quality assurance process and that code
user feedback helps them to find and correct bugs. For most commercially available and commonly
used finite element packages it is however fair to say that they can be used with a high level of trust.
The modeler also has to do his job of doing the proper convergence assessment of the model such
that the computed results does not depend on the selected finite element mesh density, the chosen
time-steps for time integration, etc.
We should note that it is a realistic assumption that:
a) no computational model will ever be fully verified, guaranteeing 100% error-free implementa-
tion,
b) a high degree of statistical certainty is all that can be reached, and that is gained as more cases
are executed to cover important modeling features,
3 code: the computer implementation of algorithms developed to facilitate the formulation and numerical solution of a

class of problems
 246 Chapter 12. Validation and Calibration Concepts

c) model verification develops as more tests are performed, errors are identified, and corrections
are made to the underlying model, often resulting in re-execution requirements to ensure code
integrity,
d) the end result of verification is technically not a verified model, but rather a model that has
passed all the verification code runs.

12.1.4 Experimental support for validation and calibration

Even under the most controlled testing conditions, the test data gathered from the experiments will
not be free from noise. By noise we mean all contributions to the acquired test data that are due
to un-modelled physics. In vibration testing, the most obvious noise source is ambient vibratory
noise that affects the test article through the surrounding air/water as uncontrolled sound waves
or ambient uncontrolled vibration transmitted through the test article’s supporting structure. Such
uncontrolled noise sources are always present and may be due to ambient traffic, vibrating fans and
motors, walking people, etc. There is also electronic noise added to the test data acquired from the
sensors. Such may be due to electromagnetic radiation in the laboratory that gets into sensor cables
or from internal electronic noise in the circuits of the sensors and the test data acquisition system.
Vibratory testing is most often set up in a condition to mimic the free condition that would
have occurred if the test piece should have been floating in the zero gravity free space or in orbit.
At the other extreme, the support conditions of the test article are sometimes mimicking the fixed
condition that would ideally be due to a perfectly rigid support. None of these conditions can be
realized in practice in earth-bound experiments. There seems to be a consensus in the vibration
community that the free condition is the one that can be realized the truest in laboratory. For this,
the test piece is mounted on soft support of rubber isolators or gas filled rubber cushions, or hangs
in very flexible bungee chords. In all cases the test pice is mounted to a support system that neither
has an ideal zero receptance or an ideal infinite receptance. In the modeling, the dynamic properties
of the supporting structure and wave propagation in the surrounding system are normally not taken
account for. Such physics are thus un-modelled and normally cause bias effects in the calibration
activity.
In the following we split the calibration deviation δ and validation deviation γ in the parts that
are due to experimental bias δ B and γ B and regular noise δ N and γ N as

δ (pp) = δ B + δ N and γ (pp) = γ B + γ N (12.2)

In the model calibration and validation activity focus must be put on minimizing the bias which
otherwise will deteriorate the parameters from their correct setting. That is done by actions such as
careful isolation of the test article from its surrounding and by proper calibration of sensors. Also
the influence of regular noise adversely affect the outcome of model calibration in the sense that
large noise gives large variances of the parameter estimates. The influence of noise can be made
smaller by proper signal processing, which is the subject of a proceeding chapter.
The purpose of validation experiments is to provide information needed to assess the accuracy
of the computational model. Therefore, all assumptions should be understood, well defined and
controlled. To assist with experimental design, preliminary calculations are recommended. These
should ideally include sensitivity and uncertainty analyses. Such analysis can be used to identify
the most efficient locations of sensing and actuation devices. These data should include not only
responses, but also measurements needed to define the model inputs and model input uncertainties
associated with loading, initial conditions, boundary conditions, etc.
The experimental activity involves the collection of raw data from various sensors used in the
experiment, such as accelerometers and load cells, and the generation of processed data such as time
integrals and averages. As necessary, the experimental data can be transformed into experimental
features that are more useful for direct comparison with simulation results. Such are the transfer
12.1 Models and model structures 247

function estimates and experimentally identified modal models. A series of repeated experiments
are generally required to quantify uncertainty due to lack of repeatability and inherent variability.
Valuable information that guide the calibration and validation process regards quantified effects
of various sources of uncertainty on the experimental data. Among these sources are measurement
error, design tolerances, manufacturing and assembly tolerances, unit-to-unit fabrication differences,
and variations in performance characteristics of experimental apparatuses and the experimenter’s
data processing. Experimental outcomes, which are the product of this uncertainty quantification
activity, will typically take the form of experimental data plus uncertainty bounds as mean and
covariance data functions of time or frequency.
 248 Chapter 12. Validation and Calibration Concepts

12.2 Problems
Problem 12.1 Illustration of overfitting with too many parameters
Eleven samples rk are taken at eleven discrete time instances tk = 0, 0.1, ..., 1.0s with equal time
increment Ts = 0.1s. The samples are: [0.97 1.114 1.171 1.33 1.324 1.39 1.38 1.375 1.273 1.143
0.972]. Identify the polynomial coefficients of a polynomial r̃(t) = a0 + a1t + a2t 2 + . . . + ant n by
use of least squares fitting. Start with polynomial order 0 and increase the order until order 12.
A validation sample rv = 1.02 is taken at time t=0.05s. Use the polynomial models to establish
the deviation metric rv − r̃(t = 0.05) for the increased model orders. Compute the deviation for
increasing polynomial order. Plot r̃ for polynomial order 10.
. .......................................................................................

Problem 12.2 Illustrates benefit of relevant data processing

Do a digital Fourier transformation (DFT) of the data given by Prob. 12.1. Keep the Fourier
coefficients of the first four harmonic components plus the static component (by setting the
remaining to zero) and do an inverse DTF back to time domain r̄k . Use the processed data to replace
the original calibration data rk and repeat the task of 12.1.
. .......................................................................................

Problem 12.3 Illustrates that bad data processing can destroy

Do a digital Fourier transformation (DFT) of data given by Prob. 12.1. Keep all but the two first
harmonic Fourier coefficients (by setting the remaining to zero) and do an inverse DTF back to time
domain r̄k . Use the processed data instead of the original data rk and repeat the task of Prob. 12.1.
NB! This is obviously a bad way of selecting Fourier coefficients for Fourier series representa-
tion (series in not complete). Anyhow it illustrates what a bad data processing can do. Other bad
data processing made in the calibration procedures can be much more delicate and still cause a lot
of hard-to-understand problems.
. .......................................................................................

Problem 12.4 Illustrate problem with dubious calibration metric

A few years ago the Balancing the Eigenvalue Equation method was proposed as a criterion function
for model calibration. Its idea was that the eigenvalue equation [KK − ω 2 M ]φφ = 0 would create a
non-zero residual δ if instead the experimentally found eigenvalue ωX2 and eigenvector φ X would
K − ωX2 M ]φφ X = δ . The quadratic criterion
take place instead of the analytical counterparts as [K
T
function δ δ should be used for calibration of K and M . Show that that criterion cannot be used
to discriminate between errors in mass and stiffness for the 1dof system shown in the figure by
proving that the Hessian is singular at K = K0 and M = M0 using that ωX2 = K0 /M0 .

. .......................................................................................

Problem 12.5 Illustrate a better criterion function for same system as in 12.4
With the Best Subspace Method the criterion can be selected to be ε = δ T δ with δ1 = (ωX2 −
ωA2 )/ωX2 and δ2 = (mX − mA )/mX . Here the eigenvectors are normalized to unity and the modal
masses are mX = φ TX M 0 φ X and mA = φ TA M φ A . Show that the Hessian is non-singular at K = K 0
and M = M 0 and thus the criterion may be better for calibration.
. .......................................................................................
12.2 Problems 249

Problem 12.6 Illustrates a crude calibration procedure

Find approximate minimum to the criterion function given in Prob. 12.5. Assume that the variables
K and M are Gaussian stochastic variables with mean at K 0 and M 0 with standard deviation
10% of the mean. With data from a Latin Hypercube Sampling (LHS, see Ch. 13.1.3) do find
approximate optimum from seven realizations. Multipliers on K 0 from the seven realizations found
from a random run of LHS are [0.9326 1.0485 0.9736 0.9891 0.7119 1.1049 1.1429] and the
corresponding multipliers on M 0 are [1.2243 1.0601 0.8939 1.0156 1.0202 0.9813 0.8306]. Take
mass and stiffness from these that minimizes the criterion.
. .......................................................................................

Problem 12.7 Two-parameter identification

One needs to estimate the Young’s and shear moduli E and G of a material for which one has a
L = 0.30m long test piece with circular cross-section with diameter d = 0.01m. In a vibration
experiment one has found the three lowest eigenfrequencies to be 23.2, 58.8 and 274.5Hz, and the
associated mass-normalized eigenmodes at dofs u1 through u3 (modal matrix Φ given in the figure).
In the experiment, the test piece has been set up as a cantilever beam (with negligible weight) with
an attached a spherical tip mass of weight M = 0.2kg and moment of inertia J = 180.106 kgm2 .
Determine E and G.

12/01/2012-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Problem 12.8 Illustrates verification challenge

Use a finite element code of your choice that has both the consistent mass and lumped mass
formulations. Do eigenvalue analysis to calculate the first 10 eigenfrequencies with consistent and
lumped mass formulations. Use a model of a 600 × 800 × 5mm steel plate for this problem. Use
quadrilateral elements with 4cm edges in a first run and then refine the mesh to 2cm elements.
Report the eigenvalues in tables. Comment any differences between the results of the two
formulations. Could you say that the models with courser meshes are verified if you require that
eigenvalues are predicted with a, say, 2% precision?.
. ...................................................................................
 13. Property Variability and Model Uncertainty

Very few computational models are made to represent only one single physical entity. On the
contrary, they often should represent a large population of mass-fabricated units. Although this is
the most common situation, exceptions can be found in various areas such as in civil engineering
for structures like dams and bridges and in astronautics for unique satellite units. Validation of
single unique units pose problems of its own, not the least the economical aspect since validation
work is not without cost. For mass-fabricated units one additional problem in model validation
is that each entity shows individually different behavior from the next. There is a spread of the
properties and characteristics between individuals. Even in the most automated production or under
otherwise very serious quality-assurance production control, the fabricated units show different
behavior to an extent that becomes more or less obvious from the outcome of vibration testing.
Reasons for the spread are plenty. One reason is that it may be caused by spread in material
properties of components that make up a physical unit. Some common reasons that give spread in
material properties are that:

a) rubber-like material used for vibration isolation is made in batches in which the constituents of
the compound is mixed to a recipe in which the batch-to-batch variation may be significant. A
spread in Young’s modulus for elastomers in excess of 15% is not uncommon even for high
quality applications,
b) products made from thin-sheet metals, such as cars, aircraft and domestic dryers, use material
that are processed by rolling. In that process, the virgin material is formed plastically to an
anisotropic crystal grain formation. The elasticity may vary significantly in orthogonal direction
of such sheets. In the production process of thin-sheet structures it is not uncommon that the
parts are cut out from rolls of metal sheets in orientations that vary from one unit to the other to
optimize the use of material from the rolls. This results in a unit-to-unit variation that may be
significant,
c) for composite components, the fibres are often laid up by hand or injected into molds in
processes that make the spread between individuals clearly observable.

Also the product assembly process leads to spread. Joining processes such as welding, riveting,
 252 Chapter 13. Property Variability and Model Uncertainty

bolting and gluing always leads to spread between individuals because such processes cannot be
quality controlled to 100%.
Another reason for variability is that in the assembly of mass-produced items, the end products
are often composed of components that vary to some extent because of other reasons. That may be
due to that components of different generations of development are used or that components from
different sub-contractors are used for the assembly. Most such variations are considered of little
importance and the simulation results of one representative finite element model of the product
should in some sense cover for such variations.
In model calibration and validation it is important to keep in mind that the models should be
validated for their intended purpose. The extent of accuracy that is expected from predictions and
simulations using the model needs to be balanced with the level of variability that is expected
between individuals. Such accuracy considerations should not be restricted to predictions and
simulation accuracy compared with the test results of one or a few test articles only. One should
consider that the ensemble of test articles, in the extreme case just one test piece, are just samples
of the full population of products for which the model is established. It is not uncommon that the
test article might even be a prototype which is not necessarily built to the exact specification of the
products in a later series production. The validation criterion chosen needs to be realistic and in
balance with these considerations.
The above considers the spread that can be anticipated between individuals of a population that
ideally should be identical. In the context of validation and calibration one also needs to consider
spread in test data that is not due to spread between individuals. One should realize that test data
could spread, also if these test data are taken from the one single sample of the population. It is
well known that test data from vibration testing vary by different reasons. A few of these are that:
a) the set-up of the same item may vary from time-to-time. One needs to support the test item but
also strive at a good isolation from the surrounding environment. This set-up may vary from
one time to the other. One time the item may hang in flexible bungee-chords and the next it
might be placed on a soft rubber cushion. The variation from time-to-time may be more subtle
than that, but small differences from set-up to set-up are unavoidable
b) the test engineer’s behavior may vary from time-to-time. Sometimes there is a great hurry to
produce test results and sometimes there is time for a more careful tests. It may also be that
the tests are conducted by different test engineers. Some of these are more skilled, careful and
fortunate than others
c) the environmental conditions change from time-to-time. Most relevant are temperature and
moisture changes. Test data taken on a hot and humid day may vary from test data taken at
another cold and dry day. Also the environmental noise is one such factor
d) if the tests are conducted at very different times also processes like ageing and wear may be
time-varying factors to consider.

13.1 Real world variability

In model calibration it is important to keep in mind that a perfect model fit to test data is not
possible due to modeling approximations and presence of noise in test data. A perfect fit to test
data from a single test may not even be desirable. That is related to the spread of the dynamic
characteristics that is presents between subjects in a population of supposedly identical individuals.
A celebrated calibration philosophy is that of the maximum likelihood parameter estimation, see
Sect. 13.2. That is that the model parameters should be adjusted to correspond to the maximum
of their probability density functions, i.e. to be at their most likely setting. This is a fully logical
approach that is in good consistency with the property spread we observe between individuals
of a population. However, in practice it is not easy to achieve a maximum likelihood calibration
 13.1 Real world variability 253

Figure 13.1: The FRF:s associated to excitation s1 and the acceleration responses r1 , r3 , r7 and r20
of eight individual wind turbine blades, see Fig. 13.3. Legend shows the blade series numbers of
the eight tested blades.
........................................................................................

since it requires that the parameter probability density functions are known or could be deduced
from testing. To achieve this is a daunting task that has never been heard of in structural dynamics
model validation. Most often testing is made on one single individual from the population which
we have reasons to believe to be a good representative of the full population. If we are fortunate
in our allocation of resources, we have the possibility to test an sub-ensemble of items from the
full population, constituting of say 2-5 individuals. To estimate good quality probability density
function estimates from these is impossible. The next-to-the-best option is then to do a statistical
evaluation and use mean value and variance estimates in the calibration process. Such data can be
used to assess the quality of the parameter calibration in the form of a parameter covariance metric,
see Sect. 13.1.2. A case study related to this subject follows.

13.1.1 Individual Spread - A Case Study

To get some sense of what can be expected from test data from an ensemble of individuals of
a mass-produced item, a case study was made by highly motivated Master’s students1 . It gave
information about spread between individuals, e.g. the frequency response functions shown in
Fig. 13.1, but also the spread that can be expected from repeated testing of the same item. It
is based on a testing of a blade of a small-scale wind turbine made of composite material, see
Fig. 13.2. The blades were made by molding of a composite fibres in a resin. The tests were
conducted on eight turbine blades in (almost) free-free conditions, see Fig. 13.3. The tests were
1 This was part of the curriculum for final year’s students in the master’s programme in Applied Mechanics at

Chalmers University of Technology in 2015

254 Chapter 13. Property Variability and Model Uncertainty

Figure 13.2: The three-bladed Ampair 600W wind turbine. The rotor diameter is 1.7m. In the test,
eight spare blades were tested under conditions mimicking a free-free boundary.
........................................................................................

Figure 13.3: The turbine blade hung in thin strings, barely visible at the green arrows. The location
of 20 accelerometers shown in yellow circles and the location of the actuator’s force transducer was
at the red arrow (at accelerometer #3). Part of the electromagnetic shaker can be seen at top right
corner.
........................................................................................
13.1 Real world variability 255

performed by 13 highly motivated students. Each student did a test on one single blade, and since
only eight blades were available, five blades were tested twice. All blades were hung in flexible
extremely thin and light-weight fishing lines to get good isolation from the ambient environment.
This provided an almost free-free condition in the sense that the flexible body modes of the blade
were practically unaffected by the blade supporting system. The blades were hung, instrumented
with accelerometers, and connected to a shaker via a force transducer. After the test, the blade was
dismounted and stored waiting for the next student to set up and do testing. This procedure was
repeated also for the blades that were tested multiple times. The test results thus gives information
about the property spread between seemingly identical blades, and also about spread in test data
between tests performed at multiple occasions.
The tests were executed using a stepped-sine excitation procedure, see Ch. 11.2, to get estimates
of the blade’s frequency response functions from one force stimulus to 20 acceleration responses.
The stepping was from 30Hz to 500Hz which covers the first eight eigenfrequencies of the free-free
blade’s flexible-body eigenmodes. The frequency increment in the stepping was 0.25 Hz. A
state-space sub-space system identification was made, using the N4SID method (see Ch. 8.3), with
further eigensolution processing to obtain modal data for the eight eigenmodes covered.

Spread between blade individuals. The frequency response functions associated to some
accelerometers can be seen in Fig. 13.1. The other 16 accelerometers showed similar frequency
response function behavior and are therefore not presented here. Fig. 13.1 (subfigure: upper right)
shows the direct accelerance as obtained from the accelerometer located at the force transducer
and oriented along the direction of the applied force through the shaker’s stinger. As can be
seen, there is a significant spread between the estimated frequency response functions. System
identifications based on the frequency response functions were made for evaluation of modal
properties. Eigenfrequencies and modal damping are listed in Tabs. 13.1 and 13.2. It can be seen
that the spread in eigenfrequencies is in the order of 3% as quantified by the coefficient if variation
(COV). The COV of the modal damping is significantly bigger at around 10%. This does not hold
for the 7:th mode that was harder to estimate precisely from the test data2 . Deviation metric such as
the COV of eigenfrequencies and dampings can give useful information to support a calibration
process. Other metric relate to the correlation of eigenvectors. Two vector correlation metric, the
modal assurance criterion (MAC) and the modal observability correlation (MOC), see Sect. 9.1.1
and 9.1.3, of the state-space models of the eight blades are presented in Tabs. 13.6 and 13.7.

Spread in test results from repeated testing on same blade. Five blades (blade series
numbers 722, 790, 841, 878 and 881) were tested twice. After each blade test, the blade was
dismounted from the test stand and put to store for a few days at the maximum. Another student
then set-up and performed a new test on the same blade. As can be expected, the variation from
test-to-test on the same blade was smaller than the variation between different blades. However,
the differences were not insignificant. Some of the frequency response functions obtained from
multiple tests of same blade are presented in Figs. 13.4 and 13.5. The frequency response functions
were estimated from discrete-time data and state-space sub-space system identifications were made
based on these for each blade test. Eigenfrequencies and modal dampings were evaluated and the
results are shown in Tabs. 13.3 and 13.4. It can be seen that the variation of eigenfrequencies is in
the order of fractions of a few percent and the variation of estimated modal dampings is in the order
of 10 percent. The variation of eigenfrequencies are thus about one order of magnitude lower than
2 Later examination revealed that the motion in the 7:th mode was more-or-less orthogonal to primary sensing

direction of the accelerometers

256 Chapter 13. Property Variability and Model Uncertainty

Figure 13.4: Bode plots of direct transfer functions associated with accelerometer #3 (see Fig. 13.3)
from two repeated tests a and b. Legends gives blade series #.
........................................................................................

Figure 13.5: Bode plots of cross transfer functions associated with accelerometer #20.
........................................................................................
13.1 Real world variability 257

Blade series #
Mode # 722 790 819 841 852 877 878 881 Mean COV (%)
1 45.2 48.4 43.6 45.2 46.5 47.5 46.7 45.3 46.05 3.3
2 126.6 128.1 128.8 128.8 128.9 133.7 128.2 126.9 128.5 2.0
3 203.5 198.7 185.1 189.2 190.7 192.7 191.3 200.0 193.9 3.2
4 241.0 240.7 247.7 248.3 251.8 254.7 245.7 244.1 246.7 2.0
5 331.1 317.9 311.7 320.5 312.4 314.8 315.3 328.5 319.0 2.3
6 387.6 394.6 393.7 391.8 386.8 400.8 390.1 393.3 392.3 1.1
7 412.0 405.5 436.9 406.5 411.3 416.7 400.5 443.1 416.6 3.7
8 471.2 464.8 454.9 462.1 435.1 467.8 469.4 464.0 461.2 2.5

Table 13.1: Blades’ eigenfrequencies with mean and coefficient of variation.

Blade series #
Mode # 722 790 819 841 852 877 878 881 Mean COV (%)
1 0.83 1.10 1.11 1.00 1.39 1.01 0.93 1.06 1.05 15.4
2 0.94 0.95 0.91 0.92 0.96 0.95 0.96 1.02 0.95 3.6
3 2.13 2.54 2.30 2.29 2.32 2.17 2.20 2.20 2.27 5.8
4 1.13 1.16 1.12 1.06 1.10 1.22 1.17 1.14 1.14 4.2
5 2.16 2.16 2.46 2.63 2.39 2.28 2.19 1.91 2.28 9.6
6 1.15 1.09 1.30 1.39 1.68 1.42 1.31 1.15 1.31 14.5
7 1.97 3.58 3.23 1.27 1.20 2.10 1.26 1.70 2.04 44.8
8 1.96 2.15 2.53 2.08 2.47 2.06 2.31 1.89 2.18 10.8

Table 13.2: Blades’ modal damping.

the variation of damping also in repeated tests of the same blades. For the vector correlation, the
MOC and MAC indices were calculated for the eight modes seen in data. It can be seen that the
modes are very well correlated with most MAC values above 99% and most MOC values above
95%.
258 Chapter 13. Property Variability and Model Uncertainty

Blade series #
Mode # 722a 722b 790a 790b 841a 841b 877a 877b 881a 881b %
1 45.18 44.98 48.44 48.34 45.18 45.26 47.61 47.48 45.31 45.28 0.23
2 124.6 124.1 128.1 127.8 128.8 128.9 134.1 133.7 126.9 126.6 0.24
3 203.5 202.2 198.7 197.9 189.2 189.9 192.3 192.7 200.0 199.4 0.38
4 241.0 239.9 240.7 240.6 248.3 248.6 255.4 254.7 244.1 243.1 0.26
5 331.1 330.0 317.9 315.6 320.5 320.8 315.1 314.8 328.5 328.0 0.28
6 387.6 386.6 394.6 392.9 391.8 392.7 401.6 400.8 393.3 392.3 0.28
7 412.0 409.7 405.5 459.0 406.5 404.3 413.0 416.7 443.1 415.3 -
8 471.2 476.2 464.8 483.0 462.1 464.0 466.9 467.8 464.0 464.8 1.13

Table 13.3: Eigenfrequency estimates from data of two tests (a and b) on same blades. Right-most
column shows average of difference in percent relative to mean. Outliers are highlighted in yellow.

Blade series #
Mode # 722a 722b 790a 790b 841a 841b 877a 877b 881a 881b %
1 0.83 1.16 1.10 1.05 1.00 0.96 0.90 1.01 1.06 1.20 13.2
2 0.94 1.06 0.95 0.99 0.92 0.90 0.92 0.95 1.02 1.00 4.5
3 2.13 1.99 2.54 2.16 2.29 2.24 2.41 2.16 2.20 1.98 9.2
4 1.13 1.08 1.16 1.19 1.06 1.13 1.06 1.22 1.15 1.15 5.7
5 2.161 2.11 2.16 2.51 2.63 2.15 2.35 2.28 1.92 2.02 9.1
6 1.15 1.29 1.09 1.26 1.39 1.16 1.25 1.42 1.15 1.27 13.2
7 1.97 1.67 3.58 2.26 1.27 1.37 1.99 2.10 1.70 1.61 - 16.0
8 1.96 1.89 2.15 14.86 2.08 2.15 2.14 2.06 1.89 2.03 -

Table 13.4: Damping estimates from data of repeated tests. Right-most column shows average of
difference in percent relative to mean. Outliers are highlighted.

Blade series #
Mode # 722a vs 722b 790a vs 790b 841a vs 841b 877a vs 877b 881a vs 881b
1 .996/.986 .997/.960 .999/.967 .998/.984 .999/.998
2 .999/.976 .999/.978 .999/.977 .999/.979 .999/.984
3 .999/.946 .998/.959 .996/.956 .999/.934 .998/.996
4 .999/.967 .995/.972 .999/.989 .999/.996 .999/.956
5 .999/.966 .989/.958 .997/.887 .998/.934 .997/.979
6 .999/.958 .983/.916 .998/.973 .998/.926 .996/.942
7 .963/.932 .055/.001 .984/.662 .891/.253 .528/.187
8 .980/.896 .397/.036 .997/.993 .997/.795 .987/.921

Table 13.5: MAC and MOC correlation indices of first eight modes from dual tests (a and b) on
same blades. Outliers are highlighted.
13.1 Real world variability 259

Blade series #
Blade # 790 819 841 852 877 878 881
1st mode at 46.05Hz
722 .997/.921 .995/.858 .997/.961 .993/.888 .994/.970 .993/.972 .992/.834
790 .995/.784 .997/.888 .992/.947 .992/.943 .993/.908 .990/.770
819 .996/.891 .993/.762 .996/.833 .993/.866 .997/.966
841 .993/.865 .994/.936 .996/.976 .993/.866
852 .996/.907 .995/.874 .988/.743
877 .997/.961 .992/.816
878 .989/.846
2nd mode at 128.5Hz
722 .972/.857 .990/.937 .994/.960 .990/.899 .994/.816 .992/.978 .985/.830
790 .988/.849 .982/.862 .990/.955 .983/.945 .977/.873 .976/.737
819 .994/.974 .994/.887 .988/.833 .982/.946 .987/.864
841 .995/.906 .993/.854 .985/.968 .979/.840
852 .996/.941 .987/.914 .986/.770
877 .995/.865 .985/.713
878 .984/.832
3rd mode at 193.9Hz
722 .982/.812 .996/.514 .978/.542 .989/.649 .978/.789 .975/.651 .963/.882
790 .978/.699 .975/.717 .989/.849 .988/.885 .990/.859 .984/.896
819 .976/.892 .987/.904 .971/.740 .966/.878 .958/.600
841 .991/.871 .974/.724 .990/.867 .984/.642
852 .988/.840 .991/.989 .987/.755
877 .986/.842 .981/.911
878 .993/.771
4th mode at 246.7Hz
722 .953/.643 .983/.822 .962/.776 .976/.662 .986/.650 .954/.697 .925/.772
790 .967/.708 .954/.710 .974/.761 .962/.651 .972/.855 .906/.717
819 .958/.932 .978/.836 .981/.855 .960/.816 .932/.884
841 .986/.870 .961/.868 .972/.846 .959/.930
852 .988/.923 .980/.928 .932/.793
877 .971/.826 .918/.777
878 .950/.841

Table 13.6: Cross correlation numbers MAC and MOC (MAC/MOC) between blade individuals for
modes 1-4.
 260 Chapter 13. Property Variability and Model Uncertainty

Blade series #
Blade # 790 819 841 852 877 878 881
5th mode at 319.0Hz
722 .960/.618 .970/.478 .963/.565 .989/.556 .980/.630 .972/.516 .952/.897
790 .922/.823 .918/.781 .953/.897 .978/.887 .981/.865 .928/.599
819 .955/.820 .976/.877 .931/.784 .931/.879 .907/.458
841 .984/.769 .957/.728 .953/.897 .953/.551
852 .977/.918 .969/.827 .959/.552
877 Symm. .986/.802 .968/.633
878 .959/.517
6th mode at 392.3Hz
722 .853/.698 .920/.847 .909/.866 .757/.612 .945/.663 .918/.780 .935/.609
790 .811/.728 .780/.684 .795/.704 .950/.890 .854/.615 .841/.658
819 .985/.967 .893/.722 .888/.741 .956/.773 .950/.668
841 .892/.729 .864/.689 .950/.787 .918/.634
852 .836/.665 .900/.618 .840/.663
877 Symm. .935/.599 .926/.701
878 .934/.533
7th mode at 416.6Hz
722 .054/.018 .482/.224 .439/.336 .439/.274 .847/.077 .484/.418 .304/.105
790 .049/.007 .206/.088 .200/.041 .006/.001 .202/.066 .066/.009
819 .542/.139 .509/.204 .370/.019 .508/.118 .147/.095
841 .959/.468 .291/.030 .975/.740 .075/.016
852 .282/.016 .934/.526 .455/.014
877 .328/.023 .455/.025
878 .111/.013
8th mode at 461.2Hz
722 .897/.501 .919/.462 .883/.472 .777/.090 .941/.601 .863/.383 .866/.750
790 .927/.784 .907/.893 .741/.211 .972/.866 .947/.714 .858/.450
819 .958/.832 .666/.266 .937/.795 .905/.539 .848/.488
841 .586/.189 .938/.811 .940/.697 .781/.405
852 .739/.162 .608/.184 .747/.113
877 .969/.666 .845/.489
878 .782/.307

Table 13.7: Cross correlation numbers MAC and MOC (MAC/MOC) between blade individuals for
modes 5-8. NB! Observe that the correlations for mode 7 are generaly very low, an observation
consistent with the large eigenvalue deviations for this mode, see Tab. 13.1.

13.1.2 Random outcomes, probability and basic statistics

We revisit the mathematical framework that treats random phenomena such that those that are
observed in vibration testing. Such testing always shows that repeated experiments are associated
with a set of possible observed results with a spread. An example was given in chapter 4.1. Items
that show this behavior are said to be random. If on the other hand experiments are performed
repeatedly with all conditions maintained as precisely as possible, and the measured results are
13.1 Real world variability 261

identical, then the item is said to be deterministic. A deterministic outcome from vibration testing
has never happened to the author’s knowledge.
Associated with each of the random outcomes is the probability of the event. It seems intuitively
reasonable that the probability is related to the expected relative frequency of occurrence of the
event in a long sequence of experiments and that the sum of probabilities of all possible events
should be unity.
All possible outcomes of the experiment, which can be represented by points called sample
points, comprise the sample space of the experiment. Any variable defined on a sample space is
called a random variable. An event is the result of one particular realization of the sample points in
the sampled space.

Random variables, probability density and statistical distribution functions. Let p be a

variable which may randomly take on any real value −∞ < p < ∞. The function pdf(p) is said to
be the probability density function (PDF) of the random variable p. We note that it is a function
that is either positive or zero since probabilities are never below zero. The probability that an
outcome of the experiment will fall in the infinitesimal range [p, p + d p] is pdf(p)d p. This define
the probability operator Pr as Pr(p) , pdf(p)d p. More generally, the notation
Z b
Pr(a ≤ x ≤ b) = pdf(p)d p (13.1)
a

is employed to indicate the probability that the random variable p has a value between the two
values a and b. A specific probability is given by the cumulative distribution function (CDF) which
is defined as
Z b
cdf(b) = Pr(p ≤ b) = pdf(p)d p (13.2)
−∞

In other words, cdf(b) is the probability that the random variable p will have a value lesser than
or equal to b. A CDF is a monotonically increasing function which possesses the following four
properties

d
cdf(−∞) = 0 0 ≤ cdf(p) ≤ 1 cdf(∞) = 1 cdf(p) = pdf(p) (13.3)
dp

In particular we note that probability density function pdf(p) is the slope of cumulative distri-
bution function cdf(p).
Frequently, more than one random variable, say the variables p1 , p2 , . . . , pn p , needs to be
involved to describe the outcome of an experiment. Then the joint behavior of these is of interest.
A joint probability density function jpdf(p1 , p2 , . . . , pn p ) = jpdf(pp) must satisfy the following
conditions

jpdf(pp) ≥ 0 (13.4)

and
Z b1 Z b2 Z bn
p
Pr(a1 ≤ p1 ≤ b1 , . . . , an p ≤ pn p ≤ bn p ) = ... jpdf(pp)d p1 d p2 . . . d pn p (13.5)
a1 a2 an p

and also
Z ∞ Z ∞
... jpdf(pp)d p1 d p2 . . . d pn p = 1 (13.6)
−∞ −∞
262 Chapter 13. Property Variability and Model Uncertainty

A joint cumulative distribution function is defined by

Z b1 Z b2 Z bn
p
jcdf(p1 ≤ b1 , . . . , pn p ≤ bn p ) = ... jpdf(pp)d p1 d p2 . . . d pn p (13.7)
−∞ −∞ −∞

where b1 , b2 , . . . , bn p are constant values. The joint CDF is non-descending with respect to all of its
random variables. Also, there is a relation between the joint CDF and the joint PDF as
∂ n p jcdf(pp)
jpdf(pp) = (13.8)
∂ p1 ∂ p2 . . . d pn p
A joint PDF provides complete probabilistic information about the PDF of any one random
variable. The latter function, when obtained from a joint PDF, is called a marginal probability
density function. A marginal PDF of the k:th random variable is obtained from the joint PDF by
integrating out the remaining variables as
Z ∞ Z ∞ Z ∞ Z ∞
pdf(pk ) = ··· ··· jpdf(pp)d p1 d p2 . . . d pk−1 d pk+1 . . . d pn p (13.9)
p1 =−∞ pk−1 =−∞ pk+1 =−∞ pn p =−∞

The joint PDF contains more statistical information than the PDFs pdf(p1), pdf(p2),..., pdf(pn p )
separately since the latter PDFs can be obtained from the former but not vice versa. The marginal
cumulative distribution function of pk is defined accordingly as
Z bk
cdf(bk ) = pdf(pk )d pk (13.10)
−∞

The random variables are said to be statistically independent if and only if

jpdf(p1 , p2 , . . . , pn p ) = pdf(p1 ) · pdf(p2 ) · · · pdf(pn p ) (13.11)

for which it then also holds for their joint cumulative distribution function that

jcdf(p1 , p2 , . . . , pn p ) = cdf(p1 ) · cdf(p2 ) · · · cdf(pn p ) (13.12)

Event expectations and statistical moments. Single variable case. Let f (p) be a continuous
function of a random variable p which means that also f is random. The expectation (also known
as the mean value) of the function f is defined as
Z ∞
E[ f ] = f (p)pdf(p)d p (13.13)
−∞

In the special case of f (p) = p the expectation becomes the expectation (or mean, or average)
of the random variable p denoted by µ p which thus is
Z ∞
µ p , E[p] = p pdf(p)d p (13.14)
−∞

Another interesting special case is where f (p) = p2 that leads to the expectation
Z ∞
E[p2 ] = p2 pdf(p)d p (13.15)
−∞
p
which is called the mean square value of the random variable p. Its positive square root E[p2 ] is
called the root mean square value. The mean and the mean square value are sometimes called the
first order moment and the second order moment of the distributions of p respectively.
13.1 Real world variability 263

The variance of p, denoted by var(p) or sometimes σ p2 is defined as the mean square value of
p about the mean as
Z ∞
var(p) = σ p2 , E[(p − µ p )2 ] , (p − µ p )2 pdf(p)d p = (13.16)
−∞
Z ∞ Z ∞ Z ∞
p2 pdf(p)d p − 2µ p p pdf(p)d p + µ p2 pdf(p)d p = E[p2 ] − µ p2 (13.17)
−∞ −∞ −∞

and thep positive square root of that variance is the standard deviation of p denoted σ p , i.e.
σ p , + var(p). Obviously,
p if the mean value µ p is zero the standard deviation σ p equals the root
mean square value E[p2 ]. For random variables with positive mean its random spread relative to
its mean value has a meaning. That quantity, named the coefficient of variation (COV) and denoted
by α p , is simply

α p = σ p /µ p (13.18)

Both the mean and variance of a stochastic variable give important information about its
distribution. The mean can be thought of as the ‘center of gravity’ of the variable’s PDF as it is
governed by the same kind of integral relations as particle mass distributed in space, while the
variance is in analogy with the mass moment of inertia for a similar same reason.

Event expectations and statistical moments. Multi-variable case. The notation of expecta-
tion applies to a set of random variables as well. The expectation of a function f (p1 , p2 , . . . , pn p ) =
f (pp) of n p random variables is defined as
Z ∞ Z ∞
E[ f (pp)] = ··· f (pp) jpdf(pp)d p1 d p2 . . . d pn p (13.19)
−∞ −∞

Certain expectations warrant special names, such as

1. The cross-correlation between p j and pk , which is denoted µ jk and is defined by

Z ∞Z ∞
µ jk , E[p j , pk ] , p j pk jpdf(p j , pk )d p j d pk (13.20)
−∞ −∞

where jpdf(p j , pk ) is the (marginal) joint PDF of p j and p j given by

Z ∞ Z ∞
jpdf(p j pk ) = ··· jpdf(pp)d p1 . . . d p j−1 d p j+1 . . . d pk−1 d pk+1 . . . d pn p (13.21)
−∞ −∞

2. The covariance between p j and pk as defined by

cov(p j , pk ) , E[(p j − µ j )(pk − µk )] ,

Z ∞Z ∞
(p j − µ j )(pk − µk ) jpdf(p j , pk )d p j d pk = µ jk − µ j µk (13.22)
−∞ −∞

We note that if p j and pk are statistically independent, then their covariance vanishes. This
results from the definition of independence of random variables, c. f . Eq. (13.11), that by the
definition of µ jk implies that µ jk = µ j µk .

3. The non-dimensional correlation coefficient between p j and pk defined by

q
α jk = cov(p j , pk )/ var(p j )var(pk ) ≤ 1 (13.23)
264 Chapter 13. Property Variability and Model Uncertainty

The random variables p j and pk are named uncorrelated if cov(p j , pk ) = 0. Hence, independent
random variables are uncorrelated. On the other hand, missing correlation between variables does
not necessarily imply statistical independence between them.
The mean vector of a set of n p variables p1 , p2 , . . . pn p with mean values µ1 , µ2 , . . . , µn p is by
definition

p̄p , {µ1 µ2 . . . µn p } (13.24)

and their covariance matrix is defined as

 
var(p1 ) cov(p1 , p2 ) · · · cov(p1 , pn p )
cov(p1 , p2 ) var(p2 ) ··· cov(p2 , pn p )
Sp ,  (13.25)
 
.. .. .. .. 
 . . . . 
symmetric var(pn p )
Let us conclude this section by quoting the Chebyshev inequality, also known as the limit
theorem. Let p be a random variable with mean µ p and variance σ p2 . Then for any real number ε
we have

Pr(|p − µ p | ≥ ε) ≤ σ p2 /ε 2 (13.26)

This shows clearly the close relation between the variance of a distribution and the dispersion
of that distribution about the mean. For proof, see Ref. [42]. The Chebyshev inequality gives a
relatively weak bound on most distributions. However, it is general and therefore applies to all
distributions whose means and variances exist.

Some commonly used probability distributions and their relation.

As an example of many useful probability density functions, and the most commonly employed
in scientific and engineering applications, is the normal distribution (also called the Gaussian
distribution). Its PDF has the following form
1 2 2
pdf(p) = √ e−(p−µ p ) /2σ p (13.27)
2πσ p
This PDF depends on two parameters, the mean µ p and the standard deviation σ p , which thus
fully characterize the Gaussian random variable. An equivalent and useful alternative to Eq. (13.27)
is obtained by defining a new variable p̆ such that p̆ = (p − µ p )/σ p which leads to the so-called
standard normal distribution N(0, 1) with PDF
1
pdf( p̆) = √ e− p̆/2 (13.28)
2π
for which the mean is zero and its standard deviation is unity. Using the error function erf(p) ,
√ R 2
2/ π 0p e−t /2 dt the associated CDF of p̆ can be expressed as
1
cdf( p̆) = (erf( p̆ + 1) (13.29)
2
The Gaussian PDF possesses a number of remarkable mathematical properties. Among them is
that functions that are linear in the Gaussian random variables also become Gaussian so that, for
instance, if p j and pk are Gaussian, then p j + pk is Gaussian as well.
The multi-variate PDF of normally distributed parameters is given by
1 T −1
jpdf(pp) = p e−{pp− p̄p} S p {pp− p̄p}/2 (13.30)
2π|SS p |
 13.1 Real world variability 265

Since many variables used in model calibration have the character of being positive and bounded
(like Young’s modulus, plate thickness and material density) the normal distribution function for
variables that can take values in the range ] − ∞, +∞ is not always suitable. The Young’s modulus,
for instance, is bounded to be larger than zero or otherwise it would contradict the second law of
thermodynamics. This is based on an energy principle. It also has an upper bound that is determined
by the cohesive force-deformation relation of the perfect atom grid of the material in question.
The thickness of a sheet metal is positive and bounded by the production tolerances. A commonly
used probability functions for such variables is the Beta distribution, see figure XXX, defined for a
standard Beta variable p ∈ [0, 1] and positive exponents r and s to be

pdf(p) = pr−1 (1 − p)s−1 /B(r, s) (13.31)

where the Beta function B(r, s) is defined as

Z 1
B(r, s) = xr (1 − x)s−1 dx (13.32)
x=0

With a transformation of variables p̆ = (phi − plow )p + plow where plow is the lower bound of
the parameter and phi is its upper bound, we have for the expectation of p̆ that

r(phi − plow )
µ p̆ = + plow (13.33)
r+s
and for the variance of p̆ that

rs(phi − plow )2
σ p̆2 = (13.34)
(r + s)2 (r + s + 1)

It should be noted that a transformation to a standard Gaussian variable q ∈ N(0, 1) from a

variable of any other distributions is possible, but the transformation is not always expressible in
a simple closed form. Symbolically however, we can state that we can relate the variable p of an
arbitrary distribution to standard normal variable q by the expression q = map(p) with the inverse
relation p = map−1 (q).
(More on mapping and uniform distribution to come)

13.1.3 Random sampling

For problems for which the system characteristics are governed by random parameters, the input-
to-output relation of the system is determined by the random parameter setting, i.e. its realization.
The response characteristics for various parameter realizations is then best described in terms of
statistical metric such as means and variances. For large scale problem solutions, common in
structural dynamics, a full statistic characterization of the responses is not feasible for reasons
of cost, time and other resources. By a full characterization we then mean a characterization
in which we utilize the true joint PDF of the parameters. Currently, the most ambitious efforts
to get statistical response data is by Monte Carlo realizations. In the Monte Carlo realizations,
the discrete value parameter settings are determined by random number generation based on the
assumed joint PDF of the parameters. Usually many thousands of realizations are required to
get useful statistical information about the responses. This can only be realized in the simulation
setting and is abandoned in the test setting. For quicker and less costly assessment of the response
spread due to parameter variability instead other approximation methods are used. One such is the
Latin Hypercube Sampling (LHS) method. This method also has a role in the parameter calibration
setting in which the outcome of simulation samples can be compared with test data.
266 Chapter 13. Property Variability and Model Uncertainty

Figure 13.6: Mapping of a uniform distribution on a normal distribution. The probability of an

outcome below S̄ jk of the uniform distribution is equal to the probability of an outcome below S jk
of the normal (Gaussian) distribution.
........................................................................................

The latin hypercube sampling scheme for planning of simulation experiments was first in-
troduced by McKay et al, see Ref. [28]. The latin hypercube sampling method was employed
because there was a strive to reduce the number of necessary realizations to study a given stochastic
problem to gain efficiency with some loss of accuracy. The inner working of the LHS method
is an initiation with the creation of a permutation matrix P for the parameters in p . The size of
the permutation matrix is nR × ns , where nR is the number of realizations required. Each of the
np columns of P contains a random permutation of the integers from 1 to nR . Another matrix U ,
called the location matrix indexlocation matrix, of independent random numbers q from a uniform
distribution q ∈ U[0, 1] is also established. Then each element S jk of the sampling matrix S is
computed as

S jk = mapu2p (S̄ jk ) = mapu2p ((Pjk +U jk − 1)/nR ) j = 1, 2, . . . , nR k = 1, 2, . . . , n p (13.35)

where mapu2p denotes the functional mapping from the uniform distribution to the true distribution
of the random variable p.
As an example, the mapping from a uniform distribution U(0, 1) to a Gaussian distribution
N(µ p , σ p2 ) is shown in Fig 13.6.
An element of the sampling matrix can be deduced from the inverse of the error function erf
and is
√
S jk = 2 erf−1 (2(Pjk +U jk − 1)/(nR − 1)) + µ p (13.36)

Each row in the sampling matrix S contains the input values for one deterministic simulation. An
example of a latin hypercube sampling with two input parameters and six realizations is illustrated
in Fig. 13.7. Subfig. 13.7c illustrates the input to the first realization according to the sampling
scheme. The position of the small strokes in the shadowed areas in Subfig. 13.7c tally with the
position of the bullet in the shadowed area of Subfig. 13.7b. The position of the strikes correspond
to the values the two random parameters p1 and p2 take in the first realization. They correspond to
the first row of the permutation matrix P . The same procedure is conducted nR times to get values
for the nR realizations.
 13.2 Parameter estimation statistics 267

Figure 13.7: (a) Latin hypercube permutation matrix P and location matrix U with two parameters
and six realizations. The permutation matrix determines the bins from which the realizations are
taken and the uniformly distributed U determines exactly where within the bin the realizations are
taken. (b) Schematic illustration of the location of the setting of two normalized parameters in
the six realizations. (c) Example of parameter values to the first realization of Gaussian variables
shown (first row of P and U ) as small bullets on abscissa. Note that the areas of the bins are equal,
i.e. A1 = A2 = . . . = A6 .
........................................................................................

When using the latin hypercube sampling there is always a chance that the covariance of the
realizations do not match the true parameter covariance and an augmented method called the LHS
with Correlation Control has been developed, see Ref. [23], to alleviate that problem.

13.2 Parameter estimation statistics

In the model calibration we estimate the model parameters from test data that we have available.
These test data are always associated with some randomness. If we repeat an experiment under
seemingly identical conditions in multiple realizations we will get different results, as we saw in
Sect. 13.1 above. If the tests are set up carefully and performed with proper hardware and skilled
test engineers in a “noise-free” environment the differences may be small and possible negligible.
However, in most practical situations there is scatter in data that probably best can be seen as the
outcome of stochastic processes. Calibrations using data stemming from different realizations with
data spread will render different calibration outcomes. The calibration parameters estimated from
test data will then seemingly be random in character because of such variations in test data even if
they represent some property that is without randomness. Here we describe some basic aspects on
the statistics obtained from the parameter estimation in the calibration process.

The maximum likelihood estimator. The topic of parameter estimation deals with the problem
268 Chapter 13. Property Variability and Model Uncertainty

of extracting information from observations that have randomness. The observations are then
realizations of stochastic variables. Suppose that the observations are represented by processed
test data as random elements in the nd -dimensional vector z . These observations may be processed
time-domain data y ∈ Rny that is the output recorded in the vibration testing. In vibration testing,
often with a high channel-count data acquisition system, high sampling rates and long duration
testing, the collected discrete-time data vector sizes may be in the order of billions. These data
are most often processed with frequency analysis and averaging into transfer function estimates
or further via system identification to invariant eigensystem data as system eigenfrequencies and
eigenvectors. This process then consists of a mapping of y into z ∈ Rnd . Although the statistical
properties of the collected output data y can be assessed without much problem, the statistical
properties of the processed data z are often less known. However, let us assume that the joint
probability density function of the elements of z is jpdf(zz|pp), i.e. the probability function of z
depend on the parameter setting p that we believe represent some real world physical properties.
But that parameter setting is hidden to us and implicitly embedded in test data. The probability of
the test outcome to be within the domain Z is then
Z
Pr(zz ∈ Z) = jpdf(zz|pp) dz1 dz2 · · · dznd (13.37)
z ∈Z

In Eq. (13.37, p is the parameter vector that quantify system properties that we want to observe.
These parameters are unknown, and the purpose of the observation is to estimate the vector p using
observation data z . This is accomplished by an estimator p̃p(zz) which is a mapping function from
Rnd to Rn p . If the observed values of z are z ∗ , than consequently the resulting estimate of the
parameters is p̃p(zz∗ ).
Many such estimator functions p̃p(zz) are possible. A particular estimator that maximizes the
probability of the observed data is the maximum likelihood estimator, see Ref. [15]. It is based
on the joint probability density function for the random observations z . The probability that
the observed realization indeed should take value z ∗ is thus proportional to jpdf(zz|pp). This is a
deterministic function of p once the numerical values z ∗ are inserted. This function is called the
likelihood function. It reflects the likelihood that the observed realization should indeed take place.
A reasonable estimator of p could then be to select it so that the observed realization becomes as
likely as possible. That is to seek an estimator

p̃p(zz∗ ) , p ML = argmax jpdf(zz∗ |pp) (13.38)

p

where the maximization is performed for fixed z ∗ , i.e. the given processed data. This function is
known as the maximum likelihood estimator (MLE) for p .

The Cramer-Rao Inequality. The credibility of an estimator p̃p can be assessed by its mean-
square error matrix

C , E[{ p̃p − p ? }{ p̃p − p ? }T ] (13.39)

Here p ? denotes the hidden true value of p , and C is evaluated under the assumption that the
joint PDF of z is jpdf(zz|pp? ).
In the selection of a good estimator from a set of possible estimators, an estimator that make C
small seems to be a natural choice. It is then interesting to note that there is a lower limit to the
values of C that can be obtained with various unbiased estimators. This is given by the Cramer-Rao
inequality that states
13.2 Parameter estimation statistics 269

p(zz) be an estimator of p with mean

Theorem 13.2.1 — The Cramer-Rao theorem. Let p̃
E[ p̃p] = p ? and assume that the joint PDF of z is jpdf(zz|pp? ). Further assume that z is real-valued
and bounded within a boundary that does not depend on p. Then the mean square error is

E[{ p̃p − p ? }{ p̃p − p ? }T ] ≥ F−1 (13.40)

for which

∂ log jpdf(zz|pp) T ∂ 2 log jpdf(zz|pp) 

  
∂ log jpdf(zz|pp) 
F = E[
∂p  ? ∂p  ? ] = −E[ ∂ p2  ? ] (13.41)
p =pp p =pp p =pp

Proof. See [27]. 

Since p is an n p dimensional vector, the gradient of the scalar value log(jpdf(zz|pp)) is an n p

dimensional column vector and the Hessian in Eq. (13.41) is an n p × n p matrix F. This matrix
is known as the Fisher information matrix. Notice that the evaluation of F normally requires
knowledge of p ? , so the exact value of F may not be available. It is then common to use the
assumption that for an estimator p̃p of p one has that F(pp? ) ≈ F( p̃p).

The Asymptotic Properties of the Maximum Likelihood Estimator. It is usually very

difficult to exactly calculate the statistical properties of the parameter estimates of any given
estimator. However, for at least one estimator - the MLE estimator, there is a classical theorem that
is valid for data that are statistically independent and in a number tend to infinity. The theorem
states that:
Theorem 13.2.2 — Maximum Likelihood Bound Theorem. Suppose that random data z ∈ Rnd
are statistically independent so that
nd
jpdf(zz|pp) = ∏ pdf(zk |pp) (13.42)
k=1

Suppose also that the distribution of z for a given setting of the parameters p = p ? is given by
jpdf(zz|pp? ). Then the parameter estimate

p ML = argmax jpdf(zz∗ |pp) (13.43)

p
√
tends to p ? with probability one as nd tends to infinity and the parameter vector nd (ppML − p ? )
converges in distribution to the normal distribution with zero mean and the covariance matrix
being the inverse of the Fisher information matrix and is thus approaches the Cramér-Rao lower
bound.

Proof. See [11]. 

Thus, when the number of data nd tends to infinity, the MLE p ML is distributed N(pp? , F−1 ).
According to the Cramer-Rao theorem, this is the best an estimator can do and therefore it is often
said that the MLE is an efficient estimator.

The case of normally distributed data. Let us consider the special case when the data zk
as obtained from evaluation of test data are quantities that can be predicted without bias with a
 270 Chapter 13. Property Variability and Model Uncertainty

parameterized model in its calibrated setting p ? . Let the predicted data from the model be z̆k . We
can thus write

zk = z̆k (pp? ) + εk (13.44)

with εk being a residual that cannot be explained by the model. Let us assume that the residuals are
statistically independent variables distributed N(0, σk2 ) with known standard deviation σk . In that
case the parameter covariance lower bound can be shown [46] to be
np
1 ∂ z̆k (pp)  ∂ z̆k (pp) T
 
F(pp) = ∑ 2 (13.45)
k=1 σk ∂ p p ∂ p p

This result simplifies matters much. As can be seen, test data are not explicitly part of the
equation, but only implicitly through the data variance σk2 . The identifiability of the parameters
of a given model can thus be evaluated provided assumptions on the residual variance. Different
model structures can thus be compared against each other to find out which gives the best parameter
identifiability.

The Akaike information theoretic criterion. Let M be a set of model structures that compete
for the best description of an observed test outcome such that the set is

M = {M1 M2 · · · Mnm } (13.46)

The model set may for example correspond to structures of the same type with increasing
number of parameters. With each of these structures Mk (ppk ) is associated a parameter vector p k .
The Akaike information theoretic criterion (AIC) is the most well-known of criteria that can be
employed to select the the model structure M ? and the associated parameters p ? that is optimal
from the perspective of statistical considerations. It suggest choosing

{M ? , p ? } = argmin(argmin QAIC (Mk , p k )) (13.47)

Mk ∈M p

where

QAIC (Mk , p k ) = N −1 (−log(jpdf(zz? |ppk )) + dim(ppk )) (13.48)

When the model structure is fixed, it corresponds to the maximum likelihood estimation of its
parameters. Conversely, if one hesitates between several structures the ones that involves the most
model parameters are most penalized by the second term of the criterion function. The Akaike
criterion thus favours models with few parameters, i.e. models that are as simple as possible.

13.2.1 Parameter identifiability and model distinguishability

Once a model structure has been chosen its properties should be studied as independently as possi-
ble of the values taken by its parameters, i.e. as independent as possible from the various models
that are contained within the given model structure. As one cannot be certain about what model
structure that fits best to test data, also a set of competing model structures should ideally be studied
to find out which has the best potential to the calibration and validation process. Preferably, these
studies should take place before testing is done to detect potential problems before collection of
data. Two generic properties of interest to these are of importance in modeling, namely parameter
identifiability and model distinguishability. These properties are said to be generic (or structural),
since if they are true, they are true for the entire domain of parameters or model structures they span.
13.2 Parameter estimation statistics 271

Parameter identifiability. We consider both a dynamic real-world system and parallel to

that a model structure M (pp) with a parameter setting that is to be determined by test data from
the system. Before starting the test and doing a parameter calibration, it seems natural to ask if
a successful calibration will be possible. That is whether the planned test will contain enough
information for the estimation of the optimal parameter setting p ? . Formulated in such vague terms,
the question has no answer, so we need to consider a generalized framework in which
• the real-world system and the model have identical structure, i.e. they are consistent,
• the test data are free from noise,
• the character of the input s and the measurement duration can be chosen at will.
Under these idealized conditions it is possible to calibrate the model parameters such that the
input-output relation, at any time and for any input, of the model is identical to that found for the
real-world system in the test. Let the true parameter setting of reality be p ? (= the oracle). We wish
to know whether this identical input-output behavior implies that the calibrated parameters p̃p of
the model equal those of reality p ? . More precisely, the parameter vector p is structurally globally
identifiable (s.g.i.) if for almost any possible setting of we have that
M ( p̃p) = M (pp? ) ⇒ p̃p = p ? (13.49)
When one cannot prove that the model structure considered is globally identifiable, one may
try to establish that it is at least locally identifiable. The parameters p will be structurally locally
identifiable (s.l.i.) if for almost any possible setting of p ? there exists a neighborhood domain
D(pp? ) such that
p̃p ∈ D p ? )

⇒ p̃p = p? (13.50)
M ( p̃p) = M (pp? )
Local identifiability is therefore a necessary condition for global identifiability. The model struc-
ture M (pp) is considered to be structurally locally identifiable if all its parameters are structurally
locally identifiable. If the parameters p are not locally identifiable they are said to be structurally
unidentifiable (s.u.i) and the corresponding model structure M is considered to be structurally
unidentifiable if any of its parameters are structurally unidentifiable. A method for testing for
structural identifiability of a linear state-space model structure is described next.

Identifiability testing by a similarity transformation approach. Assume that the real-world

system generating the test data can be fully described by the state-space realization
ẋx = A (pp? )xx + B (pp? )ss(t) (13.51a)
r (t) = C (pp? )xx + D (pp? )ss(t) (13.51b)
and the model M (pp? ) is thus the matrix quadruple {A
A(pp? ), B(pp? ),C
C (pp? ), D(pp? )}. Let p̃p = T p? ,
in which T is a non-singular square matrix. Then the transformed equations
T −1 x̃x + T B (pp? )ss(t)
x̃x˙ = T A (pp? )T (13.52a)
T −1 x̃x + D (pp? )ss(t)
r (t) = C (pp? )T (13.52b)
will obviously have the same input-output behavior as M (pp? ). They will correspond to a model if
and only if
A ( p̃p) = T A (pp? )TT −1 (13.53a)
B ( p̃p) = T B (pp? ) (13.53b)
?
D ( p̃p) = D (pp ) (13.53c)
 272 Chapter 13. Property Variability and Model Uncertainty

which is a sufficient set of conditions for M (pp? ) = M ( p̃p). From Kalman’s algebraic equivalence
theorem, this set of conditions is also necessary, provided that M (pp? ) is observable and control-
lable. The structural identifiability of M can then be tested by looking for all solutions for ( p̃p, T )
of these equations [6] . If for almost any p ? the only solution is ( p̃p, T ) = (pp? , I ), the model struc-
ture M is s.g.i. If for almost any p ? the set of solutions for p̃p is finite, the model structure M is s.l.i.

Model distinguishability. In structural dynamics modeling of complex systems, the possible

to vary the model structures seems endless. It is then natural to ask whether the test to be performed
on the real world item will make it possible to decide which one is best. This question is that of
the distinguishability of model structures, which receives a partial answer in the same idealized
framework as model identifiability. One thus assumes that the real world item is a model with
model structure M˜ while its model has the model structure M , which now differs from M˜. The
parameter vector associated with M is p and that associated with M˜ is p̃p. The parameter vectors
p and p̃p are not necessary of the same dimension. Since the real world item and its model no longer
necessarily are consistent to have the same model structure, it may become impossible to tune the
parameters p of the model so as to obtain the same input-output behavior as that of the real world.
It is this impossibility that may permit the falsification of model structure M in favour of model
structure M˜.
More precisely, M will be structurally distinguishable (s.d.) from M˜ if, for almost any feasible
values of p̃p, there are no feasible values of p such that M (pp) = M˜( p̃p).
Note the asymmetry of the previous definition. The fact that M is s.d. from M˜ does not
imply that the converse is true. One class of model structures may include the other, see Fig.
12.2. Whenever M is s.d. from M˜ and M˜ is s.d. from M , M and M are said to be mutually
structurally distinguishable.
The techniques to test pairs of model structures for distinguishability are quite similar to those
used for identification testing [47] . Note, however, that one now hopes to prove the non-uniqueness
of a solution for p , whereas in identifiability studies one hope to prove the uniqueness of the
parameter solution. For more details, see Ref. [47].

13.3 Model structure selection

(Not written yet. Should also include discussion on parameterization and error localization)
 14. Model Calibration Procedures

Before commencing a calibration one needs access to processed test data z X and a parameterized
FE model that can be used to calculate the related entities z A (pp). We may split the data set into one
calibration set with data {zzδ X , zδ A } and a validation set with data {zzγX , zγA }. For calibration we
define a vector-valued deviation metric δ such that
δ (pp) = z δ A (pp) − z δ X (14.1)
from which we form a scalar quadratic deviation metric as
Q(pp) = ||δδ (pp)||22 = δ T δ (pp) (14.2)
It is the task of the calibration procedure to minimize the metric Q(pp), i.e. to search for the
parameter values p = p ∗ that brings the model to best agreement with experimental data. This
is normally a nonlinear minimization procedure in the sense that the metric is nonlinear in the
parameters p .
This chapter treats some well-known procedures to do such minimization. The chapter starts
with a brief description of the numerical minimization of a function of a single parameter on which
the multivariate minimization methods rest. It leads up to the celebrated Levenberg-Marquardt
algorithm that is the favorite for many persons active in the field of model calibration.

14.1 Minimizing a single-variable-function

Many method for the numerical minimization of a multivariate function rely on fundamental
methods for minimizing a continuous function of a single variable. Therefore, before turning to
the multivariate minimization problem a brief summary of one such method, the Newton-Raphson
method, is given here. The outset is the truncated Taylor series of the function f (x) in which x is the
scalar variable and f is the function assumed to be continuous on the interval [a, b] and also have
first and second order derivatives with respect to x that are continuous. Let the truncated Taylor
series approximation f˜ ≈ f of f in the neighbourhood of xk ∈ [a, b] be
1
f˜ = f (xk ) + (x − xk ) f 0 |x=xk + (x − xk )2 f 00 |x=xk (14.3)
2
276 Chapter 14. Model Calibration Procedures

with a gradient that is zero at the minimum, i.e. that fulfills the criterion

f˜0 (x) = f 0 |x=xk + (x − xk ) f 00 |x=xk = 0 (14.4)

An extreme point x̃ of this approximating function is thus

x̃ = xk − f 0 |x=xk / f 00 |x=xk (14.5)

This is the foundation of the Newton-Raphson iterations for the convergence to an extreme
point of the true function in which successively better approximations to the extreme point than
xk are taken as the extreme point of the approximating function f˜, i.e. the update xk+1 for the
extreme point is xk+1 = x̃. One notes that the quotient − f 0 |x=xk / f 00 |x=xk is the correction from xk
that leads towards the better approximation to the extreme point. On the other hand the quotient
+ f 0 |x=xk / f 00 |x=xk would load away from the extreme point provided that xk is not at the extreme for
which f 0 |x=xk = 0. Since at a minimum the second order derivative f 00 |x=xk is positive this attraction
and repulsion from the minimum can therefore be used to the advantage for an iteration scheme
that converges to a minimum for which updates are evaluated as

xk+1 = xk − f 0 |x=xk /| f 00 |x=xk | (14.6)

and the sequence x0 , x1 , x2 , . . . converges to a minimizing argument xmin of f (x) provided that no xk
are maximizing arguments of f (x). Since at the closed interval [a, b] the function minimum may
also be at the extreme ends of the interval without that its gradient is there zero, the function values
at those ends also need to be computed and compared to f (xmin to find the true minimum in the
interval.
For practical reasons it may be infeasible to access the function gradient f 0 (x) and another
computational route by numerical differentiation of the function needs to be taken. Using the finite
difference numerical schemes for the first and second order derivatives we have for a small numeric
perturbation ∆x that

f (xk + ∆x) − f (xk − ∆x) f (xk + ∆x) − 2 f (xk ) + f (xk − ∆x)

f 0 (xk ) ≈ f 00 (xk ) ≈ (14.7)
2∆x ∆x2
Using that approximation for the gradients in Eq. (14.5) we have the gradient-free update of the
minimizing algorithm

f (xk + ∆x) − f (xk − ∆x)

xk+1 = xk − ∆x (14.8)
2| f (xk + ∆x) − 2 f (xk ) + f (xk−1 − ∆x)|

An iterative scheme can thus be devised as the following algorithm:

N EWTON -R APHSON M INIMIZATION A LGORITHM

1: procedure N EW R AP(F , x0 , a, b, ∆x, δ , x∗ ) . In: Function, guess, bounds [a, b], perturbation,
absolute tolerance
2: xk = x0 , D = 2δ
3: while D > δ do . Loop until convergence on x
4: D ← xk
5: f = F (xk ), f+ = F (xk + ∆x), f− = F (xk − ∆x)
6: f 0 = ( f+ − f− )/∆x
7: f 00 = ( f+ − 2 f + f− )/∆x2
8: xk ← xk − f 0 /| f 00 |
 14.2 Minimizing a quadratic functional 277

9: D ← D − xk
10: end while
11: x ∗ = xk
12: if F (a) < F (x∗ ) then
13: x∗ ← a
14: end if
15: if F (b) < F (x∗ ) then
16: x∗ ← b
17: end if
18: return x∗ . The minimizing x on the range [a, b]
19: end procedure

14.2 Minimizing a quadratic functional

In all calibration situations, it is possible to formulate a quadratic deviation functional Q from the
deviation metric δ (pp) as
1 nd 2 1
Q(pp) = ∑ δn (pp) = δ T δ (pp) (14.9)
2 n=1 2

Calibration problems of the type p ∗ = argmin Q(pp) with the quadratic structure of are nonlinear
parameter estimation problems. If the calibrated model have credibility, we can expect Q(pp∗ ) to be
small. We will require that the number of data in the deviation vector nd is greater than the number
of calibration variables np and thefore it will generally not be possible to obtain Q(pp∗ ) = 0.
Although the function Q in Eq. (14.9) can be minimized by a general unconstrained minimiza-
tion method, in most situations the properties of Q make it worthwhile to use methods designed
specifically for least-squares minimization problems. In particular, the gradient vector (the Jaco-
bian) and the second order derivative matrix (the Hessian) of (14.9) have special structures that can
be exploited. Let the nd × np Jacobian matrix of δ (pp) be denoted J δ (pp), and let the matrix H nδ
denote the np × np Hessian of the deviation vector element δn (pp). Then the Jacobian of Q, i.e. J Q ,
can be written

J Q (pp) = J Tδ (pp)δδ (pp) (14.10)

and the Hessian of Q, i.e. H Q , is

nd
H Q (pp) = J Tδ J δ (pp) + ∑ δn (pp)H
H nδ (pp) , J Tδ J δ (pp) + S Q (pp) (14.11)
n=1

From Eq. (14.11) we observe that the Hessian of the objective function Q consist of a special
structure of first-order and second-order information. That structure is utilized by the iterative
non-linear least squares methods. The least-squares methods are typically based on the premise
that eventually, as the iterative search for the minimum goes on, the first-order term J Tδ J δ of Eq.
(14.11) will dominate over the summation term S Q as the deviations δ n shrink. This assumption
is not justified when the metric Q(pp∗ ) at the solution p ∗ is large which will happen when one or
more δn (pp∗ ) are large and when Q(pp∗ ) is in the order of the largest eigenvalue of J Tδ J δ (pp∗ ) or
larger. That, in turn, might be due to the use of very noisy test data to construct δ or a poor model
structure. In such cases, one might as well use a general unconstrained minimization method and
then judge if the calibrated model is reliable. However, for many calibration problems the metric at
the termination of the calibration iterations is indeed small enough to justify the use of a special
method.
 278 Chapter 14. Model Calibration Procedures

14.2.1 Newton’s Method

Newton’s method is an iterative method that converges to a (local or global) minimum of Q from a
given guess p0 of the minimizing parameter setting. Let pk , k = 0, 1, 2, . . . be the current estimate
of the solution at the k:th algorithmic iteration step at which the deviation function is Qk , Q(ppk ).
Any quantity subscripted by k will indicate that it is evaluated at this step. If we do a Taylor
series expansion of Q to second order in the perturbation q about the current estimate p k in which
Q(ppk ) , Qk such that p = p k + q we have
1
Q(ppk + q ) = Qk + J TkQ q + q T H kQ q (14.12)
2
We note that the search is for the optimizing perturbation q ∗ that minimizes Q and gives the
optimum parameter setting p ∗ = p k + q ∗ . The minimum of Q in Eq. (14.12) will be achieved if q k
is a minimizer of the quadratic function
1
R(qq) = J TkQ q + q T H kQ q (14.13)
2
where the stationary point satisfy the linear equation

dR/dqq|q =qqk = J kQ + H kQ q k = 0 (14.14)

A minimization algorithm in which the perturbation q k is defined as the solution of Eq. (14.14) is
called a Newton method, and its solution is called the Newton direction.
If H kQ is positive definite, only one iteration is required to reach the minimum of the function
in (14.13) from any starting point p 0 , i.e. p ∗ = p 0 + λk q ∗k with the iteration step length parameter
λk = 1 and q ∗k is the solution to Eq. (14.14). Therefore, we can expect good convergence from
the Newton method when the quadratic descriptor (14.12) is accurate and no higher order Taylor
series terms are required for a good approximation. For a general nonlinear function Q, the Newton
method converges quadratically to p ∗ if p 0 is sufficiently close to p ∗ but convergence also requires
that the Hessian matrices H Q are positive definite at p ∗ and the step lengths λk converge to unity.
However, for many problems in model calibration, the deviation metric Q is highly nonlinear in
the parameters p and other higher-order methods suit better. A first step to these is through the
Gauss-Newton method that is described next.

N EWTONS ’ S N ONLINEAR P ROGRAMMING A LGORITHM

1: procedure N EWTON(pp0 , z X , M , ε, p ∗ ) . Inputs: start guess, data, model and stop value
2: ∆Q = 2ε; p k = p 0 ; k = 0
3: while ∆Q > ε do . Loop until convergence on Q
4: δ ← D EVIATION(zzX , p k , M ) . Compute deviation δ at p k
5: J δ ← JACOBIAN(ppk , M ) . Compute Jacobian of δ at p k
6: H nδ ← H ESSIANS(ppk , M ) . Compute Hessians of δn at pk
7: q ← −[JJ Tδ J δ + ∑nn=1
d
δn H nδ ]−1 J Tδ δ . Update search direction
8: ∗
λ ← argmin Q(ppk + λ q ) . One-dimensional search for minimum of Q
λ
9: p k+1 ← p k + λ ∗ q
10: ∆Q ← Q(ppk+1 ) − Q(qqk )
11: p k ← p k+1
12: end while
13: return p ∗ . The calibration parameter setting
14: end procedure
 14.2 Minimizing a quadratic functional 279

14.2.2 Gauss-Newton’s Method

Gauss developed the Newton method further for computational efficiency. From Eqs. (14.10) and
(14.11), the Newton equations (14.14) become

[JJ Tkδ J kδ + S Q ]qqk = −JJ Tkδ δ k (14.15)

The solution of Eq. (14.15) gives the Newton direction q N = q k . The k:th step iterative
minimization of the function Q is then made by a one-dimensional search for the minimum along
the direction p = p k + λ p N with scalar parameter λ for which the solution is p k+1 = p k + λ ∗ q N
determined as λ ∗ = argmin Q(pp). If Q(ppk ) tends to zero as pk approaches the minimizing solution,
the sum of data Hessians ∑21 also tends to zero. Thus, the Newton direction can be approximated by
the solution of the equations

J Tkδ J kδ q k = −JJ Tkδ δ k (14.16)

Note that the Eq. (14.16) involves only first derivatives of the deviation vector δ . The solution
of Eq. (14.16) is a solution of the linear least-squares problem

q ∗k = argmin||JJ kδ q k + δ k ||22 (14.17)

pk

which is unique if J kδ has full rank. The perturbation vector that solves (14.17) is called the
Gauss-Newton direction, and will be denoted q GN , q ∗k . The method in which q GN is used as a
search direction is known as a Gauss-Newton method.
If J kδ is of full rank, the Gauss-Newton direction approaches the Newton direction as sum tends
to zero in the sense that if for a sufficiently small positive scalar ε, then ||qqN − qGN ||/||qqN − qN || =
O(ε).
Consequently, if δ (pp∗ ) is very small and the columns of J ? are linearly independent, the
Gauss-Newton method can ultimately achieve a quadratic rate of convergence, despite the fact that
only first derivatives are used to compute q GN .
In implementations of the Gauss-Newton method great care is taken to estimate the rank of J kδ .
It is seen in Eq. (14.16) that a rank deficient J kδ cause a singular equation system, and therefore a
nearly rank-deficient matrix J kδ will make the equation system ill-conditioned rendering errors in
determining the search direction.
Ill-conditioning is a common feature of nonlinear least-squares parameter identification prob-
lems if parameter identifiability has not been ascertained. It often manifests itself by that the
deviation metric is practically independent of variation of one or more model parameters or along a
variation of a combination of parameters. Algorithm robustification is normally made by involv-
ing the singular value decomposition in the solution of Eq. (14.16) in which the determination
of the rank of J kδ plays an important role for estimation of singular value rejection. The rank
estimation is determined by approximation methods. When Q is actually close to an ill-conditioned
quadratic function, the best strategy is normally to allow the maximum possibly estimation of
the rank. However, when J kδ is nearly rank-deficient, a generous estimate of the rank tends to
cause very large elements in the solution of Eq. (14.16) for the search direction. This causes
large parameter variation for even small steps along the search direction along which the quadratic
deviation function vary very little. This is an unwanted feature in the parameter estimation process
in which we want the solution to stay at the nominal configuration for insensitive parameters. This
has motivated the introduction of parameter regularization as in the Levenberg-Marquardt method
that is described next.
 280 Chapter 14. Model Calibration Procedures

14.2.3 Levenberg-Marquardt’s Method

A good alternative to the Gauss-Newton method for minimization problems with ill-conditioned
Hessian is the Levenberg-Marquardt method. It is also an alternative to the commonly used
regularization method of augmenting the criterion function by penalizing deviation from the
nominal parameter setting p 0 by the modified target function

Qreg = Q + κ{pp − p0 }T {pp − p0 } (14.18)

where κ is known as a regularization parameter, a positive real scalar. The downside of modifying
the criterion function by augmentation as in Eq. (14.18), and thus minimizing another function Qreg
that is not the primary target, is avoided by the Levenberg-Marquardt method. However, the upside
of regularization in the form of a modified search direction that penalize line-searches away from
p 0 is still utilized by the method. The Levenberg-Marquardt search direction in the k:th iteration
step q LM , q k is defined as the solution of the equations

[JJ Tkδ J kδ + κII ]qqk = −JJ Tkδ δ k (14.19)

We may note that it differs from the calculation of the Gauss-Newton search direction of Eq.
(14.16) in that the approximation to the Hessian is augmented with the Hessian of the regularizing
second term in Eq. (14.18). A unit step for λk , i.e. q ∗ = λk q k with λk = 1, is often taken along
q k that leads to p k+1 = p k + q k . Such procedure eliminates the need for the one-dimensional line
search and might affect the convergence rate. It can be shown that, for some scalar related to , the
vector is the solution of the constrained sub-problem

By that a unit step in the Levenberg-Marquardt direction is taken at each iteration step, it
makes it a so-called trust-region method. As such a good value of κ must be chosen in order to
ascertain descent. If κ is zero, q k is the Gauss-Newton direction and as κ → ∞ the ||qqk || → 0 and
q ∗ becomes identical to the search direction of the well-known steepest-descent method, see e.g.
Ref. [18]. This implies that Q(ppk + q k ) < Qk for sufficiently large regularization parameter κ. As
an alternative, the regularization parameter κ may be fixed to and the iterate minimum be found by
one-dimensional line-search from along the direction .
The usefulness of the Levenberg-Marquardt algorithm for the calibration of computational
structural model is because that these models are often overparameterized. This, or other reasons,
normally makes some parameters very little identifiable from test data. That manifests itself by
large parameter covariance estimates and is related to the Fisher information. By the use of the
Levenberg-Marquardt method, the marginally identifiable parameters do not change much from
iteration to iteration and the calibrated solution is close to the initial parameter setting for such
parameters. That is, by many, considered to be a sympathetic property of the method. This is in
contrast to the results obtained by the Gauss-Newton method under the same circumstances.
All methods here considered, are seen to use function value and gradient information only. Non
uses the true Hessian and are thus Hessian-free. That is a huge benefit over Hessian-based methods,
since the numerical methods of obtaining the Hessian comes with a very high cost. The calculation
of the gradient vectors are needed however, and a way of computing these are described next.

14.3 Computational aspects

TBW

14.3.1 Gradients from finite differences

TBW
 14.3 Computational aspects 281

14.3.2 Parameter settings and randomized starts

TBW
 15. Validation and Cross-Validation

A model should be developed for a specific application purpose and its validity determined with
respect to that purpose. If the purpose of a model is to answer a variety of questions, the validity of
the model needs to be determined with respect to each question. Numerous sets of experimental
conditions, so-called test frames, are usually required to define the domain of a model’s intended
applicability. A model may be valid for one set of experimental conditions and invalid in another. A
model is considered valid for a set of experimental conditions if the model’s accuracy is within its
acceptable range, which is the amount of accuracy required for the model’s intended purpose. This
usually requires that the model’s output variables of interest be identified and that their required
amount of accuracy be specified.
While model validation targets the validation of a model for its intended use, statistical cross-
validation is another animal. Model validation is evaluated towards a specified validation criterion,
sometimes addressed by statistical means such as by use of hypothesis testing. Statistical cross-
validation targets the statistical properties of the model prediction deviation (often called the
prediction error). It closely follows the observation that a model normally better predicts the
calibration data than it does the validation data. By doing multiple choices of splitting data into
calibration and validation data sets, the statistical properties of model prediction can be assessed.
This chapter also discusses the cross-validation concept.

15.1 Classical validation

The normal procedure in validation is that the test data and the FE model data counterpart are split
into one calibration set with data {zzX X zX X
δ , z δ } and a validation set with data {z γ , z γ }. For validation,
the deviation between the model outcome and the test outcome is formulated as
γ (pp) = z X A
γ − z γ (p
p) (15.1)
and a validation accuracy metric is defined from this deviation of Nγ data. Usually a quadratic
validation accuracy metric is desired because of its positiveness. Let that metric, normalized with
the number of data Nγ , be denoted Γ by where
Γ(pp) = γ T γ /Nγ (15.2)
 284 Chapter 15. Validation and Cross-Validation

and it is obvious that this is zero only when all individual validation data deviations in γ are zero.
The fulfillment of the accuracy requirement then determine whether the model is valid for its
intended purpose or not. With a given validation requirement Γ? , one thus has that the model is
valid if

Γ ≤ Γ? (15.3)

while the model is falsified when

Γ > Γ? (15.4)

While this obviously gives a sharp mathematical split between validated and falsified models,
in practice this border is not always that sharp since the matter of selecting Γ? is often set in some
subjective balance of the expected of test data accuracy and the risk of accepting a false model
as being validated. Models that give Γ ≈ Γ? are therefore in a grey-zone. This chapter discusses
validation in this context, i.e. validation for the model’s intended use from a pragmatic sense and
also from a more statistical sense.

15.1.1 Pragmatic validation

In pragmatic model validation, the validation criterion is evaluated and if met, the model is
considered to be valid. The criterion can for instance be of the type that a quadratic metric Γ
should be smaller than a given threshold Γ∗ . However, only the human imagination sets the limits
for possible validation criteria. Sometimes the validation criterion is stated on the basis on MAC
correlation numbers for a set of named eigenmodes, e.g. for aircraft model validation one criterion
could be that “the fundamental wing bending modes from testing and analysis should correlate with
a MAC number larger than 0.90”. However, many other validation formulations are in use. No one
is absolutely right, and no one is absolutely wrong but they need to be in good agreement with the
application in mind.
Pragmatic validation does not consider variability. It considers data from a predetermined set of
experimental conditions, evaluates the validation deviation metric, and decides whether the model
is validated of falsified. Variability is usually not considered because of limited resources allocated
for validation. Such variability may be due to stochastic processes that makes the test outcome
from seemingly identical test conditions non-repeatable. But it can also be due to non-unique data
handling by different test operators or be due to spread between test-piece individuals. Another
possibility is that it is due to impreciseness in the validation metric. Should such variability be
considered, the crispness of the validation criterion is lost and the picture becomes more blurred.
To make decisions on validation/falsification to stand on more firm statistical ground a hypothesis
test procedure may be of help. In that the null-hypothesis for validation and the alternate hypothesis
are set up according to the following.
The positive outcome of the model validation fulfils the hypothesis:

• H0 : The model fulfils the validation criterion and is valid under the experimental frame.

Otherwise, the model is falsified and fulfils the alternative hypothesis:

• H1 : The model does not fulfil the validation criterion and is therefore invalid under the
experimental frame.

When variability is considered, these hypotheses can be scrutinized by statistical hypothesis

testing. However, for pragmatic validation, it should be acknowledged that there are two possibilities
for making a wrong validation decision. The first one, which defines what is called the type I error,
is accepting the alternative hypothesis H1 when the null hypothesis H0 is actually true. The second
 15.1 Classical validation 285

Table 15.1: Possible outcomes of validation decisions.

Accuracy state of the model
H0 true H1 true
Action (model is valid) (model is false)
Accept truth of H0 Correct decision Model user’s risk
Accept truth of H1 Model builder’s risk Correct decision

one, the type II error, is accepting the null hypothesis when the alternative hypothesis is actually
true. The probability of making the first type of wrong decision is called the model builder’s risk
and the probability of making the second type of wrong decision is called the model user’s risk.
The sum of the probabilities of making the correct validation or falsification decision together with
the user’s and modeler’s risk is 1 (one). A Type III error occurs when the wrong problem is solved
and is committed when the formulated problem does not completely contain the actual problem.
An illustration of the possibilities are summarized in Table 15.1.

15.1.2 Validation guidelines

The substantiation that a model is valid, i.e. performing model verification and validation, is
generally considered to be a process and is usually part of the overall model development process.
The amount of model accuracy required should be specified prior to starting the development of the
model or very early in the model development process. Several versions of a model are usually
developed prior to obtaining a satisfactory valid model. It is often too costly and time consuming to
determine that a model is absolutely valid over the complete domain of its intended applicability.
Instead, tests and evaluations are conducted until sufficient confidence is obtained that a model
can be considered valid for its intended application. The cost of model validation is usually quite
significant, especially when extremely high model confidence is required. If a test determines that a
model does not have sufficient accuracy for any one of the sets of experimental conditions, then
the model is falsified. However, determining that a model has sufficient accuracy for numerous
experimental conditions does not guarantee that a model is valid everywhere in its applicable
domain.
Osman Balci, see [4] and [39], has stated 15 advice for model validation that are worthwhile
to consider to increase the probability of validation success if followed properly. They treat the
verification and validation (V&V) in the modelling and simulation (M&S) context. He calls them
golden rules, which are:

i) V&V for M&S should be integrated within the entire M&S application development life
cycle. It should not just be a step in the M&S development life cycle, but a continuous activity
throughout the entire life cycle since accuracy is not something that can be imposed upon after
the fact; it has to be assessed while the work is being performed.
ii) The VV&T outcome should not be considered as a binary variable where M&S accuracy
is either perfect or totally imperfect. A model is an abstraction and a perfect representation
of reality is never expected. Therefore, M&S validity is not binary, where valid implies perfectly
accurate and falsified implies totally inaccurate. The M&S accuracy should be judged on a
scale defined by nominal scores such as Excellent, Very Good, Satisfactory, Marginal, Deficient,
and Unsatisfactory.
iii) The M&S is for a prescribed set of intended uses and its accuracy should be judged with
respect to those uses. A model is a representation and can be created in different ways
depending on the objectives for which the model is intended. The intended uses dictate
 286 Chapter 15. Validation and Cross-Validation

how representative the M&S application should be. Sometimes the nominal score very good
accuracy may be sufficient; sometimes excellent accuracy may be required. The required
level depends of the criticality of the decisions to be made based on the M&S. The adjective
sufficiently should be used in conjunction with terms such as accuracy, verity, validity, quality,
and credibility, to indicate that the judgement is made with respect to the prescribed set of
intended uses. It is more appropriate to say the model is sufficiently valid for its intended use
than saying the model is valid.
iv) The VV&T requires independence to prevent developer’s bias. It is meaningful when
conducted in an independent manner by an unbiased agent who is independent to the M&S
application developer. The people involved in M&S application development may be biased
when it comes to VV&T, because they fear that negative VV&T results may be used against
them. VV&T should be conducted under true technical, managerial, and financial independence.
v) VV&T is difficult and requires creativity and insight. It is difficult due to many reasons
including lack of data, lack of sufficient problem domain-specific knowledge, lack of qualified
subject matter experts, many qualitative elements to assess, and inability to effectively employ
M&S developers due to their conflicts of interest. Designing an effective test, identifying test
cases, and developing a test procedure require creativity and insight. VV&T experience is
required to be able to determine which of the many V&V techniques are most effective for a
given V&V task.
vi) VV&T is situation dependent. It is applied depending on the particular accuracy assessment
task, the M&S type, size and complexity together with the nature of the artifact subjected to
VV&T. A number of most effective VV&T techniques for one situation may not be so for
another. The VV&T approach, techniques, and tools should be selected depending on the task
at hand.
vii) The VV&T accuracy can be claimed only for the intended uses for which the M&S is
tested. The M&S accuracy is assessed using VV&T for a particular intended use, under which
its input conditions are defined. An application that works for one set of input conditions under
a given intended use may produce absurd output when conducted under another set of input
conditions.
viii) Complete testing is not possible for M&S. A saying is that “The only exhaustive testing is so
much testing that the tester is exhausted!”. Exhaustive testing requires that the M&S is tested
under all input conditions possible for the application. Due to time and budget constraints, it is
impossible to test the accuracy of all these. The question is not how much test data are used,
but what percentage of the potential model input domain is covered by the test data. The higher
the percentage of coverage the higher the confidence we can gain in model accuracy.
ix) The VV&T activities should be considered as confidence-building activities. We cannot
claim 100% accuracy due to M&S complexity, lack of data, reliance on qualitative human
judgement, and lack of complete testing. The VV&T activities are conducted until sufficient
confidence in M&S accuracy is gained. Accuracy is certainly the most important quality
indicator and VV&T is conducted to assess it. However, for a large and complex M&S
application, we are unable to substantiate sufficient accuracy with 100% confidence. Assessment
of other quality indicators in the VV&T help us build up our confidence in sufficient accuracy
of the M&S application.
x) The M&S VV&T activities should be planned and documented throughout the entire
M&S development life cycle. The VV&T activities should not be conducted in an ad hoc
fashion. Planning is required for; (i) scheduling VV&T tasks throughout the entire application
development life cycle, (ii) identifying software tools to acquire data, (iii) identifying method-
ologies and techniques for validation, (iv) assigning roles and responsibilities, and (v) allocating
resources such as personnel, facilities, tools, and finances. All activities should be documented
 15.2 Cross-validation 287

for certification, regression testing, re-testing, and re-certification. All artifacts such as test
designs, test data, test cases, and test procedures should be documented and preserved for re-use
during the maintenance stage of the application life cycle.
xi) Errors should be detected as early as possible in the application development life cycle.
The M&S application development must start with problem formulation and must be carried
out process-by-process in an orderly fashion in accordance with a comprehensive blueprint
of development. Skipping the early stages of development and jumping into programming is
an build-and-fix approach that must be avoided. Detection and correction of errors as early as
possible in the development life cycle results in reduced time and assures better quality. Some
vital errors may be hard to detect in later stages of the life cycle due to increased complexity. It
is relatively easier to detect, localize and correct errors in an incremental manner as the M&S
development progresses.
xii) The double validation problem should be recognized and resolved properly. A typical
VV&T test is ideally conducted by running the M&S model with the same input data that drive
the system, and then comparing the model and system outputs to determine how similar they are.
The amount of correspondence between the model and system outputs is examined to judge the
validity of the model. However, in conducting this validation test, another validation test should
be recognized and performed before this test. That validation test deals with substantiating
that the model and system stimuli match each other with sufficient accuracy. This test is also
referred to as stimulus validation, which must be successfully performed before the model
validation test.
xiii) Successfully testing each sub-model does not imply overall model validity. Models repre-
senting subsystems can be tested individually. Each sub-model can be found to be acceptable
with respect to the intended uses with some tolerable error in its representation. However, the
allowable errors for these may accumulate to be unacceptable for the whole model. Therefore,
the whole model must be tested even if each sub-model is individually found to be acceptable.
xiv) Formulated problem accuracy greatly affects the acceptability and credibility of M&S
results. A saying is “a problem correctly formulated is half solved”. The M&S life cycle starts
with problem formulation. Based on the formulated problem, the system or domain containing
the problem is defined and its characteristics are identified. Based on the defined problem
domain, M&S application requirements are engineered and the requirements become the point
of reference for the M&S application development throughout the rest of the life cycle. An
incorrectly defined problem results in simulation results that are irrelevant. Formulated problem
accuracy greatly affects the credibility and acceptability of simulation results. Sufficient time
and effort must be expended to properly define the problem..
xv) Type I, II and III errors should be recognized and prevented. Committing Type I error.
i.e. increasing the modeler’s risk and reject a model that is sufficiently credible, unnecessarily
increases the M&S application development cost. To do Type III errors, i.e. solving a problem
that does not completely contain the actual problem, can be catastrophic especially when critical
decisions are made on the basis of M&S application results.

15.2 Cross-validation
Cross-validation is normally used as a statistical means for estimating a model’s expected prediction
deviation. It has been used for long, but since it requires intense computing it has not been used
much for large scale problems until recent years. It generally requires more calculations than
calibration, since it makes use of repeated calibrations on subsets of available data.
Cross-validation targets expected squared prediction deviation (PD) of the i:th data quantity,
 288 Chapter 15. Validation and Cross-Validation

i.e.

PDi = E(zziX − z iA (pp))2 (15.5)

in which the expectation E refers to repeated sampling of experimental data z iX .

Closely related to the PD is the residual squared deviation of the prediction defined as

RSD = ||zzX − z A (pp))||22 /Nd (15.6)

where Nd is the number of data, i.e. the length of the data vectors z X and z A . The RSD may be
evaluated separately for validation data and calibration data as

RSDγ = ||γγ (pp))||22 /Nγ (15.7)

and

RSDδ = ||δδ (pp))||22 /Nδ (15.8)

where RSDγ usually is larger than RSDδ since the model has been calibrated to minimize δ .
To evaluate the RSD we would ideally use experimental data that is independent on calibration
data. Usually however, additional test data are not always available for reasons of cost or time. To
get around this, cross-validation uses part of the available data to calibrate the model, and a different
part to evaluate the RSD of it. This splitting of data is repeated multiple times with different data
in the validation and calibration data sets each time to give useful RSD statistics. The following
sections describes four common strategies to do the data splitting.

15.2.1 K-Fold cross-validation

In K-fold cross-validating the available data set is split into K parts (also known as "folds") of
roughly the same size. Let k = 1, 2, . . . , K be the index of the k:th such fold. In the k:th fold,
the number of validation data is Nkγ and the number of calibration data is Nkδ = Nd − Nγk . The
validation deviation of the k:th set is γ k and the corresponding calibration deviation is δ k . The
union of the two sets constitutes the entire data set. The validation sets γ k are mutually unique and
the union of them spans the entire data set.
For cross-validation, calibrations are made K times using the deviation metric Qk (pp) = ||δδ ||2
to obtain a parameter estimate

p k = argmin Qk (pp) (15.9)

p

resulting in a validation RSD for the k:th fold to be

RSDkγ = ||γγ (ppk ))||22 /Nkγ (15.10)

with the cross-validation estimate of the prediction deviation being the mean value

1 K
RSDCV = ∑ RSDkγ (15.11)
K k=1

A special case of K-fold cross validation, when the K folds are taken to the extreme, is the leave-
one-out cross-validation. In leave-one-out cross-validation Nd splits are made which leaves one
single data for validation while the remaining Nd − 1 data are used for calibration in N calibration
runs. While the leave-one-out cross-validation have certain advantages, a distinct disadvantage
in FEM validation is the large number of costly calibrations that have to made which makes the
leave-one-out cross-validation practically infeasible.
 15.2 Cross-validation 289

15.2.2 Bootstrapping cross-validation

TBW

15.2.3 Monte-Carlo cross-validation

TBW

15.2.4 Cross-validation estimates of parameter statistics

The cross-validation procedures above are seen to involve multiple runs for finding the calibration
parameter setting that minimizes the calibration deviation metric for various splitting of available
data. It is natural to use the multiple parameter estimates for statistical evaluation. Evaluation
of mean and covariance of the parameter estimates p k can be done with ease. This can then be
made without explicit knowledge of the noise properties of experimental data. Compared with the
parameter covariance estimate by the Cramer-Rao bound which requires explicit knowledge of the
noise properties, this is a distinct advantage.

15.2.5 Statistical validation

In the statistical model validation, the hypothesis testing of the hypotheses H0 and H1 stated
above plays a central role. Hypothesis tests can be used in the comparison of means, variances,
distributions, and prediction of the output variables of a model and a system for each set of
experimental conditions (i.e. experimental frames) to determine if the simulation model’s output
behavior has an acceptable range of accuracy.
The two types of errors associated with testing validation hypotheses, the type I error of
rejecting the validity of a valid model, and the type II error of accepting the validity of an invalid
model are treated by Balci and Sargent[11.3]. They state on solid ground that model user’s risk
(type II error) is very important and must be kept small. Both type I and type II errors must be
carefully considered when using hypothesis testing for model validation.
The amount of agreement between a model and a system given by the validity metric Γ is
such that the amount of agreement between the model and the system decrease as the value of the
validity metric increases. The acceptable range of accuracy, determined from a given validation
threshold Γ∗ can be used to determine an acceptable validity range 0 ≤ Γ ≤ Γ∗ .
The probability of acceptance of a model being valid, Pv , can be examined as a function of the
validity measure by using an operating characteristic curve, see Johnson[11.4]. Fig. 15.1 contains
three different operating characteristic curve plots to illustrate how the number of observations
frames affects Pv , PI , PII and the probability of falsification, Pf , as a function of Γ. As can be seen,
an inaccurate model has a higher probability of being validated if a smaller number of test frames
of observations is used, and an accurate model has a lower probability of being accepted if a large
number of experimental frames is used. The reason for this probability non-uniqueness is that when
data from new experimental frames are added, there is always a probability that data from new
experimental conditions increase the validation deviation.
Once the operating characteristic curves are constructed, the intervals for the model user’s risk
PII and the model builder’s risk PI can be determined for a given Γ as follows:

0 < model builder’s risk PI < 1 − Pv = Pf (Γ∗ )

0 < model user’s risk PII < Pv (Γ∗ )

Thus there is a direct relationship among the builder’s risk, model user’s risk, acceptable
validity range, and the sample size of observations. A trade-off among these must be made in using
hypothesis tests in model validation.
290 Chapter 15. Validation and Cross-Validation

Figure 15.1: Operating Characteristic Curves of same model with different numbers of test frames.
........................................................................................

Details of the methodology for using hypotheses tests in comparing the model’s and system’s
output data for model validations are beyond the scope of this book. Such details are given by Balci
and Sargent[11.3]. The statistical hypothesis testing requires the knowledge of statistical properties
of test data to be trustworthy. The usual situation in FE model validation with vibration testing is
that there is a lack of information on data statistics. This makes the use of statistical validation in
the hypothesis testing setting impractical and is today more a research instrument.
 Bibliography

[1] Abramowitz, M., Stegun, I.A.: Handbook of mathematical functions. Dover Publications
(1972)

[2] Anderson, B.D.O.: A system theory criterion for positive real matrices. SIAM Journal on
Control and Optimization 5(2), 171–182 (1967)

[3] Antoulas, A.C.: Approximation of large-scale dynamical systems. Society for Industrial and
Applied Mathematics, Philadelphia (2005)

[4] Balci, O.: Golden rules of verification, validation, testing, and certification of modeling and
simulation applications. SCS M&S Magazine 4 (2010)

[5] Bathe, K.J.: Finite Element Procedures. Prentice-Hall (2007)

[6] Berman, M., Schoenfeld, R.: Invariants in experimental data on linear kinetics and the
formulation of models. Journal of Applied Physics 27, 1361–1370 (1956)

[7] Bolotin, V.V.: The dynamic stability of elastic systems. Holden-Day (1964)

[8] Cook, R., Malkus, D., Plesha, M.: Concepts and applications of the finite element method, 4
edn. Wiley (2001)

[9] Courant, R., Hilbert, D.: Methods of Mathematical Physics - Vol I. Interscience Publishers
(1953)

[10] Craig, R.R., Kurdila, A.: Fundamentals of Structural Dynamics, 2 edn. Wiley (2006)

[11] Cramér, H.: Mathematical methods in statistics. Princeton University Press (1946)

[12] Enns, D.F.: Model reduction with balanced realizations: An error bound and frequency
weighted generalization. Proceedings of the IEEE Conference on Decision and Control pp.
127–132 (1984)
294 BIBLIOGRAPHY

[13] Ersal, T., Fathy, H.K., Rideout, D.G., Louca, L.S., Stein, J.L.: A review of proper modeling
techniques. J Dyn Syst Measur Control Trans ASME 130(6) (2008). DOI 10.1115/1.2977484

[14] Ewins, D.J.: Modal Testing: Theory and Practice, 2 edn. Research Studies Press (2000)

[15] Fisher, R.: On an absolute criterion for fitting frequency curves. Messenger of Mathematics
41, 155–160 (1912)

[16] Fung, Y.: Foundations of solid mechanics. Englewood Cliffs (1965)

[17] Geradin, M., Rixen, D.J.: Mechanical Vibrations- Theory and Applications to Structural
Dynamics, 3 edn. Wiley (2015)

[18] Gill, P.E., Murray, W., Wright, M.H.: Practical optimization. Academic Press (1981)

[19] Glover, K.: All optimal hankel-norm approximations of linear multivariable system and their
L∞ -error bounds. International Journal of Control 39, 1115–1193 (1984)

[20] Golub, G.H., van Loan, C.F.: Matrix computations, vol. 3. John Hopkins University Press
(2012)

[21] Halmos, P.: Finite Dimensional Vector Spaces. Van Nostrand, New York (1958)

[22] Hildebrand, F.: Advanced calculus for applications, 2 edn. Prentice Hall (1976)

[23] Inman, R., Conover, W.: A distribution-free approach to introducing rank correlation among
input variables. Communications in Statistics - Simulation and Computation 11(3), 311–334
(1982)

[24] Kailath, T.: Linear Systems. Prentice Hall (1980)

[25] Kammer, D.C.: Sensor placement for on-orbit modal identification and correlation of large
space structures. Journal of Guidance, Control, and Dynamics 14(2) (1991)

[26] Liu, G.R., Quek, S.S.: The Finite Element Method - A Pactical Course. Butterworth-
Heinemann (2003)

[27] Ljung, L.: System Identification - Theory for the User. Pearson Education (1998)

[28] McKay, M., R.J., R.B., Conover, W.: A comparison of three methods for selecting values of
input variables in the analysis of output from a computer code. Technometrics - a journal of
statistics for the physical, chemical and engineering sciences 21(2), 239–245 (1979)

[29] McKelvey, T., Akcay, H., Ljung, L.: Subspace-based multivariable system identification from
frequency response data. IEEE Transactions on Automatic Control 41(7), 960–979 (1996)

[30] Meirovitch, L.: Principles and Techniques of Vibrations. Prentice-Hall (1997)

[31] Moore, B.: Principal component analysis in linear systems: Controllability, observability and
model reduction. IEEE Transaction of Automatic Control AC-26, 17–32 (1981)

[32] Newmark, N.M.: A method of computation for structural dynamics. ASCE Journal of the
engineering mechanics division 85, 67–94 (1959)

[33] Noor, A.K.: Recent advances and applications of reduction methods. App Mech Rev 47(5),
125–146 (1994)
BIBLIOGRAPHY 295

[34] van Overschee, P., de Moor, B.: Subspace Identification for Linear Systems. Kluwer Academic
Publishers (1996)
[35] Overschee, P.V., Moor, B.D.: N4sid subspace algorithms for the identification of combined
deterministic-stochastic systems. Automatica 30(1) (1994)
[36] Pernebo, L., Silverman, L.: Model reduction via state space representations. IEEE Transac-
tions on Automatic Control 27(2) (1982)
[37] Qu, Z.Q.: Model order reduction techniques: with application in finite element analysis.
Springer, New York (2004)
[38] Rahrovani, S., Vakilzadeh, M.K., Abrahamsson, T.J.S.: Modal dominancy analysis based on
modal contribution to frequency response function 2-norm. Mechanical Systems and Signal
Processing 48(1-2), 213–231 (2014)
[39] Sargent, R.G., Balci, O.: History of verification and validation of simulation models. In:
Proceedings of the 2017 Winter Simulation Conference, IEEE, pp. 292–307 (2017)
[40] Schwer, L., Mair, H., Crane, R.: Guide for verification and validation in computational solid
mechanics (2006)
[41] Shabana, A.A.: Dynamics of multibody systems. Wiley (1989)
[42] Tchebichef, P.: Des valeurs moyennes. Journal de Mathématiques Pures et Appliquées 2(12)
(1867)
[43] Vakilzadeh, M.K., Rahrovani, S., Abrahamsson, T.J.S.: Modal reduction based on accurate
input-output relation preservation. In: R. Allemang, J.D. Clerck, C. Niezrecki, A. Wicks
(eds.) Topics in Modal Analysis, vol. 7, pp. 333–342. Springer, New York (2014)
[44] Vakilzadeh, M.K., Yaghoubi, V., McKelvey, T., Abrahamsson, T.J., Ljung, L.: Experiment
design for improved frequency domain subspace system identification of continuous-time
systems. In: 17th IFAC Symposium on System Identification SYSID 2015, pp. 886–891.
IFAC-PapersOnLine (2015)
[45] Varga, A.: Enhanced modal approach for model reduction. Math Model Syst 1(2), 91–105
(1995). DOI 10.1080/13873959508837010
[46] Walter, E., Pronzato, L.: Identification of parametric models from experimental data. Springer
(1997)
[47] Walter, , Raksanyi, Y.L.A., Happel, J.: On the distinguish-ability of parametric models with
different structures. Mathematics and Computers in Biomedical Applications pp. 145–160
(1985)
[48] Wilkinson, J.H.: The Algebraic Eigenvalue Problem. Clarendon Press (1965)
[49] Wilson, D.: Convolution and hankel operator norms for linear systems. IEEE Transactions on
Automatic Control 34(2) (1989)
[50] Wittrick, W., Williams, F.: A general algorithm for computing natural frequencies of elastic
structures. Quarterly Journal of Mechanics and Applied Mathematics XXIV(3) (1971)
[51] Ziegler, H.: Principles of structural stability. Birkhäuser (1977)
[52] Zienkiewicz, O., Taylor, R.: Finite Element Method, 5 edn. Elsevier (2000)
 Index

A Best Subspace Method . . . . . . . . . . . . . . . . . 248

bi-orthogonal . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
accelerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 bi-orthonormal . . . . . . . . . . . . . . . . . . . . . . . . . . 68
accelerometer Bragg cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
alignment error . . . . . . . . . . . . . . . . . . . . 224 Bryant angles . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
calibration error . . . . . . . . . . . . . . . . . . . 224 bungee cords . . . . . . . . . . . . . . . . . . . . . . . . . . 217
caveats . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
charge mode . . . . . . . . . . . . . . . . . . . . . . . 221 C
cross-sensitivity . . . . . . . . . . . . . . . . . . . 224
IEPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 characteristic equation . . . . . . . . . . . . . . . . . . . 57
mounting . . . . . . . . . . . . . . . . . . . . . . . . . 223 characteristic polynomial . . . . . . . . . . . . . . . . 57
overview . . . . . . . . . . . . . . . . . . . . . . . . . . 221 Chebyshev inequality . . . . . . . . . . . . . . . . . . 264
surface preparation . . . . . . . . . . . . . . . . . 223 coefficient of variation . . . . . . . . . . . . . . . . . . 263
voltage mode . . . . . . . . . . . . . . . . . . . . . . 221 COMAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
additivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 compatibility relation . . . . . . . . . . . . . . . . . . . . 23
aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 completely observable . . . . . . . . . . . . . . . . . . 101
analog-to-digital . . . . . . . . . . . . . . . . . . . . . . . 219 consistent model structure . . . . . . . . . . . . . . 241
anti-resonance controllability . . . . . . . . . . . . . . . . . . . . . . . . . 100
frequency . . . . . . . . . . . . . . . . . . . . . . . . . 130 Grammian . . . . . . . . . . . . . . . . . . . . . . . . 103
general . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
approximation convergence
zero-order-hold . . . . . . . . . . . . . . . . . . . . 110 spatial type . . . . . . . . . . . . . . . . . . . . . . . . 149
assembly process . . . . . . . . . . . . . . . . . . . . . . . 26 spectral type . . . . . . . . . . . . . . . . . . . . . . 150
convolution integral . . . . . . . . . . . . . . . . . . 95, 98
B convolution solution . . . . . . . . . . . . . . . . . . . . 106
correlation coefficient
beam non-dimensional . . . . . . . . . . . . . . . . . . . 263
continuous . . . . . . . . . . . . . . . . . . . . . . . . 135 Couchy number . . . . . . . . . . . . . . . . . . . . . . . . 117
element . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Courant’s mini-max principle . . . . . . . . . . . . 60
Euler-Bernoulli . . . . . . . . . . . . . . . . . . . . 135 covariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
298 INDEX

Cramer-Rao inequality . . . . . . . . . . . . . . . . . 268 eigenvalue problem

cross-correlation . . . . . . . . . . . . . . . . . . . . . . . 263 adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
cross-validation eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
bootstrapping. . . . . . . . . . . . . . . . . . . . . .289 electromagnetic shaker . . . . . . . . . . . . . . . . . 220
leave-one-out . . . . . . . . . . . . . . . . . . . . . . 288 element
Monte Carlo . . . . . . . . . . . . . . . . . . . . . . 289 constant strain triangle . . . . . . . . . . . . . . 38
cumulative distribution error function . . . . . . . . . . . . . . . . . . . . . . . . . . 264
marginal . . . . . . . . . . . . . . . . . . . . . . . . . . 262 estimator
efficient . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
D maximum likelihood . . . . . . . . . . . . . . . 268
Euler
damping theorem on rotations . . . . . . . . . . . . . . . . 20
augmented . . . . . . . . . . . . . . . . . . . . . . . . . 65 excitation
Caughey . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 aperiodic . . . . . . . . . . . . . . . . . . . . . . . . . . 205
critical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 burst. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .208
negative . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 chirp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
overcritical . . . . . . . . . . . . . . . . . . . . . . . . . 94 cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
proportional . . . . . . . . . . . . . . . . . . . . . . . . 64 mono-frequency harmonic . . . . . . . . . . 205
Rayleigh . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 multi-frequency harmonic . . . . . . . . . . 205
undercritical . . . . . . . . . . . . . . . . . . . . . . . . 94 periodic . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
data acquisition system . . . . . . . . . . . . . . . . . 219 random . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
deficient system matrix . . . . . . . . . . . . . . . . . . 49 swept-sine . . . . . . . . . . . . . . . . . . . . . . . . 205
degrees-of-freedom . . . . . . . . . . . . . . . . . . . . . 19
active . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 F
condensed . . . . . . . . . . . . . . . . . . . . . . . . 134
diagonal form . . . . . . . . . . . . . . . . . . . . . . . . . . 48 falsified . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
digital-to-analog . . . . . . . . . . . . . . . . . . . . . . . 220 FFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
direct throughput matrix . . . . . . . . . . . . . . . . . 45 filter
discrete time transition matrix . . . . . . . . . . . 114 band-pass . . . . . . . . . . . . . . . . . . . . . . . . . 215
displacement method . . . . . . . . . . . . . . . . . . . . 26 high-pass . . . . . . . . . . . . . . . . . . . . . . . . . 215
distribution narrow-band . . . . . . . . . . . . . . . . . . . . . . 215
Beta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 notch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Gaussian . . . . . . . . . . . . . . . . . . . . . . . . . . 264 Fisher information matrix . . . . . . . . . . . . . . . 269
normal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 force transducer
standard normal . . . . . . . . . . . . . . . . . . . 264 basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
distribution function installation . . . . . . . . . . . . . . . . . . . . . . . . 226
cumulative . . . . . . . . . . . . . . . . . . . . . . . . 261 Fourier transform
joint cumulative . . . . . . . . . . . . . . . . . . . 262 fast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Doppler effect . . . . . . . . . . . . . . . . . . . . . . . . . 228 frequency response function . . . . . . . . 126, 133
Duhamel integral . . . . . . . . . . . . . . . . . . . . . . . 95 Frobenius norm . . . . . . . . . . . . . . . . . . . . . . . . 192
dynamic condensation . . . . . . . . . . . . . . . . . . 134
dynamic equation . . . . . . . . . . . . . . . . . . . . . . . 45 G
dynamic stiffness matrix . . . . . . . . . . . . . . . . 126
Gauss-Newton method . . . . . . . . . . . . . . . . . 279
E generalized eigenvectors . . . . . . . . . . . . . . . . . 51
Gerschgorin’s theorem . . . . . . . . . . . . . . . . . . . 70
eigenproblem Givens’ method . . . . . . . . . . . . . . . . . . . . . . . . . 71
shifted form . . . . . . . . . . . . . . . . . . . . . . . . 76 globally identifiable
eigensolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 structurally . . . . . . . . . . . . . . . . . . . . . . . . 271
eigenvalue matrix . . . . . . . . . . . . . . . . . . . . . . . 49 Gram-Schmidt orthogonalization . . . . . . . . . 79
INDEX 299

Guyan’s method . . . . . . . . . . . . . . . . . . . . . . . 152 M

H MAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Markov parameters . . . . . . . . . . . . . . . . . . . . 106
Hölder norms . . . . . . . . . . . . . . . . . . . . . . . . . . 191 mass matrix
Hölder’s inequality . . . . . . . . . . . . . . . . . . . . . 191 consistent . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Hankel matrix . . . . . . . . . . . . . . . . . . . . . 106, 118 lumped . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Hessian-free . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 material compliance matrix . . . . . . . . . . . . . . 26
Hinton-Rock-Zienkiewicz lumping . . . . . . . 32 material points . . . . . . . . . . . . . . . . . . . . . . . . . . 19
homogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 material stiffness matrix . . . . . . . . . . . . . . . . . 25
hydraulic shaker . . . . . . . . . . . . . . . . . . . . . . . 220 matrix iteration methods . . . . . . . . . . . . . . . . . 76
hyperstatic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
hypostatic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 MOC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Modal
I Observability Correlation . . . . . . . . . . . 181
Observability Correlation, basic . . . . . 184
identifiable Assurance Criterion . . . . . . . . . . . . . . . . 181
locally . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 modal
impulse hammer . . . . . . . . . . . . . . . . . . . . . . . 220 displacements . . . . . . . . . . . . . . . . . . . . . . 59
infinitesimal rotation . . . . . . . . . . . . . . . . . . . . 22 loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
input matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 masses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
internal variables. . . . . . . . . . . . . . . . . . . . . . .153 matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
inverse iteration . . . . . . . . . . . . . . . . . . . . . . . . . 76 relative damping . . . . . . . . . . . . . . . . . . . . 63
isostatic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 stiffnesses . . . . . . . . . . . . . . . . . . . . . . . . . . 59
modal decomposition form . . . . . . . . . . . . . . . 48
J modal matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
mode acceleration method . . . . . . . . . . . 67, 150
Jordan normal form . . . . . . . . . . . . . . . . . . . . . 51 mode displacement method . . . . . . . . . . . . . 150
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
K model distinguishability . . . . . . . . . . . . . . . . 270
model structure . . . . . . . . . . . . . . . . . . . . . . . . 241
Krylov sequence . . . . . . . . . . . . . . . . . . . . . . . . 78 model training . . . . . . . . . . . . . . . . . . . . . . . . . 244
model updating . . . . . . . . . . . . . . . . . . . . . . . . 244
L mutually structurally distinguishable . . . . . 272
laser doppler velocimeter . . . . . . . . . . . . . . . 227
N
scanning . . . . . . . . . . . . . . . . . . . . . . . . . . 227
single point . . . . . . . . . . . . . . . . . . . . . . . 227 natural frequency . . . . . . . . . . . . . . . . . . . . . . . 94
laser interferometer . . . . . . . . . . . . . . . . . . . . 229 damped . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Latin Hypercube Sampling . . . . . . . . . . . . . . 265 Newton direction . . . . . . . . . . . . . . . . . . . . . . 278
leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 Newton method . . . . . . . . . . . . . . . . . . . . . . . . 278
Levenberg-Marquardt method . . . . . . . . . . . 280 Newton-Raphson . . . . . . . . . . . . . . . . . . . . . . 275
likelihood function . . . . . . . . . . . . . . . . . . . . . 268 numerical time integration
linear in parameters . . . . . . . . . . . . . . . . . . . . 157 Newmark . . . . . . . . . . . . . . . . . . . . . . . . . 110
load transducer . . . . . . . . . . . . . . . . . . . . . . . . 220 Nyquist frequency . . . . . . . . . . . . . . . . . . . . . 210
locally identifiable
structurally . . . . . . . . . . . . . . . . . . . . . . . . 271 O
long time passive . . . . . . . . . . . . . . . . . . . . . . 197
lumping observability . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Hinton-Rock-Zienkiewicz . . . . . . . . . . . 32 observability matrix . . . . . . . . . . . . . . . . . . . . 101
300 INDEX

observable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 model builder’s . . . . . . . . . . . . . . . . . . . . 285

oracle model . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 model user’s . . . . . . . . . . . . . . . . . . . . . . 285
oracle model structure . . . . . . . . . . . . . . . . . . 241 rod
output equation . . . . . . . . . . . . . . . . . . . . . . . . . 45 continuous . . . . . . . . . . . . . . . . . . . . . . . . 139
output matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 rod element. . . . . . . . . . . . . . . . . . . . . . . . . . . . .33
over-fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
S
P
sampling frequency . . . . . . . . . . . . . . . . . . . . 113
parameter estimation . . . . . . . . . . . . . . . . . . . 244 Schatten norms . . . . . . . . . . . . . . . . . . . . . . . . 192
maximum likelihood . . . . . . . . . . . . . . . 252 sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
nonlinear . . . . . . . . . . . . . . . . . . . . . . . . . . 277 shaft element . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
parameter identifiability . . . . . . . . . . . . 245, 270 shear modulus . . . . . . . . . . . . . . . . . . . . . . . . . . 26
particle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 signal
PBH eigenvector test . . . . . . . . . . . . . . . . . . . 105 distortion . . . . . . . . . . . . . . . . . . . . . . . . . 208
PBH rank test . . . . . . . . . . . . . . . . . . . . . . . . . 105 signal conditioning device . . . . . . . . . . . . . . 219
physically inconsistent . . . . . . . . . . . . . . . . . . 197 signal filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
plane strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 signal source . . . . . . . . . . . . . . . . . . . . . . . . . . 220
plane stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 similarity transformation . . . . . . . . . . . . . . . . . 68
Poisson ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
polar radius of inertia . . . . . . . . . . . . . . . . . . . . 34 spectral
pole-residue form . . . . . . . . . . . . . . . . . . . . . . 133 aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
pole-residue representation . . . . . . . . . . . . . 130 leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
pole-zero representation . . . . . . . . . . . . . . . . 130 stability
power amplifier . . . . . . . . . . . . . . . . . . . . . . . . 220 system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 state controllability . . . . . . . . . . . . . . . . . . . . . 102
principal vectors . . . . . . . . . . . . . . . . . . . . . . . . 51 state space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
probability density function . . . . . . . . . . . . . 261 state-transition matrix . . . . . . . . . . . . . . . . . . . 99
marginal . . . . . . . . . . . . . . . . . . . . . . . . . . 262 static condensation . . . . . . . . . . . . . . . . . 31, 152
process noise . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 steepest-descent method . . . . . . . . . . . . . . . . 280
stiffness matrix
R shifted . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
stiffness method . . . . . . . . . . . . . . . . . . . . . . . . 26
Rayleigh quotient . . . . . . . . . . . . . . . . . . . . 59, 77 stimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
reachability stinger . . . . . . . . . . . . . . . . . . . . . . . . . . . 196, 220
matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 stochastic state-space description . . . . . . . . . 45
realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 strain gauge . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
receptance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 structurally distinguishable . . . . . . . . . . . . . 272
reduction superposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Craig-Bampton . . . . . . . . . . . . . . . . . . . . 155 superposition integral . . . . . . . . . . . . . . . . . . . . 98
exact static . . . . . . . . . . . . . . . . . . . . . . . . 152 system identification . . . . . . . . . . . . . . . . . . . 244
Guyan-Irons . . . . . . . . . . . . . . . . . . . . . . . 152 system matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 45
state-space . . . . . . . . . . . . . . . . . . . . . . . . 158 system poles . . . . . . . . . . . . . . . . . . . . . . . 48, 129
transfer strength based . . . . . . . . . . . . . 158
regularization parameter . . . . . . . . . . . . . . . . 280 T
representation
pole-residue . . . . . . . . . . . . . . . . . . . . . . . 130 test frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
pole-zero . . . . . . . . . . . . . . . . . . . . . . . . . 130 torsion
response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 sectional factor . . . . . . . . . . . . . . . . . . . . . 33
risk trace norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
INDEX 301

transducers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
transfer function . . . . . . . . . . . . . . . . . . . . . . . 133
transmission gain . . . . . . . . . . . . . . . . . . . . . . 130
transmission zeros . . . . . . . . . . . . . . . . . . . . . 130
true model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
true model structure . . . . . . . . . . . . . . . . . . . . 241
trust-region method . . . . . . . . . . . . . . . . . . . . 280

wave number . . . . . . . . . . . . . . . . . . . . . . . . . . 136

window
cosine-taper . . . . . . . . . . . . . . . . . . . . . . . 213
exponential . . . . . . . . . . . . . . . . . . . . . . . 213
Hanning . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Tukey . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

Young’s modulus . . . . . . . . . . . . . . . . . . . . . . . . 24

zero-order-hold assumption . . . . . . . . . . . . . 114