Substructuring in Eng Mechanics
Substructuring in Eng Mechanics
Substructuring in Eng Mechanics
Substructuring
in Engineering
Dynamics
Emerging Numerical and Experimental
Techniques
International Centre
for Mechanical Sciences
CISM International Centre for Mechanical
Sciences
Volume 594
Managing Editor
Paolo Serafini, CISM—International Centre for Mechanical Sciences, Udine, Italy
Series Editors
Elisabeth Guazzelli, IUSTI UMR 7343, Aix-Marseille Université, Marseille, France
Franz G. Rammerstorfer, Institut für Leichtbau und Struktur-Biomechanik, TU
Wien, Vienna, Wien, Austria
Wolfgang A. Wall, Institute for Computational Mechanics, Technical University
Munich, Munich, Bayern, Germany
Bernhard Schrefler, CISM—International Centre for Mechanical Sciences, Udine,
Italy
For more than 40 years the book series edited by CISM, “International Centre for
Mechanical Sciences: Courses and Lectures”, has presented groundbreaking
developments in mechanics and computational engineering methods. It covers
such fields as solid and fluid mechanics, mechanics of materials, micro- and
nanomechanics, biomechanics, and mechatronics. The papers are written by
international authorities in the field. The books are at graduate level but may include
some introductory material.
Substructuring
in Engineering Dynamics
Emerging Numerical and Experimental
Techniques
123
Matthew S. Allen Daniel Rixen
Department of Engineering Physics Lehrstuhl für Angewandte Mechanik
University of Wisconsin–Madison Technische Universität München
Madison, WI, USA Munich, Bayern, Germany
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
v
vi Preface
vii
viii Contents
Abstract “Divide and Conquer” is a paradigm that helped Julius Cesar to dominate
on the wide Roman Empire. The power of dividing systems, then analyze them as
parts before combining them in an assembly, is also an approach often followed in
science and engineering. In this introductory chapter, we shortly discuss the main idea
behind domain decomposition and substructuring applied to mechanical systems. —
Chapter Author: Daniel Rixen
pioneered by mathematicians like Courant (1943), Courant and Hilbert (1953), can in
fact be seen as a clever application of the Rayleigh–Ritz technique on small region of
the computational domain, resulting therefore in not only a function decomposition
of the problem within the element, but also a spatial decomposition of the problem
as the physical domain is broken into elements.
Today when we consider substructuring, most often we envision dividing a finite
element model into subcomponents that each contain many elements. This substruc-
turing approach was developed in the 60s, motivated by the need to solve structural
problems in aerospace and aeronautics of several tens of nodes (considered to be very
large problems at that time, given the fact that computers were just arising). Hence,
the idea was to add another level of decomposition and reduction by subdividing the
mesh of a finite element model into substructures and to represent the dynamics per
substructure in a reduced and approximate way. Some of the substructuring tech-
niques still commonly used today originate from that time where memory and CPU
power were very limited.
Later, in the 80s and especially in the 90s, solving very large problems using
domain decomposition became a very active research topic. In order to efficiently
use the computational power of new multiprocessor machines, it was essential to min-
imize the communications between processing units and to ensure that each CPU
was fed with enough work that could be done independently. Domain Decomposi-
tion is the paradigm to achieve this, by letting each CPU solve the local problems
of each domain and searching for the global solution by iterating on the interface
solutions (Toselli and Widlund 2006). Nowadays, high-performance computing tech-
niques solve engineering problems of close to a billion degrees of freedom on several
hundreds of thousands of processors thanks to that paradigm (Toivanen et al. 2018).
The last wave of intensive research in substructuring came at the beginning of
the twenty-first century, this time not directly powered by exponential grows of
computing power, but by significant progress in techniques to accurately measure
the dynamics of components. Experimental substructuring aims to build models of
assemblies from the data measured on components, possibly in combination with
numerical models. Although the theory of experimental dynamic substructuring is
essentially the same as its analytical counterpart, many novel techniques needed to
be developed in order to alleviate the adverse effects of measurement errors (even if
strongly reduced thanks to novel sensing and acquisition techniques) when building
a model based on experimentally characterized components.
In mechanical engineering practice, substructuring techniques are useful for the fol-
lowing cases:
• When working on a large project (for example, the design of an aircraft), different
groups and departments work on different components. In that case, it is essential
1.2 Advantages of Substructuring in Mechanical Engineering 3
that each team can concentrate on the modeling of their part, and then have an
efficient means for assembling to other components in order to predict the overall
dynamics and optimize their design accordingly.
• In industry, it is common that models of subsystems must be shared. A classical
example is the coupled load analysis of the assembly of a satellite and a launcher:
in that case, the satellite company needs to share its model with the launcher
operator in order to perform a global dynamic analysis and guarantee that vibration
limits will not be crossed. In that case, exchanging information can be efficient if
the different components have been modeled separately in such a way that their
complexity (number of degrees of freedom) is low and such that they can be
assembled on their interface.
• In many engineering problems, parts of the system to be analyzed are significantly
more complex than the rest. In structural dynamics, for instance, parts of the model
could be highly nonlinear whereas the remainder can be considered as linear. In that
case, it is very advantageous to substructure the problem in such a way that only
the parts that are nonlinear are treated with the appropriate techniques, whereas
the linear parts are, for instance, condensed once and for all. Such a subdivision of
the problem into substructures of different complexities can greatly improve the
analysis speed and give better insight into the dynamic behavior.
• When optimizing the design during the development of a new product, or during an
improvement of an old design, one often wants to modify only a small number of
parts (for instance, change the bushings connecting the power train of a car to the
chassis in order to modify the noise and vibration harshness). When the parts that
will not be modified are available as substructures and already precomputed (e.g.,
their dynamics on the interface is known), then recomputing the global behavior
when only the components under optimization are changed requires a significantly
reduced cost, enabling much faster design cycles.
• In experimental testing, it is not always possible to test the full system; often
system-level analysis is needed before all of the components have been manufac-
tured. In other cases, it is impossible to test the full assembly because appropriately
exciting a large structure would require forces that are beyond the capabilities of
existing shakers. It is then very advantageous to experimentally characterize the
structure part by part. This allows troubleshooting problems arising from the local
dynamics in those parts, but also to build a full model by assembling the measured
parts using experimental substructuring techniques.
• When parts of the system have not yet been built, it is possible to combine the
measured dynamics of hardware components with other parts that are modeled
only numerically (hybrid substructuring). In some cases, the combination in real
time of the transient response of a hardware component (the physical substructure)
with the dynamics of a numerical substructure is essential to predict the dynamics
of the hardware part in realistic conditions. Such hardware-in-the-loop tests (also
called Real-Time Substructuring, Hybrid Testing or Cyber-Physics) allow testing
very complex systems by having only a part in the lab, the rest of the system being
co-simulated in real time.
4 1 Introduction and Motivation
• Substructuring methods open new opportunities for efficient analyses and design,
including the development of digital twins, or a model that can be used to monitor
the health of a system.
These are only a few examples of cases in which substructuring can be advan-
tageous. For these and other reasons, substructuring has attracted a lot of attention
over the years and is yet today a very active research field.
References
Courant, R. (1943). Variational methods for the solution of problems of equilibrium and vibrations.
Bulletin of the American Mathematical Society, 49(1), 23.
Courant, R., & Hilbert, D. (1953). Methods of Mathematical Physics (Vol. 1). New York: Interscience
Publishers.
Géradin, M., & Rixen, D. (2015). Mechanical vibrations. Theory and Application to Structural
Dynamics (3rd ed.). Chichester: Wiley.
Rayleigh, J. (1896). The Theory of Sound (2nd ed.). Mineola: Dover.
Ritz, W. (1909). Uber eine neue methode zur lösung gewisser variations probleme der mathema-
tishen physik. Journal für die Reine und Angewandte Mathematik, 135, 1–61.
Schwarz, H. A. (1890). Gesammelte Mathematische Abhandlungen (Vol. 2, pp. 133–143). Berlin:
Springer; First published in Vierteljahrsschrift der Naturforschenden Gesellschaft in Zürich
(1870) (Vol. 15, pp. 272–286).
Toivanen, J., Avery, P., & Farhat, C. (2018). A multilevel feti-dp method and its performance for
problems with billions of degrees of freedom. International Journal for Numerical Methods in
Engineering, 116(10–11), 661–682.
Toselli, A., & Widlund, O. (2006). Domain decomposition methods-algorithms and theory (Vol.
34). Berlin: Springer Science & Business Media.
Chapter 2
Preliminaries: Primal and Dual
Assembly of Dynamic Models
In these lecture notes, we assume that the problem has already1 been discretized
(using, for instance, an appropriate Finite Element or Boundary Element formula-
tion). The discretized problem describing the dynamics in a substructure Ω (s) can
be written in the form
(s) (s)
M(s) ü(s) + f (s)
nl u̇ , u = f (s) + g(s) . (2.1)
The superscript (s) indicates that the equation is written for a given substructure s.
M denotes the discretized mass matrix,2 f nl the discretized nonlinear internal force
Fig. 2.1 Schematic overview of substructuring domains (van der Seijs 2016)
2.1 The Dynamics of a Substructure: Domains of Representation 7
The description outlined in (2.2) expresses the dynamics of the substructure using
the displacement at a specific nodal location and is referred to as a physical repre-
sentation. Note that representation in the physical domain can be either discretized
(as assumed in this text) or continuous. In the later case, the unknown is the physical
response field and the associated equations are partial differential equations describ-
ing the (nonlinear) continuous dynamics in a body. This will not be considered here.
where ωi(s) are the eigenfrequencies of the free–free substructure and φ i(s) the asso-
ciated eigenmodes that have the fundamental property of being mass and stiffness
orthogonal, namely,
T 2
φ i(s) K(s) φ (s) (s)
j = ωi δi j , (2.4)
T
φ i(s) M(s) φ (s)
j = δi j , (2.5)
where we have assumed that the modes amplitudes have been chosen to be mass-
normalized and where δi j is the Kronecker symbol such that
δi j = 1 if i = j.
δi j = 0 if i = j.
where n (s) is the number of degrees of freedom in substructure (s). Here, ηi(s) are
the amplitudes of the modal component of the response and are often called modal
coordinate of substructure (s). Often, we will use a matrix notation as in the second
equality of (2.6), where Φ (s) is a matrix containing in its columns the vibration
8 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
modes and η (s) is a uni-column matrix containing all modal coordinates. Usually,
only a subset of modes is considered in order to have an approximated but reduced
representation of the substructure. This will be discussed in later chapters.
In general, the response u(s) can be represented as a combination of n (s) indepen-
dent vectors and we write
(s)
n
u (s)
= vi(s) qi(s) = V(s) q(s) , (2.7)
i=1
where V(s) is a square matrix containing the basis vectors for the desired change of
T
variables. Substituting in the dynamic equation (2.2) and premultiplying by V(s) to
project the equations onto the same space, we obtain
T T T T T
V(s) M(s) V(s) q̈(s) + V(s) C(s) V(s) q̇(s) + V(s) K(s) V(s) q(s) = V(s) f (s) + V(s) g(s) ,
(2.8)
which is usually written as
(s) (s) (s) (s)
M̃ q̈(s) + C̃ q̇(s) + K̃ q(s) = f̃ + g̃(s) , (2.9)
where the tilde superscript indicates that the matrices and vectors pertain now to a
representation in a transformed space. The representation vectors stored in V(s) can
be any set of independent vectors, in particular, they can be chosen as the vibration
(s) (s)
modes Φ (s) , in which case the transformed mass and stiffness matrices M̃ and K̃
will be diagonal.
The representation (2.9) will often be referred to as the modal representation, even
when the base vectors are not vibration modes Φ (s) but general representation modes
V(s) . The associated degrees of freedom q̇(s) are then called generalized degrees
of freedom or modal coordinates and do in general not represent the solution at a
particular physical location.
In case an incomplete basis is used for the representation, namely, when fewer
modes than the number n (s) of degrees of freedom in the substructure are used, the
modal representation represents the dynamics in a reduced subspace and in general
only in an approximate way (this will be discussed in detail in Chap. 3). We will then
call the representation a reduced modal representation.
In the form of Eq. (2.2), the unknown dynamic response is considered a function of
time and we say that, from a spectral point of view, the equations are expressed in
the time domain.
2.1 The Dynamics of a Substructure: Domains of Representation 9
Similar to the decomposition of the spatial response in component modes, the time
dependency of the response can also be decomposed into a combination of time
contributions. The most classical one is the Fourier decomposition4 that writes the
time function of the response in terms of harmonic functions. Using complex number
notations, the Fourier decomposition can be written as
∞
u(s) (t) = ū(s) (ω) e−iωt dω, (2.10)
−∞
(s)
where f̄ and ḡ(s) are the Fourier components of the forces, for instance,
∞
(s) 1
f̄ = f (s) (t) eiωt dt. (2.12)
2π −∞
The dynamic equation in the frequency domain (2.14) is also often written as
(s)
Z(s) ū(s) = f̄ + ḡ(s) (2.13)
3 For an interesting matrix description of time discretization see Rixen and van der Valk (2013),
van der Valk and Rixen (2014).
4 Note that other base functions in time can be used (such as wavelets), but this will not be discussed
here.
10 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
where
Z(s) (ω) = −ω 2 M(s) + iωC(s) + K(s) (2.14)
Z(s) is a dynamic stiffness matrix and is a function of the frequency ω. In this form, ū(s)
is the complex amplitude of the harmonic displacement response (or equivalently the
Fourier component of the transient response). A similar equation can be written for the
amplitude of the harmonic velocity or accelerations in which case the operator Z(ω)
is commonly called the mechanical impedance or the apparent mass, respectively.
The dynamic relation can also be inverted and written
(s)
ū(s) = Y(s) f̄ + ḡ(s) , (2.15)
where −1
Y(s) (ω) = Z(s) (ω)−1 = −ω 2 M(s) + iωC(s) + K(s) (2.16)
Y(s) is a Frequency Response Function (FRF) matrix and is often called the admit-
tance or dynamic flexibility, or more specifically receptance, mobility, or acceler-
ance/inertance if ū(s) are displacements, velocities or accelerations respectively.
Obviously, in practice, the harmonic components are calculated only for a finite
discrete number of frequencies ω, and (2.10) is approximated by the Discrete Fourier
Decomposition
Nω
u(s) (t) = ū(s)
k eiωk (2.17)
k=−Nω
choosing a frequency range covering the spectral range of the excitation. It is note-
worthy that the decomposition in (2.17) is comparable to the decomposition of the
space function of the response in (2.6) and can also be seen as a reduction of the
transient response in the time domain. It is an approximation unless the excitation
can be exactly represented by a finite combination of harmonics.
Laplace Domain
Another often used representation of the time evolution of the dynamic response is
in terms of the Laplace components. The idea is to look for the dynamic response
when modulated with a decreasing exponential function, namely,
∞
ū(s) (s) = L u(s) (t) = e−st u(s) (t) dt (2.18)
0
Clearly, there is a similarity between Laplace and Fourier transforms since (2.18)
becomes a Fourier transform if s is taken as imaginary. The main difference is that the
inverse transform is trivial for the Fourier domain (leading to the frequency domain
decomposition (2.10) or (2.17)), whereas finding the inverse Laplace transform is far
more difficult and not general. In structural dynamics, Laplace transforms are used
for highly transient problems that cannot efficiently be represented by harmonic
superposition, such as impact responses and shock propagations (see for instance
Sect. 4.3.2 in Géradin and Rixen (2015)).
Displacement Space
In addition to representing the space and spectral behavior of the system in dif-
ferent domains as explained above, the very state of the system can be described in
mainly two different manners: either one sees the displacements as the only indepen-
dent unknowns (velocities and accelerations being dependent on the displacement
through derivatives) or the velocities are seen as additional independent variables
for which the derivative relation to the displacement is explicitly expressed in the
formulation. For the first approach, often used in structural dynamics, the dynamics
of the system are described by a single set of second-order equations as in (2.1) (for
nonlinear structures) or (2.2) (for linear ones).
State-Space Representation
In the second case, namely, the velocities are seen as additional independent variables,
the state of the system is described both by the displacements and the velocities:
u(s) (t)
x(s) (t) = (2.19)
u̇(s) (t)
and the associated linear dynamic equation can for instance be written as
(s)
I 0 (s) 0 I(s) (s) 0
ẋ = x (t) + (2.20)
0 M(s) −K(s) −C(s) f (s) + g(s) .
In this state-space representation the number of equations has doubled, but the order
of the differential is now reduced to one. This representation is often used especially
in control. This form is also commonly used in structural dynamics when strong
damping is present since the concept of modes of vibration properly generalizes only
when writing the system in the State Space (see, for instance, Sect. 3.3 in Géradin
and Rixen (2015)).
12 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
From the short summary of the formulation of the structural dynamics problem, it
is clear that many variants to describe the problem, combining a spatial, spectral,
and state representation, can be constructed. In Fig. 2.1, the different aspects are
shown graphically and relations between them are summarized. Depending on the
representation chosen, numerical and experimental techniques in substructuring can
significantly differ as will be seen in these lecture notes.
Let us consider again the linearized dynamic equilibrium equation (2.11) of a sub-
structure in the physical space and in the frequency domain5
(s)
Z(s) ū(s) = f̄ + ḡ(s) s = 1 . . . N sub , (2.21)
5 Expressing the coupling of substructures in other domains (modal, time, state space …) will be
discussed in later chapters and use exactly the same approach.
6 These lecture notes deal with structural problems. Nevertheless, the general theory is also applicable
The interface equilibrium requires that the interface forces, ḡ(s) , which are internal
forces between the substructures, sum to zero when assembled. This is merely a
manifestation of Newton’s “action–reaction" principle. Considering, for instance, an
interface Γ (sr ) between two substructures s and r , one could express this condition
as7
g(s) (r )
b + gb = 0 on Γ (sr ) , (2.24)
gi(s) =0 gi(r ) = 0, (2.25)
where the subscript b indicates a restriction of the DOF to the boundary and where
we assumed that the DOF are numbered in the same manner on both sides of the
interface. The subscript i denotes DOF that are not on a boundary and are thus internal
DOF. On internal DOF, no connecting forces should exist.
In practice, the numbering of the DOF on the interface will not match across the
interfaces and in addition more than two substructures can intersect on an interface
(so-called cross-points in 2D and 3D, and edges in 3D). Hence in general, the interface
equilibrium condition needs to be expressed using Boolean localization matrices
T
L(s) of dimension n × n s that combine the forces on either side of the interface to
satisfy force equilibrium. Interestingly, these localization matrices also map the DOF
of substructure s to a global and unique set of n global DOF, as will be elaborated
later. In general, the interface equilibrium thus is written as
sub
N
T
L(s) ḡ(s) = 0 . (2.26)
s=1
This equilibrium
s condition can also be written using the block matrix LT of dimension
n b × s n acting on the set of all substructure interface forces
sub T
LT ḡ = 0 where LT = L(1) · · · L(N )
T
(2.27)
7 Ingeneral, we will assume that the decomposition in substructures generates an interface dis-
cretization that is conforming and matching, namely, that the shape functions used on either side of
the interface are identical and that the nodes coincide. In case of nonconforming or nonmatching
interfaces, the theory used in these lecture notes are still generally applicable, but the assembly
operators are then no longer Boolean. Indeed, going back to the variational principle underlying
the discretized problem, the part related to the compatibility condition over an interface Γ can be
written as
μT vdΓ
Γ
where μ and v are the field. More details about nonmatching interfaces can be found, for instance,
in Rixen (1997).
14 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
To illustrate the notation, consider the examples in Fig. 2.2. The green num-
bers indicate the global nodes for which the interface force conditions are written
(i.e., the row number in the localization matrices).
For the beam example in Fig. 2.2a, the localization matrices are
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 100 0
L(1) = ⎣0⎦ L(2) = ⎣0 1 0⎦ L(2) = ⎣0 ⎦
T T T
0 001 1
For the second example in Fig. 2.2b, we have two degrees of freedom per
node and the localization matrix can be written as
LT = L(1) L(2) L(3)
2.2 Interface Conditions for Coupled Substructures 15
⎡⎡ ⎤ ⎡ ⎤⎡ ⎤⎤
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
⎢⎢0
⎢⎢ 1 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 0⎥⎥⎢
⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 1 0 0 0 0 0⎥⎥
⎢1
⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥ ⎢0 0 0 0 0 0 0 0⎥⎥
⎥
⎥
⎢⎢0
⎢⎢ 0 0 1 0 0 0 0⎥⎥
⎢0
⎢ 1 0 0 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 1 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥ ⎢0 0 0 0 0 0 0 0⎥⎥
⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 1 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 1 0⎥⎥
⎢0
⎢ 0 0 0 1 0 0 0⎥ ⎢
⎥ ⎢1 0 0 0 0 0 0 0⎥⎥
⎥
⎥
⎢⎢0 0 0 0 0 0 0 1⎥⎥
⎢0
⎢ 0 0 0 0 1 0 0⎥ ⎢
⎥
⎢0 1 0 0 0 0 0 0⎥ ⎥
LT = ⎢ ⎢ ⎥⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 1 0 0 0 0 0⎥ ⎢
⎥ ⎢0 0 0 0 0 0 0 0⎥⎥
⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 1 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 1 0⎥ ⎢
⎥ ⎢0 0 1 0 0 0 0 0⎥⎥
⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 1⎥ ⎢
⎥
⎢0 0 0 1 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥ ⎢0 0 0 0 1 0 0 0⎥⎥
⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥⎥
⎢0
⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 1 0 0⎥⎥⎥
⎥
⎣⎣0 0 0 0 0 0 0 0⎦ ⎣0 0 0 0 0 0 0 0 ⎦ ⎣0 0 0 0 0 0 1 0 ⎦
⎦
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
These matrices are in fact identical to the localization matrices used in Finite
Element codes to assemble elementary matrices in the global system, however here
the localization is not written for one element but for one substructure (that one
could consider as a super-element or macro-element). Obviously, it is not efficient
to store these Boolean matrices as written above, but rather one should store them
as sparse matrices, or even better one should construct the mapping tables based on
the connectivity of the substructures over the interfaces.
The second condition that needs to be satisfied on the interface is that DOF pertaining
to some structural node have the same response on both sides of the interface, or in
other words that the DOF are compatible on the interface. Considering the DOF of
two substructures s and r coupled on the interface Γ (sr ) , the compatibility condition
becomes
ū(s) (r )
b − ūb = 0 on Γ (sr )
where, as before, the subscript b indicates that the compatibility is written for the
boundary DOF and where we assumed that the DOF are numbered identically on
both sides of the interface.
In general, the numbering on the interface does not coincide and therefore the
compatibility conditions are expressed using signed Boolean matrices B(s) . When
operating on ū(s) , these operators extract the interfaces DOF and give them an oppo-
site sign on each side of the interface. The interface compatibility can then be written
16 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
sub
N
B(s) ū(s) = 0 (2.29)
s=1
One can use a block matrix notation to write this condition also in the form
sub
Bū = 0 where B = B(1) · · · B(N ) . (2.30)
The side on which the entry in B is positive and negative can be chosen freely.
The interpretation of B and its associated Lagrange multipliers is depicted in
Fig. 2.3a.
For the second example in Fig. 2.2, the constraint matrix can be written as
B = B(1) B(2) B(3) .
⎡⎡ ⎤⎡ ⎤⎡ ⎤⎤
0 0 1 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
⎢⎢0
⎢⎢ 0 0 1 0 0 0 0⎥ ⎢
⎥⎢0 −1 0 0 0 0 0 0⎥ ⎢
⎥⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 1 0⎥ ⎢
⎥⎢0 0 0 0 −1 0 0 0⎥ ⎢
⎥⎢0 0 0 0 0 0 0 ⎥
0 ⎥⎥⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 1⎥ ⎢
⎥
⎢0 0 0 0 0 −1 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0 0 0 0 0 0 0 0⎥ ⎢
⎥⎢0 0 0 0 1 0 0 0⎥ ⎢
⎥ ⎢−1 0 0 0 0 0 0 ⎥
0 ⎥⎥
=⎢ ⎢ ⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 1 0 0⎥ ⎢
⎥
⎢0 −1 0 0 0 0 0 0⎥⎥⎥
⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥⎢0 0 0 0 0 0 1 0⎥ ⎢
⎥⎢0 0 −1 0 0 0 0 ⎥
0 ⎥⎥⎥
⎢⎢0
⎢⎢ 0 0 0 0 0 0 0⎥ ⎢
⎥
⎢0 0 0 0 0 0 0 1⎥ ⎢
⎥
⎢0 0 0 −1 0 0 0 0⎥⎥⎥
⎥
⎣⎣0 0 0 0 0 0 1 0⎦ ⎣ 0 0 0 0 0 0 0 0⎦ ⎣−1 0 0 0 0 0 0 0⎦⎦
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0
As for the localization matrix L(s) , the constraint matrices B(s) are, in practice, not
stored as full but as sparse, or only connectivity information is used to apply them
to a vector.
To summarize this section, an assembly of substructures in the physical frequency
domain (and similarly in other domains) is obtained by imposing compatibility and
interface equilibrium conditions, leading to the set of equations
⎧ (s)
⎪
⎪ Z(s) ū(s) = f̄ + ḡ(s) s = 1 . . . N sub
⎪
⎪
⎪
⎪ (s) (s)
sub
⎪ N
⎨ B ū = 0
(2.32)
⎪ s=1
⎪ N sub
⎪
⎪
⎪
⎪ T
⎪
⎩ L(s) ḡ(s) = 0
s=1
⎧
⎨ Zū = f̄ + ḡ
Bū = 0 (2.33)
⎩ T
L ḡ = 0
The form (2.32) (or equivalently (2.33)) of the coupled problem uses two interface
fields, namely, the primal unknowns u per substructure (i.e., on each side of the
interfaces) and the substructure interface forces g called dual unknowns.8 Solving
the dynamic problem of the assembly in the form (2.32) can be expensive since many
interface unknowns need to be resolved. Hence, these equations can be rearranged in
order to eliminate the interface forces and write the problem in terms of unique inter-
face displacements (primal assembly) or by introducing interface forces satisfying
the interface equilibrium (dual assembly).
We can define as primal unknowns for the interface a set of DOF ug that are global
and uniquely defined for the entire structure. The DOF of each substructure are then
obtained by mapping the global set ug to the local DOF of each substructure u(s) .
Such a mapping was already introduced in the previous section, (2.26), to map the
local interface forces to a global set. The same mapping, but now from the global
DOF to the local ones can be used to write
If the substructure DOF are obtained from a unique set as described above, they
automatically satisfy the compatibility conditions that matching interface DOF must
be equal. Hence
BLug = 0 ∀ ug . (2.35)
This relation mathematically means that L represents the nullspace of the constraint
matrix B:
L = null(B) (2.36)
8 Inthese lecture notes, we will introduce the coupling conditions using basically a two field
approach. A more general three-field approach can also be considered but will not be discussed
here. See for instance Voormeeren (2012).
2.3 Primal and Dual Assembly 19
In the example of the beam in Fig. 2.2a for which L was found in (2.28), this
leads to
⎡ ⎤
⎡ (1) ⎤ 1 00
ū 2 ⎢ ⎥
⎢⎡ (2) ⎤⎥ ⎢⎡ ⎤⎥ ⎡ ⎤
⎢ ū 1 ⎥ ⎢ 1 0 0 ⎥
⎢ (2) ⎥ ⎢ ⎥ ū 1
⎢⎣ū ⎦⎥ = ⎢⎣0 1 0⎦⎥ ⎣ ⎦
⎢ 2 ⎥ ⎢ ⎥ ū 2 (2.37)
⎢ ū (2) ⎥ ⎢ 0 0 ⎥
1 ⎥ ū 3
⎣ 3 ⎦ ⎢
(3) ⎣ ⎦
ū 1
0 0 1
Clearly, it can be seen that the substructure DOF are “drawn" from a global and
unique set of DOF and thus automatically satisfy the compatibility constraint
on the interface: with the compatibility constraint matrix B obtained in (2.31),
one verifies that
⎡ ⎤
100
⎡ ⎤⎥
⎢ ⎢ 100 ⎥
1 −1 0 0 0 ⎢⎣ ⎥
Bu = BLug = ⎢ 0 1 0⎦⎥ ug = 0
0 0 0 1 −1 ⎢ ⎥
⎣ 001 ⎦
001
When choosing a unique set of DOF as in (2.34), the coupled problem (2.33) can
be simplified and written as
ZLug = f̄ + ḡ
(2.38)
LT ḡ = 0,
This form is generally referred to as the primal assembly of the coupled problem
and matrix LT ZL is the primal-assembled impedance of the global system. This
assembly is similar to the assembly of finite elements.
20 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
To illustrate this process, consider again the first example in Fig. 2.2a, with
the matrix L given in (2.28). The assembled impedance matrix for this example
then writes
Zg = LT ZL
⎡ ⎤
⎡ (1)
⎤ 10 0
Z 22 0 0 ⎢ ⎥
⎡⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎤ ⎢ ⎡ (2) (2) (2) ⎤ ⎥⎢ ⎡ ⎤⎥
1 100 0 ⎢ Z 11 Z 12 Z 13 ⎥⎢
⎢ 10 0 ⎥ ⎥
⎢ ⎥ ⎣
= ⎣⎣0⎦ ⎣0 1 0⎦ ⎣0⎦⎦ ⎢ 0 ⎣ Z (2) (2)
Z 22 (2) ⎦
Z 23 0 ⎥⎢ 0 1 0⎦⎥
⎢ 21 ⎥⎢ ⎥
0 001 1 ⎣ (2) (2) (2)
⎦⎢ 0 0 1 ⎥
Z 31 Z 32 Z 33 (3) ⎢
⎣
⎥
⎦
0 0 Z 11
00 1
Zg ug = = f g =⇒
⎡ ⎤ ⎡ ⎤ ⎡ (1) ⎤
(1)
Z 22 (2)
+ Z 11 (2)
Z 12 (2)
Z 13 ū 1 f¯2 + f¯1(2)
⎣ (2)
Z 21 (2)
Z 22 (2)
Z 23 ⎦ ⎣ū 2 ⎦ = ⎣ f¯2(2) ⎦
(2)
Z 31 (2) (2)
Z 32 Z 11 + Z 33(3) ū 3 f¯3 + f¯1(3)
(2)
Let us start again from the coupled form (2.32) (or equivalently (2.33)), but instead
of satisfying a priori the interface compatibility as in the primal assembly, we satisfy
a priori the interface equilibrium.9 This can be achieved by choosing coupling forces
as
9 For the description of the problem, it does obviously not manner if one first satisfies the equilibrium
or the compatibility. This only determines which equation in the three-field formulation (2.32) is
2.3 Primal and Dual Assembly 21
g = −BT λ (2.40)
LT BT λ = 0 ∀λ (2.41)
The minus in this definition is a choice to obtain a symmetric form later. The
expression (2.40) can also be written substructure wise as
⎡ ⎤ ⎡ (1) ⎤
g(1) B
⎢ .. ⎥ ⎢ .. ⎥
⎣ . ⎦ = −⎣ . ⎦λ (2.42)
g(N ) B(N )
sub sub
showing clearly that the coupling forces are generated from a unique set of interface
forces, whose intensities are given by λ and are attributed, with the appropriate
positive or negative sign, to the substructure interface. The interface force intensities
λ are unknowns pertaining to the interface. The size n λ of λ is equal to the number
of interface compatibility conditions on the interface.
Substituting the coupling forces by their form (2.40), the coupled problem can be
written as ⎧
(s)
⎪ T
⎪ Z(s) ū(s) = f̄ + B(s) λ s = 1 . . . N sub
⎨
sub
N
(2.43)
⎪
⎪ B(s) ū(s) = 0,
⎩
s=1
One observes that λ can be interpreted as the Lagrange multipliers associated with the
compatibility constraint (see the general theory of constrained systems in dynamics,
for instance, in Géradin and Rixen (2015)).
eliminated. The end result will be identical, but the mathematical form and the approximations that
can be applied will be depending on the form that will be solved.
22 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
To illustrate this form, consider again the first example in Fig. 2.2a, with the
matrix B given in (2.31). The dually assembled form is then written as
⎡ ⎤ ⎡ (1) ⎤ ⎡ (1) ⎤
(1)
Z 22 0 0 1 0 ū 2 f¯
⎢ ⎡ ⎤ ⎡ ⎤⎥ ⎢⎡ (2) ⎤⎥ ⎢⎡ 2(2) ⎤⎥
−1 0 ⎥ ⎢ ū 1 ⎥ ⎢ f¯1 ⎥
(2) (2) (2)
⎢ Z 11 Z 12 Z 13
⎢ ⎣ 0 0 ⎦⎥ ⎢ ⎦⎥ ⎢ ⎥
⎢ 0 ⎣ Z 21
(2) (2)
Z 22 (2) ⎦
Z 23 0 ⎥ ⎢⎣ū (2) ⎥ ⎢⎣ f¯(2) ⎦⎥
⎢ (2) (2) (2) ⎥ ⎢ 2(2) ⎥ ⎢ 2(2) ⎥
⎢ Z 31 Z 32 Z 33 0 1 ⎥ ⎢ ū 3 ⎥ ⎢ f 3 ⎥
⎢ ⎥ ⎢ (3) ⎥ = ⎢ (3) ⎥
⎢ 0 0 (3)
Z 11 0 −1 ⎥ ⎢ ⎥ ⎢ ¯ ⎥
⎢ ⎥ ⎢ ū 1 ⎥ ⎢ f 1 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ 1 −1 0 0 0 ⎦ ⎣ λ1 ⎦ ⎣ 0 ⎦
0
0 0 0 1 −1 λ2 0
Notice that this representation of the system is larger than the primally assem-
bled counterpart, but in this representation the interface forces, λ are explicitly
available. In future sections we shall see that this can be useful.
In this section, we have expressed the coupled problem in three different forms that
are recalled in their block matrix form below for clarity:
⎧
⎨ Zū = f̄ + ḡ
General form Bū = 0 (2.33)
⎩ T
L ḡ = 0
Z BT ū f̄
Dual Assembly = (2.45)
B 0 λ 0
To illustrate the primal and dual assembly, Fig. 2.4 shows the assembly of components
with forces or with displacements, and the corresponding matrices.
Each of these representations is mathematically equivalent and, hence, one might
wonder why considering one form or the other might be advantageous. From a
mechanical point of view, the interpretation of these forms is different, and so each
provides different insights. The general form considers each substructure as inde-
pendent (the DOF and coupling forces are defined separately per substructure) and
therefore requires writing all of the interface conditions explicitly. In the primal form,
coupling forces are no longer part of the problem: the only interface unknowns are
2.3 Primal and Dual Assembly 23
the interface DOF, which are unique for the entire structure. In the dual form, it is
the interface force amplitudes λ that are unique for the entire structure and represent
the interface unknowns (the interface DOF remaining defined per substructure).
The different mechanical interpretations of the assembly can be used to develop
different numerical and experimental techniques in substructuring. This will hope-
fully become clear in the other chapters of these lecture notes, but here we provide
a brief preview to the various uses of these forms.
Model Reduction (Chap. 3) In model reduction by substructuring, the dynamics in
a substructure are approximated by representing the behavior of its physical DOF
in a subspace of representative modes.
– When the primal-assembled form (2.39) is considered, a substructure is seen as
receiving information from its neighboring substructures over the interface DOF
they share. Hence, a natural set of modes for the reduction might correspond to
the static response of the substructure to unit displacements at the interface, and
these could be augmented by vibration modes in which the interface is fixed (i.e.,
the Hurty/Craig–Bampton family of methods naturally arise).
– When the dual assembly form (2.45) is considered, the substructure receives excita-
tion from the neighbors as imposed interface forces. A natural set of representation
modes, in that case, would consist of the static response of the substructure to unit
interface forces and the additional dynamics would naturally be represented by
vibration modes computed with the interface free.
The reduction methods will thus be based on different approximated representa-
tions depending on which mechanical interpretation of the coupling is considered.
Experimental Substructuring (Chap. 4) In experimental substructuring, the dynam-
ics of (some) substructures are obtained by measurements. Since in experimental
dynamics one typically measures a component with free interface, applying exci-
tations and measuring the dynamic response, one typically measures dynamic
flexibilities (FRFs), namely the response of the component to excitations. There-
fore, considering the dually assembled form seems more natural. Obviously, the
24 2 Preliminaries: Primal and Dual Assembly of Dynamic Models
References
Brüls, O., Arnold, M. (2008). The generalized-α scheme as a linear multistep integrator: Towards
a general mechatronic simulator. Transactions ASME, 3, 041,007–1/041,007–10.
Chung, J., & Hulbert, G. M. (1993). A time integration algorithm for structural dynamics with
improved numerical dissipation: The generalized-α method. ASME Journal of Applied Mechan-
ics, 60, 371–375.
de Klerk, D., Rixen, D. J., & Voormeeren, S. N. (2008). General framework for dynamic substruc-
turing: History, review and classifcation of techniques. AIAA Journal, 46(8), 1169–1181. https://
doi.org/10.2514/1.33274.
Géradin, M., & Cardona, A. (2001). Flexible multibody dynamics: A finite element approach.
Hoboken: Wiley.
Géradin, M., & Rixen, D. (2015). Mechanical vibrations. Theory and application to structural
dynamics (3rd ed.). Chichester: Wiley.
Lew, A., Marsden, J., Ortiz, M., & West, M. (2004). Variational time integrators. International
Journal for Numerical Methods in Engineering, 60(1), 153–212.
Newmark, N. (1959). Method of computation for structural dynamics. Journal of Engineering
Mechanics, 85, 67–94.
Rixen, D. (1997). Substructuring and dual methods in structural analysis. Ph.D. thesis, Université
de Liège, Belgium, Collection des Publications de la Faculté des Sciences appliquées, n◦ 175
Rixen, D. J., & van der Valk, P. L. (2013). An impulse based substructuring approach for impact
analysis and load case simulations. Journal of Sound and Vibration, 332, 7174–7190.
van der Valk, P. L., & Rixen, D. J. (2014). An impulse based substructuring method for coupling
impulse response functions and finite element models. Computer Methods in Applied Mechanics
and Engineering, 275, 113–137.
van der Seijs, M.V. (2016). Experimental dynamic substructuring: Analysis and design strategies
for vehicle development. Ph.D. thesis, Delft University of Technology. https://doi.org/10.4233/
uuid:28b31294-8d53-49eb-b108-284b63edf670.
Voormeeren, S. (2012). Dynamic substructuring methodologies for integrated dynamic analysis of
wind turbines. Ph.D. thesis, Delft University of Technology
Chapter 3
Model Reduction Concepts and
Substructuring Approaches for Linear
Systems
Abstract In this chapter, we give an overview of some of the most common reduc-
tion techniques based on substructuring. Although all techniques follow a similar
approach, the main difference between the methods lies into the basis vectors used
in the approximation subspace to represent the dynamics of each substructure and the
manner in which the substructures are a couple. — Chapter Authors: Daniel Rixen,
Matt Allen, and Thomas Abrahamsson
Often, a finite element model is first built for static analysis in order to assess static
deformations and stress levels which might be high especially in small details (stress
concentration areas). Therefore, it is very common to find static models that have
a very refined mesh, hence a high number of degrees of freedom (typically up to
several million). Such very large static models can be solved by means of efficient
solvers, but when it comes to computing free vibration modes, harmonic responses
and/or transient responses, it is required to solve many static-like problems (i.e.,
in the inverse iterations of the eigensolvers or in the time-stepping loop) and the
computing time required is often unacceptable.
Fortunately, for most dynamical problems, such highly refined meshes are not
needed to capture the dynamic behavior. For instance, if we want to compute the
fundamental modes, we know that the first free vibration modes have a rather smooth
deformation and thus a coarser mesh would be sufficient to compute these modes.
Similarly, when computing the dynamic response to external loads which are no
shocks (i.e., which have no high spectral content compared to the eigenfrequency
spectrum of the structure), a coarse mesh yields in most cases accurate results. In
summary, using a coarser representation of the displacement field in the dynamic
model is often acceptable in terms of accuracy and is required in order to perform
dynamic analyses in a reasonable time.
The structural analyst should thus in principle build two models, one for the
static simulations, one for the dynamic analysis. Knowing that building a model is a
significant part of the entire study, it would be very useful to device a procedure that
reduces the size of the dynamic problem without modifying the mesh grid.
Such procedures indeed exist and are known as reduction methods. Similarly to
the way one approximates continuous fields by a set of shape functions in a Rayleigh–
Ritz approach (e.g., in Finite Elements), the driving idea in reduction techniques is to
replace the DOF by a set of global variables representing the amplitudes of possible
displacement modes (see also Sect. 2.1.1):
where u are all the DOF of the system and T is a reduction matrix of dimension
n × r , r < n and where q of dimension r × 1 is a set of reduced generalized degrees
of freedom. The columns of T define possible displacement shapes for the degrees
of freedom u, and q are the amplitude of these displacement shapes.
In general, replacing u by (3.1), only an approximate solution can be found and
the accuracy of the approximation will depend on how well the assumed modes in
T can represent the exact solution. Introducing (3.1) in the linear dynamic problem
of motion
Mü + Cu̇ + Ku = f(t), (3.2)
one obtains
MTq̈ + CTq̇ + KTq̇ = f(t) + r, (3.3)
where r is a residual force that is a remainder for the dynamic equilibrium equation:
since the reduction subspace T can in general not represent the exact solution, there
will always be an equilibrium error, whatever the choice of q. Indeed, one now has
only r unknowns q to satisfy n > r equations. Following then the idea of virtual
work, one requires that the reduced DOF q be chosen such that the residual force r
does not contribute to the dynamics in the representation space T or, in other words,
that the residual force does not produce any work for the possible shapes of motion
contained in T:
TT r = 0 (3.4)
With this condition, we can find the equations to determine q by projecting the
dynamic equilibrium equations (3.3) onto the subspace T:
where the tilde superscript indicates that the matrices and vectors pertain now to a
representation in a reduced space.
3.1 Model Reduction—General Concepts (Reduced Basis) 27
After having solved the reduced problem (3.6) for q, one can build the solution
u for the physical DOF by substituting in (3.1). The residual force r for the full
problem can be computed by substitution in the original problem (3.3): it provides a
way to monitor, a posteriori, the error on the equilibrium.
In what follows, we will drop the damping term in order to clarify the presentation.
Reduction of the damping matrix, in particular for substructure reduction, is discussed
in Gruber and Rixen (2018a, b) and the references therein.
or
One may imagine separating the condensed coordinates u1 into two contributions
u1,stat = −K−1
11 K12 u2 . (3.11)
S = −K−1
11 K 12 (3.13)
is the static condensation matrix the columns of which contain the so-called static
modes representing the static response of u1 for unit u2 displacements. They represent
the static deformation induced in the inner part of the substructure when a unit
displacement is given at an interface DOF.
The reduced matrices are then
We observe that the reduced stiffness matrix is the stiffness matrix statically con-
densed on u2 . The reduced mass matrix is the mass matrix associated to u2 and
augmented by the inertia of u1 assumed to respond statically. The dynamic problem
is then reduced to
MG I u¨2 + K G I u2 = f 2G I . (3.16)
If the static condensation algorithm is applied to static problems, the exact solution
is found. But when applied to dynamic problems, an approximation is introduced by
neglecting the dynamic response of the interior of the substructure and thereby assum-
ing that all internal nodes respond quasi-statically to the interface displacements. The
validity of the condensation algorithm thus depends on the extent to which correction
u1,dyn is negligible. It is possible to show (see, for instance, Géradin and Rixen 2015)
that the static condensation technique is valid if
ω2 μ21 , (3.17)
where ω is the highest eigenfrequency that one wants to compute for the complete
structure and μ1 is the lowest eigenfrequency of the structure when u2 are clamped.
The complete analysis allows to show that static condensation always leads to over-
estimating the eigenvalues compared to the full model. This is natural if we recall
that a model obtained by applying consistently a Rayleigh–Ritz approach is always
stiffer when further restrictions are introduced in the discretization field (see, for
instance, Géradin and Rixen 2015).
The algorithm described above is very frequently used in the context of finite ele-
ment structural analysis. Although originally proposed independently by Guyan and
Irons, it is commonly known as the Guyan reduction method. In commercial codes,
this method is often implemented and used to reduce the complexity of the problem
when no substructuring approach is used. The choice of an appropriate set of master
DOF u2 is sometimes done automatically by the software. One simple heuristics
consists of ranking the DOF u according to their individual pseudo-frequency com-
3.1 Model Reduction—General Concepts (Reduced Basis) 29
puted by the ratio of the diagonals of the stiffness and mass matrices, namely for a
DOF k νk2 = K kk /Mkk , and choosing as master DOF u2 the ones with the lowest νi2 .
(s) (s) u(s)
u =T b (3.18)
ζ (s)
These reduced matrices of the substructures can easily be shared between different
design teams. Such methods are sometimes called Component Mode Synthesis or
CMS.
3. The interface boundary degrees of freedom u(s)b are assembled on the interface of
the substructures exactly as if one would assemble the macro-elements.
quencies of the substructures fixed on the interface are not small (see criteria (3.17)).
Hence, additional information about the substructure vibrations should be added to
the reduction basis as explained in the next Section.
The substructure reduction method discussed in this section is one of the most com-
monly used substructuring technique in engineering practice. It was proposed by Roy
Craig (1968), writing in a more intuitive form ideas previously published by Hurty
(1965). Hence, although most commonly known as the Craig–Bampton method, we
will call it the Hurty/Craig–Bampton method.
The FE discretized system of equations of motion for a substructure reads
where M(s) and K(s) are the mass and stiffness matrices, respectively, u(s) is the
displacement vector, f (s) is the forcing vector and the superscript s denotes the sth
substructure. The substructure equations of motion are partitioned into interior and
boundary (or interface) DOF (denoted by subscripts i and b’, respectively), as
(s) (s) (s) (s) (s) (s)
Mii Mib üi Kii Kib ui 0
+ = (3.21)
(s)
Mbi Mbb(s)
üb(s) (s)
Kbi (s)
Kbb ub(s) fb(s) ,
where the subscripts b and i are indexes referring to the boundary and internal
component of the matrices and vectors.
The representation modes, commonly called Component Modes (CMs) in sub-
structuring, are formed by computing the static response of the interior of the sub-
structure when one interface DOF is given a unit displacement and all the other DOF
are held fixed. The CM set for the entire interface is expressed as
(s)−1 (s)
(s)
Ψ = −K ii Kib . (3.22)
I
The resulting basis Ψ (s) is used to statically eliminate all interface DOF from the
model, retaining only the boundary DOF. The resulting reduced system is usually of
small size since only boundary DOF are remaining unknowns. In order to capture
the dynamics of the system, the static modes are augmented with a set of dynamic
modes which are obtained by fixing the interface DOF and solving the following
eigenvalue problem:
3.2 Numerical Techniques for Model Reduction of Substructures 31
Kii(s) − ωr2 Mii(s) φ i,r
(s)
= 0. (3.23)
The eigenvectors obtained from this equation are referred to as fixed-interface modes.
A truncated set of m of these mass normalized eigenvectors are collected into a fixed-
interface mode matrix
(s) (s)
(s)
φ , . . . , φ i,m Φi
Φ (s) = i,1 = . (3.24)
0 0
These modes provide a normal basis for the interior DOF of the substructure. The
fixed-interface modes and the CMs are combined to form the HCB reduction matrix as
T(s)HCB = Φ (s) Ψ (s) , (3.25)
which provides a transformation from the substructure physical DOF to the HCB
generalized DOF, (s) (s)
ui HCB(s) qi
≈T , (3.26)
ub(s) ub(s)
where qi(s) represents the modal coordinate vector associated with the fixed-interface
modes. The meaning of the static modes and of the fixed-interface modes as used in
the HCB method is illustrated in Fig. 3.1.
The uncoupled substructure-reduced mass and stiffness matrices are now formed
by applying the HCB transformation to these matrices as
T
T
M(s)HCB = T(s)HCB M(s) T(s)HCB , K(s)HCB = T(s)HCB K(s) T(s)HCB . (3.27)
Fig. 3.1 Static modes and internal vibration modes of a substructure. In this example, the interface
is the far left edge of the plate; the constraint modes show motion of a single node on the interface
while the fixed-interface modes show the motion of the rest of the structure when the interface is fixed
32 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
with Ω (s)
2
m the diagonal matrix of the eigenfrequencies of the fixed- interface modes
and with the full submatrices
(s) −1
K̃bb = K(s) (s) (s) (s)
bb − Kbi K bb Kbi (3.29)
(s) −1 −1
M̃bb = M(s) (s) (s)
bb − Mbi Kii
(s)
Kib − K(s) (s)
bi Kii
(s)
Mib
−1 −1
+ K(s)
bi Kii
(s)
Mii(s) Kii(s) Kib(s)
(3.30)
−1
T
(s)
M̃ib = Φ (s) Mib − Mii(s) Kii(s) Kib
(s)
T
= M̃bi (3.31)
The HCB reduced-order models are typically coupled using a primal assembly by
defining a transformation between the uncoupled and coupled DOF that selects the
substructure boundary DOF from a unique global set of boundary DOF (for an
explanation of the Boolean assembly matrix, see Sect. 2.2). The assembled HCB
mass and stiffness matrices are formed by adding the coupled contribution from
each substructure as
n
(s)HCB T (s)HCB (s)HCB
MHCB = L M L (3.32)
j=1
n
(s)HCB T (s)HCB (s)HCB
KHCB = L K L , (3.33)
j=1
where the sth substructure coupling matrix, L(s)HCB , is formed by extracting the
corresponding rows of LHCB .
Example: W-bracket
5
4
Interface Γ2
3 2
1 Interface Γ3
Interface Γ1
Fig. 3.2 Exploded view of the W-Bracket system, inspired by the model in Hong et al. (2013),
consisting of five plate substructures arranged in a “W” configuration with three interface sets;
the interface sets i , i = 1, 2, 3 are highlighted and labeled
For example, consider the W-bracket system from Krattiger et al. (2019)
and shown in Fig. 3.2. This consists of five plates assembled into a shape that
resembles the letter “W”. For this example, the coupling matrix can be written as
3.2 Numerical Techniques for Model Reduction of Substructures 33
⎧ ⎫ ⎡ ⎤
⎪
⎪ qi,1 ⎪ ⎪ I 0 0 00 0 0 0
⎪
⎪ ⎪
⎪ ⎢0 0⎥
⎪
⎪ u ,1 ⎪
⎪ 0 0 00 I 0
⎪u ⎪ ⎢ ⎥⎧ ⎫
1
⎪
⎪ 2 ,1 ⎪
⎪ ⎢0 0⎥
⎪
⎪ ⎪
⎪ ⎢ 0 0 00 0 I ⎥⎪⎪
⎪
qi,1 ⎪
⎪
⎪
⎪
⎪ qi,2 ⎪ ⎪ ⎢0 I 0 00 0 0 0⎥ ⎪
⎪ q ⎪
i,2 ⎪
⎪
⎪ ⎪
⎪ ⎢ ⎥⎪⎪ ⎪
⎪
⎪
⎪ u2 ,2 ⎪⎪ ⎢0 0 0 00 0 I ⎥
0⎥⎪ ⎪ qi,3 ⎪
⎪
⎪
⎨u ⎬ ⎢
⎪ ⎪ ⎪
⎢0 0 0 00 0 0 I ⎥ qi,4 ⎬
⎥ ⎨
=⎢
3 ,2
⎢0 , (3.34)
⎪
⎪ qi,3 ⎪ ⎪ ⎢ 0 I 00 0 0 0⎥⎥⎪⎪ qi,5 ⎪
⎪
⎪ ⎪
⎪ ⎢0 ⎥ ⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ u1 ,3 ⎪⎪ 0 0 00 I 0 0⎥⎪ u ⎪
⎪
⎪ ⎪ ⎢ ⎪
⎪ 1⎪ ⎪
⎪
⎪ q ⎪ ⎢ ⎪ 0 0 0 I0 0 0 0⎥ ⎪
⎪ u ⎪
⎪
⎪ i,4 ⎪
⎪ ⎪ ⎢
⎪ ⎢0
⎥⎪
⎥ ⎩ 2⎪ ⎭
⎪
⎪ u ⎪ 0 0 00 0 I 0 ⎥ u3
⎪ 2 ⎪
⎪
,4
⎪ ⎢
⎪
⎪ qi,5 ⎪ ⎪
⎪
⎣0 0 0 0I 0 0 0 ⎦
⎩ ⎭
u3 ,5 0 0 0 00 0 0 I
LHCB
where the boundary DOF of each substructure are further partitioned according
to the interface sets 1 , 2 , and 3 as shown in Fig. 3.2.
The choice of the number of fixed-interface modes kept in the reduction basis for
each substructure can be made based on different criteria:
• A criterion often used in practice is based on the eigenfrequencies of the kept
modes. Typically, one chooses all fixed-interface modes having an eigenfrequency
lower than 1.8 or 2 times the highest frequency of interest in the assembly.
• One can also choose the fixed-interface modes based on how easily they can be
excited through the support, for instance, by defining a measure of the interface
reaction force associate to a mode. Following a similar reasoning, one evaluates
how complete the base of the fixed- interface modes need to be by specifying how
much of the mass of a substructure mass be represented in its reduced matrices.
This can be evaluated using the concept of effective modal mass (see, for instance,
Géradin and Rixen 2015).
• The number of modes can also be chosen based on a posteriori error estimators
(Jakobsson et al. 2011), later reformulated in a less mathematical form and used
for adaptive selection strategies in Voormeeren et al. (2013).
The discussion in this Section is based on the overview in Gruber and Rixen (2016).
Considering the equation of motion (2.2) of substructure s,
every substructure can be seen as being excited by the interface connection forces and
the external forces (contrary to the paradigm underlying the HCM method, where
the substructures are considered as being excited by interface displacements).
This indicates that the displacements of each substructure u(s) can be expressed
in terms of local static solutions u(s)
stat and in terms of eigenmodes associated to the
entire substructure matrices K(s) and M(s) (hence the free interface modes):
n (s)
−r (s)
u (s)
= u(s)
stat + θ (s) (s)
j ηj , (3.36)
j=1
where n (s) and r (s) are the number of DOF and of rigid-body modes for substructure
s, and where the free interface modes are solutions of the eigenvalue problem
K(s) − ω(s) θ (s)
2
(s)
j M j = 0. (3.37)
The static solution is written as follows (assuming no external forces are applied for
simplicity):
r (s)
(s) + (s)
u(s)
stat = −K g + r(s)
j αj
(s)
(3.38)
j=1
The static solution arises from solving Eq. (3.35) under the assumption that there are
+
no inertia forces and no external forces acting on the substructure. The notation K(s)
(s)
denotes the inverse of K when there are enough boundary conditions to prevent
the substructure with a free interface from floating Rixen (2004). If a substructure is
+
floating, K(s) is the generalized inverse of K(s) and r(s)j are the rigid-body modes
of s (see, for instance, Géradin and Rixen 2015 on how to compute a generalized
inverse and the rigid-body modes).
If only a limited number of free interface modes are used for the substructure
dynamics, Eqs. (3.36), (3.38) results in the approximation
+
u(s) ≈ −K(s) B(s) λ + R(s) α (s) + Θ (s) η(s)
T
(3.39)
The vector α (s) contains the amplitudes of the rigid-body modes and the vector η (s)
contains the amplitudes of the retained n (s)
θ local free interface modes eigenmodes.
The matrices R(s) and Θ (s) contain all rigid-body modes and the retained eigenmodes.
+
Since a part of the subspace spanned by Θ (s) is already included in K(s) , the
+
residual flexibility matrix Gr(s) can be used instead of K(s) , which is defined by
n (s) −r (s)
θ (s) (s) T
j θj
Gr(s) = (3.40)
ω(s)
2
j=n (s)
θ +1
j
3.2 Numerical Techniques for Model Reduction of Substructures 35
(s)
θ (s) (s) T
nθ
j θj
(s) +
=K − .
ω(s)
2
j=1 j
Note that, by construction, Gr(s) = Gr(s) , which is computed using the second equal-
T
ity in Eq. (3.40). For further properties of Gr(s) , see Rixen (2004). As a result, the
approximation of one substructure writes
⎡ (s) ⎤
(s) (s) (s) (s)T α(s)
u(s) ≈ R Θ Gr A ⎣η ⎦ . (3.41)
g(s)
b
T(s)
1
Gr(s) A(s) is the matrix containing the residual flexibility attachment modes of sub-
T
structure s, since the Boolean localization matrix A(s) simply picks the columns of
T
The approximation (3.41) can now be used to reduce the substructure DOF. Using
the orthogonality properties of the modes in Eq. (3.41), the equation of motion of
one substructure becomes
⎡ (s) ⎤ ⎡ (s) ⎤
α̈ α
T
(s) ⎣ (s) ⎦ (s) ⎣ (s) ⎦
M f r ee η̈ + K f r ee η = T(s)
1 f (s) + g(s) (3.42)
(s)
g̈(s)
b g b
where
(s)
= A(s) Gr(s) A(s)
T
Gr,bb (3.44)
and
⎡ ⎤
I0 0
M(s) (s) (s) (s)
= ⎣0 I 0 ⎦
T
f r ee = T1 M T1 (3.45)
(s)
0 0 Mr,bb
36 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
where
(s)
= A(s) Gr(s) M(s) Gr(s) A(s) .
T
Mr,bb (3.46)
(s) (s)
Gr,bb is the residual flexibility and Mr,bb is the interface inertia associated to the resid-
ual flexibility related to the boundary DOF, respectively, and Ω (s) being a diagonal
matrix containing the retained n (s)θ eigenvalues ω j .
(s)
The Rubin method was proposed by Rubin (1975) and, with the definitions above,
can be explained as follows. In order to assemble in a primal manner the substructure
equation of motion (3.42) in the global system, a second transformation is applied by
the RM. The force DOF g(s)b are transformed back to the boundary displacements ub
(s)
R(s) (s)
b and Θ b are the subparts of R
(s)
and Θ (s) related to the boundary DOF, respec-
tively. From this equation, the interface force DOF can be solved as
g(s) (s) (s) (s) (s)
b = Kr,bb ub − R b α − Θ (s)
b η
(s)
(3.48)
−1
(s) (s)
with Kr,bb = Gr,bb . The transformation matrix T(s) (s)
2 from force DOF gb back to the
(s)
boundary displacements ub leaving α (s) and η(s) unchanged writes then
⎡ ⎤
I 0 0
T(s)
2 =
⎣ 0 I 0 ⎦ (3.49)
(s)
−Kr,bb R(s) (s) (s)
b −Kr,bb Θ b
(s)
Kr,bb .
Application of this transformation to the matrices of Eqs. (3.45) and (3.43) gives
the RM-reduced matrices of one substructure:
ed,R = T2 K f r ee T2 = T R K T R , (3.51)
Mr(s) T(s) (s) (s)
T(s) (s) (s)
T T
ed,R = 2 M f r ee T2 = R M TR . (3.52)
3.2 Numerical Techniques for Model Reduction of Substructures 37
These matrices can be directly assembled using primal assembly to get the RM-
reduced matrices Kr ed,R and Mr ed,R of the global system. The RM applies the reduc-
tion matrix T(s)
R consistently to the mass and stiffness matrix resulting in a true
Rayleigh–Ritz method as was observed in Craig and Chang (1977).
The MNM MacNeal (1971) is nearly identical to the RM except for a small change.
(s) (s)
First, we will derive the preliminary MNM-reduced matrices K̃r ed,M N and M̃r ed,M N
following the derivation of the RM to show the similarities between these two meth-
ods. The reduced stiffness matrix of both the RM and the MNM are identical (given
in Eq. (3.51))
(s)
K̃r ed,M N = Kr(s)
ed,R (3.53)
(s)
but the MNM-reduced mass matrix M̃r ed,M N is obtained differently. The residual
(s)
mass term Mr,bb of the matrix M(s)
f r ee in Eq. (3.45) is neglected, resulting in a modified
matrix labeled ⎡ ⎤
I00
(s)
M f r ee,M N = ⎣0 I 0⎦ (3.54)
000
instead of M(s)
f r ee for the MNM Voormeeren et al. (2011). The MNM-reduced mass
matrix writes now
(s)
M̃r ed,M N = T(s) (s) (s) (s)
T
This gives in fact inconsistent equations of motion since the mass and stiffness
matrices are not reduced with the same basis. The assembly of the MNM-reduced
(s) (s)
matrices K̃r ed,M N and M̃r ed,M N in the global system proceeds in the same manner
as for the RM. Observing that the boundary DOF ub have no associated inertia
in Eq. (3.55), those DOF can be condensed out of the equation of motion of the
assembled problem and the final MNM-reduced matrices Kr ed,M N and Mr ed,M N are
obtained (MacNeal 1971). Thus, the size of the assembled MNM system is reduced
further by the number of DOF of ub .
The reduction transformation (3.41) can also be used directly, without transforming
A(s) g(s) (s)T
T
b =B λ in interface DOF (as done for the RM or for the MNM) if one
considered the dually assembled problem (see Sect. 2.3.2). In the physical domain,
the dual assembly writes
38 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
⎡ ⎤⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
.. .. .. .. .. .. .
⎢ . 0 .⎥ ⎢ ⎥ ⎢ . . 0 . ⎥ ⎢ . ⎥ ⎢ .. ⎥
⎢0 M(s) 0⎥ ⎢ (s) ⎥ ⎢ K(s) B(s) ⎥
T
⎢ (s) ⎥ ⎢ (s) ⎥
⎢ ⎥ ⎢ü ⎥ + ⎢ 0 ⎥ ⎢u ⎥ = ⎢f ⎥
⎢ . . .. ⎥ ⎢ . ⎥ ⎢ .. . ⎥ ⎢ . ⎥ ⎢ . ⎥
⎣ . . ⎦ ⎣ .. ⎦ ⎣ . .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦
··· 0 ··· 0 λ̈ · · · B(s) · · · 0 λ 0.
(3.56)
Assembling all substructures N in a dual fashion by keeping the interface forces λ
as unknowns, the entire structure can consequently be approximated by
⎡ ⎤
α (1)
⎢ η(1) ⎥
⎢ ⎥
⎢ .. ⎥
u ⎢ . ⎥
≈ T DC B ⎢ ⎥ (3.57)
λ ⎢α (N ) ⎥
⎢ ⎥
⎣ η(N ) ⎦
λ
The approximation of the dynamic equations of the dual assembled system (3.56) is
⎡ ⎤ ⎡ (1) ⎤
α̈ (1) α
⎢ η̈(1) ⎥ ⎢ η(1) ⎥
⎢ ⎥ ⎢ ⎥
⎢ .. ⎥ ⎢ .. ⎥
⎢ . ⎥ ⎢ . ⎥
Mr ed,DC B ⎢ ⎥ + Kr ed,DC B ⎢ ⎥ = TTDC B f (3.59)
⎢α̈ (N ) ⎥ ⎢α (N ) ⎥
⎢ ⎥ ⎢ ⎥
⎣ η̈(N ) ⎦ ⎣ η(N ) ⎦
λ̈ λ
with
N
B(s) Gr(s) M(s) Gr(s) B(s) .
T
Mr = (3.62)
s=1
3.2 Numerical Techniques for Model Reduction of Substructures 39
Mr ed,DC B and Kr ed,DC B are diagonal for the parts related to the different substruc-
tures. The coupling between the substructures is only achieved by the rows and
columns related to λ. The DCBM applies the reduction matrix T DC B consistently to
the mass and stiffness matrix resulting in a true Rayleigh–Ritz method.
The DCBM enforces only a weak compatibility between the substructures and
does not enforce a strong displacement compatibility between the interfaces com-
pared to many other common reduction methods (Rixen 2004). Considering the
system of Eqs. (3.56) and (3.35) multiplied by the reduction matrix TTDC B , the last
row of Eq. (3.59) results from
⎡ ⎤
M(1) ü(1) + K(1) u(1) + B(1) λ = f (1)
T
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢M(N ) ü(N ) + K(N ) u(N ) + B(N )T λ = f (N ) ⎥ (3.63)
⎢ ⎥
⎣ N ⎦
(s) (s)
B u =0
s=1
Replacing the strong interface compatibility condition of Eq. (3.56) by the weak
form according to the multiplication of Eq. (3.63) by Eq. (3.64) can be interpreted as
follows. Denote Δf (s) the residual forces of substructure s resulting from the weak
satisfaction of the local equilibrium of the substructure approximating the dynamics
by a small number of free interface normal modes. Name Δu(s) = Gr(s) Δf (s) the
displacements these residual force Δf (s) would create locally. Then the weak interface
compatibility condition Eqs. (3.63) and (3.64) states that a compatibility error (i.e., an
interface displacement jump) equal to the incompatibility of Δu(s) is permitted Rixen
(2004).
Compared to MacNeal’s and Rubin’s method (1971, 1975), the weak interface
compatibility of the DCBM avoids locking problems occurring during the appli-
cation of the aforementioned methods. Therefore, the approximation accuracy is
improved Rixen (2004). But the fact that a weak interface compatibility is allowed
in the DCBM implies that the infinite eigenvalues related to the Lagrange multipli-
ers λ in the non-reduced problem (3.56) are now becoming finite and negative Rixen
(2011). In practice those negative eigensolutions will appear only in the higher eigen-
value spectrum if the reduction space is rich enough (Rixen 2011). Nevertheless, the
reduction basis has to be selected with care avoiding potential nonphysical effects of
the possibly occurring negative eigenvalues.
If Mr in Eq. (3.60) is neglected strong interface compatibility is enforced again
and the DCBM reduced system with Mr = 0 is equivalent to the MNM (Rixen 2004).
Then static condensation can be applied again to remove λ (as it was done for ub at
the end of the derivation of the MNM in section “MacNeal Method (MNM)”) from
40 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
the assembled system since no mass is associated. Thus, the size of the assembled
system is reduced again by the number of DOF of λ.
The Benfield truss (Benfield and Hruda 1971) of Fig. 3.3 is used to compare the
results obtainable by the HCB, the MNM, the RM, and the DCBM. The planar truss
consists of two substructures having uniform bay section whereas all members have
constant area and uniform stiffness and mass properties. The left component consists
of five equal bays and has a total of 18 joints and the right component consists of
four equal bays and has a total of 15 joints (Benfield and Hruda 1971). The lowest
eigenfrequencies ω of the entire structure shall be approximated by the different
methods.
The relative error εr el, j = |ωr ed, j − ω f ull, j |/ω f ull, j of the jth eigenfrequency is
used as a criterion to assess the accuracy of the different methods. Thereby, ω f ull, j
is the jth eigenfrequency of the full (non-reduced) system and ωr ed, j represents the
jth eigenfrequency of the reduced system obtained by each method.
Using 5 elastic (fixed or free interface normal modes) per substructure the relative
errors εr el depicted in the semi-log graph in Fig. 3.4 are resulting. Since all methods
give the correct rigid-body modes only the relative errors of the elastic modes are
plotted. All methods give a relative error of less than 1% for the first six eigenfre-
quencies. Comparing the free interface methods for this example, the RM performs
always better than the DCB and the DCB performs again always better as the MNM.
The HCB and the DCB result in similar frequency errors.
The sparsity pattern of the reduced stiffness matrix Kr ed and reduced mass
matrix Mr ed of the HCB (Fig. 3.5), the MNM (Fig. 3.7), the RM (Fig. 3.8) and
the DCB (Fig. 3.6), respectively, illustrate the differences of the assembled reduced
structures. Both the reduced stiffness matrix Kr ed and the reduced mass Mr ed matrix
applying the HCB and the DCB, respectively, have only diagonal entries for the sub-
parts of each substructure. On the one hand, the coupling between the substructures
using the HCB is entirely achieved by the last rows and last columns in the mass
matrix Mr ed,C B (Fig. 3.5b) and the remaining part is diagonal (Bampton and Craig
1968). On the other hand, the coupling applying the DCB is entirely achieved by the
last rows and last columns in the stiffness matrix Kr ed,DC B (Fig. 3.6a) and again the
remaining part is diagonal (Rixen 2004). The corresponding degrees of freedoms are
either the interface displacements ub or the interface forces λ but no direct coupling
between the modal parameters of adjacent substructures occurs which ensures the
sparse structure.
In contrast, the sparsity pattern of stiffness matrix Kr ed and the reduced mass
matrix Mr ed obtained by the MNM and the RM, respectively, show full matrices.
The MNM gives indeed an entirely diagonal reduced mass matrix Mr ed,M N (Fig.
3.7b) but causes always a full coupling between all DOF of all substructures via the
reduced stiffness matrix Kr ed,M N (Fig. 3.7a). This makes the reusability of reduced
models obtained by the MNM very inefficient and therefore nearly impossible from
a practical point of view.
The RM also causes a coupling between the substructures via interface displace-
ments ub in the reduced stiffness matrix Kr ed,R (Fig. 3.8a) as well as in the reduced
mass Mr ed,R (Fig. 3.8b). Moreover, all DOF belonging to one reduced substructure
are coupled with all other DOF of the same substructure which is why the reduced
matrices of the RM are full for the substructure blocks and not diagonal.
42 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
This result concerning the sparsity of the reduced matrices is outlined in Table 3.1,
which shows the number n of nonzero elements in the reduced matrices Kr ed and Mr ed
and the sum n total of both obtained by the different methods for this example. The
reduced matrices of the HCB, the MNM, and the DCB contain a similar number
of entries while the RM causes even for such a simple example a remarkable high
number of entries. The number of entries of the MNM are comparable to the HCB and
the DCB but will increase dramatically if the number of substructures is increased
since Kr ed will always be completely full.
3.2 Numerical Techniques for Model Reduction of Substructures 43
Table 3.1 Number n of nonzero elements in the reduced matrices obtained by the different methods
for the Benfield truss using 5 normal modes per substructure
HCB MNM RM DCB
n in Kr ed 40 216 314 196
n in Mr ed 118 16 354 50
n total 158 232 668 246
In the previous sections, we have outlined the classical substructure reduction meth-
ods (or Component Mode Synthesis) using either free or fixed interface modes.
Other methods, are not outlined here, like for instance the Craig–Chang approach
which uses also free interface modes, but computes the interface forces from a global
(assembled) problem (Craig and Chang 1977).
Many other variants of CMS methods were published over the past years and they
can be classified as follows:
• Loaded Interface Modes Some authors have proposed to compute the quasi-static
modes around a central frequency (Shyu et al. 1997) or to use vibration modes
with an impedance attached to the interface. In certain cases, such modifications
can improve the accuracy of the reduced model.
• Modal Truncation Augmentation and Moment Matching The vibration modes
used, for instance, in the HCB or in the RM approaches are not specifically tuned
for the excitations coming through the interface, whereas the static modes are. It is
possible to enrich the static modes with higher order static contribution computed
over a Krylov series. This leads to a method original called Modal Truncation
Augmentation (or MTA) (Dickens and Stroeve 2000) and was generalized to higher
orders corrections in the HCB in Rixen (2002). Later, techniques called Moment
Matching (Su and Craig 1991), that basically used the same ideas as MTAs, were
applied for model reduction mainly in the control community (see, for instance,
(Antoulas 2005, Holzwarth and Eberhard 2015, Stoev and Möhring 2010). The
idea of MTA was also applied for the DCBM (Rixen 2009, Kim et al. 2017).
• Balanced Truncation In the control community, reduction is seen from the point
of view of controllability and observability. This leads to representation modes
derived from so-called Grammians that were used for reducing substructure models
in Holzwarth and Eberhard (2015), Stoev and Möhring (2010). Balancing of State-
Space models will be discussed in Sect. 3.4.4.
• Mixed Methods Several methods were published where different types of vectors
(attachment modes, static modes ... ) were combined. One of the most commonly
used in major Finite Element software (although not cited much in the community,
maybe because the explanations in the publication are poor) is the one from Herting
(Herting 1985). Note also that it is possible to mix primal and dual assembly
44 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
but are combined to minimize the compatibility error between connecting substruc-
ture interfaces. Aoyama et al. (2001) presented a method similar to S-CC except
that each interface is assembled and then reduced separately. The recently presented
Hybrid-level Characteristic Constraint (H-CC) mode method blends these concepts,
performing an S-CC type analysis on each pair of interfaces. In a related work,
Balmès (1996) explored a CMS basis that defined arbitrary interface deformations
to describe a set of generalized DOF along an interface. A few other less commonly
known methods are reviewed in (Krattiger et al. (2019)).
The following reviews three different interface reduction strategies: the S-CC
method (Castanier et al. 2001), the Exact-Compatibility (EC) (Hong et al. 2013), the
Weak-Compatibility (WC) method (Kuether et al. 2017) and the Virtual Node (VN)
method. Additional methods are reviewed in Krattiger et al. (2019). The phrase
“system-level” indicates that interface reduction is applied to the fully assembled
HCB model, whereas “local-level” indicates that interface reduction is performed on
the individual substructures before they are assembled.
As done by Craig and Chang (1977), and detailed by Castanier et al. (2001), the
HCB models for all subcomponents are first assembled, the interface partition of the
assembled system is extracted and then an eigenvalue analysis is performed
S-CC
HCB
Kbb − ωr2 Mbb
HCB
φr = 0. (3.65)
The eigenvectors obtained are called S-CC modes and the first n vectors are collected
into a matrix,
Φ S-CC = φ S-CC
1 , . . . , φ S-CC
n . (3.66)
The matrix Φ S-CC is used to replace the interface DOF of the HCB model into
amplitudes of the S-CC modes,
ub ≈ Φ S-CC qb . (3.67)
The S-CC reduced mass (and similarly stiffness) matrices are obtained by pre-
and post- multiplying the assembled HCB matrices in Eq. (3.33) with the S-CC
transformation matrix to obtain
T HCB HCB
I 0 Mii Mib I 0
MS-CC = . (3.68)
0 Φ S-CC HCB
Mbi HCB
Mbb 0 Φ S-CC
TS-CC
46 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
This reduces the size of the interface partition of the HCB model from the number
of physical DOF on the interface to the number of S-CC modes in the truncated
eigenvector set, n Φ b . It is interesting to note that static reduction modes were proposed
for the interface of a dual Craig–Bampton model in Rixen (2011).
As mentioned previously, the prior method requires that the complete HCB model of
each component be assembled prior to reduction. This may not provide the desired
computational gains in many applications, and it requires knowledge of what the
component will connect to in order to perform the reduction. An alternative is the
local-level interface reduction techniques, where the interface is reduced before the
subsystems are assembled.
The EC L-CC interface reduction was first proposed by Hong et al. (2013) and is
uses a similar eigenvalue analysis on the interface partition, only this is performed
on each individual substructure prior to assembly. For the jth subcomponent, the
following eigenvalue problem is solved:
HCB ! L-CC "
Kbb, j − ωr2 Mbb,
HCB
j φj r
=0 (3.69)
The L-CC modes are denoted Φ L-CC j for the jth substructure. The L-CC modes for
each substructure are then used to reduce the interface DOF of that substructure.
Once again the W-bracket system in Fig. 3.2 can be used to illustrate the process.
The L-CC modes in Φ L-CC
1 from Eq. (3.69) have interface DOF for boundary sets
1
and
2 , while the modes in Φ L-CC
2 have interface DOF for boundary sets
2 and
3 .
Substructures 3, 4, and 5 only have one interface
1 ,
2 , and
3 , respectively. The
L-CC modes for each subcomponent (1 through 5) are expressed then as
L-CC L-CC
Φ 1 ,1 Φ 2 ,2
Φ L-CC = , Φ2
L-CC
= L-CC ,
1 Φ L-CC
2 ,1 Φ 3 ,2 (3.70)
Φ L-CC
3 = Φ L-CC
1 ,3 , Φ 4
L-CC
= Φ L-CC
2 ,4 , Φ 5
L-CC
= Φ L-CC
3 ,5 .
Following what was done in Krattiger et al. (2019), for each substructure the
interface modes are computed for the entire interface set simultaneously, even if the
set of interface nodes comprises multiple distinct interfaces. Then, for each distinct
interface, the modes from every substructure are concatenated to create a complete
aug
basis for each interface. The result is called the augmented set, e.g., Φ
1 . For the
W-bracket system in Fig. 3.2, we have the following augmented sets:
3.3 Interface Reduction with the Hurty/Craig–Bampton … 47
aug (3.71)
3 ,2 Φ 3 ,5 .
Φ
3 = Φ L-CC L-CC
The augmented interface L-CC modes for each interface will generally not be linearly
independent and this must be corrected before they can be used. Indeed, the matrices
may be rank deficient to numerical precision such that they cannot be used as a basis.
To address this, Hong et al. used a Singular Value Decomposition (SVD) on the
augmented L-CC modes,
T
aug aug aug aug
U j S j V j = Φj , (3.72)
aug aug
where U j and S j are the left singular vectors and the diagonal singular value
matrix for the jth interface set. If some singular values are very small then the basis
is ill-conditioned and the corresponding singular vectors must be truncated. In Hong’s
work (Hong et al. 2013), the truncation tolerance used was σ R = 1 × 10−4 = 0.01%,
where σ R is the ratio of the minimum singular value to the maximum for the retained
modes. In Krattiger et al. (2019), the effect of σ R on the model performance was
evaluated, revealing that σ R = 1 × 10−4 = 0.01% was optimal for the W-bracket,
but this can be expected to be problem dependent.
The resulting matrices U
j can then be grouped for each subsystem j to form
a block-diagonal matrix that contains all the interface sets for substructure j. Take
substructures 1 and 3 as examples. The orthogonal L-CC basis is
U1 0
U1 = , U3 = U1 . (3.73)
0 U2
Then, the reduced mass and stiffness matrices for each subcomponent j are
T HCB HCB
I 0 Mii, j Mib, j I 0
MEC
j
L-CC
= HCB HCB
. (3.74)
0 Uj Mbi, j Mbb, j
0 Uj
Weak Compatibility
The WC L-CC method also begins by computing local interface modes for each
substructure using Eq. (3.69). This transformation expresses the physical interface
DOF as a linear combination of the interface modes,
ub, j ≈ Φ L-CC
j qb, j . (3.75)
48 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
The models for each substructure are not statically complete at the interface, so
they cannot simply be assembled. Furthermore, because the basis for each interface
may not span the same space, they could lock if strict constraints are enforced.
Instead, the following approach is used. The constraint equations for the interface
are written as
Bqcat = 0 (3.77)
Collecting all of the constraint equations into a single matrix and expressing them in
modal coordinates gives
⎧ ⎫
⎡ ⎤⎪
⎪ ⎪
⎪
⎪
⎪
qb,1 ⎪
⎪
Φ L-CC −Φ L-CC ⎪
⎪ ⎪
⎪q ⎪
0 0 0
⎢ 1 ,1 1 ,3 ⎥⎪ ⎪
⎢ L-CC ⎥⎪ b,2 ⎪
⎪
⎢ Φ 2 ,1 −Φ L-CC 0 0 0 ⎥⎨ ⎬
⎢ 2 ,2 ⎥ q = 0. (3.79)
⎢ L-CC ⎥⎪ b,3
⎪
⎢ Φ 2 ,1 0 0 −Φ L-CC 0 ⎥⎪
⎪ ⎪
⎪
⎣ 2 ,4 ⎦⎪
⎪ qb,4 ⎪
⎪
⎪
⎪ ⎪
⎪
0 Φ 3 ,2
L-CC
0 0 −Φ L-CC ⎪
⎪ ⎪
⎪
q ⎭
⎩
3 ,5
b,5
B
Then, the transformation Lb is sought that minimizes BLb . Specifically, the SVD,
B = UB SB VBT is computed and the columns of VB are retained if their singular values
are less than a null space tolerance, σ N . The compatibility will be good (although
typically not exact) as long as the singular values are small. Those motions at the
interface that are compatible will govern the motion of the coupled system, while
those displacements that are deemed incompatible are discarded.
Once an approximate null space has been obtained, it is placed into the boundary
partitions of the overall coupling matrix. For the W-bracket model, this gives
3.3 Interface Reduction with the Hurty/Craig–Bampton … 49
⎧ ⎫ ⎡ ⎤
⎪
⎪ qi,1 ⎪ I 0 0 00 0
⎪q ⎪
⎪ ⎪
⎪ ⎢0 N1 ⎥
⎪
⎪ ⎪ 0 0 00
⎪q ⎪ ⎪ ⎢ ⎥⎧ ⎫
b,1
⎪
⎪ i,2 ⎪
⎪ ⎢0 0 ⎥
⎪
⎪ ⎪
⎪ ⎢ I 0 00 ⎥⎪⎪
⎪
qi,1 ⎪
⎪
⎪
⎪ ⎪
qb,2 ⎪ ⎢0 N2 ⎥ ⎪ qi,2 ⎪
⎪
⎪
⎨ ⎬ ⎢
⎪ 0 0 00 ⎥⎪⎪
⎨
⎪
⎪
⎬
qi,3 ⎢0 0 I 00 ⎥
0 ⎥ qi,3
=⎢
⎢0 ⎥ , (3.80)
⎪
⎪ qb,3 ⎪
⎪ ⎢ 0 0 00 N3 ⎥ ⎪
⎪ qi,4 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎢0 ⎥ ⎪
⎪ ⎪
⎪
⎪
⎪ qi,4 ⎪
⎪ 0 0 I0 0 ⎥⎪ qi,5 ⎪
⎪
⎪ ⎪ ⎢ ⎪
⎩ ⎪
⎭
⎪
⎪ q ⎪ ⎢⎪ 0 0 0 00 N4 ⎥
⎥ qb
⎪ b,4 ⎪
⎪ ⎪ ⎢
⎪ ⎣ ⎦
⎪
⎪ q ⎪ 0 0 0 0I 0
⎩ i,5 ⎪⎭
qb,5 0 0 0 00 N5
LWC L-CC
where N j denotes the rows of N corresponding to the interface DOF of the jth
substructure. Finally, the reduced-order system is assembled by summing over the
substructures.
n
WC L-CC T WC L-CC WC L-CC
MWC L-CC = Lj Mj Lj , (3.81)
j=1
where LWCj
L-CC
contains the rows of LWC L-CC corresponding to the jth uncoupled
substructure.
Virtual Node
Lindberg et al. (2013) presented an interface reduction method that formalizes some
concepts that have been used in industry for many years. This method was proposed
as a way to connect stiff and soft substructures at a single Virtual Node (VN) having
six DOF. In Krattiger et al. (2019), this approach was generalized to allow n virtual
nodes along an interface. Each virtual node is connected to neighboring nodes with
multi-point constraints, and so the node set can collectively move in six shapes defined
by three orthogonal translations and three rotations around the VN’s position.
This type of approach is used very commonly in industry, although there are
various flavors. For example, the RBE3 elements that are commonly used in Nastran
perform a similar operation, although they do not eliminate the interface nodes merely
compute the motion of a virtual node that tracks the average motion of the interface.
The experimental counterpart, operating on measured FRFs, can be found in the
virtual point transformation, which is discussed in Sect. 4.6.
Other Approaches
displacement patches used in the VN approach, thus allowing the interfaces to deform
in a natural way but reducing the number of interface DOF (Coppolino 2018).
Other extensions to the interface reduction methods proposed here are certainly
possible. One notable addition is the Hybrid method (H-CC), which is described in
detail in Krattiger et al. (2019).
x = Ax + Bs + Ew, (3.82)
r = Cx + Ds + v.
Here x ∈ Rn is the state vector, s ∈ Rn s is the vector of known stimuli and r ∈ Rnr
is the output vector. The system order is n, the number of independent stimuli ele-
ments of the excitation vector is n s , and the number of response elements of the output
vector is n r . The column vectors w ∈ Rn w and v ∈ Rn v represent unknown process
noise and output noise, respectively. The matrices A, B, C, D, and E are state-space
coefficient matrices out of which B ∈ Rn×n s is the output matrix, C ∈ Rnr ×n is the
input matrix, and D ∈ Rnr ×n s is the direct throughput matrix. In general, for a non-
linear 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, the system of interest is often put on soft supports to isolate
it from ambient vibrations. The system is excited by some measured stimuli s and
its responses r are picked up as sensor signals which are filtered, digitized and
processed in a data acquisition system. In reality, the registered stimuli s is not the
true system stimuli since sensors are not ideal and the support system does not fully
isolate the test object from ambient vibrations which lead to unknown vibration
stimuli w transmitted to the system. However, in many cases, the stochastic terms
3.4 State-Space Model Reduction 51
x = Ax + Bs (3.83)
r = Cx + Ds
for which there is an initial value time domain solution given by a convolution integral
expression, see Eq. (3.89) below.
The frequency domain counterpart for the steady-state dynamics of the determin-
istic time domain model follows from the assumption that a harmonic stimulus with
amplitude ŝ and frequency ω, i.e., s(t) = Re{ŝeiωt }, results in a harmonic response
Re{r̂eiωt }. This gives the relation between input and output as
which defines the system’s transfer function H, also known as its frequency response
function as
H(ω) = C(iωI − A)−1 B + D (3.85)
An important special case is for the impact stimuli s(t) = [1 1 . . . 1]T δ(t) of a
system at rest with x0 = 0 and δ(t) being the Dirac delta function. For such an ideal
impact (hit), the convolution integral in (3.89) evaluates to
rhit (t) = [Dδ(t) + CA0 B + tCAB/1! + t 2 CA2 B/2! + . . .][1 1 . . . 1]T . (3.92)
The terms of the system’s impact response series are associated to its Markov
parameters defined as
h0 = D hk = CAk−1 B, k = 1, 2, . . . (3.93)
The infinite-size Hankel matrix H of the system is constructed from its Markov
parameters as follows:
⎡ ⎤ ⎡ ⎤
h1 h2 h3 h4 . . . CB CAB CA2 B CA3 B ...
⎢ .. ⎥ ⎢ .. ⎥
⎢h2 h3 h4 h5 . ⎥ ⎢ CAB CA2 B CA3 B CA4 B . ⎥
⎢ ⎥ ⎢ ⎥
H =⎢ . ⎥=⎢ .. ⎥ (3.95)
⎢h3 h4 h5 h6 .. ⎥ ⎢CA2 B CA3 B CA4 B CA5 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
Ak = n−1 j=0 α jk A . This, in turn, leads to that the number of linearly independent
k
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 with the largest impulse response of the system.
Discrete-Time State-Space Modeling. The convolution solution (3.89) is utilized
for deriving efficient computational schemes for discrete-time solutions. Consider
the solutions at two consequent discrete times kT and kT + T , where T = 1/ f s and
f s is referred to as the sampling frequency. The solutions are
# kT
xk = eAkT x0 + eA(kT −τ ) Bs(τ )dτ (3.96)
0
and
# kT
A(kT +T )
xk+1 = e x0 + eA(kT +T −τ ) Bs(τ )dτ (3.97)
0
# kT # kT +T
A(kT −τ )
=e AT
(e AkT
x0 + e Bs(τ )dτ ) + eA(kT +T −τ ) Bs(τ )dτ
0 kT
# kT +T
= eAT xk + eA(kT +T −τ ) Bs(τ )dτ.
kT
It can be shown that for stable systems ||Â|| < 1, which is a requirement for the
time-stepping numerical integration scheme (3.100) to be stable.
Combined with the trivial equation Iq̇ − Iq̇ = 0, Eq. (3.104) can be written as the
state-space dynamic equation
0 I 0
ẋ = x+ s ≡ Ax + Bs
−M−1 (V + G) −M−1 (K + λK g + L) M−1 Ps
(3.105)
where the state vector x has been introduced so that
$ % $ %
u u̇
x= and thus ẋ = (3.106)
u̇ ü
Let further Pa be the picking matrix that picks out the accelerations of interest ra
from the acceleration partition of the state vector’s time derivative ẋ. Then the output
equation for accelerations becomes
r ≡ ra = 0 Pa ẋ = 0 Pa [Ax + Bs] ≡ Ca x + Ds. (3.108)
0 −I 0
A=− B = (3.110)
M−1 (V + G) M−1 (K + λK g + L), M−1 Ps ,
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
Cd Pd 0 Dd 0
C = ⎣ Cv ⎦ = ⎣ 0 P v ⎦ D = ⎣ Dv ⎦ = ⎣ 0 ⎦
Ca , [0 Pa ]A Da [0 Pa ]B.
Here it may be noted that only acceleration output has static contribution through
the direct throughput term Da s, i.e., direct contribution from the excitation besides the
dynamics of the system via the dynamic equation ẋ = Ax + Bs. 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 ẋ = Ax + Bs, x(0) = x0 . The velocity output
is also related to displacement output by rv = drd /dt and therefore it is also given
that
rv = Cd [Ax + Bs] = Cd Ax + Cd Bs = Cd Ax (3.111)
since it can be easily verified from above that Cd B = 0 and thus no direct throughput
from s occur to velocity in this formulation either.
For other output quantities which relate linearly to the above quantities, e.g.,
strains and stresses, the corresponding state-space description is straightforward.
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.
The observability of a state-space model {A, B, C, D} can be investigated pro-
vided the model 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 = Cx + Ds, and therefore the sequence
R = Ox + GS (3.113)
with R being the vectorial concatenation of r,ṙ, r̈, . . . and S being the vectorial
concatenation of s, ṡ, s̈, . . .. The observability matrix of the system O is
⎡ ⎤
C
⎢ ⎥CA
⎢ ⎥
⎢ ⎥CA2
O=⎢ ⎥ (3.114)
⎢ ⎥ ..
⎣ ⎦ .
CAn−1
and matrix G is irrelevant for observability. If the n r n-by-n observability matrix has
a rank of less then n, some linear combination of the n columns add to zero, and
therefore there are (transformed) states that do not contribute to R. 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 equation (3.113) with the transpose of O to obtain
OT R = OT Ox + OT GS. (3.115)
If O is of full rank then OT O is nonsingular and thus the state vector x can be
determined as its unique solution, which is
Is this also the solution to Eq. (3.113)? If it is not, for the two different solutions x1
and x2 we would have O(x1 − x2 ) = 0 which means that some linear combination of
the columns of O is zero, which contradicts the assumption that O is of full column
58 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
with ||Â|| ≤ 1 does not suffer from this problem and is thus better suited for numerical
evaluation.
State Controllability. The concept of controllability relates 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: The system
ẋ= Ax + Bs is said to be state controllable at time t = t0 if there exists a piecewise
continuous inputs(t) that will drive the initial statex(t0 ) to any final statex(t f ) within
a finite time intervalt 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 # tf
x(t f ) = eAt f x(0) + eA(t f −τ ) Bs(τ )dτ, (3.118)
0
n−1
The Cayley–Hamilton theorem giving Ak = J =0 α jk Ak together with the defi-
nition exp(At) = I + tA + t 2 A2 /2! + . . . gives
n−1
e−Aτ = α j (τ )A j (3.120)
j=0
3.4 State-Space Model Reduction 59
with
∞
α j (τ ) = (−1)k α jk τ k /k! (3.121)
k=0
n−1 # tf
n−1
Δx(t f ) = Aj B α j (τ )s(τ )dτ ≡ Aj Bc j (t f ) (3.122)
j=0 0 j=0
with # tf
c j (t f ) = α j (τ )s(τ )dτ (3.123)
0
and represents a set of n equations and nn u unknowns. The matrix equation has a
solution for any Δx(t f ) provided that the n × nn s matrix
C = [B AB A2 B . . . An−1 B] (3.125)
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.
Similar to the observability matrix of continuous-time systems, the controllability
matrix has bad numerical properties when the system order gets large and ||A|| > 1.
The observability matrix of the associated discrete-time model
2 n−1
C = [B̂ ÂB̂ Â B̂ . . . Â B̂] (3.126)
of the stable system is then better suited. This matrix related to discrete-time entities
is sometimes called the reachability matrix.
Controllability and Observability Grammians. Suppose a realization is given
by ẋ= Ax + Bs, x(t0 ) = x0 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
# tf
x(t f ) = eA(t f −t0 ) x(0) + eA(t f −τ ) Bs(τ )dτ (3.127)
t0
This is an integral equation for the sought stimuli s. Its solution can be shown to
be
(t0 −τ )
G−1
T
s(τ ) = −BT eA c (t0 , t f )x0 , (3.129)
where # tf
eA(t0 −τ ) BBT eA (t0 −τ )
T
Gc (t0 , t f ) = dτ (3.130)
t0
is the controllability Grammian. To see that Eq. (3.130) really gives the solution, let
it enter into the integral of Eq. (3.128) which renders
# tf
eA(t0 −τ ) BBT eA (t0 −τ )
G−1
T
− c (t0 , t f )x0 dτ = (3.131)
t0
# tf
eA(t0 −τ ) BBT eA (t0 −τ )
dτ ]G−1
T
−[ c (t0 , t f )x0 =
t0
−Gc (t0 , t f )G−1
c (t0 , t f )x0 = −x0
since Gc (t0 , t f ) is constant. The process requires that the controllability Grammian
is nonsingular and therefore invertible, which is, therefore, the Grammian-related
condition for controllability.
Similarly, for the realization ẋ= Ax + Bs, r = Cx, the corresponding observ-
ability Grammian
# tf
(t0 −τ )
CT CeA(t0 −τ ) dτ
T
Go (t0 , t f ) = eA (3.132)
t0
must be nonsingular (Kailath 1980) for the realization to be observable from the
output during over times t0 ≤ t ≤ t f .
The Grammian singularity test is known to be *a test for linear dependence of
t
functions l j and lk for which the Grammian G jk = t01 l j lk dτ should be nonzero 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
# tf
d d
eA(t−τ ) BBT eA (t−τ ) dτ =
T
Gc (t, t f ) = (3.133)
dt dt t
# tf
d A(t−τ ) T AT (t−τ )
[e BB e ]dτ − e0 BBT e0 =
t dt
AGc (t, t f ) + Gc (t, t f )AT − BBT .
3.4 State-Space Model Reduction 61
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 Gramian derivative with respect to initial
time variation is zero. The infinitely long time Grammian G∞ c is thus governed by
the Lyapunov equation
AG∞ ∞
c (t, t f ) + Gc (t, t f )A − BB = 0.
T T
(3.134)
On the other hand, the corresponding long time observability Grammian G∞ o can
be shown (see Kailath 1980) to be the solution of another Lyapunov equation
A T G∞ ∞
o (t, t f ) + Go (t, t f )A − C C = 0.
T
(3.135)
Numerical methods for solving the Lyapunov equations exist but are computa-
tionally expensive. For large order systems (say n > 10.000), the solution may take
hours to obtain even with a fast computer (as of year 2018). Recall that the Grammi-
ans obtained by the Lyapunov equation are for infinitely long control and observation
times. The general controllability Grammian, as defined by Eq. (3.130), 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. (3.86) normally gives other Gramians Gc and Go . This is used to the advantage
for model reduction using the balancing realization, one important realization form
discussed below.
Under the assumption that the free decay is governed by the solution x(t) = Φeλt
one has
AΦ = λΦ, r(t) = CΦeλt (3.137)
for which there are nontrivial solutions (λ, Φ) provided that λ is a root of the char-
acteristic polynomial det(A − λI) = 0. Such roots are also called system poles and
the associated solution vectors Φ are the eigenvectors of A. Assuming that A and C
stem from a physically realizable system (for which the response r to a real-valued
stimulus r is real), a real form of A ∈ Rn×n and C ∈ Rnr ×n is possible. Thus the char-
acteristic polynomial has real-valued polynomial coefficients and the poles are thus
62 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
AT = TΛ, (3.138)
where the modal matrix T has all system eigenvectors as columns and the eigenvalue
matrix Λ has the associated system poles, normally sorted in increasing magnitude
order, as elements along its diagonal. It has been shown, see e.g., Golub and van
Loan (2012), that Λ is fully diagonal provided that all system poles are unique.
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 Λ 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 T−1 AT = T−1 TΛ = Λ together with the state transformation x = Tz,
the realization (3.87) becomes
and since Λ is diagonal the first-order differential equation system thus becomes
fully decoupled. In free vibration in one mode only, i.e., z k (t) = 0, z m = 0 ∀m = k,
one notes that the response is
r(t) = C̄.k z k (t) (3.140)
and the k:th column of C̄ (denoted C̄.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, C̄, D} is pos-
sible. For each complex-conjugate pair of eigenvalues λk = Re{λk } ± Im{λk }, the
corresponding 2 × 2 block of the system matrix becomes
Re{λk } −Im{λk }
(3.141)
Im{λk } Re{λk }
Since the state matrix Λ is diagonal, with the kth eigenvalue λi on its kth row and
column, one notes in particular that a transfer function element Hi j can be expressed
as
1
Hi j (ω) = C̄i. diag( )B̄. j + Di j (3.143)
iω − λk
n
C̄ik B̄k j
Hi j (ω) = + Di j , (3.144)
k=1
iω − λk
where 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.
A two-degree-of-freedom example
for which the eigenvalues to three significant digits are λ1,2 = −.0527 ± 6.18i
and λ3,4 = −.947 ± 16.14i [rad/s] and the transformed system on diagonal
form is
⎡ ⎤
−.0527 − 6.18i 0 0 0
⎢ 0 −.0527 + 6.18i 0 0 ⎥
Ā = ⎢
⎣
⎥
⎦
0 0 −.947 − 16.1i 0
0 0 0 −.947 + 16.1i
64 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
⎡ ⎤
.269 − .00555i
⎢ .269 + .00555i ⎥
B̄ = ⎢
⎣−.427 + .00779i ⎦
⎥
−.427 − .00779i
C̄ = −.00116 + .136i −.00116 − .136i .000435 + .0326i .000435 − .0326i
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
Fig. 3.10 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
3.4 State-Space Model Reduction 65
⎡ ⎤ ⎡ ⎤
.5 0 +.0177i −.0177i 0 1 0 0
⎢.5 0 −.0177i +.0177i ⎥ ⎢0 0 0 0 ⎥
T=⎢
⎣0
⎥ Ā = ⎢ ⎥
0 +.25 +.25 ⎦ ⎣0 0 −14.1421i 0 ⎦
0 0 −.25 −.25 0 0 0 14.1421i
(b) ⎡ ⎤
−.0378 −.0015 +.0222 − 0.1153i +.0222 + .1153i
⎢+.0545 +.0043 −.0111 − 0.1500i −.0111 + .1500i ⎥
T=⎢
⎣+.5685
⎥
−.0158 +.7537 + 0.2039i +.7537 − .2039i ⎦
−.8200 −.0106 1 1
66 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
⎡ ⎤
−15.0372 1 0 0
⎢ 0 −15.0372 0 0 ⎥
Ā = ⎢
⎣
⎥
⎦
0 0 −.4931 + 6.6319i 0
0 0 0 −.4931 − 6.6319i
(c) ⎡ ⎤
+.0499 +.0035 −.7071 0
⎢−.0499 −.0035 −.7071 0 ⎥
T=⎢
⎣−.7053 +.0002
⎥
0 −.7071⎦
−.0002 0 −.7071
⎡ ⎤
−14.1421 1 0 0
⎢ 0 −14.1421 0 0⎥
Ā = ⎢
⎣
⎥
0 0 0 1⎦
0 0 0 0
ż = T−1 −1
b ATb z + Tb Bs (3.146)
r = CTb z + Ds
is called a balanced realization, with Grammians balanced over the control and obser-
vation 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
R T R = Q. (3.148)
3. Compute a singular value decomposition to obtain the singular value matrix Σ for
singular values in decreasing order in the relation
3.4 State-Space Model Reduction 67
Tb = R T UΣ −1/2 . (3.150)
Many model order reduction techniques have been developed in order to balance
the accuracy and simplicity of the systems, see Refs. (Antoulas 2005, Noor 1994,
Ersal et al. 2008). 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
(Varga 1995, Qu 2004). 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. Skelton (1980, 1983) proposed a component cost analysis method for first-
order continuous-time systems which assigns a cost function associated with the
mean-squared system output under white noise stimuli. This method is best suited
for small-scale systems since it requires the computationally expensive solution of
a Lyapunov equation for the evaluation of the cost function. Another metric based
on the unit-step response was proposed by Aguirre (1993) and is based on a simple
formulation for either transfer function or state-space representation of the system.
Balancing Truncation. A reduction scheme that is popular in control engineering
is based on the balanced input–output relation M̄ = {T−1 −1
b ATb , Tb B, CTb , D} ≡
{Ā, B̄, C̄, D}. It has the 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
68 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
The model M̄r = {Ā11 , B̄1 , C̄1 , D} is then a reduced-order system obtained from
{Ā, B̄, C̄, D} by balanced truncation. It has some guaranteed properties related to
stability, controllability, observability, and impulse response which are given by the
following theorem.
Balancing Truncation Theorem: Given the controllable, observable, and stable
system M̄, the reduced-order system M̄r obtained by balanced truncation have the
following properties:
1. M̄r is balanced and has no unstable poles.
2. If λ p = eig(A11 ) = λq = eig(A22 )/∀ p, q then M̄r is controllable and observ-
able.
3. Let the ordered singular values of Sbeσi , i = 1, 2, . . . , q and let f ur ther the
number o f singular values o f Sr be k with k < q. The Hankel H∞ -norm of the
difference between the full-order systemS and the reduced-order systemSr is then
twice the sum of neglected Hankel singular values as: ||S − Sr ||H ∞ ≤ 2(σk+1 +
. . . + σq )
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 (1984) and Enns (1984).
Modal Truncation. 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. In the simplest form given by Davison (1966), the
lowest damped modes has been considered as the dominant subset to be kept while
the complement set was eliminated. The shortcoming of this reduction approach
is that the retained modes can be selected from the uncontrollable or unobservable
eigenvalue subspace of the original system. Such modes do not contribute to the
input–output relation of the system at all. Thus, the created reduced-order model is
not minimal. In Varga (1995), the original system is divided to several subsystems
which are in the balanced form. In each subsystem, eigenmodes with largest Hankel
singular values (the square root of the eigenvalues of the product of the observability
and controllability gramians) are considered as dominant modes to retain.
3.4 State-Space Model Reduction 69
ż1 Λ1 0 z1 B̄
= + 1 s
ż2 0 Λ2 z2 B̄2
z1
r = C̄1 C̄2 + Ds, (3.153)
z2
ns
c̄k b̄k
G(S, ω) − G r (Sr , ω) = (3.154)
k=nr +1
iω − λk
Thus, the H∞ norm of the error system is upper bounded by the following expres-
sion
ns
σ̄ (c̄k b̄k )
||G(S, ω) − G r (Sr , ω)||∞ = sup[σ̄ (G(S, ω) − G r (Sr , ω))] ≤ −
ω∈ k=nr +1
Re(λk )
(3.155)
where σ̄ (.) 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 (Ersal et al. 2008), e.g., the modal truncation
preserves the stability property of the full system.
The upper bound of Eq. (3.155) hints to a reduction strategy to yield the low-
order model. It remains to be determined which set of eigensolutions to be used in
70 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
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
ż k = λk z k + b̄k s
Δrk = c̄k z k , (3.156)
With a unit-impulse signal (S(ω) = 1∀ω) and a substitution of Eq. (3.159) into
Eq. (3.158) then leads to
# +∞
π b̄kH c̄kH c̄k b̄k
Mk = b̄kH (iω − λk )−1 c̄kH c̄k (iω − λk )−1 b̄k dω = − . (3.160)
−∞ 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 S(ω) = 1 ∀ω = [ω1 , ω2 ] and S(ω) = 0 elsewhere. That
leads to the following frequency-weighted metric
# ω2
M̄k = (ΔRk (ω) H ΔRk (ω))dω. (3.161)
ω1
References
Aguirre, L. A. (1993). Quantitative measure of modal dominance continuous systems. In Anon (ed.)
Proceedings of the 32nd IEEE conference on decision and control (vol. 3). IEEE. https://doi.org/
10.1109/CDC.1993.325629.
Antoulas, A. C. (2005). Approximation of large-scale dynamical systems. Philadelphia: Society for
Industrial and Applied Mathematics.
Aoyama, Y., & Yagawa, G. (2001). Component mode synthesis for large-scale structural eigenanal-
ysis. Computers and Structures, 79(6), 605–615.
Balmès, E. (1996). Use of generalized interface degrees of freedom in component mode synthesis.
In Proceedings of the 14th International Modal Analysis Conference, 1.
Balmès, E. (1997). Modèles expérimentaux complets et modèles analytiques réduits en dynamique
des structures. Ph.D. thesis, Université Pierre et Marie Curie, Paris VI, ONERA, Directions des
Structure, DRET/AC.
Bampton, M. C. C., & Craig, J. R. R. (1968). Coupling of substructures for dynamic analyses. AIAA
Journal, 6(7), 1313–1319.
Bathe, K. J., & Dong, J. (2014). Component mode synthesis with subspace iterations for controlled
accuracy of frequency and mode shape solutions. Computers and Structures, 139, 28–32.
Benfield, W., & Hruda, R. (1971). Vibration analysis of structures by component mode substitution.
AIAA Journal, 9(7), 1255–1261.
Castanier, M. P., Tan, Y. C., & Pierre, C. (2001). Characteristic constraint modes for component
mode synthesis. AIAA Journal, 39(6), 1182–1187.
Coppolino, R.N. (2018). Consolidation of weakly coupled experimental system modes. In Proceed-
ings of IMAC-XXXVI: A Conference on Structural Dynamics. Bethel, CT: Society for Experi-
mental Mechanics.
Craig, R. R., & Chang, C. J. (1977). On the use of attachement modes in substructure coupling for
dynamics analysis. 18st Structures (pp. 89–99). Structural Dynamics and Material Conference
San Diego: AIAA/ASME.
Craig, R. R, Jr., & Chang, C. J. (1977). Substructure Coupling for Dynamic Analysis and Testing.
Report, National Aeronautics and Space Administration.
Davison, E. (1966). A method for simplifying linear dynamic systems. EEE Transactions on Auto-
matic Control, 11(1), 93–101. https://doi.org/10.1109/tac.1966.1098264.
Dickens, J.M., & Stroeve, A. (2000). Modal truncation vectors for reduced dynamic
substructure models. In Structures, Structural Dynamics and Material Conference. 41st
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference and
Exhibit, Atlanta
72 3 Model Reduction Concepts and Substructuring Approaches for Linear Systems
Enns, D. F. (1984). Model reduction with balanced realizations: An error bound and frequency
weighted generalization. In Proceedings of the IEEE Conference on Decision and Control (pp.
127–132).
Ersal, T., Fathy, H.K., Rideout, D.G., Louca, L.S. & Stein, J.L. (2008). A review of proper modeling
techniques. Journal of Dynamic Systems, Measurement, and Control130(6). https://doi.org/10.
1115/1.2977484.
Géradin, M., & Rixen, D. (2015). Mechanical Vibrations: Theory and Application to Structural
Dynamics (3rd ed.). Chichester: Wiley.
Glover, K. (1984). All optimal hankel-norm approximations of linear multivariable system and their
L ∞ -error bounds. International Journal of Control, 39, 1115–1193.
Golub, G.H., van Loan, C.F. (2012). Matrix computations (Vol. 3). John Hopkins University Press.
Gruber, F., & Rixen, D. (2018a). A comparison of craig-bampton approaches for systems with
arbitrary viscous damping in dynamic substructuring. In IMAC-XXXVI: International Modal
Analysis Conference, Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Gruber, F. M., & Rixen, D. J. (2016). Evaluation of substructure reduction techniques with fixed
and free interfaces. Strojniški vestnik - Journal of Mechanical Engineering, 62(7–8), 452–462.
Gruber, F. M., & Rixen, D. J. (2018b). Dual craig-bampton component mode synthesis method for
model order reduction of nonclassically damped linear systems. Mechanical Systems and Signal
Processing, 111, 678–698.
Guyan, R. (1965). Reduction of mass and stiffness matrices. AIAA Journal, 3(2), 380.
Halmos, P. (1958). Finite dimensional vector spaces. New York: Van Nostrand.
Herting, D. (1985). A general purpose, multi-stage, component modal synthesis method. Finite
Elements in Analysis and Design, 1(2), 153–164.
Holzwarth, P., & Eberhard, P. (2015). Svd-based improvements for component mode synthesis in
elastic multibody systems. European Journal of Mechanics-A/Solids, 49, 408–418.
Hong, S. K., Epureanu, B. I., & Castanier, M. P. (2013). Next-generation parametric reduced-order
models. Mechanical Systems and Signal Processing, 37(12), 403–421.
Hurty, W. C. (1965). Dynamic analysis of structural systems using component modes. AIAA Journal,
3(4), 678–685.
Irons, B. (1965). Structural eigenvalue problems-elimination of unwanted variables. AIAA Journal,
3(5), 961–962.
Jakobsson, H., Bengzon, F., & Larson, M. G. (2011). Adaptive component mode synthesis in linear
elasticity. International Journal for Numerical Methods in Engineering, 86, 829–844.
Kailath, T. (1980). Linear systems. Wiley.
Kim, J. H., Kim, J., & Lee, P. S. (2017). Improving the accuracy of the dual craig-bampton method.
Computers and Structures, 191, 22–32.
Krattiger, D., Wu, L., Zacharczuk, M., Buck, M., Kuether, R. J., Allen, M. S., et al. (2019). Interface
reduction for hurty/craig-bampton substructured models: Review and improvement. Mechanical
Systems and Signal Processing, 114, 579–603.
Kuether, R. J., Allen, M. S., & Hollkamp, J. J. (2017). Modal substructuring of geometrically
nonlinear finite element models with interface reduction. AIAA Journal.
Lindberg, E., & Höransson, P. (2013). Component mode synthesis using undeformed interface
coupling modes to connect soft and stiff substructures. Shock and Vibration, 20, 157–170.
MacNeal, R. H. (1971). Special issue on structural dynamics a hybrid method of component mode
synthesis. Computers and Structures, 1(4), 581–601.
Moore, B. (1981). Principal component analysis in linear systems: Controllability, observability
and model reduction. IEEE Transaction on Automatic ControlAC-26, 17–32 (1981).
Noor, A. K. (1994). Recent advances and applications of reduction methods. Applied Mechanics
Reviews, 47(5), 125–146.
Qu, Z. Q. (2004). Model order reduction techniques: with application in finite element analysis.
New York: Springer.
References 73
Rahrovani, S., Vakilzadeh, M. K., & Abrahamsson, T. J. S. (2014). Modal dominancy analysis based
on modal contribution to frequency response function 2-norm. Mechanical Systems and Signal
Processing, 48(1–2), 213–231.
Rixen, D. (2002). High order static correction modes for component mode synthesis. In Fifth World
Congress on Computational Mechanics. Vienna, Austria.
Rixen, D.J. (2009). Dual craig-bampton with enrichment to avoid spurious modes. In IMAC-XXVII:
International Modal Analysis Conference, Orlando, FL. Bethel, CT: Society for Experimental
Mechanics.
Rixen, D.J. (2011). Interface reduction in the dual craig-bampton method based on dual interface
modes. In IMAC-XXIX: International Modal Analysis Conference, Jacksonville, FL: International
Modal Analysis Conference, Jacksonville, FL. Bethel, CT: Society for Experimental Mechanics.
Rixen, D. (2004). A dual Craig-Bampton method for dynamic substructuring. Journal of Compu-
tational and Applied Mathematics, 168(1–2), 383–391.
Rubin, S. (1975). Improved component-mode representation for structural dynamic analaysis. AIAA
Journal, 13(8), 995–1006.
Rubin, S. (1975). Improved component-mode representation for structural dynamic analysis. AIAA
Journal, 13(8), 995–1006.
Shyu, W. H., Ma, Z. D., & Hulbert, G. M. (1997). A new component mode synthesis method:
Quasi-static mode compensation. Finite elements in Analysis and Design, 24(4), 197–298.
Skelton, R. E. (1980). Cost decomposition of linear systems with application to model reduction.
International Journal of Control, 32(6), 1031–1055.
Skelton, R. E., & Yousuff, A. (1983). Component cost analysis of large scale systems. International
Journal of Control, 37(2), 285–304.
Stoev, J., & Möhring, J. (2010). Hybrid reduction of non-symmetric mimo systems. In P.S. et al.
(Ed.) ISMA. K.U. Leuven.
Su, T. J., & Craig, R. (1991). Model reduction and control of flexible structures using krylov vectors.
Journal of Guidance, Control, and Dynamics, 14(2), 260–267.
Vakilzadeh, M. K., Rahrovani, S., & Abrahamsson, T. J. S. (2014). 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). New York: Springer.
Vakilzadeh, M. K., Yaghoubi, V., McKelvey, T., Abrahamsson, T. J. & Ljung, L. (2015). Experi-
ment design for improved frequency domain subspace system identification of continuous-time
systems. IFAC-PapersOnLine, 886–891.
Varga, A. (1995). Enhanced modal approach for model reduction. Mathematical Modelling of
Systems, 1(2), 91–105. https://doi.org/10.1080/13873959508837010.
Voormeeren, S., Nortier, B., & Rixen, D. (2013). Error estimation and adaptive model reduction
applied to offshore wind turbine modeling. In IMAC-XXXI: International Modal Analysis Con-
ference, Error Estimation and Adaptive Model Reduction Applied to Offshore Wind Turbine
Modeling. Bethel, CT: Society for Experimental Mechanics.
Voormeeren, S. N., van der Valk, P. L. C., & Rixen, D. J. (2011). Generalized methodology for
assembly and reduction of component models for dynamic substructuring. AIAA Journal, 49,
1010–1020.
Chapter 4
Experimental Substructuring
4.1 Introduction
⎧
⎨ Zū = f̄ + ḡ
General form Bū = 0 (2.33)
⎩ T
L ḡ = 0
Z BT ū f̄
Dual Assembly = (2.45)
B 0 λ 0
The key consideration now is, what domain should we use to describe Z (i.e.,
modal domain, time domain, etc.) and what measurements can we obtain? A few
common options are reviewed briefly here:
© CISM International Centre for Mechanical Sciences 2020 75
M. S. Allen et al., Substructuring in Engineering Dynamics,
CISM International Centre for Mechanical Sciences 594,
https://doi.org/10.1007/978-3-030-25532-9_4
76 4 Experimental Substructuring
If we apply a force (i.e., with a shaker, impact hammer, etc.) or input at point j and
measure the response (i.e., with an accelerometer, laser vibrometer, etc.) or output at
point i, then the element Yi(s)
j (ω) can be measured at the excitation frequency, ω. In
practice, a force is applied over some range of frequencies, transient measurements
are acquired in the time domain, and FFT techniques are used to obtain the FRF over
a range of frequencies.
Typically, a set of measurements is obtained by including multiple response trans-
ducers so i = 1...No where No is the number of output points and multiple inputs
(or drive points) can also be used so that j = 1...Ni . Hence, one can eventually
obtain a matrix of FRFs, Y(s) (ω) at enough points so that an adequate model of the
substructure can be obtained and coupled to other substructures, as elaborated later.
Measurements may be used to create a model for one of the substructures (e.g., for
s = 1) and the others would presumably be modeled numerically.
Measuring Modal Parameters
In the physical domain, a system model can be described in terms of the mass,
damping, and stiffness matrices, but as was the case with dynamic stiffness, these
matrices cannot typically be measured.
where q ∈ R Nm with Nm the number of vibration modes that are extracted from
the measurements. Assuming that the modal masses have been mass normalized,
(s) (s)
(s)
1. In classical frequency-based substructuring, one requires the drive point and trans-
fer FRFs to be measured for every connection DOF. For this case, if one considers
one connection node per bolt, and three translations and three rotations per node,
then a matrix of 60 × 60 connection FRFs must be measured for each beam and
there are two beams. Also, in order to measure the rotations, we would need a rota-
tional accelerometer in three directions, mounted exactly at the connection point
and a hammer or shaker that applies a pure moment (i.e., no translational force)
as an input; hardware does not exist that can reliably meet these requirements.
Hence, one may decide, for logistical reasons, to measure only the translational
components of force and motion. However, this will eliminate 3/4 of the FRF
matrix, and in many cases one cannot expect that 3/4 of the FRF matrix is negli-
gible.
Fig. 4.1 Sample substructuring problem where two beams are joined by 10 bolts, represented by
vertical lines
78 4 Experimental Substructuring
a. How does one place a sensor or actuator at every DOF. Remember there is a
“hole” at the idealized connection point. Should the bolt be left in and apply
forces and mount sensors on that? How does one apply a perfect shear force,
or measure a shear acceleration right at the idealized point of connection?
b. If the two 60 × 60 FRF matrices are successfully measured, how does one
account for the compliance and damping of the interface? Traditional meth-
ods “weld” the connection DOF together with no interface compliance or
damping, but actual joints are known to have finite stiffness and often have
non-negligible damping.
c. Drive point FRFs are already ill-conditioned for the required inversion,
and measurement errors are known to make the inversion even more ill-
conditioned often producing spurious resonances.
d. This example has idealized connection to 10 nodes, but in reality, connec-
tions between parts are continuous and span the area of the lap joint.
Some of the difficulties enumerated above can be addressed by performing sub-
structuring in the modal domain, i.e., CMS substructuring.
2. Consider the requirements associated with classical CMS substructuring.
a. Here, there is not a requirement to apply a force to every connection DOF, but
only to excite all the modes active in the required frequency band. With this
approach, at least, the pesky requirement to apply moments is eliminated.
b. Again it is difficult to make the translation and rotation measurements and
accurately extract all the connection DOF mode shapes.
c. For a free-mode basis set, extremely accurate rigid body mode shapes are
required as well. (In fact, the rigid body mode shapes are often more impor-
tant than any of the elastic mode shapes.) These are often not easy to measure.
Often these analytical estimates or mass properties’ measurement provide
better estimates of these than low-frequency modal methods.
d. The extracted mode shapes eliminate the random noise on FRFs but do not
eliminate gage bias errors or bias errors introduced by the modal extraction.
e. In general, fixed base modes are difficult to obtain by testing simply because
not truly fixed base exists, so one extracts some sort of “free” set.
f. The free-mode shapes of substructures in the bandwidth of interest are often
not an adequate basis set to span the space of the connected motion. In the
analytical realm, free modes are augmented with residual flexibilities and
then they often form a suitable basis. Efforts have been made to measure
these residuals, but this once again requires force and moment inputs at
all of the connection DOF. Since the residuals represent the flexibility of
out of band modes, one might think to instead extract several modes above
the frequency bandwidth of interest, but this is typically difficult and often
many, many modes are needed to obtain convergence.
4.2 Why Is Experimental Substructuring So Difficult? 79
An example is shown in Fig. 4.2. A finite element model was used to represent
the red cylindrical substructure and measurements were used to obtain the free
modes of the flange with the beam attached (i.e., the plate-beam system in cyan
and green). The residual flexibilities were not measured, and the CMS method
was used to couple a free-mode model of the plate beam with the FEM of the
cylinder and then the modes of that model were used to reconstruct the drive
point FRF at the end of the beam. The substructuring prediction is shown in red,
with the test data measured from an actual assembly in blue. The plate attaches
to the cylinder by eight bolts and all rotations and translations are included in
the connection DOF. 100 modes of the cylinder were used and all the modes
of the plate and beam to 4000 Hz. The goal was to produce an accurate model
up to 2000 Hz. This example clearly shows that free modes are not adequate
to obtain a reasonable match for this drive point in the combined system. It
should be noted that when two components are coupled together loosely, i.e.,
using rubber bushings, then a free-mode model often can prove effective. The
discussion here focuses primarily on substructures that connect move rigidly.
Fig. 4.2 Substructuring prediction of response at point 1000y (axial direction) for the cylinder-
plate-beam system
In summary, the CMS approach that is traditionally used in finite element analysis
does not generally address the interface compliance, interface damping, or continuous
connection issues that are problematic in experiments. Techniques to address these
several problems in the physical DOF space will be addressed in this lecture note.
80 4 Experimental Substructuring
Let us consider again the dual form of an assembled system made of several compo-
nents (see (2.45), Sect. 2.3), dropping for simplicity the bar superscript introduced
earlier to indicate that the DOF is to be interpreted as complex amplitudes in the
frequency domain:
Z BT u f
= , (4.2)
B 0 λ 0
where
⎡ ⎤
Z(1) 0
⎢ .. ⎥
Z=⎣ . ⎦
Z(N )
sub
0
⎡ ⎤ ⎡
⎤ ⎡ ⎤
u(1) f (1) λ1
⎢ . ⎥ ⎢ . ⎥ ⎢ ⎥
u = ⎣ .. ⎦ f = ⎣ .. ⎦ λ = ⎣ ... ⎦
u(N ) f (N ) λn λ
sub sub
are the block diagonal matrices containing the impedances, DOF, and external forces
of the unassembled substructures, where B is the signed Boolean matrix of suit-
able dimension, which expresses the interface compatibility constraints (see (2.29),
(2.30)) and λ the associated Lagrange multipliers representing the internal forces
in the interface between substructures. An interpretation of the B and the Lagrange
multipliers is given Fig. 2.3.
The substructure displacements, u, in this dual form can be solved to obtain the
substructure displacements as functions of the forces, λ, using the dynamic equilib-
rium:
is the block matrix containing the substructure admittances. Equation (4.3) contains
the unassembled substructure dynamic equilibrium equations
82 4 Experimental Substructuring
T
u(s) = Y(s) (f (s) − B(s) λ), (4.5)
where the coupling is guaranteed by the Lagrange multipliers, unique between match-
ing DOF across the interface, representing the interface forces and the amplitude of
which is determined by the compatibility condition contained in the last line of (4.2).
Substituting (4.3) into the compatibility condition yields
BY(f − BT λ) = 0
or, substructure-wise,
⎛ ⎞
sub sub
N N
⎝ (s)
B Y B (s) (s) T ⎠λ = B(s) Y(s) f (s) . (4.7)
s s
Problem (4.6) (or its form (4.7)) is commonly referred to as the Dual Interface
Problem: it contains n λ compatibility equations for the unknown internal Lagrange
multipliers on the interface. It can be interpreted as follows:
The external forces f (s) are applied to the substructures, creating substructure
responses Y(s) f (s) and generating interface incompatibilities BYf (Fig. 4.3b). In
order to prevent those interface “gaps”, internal forces must exist on the interface
T
and have amplitudes such that the response YB(s) λ that they generate on each side
of the interface closes the gap originating from the external forces (Fig. 4.3c).
The dual interface problem for the amplitudes of the Lagrange multiplier is a
global problem since a Lagrange multiplier between two substructures will affect
the response of those substructures also on the interface with other substructures.
The operator
Fig. 4.3 Interpretation of the dual interface problem (4.7): a Substructured problem, b Effect of
local force in Ω (1) , c Global effect of λ to enforce compatiblity
4.3 Basics of Frequency-Based Substructuring (FBS) 83
⎛ sub ⎞
N
F I = (BYBT ) = ⎝ B(s) Y(s) B(s) ⎠
T
(4.8)
s
in the dual interface problem is often called the dually assembled interface flexibility.
The dual interface problem (4.48) can be used to compute the interface Lagrange
multipliers, namely,
λ = (BYBT )− BYf,
1
(4.9)
and substituting this result in the dynamic equation (4.3) for the substructure response
finally yields
u = Y f − BT (BYBT )− BYf = Y − YBT (BYBT )− BY f.
1 1
(4.10)
The different terms in this formula can easily be interpreted if one considers the way
it was found:
coupling
uncoupl.
we can write
Yg,dual = Y − YBT (BYBT )− BY .
1
(4.11)
This is the final formula for the dual assembly of substructure admittances.
Note: this admittance operates on the force vector f, which contains the external
forces defined on each uncoupled substructure, and returns the response, u, for all
DOF of the separated substructures. In other words, the dually assembled admittance
is defined for all DOF on each side of an interface. Obviously, by construction, Yg,dual
84 4 Experimental Substructuring
is identical for all corresponding DOFs on the interface and thus the information in
Yg,dual is redundant. This redundancy can easily be eliminated by keeping in Yg,dual
only the row and line corresponding to one of the redundant DOFs on the interface.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
u (1)
2 f 2(1) (1) (1)
Y22 Y23 0 0
⎢ (1) ⎥ ⎢ (1) ⎥ ⎢ (1) (1) ⎥
⎢u 3 ⎥ ⎢ f3 ⎥ ⎢Y32 Y33 0 0 ⎥
u=⎢ ⎥
⎢ (2) ⎥ f =⎢ ⎥
⎢ (2) ⎥
⎢
Y=⎢ ⎥
(2) (2) ⎥
⎣u 1 ⎦ ⎣ f1 ⎦ ⎣ 0 0 Y11 Y12 ⎦
u (2)
2 f 2(2) 0 (2) (2)
0 Y21 Y22
B = 0 1 −1 0 .
where
⎡ (1) (1) ⎤
Y22 Y23 0 0
⎢ (1) (1) ⎥
⎢Y32 Y33 0 0 ⎥
Yassemb. ⎢
=⎢ ⎥
(2) (2) ⎥
⎣ 0 0 Y11 Y12 ⎦
(2) (2)
0 0 Y21 Y22
⎡ (1) ⎤
Y23
⎢ (1) ⎥ −1
⎢ Y33 ⎥ (1)
−⎢ ⎥ (2)
⎢ (2) ⎥ Y33 + Y11
(1) (1)
Y32 (2)
Y33 −Y11 (2)
−Y12 .
⎣−Y11 ⎦
(2)
−Y21
4.3 Basics of Frequency-Based Substructuring (FBS) 85
(1) (2)
must exist, where Y33 + Y11 is the flexibility of the interface similar to the
flexibility of two springs in series.
• Since the interface force g is internal (actio-reactio):
⎡ ⎤
g2(1) ⎡ ⎤
⎢ (1) ⎥ 0
⎢g3 ⎥ ⎢ −λ ⎥
⎢ ⎥ = −BT λ = ⎢ ⎥
⎢ (2) ⎥ ⎣ λ ⎦
⎣g1 ⎦
0
g2(2)
So, the dually assembled response (4.12) can be interpreted as the combined
response to the external forces on the unassembled system and to the internal
interface forces necessary to keep the substructure together when excited by the
external forces:
uncoupl. coupling
⎡ ⎤ ⎡ ⎤
(1) (1) (1)
Y22 Y23 0 0 Y23
⎢ (1) (1) ⎥ ⎢ (1) ⎥ −1
⎢Y32 Y33 0 0 ⎥ ⎢ Y33 ⎥ (1)
and using the notation for the global primal variables as in (2.39) one can write
this also as
ug = Yg f g = Z−1
g f g.
a Differentsplittings of the external force on the interface will produce the same assembled
response, but will produce different amplitudes λ.
The LM-FBS discussed in the previous sections is based on dual assembly (also
sometimes called admittance-based substructuring) and was explained in frequency
domain. Dual assembly approaches can also, without further difficulty, be written in
the time domain. The dual assembly in the time domain (so in essence the inverse
Fourier transform of the LM-FBS) was proposed in Gordis (1995), and the general
formulation was later written in Rixen (2010), Rixen and van der Valk (2013). The
method is based on the impulse responses at the interface of the substructures and
4.3 Basics of Frequency-Based Substructuring (FBS) 87
was first developed as a simulation approach (also for multibody dynamics Géradin
and Rixen 2017). First attempts to use it as an experimental technique were only
partly successful Rixen (2010).
At this point, now that we have shown the basic algebraic techniques to find the
assembled admittance in a dual approach, one might wonder why the primal approach
(which is more commonly known by engineers and seemingly more straightforward)
is not the method of choice in experimental dynamics.
As explained in Sect. 2.3, primal and dual assembly are mathematically equiva-
lent. Since the properties of the components are obtained during testing as admit-
tances (forces applied as input, responses measured as output), the dual approach
seems more natural, although it would be not less natural to compute the component
impedances (inverting the measurements) and applying a primal assembly. The result
would be identical. So the real reason must be elsewhere.
The main reason to apply dual assembly in experimental dynamics comes from
the fact that measurements are never perfect, whereas numerical models are. (That is,
numerical models are at least consistent and physically realizable representations of
some dynamic system, even if they don’t capture the structure of interest as accurately
as might be desired.) Consistency will be further discussed in Sect. 4.9.2, but it is easy
to understand that, since measurement errors can be of any kind, there is no reason
why the measured admittance represents exactly a structural system. For instance, it
is very common that the eigenfrequencies or modal dampings as seen in the signals
of different outputs or as obtained from different inputs are not coinciding (e.g.,
additional mass effect of sensors or influence of shaker links). When using such
measurements of components, one typically observes so-called “spurious peaks” in
the FRFs of the assembled system (whether the assembly was performed in a primal
or in a dual manner, since the methods are mathematically equivalent). An example of
spurious peaks in the assembled FRF due to inconsistencies in the local admittances is
shown in Fig. 4.5. Spurious peaks were already considered in Imregun (1987) where
a smoothing procedure was proposed and a discussion of the appearance of spurious
peaks can be found for instance in Rixen (2008), Carne and Dohrmann (2006),
Nicgorski and Avitabile (2010). An analysis of error and uncertainty propagation in
the LM-FBS was presented in Voormeeren et al. (2010). It allows giving a confidence
interval to the assembled FRFs when the uncertainty of the substructure admittance
is known (for instance, from the measurements). This will, however, not be further
discussed here.
In dual assembly techniques, the compatibility condition is explicitly consid-
ered next to the substructure equilibrium equations, the interface equilibrium being
ensured by the very nature of the Lagrange multipliers. Therefore, as will be explained
in the next section (and further developed in other parts), it is possible to weaken
the interface compatibility and thereby mitigate the effect of the errors in the
measurements.
Mathematically, the effects of the measurement errors manifest themselves when
the dual interface flexibility (4.8), needed to compute the interface Lagrange multi-
pliers, is inverted in the dually assembled admittance (4.11):
⎛ sub ⎞
N
= Y − YBT F−I BY F I = (BYBT ) = ⎝ B(s) Y(s) B(s) ⎠ . (4.12)
1 T
Yg,dual
s
where we have assumed strong compatibility of the interface, meaning that we have
imposed all matching interface DOFs to be exactly equal (second set of equation
in (4.2)). If one, however, has a good idea of how the interface displacement will
look in the assembled system, one can weaken the compatibility requirement by only
requiring compatibility for the components of the interface displacements that are
important for the assembled solution.
Mathematically speaking, let us assume that we know a good representation space
VΓ r s for the displacements in the assembled system around an interface Γ r s between
two substructures s and r . The matrix VΓ r s contains, in its columns, a set of represen-
tation vectors that can well approximate the motions of the DOF around the assembled
interface. These vectors can be set up from engineering know-how (assuming, for
instance, a rigid behavior around the interface, i.e., the virtual point idea explained
later) or could be obtained from an approximate finite element model of the interface
region. In that subspace, the interface DOF of the substructures can be expressed as
u(s)
b V Γb β
rs
(s)
and u(r ) (r )
b V Γb β ,
rs (4.13)
where VΓbr s is the restriction to the interface of the common interface modes and
where β (r ) and β (s) can be seen as the corresponding displacement amplitudes on
each side of the interface. Here, we assume that the number of representation modes
in VΓ r s is smaller than the numbers of interface DOF.
The compatibility condition1
u(s) (r )
b − ub = 0 (4.14)
can be replaced by
VΓbr s β (s) − VΓbr s β (r ) = 0 (4.15)
1 Here, to simplify the presentation, we assume that the numbering on the DOF on each side of the
interface is identical. If this would not be the case, the constraint matrix B should be considered as
explained earlier.
90 4 Experimental Substructuring
or equivalently (if VΓbr s is full rank, i.e., if the vectors in the columns of VΓbr s are
linearly independent)
β (s) − β (r ) = 0, (4.16)
hence requesting that only the part of the substructure displacements living in the
subspace of VΓbr s be compatible. The remainder of the interface displacement, on
each side of the interface, that is not representable in VΓbr s is left free on the interface
and does not participate to the coupling of the substructures. This weak compati-
bility is the key to include only the important behavior in the coupling (usually well
captured by the measurements), leaving the remainder unassembled (since they may
be strongly affected or even dominated by measurement errors).
In order to write the weak compatibility as a condition on the DOF of the original
problem, the coordinates in the reduced space VΓbr s can be computed by observing
that
u(s)
b = V Γb β
rs
(s)
+ r(s) , (4.17)
where r is a remainder that will not be made compatible and that one can request to
remain orthogonal to the interface representation space:
W being any positive definite matrix defining the norm of the projection (usually
taken as the identity matrix). This condition leads to
since (VΓT r s WVΓbr s )−1 VΓT r s W can be seen as a (weighted) pseudo-inverse of VΓbr s .
b b
The pseudo-inverse can equivalently be computed using a (weighted) singular value
decomposition. Similarly, we can write the following for the other side of the inter-
face:
(r )
β (r ) = V+
Γ r s ub , (4.20)
b
The number of compatibility conditions in this form is equal to the number of repre-
sentation vector in VΓ r s , namely, less than the number of interface DOF, indicating
once more that the compatibility condition is weakened. The pseudo-inverse can
be seen as a (weighted) projection in the space of VΓ r s of the interface degrees of
freedom.
4.3 Basics of Frequency-Based Substructuring (FBS) 91
where T is a projection of the interface DOF of the substructures onto a common sub-
space (usually per local interface). The Lagrange multipliers now give the amplitude
of the generalized forces associated with each deformation in the interface represen-
tation space. For example, for a simple case where two structures are coupled using
a rigid interface with six DOFs (three translations and three rotations), the Lagrange
multipliers would give the net force in each direction and the net moment in each
direction on each interface.
Following the same procedure as earlier to find the dually assembled admittance,
one finds
T T
Yg,dual,weak = Y − YB̃ (B̃YB̃ )− B̃Y
1
B̃ = BT. (4.23)
This shows that the weak compatibility can also be interpreted as satisfying the
compatibility only when projected on a subspace, which in general can be expressed
by the weakening
Bu = 0 → VT WBu = 0.
This is strongly related to the so-called coarse space problem in dual assembly theory
(e.g., in the FETI parallel solution techniques).
Remark 2: The virtual point constraint
In order to illustrate the weakening idea, let us discuss two weak compatibility
conditions often used in practice. One possible choice for the representation of the
interface behavior consists in assuming that a part of the interface behaves rigidly (for
instance, a local bolted connection that can be considered as a point connection as
illustrated in Fig. 4.25). In that case, the representation space for the signal measured
around the connections can be taken as
VΓ r s = R Γ r s ,
92 4 Experimental Substructuring
where RΓ r s contains in its columns the six rigid body modes (three translations, three
rotations) around a virtual point on the interface. With those rigid modes, the DOF on
each side of the connection point where admittances were acquired can be described.
The weak coupling described by (4.21)–(4.23) then represents the assembly at
the interface of a virtual point having six DOFs, i.e., the coordinates β (s) , β (r ) that
are made equal on the interface represent the rigid body DOF of the interface. The
remainder of the interface behavior that cannot be described by the local rigid body
motion around the point will not be assembled, thereby avoiding the computation
of the high impedance related to deformation modes around the virtual point that is
usually badly captured in the measurement and that is often not necessary to represent
the interface behavior in the frequency range considered.
The virtual point method will be explained in more details in Sect. 4.6.
Remark 3: Interface extrapolation
Often, the interface response cannot be measured directly on the interface itself, for
instance, because sensors and actuators cannot be easily installed on the true interface
but rather in its neighborhood. This is true also for the case where the interface is
to behave rigidly (see the discussion about the virtual point in Remark 2) and where
the rigid behavior of the interface is measured away from the physical interface (see
for instance Fig. 4.25).
In that case, the interface modes VΓ r s can still be used to define generalized
interface DOF β (s) and write the interface compatibility as before in (4.16). To that
effect, taking the value of the modes VΓ r s at the measurement points m (that can thus
be different from the true interface location), one can write
β (s) − β (r ) = 0 β (s) = V+
Γmr s u
(s)
β (r ) = V+ (r )
Γmr s u , (4.25)
s r
where VΓmr ss and VΓmr sr are the values of the global mode in the vicinity of the interface
for the measured DOFs in r and in s.
The dually weakly assembled forms of the system dynamics and of the admittance
is then written like in (4.22) and (4.23), the difference being now that the substructure
matrices Y and Z refer no longer to interface DOFs but to other measurement DOFs
in the interface vicinity and that the projection matrix T contains pseudo-inverses
for VΓ r s evaluated at the measurement points.
This technique can be interpreted as using common interface modes VΓ r s to extrap-
olate measured dynamics to the interface while at the same time imposing a weak
compatibility.
admittance Y(s) of two components can be well represented by a modal fit. Assum-
ing small damping and well-separated eigenfrequencies for each substructure, the
modal fit can be used to write the admittances as (see, for instance, Géradin and
Rixen 2015)
−1
Y(s) (ω) = Φ (s) Ω (s) + 2iωΩ (s) ε(s) − ω2 I Φ (s) ,
2 T
(4.26)
where Φ (s) contains in its column the identified modes of substructure s (assumed
mass normalized), and where Ω (s) and ε (s) are diagonal matrices containing the
corresponding identified eigenfrequencies and modal damping ratios.
This modal fit has the advantage that it provides a fully consistent expression of
the admittance, removing inconsistencies and noise from the measurement.
The admittances of the substructures can now easily be obtained as
−T
−1
Z(s) (ω) = Φ (s) Ω (s) + 2iωΩ (s) ε (s) − ω2 I Φ (s)
2
(4.27)
so that the dually assembled form (4.2) (here discussed for the case of two substruc-
tures for simplicity) is written as
⎡ ⎤⎡ ⎤ ⎡ ⎤
.. .. .. ..
. .
⎢ 0 ⎥⎢ . ⎥ ⎢ . ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ (s) ⎥
⎢0 (s)
Z (ω) (s)T ⎥ ⎢u(s) ⎥
B ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎢f ⎥
⎢ ⎥⎢ . ⎥ = ⎢ . ⎥. (4.28)
⎢ ⎥⎢ . ⎥ ⎢ . ⎥
⎢ .. . ⎥⎢ . ⎥ ⎢ . ⎥
⎣ . .. ⎦ ⎣ ⎦ ⎣ ⎦
· · · B(s) · · · 0 λ 0
where η(s) are the modal coordinates in substructure s, and multiplying the local
dynamic equations in (4.28) by Φ (s) , the dual assembly now is written as
T
⎡ ⎤ ⎡.. ⎤⎡ ⎤
.. .. ..
. 0 . . .
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 (s)
Zm (ω) (s)T (s)T ⎥ ⎢η (s) ⎥
Φ B ⎥ ⎢ ⎥ ⎢Φ f ⎥ ⎢ (s)T (s) ⎥
⎢
⎢ ⎥⎢ . ⎥ = ⎢ .. ⎥, (4.29)
⎢ ⎥ ⎢ .. ⎥ ⎢ . ⎥
⎢ .. .. ⎥⎢ ⎥ ⎢ ⎥
⎣ . . ⎦ ⎣ ⎦ ⎣ ⎦
(s) (s) λ
··· B Φ ··· 0 0
94 4 Experimental Substructuring
where
(s)2 (s) (s)
Z(s)
m (ω) = Ω + 2iωΩ ε − ω 2
I .
This form can be interpreted as follows: the local dynamics of the substructures
are described in the modal space using the modal eigenfrequencies and dampings
identified from the measured FRFs. Since the compatibility still needs to be satisfied
by the physical DOF of the interfaces, the constraint matrices B(s) are now replaced
by B(s) Φ (s) , namely, the traces of the modes on the interface.2
The modes of the substructures are free interface modes and exhibit in general very
different displacements on the interface. Therefore, requiring strong compatibility
significantly limits the possible combination of modes allowed in the substructures.
Often, the number n sφ of modes Φ (s) considered for a substructure is smaller, then
the number of physical DOF u(s) b on its interface. In that case, the strong interface
compatibility in (4.29) will significantly limit the number of possible combination
of the local modes. In extreme cases where the number of interface compatibility
conditions n λ is higher than the number of substructures modes n sφ , the only
possible solution satisfying the strong constraint is η(s) = 0, leading to a full looking
of the solution space.
Even if the number of modes accounted for in the substructures is high, the
motion remaining possible for the assembled system after the strong compatibility
has been satisfied is often limited. Further, one should keep in mind that the local
modes obtained by the experimental identification are free interface modes and are in
general not well suited to represent the behavior of the substructure when assembled
to its number (i.e., when the interface is not free).3
2 It is interesting to note that this form of the dual assembly can deliver a model that is not only valid
indicating that the dual assembly of the substructure modal model yield the stiffness, damping,
and mass matrix of the assembled system for the substructure modal coordinates as DOF. These
matrices can be used, for instance, to directly write the dynamic equation in the time domain.
3 Later, in Sect. 4.4.3, we will discuss a technique to measure the substructure modes when the
interface is interacting with substructures in a way similar to what it will experience in the assembly.
This so-called transmission simulator technique will significantly enhance the basic modal-based
substructuring explained here.
4.3 Basics of Frequency-Based Substructuring (FBS) 95
From the discussion above, it is clear that it is desirable to relax the compatibility
on the interface in order to allow the free interface modes of the substructures to
properly represent the local dynamics. To that effect, the compatibility relaxation
explained in the previous section can be used as follows.
Using the technique outlined in the previous section, we can weaken the interface
compatibility by replacing the Boolean matrices B by B̃ = BT, where T contains
a projection operator on interface modes. For the case of two subdomains, calling
VΓb12 the representative modes in the vicinity of the interface, the weak compatibility
would then be written as
V+ Φ (1) η(1) − V+
Γ 12 b
Φ (2) η(2) = 0.
Γ 12 b
(4.30)
b b
In case the measurement points are not directly on the interface but in its vicinity
(see Remark 3 in the previous section), the weak compatibility conditions would be
written as
V+ (1) (1)
Γ 12 Φ m s η − V+ (2) (2)
Γ 12 Φ m r η = 0. (4.31)
ms mr
The modal-based substructuring will be further explained and used in Sect. 4.4.3.
assuming that the main contribution of the bushing in the assembly is determined
by its static stiffness and its damping, the small mass of the bushing and its internal
dynamics being neglected.
In the primal assembly setting, the dynamic equation of an assembly (see (2.39))
of two components connected by a bushing can easily be expressed as (see Fig. 4.6)
⎡ ⎤⎡ ⎤ ⎡ ⎤
Zii(1) (1)
Zib 0 0 ui(1) f i(1)
⎢ (1) (1) ⎥⎢ ⎥ ⎢ ⎥
⎢Zbi Zbb + Z(bush) Z(bush) 0 ⎥ ⎢ ub1 ⎥ ⎢ f b1 ⎥
⎢
Zg = ⎢ b1 ,b1 b1 ,b2 ⎥⎢ ⎥ = ⎢ ⎥, (4.33)
(bush) (2) ⎥ ⎢ 2 ⎥ ⎢ ⎥
⎣ 0 Z(bush)
b ,b
2 1 Z(2)
bb + Z b ,b
2 2 Zbi ⎦ ⎣ ub ⎦ ⎣ f b2 ⎦
0 0 Z(2) Z(2) ui(2) f i(2)
bi ii
96 4 Experimental Substructuring
kbush
i b b1 b2 b i
bushing
(1) (2)
cbush
where we assumed the numbering of the degrees of freedom of the bushing to fit the
numbering of the interfaces of the substructures it connects, and where the subscript
b and i denote boundary DOF on the interface of the bushing and internal DOF,
respectively.
In the experimental substructuring framework, considering a bushing element
can be challenging since, before being assembled in the system, it is a free floating
components and its impedance matrix is singular due to the fact that a rigid motion
(translation or rotation) can be given to the bushing without force (its inertia being
neglected):
Z(bush) R(bush) = 0, (4.34)
where R(bush) is a vector containing in its column a set of independent rigid body
modes of the bushing (usually 6).
which is clearly singular. There is one rigid body mode in this case
(bush) 1
R = .
1
Hence, the inverse of Z(bush) does not exist, meaning that no solution can be
computed for a set of forces applied to the bushing unless those forces are self-
equilibrated (zero resulting force and moment).4
There are several ways to introduce a bushing in the general framework of LM-
FBS, as will be shortly discussed next. First, we will consider the bushing as a
connecting element between the interfaces of two substructures that introduces an
interface incompatibility, thereby slightly modifying the dual interface problem. In
4 For a more detailed discussion on singular matrices in structural dynamics, see for instance Géradin
Bu = δ = 0, (4.35)
where δ is the interface gap allowed by the bushing. This gap introduces in the bushing
internal forces λ proportional to the dynamic stiffness of the bushing. Calling Υ the
matrix describing the dynamic flexibility of the bushing between its interface DOF,
one can write
Bu = δ = Υ λ. (4.36)
−1
Note that Υ is not equal to Z(bush) since it expresses the flexibility related to interface
gaps (hence in the space of λ) and not directly the flexibility between interface DOFs.
1
B = 0 1 −1 0 δ = u (1) (2)
b − ub δ= λ,
k bush − iωcbush
kbush
i b b i
bushing
(1) λ λ (2)
cbush
Fig. 4.7 Bushings as compatibility relaxation
where k bush − iωcbush is the dynamic stiffness of the rubber between the inter-
faces. In general, this bushing stiffness can be different for different gaps defined
across an interface (for instance, in different translational or rotational direc-
tions).
The internal forces λ in the bushing are equal in amplitude and opposite on either
side of the bushing (since the bushing has no inertia): they correspond to the Lagrange
multipliers defined in the dual assembly of structures, and the dual assembly form
98 4 Experimental Substructuring
(4.2) in the presence of a bushing is now written as (here as example for a system
with two substructures):
⎡ (1) T⎤⎡ ⎤ ⎡ (1) ⎤
Z 0 B(1) u(1) f
⎣ 0 Z(2) B(2)T ⎦ ⎣u(2) ⎦ = ⎣f (2) ⎦ . (4.37)
B(1) B(2) −Υ λ 0
So the difference compared to the usual dual assembly form consists in the fact that
the lower block diagonal term in the dual system matrix is no longer null but includes
the bushing stiffness Υ . From a mathematical point of view, Υ can also be seen as
a perturbation of the compatibility constraints.5
The dual assembly (4.37) can directly be written for the general case of N sub
substructures even if bushings are present only at some interfaces: for the interfaces
where no bushings are present, the corresponding entries in Υ are set to zero (i.e.,
no interface flexibility). In block matrix form, one thus generalizes (4.2) to
Z BT u f
= . (4.38)
B −Υ λ 0
Using the first part of (4.38) to eliminate the substructure DOF (identical to (4.3))
u = Y(f − BT λ) (4.3)
and substituting in the second part of (4.38), one finds that dual interface problem
or, substructure-wise,
⎛ sub ⎞
N
N sub
⎝ B(s) Y(s) B(s) + Υ ⎠ λ =
T
B(s) Y(s) f (s) . (4.40)
s s
5 One can transform the form (4.37) in a primal assembled problem by eliminating λ to find
u(1)
λ = Υ −1 B(1) B(2) (2)
u
! (1)T " (1)
Z(1) 0 B −1
(1) (2)
u(1) f
+ Υ B B = (2) .
0 Z(2) B(2)
T
u(2) f
In this form, the interface flexibility can be seen as an augmentation method for the compatibility
constraints. Comparing with (4.33), one observes that the bushing dynamic stiffness matrix has a
specific form induced by the connectivity between the interfaces of the bushing and the bushing
intrinsic stiffness.
4.3 Basics of Frequency-Based Substructuring (FBS) 99
Bushings as Substructures
The approach explained above considers the bushing as an interface where the com-
patibility is relaxed. Comparing the dually assembled admittance obtained by the
LM-FBS with and without bushing, we observe that the presence of bushing flex-
ibilities in (4.41) modifies the way the dual assembly is organized. If one would
like to treat the presence of bushings without changing the LM-FBS formulation,
one should apply the standard form (4.11) and consider bushings as substructures
of their own. Note that in that case the connectivity matrices are different from the
previous approach. For instance, if one considers the bushing of the simple case of
Fig. 4.6 as a third substructure,
⎡ ⎤
u(1)
0 1 −1 0 0 0
B = B(1) B(bush) B(2) = u = ⎣u(bush) ⎦ .
0 0 0 1 −1 0
u(2)
In order to apply the classical LM-FBS approach and write the assembled admit-
tance as (4.11), one needs to find a manner to write the admittance of the bushing
substructure, which is not easy since its dynamic impedance is usually singular (see
(4.34)).
A straightforward manner to build an admittance matrix for the bushing consists
in regularizing its impedance by adding virtual lumped masses to its ends, namely,
−1
Y(bush) Z(bush) − ω2 M(bush) (4.42)
In order for this modification not to affect (too much) the dynamic assembly of the
assembly, three strategies can be considered:
• Choose M(bush) as a diagonal matrix with small entries. For instance, one can
choose the mass perturbation such that the local nonzero eigenfrequencies of the
bushing are 10 times higher than the highest frequency of interest for the assembly.
• Choose any M(bush) and subtract its effect by assembling an additional (vir-
tual) substructure with the corresponding negative mass, namely, with admittance
ω2 M(bush) (see also the concept of decoupling in Sect. 4.4.1).
• Choose a diagonal M(bush) and remove the corresponding lumped masses from the
neighboring substructure admittances (for instance by decoupling, see Sect. 4.4.1).
Obviously, the first strategy is the easiest one and can be readily applied in practice.
100 4 Experimental Substructuring
has a solution only if the applied forces are self-equilibrated or, in other words, if
the applied forces have zero resulting forces and moments. Mathematically (see for
instance Géradin and Rixen 2015), this is expressed by the fact that the applied forces
must not produce any work on the rigid body modes:
R(bush) f (bush) + B(bush) λ = 0.
T T
(4.43)
Under this condition, there exist a family of solution that can be written as
+
u(bush) = Y(bush) f (bush) − B(bush) λ + R(bush) α (bush) Y(bush) = Z(bush) ,
T
(4.44)
+
where Z(bush) is a pseudo-inverse of Z(bush) that can be computed either by singular
value decomposition or, more efficiently, by a incomplete factorization (see, for
instance, Géradin and Rixen (2015) or the simple example below). The amplitudes
α (bush) of the contribution of the rigid body mode to the solution are additional
unknowns that will be determined, thanks to the self-equilibrium condition (4.43).
In other words, the overall position of the bushing needs to be such that the generated
interface forces are in global static equilibrium for the bushing.
For the simple example of Fig. 4.6, we can compute a pseudo-inverse in the
form of
+ bush
+ z bush −z bush (k − iωcbush )−1 0
Y(bush) = Z(bush) = =
−z bush z bush 0 0
The general solution can be understood as the solution obtained when the bush-
ing is fixed at the right end and when an arbitrary rigid body motion is added.
Note that fixing the bushing at one end does not introduce any additional force
to the system (since all the external forces are self-equilibrated) but merely
determines one specific solution.
The local solutions of the substructure dynamics are thus either in the general
form (4.44) for a bushing or in the regular form (4.5) for the other substructures.
This can be written in a block form similar to (4.3) as
if we define the block matrix of the rigid body modes for the bushings and their
corresponding amplitudes
⎡ ⎤
0 ··· 0
⎢ .. ⎥
⎢ . ⎥ ⎡ (bush1) ⎤
⎢ ⎥
⎢ (bush1) .. ⎥ α
⎢R 0 . ⎥ ⎢ (bush2) ⎥
R=⎢ ⎥ α=⎣ α ⎦. (4.46)
⎢ .. . ⎥ ..
⎢ . (bush2) . ⎥
⎢ R . ⎥ .
⎢ . .. .. ⎥
⎣ .. . . ⎦
0 ··· 0
The dually assembled problem can now be found as before: if we consider for the
bushing the local solution (4.45) and substitute in the compatibility condition, one
finds
Bu = B Y(f − BT λ) + Rα = 0. (4.47)
Rearranging this equation together with the self-equilibrium condition (4.43) leads
to the dual interface problem (which is an extension of the standard form (4.48)):
BYBT BR λ BYf
= . (4.48)
R T BT 0 α RT f
This extended dual problem can be solved for λ and α. Substituting the solution
into the governing equation then returns the dually assembled admittance. It can
be written as an extended form of (4.11) but will not be shown in detail here (see
Mahmoudi et al. 2019 for further details and examples).
In the literature, two types of disassembling problems are treated. In the first type of
problem, one assumes that the overall system is made of several substructures, and
the question to be answered is to find the dynamical properties of one substructure,
102 4 Experimental Substructuring
the dynamics of the rest of the structure being known. This problem will be referred
to as Substructure Decoupling.
A similar, yet different, problem arises when considering structures assembled
through bushings, i.e., structural elements that are characterized only by stiffness
and damping and having negligible mass (see Sect. 4.3.6). In this case, one knows a
priori that the forces on each side of the bushing are in equilibrium (since the internal
dynamics of the bushing are negligible) and therefore it is possible (as will be shown
later) to extract the dynamic properties of the separated bushing and substructures,
without having any pre-knowledge of the dynamics of any component. This problem
is commonly referred to as inverse substructuring in the literature.
We will first outline the basics of substructure decoupling, and then those of
inverse substructuring. In the last part, substructure decoupling will be applied in the
modal domain and the technique of transmission simulator will be outlined.
The problem of substructure decoupling has been tackled in the recent years by
several others (e.g., Ind and Ewins 2003, Sjövall 2007, Groult 2008, D’Ambrogio
2009, Voormeeren and Rixen 2009, Cloutier and Avitabile 2011, D’Ambrogio and
Fregalent 2014, D’Ambrogio and Fregolent 2015, Höller and Gibbs 2015) including
methods for simple nonlinear systems Kalaycıoğlu and Özgüven (2018), Tol and
Özgüven (2015). The presentation here is based on Voormeeren and Rixen (2012).
Let us consider a system made of two substructures called A and B in this section.
The coupling problem as discussed in the previous section consists of building the
dynamics of the assembly AB using the dynamics of the components, Fig. 4.8. The
substructure decoupling problem consists in finding the dynamic properties of sub-
structure B when the dynamics of the assembly AB and of component A is known,
Fig. 4.8.
This problem is often encountered in practice, for instance, when a component
cannot be disassembled to measure its FRF (an early application of the idea was used
for determining the dynamics of microsystems attached to a support in Epp et al.
2004).
The problem of decoupling can be solved by considering the interpretation
described in Fig. 4.9: if the part B in the assembly AB is excited by a force f i(B)
in B, the response at a DOF ui(B) in B when B would be disconnected from A can be
obtained if, in addition to the external force f i(B) , one applies to the assembly AB the
opposite of the internal force experienced by B in the assembly AB. This interface
force can be obtained by imposing on the interface of A the same displacements as
the one experienced by the assembly and by calculating the corresponding interface
forces.
Mathematically speaking, this means that the response of the uncoupled substruc-
ture B obeying
4.4 Substructure Decoupling 103
ZB uB = f B (4.49)
or ⎡ AB ⎤ ⎡ AB ⎤ ⎡ ⎤ ⎡ ⎤
Zi A ,i A ZiAB
A ,b
0 ui A 0 0
⎢ AB AB ⎥ ⎢ AB ⎥ ⎢ AB B⎥ ⎢ ⎥
⎣ Zb,i A Zb,b Zb,i B ⎦ ⎣ub ⎦ = ⎣f b = f b ⎦ + ⎣λ⎦ ,
AB
0 ZiAB B ,b
ZiAB
B ,i B
uiAB
B
f iAB
B
= f iB 0
104 4 Experimental Substructuring
where the subscripts i A and i B denote the internal DOFs in the A and B part of AB,
and where we assumed a special ordering of the DOFs.
The interface forces λ that one has to apply as external forces to AB must annihilate
the effect of the coupling on B and thus must be minus the force that AB receives
through its interface from A. The opposite of the interface force on B is the interface
force on A (actio-reactio): it can be computed as the force that A experiences when
ubAB is imposed on it, namely (see Fig. 4.8),
A
ui 0
λ such that Z A
= (4.51)
ubA λ
ubA = ubAB . (4.52)
Equations (4.4.1), (4.51), and (4.52) describing the dynamic of the uncoupled
substructure B can be summarized in the form
⎡ ⎡ ⎤ ⎤ ⎡ AB ⎤ ⎡ ⎤
0 ui A 0
⎢ Z AB ⎢ ⎥ ⎥ ⎢u AB ⎥ ⎢f B ⎥
⎢ 0 ⎣−I⎦ ⎥ ⎢ b ⎥ ⎢ b ⎥
⎢ ⎥ ⎢ AB ⎥ ⎢ B ⎥
⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ui B ⎥ ⎢f i ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥. (4.53)
⎢ 0 ⎥ ⎢ A⎥ ⎢ ⎥
⎢ ⎥ ⎢ ui ⎥ ⎢ 0 ⎥
⎢ 0 Z −
A
⎥⎢ A ⎥ ⎢ ⎥
⎣ I ⎦ ⎣ ub ⎦ ⎣ 0 ⎦
−I 0 0 I 0 0 λ 0
This relation shows that the dynamics of the isolated substructure B can be obtained
by “assembling” AB and A, enforcing the compatibility on the interface, but applying
the interface force in the opposite direction on AB. In case the numbering on the
interface of AB and of B are not corresponding, the generalized form of the problem
is written as ⎡ AB ⎤ ⎡ AB ⎤ ⎡ ⎤
Z 0 B AB u 0
⎣ 0 Z A −B A ⎦ ⎣ u A ⎦ = ⎣f B ⎦ , (4.54)
B AB B A 0 λ 0
where B AB and B A are the compatibility constraint matrices as used for the assembly.
This relation can be equivalently written as
⎡ ⎤ ⎡ AB ⎤ ⎡ ⎤
Z AB 0 B AB u 0
⎣ 0 −Z A B A ⎦ ⎣ u A ⎦ = ⎣f B ⎦ , (4.55)
B AB B A 0 λ 0
showing that the decoupling problem is equivalent to a coupling problem with A hav-
ing a negative impedance. Hence, one can apply the formula (4.11) of the frequency-
based substructuring to find the admittance of the uncoupled system if one considers
substructure A as having a negative impedance.
4.4 Substructure Decoupling 105
Assuming again that the numbering of the interface DOFs in AB and A are
matching, one can easily work out the impedance of B by solving (4.53), resulting
in Voormeeren and Rixen (2012)
B B
AB AB
AB
(B)
Yb,b Yb,i Yb,b Yb,i Yb,b AB A −1
AB AB
B
Yi,b B
Yi,i YiAB
B ,b
YiAB
b ,i B
YiAB
B ,b
(4.56)
Remark 1: Post-processing of decoupled dynamics
Note that, due to the errors inherent to the fact that the method is based on measure-
ment data, the FRFs of the uncoupled system B are usually not fully accurate and
might exhibit nonphysical properties such as non-passivity or negative mass. Several
authors have proposed different cures to improve the FRFs obtained by decoupling
Carne and Dohrmann (2006), Mayes and Ross (2012). Other techniques to enhance
the consistency of the FRFs in state space are discussed in Sect. 4.9.2.
Remark 2: Decoupling with overdetermined compatibilities
The decoupling method formalized in (4.55), exactly like for substructure coupling,
is very sensitive to small errors in the interface admittances of the components and,
therefore, it is desirable to weaken the interface compatibility using one of the tech-
niques described in Sect. 4.3.4.
However, for decoupling, there is an additional manner to relax the strong com-
patibility (4.52) required between the assembly AB and the substructure A to be
removed. Indeed, instead of requiring that λ applied to AB should be such that only
the interface DOFs of A and AB are equal (see (4.52)), one could require that all
DOFs measured in A and those measured in the part A of AB should be equal when
they experience the same λ on the interface.
A AB
ub u
= bAB . (4.57)
uiA ui B
Requiring that also the DOFs inside A must be identical in A and AB is, in theory,
unnecessary since if all interface DOFs in AB and A are identical, the internal ones
will automatically be identical. However, in practice, measurements are not perfect
and thus the data used for Y A might not be fully consistent with the information
about A contained in AB. Hence, requiring also explicitly the compatibility inside
A helps mitigating inaccuracies when computing the interface forces λ.
With the extended compatibility requirement (4.57), the decoupling problem
(4.58) is written as
106 4 Experimental Substructuring
⎡ ⎤⎤ ⎡ ⎡ ⎤
0 ⎡ AB ⎤ 0
⎣−I⎦ ⎥ iABu
⎢ Z AB 0 A ⎢f bB ⎥
⎢ ⎥ ⎢u b ⎥ ⎢ B ⎥
⎢ 0 ⎥ ⎢ AB ⎥ ⎢f i ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢u i B ⎥ ⎢ ⎥
⎢ 0 ⎥⎢ A ⎥ = ⎢ 0 ⎥. (4.58)
⎢ 0 Z −
A
⎥⎢ u ⎥ ⎢ ⎥
⎢ I ⎥ ⎢ iA ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎣ ub ⎦ ⎢ ⎥
⎣ −I 0 0 I 0 ⎦ ⎣0⎦
0 λ
0 −I 0 0 I 0
Note that, although compatibility is required for all DOFs of A, only interface forces
λ are considered. Therefore, the interface problem will be overdetermined and will
be solved using a pseudo-inverse, meaning that one computes the interface forces
that guarantees the best compatibility for all DOFs on A in a least square sense.
The admittance of B, given by (4.56) for the basic decoupling form, now becomes
Voormeeren and Rixen (2012)
B B
AB AB Y AB !Y AB A "+ Y AB Y AB
Yb,b Yb,i Yb,b Yb,i i A ,b Yi,b i A ,b i A ,i B
Y(B) =
b,b
= B
− − .
B YB
Yi,b i,i YiAB,b YiAB,i YiAB,b AB
Yb,b A
Yb,b AB Y AB
Yb,b b,i B
B b B B
(4.59)
In that way, inaccuracies in Y AB and Y A will be mitigated. This idea was already
proposed in Sjövall (2006), although in a different form.
Here, only one variant of the different possibilities to impose compatibility and
equilibrium for a decoupling problem was outlined. Other approaches modifying the
basic decoupling form (4.55) are discussed in Voormeeren and Rixen (2012), where
also application on an analytical example and on a hardware example can be found.
Remark 3: Transmission simulator and interface substitute
The substructure decoupling technique is very useful also when applying the FBS
methods to assemble measured components. Indeed, the results obtained from the
FBS techniques are significantly improved when the components are not measured
with free boundary conditions on the interface, but when the measurement is per-
formed with a so-called transmission simulator introduced in Allen et al. (2010);
Mayes and Arviso (2010), Mayes and Ross (2012) for the modal domain coupling (see
Sect. 4.4.3). The methodology is also sometimes called interface substitute van der
Seijs (2013) when used in the physical domain. The idea is illustrated in Fig. 4.10. The
coupling substitute must be removed from the assembled system using a decoupling
method as discussed in this section, assuming that the substitute is a well-defined
system for which a very good (numerical or measured) model exists.
Such a procedure seems useless since in theory adding a coupling substitute to
obtain the component dynamics, then removing it in the assembly does not affect
the final result. In practice, however, this can make a very significant difference and
dramatically improve the accuracy of the assembled system obtained by FBS, the
two main reasons being the following:
• the dynamics of the component measured with a substitute on its interface exhibits,
during the component measurement, dynamics that are similar to the dynamics it
4.4 Substructure Decoupling 107
Fig. 4.10 The idea of the transmission simulator (coupling substitute) van der Seijs (2013)
will exhibit in the assembly later. Therefore, when inverting the interface flexibility
to compute the assembled dynamics (see 4.11), the local admittance has much
better accuracy for the problem to be solved.
• the model of the component, when the substitute has been decoupled, contains
in practice not only the dynamics of the isolated component but includes also
the dynamics of the interface as it was present in the assembly with the substitute.
Hence, this interface dynamics (stiffness and damping) will also be included in the
model of the final assembly. This is essential, for instance, for assemblies where
a significant part of the damping comes from joints.
In practice, the substitute should be designed such that it exercises the interface of
the component in a manner similar to what the component will experience in the
real assembly, i.e., the substitute should have similar dynamics as the components
connected in the final assembly and it should be attached through the same physical
connections (e.g., bolted, riveted..).
Meggitt et al. (2015), Meggitt et al. (2018), Meggitt and Moorhouse (2018). We
shortly outline the method based on the approach of those latter publications.
Referring to the schematic representation of Fig. 4.6, let us consider the primal
form (4.33) of the dynamic problem where two substructures (called 1 and 2) are
connected by a bushing, recalled here for convenience:
⎡ (1) (1) ⎤⎡ ⎤ ⎡ ⎤
Zii Zib 0 0 (1) (1)
⎢ (1) (1) ⎥ ui fi
⎢Z Z + Z(bush) (bush) ⎢ ⎥ ⎢ ⎥
⎢ bi bb Z 1 2 0 ⎥ ⎥⎢ ub 1 ⎥ ⎢ ⎥
b1 ,b1 b ,b ⎥⎢ ⎥ ⎢ f b1 ⎥
Z(1+bush+2) u(1+bush+2) = ⎢
⎢ ⎢ ⎥ = ⎢ ⎥.
(bush) (2) ⎥ ⎢ u 2 ⎥
Zbb + Z 2 2 Zbi ⎥ ⎣ b ⎦ ⎣ f b2 ⎥ ⎢
(bush) (2)
⎢ 0 Z 2 1 ⎦
⎣ b ,b b ,b ⎦
(2) (2)
(2) (2) ui fi
0 0 Zbi Zii
(4.60)
If a substructure decoupling approach as explained in Sect. 4.4.2 would be used,
one would consider the bushings as a third component, measure Y(1+bush+2) and the
admittances Y(1) and Y(2) . Then one would decouple the admittances of substructure
1 and 2 by assembling to Y(1+bush+3) the negative admittances −Y(1) and −Y(2) .
In the inverse substructuring, only the measurement of the admittance Y(1+bush+2)
of the assembly is used and inverted to obtain
−1
Z(1+bush+2) = Y(1+bush+2) . (4.61)
Inverse substructuring then uses the fact that the off-diagonal terms in the measured
impedance (4.61) (comparing with the theoretical form (4.60)) are properties of the
bushing alone, namely,
Z(1+bush+2)
b1 ,b2 = Z(bush)
b1 ,b2 . (4.62)
Let us assume now that the DOFs at the two sides of the bushing are connected
with the topology shown in Fig. 4.11, i.e., each DOF on one side of the interface is
only coupled to one DOF on the other side and there are no cross couplings with any
other DOF. Additionally, it is assumed that the joint element (i.e., the rubber) has
negligible mass. If these assumptions hold, then the stiffness matrix of the rubber
isolator has the following properties:
where Z(bush)
b1 ,b1 is actually a diagonal matrix. Thus, by simply inverting the measured
FRF matrix of the assembly, the whole rubber isolator can be identified from just tak-
ing the off-diagonal blocks of the matrix (provided that the above-stated assumptions
hold). Hence, the name “inverse substructuring”. In this case, it is straightforward
to derive the dynamic stiffnesses of the two separate substructures 1 and 2 from the
measurements:
⎡ (1) (1) ⎤ ⎡ ⎤
Zii Zib 0 0 0 0 0 0
⎢ (1) (1) ⎥ ⎢ (bush) (bush) ⎥
⎢Z Z ⎥
⎢ bi bb 0 0 ⎥ (1+bush+2) −1 ⎢ 0 Zb1 ,b1 Zb1 ,b2 0⎥
⎢ ⎥= Y −⎢⎢ (bush) (bush) ⎥
⎥ . (4.64)
⎢ 0 0 Z Z ⎥ (2) (2)
⎣ 0 Z Z 0 ⎦
⎣ bb bi ⎦ b2 ,b1 b2 ,b2
(2) (2)
0 0 Zbi Zii 0 0 0 0
This comes with the great practical advantage that it is not necessary to know anything
about the dynamic properties of the two structures A and B which are joint by
the isolator. Thus, one is able to identify the dynamic properties of all involved
substructures 1, 2 and the bushing separately from only one set of measurements
performed on the assembly.
1
q1 , m1 q1 , m1
Bush
q2 , m2 q2 , m2
2
Fig. 4.12 Test and identification of a rubber bushing Haeussler et al. (2018)
The idea is to attach two well-defined and modeled rigid crosses on each side
of the rubber and consider them as substructures 1 and 2 connected by the
bushing. The crosses can be considered as very rigid for the frequency range
considered and, since the masses and inertias of the crosses are accurately
known (from a CAD model), their admittances are accurately known so that
the substructure decoupling techniques can be applied (using the rigid body
110 4 Experimental Substructuring
behavior of the crosses to define virtual points and assemble the model of the
system, see Sect. 4.6). The results obtained from the substructure decoupling
method for the dynamic properties of the bushing are then compared to the ones
obtained by applying the inverse substructuring technique.
The experiment is further described in Fig. 4.12 which shows (a) the actual
measurement setup of the assembly: rubber isolator between upper and lower
fixture hung up with rubber bands, (b) the schematics of the free-free measure-
ment of bushing between fixtures 1 and 2, and (c) the substructure of the rubber
isolator which is to be determined. The DOFs q1 and q2 represent the rigid
body DOF on the top and bottom of the rubber isolator, respectively, m1 and
m2 , containing the translational and rotational resulting forces on each side.
The results of the substructure decoupling and inverse decoupling are illustrated
in Fig. 4.13. It shows the impedance values obtained for the axial DOF of the
rubber. One clearly sees that both methods give very similar results up to 100 Hz,
but differ significantly for higher frequencies. In fact, the substructure decou-
pling yields the dynamic properties of the rubber including the bolts attached to
its ends (see Fig. 4.12c), whereas the inverse substructuring approach assumes
that the entire bushing is massless (Fig. 4.13).
Dynamic Stiffness [N/m]
I
Decoupling ZZ(bush)
b111,b1
I
Decoupling ZZ(bush)
b222,b2
I
Inv. Substr. ZZ(bush)
b21
1 ,b 1
(bush)
Inv. Substr. ZZbI12
2 ,b2
0
−90
−180
101 102 103
Frequency [Hz]
Fig. 4.13 Example of results from testing a rubber bushing Haeussler et al. (2018)
4.4 Substructure Decoupling 111
(The material in this section was taken from Paper 161 at the 2012 IMAC.)
The modal constraint for fixture and subsystem (MCFS) method was introduced at
IMAC in 2007 for the component mode synthesis approach Allen and Mayes (2007)
and the frequency-based substructuring approach Mayes and Stasiunas (2007). It uses
a fixture in the experimental dynamic substructure called the transmission simulator,
so is also called the transmission simulator method. An analytical model of the
transmission simulator is generally required with this method (or a very accurate
modal model).
Physically, the transmission simulator is mounted to the experimental substructure
with exactly the same joint geometry and material as that to which the experimental
substructure will ultimately be connected. Because of this, the resulting experimental
substructure inherently includes the linearized stiffness and damping in the joint,
which classical methods neglect, to their peril. The fixture is also designed to provide
accessible locations to mount response sensors and to apply input forces. Often
the actual connection points are poor locations to mount sensors. For example, the
connection point may be at a bolt or a threaded screw interface. Special features can
also be included to provide good driving point measurements, which are extremely
important to obtain accurate modal mass for scaling the mode shapes.
The analytical model of the transmission simulator is used in multiple ways. It is
generally mounted with an assumed welded connection to the analytical substructure
(often a finite element model) to which the experimental substructure is to be con-
nected. By constraining the transmission simulator on the experimental substructure
to have the same motion as the transmission simulator on the analytical substructure,
the systems are joined. Then the mass and stiffness of the analytical and experimental
transmission simulators are subtracted from the assembled system. The analytical
model of the transmission simulator can also be thought of as an aid to interpolate
from the measurement sensor locations back to the actual connection degrees of
freedom (DOFs). A truncated set of the mode shapes of the transmission simulator is
used that spans the frequency bandwidth of interest. The sensor set can be chosen to
provide a set of sensor locations for which all chosen transmission shapes are linearly
independent. These sensors can all be translational—no rotations are required. The
rotations at the connection points are inherently carried out in the modal coordinates
of the transmission simulator. The connections can actually be continuous, not just
discrete, as long as the transmission simulator mode shape set spans the space of the
connection motion reasonably well. In addition, the transmission of simulator mass
stresses the joint. This stress across the joint provides a much better Ritz vector shape
basis than simple free modes where there is no stress at the joint. It provides enough
improvement in the basis vectors that residuals, which are difficult to measure, do
not need to be added to the basis set.
Although this method requires fabrication of a fixture and generation of its asso-
ciated analytical model, it provides extensive benefits for the investment. One does
112 4 Experimental Substructuring
not have to measure exactly at the connection points. One does not have to measure
rotations because they are inherently included in the analytical model modal coor-
dinates (so they are not just being neglected). The stiffness and damping of the joint
are inherently included. No residual measurements are required.
One problem that can result in this method is that the mass matrix can be indefinite
when the transmission simulator is subtracted. However, methods to correct the mass
matrix have been developed recently Mayes et al. (2012). This has cleared the way
so that in practice, refinements can be made without theoretical roadblocks.
An example from Mayes et al. (2012) shows conceptually how the method was
implemented for one actual hardware case in Fig. 4.14. The transmission simulator
hardware, which is a ring with tabs, is mounted to the experimental substructure with
eight bolts, just as it will be attached in the real system to the cylinder substructure,
so it contains the joint. An analytical model of the transmission simulator is welded
into the flange of the cylinder analytical model. The transmission simulator and the
cylinder flange actually occupy the same space, which can be done with an analytical
model. Then transmission simulators for each substructure are forced to have the same
motion, which connects them. Finally, the stiffness and mass of the two transmission
simulators are analytically subtracted.
The free-free experimental modal test setup is shown in Fig. 4.14 with 12 triaxial
accels on the transmission simulator and 2 triaxial accels at points of interest on the
beam. Twenty-five rigid body and elastic modes were extracted out to 4 kHz. This
structure had eight discrete bolted attachments, which would require 48 discrete
constraints for the three rotations and three translations at each connection. Twelve
Fig. 4.14 Coupling of experimental substructure C with analytical substructure D to generate full
system E after transmission simulators (Aan1 and Aan2 ) are subtracted. Reproduced from Allen
et al. (2010)
4.4 Substructure Decoupling 113
Fig. 4.15 Experimental substructure free modal test setup with 36 sensors on the transmission
simulator ring and a few sensors at points of interest on the substructure. Reproduced from Allen
et al. (2010)
triaxial accelerometers were placed on the TS, as shown in Fig. 4.14 none of which
were at the eight connection points. The sensor locations are shown in Allen et al.
(2010). Then, 18 modes of the transmission simulator analytical model were used
to span the connection space motion, using 36 measurement points to describe the
mode shapes. The method thus reduces the number of constraints down from the
classic 48 connections to 18. The 18 modes covered a frequency bandwidth of 2 kHz
(Fig. 4.15).
Component Mode Synthesis Theory Using Primal Formulation In A General-
ized Framework
Following the notation outlined previously and in Klerk et al. (2008), and assuming
that each substructure has been approximated with a reduced model, whether experi-
mental or analytical, the displacements are approximated with the modal substitution
as
u∼ = Rη, (4.65)
M(s) Rη̈(s) + C(s) Rη̇(s) + K(s) Rη(s) = f (s) + g(s) + r(s) , (4.66)
114 4 Experimental Substructuring
where M, C, and K are mass stiffness and damping matrices, the superscript (s)
denotes the particular substructure, f (s) denotes the external forces applied to the
substructure, g(s) represents the equilibrium joining forces from another substruc-
ture that will be applied to the substructure, and r(s) is the residual force due to the
mismatch caused by the approximation of the displacements in Eq. (4.65). Premul-
tiplying by R T yields
which leaves
or in a renamed form
where subscript m denotes modal quantities. The matrices are, if the mode shapes
are mass normalized,
Mm(s) = R T M(s) R = I
C(s) T (s)
m =R C R =
\
2ζr ωr \
At this point, let us concatenate the various substructures together in the uncoupled
form as
Mm η̈ + Cm η̇ + Km η = fm + gm . (4.72)
Buc = 0, (4.73)
where B is a Boolean matrix of ones, zeros, and negative ones, and for convenience
we will include only connection DOF in the displacement vector. Take the partition
of Eq. (4.65) for only the connection DOF and again make the modal substitution
BRc η ∼
= 0, (4.74)
where the subscript c indicates taking only the partition of R necessary for the
connection DOF.
4.4 Substructure Decoupling 115
Up to this point, we have followed the generalized framework rather strictly, but
at this point, the transmission simulator method affects the rest of the development.
It is assumed that there is an analytical model of the transmission simulator; one uses
a truncated set of its mode shapes, Ψ c , as a basis to span the space of the connection
motion for each substructure. Now we use the pseudo-inverse (denoted with a super-
script +) to project the constraint on the space of the transmission simulator vector
space by premultiplying both sides by the block diagonal pseudo-inverse as
Ψ cB D+ BRc η ∼
= Ψ cB D+ 0, (4.75)
where ⎡ ⎤
Ψ+c 0 0
Ψ cB D+ = ⎣ 0 ... 0 ⎦ (4.76)
0 0 Ψ+ c
will have as many block rows as there are substructures. The right-hand side of
Eq. (4.75) is still a vector of zeros, although the number of constraints (rows) is
reduced since the matrix Ψ c is selected so that it always has more DOF than modes.
The final constraint is
B̃η = 0, (4.77)
where
B̃ = Ψ cB D+ BRc . (4.78)
There are multiple reasons for premultiplying by Ψ cB D+ . First, it softens the con-
straint (reduces the number of constraints). The advantage of this is that it gives a least
squares fit through the measured motions of the transmission simulator at the DOF to
which the constraints will be applied. The mode shapes of the transmission simulator
provide a smoothing effect through the measured motions, which always have exper-
imental error. The modified matrix greatly improves the conditioning of the problem.
Also, one does not HAVE to use motions measured directly at the attachment points
and also does not HAVE to measure rotations, which are inherently carried along in
the generalized DOF of the transmission simulator. The generalized modal DOF of
the transmission simulator is γ in the following
Ψ cγ ∼
= Rc(s) η(s) , (4.79)
which can be seen in the left-hand side of Eq. (4.75) as converting the constraint to
the transmission simulator modal coordinates.
Now continue the development in the general framework from Eq. (4.77). We
perform another modal-like substitution with
116 4 Experimental Substructuring
η = L̃q. (4.81)
B̃L̃q = 0. (4.82)
If one chooses L̃ such that it is in the null space of B̃, then Eq. (4.82) is guaranteed
to be satisfied because
B̃L̃ = 0. (4.83)
All the rows of B̃ are orthogonal to all the columns of L̃. Since B̃ is known, a one
line command in MATLAB can provide L̃. Substituting Eq. (4.81) back into the
uncoupled equations of motion in Eq. (4.72) and premultiplying by L̃T gives
which couples the equations of motion, reducing the number of rows in Eq. (4.72)
by the number of constraints (rows) in Eq. (4.77). This leads to the primal coupling
formulation in this framework. In this formulation, L̃T gm = L̃T R T g = 0, since the
rows of L̃T are orthogonal to a linear combination of the columns of R T , leaving
where
M̃m = L̃T Mm L̃
C̃m = L̃T Cm L̃
. (4.86)
K̃m = L̃T Km L̃
f̃m = L̃T fm
Bu = 0 (4.88)
LT g = 0, (4.89)
where M, C, and K are block diagonals with as many blocks as substructures. The
constraint forces g between the substructures can be written as
g = −BT λ (4.90)
4.4 Substructure Decoupling 117
Taking the Fourier transform to bring this into the frequency domain, where each
quantity is now a function of frequency, gives
Z BT ū f̄
= , (4.92)
B 0 λ̄ 0
where Z is the block diagonal impedance matrix resulting from the mass, stiffness,
and damping. The frequency response function matrix, H, which is often measured
experimentally is the inverse of Z. The dual formulation coupled formulation in
terms of H is derived from Eq. (4.92) by eliminating λ̄, yielding
Ψ B D+ Bū ∼
=0 (4.94)
So one has
B̂ū = 0, (4.95)
where
B̂ = Ψ B D+ B. (4.96)
Now simply substitute Eq. (4.96) into Eq. (4.93) to give the frequency-based trans-
mission simulator equation as
The B̂ matrix transforms the physical connection DOF frequency response functions
(FRFs) into generalized DOF FRFs cast on the space of the transmission simulator
mode shapes. This collapses the size of the physical connection DOF FRFs down
to the size of the number of modes of the transmission simulator, providing some
least squares smoothing, and makes the matrix inversion in Eq. (4.97) much better
conditioned.
where R are a truncated set of the mass normalized mode shapes of each sub-
structure coming from experiment or analysis. The motion of the transmission
simulator in all three substructures should be the same when all the substructures
are coupled, so two sets of physical constraints can be written
Just consider the first of these constraints, invoke the modal substitution from
Eq. (4.99), and premultiply by the pseudo-inverse of transmission simulator
mode shapes, RT+S to give
A similar process can be applied to the second constraint of Eq. (4.100). The
constraints can now be written in the form given in Eqs. (4.76)–(4.78) as
4.4 Substructure Decoupling 119
⎡ ⎤⎧ ⎫
R F Emeas ⎪ ηF E ⎪0 ' ( 0
RT+S 0 I −I 0 ⎢ ⎨ ⎬ 0
⎥
+ ⎣ 0 R E X Pmeas 0 ⎦ η E X P =
0 RT S 0 I −I ⎪
⎩ ⎪
⎭ 0,
0 0 RT S ηT S
(4.103)
where Ψ cB D+ in Eq. (4.75) is RcB D+ and B is the Boolean matrix here, thus
⎡ ⎤
R F Emeas 0 0
RT+S 0 I −I 0 ⎢ ⎥
B̃ = ⎣ 0 R E X Pmeas 0 ⎦ . (4.104)
0 RT+S 0 I −I
0 0 RT S
η = L̃q (4.81)
and
B̃L̃q = 0 (4.82)
and L̃ is the null space of B̃, which is known. Substitute Eq. (4.81) into Eq. (4.98)
and premultiply by L̃T to give
⎡ ⎤ ⎡ ⎤
ω2F E 0 0 IF E 0 0
⎢ ⎥ ⎢ ⎥
LT ⎣ 0 ω2E X P 0 ⎦ Lq − ω2 LT ⎣ 0 I E X P 0 ⎦ Lq̈ = 0,
0 0 −2ω2T S 0 0 −2IT S
(4.105)
providing the coupled equations from which the eigenvalue problem can be
solved yielding. The solution will provide a set of eigenvectors, Φ, frequencies,
ω2 , and modal coordinates, β. Now the coupled displacements will be
y = RLΦβ. (4.106)
from which we usually just take the diagonal values to give 2ζnew ωnew (Fig. 4.16).
120 4 Experimental Substructuring
Fig. 4.16 Transmission simulator method FRF of coupled system (red) versus truth model
(blue)—Driving point response at tip of beam in Fig. 4.14
In Fig. 4.16, one can see the axial FRF at the tip of the beam (see Fig. 4.14).
The red FRF was constructed from the modal parameters of the coupled sys-
tem mode shapes, frequencies, and damping using the transmission simulator
method with the CMS approach. The blue FRF represents the truth data con-
structed from the modal parameters of a highly validated finite element model
of the full system.
where the force input is on substructure C and the response output is on sub-
structure D, and
where the force input is on substructure C and the response output is also
on substructure C. Here, the subscript r represents the output response and
the subscript i represents the input force, and the subscript c represents the
connection DOF between the two substructures. With the transmission simulator
method, the Boolean matrix B is replaced with B̂ to convert all the connection
DOF to modal DOF of the transmission simulator in Eq. (4.97), so that
from this. If a frequency line is supposed to have a resonant response in the coupled
system, then the real part of the coupled FRF has to go to zero, and the imaginary
part of the FRF will be a drastic amplification (on the order of the quality of the
resonance, Q = 1/(2 ∗ ζ )). Restating this, if one is to get the coupled natural fre-
quency correctly, the real part of the connection FRFs must be accurate. If one is to
get the coupled amplitude of the resonant FRF correctly, the imaginary part of all the
FRFs of concern must be accurate. Consider the real part of a FRF obtained from an
impact test. On first consideration in the left figure below, it look good. Let us assume
that when the substructures are coupled, the first resonance will drop down to about
100 Hz. When one zooms in on the real part between 0 and 300 Hz, one sees that there
is a significant amount of uncertainty in the real part of the FRF. This could affect
the substructured FRF so that it crosses zero at the wrong place, or even multiple
places, yielding the resonant response at wrong frequencies (Figs. 4.17 and 4.18).
Consider the imaginary portion of the drive point FRF below. The response at left
looks good. However, if a new mode is destined for the region just above 3000 Hz,
look at the zoomed response in that bandwidth on the right. In theory, a drive point
FRF should never have an imaginary part less than zero. Clearly, there are many
frequencies with response below zero. This could not only give inaccurate estimate
of the amplitude of the imaginary portion of the FRF, but even the wrong sign! For
this reason, structures with higher damping are typically easier to perform FBS than
lightly damped structures as shown here.
1. Response Sensor Measurement Errors and their Sources
Here is a list of common sources of measurement error, approximately ranked for
probability of occurrence. Also included are some recommendations for correct-
ing these.
a. Bookkeeping errors
(Wrong sign, wrong DOF, crossed wires, wrong sensitivity/gain, local coor-
dinate system errors) Administrative double checks by multiple people on
the mounting, channel table, label checks, and coordinate system definitions
can help eliminate these. For “free” configuration tests, a great engineer-
ing check to identify such errors is to compare the low-frequency real part
response of each FRF with the analytical response for that DOF using just
the rigid body modes. Usually, they will be within about 10% of each other.
b. Choosing too wide a bandwidth
In general, the difficulty of developing an accurate experimental model
increases with frequency. Minimize the bandwidth as much as possible to
meet the requirement. Push back when customers ask for “the moon.”
c. Undetected overloads, especially in drive point sensors
Undetected overloads where the sensor response, amplifier response, or the
band edge of the DAQ hardware is beyond, the rated range can ruin or pollute
the data, but the digital anti-aliasing filter can disguise it so that it looks
“fine”. Set the DAQ and filter bandwidths wide open, and collect at-level
data in the time domain to verify that the sensors or signal conditioning
equipment is not being driven beyond its rated voltage range (a typical
integrated electronics sensor maximum is 5 volts)
d. Cross-axis sensitivity
In very lightly damped systems, the out-of-axis response to a strong res-
onance will have a peak in the real part and the imaginary part will cross
zero, out of phase with a normal acceleration response. This is not easily
corrected.
e. Base strain sensitivity
Base strain sensitivity allows strain in the surface of the structure to which the
accelerometer is mounted to produce false acceleration voltage output. It can
sometimes be observed as a large uncharacteristic low-frequency response
sloping downward as frequency increases. This can be remedied by putting
the accelerometer on a block, or (less preferably) on tape with an adhesive
layer.
f. Mounting or mounting block resonances
At some high frequency, tape-mounted accelerometers will see phase shift
from the true response and block-mounted accelerometers will have a mode
of the block mass oscillating locally on the structure. If possible, testing an
accelerometer mounted without tape and comparing the FRF mounted on
124 4 Experimental Substructuring
the tape will provide observation of this error. Mounting block resonances
will change frequency if the size of the mounting block is changed. Beware
of mounting accelerometers on a screw head or a nut as one may measure
the fastener response instead of the desired structure response.
g. Sensors with wrong amplitude sensitivity or broken sensors
Sensors that are not responding can often be diagnosed by observing their
time- or frequency-domain response, especially in relation to the noise floor
with no stimulus applied. Sensors that have wrong sensitivity, for example,
from a cracked crystal that is still partially functional, are difficult to identify
by simple observation of the response. The check in the Bookkeeping errors
section above is of value here for free systems. Another similar check for
free systems is to excite the structure at a low frequency below the elastic
modes in each rigid body direction and extract the real part of the FRF at
that frequency as a shape. The shape should be an appropriate rigid body
response shape.
2. The Most Critical Sensor Locations for Substructuring
By far the most critical sensors for experimental substructures are the connection
DOF sensors. Those are the ones used in computing inverses or pseudo-inverses
and ultimately determine the coupled modal frequencies as well as having large
bearing on the amplitude of the coupled responses.
a. Drive point sensors are the most critical
In FBS, all the connected DOFs require drive point responses. In CMS sub-
structuring, the drive point FRFs determine the scaling of the mode shapes,
so they are very important. A low-frequency check of each connection DOF
response as previously mentioned above in bookkeeping errors is a great
first check of these FRFs for free substructures. If one performs the sub-
structuring and the rigid body modes are appropriately coupled, one should
obtain the correct coupled low-frequency response. There is no approxima-
tion with coupling of rigid body modes and no singularities if mass properties
are appropriate, so this should always work. As mentioned before, if there
are errors in this coupling, it may help identify bookkeeping or some other
error in the setup.
b. Neglect connection DOF (including rotations) to your peril
As previously mentioned, 3/4 of the FRF matrices contain rotation terms. In
many cases, rotations cannot be neglected without significant degradation of
the results. Approximation of the connection DOF is better than elimination.
The transmission simulator method inherently includes rotation estimates.
The virtual point transformation is another approach to include the estimate
of rotations.
c. For CMS substructuring check synthesized FRFs against measured
data for each substucture
Once the measured FRFs are obtained, the modal parameters must be
extracted for CMS substructuring. Errors in the extraction will cause detri-
ment to the substructured results. A good first check is to compare the ana-
4.5 Measurement Methods for Substructuring 125
lytical Complex Mode Indicator Function (CMIF) with the CMIF from the
measured data. They should compare “well”. One can compare each ana-
lytical FRF with its measured counterpart as well.
In particular, the drive point FRF from which certain modes are extracted is
very important. The drive point mode shape and modal mass are estimated
from the drive point FRF. If the modal mass is not estimated correctly, the
drive point mode shape will be estimated inversely proportional in error. The
same is true if the damping is estimated incorrectly. If the drive point shape
is in error, generally all the cross point mode shapes will be inversely pro-
portional in error to the drive point mode shape error. This can be observed
by considering the fit to the FRF with an error considered in each of the
various terms singly (modal mass, damping, drive point shape, cross point
shape). See the standard real modes FRF equation below:
N
−ω2 φi φ j
Hi j (ω) = .
r =1
m r (ωr2 − ω2 + 2 jζr ωωr )
Fig. 4.19 Complex mode indicator function example. The figure compares the first singular value
in a CMIF for the experimental data, the fit modal model, and a substructuring prediction
however, components are often connected by bolts, welds, or adhesives, that physi-
cally show more resemblance to a line or surface connection than to a single point.
Modeling such connectivity by experimental measurements requires more thoughtful
contemplation of what is really going on at the interfaces between two structures.
In general, one can distinguish two categories of connectivity between substruc-
tures, of which one is normally the most appropriate for a given case:
1. The components are connected at a handful of points that can be regarded as
discrete nodes (Fig. 4.23a). These points are permitted to displace with respect to
each other, thereby allowing for significant deformations in the structures. The
connection points themselves, however, behave fairly rigid in the area near the
interfaces. Deformations in the connecting points are of minor order and not
contributing to the coupled vibrations. Examples are truss structures in bridges
and a steering system mounted in a vehicle subframe.
4.6 Virtual Point Transformation 129
2. The interface resembles a continuous line or surface (Fig. 4.23b). The deformation
of the interface is strongly coupled to the internal vibrations of the structure, and
can therefore not be discretized to a small number of nodes. Examples include
the coupling of a gearbox to an engine, glued carbon-fiber panels of a vehicle’s
bodywork and the partitions of an aircraft’s fuselage.
This next section focuses on the discrete interface situation. Each connection point
can be appointed linear and rotational DOFs for the nodal displacement and forces,
which have a clear physical interpretation. In order to solve the second situation, one
would instead define a set of deformation modes for the interface region and assign
DOFs to their modal amplitudes and forces.
Discrete Interface Modeling
Let us consider a single bolted connection as illustrated in Fig. 4.24. The pretension
in the bolt causes friction to occur at the interfaces, which prohibits the two structures
to have relative motion with respect to each other. If both structures were represented
by an FE model, the substructuring task amounts to coupling the displacements and
interface forces of the coinciding nodes, as illustrated in Fig. 4.24b. By coupling a
sufficient number of nodes over a larger area, any rotational coupling is implicitly
accounted for.
Let us now consider the procedure of experimental modeling. An admittance
measurement needs to be conducted on each substructure in order to obtain the
respective FRF matrices. To facilitate substructure coupling, the simplest way would
be to mount a single triaxial accelerometer on the connection point of each substruc-
ture and excite in X-, Y-, and Z-directions close to the sensor. This yields a 3-DOF
description which is sometimes called a Single-Point Connection (SPC). In this way,
one obtains a 3 × 3 FRF matrix for the substructure’s interfaces, which can be used
for substructure coupling.
130 4 Experimental Substructuring
(a) Friction at the full contact area. (b) Discrete coupling with coinciding
nodes.
(c) EMPC: multiple points, only trans- (d) Virtual point transformation: 6 col-
lational directions (9 DoFs). located translations and rotations (9 →
6 DoFs).
Fig. 4.24 A bolted connection: four ways to consider the interface problem
The SPC method does not account for rotational coupling, which is known to be
essential in many cases of substructuring Liu and Ewins (1999), Montalvao et al.
(2004). Measurement and excitation (or in fact observation and controllability) of
rotations has been an active topic of research among experimentalist and has led to
several approaches:
1. Direct measurement. Rotation sensors and transducers can be thought of, but
have not been the practical standard over the past. Some examples of unconven-
tional hardware for rotation measurement or excitation can be found in Petersson
(1987), Bello et al. (2003), Liu et al. (2013).
2. Modal expansion. If modal information of the structure is available, for instance,
from static FEM analysis, rotations can be obtained from extrapolation of the
measured data. This is the concept of the system equivalent reduction and expan-
sion process (SEREP) O’Callahan et al. (1989).
3. Finite differences. Rotations can be approximated from translations measured
at known distances from each other Duarte and Ewins (2000), Bregant and
Sandersons (2000). Simple finite difference approaches such as θ = (u 0 − u d )/d
may suffice for simple plate or beam problems Sanderson and Fredö (1995),
Elliott et al. (2012), but can be cumbersome for obtaining full 6-DOF kinematics.
4. Multiple translations. The equivalent multi-point connection (EMPC) method
Pasteuning (2007), de Klerk et al. (2008) accounts for the rotations implicitly,
by coupling translational FRFs of multiple points in the proximity of the inter-
face. This is illustrated in Fig. 4.24c. In practice, a minimum of three triaxial
accelerometers (not in line) is sufficient to couple all translational and rotational
4.6 Virtual Point Transformation 131
For the derivation of the virtual point transformation, let us consider the coupling
problem of two experimental substructures A and B as illustrated in Fig. 4.25. The
substructures have non-collocated interface DOFs by nature: neither sensors nor
excitation points appear identically. The measured FRF matrices are denoted by
YA (ω) and YB (ω). The governing equation of the uncoupled dynamics is of the
admittance FRF form:
fB2
uA
2
B B
qA A
2 , m2 ⇐⇒ q2 , m2 uB2
fA
2
Fig. 4.25 The virtual point transformation workflow for coupling of substructure A (left) and B
(right). Each substructure is instrumented by three acceleration sensors (blue, only four of the six
sensors are visible here). The excitations of the FRF measurements are depicted by red arrows
4.6 Virtual Point Transformation 133
2 ⇔ g2 . In what follows, the assumption is used that the structures behave rigidly
gA B
in the area close to the coupling point. The derivation focuses on the interface DOFs
of a single substructure, denoted for simplicity by u ∈ Rn .
Interface Displacement Reduction
Let us express the n interface displacements u by m < n interface deformation modes
(IDMs) q. The IDMs are contained in the columns of the n × m matrix R, which
is a frequency-independent mode shape matrix (the actual construction of the IDMs
is discussed in Sect. 4.6.3). As the number of IDMs is smaller than the number of
interface DOFs, a residual on the displacements is added, denoted by μ. This residual
captures all displacements not in the subspace of the IDMs, which is normally the
“flexible” motion.
u = Rq + μ q ∈ Rm . (4.116)
⎧ T −1 T
⎨ q = R WR R Wu
⎪ (4.118a)
ũ = R RT WR −1 RT Wu (4.118b)
⎪
⎩
μ = I − R RT WR −1 RT W u (4.118c)
⎪ ũ = Fu with F RT (4.119b)
⎩
μ = (I − F) u (4.119c)
9 Todetermine a 6-DOF virtual point with just n = 6 translations, one should consider a triad of
points and register, for instance, the directions XYZ, XY, and Z, respectively.
4.6 Virtual Point Transformation 135
However, one can always calculate a resultant moment m in the virtual point
due to the applied forces f, if their locations and directions are known.
Hence, the reduction of forces is performed inversely. A similar IDM matrix can
be constructed, relating the measured forces to virtual point loads. This is again
denoted by R; distinction with the displacement IDMs will be made later on. Note,
however, that the matrix needs to be transposed in order to write the relation between
m virtual loads and n (interface) forces10 :
m = RT f. (4.121)
⎪ f̃ = Ff with F T R = RT
T T
(4.123b)
⎩
ν = (I − F) f. (4.123c)
Equations (4.123b) and (4.123c) allow to calculate the IDM-filtered forces and
residuals (here denoted by ν) in case that the forces are the input. For the virtual
point transformation of admittance, these are normally not used.
Virtual Point Assembly
The interface reduction steps as derived above allow us to assemble only the part of
the dynamics of substructures A and B that can be described by the subspaces of their
respective IDMs. The remaining displacements and forces are left free. From here
on, the transformation matrices for the displacements and forces are clearly denoted
by Tu and Tf .
Let us now continue the assembly of (4.115) by coupling only the part of the
displacements and forces that is admissible with respect to the IDMs. The following
definitions are used:
A A A
u 2 g 2 Y 22 0
u= B ; g= B ; Y=
u 2 g 2 0 YB22
10 The calculation of m in (4.121) shows resemblance with the modal participation factor calculated
The virtual point transformations of (4.119a) and (4.123a) are now used to write com-
patibility and equilibrium. Whereas the measured displacements uA 2 , u2 are incom-
B
patible by nature, the virtual points are in fact compatible so it is permitted to require
qA2 = q2 . This means that the IDMs used to express the kinematics of substructure
B
A and B are defined the same way, for instance: both describe three global-frame
translations and three rotations with respect to a unique point in space, i.e., the vir-
tual point. The virtual point DOFs are collocated and can thus be used to write
compatibility and equilibrium.
The same holds for the virtual interface loads, allowing to write the following
dual coupling equations:
'
u = Y (f + g) = Y f − TfT BT λ
with B −I I. (4.123)
Bq = BTu u = 0
Lagrange multipliers are substituted for the virtual point interface forces of A and
B, enforcing equilibrium on the (rigid) part of the forces in the subspace of the
IDMs. Likewise, compatibility is enforced by the second equation onto the admissible
displacements. The assembly procedure continues in a familiar LM-FBS fashion,
namely, by solving for λ:
BTu Y f − TfT BT λ = 0
BTu Yf = BTu YTfT BT λ
λ = BTu YTfT BT −1 BTu Yf (4.124)
Now that the virtual point coupling forces are known, the full virtual point coupled
solution is found by substitution back into (4.123):
u = Yf − YTfT BT BTu YTfT BT −1 BTu Yf. (4.125)
Let us now examine (4.125). At first sight, one recognizes the same ingredients as
for standard LM-FBS coupling. The response consists of an uncoupled response,
i.e., the first term on the right-hand side, and a coupled response to the applied
forces f. The coupling, however, takes a “detour” through a total of four virtual point
transformation matrices. To better comprehend this procedure, let us expand (4.125):
virt. point
admittance
Yf − YTfT BT B Tu YTfT BT −1 BTu Yf .
u =
uncoupled response to virt. point incomp.
response
virt. point λ int. admittance in q
4.6 Virtual Point Transformation 137
Reading from right to left, we first recognize the incompatibility caused by the
uncoupled responses. These are transformed to the virtual point space by the operator
Tu . Between the brackets, we observe the combined interface admittance. This is put
in the virtual point space by pre- and post-multiplication of the measured FRFs by
Tu and TfT . Hence, the term in the center constitutes the virtual point admittance
of the interface, which will be elaborated on later. The Boolean matrices on both
sides of this term function as expected, namely, to select and sum up the admittance
at the interface DOFs in the virtual point space. Lastly, the term to the left of the
inverted admittance governs the response to the virtual point interface forces that are
in front of it. These forces are transformed back to the measured DOF space by the
operator TfT .
This section has demonstrated how interface FRFs, incompatible by the nature
of experimental measurement, can be coupled by use of appropriate virtual point
transformations. Having clarified all the terms in (4.125), let us now define two
approaches to proceed.
Virtual Point Coupling
The coupling detour explained above can be simplified by grouping the respective
pairs of B and T matrices. This yields two new, yet familiar operators that carry out
the actions of virtual point transformation and DOF selection at once:
)
Bc BTu compatibility at virtual point (4.127a)
Be BTf equilibrium at virtual point (4.127b)
Note that these matrices can be fed straight into the LM-FBS algorithm. The sub-
structures’ admittance matrices can be left untouched and, as a consequence, the
resulting coupling FRFs are found in the original dual space of the measured DOFs.
Virtual Point FRFs
In the introduction of this chapter, the need for a nodal model was highlighted: a
dynamic model, composed from collocated FRFs, that is compatible for coupling
with FE or other experimental models. Indeed, for the purpose of dynamic sub-
structuring or transfer path analysis, it may be preferred to transform the entire
measurement into a virtual point admittance matrix. Such matrix can be stored in a
database and easily retrieved for modeling purposes, in the capacity of a dynamic
super-element.
The virtual point admittance was already identified above, namely, as the com-
bined subsystems’ admittances pre- and post-multiplied with their respective trans-
formation matrices. This may as well be computed on a substructure level, for
instance, to obtain the nodal admittance FRFs of substructure s:
The subscript qm is used here to denote the virtual point DOF space of the new
admittance matrix. As will be demonstrated later, the IDM matrices R have a block
138 4 Experimental Substructuring
diagonal structure and so do their resulting T matrices. This means that, given an
arbitrary structure s, the subset of FRFs used for transformation of coupling point
1 is not being reused for transformation of coupling point 2. In addition, the block
diagonal structure makes it easy to keep measured points of interest or additional
excitation points untouched, by mixing in identity elements in the substructure’s
transformation matrices. Let us demonstrate this for the DOFs of the substructures
A and B:
uA
1
I 0 uA
= 1 =⇒ qA = TAuu
A
qA
2
0 TAu2 uA
2
B B B
q2 Tu2 0 u2
= =⇒ qB = TBu uB .
uB3 0 I uB3
The coordinate sets qA and qB now contain a mix of original internal DOFs and
virtual point interface DOFs. With respect to the interface conditions, one can now
define Boolean matrices in the traditional fashion, as all interface DOFs are already
collocated due to their matching IDMs. The principles as demonstrated here can be
employed in a customized fashion to serve the needs of the substructuring or TPA
job in prospect.
In the preceding section, it was shown that the virtual point transformation relies on a
spatial reduction of coordinates. The interface connectivity is reduced to virtual points
using a predetermined set of interface displacement modes (IDMs) R. This section
elaborates on the construction of the 6-DOF IDMs for a virtual point. The example
of Fig. 4.26 is used to illustrate the theory. Virtual point v is surrounded by n k = 3
triaxial acceleration sensors, registering a total of nine translational displacements11
in the local (x, y, z) frame of the sensors. A hammer impact is indicated by a thick
red arrow.
Displacement IDM
The sensor IDM matrix R (the subscript is dropped in this section) translates the local-
frame displacements to six virtual point displacements and rotations plus a residual.
Let us write out this equation for a single triaxial sensor k. The orientation of the
sensor is determined by three measurement directions, specified as orthonormal unit
vectors:
[ ekx eky ekz ] = Ek
11 Infact the sensors measure accelerations, but for simplicity of notation displacements are con-
sidered here.
4.6 Virtual Point Transformation 139
eky
ukx ekx
uky
eh fh
rk
qYv
rh
Y qvX
X qv Z
Fig. 4.26 The IDMs associated with the virtual point (green) can be constructed from the positions
and directions of the measured displacements (blue) and impacts (red)
The distance from the sensor to the virtual point is given by vector rk . The respective
local displacements along these directions are denoted by uk . The six DOFs of virtual
point v comprised of the set qv . The following kinematic relation can be established
between the virtual point qv and the sensor displacements uk :
⎤ ⎡
q Xv
⎡ k⎤ ⎡ k ⎤⎡ ⎤⎢ ⎥
⎢ qv ⎥ ⎡ k ⎤
ux k
ex,X ex,Y k
ex,Z 1 0 0 0 r Zk −rYk ⎢ Y ⎥ μx
⎢ k⎥ ⎢ k k ⎥⎢ ⎥⎢ q v⎥
⎢ k⎥
⎣u y ⎦ = ⎣e y,X e y,Y
k
e y,Z ⎦ ⎣ 0 1 0 −r Z 0 r X ⎦ ⎢
k k Z ⎥
⎢ v ⎥ + ⎣μ y ⎦ . (4.129)
⎢ ⎥
q
u kz k
ez,X k
ez,Y k
ez,Z 0 0 1 rYk −r Xk 0 ⎢ θ X ⎥ μkz
⎣qθv ⎦
Y
qθvZ
This relation states how much displacement is measured on the channels of sensor
k for a given displacement and rotation of the corresponding virtual point v. Any
residual displacement ends up in μ.
Introducing R̄kv as the 3 × 6 global-frame IDM matrix associated with sensor k
and virtual point v, we can write this relation for each sensor:
T
uk = Ek R̄kv qv + μk
T
uk = Rkv qv + μk with Rkv Ek R̄kv . (4.130)
Rkv is the local-frame IDM matrix. For a typical configuration of three sensors
per virtual point, the Rkv matrices can be stacked to form the following system of
equations:
140 4 Experimental Substructuring
⎡ ⎤
R1,1
⎢ R2,1 ⎥
⎢ 3,1 ⎥
⎢R ⎥
⎢ ⎥
⎢ R4,2 ⎥
⎢ ⎥
u = Rq + μ with R = ⎢ R5,2 ⎥. (4.131)
⎢ ⎥
⎢ R6,2 ⎥
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
Rn k ,n v
Equation (4.131) has the same form as (4.116). Alternatively, one may prefer to
describe the virtual point relation for each measurement channel individually. A
single displacement DOF u ik with local direction vector eik can be related to the
associated set of virtual point coordinates in qv :
* T T + v
u ik = eik rik × eik q + μik = Rikv qv + μik . (4.132)
The single-DOF row vector between the block brackets is of size 1 × 6. The full
matrix R can be constructed row-wise per sensor channel i and column-wise for the
different virtual points.
Regardless of the construction method, the resulting IDM matrix R is block diag-
onal, or at least contain IDMs that are uncoupled with respect to the various virtual
points and sensor groups. Although not treated here, R may be augmented with flex-
ible IDMs if desired.12 In any case, care should be taken that R is full rank, which
implies that sufficient measurement points are considered to compute an inverse or
pseudo-inverse T using (4.119a).
Force IDM
The transformation of the forces can be performed in an analogue way. The following
relation can be written for the set of virtual forces mv as a result of a single impact
f h in the direction of eh and at a distance rh from the virtual point v:
⎡ v⎤ ⎡ ⎤
mX 1 0 0
⎢ m vY ⎥ ⎢ 0 0 ⎥
⎢ ⎥ ⎢ 1 ⎥⎡ h ⎤
⎢ mv ⎥ ⎢ 0 1 ⎥ e
⎢ Z⎥ ⎢ 0 ⎥ ⎣ Xh ⎦ h
⎢ v ⎥=⎢ ⎥ e f (4.133)
⎢m θ X ⎥ ⎢ 0 −r Zh rYh ⎥ Yh
⎢ v ⎥ ⎢ h ⎥ eZ
⎣m θ ⎦ ⎣ r Z 0 −r Xh ⎦
Y
m vθZ −rYh r Xh 0
Note that the virtual point forces are indeed the direct result of applied forces; hence,
we do not need to add a residual.
12 For a definition and application of flexible IDMs, see Pasma et al. (2018).
4.6 Virtual Point Transformation 141
Every impact adds a single column to the transposed IDM matrix RT . Assuming
nine impacts per virtual point, the system for the complete set of DOFs reads
⎡ ⎤
R1,1
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢ 9,1 ⎥
⎢R ⎥
⎢ ⎥
⎢ R10,2 ⎥
m = R f with R = ⎢
T
⎢ .. ⎥.
⎥ (4.135)
⎢ . ⎥
⎢ ⎥
⎢ R18,2 ⎥
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
Rn h ,n v
Note that the IDM matrix of the forces takes the same form as the IDM matrix of the
displacements from (4.131). In fact, if one decides to excite only on the sensor faces
(which is generally not advised, see Sect. 4.6.5), one has obtained the exact same
IDM matrix, apart from some sign changes.
With more than six independent excitations per virtual point, (4.135) becomes
underdetermined, which means that there is no unique combination of the excitation
forces f from which a certain m can be constructed. The solution is found using the
pseudo-inverse of (4.123a). Instead of minimizing a residual (which was the case for
the displacements), this operation finds a set of excitation forces that minimizes the
W−1 -norm of f:
f = argmin fT W−1 f .
This again has advantageous properties for interface weakening and error suppres-
sion, as discussed in Sect. 4.6.2.
Sensor Consistency
The sensor consistency function is an indicator of the consistency of the measured
response channels with respect to a certain load case. It compares the measured
responses with the responses after projection onto the IDMs. It was originally pro-
posed in de Klerk et al. (2008) as an indicator of the “rigidness” of the interface, but
has additional, generally useful properties.
For the evaluation of the consistency of the sensors around virtual point 1, let
us consider their responses to a load case f2 , denoted by u1,2 . This load vector is
composed from one or more excitations at coupling point 2, reasonably distant from
virtual point 1. This way, the responses in u1 due to f2 are probably more global
(showing some signal in all directions) than when considered for one ore more
excitations near the sensors around point 1. The filtered responses are found by
premultiplying the responses with the IDM projection matrix Fu for the respective
coordinates:
'
u1,2 u1,f2 = Y12 f2 (4.136a)
ũ1,2 ũ1,f2 = F11 Y12 f2 . (4.136b)
ũ1,2 (ω)
ρu1,2 (ω) = . (4.137)
u1,2 (ω)
A function value ρ(ω) = 1 means that the amplitude of all filtered sensor channels
is left unchanged, suggesting that all sensor responses can be fully described by
the m IDMs (6 for a rigid interface) at the specified frequency. Mathematically, it
implies that the singular value decomposition Y12 = U11 Σ 12 V22 T
only comprises m
significant singular values and that the associated left singular modes U11 are all in
the subspace of Ru . In practice, it indicates that all sensors are correctly positioned,
aligned, and calibrated, as otherwise at least one channel would “swing” out of line
and reduce the overall value of ρ.
If the value drops down above a certain frequency, it is often an indication of the
presence of a flexible interface mode. This can be verified by looking at mode m + 1
4.6 Virtual Point Transformation 143
in U11 . If the associated singular value is significant, one could consider to augment
the IDM basis by a flexible mode, e.g., based on mode m + 1 in U11 .
Specific Sensor Consistency
For evaluation of the consistency per measurement channel, we introduce a variation
of the spectral coherence function. Essentially, this is an expansion of the spectral
coherence function Bendat and Piersol (1980) when evaluated for a total of two
spectra13 :
⎧
⎪
⎨1 if x = y
(x + y)(x ∗ + y ∗ )
coh(x, y) x, y ∈ C =⇒ 1
if x ⊥ y
2(x x ∗ + yy ∗ ) ⎪
⎩
2
0 if x = −y.
(4.138)
This equation compares two values with respect to their amplitude and phase. It is
again bounded between zero and one and can thus be used to objectively compare two
complex response spectra with each other. For evaluation of the specific consistency
of measurement channel u i , we write
ρu i,2 (ω) = coh ũ i,2 (ω), u i,2 (ω) u i ∈ u1 . (4.139)
If the overall sensor consistency is poor over the full frequency range, it may very
well be a problem of just one incorrect sensor direction or calibration value.14 (4.139)
allows to evaluate the consistency of each measurement channel with respect to the
full transformation. This way, the problematic sensor can be identified and the entries
in Ru corrected if needed.
Impact Consistency
Similar to the sensor consistency, it is possible to define a consistency function for the
applied forces. This value is particularly useful to assess the accuracy of the excitation
positions and directions—crucial aspects of impact hammer measurement.
We now consider the effect of the force filter matrix Ff on the responses. In order
to provide a methodology analogue to the sensor consistency, a linear combination
of responses is considered for the evaluation, denoted by scalar y2 . Vector w2 is
introduced to compute a (weighted) sum over the response channels of virtual point
2:
y2 = w2T u2 w2 ∈ Rn 2 .
Note that the quantity y2 has very little dynamical significance; it only allows to
assess the effect of the force filtering with respect to some responses in u2 , selected
13 Note that other similarity functions may be used here that penalize differences in amplitude and
phase of the two spectra.
14 For instance, if one sensor is rotated by 90◦ along the z-axis, a rigid interface motion in the
x, y-plane will appear as a partly flexible motion. The effect will be visible as a low consistency
over the full frequency range.
144 4 Experimental Substructuring
and combined using w2 . Let us again define an original and force-filtered set of
responses, namely, for all forces f j ∈ f1 corresponding to virtual point 1:
'
y2,1 [ . . . y2, f j . . . ] = w2T Y21 (4.140a)
ỹ2,1 [ . . . ỹ2, f j . . . ] = w2T Y21 .F11 (4.140b)
Here, y2,1 is a row vector comprising the summed responses as a result of all singleton
impacts associated with the columns of Y21 . Row vector ỹ2,1 comprises a similar set
of responses, yet for the forces filtered by the force IDM matrix. This can be reasoned
as follows: for every single impact f j ∈ f1 , F11 looks for an equivalent combination
of impacts f1 that produces the same load on the virtual point. This allows to evaluate
if there is any difference in response between the direct application of a single impact
and the compound excitation suggested by the IDM matrix.
Overall Impact Consistency
Comparing the norm of the original and filtered set of responses, a measure is obtained
for the overall consistency of the impacts:
ỹ2,1
T
(ω)
ρf1,2 (ω) = . (4.141)
y2,1
T
(ω)
This frequency-dependent function quantifies how well the full set of impact forces
f1 can be represented by three translational forces and three moments (in case of a
6-DOF IDM matrix). If some excitations points have been positioned too far away
from the virtual point, they may load the interface in a flexible manner, which is
reflected by a low value. More often actually, the value drops because of position
and direction errors in the impacts.
Specific Impact Consistency
The specific impact consistency is used to spot individual impacts that are “non-
consistent” with respect to the transformation. Using the definitions of (4.138) and
(4.140a–4.140b), the following function can be defined:
ρ f j,2 (ω) = coh ỹ2, j (ω), y2, j (ω) f j ∈ f1 . (4.142)
This function quantifies the similarity in both amplitude and phase between the two
reciprocal virtual point FRFs. Note that for the diagonal of the matrix, i = j and
χii = 1 by definition. The reciprocity function is therefore only useful for evaluation
of the off-diagonal terms.
Driving Point Passivity
One of the lesser known properties of the FRF matrix is driving point passivity.
Driving point FRFs, i.e., the FRFs on the diagonal of the matrix, need to be minimum-
phase functions. This means that an applied force at DOF i shall always result in a
displacement in the same direction at that point.15 If the function is not minimum
phase, it implies that energy is added after the impulse, which contradicts system
passivity Sjövall and Abrahamsson (2007), Liljerehn and Abrahamsson (2014). As
a consequence, the phase of driving point FRFs shall be bounded as follows:
⎧
⎪
⎨∈ [−180, 0] for receptance FRFs
∠ Yii ∈ [−90, 90] for mobility FRFs (4.144)
⎪
⎩
∈ [ 0, 180] for accelerance FRFs
This property allows to assess the diagonal of the virtual point FRF matrix, which
are the only entries that must demonstrate minimum-phase responses. Conveniently,
those are also the only entries for which the reciprocity criterion does not apply.
Hence, the two criteria are complementary, in some practical way.
15 Signchanges only occur at a minimum distance of half the wavelength of the underlying modes.
This can be understood from modal synthesis.
146 4 Experimental Substructuring
This section discusses some more practical issues related to the virtual point exper-
iment. Proper positioning of the measurement and excitation points is essential to
obtain a high-quality model. Care should be taken to ensure that all six DOFs per
virtual point can be described independently. This has implications for both sensor
placement and impact positions. The following preparation and post-processing steps
are presented in the order as they normally appear in time.
Sensor Placement
The use of triaxial accelerometers has become standard practice in experimental
testing. A single accelerometer hosts three sensing devices in orthogonal directions,
which can be assumed to be collocated. Although one pair of those sensors measures
a total of six channels, it is normally not sufficient to describe all six DOFs of
the virtual point. One linear dependence will appear in the virtual point DOFs: the
triaxial sensors are unable to describe the rotation over the axis spanned between
both sensors, as illustrated in Fig. 4.27b. This is regardless of the position of the
sensors relative to the virtual point. Introduction of a third triaxial sensor, such that
the three sensors span a surface (Fig. 4.27c), enables the three sensors with a total of
nine DOFs to describe all six DOFs of the virtual point properly.
The additional benefit of the third triaxial sensor is the overdetermination of
the interface problem. Applying the virtual point transformation as discussed in
Sect. 4.6.2, the effects of uncorrelated measurement noise, as well as bias errors due
to misalignment of the sensors are reduced, which is generally considered as good
practice. The use of at least three triaxial sensors per virtual point can therefore be
held as rule of thumb, if the aim is to describe the virtual point by six independent
coordinates.
Finally, considering the rigid interface assumption, it seems logical to place the
sensors as close as possible to the virtual point. The smaller the distances are, the
lesser the effects of flexible interface motion compared to the rigid interface motion.
(a) Two sensors spanning a line: one dependency (b) Three sensors spanning a surface: all
exists between the rotational axes. rotations are fully determined.
On the other hand, for smaller distances, the virtual point transformation gets more
sensitive to absolute errors on the position. In general, one should approach this
aspect with some engineering judgment or a priori knowledge about the system and
measurement equipment.
Excitation Positions
Unlike the six accelerations measured by two triaxial sensors, six well-positioned
hammer impacts are potentially sufficient to fully determine the six generalized loads
of the virtual point. Still, for the same reason, it is preferable to use more excitation
points and overdetermine the force transformation. Similar to the sensor placement,
three impacts in each direction (x, y, z) can be used as minimum, creating nine
columns in the FRF matrix for each coupling point. However, as one is not restricted
by the available measurement equipment (as is the case for the simultaneous sensor
positions), it is advised to excite at much more points. Care should be taken to include
excitation directions that do not point straight to the virtual point, in order to generate
“moment” along the rotational axes.
In some studies, the faces of the acceleration sensors were suggested as possible
impact locations de Klerk (2009), van der Seijs et al. (2013). As the impact locations
and directions are equal to the locations and orientations of the measured responses
of the sensor, the same IDM matrix can potentially be used for both transforma-
tions. However, practice shows that FRFs obtained at the sensors’ faces exhibit poor
coherence, especially for the cross-directional FRFs of one sensor. Also, the sensor
is easily driven in overload. As the virtual point transformation does not require
physical driving point measurement (driving point FRFs are correctly rendered by
the two transformations), this type of excitation is discouraged.
With respect to the distance of impact points from the virtual point, a similar
reasoning applies as for the sensors. However, remind that the uncertainty of hammer
impacts is highly subject to the skill of the experimentalist.16 Position errors are
always made, typically in the order of a few millimeters. To minimize their effect
for the transformation, it is advised to excite at some more distance from the virtual
point. This is possible as long as the local rigidness assumption is still justified, which
is normally a matter of the targeted frequency bandwidth.
16 This only applies to manual impact hammer testing. Automated hammer or shaker measurements
probably reduce these uncertainties, but are far more elaborate in practice. Considering the sub-
stantial amount of measurement points for our applications, impact hammer measurement is the
preferred method for using the virtual point transformation.
148 4 Experimental Substructuring
In this section, nodal FRF models will be obtained from impact hammer measure-
ments on the two substructures as depicted in Fig. 4.28. Let us introduce them one
by one and briefly touch upon some design considerations:
• Substructure A is welded together from three pieces of solid aluminum
(30×30 mm). It forms an evenly sided triangle and loosely resembles the character
“A”, but was made asymmetric to avoid double resonance modes. It comprises a
number of 10 mm diameter holes, at the corner points and along the length of the
members, evenly spaced at distances of 75 mm. It hosts a vibration source (not
further discussed here) and can therefore represent the active source system in a
TPA or source characterization problem. The combined weight is circa 2.5 kg.
• Substructure B is constructed from two plates of stainless steel with a solid piece of
steel welded in between. The plates are produced using precise laser cutting. Five
holes are placed spanning a total distance of 300 mm, again with 75 mm spacing
in between. A honeycomb-like pattern of cuts was introduced to reduce weight, as
well as to provide a pattern to align sensors for an observability investigation. As
such, substructure B represents a receiving side into which the source vibrations
of substructure A may propagate. The total weight is circa 10 kg.
FRF Measurement
All substructure FRFs have been obtained by impact hammer testing. Figure 4.28
depicts how hammer impact points (red arrows) and triaxial accelerometers (gray
cubes) have been positioned and oriented on substructures A and B. Besides some
internal points, the main interests for both substructures are the three coupling points.
VP1
VP3
VP2
VP1
VP3 VP2
(a) Substructure A. The three coupling points (b) Substructure B. The three coupling points
are each instrumented by 3 tri-axial ac- are each instrumented by 3 tri-axial ac-
celerometers and 16 impact points. celerometers and 16 impact points; 2 addi-
tional sensors register target responses in the
structure.
Fig. 4.28 Acceleration sensors (indicated by gray cubes) and impact locations (red arrows) visu-
alized on the substructures
4.7 Real Applications 149
50%
0%
0 1000 2000 3000 4000 5000
Frequency (Hz)
(a) Overall sensor consistency of 9 sensor (b) Overall impact consistency of 16 (light
channels around VP2 with respect to excita- blue) and 13 (blue) impacts out of 16 around
tions around VP1. VP1 with respect to responses around VP2.
Each coupling point has been instrumented by three triaxial accelerometers of type
PCB 356B21. To determine forces and moments, 16 impact hammer positions are
chosen per coupling point. Altogether, this results in sufficient overdetermination of
the virtual point transformations.
FRF Consistency
In order to evaluate the above assumption on rigidity and obtain insight into the
contribution of single force impacts or displacements to the VP dynamics, several
consistency checks can be done. Let us illustrate the various consistency checks
for substructure A. Figure 4.29a shows the overall sensor consistency of VP2 for
excitations around VP1. This operation takes the accelerances of all nine sensor
channels (FRF matrix rows) around VP2 (u) for a combination of hammer impacts
(FRF matrix columns) around VP1, transforms these to the virtual point q, and
expands the accelerances back to their original sensor channels (ũ). The score of
100% over the full bandwidth of 5000 Hz indicates that all sensor channels are
perfectly consistent, i.e., ũ = u. This is obvious as the region between the three
sensors is very stiff; values below 100% would probably indicate incorrect placement
of a sensor.
Figure 4.29b shows the overall impact consistency for VP1 with respect to
responses around VP2. The light-blue area was computed for all 16 impact points,
which is clearly not optimal. Looking into the specific impact consistency for each
16 impacts, three impacts had significant lower score than average. By discarding
these 3 from the set of 16, the full 6-DOF set of virtual point forces/moments can
still be determined. The dark-blue area was computed for the optimized set, clearly
showing an improved overall impact consistency.
FRF Reciprocity
The VP transformation allows to validate reciprocity of the obtained virtual point
FRFs, as computed by Eq. (4.128). Note that this is possible as the VP displacements
(i.e., translational and rotational accelerations) are perfectly “vectorially associated”
with the corresponding VP loads (i.e., forces and moments). In other words, the
virtual point FRFs behave as if they were computed for nodes of an FE model.
Figure 4.30 shows two typical virtual point FRFs: response VP2Y over force
VP3Y of substructure A (left) and response VP2Z over force VP3Z of substructure
150 4 Experimental Substructuring
101 101
100 100
10-1 10-1
10-2 10-2
-3
10 10-3
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
180 180
Phase (deg)
Phase (deg)
90 90
0 0
-90 -90
-180 -180
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
Frequency (Hz) Frequency (Hz)
(a) Substructure A: Y-direction of VP2 to VP3 (b) Substructure B: Z-direction of VP2 to VP3
and its reciprocal. and its reciprocal.
Fig. 4.30 Reciprocity of the virtual point FRFs of the experimental models of substructures A and
B
B (right). The FRFs reciprocal FRFs are displayed in red. It can be observed that
reciprocity is indeed satisfied, especially up to 2 kHz.
Now that VP-transformed FRFs are available for substructures A and B, both struc-
tures are coupled using the LM-FBS algorithm. To do so, the substructure FRF
matrices of A and B are put in block diagonal form and an appropriate Boolean
matrix B is written as follows (not discussed here, see Sect. 2.3 for guidelines):
A
Y 0
Ỹ = Y − YBT BYBT −1 BY Y . (4.145)
0 YB
17 For an accelerance driving point FRF, the phase should be bounded by 0 and +180◦ .
4.7 Real Applications 151
DS coupling - Driving point FRF VP2Z / VP2Z DS coupling - Transfer FRF 10Z / 3Z
103 103
Accelerance ((m/s 2 )/N)
10-3
0 200 400 600 800 1000 1200 1400 1600 10-1
180
Phase (deg)
90
0 10-2
-90
-180 10-3
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
(a) Driving-point FRF for VP3 in Z-direction. (b) Transfer FRF for an internal impact point
on A to an internal acceleration response on B,
both in Z-direction.
substructured FRFs (blue) are compared the measured FRFs of the coupled structure
AB.
The first substructuring results, although not perfect yet, are by all means encour-
aging. It can be seen how resonance frequencies are created at roughly the right
frequencies. The phase around antiresonances is not fully stable; however, the over-
all amplitude of the FRFs matches quite well. Note that no filtering or processing has
been applied to the measured FRF data, except for transformation to virtual points.
with the finite element method. However, the modeling becomes much harder when
parts (substructures) are put together into an assembly by riveting, bolting, gluing,
or welding. It is often at the modeling of joints and component interfaces that the
modeler’s skill and experience come to a severe test. A vast part of the contribu-
tion to system damping is often attributed to the physical processes occurring in the
interfaces between parts. This poses one challenging problem in the dynamic sub-
structuring since it is at such interfaces the system is normally divided to make the
submodels more manageable. 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. The system damping is therefore most often obtained by testing. After a
model calibration, this damping can usually be mapped to the system model without
precise knowledge about the physics of the damping processes.
This section describes to some detail a frequency response-based calibration pro-
cess that has been found to work on large-scale finite element models and an industrial
example will be given. It starts with defining the concept of a model structure which
regards the parameterization of the calibration problem and some concepts related to
that. It continues with the imprecision of the estimation of the calibration parameters
that is the result of unavoidable noise in testing. Thereafter, it briefly treats the data
processing that leads to the data split that is used for validation with calibration and
cross-validation for assessment of model prediction accuracy. It finishes with the
specifics of calibration and cross-validation, and in particular treats some concepts
and tricks that have been found useful in the calibration of large-scale FE models
with many calibration parameters. However, it does not cover the important topics
of best preparing for an accurate and informative validation test, how to best conduct
the test or to best process the test data for possible inaccuracies and outliers.
Finite element model structure. A finite element model is often specified by an
input file with a given structure with data for the analysis. In that file, the system’s
physical properties are assigned numerical values. The constitutive models for the
used materials are given material property data. The structure’s geometry is given by
nodal point locations of the established discretization grid and the selected element
types. The modeler might have chosen lower dimension elements to represent all or
parts of the structure. Such may by simple scalar elements such as discrete springs
or lumped masses, it may be one-dimensional rod or beam elements, and it may
be two-dimensional plate or shell elements. Such simplistic elements also need to
be described in the input file as plate thicknesses, beam cross-sectional properties,
lumped mass weight, etc. All such numerical data are required to describe the struc-
tural model. The numerical data given can be seen as one specific realization from
a general parametric model structure given by the structure of the input file. Let us
call that model structure M(p) with parameter vector p. Two models M1 and M2
are meant to be equal, i.e., M1 (p1 ) = M2 (p2 ), if they share the same input-to-output
relation but not necessarily the same parameters. If the parameters of the model
parameter vector p were let free, they may be adjusted to the parameter setting that
gives a better fit to the input-to-output relation obtained as experimental data. If the
ideal situation would occur in which the experimental data were free from noise and
bias errors, an exact fit of a model of a proper model structure would be possible. If
4.8 Finite Element Model Updating for Substructuring 153
that situation would occur (in practice it never will), it is said that a consistent model
structure or a true model structure is at hand and the associated parameter setting
of the associated perfect model is denoted p . This parameter setting is sometimes
referred to as the oracle parameter setting.
An optimization scheme can be used to adjust the model parameters to best fit
the test data, that is, the process of model calibration. The calibrated model is thus
one specific realization from the given model structure. If it is found, then find,
eventually after unsuccessful calibration or validation attempts, that it was impossible
to calibrate or validate the model to an acceptable accuracy one needs to choose
another model structure or give up and accept defeat. A validation attempt that
resulted in a negative outcome can often be scrutinized to give some hints to what
model structure changes that are required. A change to another constitutive model
for the material constituents, such as changing from an isotropic material description
to an orthotropic, is an example of changing the model structure. Another possibility
might be to do a more precise modeling of parts that join structural components
together. There are endless other possibilities. However, all such model structure
changes have to be made with good scientific/engineering judgement to be truly
successful. Besides data of the structural model, the input file of the finite element
simulation code also includes a load model and a boundary condition model. The
process of validating the load model is a topic of its own and is not further treated here.
It is assumed that the loading in the associated calibration and validation experiments
are measured with such high precision resulting in load uncertainties that are too small
to matter.
However, the boundary condition model may require some considerations. The
boundary conditions for the structure under study might be different in the test labo-
ratory than in the real-world application for which the model should be made valid.
In commonly used finite element models, the structural boundaries are most often
considered to be perfectly free or perfectly rigid. These are conditions that are never
met in practice or in the test laboratory. Any nonideal boundary condition may cause
bias effects for the model calibration if not properly accounted for. An assessment
of the mobility of the structural boundary that is connected to the surrounding via
hardware may be required. The remaining surfaces of the structure that tend toward
open-air normally play an insignificant role to the observed structural behavior, but
might be of importance for very lightweight (meaning small weight in relation to the
enveloped volume) structures. For submerged structures, the wet surface boundary
conditions can have a great impact on the structural response and should be modeled
properly.
In the model calibration, the test data that are available are used in the estimation of
model parameters. These test data are always associated with some randomness. If an
154 4 Experimental Substructuring
i.e., the probability function depends on the parameter setting p that is believed
to represent real-world physics. But that parameter setting is hidden and implicitly
embedded in test data. The probability of the test outcome to be within the domain
Z (z ∈ Z) is then
,
P(z ∈ Z) = pdf(z 1 , z 2 , . . . , z N |p)dz 1 dz 2 . . . dz N . (4.147)
z∈Z
It is based on the joint probability density function (4.146) for the random observa-
tions z. The probability that the observed realization indeed should take value z∗ is
thus proportional to jpdf(z 1∗ , z 2∗ , . . . , z ∗N |p). 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
where the maximization is performed for the processed test data z∗ , which is one set
of data of the infinitely many realizations possible. The function pML is known as
the maximum likelihood estimator (MLE) for p.
The Cramer–Rao Inequality. The credibility of an estimator can be assessed
by its mean-square error matrix
Here, p0 denotes the true (but hidden) value of p, and (4.149) is evaluated under
the assumption that the joint probability density of z is pdf(z|p0 ).
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 the
following.
where
Theorem 5.2 Suppose that the random data p ∈ R N are statistically independent
so that
-
N
jnf(z 1 , z 2 , . . . , z N |p) = jnf(z n |p). (4.152)
n=1
Suppose also that, in particular, the distribution of z for a given setting of the param-
eters p = p0 is given by jnf(z|p0 ). Then the random parameter variable p M L (z)
tends to p0 with probability 1 as N tends to infinity, and the random variable
√
N (p M L − p0 ) converges in distribution to the normal distribution with zero mean
and covariance matrix given as the Cramér–Rao lower bound, i.e., the covariance
matrix is the inverse of the Fisher information matrix.
Thus, when the number of data N tends to infinity, the MLE p M L are distributed
N (p0 , F−1 ). According to the Cramer–Rao theorem, this is the best an estimator can
do and therefore it is said that the MLE is an efficient estimator.
The Case of Normally Distributed Data. Consider the special case when the data
z nX , as obtained from evaluation of test data (eXperiment), can be predicted without
bias with a parameterized model in its calibrated setting p∗ . Let the predicted data
from the model be z̆ n . The relation
z nX = z̆ n (p∗ ) + εn (4.153)
introduces εn being the residual that cannot be explained by the model. Assume that
the residuals are statistically independent variables distributed εn ∈ N (0, σn2 ) with
known standard deviation σn . In that case, the Fisher information matrix, and thus
the parameter covariance lower bound, can be shown to be
N
∂ z̆ n (p) ∂ z̆ n (p) T
F(p) = σn−2 [ ][ ] . (4.154)
n=1
∂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 σn2 . The identifiability
of the parameters of a given model can thus be evaluated provided assumptions
on residual variance. Different model structures can thus be compared against each
other to find out which gives the best parameter identifiability. For a derivation of
Eq. (4.154), see Ljung (1998).
4.8 Finite Element Model Updating for Substructuring 157
The expectation refers to the mean value obtained after repeated sampling of
same experimental data entity z nX . Closely related to the P D is the residual squared
deviation (RS D) of the prediction defined as
where Nz is the number of data, i.e., the length of the data vectors z X and z̆. The
RS D may be evaluated separately for validation data and calibration data as
and
RS Dδ = δ T (p)δδ (p)/Nδ , (4.158)
where usually RS Dγ is larger than RS Dδ since the model has been calibrated with
an effort to minimize δ T δ . To truly evaluate the P Di would require experimental
data that are fully independent on calibration data and thus require repeated tests.
Usually, however, test repetitions are not always available for reasons of cost or time.
To circumvent that and give a reasonable substitute of the true prediction deviation,
XV 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 describes one possible strategy to do the data splitting.
K-Fold Cross-Validation. In K-fold XV, the available data set is split into K parts
(folds) of roughly the same size. Let k = 1, 2, . . . , K be the index of the k:th such
fold with the Nz data values divided into Nγ k validation data and Nδk = Nz − Nγ k
calibration data. The union of the two data sets thus constitutes the entire data set. The
validation sets are mutually unique and the union of them spans the entire data set.
The validation deviation of the kth set is RS Dγ k and the corresponding calibration
deviation is RS Dδk .
For the XV, the calibration is made K times using the deviation metric Q k (p) =
δ kT δ k to obtain a parameter estimate associated with the kth split as
1
K
RS DXV = RS Dγ k . (4.161)
K k=1
Fig. 4.32 Normalized deviation metric versus parameter variation from nominal of a stiffness
parameter k and a mass parameter m for (upper) three system damping levels at 0.01, 0.1, and 1%,
and (lower) various frequency sampling rates in number of samples per half-bandwidth (p1/2bw)
4.8 Finite Element Model Updating for Substructuring 161
model using first principles, it is most often assigned a simple representation for mod-
eling 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 suf-
ficient for its intended purpose. In case of modal damping modeling, the model’s
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 similar systems. The modal damping found in exper-
iments is normally used for FE simulation without further attempts to understand
their physical background. Physically based parameterization of damping phenom-
ena 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 map-
ping of modal damping relies on mode shape pairing, meaning that the same amount
of modal damping should be assigned to modes that are similar with respect to their
deformation pattern.
Mode pairing of experimentally found modes and FEM modes are usually made
through correlation analysis using MAC correlation quantification. Such MAC-based
pairing is normally not straightforward, especially for systems with high modal den-
sity and with sparsely distributed experimental sensor layout. Eigenmode pairing for
the purpose of damping mapping would be unnecessary if the modal damping was the
same for all modes. To overcome the problem of mode pairing, a method of damping
equalization has therefore been suggested, see Abrahamsson and Kammer (2015).
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.
Using efficient system identification methods, such as the state-space subspace
method N4SID, these experimental data can be used to obtain a mathematical state-
X
space model. The experimentally found system transfer function Hraw can then be
represented by the identified system
such that
H X = C(iωI − A)−1 B + D. (4.164)
with Σ = diag(σi ) in which σi are the complex-valued system poles as given by the
experimental data. The relative modal damping ζi , 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 ζ0 , i.e.,
ζi = ζ0 . (4.167)
The effect of such damping equalization is that the oscillatory imaginary part of
the poles is preserved and the real damping part is modified such that the perturbed
system poles are now
σ̃i = −ζ0 |Re(σi )| + iIm(σi ), (4.168)
with
à = diag(σ̃i ). (4.170)
This in turn gives 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
X
procedure on the raw test data Hraw has led to a mathematical model which can
be evaluated for any frequency ω. In particular, it means that equal log-frequency
increments can be used for transfer function evaluation. In addition to that, also
fictitious modifications of the system under test can be made. A particular use of such
modification is that the system damping level can be adjusted, 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 toward 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, the
viscous damping matrix V is (see Craig 2000)
with eigenfrequencies ωi , modal masses m i , and the modal matrix X given by the
undamped system’s eigenvalue problem
4.8 Finite Element Model Updating for Substructuring 163
M
H(p) = J(p)T J(p) + δi (p)Hi (p) ≡ J(p)T J(p) + S. (4.175)
i=1
From (4.175), we observe that the Hessian of the least squares objective function
consists of a special combination of first- and second-order information. Least squares
methods are typically based on the premise that eventually, as the iterative search for
the minimum go on, the first-order term J(p)T J(p) of (4.175) will dominate over the
second-order term S as the deviations δi shrink. This assumption is not justified when
the residual at the solution Q(p) is large, i.e., in the order of the largest eigenvalue of
J(p)T J(p) or greater. In such a case, which might be due to a poor model structure or
164 4 Experimental Substructuring
noisy test data, one might as well use a general unconstrained minimization method.
However, for many calibration problems, the residual is indeed small enough to
justify the use a special method that uses the smallness of S(p) to the advantage.
The Newton Method. Let pk be the current estimate of the solution at the kth
algorithmic iteration step. Any quantity subscripted by k will indicate that it is evalu-
ated at this step. A Taylor series expansion of to second order about the current point
pk with a perturbation q gives
1
Q(pk + q) ≈ Q k + JkT q + qT Hk q. (4.176)
2
The search is for the optimizing step q∗ that minimizes Q, i.e., p∗ = pk + q∗ . The
minimum will be achieved if q∗ is a minimizer of the quadratic function
1
Φ(q) = JkT q + qT Hk q, (4.177)
2
where q∗k satisfies the linear system
Let qk denote the solution of (4.178), also called the Newton direction qN . The
kth step minimization of the function is then made by a one-dimensional search for
the minimum along p = pk−1 + λpN for the search factor λ. The minimum at pk−1 +
λ∗ pN is then given by λ∗ = argmin Q(p(λ)). If Q k tends to zero as pk approaches
the minimizing solution, the matrix Sk also tends to zero. Thus, the Newton direction
can be approximated by the solution of the equation
Note that Eq. (4.180) involves only the first derivatives of δ . The solution of (4.180)
is the solution of the least squares problem
and is unique if Jk has full rank. The vector that solves (4.181) is called the Gauss–
Newton direction and will be denoted qGN . A method in which this vector is used as
a search direction is known as a Gauss–Newton method.
If Jk is of full rank, the Gauss–Newton direction approaches the Newton direction
as Sk tends to zero in the sense that if ||Sk || < ε for a sufficiently small positive scalar
ε, then
||qN − qGN ||/||qN || = O(ε). (4.182)
Consequently, if ||δδ (p∗ )|| is zero and the columns of J(p∗ ) are linearly indepen-
dent, the Gauss–Newton method can ultimately achieve a quadratic rate of conver-
gence, despite the fact that only first derivatives are used to compute qGN .
In implementations of the Gauss–Newton method, great care is taken to estimate
the rank of Jk . It is seen in Eq. (4.180) that a rank-deficient Jk cause a singular equa-
tion system, and a close-to rank-deficient matrix Jk 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 problems 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 robus-
tification is normally made by involving the singular value decomposition in the
solution of (4.180) in which the determination of the rank of Jk plays an important
role in estimation good 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 Jk is nearly rank-deficient, a generous estimate of the rank
tends to cause very large elements in the solution of (4.180) 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 fea-
ture in the parameter estimation process in which the solution should preferably stay
at the nominal configuration for insensitive parameters. This has motivated the intro-
duction of parameter regularization as is done in the Levenberg–Marquardt method
that is described next.
The Levenberg–Marquardt Method. A good alternative to the Gauss–Newton
method for 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 p0 into
where the regularizing parameter κk is a nonnegative real scalar. It can be noted that it
differs from the calculation of the Gauss–Newton search direction of (4.180) in that
the approximation to the Hessian is augmented with the Hessian of the regularizing
term in (4.183).
A unit step is almost always taken along qk , i.e., pk+1 is given by pk+1 = pk + qk .
It can be shown that, for some scalar κ0 related to κk , the vector qk is the solution of
the constrained subproblem
State-space system identification (SSSI) methods are usually split in two categories
out of which one is the stochastic SSSI category (a.k.a. output-only system identifi-
cation) relating to a sub-calls of state-space models of the full state-space model of
Eq. (3.82). For this subclass, one has
x = Ax + Ew (4.186)
r = Cx + v
x = Ax + Bs (4.187)
r = Cx + Ds
for which both the system stimuli s and response r are considered being fully known
test data. For the deterministic SSSI category, the system transfer function Hx (ω) can
be estimated from experimental data. Central to the two categories is the estimation
of the system’s observability matrix
⎡ ⎤
C
⎢ CA⎥
⎢ ⎥
⎢ ⎥
CA2 C O
O =⎢
x
⎥≡ ≡ . (4.188)
⎢ ⎥
.. O CAn x −1
⎣ ⎦
.
CAn x −1
168 4 Experimental Substructuring
Since the true model order n s of the tested system is hidden (for a continuous sys-
tem, the order is infinite), a system identification user needs to specify the model order
n x that he believes best capture the model as seen through test data. Discrete-time
data for the SSSI are from the n t steps of the stimuli sequences (sk , k = 1, 2, . . . , n t )
and response sequences (rk , k = 1, 2, . . . , n t ). 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 van Overschee and de Moor (1996). For a given
Ox , an estimation of the system’s output matrix C is given by the upper n r rows of
Ox . The shift property O = OA is exploited to give an estimation of the system’s
dynamic matrix A as
†
A = O O, (4.189)
where ()† denotes the Moore–Penrose pseudo-inverse. For a given couple {A, C},
one notes that the system transfer function (3.85) is linear in {B, D} and estimates
{B∗ , D∗ } for these can be obtained by the least squares solution from data given at
discrete frequencies ωk , k = 1, 2, . . . , K as
⎡ ⎤
C(iω1 I − A)−1 B + D − Hx (ω1 )
⎢ C(iω2 I − A)−1 B + D − Hx (ω2 ) ⎥
⎢ ⎥ 2
{B∗ , D∗ } = arg min || ⎢ .. ⎥ ||2 . (4.190)
{B,D}|{A,C} ⎣ . ⎦
C(iω K I − A)−1 B + D − Hx (ω K )
Since, for a given set {A, B}, the system transfer function (3.85) is linear in {C, D}
a better estimation of these {C∗ , D∗ } can be obtained by solving the complementary
least squares problem
⎡ ⎤
C(iω1 I − A)−1 B + D − Hx (ω1 )
⎢ C(iω2 I − A)−1 B + D − Hx (ω2 ) ⎥
⎢ ⎥ 2
{C∗ , D∗ } = arg min || ⎢ .. ⎥ ||2 . (4.191)
{C,D}|{A,B} ⎣ . ⎦
C(iω K I − A)−1 B + D − Hx (ω K )
The procedure using Eqs. (4.190) and (4.191) in an iterative fashion usually ren-
ders a quick convergence to a good model {A, B, C, D} = {A, B∗ , C∗ , D∗ }.
A second-order structural dynamics finite element model as that given in Eq. (3.103)
is physically sound. Its bases are well-known first principles and well-established
discretization techniques. The corresponding state-space companion can therefore
also be considered as physically sound.
4.9 Substructuring with State-Space Models 169
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, nonideal sensors, and others. Such errors
may lead to state-space model behavior that is not motivated by the physics of the
tested real-world system. Typical such errors are related to stimuli measurements.
Vibration testing is 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 toward 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
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
follows.
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 con-
verges 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 nei-
ther asymptotically stable nor unstable. An undamped structural dynamics system
Mü + (K + λK g )u = 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 Mü + Vu̇ + Ku = F with a symmetric positive definite viscous
damping matrix V is asymptotically stable. An identified state-space model, stem-
ming 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 bring 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 passivity relates to the flow of energy over a system bound-
ary. The energy conservation principle stipulates that
170 4 Experimental Substructuring
with E s being the system’s internal energy and E in and E out 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 encap-
sules the inside (the system) from its outside. For a system to be long time passive in
the mechanical sense (LTP) the outside energy influx E in 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
E s (t + nT ) + E m (t + nT ) − E out (t + nT ) =
E s (t + nT + T ) + E m (t + nT ) + ΔE m (T ) − E out (t + nT ) − ΔE out (T )
(4.193)
with E m = E in being the energy provided by the mechanical input. Over one full
cycle T , the internal energy has thus been changed the amount of ΔE m (T ) −
ΔE out (T ). For the LTP system, the energy is bounded 0 ≤ E s (0) < ∞ and thus
it is required for long time stationarity that this amount of change needs to be posi-
tive and thus
ΔE m (T ) > ΔE o (T ) ≥ 0. (4.194)
H
where H is the system transfer function for force-to-velocity. The real part of F̂ HF̂
evaluates to
which is seen to be positive for all positive definite Re{H}. For an 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
4.9 Substructuring with State-Space Models 171
C(iωI + A)−1 B is positive definite, the passivity constraint is related to the positive
real (PR) lemma, see Anderson (1967). 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
holds. It has been experienced to be far from trivial to find a solution to (4.196) and
(4.197) 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. The general problem is hard to solve since it
requires the search for a solution {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 differ-
ential 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 Hr (ω), is then to make
averaging of non-diagonal elements of the transfer function matrix as
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 ra = Ca x + Da s with Da = 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 ẋ = Ax + Bs. The output
equation for displacement output thus must read rd = Cd x. Since the velocity output
is related to displacement output as rv = ṙd = Cd (Ax + Bs) = Cd Ax + Cd Bs, it
must also hold that Cd B = 0. With Cd B = 0, one also notes that the acceleration
output may be written ra = ṙv = Cd A(Ax + Bs) = Cd A2 x + Cd ABs which leads
to the conclusion that Cd AB is not necessarily zero.
172 4 Experimental Substructuring
For accelerance data H x , the procedure to improve the estimates of the pair {B, C}
described in Sect. 4.9.1 can be modified to take account for such physically motivated
constraint by iteratively improve the estimates of first B and then Cd in a augmented
least squares sense in the sequence
⎡ ⎤
Cd A2 (iω1 I − A)−1 B + Cd AB − Hx (ω1 )
⎢Cd A2 (iω2 I − A)−1 B + Cd AB − Hx (ω2 )⎥
⎢ ⎥
⎢ .. ⎥ 2
B∗ = arg min|| ⎢ . ⎥ ||2 (4.200)
B|{A,Cd } ⎢ ⎥
⎣Cd A2 (iωk I − A)−1 B + Cd AB − Hx (ωk )⎦
Cd B
and
⎡ ⎤
Cd A2 (iω1 I − A)−1 B + Cd AB − Hx (ω1 )
⎢Cd A2 (iω2 I − A)−1 B + Cd AB − Hx (ω2 )⎥
⎢ ⎥
⎢ .. ⎥ 2
C∗d = arg min|| ⎢ . ⎥ ||2 . (4.201)
Cd |{A,B} ⎢ ⎥
⎣Cd A2 (iωk I − A)−1 B + Cd AB − Hx (ωk )⎦
Cd B
and
⎡ ⎤
Cd A(iω1 I − A)−1 B + Cv B − Hx (ω1 )
⎢Cd A(iω2 I − A)−1 B + Cv B − Hx (ω2 )⎥
⎢ ⎥
⎢ .. ⎥
∗ ⎢ . ⎥ 2
Cd = arg min|| ⎢ ⎥ ||2 . (4.203)
Cd |{A,B} ⎢Cd A(iωk I − A)−1 B + Cv B − H x (ωk )⎥
⎢ ⎥
⎣ Cv A−1 B ⎦
Cv B
4.9 Substructuring with State-Space Models 173
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 ∈ n c . To the displacement, the vectorially associated (energy-conjugate)
stimuli vectors si ∈ n c that act on the two components are s I and s I I . For dis-
placement output, a proper state-space representation of the two models i = I, I I
is
ẋi = Ai xi + Bi si (4.204)
r=Cx. i i
Proof With the response r = Cx + Ds, the response rate is ṙ = Cẋ + Dṡ =
CAx + CBs + Dṡ and thus the instantaneous power is P = sT (CAx + CBs) +
sT Dṡ. To be independent on the stimuli rate ṡ, it requires that sT 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, B, C} is given by the next lemma.
Proof For a second-order system, one possible state-space realization pair is, see
Eq. (3.110),
0
B= −1 , C = Pd I 0 ,
M Ps
from which it follows that, also after an arbitrary similarity transformation x = Tz,
the relation of the pair {T−1 B, CT} is such that
−1
0
(CT)(T B) = Pd I 0 = 0,
M−1 Ps
174 4 Experimental Substructuring
which proofs part (a). With the structure of A given by Eq. (3.110), it can eas-
ily be verified that CAB = Pd M−1 Ps . Since M is full rank thus CAB = 0, unless
under the trivial condition that Pd = 0 and/or Ps = 0, which concludes the proof
of (b).
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 s̄. The interface of the
two components is then loaded by the total external stimuli s = α I s + α I I s (thus
α I + α I I = 1). For the interaction forces β I s̄ and β I I s̄, it holds that β I = −β I I
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
⎢ ⎥ ⎢ ⎥
⎛ ⎞ ⎡ ⎤ ⎡ ⎤
ṙ 0 I 0 0
⎜ −I ⎟ ⎢ −I I −I I ⎥ I ⎢ ⎥
⎝Bvv r̈⎠ = ⎣Bvv Avd B−I
vv Avv Bvv Avb ⎦ x + ⎣α I s + β I s̄⎦
I
(4.207)
ẋbI AbdI
AbvI
AbbI
0
r = I 0 0 xI
and
⎛ ⎞ ⎡ ⎤ ⎡ ⎤
ṙ 0 I 0 0
⎜ −I I ⎟ ⎢ −I I I I −I I I I −I I I I ⎥ I I ⎢ ⎥
⎝Bvv r̈⎠ = ⎣Bvv Avd Bvv Avv Bvv Avb ⎦ x + ⎣α I I s + β I I s̄⎦ (4.208)
ẋbI I II
Abd II
Abv AbbII
0
II
r= I00 x .
Adding the two systems together results in the coupled system equations with the
expanded state vector x̄ T = [r ṙ xbI xbI I ] as
4.9 Substructuring with State-Space Models 175
⎛ ⎞
ṙ
⎜(B−I + B−I I )r̈⎟
⎜ vv vv ⎟
⎜ ⎟= (4.209)
⎝ xbI ⎠
xbI I
⎡ ⎤
0 I 0 0
⎢B−I A I + B−I I A I I B−I −I I I I −I I −I I I I ⎥
vv Avv + Bvv Avv Bvv Avb Bvv Avb ⎥
I
⎢ vv vd vv vd
⎢ ⎥ x̄
⎣ I
Abd I
Abv I
Abb 0 ⎦
II II II
Abd Abv 0 Abb
⎡ ⎤
0
⎢(α + α )s + (β + β )s̄⎥
⎢ I II II II ⎥
+⎢ ⎥
⎣ 0 ⎦
0
r = I 0 0 0 x̄.
Āvd = Γ [B−I −I I I I
vv Avd + Bvv Avd ]
I
(4.211)
Āvv = Γ [Bvv Avv + B−I
−I I
vv Avv ]
I II
(4.212)
I
Āvb = Γ B−I I
vv Avb (4.213)
II
Āvb = Γ B−I I II
vv Avb (4.214)
B̄vv = Γ. (4.215)
to a system with state-space triple {Ā, B̄, C̄} on the coupling form according to Eq.
(4.205) provided that [C; CA] is of full row rank 2n c and that n c ≤ n.
Proof Let T be a transformation for {A, B, C} with inverse Z be such that
⎡ ⎤
C
The nullspace NC is of full row rank n − 2n c and thus the combined nullspace
N is also of full row rank n − 2n c because B is not in the nullspace of [C; CA]
since CAB = 0. Since both [C; CA] and N are full row rank and N is a nullspace of
[C; CA], it implies that T is full rank and thus nonsingular. Using that TZ = I, one
has ⎡ ⎤
CZ1 CZ2 CZ3
⎣CAZ1 CAZ2 CAZ3 ⎦ = I,
NZ1 NZ2 NZ3
which, in particular, imply that CAZ1 = 0, CAZ2 = I, and CAZ3 = 0. The first
block row of the system matrix on coupling form Ā = TAT−1 = TAZ is thus
which concludes the proof of the coupling form structure of Ā. The structure of B̄ is
given by
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
C CB 0 0
B̄ = TB = ⎣CA⎦ B = ⎣ CAB ⎦ = ⎣CAB⎦ ≡ ⎣B̄vv ⎦ ,
N NC NTB (N B NTB )−1 N B B 0 0
since CB = 0 and N B B = 0 which thus proves the coupling form structure of B̄. The
structure of C̄ on the other hand is given from the definition of Z by
References
Abrahamsson, T., & Kammer, D. C. (2015). Finite element model calibration using frequency
responses with damping equalization. Mechanical Systems and Signal Processing, 62–63, 218–
234.
Allen, M. S., & Mayes, R. L. (2007). Comparison of FRF and modal methods for combining
experimental and analytical substructures. In 25th international modal analysis conference (IMAC
XXV).
Allen, M. S., & Mayes, R. L. (2007). Comparison of frf and modal methods for combining experi-
mental and analytical substructures.
Allen, M. S., Kammer, D. C., & Mayes, R. L. (2010). Uncertainty in experimental/analytical sub-
structuring predictions: A review with illustrative examples. In Proceedings of the international
conference on noise and vibration engineering (ISMA) (pp. 1833–1850). Belgium: Leuven.
Allen, M. S., Mayes, R. L., & Bergman, E. J. (2010). Experimental modal substructuring to couple
and uncouple substructures with flexible fixtures and multi-point connections. Journal of Sound
and Vibration, 329(23), 4891–4906.
Anderson, B. D. O. (1967). A system theory criterion for positive real matrices. SIAM Journal on
Control and Optimization, 5(2), 171–182.
Avitabile, P. (2003). Twenty years of structural dynamic modification – a review. Sound & Vibration
MagazineJanuary, 14–25.
Beck, J. L., & Yuen, K. V. (2004). Model selection using response measurements: Bayesian prob-
abilistic approach. Journal of Engineering Mechanics, 130, 192–203. https://doi.org/10.1061/
(ASCE)0733-9399.
Bello, M., Sestieri, A., D’Ambrogio, W., & La Gala, F. (2003). Development of a rotation transducer
based on bimorph PZTs. Mechanical Systems & Signal Processing, 17(5), 1069–1081. https://
doi.org/10.1006/mssp.2003.1607.
Bendat, J. S., & Piersol, A. G. (1980). Engineering Applications of Correlation and Spectral Anal-
ysis. New York: Wiley.
Bregant, L., & Sandersons, M. (2000). Rotational degrees of freedom: an historical overview on
techniques and methods. In Proceedings of the international conference on noise and vibration
engineering (ISMA) (pp. 973–980). Belgium: Leuven.
Carne, T. G., & Dohrmann, C. R. (2006). Improving experimental frequency response function
matrices for admittance modeling.
Cloutier, D., & Avitabile, P. (2011). Dynamic uncoupling of a system model for component iden-
tification. In IMAC-XXIX: International Modal Analysis Conference, Jacksonville, FL. Bethel,
CT: Society for Experimental Mechanics.
Craig, R. R. (2000). Coupling of substructures for dynamic analyses: An overview. In Proceedings
of AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics, and materials conference and
exhibit (pp. 1573–1584).
Crowley, J. R., Klosterman, A. L., Rocklin, G. T., & Vold, H. (1984). Direct structural modification
using frequency response functions. Proceedings of IMAC II.
D’Ambrogio, W., & Fregalent, A. (2009). Decoupling procedures in the general framework of fre-
quency based substructuring. In IMAC-XXVII: International modal analysis conference, Orlando,
FL. Bethel, CT: Society for Experimental Mechanics.
D’Ambrogio, W., & Fregalent, A. (2014). Are rotational dofs essential in substructure decoupling?
In IMAC-XXXII: International modal analysis conference, Orlando, FL. Bethel, CT: Society for
Experimental Mechanics.
D’Ambrogio, W., & Fregolent, A. (2015). Ignoring rotational dofs in decoupling structures con-
nected through flexotorsional joints. In IMAC-XXXIII: International modal analysis conference,
Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
de Klerk, D. (2009). Dynamic response characterization of complex systems through operational
identification and dynamic substructuring. Ph.D. thesis, Delft University of Technology, The
Netherlands. http://resolver.tudelft.nl/uuid:2e4dd63e-68d5-41a4-8201-1e784f2f752c.
178 4 Experimental Substructuring
de Klerk, D., & Voormeeren, S. (2008). Uncertainty propagation in experimental dynamic substruc-
turing. In Proceedings of the XXVI international modal analysis conference (IMAC), Orlando,
FL. Bethel, CT: Society for Experimental Mechanics.
de Klerk, D., Rixen, D. J., de Jong, J. (2006). The frequency based substructuring (FBS) method
reformulated according to the dual domain decomposition method. In International modal anal-
ysis conference, IMAC-XXV. St Louis, MO: SEM.
de Klerk, D., Rixen, D., Voormeeren, S., & Pasteuning, F.: Solving the RDoF problem in experimen-
tal dynamic substructuring. In: Proceedings of the XXVI international modal analysis conference
(IMAC), Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Duarte, M. L. M., & Ewins, D. J. (2000). Rotational degrees of freedom for structural coupling
analysis via finite-difference technique with residual compensation. Mechanical Systems & Signal
Processing, 14(2), 205–227. https://doi.org/10.1006/mssp.1999.1241.
Elliott, A. S., Moorhouse, A. T., & Pavić, G. (2012). Moment excitation and the measurement of
moment mobilities. Journal of Sound & Vibration, 331(11), 2499–2519. https://doi.org/10.1016/
j.jsv.2012.01.022.
Epp, D. S., Ozdoganlar, O. B., Chaplya, P. M., Hansche, B. D., & Carne, T. G. (2004). A base
excitation test facility for dynamic testing of microsystems. In IMAC-XXII: International modal
analysis conference, Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Ewins, D. J., & Inman, D. J. (2001). Structural dynamics @ 2000: current status and future direc-
tions. Baldock, Hertfordshire, England: Research Studies Press.
Géradin, M., & Rixen, D. (2015). Mechanical vibrations (3rd ed.). Theory and application to
structural dynamics. Chichester: Wiley.
Géradin, M., & Rixen, D. J. (2017). Impulse-based substructuring in a floating frame to simulate high
frequency dynamics in flexible multibody dynamics. Multibody System Dynamics (pp. 1–31).
Gialamas, T., Tsahalis, D., Otte, D., Van der Auwaraer, H., & Manolas, D. (2001). Substructuring
technique: improvement by means of singular value decomposition (svd). Applied Acoustics,
62(10), 1211–1219.
Gordis, J. H. (1995). Integral equation formulation for transient structural synthesis. AIAA Journal,
33(2), 320–324.
Groult, B. (2008). Extension d’une mèthode de modification structurale pour la conception de
dispositifs dissipatifs intègrant des matèriaux viscoèlastiques. Ph.D. thesis, Ecole Centrale Paris,
Laboratoire de MÈcanique des Sols, Structures et MatÈriaux.
Haeussler, M., Klaassen, S., & Rixen, D. (2018). Comparison of substructuring techniques for
experimental identification of rubber isolators dynamic properties. In P. S. et al. (ed.) International
conference on noise and vibration engineering, ISMA. Belgium: KUL, Leuven.
Harvie, J., & Avitabile, P. (2013). Effects of precise FRF measurements for frequency based sub-
structuring. In Topics in experimental dynamic substructuring, Proceedings of the 31st IMAC, a
conference on structural dynamics (vol. 2, Chap. 22, pp. 277–285). New York: Springer. https://
doi.org/10.1007/978-1-4614-6540-9_22.
Helderweirt, S., Van der Auweraer, H., Mas, P., Bregant, L., & Casagrande, D. (2001). Application of
accelerometer-based rotational degree of freedom measurements for engine subframe modelling.
In Proceedings of the XIX international modal analysis conference (IMAC), St. Louis, MO (vol. 2,
pp. 1298–1304). Bethel, CT: Society for Experimental Mechanics.
Höller, C., & Gibbs, B. M. (2015). Indirect determination of the mobility of structure-borne sound
sources. Journal of Sound and Vibration, 344, 38–58.
Imregun, M., Robb, D., & Ewins, D. (1987). Structural modification and coupling dynamic analysis
using measured FRF data. In IMAC-V: international modal analysis conference, London, England
(pp. 1136–1141). Bethel, CT: Society for Experimental Mechanics.
Ind, P., & Ewins, D. (2003). Impedance based decoupling and its application to indirect modal
testing and component measurement: A numerical investigation. In IMAC-XXI: International
modal analysis conference, Kissimmee, FL. Bethel, CT: Society for Experimental Mechanics
Jetmundsen, B., Bielawa, R. L., & Flannelly, W. G. (1988). Generalized frequency domain sub-
structure synthesis. Journal of the American Helicopter Society, 33(1), 55–64.
References 179
Kalaycıoğlu, T., & Özgüven, H. N. (2018). Experimental verification of a recently developed frf
decoupling method for nonlinear systems. In IMAC-XXXVI: International modal analysis con-
ference, Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Keersmaekers, L., Mertens, L., Penne, R., Guillaume, P., & Steenackers, G. (2015). Decoupling of
mechanical systems based on in-situ frequency response functions: The link-preserving, decou-
pling method. Mechanical Systems and Signal Processing.
Klerk, D. D., Rixen, D. J., & Voormeeren, S. N. (2008). General framework for dynamic substruc-
turing: History, review and classification of techniques. AIAA Journal, 46(5), 1169–1181.
Klosterman, A. L. (1971). A combined experimental and analytical procedure for improving auto-
motive system dynamics. Ph.D. thesis, University of Cincinnati, Department of Mechanical Engi-
neering.
Liljerehn, A., & Abrahamsson, T. (2014). Dynamic sub-structuring with passive state-space compo-
nents. In Proceedings of the international conference on noise and vibration engineering (ISMA)
(pp. 3879–3890). Belgium: Leuven.
Lim, T.C., & Steyer, G. C. (1992). An improved numerical procedure for the coupling of dynamic
components using frequency response functions. In: IMAC-IX: international modal analysis
conference, Florence, Italy (pp. 355–355). SEM Society for Experimental Mechanics INC.
Liu, W., & Ewins, D. (1999). The importance assessment of RDOF in FRF coupling analysis. In
Proceedings of the XVII international modal analysis conference (IMAC), Kissimmee, FL (pp.
1481–1487). Bethel, CT: Society for Experimental Mechanics.
Liu, J., Li, M., Qin, L., & Liu, J. (2013). Principle research on a single mass piezoelectric six-degrees-
of-freedom accelerometer. Sensors, 13(8), 10844–10855. https://doi.org/10.3390/s130810844.
Ljung, L. (1998). System identification: theory for the user. Pearson Education.
Mahmoudi, A. E., Meyer, C. H., & Rixen, D. J. (2019). Comparison of different approaches to
include rubber bushings into frequency-based substructuring coupling process. Experimental
Techniques submitted.
Martinez, D.R., Carne, T.G., Gregory, D.L., Miller, A.K.: Combined experimental/analytical mod-
eling using component mode synthesis (1984). Compilation and indexing terms, Copyright 2007
Elsevier Inc. All rights reserved 84060103006 COMPUTED MODAL FREQUENCY ACCU-
RACY MODE SHAPES SIMPLE BEAM STRUCTURE FREE-FREE MODES EXPERIMEN-
TAL MODAL DATA BASE FINITE ELEMENT MODEL
Martinez, D. R., Miller, A. K. & Carne, T. G. (1985). Combined experimental and analytical
modeling of shell/payload structures.
Mayes, R. L., & Arviso, M. (2010). Design studies for the transmission simulator method of exper-
imental dynamic substructuring. In International seminar on modal analysis, ISMA. Belgium:
KUL, Leuven.
Mayes, R. L., & Stasiunas, E. C. (2007). Lightly damped experimental substructures for combining
with analytical substructures.
Mayes, R. L., Allen, M. S., & Kammer, D. C. (2012). Eliminating indefinite mass matrices with the
transmission simulator method of substructuring.
Mayes, R., & Ross, M. (2012). Advancements in hybrid dynamic models combining experimental
and finite element substructures. Mechanical Systems and Signal Processing, 31, 56–66.
Meggitt, J., & Moorhouse, A. (2018). In-situ dynamic sub-structure decoupling of resiliently coupled
assemblies. Mechanical Systems and Signal Processing. In press (2018).
Meggitt, J. W. R., Elliott, A. S., Moorhouse, A. T., & Lai, H. K. (2015). In situ determination of
dynamic stiffness for resilient elements. In Proceedings of the institution of mechanical engineers,
Part C: Journal of mechanical engineering science (p. 0954406215618986).
Meggitt, J., Moorhousea, A. T., & Elliott, A. S. (2018). On the problem of describing the coupling
interface between sub-structures: An experimental test for ‘completeness’. In IMAC-XXXVI:
International modal analysis conference, Orlando, FL. Bethel, CT: Society for Experimental
Mechanics.
180 4 Experimental Substructuring
Montalvao, D., Ribeiro, A., Maia, N., & Silva, J. (2004). Estimation of the rotational terms of the
dynamic response matrix. Shock and Vibration, 11(3–4), 333–350. https://doi.org/10.1155/2004/
780837.
Moorhouse, A., Elliot, A. S., & Heo, Y. (2013). Intrinsic characterisation of structure-borne sound
sources and isolators from in-situ measurements. In Proceedings of Meetings on Acoustics
ICA2013 (vol. 19, p. 065053). ASA.
Moorhouse, A. T. (2003). Compensation for discarded singular values in vibro-acoustic inverse
methods. Journal of Sound & Vibration, 267(2), 245–252. https://doi.org/10.1016/S0022-
460X(02)01432-3.
Nicgorski, D. (2008). Investigation on experimental issues related to frequency response function
measurements for frequency based substructuring. Master’s thesis, University of Massachusetts
Lowell.
Nicgorski, D., & Avitabile, P. (2010). Experimental issues related to frequency response function
measurements for frequency-based substructuring. Mechanical Systems and Signal Processing,
24(5), 1324–1337.
Nicgorski, D., & Avitabile, P. (2010). Conditioning of FRF measurements for use with frequency
based substructuring. Mechanical Systems & Signal Processing, 24(2), 340–351. https://doi.org/
10.1016/j.ymssp.2009.07.013.
Nicgorski, D., & Avitabile, P. (2010). Experimental issues related to frequency response function
measurements for frequency-based substructuring. Mechanical Systems & Signal Processing,
24(5), 1324–1337. https://doi.org/10.1016/j.ymssp.2009.09.006.
O’Callahan, J. C., Avitabile, P.,& Riemer, R. (1989) System equivalent reduction expansion process
(SEREP). In Proceedings of the VII international modal analysis conference (IMAC), Boston,
MA (vol. 1, pp. 29–37). Bethel, CT: Society for Experimental Mechanics.
Otte, D., Leuridan, J., Grangier, H., & Aquilina, R. (1991). Prediction of the dynamics of structural
assemblies using measured FRF-data: Some improved data enhancement techniques. In IMAC-
IX: international modal analysis conference, Florence, Italy (pp. 909–918). Bethel, CT: Society
for Experimental Mechanics.
Pasma, E. A., van der Seijs, M. V., Klaassen, S. W. B., & van der Kooij, M. W. (2018). Frequency
Based Substructuring with the Virtual Point Transformation, Flexible Interface Modes and the
Transmission Simulator.
Pasteuning, F. (2007). Experimental dynamic substructuring and its application in automotive
research. Master’s thesis, Delft University of Technology, The Netherlands.
Petersson, B. (1987). On the use of giant magnetostrictive devices for moment excitation. Journal
of Sound & Vibration, 116(1), 191–194. https://doi.org/10.1016/S0022-460X(87)81332-9.
Rixen, D. J. (2008). How measurement inaccuracies induce spurious peaks in frequency based
substructuring. In IMAC-XXVII: International modal analysis conference, Orlando, FL. Bethel,
CT: Society for Experimental Mechanics.
Rixen, D. (2008). How measurement inaccuracies induce spurious peaks in frequency based
substructuring. In Proceedings of the XXVI international modal analysis conference (IMAC),
Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Rixen, D. J. (2010). Substructuring technique based on measured and computed impulse response
functions of components. In P.S. et al. (Ed.) ISMA. K.U. Leuven.
Rixen, D. J. (2010). Substructuring using impulse response functions for impact analysis. In IMAC-
XXVIII: International modal analysis conference, Jacksonville, FL. Bethel, CT: Society for Exper-
imental Mechanics.
Rixen, D., Godeby, T., Pagnacco, E. (2006). Dual assembly of substructures and the fbs method:
Application to the dynamic testing of a guitar. In International conference on noise and vibration
engineering, ISMA. Belgium: KU Leuven.
Rixen, D. J., & van der Valk, P. L. (2013). An impulse based substructuring approach for impact
analysis and load case simulations. Journal of Sound and Vibration, 332, 7174–7190.
Sanderson, M. A., & Fredö, C. R. (1995). Direct measurement of moment mobility: Part I: A
theoretical study. Journal of Sound & Vibration, 179(4), 669–684. https://doi.org/10.1006/jsvi.
1995.0043.
References 181
Sjövall, P., & Abrahamsson, T. (2007). Substructure system identification from coupled system test
data. Mechanical Systems and Signal Processing, 22, 1–260.
Sjövall, P., & Abrahamsson, T. (2007). Component system identification and state-space model
synthesis. Mechanical Systems & Signal Processing, 21(7), 2697–2714. https://doi.org/10.1016/
j.ymssp.2007.03.002.
Sjövall, P., McKelvey, T., & Abrahamsson, T. (2006). Constrained state-space system identification
with application to structural dynamics. Automatica, 42, 1539–1546.
Tol, Ş., & Özgüven, H. N. (2015). Dynamic characterization of bolted joints using frf decoupling
and optimization. Mechanical Systems and Signal Processing, 54, 124–138.
Urgueira, A. P. V. (1989). Dynamic analysis of coupled structures using experimental data. Ph.D.
thesis, Imperial College, London.
van Overschee, P., & de Moor, B. (1996). Subspace Identification for Linear Systems. Kluwer
Academic Publishers.
van der Seijs, M., de Klerk, D., Rixen, D., & Rahimi, S. (2013). Validation of current state frequency
based substructuring technology for the characterisation of steering gear – vehicle interaction.
In IMAC-XXXI: International modal analysis conference, Garden Grove, California USA (pp.
11–14). Bethel, CT: Society for Experimental Mechanics.
van der Seijs, M. V., de Klerk, D., Rixen, D. J., & Rahimi, S. (2013). Validation of current state
frequency based substructuring technology for the characterisation of steering gear – vehicle
interaction. In Topics in experimental dynamic substructuring, Proceedings of the 31st IMAC, a
conference on structural dynamics (Vol. 2, Chap. 20, pp. 253–266). New York: Springer. https://
doi.org/10.1007/978-1-4614-6540-9_20.
van der Seijs, M. V., Pasma, E. A., van den Bosch, D. D., & Wernsen, M. W. M. (2017). A
benchmark structure of validation of experimental substructuring, transfer path analysis and
source characterisation techniques. In Dynamics of coupled structures, Proceedings of the 35th
IMAC, a conference and exhibition on structural dynamics (Vol. 1, Chap. 6, pp. 73–82). New
York: Springer.
van der Seijs, M. V., van den Bosch, D. D., Rixen, D. J., & de Klerk, D. (2013). An improved
methodology for the virtual point transformation of measured frequency response functions in
dynamic substructuring. In M. Papadrakakis, V. Papadopoulos, V. Plevris (Eds.) 4th ECCOMAS
thematic conference on computational methods in structural dynamics and earthquake engineer-
ing (COMPDYN) (pp. 4334–4347). Kos Island, Greece. https://doi.org/10.13140/RG.2.1.2715.
3126.
Voormeeren, S., & Rixen, D. (2009). Substructure decoupling techniques - a review and uncertainty
propagation analysis. In IMAC-XXVII: International modal analysis conference, Orlando, FL.
Bethel, CT: Society for Experimental Mechanics.
Voormeeren, S., & Rixen, D. (2012). A family of substructure decoupling techniques based on a
dual assembly approach. Mechanical Systems and Signal Processing, 27, 379–396.
Voormeeren, S., de Klerk, D., & Rixen, D. (2010). Uncertainty quantification in experimental
frequency based substructuring. Mechanical Systems and Signal Processing, 24(1), 106–118.
Wang, Z. W., & Wang, J. (2011). Inverse substructure method of three-substructures coupled system
and its application in product-transport-system. Journal of Vibration and Control, 17(6), 943–951.
Wang, Z., Zhu, P., & Liu, Z. (2018). Relationships between the decoupled and coupled transfer
functions: Theoretical studies and experimental validation. Mechanical Systems and Signal Pro-
cessing, 98, 936–950.
Zhen, J., Lim, T.C., & Lu, G. (2004). Determination of system vibratory response characteris-
tics applying a spectral-based inverse sub-structuring approach. Part i: Analytical formulation.
International Journal of Vehicle Noise and Vibration1(1), 1–30.
Zhen, J., Lim, T.C., & Lu, G. (2004). Determination of system vibratory response characteris-
tics applying a spectral-based inverse substructuring approach. Part ii: Motor vehicle structures.
International Journal Vehicle Noise and Vibration1(1/2), 31–67.
Chapter 5
Industrial Applications & Related
Concepts
Transfer Path Analysis (TPA) has been a valuable engineering tool for as long as
noise and vibrations of products have been of interest. An TPA concerns a product’s
actively vibrating components (such as engines, gearing systems or turbochargers)
and the transmission of these vibrations to the connected passive structures. TPA is
particularly useful when the actual vibrating mechanisms are too complex to model
or measure directly, as it allows to represent a source by forces and vibrations dis-
played at the interfaces with the passive side. In this way, the source excitations
can be separated from the structural/acoustic transfer characteristics, allowing to
troubleshoot the dominant paths of vibration transmission. The engineer can then
anticipate by making changes to either the source itself or the receiving structures
that are connected to it.
An TPA often rises from the need to reduce some sort of undesired noise or
vibration, for instance, to improve product comfort or lifetime, ensure safety, or
preserve stealthiness. Aside from automotive development, applications are also
seen in industries such as marine and airplane engineering, building acoustics, and
acoustic modeling of musical instruments. An TPA is generally motivated by one of
the following desires:
1. Secrecy: perhaps, the earliest TPA studies were triggered by the need to reduce
the transmission of engine vibrations in military ships and submarines in order
to make them stealthy. Many publications in the 1950s and 1960s document on
isolation of ship engines by means of absorbers and decoupling mechanisms (Har-
rison et al. 1952; Ungar and Dietrich 1966; Plunkett 1958; Soliman and Hallam
1968; Snowdon 1979) to minimize the transmission through the interfaces.
2. Safety: along with the rapid development of airplanes and spacecraft in the 1960s,
TPA concepts started to be of use to study fatigue and stability (flutter) problems
due to active or induced vibrations. As sources of vibrations are much more per-
sistent in aeronautics—think of vortex-induced vibrations—focus was on char-
acterizing the passive transfer paths by means of modal analysis (Bisplinghoff
et al. 1955; On 1967).
3. Comfort: over the past decades, TPA tends to be particularly associated with
Noise, Vibration, and Harshness (NVH) engineering as commonly encountered
in the automotive industry. The majority of recent developments and commercial
solutions have been tailored toward this engineering society or related industries,
driven by the increasing customer expectations on acoustic comfort (Penne 2004;
Plunt 2005; van der Auweraer et al. 2007; Zeller 2009; Cerrato 2009).
In response to the evolving demands, TPA methods have been under continuous
development and their family members have grown numerous. Some designations
that found their way into literature include Operational TPA (OTPA), Operational
Path Analysis with exogenous inputs (OPAX), blocked-force TPA, Gear Noise Prop-
agation, in situ Source Path Characterization and Virtual Acoustic Prototyping. Very
often, those methods are presented from highly case-specific derivations. Not surpris-
ingly, as the underlying physical concepts are similar, some of the abovementioned
show strong similarities or are even identical.
An TPA workflow can typically be subdivided into the following steps:
1. operational measurement on the active component;
2. determination of the passive (sub)system characteristics (commonly by means of
FRFs);
3. identification of interface loads;
4. calculation of path contributions.
The steps are shown schematically in Fig. 5.1. Depending on the TPA method at
hand, some or all of these steps may be performed in arbitrary order. The opti-
mization actions that follow from such an analysis are generally not considered part
of the workflow.
The next section presents a unified framework for derivation of a large range of
TPA methods. It is chosen to present and classify the methods separate from their
typical fields of application, such that the underlying physical concepts are exposed
and can be compared. In Sect. 5.2, a general framework for TPA is introduced, start-
ing by depicting the transfer problem using the Dynamic Substructuring paradigm
(de Klerk et al. 2008). Hereafter, the TPA methods are derived and classified along
three families, namely the classical (Sect. 5.3), component-based (Sect. 5.4), and
transmissibility-based (Sect. 5.5) TPA methods, as depicted vertically in Fig. 5.1.
5.1 Introduction to Transfer Path Analysis 185
(a) operational test (b) FRF measurement (c) interface loads (d) path contributions
component-
based TPA active subsystem A assembly AB product of eq. forces
for subsystem A for A and FRFs of AB
transmissibility-
based TPA assembly AB transmissibility from operational responses on AB partial responses from
transmissibility of AB
Fig. 5.1 The TPA workflow, depicted stepwise for the three TPA families
Let us consider the dynamic system AB as schematically depicted in Fig. 5.2a. Two
subsystems can be distinguished: an active subsystem A containing an excitation
at node 1 and a passive subsystem B comprising the responses of interest at node
3. The subsystems are rigidly interconnected at the interface node 2. For simplicity
of derivation, the Degrees of Freedom (DoFs) in this example is restricted to three
distinct nodes. These may, however, represent a larger set of DoFs, representing,
respectively:
(1) source: internal DoFs belonging to the active component that causes the opera-
tional excitation but are unmeasurable in practice;
(2) interface: coupling DoFs residing on the interface between the active and passive
component;
(3) receiver: response DoFs at locations of interest at the passive component, pos-
sibly including acoustic pressures and other physical quantities.
Hence, the example of Fig. 5.2a is illustrative for a wide range of practical problems,
provided that the structure of interest can be decomposed into an active and passive
part. In what follows, all methods assume that the operational excitation at node 1
is unmeasurable in practice, but transmits vibrations through the interfaces at node
2 to receiving locations at node 3. The responses shall then be built up from a
certain description of the vibrations measured at the interface (node 1 → 2) and
an appropriate set of transfer functions relating these vibrations to the receiving
responses (node 2 → 3). The fundamental choices herein dictate to which TPA family
the method is classified.
Let us first approach the transfer problem top-down for the assembled system AB, see
Fig. 5.2a. We are interested in the response spectra at the receiving locations u3 (ω)
186 5 Industrial Applications & Related Concepts
u3
u2
u1 YAB
31
f1
(a) based on the admittance of assembly AB.
subsystem A subsystem B
u3
uA
2 uB2 YB32
u1
YA gA
2
λ gB2
21
f1
(b) based on the admittances of subsystems A and B.
for source excitations at node 1, given by the force spectra f1 (ω). For the assembled
problem, this is simply obtained from a superposition of the individual contributions,
i.e. the excitation force spectra multiplied with their respective linear(ized) transfer
functions, contained in the columns of admittance FRF matrix YAB (ω).
u i (ω) = j (ω) f j (ω)
YiAB =⇒ u3 (ω) = Y31
AB
(ω) f1 (ω). (5.1)
j
In the equations that follow, the explicit frequency dependency (ω) will be omitted
to improve readability. Also note that the response set u can include displacements,
velocities, accelerations, or any other quantity, provided that the rows of the FRF
matrices are obtained accordingly. Furthermore, in order to keep the derivations
brief and understandable, it is chosen to only consider the structure-borne paths.
Nevertheless, Eq. (5.1) can easily be extended to include contributions of airborne
paths if the application so requires. In that case, Y and f need to be augmented with
a set of (responses to) acoustic loads such as volume velocities (m3 /s) (van der Seijs
et al. 2016).
The same transfer function is now derived for an assembly of the individual subsys-
tems, as depicted in Fig. 5.2b. Let us first put the subsystem’s FRF matrices YA and
5.2 The Transfer Path Problem 187
The following physical explanation can now be reasoned to solve Eq. (5.2). The
excitation force at node 1 induces a motion at node 2 of subsystem A. As subsystem
B is not directly affected by forces at A (due to the block-diagonal form of the global
FRF matrix), an incompatibility is caused between u2A and u2B . This is denoted by the
interface “gap” δ, which can conveniently be written using a signed Boolean matrix1
B as shown by Eqs. (5.3a) and (5.3c). Next, assuming that no additional mass is
present between the subsystems, the equilibrium condition is satisfied, requiring the
interface forces g2 on both sides to be equal in magnitude and opposing in sign. The
interface forces are expressed by Eqs. (5.3b) and (5.3c), using a Lagrange multiplier
λ for the magnitude and the transposed Boolean matrix to account for the interface
force direction.
The interface forces λ that pull the two subsystems together can be determined from
Eq. (5.3a) by requiring δ = 0, which enforces the compatibility condition u2A =
u2B . Considering Eq. (5.3b), the interface forces that ensure compatibility can be
determined by equating the second and third line of Eq. (5.2) and solving for λ:
A
Y21 f1 + Y22
A A
g2 = Y22
B B
g2
A
Y22 + Y22 λ = −Y21
B A
f1
A
B −1 A
λ = − Y22 + Y22 Y21 f1
−1
g2B = Y22
A
+ Y22
B A
Y21 f1 . (5.4)
Equation (5.4) provides the interface forces at the coinciding interface DoFs caused
by the operational excitation f1 inside subsystem A. The response at the receiving
side u3B is found by substituting Eq. (5.4) into the last line of Eq. (5.2):
1 Thesigned Boolean matrix B establishes the relations for all interface DoFs of A and B that are
vectorially associated, e.g. u A B
2x and u 2x . Guidelines on the construction and properties of the signed
Boolean matrix are found in de Klerk et al. (2008).
188 5 Industrial Applications & Related Concepts
A
B −1 A
u3 = Y32
B B
g2 = Y32
B
Y22 + Y22 Y21 f1 . (5.5)
AB
Y31
Comparing with Eq. (5.1), it follows that the terms between the brackets indeed
AB
represent the admittance of the assembly Y31 by coupling of the subsystems’ admit-
tances. It can be verified that this result is in accordance with Lagrange Multiplier Fre-
quency Based Substructuring (LM-FBS) assembly (de Klerk et al. 2008) as detailed
in Chap. 3.
So far, it has been assumed that the excitation at node 1 is measurable. In reality,
however, it is impossible or impractical to identify the exact force loading. This
is solved in TPA by assuming that the dynamics at the interface node 2 due to
this excitation are measurable and may very well represent the source excitation.
In what follows, different approaches are examined to describe the transmission of
vibrations, or rather, the response at the passive subsystem B for a nonmeasurable
excitation somewhere inside (or on) the active subsystem A. The notation of Dynamic
Substructuring is used to reveal the relations between the different approaches.
The family of Classical TPA methods is essentially intended to identify transfer path
contributions in existing products. They have, nowadays, become standard prac-
tice to troubleshoot NVH problems in automotive engineering (Plunt 2005; van der
Auweraer et al. 2007). A classical TPA performs operational tests on the assembled
product AB to obtain interface forces between the active and passive side, namely λ
in Eq. (5.4). It can be verified from Eqs. (5.2) and (5.5) that these interface forces
fully determine the responses at the passive side and are thus representative for the
effects of the source vibrations at the receiver locations u3 . To calculate the receiver
responses,2 , the passive side interface forces g2B = −λ are applied to the interfaces
of subsystem B, as shown in Fig. 5.4a.
u3 = Y32
B B
g2 . (5.6)
Both steps pose some challenges in practice. The FRFs of the passive side are typically
determined from impact or shaker tests, or in a reciprocal fashion using, for instance,
an acoustic source at the receiving location and accelerometers at the interface nodes
(Ten Wolde 1973; ten Wolde et al. 1975; Fahy 1995, 2003; ten Wolde 2010). Either
way, it requires dismounting of the active part(s) from the passive side. With respect
to the determination of operational interface forces g2B , it could be impractical to
2 Inthis framework, the resulting responses u3 are formulated as a matrix-vector product, namely
the sum of the partial responses. Techniques to evaluate the individual transfer paths contributions
are discussed in van der Seijs et al. (2016).
5.3 Classical TPA 189
mount force sensors between the active and passive part. Therefore, a number of
indirect methods have been developed to circumvent direct force measurement.
Hence, the variants of classical TPA are defined according to how g2B is obtained
from operational tests, which are discussed next.
The most straightforward technique to obtain the interface forces is from force trans-
ducers mounted directly between the active the passive side, as depicted in Fig. 5.3a.
It was demonstrated by Eq. (5.4) that the interface force caused by the operational
excitation is given by
A
B −1 A
λ = − Y22 + Y22 Y21 f1 =⇒ g2B = −λ. (5.7)
This is valid under the assumption that u2B = u2A , which requires the stiffness of the
transducers to be high enough (relative to the stiffness of the actual subsystems)
in the frequency range of interest. In fact, the main drawback of the method is the
inconvenience of placing the transducers between the active component and the
receiving structure. Lack of space, distortion of the original mounting situation,
and the incapability to measure all desired degrees of freedom at a connection point
render the method impractical, especially for typical automotive applications. In case
of large-scale systems such as ship machinery, this method may still be preferred
(Verheij 1982).
g Aj 1 −1 u Aj
= −z mt
j j −1 1 =⇒ g Bj = z mt
j j (u j − u j )
A B
j ∈ 1, . . . , m.
g Bj u Bj
190 5 Industrial Applications & Related Concepts
λ
gB2 = −λ gB2 = Zmt uA
2 − u2
B
λ u3 Zmt u3
uA
2 uA
2
? ?
uB2 uB2
+
gB2 = YB42 u4
u4 u3
uA
2
? YB42
u4
Fig. 5.3 Three approaches to determine the operational interface forces in Classic TPA
A spring-like stiffness matrix can be recognized, however, with a minus sign because
the interface forces g j act on the connected subsystems A and B instead of the mount.
Introducing the diagonal mount stiffness matrix Zmt , the full set of m interface forces
g2B can be estimated from the differential interface displacements between the source
and receiver side, i.e. the measured displacements at both sides of the mounts:
In most cases, the flexible mounts are already integrated in the design to attenuate
the vibration transmission. If they are, however, placed in the system for the mere
purpose of TPA, it can be shown that the interface forces and thus the vibrations of
the coupled system are altered significantly (Reuss et al. 2012; Barten et al. 2014),
namely:
A −1 A −1
g2B = Y22 + Y22
B
+ Ymt Y21 f1 with Ymt = Zmt .
Although the mount stiffness method can be powerful and easy to conduct, the
accuracy is highly subject to the terms3 in Zmt . Typical absorbers exhibit amplitude-
dependent nonlinearities and directional characteristics (Harrison et al. 1952; Ungar
and Dietrich 1966; Thompson et al. 1998).
3 Note that the terms in the dynamic stiffness matrix Zmtcorrespond to differential displacements of
the associated interface DoFs u2A − u2B and not the coordinates of both A and B. Some implications
for the terms in Zmt are discussed in Voormeeren (2012), Barten et al. (2014).
5.3 Classical TPA 191
The third and perhaps most popular classical TPA member is the matrix-inverse
method (Powell and Seering 1984; Dobson and Rider 1990; Thite and Thompson
2003a, b). It was observed from Eq. (5.6) that responses at the passive side are found
from application of the interface forces to the passive subsystem’s FRFs. Recalling
these responses from Eq. (5.2):
B B
u2 Y22 B
= B g2 .
u3 Y32
This problem can be inverted if the left-hand side contains sufficient independent
responses to describe all m interface forces and moments in g2B . The set of receiver
responses u3 is typically too small in number and too distant from the interfaces to be
suitable for inversion. Inversion of the first row is theoretically sufficient, but requires
complete instrumentation of the assembled structure’s interfaces to measure all DoFs
u2B associated with g2B . In addition, a symmetric FRF matrix Y22 B
would be required
for the passive subsystem’s interfaces, which is challenging to obtain accurately.
In practice, the passive side is equipped with so-called indicator responses u4 as
shown in Fig. 5.3c. An amount of n ≥ m responses shall be located in the proximity
of the interfaces, such that the full set of m interface forces is properly observable
from these points (this is addressed below). As these indicator DoFs merely assist
in the determination of the interface forces, the sensor type can be chosen rather
arbitrarily, although (triaxial) accelerometers are the most common choice.
Two sets of measurements are now required to reconstruct the interface force
spectra. First, responses u4 are measured on the assembled system AB during oper-
ational tests. These can be expressed in terms of subsystem admittances, similar to
Eq. (5.5):
A
B −1 A
u4 = Y41
AB
f1 = Y42
B
Y22 + Y22 Y21 f1 . (5.9)
g2B
Next, FRFs need to be measured for the passive subsystem B, relating the motion at
B
these indicator points to forces at the interface, namely Y42 . Note that this requires
B
dismounting of the active components from the assembly. Nevertheless, the FRFs Y42
B
(needed for the matrix-inverse force determination) and Y32 (to calculate responses at
the target locations, Eq. (5.6)) can be obtained from the same FRF measurement cam-
paign, as it only involves mounting of additional sensors. The operational interface
forces can be reconstructed from a pseudoinverse of the indicator response spectra
u4 with the subsystem FRFs:
B +
g2B = Y42 u4 . (5.10)
192 5 Industrial Applications & Related Concepts
B + B
B
If Y42 is full rank, it holds that Y42 Y42 = I, such that in theory the correct interface
forces from Eq. (5.4) are obtained by subsequent application of Eqs. (5.9) and (5.10).
Hence, the conditioning of the FRF matrix is crucial, which should have
1. sufficient rank to describe all interface forces g2B independently from the set u4
and;
2. a reasonably low condition number in order not to amplify measurement errors
in the inversion.
As a rule of thumb, it is common to use at least twice as many response DoFs as
strictly required to fully determine the interface forces. Much attention has been
devoted to improving the conditioning of the inverse problem by means of singular
value decomposition, see for instance, the work of Thite and Thompson (2003a, b,
2006), Dobson and Rider (1990), Choi et al. (2007) or an early review of techniques
(Mas and Sas 1994). Time domain implementations have also been developed, such
as the inverse structural filtering methods described in Powell and Seering (1984),
Kammer (1998).
Recently, there has been an interest in the application of strain gages, instead of
the commonly used accelerometers at the indicator points. It is argued that strain
responses possess a more direct relation to interface forces and are better able to
capture the local phenomena of the structure. Consequently, it is expected that strain
measurements lead to better conditioning of the matrix to be inverted (Sachse et al.
2013; dos Santos et al. 2014). More research is currently needed to further verify
this assumption.
assembly AB
subsystem A subsystem B
u3
u3 YAB
32
YB32 eq
f2 u2
uB2
2 g2
gA B
Fig. 5.4 Application of forces representing the operational excitation: classical TPA and
component-based TPA
The response u3 , as a result of the equivalent forces, should equal the response
obtained when the active component is running in the same assembly, caused by f1
in Eq. (5.5). Comparing Eqs. (5.11) with (5.5), it follows naturally that the equivalent
forces should take the form
A
B −1 A eq
A
B −1 A
B
Y32 Y22 + Y22 Y22 f2 = Y32
B
Y22 + Y22 Y21 f1 =⇒
194 5 Industrial Applications & Related Concepts
eq A −1 A
f2 = Y22 Y21 f1 . (5.12)
Equation (5.12) shows that the equivalent forces are indeed a property on the active
component A only. The existence of such an equivalent source problem offers tremen-
dous potential for practical component-based TPA methods. There is, however, one
important limitation: the equivalent forces only properly represent the operational
excitations for responses on the receiving structure or on the interface. Responses
obtained on the source will be different and therefore of no use. This limitation
was already mentioned in de Klerk (2009), de Klerk and Rixen (2010) and can be
understood by examining the system of Eq. (5.2): responses at the passive side B
are caused only by forces through or onto the interface, whereas the responses at
the source side A are a result of both the direct contribution of f1 and its reflection
through the coupled subsystem B.
In the next section, several approaches are discussed that yield a set of equivalent
forces from operational tests, following the definition of Eq. (5.12).
Consider the case where the boundary of subsystem A is rigidly fixed, as depicted in
Fig. 5.5a. The operational excitation f1 is entirely portrayed by the reaction forces at
the “blocked” interface g2bl , such that the interface displacements u2A = 0. A direct
relation is found, leading to the following equivalent force:
A A A −1 A
u1 Y11 Y12 f1 g2bl = Y22 Y21 f1
= =⇒
u2 = 0 A
Y21 A
Y22 g2A = −g2bl eq
f2 = g2bl .
(5.13)
The so-called blocked-force TPA is perhaps the most commonly known variant of
component-based TPA methods because of its straightforward applicability. It can be
seen as an application of the Thévenin equivalent source problem (Thévenin 1883),
that found its way into popular mechanics halfway the twentieth century (Hixson
1961; Ungar and Dietrich 1966; Rubin 1967). Mathematically, one can regard the
blocked-force method as imposing a Dirichlet boundary condition on the active
subsystem’s interface.
The blocked-force method assumes the boundary to be infinitely stiff in all direc-
tions, which is in practice rarely the case. Hence, the accuracy of the blocked forces
is highly subject to the stiffness of the boundary relative to the component at hand
(de Klerk 2009; van den Bosch et al. 2014; van der Seijs et al. 2014). An additional
difficulty is the measurement of rotational moments, as most commonly used sen-
sors are unable to measure collocated 6-DoF interface loads. As a consequence, the
blocked-force method is expected to perform best at the low-frequency end for which
the rigid boundary assumption is most valid and rotational effects are in practice least
prominent.
5.4 Component-Based TPA 195
eq eq −1
f2 = gbl
2 f2 = YA
22 ufree
2
u2 YA
22
? ? ufree
2
gbl
2
eq −1 eq +
f2 = gR2 + YA
22 u2 f2 = YAB
42 u4
YA YR22 u4
22 u2
? u2 ? YAB
gR2 42
u4
Fig. 5.5 Component-based TPA methods: various approaches to obtain equivalent forces repre-
senting the excitation
The direct counterpart of the blocked-force method is the free velocity TPA as depicted
in Fig. 5.5b. In this case, the component’s interfaces are left free, such that all vibra-
tions are seen as “free displacements” u2free at the interface DoFs. From here on,
equivalent forces4 can be calculated by inverting the admittance measured at the free
subsystem’s interfaces, which can be understood by comparing the free displace-
ments with the blocked force definition of Eq. (5.12):
A A
u1 Y11 Y12 f1 u2free = Y21
A
f1
= =⇒ A −1 free (5.14)
u2A = u2free A
Y21 A
Y22 g2A = 0 eq
f2 = Y22 u2 .
Analogue to the blocked-force TPA, this method can be seen as a strict application
of a Neumann boundary condition and is furthermore related to Norton’s equivalent
source theorem for electric networks (Norton 1926). Again, imposing free boundary
conditions can be troublesome as the vibrating system often needs to be mounted at
one or more connection points to be able to run in operation. Therefore, the method
is in practice expected to perform best for frequencies well above the rigid body
modes of the structure. Also note that effects such as friction and damping, which
4 If the admittances of the subsystems are available separately, one may also apply the free velocities
would occur at the interfaces of the assembled system, are absent in the operational
measurement on the free component.
The fact that the method makes reference to free velocities rather than free dis-
placements has a historical motive, as the former quantity is commonly applied
in acoustical engineering in combination with acoustic pressure. For such acous-
tic problems, dynamic coupling of admittance is normally discarded. This is a fair
assumption, considering that the impedance of the radiating source is much larger
than the impedance of the receiving surrounding fluid, i.e. air. In structure-borne
vibrations, the source and receiving system often exhibit similar dynamics, hence
explicit coupling is needed. This was realized for a single-DoF problem by means
of a nondimensional coupling function in Mondot and Petersson (1987). Extensions
to multi-DoF systems and further discussion of this topic are found in Petersson and
Gibbs (1993), Petersson and Gibbs (2000), Moorhouse (2001). As a final note, the
free velocity concept has been standardized into an ISO norm for characterization of
structure-borne sound sources (ISO Technical Committee 1996).
As both abovementioned methods pose their limitations in practice, one often prefers
to conduct operational tests on an appropriate support structure. Such a coupled
structure unavoidably displays its own dynamics on the interfaces, which need to be
accounted for in order to render the equivalent forces independent of any connected
part. Let us, therefore, imagine the active component fixed onto a test bench or test
rig R as illustrated in Fig. 5.5c. Denoting the interface admittance of the test rig by
R
Y22 , we obtain for the extended system of equations:
⎡ ⎤ ⎡ A A ⎤⎡ ⎤
u1A Y11 Y12 0 f1
⎣u2 ⎦ = ⎣Y21 Y22 0 ⎦ ⎣g2 ⎦ with
A A A A u2A = u2R
(5.15)
u2R 0 0 Y22R
g2R g2A = −g2R .
After enforcing the compatibility and equilibrium conditions on the interface DoFs, a
derivation similar to Sect. 5.2.2 can be followed to find the operational interface forces
g2R . These forces are now dependent on the properties of both the active component
A and the test rig R. Substituting the forces back into the second row of Eq. (5.15),
the corresponding interface displacements u2 are obtained as well (the superscript is
dropped because of compatibility):
⎧ A
⎨ g2R = Y22 R −1 A
+ Y22 Y21 f1 (5.16a)
⎩ u2 = I − YA YA + YR −1 YA f1 . (5.16b)
22 22 22 21
R
Equations (5.16a), (5.16b) hold for any Y22 , as long as no external force is acting
on the test rig. The desired set of equivalent forces Eq. (5.12) is obtained by com-
R
bining Eqs. (5.16a) and (5.16b) in such way that the dynamics of the test rig Y22 are
5.4 Component-Based TPA 197
eliminated.
eq A −1 A eq A −1
f2 = Y22 Y21 f1 =⇒ f2 = g2R + Y22 u2 . (5.17)
As it turns out, Eq. (5.17) forms the sum of contributions of both the blocked force
and the free velocity experiment, respectively Eqs. (5.13) and (5.14). One could
therefore speak of a hybrid interface condition, or Robin boundary condition in a
mathematical sense. It can indeed be verified that Eqs. (5.16a)–(5.16b) converge5 to
the blocked forces for YR → 0 and to free velocities when YR → ∞.
The hybrid interface approach combining force and motion was originally pub-
lished in de Klerk and Rixen (2010); the displacement term was regarded in this
work as the “non-rigid test bench compensation” to the imperfect blocked forces.
Although generally applicable in theory, it should be mentioned that the method can
be rather costly and time consuming in practice, as one needs to explicitly measure
collocated forces and motion in all directions for every interface node (van den Bosch
et al. 2014).
In lack of force measurement, one may substitute the third row of (5.15) for g2B .
The so obtained approach was suggested in van der Seijs et al. (2014) and takes
displacement measurement only:
eq R −1 A −1
f2 = Y22 u2 + Y22 u2 . (5.18)
The price for not having to measure interface forces is that separate FRF measure-
ments should now be conducted, to obtain the subsystem admittances of the active
component and the test rig.
Looking again at Eq. (5.18), we observe that the two inverted admittance FRF matri-
ces, in fact, represent the dynamic stiffness matrices of respectively component A
and R for the same set of collocated interface DoFs. They can be combined by simple
impedance addition6 :
eq
f2 = ZR22 + ZA
22 u2 = Z22 u2 .
AR
(5.19)
The result of Eq. (5.19) is indeed indifferent to the dynamics of R or any other mount-
ing structure that component A is connected to. Transforming back to admittance
notation, we find that the blocked forces can be calculated inversely from the admit-
tance of the assembly’s interface or, in an overdetermined fashion, using a sufficient
set of indicator points u4 on the receiving subsystem:
5A intuitive presentation of the range between the two limit cases is given in Moorhouse (2001).
6 This is standard practice for FEM assembly.
198 5 Industrial Applications & Related Concepts
eq AR −1 eq AR +
f2 = Y22 u2 or f2 = Y42 u4 . (5.20)
Obtained from a different derivation, the approach of Eq. (5.20) has first been pro-
posed by Moorhouse and Elliott (Elliott and Moorhouse 2008; Moorhouse et al.
2009) as the in situ method. As implied by the name, the operational measurement
may even be conducted in the target assembly AB, avoiding dismounting of any part,
namely:
eq AB −1 eq AB +
f2 = Y22 u2 or f2 = Y42 u4 . (5.21)
The equivalent forces resulting from application of Eq. (5.21) are nonetheless a
property of component A only and thus transferable to an assembly with another
passive side.
The in situ force determination procedure is illustrated in Fig. 5.5d. Indeed, Eq.
(5.21) represents the inverse of Eq. (5.11), yet with an extended set of response DoFs
in order to render the system (over)determined and thus invertible. As a consequence,
the method shows great resemblance with the classical matrix-inverse method of Eq.
(5.10), the difference being the choice for the assembled admittance instead of the
passive subsystem’s admittance. Similar techniques regarding matrix conditioning
(e.g. overdetermination, singular value rejection) apply as well to the in situ method.
Most remarkable is that the receiving side can be chosen arbitrarily, as the equivalent
forces identified by Eq. (5.20) or (5.21) are theoretically invariant of any subsystem
coupled to it. In that respect, two important conditions need to be kept in mind:
1. Operational excitations f1 may only originate from the domain of component A.
Any excitation coming from the passive side will disturb the determination of
equivalent forces.
2. Although the responses used for the matrix inversion (u2 or u4 ) can be positioned
rather arbitrarily, they are bounded to the domain of the interface and the passive
side. This relates back to the remark made after Eq. (5.12): vibrations at the
source structure are not only caused by the interface forces but also by the source
excitations directly.
Physically, one could interpret the in situ method as follows: knowing the transfer
functions from the interface DoFs to several points on the passive side, Eq. (5.20) or
eq
(5.21) seeks for a set of equivalent forces f2 that, applied to the interface DoFs of
the assembled structure, generates the same responses u4 at the passive side. Given
that this response set is overdetermined, the equivalent forces are calculated such
that they minimize the sum of the squared errors (or l 2 -norm) in the extended set7 u4 .
Several numerical and experimental studies of in situ source characterization have
been reported (Elliott 2009; Bonhoff 2010; Elliott et al. 2013; Madsen 2014; van der
7 Similar ideas are used in the field of experimental substructure decoupling: the identification of the
force that decouples the residual substructure can be improved by defining an “extended interface”,
adding some additional DoFs on the structure of interest distant from the interface (Sjövall and
Abrahamsson 2008; D’Ambrogio and Fregolent 2010; Voormeeren and Rixen 2012).
5.4 Component-Based TPA 199
+
u2 = YA A
21 f1 − Y2ps fps = 0 fps = YAB
4ps u4
fps fps
u4
u2 = 0 u4
f1 f1
u4
u3 = YAB
3ps fps
fps u3
YAB
3ps
Seijs et al. 2014, 2015). A time domain force reconstruction algorithm was proposed
by Sturm et al. (2012, 2013). Further generalization of the in situ concept can be
recognized in the pseudo-forces method that is discussed next.
The last member of the component-based TPA family to discuss is the pseudo-
forces method as proposed by Janssens and Verweij in the late 1990s (Janssens et al.
1999; Janssens and Verheij 2000; Janssens et al. 2002). It assumes the existence of a
nonunique set of pseudo-forces acting on the outer surface of the active component,
canceling out the effect of the operational vibrations transmitted through the interface
to the passive side. This is illustrated in Fig. 5.6a. If those forces are now applied
in the opposite direction to the assembly with the source shut down, an identical
response at the receiving side should be obtained, see Fig. 5.6c. In other words, the
pseudo-forces are supposed to represent the source excitation for responses at the
passive side.
The fact that it can be regarded as a component TPA method lies in the former
assumption: if there exists a set of pseudo-forces that cancels out the operational
dynamics at the interface, the responses beyond this interface shall be zero as well,
200 5 Industrial Applications & Related Concepts
hence, these forces are invariant of any structure attached to it. A similar reasoning
may as well be applied to the previously discussed equivalent force methods, yet the
locations of the pseudo-forces are not bounded to the interface but extend to the full
domain of the active component (see Fig. 5.6).
The actual determination of the pseudo-forces is carried out slightly differently.
The first step is to define a set of s pseudo-forces fps on the active component that is
(a) sufficient to represent the excitation source and (b) easily accessible for impact
hammer or shaker measurement. One could think of a minimum of s = 6 forces
when the interface behavior can be considered as rigid, or a larger number in case of
more intricate connectivity. Second, a set of n ≥ s indicator response DoFs has to be
chosen on the passive side from which the operational excitation is monitored. These
responses are denoted by u4 and can be written as a result of f1 , similar to Eq. (5.9):
A
B −1 free
u4 = Y41
AB
f1 = Y42
B
Y22 + Y22 free
u2,f1 with u2,f1
Y21
A
f1 . (5.22a)
A substitution is made here using the free velocity caused by the excitation f1 , or in
other words: the theoretical motion at the interface of component A if the interfaces
were left free, see also Sect. 5.4.3. This is by no means a quantity that needs to be
measured, but will prove useful for the derivation later on.
The next step is to determine pseudo-forces fps that best recreate the operational
responses at the indicator DoFs u4 when the source excitation is shut down. Note
that the assumption is similar to the equivalent forces problem stated in Sect. 5.4.1,
yet the pseudo-forces act on the outer surface of the active component rather than its
interfaces. Similar to Eq. (5.22a), the response at the indicator DoFs as a result of
the pseudo-forces reads:
A
B −1 free
u4 = Y4ps
AB
fps = Y42
B
Y22 + Y22 free
u2,fps with u2,fps
Y21
A
fps . (5.22b)
Now, the pseudo-forces fps that best describe the operational source excitations are
found from solving an overdetermined system with the response set u4 that was
measured under operation (see Fig. 5.6b):
+
fps = Y4ps
AB
u4 . (5.23)
It can be shown that these forces, under certain conditions, are a property of com-
ponent A only. A sufficient condition is that the free interface velocities u2free as a
result of the original source excitation can be fully represented by the set of pseudo-
forces determined from Eq. (5.23). In that case the definitions of Eqs. (5.22a) and
(5.22b) may be equated and one finds that the pseudo-forces are specific for the
source component:
+
free
u2,fps
= u2,f
free
1
=⇒ fps = Y2ps
A A
Y21 f1 . (5.24)
5.4 Component-Based TPA 201
The pseudo-forces determined from Eq. (5.23) are indeed transferable to an assembly
with another passive side. Equation (5.25) yields the responses for this new assembly,
AB
provided that the columns of the FRF matrix Y3ps correspond to the same pseudo-
forces, i.e. excitation points at the source.
The previously discussed in situ method can be regarded as a special case of the
pseudo-force method, namely for the case where the pseudo-forces are located at
the interfaces. The pseudo-forces calculated from Eq. (5.23) shall then equal the
eq
equivalent (blocked) forces from Eq. (5.12), namely fps = f2 . With regard to the
positioning of the indicator DoFs u4 , the same restriction holds as for the in situ
method, namely that they must be located at the passive system B or at the interface.8
The two previously discussed families of TPA have in common that they attempt to
model the vibration transmission in a physically correct sense, namely by determining
as many forces and moments as required to describe the subsystem connectivity in
full. Consequently, both families explicitly require the corresponding FRFs for the
interface DoFs to the receiving response locations. It is evident that this approach
ultimately reveals a wealth of information on the particular functioning of the active
component, force distribution over the interface, resonances in the structure, etc. If,
however, the mere purpose of an TPA is to identify the dominant path contributions
in the assembled product, the efforts to set up and conduct the full experiment can
be fairly disproportionate and costly. This is, especially, the case when multiple
incoherent vibration sources are to be investigated.
The last family to discuss avoids the stage of explicit force determination. Instead,
the path contributions are determined from so-called “transmissibilities” between
sensors, possibly calculated from operational measurements. Various methods are
discussed in the following sections that share the following properties:
1. Measurements are conducted on the assembled product only, saving time to dis-
mount the active components. In fact, the interfaces between the active and passive
components are no longer of principal interest.
2. Path contributions are determined from well-chosen indicator points around the
sources or connections. These indicator points function as inputs to the TPA.
3. The result of the analysis is highly subject to the choice for these indicator points;
care should, therefore, be taken to include all transmission paths.
8 This was erroneously stated in the original work (Janssens and Verheij 2000).
202 5 Industrial Applications & Related Concepts
u3 fps u3
u4 u4
f1
u4 u4
(a) Transmissibility from source FRFs. (b) Transmissibility from alternative FRFs.
Fig. 5.7 The transmissibility concept for an example with two connection points
The family of transmissibility-based TPA methods indeed departs from the traditional
source–transfer–receiver model that assumes a physically meaningful set of loads,
FRFs, and responses. Although potentially less accurate, methodologies derived from
this concept tend to be easy to set up, versatile concerning sensor type and particu-
larly effective for ranking contributions from several sources. From a practical point
of view, transmissibility-based TPA tries to outrun the physically correct methods
by its ability to conduct multiple cycles in a shorter time. Nevertheless, under cer-
tain conditions, results of similar (or even equivalent) accuracy can be achieved as
compared to classical and component-based TPA methods.
9 An example with two connection points was chosen here merely to provide better insight into some
The so obtained transmissibility matrix TAB 34,f1 relates motion at the indicator DoFs u4
(the inputs) to the receiver DoFs u3 (the outputs) for excitation forces f1 . Interestingly,
the size of the transmissibility matrix has become p × n, obfuscating the o excitations
associated with the original FRFs. This raises the question which excitations are
really needed to construct the transmissibility matrix and to what extend this matrix
is generally valid for the problem (u4 → u3 ) under different excitations of the source
structure.
To gain more insight in the transmissibility problem stated by Eq. (5.27), let us
expand Eqs. (5.26a) and (5.26b) in terms of the subsystems’ admittances. As seen
in previous derivations, the terms that couple the subsystems are identical for all
responses (u3 , u4 ) at the passive side:
AB B
u3 Y31 Y32 A
B −1 A
= AB f1 = B Y22 + Y22 Y21 f1 .
u4 Y41 Y42
Recalling now the expressions for the interface force and free velocity, respectively
Eqs. (5.4) and (5.14), the following substitutions can be made:
B
u3 Y32 B
= B g2
u4 Y
42
B −1 free (5.28)
g2B = Y22 A
+ Y22 u2 for m interface forces
width
u2 = Y21 f1
free A
for m free velocities.
Hence, Eq. (5.28) shows that the transmission of vibrations from o forces f1 to n
responses u4 is limited by the number of DoFs of the interface forces/displacements
m. This means that the interface acts as a bottleneck: it limits the effective rank of
the transmissibility problem to a maximum of m. Furthermore, Eq. (5.28) exposes
two interesting properties of the transmissibility concept:
1. Regarding the source excitation, Eq. (5.28) shows a direct relation between the
interface forces g2B of the coupled system and the theoretical free velocities u2free at
the disconnected interfaces of component A. As understood from the component-
based TPA methods, various sets of forces can be defined that equivalently produce
these free interface velocities u2free , such as the pseudo-forces of Eq. (5.22b). This
is illustrated in Fig. 5.7b. In fact, any set of forces on the source that renders
the interface fully controllable will guarantee convergence to a transmissibility
matrix that generally captures the transmission of vibrations caused in component
A. This property is used in the operational TPA method as discussed in the next
section.
2. Regarding the transmissibility, the problem (u4 → u3 ) is in fact specific to the
passive side only, provided that the interface forces g2B are fully observable from
u4 . That means that the second line of Eq. (5.28) can be inverted, which is equiv-
alent to Eq. (5.10) from the classical matrix-inverse method. With respect to
204 5 Industrial Applications & Related Concepts
An elegant duality can be observed here with the component-based TPA concept
that characterizes the excitation as a property of the source component. More
properties of the transmissibility matrix are discussed in Guasch et al. (2013),
Weijtjens et al. (2014).
Theoretically, it can be stated that the controllability is a property of the source
component A, whereas the observability is specific to the receiving structure B. If
both conditions are satisfied, Eqs. (5.27) and (5.29) generate the same transmissibility
matrix, hence TAB34 = T34 or simply T34 . The pseudo-inversion step involved in Eq.
B
This process is often referred to as Cross-Talk Cancelation (CTC). Note that this
concerns a spatial or “modal” decorrelation10 only, just like other methods involving
FRF matrix inversion. It should therefore by no means be understood as some special
signal processing step.
In practice, to obtain T34 from FRFs as illustrated by Fig. 5.7a or b, one requires
sufficient excitations and indicator DoFs to respectively control and observe the
dynamics at the interface. However, the foremost reason for using a transmissibility-
based TPA method is to identify source path contributions without the need to conduct
a tedious FRF measurement campaign. The next section discusses how to obtain T34
from operational responses only.
The transmissibility matrix T34 can be estimated statistically from the correlation
between u3 and u4 when monitored under a variety of operational test conditions.
The fundamental assumption is that, by testing the vibration source under different
10 Thisdecorrelation has been approached in several ways: for instance, based on so-called Global
and Direct Transmissibility Functions (GTDT) in the work of Magrans (1981), Guasch and Magrans
(2004); Guasch (2009); Guasch et al. (2013) or along conventional FRFs as seen in Varoto and
McConnell (1998), Ribeiro et al. (1999, 2000), Maia et al. (2001).
5.5 Transmissibility-Based TPA 205
f1 (, , . . . )
Fig. 5.8 Operational TPA:
the source excitation f1 u4
depends on operational
u3
parameters (α, β, . . . ) such
T34
as engine speed or torque f1
u4
11 In Roozen and Leclère (2013), it is suggested that external excitation on the source structure,
e.g. by use of a non-instrumented hammer, leads to better conditioning of the transmissibility
matrix compared to a sequence of operational excitations. This resembles the approach depicted by
Fig. 5.7b.
12 The H estimator is a well-known principle in experimental modal analysis to determine FRFs
1
from a Multiple Input–Output (MIMO) test, see for instance, Bendat and Piersol (1980), Ewins
(2000). Alternative ways to obtain the transmissibility matrix have recently been explored, such as
the H2 or Hs estimator to balance the error contributions between the inputs and outputs (Leclère
et al. 2012).
206 5 Industrial Applications & Related Concepts
Equation (5.31) performs cross-talk cancelation similar to Eq. (5.27). Now that the
transmissibility matrix T34 has been estimated globally, the actual path contributions
can be analyzed for each measurement cycle. If performed correctly (i.e. all transfer
paths were included), the reconstructed response spectra equal the response spectra
that were measured directly.
The approach governed by Eqs. (5.30) and (5.32) is generally known as Oper-
ational Transfer Path Analysis (OTPA) (Noumura and Yoshida 2006). Clearly, the
performance of this method is subject to the choice and positioning of the indicator
sensors u4 , that function as “inputs” to the analysis. Too few sensors (or too distant
from the source connections) could lead to neglecting important transmission paths,
resulting in unrealistic prognoses (Gajdatsy et al. 2008). On the other hand, having
too many sensors (or too near to each other) complicates the inversion of S44 due to
poor conditioning, resulting in amplification of measurement noise (Gajdatsy et al.
2008).
Theoretically, the rank of U4 is limited by the number of DoFs at the interface m.
This implies that U4 and thus S44 is rank deficient, even if r > n. It is therefore com-
mon practice to perform a Principal Component Analysis (PCA) in order to identify
how many independent modes of vibration are present in the system and transferred
across the interfaces. Principal components are calculated from the singular value
decomposition of U4 . The left-singular vectors represent the vibration mode shapes
that build up to the vibrations in the indicator DoFs, sorted from the largest contribu-
tion to the smallest. By controlling the amount of principal components that are used
in the computation of the transmissibility by Eq. (5.32), the condition number can be
kept low such that one balances between the completeness of path descriptions and
attenuation of measurement noise. This is not discussed in detail here, guidelines can
be found in specific literature on OTPA (Noumura and Yoshida 2006; de Klerk and
Ossipov 2010; Putner et al. 2012).
Another strong advantage of OTPA is the ease of combining various types of
sensors, both for the input (indicator) and output (receiver) DoFs. Quantities such as
accelerations, velocities, sound pressures and even forces and strains can be used in
a mixed fashion, as long as proper scaling (unit normalization) is taken into account
(Noumura and Yoshida 2006). This makes OTPA a suitable method to quickly inves-
tigate the proportion of structure-borne and airborne contributions (de Klerk et al.
2009). Other related extensions or applications include Response Modification Anal-
ysis (RMA) for providing structural sensitivities to the target responses (Scheuren
and Lohrmann 2014), pass-by analysis of road vehicle noise (Putner et al. 2013;
ISO Technical Committee 2015) and time domain auralization of OTPA, sometimes
referred to as Transfer Path Synthesis (TPS) (Vorländer and Dietrich 2008).
5.6 Substructuring to Estimate Fixed-Interface Modes 207
where qfixt denotes the modal coordinates of the fixture. This expression can be
+ fixt
inverted to find the participation of each fixture mode, as qfixt ≈ φ fixt
f u , where
()+ denotes the Moore–Penrose pseudoinverse of the matrix. The estimate of the
fixture modal amplitudes is only meaningful here if one has at least as many mea-
surement points as modes of interest and if the measurement locations on the fixture
are chosen such that φ fixt
f has full column rank.
208 5 Industrial Applications & Related Concepts
If the modal basis of the fixture is sufficiently rich, then it will be adequate to span
the space of the motion of the fixture even after the structure is attached (i.e., of the
fixture+structure). If this is the case, then the fixture modal motions can be estimated
from the measured motions, ufixt , as follows.
+ +
qfixt = φ fixt
f u f = φ fixt
f φ f q, (5.34)
where q denotes the modal coordinates of the assembled fixture+structure. Our desire
is to estimate the modal parameters that the structure would have if it were attached
to a truly rigid fixture. One way to do this is to apply the constraints,
qfixt = 0 (5.35)
to the modes of the fixture+structure using the methods outlined in this course. In
terms of the modal coordinates of the fixture+structure, the constraint equations are
fixt +
φf uf = 0 (5.36)
or +
φ fixt
f φfq = 0 (5.37)
In some cases, it may not be feasible to perform a modal test on the fixture alone in
order to estimate φ fixt
f . There also may be situations in which the free modes of the
fixture do not form an efficient basis for the motion of the fixture+structure, so some
other basis must be considered. In these cases, one alternative is to use the dominant
singular vectors of the fixture mode shape matrix to form the constraints. The first
step is to perform a Singular Value Decomposition (SVD) of the fixture motions
observed on the fixture+structure
φ f = USVT (5.38)
u f = U N SV D S N SV D (V N SV D )T q, (5.39)
where the matrices now contain only the first N SV D columns or rows or both. The
following constraint will set this approximation to the fixture’s motion to zero,
(U N SV D )T u f = 0. (5.40)
This is an alternative to the constraints in Eq. 5.38 and avoids having to measure the
modes of the fixture alone.
The methods in the prior section can be extended and used to estimate Hurty/Craig–
Bampton models of subcomponents from measurements. In particular, for small
components a quantity of interest is the effective mass. In Sedaghati et al. (2003),
Mayes (2015), Mayes et al. (2013) and Mayes and Linehan (2014), a methodology
was developed and validated on several structures.
Consider a component mounted on a larger structure. If the structure is large, we
can consider the component to be subjected to a base acceleration that is independent
of the component’s motion. Then, the HCB model allows one to compute the effective
mass, which describes how much each of the fixed-base modes of the structure is
excited by the base acceleration. For example, if one considers a clamped–clamped
beam on a rigid foundation, then each symmetric mode would be excited by a uniform
210 5 Industrial Applications & Related Concepts
motion of the base but asymmetric modes would not be. Hence, the asymmetric
modes would have zero effective mass and the symmetric modes’ effective mass
would vary depending on their propensity to be excited; typically, the lower modes
have higher effective mass. Such a model can be effective when seeking to determine
if a subcomponent survives a given environment.
Effective mass models are usually derived from finite element models, but methods
were recently presented that can derive them experimentally (Mayes et al. 2013). To
do so, a test is performed with the component on a transmission simulator, with the
assembly free–free. Then the substructuring method in the prior section can be used to
constrain the transmission simulator to ground and estimate the fixed-interface modes
of the test article. Further processing steps yield the effective mass. See Sedaghati
et al. (2003), Mayes (2015), Mayes et al. (2013) and Mayes and Linehan (2014) for
further details.
Fig. 5.9 Rear subframe location in the car and two bushing types of particular interest
shown and the mass loading component can be seen in Fig. 5.10 together with a
small region of an FE representation of the mass-loaded bushing.
Two finite element models are used, one with and the other without mass-loaded
bushings. The FE models of both configurations each consist of more than half
a million degrees of freedom. The calibration procedure minimizes the deviation
between the FE model’s frequency response functions and the experimentally iden-
tified model’s ditto. Equalized damping is imposed on both models to circumvent
the mode pairing problem. A bootstrapping cross-validation is further used to quan-
tify the parameter uncertainties. A SIMO stepped-sine testing procedure was used
to obtain test data. The FE model with mass loading was calibrated first to obtain
good parameter estimation of the rubber’s Young’s modulus, i.e., all other model
parameters were then fixed at their nominal values. The configuration without mass
loading was then calibrated with fixed calibrated rubber parameters for an overall
good parameter estimation of the rear subframe FE model. In the final step, a subset
of all parameters was updated toward a higher frequency region for the configuration
without mass loading to gain further physical understanding of the behavior of the
component.
Experimental modal analysis test setup. The vibration tests were performed
in the Vibration Lab at the Chalmers University of Technology. Three nominally
identical rear subframes of a Volvo XC90 with and without mass loading were used to
obtain data for the spread between individuals. The testing procedure was the same for
all components. The rear subframe with and without mass loading is illustrated in Fig.
5.11. The components were hung in long thin lines (red and thickened for visibility
in Fig. 5.10) attached to a support structure via soft springs. The support modes were
all below 5 Hz and were thereby well below the first flexible eigenfrequency that
appeared above 40 Hz in both configurations.
With the target to obtain particularly good test data for the first 20 flexible body
modes, the position of 20 uniaxial sensor locations was found from a set of 170
candidate positions with a sensor placement pretest planning method (Kammer 2005;
212 5 Industrial Applications & Related Concepts
Fig. 5.10 Rear subframe with additional weights fitted to four bushings in instrumented test con-
figuration (left). CAD illustration of additional weight (right)
Fig. 5.11 Top view of configuration without additional weights in a and bottom view in b. Top view
of mass-loaded configuration in c and bottom view in d. Circles mark accelerometer locations and
a rectangle marks the excitation position. Excitation was normal to the surface, with a direct accel-
erance force and accelerometer sensor configuration. Green (uniaxial) and yellow (triaxial) circles
represent accelerometer locations common for both configurations. Cyan (without weights) and red
(with weights) markings indicate triaxial sensor positions that differ between the configurations
5.8 Calibration of an SUV Rear Subframe 213
Gibanica et al. 2016). Six more uniaxial accelerometer positions were added for
visualization purpose. Ten additional triaxial accelerometers were placed on the
bushings. All sensor placements can be seen in Fig. 5.11. In the colocated and colinear
force/acceleration sensor configuration, the accelerometer and force transducer was
placed on the opposite side of a thin sheet metal part of the subframe. Accelerometers
9–36 (numbering; see Fig. 5.11) were placed identically on both configurations while
accelerometers 1–8 were placed directly on the bushings in the configuration without
added weights and these were instead placed on the additional weights in the mass-
loaded configuration. The triaxial accelerometers were PCB 356A03 (1 gram, glued)
and the uniaxial were PCB 352C22 (0.5 gram, bees wax fastened). The accelerometer
masses were included in the FE models. The shaker was an LDS V201 attached to the
subframe via a metallic stinger of approximately 5 mm length. The excitation force
was measured with a Brüel&Kjær 8203 force sensor together with an B&K IEPE
converter 2647B, attached to the component through a glued stinger attachment plate
with 0.2 grams weight. The force transducer and stinger attachment plate masses were
not included in the FE models.
Two excitation methods were used for both configurations. Periodic chirp tests at
various amplitudes were performed to assess the linearity of the systems to obtain
proper excitation loads. Most test data used for calibration were collected with multi-
sine tests for diminishing the influence of noise in the model updating procedure.
For the mass-loaded configuration, a frequency range (40,500) Hz with 2000 discrete
frequencies was deemed sufficient due to the relatively strong damping. In the con-
figuration without added weights, a frequency range (60,500) Hz with 3000 discrete
frequency were used instead. The discrete frequencies were selected based on the
method specified in Vakilzadeh et al. (2015) to minimize system identification model
bias.
The calibration procedure used gives a parameter setting that minimizes the devi-
ation between the FE model and an identified state-space system model, and thus
indirectly from raw test data through the identified model. It is therefore crucial for
the system identification model to be of high quality. For the configuration without
mass loading, all accelerometers were used in the identification, from 60 to 350 Hz
with 20 states. Low- and high-frequency residual modes were added to the obtained
system and the state-space model’s B and C matrices were augmented with additional
data from chirp measurements for the low-frequency region from 20 to 60 Hz and
stepped-sine data for the high-frequency region from 350 to 470 Hz. This produced a
system that accurately predicted the system’s resonance and antiresonance behavior.
For the mass-loaded configuration, only the bushing accelerometers, accelerom-
eters 1 through 8 in Fig. 5.11, were used initially for the system identification due to
the many local bushing modes identified from the measurements, not visible in other
channels. Thus an identified system with 40 states from 40 to 225 Hz was identified
for 24 channels (8 first triaxial accelerometers). A low- and high-frequency resid-
ual system pole was added to the system and additional data from a chirp test from
20 to 40 Hz and stepped-sine test from 225 to 250 Hz was used to reestimate the
system’s B and C matrices by the procedure given in Sect. 4.9.1. In addition, all 36
214 5 Industrial Applications & Related Concepts
Fig. 5.12 Parameterization of the FE models with E representing stiffness and t thickness, in a
for the mass loaded configuration and in b, c, and d for the configuration without mass loading.
Parameters p1 through p4 were also parameterized in the configuration without mass loading
other parameters were at this stage fixed to their nominal values, where parameters
p12 to p14 were set to 1.20 mm. These four parameters were manually updated to
a nominal setting of 1.40 mm in the calibration of the configuration without mass
loading.
The configuration without mass loading was first calibrated from 60 to 265 Hz,
with updated bushing parameters, by updating parameters p5 to p11 . Updated param-
eters p5 to p11 were then used as nominal parameters in the final calibration from 60
to 350 Hz where parameters p1 , p2 , p10 , p13 , p14 and p15 were updated. Parameter
p12 was kept fixed in all three calibrations.
In Table 5.1, it can be seen that the mean value of the cross-validation results, based
on 20 data splits with raw test data and unconstrained minimization, differ somewhat
from the deterministic calibrated estimates. There are significant noise levels in the
raw test data, especially noticable around the low-frequency antiresonances for some
channels (see Figs. 5.13 and 5.14) that might affect the cross-validation results.
Furthermore, the identified model used in the calibration does not perfectly fit test
data in all channels, especially around the antiresonances (see again Fig. 5.13), and a
difference between calibration and cross-validation results is therefore to be expected.
The parameters show little variation, as indicated by the COV in Table 5.1, with
parameters p9 , p10 , and p11 for calibration FE265 varying more than than the others,
approximately from 3 to 6%.
It can be seen in Table 5.1 that the two calibrated stiffness parameters p1 and
p2 in the mass-loaded configuration, (see Fig. 5.12a), are considerably higher than
p3 and p4 . This is a consequence of making the FE model fit experimental data so
that the FE modes as shown in Fig. 5.15 fit the eigenfrequency order of that was
216 5 Industrial Applications & Related Concepts
Table 5.1 Nominal parameters (Nom) along with three calibration results. Calibrated parame-
ters for the mass-loaded configuration denoted F E mass , configuration without mass loading with
calibration up to 265 Hz denoted F E 265 and up to 350 Hz denoted F E 350 . The mean value and coef-
ficient of variation (%) of the cross-validation is denoted with superscripts μ and COV. Empty fields
(–) represent fixed parameters in that particular calibration and left arrows (←) point to numerical
data used in that calibration. E represent Young’s modulus and t denote plate thickness
μ COV F E μ C O V F E COV F E COV
# Param Nom F E mass F E 265 F E 350 F E mass F E mass 265 F E 265 350 350
p1 E 5.00 4.01 ← 4.62 3.67 0.72 – – 4.98 0.39
(MPa)
p2 E 5.00 3.82 ← 7.01 3.48 0.85 – – 5.25 0.11
p3 E 5.00 2.83 ← 5.83 2.69 0.94 – – – –
p4 E 5.00 2.41 ← 5.41 2.94 1.19 – – – –
p5 E 210. ← 192.3 ← – – 196.4 0.99 – –
(GPa)
p6 E 210. ← 208.0 ← – – 198.9 0.13 – –
p7 E 210. ← 240.7 ← – – 228.6 0.09 – –
p8 E 210. ← 222.4 ← – – 249.1 1.73 – –
p9 E 210. ← 250.8 ← – – 264.4 3.86 – –
p10 E 210. ← 166.6 ← – – 162.6 3.10 189.2 0.10
p11 t (mm) 1.80 ← 1.46 ← – – 1.35 5.81 – –
p12 t 1.40 1.20 1.40 ← – – – – – –
p13 t 1.40 1.20 1.40 3.67 – – – – 2.56 0.71
p14 t 1.40 1.20 1.40 1.81 – – – – 1.75 1.12
p15 t 1.40 1.20 1.40 1.61 – – – – 1.80 0.74
Fig. 5.13 Cross-accelerance for input at position 24 and output at 4x. Note that raw test data and
data from identified model almost are almost coincident
5.8 Calibration of an SUV Rear Subframe 217
Fig. 5.15 The nominal model with mass-loaded bushings (a) along with the 11 first calibrated
flexible modes, from mode 1 (b) to mode 11 (l). Blue indicate small motion and red large motion
In Table 5.2, the eigenfrequencies of the first eleven flexible modes are shown
for the configuration with and without mass loading, along with their masses. It can
be seen that for the mass-loaded configuration the nominal results are very different
from what is experimentally identified. It should be noted that mode 2 is very close
to mode 1 in the experimental data and could not be identified, as indicated by the
FE modes show in Fig. 5.15b, c. The same holds for modes 6 and 7, shown in Fig.
5.15g and h. FE results indicate that modes 4 and 5, and also modes 8 and 9 almost
coalesce, as seen from their mode shapes shown in Fig. 5.15. The calibrated F E mass
has a much better correspondence to experimental data compared to the nominal
model F E nom mass which is also indicated by the deviation metric. The deviation
metric between the experimental data E x pmass and the nominal FE model F E nom
is 0.90 and 1.02 for the calibration and cross-validation models, respectively. The
calibrated model gives a deviation metric of 0.44 and 0.50 for the calibration and
cross-validation data, respectively. Thus an improvement of 51% is seen in both the
calibration and cross-validation metrics.
Table 5.2 indicate that eigenfrequencies of the nominal configuration without
mass loading, F E nom , correlates rather well with those of the experimental model.
However, the modal assurance criterion (MAC, see Ref. Allemang and Brown 1982)
analysis results reported in Fig. 5.17 show that there are mode switches between
modes 3 and 4, and modes 5 and 6. Also, some modes show a very low MAC
correlation and the last two modes have a high cross-correlation. The updated model
F E 350 , for which the MAC analysis is shown in Fig. 5.17b, shows a much better MAC
correlation, even for the modes above 300 Hz. It can be noted that the FE analysis
indicates that there is a mode, not identified in the experiments, around 290 Hz. It
is shown in Fig. 5.16j where it can be seen that in that mode the motion is strongly
5.8 Calibration of an SUV Rear Subframe 219
Fig. 5.16 The nominal model without mass-loaded bushings (a) along with the 11 first calibrated
flexible modes, from mode 1 (b) to mode 11 (l)
Fig. 5.17 MAC correlation between the experimentally identified system and the nominal FE model
F E nom (left) and identified system and the calibrated FE model F E 350 (right)
localized to one bushing (upper left). It is therefore just marginally controllable from
the input position far away from the bushing. Also, due to strong damping in the
bushing modes, it would be very hard to identify it from the raw frequency response
data of the used sensor set. Further in Fig. 5.16, it can be seen that in the two modes
above 300 Hz that the bushings are considerably activated. In order to capture the
behavior of the these modes, the stiffness parameters p3 and p4 of the rubber bushings
were again used as free parameters in the last calibration step, F E 350 . A justification
for this is that rubber has frequency dependent stiffness properties, i.e., its stiffness
increases with frequency (see Ref. Jones 2001). In the mass-loaded configuration,
the calibration focus is more on the lower frequency spectrum as compared to that
of the calibration without mass loading.
220 5 Industrial Applications & Related Concepts
The intermediate calibration step F E 265 does not show a good correlation for
modes above the 8:th flexible modes, as seen in Table 5.2. This is due to that modes
above 265 Hz are not included in the model calibration.
The deviation metric for the intermediate model F E 265 was improved by 60%
(from 0.91 to 0.38) after the calibration. The validation metric was improved with
23% (from 1.25 to 0.97). It should be noted that the model from which F E 265 was
calibrated started from a model with updated bushing parameters from the previous
calibration step, and that this is reflected in the deviation metric. The deviation metric
for the final model calibration F E 350 was improved with 71% (from 0.88 to 0.26)
for the calibration. The calibration of this model started from the parameter settings
of model F E 265 , with exception of parameters p3 and p4 that was set according
to above. The validation metric was improved with 33% (from 1.20 to 0.81). The
improvement in the validation metric is lower compared to that of the calibration
metric. One reason for this is likely that poorer frequency response test data were
used as validation data. In Fig. 5.13, one of the better channels is shown for raw test
data, the identified model, the nominal FE model, and the calibrated FE model up to
265 and 350 Hz. Figure 5.14 depicts one of the poorer channel estimates.
Fig. 5.18 Body in White chassis of a Volvo XC90 resting on rubber cushions (at yellow arrows)
ready for vibration testing. Loose rear subframe resting on lab floor in front
Fig. 5.19 Location of dummy adapters C1 through C4 under rear part of BiW (left). Location of
virtual points C1 through C4 on RSF fitted to BiW (right)
Fig. 5.20 Location of BiW-mounted accelerometers 1 through 22. Coupling adapters C1 through
C4 have three triaxial accelerometers and are subjected to three successive SIMO stimuli tests each.
Accelerometer numbers in rectangles also indicate location of validation stimuli
222 5 Industrial Applications & Related Concepts
Fig. 5.21 Subframe with positions of accelerometers marked. Circle markings are for triaxial
accelerometers and the rectangular marking is for one uniaxial accelerometer and the input force
position
Fig. 5.22 Adapters C1 through C4 equipped with triax accelerometers (left). Location vectors r1
through r3 indicate kinematical transformation to virtual point C1 (right)
each to provide sufficient information about the motion of the stiff adapters in a
stepped-sine vibration testing. Besides the 4 × 3 × 3 = 36 sensor channels of the
triaxial accelerometers, an additional 22 uni-axis accelerometers were spread out
over the chassis (see Fig. 5.20). The signals of the triaxial accelerometers were trans-
formed to provide estimates of the translational and rotational motion of so-called
virtual points associated with each adapter (see Fig. 5.22). Its three translation and
three rotations were obtained by least squares fitting and kinematic transformations
on the nine signals from its sensors. The system stimulus was provided with a roving
shaker that provided 4 × 3 = 12 force inputs to the adapters in 12 repeated SIMO
experiments to give frequency response data. Two additional excitation points were
used for validation purpose. The chassis was put on soft rubber cushions to mimic
free–free boundary conditions during testing. Test data indicated at least 136 flexible
body modes below 200 Hz for the BiW.
5.9 Coupling of Two Major SUV Components 223
Fig. 5.23 Experimental FRF and ditto synthesized from identified state-space model of BiW
The testing procedure outlined in Sect. 5.8 was used to get data for the RSF in two
configurations, one with free bushings and one with mass-loaded bushings. Testing
with mass-loaded rubber bushings was crucial for obtaining good calibration for the
rubber material properties. Since the main purpose of the bushings is to provide good
vibration isolation it is safe to say that the material parameters of the rubber are of
utmost importance for a successful coupling procedure. Since the rubber bushings
are parts of the RSF, the translational and rotational dynamic stiffnesses of these parts
are accounted for. The virtual points for coupling are included as nodes in the finite
element model and provide input and output information of the FE-based state-space
model of this component. Test data indicated about 40 flexible body modes below
200 Hz for the RSF.
Make state-space system identification of BiW. Preliminary identification
results indicated that direct accelerance FRF data were particularly troublesome
in the state-space model fitting. As the direct dynamical stiffness associated to the
coupling degrees of freedom plays an important role in the coupling, increased effort
was put in obtaining a good state-space model for these. A procedure given in Ref.
Gibanica et al. (2018) was followed. The procedure is based on observation from
direct accelerance measurements. These observations indicate that strong outside-
band eigenmodes strongly affects the frequency response within the frequency band
of interest in the testing. In comparison with more traditional approaches that com-
pensate for outside-band low frequency modes with a residual mass model and high
frequency modes with a residual stiffness model, the procedure is to add virtual sys-
tem poles (residual states) just outside the frequency range of interest to compensate
for the effect of outside-band modes just above the upper frequency of interest. The
effect of residual states was identified by linear regression to test data. The influ-
ence of these states was first removed from test data before a traditional State-space
Subspace Identification method (N4SID) was applied and then superimposed to the
identified model to get a higher order state-space model with a very good fit to data,
see Fig. 5.23 for one specific transfer function. An assessment of all transfer functions
is reported in Fig. 5.24 which utilizes the FRAC correlation metric defined as
K
K
K
FRACi j = | HiAj (ωk )(HiXj (ωk )) H |2 /( |HiAj (ωk )|2 |HiXj (ωk )|2 ) (5.41)
k=1 k=1 k=1
224 5 Industrial Applications & Related Concepts
Fig. 5.24 Model fit (by FRAC) of BiW state-space model in range (30,200) Hz. Stimuli locations
are three forces at each cylinder C1 through C4 and at locations 21 and 22. Responses stem from
same locations but also from accelerometers 1 through 20
Fig. 5.25 Model fit (by FRAC) of coupled model in range (30,200) Hz
for which a perfect fit between test data from the tested assembly and the synthesized
system would render a value of 1 (one).
Physically motivated constraints on reciprocity were imposed by averaging off-
diagonal elements of the raw test data transfer function matrices. Also, in order to
fully couple components that share the same translation and rotational motion at
their mutual interfaces, all translation and rotation outputs and corresponding force
and moment inputs must be available at the coupling points. Since the test lacks
moment input, the transfer functions related to moment input have to be estimated
by through a modal system expansion, see paper E of Liljerehn’s doctoral thesis
(Liljerehn 2016). For comparison, the coupling was performed with and without the
rotational couplings, see Fig. 5.26.
Make state-space system identification of RSf and calibrate its FE model. The
system identification from RSf data and the proceeding calibration of the RSf finite
element model were conducted along the lines described in Sect. 5.8. In particular,
the stiffness parameters of the rubber bushings were much reduced in comparison to
the nominal parameter configuration.
References 225
Fig. 5.26 An example of an FRF of the coupled system; the receptance from input at 22 to response
at 12. TR denotes results obtained with full coupling (translations and rotations) while T denotes
results obtained with translational coupling only
References
Cerrato, G. (2009). Automotive sound quality – powertrain, road and wind noise. Sound and Vibra-
tion, 43(4), 16–24. http://www.sandv.com/downloads/0904cerr.pdf.
Choi, H. G., Thite, A. N., & Thompson, D. J. (2007). Comparison of methods for parameter selection
in Tikhonov regularization with application to inverse force determination. Journal of Sound and
Vibration, 304(3), 894–917. https://doi.org/10.1016/j.jsv.2007.03.040.
D’Ambrogio, W., & Fregolent, A. (2010). The role of interface DoFs in decoupling of substructures
based on the dual domain decomposition. Mechanical Systems and Signal Processing, 24(7),
2035–2048. https://doi.org/10.1016/j.ymssp.2010.05.007.
de Klerk, D., & Ossipov, A. (2010). Operational transfer path analysis: Theory, guidelines and tire
noise application. Mechanical Systems and Signal Processing, 24(7), 1950–1962. https://doi.org/
10.1016/j.ymssp.2010.05.009. Special Issue: ISMA 2010.
de Klerk, D., & Rixen, D. J. (2010). Component transfer path analysis method with compensation
for test bench dynamics. Mechanical Systems and Signal Processing, 24(6), 1693–1710. https://
doi.org/10.1016/j.ymssp.2010.01.006.
de Klerk, D., Rixen, D. J., & Voormeeren, S. N. (2008). General framework for dynamic substruc-
turing: History, review and classifcation of techniques. AIAA Journal, 46(8), 1169–1181. https://
doi.org/10.2514/1.33274.
de Klerk, D. (2009). Dynamic response characterization of complex systems through operational
identification and dynamic substructuring. Ph.D. thesis, Delft University of Technology, The
Netherlands. http://resolver.tudelft.nl/uuid:2e4dd63e-68d5-41a4-8201-1e784f2f752c
de Klerk, D., Lohrmann, M., Quickert, M., & Foken, W. (2009). Application of operational transfer
path analysis on a classic car. In DAGA. Rotterdam.
Dobson, B. J., & Rider, E. (1990). A review of the indirect calculation of excitation forces from
measured structural response data. Proceedings of the Institution of Mechanical Engineers,
Part C: Journal of Mechanical Engineering Science, 204(2), 69–75. https://doi.org/10.1243/
PIME_PROC_1990_204_080_02.
dos Santos, F. L. M., Peeters, B., Lau, J., & Desmet, W., Góes, L. C. S. (2014). An overview of exper-
imental strain-based modal analysis methods. In Proceedings of the International Conference on
Noise and Vibration Engineering (ISMA). Leuven, Belgium
Elliott, A. S. (2009). Characterisation of structure borne sound source in-situ. Ph.D. thesis, Univer-
sity of Salford. http://usir.salford.ac.uk/id/eprint/26661.
Elliott, A. S., & Moorhouse, A. T. (2008). Characterisation of structure borne sound sources from
measurement in-situ. Journal of the Acoustical Society of America, 123(5), 3176. https://doi.org/
10.1121/1.2933261.
Elliott, A. S., Moorhouse, A. T., Huntley, T., & Tate, S. (2013). In-situ source path contribution
analysis of structure borne road noise. Journal of Sound and Vibration, 332(24), 6276–6295.
https://doi.org/10.1016/j.jsv.2013.05.031.
Ewins, D. J. (2000). Modal testing: Theory, practice and application (2nd ed.). New York: Wiley
Inc.
Fahy, F. J. (1995). The vibro-acoustic reciprocity principle and applications to noise control. Acta
Acustica United with Acustica, 81(6), 544–558.
Fahy, F. J. (2003). Some applications of the reciprocity principle in experimental vibroacoustics.
Acoustical Physics, 49(2), 217–229. https://doi.org/10.1134/1.1560385.
Gajdatsy, P., Janssens, K., Gielen, L., Mas, P., & Van der Auweraer, H. (2008). Critical assessment
of operational path analysis: effect of neglected paths. In Proceedings of the XV International
Congress on Sound and Vibration (ICSV) (pp. 1090–1097). Daejeon, Korea.
Gajdatsy, P., Janssens, K., Gielen, L., Mas, P., & Van Der Auweraer, H. (2008). Critical assessment
of operational path analysis: Effect of coupling between path inputs. Journal of the Acoustical
Society of America, 123(5), 3876. https://doi.org/10.1121/1.2935777.
Gaudin, A., & Beniguel, J. F. (2012). Low frequency road noise decomposition at wheel center on
an roller bench. In Acoustics 2012. Société Française d’Acoustique, Nantes, France. https://hal.
archives-ouvertes.fr/hal-00811236.
References 227
Gibanica, M., & Abrahamsson, T. J. S. (2017). Parameter estimation and uncertainty quantification
of a subframe with mass loaded bushings. In R. B. (Ed.), Proceedings of 32nd IEEE Conference
on Decision and Control (Vol. 3, pp. 61–76). Berlin: Springer International Publishing.
Gibanica, M., Abrahamsson, T. J. S., & Allemang, R. J. (2018). Residual states for modal models
indentified from accelerance data. In: M. Maines & B. Dilworth (Eds.), Topics in modal analysis
and testing (Vol. 9). Berlin: Springer International Publishing.
Gibanica, M., Abrahamsson, T. J. S., Kammer, D. C. (2016). Redundant information rejection in
sensor localisation using system gramians. In: M. Mains (ed.) Topics in modal analysis and testing
(Vol. 10). Berlin: Springer International Publishing.
Ginsberg, J. H. (2001). Mechanical and structural vibrations (1st ed.). New York: Wiley.
Guasch, O. (2009). Direct transfer functions and path blocking in a discrete mechanical system.
Journal of Sound and Vibration, 321(3), 854–874. https://doi.org/10.1016/j.jsv.2008.10.006.
Guasch, O., & Magrans, F. X. (2004). The global transfer direct transfer method applied to a finite
simply supported elastic beam. Journal of Sound and Vibration, 276(1), 335–359. https://doi.
org/10.1016/j.jsv.2003.07.032.
Guasch, O., García, C., Jové, J., & Artís, P. (2013). Experimental validation of the direct transmis-
sibility approach to classical transfer path analysis on a mechanical setup. Mechanical Systems
and Signal Processing, 37(1), 353–369. https://doi.org/10.1016/j.ymssp.2013.01.006.
Harrison, M., Sykes, A. O., & Martin, M. (1952). Wave effects in isolation mounts. Journal of the
Acoustical Society of America, 24(1), 62–71. https://doi.org/10.1121/1.1906850.
Hixson, E. L. (1961). Mechanical impedance and mobility. In: C. M. Harris & C. E. Crede (Eds.),
Shock and vibration handbook (1st ed., pp. 1–59), Chap. 10. New York: McGraw-Hill.
ISO Technical Committee 43/Subcommittee 1/Workgroup 22 (ISO/TC43/SC1/WG22) (2015).
Measurement of noise emitted by accelerating road vehicles – Engineering method – Part 1:
M and N categories. ISO 362-1, International Standards Organisation.
ISO Technical Committee 43/Subcommittee 1/Workgroup 22 (ISO/TC43/SC1/WG22). (1996).
Acoustics – Characterization of sources of structure-borne sound with respect to sound radi-
ation from connected structures – Measurement of velocity at the contact points of machinery
when resiliently mounted. ISO 9611, International Standards Organisation.
Janssens, M. H. A., & Verheij, J. W. (2000). A pseudo-forces methodology to be used in character-
ization of structure-borne sound sources. Applied Acoustics, 61(3), 285–308. https://doi.org/10.
1016/S0003-682X(00)00035-9.
Janssens, M. H. A., Verheij, J. W., & Thompson, D. J. (1999). The use of an equivalent forces
method for the experimental quantification of structural sound transmission in ships. Journal of
Sound and Vibration, 226(2), 305–328. https://doi.org/10.1006/jsvi.1999.2303.
Janssens, M. H. A., Verheij, J. W., & Loyau, T. (2002). Experimental example of the pseudo-forces
method used in characterisation of a structure-borne sound source. Applied Acoustics, 63(1),
9–34. https://doi.org/10.1016/S0003-682X(01)00023-8.
Jones, D. I. G. (2001). Handbook of viscoelastic vibration damping. Chichester: Wiley.
Kammer, D. C. (1998). Input force reconstruction using a time domain technique. Journal of Vibra-
tion and Acoustics, 120(4), 868–874. https://doi.org/10.1115/1.2893913.
Kammer, D. C. (2005). Sensor set expansion for modal vibration testing. Mechanical Systems and
Signal Processing, 19, 700–713.
Kay, S. M. (1993). Fundamentals of statistical signal processing (Vol. I: Estimation Theory, 1st
ed.). Englewood Cliffs: Prentice Hall.
Klerk, D. D., Rixen, D. J., & Voormeeren, S. N. (2008). General framework for dynamic substruc-
turing: History, review and classification of techniques. AIAA Journal, 46(5), 1169–1181.
Leclère, Q., Roozen, N. B., & Sandier, C. (2012). Experimental estimation of transmissibility
matrices. In Proceedings of the International Conference on Noise and Vibration Engineering
(ISMA). Leuven, Belgium.
Liljerehn, A.: Machine tool dynamics - a constrained state-space substructuring approach. Ph.D.
thesis, Chalmers University of Technology, Dept Applied Mechanics (2016)
228 5 Industrial Applications & Related Concepts
Madsen, M. B. (2014). Electrical power assisted steering – dynamic source strength characteristic
and vehicle NVH prediction. Master’s thesis, University of Southern Denmark.
Magrans, F. X. (1981). Method of measuring transmission paths. Journal of Sound and Vibration,
74(3), 321–330. https://doi.org/10.1016/0022-460X(81)90302-3.
Maia, N. M. M., Silva, J. M. M., & Ribeiro, A. M. R. (2001). The transmissibility concept in multi-
degree systems. Mechanical Systems and Signal Processing, 15(1), 129–137. https://doi.org/10.
1006/mssp.2000.1356.
Mas, P., Sas, P. (1994). Indirect force identification based upon impedance matrix inversion: A study
on statistical and deterministical accuracy. In: Proceedings of the International Conference on
Noise and Vibration Engineering (ISMA). Leuven, Belgium.
Mayes, R. (2015). A modal Craig-Bampton substructure for experiments, analysis, control and
specifications.
Mayes, R., & Linehan, D. (2014). Measuring effective mass of a circuit board.
Mayes, R., Schoenherr, T., Blecke, J., & Rohe, D. (2013). Efficient method of measuring effective
mass of a system.
Mondot, J. M., & Petersson, B. A. T. (1987). Characterization of structure-borne sound sources: The
source descriptor and the coupling function. Journal of Sound and Vibration, 114(3), 507–518.
https://doi.org/10.1016/S0022-460X(87)80020-2.
Moorhouse, A. T., Elliott, A. S., & Evans, T. A. (2009). In situ measurement of the
blocked force of structure-borne sound sources. Journal of Sound and Vibration, 325(4–
5), 679–685. https://doi.org/10.1016/j.jsv.2009.04.035. http://www.sciencedirect.com/science/
article/pii/S0022460X09003794.
Moorhouse, A. T. (2001). On the characteristic power of structure-borne sound sources. Journal of
Sound and Vibration, 248(3), 441–459. https://doi.org/10.1006/jsvi.2001.3797.
Norton, E. L. (1926). Design of finite networks for uniform frequency characteristic. Technical
Report TM26–0–1860, Bell Laboratories.
Noumura, K., & Yoshida, J. (2006). A method of transfer path analysis for vehicle interior sound
with no excitation experiment. In Proceedings of FISITA 2006 World Automotive Congress,
F2006D183. JSAE. http://www.fisita.com/publications/papers?id=4748
On, F. J. (1967). Mechanical impedance analysis for lumped parameter multi-degree of
freedom/multi-dimensional systems. Technical report, NASA Technical Note D-3865 (1967)
Penne, F. (2004). Shaping the sound of the next-generation BMW. In Proceedings of the Interna-
tional Conference on Noise and Vibration Engineering (ISMA) (pp. 25–39). Leuven, Belgium.
Petersson, B. A. T., & Gibbs, B. M. (1993). Use of the source descriptor concept in studies of
multi-point and multi-directional vibrational sources. Journal of Sound and Vibration, 168(1),
157–176. https://doi.org/10.1006/jsvi.1993.1367.
Petersson, B. A. T., & Gibbs, B. M. (2000). Towards a structure-borne sound source characterization.
Applied Acoustics, 61(3), 325–343.
Plunkett, R. (1958). Interaction between a vibratory machine and its foundation. Noise Control,
4(1), 18–22.
Plunt, J. (2005). Finding and fixing vehicle NVH problems with transfer path analysis. Sound and
Vibration, 39(11), 12–16.
Powell, R. E., & Seering, W. (1984). Multichannel structural inverse filtering. Journal of Vibration
and Acoustics, 106(1), 22–28. https://doi.org/10.1115/1.3269147.
Putner, J., Fastl, H., Lohrmann, M., Kaltenhauser, A., & Ullrich, F. (2012). Operational transfer
path analysis predicting contributions to the vehicle interior noise for different excitations from
the same sound source. In InterNoise12, no. 3 in INTER-NOISE and NOISE-CON Congress and
Conference Proceedings (pp. 2336–2347). New York City, NY. http://www.ingentaconnect.com/
content/ince/incecp/2012/00002012/00000009/art00034.
Putner, J., Lohrmann, M., & Fastl, H. (2013). Contribution analysis of vehicle exterior noise with
operational transfer path analysis. Journal of the Acoustical Society of America, 133(5), 3323.
https://doi.org/10.1121/1.4805558.
References 229
Reuss, P., Zeumer, B., Herrmann, J., & Gaul, L. (2012). Consideration of interface damping in
dynamic substructuring. In Topics in Experimental Dynamics Substructuring and Wind Turbine
Dynamics, Proceedings of the 30th IMAC, A Conference on Structural Dynamics (Vol. 2, pp.
81–88) Chap. 10. New York: Springer. https://doi.org/10.1007/978-1-4614-2422-2_10.
Ribeiro, A. M. R., Maia, N. M. M., & Silva, J. M. M. (1999). Experimental evaluation of the
transmissibility matrix. In Proceedings of the XVII International Modal Analysis Conference
(IMAC). Kissimmee, FL
Ribeiro, A. M. R., Silva, J. M. M., & Maia, N. M. M. (2000). On the generalisation of the trans-
missibility concept. Mechanical Systems and Signal Processing, 14(1), 29–35. https://doi.org/
10.1006/mssp.1999.1268.
Rixen, D. J., Boogaard, A., van der Seijs, M. V., van Schothorst, G., & van der Poel, T. (2015).
Vibration source description in substructuring: A theoretical depiction. Mechanical Systems and
Signal Processing, 60–61, 498–511. https://doi.org/10.1016/j.ymssp.2015.01.024.
Roozen, N. B., & Leclère, Q. (2013). On the use of artificial excitation in operational transfer path
analysis. Applied Acoustics, 74(10), 1167–1174. https://doi.org/10.1016/j.apacoust.2013.04.011.
Rubin, S. (1967). Mechanical immittance- and transmission-matrix concepts. Journal of the Acous-
tical Society of America, 41(5), 1171–1179. https://doi.org/10.1121/1.1910455.
Sachse, D., Geluk, T., van Wayenberge, T. (2013). Accuracy of inverse load identification techniques
for transfer path analysis. In Automotive Acoustics Conference 2013, 2nd International ATZ
Conference. ATZlive
Scheuren, J., & Lohrmann, M. (2014). Transfer path analysis – experiences, expectations and
perspectives. In International Noise and Vibration Colloquium. SAE Brazil.
Sedaghati, R., Soucy, Y., & Etienne, N. (2003). Experimental estimation of effective mass for
structural dynamics and vibration applications.
Sjövall, P., & Abrahamsson, T. (2008). Substructure system identification from coupled system test
data. Mechanical Systems and Signal Processing, 22(1), 15–33. https://doi.org/10.1016/j.ymssp.
2007.06.003.
Snowdon, J. C. (1979). Vibration isolation: use and characterization. Journal of the Acoustical
Society of America, 66(5), 1245–1274. https://doi.org/10.1121/1.383546.
Soliman, J., & Hallam, M. (1968). Vibration isolation between non-rigid machines and non-
rigid foundations. Journal of Sound & Vibration, 8(2), 329–351. https://doi.org/10.1016/0022-
460X(68)90236-8.
Sturm, M., Moorhouse, A. T., Alber, T., & Li, F. F. (2012). Force reconstruction using an adaptive
algorithm in time domain. In Proceedings of the International Conference on Noise and Vibration
Engineering (ISMA) (, pp. 17–19). Leuven, Belgium.
Sturm, M., Moorhouse, A. T., Kropp, W., & Alber, T. (2013). Robust calculation of simultaneous
multi-channel blocked force signatures from measurements made in-situ using an adaptive algo-
rithm in time domain. In Proceedings of the XX International Congress on Sound and Vibration
(ICSV) (, vol. 2, pp. 1610–1617). Bangkok, Thailand.
Ten Wolde, T. (1973). Reciprocity experiments on the transmission of sound in ships. Ph.D. thesis,
Delft University of Technology, The Netherlands. http://resolver.tudelft.nl/uuid:6d3d9f26-e631-
4043-a3ab-04bd45edde98
Ten Wolde, T. (2010). Reciprocity measurements in acoustical and mechano-acoustical systems.
Review of theory and applications. Acta Acustica United with Acustica, 96(1), 1–13. https://doi.
org/10.3813/AAA.918250
Ten Wolde, T., Verheij, J. W., & Steenhoek, H. F. (1975). Reciprocity method for the measurement
of mechano-acoustical transfer functions. Journal of Sound and Vibration, 42(1), 49–55. https://
doi.org/10.1016/0022-460X(75)90301-6.
Thévenin, M. L. (1883). Sur un nouveau théoreme d’électricité dynamique [On a new theorem
of dynamic electricity]. Comptes Rendus hebdomadaires des des Séances de l’Académie des
Sciences, 97, 159–161.
230 5 Industrial Applications & Related Concepts
Voormeeren, S., & Rixen, D. (2012). A family of substructure decoupling techniques based on a
dual assembly approach. Mechanical Systems and Signal Processing, 27(18), 379. https://doi.
org/10.1016/j.ymssp.2011.07.028.
Vorländer, M., & Dietrich, P. (2008). Transfer path analysis and synthesis for auralization. In 39th
Congreso Espanol de Acústica 2008. Coimbra, Portugal.
Weijtjens, W., De Sitter, G., Devriendt, C., & Guillaume, P. (2014). Operational modal parameter
estimation of mimo systems using transmissibility functions. Automatica, 50(2), 559–564. https://
doi.org/10.1016/j.automatica.2013.11.021.
Zeller, P. (2009). Handbuch Fahrzeugakustik. Springer,. https://doi.org/10.1007/978-3-8348-8657-
6.
Chapter 6
Model Reduction Concepts
and Substructuring Approaches
for Nonlinear Systems
Abstract This chapter reviews common nonlinearities that are encountered in engi-
neering structures, with a particular emphasis on geometric nonlinearity. Popular
ways to construct reduced order models for geometrically nonlinear problems are
discussed. The concept of nonlinear normal modes is presented to help understand the
dynamics of these structures, and some recently presented substructuring methods
are reviewed. —Chapter Authors: Matt Allen & Paolo Tiso.
Up to here, all the developed theory considered systems in their linearized form—that
is, the interest focused on small dynamic deviations from the equilibrium. In fact,
any engineering system is inherently nonlinear, and is often just designed to operate
in the linear regime. In other cases, one might want to exploit nonlinear responses to
achieve a certain performance. In either case, the nonlinearities need to be properly
assessed with the aid of numerical models. Nonlinearities can be classified into three
broad categories, namely
1. Boundary condition nonlinearities, arising from changing constraints to the sys-
tem. Typically, this involves contact between subcomponents of the same system,
between two or more different systems, or the system and an external boundary.
2. Material nonlinearities, which are due to any nonlinearity present in the consti-
tutive law, which relates strain and stress rates. Typical examples are plasticity,
hyperelasticity, progressive damage, etc.
3. Geometric nonlinearities, arising from the redirection of the internal stresses
due to finite rotations. This type of nonlinearity is particularly present in thin-
walled, lightweight constructions, where geometric nonlinearities cause coupling
between bending and stretching deformations. Likewise, geometric nonlinearities
arise in multibody systems, where the components are subjected to holonomic
constraints that establish nonlinear mappings between displacements and gener-
alized coordinates.
Being the realm of nonlinearities so vast, one cannot devise a general method for
the efficient reduction of any given nonlinear system. Here, we limit ourselves to
the relevant case of geometric nonlinearities. Apart from the richness in dynamic
behavior, one could expect from a geometrically nonlinear system, it is important to
highlight here that geometric nonlinearities are smooth with respect to the generalized
displacements. This important property is a key aspect for many of the reduction
methods discussed in this chapter.
6.1.1 2D Beam
In this section, we briefly review the nonlinear beam theory in 2D, and discuss
a widely used approximation, namely the von Karman kinematic model (i.e. see
Figs. 6.1 and 6.2). We start by assuming a straight, two-dimensional beam, with a
rigid cross section, able to undergo arbitrary large displacements and rotations. The
material is assumed linear and elastic, and the elastic strains small. The interested
reader could refer, for instance, to Woinowski-Krieger (1950).
The displacement field is indicated by U = [u; w], where u and w are axial and
transversal displacements, respectively. Upon an arbitrary deformation of the beam,
we can write
1 T
U = uex + wez + [R(θ ) − I]zez = F F−I , (6.1)
2
where R(θ ) is the rotation matrix of the cross section, I is the identity matrix, and F
is the deformation gradient. The Green–Lagrange strain tensor is then written as
1
= ∇U + ∇ T U + ∇ T U · ∇U . (6.2)
2
6.1 Geometric Nonlinearities 235
The polar decomposition theorem allows to split the deformation gradient F into
rotations and deformations as
1
F = R(θ )(I + L) ⇒ = L + LT + LT L . (6.3)
2
Linearizing the strains leads to
1 e − zk γ
≈ L + LT = , (6.4)
2 γ 0
where the axial strain e, the shear strain γ and the curvature k are, respectively, given
by
1+ν ν
= σ − tr(σ )I (6.8)
E E
the resulting cross sectional forces are
N = E Ae, (6.9)
T = G Aγ , (6.10)
M = E I kθ , (6.11)
where N , T , and M are the axial force, shear force, and bending moment, respectively.
The Young’s modulus is indicated by E, the shear modulus by G, the cross-sectional
area by A and the bending moment of inertia by I .
By using the virtual work principle, the equations of motion are found as
where q is the distributed transversal load per unit length. Note that these equations
are valid for any arbitrary cross section rotation θ . This set of equations is highly
nonlinear and therefore of difficult numerical solution. In many applications, the
system operates in the mild nonlinear range—that is to say, that the order of the
nonlinear terms is the same as that of the linear terms. If this holds, the above model
can be expanded and only the significant terms of the expansion need to be retained.
This is discussed in the next section.
This is to say that the axial strain e is not only determined by the derivative of the
axial displacement u but also by the square of the rotation θ of the cross section.
Note also that by assuming ü ≈ 0, (6.12) yields
N = 0, (6.20)
that is to say that the axial force is constant along the length of the beam, and given by
1 2
N = EA u + w . (6.21)
2
N = 0 (6.22)
ρ Aẅ − E I w − N w = q (6.23)
T = −M (6.24)
M = E I w (6.25)
6.1 Geometric Nonlinearities 237
Fig. 6.3 The bending–stretching coupling effect resulting from finite bending displacements
1 2
N = EA u + w , (6.26)
2
equipped with boundary conditions. Note that the presence of N creates a nonlinear
restoring term that depends both on u and w. This effect is known as string effect, and
it is illustrated in Fig. 6.3. Here we can distinguish between two situations, namely:
1. axially unmovable ends: u(0) = u(L) = 0;
2. free ends: N (0) = N (L) prescribed.
Unmovable Ends
In this case, the axial displacement can be integrated along the length of the beam,
to give the axial force N (t) as
L
E A
N (t) = w w2 d x (6.27)
2L 0
when plugged into the governing equation for the transverse displacement w, one
obtains
E A L 2
ρ Aẅ − E I w − w w d x = q. (6.28)
2L 0
Note that the nonlinear term is a cubic function of w , and represents the string effect.
This term is responsible for the hardening of softening behavior of the structure when
subjected to external loads.
Free Ends
In this case, the value of N is prescribed, and therefore the additional stiffness term
in the governing equation is linear:
Consider the 2-node, 2D beam element sketched in Fig. 6.4. The nodal degrees of
freedom can be partitioned into axial and out-of-plane as u = [u 1 , y2 ]T and u =
[w1 , θ1 , w2 , θ1 ]T . The value of u and w within the element are given by the shape
functions as
1
huT = [1 − ξ, 1 + ξ ] , (6.31)
2
1
hwT = 4 − 6ξ + 2ξ 3 , L(ξ 2 − 1)(ξ − 1), . . .
8
and ξ = 2x
L
− 1. The axial force N is then given by
1 T 2
N = E Ae = E A buT u + (bw w) , (6.33)
2
where bu = dhu
dξ
and bw = dhw
dξ
. Likewise, the bending moment M is given by
M = E I k = E I c T w, (6.34)
2
where c = ddξh2w . By using the virtual work principle, the nodal internal forces can be
found. The axial contribution Ui is given by
L 1
L
Ui = N bu d x = N bu dξ, (6.35)
0 −1 2
Note that the axial forces Ui is quadratic and depends both on in-plane and out-
of-plane deformations, while Wi is cubic. The tangent stiffness matrix is readily
obtained by differentiating the internal nodal forces with respect to the nodal dofs,
as ⎡ ⎤
∂Ui ∂Ui
⎢ ∂u ∂w ⎥
⎢ ⎥ Kuu Kuw
K=⎢ ⎥= (6.37)
⎣ ∂Wi ∂Wi ⎦ Kwu Kww ,
∂u ∂w
where
L
Kuu = E Abu buT d x (6.38)
0
L
Kuw = Kwu
T
= E A(bwT w)bu bwT d x (6.39)
0
L
Kww = E I (cc T + E AbwT w)2 bw bwT + N bw bwT d x. (6.40)
0
where r(t) is a residual vector. The ROM is then obtained using Galerkin projection,
which requires the residual to be orthogonal to the subspace span V spanned by V,
i.e., posing VT r = 0. This leads to
T
V CV q̇(t) + VT f(Vq(t)) = VT g(t),
MV q̈(t) + VT
M̃ C̃
where M̃, C̃ ∈ Rm×m are the reduced mass and damping matrices, respectively. Often,
the internal force can be split into its linear and nonlinear contributions as f(u) =
Ku + f nl (u), to obtain a reduced stiffness matrix as well as
M̃q̈(t) + C̃q̇(t) + V T
KV q(t) + VT f nl (Vq(t)) = VT g(t). (6.44)
K̃ f̃(q(t))
As for the case of linear models, the reduced matrices M̃, C̃ and K̃ can be pre-
computed in the offline stage prior to time integration. In other words, during time
integration (online phase), the computational cost associated to the evaluation of the
linear terms in (6.44) scales only with the number of reduced variables m. Unfortu-
nately, this is not the case for the computation of the nonlinear term f̃(q(t)).
For FE applications, this nonlinear forces are usually evaluated as follows:
ne
f̃(q) = VT f nl (Vq) = VeT fe (Ve q), (6.45)
e=1
where fe (ue ) ∈ R Ne is the contribution of the element e toward the vector f nl (u) (Ne
being the number of DOFs for the element e), Ve is the restriction of V to the rows
indexed by the DOFs corresponding to e, and n e is the total number of elements
in the mesh. Since the reduced nonlinear term f̃(q) is evaluated in the space of full
variables, the computational cost associated to its evaluation does not scale with m
6.3 Galerkin Projection 241
alone. Indeed, (6.45) shows that this cost scales linearly with the number of elements
in the structure, and can hence be high for large systems. Thus, despite the reduction in
dimensionality achieved in (6.44), the evaluation of the reduced nonlinear term f̃(q)
hinders any fast prediction of system response using the ROM. Different strategies
are available to overcome this problem, and will be addressed in the following section
in details. Here, it is sufficient to realize the following:
• In the case of polynomial nonlinearities, as the one arising from von Karman
kinematics of linear elastic bodies, it is possible to write the nonlinear reduced
force vector f̃(q) directly in modal coordinates by constructing reduced tensors
that contract on the modal coordinate directly, see for instance,(Touzé et al. 2014).
• When the nonlinearity is not polynomial, one is facing essentially two possibilities:
1. assume a certain form of f̃(q) and identify its parameters or
2. try to compute f̃(q) affordably, but still querying the underlying HFM mesh.
This strategy is addressed by the so-called hyper-reduction methods, which are
discussed, at least to some extent, in Sect. 6.8.
Clearly, any effort in efficiently computing the reduced nonlinear terms is useless
in the reduction basis V does not properly span the sought solution. In summary, a
successful ROM for (6.41) requires a suitable basis V and an efficient strategy to
scale the evaluation cost of f̃ to the size of the ROM. These aspects will be addressed
in the following sections.
When a structure’s nonlinear behavior is not negligible, the traditional linear modal
framework, which is fundamental to our insight into the behavior of linear systems,
is no longer valid. We also lose the convenience of linear superposition and the
associated computational savings. The Nonlinear Normal Mode (NNM) offers a
different theoretical definition of a vibration mode that can be used to understand the
behavior of a system, while maintaining a conceptual similarity to linear vibration
modes. A structure’s nonlinear normal modes can be very useful as a metric to
evaluate a candidate Reduced Order Model (ROM) and to compare various ROMs.
The NNMs decompose the structure’s response into a set of modes, which can each
be described by an amplitude (or energy)-dependent natural frequency and mode
shape. However, they usually do not form a basis that can be used to uncouple the
equations of motion and solve for the response.
The nonlinear normal mode was originally defined by Rosenberg (1960) as a syn-
chronous, periodic motion, or a vibration in unison, of the nonlinear equations of
motion. More recently, Vakakis, Kerschen, and others (Vakakis 1997; Kerschen et al.
2009) have extended this definition to include periodic motions where modal inter-
actions may also occur by relaxing the restriction to synchronous motion. According
to their definition, which is the one used throughout this paper, a nonlinear normal
242 6 Model Reduction and Substructuring for Nonlinear Systems
The external loads are applied through the N × 1 force vector f(t). The N × 1
nonlinear restoring force vector, f N L (x), accounts for the nonlinearity in the physical
system.
The undamped NNM definition used here comes from the works of Kerschen et
al. (2009), who defined an NNM as a not necessarily synchronous periodic motions
of the conservative equations of motion. By this definition, the periodic motions that
occur when two or more modes interact are still considered NNMs since they relaxed
the restriction of synchronous motion defined originally by Rosenberg (1960). Many
new features emerge with this definition of a vibration mode that cannot be described
with linear modal analysis, such as frequency–energy dependence, bifurcations,
localization, and modal interactions. Typically, the NNMs are represented on the
frequency–energy plane, showing how the fundamental frequency of the periodic
motion changes as the energy of the system (which is conserved over the period of
the response) evolves.
Mathematically, the NNMs are solutions to the undamped, unforced equation of
motion given below.
6.4 Nonlinear Normal Modes 243
A variety of algorithms exist for computing NNMs, for example, the shooting algo-
rithm in Kerschen et al. (2009) has proven very effective. Typically, the algorithm is
initialized with the linear natural frequency and mode shape as an initial guess, and
continuation methods are used to trace the solution as amplitude increases.
If we consider again the forced, damped system, we can see that the NNM is a
solution to the forced, damped system only if
This corresponds to the resonant case for the system, because only the damping
forces are able to resist the applied force. Hence, the NNM can be thought of as a
collection of worstcase forced responses for the system.
As an example, consider the curved beam shown in Fig. 6.5. The beam is
3D printed from PLA plastic onto a stiff, C-shaped backing structure that is
constructed from the same material. The structure is excited at its base by a
slowly sweeping sinusoidal input (so that it nearly reaches steady state at any
instant) over a range of amplitudes.
Fig. 6.5 A curved beam is an example of a structure that can exhibit geometric nonlinearity.
As the vibration amplitude increases, stresses that build up in the beam start to cancel the
bending stiffness, so the effective stiffness and hence effective vibration frequency goes down.
At large amplitudes, the structure stiffens since the beam must stretch axially to accommodate
higher amplitude motion
The response of the beam is shown in Fig. 6.6. As the amplitude of the excita-
tion increases, the frequency at which the peak response occurs first decreases
from about 145 Hz to about 138 Hz. The NNM is also shown, as measured
experimentally over a range of amplitude. The NNM correlates with the peak
244 6 Model Reduction and Substructuring for Nonlinear Systems
Fig. 6.6 Response of the curved beam to swept-sine input forces at various amplitudes. The
NNM is overlaid, expresses as the velocity of the center point of the beam versus frequency
Considering again Eq. 6.48, and noting that the damping forces are generally
distributed over the entire structure, it seems that it would be rare that a structure
would be excited into an NNM. However, one can use force balance to show that a
force may be distributed spatially in a variety of ways, including a simple concentrated
force at a point, and still excite an NNM.
Specifically, one can use the energy balance technique in Hill et al. (2014, 2015)
to show that the total energy dissipated by the nonlinear system over one cycle is
T
E diss/cyc = Pdiss dt. (6.49)
0
If the system is excited by an arbitrary forcing function f(t), the energy input into
the system as
T
E =
in/cyc ẋ(t)T f(t) dt. (6.50)
0
At steady state, the energy dissipated by the damping forces must match the total
energy input to the system over the period T . The balance is enforced by setting
E diss/cyc = E in/cyc . Then, if the response, ẋ(t) is assumed to take the shape of
the NNM that is nearest in frequency, and the spatial distribution of the force is
6.4 Nonlinear Normal Modes 245
known, then one can solve for the force amplitude. For example, if a single-point
co-sinusoidal force is applied with complex amplitude A + i0, the scaling on A can
be computed by satisfying
T T
A ẋ(t)T en eiωt dt = ẋ(t)T Cẋ(t) dt (6.51)
0 0
where en is a vector of zeros with a value of one at the location n, which is the point
at which the force is applied. If both the NNMs x(t) and the damping C in the system
are known, Eq. (6.51) can be readily used to estimate the forcing amplitude A that
would excite the system at resonance in the associated NNM x(t). In Kuether et al.
(2015), this approach was demonstrated and found to be very effective for predicting
the forcing required to drive a system near a known NNM. Furthermore, it also
proved useful in providing the initial guesses needed to compute isolated resonance
branches.
By nonintrusive methods, we mean procedures that are able to construct the ROM
(6.44) without requiring access and manipulation of elemental quantities, as for
instance internal forces and tangential stiffness matrices. As such, the ROM is
formed by properly probing the model by use of general purpose FE programs (e.g.,
ABAQUS, NASTRAN) and match an a priori determined form for the ROM to the
obtained data.
The discretized system of equations for an N Degree of Freedom (DOF) linear
elastic, geometrically nonlinear finite element model can be written as
The reduced coordinate space q is significantly smaller than the physical coordinate
space X (i.e., m << N ).
In order to reduce the full order EOMs, Eq. (6.53) is substituted into Eq. (6.52) and
premultiplied by ϕmT to assure that the residual error is orthogonal to the reduction
basis (note that ( )T is the transpose operator). The reduced, nonlinear modal form
of the r th equation becomes
where ωr is the linear natural frequency, and qr is the r th modal displacement. The
nonlinear modal restoring force is given as
The nonlinear restoring forces are a nonlinear function, θr (q), of the modal dis-
placements. If the finite element model is linear elastic with geometric nonlinear-
ities derived using quadratic strain–displacement relationships, the function θr (q)
becomes a quadratic and cubic polynomial, and is given as
m
m
m
m
m
θr (q1 , q2 , . . . qm ) = Br (i, j)qi , q j + Ar (i, j, k)qi , q j , qk .
i=1 j=1 i=1 j=1 k=1
(6.56)
The coefficients and are the quadratic and cubic nonlinear stiffness terms, respec-
tively, for the r th nonlinear modal equation. If the full order EOM in Eq. (6.52) are
known in closed form, then the nonlinear coefficients in Eq. (6.56) can be directly
computed (referred as direct evaluation in Hollkamp and Spottswood 2005). How-
ever, in cases where the closed form equations are not explicitly available (e.g.,
within commercial finite element packages), an indirect approach must be used. The
coefficients are approximated using either the ED or ICE methods using a series
of nonlinear static solutions. Once these nonlinear stiffness terms are identified, the
NLROM equations are given by Eqs. (6.54) and (6.56), providing a significant com-
putational savings for response prediction compared to direct time integration.
Typically, the linear term is set to ωr2 and treated as a known value during the
identification of Br and Ar .
Xc = ϕ1
q1 + ϕ2
q2 + · · · + ϕm
qm , (6.57)
6.5 Nonintrusive Reduced Order Modeling (ROM) … 247
where Xc is an N × 1 displacement vector and qr is the scaling factor for the r th mode
in the basis. In order to accurately estimate the ROM, the scaling factors should scale
the deformation shape of each mode to a level that exercises the geometric nonlinear-
ity in the structure. Using a commercial finite element software package, the resulting
reaction forces, Fc , are computed in response to the given displacement field. Then,
the quasi-static representation of the NLROM for the r th modal equation becomes
m
m
m
m
m
ωr2
qr + Br (i, j)qi , q j + Ar (i, j, k)qi , q j , qk = ϕrT Fc . (6.58)
i=1 j=1 i=1 j=1 k=1
The nonlinear stiffness coefficients are then solved for using the approach pre-
sented in Muravyov and Rizzi (2003) with a set of displacement load cases that are
in a combination of one, two or three modes in the basis set. The number of nonlinear
static solutions, for m > 3, required by the ED method (Mignolet and Spottswood
2013) is
3m! m!
2m + + . (6.59)
2(m − 2)! 6(m − 3)!
The applied loads procedure uses a set of static forces in the shape of the linear modes
in order to find the nonlinear stiffness coefficients of the NLROM. For example, a
multimode static force can be a combination of those forces that would excite any
of the modes in the reduced basis, given as
Fc = M(ϕ1
f 1 + ϕ2
f 2 + · · · + ϕm
f m ), (6.60)
where Fc is the vector of applied forces and fr is the force scaling factor for the r th
mode. Note that the previous works on the applied loads method did not mention the
use of the mass matrix when computing the force vector, but it was used here and
is needed to obtain a force that exactly isolates a single mode for a linear system.
The force scaling factors can be varied to exercise the desired amount of geometric
nonlinearity in the structure, as will be discussed further in Section III. The resulting
static deformation, Xc , can then be projected onto the linear modal coordinates as
follows.
qr = ϕr Mxc . (6.61)
Using the quasi-static force–displacement relationship in Eq. (6.58), one can then
solve for the nonlinear stiffness coefficients with the constrained approach (Gordon
and Hollkamp 2011) using a set of these applied loads exercising the different modes
in the basis. With the applied loads approach, the number of load permutations (for
248 6 Model Reduction and Substructuring for Nonlinear Systems
m! 4m!
2m + 2 + . (6.62)
(m − 2)! 3(m − 3)!
The number of loads differs from that of the Enforced Displacement in Eq. (6.59)
due to differences in how the fits are performed; with the ED method sets of two
and three load cases can be used to solve algebraic equations for the coefficients,
whereas with ICE a larger least squares problem is solved.
Vibration Modes
The system in (6.41) can be linearized around its static equilibrium position. Without
loss of generality, we assume the equilibrium configuration is ueq = 0 to obtain
where
∂f(u)
K|eq = . (6.64)
∂u u=0
This linearized system would be a good approximation to the original system in (6.41)
for small enough displacements from the linearization point. It is well established
that for such system, the system response can be written as a linear combination of
constant eigenmodes (or Vibration Modes (VMs) in the structural dynamics context)
which form a basis of Rn as follows:
n
u(t) = φ i qi (t), (6.65)
i=1
6.6 Intrusive Methods 249
where the eigenmodes φ i ∈ Rn are the solution of the generalized eigenvalue prob-
lem1
(K|eq − ωi2 M)φ i = 0. (6.66)
m
u(t) ≈ φ i ηi (t) = Φq(t), (6.67)
i=1
where Φ ∈ Rn×m , q(t) = [q1 (t) q2 (t) · · · qm (t)]T ∈ Rm , m n. Note that this is
equivalent to performing a Galerkin projection as shown in Sect. 3.1 for a lin-
earized system, where Φ is the reduction basis and q are the corresponding reduced
unknowns.
Modal Derivatives
When the deviation from the linearization point increases, the response of (6.63)
can no longer be considered as a good approximation for the original nonlinear
counterpart. One might still think of using the VMs obtained from the linearized
model to form a reduction basis for the reduction of the nonlinear set of equations.
A basis composed of a few dominant VMs, however, is typically not sufficient for
reduction since it does not feature the dominant coupling effects (e.g., membrane-
bending), typical of geometrically nonlinear structures.
Earlier work in Idelsohn and Cardona (1985a, b) and more recent work in Weeger
et al. (2016); Witteveen and Pichler (2014) discuss the use of the so-called Modal
Derivatives (MDs) to capture the response of the nonlinear system upon departure
from the linear behavior. If (6.66) is differentiated with respect to the modal amplitude
q j (assuming M to be a constant mass matrix) and then evaluated at equilibrium, one
obtains
∂φ i ∂K ∂ωi2
(K|eq − ωi |eq M)
2
+ − M φ i |eq = 0, (6.68)
∂q j eq ∂q j eq ∂q j eq
where the MD ∂φ i
∂q j
denotes the derivative of the ith mode in the jth modal direction.
The tangent stiffness matrix derivative with respect to η j is obtained by imposing t
a displacement in the direction of φ j |eq to the system, i.e.,
1 Here, we neglect the damping contribution in eigenvalue problem to avoid complex eigenvalues
and vectors. Note that for damped linear systems with low damping or modal/Rayleigh damping,
the eigenvectors for an undamped system are a good approximation for the damped counterpart and
still form a good basis for linear modal superposition.
250 6 Model Reduction and Substructuring for Nonlinear Systems
∂K ∂K(u = η j φ j |eq ))
= . (6.69)
∂q j eq ∂q j
q j =0
By differentiating the equation above with respect to the modal amplitude one obtains
∂φ i ∂φ
φ iT M + φ iT MT i = 0 ∀i, j ∈ {1, 2, . . . , m}. (6.71)
∂q j ∂q j
The following direct approach to calculate the MDs can then be formulated using
(6.68), (6.72):
⎤ ⎡
⎡ i ∂K
⎤
[K|eq − ωi2 |eq M]n×n − Mφ i |eq n×1 [1] ∂φ
T ⎣ ⎦ = ⎣− ∂q j φ i |eq ⎦ .
∂q j eq
The above non-singular system can be used to solve for MDs. This method, however,
destroys the band structure of the original system. Nonetheless, it is rigorous and
accurate. Apart from this direct approach, the pseudoinverse technique (cf. G’eradin
and Rixen 1997) and the Nelson’s method (Nelson 1976) are some techniques that
preserve the band structure of the matrices.
Regardless of the method adopted to solve (6.68), a high-dimensional matrix needs to
be factorized for each ωi . In Idelsohn and Cardona (1985b), a method to approximate
the problem (6.68) is discussed. They propose neglecting the inertial contribution as
∂φ i s ∂K
K|eq =− φ |eq . (6.73)
∂q j eq ∂q j eq i
6.6 Intrusive Methods 251
s
∂φ i
∂q j eq
We call the obtained MDs as Static MDs (SMDs). The superscript s in stands
for static. s
i
Note that the Static Modal Derivatives given in (6.73) are symmetric, i.e., ∂φ∂q j eq
=
∂φ j s
∂qi
. In fact, it is easy to see that the stiffness matrix derivative given by (6.69)
eq
can be written as
∂K ∂K(u) ∂ 2 f(u)
= · φ j |eq = · φ j |eq . (6.74)
∂q j eq ∂u u=0 ∂u∂u u=0
While the SMDs have been shown to be symmetric, such a claim cannot be made
for the modal derivatives as given by (6.68).
(S)MDs in a Reduction Basis
A linear basis (Ψ ) consisting of VMs augmented with the corresponding MDs could
be used to reduce the nonlinear system (Idelsohn and Cardona 1985a, b).
Another popular way of constructing the reduction basis V is given by the Krylov
vectors. This technique is also known under different denominations as Wilson vec-
tors (Ming-Wu 1982), Arnoldi sequence (Arnoldi 1951) and load dependent vectors
(Idelsohn and Cardona 1985). Unlike vibration modes, it consists of finding a proper
reduction basis by using the information of the applied load distribution. Consider
the linear problem
Mü + Ku = g, (6.79)
where damping has been neglected. We first start by computing the linear solution
1
u0 = ũ0 (6.81)
ũ0T Mũ0
The procedure can be made recursive, making sure that the newly computed vec-
tor is orthogonalized with respect to the previously calculated ones, and properly
orthogonalized. A generic vector ui of the sequence is first formed by computing
ûi as
Kûi = Mui−1 , (6.83)
where
c j = uTj Mûi for j = 1, . . . , i − 1. (6.85)
6.6 Intrusive Methods 253
1
ui = ûi . (6.86)
ũiT Mũi
The so obtained vectors represent a good basis for the linearized problem (6.79).
With respect to the vibration modes, they do not require the solution of an eigenvalue
problem, but only the solution of a sequence of linear problems sharing the same
matrix of coefficients, which can be then factorized once for all. Note also that the
sequence is initiated by the static solution u0 , which is directly related to the shape
of the applied load. In contrast to vibration modes, Wilson vectors do account for the
particular load shape applied. As such, they tend to show faster convergence with
respect to vibration modes, but would need to be recomputed in case another load
shape is applied to the system.
Since the so obtained basis relates to the linearized problem, it is likely to per-
form poorly in the case of geometric nonlinearities. To overcome this limitation,
modal derivatives of Wilson vectors can be computed in a way analogous to the one
described in the previous section. In particular, say VW = [u1 , . . . , um ] the basis of
linear Wilson vectors, then one can compute the modal derivatives
∂K
θ Wi j = K−1 uj (6.87)
∂qi
This is a least squares problem and the vectors in V are the left singular vectors of
U, obtained by the Singular Value Decomposition (SVD) of U of zero mean as
U = LSR T , (6.89)
Thus, the left singular vectors li corresponding to the highest singular values are
the most relevant for constructing a reduction basis using POD. for our case, each
(normalized) left vector li is scaled by the corresponding singular value σi , and
associated to the time history contained in riT . Since an SVD can be performed over
any solution snapshots of any general nonlinear problem, POD is seen as a versatile
method. It should be noted, however, that such a reduction basis is optimal only
for capturing the solution used to collect the snapshots. The procedure needs to be
repeated (or a database of full solutions need to be created in the first place) to take
other types of loading or boundary conditions into account.
Once a reduction basis is obtained, the same ensemble of full solution vectors
can be used as training vectors for hyper-reduction to accelerate the computation of
nonlinearity during the online stage. This is discussed in the following sections.
6.8 Hyper-reduction 255
6.8 Hyper-reduction
With the term hyper-reduction we indicate techniques that aim at providing a com-
putationally affordable computation of the reduced nonlinear terms f̃(q). In a FE
context, f̃(q) requires assembly over the whole FE mesh, as
ne
f̃(q) = VeT fe (Ve q), (6.90)
e=1
and, as such, is computationally intensive, as it scales with the mesh size and not
with the size of the reduction basis. As already seen, one possibility is to write f̃(q)
directly in terms of q as
m
m
m
m
m
f̃(q) = Br (i, j)qi , q j + Ar (i, j, k)qi , q j , qk , (6.91)
i=1 j=1 i=1 j=1 k=1
where the coefficients Br (i, j) and Ar (i, j, k) are identified by means of nonintrusive
methods, or computed directly in case of polynomial nonlinearities. In more general
cases, alternatives to this approach are available. In the following sections, we briefly
present the Discrete Empirical Interpolation (DEIM) and some useful variants for the
FE context, and the Energy Conserving Sampling and Weighting (ECSW) method.
They differ in their philosophy, but they both aim at constructing f̃(q) cheaply and
independently of the actual form of f.
The main idea behind DEIM is to provide and approximation for the nonlinear
forces f(u) as a combination of few vectors forming a base F = [f1 , . . . , fm ] ∈ Rn×m
spanning a subspace of dimension m < n, as
f ≈ Fc, (6.92)
where c ∈ Rm×1 the vector of unknown amplitudes of such force vectors. The system
of Eqs. (6.92) cannot be solved, as it is overdetermined. In order to solve it, one option
is to choose m distinct rows of f. This can be done by premultiplying by the transpose
of a boolean matrix P ∈ Rn×m as
P T f = P T Fc. (6.93)
c = (P T F)−1 P T f, (6.94)
This is to say, the term P T f specifies that the nonlinear vector f needs to be calcu-
lated only at the entries specified by P T . The precomputed term F(P T F)−1 provides
approximated values at the other entries not considered by P. The projection basis F
is computed from a POD of an appropriate set of snapshots of the nonlinear forces
F = [f(u1 ), . . . , f(un s )]. The selection points are determined from the greedy Algo-
rithm 1. Details are given in Chaturantabut and Sorensen (2010).
Unfortunately, this does not hold when f comes from a FEM discretization. In a FE
framework, each component f i represents the generalized force component at one
specific node. This entry depends on the generalized displacements of all the nodes
belonging to the neighboring elements, i.e., a given degree of freedom is supported
by a set of elements. As a consequence, P cannot be brought inside the function
Fig. 6.7 Scalar-valued function (left) and vector valued function (right). In the first case, each
component Fi of F depends on only one component u j of u. In the second case, Fi depends on
several components of u. This latter case depicts the FE discretization
Fig. 6.8 Sparsity comparison for the evaluation of f at the DEIM points. In the case of assembled
meshes (left), one evaluation implies an element function call for all the neighboring elements. If
the mesh is treated as unassembled (left), only one element per DEIM point needs to be queried
This issue is illustrated in Fig. 6.7b and is responsible for the diminished inefficiency
of DEIM when applied on FE schemes. In general, a generic component f ℘i depends
on a subset of the displacements ū℘i ∈ R p , where p depends on the type of element
and the mesh connectivity. This implies that the evaluation of a single component f i
of the nonlinear force vector f requires a call to the element functions relative to all
the neighboring elements.
Alternative DEIM Formulation
A possible workaround of the inefficiency discussed in the previous section is to apply
DEIM on an unassembled force vector, i.e., on the stacking of all the elemental forces
without performing element connection. In this way, the evaluation of a single entry
of such vector will require just one element call. The concept is illustrated in Fig. 6.8.
This modification is called Unassembled DEIM (UDEIM).
Unassembled DEIM
where feTi (yei ) ∈ R p×1 is the element force vector relative to the element i, i =
1, . . . , n e and p is the number of elemental degrees of freedom. It follows from (6.98)
that fuT ∈ Rn u ×1 , where n u = pn e is the size of the unassembled force snapshots. The
vector fu is projected onto a basis Uu obtained by a SVD of the unassembled force
258 6 Model Reduction and Substructuring for Nonlinear Systems
snapshots. The DEIM algorithm is then applied to Uu and the obtained indices will
point to elemental degrees of freedom, instead of mesh ones. Once the amplitudes
cu have been obtained, an assembled approximation of f can be formed by
Algorithm 2 UDEIM
Input: unassembled snapshots Fu = [fu1 , . . . , fun s ] ∈ Rn u ×n s
1: SVD(Fu ) : Fu = Ũu Σ u VuT
2: select first m columns of Ũu : Uu = Ũu (:, 1 : m), with m < n s
3: DEIM(Uu ) → Pu (Algorithm 1)
4: ecompute (PuT Uu )−1
5: assemble Ua ∈ Rn×m : Ua := assemble(Uu )
6: f ≈ Ua (PuT Uu )−1 (PuT fu )
The unassembled basis Uu contains the most relevant force modes where each ele-
ment contribution is left unassembled. A DEIM selection on Uu (UDEIM) as outlined
in Algorithm 1 then picks the most relevant elemental forces to form the interpolation.
This way to proceed improves the sparsity of the force–displacement dependency
but increases the effort of performing the SVD of the snapshots.
The main idea is that each elemental contribution can be represented by a surrogate
quantity rather than the complete elemental force vector. If this holds, the size of the
unassembled snapshot vectors can be reduced while preserving the advantage of
operating on elements directly.
Let us assume that the contribution of element i to the nonlinear behavior can be
represented by the sum sie of all its generalized force components f ie , as
p
sie = f ej . (6.100)
j=1
In case the generalized element forces do not share the same units (as in the shell
element case here considered), the surrogate can be extended as
6.8 Hyper-reduction 259
⎡ "p f e
⎤
j=1 f j
sie = ⎣ " ⎦, (6.101)
pm
j=1 mej
where m ie are the nodal elemental moments, and p f and pm are the number of
element
forces and
moments, respectively. Then, a collection of surrogates snapshots
S = S1 , . . . , Sn s can be built, where each vector Si is given by
and S ∈ Rs×n s , s being the size of the surrogate snapshots, and s n u . If the sur-
rogate snapshots Si are supposed to feature the same overall contribution as the
corresponding unassembled snapshots Su , then we can assert that a SVD decompo-
sition of S and Fu should yield very similar right singular vectors V, at least for the
dominant left singular vectors Uu . This request can be written as
SVD(Fu ) : Fu = Uu Σ u VT
(6.103)
SVD(S) : S = US Σ S VT .
It is useful to remind here that the right vectors V represent the time history of the
corresponding left singular vectors Uu . By claiming the decomposition (6.103) holds,
we imply that the unassembled singular vectors Uu and the surrogates vectors US
share the same time behavior. Our numerical results indeed show that this indeed
holds for the most significant modes. Since we are interested in only the spatial basis
Uu , the expensive decomposition SVD(Fu ) can be avoided and the basis Uu can
be formed by using the vectors V obtained by the less computationally expensive
SVD(S), as
Fu V = Uu Σ u VT V = Uu Σ u = Ūu ,
(6.104)
Uu = orth(Ūu ),
Algorithm 3 SUDEIM
Input: unassembled snapshots Fu = [Fu1 , . . . , Fun s ] ∈ Rn u ×n s ,
1: surrogate snapshots S = [S1 , . . . , Sn s ] ∈ Rs×n s , s < n
2: SVD(S) : S = ŨS Σ S VT
3: Ūu ≈ Fu V and Ũu = orth(Ūu )
4: select first m columns of Ũu : Uu = Ũu (:, 1 : m), with m < n s
5: DEIM(Uu ) → Pu (Algorithm 1)
6: compute (PuT Uu )−1
7: assemble U ∈ Rn×m : U := assemble(Uu )
8: F ≈ U(PuT Uu )−1 (PuT Fu )
Surrogate DEIM
It follows from the discussion above that the concept of the surrogate can be applied
also to the assembled basis for the nonlinear forces U, as outlined in Algorithm 4.
The computational saving lies in this case only in the cheaper SVD decomposition of
the surrogate, while the sparsity of the force–displacement mapping is not improved.
This variant is called Surrogate DEIM (SDEIM) (see Algorithm 4). Numerical results
comparing these different variants of DEIM are presented and discussed in Tiso and
Rixen (2013).
Algorithm 4 SDEIM
Input: snapshots F = [F1 , . . . , Fn s ] ∈ Rn×n s ,
1: surrogate snapshots S = [S1 , . . . , Sn s ] ∈ Rs×n s , s < n
2: SVD(S) : S = ŨS Σ S VT
3: Ū ≈ FV and Ũ = orth(Ū)
4: select first m columns of Ũ: U = Ũ(:, 1 : m), with m < n s
5: DEIM(U) → P (Algorithm 1)
6: compute (P T U)−1
7: F ≈ U(P T U)−1 (P T F)
The ECSW method (Farhat et al. 2014, 2015) is based on a different principle as
compared to DEIM. Essentially, ECSW aims to identify a small set of elements E
of the structure (|E| n e ) to cheaply approximate f̃(q) as (cf. (6.45))
ne
f̃(q) = VeT fe (Ve q) ≈ ξe VeT fe (Ve q), (6.105)
e=1 e∈E
pared to the cost associated to to (6.45). In doing so, ECSW approximates the virtual
work done by the internal forces on the set of vectors in the basis V. As a consequence,
the ECSW preserves the structure and the stability properties of the underlying full
model (cf. Farhat et al. 2015).
The elements and weights are determined to approximate virtual work over the
chosen training sets, which generally come from full solution run(s). If there are n t
training displacement vectors in the set with u(i) representing the ith vector, then
corresponding reduced unknowns q(i) can be calculated using least squares as
The element level contribution of projected internal force for each of the training
vectors can be assembled in a matrix G as
⎡ ⎤ ⎡ ⎤
g11 . . . g1n e b1
⎢ .. . . . ⎥ ⎢ ⎥
G=⎣ . . .. ⎦ ∈ R t e ,
mn ×n
b = ⎣ ... ⎦ ∈ Rmn t ,
gn t 1 . . . gn t n e (6.106)
(i)
(i) "bnnet
gie = Ve fe (Ve q ),
T
bi = f̃ q = e=1 gie ,
∀i ∈ {1, . . . , n t }, e ∈ {1, . . . , n e } .
The set of elements and weights is then obtained by a sparse solution to the following
Non-negative Least Squares (NNLS) problem
A sparse solution of (6.107) (i.e., a solution vector with lots of zeros) returns a sparse
vector ξ , the nonzero entries of which form the reduced mesh E used in (6.105) as
E = {e : ξe > 0}.
1: E ← ∅, Z ← {1, . . . , n e }, ξ ← 0 ∈ Rn e
2: while Gξ − b2 > τ b2 do
3: μ ← GT (b − Gξ )
4: [|ν|, e] = max (μ) max: returns maximum value in a vector followed by its location
(index)
5: E ← E ∪ {e}, Z ← Z \{e}
6: while true do
7: ζ E ← G†E b † represents pseudoinverse
8: ζZ ← 0
9: if ζ E > 0 then
10: ξ ←ζ
11: break
12: end if
13: η = min ξk /(ξk − ζk )
k∈E
14: ξ ← ξ + η(ζ − ξ )
15: Z ← {i|ξi = 0}
16: E ← {1, . . . , n e }\Z
17: end while
18: end while
The material reviewed here comes primarily from Kuether et al. (2016) and Kuether
et al. (2017). Several nonlinear substructuring methods have been developed to pre-
dict the dynamics of an assembly based on the dynamics of its subcomponents. For
a review of the relevant literature, see Kuether et al. (2017).
This section focuses on geometric nonlinearities that are distributed throughout
all the elements in the subcomponent FEA model. Recently, Wenneker (2013) used a
basis defined by the Craig–Bampton and Rubin approach and augmented these with
modal derivative vectors in order to account for the effects of geometric nonlinearity.
The number of modal derivatives required scales quadratically with the number of
component modes used, resulting in a rather large order reduced system. Perez (2012)
was the first to explore the use of CB modes and a reduced set of constraint modes
in conjunction with an indirect reduced-order modeling approach. He presented a
thorough analysis of a complicated multi-bay frame reducing the linear model from
96,000 to 232 DOF that were a combination of fixed-interface modes and constraint
modes reduced using proper orthogonal decomposition. Unfortunately, this model
was still 2.6 times larger than an 89 mode ROM that he had created for the assembled
system, so nonlinear modal substructuring was not pursued further.
6.9 Hurty/Craig–Bampton Substructuring with ROMs … 263
the N × 1 nonlinear restoring force vector, f N L (x), captures the interior forces intro-
duced by the geometric nonlinearity and depends on the strain model used to model
the large deformations. For the reduction procedure that follows, it is assumed that
the FEA model is linear elastic with a quadratic strain–displacement relationship.
The N × N linear mass and stiffness matrices, M and K, respectively, are generated
using linear finite element analysis, and the external loads are applied through the
N × 1 vector f(t).
As elaborated previously, a Hurty/Craig–Bampton model can be written as fol-
lows. Below, the nonlinear forces due to geometric nonlinearity are also included.
264 6 Model Reduction and Substructuring for Nonlinear Systems
# $ # $ # $ # $
Mii Mib ẍi Kii Kib xi f N L ,i (x) 0
+ + = . (6.109)
Mbi Mbb ẍb Kbi Kbb xb f N L ,b (x) f(t)
where
Mii Mib Ikk M̂kb
M̂C B = TCT B TC B = , (6.112)
Mbi Mbb M̂bk M̂bb
Kii Kib Λkk 0kb
K̂C B = TCT B TC B = . (6.113)
Kbi Kbb 0bk K̂bb
The linear portion is exactly that of the CB reduced order model for any linear
subsystem. Most commercial finite element packages are not able to directly extract
the functional form of the nonlinear restoring force, f N L (TCT B q), in terms of the modal
coordinates. However, because the geometric nonlinearity is known to be expressible
in terms of quadratic and cubic polynomials for the common large displacement
strain models, the same functional form will hold for the fixed-interface modes of
the structure. As a result, the nonlinear term in Eq. (6.111) becomes
# T $
f T q
TCT B N L ,i CT B = θ (q) (6.114)
f N L ,b TC B q
and each row of the nonlinear function θ (q) defines the nonlinear portion of the r th
modal equation as
m
m
m
m
m
θr (q) = Br (i, j)qi , q j + Ar (i, j, k)qi , q j , qk (6.115)
i=1 j=1 i=1 j=1 k=1
The nonlinear stiffness coefficients can then be determined using the Implicit
Condensation method as explained in Sect. 6.5. The resulting nonlinear model can
also be expressed in the Nash form, which is convenient for what follows.
∂β !
∂β ∂β ∂β
N1 (q) = = ∂q1 ∂q2
... ∂qm , (6.116)
∂q
6.9 Hurty/Craig–Bampton Substructuring with ROMs … 265
∂α !
∂α ∂α ∂α
N2 (q) = = ∂q1 ∂q2
... ∂qm . (6.117)
∂q
The CB-NLROM in Eq. (6.111) is then rewritten in matrix form with the nonlinear
stiffness matrices in the equations above as
# $
1 1 0
M̂C B q̈ + K̂C B q + N1 (q) q + N2 (q) q = TC B
T
(6.118)
2 3 f (t)
These ROMs can then be coupled using the methods outlined earlier, and their
behavior depends on the adequacy of the HCB basis for the modes of interest as
discussed earlier.
As an alternative to Implicit Condensation, one can also use modal derivatives
with an enforced displacements procedure to create a ROM for a substructure. This
is detailed in Chap. 3 of Wu (2018). Substructuring can also be extended to flexible
multibody systems, as detailed in Chaps. 4 and 5 of Wu (2018). The content of Chap. 4
in Wu (2018) was published in Wu and Tiso (2016).
References
Arnoldi, W. E. (1951). The principle of minimized iterations in the solution of the matrix eigenvalue
problem. Quarterly of Applied Mathematics, 9, 17–29.
Chaturantabut, S., & Sorensen, D. (2010). Nonlinear model reduction via discrete empirical inter-
polation. SIAM Journal on Scientific Computing, 32(5), 2737–2764. https://doi.org/10.1137/
090766498.
Farhat, C., Avery, P., Chapman, T., & Cortial, J. (2014). Dimensional reduction of nonlinear finite
element dynamic models with finite rotations and energy-based mesh sampling and weighting for
computational efficiency. International Journal for Numerical Methods in Engineering, 98(9),
625–662. https://doi.org/10.1002/nme.4668.
Farhat, C., Chapman, T., & Avery, P. (2015). Structure-preserving, stability, and accuracy properties
of the energy-conserving sampling and weighting method for the hyper reduction of nonlinear
finite element dynamic models. International Journal for Numerical Methods in Engineering,
102, 1077–1110. https://doi.org/10.1002/nme.4820.
G’eradin, M., & Rixen, D. (1997). Mechanical Vibrations: Theory and Application to Structural
Dynamics (2nd Ed.). Hoboken: Wiley. ISBN: 0-471-97524-9
Gordon, R.W., & Hollkamp, J. J. (2011). Reduced-order models for acoustic response prediction.
AFRL-RB-WP-TR-2011-3040
Hill, T. L., Cammarano, A., Neild, S. N., & Wagg, D. (2014). An analytical method for the opti-
misation of weakly nonlinear systems. In Proceedings of the 9th International Conference on
Structural Dynamics-EURODYN. Porto, Portugal.
Hill, T. L., Cammarano, A., Neild, S. N., & Wagg, D. (2015). Interpreting the forced response
of a two-degree-of-freedom nonlinear oscillator using backbone curves. Journal of Sound and
Vibration, 349, 276–288.
Hollkamp J. J., G.R.W., Spottswood, S.M. (2005). Nonlinear modal models for sonic fatigue
response prediction: a comparison of methods. Journal of Sound and Vibration, 284, 1145–1163.
Idelsohn, S. R., & Cardona, A. (1985) A load-dependent basis for reduced nonlinear struc-
tural dynamics. Computers & Structures, 20(1), 203 – 210. https://doi.org/10.1016/0045-
7949(85)90069-0. Special Issue: Advances and Trends in Structures and Dynamics.
266 6 Model Reduction and Substructuring for Nonlinear Systems
Idelsohn, S. R., & Cardona, A. (1985a). A load-dependent basis for reduced nonlinear structural
dynamics. Computers & Structures, 20, 203–210. https://doi.org/10.1016/0045-7949(85)90069-
0.
Idelsohn, S. R., & Cardona, A. (1985b). A reduction method for nonlinear structural dynamic
analysis. Computer Methods in Applied Mechanics and Engineering, 49(3), 253–279. https://
doi.org/10.1016/0045-7825(85)90125-2.
Kerschen, G., Peeters, M., Golinval, J. C., & Vakakis, A. F. (2009) Nonlinear normal modes. part
i. a useful framework for the structural dynamicist. Mechanical Systems and Signal Process-
ing23(1), 170–194. Copyright 2008, The Institution of Engineering and Technology 10330862
0888-3270 nonlinear normal modes NNMs large-scale engineering structures time-frequency
analysis structural dynamics nonlinear vibrating structures wavelet transform modal analysis.
Kuether, R. J., Allen, M. S., & Hollkamp, J. J. (2016). Modal substructuring for geometrically non-
linear finite element models. AIAA Journal, 54(2), 691–702. https://doi.org/10.2514/1.J054036.
Kuether, R. J., Allen, M. S., & Hollkamp, J. J. (2017). Modal substructuring of geometrically
nonlinear finite element models with interface reduction. AIAA Journal (accepted).
Kuether, R. J., Renson, L., Detroux, T., Grappasonni, C., Kerschen, G., & Allen, M. S. (2015)
Nonlinear normal modes, modal interactions and isolated resonance curves. Journal of Sound
and Vibration, 351(299-310). https://doi.org/10.1016/j.jsv.2015.04.035.
Mignolet M. P., Przekop, A., Rizzi, S. A., & Spottswood, S. M. (2013). A review of indirect/non-
intrusive reduced order modeling of nonlinear geometric structures. Journal of Sound and Vibra-
tion, 332(10), 2437–2460.
Muravyov, A. A., & Rizzi, S. A. (2003). Determination of nonlinear stiffness with application to
random vibration of geometrically nonlinear structures. Computers &Structures, 81(15), 1513–
1523.
Nelson, R. B. (1976). Simplified calculation of eigenvector derivatives. AIAA Journal, 14(9), 1201–
1205. http://arc.aiaa.org/doi/abs/10.2514/3.7211?journalCode=aiaaj
Peharz, R., & Pernkopf, F. (2012). Sparse nonnegative matrix factorization with 0 -constraints.
Neurocomputing, 80, 38–46.
Perez, R. A. (2012). Multiscale reduced order models for the geometrically nonlinear response of
complex structures.
Rosenberg, R. M. (1960). Normal modes of nonlinear dual-mode systems. Journal of Applied
Mechanics, 27, 263–268.
Siddhi, V. (2005). A generalized approach for calculation of the eigenvector ensitivity for various
eigenvector normalizations. Master’s thesis, University of Missouri - Columbia. https://mospace.
umsystem.edu/xmlui/bitstream/handle/10355/4244/research.pdf?sequence=3
Tiso, P., & Rixen, D. J. (2013). Discrete empirical interpolation method for finite element structural
dynamics. In G. Kerschen, D. Adams, & A. Carrella (Eds.), Topics in nonlinear dynamics (Vol.
1, pp. 203–212). New York: Springer.
Touzé, C., Vidrascu, M., & Chapelle, D. (2014). Direct finite element computation of non-linear
modal coupling coefficients for reduced-order shell models. Computational Mechanics, 54(2),
567–580. https://doi.org/10.1007/s00466-014-1006-4.
Vakakis, A. F. (1997). Non-linear normal modes (nnms) and their applications in vibration theory:
an overview. Mechanical Systems and Signal Processing11(1), 3–22. Compilation and indexing
terms, Copyright 2009 Elsevier Inc. Compendex 1997163536841 08883270 Nonlinear normal
mode Shock isolation designs Vibrational energy.
Weeger, O., Wever, U., & Simeon, B. (2016). On the use of modal derivatives for nonlinear model
order reduction. International Journal for Numerical Methods in Engineering, pp. n/a–n/a. https://
doi.org/10.1002/nme.5267.
Wenneker, F. (2013). Component mode synthesis for geometrically nonlinear structures.
Wilson, E. L., Yuan, M. W. & Dickens, J. M. (1982). Dynamic analysis by direct superposition of
Ritz vectors. Earthquake Engineering & Structural Dynamics10(6), 813–821. https://doi.org/10.
1002/eqe.4290100606.
6.9 Hurty/Craig–Bampton Substructuring with ROMs … 267
Witteveen, W., & Pichler, F. (2014). Efficient model order reduction for the nonlinear dynamics
of jointed structures by the use of trial vector derivatives. In IMAC-XXXII: International Modal
Analysis Conference, Orlando, FL. Bethel, CT: Society for Experimental Mechanics.
Woinowski-Krieger, S. (1950). The effect of axial force on the vibration of hinged bars. Journal of
Applied Mechanics, 17.
Wu, L. (2018). Model order reduction and substructuring methods for nonlinear structural dynam-
ics. Dissertation, Delft University of Technology.
Wu, L., & Tiso, P. (2016). Nonlinear model order reduction for flexible multibody dynamics: a
modal derivatives approach. Multibody System Dynamics, 36(4), 405–425. https://doi.org/10.
1007/s11044-015-9476-5.
Chapter 7
Weakly Nonlinear Systems: Modeling
and Experimental Methods
Abstract The prior chapter presented rigorous theory and methods for nonlinear
systems, which is necessary in general because many nonlinear systems exhibit
strong modal coupling due to the nonlinearity; this is commonly the case for the
geometrically nonlinear structures that were the focus of that chapter. However, one
of the most common sources of nonlinearity in built-up structures is the joints, and
in many cases, these introduce only a weak stiffness nonlinearity together with a
significant damping nonlinearity. In this case, and in many others that are relevant
to industry, one can obtain good estimates of the response of the structure using a
weakly nonlinear model in which the linear modes of the structure are presumed to
be preserved and coupling between modes is neglected. This chapter provides a brief
introduction to these concepts.—Chapter Authors: Randall Mayes and Matt Allen
Much of the uncertainty in finite element models for built-up structures comes from
the interfaces (e.g., bolted joints, riveted connections, press-fits). They often cause
the structure to behave nonlinearly, making experiments more challenging, and to
model them accurately requires a nonlinear model that can be dramatically more
expensive. See Brake (2018), Bhushan (2013), Ferri (1995) for a review of these
issues. This work focuses on methods by which joints are represented in a simplified
way, so that the structure’s response can be simulated quickly and the model updated
to correlate with experimental measurements.
In most industrial applications, the bolts are designed to retain integrity, so most of
the joints nonlinearity is attributed to micro-slip friction in which the joint remains
intact but there is partial slipping for some material near the edges of the contact
(Groper 1985; Gaul and Lenz 1997). Commercial finite element codes typically
include the ability to simulate the static and dynamic response of flexible structures
with Coulomb friction between contacts. For example, Abaqus (2014) includes the
capability to model friction using Penalty and Lagrange methods. However, once the
contact interfaces are meshed with sufficient resolution to capture micro-slip and the
© CISM International Centre for Mechanical Sciences 2020 269
M. S. Allen et al., Substructuring in Engineering Dynamics,
CISM International Centre for Mechanical Sciences 594,
https://doi.org/10.1007/978-3-030-25532-9_7
270 7 Weakly Nonlinear Systems: Modeling and Experimental Methods
penalty stiffness and stabilization are set to acceptable levels, the simulations become
very expensive for static response (Jewell et al. 2017), and completely prohibitive
when a dynamic response is desired that spans many cycles of oscillation. To seek
to address this, harmonic balance methods have been developed that computes the
frequency response directly (Petrov and Ewins 2003; Bograd et al. 2011). In order
to model micro-slip behavior accurately, such models typically require hundreds
of friction elements per interface (Di Maio et al. 2016). The resulting simulations
are still quite expensive and may be impractical if the model contains many joints.
Furthermore, when updating a model to correlate with measurements, one must
typically run many simulations while varying the parameters for the joint, a daunting
prospect with current models.
A recent work by Festjens et al. (2013) presented an alternative to traditional
time and frequency domain methods. They computed the quasi-static response of
the structure to a loading representative of a certain mode, and then used the results
of that simulation to compute the effective natural frequency and damping for the
structure when it vibrates in that mode. Their predictions were validated by comparing
them to the transient response computed by integrating a full-order nonlinear finite
element model, which included the preload in the fastener and the resulting Coulomb
friction between the contacting surfaces. Their methodology was found to provide
good estimates of the effective damping and natural frequency, however the costs are
still quite high if many iterations on the model must be run, i.e., for model updating
applications.
One way of reducing the computational burden associated with joints is to replace
the joint region with a simpler element that captures its effective stiffness and damp-
ing. Then, instead of enforcing frictional contact between individual nodes or element
faces, those degrees of freedom can instead be tied (e.g., by multi-point constraints)
to two ends of a single nonlinear element whose constitutive formulation is designed
to reproduce micro-slip behavior internally. Such a modeling approach is the premise
of Segalman’s “whole-joint” model (Segalman 2006). The nonlinear element may
be a hysteretic model such an Iwan (1966), Segalman (2005) model or a Bouc–Wen
(Ismail et al. 2009) or LuGre model. This is appealing because joints are often only
one small part of a complicated, assembled structure, and various experimental stud-
ies have shown that the presence of one or more bolted joints induces only a slight
softening in the natural frequencies, accompanied with a large increase in the damp-
ing as the structure is excited to higher response amplitudes (Di Maio et al. 2016;
Smallwood et al. 2000; Hartwigsen et al. 2004; Abad et al. 2012; Dossogne et al.
2017). Unfortunately, no methods exist in literature that can predict the parameters of
these hysteresis elements based on joint geometry, surface roughness, material, etc.
So, one must use a model updating approach to determine the parameters. For exam-
ple, an experiment may be performed in which the transient-free response is mea-
sured and then the instantaneous natural frequency and damping for each mode are
extracted from those measurements (Song et al. 2004; Deaner et al. 2015; Roettgen
et al. 2017; Lacayo et al. 2017). Then, model updating can then be used to find the
joint parameters that bring the model into an agreement with the measurements as
was done in Lacayo and Allen (2018).
7.2 Test Methods for Identifying Nonlinear Parameters … 271
The nonlinear modal models described in the previous section are shown schemat-
ically in Fig. 7.1, where the oscillator representing each mode contains nonlinear
elements in parallel to the linear spring and damper for each mode in order to pro-
duce the desired deviations from linear behavior. This method assumes that stiffness
is almost constant, but damping can be quite nonlinear. Modes are assumed not to
interact and mode shapes are assumed to not change. The Complex Mode Indica-
tor function plot in Fig. 7.2 shows the difference between a traditional linear modal
model response taken from low input force modal data and response of the system
from a higher level input. One can see that the frequencies appear to drop slightly but
the damping changes drastically at the higher amplitude response. With this approach
applied to such data, a modal filter can be utilized to effectively separate each modal
response (Mayes et al. 2016; Roettgen et al. 2017).
Modal Filter
A modal filter is a spatial filter that uses the mode shapes to pull out the response
of a single mode and rejects the responses of all other modes. By using such a filter,
one can analyze the response associated with that single shape and determine how
nonlinear it is or assume a functional form for the frequency and damping and fit
parameters to identify the model. It should be emphasized that if the response of other
modes seeps through the filter, the estimates of stiffness or damping as a function
of amplitude can be polluted. As a result, the modal filter must be very effective in
order to reliably quantify the nonlinearity associated with the mode shape of interest.
The standard modal filter is a vector of constants, here denoted as Ψ , multiplied
by a vector of actual responses, x, which produces a single modal response q for
mode i, as
Ψ T x = qi . (7.1)
The traditional full modal filter may be derived from the standard modal substitu-
tion as
x = Ψ q. (7.2)
Assuming that one has more responses than modes and the responses are appropri-
ately place one can use
q = Ψ + x, (7.3)
where each row of the pseudoinverse of Ψ can be used as a modal filter for the
associated mode. A variation on this is to use only a single mode of phi, but it is not
as effective as the full modal filter. The SMAC modal filter (Mayes and Johansen
1998) assumes that one has found the best linear estimate of frequency and damping
for a mode in a standard modal test. In this case, one can analytically write the
standard real modes FRF using the associated frequency, damping and drive point
mode shape, and use the measured FRFs from the test in the following equation:
Ψ T Hx = Hqi , (7.4)
where the only unknown is the vector Ψ . This is solved in the frequency domain and
transposed so that the frequency lines are associated with the rows with
Ψ = HxT + Hqi
T
. (7.5)
In our experience, neither of the approaches above always performs better than the
other, however, usually the SMAC modal filter is more robust. A typical set of filtered
data for an attempt to extract a modal response in the mode above 550 Hz for all
three methods is shown in Fig. 7.3.
7.2 Test Methods for Identifying Nonlinear Parameters … 273
Test Methods
The most attractive response for this approach is a modal filtered response from an
impact. Immediately after the impact, there are no external forces acting, so any
changes in the frequency or damping as a function of amplitude can be observed in a
straightforward analysis of the free decay time response. A version of the Restoring
Force Surface method can be used to fit the parameters in the assumed nonlinear
form. Below is the equation that is utilized from this method for an assumed form of
quadratic polynomials to fit the frequency and damping as a function of amplitude.
In this case, the entire right-hand side is known from the modal filtered data and the
frequency and damping extracted from a low level modal test. The functions of q on
the left-hand side are formed from the modal filtered q. This leaves only the quadratic
and cubic constants of the damping and stiffness polynomial to be extracted from
a least squares problem. The rows can be either time response, or response in the
frequency domain.
⎡ ⎤
c1
⎢ c2 ⎥
|q̇| q̇ q̇ 3 |q| q q 3 ⎢ ⎥
⎣ k1 ⎦ = f − q̈ − c0 q̇ − k0 q.
k2
Impact Methods
As stated above, the free decay from an impact is very attractive from an analysis
point of view. This makes and impact test desirable. The first one does the low- level
modal test with impact to get the linear estimates of frequency and damping for the
right-hand side above. Then one simply does a high-level impact and extracts the
274 7 Weakly Nonlinear Systems: Modeling and Experimental Methods
modal response q with a modal filter to obtain the response for other amplitudes.
This is very easy to implement if the modal test has already been set up.
A disadvantage of the impact method has been observed in that since all modes
active in the drive point are excited at once, the drive point accelerometer can overload
at some impact levels, which provides the limit to how much the modes can be excited.
Also, because many modes are excited, the modal filter is taxed to the greatest extent.
A tailored shaker input can overcome these disadvantages and will be discussed in
the next section.
Shaker Windowed Sinusoid
A way to overcome the disadvantages of the impact method is to use a shaker with an
input tailored to the local frequency band of a single mode. This has two benefits. One
is that there are many modes that are not excited at all, so the modal filtering is easier
in the sense that many fewer modes must be filtered out. The second is that all the
energy of the shaker and amplifier can be poured into a single mode to drive it to higher
amplitudes so that the nonlinear parameters are calibrated for the higher amplitudes.
A windowed sinusoid signal has been developed (Pacini et al. 2017), which allows
easy tailoring of the input to excite the mode of interest. The sinusoid frequency
is set to that of the resonant frequency of the mode. Then a triangular window is
multiplied by the sinusoid signal as seen in Fig. 7.4. A narrow triangular window
provides a broad input around the resonant frequency in the frequency domain. A
wide triangular window provides a narrow input around the resonant frequency in the
frequency domain. The frequency domain plot is shown in Fig. 7.5 (Other windows
are available as well.).
The signal is fed into the amplifier and the force is measured. A frequency response
function between the force and input voltage can be calculated. If the force drops
out too much (a common problem in shaker testing), then the FRF can be used to
calculate a voltage input that will fill in the dropout of the force at resonance with a
force correction. The corrected force is plotted in red in Fig. 7.6 with the desired force
in green and the initial force from the windowed sinusoid in blue. A time domain
7.2 Test Methods for Identifying Nonlinear Parameters … 275
modal filtered response is shown from a shaker input in Fig. 7.7. Figure 7.8 shows
hardware which was tested for nonlinearities of the type described here.
After the high-level burst signal is applied, the data are reduced to one q through
the modal filter. The frequency domain RFS nonlinear parameter fitting process
described earlier can then be applied. Experience has shown that if the nonlinearity
excited by this method becomes severe enough to create impact and rattle, the method
is not adequate for such cases. In other cases, it has worked well.
276 7 Weakly Nonlinear Systems: Modeling and Experimental Methods
References
Abad, J., Franco, J. M., Celorrio, R., & Lezáun, L. (2012). Design of experiments and energy
dissipation analysis for a contact mechanics 3D model of frictional bolted lap joints. Advances
in Engineering Software, 45(1), 42–53. https://doi.org/10.1016/j.advengsoft.2011.09.021.
Abaqus analysis user’s guide, User Manual, Simulia (2014).
Bhushan, B. (2013). Concepts and applications of tribology (2nd ed.). Hoboken: Wiley.
Bograd, S., Reuss, P., Schmidt, A., Gaul, L., & Mayer, M. (2011). Modeling the dynamics of
mechanical joints. Mechanical Systems and Signal Processing, 25(8), 2801–2826. https://doi.
org/10.1016/j.ymssp.2011.01.010.
Brake, M. R. (2018). The mechanics of jointed structures: Recent research and open challenges for
developing predictive models for structural dynamics (1st ed.). Berlin: Springer.
Deaner, B. J., Allen, M. S., Starr, M. J., Segalman, D. J., & Sumali, H. (2015). Application of
viscous and Iwan modal damping models to experimental measurements from bolted structures.
Journal of Vibration and Acoustics, 137(2), 021,012. https://doi.org/10.1115/1.4029074.
Di Maio, D., Schwingshackl, C., & Sever, I. A. (2016). Development of a test planning methodology
for performing experimental model validation of bolted flanges. Nonlinear Dynamics, 83, 983–
1002. https://doi.org/10.1007/s11071-015-2382-9.
Dossogne, T., Jerome, T. W., Lancereau, D. P. T., Smith, S. A., Brake, M. R. W., Pacini, B., et al.
(2017). Experimental assessment of the influence of interface geometries on structural dynamic
response. In M. Allen, R. Mayes, & D. Rixen (Eds.), Dynamics of coupled structures (Vol. 4,
pp. 255–261)., Conference Proceedings of the Society of Experimental Mechanics Series Berlin:
Springer. https://doi.org/10.1007/978-3-319-54930-9_22.
Ferri, A. A. (1995). Friction damping and isolation systems. Journal of Mechanical Design, 117,
196–206. https://doi.org/10.1115/1.2836456.
Festjens, H., Chevallier, G., & Dion, J. L. (2013). A numerical tool for the design of assembled
structures under dynamic loads. International Journal of Mechanical Sciences, 75, 170–177.
https://doi.org/10.1016j.ijmecsci.2013.06.013.
Gaul, L., & Lenz, J. (1997). Nonlinear dynamics of structures assembled by bolted joints. Acta
Mechanica, 125(1), 169–181. https://doi.org/10.1007/BF01177306.
Groper, M. (1985). Microslip and macroslip in bolted joints. Experimental Mechanics, 25(2), 171–
174. https://doi.org/10.1007/BF02328808.
References 277
Hartwigsen, C. J., Song, Y., McFarland, D. M., Bergman, L. A., & Vakakis, A. F. (2004). Experi-
mental study of non-linear effects in a typical shear lap joint configuration. Journal of Sound and
Vibration, 277(1–2), 327–351. https://doi.org/10.1016/j.jsv.2003.09.018.
Ismail, M., Ikhouane, F., & Rodellar, J. (2009). The hysteresis Bouc-Wen model, a survey. Archives
of Computational Methods in Engineering, 16(2), 161–188. https://doi.org/10.1007/s11831-009-
9031-8.
Iwan, W. D. (1966). A distributed-element model for hysteresis and its steady-state dynamic
response. Journal of Applied Mechanics, 33(4), 893–900. https://doi.org/10.1115/1.3625199.
Jewell, E., Allen, M. S., Lacayo, R. (2017). Predicting damping of a cantilever beam with a bolted
joint using quasi-static modal analysis. In Proceedings of the ASME 2017 International Design
Engineering Technical Conference IDETC 2017, Cleveland, OH, USA.
Lacayo, R. M., & Allen, M. S. (2018) Updating structural models containing nonlinear iwan joints
using quasi-static modal analysis. Mechanical systems and signal processing. Submitted 2017.
Lacayo, R. M., Deaner, B. J., & Allen, M. S. (2017). A numerical study on the limitations of modal
Iwan models for impulsive excitations. Journal of Sound and Vibration, 390, 118–140. https://
doi.org/10.1016/j.jsv.2016.11.038.
Mayes, R. L., Johansen, D. D. (1998). A modal parameter extraction algorithm using best-fit recip-
rocal vectors. Requires guessing initial frequencies, mode vectors and damping ratios and using
optimization to improve the result. Sounds pretty weak, although it allows one to play around
with the result. Check future papers. Interesting links to papers on “Modal Filters”.
Mayes, R. L., Pacini, B. R., & Roettgen, D. R. (2016). A modal model to simulate typical structural
dynamic nonlinearity. Dynamics of coupled structures (pp. 57–76). Berlin: Springer.
Pacini, B. R., Mayes, R. L., Owens, B. C., & Schultz, R. (2017). Nonlinear finite element model
updating, part i: Experimental techniques and nonlinear modal model parameter extraction.
Dynamics of coupled structures (pp. 263–274). Berlin: Springer.
Petrov, E. P., & Ewins, D. J. (2003). Analytical formulation of friction interface elements for
analysis of nonlinear multi-harmonic vibrations of bladed disks. Journal of Turbomachinery,
125(2), 364–371. https://doi.org/10.1115/1.1539868.
Roettgen, D. R., & Allen, M. S. (2017). Nonlinear characterization of a bolted, industrial structure
using a modal framework. Mechanical systems and signal processing (Vol. 84, Part B, pp. 152–
170). https://doi.org/10.1016/j.ymssp.2015.11.010.
Roettgen, D. R., Allen, M. S., Kammer, D. C., & Mayes, R. L. (2017). Substructuring of a nonlinear
beam using a modal iwan framework: Part i: Nonlinear modal model identification. Dynamics of
coupled structures (pp. 165–178). Berlin: Springer.
Segalman, D. J. (2005). A four-parameter Iwan model for lap-type joints. Journal of Applied
Mechanics, 72(5), 752–760. https://doi.org/10.1115/1.1989354.
Segalman, D. J. (2006). Modelling joint friction in structural dynamics. Structural Control and
Health Monitoring, 13(1), 430–453. https://doi.org/10.1002/stc.119.
Smallwood, D. O., Gregory, D. L., Coleman, R. G. (2000). Damping investigations of a simplified
frictional shear joint. In 71st Shock and Vibration Symposium, Arlington, Virginia.
Song, Y., Hartwigsen, C., McFarland, D., Vakakis, A., & Bergman, L. (2004). Simulation of dynam-
ics of beam structures with bolted joints using adjusted Iwan beam elements. Journal of Sound
and Vibration, 273, 249–276. https://doi.org/10.1016/S0022-460X(03)00499-1.