See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/253671084

Sensors, Degrees of Freedom, and Generalized Modeshape Expansion

Methods, #8

Article  in  Proceedings of SPIE - The International Society for Optical Engineering · March 1999

CITATIONS READS

26 116

1 author:

Etienne Balmes
Ecole Nationale Supérieure d'Arts et Métiers
123 PUBLICATIONS   1,430 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Dynavoie : reduced FEM modeling of train tracks View project

OSCAR pantograph catenary interaction View project

All content following this page was uploaded by Etienne Balmes on 16 September 2014.

The user has requested enhancement of the downloaded file.

Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 1

SENSORS, DEGREES OF FREEDOM, AND GENERALIZED

MODESHAPE EXPANSION METHODS

Etienne Balmès

École Centrale Paris, MSSMat

92295 Ch âtenay-Malabry, France
[email protected]

ABSTRACT duction [1] ) and modal/SEREP [2, 3] expansion methods. Dy-

Modeshape expansion techniques deal with the spatial incom- namic expansion [4] is often acknowledged as the best exten-
patibility linked to the measurement of modeshapes through sion of static expansion but its use has been limited because
a limited set of physical sensors and their analytical predic- of its high numerical cost. The limitations linked to the lack
tion at a (larger) number of finite element degrees of free- of residual terms in modal expansion have shown the need
dom. Expansion methods are formulated here with the mini- to create hybrid methods [5, 6] . But such combinations imply
mal assumption that sensor measurements be linearly related the setting of parameters (coefficients or selection of target
to finite element DOFs. This allows the localization of experi- modes) which leads to a possibly dangerous form of user in-
mental points independently from the FE mesh, the measure- volvement.
ment of translations in arbitrary directions, as well as the com- The second part of this paper shows how observation equa-
bined use of translation and strain sensors. Various methods tions, static/dynamic expansion, and finite element model re-
to interpolate sensor motion from DOF motion are discussed. duction methods can be combined to formulate generalized
Generalized static, dynamic, and minimum residual expansion static and dynamic expansion methods. The discussion of
methods are introduced and it is shown how reduction meth- the assumptions underlying the proposed Reduced Basis Dy-
ods traditionally used in Component Mode Synthesis applica- namic Expansion (RBDE), then leads to the introduction of
tions allow the use of the proposed methods for industrial size a Minimum Residual Expansion (MRE) and its extension us-
finite element models. ing quadratic inequality constraints on measurement errors
(MRE-QI) in a fashion similar to the QI versions of standard
1 INTRODUCTION methods discussed in [7] . The proposed methods are then
briefly illustrated.
Expansion methods seek to estimate the motion at all DOFs of
a finite element model based on measured information (mode-
shapes or frequency response functions). While a signifi- 2 SENSORS AND DOFS
cant literature exists in the area, well established method still
2.1 Motivation
present major shortcomings.
To illustrate the distinction between sensors and DOFs let us
The theory is always presented using the assumption that
consider the laser vibrometer test of a harp resonator shown
measurements corresponds to DOFs which leads to major
in figure 1.
and unnecessary limitations. DOFs of finite element models
are translations/rotations at finite element mesh nodes. Con-
fusing sensors and DOFs thus imposes that the sensors be
located at mesh nodes. Furthermore, using sensors that cor-
respond translations in arbitrary possibly non-orthogonal di-
rections or strain measurements is either impossible or very
cumbersome to implement. The present paper shows that the
use of an observation equation giving a linear description of
the relation between measurements and DOFs gives a sound
theoretical basis to solve expansion problems even for very
complex test configurations.
Most expansion methods can be categorized by how they re-
Figure 1: Sensors, test nodes, DOFs
late to, modify, or combine the static (based on Guyan re-
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 2

When doing the test, one measures translations in the line of The last item comes from the fact that correlation is gener-
sight of the measurement head. The measurements made de- ally performed using software packages that differ from the
fine a set of sensors and one can build an input/output model general purpose finite element codes used for analysis. It
from the exciter(s) to the sensors. Sensors typically used for is important to provide methods that will work without insider
modal analysis are knowledge of how finite elements are formulated for two rea-
sons. First detailed element formulation and source code,
• accelerometers which, at a given point, measure the accel- is not publicly available for major commercial finite element
eration in one or more directions codes. Second, methods based on particular elements would
need to be adapted to the countless elements available in
• laser vibrometers which measure velocity or displacement most codes.
of a surface in the line of sight
• strain gauges which measure local deformation in one or 2.2 Practical methodology
more directions.
The distinction of DOFs {q} and outputs {y} through the use
In the testing process, it is important to visualize test results of linear observation equations of the form
using a wire frame representation of the structures linking
physical points where a measurement is made (these will be
{y(t)} = [c] {q(t)} (1)
called test nodes). Here the motion is only measured in a
single sensor direction assumptions must thus be made to re- is the key of the proposed methodology. One thus considers
construct the 3-D motion of test nodes. Setting motion orthog- that the dynamics of a system are described by an evolution
onal to the line of sight measurement to zero is easy enough equation (2) and a set of observation equations (1). This de-
and acceptable for intermediate verifications of test results. scription is common in control theory (state-space models are
For a laser head positioned close to the harp, this however composed of two sets of equations) but rarely used in me-
significantly differs from true motion and is not acceptable for chanical applications. Its usefulness will be shown here.
test/analysis correlation.
In the simplest case where sensors are positioned at FE
To go beyond the wire-frame representation of the test config- nodes and measure in DOF directions, the observation ma-
uration, one needs to start using a mechanically meaningful trix [c] is just a Boolean matrix often called a localization ma-
model which is generally a finite element model of the struc- trix. An objective of this section is to show that in general, one
ture. Finite element models of linear structures lead to second should use observation matrices that are not just Boolean.
order differential equations with a number of Degrees of Free-
dom (DOF) which generally correspond to translations and For translation measurements, which is the common case, it is
possibly rotations of finite element mesh nodes. For this ex- useful to consider two levels of observation. First, one relates
ample and most industrial structures, the mesh is generated the 3-D motion of test nodes to DOFs. Then, one projects this
by a CAD system and the number of nodes is strongly depen- motion along an arbitrary sensor direction to obtain the sen-
dent on the geometric complexity of the structure. sor measurement (this allows the use of an arbitrary number
of possibly non-orthogonal sensors). The first level of obser-
From the analysis of this example it appears that a complete vation is clearly the difficult part.
methodology for test/analysis correlation
For test nodes that do not coincide with finite element nodes
• must allow arbitrary numbers of sensors (scalar measure- (second requirement), the optimal approach would probably
ments of translation, rotation, strain) at each test node be to get inside the formulation of the element to which the
sensor is connected and to use its shape functions to deter-
• must allow an independent selection of test and mesh mine the displacement of a particular node in physical space.
nodes We have however already mentioned that this is not practical
(third requirement).
• should be independent from the element formulation
The simplest alternative is to consider the nearest node. For
The first item is needed for non triaxial measurements. Such coarse meshes or movements with significant rotational com-
measurements are often used to overcome limitations of the ponents, the error made by neglecting the relative motion of
correlation methodology in cases where mono or biaxial mea- the nearest and physical nodes can be significant.
surements would be sufficient. Cost effectiveness dictates an
The natural extension is thus to take rotations into account
optimal repartition of measurements and thus the use of non-
by imposing a linearized rigid connection between the two
triaxial sensors.
nodes. Rotations are typically not defined for solid elements
Non coincidence of test and FE nodes is a practical constraint. and must be handled with caution in the case of plates
The FE mesh and test nodes are often positioned by different and shells. Typical finite elements either eliminate the ro-
teams, at different times, with different objectives. While bring- tation around axes normal to the shell (this DOF is often
ing the test and analysis people to cooperate better is certainly called a drilling DOF) or use it to improve the convergence
desirable, one must provide methods that do not require full of the membrane properties of the element. Thus, the popu-
coordination. lar MSC/NASTRAN QUAD4 element uses drilling DOFs that
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 3

have no physical meaning, while the QUADR uses this rotation frequency range is typically small. This principle is the ba-
to represent motion at the element mid-sides [8] . The vector sis of reduction methods which are briefly summarized before
of rotational DOFs thus often does not give an accurate in- introducing generalized expansion methods.
dication of the local rotation of the physical vector linking the
test/FE node pair. 3.1 Short reminder on model reduction
A third method acknowledges the problems linked to rotations The fundamental approach for model reduction is to project
and thus uses a rigid link with rotations estimated based on the FE mass and stiffness matrices on the vector space
the translations of two additional non–collinear nearby nodes. spanned by the columns of a reduction matrix [T ] which has
Figure 2 illustrates the nodes used to build the observation less columns than rows
matrix of an engine block cover (see details in section 4). The
rigid links connect the circles, while the additional nodes used [MR ] = [T ]T [MF E ][T ] and [KR ] = [T ]T [KF E ][T ] (3)
to infer rotations are shown as pluses linked to the nearest
FEM node used. The plot and the observation matrix was
The choice of the reduction basis has been the object of nu-
here automatically generated using the Structural Dynamics
merous publications and categories of methods known as
Toolbox [9] with an additional effort to ensure that links only modal analysis, condensation, component mode synthesis,
use nodes of the physical component on which the sensor is
sub-structuring [10, 11] . For the present paper, vectors con-
located.
sidered for reduction will combine

• eigenvectors of the nominal model. They are used to en-

sure that the reduced model will be valid over a certain fre-
quency range. Boundary conditions need not correspond
to those of the test of interest.
• static responses to particular imposed displacements or
loads. They ensure spatial completeness for the consid-
ered loads.
Figure 2: Nodes used to build the observation matrix of an
engine block cover A key property for the present paper is the fact that the dynam-
ics of reduced models only depend on the subspace spanned
by [T ]. More precisely, the dynamics of the system are char-
The third method (rigid link with inferred rotation), is currently acterized by the dependence of outputs {y(t)} on the inputs
considered as giving a good trade-off between cost and ac- {u(t)}. This relation is described by two equations (evolution
curacy. Improvements to this approach will obviously be in- and observation) given for the reduced model by
troduced, but results obtained so far have always been satis-
factory (which is not the case of the two other methods men- [MR ]{q̈R } + [KR ]{qR } = [T ]T [b]{u(t)}
(4)
tioned). {y(t)} = [c][T ] {qR (t)}
Finally, observation equations provide a theoretical basis al- or in the frequency (Laplace) domain
lowing to deal with measurements of mixed nature (transla-
tions, rotations, strains, ...), this will not be emphasized in the
examples but is another important reason to distinguish sen-


[ZR ]{qR } = MR s2 + KR {qR } = T T b {u(s)}
sors and DOFs. (5)
{y(s)} = [c][T ] {qR (s)}

3 EXPANSION OF TEST DATA

The u, y relationship

 is clearly identical for a model projected
We will now consider that the FE model leads to equations of on [T ] or on T̃ = [T ][A] with A non-singular
motion of the form






[cT ] T T ZF E (s)T T T b = cT̃ T̃ T ZF E T̃ T̃ T b (6)
[MF E ]{q̈} + [KF E ]{q} = [b]{u(t)} (2)

but instead of the traditional partition of q into measured and We will refer to this property as the equivalence of models
unmeasured DOFs, we will use the observation equation ( 1). projected on various bases of the same subspace.
Expansion methods estimate the motion of DOFs based on
the observation of particular measurements at sensors. The 3.2 Generalized Static Reduction/Expansion
basis for expansion is the fact that while a finite element model
has many DOFs, the dimension of a subspace accurately rep- The standard Guyan or static reduction method [1] partitions
resenting the response to a limited sets of loads in a limited the DOFs in two sets of active and complementary DOFs. The
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 4


−1
active DOFs, which correspond to interface DOFs in Compo- so that T̃ cT̃ is indeed the static expansion considered
nent Mode Synthesis applications, are for tests assumed to in (8) and the proposed method can be called a generalized
correspond directly to sensor measurements static expansion.

{qa } = {yT } (7) 3.3 Generalized Dynamic Expansion

The assumption that inertia forces are negligible is often poor
Assuming further that the inertia forces acting on the comple-
for some deformations. Dynamic expansion was thus in-
mentary DOFs are negligible, there exists an exact relationship
troduced to take into account that in many applications the
between active and complementary DOFs given by
considered deformation is associated to a known frequency
    (modeshapes, measured FRFs). The Kidder dynamic ex-
qa I
= −1 {qa } = [T ]{qa } (8) pansion [4] thus generalizes static expansion by considering
qc −Kcc Kca
the exact steady-state response to an imposed harmonic dis-
which can be used as an expansion method (given {q a }, the placement at sensors that correspond to DOFs
relation above gives an interpolation for {q c }) or a reduction    
method (application of (4) with T given by (8)). qa (ω) I
= {qa (ω)} (12)
qc (ω) −Z(ω)−1
cc Z(ω)ca
When using a non-boolean observation equation, this method
is not directly applicable. The fundamental assumption made This is a simple generalization of the static expansion but
in the classical method is that the only significant forces (iner- guarantees that the exact modes will be found if {q a (ω)} is a
tial and external) are applied on active DOFs. One thus intro- restriction of the shape and the associated frequency are ex-
duces a generalized static expansion defined as the static re- act. While this method is clearly much more accurate than the
sponse to forces collocated to the sensors such that the predicted static expansion, it uses a different reduction basis for each
response at sensors corresponds exactly to the measurement. frequency. For industrial FE models in an model updating
For a set of sensors observed through (1), unit collocated phase, dynamics responses to unit loads at sensors have to
forces (those associated with the sensors by the reciprocity be computed for each target frequency and each step of the
assumption) are given by the columns of [c] T . The considered updating procedure. Numerically this is often very expensive
reduction basis is thus defined by and difficult to justify for a method that typically only has an
effect on few modes.


[KF E ] T̃ = [c]T (9) For an arbitrary reduction basis T̃ with more vectors than sen-
sors, one can define a generalized dynamic expansion (GDE)
as the search for harmonic inputs {u(ω)} collocated to the
In many cases, the finite element model will have rigid body
considered sensors and such that the predicted response cor-
modes so that [KF E ] is singular. This is a standard difficulty
responds to the measurement
in structural dynamics and the two standard approaches to
solving this problem are to orthogonalize the response with

T
T̃ T ZF E (ω)T̃ {qR (ω)} = cT̃ {u(ω)}
respect to rigid body modes [c] T [12] or to use a mass-shifted

and (13)
stiffness matrix [11] . cT̃ {qR (ω)} = {yT est (ω)}
The second step of the generalized static expansion takes ad-
vantage of the invariance property ( 6). To go back to an equa- The statement of this problem is clearly invariant by any
tion of the form (8), one wants to find a new basis T of the change of basis
in the subspace spanned by T̃ . Assuming
subspace spanned by T̃ such that cT = I. Assuming that the that the rank of cT̃ is equal to the number of sensors, it is
observations of the vectors of the reduction basis are inde- possible to build a transformation A such that



pendent (c T̃ is non-singular, this assumption will be relaxed in cT̃ A = [cT ] = [I]NS×NS [0]NS×NR−NS (14)
the next section), a generalized static expansion will thus be
given by The last N R −
N S columns of A correspond to any basis
of the kernel of cT̃ (which can be found using the singular


−1
{q} = T̃ cT̃ {y} (10) value decomposition of this matrix for example [9] ). Given a
basis of the kernel, the first N S column of A can be found
For a structure without rigid body modes and a boolean ob- by applying (10) to a basis of a subspace orthogonal to the
servation matrix ([c]{q} = {q a }), one can verify that kernel.
With a reduction basis T verifying (14), the generalized
 DOFs

 −1   of the reduced model are of the form {q R }T = yTT qRc T
and

 Kaa Kac I
T̃ = the solution of problem (13) is
Kca Kcc 0
  (11)  
I
−1
−1 I
= −1 Kaa − Kac Kcc Kca {q} = [Te ]{yT } = [T ] {yT (ω)} (15)
−Kcc Kca −Z(ω)−1
RcRc ZRcT
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 5

which is very similar to (12) but only requires the inversion of imization which has a solution of the form
Z(ω)RcRc which has dimensions N R−N S rather than N −N S
which carried an unacceptable numerical cost.
−1
{q} = [T ]{qR } = [T ] [cT ]T [cT ] [cT ]T {yT est } (17)
Note that the invariance of problem ( 13) by any change of ba-
sis in the subspace spanned by T̃ , implies that the solution
found in (15) is independent of the choices made for the coor-
It can easily be verified that for Boolean [c] selecting measured
dinate change leading to ( 14).
DOFs and a reduction basis containing a set of target modes,
The proposed generalized dynamic expansion (GDE) has the the application of (17) is known as the SEREP method [3] .
advantages of the standard dynamic expansion (no sensitiv-
ity to mass effects) while allowing low cost computations by Ref. [7] showed how classical expansion methods can be re-
using a model reduction (RBDE : reduced basis dynamic ex- formulated in terms of minimization problems and how the
pansion). A typical reduction basis would combine static re- classical least squares problem with quadratic inequalities
sponses to unit sensor loads (to ensure that the result is at (LSQI) [15] allows to account for errors in test results. The
least as good at static expansion), analytical target mode- same work can and should be done for the extensions pro-
shapes (to guarantee exact expansion for these modes), and posed in this paper.
possibly other vectors (target modeshapes of other FE con-
The principle of LSQI based expansions is as follows. As-
figurations, modeshape sensitivities, etc. [13, 14] ). It should
suming that the measurement {y T } is inexact, one seeks
be noted that this method can be viewed as the definition of a
a smoothed {ŷ T } close to the measurement (verifying the
projector in the general class of hybrid methods discussed in
quadratic inequality ||{ŷ T − yT }|| ≤ α||{yT }||) such that the
Refs. [5, 6] expanded vector is more realistic. Various definitions of realis-
tic lead to different expansion methods.
3.4 Minimum dynamic residual expansion Thus, the minimization of the strain energy of the dynamic
The assumption that dynamic loads used for the expansion residual associated to a modeshape leads to a MDRE-QI
are only applied at sensor locations is not particularly realistic. method where one solves
In particular for finite element updating procedures where the
model is known to be incorrect, the dynamic residual (R j =  T  
Z(ωj )φj for modeshapes or R j = Z(ω)q − F for frequency ŷT
−1 ŷT
min || [Z(ωj )] K̂ [Z(ωj )] || (18)
response to the harmonic load F) should be non zero at most qRc qRc qRc
DOFs.
One thus defines here a minimum dynamic residual expansion with
(MDRE) which seeks to minimize the strain energy associ- ||{ŷT } − {φT }|| ≤ α||{φT }|| (19)
ated to the dynamic residual. For the case of a modeshape
and a reduced
  basis coordinate change such that {q R }T =
T T
yT qRc , one thus seeks the solution of
4 ILLUSTRATIONS
 T  
φT
−1 φT
4.1 Observation equations
min || [Z(ωj )] K̂ [Z(ωj )] || (16)
qRc qRc qRc The first part of this paper motivated the need to use observa-
tion matrices to distinguish sensors and DOFs and discussed

 various methods to interpolate test node displacement from fi-
where K̂ is a mass shifted stiffness for cases with rigid body
modes and the standard stiffness otherwise. nite element DOFs. Figure 3 illustrates this aspect for the case
of an engine block cover (data kindly provided by Renault-
The numerical cost associated with this expansion method is DR). The test configuration uses 182 sensors distributed at
only acceptable for cases with q Rc not exceeding a few hun- 91 nodes with measurements made in two directions at each
dred DOFs. The MDRE should thus only be considered for re- node. The (coarse) finite element model uses 2218 nodes,
duced basis dynamic expansion using bases similar to those 7758 DOFs (solid elements have no rotational DOF), 444
considered in the previous section. plate/shell elements for the cover and 1386 solid elements for
the base.
3.5 Estimation error and smoothing Test DOF motion was interpolated based on analytical mode-
shapes computed with MSC/NASTRAN using the nearest
As a result of measurement and estimation errors (bias and
node, linearized rigid link and alternate rigid link (interpolated
variance), identified modeshapes are never exact. The obser-
rotations) methods discussed in section 2. The plot in figure 3
vation of the expanded vector should thus be allowed to differ
illustrates the difference between the first 2 methods and the
somewhat from the measurement.
3rd (assumed to be ”exact”). The plot clearly indicates that
For cases with a reduction basis with less vectors than sen- the nearest node method shows less difference which, at first,
sors, the problem is easily solved through a least squares min- seems surprising.
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 6

TABLE 1: CPU times for modeshape expansion of the

engine block cover test (on an SGI-R10000 processor
running MATLAB 5.2.1 and SDT 3.1) .

Model assembly 72 s
Eigenvalue solution (20 modes) 21 s
Full order Dynamic Expansion (14 modes) 77 s
Reduction for RBDE (202 shapes) 60 s
Reduced Basis DE (14 modes) 4s

1 block extraction and factorization). This cost is directly pro-

portional to the number of modes to be expanded.
M−MAC

0.9
Near / A.Rigid
Rigid / A.Rigid The cost of reduced basis version of the generalized dynamic
0.8
expansion (RBDE) is composed of an up-front cost linked to
0.7 the reduction and an additional cost linked to each expansion.
0 5 10 15
Mode # The later part is very small for this model and only depends
on the reduction basis size so that it would not grow for larger
Figure 3: Wire frame representation of test configuration, models. The reduction is here fairly expensive but it is driven
finite element model, MAC comparisons of different by the number of shapes in the reduction basis (here 182
methods for test node interpolation static responses associated to each sensor + 20 modeshapes
already computed).
This case was chosen because it has many sensors and few
modes to be expanded, the full and reduced order methods
are thus found to have similar computational costs. The ad-
vantage of the reduced order method becomes significant
when the model size or number of expanded modes is in-
z Near
Rigid creased, the number of sensors is decreased, or if the reduc-
y x A. Rigid tion basis is already computed for other reasons (estimation
of modeshape sensitivities, ...).
Figure 4: Interpolations of first analytical torsion mode on
test mesh. 4.3 Selection of expansion method
The relative merits of various expansion methods discussed in
this paper are difficult to establish since they all work in many
The origin of this difference is actually linked to drilling DOFs cases. This section will thus seek to illustrate typical difficul-
of the QUAD4 elements used in the NASTRAN model. At the ties of that have motivated the development of the MDRE-QI
joint between the cover and the support block (top in figure method which is currently considered as best by the author.
4), these DOFs are mostly oriented around the y direction so
that off-sets of test nodes in the xz plane will, for the standard The illustrations will be made using the example of the GAR-
rigid link method, lead to non-physical corrections linked to the TEUR SM-AG-19 testbed [16, 17] . This test article is a simple
drilling rotation. In figure 4, when comparing the response at structure with publicly available test results (contact the author
the joint level (top of the structure in the figure), one indeed for more information). The simple 816 DOF/90 element model
sees that the basic rigid method significantly differs from the of the structure shown in figure 5, the nominal 24 sensor con-
two other results. figuration, the first 14 modeshapes measured by participant C
will be used here.
4.2 Computational times
Figure 5 compares, for mode 11, static and dynamic expan-
Computational times are always dependent on many factors sion to the finite element modeshape. For the mode shown,
including computer, level of software optimization, ... All com- the 3 sensors available in the x direction (shown as arrows on
putations and illustrations are here performed using the Struc- the plot), cannot capture the bending of the fuselage. Static
tural Dynamics Toolbox for use with MATLAB [9] . The only expansion thus gives a significant rigid body contribution for
area where version 3.1 of the SDT is known to be fundamen- the fuselage and wings which is very unrealistic.
tally slower than fully compiled codes is in the computation of
element matrices. The comparisons made in table 1 are thus
quite representative of the real cost of expansion methods.
Full order dynamic expansion has a high cost (linked to the
Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 7

Mode 11 (103 Hz)

1

0.8

Static 0.6

M−MAC
Dynamic Modal
FE mode 0.4 Static
RBD
0.2 MDR
MDR−QI
0
2 4 6 8 10 12 14
Flex Mode #

Figure 7: Mass weighted MAC of modal expansions of

GARTEUR test data versus FE model

constitutive law criterion) is displayed for the first 4 modes.

Figure 5: Modeshape expansion for the GARTEUR
SM-AG-19 testbed. Mode 1 at 6.376 Hz Mode 2 at 16.1 Hz

Most modal test have enough sensors for static expansion to

give correct results for many modes. But this example shows
that even in simple tests, it can occasionally fail miserably
to represent inertia effects whereas dynamic expansion gives
much more robust results.
Figure 6 analyzes modal/SEREP results for various selec-
tions of modeshapes. Keeping the 6 rigid body modes works
Mode 3 at 33.12 Hz Mode 4 at 33.53 Hz
well with only 8 flexible modes (1:14 case) and poorly with
14 (1:20). Keeping 14 flexible modes (7:20) works well but
adding more 7:26 and 7:30 shows clear deterioration. The
modal method thus shows a lack of robustness which elimi-
nates it as a good alternative to static expansion.

Modal expansion
1

0.8
1:14
Figure 8: Strain energy distribution of dynamic residual for
7:20 RBDE
M−MAC

0.6
1:20
0.4
7:26
7:30
0.2 The experimental modeshapes used show a slight calibration
0
error of the sensors at the middle of the drums so that they
2 4 6 8 10 12 14
appear to bend. The direct MDRE result thus indicates large
error levels on the drums which is indication of experimental
Figure 6: Mass weighted MAC of modal expansions of
error and not of model errors which are of interest.
GARTEUR test data versus FE model
When gradually increasing α (how to do this is really the
problem with the method), the MDRE-QI method accounts for
measurement errors and gives the correct result that the first
Figure 6 confirms the result mentioned previously that dy-
mode is mostly in error because of a poor representation of
namic expansion is more accurate than static (modes 11:14).
Modal results (given for the case with 8 flexible modes) is good the viscoelastic constraining layer [18] , mode 2 shows an er-
in this range. No distinction is made between dynamic and re- ror on the tail connection, ...
duced basis dynamic expansion as differences are minimal.
The RBDE, MDRE and MDRE-QI methods give an increas- 5 CONCLUSION
ingly good correlation which could be expected but is not nec-
essarily a good indication of true correlation. The use of observation matrices gives a sound theoretical ba-
sis extend expansion methods to non trivial test configurations
Finally, the difference between MDRE and MDRE-QI is (non coincident test and FEM nodes, non-orthogonal sensors,
illustrated in figures 8-9 where the strain energy dis- strain measurements, ...). Combining observation matrices
tribution associated to the dynamic residual {R j } =

−1
 and model reduction gives a good framework to extend tradi-
K̂ K − ωj2id M {φj ex } (a particular case of the error in tional modeshape expansion methods and provide computa-
 Presented at IMAC 1999, reproduced with permission from SEM : www.sem.org 8

Mode 1 at 6.376 Hz Mode 2 at 16.1 Hz [4] Kidder, R., Reduction of Structural Frequency Equa-
tions, AIAA Journal, Vol. 11, No. 6, 1973.

[5] Kammer, D., A Hybrid Approach to Test-Analysis Model

Development for Large Space Structures, Journal of Vi-
bration and Acoustics, Vol. 113, No. 3, pp. 325–332,
1991.

[6] Roy, N., Girard, A. and Bugeat, L.-P., Expansion of Ex-

perimental Modeshapes - An Improvement of the Projec-
tion Technique, IMAC, pp. 152–158, 1993.
Mode 3 at 33.12 Hz Mode 4 at 33.53 Hz

[7] Levine-West, M., Kissil, A. and Milman, M., Eval-

uation of Mode Shape Expansion Techniques on the
Micro-Precision Interferometer Truss, IMAC, pp. 212–
218, 1994.

[8] MacNeal, R., Finite Elements: their Design and Perfor-

mance, Marcel Dekker, Inc., New York, 1994.

Figure 9: Strain energy distribution of dynamic residual for [9] Balmès, E., Structural Dynamics Toolbox 3.1 (for
RBDE-QI with α = 3% use with MATLAB), Scientific Software Group,
http://www.sdtools.com, 1998.

[10] Craig, R. R. J., A Review of Time-Domain and Fre-

tionally efficient methods.
quency Domain Component Mode Synthesis Methods,
While all methods work on many cases, some are more ro- Int. J. Anal. and Exp. Modal Analysis, Vol. 2, No. 2,
bust than others. Modal based methods tend to be sensitive pp. 59–72, 1987.
for mode selection. Static expansion gives poor results in con-
figurations with few sensors in certain key areas. Dynamic [11] Géradin, M. and Rixen, D., Mechanical Vibrations. The-
and minimum residual expansions combine static and modal ory and Application to Structural Dynamics., John Wiley
results and their reduced basis versions have relatively low & Wiley and Sons, 1994, also in French, Masson, Paris,
numerical costs, they thus seem to be the best. One of the 1993.
advantages of the modal approach was to allow some level [12] Farhat, C. and G éradin, M., On the General Solution
of smoothing trough a least-squares formulation. Accepting by a Direct Method of a Large-Scale Singular System of
that measurements are not exact seems a very important fac- Linear Equations: Application to the Analysis of Floating
tor in any expansion method. The proposed formulation using Structures, International Journal for Numerical Methods
quadratic inequalities seems a promising possibility but ex- in Engineering, Vol. 41, pp. 675–696, 1998.
perience on how to set the error level parameter and how
to choose a proper norm for the measured deformation still [13] Balmès, E., Parametric families of reduced finite ele-
needs to be gained. ment models. Theory and applications, Mechanical Sys-
ACKNOWLEDGMENTS tems and Signal Processing, Vol. 10, No. 4, pp. 381–394,
1996.
This work is supported by Arospatiale and Eureka project no
1562 Sinopsys (Model based structural monitoring using in- [14] Balmès, E., Efficient Sensitivity Analysis Based on Finite
operation system identification). The finite element model of Element Model Reduction, IMAC, pp. 1168–1174, 1998.
the engine cover was kindly provided by Renault-DR.
[15] Golub, G. and Van Loan, C., Matrix computations,
REFERENCES Johns Hopkins University Press, 1983.

[1] Guyan, R., Reduction of Mass and Stiffness Matrices, [16] Balmès, E., GARTEUR group on Ground Vibration Test-
AIAA Journal, Vol. 3, pp. 380, 1965. ing. Results from the test of a single structure by 12 lab-
oratories in Europe., IMAC, pp. 1346–1352, 1997.
[2] Kammer, D., Test-Analysis Model Development Using
an Exact Modal Reduction, International Journal of An- [17] Balmès, E., Predicted Variability and Differences Be-
alytical and Experimental Modal Analysis, pp. 174–179, tween Tests of a Single Structure, IMAC, pp. 558–564,
1987. 1998.

[18] Plouin, A. and Balm ès, E., A test validated model of

[3] O’Callahan, J., Avitabile, P. and Riemer, R., Sys-
plates with constrained viscoelastic materials, IMAC, pp.
tem Equivalent Reduction Expansion Process (SEREP),
1440–1446, 1999.
IMAC VII, pp. 29–37, 1989.

View publication stats