Test-Based Computational Model

Updating of a Car Body in White

C. Schedlinski and F. Wagner, ICS Solutions, Langen, Germany
K. Bohnert, Porsche AG, Weissach, Germany
J. Frappier, Volkswagen AG, Wolfsburg, Germany
A. Irrgang, BMW AG, Mnchen, Germany
R. Lehmann, DaimlerChrysler AG, Sindelfingen, Germany
A. Mller, Audi AG, Ingolstadt, Germany
To validate finite-element models, test data from experimental modal analyses may be utilized. The model data must be
highly accurate since they form the basis for subsequent validation efforts. An integrated validation strategy is presented
that takes into account the complete process chain from
model-based test design through modal testing, data evaluation, test/analysis correlation to computational model updating. By means of a real car body in white, the single steps of
the validation strategy are highlighted, and it is shown that
very encouraging results can be obtained even for very complex systems.
The fidelity of structural mechanical finite-element analyses (FEA) can be evaluated by using data from static or dynamic
tests. Eigen frequencies and eigen vectors are employed that
can be identified from vibration tests by means of experimental modal analysis (EMA).1,2 The deviations between test and
analysis allow for evaluating the quality of the target finiteelement model. If the deviations between test and analysis are
not acceptable, the elastomechanical system needs to be reviewed and eventually updated to validate the finite-element
model. It is vital to keep uncertainties from the experimental
investigations as small as possible and to generate an optimal
data base for subsequent model validation by thorough test
planning and test execution.
If the structure of the finite-element model with respect to
discretization, chosen element types, etc., is correct,3 the test/
analysis deviations can be minimized by changing appropriate parameters based on the experience of the engineer in
charge. However, this procedure has limitations for real
elastomechanical systems due to the large number of parameters to be considered. Here,techniques for computational
model updating (CMU) need to be applied to allow for simultaneous updating of multiple model parameters.4,5 These techniques minimize the test/analysis deviations and enable validation of the finite-element model.
In practical applications, the finite-element model structure
is often incorrect. Computational model updating techniques
can still be utilized, but final parameter changes often do not
allow for a physical interpretation. They are mathematical substitutes for simply reducing the deviations between test and
analysis. Whether individual parameter changes are acceptable
or if the finite-element model is to be revised often depends
on the intended application of a particular model. A large and
physically uninterpretable increase in shell thickness might be
irrelevant for vibration analyses but would be unacceptable for
stress analyses.
For the project presented in this article, test planning and
computational model updating is supported by ICS.sysval, a
special MATLAB software package for model validation, developed by ICS and Professor Michael Link of the University
of Kassel.6 Among other things, this software tool allows for
direct update of large-scale MSC.Nastran finite-element
models from experimental modal data (eigen frequencies and
Based on a paper presented at ISMA2004, the 2004 International
Conference on Modal Analysis, Noise and Vibration Engineering,
Leuven, Belgium, September 2004.

SOUND AND VIBRATION/SEPTEMBER 2005

mode shapes). The MSC.Nastran solvers are used for eigen

value and eigen vector sensitivity analysis under Solution
200 (optimization).

Theory Overview
The foundation for updating physical stiffness, mass, and
damping parameters is a parameterization of the system matrices according to Equation 1:4,5
(1a)
K = K A + a i K i , i = 1 ... na

M = M A + b jM j ,

j = 1 ... n b

D = DA + g k Dk ,

k = 1 ... ng

(1b)

(1c)
where:
K A,M A,DA = initial analytical stiffness, mass and damping matrices, respectively
K i,M j,Dk = given substructure matrices defining location and
type of model uncertainties
[ai,bj ,g k] = unknown design parameters.
This parameterization permits local adjustment of uncertain
model regions. By utilizing Equation 1 and appropriate residuals, which consider different test/analysis deviations, the following objective function can be derived:

J(p) = DzT W Dz + pT Wp p min

(2)
where:
p = [a i,b j ,g k] = vector of unknown design parameters
Dz = residual vector
W,Wp = weighting matrices.
The minimization of the objective function yields the desired
design parameters p. The second term on the right-hand side
of Equation 2 is used for constraining the parameter variation.
The weighting matrix must be carefully selected, as for W p >>
0 no parameter changes will occur. 4
The residuals Dz = zT z(p), (zT = test data vector, z(p) = corresponding analytical data vector) are usually nonlinear functions of the design parameters. Thus, the minimization problem is also nonlinear and has to be solved iteratively. One
solution is the application of the classical sensitivity approach. 5 Here, the analytical data vector is linearized at point
0 by means of a Taylor series expansion truncated after the linear term. Proceeding this way leads to:
(3)
Dz = Dz0 - G0 Dp
where:
Dp = p p 0 = design parameter changes
Dz 0 = z T z(p 0) = test/analysis deviations at linearization
point 0
G 0 = z/p| p = p = sensitivity matrix at linearization point 0
0
p 0 = design parameters at linearization point 0
As long as the design parameters are not bounded, the minimization problem (Equation 2) yields the linear problem (Equation 4). The latter is to be solved in each iteration step for the
current linearization point:
(G0T W G0 + Wp) p = G0T Wz0

(4)

For Wp = 0, Equation 4 represents a standard weighted leastsquares approach. Of course, any other mathematical minimi-

19

FE
modeling

0, X
experimental
modal analysis

automatic update

analytical modal
analysis

test planning

0T, XT

N.O.K.

correlation

update of selected
inertia and stiffness
parameters

O.K.

end

Figure 1. Model validation strategy.

zation technique can be applied for solving Equation 2.

In contrast to the assembly of the analytic stiffness and mass
matrix, generating the analytic damping matrix is usually a difficult task. Modal damping parameters can be utilized alternatively for treating system damping in an update process. See
the literature for further discussions on this topic. 4,8
Typically, the eigen value and the eigen vector residuals are
employed. Here, the analytical eigen values (squares of the
eigen frequencies) and eigen vectors are subtracted from the
corresponding experimental results. The residual vector in this
case becomes:
lTi - li
Dz0 =
, i = 1,...,n
xTi - x i 0

(5)

where:
lTi , l i = test/analysis vectors of eigen values
xTi, xi = test/analysis mode shape vectors
The correlation between analytical data and test data is accomplished by means of the MAC value of the eigen vectors:
MAC: =

(x x )
x
( x ) (x x )
T
T

T
T

(6)

which states the linear dependency of two vectors xT, x. A MAC

value of 1 denotes that two vectors are collinear; a MAC value
of 0 indicates that two vectors are orthogonal.
The sensitivity matrix for the residual vector introduced in
Equation 5 is given by Equation 7. The calculation of the partial derivatives can be found in References 4 and 5.
l i
p
, i = 1,...,n
G0 =
(7)
x i
p

If real eigen values and eigen vectors are employed, the adjustment of damping parameters is not possible. The corresponding sensitivities equal zero, since the real eigen values
and eigen vectors depend solely on the stiffness and mass parameters of the system.

Model Validation Strategy

The model validation is accomplished through computational model updating of physical parameters (stiffness and
inertia parameters) of the finite-element model by minimizing
the deviations between the identified and the analytical eigen
values and mode shapes. All deviations between test and analysis are presumed to be exclusively based on uncertainties of
the finite-element model. An integral part of the validation
strategy is keeping the inevitable uncertainties from the test
side as small as possible and to create a reliable data base for
the following validation tasks by thorough test planning and
test execution. Figure 1 shows the principle proceeding.
The test planning utilizes the given finite-element model,
which enables not only the test design but also considerably
simplifies the later correlation with the analytical results (finite-element model and test model match). Test planning
20

Figure 2. Finite-element model of the body in white.

should cover the following aspects:

y Selecting relevant target modes
y Selecting measurement degrees of freedom with respect to:
- Essential test information
- Sufficient spatial resolution of target modes
- Coincidence of measurement and finite-element nodes
- Accessibility of measurement nodes
- Redundancy of the measurement degrees of freedom
- Robustness of the test model
y Selecting exciter positions (if possible, simultaneous excitation of all target modes)
y Sufficient frequency resolution (for following identification
methods)
Test planning and computational model updating was conducted using ICS.sysval. Necessary parameter changes are directly applied to the so called bulk data section of the
MSC.Nastran input file. Typical parameters are shell thicknesses, beam section properties, Youngs moduli, and densities. However, virtually all physical parameters, which can be
considered in an eigen value and eigen vector sensitivity analysis by MSC.Nastran, can be used for model updating.
After successfully updating the stiffness and inertia properties (physical parameters), modal damping parameters (modal
parameters) can be adjusted subsequently by minimizing the
deviations in the resonance regions between measured and
simulated frequency response functions.8
A common difficulty in computational model updating is the
proper choice of appropriate model parameters. Automated
methods may be applied7 in addition to selection based on engineering experience, but currently do not deliver reliable predictions. Another possibility is to select parameters based on
a sensitivity analysis. Here the sensitivity matrix according to
Equation 7 is computed for a set of suitable parameters. In a
subsequent investigation, those parameters that have a significant influence on results are identified. However, the sensitivity analysis does not supply any information on the physical
relevance of a particular parameter but merely detects its potential to change analysis results.

Example Car Body in White

The validation strategy that has been discussed will be demonstrated by means of the car body in white depicted in Figure 2. This is currently investigated under the auspices of the
work group 6.1.19 structure optimization and acoustics of the
German car industry.
The finite-element model consists of about:
y 142,000 nodes
y 130,000 elements
y 3,500 spot weld elements

SOUND AND VIBRATION/SEPTEMBER 2005

Figure 3. Test model of the body in white with measurement degrees of

freedom.

Figure 5. Mode indicator values of the body in white.

Figure 6. Test model of the body in white with exciter degrees of freedom.
Figure 4. Auto MAC matrix of analytical mode shapes at measurement
degrees of freedom.

The spot welds are modeled by means of MSC.Nastran

CWELD elements. These elements are currently used by some
automotive companies. The goal of the model validation is to
correctly predict the structural dynamics of the body in white
up to a frequency of about 100 Hz.
As already noted, test planning provides an important basis
for all subsequent investigations. It is especially important that
all required information be collected during the tests, which
is necessary for the following validation steps.
For the analyzed body in white, the following planning steps
are performed:
Select Target Mode Shapes. First, the boundary conditions,
frequency range and the relevant mode shapes are determined.
The body in white will be investigated in a free/free configuration that can be relatively easily realized in the test setup via
bungee cords or air springs. Since the goal of the model validation is to predict the structural dynamics of the body in white
up to a frequency of about 100 Hz, all analytic mode shapes in
this frequency range are to be considered in test planning.
Select Measurement Degrees of Freedom. Since a reliable
orientation of the accelerometers on the body in white is difficult due to the curvature of the car body, only measurement
degrees of freedom normal to the sheet surfaces are considered
in the test design. This provides the possibility to perform a
roving-hammer structural excitation. Of course, selection of the
measurement degrees of freedom should also ensure the unique
classification of individual mode shapes.
All measurement considerations noted under Model Validation Strategy should be taken into account. These include: essential test information; sufficient spatial resolution of the tar-

SOUND AND VIBRATION/SEPTEMBER 2005

get modes (preferably diagonal shape of the auto MAC matrix);

the coincidence of measurement and finite-element nodes (vital for correlation); redundancy of the measurement degrees of
freedom (for uncertainties in some measurement degrees of
freedom); and the robustness of the test model (due to uncertainties in the finite-element model used for test planning). 9
The final test model resulting from test planning is shown
in Figure 3, and the corresponding auto MAC matrix of the
analytical mode shapes at the measurement degrees of freedom
is depicted in Figure 4. The overall spatial resolution is sufficient. Only four mode shapes (No. 24 and 27 as well as No. 32
and 35) exhibit off diagonal coupling larger than 60%. These
are relatively local mode shapes that require a considerably
higher resolution. But to keep the measurement efforts manageable, a higher resolution of the measurement mesh is not
applied.
Select Exciter Degrees of Freedom. Identification of the exciter degrees of freedom is done in two steps. First, a preliminary subset of appropriate exciter degrees of freedom is calculated using a special automatic method.9 Next, the final exciter
degrees of freedom are defined by means of mode indicator
values. For a given exciter position, a mode indicator value
of 0 at a certain eigen frequency states that the corresponding
mode shape can be fully excited (satisfaction of the so called
phase-resonance criteria). In contrast, a value of 1 means that
the corresponding mode shape cannot be excited at the chosen exciter position.
The strategy is to select several exciter positions so that each
mode shape can be sufficiently excited at least at one exciter
position. The accessibility of the exciter position for a modal
exciter test is also taken into account. Figure 5 shows the mode
indicator values for the first 20 target modes at the four selected
references according to Figure 6. Obviously, each mode shape

21

can be excited at least by one of the defined exciter positions.

Select Frequency Resolution. The applied techniques to
identify the experimental modal data require a sufficient frequency resolution, especially in the lower frequency range.
Therefore, a minimal frequency resolution based on the first
elastic eigen frequency and the expected modal damping is
determined. Since the real damping behavior is usually not
known in advance, these results are to be verified before the
final test execution.

Test Setup and Preliminary Investigations

For approximating the free/free boundary conditions of the
finite-element model, the car body is mounted on four air
springs as shown in Figure 7. If necessary, the suspension can
be accommodated in the finite-element model by means of
springs. The spring properties can be estimated from the experimental rigid-body modes.
Different preliminary studies with impact-hammer and
modal exciters are performed to investigate the real behavior
of the body in white. Specifically, the linearity is reviewed by
means of modal exciter tests with different levels of excitation
and by coherence and reciprocity examinations. Altogether, the
investigated car body shows a sufficiently linear behavior.
Since test results with the impact hammer exhibit a very good
consistency with the results of the modal exciter tests, the final test was executed as a roving-hammer test with fixed-reference accelerometers. Proceeding this way also avoids mass
loading caused by moving accelerometer masses. Mass loading effects easily arise at large surface areas and can generate
frequency shifts of the resonance peaks. In regions with large
modal density, these effects produce unreliable test results.
Since the measurement degrees of freedom are manually
mapped onto the car body, an uncertainty regarding the real
position of the measurement points exists. For quantifying
possible deviations, the measurement points are digitized usTable 1. Initial correlation.
No. EMA1 FEA f, % MAC, % No. EMA1 FEA f
1
7
1
7.04
98.24
2
8
2
0.19
87.42
3
9
3
6.17
63.55
4
13
5
0.83
73.20
5
11 11.72 61.13
6
6
14 3.39
8
79.49
7
15 6.55
10
66.51
1) without rigid body modes

8
9
10
11
12
13
14

12
13
14
15
17
18
19

18
19
20
22
24
26
27

3.07
0.02
1.85
1.39
0.07
1.07
1.49

MAC, %
66.41
72.49
86.89
78.18
6788
96.49
76.60

Table 3. Correlation after computational model updating.

No. EMA1 FEA f, % MAC, % No. EMA1 FEA f
1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
9

7
8
9
10
11
12
13
14
15

4.35
0.75
1.49
1.10
0.50
1.83
2.67
1.26
1.55

98.69
97.54
95.49
94.38
93.86
65.05
90.49
95.39
80.52

10
11
12
13
14
15
16
17

10
11
12
13
14
15
17
18

16 0.21
17 0.09
18 0.90
19 0.15
20 1.35
22 1.49
24 0.85
25 0.71

MAC, %
82.22
69.57
85.80
75.85
92.13
63.63
85.47
98.09

1) without rigid body modes

Table 2. Correlation after remodeling.
No. EMA1 FEA f, % MAC, % No. EMA1 FEA f
1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
10

7
8
9
10
11
12
13
15
14

4.51
3.01
1.49
0.16
1.26
0.65
0.71
0.22
6.35

98.62
97.51
96.22
90.12
90.92
68.55
59.68
94.88
52.61

1) without rigid body modes

22

10
11
12
13
14
15
16
17

11
12
13
14
15
16
17
18

17
18
19
20
21
22
24
25

0.14
0.57
0.19
1.31
0.83
0.18
0.48
0.98

MAC, %
68.37
88.51
72.25
89.43
65.52
66.29
86.01
98.19

Figure 7. Mounting of the body in white using air springs.

ing supersonic triangulation. Acquired real measurement point

positions agreed well with the finite-element model.

Experimental Modal Analysis

The analysis of the acquired test data is conducted by two
completely different identification algorithms, one working in
the time domain (polyreference) and one working in the frequency domain (direct parameter estimation). The identified
sets of modal data are subsequently correlated, allowing for an
assessment of the individual quality of the measured eigen frequencies, mode shapes, etc. This is of particular importance
for subsequent model validation, since only experimental results with sufficient quality should be used.

Initial Correlation
For evaluating model quality, frequency deviations between
test and analysis as well as the MAC values of the corresponding mode shapes according to Equation 6 are employed. Besides
the correlation table, the MAC matrix will also be posted for
assessing the quality of the correlation.
The initial correlation for the investigated body in white is
given by Table 1 and Figure 8 and consider only mode shapes
with a frequency deviation less than 30% and a MAC value
larger than 50%. Usually, a tolerance limit of 70% is acceptable to ensure a consistent correlation of the mode shapes used
for updating. In this case, however, a decrease is necessary for
considering all relevant mode shapes in the initial correlation
and especially in subsequent computational model updating.
Even by reducing the tolerance limit to 50%, not all mode
shapes can be correlated (gaps on the main diagonal of the
MAC matrix). Moreover, some couples show frequency deviations larger then 10%.

Sensitivity Analysis
For reducing the multitude of potential parameters to a subset of parameters, which have a significant influence on the
model behavior, a sensitivity analysis is performed by computing the eigen value and eigen vector sensitivities for all potential parameters. Subsequently, the most promising parameters
for each mode shape are determined using the sensitivity module of ICS.sysval. These parameters constitute the basis for the
following computational model updating.

Remodeling and Computational Model Updating

In the investigated example, dividing the updating task into
two individual steps proves very efficient. In both cases
Youngs moduli are used for updating (for preserving the total
mass of the car body).
In the first step, areas with large, physically not interpret-

SOUND AND VIBRATION/SEPTEMBER 2005

Figure 8. MAC matrix, initial correlation.

able, parameter changes are identified by applying computational model updating methods. These areas are then remodeled using the CAD data of the geometry, which already leads
to an improvement of model quality. The results after remodeling are collected in Table 2 and Figure 9, respectively.
The remodeling already yields a significant increase of model
quality. On the one hand, the MAC values can be increased to
more than 90%, especially in the lower frequency range. On
the other hand, frequency deviations are noticeably reduced.
However, the correlation in the range of the sixth, seventh,
ninth, and tenth measured mode is still not fully satisfactory.
For further increasing the model quality, multiple updating
runs are performed in a second step using the remodeled finite-element model. The final results are summarized in Table
3 and Figure 10.
Comparing these results with Table 2 and Figure 9 shows an
additional increase in the model quality achieved by subsequent computational model updating. Both the mentioned gaps
on the main diagonal of the MAC matrix can be filled, and the
frequency deviations can be further reduced. Besides the first
eigen frequency, all frequency deviations are now below 3%.
However, note that the total deviation for the first eigen frequency is less than 1 Hz.

Summary
We examined validation of the finite-element model of a
body in white using a special software package for computational model updating. It enables direct updating of large-scale
MSC.Nastran finite-element models.
For keeping the inevitable uncertainties from the test side
as small as possible, thorough test planning is essential. By
using the initial finite-element model, both measurement degrees of freedom and exciter positions are virtually defined and
then transferred onto the body in white. For avoiding mass
loading effects, data acquisition is performed using a rovinghammer excitation with fixed accelerometer positions.
Besides the direct increase of model quality, the computational model updating technique provides the additional opportunity to identify regions where remodeling already yields
an improvement of model quality. Test/analysis correlation before and after computational model updating as well as remodeling exhibits a noticeable reduction of frequency deviations
along with an increase of MAC values over a broad frequency
range.
These investigations are currently expanded towards attachment parts like doors and subassemblies under the auspices of
the work group 6.1.19 structure optimization and acoustics
of the German car industry. A first goal is to sufficiently represent the structural dynamics as a thorough basis for subsequent
investigations on acoustic phenomena.

SOUND AND VIBRATION/SEPTEMBER 2005

Figure 9. MAC matrix after remodeling.

Figure 10. MAC matrix after computational model updating.

Acknowledgments
This paper is dedicated to Mr. Jrn Frappier, who died in a
tragic accident at the beginning of 2004.

References
1. Allemang, R. J., Vibrations: Experimental Modal Analysis, Structural Dynamics Research Laboratory, University of Cincinnati, UCSDRL-CN-20-263-663/664, Cincinnati, OH, 1995.
2. Ewins, D. J., Modal Testing: Theory And Practice, Research Studies
Press Ltd., Taunton, Somerset, England, 1995.
3. Link, M., and Hanke, G., Model Quality Assessment and Model
Updating, NATO Advanced Study Institute, Modal Analysis & Testing, Sesimbra, Portugal, 1998.
4. Link, M., Updating of Analytical Models Review of Numerical Procedures and Application Aspects, Structural Dynamics Forum SD
2000, Los Alamos, NM, April 1999.
5. Natke, H. G., Einfhrung in die Theorie und Praxis der Zeitreihenund Modalanalyse, 3rd, revised edition, Vieweg Verlag, Braunschweig, Wiesbaden, 1992.
6. Schedlinski, C, Information about ICS.sysval Software, www.icssolutions.de, ICS Langen, 2004.
7. Lallement, G., Localisation Techniques, Proc. of Workshop Structural Safety Evaluation Based on System Identification Approaches,
Braunschweig/Wiesbaden, Vieweg, 1988.
8. Schedlinski, C., Computational Model Updating of Large-Scale Finite-Element Models, Proc. of the 18th International Modal Analysis Conference, IMAC, San Antonio, TX, 2000.
9. Schedlinski, C., An Approach to Optimal Pick-up and Exciter Placement, Proc. of the 14th International Modal Analysis Conference,
IMAC, Dearborn, MI, 1996.
The author can be contacted at: sched@ics-solutions.de.

23