Measurement 102 (2017) 121–130

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Comparison of experimental and operational modal analysis on a

laboratory test plate
Esben Orlowitz ⇑, Anders Brandt
Department of Technology and Innovation, University of Southern Denmark, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: Operational modal analysis (OMA) is widely used whenever the dynamic characteristics of structures that
Received 10 November 2016 do not fit into a laboratory are desired. In addition, OMA offers a test of the structure under its real bound-
Received in revised form 27 January 2017 ary conditions which may sometimes be preferable for validation of numerical models. Theoretically, the
Accepted 1 February 2017
natural frequencies and damping ratios should be identically estimated by an OMA test and an experi-
Available online 4 February 2017
mental modal analysis (EMA) test. However it is still often reported that EMA tests are more reliable.
The present paper presents a thorough comparison of EMA and OMA tests of a Plexiglas plate. The exper-
Keywords:
iments were carefully designed, to ensure that the plate was tested under similar boundary conditions.
Operational modal analysis
Experimental modal analysis
Estimated modal parameters from the EMA test and OMA test are presented and compared, for the first
Damping estimation ten modes of the plate. It is found that natural frequencies are deviating by less than 0.3%, damping ratios
Experimental test by less than 7%, whereas cross-MAC values between the mode shapes of the two tests are found to be
above 0.99. The experimental test was conducted first by an EMA test, followed by the OMA test and
finally another EMA test was conducted in order to catch any time-variance. It is concluded that no sig-
nificant differences were found between modal parameters obtained by OMA and EMA.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction all inputs (excitation forces) are measured, which is practically

As simulation of vibration responses based on numerical mod-
els are increasingly used to investigate dynamic behavior of struc-
tures, experiments are increasingly required for validation or
correction of the numerical models. As an example, a dynamic
model of a structure with incorrect damping assumptions could
give misleading life time estimations [1]. Because of the complex-
ity of damping effects, it is usually not possible to give a proper
analytical estimate, hence experimental tests are needed. How-
ever, damping can be a notoriously difficult parameter to consis-
tently estimate from experiments, and the accuracy of estimates
of damping from, particularly, operational modal analysis (OMA)
is currently receiving a lot of research interest.
Experimental modal analysis (EMA) has been used for decades
to extract information about the structural damping. EMA is con-
sidered reliable because it is based on input-output system identi-
fication, which allows validation e.g. of the estimated frequency
response functions (FRFs) by coherence functions. However, the
strength of EMA is also limiting its applicability as it requires that
⇑ Corresponding author.
E-mail address: esbenorlowitz@gmail.com (E. Orlowitz).
unfeasible for many structures.
The last couple of decades, OMA, by which only the structural
responses are used, has been widely applied and described in liter-
ature. OMA is attractive in many situations because it can be
applied to structures in operation, and does not require excitation,
which is practical for many large structures, see for example
[25,19,24,4,27]. In addition, OMA can sometimes be preferable
for validation of numerical models because the effects of boundary
conditions are included in the results, as opposed to EMA where
the structure is usually tested under free-free conditions. OMA
methods are based on some assumptions about the nature of the
loads exciting the structure; the loads being the result of random
processes, that may be colored as long as the poles of the force col-
oring are different from those of the structure. In addition the
structure should be linear and time-invariant [2–4].
Theoretically, modal parameters should be identically esti-
mated via an OMA test and a classical EMA test [5]. However, it
has still sometimes been reported that an EMA test is more reliable
because of the available information and controlled environment
[6]. Furthermore it has also been reported that the damping esti-
mates from OMA are over estimated, hence suggesting a critical
bias error [7,8].

http://dx.doi.org/10.1016/j.measurement.2017.02.001
0263-2241/Ó 2017 Elsevier Ltd. All rights reserved.
122 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130
Little comparison between EMA and OMA has been published
where exactly the same structure and experimental setup have
been used for both cases. In [9] however, a comparison of EMA
and OMA was presented for some case studies of structures with
simple to more complex geometries and it was shown that for sim-
ple geometries the differences of the extracted modes were small.
However, difficulties to extract all modes from OMA tests were
reported and in addition it was reported that choosing the loca-
tions of the excitations in an OMA laboratory test is non-trivial.
For a simple geometry with random spatial excitation, problems
with finding all modes from the OMA test were reported, and it
was found that the mode shapes from the two tests were not iden-
tical. The problem of undiscovered modes increased when the exci-
tation was localized to a single point for the OMA test.
For a more complex structure also investigated in [9], a consid-
erable number of modes were not well estimated or not estimated
at all. Damping ratios were not reported and mode shapes were in
general showing poor similarity between the EMA and OMA tests.
However, it should be noted that only a single reference DOF was
used for the calculation of spectra and the measurement time
was not reported. Both the number of reference channels and the
measurement time are crucial for OMA [6,10].
In [11] a comparison of EMA (using impact testing) and OMA
was carried out, which showed large deviations in the estimated
damping ratios. However, the methods used for parameter estima-
tion for OMA and EMA were not the same and thus the applied set-
tings for estimation are not comparable. The authors of [11]
pointed out that the difference in vibration levels between the
two tests could be the reason for the large deviations of the damp-
ing ratios.
In [8] a comparison of EMA and OMA was performed on the
same data set, acquired from excitation with two shakers and rov-
ing accelerometers over a test structure. For OMA the measured
input was simply ignored, using only the responses. It was
observed that the damping ratios estimated by OMA were higher
than the ones from EMA and the same was observed for the natural
frequencies. However, in this study the dynamical characteristics
of the shakers seems too have been included in the OMA, but not
in the EMA, results. For EMA the influence of the shakers is
excluded by using frequency responses relative to the inputs, but
this is not the case for OMA. Thus the boundary conditions were
very different in the OMA test compared to the EMA test, which
explains the higher damping in the OMA estimates.
It is worth noticing that one of the attractive features of OMA is
that it indeed includes the boundary conditions. This means, how-
ever, that in order to compare EMA and OMA, great care must be
taken in the design of the experiment to ensure that the boundary
conditions are identical in both tests.
The present paper presents an experimental study of EMA and
OMA tests on a Plexiglas plate where the EMA and OMA tests have
been carefully designed so that they test the structure under very
similar boundary conditions. The plate, which is a simple geome-
try, is designed to have closely spaced modes and is considered a
good structure for comparison. Modal parameters are estimated
and compared between the EMA and OMA tests.
Damping ratio estimates are the primary focus in the present
work as they are known to be the most challenging, although both
natural frequencies and modes shapes are also compared for
completeness.
2. Experimental setup
The Plexiglass (PMMA) plate that has been experimentally
tested can be seen in Fig. 1. The dimensions of the plate are
533 mm
321 mm
20 mm. The plate is similar to the so-called
IES-Plate proposed in [12], although, for practical reasons, the plate
thickness was chosen slightly different from the IES-Plate, by
choosing 20 mm thickness, which is a standard thickness in
Europe.
The measurement grid consisted of 35 out-of-plane DOFs dis-
tributed uniformly over the plate as shown in Fig. 2 and responses
from all DOFs were measured simultaneously. The accelerometers
and cables significantly mass load the structure, and it is also likely
that there is additional damping added. This is, however, irrele-
vant, as the present experiments are designed so that it is the plate
together with the instrumentation that is investigated. The condi-
tions for the EMA and OMA tests should therefore be equal, albeit
not being those of the free-free plate.
The following data acquisition (DAQ) and sensor equipment
were used:
 3 National Instrument 4497, 16 channel, 24bit analog inputs
cards.
 35 Dytran 3097A2 accelerometers, 100 mV/g, IEPE, 4.3 grams.
 1 Dytran 5800B4 impulse hammer, 10 mV/N, IEPE, with plastic
tip.
 In-house software for DAQ control, based on MATLAB DAQ-
toolbox.
The plate was suspended by springs via a thin fishing line, see
Fig. 1, giving it relatively low rigid body modes (<3 Hz). For all tests
a sampling frequency of 5 kHz was used.
For the EMA test, frequency response functions (FRFs) were
measured by impact testing. FRFs relative to two different excita-
tion DOFs were measured, in order to have multi-reference data
for the parameter extraction. DOFs 1 and 29 were thus excited,
and time sequences of 20 impacts with 20 s in-between each
impact were acquired for each reference point.
For the OMA test, 300 s of data were acquired from all DOFs
simultaneously, corresponding to approx. 40000 periods of the
lowest natural frequency. This rather long measurement time
was chosen to reduce possible effects on modal parameters by lim-
ited measurement time, see e.g. [10,4]. The excitation was applied
by gently tapping the tip of a pencil randomly around the plate.
The idea of tapping rather than scratching was chosen in order to
minimize the interaction with the plate as much as possible, as it
was initially experienced that scratching the plate influenced its
dynamics in a significant way, both on natural frequencies and
damping ratios.
Two impact tests were performed; one test was performed
before and one after the OMA test, in order to be able to detect
any time-variance. The impact test before the OMA test is referred
to as EMA1 and the one after is referred to as EMA2 in the following.
All three tests were conducted consecutively within approximately
one hour, under which no interaction with the experimental setup
(except the excitation) was allowed. Between the end of the EMA1
test and the start of the OMA test, 15 min elapsed, and from the
end of the OMA test to the start of EMA2, 5 min elapsed, all due
to initial quality checks. Each of the impact tests took approxi-
mately 15 min from start to end.
3. Data processing and analysis
3.1. Impact tests
The acquired time sequences from the impact test were pro-
cessed to obtain FRFs. FRFs and coherence functions were esti-
mated by averaging the spectra from two selected impacts of the
time sequences (the input and all the responses) using the H1 esti-
mator in the MATLAB toolbox ABRAVIBE [13]. The two impacts
 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130
Fig. 1. Experimental setup. Plexiglas plate suspended vertically in two springs with 35 accelerometers mounted. In the lower right corner the impact hammer and pencil used
for excitation in the EMA and OMA measurements, respectively, can be seen.
Fig. 2. Measurement grid on the plate consisting of 35 measurement points with a
separation of 75
85 mm. The full dimensions of the plate are 533
321 mm and
thickness 20 mm.
producing a FRF estimate with optimum coherence were selected
by an automatic process as described in [14].
The blocksize for the estimation of FRFs was 64 k. This blocksize
was selected by a comparison of FRF estimates using increasing
blocksizes, and selecting the FRF estimate for which a further
increase in blocksize does not have higher peaks of the FRFs. This
procedure ensures that bias errors in the FRF estimates are negligi-
ble. For noise suppression, a force window of 1% of the blocksize
was applied to the force signal and an exponential window
decreasing to 0.001% at its end was applied to all signals. The
effects on the damping estimates due to the exponential window
were removed after the parameter extraction as described in [14].
The peak force levels of the force impacts were approximately
15 N. In Fig. 3, a comparison of two FRFs from two different pairs
of impacts of different force levels are shown. The similarities
between the FRFs indicate that linearity can be assumed within
the current force range.
Linearity is a fundamental assumption for both EMA and OMA
and as the excitations of the two tests are different in nature,
any amplitude dependent damping would make the comparison
of EMA and OMA invalid.
As stated earlier, two impact tests were conducted, one before
the OMA test and one immediately after; the EMA1 test and
123
EMA2 test, respectively. For EMA1 the FRFs and coherence func-
tions between all response locations and the references (70 in
total) are shown in Fig. 4 and similar results were found for EMA2.
As can be seen in Fig. 4(b), the coherence functions are very
close to unity at frequencies above, say, 100 Hz, which is well
below the first natural frequency. Above this frequency, no signif-
icant dips in the coherence functions are present.
Reciprocity plots for EMA1 are shown in Fig. 5 and are also indi-
cating good measurements; the slight differences that can be
observed may be caused by impacts chosen not being exactly in
the point where the accelerometer in the excitation points are
located, or slight sensitivity differences between the accelerom-
eters used. Neither of these causes of the differences will affect the
pole estimates, but only the mode shapes, and the reciprocity is
therefore acceptable in the present work where the primary
parameter of interest is damping. A similar examination of the
reciprocity of the EMA2 test showed similar results. The reciprocity
plot is one of the validation tools that are available when input and
output data are measured in EMA applications.
3.2. OMA test
For the OMA test, correlation functions (CFs) were estimated as
substitutes for impulse responses. A requirement for estimating
CFs is that the responses are stationary. The reverse arrangements
test was therefore applied on the rms-levels of frames of the
responses, to test for (wide sense) stationarity [15,14]. A signifi-
cance level of 2% and 100 frames were used as recommended in
[16], and all responses passed the reverse arrangements test, indi-
cating that the signals are likely to be stationary.
The rms-levels of the responses ranged from 113 to 339 mm/s2
with a median of 207 mm/s2. Maximum peak values for the
responses ranged from 6 to 26 m/s2 which is significantly lower
than the first part of the transient responses from the impact tests,
which ranged between 50 and 150 m/s2. It is not possible to com-
pare the transient decaying responses of the impact tests directly
with the continuous responses from the OMA test, but as linearity
has already been verified, as described above, this should not be an
issue.
In Fig. 6 two typical power spectral densities (PSDs) are shown
from which it can be observed that the dynamic range of the mea-
surements is sufficient.
124 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130

2
10

FRF [(m/s )/N]

1
10

2
0
10

-1
10
0 200 400 600 800 1000
Force spec. [Ns]

10 0

10 -2

10 -4
0 200 400 600 800 1000

Frequency [Hz]

Fig. 3. Comparison of frequency response function (upper) from different force levels, see the force spectra (lower).
FRF [(m/s 2 )/N]

0
10

0 200 400 600 800 1000

Frequency [Hz]
(a) Frequency response functions
1
Coherence

0.5

0
0 200 400 600 800 1000
Frequency [Hz]
(b) Coherence functions
Fig. 4. Frequency response functions (a) and coherence functions (b) from EMA1, the first impact test. In total 70 functions are plotted in both graphs, which shows that the
coherence functions are very good above 100 Hz, which is well below the first natural frequency. Thus the FRF estimates are very good in the frequency range of interest.

The CFs were computed as unbiased estimates by the peri-
odogram approach, see [17]. Thus, the discrete Fourier transforms
of zero padded sequences are computed, to avoid cyclic effects of
the DFT. To obtain, for example, the cross-correlation between
two responses y1 and y2 , with the first signal as the reference,
the DFT of y2 is multiplied by the complex conjugate of the DFT
of y1 in the frequency domain, whereafter the inverse DFT of this
product produces the biased correlation function. The unbiased
estimate is then formed by proper weighting so that the correla-
tion function between signal y1 ðtÞ and y2 ðtÞ becomes
X
Dt Lm1
R21 ðmÞ ¼ y ðnÞy1 ðn  mÞ
L  m n¼0 2
where m is the discrete lags corresponding to time m=f s , and L is the
length of the time sequences xðnÞ and yðnÞ sampled by time incre-
ment Dt. Finally, the negative lags of the correlation functions are
discarded.
3.3. Parameter extraction
The estimated impulse response functions and correlation func-
tions were processed by the multi-reference Ibrahim time domain
(MITD) method for modal parameter extraction [18]. The MITD
method is in principle equivalent to the more well known stochas-
ð1Þ tic subspace identification (SSI) method [19]. In the MITD method,
a block Hankel matrix of impulse responses or correlation func-
 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130 125

2
10

1
10

FRF [(m/s )/N]

0
2 10

-1
10

-2
10
In=1, Out=29
In=29, Out=1
-3
10
0 200 400 600 800 1000
Frequency [Hz]
Fig. 5. Reciprocity plot for the EMA1 test.

10 -3

DOF 1-1
Auto and Cross PSD [(m/s ) /Hz]

DOF 1-29
-4
10
2 2

10 -5

10 -6

-7
10

10 -8
0 200 400 600 800 1000
Frequency [Hz]
Fig. 6. Typical (auto and cross) power spectral densities for DOF 1–1 and 1–29. A blocksize of 2048 (sampling frequency 5k), Hanning window and an overlap of 50% has been
used, resulting in 1463 averages.

tions is constructed, which is decomposed by singular value
decomposition (SVD). From the SVD compressed block Hankel
matrix, an eigenvalue problem can be established for different sys-
tem order (number of poles), depending on the number of singular
values used. A detailed description of the method can be found in
[18].
The modal parameter estimates (poles and eigenvectors) from
different levels of compression by the SVD, depending on the sys-
tem order, were plotted in a stabilization diagram, sometimes
referred to as a consistency diagram, as a tool to help finding the
reliable physical modes. The natural frequencies and modal damp-
ing ratios were then calculated from the stable poles selected from
the stabilization diagram. The mode shapes were calculated from
the eigenvectors corresponding to the selected poles by expansion
via the left singular vectors from the SVD compression, producing
unscaled mode shapes.
The FRFs from the impact tests were converted to impulse
response functions (IRFs) before parameter estimation by the MITD
method. The first 5 samples of the IRFs were neglected as these
samples are contaminated by noise on the responses [20]. The mul-
tivariate mode indicator function [21] of the EMA1 data are shown
in Fig. 7 together with the stabilization diagram from the MITD
method showing the estimated poles at increasing model orders
up to 100, a similar stabilization diagram was obtained for EMA2.
As can be observed from the mode indicator function there are
two closely spaced modes around 140 Hz, which is the reason for
using two references in the modal analysis.
For both EMA and OMA, 200 samples of the impulse response/-
correlation functions were used for the estimation of modes, corre-
sponding to 40 ms, or approximately 5.5 periods of the lowest
natural frequency of the plate. This amount was found to give
the best stabilization of the first ten estimated modes.
126 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130

80

70

60

Number of Poles
50

40

30

20

10

0
0 200 400 600 800 1000
Frequency [Hz]

Fig. 7. Stabilization diagram of impact test data from EMA1 with overlaid multivariate mode indicator function. Crosses and circles indicate stable and unstable poles,
respectively.

In Fig. 8 the stabilization diagram from the OMA test is shown
together with an averaged PSD function as a simple mode indicator
function.
4. Experimental results
Modal parameters were extracted for the first 10 modes
(<1000 Hz) by the method described in Section 3, in the following
manner. All modes fulfilling the stabilization criteria of deviations
of 0.1% for natural frequencies, 5% for damping ratios and having
modal assurance criterion (MAC) values exceeding 0.9 were
extracted from the stabilization diagrams. Also the mode shapes
of the selected modes were averaged. This resulted in at least 26
pole estimates of each mode from different model orders, which
were averaged and the corresponding scatter calculated (calcu-
lated as the normalized random error er , i.e., the standard deviation
divided by the average value). The resulting mean natural frequen-
cies for all three experiments are presented in Table 1 and, corre-
spondingly, damping ratios are presented in Table 2. The auto-
MAC matrix of the averaged mode shapes are shown in Fig. 9 for
the impact test EMA1 and the OMA test; a similar auto-MAC matrix
was obtained for EMA2 but is not shown.
From Table 1 it can be observed that the natural frequencies
decrease slightly between the first test (EMA1) and the last
(EMA2). A comparison of the differences of the results of the impact
tests with respect to the OMA test are presented in Table 3 and
graphically presented in Fig. 10. It can be observed that the rate
of decrease in natural frequencies is smaller between the last
two tests (OMA and EMA2) than between the first two tests
(EMA1 and OMA).
This could suggest that the decrease is due to some environ-
mental change in the surroundings, possibly temperature increas-
ing in the laboratory during the measurements, causing the
stiffness of the plate to decrease. It should also be noted, however,

80

70

60
Number of Poles

50

40

30

20

10

0
0 200 400 600 800 1000

Frequency [Hz]

Fig. 8. Stabilization diagram from the OMA test. Crosses and circles indicate stable and unstable modes, respectively, and the background curve is the power spectral density
of the response of DOF 1.
 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130 127

Table 1
Estimated (averaged) natural frequencies, f n , and corresponding normalized random errors, er , for all three experimental tests. The random errors are calculated based on at least
26 extracted modes.

EMA1 OMA EMA2

f n [Hz] er ðf n Þ [%] f n [Hz] er ðf n Þ [%] f n [Hz] er ðf n Þ [%]

Mode 1 142.0 0.008 141.9 0.019 141.8 0.015
Mode 2 142.8 0.004 142.6 0.011 142.5 0.003
Mode 3 323.6 0.055 323.3 0.005 323.1 0.031
Mode 4 393.4 0.007 393.0 0.001 392.7 0.009
Mode 5 409.1 0.008 408.8 0.003 408.5 0.002
Mode 6 504.5 0.021 503.8 0.005 503.7 0.017
Mode 7 584.2 0.038 583.6 0.007 583.3 0.040
Mode 8 713.1 0.010 711.3 0.003 712.2 0.004
Mode 9 800.9 0.004 798.8 0.002 799.9 0.008
Mode 10 941.6 0.014 940.3 0.010 940.5 0.032

Table 2
Estimated (averaged) damping ratios, fn , and corresponding normalized random errors, er , for all three experimental tests. The random errors are calculated based on at least 26
extracted modes.

EMA1 OMA EMA2

fn er ðfn Þ [%] fn er ðfn Þ [%] fn er ðfn Þ [%]
Mode 1 3.28 0.242 3.25 0.529 3.34 2.448
Mode 2 3.09 0.279 3.06 0.297 3.13 0.370
Mode 3 2.68 2.114 2.68 0.206 2.70 1.134
Mode 4 2.56 0.305 2.54 0.151 2.61 0.632
Mode 5 2.47 0.443 2.45 0.252 2.50 0.322
Mode 6 2.48 0.420 2.42 0.277 2.52 0.290
Mode 7 2.46 2.210 2.38 0.441 2.53 2.522
Mode 8 2.38 0.560 2.41 0.085 2.42 0.578
Mode 9 2.27 0.367 2.28 0.373 2.33 0.442
Mode 10 2.44 0.793 2.57 0.544 2.53 0.841

1 1

1 0.8 1 0.8

0.5 0.6 0.5 0.6

0.4 0.4
0 0
12 12
34 0.2 34 0.2
10 10
56
7 89 56
7 89
78 56 78 56
9 34 9 34
10 12 10 12
Mode # Mode #

(a) EMA1 (b) OMA

Fig. 9. Auto MAC matrices from the averaged mode shapes of the EMA1 and OMA tests. Similar results were obtained for EMA2.

Table 3
Difference of the results of the impact tests (EMA1 and EMA2) with respect to the OMA test.

Difference of: EMA1 wrt. OMA EMA2 wrt. OMA

f n [%] fn [%] Cross-MAC f n [%] fn [%] Cross-MAC

Mode 1 0.101 1.089 0.999 0.059 2.903 0.999

Mode 2 0.136 0.847 0.999 0.043 2.176 0.999
Mode 3 0.119 0.058 1.000 0.047 0.696 1.000
Mode 4 0.079 0.708 0.997 0.082 2.742 0.997
Mode 5 0.071 0.933 1.000 0.078 2.061 1.000
Mode 6 0.138 2.588 1.000 0.012 4.390 1.000
Mode 7 0.105 3.144 1.000 0.054 6.286 1.000
Mode 8 0.248 1.275 0.994 0.117 0.500 0.994
Mode 9 0.274 0.312 0.992 0.148 2.259 0.992
Mode 10 0.136 5.075 0.997 0.020 1.674 0.997

that the changes are very small; less than 0.3%. As stated in Sec- the smaller decrease of natural frequencies between the latter
tion 2, the time elapsed between EMA1 and OMA was larger than tests. A divergence from this can be observed for modes 8 and 9
the time between OMA and EMA2 which is in compliance with from the OMA test, see Table 1, that cannot be explained only by
128 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130

Frequency wrt. OMA

Relative deviation [%]

0.2

-0.2 EMA1
EMA2
-0.4
1 2 3 4 5 6 7 8 9 10
Mode #
Damping wrt. OMA
Relative deviation [%]

10

-10
1 2 3 4 5 6 7 8 9 10
Mode #

Fig. 10. Relative deviations of EMA1 and EMA2 results with respect to the OMA results.

the observed scatter. No obvious reason has been found for this
divergence, but all errors are small.
The damping estimates, see Table 3 and Fig. 10, show larger
deviations between the three tests than for the natural frequencies,
which is expected, since the inaccuracy of damping estimates are
well known to be larger than of frequency estimates. The results
of the two EMA tests are, however, in better agreement with each
other than with the OMA results, although the estimated damping
ratios are found consistently larger for EMA2 compared to EMA1.
No consistency is found in the comparison of the damping esti-
mates from the OMA test and the EMA tests and again from a prac-
tical point of view the observed errors are small.
The observed errors are influenced by the choice of evaluation
of the results in the present work, where all stabilized modes are
averaged together. For example, if the estimated damping is plot-
ted as a function of model order, as presented for modes 3 and
10 in Fig. 11, it can be observed that the averaging over increasing
model order influences the results in some cases, e.g. for mode 3 as
seen in Fig. 11(a). For mode 10, however, as shown in Fig. 11(b), the
observed deviation is more consistent as a function of model order,
but as stated earlier there is a difference in the OMA and EMA tests
due to the excitation which could influence the observed deviation
between tests.
In addition to the influence of the model order at which the
modes are estimated, as discussed above, other parameters in the
estimation process may also be of importance. To make the com-
parison between the OMA and EMA results as independent as pos-
sible of any such differences, the analysis was made in this study

0.029
Damping ratio [%]

0.028

0.027

0.026
10 20 30 40 50 60 70 80
Model order
(a) Mode3
0.026
Damping ratio [%]

0.025

0.024
40 50 60 70 80
Model order
(b) Mode10
Fig. 11. Damping ratio estimates depending on model order in the estimation for mode 3 and mode 10. Comparing the estimates from for OMA ‘+’ and EMA1 ‘’.
 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130
with the same parameter estimation algorithm, which should min-
imize any differences in parameter estimates due to uncertainties.
It should further be considered that the relatively long measure-
ment time used for the present OMA analysis, results in small ran-
dom errors in the correlation functions estimates, which in turn
results in less uncertainties in the modal parameter estimates.
The impulse responses used for the EMA parameter estimation
have similarly small random errors due to the high coherence func-
tion values. A thorough study of the influence of various parameter
estimation choices can be found in e.g. [6,26].
5. Discussion
From Table 2 it is clear that the differences in damping esti-
mates between the three tests are insignificant, which is in agree-
ment with the findings of [6,22]. From the tests in Tables 1 and 2, it
can also be observed that the random errors for both natural fre-
quencies and damping ratios are comparable between the OMA
test and the EMA tests.Thus, the results support the hypothesis
that damping ratios estimated by OMA are as reliable as those esti-
mated by EMA.
The possibility of different response levels and non-linear
damping could be a reason for the small deviations in the damping
estimates from the experimental EMA and OMA tests, but it was
shown in Section 3.1 that, at least within a given force range of
impacts, no indications of non-linear behavior was observed. This
strongly suggests that the small differences found here between
EMA and OMA damping estimates can be explained by the slight
differences in boundary conditions between the tests, as the tap-
ping of the plate during the OMA test is certainly not identical to
the impact used for the EMA test.
The setup used in the present work where impact testing was
used has the advantage that no interaction with the plate is done,
except the excitation. In further work acoustic excitation could
perhaps be an opportunity to decrease the interaction for OMA
excitation in the laboratory, see e.g. [23]. However, the best excita-
tion for OMA is a load distributed over the structure, which is elab-
orate to implement for acoustic excitation.
Finally, it is noted that the cross-MAC values in Table 3 show
that the mode shapes of the OMA and EMA tests, are very similar.
Thus, for the present experiments where the boundary conditions
were very similar between the tests, there is no indication that the
OMA and EMA methods would result in any different modal
parameters.
6. Conclusions
The present paper presents a comparison of an Experimental
Modal Analysis (EMA) test and an Operational Modal Analysis
(OMA) test on a Plexiglas plate. The experimental setup was care-
fully designed so that both tests would test the plate under very
similar boundary conditions, in order to make a proper comparison
of modal parameters extracted by EMA and OMA possible. The
EMA test applied in the present work was based on impact testing
and in the OMA test the excitation was applied by gently tapping
the tip of a pencil randomly around the plate.
Two impact tests were performed, one test was performed
before and one after the OMA test, to be able to detect any time-
variance. In addition, the three tests were conducted within a nar-
row time frame, to ensure no significant environmental changes
would occur. The two EMA tests, before and after the OMA test,
gave similar results with deviations between the modal parame-
ters obtained from the two tests of less than 0.2% for natural fre-
quencies, less than 4% for the damping ratios of the first ten
modes, and cross-MAC values between the mode shapes larger
than 0.98.
129
Comparison of the results from the OMA test with the EMA tests
showed overall very good agreement. The natural frequencies
found by the OMA test were consistently lower than the ones of
the first EMA test, with deviations less than 0.3%, and deviated
from the later EMA test with less than 0.2%. All cross-MAC values
between the OMA test and the EMA tests were above 0.99, thus
showing close similarity of the mode shapes. The estimated damp-
ing ratios between the EMA tests and the OMA test were less than
7% and typically around 2%. The estimated damping ratios from the
OMA test were both higher and lower than the EMA estimates,
hence no systematic error was observed.
The results of the present work strongly indicate that there is no
significant difference in estimated modal parameters obtained by
OMA and EMA provided that the conditions of the tests are compa-
rable. We would like to emphasize, however, that OMA tests in
most engineering applications are not comparable to typical EMA
tests. During a typical OMA test, the structure has different bound-
ary conditions than the typical free-free conditions of an EMA test.
Therefore, it can be expected that OMA results in many (or even
most) engineering applications will show higher damping values
than a free-free EMA test. But this result is not due to differences
between the signal processing and modal parameter extraction,
but to the actual boundary conditions of the tests.
References
[1] A. Linderholt, Y. Chen, E. Orlowitz, A. Brandt, A study of the coupling between
test data accuracy and life prediction, in: Proceedings of Conference on Noise
and Vibration Engineering, Leuven, 2014, pp. 415–428.
[2] S.R. Ibrahim, R. Brincker, J.C. Asmussen, Modal parameter identification from
responses of general unknown random inputs, in: Proceedings of XIV
International Modal Analysis Conference, Dearborn, Michigan, 1996, pp.
446–452.
[3] W. Weijtjens, J. Lataire, C. Devriendt, P. Guillaume, Dealing with periodical
loads and harmonics in operational modal analysis using time-varying
transmissibility functions, Mech. Syst. Signal Process. 49 (12) (2014) 154–164.
[4] R. Brincker, C. Ventura, Introduction to Operational Modal Analysis, Wiley,
2015.
[5] G. James, T. Carne, J. Lauffer, The natural excitation technique (next) for modal
parameter extraction from operating structures, Int. J. Anal. Exp. Modal Anal.
10 (1995) 260–277.
[6] F. Magalhães, .A. Cunha, E. Caetano, R. Brincker, Damping estimation using free
decays and ambient vibration tests, Mech. Syst. Signal Process. 24 (5) (2010)
1274–1290, Special Issue: Operational Modal Analysis.
[7] P. Avitabile, Modal space: someone told me that operating modal analysis
produces better results and that damping is much more realistic, Exp. Tech. 30
(2006) 25–26.
[8] T. Lauwagie, R.V. Assche, J.V. der Straeten, W. Heylen, A comparison of
experimental, operational, and combined experimental-operational parameter
estimation techniques, in: Proceedings of Conference on Noise and Vibration
Engineering, Leuven, 2006.
[9] L. Thibault, T. Marinone, P. Avitabile, C. Van Karsen, Comparison of modal
parameters estimated from operational and experimental modal analysis
approaches, in: Proceedings of International Modal analysis XXX, Jacksonville,
Florida, 2012.
[10] M. Ozbek, D.J. Rixen, A new analysis methodology for estimating the
eigenfrequencies of systems with high modal damping, J. Sound Vib. 361
(2016) 290–306.
[11] D.F. Giraldo, W. Song, S.J. Dyke, J.M. Caicedo, Modal identification through
ambient vibration: comparative study, J. Eng. Mech. 135 (2009) 759–770.
[12] D. Smallwood, D. Gregory, A rectangular plate is proposed as an ies modal test
structure, in: Proceedings of International Modal analysis IV, Los Angeles,
California, 1986.
[13] A. Brandt, ABRAVIBE – a MATLAB toolbox for noise and vibration analysis and
teaching, Department of Technology and Innovation, University of Southern
Denmark, 2011, July http://www.abravibe.com.
[14] A. Brandt, Noise and Vibration Analysis – Signal Analysis and Experimental
Procedures, Wiley, 2011.
[15] J. Bendat, A.G. Piersol, Random Data: Analysis and Measurement Procedures,
Wiley Interscience, 1986.
[16] H. Himelblau, J.W. Wise, A.G. Piersol, M.R. Grundvig, Handbook for Dynamic
Data Acquisition and Analysis – IES Recommended Practices 012.1, Institute of
Environmental Sciences, 1994.
[17] E. Orlowitz, A. Brandt, Influence of correlation estimation methods on damping
estimates, in: Proceedings of 5th International Operational Modal Analysis
Conference, Guimaraes, Portugal, 2013.
[18] R. Allemang, D. Brown, Experimental modal analysis and dynamic component
synthesis – vol 3: Modal parameter estimation, Tech. rep., USAF, 1987.
130 E. Orlowitz, A. Brandt / Measurement 102 (2017) 121–130
[19] B. Peeters, G. De Roeck, Reference-based stochastic subspace identification for
output-only modal analysis, Mech. Syst. Signal Process. 13 (1999) 855–878.
[20] H. Vold, K. Napolitano, D. Hensley, M. Richardson, Aliasing in modal parameter
estimation, SOUND&VIBRATION, 2008.
[21] R. Williams, J. Crowley, H. Vold, The multivariate mode indicator function in
modal analysis, in: Proceedings of the III International Modal Analysis
Conference, Orlando, 1985.
[22] E. Orlowitz, A. Brandt, Producing simulated time data for operational modal
analysis, in: Proceedings of International Modal analysis XXXIII, Orlando,
Florida, 2015.
[23] Y.F. Xu, W.D. Zhu, Operational modal analysis of a rectangular plate using
noncontact acoustic excitation, in: Proceedings of International Modal analysis
29th, Los Angeles, California, Springer, New York, New York, NY, 2011, pp.
359–374.
[24] C. Rainieri, G. Fabbrocino, Operational Modal Analysis of Civil Engineering
Structures, Springer, 2014.
[25] G.H. James, T.G. Carne, J.P. Lauffer, A.R. Nord, Modal testing using natural
excitation, in: Proceedings of International Modal analysis X, San Diego, 1992.
[26] C. Rainieri, G. Fabbrocino, Development and validation of an automated
operational modal analysis algorithm for vibration-based monitoring and
tensile load estimation, Mech. Syst. Signal Process. (2015) 512–534.
[27] C. Rainieri, A. Dey, G. Fabbrocinoa, F.S. Magistrisa, Interpretation of the
experimentally measured dynamic response of an embedded retaining wall by
finite element models, Measurement (2016).