Presented at IMAC XXII January 26-29,2004

Measurements Required for Displaying Operating Deflection Shapes

Brian Schwarz & Mark Richardson

Vibrant Technology, Inc.
Jamestown, California

ABSTRACT When collecting data using multiple measurement sets, two

issues must be addressed. First, the phase relationship be-
In order to identify and display the operating deflection
tween measurement sets must be maintained. Second, if the
shapes or operating mode shapes of a machine or structure,
vibration levels change, possibly due to variation in loads
data must be acquired and processed so that all shape DOFs
during testing, the effects of variations in amplitude must be
have the correct magnitudes and phases relative to one an-
addressed.
other. If the data is simultaneously acquired the correct
magnitudes and phases are guaranteed. However, in the Cross Spectra
majority of testing situations, data is acquired in multiple
The Cross spectrum is computed by multiplying the Fourier
measurement sets, and only the data in each measurement
Spectrum of a measured response by the complex conjugate
set is simultaneously acquired. When using multiple meas-
of the Fourier Spectrum of a (fixed) reference response
urement sets, the data must be processed to correct the mag-
measurement. The result is a complex valued measurement
nitudes and phases between sets.
that has the magnitude of the (roving) response times the
In this paper several processing methods involving the magnitude of the reference response. The resulting phase is
Transmissibility, Auto spectrum, and Cross spectrum are the phase difference between the (roving and reference re-
presented for calculating the response spectra required for sponses.
displaying ODS’s. The mathematics behind the processing
Cross Spectrum: G xy (ω) = Fx (ω)Fy (ω)
*
methods is shown, and the sensitivity to noise and structural (1)
changes of the methods is also compared. Various ways of
scaling data from multiple measurement sets is also dis- The reference response is a measurement taken at the same
cussed. DOF for all of the measurement sets. To ensure a good sig-
nal, the reference DOF should be chosen at a DOF where
In addition, it is shown how operating mode shapes are ob- the machine or structure has lots of motion. A DOF on a
tained by curve fitting a set of response spectrum data. Re-
fixed machine base is a poor choice, since Fy ( ω) = 0 and
*
sults from the various spectrum estimates are compared us-
ing data from several bridge tests. therefore the cross spectra will also be zero.
MEASUREMENTS A Cross spectrum is good for ensuring that phases match
between measurements in multiple measurement sets.
ODS’s are displayed from response only measurements tak-
Measurement noise can also be reduced by averaging to-
en from two or more degrees-of-freedom (DOFs) of a ma-
gether several Cross spectra taken from the same pair of
chine or structure. This data is typically a measure of sur- DOFs.
face motion, and can be in displacement, velocity, or accel-
eration units. However, the Cross spectrum is very sensitive to changes in
load levels (and therefore amplitude levels) between meas-
A variety of measurements can be used to display ODS data.
urement sets. Since a load level change affects both the
If ODS’s are displayed directly from time waveform meas-
roving and reference responses, any change in load causes a
urements, they can only be used if the data is simultaneous-
squared change in the Cross spectrum amplitude. Other
ly acquired, or when the measurement is repeatable. methods are less sensitive to load changes.
The majority of ODS analysis is done using a multi-channel
Transmissibility
analyzer or data acquisition system that doesn’t have a suf-
ficient number of channels to collect all of the data simulta- A Transmissibility measurement is similar to a Frequency
neously. In this case, data must be collected using multiple Response Function (FRF) measurement, but uses roving and
measurement sets. This places some limits on the way the reference response signals instead of a force and response
data can be collected. signal pair. The Transmissibility is defined as the Fourier
spectrum of the roving response divided by the Fourier
spectrum of the reference response. It is actually computed

Page 1 of 6
Presented at IMAC XXII January 26-29,2004

by dividing the Cross spectrum between the roving and ref- Transformed to provide a consistent set of response time
erence responses by the Auto spectrum of the reference re- waveforms.
sponse. The resulting measurement provides the motion of
each roving DOF normalized by the motion of the reference
DOF.
Transmissibility:
Fx (ω)Fy* (ω)
H xy (ω) =
Fy (ω)Fy* (ω) (2)
= G xy (ω) / G yy (ω)
Where the Auto spectrum of the reference response is:

G yy (ω) = Fy (ω)Fy* (ω) (3)

Transmissibility’s ensure a phase match between measure-

ment sets through the use of the Cross spectrum. Variations
in load levels are accounted for as well because the roving
response is normalized with respect to the reference motion.
Noise in the roving response can also be averaged out of
Transmissibility through spectrum averaging, as described
before.
Problem with Transmissibility’s
One difficulty with a Transmissibility measurement is that
the peaks in Transmissibility are not evidence of resonances.
Figure 1. Roving Auto Spectrum & Transmissibility.
Rather, resonant frequencies are located at "flat spots" in the
Transmissibility. ODS FRF
The figure below shows a response Auto spectrum plotted An ODS FRF is another frequency domain function that can
above Transmissibility. The response Auto spectrum con- be calculated from response only or operating, data. An
tains resonance peaks. At the frequency of the resonance ODS FRF is computed by replacing the magnitude of the
peak in the Auto spectrum (inside the Band cursor), the Cross spectrum between a roving and reference responses
Transmissibility has a "flat spot", not a peak. Moreover, the pair, with the square root of the magnitude of the roving
peaks in the Transmissibility do not correspond to resonanc- response Auto spectrum. Mathematically, this is equivalent
es but are merely the result of the division of the roving re- to:
sponse spectrum by the reference response spectrum at fre-
quencies where the reference response is relatively small. G xy (ω)
ODS FRF(ω) = G xx (ω)
Responses from Transmissibility’s & Auto Spectra G xy (ω)
Roving responses can be computed by multiplying a set of G xy (ω)
Transmissibility’s by a reference Auto spectrum. = Fx (ω) (6)
G xy (ω)
Fx ( ω) = H xy ( ω) G yy ( ω) (4)
= Fx (ω)
This method gives the true amount of motion at each point,
along with the relative phase information. Noise can be an The ODS FRF measures the true amount of motion at each
issue here because the reference Auto spectrum can contain DOF, along with the correct relative phase information be-
noise. tween multiple roving responses. Furthermore, the ODS
FRF has peaks at resonant frequencies, which makes it
An alternative method is to use a reference Fourier Spec- easier to identify mode shapes from the response only data.
trum instead of the Auto spectrum,
Also, the ODS FRF has better noise characteristics than a
Fx (ω) = H xy (ω)Fy (ω) (5) roving response spectrum that is calculated by multiplying a
reference Auto spectrum by the Transmissibility. The noise
This yields a consistent set of Fourier Spectra for all of rov- in the roving and reference Auto spectra will generally be
ing response DOFs. This data can also be Inverse Fourier

Page 2 of 6
Presented at IMAC XXII January 26-29,2004

about the same. The difference lies in the Cross spectra In order to curve fit operating data using an FRF model, the
which will have less noise than a Transmissibility. following assumptions must be met;
Scaling ODS FRFs 1. The dynamic behavior of the machine or structure
adequately satisfies a set of linear, stationary (non-
The ODS FRF does not account for variations in load level
time varying), 2nd order differential equations.
as the Transmissibility does. However, each measurement
set of ODS FRFs can be re-scaled to account for variations 2. Maxwell’s reciprocity, or symmetry is valid.
in load between measurement sets. An effective way of
In the frequency domain, meeting the above assumptions is
doing this is to create a scale factor (SFi) for the ith meas-
equivalent to satisfying an FRF matrix model. The FRF ma-
urement set as the ratio of the average reference Auto spec-
trix model contains FRFs, Fourier spectra of applied forces,
trum for all measurement sets divided by the reference Auto
and Fourier spectra of resulting displacement responses.
spectrum for the ith measurement set.
Each element of the FRF matrix model can be expressed as
#Meas Sets  ω2 
∑  ∑ G yy (ω j )k 
 j=ω 
Fx ( jω) = H x ,f ( jω)Ff ( jω) (8)
SFi =
k =1  1  (7)
ω2 where:
∑G
j= ω1
yy ( ω j )i
Fx ( jω) = Fourier spectrum of the response.

This scale factor can be calculated for any frequency range Ff ( jω) = Fourier spectrum of the force.
(ω1, ω2) of interest, or just at a specific frequency ω1 = ω2.
H x ,f ( jω) = FRF between the force and response DOFs.
Coherence
The Coherence function can also be used with operating If it is assumed that the Auto spectrum of the excitation
data to ensure that phases in the measurements are valid. A force (or forces) is flat in the frequency range surrounding a
valid Coherence measurement requires that at least two av- mode, then the magnitude of the Fourier spectrum of the
erages of spectrum data are taken. A Coherence value close force can be represented by a constant value.
to one (1) at all frequencies is an indication that;
Ff ( jω) = C (9)
1. The roving and reference responses are linearly re-
lated. The structural response then becomes proportional to the
FRF,
2. The roving and reference spectra have minimum
leakage. Fx ( jω) ∝ CH x ,f ( jω) (10)
3. Measurement noise is low, meaning that the signal to
If the Spectrum of the applied forces is flat in the region of a
noise ratio is high.
mode, then the frequency, damping and magnitude of the
Wherever the Coherence is less than one (1), one (or more) mode is preserved between the FRF and the ODS.
of the above conditions is not being met.
The simplest techniques used to estimate modal parameters
Modal Parameter Estimation using Operating Data are SDOF methods. These methods assume only one mode
dominates at a frequency and there are a variety of SDOF
A set of ODS FRFS can be curve fit to estimate modal pa-
methods that can be used to estimate the modal parameters
rameters, provided that the following assumption is made.
that work with both FRFs and ODS FRFs. Reference [1]
Flat Force Spectrum: If the excitation force spectrum compares curve fitting results using two popular MDOF
matrix can be assumed to be “relatively flat” over in the methods, the Complex Exponential method and the Poly-
frequency range of the modes of interest, then ODS FRFs nomial method.
can be curve fit using an FRF curve fitting model. Example 1: Bridge Data
Modal parameter estimation is based upon the use of an In this example, some experimental data taken from a high-
analytical model for an FRF. During curve fitting, the mod- way Bridge, that has been documented in previous refer-
al frequency, damping and mode shape component (residue) ences [1]-[3], will be post processed to calculate the force
are estimated for each mode by matching the FRF model to spectra of the unmeasured forces, and to compare examine
experimental data, usually in a least squared error sense. the noise content the ODS FRF versus Transmissibility.
(Equivalently, and analytical expression for the Impulse
Response Function, or IRF, can be curve fit to experimental Three tests were performed on the bridge, as documented in
impulse response data.) [3]. The first test was a traditional modal test using two
shakers driven by random signals for excitation. FRFs were
calculated and curve fit to obtain experimental modal pa-

Page 3 of 6
Presented at IMAC XXII January 26-29,2004

rameters. In the second test, excitation was provided by

impacting the bridge (separately at three different locations),
but only bridge responses were measured. In the third test,
only the ambient response to bridge traffic was measured.
Figure 1 shows a 3 by 2 matrix of FRF measurements taken
from the bridge during the multi-shaker test.

Figure 3. Calculated Impact Force Spectra.

Figure 2 shows typical impact responses taken during the
one of the impact tests. Figure 3 shows the (unmeasured)
force spectra at DOFs 1Z and 2Z calculated from the data n
Figures 1 & 2.
Clearly, the force spectra are not “flat” over the entire fre-
quency span, but they do have sufficient levels of the band
(3 to 30 Hz) where the modes where excited during the
shaker test.

Figure 1. 3 by 2 FRF Matrix from Z24 Bridge.

Figure 4. Calculated Ambient Force Spectra.

Figure 4. shows the calculated force spectra of the ambient
forces, using ambient response data and the FRF model in
Figure 1.
Comparing these to the impact spectra, it is clear that the
ambient force spectra have more noise and also violate the
Figure 2. Impact Responses. “flat spectrum” assumption.
Nevertheless, the ODS FRF and Transmissibility can still be
calculated from the response only data. Some results are
shown in Figure 5. The ODS FRFs are on the left, and
Transmissibility based responses are on the right. Although
the results are the same for 1Z:1Z, in the other cases, the
Transmissibility based responses have more noise in them.

Page 4 of 6
Presented at IMAC XXII January 26-29,2004

in Figure 7. Its spectrum has unit amplitude and random

phase. The time waveform has 82000 samples over a time
length of

Figure 7. Random Excitation Force Signal.

Figure 5. ODS FRFs versus Transmissibility Responses.
99 forced response time waveforms were calculated using
Example 2: Modes From Response Only Data the data in Figures 6 & 7. These random responses were
In this example, a set of single reference experimental FRFs then used to calculate a set of ODS FRFs using the response
is used together with a synthesized random time domain at DOF 15Z as the reference response.
force signal to calculate a set of forced responses for a beam
structure. Then, a set of ODS FRFs is calculated from the
random response only data. Finally, the set of ODS FRFs is
curve fit and the response only modal parameters are com-
pared with the experimental modes obtained from the FRFs,
Figure 6 contains some of the typical FRFs measured on the
beam.

Figure 8. Bode Plots of Typical ODS FRFs.

Some of the 99 ODS FRFs are shown in Figure 8. Before
curve fitting, these measurements were inverse FFT’d and
multiplied by an exponential window to insure that they
took the form of decayed sinusoidal functions, similar to
impulse responses. This was followed by an FFT operation
to yield a set of measurements that are similar to FRFs,
Figure 6. Typical FRFs from the Beam. hence can be curve fit using an FRF model. Table 1 contains
the modal frequency estimates for the two sets of experi-
A total of 99 FRFs were measured, in three directions at 33
points. The synthesized random excitation signal is shown

Page 5 of 6
Presented at IMAC XXII January 26-29,2004

mental data. Table 2 contains the MAC values of the mode

CONCLUSIONS
shape estimates.
It was first shown that ODS FRF estimates are more noise-
free than Transmissibility’s multiplied by the reference Auto
Mode FRF ODS FRF spectrum. This is because the denominator of the Transmis-
Frequency (Hz) Frequency (Hz) sibility also contains the measurement noise of the reference
Auto spectrum, which adds more noise to the final result.
1 165.0 165.0 The ODS FRF only contains the additive noise of the roving
response Auto spectrum.
2 224.6 224.5
Next, it was shown that modal parameters can be obtained
3 347.9 346.7 by using an FRF model to curve fit a set of ODS FRFs.
These results were compared to the parameter estimates
4 460.4 473.0 obtained from a set of FRFs from the same structure.
Obtaining modal parameter estimates from response only
5 493.0 493.6 data has been dubbed Operational Modal. It was shown
that this approach has limitations, and produces noisier
6 635.5 633.7 measurements and consequently less accurate modal param-
eters than the more traditional FRF based technique.
7 1109.0 1111.0
Nevertheless, in situations where the excitation forces can-
8 1211.0 1215.0 not be measured, the use of response only data can still pro-
vide some useful results.
9 1323.0 1319.0 REFERENCES
10 1557.0 1547.0 1. Schwarz, B. and Richardson, M. "Modal Parameter
Estimation from Operating Data" Sound and Vibration
Table 1. Modal Frequency Comparison. Magazine, January 2003.
2. Schwarz, B. & Richardson, M. “Modal Parameter Es-
timation from Ambient Response Data” International
Mode Modal Assurance
Modal Analysis Conf. (IMAC XXII), February 5-8,
Criterion
2001. (Vibrant Tech. Paper No. 33 www.vibetech.com)
1 0.965 3. Schwarz, B. & Richardson, M. “Post-Processing Ambi-
ent and Forced Bridge Data to Obtain Modal Parame-
2 0.024 ters” International Modal Analysis Conf. (IMAC XXII),
February 5-8, 2001. (Vibrant Tech. Paper No. 34
3 0.986 www.vibetech.com)

4 0.527
4. Kramer, C., de Smet, CAM, de Roaek, G. “Z24 Bridge
Damage Detection Tests”, IMAC, 1999.
5 0.964 5. Kramer, C., de Smet, CAM, Peters, B. “Comparison of
Ambient and Forced Vibration Testing of Civil Engi-
6 0.932 neering Structures”, IMAC, 1999.

7 0.964 6. Richardson, M. "Is It A Mode Shape Or An Operating

Deflection Shape?" Sound and Vibration Magazine,
8 0.955 February, 1997. (Vibrant Tech. Paper No. 10
www.vibetech.com)
9 0.961 7. Vold, H. Schwarz, B. & Richardson, M. “Measuring
Operating Deflection Shapes Under Non-Stationary
10 0.918 Conditions” International Modal Analysis Conf. (IMAC
XXI), February 7-10. 2000. (Vibrant Tech. Paper No.
Table 2. Modal Shape Comparison. 32 www.vibetech.com)

Page 6 of 6