Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL

Curve Fitting Analytical Mode Shapes to Experimental Data

Brian Schwarz, Shawn Richardson, Mark Richardson

Vibrant Technology, Inc.
Scotts Valley, CA

ABSTRACT A mode of vibration is a mathematical representation of a struc-

tural resonance. Each mode is defined by three distinct numer-
In this paper, we employ the fact that all experimental vibration
ical parts; a natural frequency, a damping or decay constant, and
data, whether in the form of a set of FRFs or a set of output-only
a mode shape. A mode shape represents the “standing wave
spectra, is a summation of resonance curves, each curve due to
deformation” of the structure at the natural frequency of a res-
a mode of vibration. We also use this superposition property of
onance. This standing wave behavior is caused when energy be-
modes to calculate a modal participation matrix, a measure of
comes trapped within the material boundaries of the structure
the participation of each mode in the experimental vibration data
and cannot readily escape.
First we show how this superposition property can be used to
Modal Participation
curve fit a set of FEA mode shapes to EMA mode shapes or
ODS’s. The modal participation matrix is calculated as a least- When the dynamic response of a structure is expressed in terms
squared-error solution, so any number of FEA mode shapes can of modal parameters, every solution is a summation of contri-
be curve fit to any number of EMA mode shapes or ODS’s. butions from all of the modes. Another way of expressing this
Next we show how an expanded and enhanced set of FRFs, superposition property is that all modes participate in or con-
Cross spectra or ODS FRFs is obtained by curve fitting FEA tribute to the dynamic response of a structure when it is excited
mode shapes to experimental data. by applied forces. Ideally, all structures have an infinite num-
ber of modes, but in a practical sense only a few low frequency
This approach in an alternative to FEA Model Updating, where
modes participate significantly in their response.
an FEA model is modified so that its modes more closely corre-
late with experimental data. By curve fitting FEA shapes to ex- If a structure’s overall dynamic response is represented by time
perimental data, an extending and enhanced dynamic model is waveforms, these waveforms can be decomposed into a sum-
obtained which is more suitable for machinery & structural mation of modal contributions. Likewise, if the dynamic re-
health monitoring, and for troubleshooting noise & vibration sponse is represented by a set of frequency functions or spectra,
problems using SDM and MIMO methods. the overall response can also be decomposed into a summation
of resonance curves.
KEY WORDS
FEA Modes
Frequency Response Function (FRF)
Auto power spectrum (APS) FEA modes are solutions to a set of time domain equations of
Cross power spectrum (CPS) motion. The equations are a statement of Newton’s second law
ODSFRF [3]-[5], a force balance between internal and external forces.
Operating Deflection Shape (ODS) This force balance is written as a set of differential equations,
Finite Element Analysis (FEA)
Experimental Modal Analysis (EMA) [M ]{ x(t )} + [C]{ x (t )} + [K ]{ x(t )} = {f (t )} (1)
FEA Mode Shapes where,
EMA Mode Shapes
Modal Assurance Criterion (MAC) [M ] = (n by n) mass matrix
Shape Difference Indicator (SDI)
[C] = (n by n) damping matrix
Structural Dynamics Modification (SDM)
Multi-Input Multi-Output (MIMO) Modeling & Simulation [K ] = (n by n) stiffness matrix
INTRODUCTION {x(t )} = Accelerations (n-vector)
{x (t )} = Velocities (n-vector)
When resonances are excited by dynamic forces in a machine,
or in a mechanical or civil structure, response levels can far ex- {x(t )} = Displacements (n-vector)
ceed deformation levels due to static loads. Moreover, high lev- {f (t )} = Externally applied forces (n-vector)
els of resonance-assisted dynamic response can cause rapid and
unexpected failures, or over long periods of time, structural fa- This set of differential equations describes the dynamics be-
tigue and material failure can often occur. tween n-discrete degrees-of-freedom (DOFs) of the structure.
Equations can be created for as many DOFs of a structure as
necessary to adequately describe its dynamic behavior.

Page 1 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
Finite element analysis (FEA) is used to create the coefficient ωk = Natural frequency of the k th mode
matrices of the differential equations (1). The mass & stiffness
matrices can be generated in a straightforward manner but the {u k } = Mode shape for the k th mode (n-vector)
damping matrix cannot, hence it is usually left out of the FEA A k = Scaling constant for the k th mode
model. The equations are then solved for the analytical FEA
mode shapes and their corresponding natural frequencies. Figure 1 shows a transfer function for a single mode of vibra-
tion, plotted over half of the s-plane.
FEA modes are a mathematical eigen-solution to the homoge-
neous form of equations (1), where the right-hand side is zero.
Each natural frequency is an eigenvalue, and each mode shape
is an eigenvector.
EMA Modes
EMA mode shapes are obtained by curve fitting a set of exper-
imentally derived FRFs [3]-[5]. FRF-based curve fitting is a nu-
merical process by which an analytical parametric model with
unknown modal parameters in it is matched to experimental
FRF data over a band of frequencies. Equation (3) is an expres-
sion of the analytical FRF model.
In the frequency domain, the equations of motion are written as
algebraic equations, in a form called a MIMO model or transfer
function model. Like equation (1), this model also describes the
dynamics between n-DOFs of the structure. It contains transfer
functions between all combinations of DOF pairs,
{ X(s )} = [H(s )]{ F(s )} (2)

where,
Figure 1. Transfer Function & FRF for a Single Mode in the s-plane
s = Laplace variable (complex frequency)
Experimental FRFs
[H(s )] = (n by n) matrix of transfer functions
{ X(s )} = Laplace transform of displacements (n-vector) An FRF is defined as the values of a transfer function along the
jω -axis in the s-plane, as shown in Figure 1. FRFs can only be
{F(s )} = Laplace transform of external forces (n-vector)
calculated from experimental data when all of the excitation
These equations can be created between as many DOF pairs of forces and responses caused by them are simultaneously ac-
the structure as necessary to adequately describe its dynamic be- quired. Equation (3) is the analytical form of an FRF that is
havior. used for FRF-based curve fitting of experimental data. The out-
come is an EMA pole & mode shape for each term that is used
When a transfer function is represented analytically as the par- in the summation during curve fitting.
tial fraction expansion shown in equations (3) & (4) below, it is
clear that its value at any frequency is a summation of reso- Mode Shape Curve Fitting
nance curves, one for each mode of vibration. Equation (4) also tells us that each FRF can be represented as a
m * summation of resonance curves, hence a set of FRFs can be
[rk ] [r ]
[H(s )] = ∑ − k
(3) decomposed into their resonance curves by curve fitting them
k =1 2 j (s − p k ) 2 j (s − p *k ) one frequency at a time with a set of mode shapes.

or, Output-Only Frequency Spectra

In cases where the excitation forces are not measured and there-
m
A {u }{u k }t A *k {u *k }{u *k }t
[H(s )] = ∑ k k − (4) fore FRFs cannot be calculated, three other types of output-only
k =1 2 j (s − p k ) 2 j (s − p *k ) frequency spectra can be calculated from acquired response time
waveforms; Fourier spectra, Cross spectra, and ODSFRFs.
where,
A Fourier spectrum is simply the Digital Fourier Spectrum
m = Number of modes of vibration (DFT) of a digital response time waveform. It is calculated with
[rk ] = (n by n) residue matrix for the k th mode the FFT algorithm. A Cross spectrum is calculated between
two simultaneously acquired responses. It is the DFT on one
p k = − σ k + jω k = Pole location for the k th mode signal multiplied by the complex conjugate of the other. An
σ k = Damping decay of the k th mode ODSFRF is the Auto spectrum of a roving response combined

Page 2 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
with the phase of the Cross spectrum between the roving re- EXPANDING EMA MODES USING FEA MODES
sponse and a (fixed) reference response.
The aluminum plate shown in Figure 2 was tested using a roving
When it can be assumed that these output-only measurements impact hammer. During the test, a 5 by 6 grid of points was im-
can also be represented as a summation of resonance curves, pacted in the vertical direction. A tri-axil accelerometer located
they too can be decomposed into a set resonance curves by curve near one corner of the plate was used to measure the response
fitting them one frequency at a time with a set of mode shapes. due to the impacts.
Operating Deflection Shape (ODS)
When a vibration response is analytically modeled or experi-
mentally measured at two or more points & directions on a ma-
chine or structure, this data is called an Operating Deflection
Shape (or ODS) [4]. Three different types of ODS’s are possi-
ble; time-based ODS’s, frequency-based ODS’s, and order-
based ODS’s.
Time-Based ODS
A time-based ODS is simply the values of two or more simulta-
neously acquired or calculated time waveforms for the same
time value. A time-based ODS is the true overall response of
the structure at any moment of time.
Frequency-Based ODS
A frequency-based ODS is the values of two or more frequency
domain functions (FRFs or spectra) at the same frequency. A Figure 2. Impact Test of Aluminum Plate
frequency-based ODS is the true overall response of the struc-
ture for any frequency for which the measurement functions
were calculated
All frequency domain functions, (except an Auto spectrum) are
complex valued (with magnitude & phase), so all frequency-
based ODS’s are also complex valued.
Order-Based ODS
In a rotating machine, the dominant forces are applied at multi-
ples of the machine running speed, called orders. An order-
based ODS is assembled from the peak values at one of the or-
der frequencies in a set of output-only frequency spectra. These
spectra are calculated from response data that is acquired while
the machine is running. When displayed in animation on a 3D Figure 3. FRF-based Curve Fit of a Measurement
model of the machine, an order-based ODS is a convenient way
of visualizing distributed vibration levels. These distributed lev- FRFs were calculated from simultaneously acquired force & re-
els can also be used for monitoring the health of the machine. sponse data, and the modal parameters of 14 EMA modes were
extracted by curve fitting the FRFs using FRF-based curve fit-
ODS Expansion ting. A typical curve fit is shown in Figure 3.
In a previous paper [1], it was shown how modes participate in Plate FEA Model
an order-based ODS of a rotating machine, and how they partic-
ipate differently at different operating speeds. It another previ- Also, an FEA model of the plate was created by adding 180 FEA
ous paper [2], it was also shown how the modal participation brick elements to the same model used to display the EMA mode
can be used to expand an order-based ODS so that it is a valid shapes. The first flexible body FEA mode is shown in Figure 4.
representation of the ODS for all DOFs of the machine, both The 30 impact points are also labeled.
measured & un-measured.
Measurement Expansion
In this paper the same curve fitting and expansion equations that
were used for ODS expansion will be used to decompose and
expand a set of FRFs, and sets of output-only Cross spectra and
ODSFRFs.

Page 3 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
Shape Expansion
The participation factors in Figure 5 were used to expand the
EMA modes from 30 DOFs to 1248 DOFs, the number of FEA
mode shape DOFs. The MAC values between the expanded &
original EMA mode shapes are shown in Figure 6. The SDI
values between the expanded EMA and the original EMA mode
shapes are shown in Figure 7.
Both Figures 6 & 7 indicate that the FEA mode shapes were ac-
curately curve fit to the EMA shapes. MAC indicates the co-
linearity of each shape pair. SDI more strongly indicates that
each shape pair has nearly the same values for the 30 DOFs that
are common among the original and expanded shapes.

Figure 4. FEA Mode Shape in Animation

Modal Participation
The bar chart of the participations of the first 14 flexible body
FEA modes in the 14 EMA modes is shown in Figure 5. The
participation values reflect the different scaling of the FEA
mode shapes versus the EMA modes. The FEA mode shapes
were scaled to Unit Modal Masses, while the EMA mode shapes
contained the residues (numerators) obtained from using equa-
tion (3) to curve fit the FRFs.

Figure 7. SDI Values between Expanded & Original EMA modes

FRF DECOMPOSITION
The 14 FEA mode shapes were then used to decompose the 30
FRF measurements for the aluminum plate at each frequency.
The values of the 30 FRFs at each frequency are a frequency-
based ODS. The FEA mode shapes participate differently in the
ODS at each frequency, and when the participations are plotted
for all frequencies, they create a separate resonance curve for
Figure 5. Participation of FEA modes in EMA modes each FEA mode. Figure 8 is a plot of the 14 resonance curves
obtained from decomposing the 30 FRFs using the FEA modes.

Figure 6. MAC Values between Expanded & Original EMA modes Figure 8. FRF Decomposition using 14 FEA modes

Page 4 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
Notice that each resonance curve in Figure 8 has a single domi-
nate peak in it at the natural frequency of one of the 14 reso-
nances. These resonance curves clearly illustrate the superposi-
tion property of modes.
FRF RECONSTRUCTION
A set of reconstructed FRFs can then be calculated by multi-
plying the resonance curves in Figure 8 by the FEA mode
shapes. Figure 9 shows a reconstructed FRF overlaid on its cor-
responding original experimental FRF.
FRF EXPANSION
Not only are the original 30 FRFs reconstructed, but an ex-
panded set of 1248 FRFs can be calculated using all the DOFs
of the FEA mode shapes. This expanded set of FRFs can then
be curve fit using FRF-based curve fitting to obtain a set of EMA Figure 10. Cross spectra Decomposition using 14 FEA modes
modes with frequency, damping, and mode shapes with 1248
DOFs in them.
NOTE: Only FEA mode shapes are required to perform de-
composition, reconstruction, and expansion of experimental vi-
bration data.
Decomposition, reconstruction and expansion can be applied to
either time or frequency domain vibration data.

Figure 11. Reconstructed & Experimental Cross spectra Overlaid

ODSFRF DECOMPOSITION
The 14 FEA mode shapes were also used to decompose 30 ODS-
FRF measurements for the aluminum plate at each frequency.
Figure 12 is a plot of the participations (or resonance curves) of
the 14 FEA modes in the ODSFRF measurements.

Figure 9. Reconstructed & Experimental FRFS Overlaid

CROSS SPECTRUM DECOMPOSITION
The 14 FEA mode shapes were also used to decompose 30 Cross
spectrum measurements for the aluminum plate at each fre-
quency. Figure 10 is a plot of the participations of the 14 FEA
modes in the Cross spectra measurements.
CROSS SPECTRUM EXPANSION
Figure 11 shows a reconstructed Cross spectrum overlaid on an
original experimental Cross spectrum. The reconstructed Cross
spectra were calculated by multiplying the resonance curves in
Figure 10 by the FEA mode shapes. Again, not only can the
original 30 Cross spectra be reconstructed, but an expanded set Figure 12. ODSFRF Decomposition using 14 FEA modes
of 1248 Cross spectra can be calculated using the DOFs of the
FEA mode shapes.

Page 5 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
ODSFRF EXPANSION to lowest MAC value. Figure 15 is an ordered magnitude plot
of SDI values.
Figure 12 shows a reconstructed ODSFRF overlaid on an origi-
nal experimental ODSFRF. The reconstructed ODFRFs were Both MAC & SDI have their lowest values for the reconstructed
calculated by multiplying the resonance curves in Figure 12 by & experimental FRF pair at DOF 14Z. These low values indicate a
the FEA mode shapes. Again, not only can the original ODS- mismatch between the reconstructed & experimental FRFs. The over-
FRFs be reconstructed, but an expanded set of 1248 ODSFRFs laid reconstructed & experimental FRF pair is shown in Figure 16.
can be calculated using all DOFs of the FEA mode shapes.

Figure 15. Ordered SDI for Reconstructed & Experimental FRF Pairs

Figure 13. Reconstructed & Experimental ODS FRFs Overlaid

MEASUREMENT ERRORS
One of the most useful applications of measurement curve fit-
ting and expansion using mode shapes is that experimental er-
rors are quickly spotted. Bad measurements are found by over-
laying each reconstructed measurement on its corresponding ex-
perimental data. One pair of each of the measurement types is
overlaid in Figures 8, 11, & 13.
Measurement Comparison Using MAC & SDI
A numerical method for comparing reconstructed & experi-
mental data is to use MAC & SDI values. Both MAC & SDI
have values between 0 & 1. When MAC is applied to pairs of Figure 16. Reconstructed & Experimental 14Z:5Z FRFs Overlaid
FRFs, it is also called the Frequency Response Assurance Cri- Since all of the other FRF pairs have high MAC & SDI values (close to
terion (or FRAC). 1), this is strong evidence that curve fitting the FEA mode shapes to the
experimental data provided accurate reconstructing FRFs, except in a
few cases like the pair shown in Figure 16. Since the reconstructed
FRFs were the result of a least-squared-error curve fit to all 30 exper-
imental FRFs, the strongest conclusion from the mismatch in Figure 16
is that the experimental FRF data is in error.
Since the aluminum plate was tested using as roving impact test, it can
be concluded that Point 14 on the plate was impacted at a different
point than Point 14 on the FEA model.
CONCLUSIONS
It was shown how the FEA mode shapes of a structure can be
used to decompose, reconstruct, and expand a set of FRFs, Cross
spectra, and ODSFRFs. FRFs are normally calculated when all
of the excitation forces causing a structure to vibrate are meas-
Figure 14. Ordered MAC for Reconstructed & Experimental FRF Pairs ured. Cross spectra & ODSFRFs are calculated from output-
only data when the excitation forces are not measured.
Figure 14 is a magnitude plot of the MAC values between the
30 reconstructed & experimental FRFs, ordered from the highest It was demonstrated in each case that these three types of exper-
imental data can be decomposed into a summation of resonance

Page 6 of 7
Presented at IMAC XXXVI January 25-28, 2016 Orlando, FL
curves by curve fitting a set of mode shapes to them, one fre- REFERENCES
quency at a time. Then it was shown that the participations of
1. B. Schwarz, M. Richardson, “Linear Superposition and Modal
the modes in the experimental data can be used to construct an Participation” IMAC XXXII, February 3-6, 2014.
expanded set of measurements using the DOFs of the mode
shapes, including the DOFs that were common with the experi- 2. B. Schwarz, Shawn Richardson, Mark Richardson, “Using Mode
mental data. Shapes for Real Time ODS Animation” IMAC XXXIII
February 3-5, 2015.
A key advantage of this approach is that only the mode shapes
are required in these calculations. Accurate mode shapes can be 3. S.C. Richardson, M.H. Richardson “Using Photo Modeling
easily obtained from a simple FEA model. Accurate FEA fre- to Obtain the Modes of a Structure” Proceedings of the
quencies that match the EMA frequencies usually require a more International Modal Analysis Conference, 2008.
accurate FEA model, but frequencies are not required for this
calculation. 4. M. Richardson, “Is It a Mode Shape or an Operating Deflection
Shape?” Sound and Vibration magazine, March, 1997.
Another advantage of this approach is that complex data can be
decomposed using normal modes. Modes are called normal 5. M. Richardson "Modal Analysis versus Finite Element Analysis"
when the FEA model they are derived from has no damping Sound and Vibration Magazine, September 2005.
terms, and hence the mode shapes are real valued or normal.
Normal modes can be used to expand complex mode shape or
ODS data because the modal participation factors are complex
valued.
This decomposition and expansion capability is useful not only
for creating measurements for all of the un-measured DOFs on
a structure, but also for identifying bad measurements. The ex-
panded set of measurements can also be curve fit using FRF-
based curve fitting to obtain EMA modes with frequency,
damping, and expanded mode shapes in them. This EMA
modal model can then be used for SDM and MIMO Modeling
& Simulation studies involving un-measured DOFs of the
structure.
This combined use of an analytical model with experimental
data provides a more complete characterization of the dynamic
behavior of a structure from a relatively small number of meas-
urements. This means that less time & expense are required to
obtain meaningful data for use in machinery & structural health
monitoring, or for troubleshooting noise & vibration problems.

Page 7 of 7