Output-Only Modal Identification of a Nonuniform

Beam by using Decomposition Methods

Rickey A. Caldwell, Jr.

Department of Mechanical Engineering
Michigan State University
East Lansing, MI, USA
caldwe20@msu.edu
517-355-8310

Brian F. Feeny∗
Department of Mechanical Engineering
Michigan State University
East Lansing, MI, USA
feeny@egr.msu.edu
517-353-9451

Reduced-order mass weighted proper orthogonal decompo-
sition (RMPOD), smooth orthogonal decomposition (SOD),
and state variable modal decomposition (SVMD) are used
to extract modal parameters from a nonuniform experimen-
tal beam. The beam was sensed by accelerometers. Ac-
celerometer signals were integrated and passed through a
high-pass filter to obtain velocities and displacements, all
of which were used to build the necessary ensembles for the
decomposition matrices. Each of these decomposition meth-
ods was used to extract mode shapes and modal coordinates.
RMPOD can directly quantify modal energy, while SOD and
SVMD directly produce estimates of modal frequencies. The
extracted mode shapes and modal frequencies were com-
pared to an analytical approximation of these quantities, and
to frequencies estimated by applying the fast Fourier trans-
form to accelerometer data. SVMD is also applied to esti-
mate modal damping, which was compared to that estimated
by logarithmic decrement applied to modal coordinate sig-
nals, with varying degrees of success.
1 Introduction
Output-only modal analysis uses displacements, ac-
celerations, or velocities to determine modal information.
Output-only modal analysis is useful when inputs cannot
be recreated or are unknown. Other benefits include the
avoidance of frequency response functions (FRFs) and
FRF matrices and related testing procedures [1] which
reduces test time and amount of data needed. Output-only
modal analysis can be done in time domain or frequency
domain. Some examples of time domain methods include
∗ Address all correspondence to this author.
1
the eigensystem realization algorithm [2], Ibrahim time
domain method [3], independent component analysis [4, 5]
and the polyreference method [6]. Examples of frequency
domain methods are orthogonal polynomial methods [7, 8],
complex mode indicator function [9], and frequency domain
decomposition [10]. The methods that will be explored
in this work are in the time domain and are extensions of
proper orthogonal decomposition (POD). These include
mass-weighted proper decomposition, smooth orthogonal
decomposition, and state variable modal decomposition. The
outputs used are displacements, velocities, and accelerations.
In POD a structure is sensed with M sensors, whose
signals are processed as needed to generate displacement
time signals. An M × N ensemble matrix X is created
such that each row corresponds to a sensor and each
column is a time step. That is, X = [x1 x2 · · · xM ]T , where
xi = [xi (0) xi (ΔT ) xi (2ΔT ) · · · xi (NΔT )], and N is the
number of time samples. This ensemble matrix is used in
both POD and the other decomposition methods.
In POD, the “correlation matrix” R is formed using the
XXT
ensemble matrix, such that R = . Then the eigenvalue
N
problem (EVP) Rψ ψ = λψψ is solved. If the mass is uniform,
and the system is lightly damped, POD produces estimates
of the mode shapes from free vibration responses [11–13]
and random responses [11]. The former case has been
verified in experiments [14–16].
The decomposition methods which are generalizations
of POD reported here are reduced-order mass-weighted POD
(RMPOD), state-variable modal decomposition (SVMD),
and smooth orthogonal decomposition (SOD). These meth-
ods are applied to a nonuniform beam experiment. RMPOD
has been applied successfully in simulations [8]. SVMD
has been verified for a uniform beam [17]. SOD has been
applied in simulations [18] so an experimental applications
are of interest.

In the next two sections, the beam experiment is de-

scribed, and an approximate analytical description of the
modes is outlined. Then RMPOD, SOD, and SVMD are ap- Fig. 1: Experimental beam
plied to the data from an experimental beam. In each section,
the application of the method is introduced, and the results
from the beam experiment are presented.
removed using the detrend command. The signal was
passed through a second-order high-pass filter twice: once
EXPERIMENTAL SETUP
forward and the second time backwards, which corrected
A thin nonuniform lightly damped cantilevered beam, a
the resulting phase shift caused by the filter. The signal was
Buck Bros. tapered saw blade, shown in Fig. 1, was sensed
then integrated using the MATLAB function, cumtrapz,
with eleven accelerometers (PCB model 352B10) which
which approximates the integral using the trapezoid rule.
have a sensitivity of 10 ± 0.5 mV
g . The beam was clamped in One iteration of this process yielded velocities. The mean
a fixture such that the length was 11.5 inches (0.2921 m).
subtraction, detrending, filtering, and integration processes
The width was 3.5 inches (0.0889 m) at the clamped end,
were repeated again to produce displacements. Acceler-
tapering from 3.5 inches (0.0889 m) at a location of 1.78
ations, velocities, and displacements were used to create
inches (0.0452 m) from the clamp, to 0.80 inches (0.02032
ensemble matrices for RMPOD, SOD, and SVMD. The
m) at the free end. The beam was clamped such that the
modal assurance criterion (MAC) [19] was used to evaluate
midline of the taper was horizontal, and the flexure of the
the quality of the extracted mode shapes as referenced to the
beam was in the horizontal plane. Parameters of the beam
discretized analytical mode shapes of the nonuniform beam
are listed in Table 1. The beam material was unknown. The
model, which is discussed in the next section.
density was measured, and Young’s modulus was assumed
to be that of stainless steel.
The impact tests were conducted with a variety of im-
The accelerometers were placed at one inch intervals pulses to gain insight into their effects on modal identifica-
starting one inch from the clamped end and progressing to tion. The beam was struck with an impact hammer at sensor
the free end. Each accelerometer was attached to the beam location of 2 in (0.05 m), 6 in (0.15 m), and 11 in (0.28 m)
using wax. The resulting signals from the accelerometers from the clamped end. Two impulse magnitudes were used,
were connected through a PCB Model 481 signal condi- the first, a small amplitude impulse such that all resulting
tioner which also amplified the signal with a gain of 10. Af- free vibration accelerations were less than 10g’s. The sec-
ter the signals were passed through the signal conditioner ond, a large amplitude impulse where the free vibration ac-
they were sent to the TEAC GX-1 data recorder. The sig- celerations were greater than 20g’s and less than 50g’s. Fi-
nals were sampled at a rate of 5000 Hz and sent through a nally, the signal of resulted free vibrations were divided into
low-pass filter with a cutoff frequency of 2000 Hz. This sat- four equal time bins. The results presented in this paper
isfied the Nyquist criterion and prevented aliasing. A fast used the following parameters: a small amplitude impulse
Fourier transform (FFT) of any of the acceleration signals two inches from the clamped end and the second time bin,
during free vibration revealed the following natural frequen- t ∈ [1.3406, 1.4594] seconds.
cies: 8.45 Hz, 40.28 Hz, 107.4 Hz, 205.1 Hz, 498 Hz, and
677 (Table 2).

In the free-vibration experiments the beam was struck

two inches from the clamped end with an impact hammer. ANALYTICAL APPROXIMATION
The resulting accelerations were recorded and imported
In order to evaluate the experimental results an analyt-
into MathWorks’ MATLAB. In MATLAB further signal
ical approximation to the nonuniform Euler-Bernoulli equa-
processing was performed. Since the accelerometers had
tion was formulated. The general equation describing a
phase distortions near 8 Hz, a high-pass filter was used with
clamped-free beam with variable cross-section is
a cutoff frequency of 20 Hz. This attenuated the first mode
and consequently removed the first mode from the results ! "
of the decomposition methods. Within MATLAB the mean d2 d2
m(x)ÿ(x,t) + EI(x) y(x,t) =0 (1)
of each signal was subtracted out and linear trends were dx2 dx2

2
with boundary conditions Material Property Value
Young’s Modulus (E) 190 × 109 Pascals
y(0,t) = 0
d *Density (ρ) 7035 kg/m3
y(0,t) = 0
dx Height (h) 0.00066 m
d2
[EI(x)y(L,t)] = 0
dx2
d3 0.08787 m 0 ≤ 0.0456 m
[EI(x)y(L,t)] = 0
dx3 Width (w) −0.27409x + 0.1004
0.0457 m ≤ x ≤ 0.2921 m
where y(x,t) is the transverse displacement at location x of
the beam, m(x) is the mass per unit length, E is Young’s Length (L) 0.2921 m
modulus, and I is the cross-sectional area moment of interia.
Table 1: Material properties for the beam. *Estimated den-
sity includes the mass of the accelerometers
We discretized this equation by approximating y(x,t) ∼
=
M
∑ qi (t)ui (x), where ui (x) are assumed modes and qi (t) are Solving this EVP leads to estimates λi ∼ = ω2i of the modal
i=1
the assumed modal coordinates. Inserting into Eqn. (1), mul- frequencies of the beam model, and a modal matrix P for the
tiplying by u j (x), and integrating (the second term by parts system of equations (2). Applying the transformation q = Pr
twice) yields diagonalizes Eqn. (2). As such, composing q = Pr and
y = Uq, we find that y ∼= UPr transforms system Eqn. (3) in
original coordinates, to the diagonal system in r. Then the
Mq̈ + Kq = 0 (2)
discretized mode shapes are approximated by the columns
of the composite modal matrix UP.
where the elements mi j of M and ki j of K are
In application, a matrix U is created such that
! L U = [u1 u2 · · · uM ] where ui ’s are the discretized assumed
mi j = m(x)ui (x)u j (x) dx modal functions. We obtain these from the true modal
0 functions of the damped-free uniform Euler-Bernoulli
(2.5) equation. We then build the associated mass and stiffness
! L
matrices M and K using Eqn. (2.5). Matrix P is created
ki j = EI(x)u%%i (x)u%%j (x) dx such that P = [p1 p2 · · · pM ] where pi ’s are from the
0
resulting eigensystem in Eqn. (4). Then, the LNMs for
the nonuniform beam are approximated as columns of UP.
These modes will be compared to experimentally estimated
modes in the following section.
The spatial discretization of y(x,t) can be expressed as
y∼= Uq, where U is a modal matrix made of column vectors When this model was applied to the experimental saw
that are spatial descretizations of the ui (x). As such, the ele- blade, the estimated modal frequencies were 9.02 Hz, 43.51
M
Hz, 112.00 Hz, 214.46 Hz, 350.50 Hz, 520.93 Hz, and
ments of y are yi = y(xi ,t) ∼
= ∑ u j (xi )q j (t). We can assume
j=1 725.83 Hz, as listed in Table 2.
that there exists a discretized system of equations
SOURCES OF ERROR
# + Ky
Mÿ # =0 (3) Sources of error for the frequency estimation are dis-
cussed here. The ratios between analytical and experimental
frequencies, shown in Table 2, are uniformly between 1.04
that faithfully discretizes the original system Eqn. (1), such
# and and 1.08. This could be caused by an error in the effective
that the system matrices are related by M = UT MU
# parameter group, for example due to errors in E and h,
K = UT KU.
which would systematically scale the estimated frequencies.
The error due to measured dimensions and density are likely
Assuming synchronous motion, such that q(t) = pr(t), to be small. Discretized models sometimes are slightly
Eqn. (2) leads to the eignevalue problem (EVP) stiffened by the “constraint” associated with discretization.

Another source of error may be in the assumptions

λMp = Kp. (4) embedded in the beam model. Since the beam is wide, there

3
 Mode no. Analytical Beam FFT Ratio is solved: RMr ψ = λψ ψ. The eigenvectors ψ correspond to
Approx. LNMs and the eigenvalues are mass-weighted mean squared
values of the modal coordinates [20].
1 9.02 Hz 8.55 Hz 1.06
2 43.51 Hz 40.28 Hz 1.08 When the dimension of M is larger than that of R, an in-
terpolation scheme can be used to reduce the effective order
3 112.00 Hz 107.40 Hz 1.04 of the system to that of the number of sensors used in build-
4 214.46 Hz 205.10 Hz 1.05 ing R [20]. For the case of a one-dimensional distributed
parameter system, using linear interpolation represented by
5 350.50 Hz not seen in N/A interpolation functions ηi (x) leads to a reduced mass matrix
FFT Mr whose elements are
6 520.93 Hz 498.00 Hz 1.05
! L
7 725.83 Hz 677.50 Hz 1.07 Mi j = M ji = m(x)ηi (x)η j (x)dx. (5)
0
Table 2: Second column: Frequencies predicted by the ana-
lytical approximation. Third column: Frequencies estimated We used ηi (x) and η j (x) as tent-shaped interpolation func-
from FFTs of experimental beam accelerations tions of the form ηi (x) = 1h (x−(i−1)h) for (i−1)h ≤ x < ih,
. ηi (x) = 1h (x − (i + i)h) for ih ≤ x < (i + 1)h, and ηi (x) = 0
otherwise, where h is the spatial interval of the sensors on
the beam.
may be some influence of plate characteristics. The infinite
∂2 w
uniform 1-D plate equation is 2 + Dw%%%% = 0, while that In this work, we inserted an expression representing the
∂t
∂2 w EI %%%% mass distribution m(x), as defined by the taper in the saw
of the uniform beam is 2 + u = 0. The ratio between blade, into Eqn. (5) to obtain the Mr matrix for weighting
∂t m
EI the RMPOD process.
parameter groups is = 1 − ν2 , where ν is Poisson’s
mD
ratio, which bounds the deviation between the infinite 1-D As such, RMPOD was applied. A permutation of input
plate and the Euler-Bernoulli beam. Using ν = 0.3, this parameters and ensembles was used in the RMPOD decom-
leads to an decrease of up to about 5%, in the analytically position to gain some experience regarding their effects on
estimated frequencies, which would increase the difference the modal decomposition results. Results were evaluated
between the frequencies of the approximated model and the using the modal assurance criterion (MAC) [19] values with
experiments. discretized analytical approximation mode shapes as the
reference. In this case, the RMPOD based on acceleration
ensembles, with small impulses located 2 inches (0.0508 m)
from the clamped end gave the best overall performance,
REDUCED-ORDER MASS-WEIGHTED PROPER although only marginally better than other permutations of
ORTHOGONAL DECOMPOSITION the testing input parameters. Figures 2, 3, and 4 show the
It was noted above that POD produces estimates of plots of the extracted modes, modal coordinate accelerations
mode shapes when the mass distribution is uniform. When of the extracted modes, and the magnitudes of the FFTs of
the mass distribution is nonuniform and known, the weighted the modal coordinate accelerations for the second, third, and
EVP RMψ ψ = λψ ψ produces estimates of the normal modes. fourth mode respectively.

Mass-weighted POD is straight forward when the
mass matrix is of the same dimension as the correlation
matrix. The challenge addressed by RMPOD is when the
mass distribution is not dimensionally compatible with the
correlation matrix R. A common reason for the mass matrix
to be larger than R is that the number of available sensors,
M, may be limited. For example the mass matrix produced
by a finite-element program is easily greater than 100 × 100.
Thus by mathematical necessity in order to create a cor-
relation matrix whose dimensions match the mass matrix
the experimenter needs at least 100 sensors. Likewise,
a continuous structure will always have more degrees of
freedom than sensors. RMPOD uses a reduced-order mass
matrix Mr of dimension M × M, such that the matrix
multiplication RMr is possible. Then the following problem
We used the modal coordinates to further evaluate the
decompositions. The modal coordinates are defined through
the transformation X = Ψ Q, where the jth column of the
modal matrix Ψ is ψ, from the EVP and Q is an ensem-
ble of modal coordinate time histories. Each row is a sam-
pled time history corresponding to the associated column in
Ψ . Then the modal coordinate accelerations are given by
Qa = Ψ −1 A (here A is the acceleration ensemble). The mag-
nitude of the FFT of the modal coordinate accelerations for
the second mode showed a single peak at 39.14 Hz. The
third modal coordinate acceleration had a maximum peak at
107.6 Hz and smaller peak at 39.14 Hz. This shows some
pollution [21] from the second mode into the third modal. A
similar phenomenon occurs for the fourth modal coordinate
acceleration, which had a maximum peak at 205.5 Hz, fol-

4
lowed by 39.14 Hz, and finally, 107.6 Hz. Despite this modal
pollution, the extracted mode shapes were strong approxi-
mations to linear normal modes, which is evident by MAC
values close to unity for these mode shapes. Those values
are 0.986, 0.852, and 0.912 for the second, third, and fourth
mode, respectively. The extracted modes shapes are shown
in comparison to the discretized analytically approximated
mode shapes in Figs. 2, 3, and 4. Figure 5 was included as
an example of a poor extraction.
Fig. 2: Top: Second mode shape extracted by RMPOD (◦)
plotted with the analytical approximation’s discretized mode
shape. Middle: Second mode, modal acceleration coordinate
from RMPOD. Bottom: FFT of the modal coordinate accel-
eration
SMOOTH ORTHOGONAL DECOMPOSITION
In the case of SOD two correlation matrices are created.
One is the displacement correlation matrix R, such that
5
Fig. 3: Top: Third mode shape extracted by RMPOD (◦)
plotted with the analytical approximation’s discretized mode
shape. Middle: Third mode, modal acceleration coordinate
from the RMPOD. Bottom: FFT of the modal coordinate
acceleration
XXT
R= and the other is the velocity correlation matrix
N
VV T
S= , where V is an ensemble of velocity measure-
N
ments. R and S must be of the same dimensions. Next,
R and S are used in the generalized eigenvalue problem
described by λRφφ = Sφφ√. The natural modal frequencies
are estimated as ωi = λi , i = 1, · · · M, and LMNs are
approximated by columns of Ψ = Φ −T , where Φ is a matrix
containing the eigenvectors of the generalized eigenvalue
problem. Derivations of the above ideas can be found
in [17, 18]. SOD has been shown to extract approximations
to LNMs and natural frequencies from simulated discrete
and continuous systems [18].
Chelidze et al. did extensive simulations comparing
Fig. 4: Top: Fourth mode shape extracted by RMPOD (◦)
plotted with the analytical approximation’s discretized mode
shape. Middle: Fourth mode, modal acceleration coordinate
form RMPOD. Bottom: FFT of modal the acceleration coor-
dinates
SOD to POD [18]. Our work contributes to the field by first
using experimental data, and second by using a nonuniform
beam.
The results of applying SOD to the free vibrations of the
nonuniform beam of this study are included next. The natu-
ral frequencies estimated by SOD were 43.72 Hz, 107.77 Hz,
and 203.53 Hz. The extracted mode shapes had MAC values
of 0.999, 0.820, and 0.937. From these it is suggested that
the SOD can extract the lower modes of a lightly damped
nonuniform beam. Figures 6, 7, and 8 show the SOD ex-
tracted modes for the 2nd, 3rd, and 4th modes respectively.
These modes are plotted with the analytical approximations
of a nonuniform Euler-Bernoulli beam. This research shows
that SOD can be used to extract modal information from
a lightly damped freely vibrating nonuniform cantilevered
6
Fig. 5: Seventh mode shape extracted by RMPOD of instruc-
tive purpose as an example of poor extraction
beam.
STATE VARIABLE MODAL DECOMPOSITION
In SVMD [22] the outputs of the freely vibrating, lightly
damped beam were used to estimate the mode shapes, nat-
ural frequencies, and in some cases modal damping of
the beam. When applying SVMD one must first create a
state-variable ensemble matrix, Y = [VT XT ]T , where V
is the velocity ensemble matrix and Xis the displacement
ensemble matrix [22]. As such Y = [y(t1 ) y(t2 ) · · · y(tN )],
where y(t) = [ẋ1 (t), · · · ẋM (t); x1 (t), · · · xM (t)]T . A
YYT
2M × 2M correlation matrix is created such that R = ,
N
and unique to SVMD, a second 2M × 2M nonsymmetric
YWT
correlation matrix is created, N, such that N = , where
N
W = [A V ] = [ẏ(t1 ) · · · ẏ(tN )].
T T T
Fig. 6: Top: Second mode shape extracted by SOD (◦)
plotted with the analytical approximation’s discretized mode
shape. Middle: Second mode, modal acceleration coordinate
form RMPOD. Bottom: FFTs of modal coordinates
Once the two correlation matrices are computed then
an eigenvalue problem is cast as λRφφ = Nφφ. This problem
can be solved for 2M eigenvalues λ and eigenvectors φ .
If this eigensystem is solved in MATLAB using the eig
command, it produces two matrices Λ and Φ in matrix form
ΦΛ = SΦ
satisfying, RΦ Φ. The eigenvalue matrix Λ is diagonal
and contains information about the natural frequencies and,
in theory, modal damping. The real part of eigenvalue λi
indicates the exponential decay rate of the mode, while the
imaginary part represents the damped modal frequency.
The eigenvector matrix Φ contains modal information
but the inverse transpose of this matrix must be taken to
extract the mode shapes [22]. So the matrix of eigenvectors
is Ψ = Φ −T and each 2M × 1 column of Ψ contains
information about the mode shapes of the beam; the bottom
M × 1 rows will contain the mode shapes from the displace-
ments and approximate LNMs. These mode shapes may
be complex. If damping is approximately Caughey [23]
7
Fig. 7: Top: Third mode shape extracted by SOD (◦) plotted
with the analytical approximation’s discretized mode shape.
Middle: Third mode, modal acceleration coordinate form
RMPOD. Bottom: FFTs of modal coordinates
or Rayleigh (proportional) then the real parts dominate the
complex modes and approximately correspond to the LNMs.
Using SVMD on the experimental beam, approxima-
tions to LNMs were extracted as as shown in Figs 9, 10, and
11 below. The extracted damped modal frequencies from the
imaginary parts of the eigenvalues were 40.08 Hz, 106.42
Hz, and 205.08 Hz for the second, third, and forth mode,
respectively. The MAC values for these modes when com-
pared to the analytical approximation were 0.9921, 0.9729,
and 0.9865 for the second, third, and fourth mode respec-
tively.
DAMPING RATIO ESTIMATION
In the limit of modal (Caughey) damping, the eigenval-
ues associated with the underdamped modes are expected
to be complex in the form λi = −ζi ωi ± ωd ı̇. Where ζi is
the$damping ratio, ωi is the circular frequency, and ωdi =
ωi 1 − ζ2i is the damped circular frequency. Noting that

Fig. 8: Top: Fourth mode shape extracted by SOD (◦) plotted
with the analytical approximation’s discretized mode shape.
Middle: Fourth mode, modal acceleration coordinate form
RMPOD. Bottom: FFTs of modal coordinates
Fig. 9: Top: Second mode shape extracted by SVMD (◦)
plotted with the analytical approximation’s discretized mode
shape (-). Middle: Second modal coordinate displacement
from SVMD. Bottom: FFT of modal coordinate displace-
ment

Mode ζSV MD ζlogdec Table 3 for the second through six modes. The second col-
umn shows the SVMD extracted damping ratio and the third
2 0.0130 0.0126 column shows the damping ratio computed from the SVMD
3 0.0036 0.0116 modal coordinates using logarithmic decrement on all the
local maxima of the modal coordinate. The second mode
4 0.0106 0.0064 shows great agreement between the SVMD extracted damp-
5 -0.0022 0.003 ing and the logarithmic decrement. The agreement between
the two values deteriorate after the second mode. A simi-
6 0.0141 0.013 lar trend was mentioned in [17], and based on insight gained
from this paper attempts were made to extract better damping
Table 3: SVMD extracted damping ratios and damping ra- estimates.
tios computed from the log decrement of the SVMD modal
coordinates, using all peaks in the modal coordinates
IMPROVING DAMPING ESTIMATE
SIGNAL SEGMENTS
%
|λi | = λi λi = ωi and λi + λi = 2ζi ωi , where the overbar In some situations, when damped structures undergo
indicates a complex conjugate, then free vibrations higher frequency modes decay faster than
lower frequency modes. If a time segment of the free vibra-
tion signal was chosen based on the settling time, Ts , of the
λi + λi highest frequency of interest, for example the fourth mode,
ζi = (6)
2|λi | it maybe possible to improve the extraction of modal damp-
ing using SVMD. The damping estimates in Table 3 used
the time segment t ∈ [1.1142 2.2000] seconds and all the
The modal damping ratios extracted by SVMD are in included peaks in modal coordinates were used to compute

8
Fig. 10: Top: Third mode shape extracted by SVMD (◦)
plotted with the analytical approximation’s discretized mode
shape (-). Middle: Third modal coordinate displacement
form SVMD. Bottom: Fast Fourier Transform of modal co-
ordinate displacement
the logarithmic decrement. Figure 12 shows the modal co-
ordinate in this time window. The settling time computed
using Ts = 4/ζω where ζ and ω and are taken from the
initial SVMD estimates. In this calculation ζ4 = 0.01 and
ω4 = 1275.66 rad/s which results in Ts4 = 0.2929 seconds.
Using this as a guide for capturing enough cycles of the lower
modes for SVMD a time range of t ∈ [1.1142 1.5330] sec-
onds was used.
Upon close inspection of the modal coordinates it was
noticed that the first few peaks showed some transient distor-
tions, which may be a result of the filtering. As such the first
couple of peaks are excluded in future logarithmic decrement
calculations. In the long time window, t ∈ [1.1142 2.2000]
seconds, peaks 3 and 40 were used in computing the loga-
rithmic decrement of the second mode, peaks 3 and 40, and
3 and 60 were used for the third and fourth modes respec-
tively. The reason for the difference is that the second mode
has a lower frequency and therefore fewer cycles this can
be seen in Fig. 12 and the resulting logarithmic decrement
calculations can be seen in Table 4 for t ∈ [1.1142 2.2000]
seconds and Table 5 for t ∈ [1.1142 1.5330] seconds.
The shorter time segment showed improved damping ra-
tio extraction. Next, proportional modal damping in addition
9
Fig. 11: Top: Fourth mode shape extracted by SVMD (◦)
plotted with the analytical approximation’s discretized mode
shape (-). Middle: Fourth modal coordinate displacement
from SVMD. Bottom: FFT of modal coordinate displace-
ment
to the shorter time segment was considered as a means for
better modal damping extraction.
PROPORTIONAL DAMPING
If the damping is assumed to be proportional then the
damping matrix is C = αM M + βK
K . Then the modal damp-
ing coefficients are 2ζi ωi = α + βΛ
Λ. Three cases were ex-
amined: β = 0, α = 0 and least squares solution of α and
β using SVMD extracted data from Table 3. The result-
ing modal damping extractions are shown in Table 4 for
t ∈ [1.1142 2.2000] seconds, Table 5 for t ∈ [1.1142 1.5330]
seconds and Figs. 13, 14 and 15.
CONCLUSION
We applied three output-only modal analysis decom-
position methods to a nonuniform beam. The fast Fourier
transform was applied to the beam’s accelerations to identity
the modal frequencies. The first mode was filtered out
since it was below the range of reliable accelerometer
performance. The beam was modeled as a nonuniform
Euler-Bernoulli beam. An analytical approximation of the
 Fig. 13: Top: SVMD modal coordinates for the second
mode 1.1142 ≤ t ≤ 2.2000 (solid curve), standard SVMD
damping(!), logarithmic damping (◦) for peak 3 to peak
20, least square solution to 2ωi ζi = α + βΛ Λ damping (*)
Bottom: SVMD modal coordinates for the second mode
1.1142 ≤ t ≤ 1.5330 (solid curve), standard SVMD damp-
ing (!), logarithmic damping (◦) for peak 3 to peak 14, least
square solution to 2ωi ζi = αII + βΛ
Λ damping (*)

Fig. 12: SVMD modal coordinates plotted with maximum

peaks (◦), and boundaries for log decrement calculations
(• ←). Time segment 1.1142 ≤ t ≤ 2.2000 Top: Second
mode, Middle: Third mode, Bottom: Fourth mode

1.1142 ≤ t ≤ 2.2000
Mode ωSVMD ζSVMD Log Dec (n)
2 251.45 0.0130 0.0122 (20)
3 644.96 0.0036 0.0126 (40)
4 1275.66 0.0106 0.0063 (60)

Table 4: SVMD modal damping long time segment esti-

mates. 2nd column: Log decrement was computed using
peaks [3,(n)].

beam was created and its mode shapes were determined and
used to predict the natural frequencies of experimental beam
and showed agreement with the FFT frequencies.
The reduced-order mass-weighted POD was applied
using a permutation of conditions involving impulse lo-
cation, impulse amplitudes, and ensembles created from
displacement, velocity and acceleration signals. RMPOD
Fig. 14: Top: SVMD modal coordinates for the third mode
1.1142 ≤ t ≤ 2.2000 (solid curve), standard SVMD damp-
ing (!), logarithmic damping (◦) for peak 3 to peak 60, least
square solution to 2ωi ζi = αI + βΛΛ damping (*), Bottom:
SVMD modal coordinates for the third mode 1.1142 ≤ t ≤
1.5330 (solid curve), standard SVMD damping (!), logarith-
mic damping (◦) for peak 3 to peak 40, least square solution
to 2ωi ζi = αII + βΛ
Λ damping (*)

10

Fig. 15: Top: SVMD modal coordinates for the fourth mode
1.1142 ≤ t ≤ 2.2000 (solid curve), standard SVMD damp-
ing (!), logarithmic damping (◦) for peak 3 to peak 60, least
square solution to 2ωi ζi = αI + βΛΛ damping (*) Bottom:
SVMD modal coordinates for the fourth mode 1.1142 ≤ t ≤
1.5330 (solid curve), standard SVMD damping (!), logarith-
mic damping (◦) for peak 3 to peak 40, least square solution
to 2ωi ζi = αII + βΛ
Λ damping (*)
1.1142 ≤ t ≤ 1.5330
Mode ωSVMD ζSVMD Log Dec(n)
2 253.34 0.0058 0.0123 (14)
3 663.17 0.0097 0.0113 (40)
4 1275.53 0.0030 0.0037 (40)
Table 5: SVMD modal damping short time segment esti-
mates. Logarithmic decrement was computed using peaks
[3,(n)].
extracted approximations to the second, third, and fourth
LNMs suggested by MAC values of 0.986, 0.852, and
0.912 between extracted, modes and analytical modes of
the second, third, and fourth modes, respectively. Further
confirmation on the quality of the modes was provided from
computing the modal coordinates and taking their FFTs. The
peak frequency for the lowest extracted mode was dominant.
For increasingly higher modal coordinates, frequencies of
other modes leaked in. Apparently, the pollution of these
modes did not greatly affect the approximation to the LNMs.
SOD and SVMD were also used on the same data to
extract the modal frequencies and approximations to the
LNMs. The SOD extracted frequencies agreed with the
FFT of the experimental beam. SOD extracted mode shapes
agreed with the analytical approximation. Similar agreement
11
1.1142 ≤ t ≤ 2.2000 α = 2.53 β = 1.45 × 10−6
Mode SVMDζ
2 0.0053
3 0.0025
4 0.0019
Table 6: SVMD modal damping long time least square pro-
portional estimates
1.1142 ≤ t ≤ 1.5330 α = 6.946 β = 1.23 × 10−6
Mode SVMDζ
2 0.0139
3 0.0056
4 0.0035
Table 7: SVMD modal damping short time least square pro-
portional estimates
was found when SVMD was applied.
SVMD was able to extract the modal damping ratio
directly without the use of modal coordinates. SVMD
extraction of the lower modes’ modal damping was better
with longer time signals. This could be a result of the
higher modes decay faster and as time increases the signal is
dominated by the lower modes. Likewise with short signal
segment near the time of impact SVMD’s ability to extract
high modes improved. With similar reasoning at the shorter
signal near the impact are not as dominated by the lower
frequencies modes Assuming proportional damping should
best performance with taking using α and β determined
from the least squares solution.
These tests suggest that RMPOD, SOD, and SVMD
can be reliable methods of modal identification, at least for
the lower modes of a structure and are easy to implement.
The only signal processing needed is in integrating the ac-
celerometer signals into the desired states (displacement, ve-
locity, or acceleration), and high pass filtering used to pre-
vent integrator drift. The relative performance of these three
methods cannot determined from the results presented in this
paper. Several impulses was used to excited the beam at a
variety of locations and only one permutation of those inputs
are reported. The data set reported in this paper optimized
the results of RMPOD only.When these excitation parame-
ters are applied to other data sets, the methods seem to be
equally reliable and in the absence of an analytical approxi-
mation, the methods could be used in concert to cross check
results. Table 8 summarizes some of the benefits and draw-
backs of each method. However, the drawbacks of each de-
composition method are not serious.
 [2] Juang, J., and Pappa, R., 1985. “An eigensystem re-
PROS and CONS
alization algorithm for modal parameter identification
and model reduction”. Journal of Guidance, Control
RMPOD
and Dynamics, 8 (5), p. 620627.
can estimate mode shapes frequencies not directly es- [3] Ibrahim, S., and Mikulcik, E., 1977. “A method for
timated (need Q ) the direct identification of vibration parameters from
the free response”. Shock and Vibration Bulletin, 47
RMPOVs estimate modal need to compute the re- (4), pp. 183–198.
strength duced mass matrix [4] Kerschen, G., Poncelet, F., and Golinval, J., 2007.
requires single R “Physical interpretation of independent component
analysis in structural dynamics”. Mechanical Systems
requires X only and Signal Processing, 21 (4), p. 15611575.
input signal not needed [5] Poncelet, F., Kerschen, G., Golinva, J., and Verhelst,
D., 2007. “Output-only modal analysis using blind
SVMD source separation techniques.”. Mechanical Systems
and Signal Processing, 21, p. pp. 23352358.
can estimate mode shapes no modal strength, except [6] Vold, H., Kundrat, J., Rocklin, G., and Russel, R., 1982.
by Q “A multi-input modal estimation algorithm for mini-
computer”. SAE Technical Papers Series, No 820194,
estimate modal frequencies need X , V , and A 91, pp. 815–821.
directly [7] Vold, H., 1986. “Orthogonal polynomials in the
possibility of modal damp- polyreference method”. In Katholieke University of
ing directly Leuven, Belgium, Katholieke University of Leuven,
Belgium.
mass not required [8] Richardson, M., and Formenti, D. L., 1982. “Parame-
input signal not needed ter estimation from frequency response measurements
using rational fraction polynomials”. In Proceedings of
SOD the International Modal Analysis Conference, pp. 167–
182.
can estimate mode shapes no modal strength except [9] Shih, C., Tsuei, Y., Allemang, R., and Brown, D., 1988.
by Q “Complex mode indication function and its application
to spatial domain parameter estimation.”. Mechanical
estimate modal frequency need X and V
System and Signal Processing, 2, pp. 367–377.
directly
[10] Brincker, R., Zhang, L., and Andersen, P., 2001.
mass not required “Modal identification of output-only systems using fre-
quency domain decomposition.”. Smart Materials and
input signal not needed
Structures, 10, p. 441445.
[11] Feeny, B. F., and Kappagantu, R., 1998. “On the physi-
Table 8: Pros and cons of each decomposition method cal interpretation of proper orthogonal modes in vibra-
tions”. Journal of Sound and Vibration, 211(4), pp. 607
– 616.
ACKNOWLEDGMENTS [12] Feeny, B. F., 2002. “On the proper orthogonal modes
This work was supported by the National Science and normal modes of continuous vibration systems”.
Foundation grant numbers CMMI-0943219 and CMMI- Journal of Vibration and Acoustics, 124(1), pp. 157–
0727838. Any opinions, findings, and conclusions or 160.
recommendations are those of the authors and do not nec- [13] Kerschen, G., and Golinval, J. C., 2002. “Physical in-
essarily reflect the views of the National Science Foundation. terpretation of the proper orthogonal modes using the
singular value decomposition”. Journal of Sound and
Additional support was received from the Diversity Pro- Vibration, 249(5), pp. 849–865.
grams Office and the College of Engineering at Michigan [14] Riaz, M. S., and Feeny, B. F., 2003. “Proper orthogo-
State University. nal decomposition of a beam sensed with strain gages”.
Journal of Vibration and Acoustics, 125((1)), pp. 129–
131.
References [15] Han, S., and Feeny, B. F., 2003. “Application of
[1] Farooq, U., and Feeny, B. F., 2008. “Smooth orthogo- proper orthogonal decomposition to structural vibration
nal decomposition for modal analysis of randomly ex- analysis”. Mechanical Systems and Signal Processing,
cited systems”. Journal of Sound and Vibration, 316(1- 17((5)), pp. 989–1001.
5), pp. 137 – 146. [16] Iemma, U., Morino, L., and Diez, M., 2006. “Digi-

12
 tal holography and Karhunen-Loeve decomposition for
the modal analysis of two-dimensional vibrating struc-
tures”. Journal of Sound and Vibration, 291, pp. 107–
131.
[17] Farooq U., Feeny, B. F., to appear. “An experimental
investigation of a state-variable modal decomposition
method for modal analysis”. Journal of Vibrations and
Acoustics.
[18] Chelidze, D., and Zhou, W., 2006. “Smooth orthog-
onal decomposition-based vibration mode identifica-
tion”. Journal of Sound and Vibration, 292, pp. 461–
473.
[19] Allemang, R., 2003. “The modal assurance criterion
(MAC): Twenty years of use and abuse”. Sound and
Vibration, August, pp. 14–21.
[20] Yadalam, V. K., and Feeny, B. F., 2011. “Reduced mass
weighted proper decomposition for modal analysis”.
Journal of Vibration and Acoustics, 133(2), p. 024504.
[21] Feeny, B. F., 2002. “On proper orthogonal coordinates
as indicators of modal activity”. Journal of Sound and
Vibration, 255(5), pp. 805–817.
[22] Feeny, B. F., and Farooq, U., 2008. “A nonsymmetric
state-variable decomposition for modal analysis”. Jour-
nal of Sound and Vibration, 310(4-5), March, pp. 792–
800.
[23] Caughey, T. K., 1960. “Classical normal modes in
damped linear systems.”. Journal of Applied Mechan-
ics, 27, pp. 269–271. Transactions of the ASME 82,
series E.

13