Operational Modal Analysis –

Another Way of Doing Modal Testing

Mehdi Batel, Brüel & Kjær, Norcross, Georgia

Operational modal analysis (often called output-only or Combined Ambient System

ambient modal analysis) is described in this article. Modal
testing is performed on a plate structure with well-defined Stationary
Zero Mean Loading Structural System
modes, resonance frequencies and damping values. Frequency Gaussian System (Linear, Time-Invariant)
Responses
Domain Decomposition (FDD) and Enhanced Frequency Do- White Noise
main Decomposition (EFDD) concepts are presented and ap-
plied to a plate structure. This article details the signal pro- Unknown Excitation Forces
cessing mathematical background and presents alternative
curve-fitting processes. Figure 1. Combined ambient model.

An alternative modal analysis technique is presented in this

article. A typical modal test of a structure is performed by
measuring the input forces and output responses for a linear,
time-invariant mechanical system. The excitation is either tran-
sient (impact hammer testing), random, burst-random or sinu-
soidal (shaker testing). The advanced signal processing tools
used in operational modal analysis techniques allow the inher-
ent properties of a mechanical structure (resonance frequen-
cies, damping ratios, mode patterns) to be determined by only
measuring the response of the structure without using an arti-
ficial excitation. This technique has been successfully used in
civil engineering structures (buildings, bridges, platforms, tow-
ers) where the natural excitation of the wind is used to extract
modal parameters. 1, 2, 3 It is now being applied to mechanical
and aerospace engineering applications (rotating machinery,
on-road testing, in-flight testing). 4,5,6
The advantage of this technique is that a modal model can
be generated while the structure is under operating conditions.
Figure 2. PULSE ™ multi-analysis portable system connected to test
That is, a model within true boundary conditions and actual plate; hand-tapping excitation is demonstrated.
force and vibration levels. Another advantage of the technique
is the ability to perform modal testing in-situ, i.e., without re- and modal extraction procedures were performed.
moving parts under test. The test can be performed with other
applications or activities in parallel and does not affect or in- Measurement Procedure
terrupt daily use of the machine. The measurement technique The specimen used is a rectangular plate (29 cm´25 cm) rest-
is identical to Operational Deflection Shape (ODS) measure- ing on a foam pad to simulate free-free boundary conditions.
ment procedures where one accelerometer is used as a refer- The measurements were made using 4 accelerometers (1 for the
ence and a series of accelerometers for the responses at all De- reference and 3 roving accelerometers for the responses at 36
grees of Freedom (DOFs) of interest. DOFs). The data acquisition system was a portable PULSE™
Figure 1 shows a schematic description of an ambient re- analysis platform (see Figure 2), composed of a 4-channel por-
sponse system. The inputs to the system (that represent the table front-end (4 inputs/2 outputs) and a laptop computer for
excitation forces) are assumed to have a Gaussian amplitude the software.
distribution. In civil engineering applications, Gaussian exci- Hand tapping the plate for each set of measurements pro-
tation is typically provided by waves (offshore applications), vided enough energy to the structure. The PULSE Modal Test
wind or traffic load. In mechanical engineering, loads are typi- Consultant™ was used to set-up the hardware, create the ge-
cally generated from bearings, vibration from the road or the ometry and assign the measurements to each DOF. The refer-
air, rotating components or the engine. To define all modes, the ence accelerometer was maintained at a well-chosen point on
excitation should be broadband. Practically, this may involve the plate. The reference point selection has a significant effect
running up an engine or modifying the frequency of excitation. on measurement results. It has to be placed such that all modes
In cases where the loading forces cannot be modified, the end- contribute to the reference accelerometer. A preliminary idea
user needs to have at least some idea about the excitation fre- of the mode shapes to be measured definitely helps in under-
quencies that are generated to interpret the results and use the standing where to place the reference points. Typically, points
appropriate modal extraction method. For nonbroadband ex- that are not nodal or peak deflection points are good choices
citation, modal extraction becomes difficult and would result (atypical degrees of freedom). Each data set is then composed
in poor model descriptions. The art of operational modal analy- of the reference accelerometer signal and the 3 accelerometers
sis is then to distinguish real structural behavior from noise measuring the responses at the specified DOFs. Twelve data
and other measured sources. sets were then collected for the plate. The raw time histories
In this study, measurements were made with a Brüel & Kjær were captured by a “Time Capture Analyzer” for each measure-
PULSE™ Multi-Analyzer System and the Modal Test Consult- ment set. A pretest measurement indicated that the lowest fre-
ant ™ (Type 7753) to create the test structure geometry, assign quency of interest was about 350 Hz. In that case only 2 sec of
measurement points and capture the data. The analysis was data capture would be enough to represent more than 500
performed using the Brüel & Kjær Operational Modal Analy- cycles at the lowest frequency of interest. A sample data set is
sis ™ software (Type 7760) where advanced signal processing shown in Figure 4. This figure also shows a Short Time Fou-

22 SOUND AND VIBRATION/AUGUST 2002

 Gyy ( jω ) 12

Gyy ( jω ) 11

Gyy ( jω ) 3

Gyy ( jω ) 2 Measurements

Gyy ( jω ) 1

Figure 3. Two views of the plate geometry.

Figure 5. Cross spectral density matrix elements between the 4 re-
sponses.

Figure 6. Magnitude of the spectral density between the reference point

and response 4 for dataset 10.

overlap of the windowed segments before averaging them to-

Figure 4. Sample output from the “Time Capture Analyzer” for measure-
ment 10. gether. This technique minimizes spectral noise and the effects
of other artifacts. Using the averaged spectrum for frequency
rier Transform (STFT) analysis that provides a time/frequency peak-picking reduces possible misinterpretation of spectral
representation of all the responses captured. components.
The STFT analysis is performed by a traveling Fast Fourier The spectral density matrices are then calculated for all the
Transform (FFT) window with user defined parameters. Fig- series of measurements. The size of the matrix is n´n, n being
ure 4 exhibits straight lines showing up at specific frequencies the number of transducers (4 in this case, i.e., 4 measured
during the entire capture that correspond to structural reso- DOFs). In this example, 12 matrices (of a size 4´4) were calcu-
nances. lated for each frequency. Each element of those matrices is a
All the raw time data, the geometry and the series of mea- spectral density function. The diagonal elements of the matrix
surements are then directly exported from the data acquisition are the magnitudes of the spectral densities between a response
system to the operational modal analysis curve-fitter for sig- and itself (power spectral densities). The off-diagonal elements
nal processing calculations and modal extraction. are the cross spectral densities between the 4 responses (Fig-
ure 5). All those matrices are Hermitian (symmetric with com-
Signal Processing and Decomposition plex conjugate elements around the diagonal).
Preliminary Signal Processing. The first step of the analy- Each matrix is expressed in terms of power and cross spec-
sis is to perform a Discrete Fourier Transform (DFT) on the raw tral densities as follows:
time data to obtain acceleration spectral density matrices that
will contain all the modal information. Since the excitation is  PSD11( jω ) CSD12( jω ) CSD13( jω ) CSD14( jω ) 
CSD ( jω ) PSD22( jω ) CSD23( jω ) CSD24( jω ) 
broadband and has a continuous spectrum, the best spectral [Gyy ( jω )]i = 
21
(1)
descriptor is the acceleration spectral density (g2/Hz) that nor- CSD31( jω ) CSD32( jω ) PSD33( jω ) CSD34( jω ) 
 
malizes the measurements with respect to the bandwidth of the CSD41( jω ) CSD42( jω ) CSD43( jω ) PSD44( jω )i
frequency analysis filter (FFT).
The analysis is performed by specifying the order of deci- PSD(jw ) denotes the power spectral density (magnitude of the
mation (fraction of the original sampling frequency) and the auto spectral density) and CSD(jw ) denotes the cross spectral
number of spectral lines for the Fourier analysis. The software density. Since the matrices calculated are Hermitian we have
has the capability of applying a filter (bandpass, bandstop,
CSDpq ( jω ) = CSDqp*( jω ), p ≠ q (2)
highpass or lowpass) on the data to remove unwanted compo-
nents that may obscure any curvefitting process in the analy- The ‘*’ symbol denotes a complex conjugate value. The PSD pq
sis. No decimation process was chosen since an appropriate (jw ) are all real valued elements, and the CSD qp(jw ) take com-
frequency range was already based on the pretest (a Nyquist plex values, carrying the phase information between the mea-
frequency of 4096 Hz and a sampling frequency of 8192 Hz). surement and the reference degree of freedom.
The spectral estimation was performed using the modified av- Figure 6 shows the result obtained of the spectral density
eraged periodogram method (Welch’s technique) with an over- calculation between response 4 and the reference accelerom-
lap of 66.7% and a Hanning weighting function. This ensures eter at measurement 10 (cross spectral density).
that all data are equally weighted in the averaging process, Frequency Domain Decomposition Theory Background. Fre-
minimizing leakage and picket fence effects. The Welch quency Domain Decomposition (FDD) is an extension of the
method performs a splitting of the time series and then an Basic Frequency Domain (BFD) technique or more often called

SOUND AND VIBRATION/AUGUST 2002 23

the peak-picking technique. This approach uses the fact that The contribution of the residue has the following expression:
modes can be estimated from the spectral densities calculated
with the assumption of white noise input and a lightly damped [Rk ]C[Rk ]H
[Ak ] = (12)
structure. This nonparametric technique estimates modal pa- 2σ k
rameters directly from signal processing calculations. Considering a lightly damped model, we have the following
The FDD technique estimates the modes using a Singular relationship:
Value Decomposition (SVD) of each of the spectral density
matrices. This decomposition corresponds to a Single Degree lim [Ak ] = [Rk ]C[Rk ]T = ψ k γ kT C γ kψ kT = dkψ kψ kT (13)
damping →light
of Freedom (SDOF) identification of the system for each sin-
gular value. where d k is a scalar constant.
The relationship between the input x(t) and the output y(t) The contribution of the modes at a particular frequency is
can be written in the following form: 7 limited to a finite number (usually 1 or 2). The response spec-
[Gyy ( jω )] = [H ( jω )]*[Gxx ( jω )][H ( jω )]T (3) tral density matrix can then be written as the following final
form:
dkψ kψ H dk*ψ k*ψ k H
where Gxx(j w ) is the input power spectral density matrix that [Gyy ( jω )] = ∑ + (14)
is constant in the case of a stationary zero mean white noise k ∈Sub(ω ) jω − λk jω − λ k *
input. This constant will be called C in the rest of the math-
ematical derivation. Gyy(jw ) is the output PSD matrix and H(jw ) where Sub(w ) is the set of modes that contribute at the particu-
is the frequency response function (FRF) matrix. As seen in lar frequency.
Equation (3), the output G yy will be highly sensitive to the in- This final form of the matrix is then decomposed into a set
put constant C. The rest of the equation derivations and single of singular values and singular vectors using the Singular Value
degree of freedom identification will provide relevant results, Decomposition technique (SVD). This decomposition is per-
only by assuming that the input is effectively represented by a formed to identify single degree of freedom models of the prob-
constant value (mean Gaussian). It is therefore important to lem.
realize how this input assumption will be crucial to the tech- Singular Value Decomposition. The singular value decom-
nique. position of an m´n complex matrix A is the following factor-
The FRF matrix can be written in a typical partial fraction ization:
form (used in classical modal analysis), in terms of poles and A = U ΣV H (15)
residues
[Y (ω )] m
[Rk ] [Rk ]* where U and V are unitary and S is a diagonal matrix that con-
[H ( jω )] = = ∑ + (4) tains the real singular values.
[H (ω )] k =1 jω − λk jω − λk*

with Σ = diag (s1,......., sr ) (16)

λk = −σ k + jωdk (5) r = min(m, n) (17)

m being the total number of modes, lk being the pole of the kth The superscript H on the matrix V denotes a Hermitian trans-
mode, s k the modal damping and w dk the damped natural fre- formation (transpose and complex conjugate). In the case of real
quency of the kth mode: valued matrices, the V matrix is only transposed. The s i ele-
ments in the matrix S are called the singular values and their
ωdk = ω0k 1 − ς k 2 (6) following singular vectors are contained in the matrices U and
with V.
σk
ςk = (7) This singular value decomposition is performed for each of
ω 0k
the matrices at each frequency and for each measurement (Fig-
zk being the critical damping and w 0k the undamped natural ure 7). The spectral density matrix is then approximated to the
frequency, both for mode k. following expression after SVD decomposition:
[R k ] is called the residue matrix and is expressed in an outer
product form: [Gyy ( jω )] = [Φ][Σ][Φ]H (18)

[Rk ] = ψ k γ k T (8) with [Φ]H [Φ] = [I ] (19)

where y k is the mode shape and gk is the modal participation S being the singular value matrix and F the singular vectors
vector. All those parameters are specified for the k th mode. unitary matrix:
The transfer function matrix [H] is symmetric and an element s1 0 0 . . 0
H pq(jw ) of this matrix is then written in terms of the compo- 0 s2 0 . . . 
nent r kpq(jw ) of the residue matrix as follows: 
0 . s3 . . .
m
r ( p, q) rk ( p, q)* [Σ] = diag(s1,......., sr ) =   (20)
H pq ( jω ) = ∑ jωk − λ +
jω − λ k *
(9) . . . . . 0
k =1 k . . . . sr 0
 
Using expression (3) for the matrix Gyy and the heaviside  0 . . 0 0 0
partial fraction theorem for polynomial expansions, we obtain
the following expression for the matrix output PSD matrix G:
[Φ ] = [{ϕ1 } {ϕ2 } {ϕ3 } . . .{ϕ r }] (21)
m
[Ak ] [Ak ]* [Bk ] [Bk ]*
[Gyy ( jω )] = ∑ + + (10) The number of nonzero elements in the diagonal of the singu-
k = 1 jω − λ k jω − λ k − jω − λ k − jω − λ k *
*
lar matrix corresponds to the rank of each spectral density
matrix. The singular vectors correspond to an estimation of the
where [A k ] is the kth residue matrix of the matrix [Gyy]. The mode shapes and the corresponding singular values are the
matrix Gxx is assumed to be a constant value C, since the exci- spectral densities of the SDOF system expressed in Equation
tation signals are assumed to be uncorrelated zero mean white (14).
noise in all the measured DOFs. This matrix is Hermitian and Figure 8 shows the result of the singular value decomposi-
is described in the form: tion of the spectral density matrix at measurement 10. We ob-
m
[Rs ]H [Rs ]T tain 4 singular values and 4 singular vectors for each of the
[Ak ] = [Rk ]C ∑ + (11) spectral density matrices. The singular values and their corre-
s =1 −λk − λs
* −λk − λs
sponding singular vectors are ranked in singular value de-

24 SOUND AND VIBRATION/AUGUST 2002

 Gyy (jω 1) Gyy (jω n)
Meas. 12 Meas. 12

SVD SVD

Figure 9. Peak-picking of the average normalized singular values of the

Gyy (jω 1) Gyy (jω n) complete PSD matrix.
Meas. 1 Meas. 1

SVD SVD

Figure 7. Singular value decomposition procedure for the spectral den-

sity matrix at each frequency.

Figure 8. Result of the SVD for measurement 10.

scending order for each of the spectral density matrices, mean-

ing that the first singular value will be the largest.
The green curve shows the first singular value and the other
curves show the 3 remaining singular values. There are as many Figure 10. Three dimensional renderings of plate deformation patterns.
singular values as measured DOFs (4 in this case) for a particu-
lar measurement (measurement 10 in the figure). This tech- is important to note that the deformations obtained are not true
nique allows us to identify possible coupled modes that are mode shapes. The residues obtained in the mathematical deri-
often indiscernible as they appear on the spectral density func- vation are not scaled to the input force and therefore will not
tions. If only one mode is dominating at a particular frequency, provide scaled shape vectors.
then only one singular value will be dominating at this fre- It is also possible to obtain damping characteristics of each
quency. In the case of close or repeated modes, there will be mode and more precise resonance frequencies by using the
as many dominating singular values as there are close or re- enhanced frequency domain decomposition based on the de-
peated modes. termination of the correlation functions.
Frequency Domain Decomposition – Peak-Picking and Mode Enhanced Frequency Domain Decomposition (EFDD). The
Determination. Each of the SDOF systems obtained by the sin- enhanced FDD technique allows the resonance frequency and
gular value decomposition allows us to identify the natural damping of a particular mode to be extracted by computing the
frequency and mode shape (unscaled) at a particular peak. auto- and cross-correlation functions. The SDOF power spec-
Using the operational modal analysis software, we perform the tral density function identified around a resonance peak is
peak-picking technique (similar to the quadrature-picking in returned to the time domain using the Inverse Discrete Fou-
classical modal analysis) for each resonance on the average of rier Transform (IDFT). The resonance frequency is obtained by
the normalized singular values for all data sets (see Figure 9). determining zero crossing times and damping by the logarith-
The FDD technique provided the resonance frequencies and the mic decrement of the corresponding SDOF normalized auto
mode shapes listed in Table 1. The very well defined deforma- correlation function.
tion patterns were animated using the software and exhibit very The free-decay time domain function (the correlation func-
clear modal deformations (Figure 10). tion of the SDOF system) is used to estimate damping for mode
The frequency domain decomposition provided very good k:
results for the resonance frequencies and the mode shapes. It 2 r 
δ k = ln  0k  (22)
p  rpk 
Table 1. First six plate modes.
where r 0k is the initial value of the correlation function and
Mode Resonance Frequency
r pk is the p th extrema. The critical damping ratio for mode k is
1 – Torsional (1,1) . . . . . . . . . . . . . . . . . . . . . . . . . 352.0 Hz – FDD obtained with the formula:
2 – Flexural (2,0) . . . . . . . . . . . . . . . . . . . . . . . . . . 492.0 Hz – FDD
δk
3 – Flexural (0,2) . . . . . . . . . . . . . . . . . . . . . . . . . . 716.0 Hz – FDD ςk = (23)
4 – Torsional (2,1) . . . . . . . . . . . . . . . . . . . . . . . . . 868.0 Hz – FDD ξk + 4π 2
2
5 – Torsional (1,2) . . . . . . . . . . . . . . . . . . . . . . . . . 972.0 Hz – FDD
6 – Flexural (3,0) . . . . . . . . . . . . . . . . . . . . . . . . . 1424.0 Hz – FDD The damped natural frequency is obtained by linear regression
of the crossing times corresponding to the extrema of the cor-

SOUND AND VIBRATION/AUGUST 2002 25

Figure 11. Singular value spectral bell identification for measurement Figure 13. Damping ratio estimation from the decay curve of the corre-
10. lation function.

Figure 14. Natural frequency estimation.

Figure 12. Normalized correlation function for measurement 10.
algorithm. In that example, the modes are well spaced in the
relation function. The undamped natural frequency for mode frequency domain and will provide leakage-free correlation
k is then: fdk functions. In cases where frequency peaks are not clearly
f0k = (24) spaced, leakage will definitely affect the Inverse Fourier pro-
1 − ξk 2
cess since only a limited frequency range is used for the Fou-
Both parameters and an improved version of the mode shapes rier calculations.
are estimated from the SDOF Bell functions. The SDOF Bell Damping is estimated by the logarithmic decrement tech-
function is estimated using the mode determined by the pre- nique of Equation 22 and the logarithmic envelope of the cor-
vious FDD peak-picking operation. The latter is used as a ref- relation function. The estimation is performed with a linear re-
erence vector in a correlation analysis based on the Modal As- gression technique (red part of the curve in Figure 13). The
surance Criteria (MAC). A MAC value is computed between the resonance frequency is simply obtained by counting the num-
reference FDD vector and a singular vector for a particular fre- ber of times the correlation function crosses the zero axis (see
quency region. The MAC value describes the degree of corre- Figure 14). This gives more accurate results than the simple
lation between 2 modes (it takes a value between 0 and 1) and FFD peak picking method.
is defined as follows for 2 vectors f and y : The result of the regression is shown as a red line. It is im-
portant to note that the Bell functions and the estimation of the
2
{Φ}*{Ψ} (25) damping and natural frequency is performed for each set of
MAC({Φ},{Ψ}) = measurements. The result is then obtained by averaging all the
{Φ} . {Ψ}
data sets together. Both the average value and the standard
If the largest MAC value of this vector is above a user-speci- deviation of damping and natural frequency are calculated
fied MAC Rejection Level, the corresponding singular value is from the data sets.
included in the description of the SDOF Spectral Bell function. Validation between FDD and EFDD Technique. Table 2 pre-
The lower this MAC Rejection Level the larger the number of sents a comparison between the two techniques for the deter-
singular values included in the identification of the SDOF Bell mination of resonance frequencies. It also presents the damp-
function. A good compromise value for this rejection criteria ing ratios calculated for each mode in the case of the EFDD
is 0.9. An average value of the singular vector (weighted by the analysis. The Mode shapes are also compared using the soft-
singular values) is then obtained. ware and a MAC value is presented in the MAC matrix shown
Figure 11 shows the estimated SDOF Bell function with a in Figure 15. The MAC values are very close to 1 for all modes
MAC rejection level of 0.9. The value of the MAC rejection cri-
teria has to be chosen so that we obtain a good representation
Table 2. Comparison of FDD and EFDD techniques for determining
of the Bell function around the peak chosen and not include resonance frequencies.
any noise around it. Noise is often present in an ambient modal
analysis. EFDD
Using this SDOF Bell function, we perform an inverse Fou- Mode FDD (Hz) EFDD (Hz) Damping (%) MAC
rier transform for the determination of damping and natural (1,1) . . . . . . . . . . . . . . . . . 352 373.34 0.75 0.988
frequency. The obtained normalized correlation function is (2,0) . . . . . . . . . . . . . . . . . 492 486.80 0.64 0.986
shown in Figure 12. (0,2) . . . . . . . . . . . . . . . . . 716 712.10 0.41 0.998
(2,1) . . . . . . . . . . . . . . . . . 868 857.53 0.48 0.968
Figure 12 exhibits a typical response of a resonating system
(1,2) . . . . . . . . . . . . . . . . . 972 969.68 0.48 0.996
that decays exponentially. The scattered region indicates the (3,0) . . . . . . . . . . . . . . . . . 1424 1418.88 0.48 0.996
part of the correlation function that is used for the estimation

26 SOUND AND VIBRATION/AUGUST 2002

 The idea behind the SSI technique is to be able to represent
the system in Equation (28) in the frequency domain in terms
of a Transfer Function that involves the matrices A, C, K, and
the identity matrix. The eigenvalue decomposition of the ma-
trix A leads to a representation of the transfer function matrix
that contains the modal parameters (natural frequencies and
damping ratios). The mode shapes are extracted from the eigen-
vectors of the matrix A and the observation matrix C.
This technique can be used to compare and validate the re-
sults obtained by the nonparametric techniques (FDD and
EFDD). It is possible to calculate the MAC and confidence fac-
tor values for 2 modes calculated by the two techniques. Fur-
thermore, this comparison can be used to detect and differen-
tiate the actual structural resonances and the harmonic
components that often appear in rotating machinery analysis.
Comparing this technique with the nonparametric methods is
also very useful to determine the modes that actually corre-
spond to a physical solution of the problem related to a logi-
cal motion of the structure. Indeed the peak-picking technique
Figure 15. Mode shape comparison – MAC values. assumes that the peak detected and extracted is due to a modal
structural motion and not to excessively high operation forces.
revealing a strong correlation between the two modal extrac-
tion techniques. Conclusions
Further Investigations: SSI Parametric Technique. The FDD The operational modal analysis technique allows a scientist,
and EFDD techniques can be correlated with a parametric tech- technician or engineer to perform a modal investigation eas-
nique called the Stochastic Subspace Identification (SSI) tech- ily, quickly and accurately. It can be accomplished by only mea-
nique. 8,9,10 This technique has been widely used in the domain suring the response of the structure subjected to unknown and
of structural mechanics and allows the user to compare results unmeasured input forces (still assumed to be broadband and
obtained from both signal processing calculations and time- stationary).
domain model parameterizations. The SSI technique involves It is very important to realize that this technique provides
the use of statistics, optimal prediction, linear system theory operating deformation patterns and not scaled mode shapes (no
and stochastic processes. A brief description of the principles identification of modal masses). Modal analyses performed
of this technique will be made here. with known excitations will provide modes that are scaled to
The dynamic system is expressed in terms of inertial (mass), the input force and therefore can be utilized for other tech-
dissipative (damping) and restoring (stiffness) matrices. It is niques like structural dynamics modifications or frequency
written in terms of a linear set of differential equations of the response simulations. In order to obtain a full description of
type: all relative modal parameters, it is important to note that known
.. .
[M ]{x(t )} + [C ]{x(t )} + [K ]{x(t )} = { f (t )} (26) energy force levels (provided by a force transducer and a well-
defined excitation) have to be determined, regardless of the so-
By rewriting the equations of motion in a classical state-space phistication of signal processing techniques. Impact or shaker
formulation often used in modern control theory, the dynamic testing are still excellent ways to perform modal testing and
system is expressed as follows: remain the only modal techniques that will provide scaled
 xt +1 = [A]xt + wt mode shapes and modal masses for simulation purposes.

 y t = [C ]xt + v t (27)
References
where xt represents the Kalman sequences found by an orthogo- 1. Brincker, R., Zhang, L., and Andersen, P., “Modal Identification from
Ambient Response Using Frequency Domain Decomposition,” Proc.
nal-projection technique. The first equation (state equation) of the 18th International Modal Analysis Conference, San Antonio,
represents the dynamic behavior of the physical system and the TX, February 7-10, 2000.
second equation (observation equation) is called the output 2. Schwarz, Brian, and Richardson, M. H., “Modal Parameter Estima-
equation. The measured response y t is generated by 2 stochas- tion from Ambient Response Data,” presented at IMAC 2001, Feb-
ruary 5-8, 2001.
tic processes w t and v t that represent the unmeasured and un-
3. Schwarz, Brian, and Richardson, M. H., “Post-Processing Ambient
known noise processes. The matrix A is called the state ma- and Forced Response Bridge Data To Obtain Modal Parameters,” Pro-
trix and the matrix C is called the observation matrix. ceedings of the IMAC XIX Conference, Orlando, FL, Feb. 5-8, 2001.
This technique uses a mathematical framework based on 4. Herlufsen, H., and Møller, N., “Operational Modal Analysis of a
Wind Turbine Wing Using Acoustical Excitation,” Brüel & Kjær Ap-
stochastic processes. A stochastic process is a mathematical plication Note, 2002.
modelization of a physical phenomenon that is not determin- 5. Møller, N., Brincker, R., Herlufsen, H., Andersen, P., “Modal Test-
istic and is somehow not predictable from the knowledge of ing of Mechanical Structures Subject to Operational Forces,” IMAC
the present state of the system. XIX.
6. Brincker, R., Andersen, P., and Møller, N., “Output-Only Modal Test-
The intrinsic randomness of the operational modal analysis ing of a Car Body Subject to Engine Excitation,” Proc. of the 18th
technique makes stochastic techniques very suitable for mod- International Modal Analysis Conference, San Antonio, TX, Febru-
eling a physical system. The SSI technique works on the raw ary 7-10, 2000.
time data and tries to fit a model to the data captured from the 7. Bendat, Julius S., and Piersol, Allan G., Random Data, Analysis and
Measurement Procedures, John Wiley and Sons, 1986.
responses at the degrees of freedom.
8. Brincker, R., and Andersen, P., “ARMA (Auto Regressive Moving Av-
The system represented in Equation (27) can be rewritten in erage) Models in Modal Space,” Proc. of the 17th International
the following form: Modal Analysis Conference, Kissimmee, FL, 1999.
9. Peeters, Bart, and DeRoeck, Guido, “Reference-Based Stochastic
Subspace Identification for Output-Only Modal Analysis,” Mechani-
 xˆ t +1 = [A]xˆ t + [K ]et
 (28) cal Systems and Signal Processing, July 1999.
 y t = [C ]xˆ t + et 10. Van Overschee, P., and DeMoor, B., Subspace Identification for Lin-
ear Systems, Kluver Academic Publishers, 1996.
The matrix K is the non-steady state Kalman gain (covariance
matrix), and e t is the innovation Gaussian process. X t and xt+1
are the corresponding prediction state vectors for Equation (27). The author can be contacted at: [email protected].

SOUND AND VIBRATION/AUGUST 2002 27