See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/268048253

PolyMax: a revolution in operational modal analysis

Article · January 2005

CITATIONS READS

50 1,897

2 authors:

Bart Peeters Herman Van der Auweraer

Siemens Industry Software Siemens Industry Software NV
234 PUBLICATIONS 4,605 CITATIONS 321 PUBLICATIONS 3,379 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Unvico View project

Virtual Sensing View project

All content following this page was uploaded by Bart Peeters on 29 April 2015.

The user has requested enhancement of the downloaded file.

 POLYMAX: A REVOLUTION
IN
OPERATIONAL MODAL ANALYSIS

Bart Peeters, Herman Van der Auweraer

LMS International, Belgium
[email protected]

Abstract
Recently the superior performance of the PolyMAX modal parameter estimation method using
frequency response functions was demonstrated. The method can be implemented in a very similar
way as the industry standard polyreference (time-domain) least-squares complex exponential
method: in a first step a stabilisation diagram is constructed containing frequency, damping and
participation information. Next, the mode shapes are found in a second least-squares step, based on
the user selection of stable poles. One of the specific advantages of the technique lies in the very
stable identification of the system poles and participation factors as a function of the specified
system order, leading to easy-to-interpret stabilisation diagrams. This implies a potential for
automating the method and to apply it to “difficult” estimation cases such as high-order and/or
highly damped systems with large modal overlap. Furthermore, PolyMAX is computationally
extremely efficient. In this paper the method will be extended to the operational case: only the
structural response to some unknown inputs is available for system identification. Attention will be
given to the pre-processing of the raw operational time data and the post-processing of the poles
and operational reference factors to obtain the mode shapes. The paper also gives an overview of
existing Operational Modal Analysis methods and discusses the positioning of PolyMAX herein.

1 Introduction
The application of system identification to vibrating structures yielded a new research domain in
mechanical engineering, known as Experimental Modal Analysis. In this case, a modal model,
consisting of eigenfrequencies, damping ratios, mode shapes and modal participation factors, is
identified from vibration data. Classically, one applies an artificial, measurable input to the system
and one measures the output. From these measurements, the experimental model can be obtained
by a variety of parameter estimation methods. However, cases exist where it is rather difficult to
apply an artificial force and where one has to rely upon available ambient excitation sources. It is
practically impossible to measure this ambient excitation and the outputs are the only information
that can be passed to the system identification algorithms. In this case one speaks of Operational
Modal Analysis.
Operational Modal Analysis proved to be very useful in civil engineering, where it is very difficult
and expensive to excite constructions such as bridges and buildings with a hammer or shaker and to
obtain artificially induced vibration levels that exceed the natural vibrations due to traffic or wind.
Also in mechanical engineering, Operational Modal Analysis is successfully applied to obtain data-
based dynamic models of, for instance, a car during road testing or an aircraft during flight tests
[1].

In Proceedings of the 1st International Operational Modal Analysis Conference,

Copenhagen, Denmark, 26-27 April 2005
2 Operational modal parameter estimation
The simplest method to estimate the modal parameters of a structure subjected to ambient loading
is the so-called peak-picking method. The method is named after the key step of the method: the
identification of the eigenfrequencies as the peaks of power spectrum plots. A violation of the basic
assumptions (low damping and well-separated frequencies) leads to erroneous results. In fact the
method identifies operational deflection shapes instead of mode shapes and for closely spaced
modes such an operational deflection shape will be the superposition of multiple modes. Another
disadvantage is that the selection of the eigenfrequencies is a rather subjective task and that the
method does not yield any damping estimates.
A more advanced method consists of computing the Singular Value Decomposition (SVD) of a
matrix with the auto and cross spectra between the outputs. The method is discussed in many
references [2][3][4] and received different names: Principal Component Analysis (PCA), Complex
Mode Indication Function (CMIF), and Frequency Domain Decomposition (FDD). The method is
able to detect mode-multiplicity (i.e. different modes at the same frequency), but is still rather
subjective. The PP and the SVD-based method are both so-called nonparametric methods: the
eigenfrequencies are determined by just looking at signal-based features without fitting or
estimating a parametric model. On the other hand, methods in which a parametric model is
determined from the data allow the use of objective criteria to assess the modelling quality (e.g. the
use of stabilisation diagrams or other convergence criteria). This kind of methods will be discussed
in the following.
An interesting, more advanced alternative for peak picking is the stochastic subspace identification
method. In stochastic subspace identification a so-called stochastic state space model is identified
directly from measured output data [5] or output correlations. It can be shown that a stochastic state
space model is a good representation of a vibration structure excited by some unknown forces,
which are assumed to be white noise signals. Successful real-life modal analysis applications of
stochastic subspace identification can be found in [6][7].
In [8], a frequency-domain maximum likelihood approach for the extraction of modal parameters
from output-only data was proposed. Originally intended for application to frequency response
functions (FRFs) [9], the method was extended to use output spectra as primary data. Maximum
likelihood identification is an optimisation-based method that estimates the parameters of a model
by minimising an error norm. A discussion on the use of the method to identify parametric
frequency-domain models can be found in [10][11]. During the last years attention has been paid to
the optimisation of the frequency-domain maximum likelihood method: the algorithm has been
modified to keep the memory requirements as low as possible; and using an adapted
parameterisation and fast signal processing techniques, an important reduction of the computation
time was possible. Successful industrial Operational Modal Analysis applications of frequency-
domain maximum likelihood identification are reported in [12][13].
Section 5 will introduce the new PolyMAX method, of which the roots go back to the linear least
squares method that was applied to find the initial values for the iterative maximum likelihood
method.

3 Output-only frequency-domain model

Frequency-domain Operational Modal Analysis methods, such as PolyMAX and maximum
likelihood identification require output spectra as primary data. In this subsection, it will be shown
that, under the assumption of white noise input, output spectra can be modelled in a very similar
way as FRFs. It is well known that the modal decomposition of an FRF matrix [H ( ω)] ∈ » l ×m is
[14]:

[H (ω)] = ∑ {vi }< li

n T
>
+
{v }< l
*
i i
H
>
(1)
i =1 jω − λ i jω − λ*i

where n is the number of modes; •* is the complex conjugate of a matrix; • H is the complex
conjugate transpose (Hermitian) of a matrix; {vi }∈ » l are the mode shapes; < liT >∈ » m are the
modal participation factors and λ i are the poles, which are occurring in complex-conjugated pairs
and are related to the eigenfrequencies ωi and damping ratios ξi as follows:

λ i , λ*i = −ξi ωi ± j 1 − ξi2 ωi (2)

[ ]
The input spectra [S uu ]∈ » m×m and output spectra S yy (ω) ∈ » l ×l of a system represented by the
FRF matrix [H (ω)] are related as:
[S yy (ω)] = [H (ω)][S uu ][H (ω)]H (3)

In case of operational data the output spectra are the only available information. The deterministic
knowledge of the input is replaced by the assumption that the input is white noise. A property of
white noise is that it has a constant power spectrum. Hence [S uu ] is independent of the frequency
in (3). The modal decomposition of the output spectrum matrix is obtained by inserting (1) into (3)
and converting to the partial fraction form [6][7][8]:

[S yy ( jω)] = ∑ {vijω}<−gλi > + {vi }< gi* > + {−gi j}ω<−vλi > + {g i }< vi *>
n * * * *
(4)
i =1 i jω − λ i i − jω − λ i

where < g i >∈ » l are the so-called operational reference factors, which replace the modal
participation factors in cases where only output data are available. Their physical interpretation is
less obvious as they are a function of all modal parameters of the system and the constant input
spectrum matrix. Note that the order of the power spectrum model is twice the order of the FRF
model. The goal of Operational Modal Analysis is to identify the right hand side terms of (4) based
on measured output data pre-processed into output spectra (Section 4).

4 Pre-processing operational data

Power spectra are defined as the Fourier transform of the correlation sequences. The most popular
non-parametric spectrum estimate is the so-called weighted averaged periodogram (also known as
modified Welch’s periodogram). Weighting means that the signal is weighted by one of the
classical windows (Hanning, Hamming, …) to reduce leakage. Welch’s method starts with
computing the DFT of the weighted outputs:
N −1
Y ( ω) = ∑ wk yk exp( − jωk∆t ) (5)
k =0

where wk denotes the time window. An unbiased estimate of the spectrum is the weighted
periodogram:
1
S (yyi ) (ω) = N −1
Y ( ω) Y H ( ω) (6)
2
∑ wk
k =0
The variance of the estimate is reduced by splitting the signal in possibly overlapping segments,
computing the weighted periodogram of all segments and taking the average:
1 P (i )
S yy (ω) = ∑ S yy (ω)
P i =1
(7)

where P is the number of segments and superindex i denotes the segment index.
Another non-parametric spectrum estimate is the so-called weighted correlogram. It will be shown
that this estimate has some specific advantages in a modal analysis context. First the correlations
have to be estimated:
N −1
1
Ri =
N
∑ y k +i y kT (8)
k =0

High-speed (FFT-based) implementations exist to compute the correlations as in (8); see for
instance [15]. The weighted correlogram is the DFT of the weighted estimated correlation
sequence:
L
S yy ( ω) = ∑ wk Rk exp( − jωk∆t ) (9)
k =− L

where L is the maximum number of time lags at which the correlations are estimated. This number
is typically much smaller than the number of data samples to avoid the greater statistical variance
associated with the higher lags of the correlation estimates. In a modal analysis context, the
weighted correlogram has the following advantages:
• It is sufficient to compute the so-called half spectra which are obtained by using only the
correlations having a positive time lag in (9):
w0 R0 L
+
S yy ( ω) = + ∑ wk Rk exp( − jωk∆t ) (10)
2 k =1

The relation between the half spectra (10) and the full spectra (9) is the following:
+
S yy ( ω) = S yy +
( ω) + S yy ( ω)( )H
(11)

It can be shown (see for instance [8][13]) that the modal decomposition of these half
spectra only consists of the first two terms in (4):
+
S yy ( jω) = ∑
n
{vi }< g i > + {vi* }< gi* > (12)
i =1 jω − λ i jω − λ*i

The advantage in modal analysis is that models of lower order can be fitted without
affecting the quality.
• Under the white noise input assumption, the output correlations are equivalent to impulse
responses. So, just like in impact testing, it seems logical to apply an exponential window
to the correlations before computing the DFT (10). The exponential window reduces the
effect of leakage and the influence of the higher time lags, which have a larger variance.
Moreover, the application of an exponential window to impulse responses or correlations
is compatible with the modal model and the pole estimates can be corrected. This is not
the case when a Hanning window is used: such a window always leads to biased damping
estimates.
Interesting discussions on the estimation of spectra from the measured time histories in a modal
analysis context can be found in [8][12][13].
The weighted correlogram approach is illustrated using in-flight aircraft response data. Figure 1
shows the time history of the wing tip acceleration due to unmeasurable turbulence excitation.
Figure 2 (Left) shows how a few thousands time data samples are reduced to 256 correlation
samples using (8). The effect of the exponential window is visible. Figure 2 (Right) shows the half
spectrum estimate according to (10). The half spectrum looks much better if an exponential
window has been applied for the above-discussed reasons.

Figure 1: Aircraft acceleration due to unmeasurable turbulence excitation.

Figure 2: (Left) Typical output correlation; before and after applying an exponential window.
(Right) Corresponding half spectra.

5 PolyMAX modal parameter estimation

After having pre-processed output data into output spectra (Section 4), it is now the task to identify
a modal model (Section 3). By comparing (12) with (1), it is clear that FRFs and half spectra can be
parameterised in exactly the same way. By consequence, the same modal parameter estimation
methods can be used in both cases. PolyMAX is such a method. It will be reviewed in the
following.
Originally, the least-squares complex frequency-domain (LSCF) estimation method was introduced
to find initial values for the iterative maximum likelihood method [9]. The method estimates a so-
called common-denominator transfer function model [11]. Quickly it was found that these “initial
values” yielded already very accurate modal parameters with a very small computational effort
[9][16][17]. The most important advantage of the LSCF estimator over the available and widely
applied parameter estimation techniques [14] is the fact that very clear stabilisation diagrams are
obtained.
It was found that the identified common-denominator model closely fitted the measured frequency
response function (FRF) data. However, when converting this model to a modal model (1) by
reducing the residues to a rank-one matrix using the singular value decomposition (SVD), the
quality of the fit decreased [16]. Another feature of the common-denominator implementation is
that the stabilisation diagram can only be constructed using pole information (eigenfrequencies and
damping ratios). Neither participation factors nor mode shapes are available at first instance. The
theoretically associated drawback is that closely spaced poles will erroneously show up as a single
pole [18]. These two reasons provided the motivation for developing “PolyMAX”, which is a
polyreference version of the LSCF method, using a so-called right matrix-fraction model. In this
approach, also the participation factors are available when constructing the stabilisation diagram.
The main benefits of the polyreference method are the facts that the SVD step to decompose the
residues can be avoided and that closely spaced poles can be separated. PolyMAX was introduced
and validated using industrial data sets in [18][19][20]. The present paper will indicate how
PolyMAX can also be applied to operational data when appropriate pre-processing (Section 4) and
post-processing (Section 5.4) is applied.

5.1 Right matrix-fraction model

The PolyMAX method needs FRFs as primary data and identifies a right matrix-fraction model:
[H (ω)] = [B(ω)][A(ω)]−1 (13)

where H (ω) ∈ » l ×m is the FRF matrix containing the FRFs between all m inputs and all l outputs;
B( ω) ∈ » l ×m is the numerator matrix polynomial and A(ω) ∈ » m×m is the denominator matrix
polynomial. Each row of the right matrix-fraction model (13) can be written as:
∀o = 1,2,..., l : < H o (ω) > = < Bo (ω) > [A(ω)]−1 (14)

The numerator row-vector polynomial of output o and the denominator matrix polynomial are
defined as:
p
< Bo ( ω) > = ∑ Ω r ( ω) < β or >
r =0 (15)
p
[A(ω)] = ∑ Ω r (ω) [α r ]
r =0

where Ω r (ω) are the polynomial basis functions and p is the polynomial order. In the PolyMAX
method, a z-domain model is used (i.e. a frequency-domain model that is derived from a discrete-
time model) and, by consequence, the basis functions are (with ∆ t the sampling time):
jω∆t r
Ω r ( ω) = e (16)

The polynomial coefficients β or ∈ »1×m and α r ∈ » m×m are assembled in following matrices:
  β1 
 βo0   α0   
    β2 
β
 o1  ( p +1)×m
α
 1 m ( p +1)×m    ∈
( l +m )( p +1)×m (17)
βo =  ∈
( ∀o = 1, 2,..., l ), α = ∈
, θ =
   
     βl 
 β op  α p 
    α
 
The FRF model of (13) is now written as function of the coefficients H (ωk , θ) . Note that in the
PolyMAX method real-valued polynomial coefficients are assumed. It is possible to derive another
variant of the method with complex coefficients. A discussion on these aspects for the common-
denominator LSCF implementation can be found in [17].

5.2 Equation error formulation

The challenge is now to find all unknown model coefficients θ based on measurements (or better:
non-parametric estimates) of the FRFs Hˆ o (ωk ) , where •ˆ denotes a measured quantity and
ωk ( k = 1,2,..., N f ) are the discrete frequencies at which FRF measurements are available. The
coefficients θ can be identified by minimising the following non-linear least-squares (NLS)
equation errors ε oNLS (ω k , θ) ∈ »1×m :
o k o k ( o k o k ) o k (
ε NLS ( ω , θ) = w (ω ) H ( ω , θ) − Hˆ ( ω ) = w (ω ) B ( ω , β ) A −1 ( ω , α) − Hˆ ( ω ) (18) o k o k o k )
where the scalar weighting function wo (ωk ) is introduced. This frequency- and output-dependent
weighting function allows taking into account data quality differences that may exist between
different outputs. The equation errors for all outputs and all frequency lines are combined in
following scalar cost function:
l Nf

NLS ( θ) = ∑∑ tr  ε oNLS (ωk , θ)

( ) H
ε oNLS ( ωk , θ) 

(19)
o =1 k =1

where • H is the complex conjugate transpose (Hermitian) of a matrix and tr{•} is the trace of a
matrix (sum of the elements on the main diagonal). In fact, the ‘trace’ notation boils down to
computing the squared 2-norm of the row vector ε oNLS ( ωk , θ) . The cost function is minimised by
putting the derivatives of (19) with respect to the unknown model coefficients θ equal to zero. It is
obvious that this leads to non-linear equations when expression (18) is used for the equation errors.
This non-linear least-squares problem can be approximated by a (sub-optimal) linear least-squares
one by right-multiplying (18) with the numerator matrix polynomial A, yielding equation errors
ε oLS ( ωk , θ) ∈
1×m that are linear in the parameters:
o k o k (
ε LS ( ω , θ) = w ( ω ) B ( ω , β ) − Hˆ (ω ) A( ω , α)
o k o o k k )
( )
p
(20)
= wo ( ωk ) ∑ Ω r (ωk ) β or − Ω r ( ωk ) Hˆ o ( ωk ) α r
r =0
N f ×m
The equation errors at all frequency lines are stacked in a matrix EoLS ( θ) ∈
:
 ε oLS ( ω1 , θ) 
 
LS
 ε ( ω , θ)  β 
 = (X o Yo )  o 
o 2
EoLS (θ) =  (21)

 α
 εoLS ( ω N , θ) 
 f 
with:

 (
 wo ( ω1 ) Ω 0 ( ω1 ) Ω p ( ω1 ) ) 

Xo =    ∈
N f ×( p +1) (22)

 f f
(
 wo (ω N ) Ω 0 ( ω N )Ω p (ω N
f
)

)


 (
− wo ( ω1 ) Ω 0 ( ω1 )Ω p ( ω1 ) ⊗ Hˆ o ( ω1 ) ) 

Yo =    ∈
N f ×m ( p +1) (23)

 f
(
 − wo ( ω N ) Ω 0 ( ω N )Ω p ( ω N ) ⊗ Hˆ o ( ω N
f f f
) 
)

where ⊗ denotes the Kronecker product.

5.3 Reduced normal equations

Similar as (19), the following cost function can be formulated based on the linearised equation
errors:
l Nf


(

LS (θ) = ∑∑ tr  ε oLS (ω k , θ) )
H
ε oLS (ωk , θ) 

(24)
o =1 k =1

Minimising this cost function leads to a weighted linear least squares problem. Introducing the
matrix notation of (21), (22), (23), the cost function can be written as:

( )
l

LS (θ) = ∑ tr  EoLS (θ) EoLS (θ) 
H

o =1  
  β 
( )  X oH 
l
= ∑ tr  βTo α T   (X Yo ) o  (25)
YH  o
o =1   o   α 

{
= tr θT J H J θ }
lN f ×(l + m )( p +1)
where J ∈ » is the so-called Jacobian matrix given by:
 X1 0
0 Y1 
 
 0 X2
0 Y2 
J = (26)






 
 0 0
X l Yl 

In the case of real-valued coefficients θ , it can be shown that the expression J H J can be
substituted by its real part Re( J H J ) ∈ » ( l +m )( p +1)×( l +m )( p +1) and the cost function (25) becomes:
{

LS ( θ) = tr θT Re( J H J ) θ } (27)

with:
  R1 0  0 S1 
 
 0 R2  0 S2 
Re( J J ) =  
H      ∈
( l +m )( p +1)×( l + m )( p +1)
 (28)
 0 0  Rl Sl 
 T
∑o=1To 
l
 S1 S 2T  S lT

in which following notation is used:

Ro = Re( X oH X o ) ∈ » ( p +1)×( p +1)
So = Re( X oH Yo ) ∈ » ( p +1)×m ( p +1) (29)
To = Re(YoH Yo ) ∈ » m ( p +1)×m ( p +1)

The cost function (27) is minimised by putting its derivatives with respect to the unknown
polynomial coefficients θ equal to zero:
∂
LS ( θ)
= 2( Roβ o + So α) = 0 , ∀o = 1,…, l
∂β o
(30)
∂
LS ( θ) l
= 2∑ ( SoT β o + To α) = 0
∂α o =1

These equations are the so-called normal equations which can be written synthetically using (28)
and (17):
2 Re( J H J ) θ = 0 (31)

At first instance, we are only interested in the denominator polynomial α from which the poles
and modal participation factors can be computed. This information is sufficient to construct a
stabilisation diagram. By consequence, the least-squares problem can be reduced by eliminating the
β o coefficients from (30):
β o = − Ro−1S o α (32)

yielding the so-called reduced normal equations:

 l 
2∑ (To − S o Ro S o )  α = 0
T −1

 o =1 
(33)
Mα = 0

where M ∈ » m ( p +1)×m ( p +1) is defined in above equation and can be computed from the measured
FRF data. This equation can be solved for the denominator polynomial α in a least-squares sense.
To avoid finding the trivial solution α = 0 , a constraint is imposed on the parameters.

5.4 Practical implementation and operational residuals

In modal analysis applications, one is usually not only interested in a good model as such, but more
important, in the accuracy of the estimated modal parameters. When trying to estimate the modal
parameters from real data, it is generally a good idea to over-specify the model order considerably;
i.e. to try to fit high-order models that contain much more modes than present in the data.
Afterwards the true physical modes are separated from the spurious numerical ones by interpreting
a so-called stabilisation diagram [21]. The poles corresponding to a certain model order are
compared to the poles of a one-order-lower model. If their differences are within pre-set limits, the
pole is labelled as a stable one. The spurious numerical poles will not stabilise at all during this
process and can be sorted out of the modal parameter data set more easily. The use of a stabilisation
diagram solves a typical system identification problem, namely the determination of the model
order. A stabilisation diagram using stochastic subspace identification is compared to a PolyMAX
diagram in Figure 3. Although both methods are known to yield accurate modal parameter
estimates, it is obvious that PolyMAX considerably facilitates the identification process: it is much
easier to pick the stable poles from the stabilisation diagram. More examples of stabilisation
diagrams can be found in another paper of these proceedings [1].
The poles λ i and the modal participation factors < liT > (in case of FRFs) or operational reference
factors < g i > (in case of spectra) are found from the denominator polynomial (33) by computing
the eigenvalue decomposition of its companion matrix, which is rather classical [14]. A pth order
right matrix-fraction model yields pm poles. An efficient construction of the PolyMAX stabilisation
diagram is possible by formulating the least squares problem (33) for the maximum model order
p max . Considering submatrices of appropriate dimensions can then solve lower-order problems.
The interpretation of the stabilisation diagram yields a set of poles and corresponding operational
reference factors. The mode shapes can then be found from (12):
+
S yy ( jω) = ∑
n
{vi }< gi > + {vi* }< g i* > + LR + jω UR (34)
i =1 jω − λ i jω − λ*i jω

in which LR, UR ∈ » l ×m , respectively the lower and upper residuals, have been introduced to
model the influence of the out-of-band modes in the considered frequency band. The only
unknowns in (34) are the mode shapes {vi } and the lower and upper residuals. They are readily
obtained by solving (34) in a linear least-squares sense. This second step is commonly called least-
squares frequency-domain (LSFD) method [14].
The operational residues LR, UR in (34) are different from the usual residuals in case of FRFs.
They were determined by verifying the asymptotic behaviour of the output spectra of a SDOF
mechanical system excited by white noise. The formats of the residuals that have to be applied in
the LSFD method are listed in Table 1 for FRFs (1), classical full output spectra (4) and half
spectra (12). In general, they depend on the measured output quantity (displacement – velocity –
acceleration). Deriving displacements to velocities (and similarly velocities to accelerations)
implies multiplication by jω for FRFs and ( jω) 2 for full spectra. Rather surprising is that this is
not the case for half spectra: the residuals are independent of the output quantity.

Output FRFs (2) Full spectra (5) Half spectra (13)

quantity LR UR LR UR LR UR
LR LR LR
Displacement UR UR ( jω) UR
( jω) 2 ( jω) 4 jω
LR LR LR
Velocity ( jω) UR ( jω) 2 UR ( jω) UR
jω ( jω) 2 jω
LR
Acceleration LR ( jω) 2 UR LR ( jω) 4 UR ( jω) UR
jω

Table 1: Lower and upper residuals.

Figure 3: Stabilisation diagrams obtained by applying (Left) stochastic subspace identification and
(Right) the PolyMAX method applied to half spectra with an exponential window of 1%.

6 Conclusions
In this paper, the PolyMAX modal parameter estimation method that is normally applied to
frequency response functions is extended to the operational case: only the structural response to
unknown inputs is available for system identification. Attention was focused on pre-processing the
raw operational time data into auto and cross spectra. The correlogram approach for estimating half
spectra seems to be preferred in a modal analysis context: correlations with positive time lags are
computed from the time data; an exponential window is applied to reduce leakage and the influence
of the noisier higher time lag correlation samples; and finally the DFT of the windowed correlation
samples is taken. An exponential window is compatible with the modal model and therefore, the
pole estimates can be corrected for the application of such a window. In all examined cases (see
Figure 3 and [1]) it has been demonstrated that the operational PolyMAX method yields extremely
clear stabilization diagrams, easing the pole-selection process.

References
[1] B. PEETERS, H. VAN DER AUWERAER; STEVEN PAUWELS, AND JAN DEBILLE, Industrial
relevance of Operational Modal Analysis – civil, aerospace and automotive case histories, In
Proceedings of IOMAC, the 1st International Operational Modal Analysis Conference,
Copenhagen, Denmark, 26-27 April 2005.
[2] M. PREVOSTO, Algorithmes d'Identification des Caractéristiques Vibratoires de Structures
Mécaniques Complexes, PhD thesis, Université de Rennes I, France, 1982.
[3] D. OTTE, Development and evaluation of singular value analysis methodologies for studying
multivariate noise and vibration problems, PhD thesis, Dept. of Mech. Eng., K.U.Leuven,
Belgium, 1994.
[4] R. BRINCKER, L. ZHANG, AND P. ANDERSEN, Modal identification from ambient responses
using frequency domain decomposition, In Proceedings of IMAC 18, 625-630, San Antonio
(TX), USA, 2000.
[5] P. VAN OVERSCHEE AND B. DE MOOR, Subspace Identification for Linear Systems: Theory -
Implementation – Applications, Dordrecht, The Netherlands: Kluwer Academic Publishers,
1996.
 [6] B. PEETERS, System identification and damage detection in civil engineering, PhD thesis,
Department of Civil Engineering, K.U.Leuven, Belgium, www.bwk.kuleuven.ac.be/bwm,
2000.
[7] L. HERMANS AND H. VAN DER AUWERAER, Modal testing and analysis of structures under
operational conditions: industrial applications, Mechanical Systems and Signal Processing,
13(2), 193-216, 1999.
[8] L. HERMANS, H. VAN DER AUWERAER, AND P. GUILLAUME, A frequency-domain maximum
likelihood approach for the extraction of modal parameters from output-only data, In
Proceedings of ISMA23, 367-376, Leuven, Belgium, 16–18 September 1998.
[9] P. GUILLAUME, P. VERBOVEN, AND S. VANLANDUIT, Frequency-domain maximum likelihood
identification of modal parameters with confidence intervals, In Proceedings of ISMA 23,
K.U.Leuven, Belgium, 1998.
[10] PINTELON R. AND J. SCHOUKENS, System Identification: a Frequency Domain Approach,
New York: IEEE Press, 2001.
[11] P. GUILLAUME, R. PINTELON, AND J. SCHOUKENS, Parametric identification of multivariable
systems in the frequency domain - a survey. In Proceedings of ISMA 21, 1069-1082, Leuven,
Belgium, 18-20 September 1996.
[12] E. PARLOO, Application of frequency-domain system identification techniques in the field of
operational modal analysis, PhD thesis, Department of Mechanical Engineering,
V.U.Brussel, Belgium, http://avrg.vub.ac.be, 2003.
[13] B. CAUBERGHE, Applied frequency-domain system identification in the field of experimental
and operational modal analysis, PhD thesis, Department of Mechanical Engineering,
V.U.Brussel, Belgium, http://avrg.vub.ac.be, 2004.
[14] W. HEYLEN, S. LAMMENS, AND P. SAS, Modal Analysis Theory and Testing, K.U.Leuven,
Belgium, 1997.
[15] A.V. OPPENHEIM AND R.W. SCHAFER, Digital Signal Processing, Prentice-Hall, 1975.
[16] H. VAN DER AUWERAER, P. GUILLAUME, P. VERBOVEN AND S. VANLANDUIT, Application of
a fast-stabilizing frequency domain parameter estimation method, ASME Journal of Dynamic
Systems, Measurement, and Control, 123(4), 651-658, 2001.
[17] P. VERBOVEN, Frequency domain system identification for modal analysis, PhD Thesis,
V.U.Brussel, Belgium, 2002.
[18] P. GUILLAUME, P. VERBOVEN, S. VANLANDUIT, H. VAN DER AUWERAER AND B. PEETERS, A
poly-reference implementation of the least-squares complex frequency-domain estimator, In
Proceedings of IMAC 21, Kissimmee (FL), USA, February 2003.
[19] B. PEETERS, P. GUILLAUME, H. VAN DER AUWERAER, B. CAUBERGHE, P. VERBOVEN, AND J.
LEURIDAN, Automotive and aerospace applications of the PolyMAX modal parameter
estimation method, In Proceedings of IMAC 22, Dearborn (MI), USA, 26–29 January 2004.
[20] B. PEETERS, H. VAN DER AUWERAER, P. GUILLAUME, AND J. LEURIDAN, The PolyMAX
frequency-domain method: a new standard for modal parameter estimation? Shock and
Vibration, Special Issue dedicated to Professor Bruno Piombo, 11, 395–409, 2004.
[21] H. VAN DER AUWERAER AND B. PEETERS, Discriminating physical poles from mathematical
poles in high order systems: use and automation of the stabilization diagram, In Proceedings
of IMTC 2004, the IEEE Instrumentation and Measurement Technology Conference, Como,
Italy, 18–20 May 2004.
[22] J. LANSLOTS, B. RODIERS, AND B. PEETERS, Automated pole-selection: proof-of-concept &
validation, In Proceedings of ISMA 2004, Leuven, Belgium, 20–22 September 2004.

View publication stats