Bart Peeters1

LMS International,
Interleuvenlaan 68,
B-3001 Leuven,
Belgium
e-mail: [email protected]

Guido De Roeck
Department of Civil Engineering,
K.U. Leuven,
Kasteelpark Arenberg 40,
B-3001 Leuven,
Belgium
e-mail: [email protected]

Stochastic System Identification

for Operational Modal Analysis:
A Review
This paper reviews stochastic system identification methods that have been used to estimate the modal parameters of vibrating structures in operational conditions. It is found
that many classical input-output methods have an output-only counterpart. For instance,
the Complex Mode Indication Function (CMIF) can be applied both to Frequency Response Functions and output power and cross spectra. The Polyreference Time Domain
(PTD) method applied to impulse responses is similar to the Instrumental Variable (IV)
method applied to output covariances. The Eigensystem Realization Algorithm (ERA) is
equivalent to stochastic subspace identification. DOI: 10.1115/1.1410370

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. Because in these cases the deterministic knowledge of the
input is replaced by the assumption that the input is a realization
of a stochastic process white noise, one speaks of stochastic
system identification. Specializing to the identification of vibrating
structures the terms output-only modal analysis and operational
modal analysis are used. A common problem of operational modal
analysis methods is that if the white noise assumption is violated,
for instance if the input contains in addition to white noise also
some dominant frequency components, these frequency components cannot be separated from the eigenfrequencies of the system
and will be identified as such.
The need to perform output-only modal analysis probably
emerged first 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.
Nevertheless, also in mechanical engineering, operational modal
analysis proved to be very useful: for instance to obtain the modal
parameters of a car during road testing or an aeroplane during
flight tests.
Many textbooks exist that give an extensive overview of inputoutput modal parameter estimation methods 1 4. We should
also mention some recent efforts to compare input-output system
identification methods for applications in structural dynamics
5,6. Individual output-only modal parameter estimation methods
are discussed in several papers, but an overview and comparison
of different methods are missing. The present paper, which is
1
Research performed while working at the Department of Civil Engineering, K.U.
Leuven.
Contributed by the Dynamic Systems and Control Division for publication in the
JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL. Manuscript
received by the Dynamic Systems and Control Division February 7, 2001. Associate
Editor: S. Fassois.

based on 7, tries to fill that gap. Almost all methods discussed in

this paper have successfully been applied to real-life vibration
data. These applications are beyond the scope of this paper, but
can, for instance, be found in 79.

2 Vibrating Structures: Models and Measurement

Data
The aim of this section is to introduce some basic concepts that
are most helpful in understanding the similarities and differences
between stochastic system identification methods.
2.1 Models. One of the first steps of system identification is
adopting a certain model structure. Afterwards the parameters of
the chosen model are estimated from measurement data. A wide
range of model structures is proposed in system identification literature, see for instance Ljung 10. The general system identification procedure is to try out several model structures without
bothering about the underlying physical input-output relations.
The aim of this section is to discuss some models that can truly
represent a vibrating structure excited by white noise. By consequence, these models are physically meaningful.
Physical Model. The dynamic behavior of a discrete mechanical system consisting of n 2 masses connected through springs and
dampers is described by following matrix differential equation:
M q t C 2 q t Kq t f t

(1)

n 2 n 2

where M ,C 2 ,KR
are the mass, damping and stiffness matrices; q(t)Rn 2 is the displacement vector at continuous time t.
A dot over a time function denotes the derivative with respect to
time. The vector f (t)Rn 2 is the excitation force. For systems
with distributed parameters e.g., civil engineering structures, Eq.
1 is obtained as the Finite Element FE approximation of the
system with only n 2 degrees of freedom DOFs left. Although the
nearly physical model 1 is a good representation of a vibrating
structure, it is not directly useful in an experimental modelling
context. First, it is not possible and also not necessary to measure all DOFs of the FE model. Second, this equation is in
continuous-time, whereas measurements are available as discrete
time samples. And finally, there is some noise modelling needed:
there may be other unknown excitation sources apart from f (t)
and measurement noise is always present in real life.
Stochastic State-Space Model. It can be shown that, by applying model reduction, sampling and modelling the noise, Eq. 1
can be converted to following discrete-time stochastic state-space
model see for instance 7,11 for a detailed derivation:

Journal of Dynamic Systems, Measurement, and Control

Copyright 2001 by ASME

DECEMBER 2001, Vol. 123 659

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

x k1 Ax k w k

(2)

y k Cx k k
where y k R is the sampled output vector the measurements;
x k Rn is the discrete state vector; w k Rn is the process noise,
typically due to disturbances and modelling inaccuracies, but here
also due to the unknown excitation of the structure; k Rl is the
measurement noise, typically due to sensor inaccuracy, but here
also due to the unknown excitation of the structure; k is the time
instant; l is the number of outputs; n is the system order (n
2n 2 ). The matrix ARnn is the state transition matrix that
completely characterizes the dynamics of the system by its eigenvalues; and CRln is the output matrix that specifies how the
internal states are transformed to the outside world. The noise
vectors are both unmeasurable vector signals assumed to be zero
mean, white and with covariance matrices:
l

Q
wp
w Tq Tq T
p
S

S
R

pq

(3)

where E is the expected value operator; pq is the Kronecker

delta.
It is precisely such a stochastic state-space model that will be
identified when using so-called subspace identification methods
see Sections 4.2 and 5.1. In a second step, the modal parameters
are obtained from the matrices A and C. The derivation starts with
the eigenvalue decomposition of A:
A d

i , i* i i j 1 i2 i

(5)

where t is the sampling time. Finally, the mode shapes V

Cln are found as:
VC

(6)

ARMA Model. The more classical system identification methods 10 identify models that do not contain the state. It can be
shown 12,13 that the following so-called ARMA model is
equivalent to the stochastic state-space model 2:
y k 1 y k1 p y k p e k 1 e k1 p e kp

p1

where S y (z)Cll is the spectrum matrix containing the power

and cross spectra between the outputs. The power spectra are real
and located on the main diagonal. This expression can be evaluated for any number on the unit circle ze j t where rad/s
can be any frequency of interest. Matrix GRnl is the next
stateoutput covariance matrix and R 0 Rll is the zero-lag output covariance matrix:

0
0

I
1

V
V
V d
V d

d
V dp2
V dp2
V dp1
V dp1

660 Vol. 123, DECEMBER 2001

(8)

R 0 E y k y Tk

GE x k1 y Tk ,

(10)

An in structural dynamics more common form of the output

power spectrum matrix is obtained by applying the Laplace transform to a continuous-time model and introducing the modal parameters see 7 for a detailed derivation:

S y s

i1

1
l T R u
s i i i

i1

1
l T
s * i i i

s j

(11)

where s is the Laplace variable; i C is the ith modal vector;

l iT Cm is the ith modal participation vector; m is the number of
l

white noise inputs; R u Rmm is the white noise input covariance matrix.

2.2 Measurement Data. In principle output data y k are

available as discrete samples of the time signal. The identification
methods of Section 5 will be able to use directly these time signals. However, many system identification methods exist that require other types of data.
The identification methods of Section 4 require output covariances as primary data type. Output covariances are defined as:

(7)

where, as before, y k is the output vector and e k Rl is a white

noise vector sequence. The left-hand side is called the AutoRegressive AR part and the right-hand side the Moving Average
MA part, hence the name of the model. The matrices i Rll
are the AR matrix parameters; matrices i Rll are the MA matrix parameters. Sometimes, as in the present case of multiple
outputs, one speaks of ARMAV models as to stress their multivariable character. An ARMA model that is deduced from a statespace model has the same AR order as MA order. This is denoted
as p in Eq. 7. The ARMA model order is related to the statespace model order as: pln. Since it is derived from a stochastic
state-space model, also an ARMA model can truly represent a
vibrating structure.
The modal parameters can be computed from the ARMA model
by the eigenvalue decomposition of the companion matrix of the
AR polynomial:
I

S y z C zIA 1 GR 0 G T z 1 IA T 1 C T ze j t (9)

nn

i e i t ;

Frequency-Domain Model. Many identification methods identify frequency-domain models from samples of the Fourier transform of the measurement signals. Frequency-domain models are
readily obtained by applying the z-transform to the discrete-time
models 2, 7. For instance, the output power spectrum matrix of
a state-space system can be written as 14:

(4)

is the eigenvector matrix and d C

is a diwhere C
agonal matrix containing the discrete-time eigenvalues. The
eigenfrequencies i and damping ratios i are found from:
nn

The observed mode shapes V are the first l rows of the eigenvector matrix. As in the state-space case, the eigenfrequencies and
damping ratios can be computed from the discrete eigenvalues in
d ; see Eq. 5.

R i E y ki y Tk lim
N

1
N

N1

k0

y ki y Tk

(12)

where the second equation follows from the ergodicity assumption. Of course, in reality, a finite number N of data is available
and a covariance estimate is simply obtained by dropping the limit
in 12.
The identification methods of Section 3 require a frequency
domain representation of the output signals. The frequencydomain representation of stochastic signals is provided by the
power spectrum S y Cll , defined as the discrete-time Fourier
transform of the covariance sequence:

S y e

j t

R k e j kt

(13)

For details on methods to estimate covariances and spectra

from measured time data, we refer to the extensive literature that
exist on the subject; see for instance 15,16.

Frequency-Domain Spectrum-Driven Methods

The presentation order of the identification methods roughly

corresponds to the historical application of stochastic system identification: from picking the peaks in spectrum plots Section 3.1
Transactions of the ASME

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

to time-domain subspace methods that make extensively used of

concepts from numerical linear algebra Section 5.1.
3.1 The Peak-Picking Method PP. The simplest approach to estimate the modal parameters of a structure subjected
to ambient loading is the so-called Peak-Picking PP method. The
method is named after the key step of the method: the identification of the eigenfrequencies as the peaks of a spectrum plot. The
method is for instance discussed in 15,17.
Under the conditions of low damping and well-separated eigenfrequencies, the spectrum 11 around an eigenfrequency i can
be approximated by:
S y j i i i iH

(14)

where i is a scale factor depending on the damping ratio, the

eigenfrequency, the modal participation factor and the input covariance matrix. Vector i is the ith mode shape. From 14 it is
clear that each column or row of the spectrum matrix at an eigenfrequency can be considered as an estimate of the mode shape at
that frequency. Therefore it suffices to compute only 1 column or
row of the spectrum matrix. In order to obtain damping ratios, it is
often suggested to use the half-power bandwidth method, which is
able to quantify the sharpness of a resonance peak. It is, however,
widely accepted that this estimate is not a very accurate one.
Some refinements of the PP method exist. The coherence function between two channels tends to go to one at the resonance
frequencies because of the high signal-to-noise ratio at these frequencies. Consequently, inspecting the coherence function can assist in selecting the eigenfrequencies. Also, the phase angles of the
cross spectra are helpful: if real modes are expected, the phase
angles should be either 0 or 180 deg at the resonance frequencies.
A violation of the basic assumptions low damping and wellseparated 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. Other
disadvantages are that the selection of the eigenfrequencies can
become a subjective task if the spectrum peaks are not very clear
and that the eigenfrequencies have to be a subset of the discrete
frequency values of the discrete Fourier transform.
Despite these drawbacks many civil engineering cases exist
where the method is successfully applied; see for instance 18,19.
The popularity of the method is due to its implementation simplicity and its speed: the only algorithm that is needed is the Fast
Fourier Transform FFT to convert time data to spectra. The computational speed can become irrelevant though, because of the
large amount of user interaction needed to try to improve the
estimation results.
3.2 The Complex Mode Indication Function CMIF. A
more advanced method consists of computing the Singular Value
Decomposition SVD of the spectrum matrix:
S y j U j j U H j

(15)

where UCll is a complex unitary matrix containing the singular vectors as columns. The diagonal matrix Rll contains the
real positive singular values in descending order. This method
based upon the diagonalization of the spectral density matrix, as
it was called, was already used in the beginning of the eighties to
obtain the modes of a vibrating system subjected to natural excitation 20. Some years later, the method was also applied to Frequency Response Functions FRFs and became known as the
Complex Mode Indication Function CMIF. As suggested by the
name, the CMIF was originally intended as a tool to count the
number of modes that is present in measurement data. As a useful
byproduct, the CMIF also identifies the modal parameters from
FRFs 21. Recently, the spectrum-driven method received again
attention as an alternative for the PP method in civil engineering
applications 22. In this paper, the old method was given a new
name, the frequency-domain decomposition method.
Journal of Dynamic Systems, Measurement, and Control

Fig. 1 The complex mode indication function CMIF. The singular values of the spectrum matrix are plotted as a function of
the frequency. Around 2.4 Hz and 7 Hz, two singular values are
significant, indicating that there are two close modes.

The method is based on the fact that the transfer function or

spectrum matrix evaluated at a certain frequency is only determined by a few modes. The number of significantly contributing
modes determines the rank of the spectrum matrix. The SVD is
typically used for estimating the rank of a matrix: the number of
nonzero singular values equals the rank 23. The singular values
as a function of frequency ( j ) is the actual CMIF; see Fig. 1
for an example. Therefore, plotting the CMIF yields the eigenfrequencies as local maxima. The CMIF is also able to detect closely
spaced modes: more than one singular value will reach a local
maximum around the close eigenfrequencies.
If only 1 mode is important at a certain eigenfrequency i , the
spectrum approximates a rank-one matrix and can be decomposed
as 15:
S y j i 1 j i u 1 j i u H
1 j i

(16)

By comparing 16 with 14, it is evident that the first singular

vector at resonance is an estimate of the mode shape at that frequency. In case of mode multiplicity at a resonance frequency,
every singular vector corresponding to a nonzero singular value is
considered as a mode shape estimate.
In some sense, the CMIF method can be considered as an SVD
extension of the PP method. The SVD is able to resolve mode
multiplicity. The method can also be applied to a reduced spectrum matrix, where only the spectra between a chosen set of reference sensors and all outputs have to be computed. In this case,
the maximum number of detectable multiple poles cannot exceed
the smallest dimension of the reduced spectrum matrix.
Extensions of the CMIF method are possible that do estimate
eigenfrequencies and damping ratios differently as in the PP
method. After applying the SVD to the spectrum matrix, this matrix is in fact decomposed in single-DOF systems. To such a system, single-DOF modal parameter estimation methods could be
applied, extensively documented in the modal analysis literature
1 4.
3.3 Maximum Likelihood Identification ML. Contrary
to the PP method or the CMIF which considers only one mode at
a time, this method estimates the parameterized spectrum matrix
as a whole. Maximum Likelihood ML identification is an
optimization-based method that estimates the parameters of a
model by minimizing an error norm. A discussion on the use of
the ML estimator to identify parametric frequency-domain models
DECEMBER 2001, Vol. 123 661

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

can be found in 24,25. The ML method results in equations that

are nonlinear in the unknown parameters. This requires an iterative procedure. Therefore it is no surprise that often mentioned
drawbacks of ML estimators are the high computational load and
the fact that they are not suited to handle large amounts of data.
During the last years attention has been paid to the optimization of
the ML method: the algorithm has been modified to keep the
memory requirements as low as possible; and using an adapted
parameterization and fast signal processing techniques, an important reduction of the computation time was possible. It has been
shown that ML identification is a robust method to find the modal
parameters of a structure from a large and noisy data set 26,27.
Originally intended for application to FRFs, the method was extended to use spectra as primary data, so that it also could be used
in output-only cases 28.

Time-Domain Covariance-Driven Methods

It is known for some time that there exist similar mathematical

expressions for impulse responses and output covariances of a
system excited by white noise as a function of the system parameters; see for instance 12,15. In modal analysis applications, this
observation is used to feed classical impulse response based
modal parameter estimation methods with output covariances instead 29. Two such methods are discussed below.
4.1 The Instrumental Variable Method IV. In this section, a method belonging to the class of so-called Instrumental
Variable IV methods will be discussed. Although derived in a
different way, the final equations of the IV method correspond to
the Polyreference Time Domain PTD method after substituting
impulse responses by output covariances. The PTD method is
probably the most widely used traditional modal parameter estimation method. It contains the Least Squares Complex Exponential LSCE and the Ibrahim Time Domain ITD methods as
special cases. For an overview, relations between these traditional
input-output methods and the original references, see
1,3,30,31. A more generic discussion and more references on IV
methods can be found in 10.
An ARMA model 7 of suitable order can represent a vibrating
structure. Unfortunately, the application of a classical prediction
error method 10 to an ARMA model results in a highly nonlinear parameter estimation problem; see also Section 5.2. The nonlinearity is caused by the MA parameters. The advantage of the IV
method is that it identifies only the AR parameters and that this is
achieved in a linear way, while the underlying model structure
still is an ARMA model.
The idea of system identification is to fit a model to measured data y k . A good parameter estimation method should extract
the maximum information from the data, leaving residuals e k that
are uncorrelated with past data. This is formally written as:
T
T
i0 : E e k y ki
E e k E y ki
0

(17)

where the first equality says that e k and y ki are uncorrelated; and
the second equality follows from the zero-mean property of the
noise sequence. If, on the contrary the residuals are correlated
with past data, they still contain useful but unmodelled information and the model is not ideal. The derivation of the IV method
starts by imposing conditions like 17 to the ARMA model 7 in
order to get rid of the right-hand side the MA part. The oldest
noise term is e k p ; so by post-multiplying the ARMA model by
T
y k
pi for i0 and by taking the expectation we obtain:
T
T
i0 : E y k y kpi
1 E y k1 y kpi

T
p E y kp y kpi
0

(18)

T
Because of stationarity and Eq. 12, we have: E y k y ki

E y ki y Tk R i ;and the basic IV equation can be written in

terms of the output covariances R i :

662 Vol. 123, DECEMBER 2001

i0 : R pi 1 R pi1 p R i 0

(19)

By replacing the output covariances by their estimates and writing down the equation for all available time lags i, the AR parameters 1 , . . . , p can be estimated by solving the resulting overdetermined set of equations in a least squares sense. Finally, the
eigenvalues and the observed mode shapes are obtained from the
eigenvalue decomposition of the companion matrix of the AR
coefficients; see Eq. 8.
As in the previous methods of Section 3, it is also in the case of
the IV method possible to reduce the dimensions of the involved
matrices and the related computational effort by making use of a
subset of reference sensors. Only the covariances between all outputs and this subset have to be computed. Details can be found in
7. This corresponds to classical modal analysis, where the impulse response matrices are rectangular matrices having l rows
i.e., the number of outputs and m columns i.e., the number of
inputs. In output-only cases, the impulse responses are substituted by output covariances and the inputs by the reference outputs; see also 8,29.
A typical problem of estimating a parametric model from data
is the determination of the model order. A pth order ARMA model
based on l outputs contains pl poles. Consequently, the expected number of poles covered by the data gives an indication
of the model order. This expected number can be based on physical insight or counted as two times the number of peaks in the
frequency-plot of a nonparametric spectrum estimate; see also the
PP method, Section 3.1. A more accurate model order estimate is
provided by the CMIF, a frequency-plot of the singular values of
a nonparametric spectrum estimate Section 3.2.
More formal procedures estimate models of different order and
compare these models according to a quality criterion such as
Akaikes Final Prediction Error or Rissanens Minimum Description Length criterion 10.
However, in modal analysis one is usually not interested in a
good model as such, but rather in the modal parameters extracted
from that model. Practical experience with parametric models in
modal analysis applications learned that it is better to over-specify
the model order and to eliminate spurious numerical poles afterwards, so that only true physical system poles are left. The famous
stabilization diagram 1,3 is a great tool to achieve this goal. The
poles corresponding to a certain model order are compared to the
poles of a one-order-lower model. If the eigenfrequency, the
damping ratio and the related mode shape or modal participation
factor differences are within preset limits, the pole is labeled as a
stable one. The spurious numerical poles will not stabilize at all
during this process and can be sorted out of the modal parameter
data set more easily. Such a stabilization diagram is represented in
Fig. 2.
Interesting to note and very relevant for civil engineering practice is that the IV method is robust against nonstationary inputs
i.e., white noise with time-varying covariances. This does not
only follow from practical experience but has also been theoretically proven in 32.

4.2 Covariance-Driven Stochastic Subspace Identification

SSI-COV. Like the CMIF method can be considered as an
SVD-enhanced PP method, covariance-driven subspace identification cansomewhat disrespectfullybe considered as an SVDenhanced instrumental variable method. The covariance-driven
Stochastic Subspace Identification method SSI-COV is addressing the so-called stochastic realization problem, i.e., the problem
of identifying a stochastic state-space model 2 from output-only
data.
Stochastic realization is closely related to deterministic inputoutput realization, that goes back to Ho and Kalman 33 and was
Transactions of the ASME

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

Fig. 2 Stabilization diagram obtained with the IV method. The used symbols are: for
a stable pole; .v for a pole with stable frequency and vector; .d for a pole with stable
frequency and damping; .f for a pole with stable frequency and . for a new pole. Two
zooms are added that concentrate on the close modes around 2.4 and 7 Hz.

extended with the SVD to treat noisy data in 34 and 35. The
so-called Eigensystem Realization Algorithm ERA, developed
by Juang 36,37, is a modal analysis application of these deterministic realization algorithms. The stochastic output-only realization problem is solved in 12,38 40. Application of stochastic
realization to modal parameter estimation was reported by Benveniste and Fuchs 32. They also proved that their algorithm is
robust against nonstationary inputs e.g., a white noise sequence
with time-varying covariance.
Stochastic realization relies upon a fundamental property of stochastic state-space systems. It can be proven see for instance
41,42 that the output covariances R i 12 can be decomposed
as:
R i CA i1 G

(20)

where G is defined in 10. The covariance sequence R i can be

estimated from measurement data; so it remains to be solved how
to decompose the covariances as to find A, C, G. Hereto, a block
Toeplitz matrix T 1 i Rlili is formed that consists of covariances:

T 1i

Ri

R i1

R i1

Ri

R 2i1

C
CA

CA i1

R 2i2

R1

R2

Ri

A i1 GAG G lnO i i

(21)

The second equality follows from applying property 20. The

definitions of the extended observability matrix O i Rlin and the
reversed extended stochastic controllability matrix i Rnli are
obvious from 21. For lin, and in case of an observable and
Journal of Dynamic Systems, Measurement, and Control

controllable system, the rank of the lili Toeplitz matrix equals

n, since it is the product of a matrix with n columns and a matrix
with n rows.
The actual implementation of SSI-COV consists of estimating
the covariances R i , computing the SVD of T 1 i , truncate the SVD
to the model order n, estimating O i and i by splitting the SVD in
two parts and finally estimating A, C, G from O i and i . The
modal parameters are found from A and C as indicated in 4 and
6. Implementation details can be found in the above-cited references. Again, it is possible to slightly reformulate the SSI-COV
method, so that it only needs the covariances between all outputs
and a set of references 7,8.
In theory, the system order n can be determined by inspecting
the number of nonzero singular values of T 1 i 21. In practice,
however, the estimated covariance Toeplitz matrix is affected by
noise leading to singular values that are all different from zero.
As typical noise sources we have:
Modelling inaccuracies. It is possible that the true system that
generated the data cannot be modelled exactly as a stochastic
state-space model. An attempt to model this system by a
state-space model introduces an error in these cases.
Measurement noise: introduced by the sensors and the electronics of the measurement hardware.
Computational noise due to the finite precision of any
computer.
The finite number of data. The covariances have to be estimated, so that the factorisation property 20 does not hold
exactly. As a consequence, the rank of the covariance Toeplitz
matrix will not be exactly n.
Sometimes it is suggested to look at the gap between two
successive singular values. The singular value where the maximal
gap occurs yields the model order. This criterion should, however,
not be applied too dogmatically, since in most practical cases
there is no gap. Other statistical methods to determine the model
order are discussed in 43.
DECEMBER 2001, Vol. 123 663

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

Fig. 3 Stabilization diagram obtained with the SSI-COV method. By comparing this diagram with the IV diagram Fig. 2, it is clear that the IV method requires higher model
orders to find stable poles.

However, to obtain a good model for modal analysis applications, it is probably a better idea to construct a stabilization diagram, by identifying a whole set of models with different order.
The stabilization diagram was already introduced in Section 4.1.
In case of the SSI-COV method, an efficient construction of the
stabilization diagram is achieved by computing the SVD of the
covariance Toeplitz matrix only once. Models of different order
are then obtained by including a different number of singular values and vectors in the computation of O i and i . A stabilization
diagram obtained with the SSI-COV method is shown in Fig. 3.

Time-Domain Data-Driven Methods

5.1 Data-Driven Stochastic Subspace Identification SSIDATA. Recently, a lot of research effort in the system identification community was spent to subspace identification as evidenced by the book of Van Overschee and De Moor 42 and the
second edition of Ljungs book 10. Subspace methods identify
state-space models from input and output data by applying robust numerical techniques such as QR factorization, SVD and
least squares. As opposed to SSI-COV, the DATA-driven Stochastic Subspace Identification method SSI-DATA avoids the computation of covariances between the outputs. It is replaced by
projecting the row space of future outputs into the row space of
past outputs. In fact, the notions covariances and projections are
closely related. They both are aimed to cancel out the uncorrelated noise. The first SSI-DATA algorithms can be found in 44.
A general overview of data-driven subspace identification both
deterministic and stochastic is provided in 42.
Although somewhat more involved as compared to previous
methods, it is also possible with SSI-DATA to reduce the dimensions of the matrices and the memory requirements of the algorithm by introducing the idea of the reference sensors. This is
demonstrated in 7,11. It is beyond the scope of this paper to
explain the SSI-DATA method in detail. The interested reader is
referred to the above-cited literature.
664 Vol. 123, DECEMBER 2001

Covariance-Driven Versus Data-Driven Subspace Identification.

At this point it is useful to discuss the similarities and differences
between the SSI-COV Section 4.2 and the SSI-DATA method
Section 5.1. First the similarities. Both methods start with a data
reduction step. In the SSI-COV algorithm the raw time histories
y k , consisting of l channels and N data points with N, are
converted to the covariances of the Toeplitz matrix 21. The number of elements is reduced from lN to lili. In the SSI-DATA
algorithm a similar reduction is obtained by projecting the row
space of the future outputs into the row space of the past outputs.
This projection is computed from the QR factorization of a big
data Hankel matrix; see 7,11,42. A significant data reduction is
obtained because only a part of the R factor is needed in the
sequel of the algorithm. Both methods then proceed with an SVD.
This decomposition reveals the order of the system and the column space of O i 21.
Several variants of stochastic subspace identification exist.
They differ in the weighting of the data matrices before the application of the SVD. This weighting determines the state-space basis in which the identified model will be identified. Equivalent
implementations exist for both SSI-COV and SSI-DATA. More
details can be found in 40 and 42. Well-known variants are
Canonical Variate Analysis CVA, Principal Components PC or
Unweighted Principal Components UPC. This last variant is
sometimes also called Balanced Realization BR.
There are also differences between the covariance-driven and
data-driven approaches. In the SSI-COV method, the covariances
can be computed in a very fast way by using the FFT algorithm
15,16. The corresponding step in SSI-DATA is the relatively
slow QR factorisation. Therefore SSI-COV is much faster than
SSI-DATA. In favor of the data-driven method is that it is implemented as a numerically robust square root algorithm: the output
data is not squared up as in the covariance-driven algorithm.
We should add that in practical applications of any of the variants of stochastic subspace identification, no accuracy differences
can be observed when looking at the identified modal parameters.
Transactions of the ASME

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

5.2

Other Data-Driven Methods

The prediction Error Method Applied to an ARMA Model.

Prediction Error Methods PEM can be considered as a general
system identification framework 10. These methods identify the
parameters of a model by minimizing the so-called prediction errors. The straightforward application of PEM to estimate an
ARMA model 7 from data results in a highly nonlinear optimization problem with related problems as: convergence not being
guaranteed, local minima, sensitivity to initial values and a high
computational load. In contrast to the nonlinear frequency-domain
ML method see Section 3.3, nonlinear time-domain methods
such as PEM applied to an ARMA model never reached an
acceptable level of robustness and applicability for real-life data
45 47. Despite these drawbacks, some authors tried to apply the
PEM to identify the modal parameters of civil engineering structures; see for instance 48,49.
The Prediction Error Method Applied to an AR Model. The
nonlinearity of the PEM is caused by the MA part of the ARMA
model 7. By omitting the moving-average part, an autoregressive model is obtained:
y k 1 y k1 p y kp e k

(22)

and the PEM reduces to a linear least squares problem, which is

easily solved. Unfortunately, a pth-order AR model is not an
equivalent representation of a vibrating structure with pl modes.
The use of an AR model as a substitution of an ARMA model can
only be justified if the AR model order goes to infinity: p
10. In practice this means that many spurious poles will be introduced that need to be separated from the true system poles. The
use of AR models for modal parameter estimation is, for instance,
demonstrated in 50,51.

Experimental Comparison

In this section, the accuracy of the system identification results

of the different methods in terms of the modal parameters are
compared by means of a Monte-Carlo analysis consisting of 100
simulation runs.
The mast structure, shown in Fig. 4, is subjected to independent
white noise inputs at all horizontal translation DOFs. The re-

Fig. 4 FE model of the mast structure used in the Monte-Carlo

analysis

Journal of Dynamic Systems, Measurement, and Control

Fig. 5 Eigenfrequency estimation results from 100 MonteCarlo simulations. The estimates are divided by the true values
a value of 1 on the graphs indicates a perfect estimate. These
relative frequencies are shown as dots. The scatter of this
quantity gives an idea about the variance of the estimate. The
average estimate is also shown as a dashed line. The deviation of this quantity from 1 full line corresponds to the bias of
the estimate. The rows show the 3 modes; the columns represent the results of 3 identification methods: PP, IV and SSIDATA.

sponses at 6 horizontal DOFs are simulated and afterwards contaminated by white measurement noise with N/S10% N/S is
the ratio of the rms values of the noise and output signal. The
noisy outputs are then fed to 5 system identification methods: PP
Section 3.1, CMIF Section 3.2, IV Section 4.1, SSI-COV

Fig. 6 Damping ratio estimation results from 100 Monte-Carlo

simulations. The estimates are divided by the true values a
value of 1 on the graphs indicates a perfect estimate. These
relative damping ratios are shown as dots. The scatter of this
quantity gives an idea about the variance of the estimate. The
average estimate is also shown as a dashed line. The deviation of this quantity from 1 full line corresponds to the bias of
the estimate. The rows show the 3 modes; the columns represent the results of 3 identification methods; PP, IV and SSIDATA.

DECEMBER 2001, Vol. 123 665

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

Fig. 7 Mode shape estimation results from 100 Monte-Carlo

simulations. The correlation between the estimated and the
true mode shapes are shown as dots. The average correlation
is also shown as a dashed line. The rows show the 3 modes;
the columns represent the results of 4 identification methods:
PP, CMIF, IV and SSI-DATA. The scaling of the y-axis varies in
vertical direction to accommodate to the changing estimation
quality of the different modes, but not in horizontal direction,
allowing an easy comparison of the methods.

Section 4.2 and SSI-DATA Section 5.1. More details about the
structure and the simulations can be found in 7.
Figures 13, which were already introduced previously, contain
intermediate identification results of CMIF, IV and SSI-COV. As
explained in Sections 4.1 and 4.2, no order selection criteria as
such is applied. The stable poles are selected from the stabilization
diagrams and they dont have to originate from one model. The
modal parameter estimation results for the first three modes are
represented in Figs. 57. In our discussion of the CMIF method,
we did not include an alternative frequency or damping estimation
procedures as compared to the PP method. The only difference is
that the CMIF can detect closely spaced modes and finds the
eigenfrequencies in a more objective way. Therefore the CMIF
frequencies and damping ratios are not presented in the figures.
The results of SSI-COV and SSI-DATA are so close to each other,
that only SSI-DATA is presented.
The eigenfrequency estimates of the PP method can only take
the discrete values determined by the frequency resolution of the
spectrum Fig. 5. All methods yield unbiased eigenfrequency estimates. Although still small, the standard deviation of the PP
estimates is three times higher than for the other methods.
When looking at the damping estimates Fig. 6, the high bias
of the PP damping estimates is striking. It is rather a coincidence
that mode 2 and 3 have unbiased damping estimates, as the
situation changes when choosing different options for the
non-parametric spectrum estimate resolution, window, overlap,
averages.
Regarding the mode shape estimates Fig. 7, the IV estimates
for the first mode are too bad to fit into the scales. Also the
average correlation of the PP estimates of the third mode could
not be represented. The subspace methods clearly outperform the
others.

Conclusions

This paper reviewed stochastic system identification methods

for operational modal analysis.
The basic peak-picking method PP finds the eigenfrequencies
as the peaks of non-parametric spectrum estimates. This frequency
666 Vol. 123, DECEMBER 2001

selection procedure becomes a subjective task in case of noisy

operational data, weakly excited modes and relatively close eigenfrequencies. The related half-power bandwidth damping estimation method is unreliable; and operational deflection shapes are
identified instead of mode shapes.
The complex mode indication function CMIF is an SVDextension of the PP method, allowing for a more objective selection of the eigenfrequencies and the identification of closely
spaced modes. In seems however that the mode shape estimation
quality depends on the selected singular vector around resonance
and that it is not always the vector at resonance that gives the
best estimates.
The parametric methods IV, SSI-COV, SSI-DATA share the
advantage that stabilization diagrams can be constructed by identifying parametric models of increasing order. These diagrams are
very valuable in separating the true system poles from the spurious numerical poles.
The instrumental-variable method IV does not involve an
SVD and consequently suffers from the lack of a noise-truncating
mechanism. This is reflected in the fact that the mode shape estimates are less accurate than in the subspace methods and that
higher order models are required to obtain good modal parameter
estimates. A lot of additional poles are necessary for fitting the
noise.
Both covariance- SSI-COV and data-driven subspace methods
SSI-DATA seem to perform equally well concerning modal parameter estimation performance, although theoretically the numerical behavior of SSI-DATA should be better than that of SSICOV since it avoids to square up the data. The SSI-COV method
is considerably faster than the SSI-DATA method if its datareduction step is carried out by the FFT, whereas SSI-DATA requires a slower QR factorization step. Evidently, because it only
uses linear numerical algorithms, the SSI-DATA method is still
much faster than non-linear prediction error methods that are
sometimes proposed to estimate the modal parameters from operational data.
As was indicated in this paper, a lot of classical input-output
methods carry over after some modifications to the output-only
case. FRF-driven methods can be converted to spectrum-driven
methods; impulse-response-driven methods are almost identical to
output-covariance-driven methods and input-output data-driven
methods are very similar to output-only data-driven methods.

References
1 Heylen, W., Lammens, S., and Sas, P., 1995, Modal Analysis Theory and
Testing, Department of Mechanical Engineering, Katholieke Universiteit Leuven, Leuven, Belgium.
2 Maia, N. M. M., Silva, J. M. M., He, J., Lieven, N. A. J., Lin, R. M., Skingle,
G. W., To, W.-M., and Urgueira, A. P. V., 1997, Theoretical and Experimental
Modal Analysis, Research Studies Press, Taunton, Somerset, UK.
3 Allemang, R. J., 1999, Vibrations: Experimental Modal Analysis, Course
Notes, Seventh Edition, Structural Dynamics Research Laboratory, University
of Cincinnati, OH. http://www.sdrl.uc.edu/course info.html.
4 Ewins, D. J., 2000, Modal Testing: Theory, Practice and Application, Research
Studies Press, Baldock, Hertfordshire, UK.
5 Abdelghani, M., Verhaegen, M., Van Overschee, P., and De Moor, B., 1998,
Comparison study of subspace identification methods applied to flexible
structures, Mech. Syst. Signal Process., 12, No. 5, pp. 679 692.
6 Petsounis, K. A., and Fassois, S. D., 2001, Critical comparison and assessment of parametric time domain methods for the identification of vibrating
structures. Mech. Syst. Signal Process., accepted for publication.
7 Peeters, B., 2000, System Identification and Damage Detection in Civil Engineering, PhD thesis, Department of Civil Engineering, Katholieke Universiteit Leuven, Belgium, http://www.bwk.kuleuven.ac.be/bwm, Dec.
8 Hermans, L., and Van der Auweraer, H., 1999, Modal testing and analysis of
structures under operational conditions: industrial applications, Mech. Syst.
Signal Process., 13, No. 2, pp. 193216.
9 Peeters, B., and De Roeck, G., 2001, One-year monitoring of the Z24-Bridge:
environmental effects versus damage events, Earthquake Eng. Struct. Dyn.,
30, No. 2, 149171.
10 Ljung, L., 1999, System Identification: Theory for the User, Second Edition,
Prentice-Hall, Upper Saddle River, NJ.
11 Peeters, B., and De Roeck, G., 1999, Reference-based stochastic subspace
identification for output-only modal analysis, Mech. Syst. Signal Process., 13,
No. 6, pp. 855 878.

Transactions of the ASME

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

12 Akaike, H., 1974, Markovian representation of stochastic processes and its

application to the analysis of autoregressive moving average processes, Annals of the Institute of Statistical Mathematics, 26, pp. 363387.
13 Basseville, M., and Nikiforov I. V., 1993, Detection of Abrupt Changes:
Theory and Applications, Prentice-Hall, Englewood Cliffs, NJ. http://
www.irisa.fr/sigma2/kniga.
14 Caines, P., 1998, Linear Stochastic Systems, Wiley, New York.
15 Bendat, J. S. and Piersol, A. G., 1993, Engineering Applications of Correlation
and Spectral Analysis, Second Edition, Wiley, New York.
16 Stoica, P. and Moses, R. L., 1997, Introduction to Spectral Analysis, PrenticeHall, Englewood Cliffs, NJ.
17 Felber, A. J., 1993, Development of a Hybrid Bridge Evaluation System,
PhD thesis, University of British Columbia, Vancouver, Canada.
18 Felber, A. J. and Cantieni, R., 1996, Introduction of a New Ambient Vibration
Testing System-Description of the System and Seven Bridge Tests, Internal
Report 156521, EMPA, Dubendorf, Switzerland.
19 Cunha, A., Caetano, E., Calcada, R., and Delgado, R., 1999, Modal identification and correlation with finite element parameters of Vasco da Gama
Bridge, Proceedings of IMAC 17, pp. 705711, Kissimmee, FL, Feb.
20 Prevosto, M., 1982, Algorithmes dIdentification des Caracteristiques Vibratoires de Structures Mecaniques Complexes, PhD thesis, Universite de
Rennes I, France, Mar.
21 Shih, C. Y., Tsuei, Y. G., Allemang, R. J., and Brown, D. L., 1988, Complex
mode indication function and its application to spatial domain parameter estimation, Mech. Syst. Signal Process., 2, No. 4, pp. 367377.
22 Brincker, R., Zhang, L., and Andersen, P., 2000, Modal identification from
ambient responses using frequency domain decomposition, Proceedings of
IMAC 18, the International Modal Analysis Conference, pp. 625 630, San
Antonio, TX, Feb.
23 Golub, G. H. and Van Loan, C. F., 1996, Matrix Computations, Third Edition,
The Johns Hopkins University Press, Baltimore, MD.
24 Schoukens J. and Pintelon R., 1991, Identification of Linear Systems: a Practical Guideline to Accurate Modelling, Pergamon Press, London, UK.
25 Pintelon, R., Guillaume, P., Rolain, Y., Schoukens, J., and Van Hamme, H.,
1994, Parametric identification of transfer functions in the frequency
domain-a survey, IEEE Trans. Autom. Control, AC-39, No. 11, pp. 2245
2260.
26 Guillaume P., Verboven P., and Vanlanduit S., 1998, Frequency-domain maximum likelihood identification of modal parameters with confidence intervals,
Proceedings of ISMA 23, Noise and Vibration Engineering, K. U. Leuven,
Belgium.
27 Vanlanduit, S, Guillaume P., and Schoukens, J., 1998, High spatial resolution
modal parameter estimation using a parametric MLE-like algorithm, Proceedings of ISMA 23, Noise and Vibration Engineering, K. U. Leuven,
Belgium.
28 Hermans, L., Guillaume, P., and Van der Auweraer, H., 1998, A frequencydomain maximum likelihood approach for the extraction of modal parameters
from output-only data, Proceedings of ISMA 23, Noise and Vibration Engineering, K. U. Leuven, Belgium.
29 James, G. H., Carne, T. G., and Lauffer, J. P., 1995, The natural excitation
technique NExT for modal parameter extraction from operating structures,
Int. J. Anal. Exp. Modal Anal., 10, No. 4, pp. 260277.
30 Leuridan, J., 1984, Some Direct Parameter Model Identification Methods
Applicable for Multiple Input Modal Analysis, PhD thesis, University of
Cincinnati, OH.

Journal of Dynamic Systems, Measurement, and Control

31 Allemang, R. J., Brown, D. L., and Fladung, W. A., 1994, Modal parameter
estimation: a unified matrix polynomial approach, Proceedings of IMAC 12,
the International Modal Analysis Conference, pp. 501514, Honolulu, HI.
32 Benveniste, A., and Fuchs, J.-J., 1985, Single sample modal identification of
a nonstationary stochastic process, IEEE Trans. Autom. Control, AC-30, No.
1, pp. 66 74.
33 Ho, B. L., and Kalman, R. E., 1966, Effective construction of linear statevariable models from input/output data, Regelungstechnik, 14, pp. 545548.
34 Zeiger, H. P., and McEwen, A. J., 1974, Approximate linear realization of
given dimension via Hos algorithm, IEEE Trans. Autom. Control, AC-19,
No. 2, pp. 53.
35 Kung, S. Y., 1974, A new identification and model reduction algorithm via
singular value decomposition. Procedings of the 12th Asilomar Conference
on Circuits, Systems and Computers, pp. 705714, Asilomar, CA, USA, Nov.
36 Juang, J.-N., and Pappa, R. S., 1985, An eigensystem realization algorithm
for modal parameter identification and model reduction, J. Guid. Control
Dyn.
37 Juang, J.-N., 1994, Applied System Identification, Prentice Hall, Englewood
Cliffs, NJ.
38 Desai, U. B., Pal, D., and Kirkpatrick, R. D., 1985, A realization approach to
stochastic model reduction, Int. J. Control 42, pp. 821 839.
39 Aoki M., 1987, State Space Modelling of Time Series, Springer-Verlag, Berlin,
Germany.
40 Arun, K. S., and Kung, S. Y., 1990, Balanced approximation of stochastic
systems, SIAM J. Matrix Anal. Appl., 11, pp. 42 68.
41 Akaike, H., 1974, Stochastic theory of minimal realization, IEEE Trans.
Autom. Control, 19, pp. 667 674.
42 Van Overschee, P., and De Moor, B. 1996, Subspace Identification for Linear
Systems: Theory-Implementation-Applications, Kluwer Academic Publishers,
Dordrecht, The Netherlands.
43 Camba-Mendez, G., and Kapetanios, G., 2001, Testing the rank of the Hankel covariance matrix: a statistical approach, IEEE Trans. Autom. Control,
AC-46, No. 2, pp. 325330.
44 Van Overschee, P., and De Moor, B., 1993, Subspace algorithm for the stochastic identification problem, Automatica, 29, No. 3, pp. 649 660.
45 Viberg, M., 1995, Subspace-based methods for the identification of linear
time-invariant systems, Automatica, 31, No. 12, pp. 18351851.
46 Kirkegaard P. H., and Andersen, P., 1997, State space identification of civil
engineering structures from output measurements, Proceedings of IMAC 15,
The International Modal Analysis Conference, pp. 889 895, Orlando, FL.
47 Peeters, B., De Roeck, G., and Andersen, P., 1999, Stochastic system identification: uncertainty of the estimated modal parameters, Proceedings of
IMAC 17, The International Modal Analysis Conference, pp. 231237, Kissimmee, FL, Feb.
48 Piombo, B., Ciorcelli, E., Garibaldi, L., and Fasana, A., 1993, Structures
identification using ARMAV models, Proceedings of IMAC 11, the International Modal Analysis Conference, pp. 588 592, Orlando, FL.
49 Andersen, P., 1997, Identification of Civil Engineering Structures Using Vector ARMA Models, PhD thesis, Department of Building Technology and
Structural Engineering, Aalborg University, Denmark, May.
50 Pandit, S. M., 1991, Modal and Spectrum Analysis: Data Dependent Systems
in State Space, Wiley, New York.
51 De, Roeck G., Claesen, W., and Van den Broeck, P., 1995, DDS-methodology
applied to parameter identification of civil engineering structures, Vibration
and Noise 95, pp. 341353, Venice, Italy, Apr.

DECEMBER 2001, Vol. 123 667

Downloaded 17 Feb 2013 to 152.3.102.242. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm