Modal Identification through Ambient Vibration:

Comparative Study
Diego F. Giraldo1; Wei Song2; Shirley J. Dyke3; and Juan M. Caicedo4

Abstract: An analytical comparison between three techniques for the identification of modal properties of structures when subjected to
ambient vibrations is performed. The algorithms examined include the eigensystem realization algorithm with data correlations, the
prediction error method through least squares, and the stochastic subspace identification 共SSI兲 technique. Both analytical and experimental
data from a four-storey building scaled at 1:3 are used to perform these evaluations. The level of noise added to the simulated data is
varied to study the robustness of the techniques. All techniques are fully automated, allowing for assessments to be conducted through
Monte Carlo simulations. The results indicate that the SSI technique provides the most accurate identification of natural frequencies and
mode shapes even with high noise levels, all while requiring the least amount of experience for implementation.
DOI: 10.1061/共ASCE兲0733-9399共2009兲135:8共759兲
CE Database subject headings: Vibration; Modal analysis; Comparative studies; Structural analysis.

Introduction identification would succeed only if the inputs were broadband

white noise with a Gaussian distribution. Using different ap-
Identification of the global dynamic properties of civil engineer- proaches, Andersen 共1997兲 and Ibrahim et al. 共1996兲 reduced such
ing structures using vibration responses is a necessary step in concerns by showing that, as long as the input can be generated
several types of analysis including, for instance, model updating by filtering Gaussian white noise through a linear filter, the origi-
and structural health monitoring 共SHM兲. Although techniques nal system can be substituted by an alternative state-space repre-
based on forced testing approaches using impulsive loads or more sentation with Gaussian white noise excitation whose modal
general broadband excitations have long been used in several en- properties remain unchanged 共Andersen 1997兲. Experimentally,
gineering fields as reliable means to obtain modal properties, the Caicedo et al. 共2004兲 showed that the identification can be carried
use of these tests is not generally feasible in civil engineering out using ambient vibrations with questionable stationarity, as in
applications due to safety concerns and the massive size of these the case of some civil structures. However, limited bandwidth and
structures 共Wenzel and Pichler 2005兲. However, civil structures low amplitudes may severely handicap the identification in some
are invariably subjected to unmeasurable external loads such as cases. Therefore, the use of ambient responses may not be appli-
wind and soil vibrations. This factor, combined with the reduced cable in all cases.
cost and simplicity of such testing techniques, makes the use of Although a number of techniques that use ambient responses
ambient vibrations an attractive alternative for modal identifica- have been developed in the last four decades, our knowledge of
tion when considering large civil engineering structures 共Juang the relative advantages and disadvantages remains limited. Sev-
and Pappa 1985兲. eral writers have published comparisons that apply only to spe-
There are a number of challenges associated with the use of cific situations and Monte Carlo results are scarcely reported.
such ambient or operational responses because the statistical Lew et al. 共1993兲 compared the techniques derived from the
properties of the unknown inputs play a significant role in the eigensystem realization algorithm 共ERA兲 while trying to keep the
success or failure of the identification. It was believed that modal computational effort needed at the same level. Although this is an
attractive feature to use as a baseline, by restricting the methods
1 in this way, some techniques were at a disadvantage. Deforges
Graduate Engineer, Walter P. Moore and Associates, Inc., Las Vegas,
et al. 共1995兲 made a comparison in which only natural frequencies
NV 89119.
2
Doctoral Student and Research Assistant, Dept. of Civil Engineering, and damping ratios were considered. More recently, Kinkegaard
Washington Univ. in St. Louis, St. Louis, MO 63130 共corresponding and Andersen 共1997兲 compared three popular techniques using a
author兲. E-mail: [email protected] rather simple dynamic system. Other nonstatistical comparative
3
Edward C. Dicke Professor of Engineering, Dept. of Civil Engineer- studies have been published by Brinker et al. 共1996兲, Abdeghani
ing, Washington Univ. in St. Louis, St. Louis, MO 63130. E-mail: et al. 共1998兲, He et al. 共2006兲, and Peeters and Ventura 共2003兲. In
[email protected] general, complete comparisons are difficult to achieve due to the
4
Assistant Professor, Dept. of Civil and Environmental Engineering, amount of technical parameters needed to carry out the tech-
Univ. of South Carolina, Columbia, SC 29208. E-mail: caicedo@ niques. However, comparison studies are rarely redundant, pro-
engr.sc.edu
viding different insights into certain applications and conditions.
Note. This manuscript was submitted on November 28, 2006; ap-
proved on February 9, 2009; published online on July 15, 2009. Discus-
The main goal of this study is to carry out a statistical com-
sion period open until January 1, 2010; separate discussions must be parison, via Monte Carlo simulations facilitating consideration of
submitted for individual papers. This paper is part of the Journal of realistic randomness, of three cutting edge and somewhat popular
Engineering Mechanics, Vol. 135, No. 8, August 1, 2009. ©ASCE, ISSN modal identification techniques. Identification with each tech-
0733-9399/2009/8-759–770/$25.00. nique is obtained on the basis of best possible results achieved in

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 759

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 1. Benchmark structure

an entirely automated environment, while varying parameter sets third measured east-west accelerations. Ambient vibration was in-
for model orders regardless of the computational power required. duced by several factors such as wind, and ground excitation
The evaluations use analytically simulated acceleration records produced by traffic, working machinery, and people walking near
from a four-story benchmark problem developed by the IASCM- the structure. These records were taken for a period of 5 min with
ASCE Task Group on SHM. Unlike most previous comparative a sampling frequency of 200 Hz. Detailed information about
studies, the use of a finite-element 共FE兲 model allows us to com- the configuration of the structure can be found in Dyke et al.
pare the identified modal properties with the true properties ob- 共2003兲, http://wusceel.cive.wustl.edu/asce.shm/benchmarks.htm,
tained with the known mass, stiffness, and damping matrices of and Johnson et al. 共2000兲.
the model.
Although the advantages of modal identification using ambient
Analytical Model
vibrations of civil structures are undeniable, it is also important to
recognize the limitations of this approach. In an attempt to do so, 132 beam elements were used to construct the FE model 关see Fig.
experimental acceleration records from the scaled-down bench- 1共b兲兴. Geometry, cross-sectional areas, material properties, and
mark model is used here to identify its properties while exposed mass distribution of all members were based on the experimental
to typical ambient vibrations of civil structures. The identification model. Moreover, the mass associated with the steel plates at-
of the properties is performed using only the technique found to tached to the structure was distributed as lumped masses along
perform best in the statistical comparison of the simulated data the nodes. However, some assumptions were made to simplify the
and compared to those obtained with more traditional hammer analysis and reduce the computational effort needed to simulate
testing. It is important to emphasize that a comparison to exact the dynamic responses. For instance, the properties of the inter-
results is not possible here because the true dynamic properties of storey diagonals were purposely limited so that only axial forces
the structure are unknown. were developed on these members. In addition, fixed conditions
in all directions were provided at the nodes connected to the
ground. But perhaps the main assumption is that all floors behave
Benchmark Structure like rigid membranes with no in-plane deformation. This assump-
tion is based on the fact that, as seen in Fig. 1共c兲, 16 beam ele-
To enhance the understanding of SHM techniques, a task group of ments provide stiffness for each floor. Moreover, the steel plates
the IASC-ASCE developed a series of benchmark studies based attached to the experimental structure provided not only the in-
on a structure constructed at the University of British Columbia tended mass but also some additional stiffness to the plane. As a
共Dyke et al. 2003兲. Experimental data collected from the healthy result, all in-plane nodes were expected to have similar behavior.
structure and simulated data using an FE model are used herein With this in mind, displacements along each axis as well as rota-
for the examination of modal identification techniques. In the fol- tions with respect to the vertical axis of all external nodes were
lowing sections detailed descriptions of the experimental setup as constrained to be dependent on the central node. Rotations with
well as the analytical model are provided. respect to the x and y axes were allowed at all nodes. By con-
densing these degrees of freedom 共DOF兲, the FE model is simpli-
fied in such a way that 88 DOF are active. Finally, damping ratios
Experimental Setup
of 1% were introduced to each mode of the analytical model.
The structure is a four-story two-bay by two-bay steel frame, Ambient disturbance responses were simulated by introducing
scaled at 1:3 关see Fig. 1共a兲兴. It has a 2.5-m ⫻ 2.5-m plan and is three bandwidth-limited, statistically independent, and normally
3.6-m tall. Four steel plates were attached to each floor to repre- distributed random inputs acting as horizontal ground excitation
sent the mass of the building. Each floor was equipped with three to the model in the x and y directions as well as a rotation with
uniaxial sensors, two of which measured accelerations in the respect to the z axis. Similar to the experimental case, only three
north-south direction at opposite sides of the structure, and the acceleration responses per floor were recorded. Each simulation

760 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
Table 1. Theoretical Natural Frequencies and Damping Ratios
␻ = 冑a2 + b2 ;
a
␻d = b; ␨= 共4兲
Mode Frequency Damping ratio 冑a2 + b2
number Description 共Hz兲 共%兲
The mode shapes in physical coordinates can be obtained by mul-
1 East-West motion 7.827 1.0
tiplying the eigenvectors of the system matrix A by the output
2 North-South motion 8.397 1.0
matrix C.
3 Torsion 11.115 1.0 When a discretized system described by Eqs. 共1兲 and 共2兲 is
4 East-West motion 21.959 1.0 vibrating in free response 共i.e., Wk = 0兲 due to an unknown set of
5 North-South motion 23.763 1.0 initial conditions Z1, the states and output of the system at k = 1
6 Torsion 31.658 1.0 are given by Z2 = AZ1 and Y1 = CAZ1, respectively. In general,
the states and output of the system are known as the Markov
parameters and can be expressed as
Zk = Ak−1Z1 ; Yk = CAkZ1, for k = 1,2, . . . 共5兲
consisted of 6-min inputs from which data were obtained with a
Early developments on modal identification using time-domain
sampling frequency of 200 Hz. The first minute of each simula-
techniques rely heavily on free responses and Markov parameters.
tion was discarded to ensure that no transient response was being
Therefore, methods for estimating such responses using the vibra-
considered. Simulated sensor noise was added to each sensor, as
tions of civil structures as they operate under normal conditions
will be explained in later sections. The identification of the modal
have allowed researchers to implement several techniques. The
properties from this FE model is limited to the six lowest frequen-
most widely accepted of these methods is known as the natural
cies and the associated damping ratios and mode shapes. For
excitation technique 共NExT兲.
comparison purposes, the theoretical properties of the model were
calculated using its stiffness, mass, and damping matrices. These
parameters are hereafter referred to as the theoretical properties NExT
共see Table 1兲.
Although auto- and cross-correlation functions had already been
used to produce free decaying functions of systems subjected to
unknown inputs 共see for instance Bendat and Piersol 1980兲,
Time-Invariant Linear Systems James et al. 共1993兲 provided the mathematical background for
this practice. The technique is known as NExT and assumes that
Consider a linear system driven by unmeasurable stochastic dis- the excitation is weakly stationary, broadband, and uncorrelated
turbances and modeled with the discrete equation to prior system responses. Consider the n degree of freedom,
time-invariant, and linear system described by
Zk+1 = AZk + Wk 共1兲 MẌ共t兲 + CsẊ共t兲 + KX共t兲 = F共t兲 共6兲
where A = system matrix; Wk = unmeasurable disturbance; and Postmultiplying Eq. 共6兲 by the displacement of any of the DOF
Zk = vector known as the state vector, which may or may not have measured Xi共s兲 共referred to as the reference signal兲 and taking the
any physical meaning. An output equation compliments Eq. 共1兲 expected value of each side yield
and is expressed as
MRX¨ Xi共t,s兲 + CsRX˙ Xi共t,s兲 + KRXXi共t,s兲 = RFXi共t,s兲 共7兲

Yk = CZk + Vk 共2兲 where R共 兲 denotes the vector of correlation functions. For

weakly stationary processes it can be shown that 共Bendat and
where Yk = vector containing the measured outputs; C = output Piersol 1980兲
matrix; and Vk represents the noise in the system. Together, Eqs.
共1兲 and 共2兲 are known as a discrete state-space representation of ṘXXi共␶兲 = RX˙Xi共␶兲 = − RXX˙i共␶兲 共8兲
the system and can be easily analyzed assuming that the processes
Wk and Vk have zero mean, are statistically independent, and are where ␶ = t − s. Assuming that the displacement, velocity, and ac-
uncorrelated with their previous values. In other words, the cova- celeration processes are weakly stationary and uncorrelated with
riance matrices can be expressed as future disturbances 关i.e., RFXi共t , s兲 = 0兴 and taking the fourth de-
rivative of Eq. 共7兲, we obtain

E 冋冉 冊
Wp
Vp
共WTq 册冋 册
VTq 兲 =
Q S
ST R
␦ pq 共3兲
MR̈X¨X¨i共␶兲 + CsṘX¨X¨i共␶兲 + KRX¨X¨i共␶兲 = 0
which shows that the cross-correlation function of the responses
共9兲

where E关 兴 represents the expectance operator; ␦ = Kronecker of the structure with a reference signal satisfies the homogeneous
delta; and p and q indicate discrete time steps. equation of motion and can be treated as free response data.
The modal properties of a linear dynamic system described by The successful implementation of NExT highly depends on
Eqs. 共1兲 and 共2兲 can be obtained from the eigensolution of the two factors, namely, 共1兲 the capability of the unknown input to
system matrix A and the output matrix C. In the case of civil excite all modes of the structure and 共2兲 the length of the records
structures that are typically lightly damped, the eigenvalues of A used. Because cross-spectral density functions are usually ob-
occur in complex conjugates. From a conjugate pair of complex tained by “windowing” data records and averaging results in the
eigenvalues, ␭ = a ⫾ bj, the damped and undamped natural fre- frequency domain, lengthy records provide more samples to be
quencies as well as the associated damping ratio can be obtained, averaged and, therefore, better results. Moreover, in terms of
respectively, as quality of results, no benefits are gained by using data that are

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 761

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Data Management
Automation of a modal identification methodology is challenging
for several reasons. For instance, the resulting identified modal
values often may not represent the true dynamic behavior of the
structure and may simply be computational modes and not true
physical modes of the system. This phenomenon is usually a con-
sequence of noisy sensors, overestimation of the model order, or
the inability of the unknown inputs to sufficiently excite a par-
ticular vibrational mode. In some cases, the identified modes have
clear signs of not being true modes of the system 共e.g., unexpect-
edly high damping ratios兲. To systematically eliminate these
modes, some criteria have to be established. The specific criteria
developed for the elimination of doubtful modes are discussed in
the sections of this article that are devoted to the numerical imple-
mentation of each algorithm. General knowledge of the system is
fundamental to carry out this process.
Fig. 2. Estimated free response with NExT
But perhaps the most challenging issue regarding automating
the modal identification process is to deal with those cases when
the identified properties are computational rather than physcial,
but do not have clear indications of this fact. Reported solutions
sampled at a much higher rate than needed. In contrast, the com- 共Pappa et al. 1998; Lanslots et al. 2004; Lau et al. 2007兲 in auto-
putational effort increases significantly as higher rates are used. mated modal identification provide robust results in terms of the
The main drawback in the implementation of NExT is related appearance of spurious modes, facilitating a series of numerical
to the fact that one sensor is selected as reference for the entire set statistical studies possible for complex systems. Most widely
of measurements. Several potential problems result from this re- adopted is a method for determining the order of the system based
quirement. For instance, when the reference sensor is located near on stabilization diagram, using rules that can be implemented in
a modal node 共or if the motion is orthogonal to the measurements an autonomous fashion. In this paper, such an approach is applied
to determine the true modes of the systems. Because no actual
of the sensor兲, the estimated free response contains little or no
diagrams are required to automate this process, it is referred here
information of that mode. Moreover, for every channel selected as
as “stabilization of modes.”
reference, a different set of free responses is obtained. To over- A stabilization diagram is a useful tool with which the user can
come these problems, modal identification based on free re- easily visualize the frequencies that have been detected in a con-
sponses obtained from NExT have to be applied repeatedly, using sistent manner and discard those that appear to be unrelated with
several channels as reference. the true dynamic behavior of the structure. The automation of this
Because the six targeted frequencies of the benchmark’s simu- process is somewhat affected by subjectivity as thresholds have to
lated data range approximately from 7 to 31 Hz, data produced be set. Basically, the detection of similar vibrational modes is
with the analytical model is filtered through a low-pass filter, based on two criteria, namely, 共1兲 the proximity of the frequencies
down sampling the data from 200 to 100 Hz while leaving and 共2兲 the correlation of the mode shapes. The first criterion is
enough frequency content in the estimated free response to cap- simply evaluated by setting a 2% threshold. Here, two frequencies
ture the dynamics of the system adequately. Fast Fourier trans- are considered similar if their values are within 2% of each other.
forms are performed using segments of data with 512 points, and The correlation of the mode shapes is evaluated by calculating the
frequency domain averaging is performed by overlapping the modal assurance criterion 共MAC兲 of the two vectors. The estab-
samples by 75% of their length 共boxcar windows兲. As a result, the lished threshold for this criterion in this study is set equal to 0.9.
estimated free responses are 2.56 s long. All 300 s of the simu- The algorithm then recognizes two vibrational modes as similar if
lated ambient vibration is used. Fig. 2 shows, in time and fre- both criteria are met. Although the proximity of the detected
quency domains, the estimated free response for one of the damping ratios has also been used by some researchers, it is well
known that the variation of these ratios is usually high compared
sensors that measures north-south vibrations of the top floor,
to the variation of the frequencies. Therefore, the ability to accu-
using the acceleration record from the opposite side as reference
rately identify damping ratios was not emphasized in this study.
channel.
Finally, stable modes are determined by searching, in descending
order, those modes with the most similar sets, while averaging the
corresponding frequencies, mode shapes, and damping ratios.
Modal Identification Techniques

Three common methods for modal identification for ambient re- Eigensystem Realization Algorithm with Data
sponses are to be employed in this comparison. The methods are Correlation
presented in this section and the specific considerations needed to Building on the Ibrahim time-domain method, Juang and Pappa
apply each technique are discussed. Additionally one of the ob- 共1985兲 developed a technique that uses the Markov parameters to
jectives of this research is to fully automate the modal identifica- build state-space representations of linear systems. The technique
tion techniques, eliminating the subjective judgment of different is known as the ERA and has shown to be effective for lightly
users. Thus, the section begins with a discussion of the general damped systems. Juang et al. 共1988兲 proposed an alternative ap-
procedure used for automation of the various methodologies. proach to the ERA called ERA/DC, where DC indicates that it

762 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
makes use of data correlations. The improved algorithm has
shown to be at least as effective as its predecessor 共see Lew et al.
1993兲. The algorithm uses the discretized free response h 共e.g.,
RX˜ ˜Xi from NExT兲 to form a matrix known as the Hankel matrix

冤 冥
hk hk+1 ¯ hk+c
hk+1 hk+2 ¯ ]
Hk−1 = 共10兲
] ] ]
hk+r ¯ ¯ hk+c+r
Using the Markov parameters, Eq. 共10兲 can be expressed as

冤冥
C
CA
Hk−1 = Ak−1关AZ1 A2Z1 ¯ Ac+1Z1兴 = UAk−1V
]
CAr
共11兲
where A and C = system and output matrices of a state-space re-
alization of the linear system and Z1 = vector of unknown initial
conditions. Using two different Hankel matrices formed from the
same free response, the correlation matrix Rk is defined as

Rk = HkHT0 = UAkUT 共12兲 Fig. 3. Numerical implementation of ERA/DC

and a block correlation matrix is then formed as

冤 冥
Rk Rk+␥ ¯ Rk+␤␥ Successful implementation of the ERA/DC usually requires
Rk+␥ Rk+2␥ ¯ ] some knowledge of the structure as well as some experience with
Uk = 共13兲 the algorithm, making it challenging to implement successfully
] ] ] for first-time users. For instance, Juang and Pappa 共1985兲 sug-
Rk+␣␥ ¯ ¯ Rk+␣␥+␤␥ gested using a Hankel matrix whose number of columns is 20
Combining Uk with Eq. 共12兲, the block correlation matrix can be times the number of frequencies and whose number of rows is
expressed as 2–3 times the number of columns. Another commonly used prac-
tice to select these dimensions is associated with the quality of the

冤 冥
UAk free responses. Under this criterion, the Hankel matrix is built
UAk+␥ making full use of the decaying signal provided that the signal-
Uk = Ak关AkUT Ak+␥UT ¯ Ak+␣␥UT兴 = PAkQ to-noise ratio is high.
] The selection of the number of singular values used to truncate
UAk+␣␥ the matrices ⌿, R, and S is also critical. Underestimation of this
共14兲 parameter translates into overlooking vibrational modes, whereas
overestimation may lead to obtaining computational modes or
A singular value decomposition of U0 is performed yielding even modes that are linear combinations of the true modes. Al-
though, in theory, this number should be equal to the number of
U0 = G⌿ST 共15兲
desired poles from the structural system, a number of factors may
where ⌿ = diagonal matrix with the singular values on the diago- lead to undesirable results. For instance, if the free responses
nal and the matrices G and S = square and unitary. The matrices contain little contribution from one particular mode, an erroneous
⌿N, GN, and SN are obtained eliminating the rows and columns mode will be introduced. In general, however, overestimation is
corresponding to small singular values produced by computa- generally preferable as numerical tools can later be used to dis-
tional modes. Since ⌿N is a diagonal matrix, Eq. 共15兲 can be tinguish true modes from computational modes.
expressed as In an attempt to improve the quality of the modal properties
identified with ERA/DC, herein each sensor is used as a reference
U0 = 共GN⌿1/2
N 兲共⌿N SN兲
1/2 T
共16兲 channel to estimate free responses from NExT. Therefore, a total
From Eqs. 共14兲 and 共16兲, when k = 0 of 12 sets of results are available for a given set of ambient
vibration records. While this process helps in detecting all true
P = GN⌿1/2
N ; Q = ⌿1/2
N SN
T
共17兲 modes of the structure, it also leads to obtaining different sets of
values for each mode. Fig. 3 shows a flow chart of the complete
For k = 1, Eq. 共14兲 takes the form U1 = PAQ, which can be com- modal identification process.
bined with Eqs. 共11兲 and 共17兲 to obtain the system and output Here Hankel matrices are formed with 200 columns and 96
matrices rows 共8 block rows of 12 sensors each兲, whereas the block corre-
A = ⌿−1/2 lation matrices were built with three Hankel matrices as block
N G NU 1S N⌿ N ; C = 关I 0兴GN⌿−1/2 共18兲
T −1/2
N
columns and two as block rows. The sampling frequency of the
where I = identity matrix and 0 = zero matrix, both of proper estimated free responses is 100 Hz. As a result, a total of 2.31 s of
dimensions. the free response are used to build both U0 and U1. The dimen-

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 763

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
sions of these matrices are chosen to make use of the first 2.5 s of output of the system 共i.e., ek = yk − ŷk兲. In the second step, a
the decaying response, where, as seen in Fig. 2, the signal-to- pseudo-ARX 共AR with exogenous input兲 model is estimated
noise ratio appears to be high. The truncation of matrices ⌿, R, using the error obtained from the AR model as the pseudoinput.
and S is carried out using 12 singular values, so six vibrational This ARX model takes the form
modes are calculated for a given set of free responses.
As mentioned previously, once ERA/DC has been executed ŷk = − A1yk−1 − ¯ − Anayk−na + C1ek−1 + ¯ + Cncek−nc 共21兲
some of the calculated modes may be computational. The system- from which na AR and nc moving average matrix coefficients can
atic elimination of doubtful modal properties is then carried out as be estimated using, again, a least squares approach
follows:

冤 冥冤
• Modes with unexpectedly high damping ratios are rejected. A − AT1 +

冥冤 冥
10% threshold was used for this steel frame. ] yTna ¯ yT1 eTna ¯ enTa−nc+1 ynTa+1
• To avoid poorly calculated modes due to the low amplitude of
− ATna ynTa+1 ¯ yT2 enTa+2 ¯ enTa−nc+2 ynTa+2
modes at the reference sensor, those modes whose shape vec-
=
tor have an amplitude at this location that is lower than one- CT1 ] ] ] ] ] ]
tenth of the highest value are eliminated. ] yTj−1 ¯ yTj−na eTj−1 ¯ eTj−nc yTj
• Repeated modes are eliminated on the basis of lowest energy
content using the associated singular values obtained in the CnTc
decomposition. The identification of such sets is performed as 共22兲
described in the stabilization of modes. However, lower
thresholds are used in this case 共frequencies within 20% of The system and output matrices of a state-space realization
each other and MAC values above 0.8兲. The identification and associated with this ARMAv model of order 共na , nc兲 can be
elimination of modes that are linear combinations of true formed as

冤 冥
modes are, on the other hand, more challenging. No action is 0 I ... 0 0
taken in these cases.
• Because only the first six vibrational modes of the system are ] ] ] ]
A= ; C = 关I 0 ¯ 0兴
targeted 共roughly between 7 and 32 Hz兲, only frequencies 0 0 ... 0 I
below 35 Hz are taken into account. Ana Ana−1 . . . A2 A1
Once the modal properties for each of the 12 estimated free
共23兲
responses are calculated, stabilization of modes is executed to
average all values identified as corresponding to a given vibra- and the dynamic properties can be extracted as described previ-
tional mode. Only modes detected three or more times are con- ously.
sidered stable. The sizes of the system and output matrices of the estimated
state-space realization are significantly larger than those estimated
with other techniques. As a result, along with the desired modal
Prediction Error Method through Least Squares
properties, several computational modal parameters are calcu-
Prediction error methods 共PEMs兲 have been widely reported in lated. This phenomenon constitutes the main drawback in the use
the literature 共see, for instance, Andersen 1997; Ljung 1999兲. The of ARMAv models for modal identification. To overcome this
main idea is to identify a system of linear equations that, based on problem, the use of stabilization diagrams is helpful to separate
past inputs and outputs, can predict any output. For the special true structural modes from computational modes. Several models
case of multivariate output-only measurements, these models are of different orders must be obtained. Two of the most widely used
known as autoregressive with moving average vector 共ARMAv兲. techniques for selecting the order of a parametric model are
In this case, the prediction is based not only on past outputs but Akaike’s final prediction error 共FPE兲 criterion and Akaike’s infor-
also on past errors of the prediction. Several algorithms have been mation theoretic criterion 共AIC兲 共Ljung 1999兲. These criteria are
proposed in these methods. Nonlinear optimization and multi- based on monitoring the decrease in the criterion function VN共␪兲
stage least squares are among the possibilities. In this study, PEM as the order 共na , nc兲 increases. The definition of VN共␪兲 is

冉兺 冊
is carried out using a two-stage least squares approach 共PEM/LS兲. N
As a first step, a long autoregressive 共AR兲 model is fitted to the 1
measurements using least squares. The AR model is of the form VN共␪兲 = det ␧关n兩␪兴 · ␧关n兩␪兴T 共24兲
N n=1
ŷk = − A1yk−1 − A2yk−2 − ¯ − Anyk−n 共19兲 where ␧关n 兩 ␪兴 = prediction error defined as ␧关n 兩 ␪兴 = yk − yk兩k−1;␪ and
where ŷ and y = predicted and true outputs of the system, respec- ␪ stands for the parameter matrix= 关A1 A2 ¯ Ana C1 C2 ¯ Cnc兴.
tively, and Ai=ith AR coefficient of the n-order AR model. In the Here both the FPE and AIC criteria are employed to consider how
case of multiple channels, the AR coefficients can be estimated as the accuracy of the ARMAv model varies with the order of
共na , nc兲. The expressions for the FPE and AIC take the form
+

冤 冥冤 冥冤 冥
− AT1 yTn T
yn−1 ¯ yT1 T
yn+1 1 + ␷/N
FPE = VN共␪兲 ; AIC = N log关VN共␪兲兴 + 2␷ 共25兲
− AT2 T
yn+1 yTn ¯ yT2 T
yn+2 1 − ␷/N
= 共20兲
] ] ] ] ] where N = number of samples and ␷ = 共na + nc兲m2 = total number of
estimated parameters.
− ATn yTj−1 yTj−2 ¯ yTj−n yTj
Some writers have employed this technique simply to calcu-
where j = length of the recorded measurements and “ +” represents late initial values for more complex nonlinear optimization algo-
the pseudoinverse. The error of this AR model can be calculated rithms 共see, for instance, Andersen 1997兲. However, the
by subtracting the predicted output 关using Eq. 共19兲兴 from the true computational power required to execute such algorithms is con-

764 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 4. Accuracy of PEM/LS with various ARMAv model orders

siderable and grows exponentially as the number of sensors in- However, it is important to understand the effects of the order of
creases. Moreover, because the estimation of several ARMAv the ARMAv models on the accuracy of the modal properties
models of different orders is necessary, the implementation of themselves.
nonlinear optimization algorithms becomes impractical. Numeri- Fig. 4 shows the deviation of the identified fundamental fre-
cal evaluations 共not reported here兲 showed that only marginal quency of the benchmark problem from the true analytical value
improvements 共if any兲 on the quality of the detected modal prop- for several ARMAv models. Although somewhat erratic, it is clear
erties are gained when optimization algorithms are implemented. from Fig. 4共a兲 that more accurate frequencies are obtained as the
In general, lengthy records allow an accurate fitting of the data order of the AR model is increased. Similarly, the order of the
with simple least squares techniques, making the implementation moving average model affects the identified values in the same
of optimization algorithms unnecessary. manner 关see Fig. 4共b兲兴. Higher frequencies, mode shapes, and
Although not as critical as in ERA/DC, a number of param- damping ratios are also influenced positively as the order of these
eters must be selected to implement the PEM/LS algorithm and models is increased. To consider the most appropriate order of the
thus the experience of the user is important. In the case of the model to use for identification, the FPE and AIC are shown in
two-stage least squares approach used in this study, the main vari- Fig. 4共c兲, where the trends clearly indicate the fact that a higher
ables include the order of both the AR and the moving average modal order results in a better ARMAv model. From the AIC
models. In general, an increase in the AR model order translates criterion, this indicates that ARMAv models with the order of
into more eigenvalues and eigenvectors obtained. This effect also 共2s , 2s − 1兲 result in lower errors than the models with the order of
means that the selection of the true vibrational properties of the 共2s , 2s兲. Since higher-order models require more computational
structure from computational ones becomes more challenging. power, it could be concluded that the accuracy of the methodol-

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 765

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 6. Typical stabilization diagram for PEM/LS

stochastic and unmeasurable inputs, the algorithm is known as the

stochastic subspace identification 共SSI兲 technique. Here, the Kal-
man filter states of a realization are obtained through a row space
projection of Hankel matrices. Once the states are known, the
identification problem is reduced to a linear set of equations from
which the system and output matrices of the realization can be
easily estimated.
Fig. 5. Numerical implementation of PEM/LS
Consider the Hankel matrix of the form

冤 冥
y0 y1 ¯ y j−1
ogy becomes a trade off with time. However, noise, nonstationary
inputs, and nonlinearities in the system can affect the technique y1 y2 ¯ yj
Y0兩i−1 = 共26兲
and even make it lose accuracy for excessively high-order mod- ] ] ]
els. yi−1 ¯ ¯ yi+j−2
Fig. 5 illustrates the implementation of the PEM/LS as em-
ployed in this study. Note that the modal properties were not where yk = random response of the system at time k and the two
obtained through averaging. The process starts by forming a basis subscripts on Y denote the time index of the upper left and bottom
to separate the true vibrational modes of the structure, using a left elements, respectively. It is assumed herein that the number of
series of six low-order ARMAv models. The order of these mod- columns of the Hankel matrix approaches infinity 共j → ⬁兲. Let us
els were 共4,2兲, 共4,3兲, 共5,4兲, 共5,3兲, 共6,4兲, and 共6,5兲, where the first now define Ui as the orthogonal projection of the row space of
and second numbers represent the AR and moving average orders, Yi兩2i−1 共known as the future兲 onto the row space of Y0兩i−1 共known
respectively. Moreover, the order of the AR model estimated in as the past兲
the first step of the algorithm was two times the order of the final Ui = Yi兩2i−1兩Y0兩i−1 共27兲
ARMAv model. All 5 min of data sampled at 200 Hz was used to
identify these models. Through a singular value decomposition of this projection, it
The systematic elimination of questionable modal properties can be shown that 共Kirkegaard and Andersen 1997兲
was carried out using the same criteria used in the implementation U i = Q iX i 共28兲
of ERA/DC. As expected, this handful of thresholds was particu-
larly useful in the implementation of the PEM/LS algorithm, where Qi = observability matrix and X̂i = Kalman state sequence
eliminating most of the computational modes that, by nature, the
algorithm estimates with each ARMAv model. Similarly, the sta- Qi = 关CT 共CA兲T 共CA2兲T ¯ 共CA j−1兲T兴T 共29兲
bilization of the identified modes is performed using the same
criteria used with ERA/DC. However, this process was carried =关x̂t x̂t+1 ¯ x̂t+j−1兴 共30兲
out with the sole purpose of determining the stable modes of the Similarly, it can be shown that the projection
system. Subsequently, more accurate values were extracted from
a high-order model with 10 AR and 10 moving average matrix Ui+1 = Yi+1兩2i兩Y0兩i−1 共31兲
coefficients 共10,10兲. Fig. 6 shows a typical stabilization diagram can be expressed as
where the stable vibrational modes are linked with a dotted line.
Ui = Qi−1X̂i+1 共32兲
Stochastic Subspace Identification From Eqs. 共29兲 and 共32兲, the Kalman states can be obtained
Subspace identification algorithms have slowly evolved over the from output data using a singular value decomposition. As a re-
last three decades. Van Overschee and De Moor 共1996兲 provided sult, the system and output matrices A and C are estimated from
the following set of linear equations:

冋 册冋册
a common mathematical background for all subspace algorithms
existing in the literature, including those used to identify input-
X̂i+1 A
output systems, stochastic output-only systems, and the combina- = X̂i 共33兲
tion of the two. When applied to linear systems driven by Yi兩i C

766 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Table 2. Evaluation Criteria
Criteria Description Formula
␻true − ␻id
J1 Accuracy of detected natural 100
frequencies as error percentage ␻true
J2 Accuracy of detected mode 1 − MAC共␾true , ␾id兲
shapes through the MAC
␨true − ␨id
J3 Accuracy of detected damping 100
ratios as error percentage ␨true

The implementation of the SSI algorithm requires only two

Fig. 7. Accuracy of SSI with various parameters parameters to setup, namely, 共1兲 the dimensions of the Hankel
matrices used in the orthogonal projections and 共2兲 the model
order used to truncate the singular value decompositions. The
former has an important impact in the computational power re-
quired to execute the algorithm. An appropriate selection of these
dimensions must take into account the computational power
available. On the other hand, the selection of a model order be-
comes difficult when noisy measurements are employed or when
nonstationary inputs excite the structure. Depending on the qual-
ity of the measurements and the frequency content of the input, it
may be necessary to retain a large number of singular values to
identify all vibrational modes. The use of several model orders
and the stabilization of the results is recommended by some re-
searchers and implemented here.
Fig. 7 shows the average error of the identified natural fre-
quencies with respect to the true analytical values for the already
mentioned setup parameters: truncation size and Hankel matrix
size. It can be observed that using a higher-order model for sin-
gular value decomposition has a positive impact on the accuracy
of these identified properties. Mode shapes and damping ratios
are similarly influenced. The identification of modal properties is
Fig. 8. Numerical implementation of SSI then carried out using three truncation levels 共20, 22, and 24兲 and
averaging the results using stabilization diagram. Fig. 7共b兲, on the

Fig. 9. Accuracy of identified natural frequencies 共J1兲 as a function of noise level

Fig. 10. Accuracy of identified mode shapes 共J2兲 as a function of noise level

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 767

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 11. Accuracy of identified damping ratios 共J3兲 as a function of noise level

other hand, indicates that the size of the Hankel matrices has an However, the identification of mode shapes is affected negatively
erratic and weak impact on the identification. Hankel matrices as the noise level increases 共see Fig. 10兲. SSI is clearly more
with 12 block rows that make use of the entire records available robust than ERA/DC and PEM/LS for mode shape identification
are used in this study. Systematic elimination of questionable with significant sensor noise.
modal properties is executed here using the same criteria dis- The mean value and standard deviation of the evaluation cri-
cussed previously. However, this step is rarely necessary. All teria for the various algorithms are shown in Table 3. All 6,000
5 min of data sampled at 200 Hz is used here. Fig. 8 summarizes results are employed to calculate the tabulated values. Results
the implementation of the algorithm. indicate that the SSI algorithm outperforms the other two algo-
rithms in the identification of all modal properties, demonstrating
its robustness with regard to noisy measurements. Given these
Comparison of Methods results and the fact that subspace identification requires the lowest
number of parameters to setup, it is concluded here that, among
One thousand simulations are performed to investigate the capa- the algorithms evaluated, SSI is the most appropriate choice.
bilities of each algorithm. Table 2 shows the evaluation criteria
used to carry out the comparison. Here, ␻, ␾, and ␨ represent the
frequencies, mode shapes, and damping ratios, respectively, Experimental Examination
whereas the subscripts “true” and “id” stand for analytical and
identified, respectively. The effect of sensor noise is investigated Regardless of how detailed an analytical study may be, it is chal-
by varying the signal-to-noise ratios. Thus a zero-mean Gaussian lenging to truly reproduce all possible real-world sources of un-
broadband noise is added to the acceleration records from each certainty. On the other hand, the main challenge in evaluating
simulation to simulate the effects of sensor noise. This noise is modal properties identified from experimental ambient vibration
scaled by a constant such that, over all simulations, it is defined as records is that, unlike with analytical studies, the exact dynamic
a uniformly distributed random variable 关0 , 2 ⫻ 10−4兴 共this distri- properties are unknown. Therefore, to verify the feasibility of
bution results in a noise with a maximum noise level of approxi- modal identification using ambient vibration, responses measured
mately 80% of the acceleration response for all acceleration from the experimental benchmark structure are used to obtain its
records, in RMS sense兲. modal properties and compare them with those identified with
Figs. 9, 10, and 11 show all 6,000 values 共six modes in 1,000 more traditional hammer testing.
simulations兲 for J1, J2, and J3, respectively, as a function of the Because a statistical comparison is not feasible, only the SSI
added noise level. All three algorithms successfully identify the algorithm is used to carry out the identification of the modal
targeted first six vibrational modes in all 1,000 simulations. How- properties using the ambient accelerations, and its implementation
ever, ERA/DC consistently calculates some computational modes is identical to that of the analytical study. In addition, because
共up to three兲. It is clear that both PEM/LS and SSI outperform ERA/DC was essentially developed to use impulse responses, this
ERA/DC in all evaluation criteria. Additionally, none of the meth- algorithm was implemented to extract the modal properties of the
odologies are capable of identifying damping ratios with reason- structure from the impact tests. The following section describes
able accuracy. While errors rarely reach 0.5% for the identified the evaluation of hammer tests using ERA/DC, followed by a
frequencies, damping ratios are often off by more than 20% and comparison of the results.
even reach 50% 共especially when using ERA/DC兲. Hammer impacts were all applied to the benchmark structure
It can be observed in Figs. 9 and 11 that, when identifying in the southwest corner of the first floor 共between first and second
natural frequencies and damping ratios, none of the methodolo- stories兲 of the structure 共Dyke et al. 2003兲. Three impacts in the
gies shows any clear signs of sensitivity to noise in the sensors. north-south direction and three more in the east-west direction

Table 3. Statistical Properties of Evaluation Criteria

Algorithm 兩J1兩 ␴ J1 J̄2 ␴ J2 兩J3兩 ␴ J3

−3 −2 −4 −4
ERA/DC 88.79⫻ 10 12.06⫻ 10 16.57⫻ 10 26.44⫻ 10 15.76 13.97
PEM/LS 61.77⫻ 10−3 7.93⫻ 10−2 5.56⫻ 10−4 6.92⫻ 10−4 6.11 7.83
SSI 54.90⫻ 10−3 7.06⫻ 10−2 4.56⫻ 10−4 4.75⫻ 10−4 5.55 7.23

768 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
Table 4. Identified Modal Properties
Frequency 共Hz兲 Damping ratio 共%兲
Mode shapes
Ambient Hammer Ambient Hammer MAC
Mode Description 共SSI兲 共ERA兲 共SSI兲 共ERA兲 ␾ham, ␾amb
1 East-West motion 7.466 7.449 1.059 0.796 0.895
2 North-South motion 7.759 7.634 0.973 0.852 0.932
3 Torsion 14.472 14.436 0.805 0.407 0.942
4 East-West motion 19.892 19.832 0.001 0.441 0.995
5 North-South motion 20.989 20.850 0.017 0.493 0.969
6 Torsion — 22.472 — 0.634 —
7 East-West motion 25.553 25.245 0.828 0.464 0.939

were recorded. Therefore, a total of six free responses recorded at in this benchmark problem, relatively low ambient vibrations are
a sampling frequency of 1,000 Hz are available. As with the ana- not sufficient for accurate modal identification. However, in some
lytical study, the available data are “down sampled” to 100 Hz applications the limitations demonstrated in this study are not
through low-pass filters. Unlike the estimated free responses ob- going to significantly impact the success of the application. For
tained with NExT, where only a couple of seconds constitute instance, this experimental benchmark structure has been em-
meaningful data, more than 5 s is part of the decaying response ployed for SHM applications by Giraldo et al. 共2003兲. Here it was
that greatly overpowers the ambient vibration. Thus, larger Han- found that, as long as the state of the structure is analyzed while
kel matrices can be used in this case, having 400 columns and 180 vibrating under the same conditions as in the healthy state 共i.e.,
rows 共15 block rows of 12 sensors each兲. The block correlation the ambient vibration levels are similar兲, damage can be success-
matrices are formed using three Hankel matrices as block col- fully determined because the modal characteristics will be influ-
umns and two as block rows. Thus, a total of 4.18 s of the free enced by the same 共nonlinear兲 behaviors.
response are employed. Systematic elimination of questionable
modal properties, as well as the stabilization of those remaining,
is carried out using the same criteria applied previously in the Conclusions
analytical study.
Table 4 contains the modal properties identified using both This paper focused on the identification of modal properties from
ambient vibration records 共with SSI兲 and hammer testing 共with civil engineering structures as they vibrate in their natural envi-
ERA/DC兲. Natural frequencies and damping ratios, as well as the ronment. The main goal of the study was to statistically evaluate
MAC values between corresponding mode shapes, are tabulated and compare three of the most popular time-domain modal iden-
here. Unlike the analytical study, no specific numbers of vibra- tification techniques in an entirely automated environment. The
tional modes are pursued. In fact, a total of seven modes are results indicated that the SSI algorithm is both user friendly and
identifiable from the hammer tests. Note that, perhaps due to a highly robust to sensor noise. Acceptable results were also ob-
frequency content of the disturbances, the sixth mode is not iden- tained with PEM/LS with better performance at low noise levels.
tified with the ambient vibration records. However, it is observed However, as the noise in sensors increased, its accuracy seemed
that the natural frequencies identified with ambient vibrations are to be negatively affected.
in good agreement with those identified from hammer impacts. In A secondary goal of this study was to examine the perfor-
fact, only the second natural frequency differs by more than 1% mance of the most effective methodology by comparing it to more
共1.63%兲. Damping ratios, on the other hand, have much larger
errors.
Perhaps the most important parameters for some applications,
for instance, SHM, are the mode shapes. Here, only the fourth
mode has excellent agreement 共i.e., MAC of 0.995兲. Moreover,
although relatively good agreement is observed in most results,
the vectors associated with the first mode seem to differ signifi-
cantly 共MAC of 0.895兲. Although not reported here, similar re-
sults were obtained using ERA and PEM/LS. Because relatively
long ambient vibration records are analyzed and averages from
different estimated free responses are obtained, this phenomenon
cannot be attributed to noise in the sensors. It is likely due to
inconsistent behaviors in the structure when vibrating at low and
high amplitudes, which are indicative of nonlinear behavior. Al-
though the source of this phenomenon is not precisely known, it
is likely that these changes are due to friction 共e.g., in the joints兲
that overtakes the vibrations mainly when the amplitude of the
displacements is small 共e.g., ambient vibration兲. Schematically,
Fig. 12 shows the differences between the first three mode shapes.
The inconsistencies in the mode shapes obtained from ambient
vibrations and hammer tests may lead some to argue that, at least Fig. 12. Experimentally identified mode shapes

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 769

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
traditional ways of testing. Experimental data were used to com- study of system identification techniques applied to new Caquinez
pare the modal properties obtained from real ambient vibrations Bridge.” Proc., 3rd Int. Conf. on Bridge Maintenance, Safety, and
and those obtained with hammer tests. Although very reasonable Management (IABMAS), Porto, Portugal.
results were obtained with SSI, some mode shapes have consis- Ibrahim, S. R., Brinker, R., and Asmussen, J. C. 共1996兲. “Modal param-
tent errors that are likely due to nonlinearities present at low eter identification from responses of general unknown random in-
levels of vibration. The intention here was not to discuss the puts.” Proc., Int. Modal Analysis Conf., SEM, Dearborn, Mich.
stationarity of the loads acting on civil structures and their capa- James, G. H., Carne, T. G., and Lauffer, J. P. 共1993兲. “The natural exci-
bility to excite an acceptable amount of vibrational modes. These tation technique for modal parameter extraction from operating wind
considerations are unique for each case and should be evaluated turbines.” Rep. No. SAND92-1666 (UC-261), Sandia National Labo-
in the field. ratories, Albuquerque, N.M.
Johnson, E. A., Lam, H. F., Katafygiotis, L., and Beck, J. 共2000兲. “A
benchmark problem for structural health monitoring and damage de-
Acknowledgments tection.” Proc., 14th Engineering Mechanics Conf., ASCE, Reston,
Va.
The writers would like to acknowledge support in part from NSF Juang, J. N., Cooper, J. E., and Wright, J. R. 共1988兲. “An eigensystem
Grant No. CMS 02-45402. realization algorithm using data correlations 共ERA/DC兲 for modal pa-
rameter identification.” Control Theory Adv. Technol., 4共1兲, 5–14.
Juang, J. N., and Pappa, R. S. 共1985兲. “An eigensystem realization algo-
rithm for modal parameter identification and model reduction.” J.
References Guid. Control, 8共5兲, 620–627.
Kirkegaard, P. H., and Andersen, P. 共1997兲. “State space identification of
Abdelghani, M., Verhaegen, M., Van Overchee, P., and De Moor, B. civil engineering structures from output measurements.” Proc., 15th
共1998兲. “Comparison study of subspace identification methods applied Int. Modal Analysis Conf., SEM, Orlando, Fla.
to flexible structures.” Mech. Syst. Signal Process., 12共5兲, 679–692. Lanslots, J., Rodiers, B., and Peeters, B. 共2004兲. “Automated pole-
Andersen, P. 共1997兲. “Identification of civil engineering structures using
selection: Proof-of-concept and validation.” Proc., Int. Conf. on Noise
ARMA models.” Ph.D. dissertation, Univ. of Aalborg, Aalborg, Den-
mark. and Vibration Engineering, ISMA, Leuven, Belgium.
Bendat, J. S., and Piersol, A. G. 共1980兲. Engineering applications of Lau, J., Lanslots, J., Peeters, B., and Van der Auweraer, H. 共2007兲. “Au-
correlation and spectral analysis, Wiley, New York. tomatic modal analysis: Reality or myth?.” Proc., 25th Int. Modal
Brinker, R., De Stefano, A., and Piombo, B. 共1996兲. “Ambient data to Analysis Conf., SEM, Orlando.
analyze the dynamic behavior of bridges: A first comparison between Lew, J. S., Juang, J. N., and Longman, R. W. 共1993兲. “Comparison of
different techniques.” Proc., 14th Int. Modal Analysis Conf., SEM, several system identification methods for flexible structures.” J. Sound
Dearborn, Mich. Vib., 167共3兲, 461–480.
Caicedo, J. M., Dyke, S. J., and Mizumori, A. 共2004兲. “Characterization Ljung, L. 共1999兲. System identification—Theory for the user, Prentice-
of traffic excitation for numerical studies in structural health monitor- Hall, Upper Saddle River, N.J.
ing.” Proc., 17th Engineering Mechanics Conf., ASCE, Reston, Va. Pappa, R. S., James, G. H., III, and Zimmerman, D. C. 共1998兲. “Autono-
Desforges, M. J., Cooper, J. E., and Wrigth, J. R. 共1995兲. “Spectral and mous model identification of the space shuttle tail rudder.” J. Spacecr.
modal parameter estimation from output-only measurements.” Mech. Rockets, 35共2兲, 163–169.
Syst. Signal Process., 9共2兲, 169–186. Peeters, B., and Ventura, C. E. 共2003兲. “Comparative study of modal
Dyke, S. J., Bernal, D., Beck, J., and Ventura, C. 共2003兲. “Experimental analysis techniques for bridge dynamic characteristics.” Mech. Syst.
phase of the structural health monitoring benchmark problem.” Proc., Signal Process., 17共5兲, 965–988.
16th Engineering Mechanics Conf., ASCE, Reston, Va. Wenzel, H., and Pichler, D. 共2005兲. Ambient vibration monitoring, Wiley,
Giraldo, D., Caicedo, J. M., and Dyke, S. J. 共2003兲. “Experimental phase New York.
of the SHM benchmark studies: Damage detection using NExT and Van Overchee, P., and De Moor, B. 共1996兲. Subspace identification for
ERA.” Proc., 16th Engineering Mechanics Conf., ASCE, Reston, Va. linear systems: Theory, implementation and applications, Kluwer
He, X., Moaveni, B., Conte, J. P., and Elgamal, A. 共2006兲. “Comparative Academic, Dordrecht, The Netherlands.

770 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 20 Jul 2009 to 129.252.24.32. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright