You are on page 1of 22

Response-only modal identification of structures using strong

motion data
S. F. Ghahari
1
, F. Abazarsa
2
, M. A. Ghannad
1
and E. Taciroglu
3,
*
,
1
Department of Civil Engineering, Sharif University of Technology, P.O. Box 11155-9313, Tehran, Iran
2
Structural Engineering Department, Int. Institute of Earthquake Eng. and Seismology, P.O. Box 3913/19395, Tehran, Iran
3
Civil and Environmental Engineering Department, 5731E Boelter Hall, University of California, Los Angeles,
CA 90095, USA
SUMMARY
Dynamic characteristics of structures viz. natural frequencies, damping ratios, and mode shapes are
central to earthquake-resistant design. These values identied from eld measurements are useful for model
validation and health-monitoring. Most system identication methods require input excitations motions to be
measured and the structural response; however, the true input motions are seldom recordable. For example,
when soilstructure interaction effects are non-negligible, neither the free-eld motions nor the recorded
responses of the foundations may be assumed as input. Even in the absence of soilstructure interaction, in
many instances, the foundation responses are not recorded (or are recorded with a low signal-to-noise ratio).
Unfortunately, existing output-only methods are limited to free vibration data, or weak stationary ambient
excitations. However, it is well-known that the dynamic characteristics of most civil structures are amplitude-
dependent; thus, parameters identied from low-amplitude responses do not match well with those from strong
excitations, which arguably are more pertinent to seismic design. In this study, we present a new identication
method through which a structures dynamic characteristics can be extracted using only seismic response
(output) signals. In this method, rst, the response signals spatial time-frequency distributions are used for
blindly identifying the classical mode shapes and the modal coordinate signals. Second, cross-relations among
the modal coordinates are employed to determine the systems natural frequencies and damping ratios on the
premise of linear behavior for the system. We use simulated (but realistic) data to verify the method, and also
apply it to a real-life data set to demonstrate its utility. Copyright 2012 John Wiley & Sons, Ltd.
Received 7 November 2011; Revised 24 September 2012; Accepted 1 October 2012
KEY WORDS: blind system identication; modal identication; output-only techniques; strong ground
motions; soilstructure interaction; spatial time-frequency distributions
1. INTRODUCTION
Identication of dynamic characteristics of civil structures from response recorded during strong
ground shaking has been a subject of research for more than three decades [14]. System
identication techniques are not only useful for condition assessment, but also for improving future
designs, validating predictive models, and verifying retrot procedures [5]. Active and semi-active
control systems that are now being frequently employed in important structures throughout the
world also make use of system identication techniques, which provide intelligent feedback from
the structure so that the active and passive (damping) forces supplied to the system can be properly
controlled [6]. Although the diversity of applications continue to ourish, a major part of system
identication research remains focused on model validation and damage assessment. Numerous
*Correspondence to: E. Taciroglu, Department of Civil and Environmental Engineering, 5731E Boelter Hall, University
of California, Los Angeles, CA 90095-1593, USA.

E-mail: etacir@ucla.edu
Copyright 2012 John Wiley & Sons, Ltd.
EARTHQUAKE ENGINEERING & STRUCTURAL DYNAMICS
Earthquake Engng Struct. Dyn. (2012)
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/eqe.2268
studies have been undertaken in recent years to develop robust methods by which structural damage
can be detected and quantied (see, e.g., [7, 8]). In most of these studies, changes in dynamic
properties are used as indices for detecting the location and severity of damage (see, e.g., [9, 10]
and references therein).
Many of the existing system identication methods are common to mechanical, electrical, and
structural engineering elds. Quite a number of those methods have been developed for output-only
modal identication, and they are especially pertinent for civil structures, for which input excitations
are not always readily available. However, most of these techniques are designed to use free or
ambient vibration data as input [1113] wherein the identied modal properties are related typically
to small-amplitude vibrations. Moreover, in techniques devised for ambient vibrations, it is usually
assumed that the input excitation is a broadband stochastic process, which is modeled as stationary
Gaussian white noise [14]. On account of these facts, available output-only identication techniques
cannot be employed for nonstationary response signals recorded during strong ground motions.
Consequently, apart from a few notable exceptions [15, 16], nearly all of the existing techniques
employed in the identication of civil structures from their seismic responses require the input
motions to be known/measured. Yet, in most practical cases, the true foundation input motion (FIM)
is not recorded at all, or recorded with a low signal-to-noise ratio (SNR), or cannot be recorded
because of, for example, soilstructure interaction (SSI) effects. When there is signicant SSI,
the FIM is different from both the free-eld motion recorded in the structures vicinity, and
the responses recorded at the foundation level [17]. The usual contrast between the stiffness of
a (nearly rigid) foundation and the surrounding soil causes the motion experienced by the foundation
(i.e., FIM) to differ from the free-eld motion; an effect coined as kinematic interaction. Inertial
interaction, which is caused by the soils exibility and attenuation, and the structures and
its foundations masses, compounds the said difference between the foundation response and
the FIM.
Even in the absence of SSI, for many real cases, foundation responses are not always recorded, or
are recorded with low SNR. As such, recorded foundation responses are usually assumed to be input
motions in dynamic analyses, and in system identication studies. For example, Skolnik et al. [18]
used signals recorded at the ground level of a 15-story, steel-frame building (i.e., the UCLA Factor
Building) as input motions for their nite element model updating studies. However, the Fourier
spectra of those signals (Figure 1) indicate that they are affected by the buildings own dynamic
response: in the Fourier amplitude spectrum of the eastwest (EW) acceleration (cf., Figure 1(a)),
there are two dominant peaks occurring at frequencies 0.42 and 0.65 Hz, which are very close to
the reported rst natural frequency of the EW mode (0.47 Hz) and the natural frequency of the
rst torsional mode (0.68 Hz), respectively. This similarity is also seen in the northsouth (NS)
direction spectrum, wherein a dominant peak corresponding to rst natural frequency in the NS
direction (i.e., 0.51 Hz) is observed (cf., Figure 1(b)). Clearly, the motion used as input is inuenced
by the structures response. Examples such as these are not uncommon [19].
Figure 1. Fourier spectra of the ground level responses of the UCLAFactor Building during the 2004 Parkeld
earthquake.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
In this paper, we present a new system identication method through which the modal properties of
civil structures are identied from their seismic responses without having measurements of input
excitations. The method works in two steps:
i. Mode shapes and modal coordinates are extracted by applying a blind source separation (BSS)
technique to the time-frequency representation of recorded responses; and
ii. Concurrent analyses of modal coordinates are carried out using a cross-relation (CR) technique to
identify natural frequencies and damping ratios from modal coordinates extracted in the rst step.
The theoretical background needed for the rst step is presented in the next section, where the time-
frequency distribution (TFD) of a signal and also spatial TFD (STFD) of signals are briey introduced.
This is followed by detailed descriptions of the methods two steps (i.e., BSS and CR). Performance of
this new method is evaluated subsequently through simulated and real data examples.
2. THEORETICAL BACKGROUND
2.1. Time-frequency distribution of a signal
Signals are classically represented in either the time-domain or the frequency-domain. In both forms, the
time and frequency variables are treated as mutually exclusive. Consequently, these representations are
nonlocal with respect to the excluded variable [20]. As such, they are only suitable for signals with
time-invariant or frequency-invariant properties. TFDs are needed for nonstationary signals, or signals
from a nonlinear system, for which the signal characteristics are changing. The theoretical background
of TFDs is beyond the scope of this paper; and interested readers are referred to Boashashs textbook [20]
and the references cited therein. For brevity, only the pertinent fundamental formulae are presented
below.
The discrete-time form of the Cohen-class of TFDs is given by [21]
D
xx
t; f
X
1
l1
X
1
m1
m; l x t m l x

t m l e
j4pfl
(1)
where x(t) is a time signal, and t and f represent the time and frequency variables, respectively. The
kernel function, (m,l), depends on both the time (t) and the lag (l) variables. Different choices of
the kernel function lead to different TFD realizations. TFDs dened through Equation (1) are
categorized as nonlinear or quadratic TFDs. Contrary to linear TFDs, such as the short-time Fourier
transform, or the wavelet transform, the signals product with its complex conjugate is used.
Choosing the most suitable TFD for a given situation depends on which of its characteristics are
desired. Linear TFDs are sometimes preferred, because they are real-valued, simpler, and free of
interference-terms; however, their simultaneous time-frequency resolutions are limited [22].
Quadratic TFDs have higher time-frequency resolutions, but suffer from interference. Interference
terms, which are also dubbed as outer artifacts or cross-terms [20], are spurious features that
appear when representing a multicomponent signal in the time-frequency domain using one of the
quadratic methods. To wit, consider the signal x(t) = x
1
(t) + x
2
(t) in which both x
1
(t) and x
2
(t) are
analytic and monocomponent signals. Substituting x(t) into Equation (1), its TFD will be,
D
xx
t; f D
x
1
x
1
t; f D
x
2
x
2
t; f 2Re D
x
1
x
2
t; f f g (2)
in which D
x
1
x
1
t; f and D
x
2
x
2
t; f are the auto-terms, which represent energy concentration in the time-
frequency plane, whereas 2Re D
x
1
x
2
t; f f g is a cross-term that occurs in (t,f) points in which no energy
is expected at all. These cross-terms have large oscillating amplitudes on average times and frequencies
of the true components, and can make the TFD difcult to interpret. This is especially true if the
components are numerous (or they are close to each other), and also in the presence of noise,
because they can be produced between the signal and the noise components. Therefore, in most
practical applications, quadratic TFDs are not used despite their higher resolution. There are three
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
approaches to suppress the cross-term effects: (i) by multiplying the TFD with the spectrogram, which
is a cross-term free TFD [23]; (ii) by using one of the existing components as a reference signal [24];
and (iii) by attenuating the cross-terms through a properly selected kernel function, which results in a
TFD family that is referred to as reduced interference distribution. Herein, we adopt the smoothed
pseudo-WignerVille distribution (SPWVD) [25], which is an enhanced version of WignerVille
distribution (WVD) [26] and belongs to the said reduced interference distribution family.
In all TFD methods, the analytic associate of the real signal is used to eliminate the unnecessary
negative frequencies without losing information. This removal has two benecial effects: rst, it
halves the total bandwidth, allowing the signal to be sampled at half the usual Nyquist rate without
aliasing [27, 28]; second, it avoids the appearance of various interference terms generated by the
interaction of positive and negative components in quadratic TFDs [20]. The analytic associate, x
a
(t), of
a signal x(t) is dened as
x
a
t x t j^x t ; ^x t
1
p
Z
1
1
x t
t t
dt (3)
where ^x t denotes the Hilbert Transform of x(t).
2.2. Spatial time-frequency distribution
The TFD dened in the previous section conveys the energy distribution of a single signal in the time-
frequency plane, and thus, it is called an auto-TFD. In many cases, it is necessary to have the joint
energy distribution of two signals, for which the cross-TFD is used. The cross-TFD of two signals
x
1
(t) and x
2
(t) is dened in similar fashion to Equation (1),
D
x
1
x
2
t; f
X
1
l1
X
1
m1
m; l x
1
t m l x

2
t m l e
4pjfl
: (4)
On the basis of Equation (4), the STFD of a vector x containing n signals is dened as
D
xx
t; f
X
1
l1
X
1
m1
m; l x t m l x
H
t m l e
4pjfl
(5)
where D
xx
t; f
ij
D
x
i
x
j
t; f for i, j 2{1, . . .,n}; and the superscript H denotes a Hermitian transpose.
Although the same kernel function is used in Equation (5) for all pairs, it is possible to use a specic
kernel for each pair.
In the previous section, auto-terms and cross-terms were introduced as points with true and ghost
energy concentrations, that is, nonzero TFD, respectively. For an STFD, two extra terms are
introduced: The point (t
a
, f
a
) is dubbed an auto-source TF point of a source (signal) x
i
(t), if its auto-
TFD at this point, that is, D
x
i
x
i
t; f , exhibits an energy concentration. This energy concentration can
be true if x
i
(t) is monocomponent, or a ghost if the source is multicomponent. Additionally, the
point (t
c
, f
c
) is dubbed a cross-source TF point between signals x
i
(t) and x
j
(t), if their cross-TFD at
this point, that is, D
x
i
x
j
t; f , exhibits an energy concentration [29].
3. THE PROPOSED IDENTIFICATION METHOD
In this section, we present a new system identication method with which modal properties, that is,
natural frequencies, damping ratios, and mode shapes, can be extracted from a buildings response
recorded during strong ground motions. This method works without the knowledge of the input
motions, so it may be categorized as an output-only identication method. At the present time, there
appears to be no other robust technique proposed in the open literature that claims the same feat.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
The governing equations of motion for an MDOF system with n
d
DOFs, which is excited by a
unidirectional (scalar) ground acceleration, can be expressed as follows:
M x t C x t Kx t Mlx
g
t (6)
in which M, C, and K are the constant n
d
n
d
mass, damping, and stiffness matrices of the system,
respectively. The vector x(t) contains relative displacement responses of the system at all DOFs;
x
g
t is a scalar time-signal, which represents the (unknown) ground acceleration; and l is the
inuence vector [30]. In practical cases, the absolute acceleration of structure is recorded, which is
x
t
t x t lx
g
t (7)
By assuming a proportional damping matrix, the absolute acceleration response can be expressed in
modal space as
x
t
t fq t (8)
where fis an n
d
n
d
real-valued mode shape matrix whose i-th column (f
i
) is the i-th mode shape; and
q t is a vector that contains the absolute acceleration modal coordinates. The i-th modal coordinate is
the absolute acceleration response of an SDOF system that corresponds to the i-th mode, given by
q
i
t h
i
t b
i
x
g
t ; i 1; . . . ; n
d
(9)
where the operator * indicates linear convolution; b
i
f
i
T
Ml=f
T
i
Mf
i
is the modal contribution
factor [30]; and h
i
(t) is the SDOFs impulse response function, which is calculated as [31],
h t
1
o
d
e
xo
n
t
o
d
2
x
2
o
n
2
_ _
sin o
d
t 2xo
n
o
d
cos o
d
t
_
(10)
where x, o
n
, and o
d
o
n

1 x
2
p
are the damping ratio, and the undamped and damped natural
frequencies of the SDOF system, respectively. As such, the identication of a system may be divided
into two stages: (i) modal decomposition and (ii) SDOF system identication. The following sections
describe the said two stages of the proposed output-only identication method.
3.1. Blind source separation
Blind source separation is a well-established methodology in the sound separation eld [3235]. BSS
techniques are used for recovering the source signals and the unknown mixing matrix using only the
recorded signals (output). Here, we apply BSS to estimate the modal coordinates (source signal) and
the mode shape matrix (mixing matrix).
Consider again Equation (8) in which n
m
contributing modal coordinates are linearly combined to
produce n
n
response signals, with n
m
n
n
n
d
. Because Equation (8) is a single equation with two
unknowns, restricting assumptions must be considered for the source signals and the mixing matrix.
If we assume that the source signals have distinct structures and localization properties in the time-
frequency domain, and that the mixing matrix has full-column rank, then the mixing matrix can be
identied, and the source signals can be recovered up to an arbitrary scaling factor and permutation
via BSS [3638].
Before describing the details of the time-frequency domain BSS method, we present a simple
conceptual example to illustrate the implication of being localized in the time-frequency domain.
Figure 2 displays the time-frequency distributions of two synthetic signals, schematically. In
Figures 2(a) and 3(b), the types of signals that are mostly encountered in the electrical engineering
eld are depicted [29, 39, 40]. Having a distinct localization in the time-frequency domain means
that the signals have disjoint time-frequency signatures, as in Figure 2(a). However, BSS techniques
can also be applied to signals that are quasi-disjoint, that is, they have overlapping regions in the
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
Figure 2. A schematic diagram of time-frequency distributions of disjoint and quasi-disjoint signals in
typical electrical (a, b) and earthquake (c, d) engineering applications.
Figure 3. TFDs of the rst oor dynamic response of the 5-DOF system under the El Centro Array #9 accel-
erogram recorded in Imperial Valley earthquake, 1940 [50]: (a, c) WVD, and (b, d) SPWVD for systems
with stiffness-proportional and mass-proportional damping, respectively.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
time-frequency domain, as in Figure 2(b). Herein, the modal coordinates are taken as the source
signals. Thus, they must have distinct time-frequency signatures so that BSS techniques are
applicable. Fortunately, in most cases, the time-frequency distributions of modal coordinates are
similar to the schematic TFD distribution shown in Figure 2(c). In the worst-case scenario, the input
excitation also has a dominant frequency that contributes in all modal coordinates as shown in
Figure 2(d) in which the modal coordinates can be assumed as quasi-disjoint sources. Hence, time-
frequency-based BSS methods can be applied to separate these source signals blindly. This method
is described next.
Calculating the STFD of both sides of Equation (8) and neglecting the noise effects yields,
D
x
t
x
t t; f fD
qq
t; f f
H
(11)
where D
x
t
x
t and D
qq
t; f are, respectively, n
n
n
n
and n
m
n
m
matrices whose elements are the auto-
TFDs and cross-TFDs of the recorded signals and the modal coordinates. Equation (11) is similar to the
well-known BSS equation in the sound-separation eld (cf., [41]). For simplicity, consider an ideal
TFD, that is, one without any cross-terms in any time-frequency point that corresponds to an auto-
term of only a source signal. In this case, D
qq
t; f is diagonal with only one nonzero diagonal
element. Thus, at each auto-term, Equation (11) is converted to,
D
x
t
x
t t
i
; f
i
f
i
D
q
i
q
i
t
i
; f
i
f
i
H
(12)
where f
i
is i-th column of f, and D
q
i
q
i
t
i
; f
i
is i-th modes auto-TFD. Knowing the proper auto-term
of each modal coordinate, an eigenvalue decomposition (EVD) can be used to extract f
i
and
D
q
i
q
i
t
i
; f
i
. However, D
qq
t
i
; f
i
is rank-decient, and thus, f cannot be uniquely determined.
Moreover, external criteria are needed to specify the true auto-terms of each source signal and to
decide which modes auto-term should be used for EVD. The difference between the number of
modes and the number of sensors poses further difculties for EVD. Belouchrani et al. [37] showed
that it is necessary and sufcient to use all auto-terms simultaneously, without knowing to which
source they belong; hence, a joint approximate diagonalization (JAD) is employed instead of
EVD [42, 43]. In this method, several Jacobi rotations are applied on a set of matrices to diagonalize
them (details are omitted here for brevity, but may be found, for example, in [44]).
Because the JAD algorithm is restricted to nding a unitary diagonalizing matrix, a preprocessing
step that converts the problem of nding f (which is not always unitary) to nding a unitary matrix
U is needed. This step, dubbed whitening, renders Jacobi rotations applicable, and also solves the
dimensional problem mentioned earlier. As the term whitening implies, this step converts the
sensor signals to white signals, that is, their zero-lag correlation matrix becomes equal to the identity
matrix I. The whitening process reduces the determination of the n
n
n
m
mode shape matrix f to
that of a unitary n
m
n
m
matrix (U) as follows. On the basis of the aforementioned denition, an
n
m
n
n
whitening matrix (W) converts the recorded signals, x
t
t , to z t W x
t
t , such that,
R
z
0 I (13)
where R
z
(0) is a zero-lag correlation matrix of the new version of recorded signals; and is computed as
R
z
0 E z t z

t (14)
where E[.] denotes the expected value and the superscript * indicates the complex conjugate. For
discrete deterministic signals, Equation (14) can be written as,
R
z
0
1
N
X
N
k1
z k z

k (15)
where Nis the number of time samples, andk is a discrete index. Inserting x t fq t into z t W x
t
t ,
and applying the condition presented in Equation (13) yields
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
R
z
0 E z t z

t E Wfq t q

t f
H
W
H
_
WfR
q
0 f
H
W
H
I (16)
Because of the intrinsic ambiguity embedded in BSS techniques, and without any loss of generality,
it is assumed that R
q
0 I . In other words, it is implicitly assumed that the source signals are
uncorrelated at zero-lag with unit variance; thus,
Wf Wf
H
I (17)
Equation (17) implies that Wf=U, in which U is a unitary matrix. If the whitening matrix is known,
then f may be identied from U by a simple MoorePenrose pseudo-inverse [41]. For determinate
cases in which the number of sensors is equal to the number of modes, it can be easily shown that
the whitening matrix can be obtained as follows [41]:
W R
x
t 0
0:5
(18)
where R
x
t 0 is the correlation matrix of recorded signals, and R
x
t 0
0:5
is its principal square root.
For cases in which n
m
<n
n
, the rst n
m
largest eigenvalues and eigenvectors of R
x
t 0 are used as
follows [41]:
W
0:5
n
m
H
H
n
m
(19)
where
n
m
is an n
m
n
m
diagonal matrix in which diagonal entities are the n
m
largest eigenvalues of
R
x
t 0 , and H
n
m
is an n
m
n
m
matrix that contains the corresponding eigenvectors. The whitening
process must be applied on a noise-free portion of the recorded signals, which is clearly not
possible for real data. Hence, for the under-determined case, the average of (n
n
n
m
) the remaining
eigenvalues of R
x
t 0 is used as noise variance, and is subtracted from the diagonal elements of

n
m
to reduce the noise effects [41]. The aforementioned whitening process may be carried out
with other approaches, for example, Robust Whitening [45, 46], which may improve the results in
exchange for additional computation.
The STFD of the, now whitened, signals can be expressed as
D
zz
t; f UD
qq
t; f U
H
(20)
Therefore, any whitened STFD-matrix is diagonal when stated in the basis of the columns of the
matrix U. As mentioned previously, this unknown unitary matrix can be identied through a JAD of
D
zz
(t,f) at the chosen time-frequency points that are true auto-terms. After the identication of U,
the mode shape matrix f, and the modal coordinates q t , may be recovered as follows:
f W
#
U; (21)
q t U
H
W x
t
t (22)
where the superscript # denotes a MoorePenrose pseudo-inverse.
As mentioned above, the JAD procedure is to be applied on a set of D
zz
(t,f) matrices calculated at
time-frequency points that are auto-terms. Several researchers have proposed various criteria to
determine/select the said auto-terms [29, 37, 38, 47]. However, before attempting to identify the
auto-terms, it is expedient to increase the odds of choosing proper points (and also to decrease the
computational effort involved) by removing the noise effects and also by discarding the low-energy
points. To remove the time-windows with no signicant energy (e.g., those portions that usually
comprise the beginning and the end of strong motion records) the following criterion may be used:
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
max
f
D
x
t
x
t t
i
; f k k
F
max
t; f
D
x
t
x
t t; f k k
F
> E
1
(23)
where .
F
denotes the Frobenius norm. The criterion includes time slices t
i
, in which the ratio of the
maximum energy along the frequency axis to the maximum energy in the time-frequency plane is
greater than some scalar E
1
. The value of E
1
should be chosen based on the signal-to-noise ratio. Our
experiences with real-life records suggest that values greater than 1% will work well for most cases.
However, for weak earthquake motions with noisy data, it may have to be increased up to 10%.
Next, an energy criterion can be used to remove unnecessary points in each time slice as in,
D
x
t
x
t t
i
; f k k
F
max
f
D
x
t
x
t t
i
; f k k
F
> E
2
: (24)
Here, in each time slice t
i
, the auto-term points are selected if the ratio of their energy to the maximum
energy at that time window is greater than some scalar E
2
. The value of E
2
is chosen based on earthquake
energy; for weak input motions, larger values of E
2
should be used (typically greater than 1%).
To identify the auto-term points, Belouchrani et al. [38] suggested exploiting the off-diagonal
structure of the STFD matrices at cross-terms; that is, for time-frequency points that correspond to
cross-terms, the following condition can be written:
trace D
zz
t; f trace UD
qq
t; f U
H
_
trace D
qq
t; f
_
0 (25)
This equation is valid because the trace of a matrix is invariant under a unitary transform. Thus, they
suggested using the criterion, trace D
zz
t; f = D
zz
t; f k k
F
< E
3
, to detect the cross-terms and to exclude
them. Here E
3
is a positive scalar less than 1 (typically, E
3
=0.8) [29]. Although this criterion works well
for excluding the cross-terms (produced because of deciencies inherent to quadratic TFD), it cannot
detect cross-sources. In other words, at common points (i.e., points at which several sources are
contributing) the value of trace D
zz
t; f = D
zz
t; f k k
F
< E
3
may be greater than 1; and these points
will be selected as auto-source points for the diagonalization process. This scenario is highly likely
for a system that is subjected to strong ground shaking the scenario in which the input excitation
contains several dominant frequencies. If the dominant frequencies of input are observed in several
modes, then the aforementioned criterion cannot lter out the time-frequency points corresponding
to those frequencies; and another criterion would have to be used to select the best auto-sources.
Under the source time-frequency disjointness assumption, each auto-source STFD matrix is of rank
one; or at least, each matrix has one signicantly large eigenvalue compared with its other
eigenvalues. Therefore, the following criterion may be used to deselect the cross-source points [29]:
l
max
D
x
t
x
t t; f
D
x
t
x
t t; f k k
F
1

> E
4
(26)
where E
4
is a small positive scalar (typically, E
4
= 0.001) and l
max
[.] represents the largest eigenvalue of
its argument matrix. Note that in this new criterion, the STFD matrix of the original data (as opposed to
that of the whitened data) is used.
Remark 1
To reduce modal interference in systems with closely spaced modes in two directions, signals of those
two directions should be analyzed separately for symmetric buildings. For asymmetric cases, a stricter
auto-source point selection criterion (i.e., smaller values for E
4
) must be used.
By employing the aforementioned process for selecting the auto-source points, and by applying
JAD for the diagonalization of whitened auto-source STFD matrices, the mode shapes and modal
coordinates in absolute acceleration forms will have been identied. The second step is to extract
the natural frequencies and damping ratios, which is described next.
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
3.2. Simultaneous identication of the modal coordinates
On the basis of Equation (9), the i-th absolute acceleration modal coordinate q
i
t , is the response of an
SDOF system to the translational acceleration b
i
x
g
t . Applying a Z-transform to both sides of
Equation (9), the response of the system in the Z-plane can be represented by the following equation:

q
i
z

h
i
z b
i

x
g
z (27)
Here,

q
i
z ,

x
g
z , and

h
i
z denote the Z-transforms of the modal coordinate, the input motion, and the
impulse response function, respectively. Unilateral Z-transform of a discrete-time signal x[k] is dened
as [48],
x z
X
1
k0
x k z
k
(28)
where k is an integer time index and z is, in general, a complex number. The term

h
i
z is calculated in
discrete form based on Equations (10) and (29) as

h
i
z
1
T
C
i
D
i
z
1
1 A
i
z
1
B
i
z
2
_ _
(29)
where T is sampling time and the coefcients are given by
A
i
2e
x
i
w
n
iT
cos w
di
T ; B
i
e
2x
i
w
ni
T
; C
i
2x
i
w
ni
T;
D
i
o
ni
Te
x
i
o
ni
T
o
ni
o
di
1 2x
i
2
_ _
sin o
di
T 2x
i
cos o
di
T
_ _
(30)
Substituting these expressions into Equation (27) yields
T
b
i
C
i
1 A
i
z
1
B
i
z
2
_ _

q
i
z 1 D

i
z
1
_ _

x
g
z (31)
where D

i
D
i
=C
i
, that is,
D

i
e
x
i
o
ni
T
o
ni
o
di
1 2x
i
2
2x
i
_ _
sin o
di
T cos o
di
T
_ _
(32)
Transforming Equation (31) back to the time-domain, a new representation is obtained,
1
b
i
C
i
q
i
k A
i
q
i
k 1 B
i
q
i
k 2 x
g
k D

i
x
g
k 1 ; 1in
m
; 1kN (33)
with N denoting the number of samples used for discretization. This equation indicates that the
absolute acceleration modal coordinate at discrete time k, that is, q k , may be written as a linear
difference equation involving previous responses, and previous and current input excitations. In the
literature, such models are termed as auto-regressive models with eXternal/eXogeneous input. Using
a nonlinear least-squares technique, the n
m
linear equations corresponding to all of the extracted
modes may be solved simultaneously to determine the natural frequencies, damping ratios, and modal
contribution factors of the system, and the input motion. However, this approach is very expensive
computationally. An alternative is presented below that alleviates this burden.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
Equation (33) may be restated for all time instants in matrix form as,
Q
i

i
x
gD
i
for 1in
m
(34)
where
Q
i

q
i
2 q
i
1 q
i
0

q
i
N 1 q
i
N 2 q
i
N 3
2
4
3
5
(35)

i
g
i
0
g
i
1
g
i
2
_
T
; (36)
x
gD
i

x
g
2 D

i
x
g
1

x
g
N 1 D

i
x
g
N 2
2
4
3
5
; (37)
where g
i
0
1= b
i
C
i
, g
i
1
g
i
0
A
i
, g
i
2
g
i
0
B
i
and T denotes matrix transposition. On the basis of
Equation (34), n
m
(n
m
1)/2 possible relations can be established among all modes as follows:
Q
i

i
Q
j

j
D

i
D

j
_ _
x
g
for 1i; jn
m
(38)
where
x
g
x
g
1 x
g
N 2
T
: (39)
Repeating Equation (38) for a different pairs and merging those results, we may obtain a new
relation that is free from input excitation. To wit,
Q
i
Q
j
Q
k
Q
l
_
i
D

i
D

j
D

i
D

k
D

k
D

l
D

k
D

l
_ _
T
0 for i; j 6 k; l : (40)
Using any four modes in Equation (40), we may write L
n
m
2
_ _
2
0
@
1
A
relations provided that
n
m
3. These relations can be combined in a matrix form as follows:
Q

L n
m
n
m
1

n
m
n
m
1 1
0 (41)
where

n
m
D

n
m
D

n
m
1

n
m
1
D

n
m
D

n
m
1

n
m
D

n
m
D

n
m
2

n
m
2
D

n
m
D

n
m
2


2
D

2
D

1
D

2
D

1
" #
(42)
in which
Q

L
L
Q

nm
(43)
where
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
L
2
I I ; L
l

L
11
l
L
12
l
0 L
l1
_ _
for 3lL (45)
and
L
11
l
I I I
T
; L
12
l

I 0

0 I
2
4
3
5
for 3lL (46)
where I is an identity matrix with size N2.
In the presence of noise, Equation (41) can be solved in a least-squares sense through the
constrained minimization problem,

CR
arg min

k k1

Q
H

(47)
as described in [49]. Alternatively, the eigenvector associated with the smallest eigenvalue of the
matrix

Q
H

Q can also be used as a solution of Equation (47). The present equations would yield
(n
m
1) answers for each
i
, each of which containing a different denominator term. Nevertheless,
the (n
m
1) answers for the damping ratio and natural frequency of each mode can be identied from
the following equations, wherein g
i
0
, g
i
1
, and g
i
2
related to each answer have the same denominator,
which cancel out,
x
i
o
ni

1
2T
ln
g
i
2
g
i
0
; o
ni

1 x
i
2
q

1
T
cos
1
g
i
1
e
x
i
o
ni
T
2g
i
0
_ _
for 1in
m
; (48)
x
i

x
i
o
ni
o
ni
; o
ni

x
i
o
ni

2
o
ni

1 x
i
2
q
_ _
2
s
for 1in
m
: (49)
4. VERIFICATION OF THE METHOD AND ITS APPLICATION TO REAL DATA
4.1. A simulated MDOF system for method verication
Consider a 5-DOF linear shear-building model, in which the oor mass is taken to be 3 10
5
for each
story. Interstory stiffnesses, from the rst to the fth story, are 7k, 5k, 3k, 2k, and k, respectively, where
k = 5 10
7
. This structure mimics the dynamic characteristics of a typical 5-story building fairly well.
Two separate cases are investigated: in the rst case, the damping ratio, x
m
(%), is set to be mass-
proportional with a 10% rst mode damping ratio; in the second case, the damping ratio, x
s
(%), is
set to be stiffness-proportional with a 0.5% rst mode damping ratio. Horizontal accelerogram
recorded in El Centro Array #9 during Imperial Valley earthquake, 1940 [50] is used as input
motion for generating the dynamic response of the system. This dynamic analysis is carried out
using the lsim command in MATLAB [51] with a 100 Hz sampling frequency. Figure 3 displays the
rst oors absolute acceleration responses using two different kinds of TFDs, viz. WVD and
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
SPWVD. Here we see that the SPWVD has better time and frequency resolutions, because of the
reduced cross-terms; thus, we adopt it for all subsequent analyses.
Remark 2
As expected, the higher mode energies of the system with stiffness-proportional damping are much
lower than those of the lower modes. Their small contributions, which occur during a limited time
window, render traditional frequency methods (e.g., Fourier transform) to be insufcient for system
identication. The proposed method, on the other hand, can make use of this limited information
and yield the structures dynamic properties in its higher modes, as will be demonstrated below.
Application of the BSS time-frequency decomposition presented earlier with E
4
=0.001 (E
1
and E
2
are
not used, as data are noise free in this example) yields the auto-sources shown in Figure 4, where selected
points at both ends of the time axis are discarded because of edge effects. As this gure indicates, the 4th
and 5th modes of the stiffness-proportional system are contributing only around the arrival-time of the
excitation waves. However, because of the light damping in higher modes, the system with mass-
proportional damping responds in all modes at all times during the excitation. Comparing the identied
mode shapes with their analytical counterparts conrms this observation. This comparison can be
quantitatively made with the modal assurance criterion (MAC) values calculated through,
MAC
i

f
a
i
:f
i
i
_ _
2
f
a
i
_
_
_
_
2
f
i
i
_
_
_
_
2
(50)
where f
i
i
and f
a
i
denote the i-th identied and the analytical mode shapes, respectively. The MAC
indices for all modes are presented in Table I for the two systems. The MAC values for all modes of
both systems are higher than 90%, which favorably suggests that the proposed method is highly
accurate. Nevertheless, the fourth and fth modes of the structure with stiffness-proportional damping
have distinctly lower MAC values. To investigate this, we applied the BSS-TFD technique to eight
equal-sized time windows, which yielded the variation of MAC value with respect to time. Figure 5
Figure 4. Selected points for BSS of systems with (a) stiffness-proportional and (b) mass-proportional damping.
Table I. Comparison of the analytical and the identied modal properties.
Identied
Stiff. prop. Mass prop. Analytical
Mode no. f
n
(Hz) x
s
(%) MAC f
n
(Hz) x
m
(%) MAC f
n
(Hz) x
s
(%) x
m
(%)
1 1.12 0.50 1.000 1.11 10.36 1.000 1.11 0.50 10.00
2 2.56 1.13 1.000 2.53 4.28 0.996 2.56 1.15 4.34
3 4.02 1.69 0.998 3.99 2.87 0.994 4.05 1.82 2.75
4 5.63 2.72 0.928 5.74 1.66 0.982 5.66 2.55 1.96
5 7.35 3.33 0.929 8.03 1.58 0.994 8.13 3.66 1.37
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
displays the said time variation of MAC values for both systems (markers in these plots are located at the
end-time of each time window). As this gure indicates, the 4th and 5th modes can be more accurately
identied only during the rst time window for the system with stiffness-proportional damping; whereas,
for the system with mass-proportional damping, the proposed method yields nearly the same results at all
time windows.
Remark 3
It is worth mentioning here that the use of all of the data can increase the MAC values, as it happened
for the 4th mode of the system with stiffness-proportional damping. The primary reason for this
phenomenon is related to the JAD procedure, which seeks orthogonal directions. Thus, it extracts the
estimates of inactive modes even at time-frequency points at which they are not contributing.
The application of the second step of the identication process yields the natural frequencies and
damping ratios of the two systems as shown in Table I. The exact/analytical values are shown here
also for comparison. As can be seen, for both systems, the identied modal values are nearly the same,
especially for the lower modes. It is important to note here that the inaccuracies in the identied frequency
and damping ratio values primarily belong to the previous step, which produced inaccurate modal
coordinates. The natural frequencies and damping ratios presented in Table I are the average of (n 1)
solutions (their variations are not presented here for brevity). Results indicate that variations of the identied
natural frequencies are negligible, while the identied damping ratios for the mass-proportional system
display some variations. However, these variations are mostly due to errors accrued during modal
decomposition, because the second step of the proposed method works perfectly for exact modal coordinates.
To prove this statement, the identication technique was applied to the exact modal coordinates,
which resulted in exact natural frequencies and damping ratios (again, omitted here for brevity).
4.2. A simulated soilstructure system for method verication
The 5-DOF structure introduced in the previous section is now placed on a exible foundation to
model a new 7-DOF soilstructure system with two additional DOFs (foundation sway and system
rocking). The foundation mass, and the horizontal and rocking soil stiffnesses are as m
f
= 3 10
5
,
k
h
= 9 10
8
, and k
r
= 3 10
10
, respectively. Mass moment of inertia of all stories and the foundation
are set as I
i
= 25m
i
, wherein m
i
denotes story mass. Also, a constant story height equal to 4 m is
considered for all stories. For the sake of simplicity, only mass-proportional damping is considered
so that the rst mode has a 10% damping ratio. Similar to the xed-base example, dynamic analysis
with the same horizontal input motion is carried out in MATLAB [51] with a 100 Hz sampling
frequency. Figure 6 displays the SPWVD graphs of foundations and roofs absolute acceleration
responses along with rocking response of the system. As can be seen, all modes are not detectable
in all DOFs. For example, the highest mode is only observed in foundations response, while the
Figure 5. Time variation of MAC for systems with (a) stiffness-proportional and (b) mass-proportional
damping.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
roofs response is composed of the rst three modes. This gure also shows that rocking response is
mainly affected by the fourth mode.
Using all DOFs absolute accelerations, mode shapes are identied by applying the BSS time-
frequency decomposition method with E
4
= 0.001. The calculated MAC indices for all modes are
1.000, 0.993, 0.997, 0.737, 0.981, 0.998, and 0.996, which indicates that the 4th mode is not
identied accurately. However, this mode is a rocking mode, and the MAC index is low because the
order of rotation is small compared with the horizontal displacement. For this reason, the
COordinate Modal Assurance Criterion [52] index is calculated for all DOFs, which is dened as
COMAC
l

X
7
k1
f
a
kl
f
i
kl

_ _
2
X
7
k1
f
a
kl
2
_ _
X
7
k1
f
i
kl
2
_ _ (51)
where f
a
kl
and f
i
kl
denote the analytical and the identied mode shapes, respectively, at the l-th DOF in
the k-th mode. This index for foundation sway and rocking and the other ve structural DOFs are
0.8271, 0.9226, 0.8795, 0.9201, 0.9207, 0.9361, and 0.9505, respectively. As seen, the rocking
DOFs deformation, which is the dominant deformation in the fourth mode, is identied well.
Repeating the decomposition process with only ve signals (excluding the foundations horizontal
and rocking response signals), ve modes are identied. A comparison of the identied and analytical
mode shapes indicates that the identied modes are the 1st, 2nd, 3rd, 5th, and 6th modes. MAC indices
for the identied mode shapes are 1.000, 0.996, 0.998, 0.978, and 0.996. Thus, although the exclusion
of two response signals converts this case to an under-determined problem, ve modes are still
identiable. Figure 7 displays the time variation of MAC values for both cases. As this gure
indicates, the MAC index for mode 4 is high between 25 to 30 s, because during this time window,
the contribution of mode 4 relative to other modes is signicant (cf., Figure 6). Moreover, as can be
seen in Figure 7(b), all identied modes are nearly accurate in all time segments, even with only
Figure 6. SPWVDs of foundation (a) translation and (b) rocking, and (c) roof translation responses under the
El Centro Array #9 accelerogram recorded during the 1940 Imperial Valley earthquake [50].
Figure 7. Time variation of MAC index for two cases: (a) using all DOFs response signals and (b) excluding
foundation horizontal and rocking response.
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
ve sensors. The averages of the identied natural frequencies and damping ratios using two sets of
sensors are presented in Table II. This table shows that the identied values using ve sensors are
fairly as accurate. Nevertheless, excluding the foundation and rocking sensors results in identication
errors, especially for the rst mode damping ratio.
4.3. Application of proposed method to data recorded at the Van Nuys Hotel
Herein, we apply the proposed method to real-life data recorded at the Van Nuys Hotel. Data are
provided by the Centre for Engineering Strong Motion Data website (http://strongmotioncenter.org).
This building is ideally suited for the present study, because there are several structural vibration
records from damaged states with extents that range from moderate to severe [19]. Not surprisingly,
quite a number of studies have been conducted on data from this building. A brief summary of
those studies along with a description and history of the building are presented in what follows.
The Van Nuys Hotel is a 7-story building with a 6200 m
2
oor plan, located in the San Fernando
Valley of Los Angeles County, California (34.2203 N, 118.4713 W). It was designed in 1965 and
built during 1966 [53]. The structure is essentially symmetric in both directions. Spandrel beams and
exterior columns form the primary frame that resists the lateral loads in both directions. The oors
are reinforced concrete at slabs. The structure sits atop a group of pile foundations on recent
alluvial soils that primarily consist of ne silty sands [54]. The building was rst lightly, and then
severely damaged during the 1971 San Fernando (M6.6), and the 1994 Northridge earthquakes
(M6.7), respectively. After the 1994 earthquake, the building was retrotted with new reinforced
concrete shear walls in the eastwest direction at exterior frames [53].
The building was instrumented in 1980 through the California Strong Motion Instrumentation
Program (CSMIP Station No. 24386). Of the 16-channel recording system, which comprises uni-
axial, bi-axial, and tri-axial instruments, only ve channels are devoted to the eastwest direction.
These ve channels are mounted on the ground, second, third, sixth, and roof levels. Several
researchers have used acceleration data provided by these instruments to determine the natural
frequencies and damping ratios of the building. Ambient tests conducted soon after construction in
1967, and following the 1971 San Fernando earthquake before and after a seismic retrot revealed
that the rst natural frequency of the eastwest response decreased from 1.89 Hz before the
earthquake to 1.39 Hz after the earthquake, and then increased to 1.56 Hz following repair [55]. Two
detailed ambient vibration tests were also conducted by Ivanovic et al. [54] after the 1994
Northridge earthquake. They placed sensors on one of the interior eastwest frame at all oors to
determine the lateral frequencies. They reported the frequencies of the system as 1.0, 3.5, and 5.7 Hz
for three translational modes in the EW direction. Note that all of the aforementioned frequencies
are apparent (or pseudo-exible base) frequencies, because they are calculated from relative
spectra of the response signals to the ground oor signals, while rocking response is not excluded [56].
Owing to the existence of large sets of recorded earthquake data, there were also numerous attempts
to estimate apparent natural frequencies from recorded seismic responses. Trifunac et al. [57, 58]
studied the changes in natural frequencies of the building from earthquake to earthquake, using
Table II. Comparison of the identied and analytical modal properties.
Identied
5 Sensors 7 Sensors Analytical
Mode no. f
n
(Hz) x(%) f
n
(Hz) x(%) f
n
(Hz) x(%)
1 0.99 28.00 0.90 8.02 0.91 10.00
2 2.42 3.42 2.43 3.76 2.45 3.70
3 3.87 2.34 3.86 2.29 3.87 2.34
4 4.81 1.83 4.91 1.85
5 5.44 1.53 5.30 1.59 5.43 1.67
6 7.72 1.17 7.59 1.21 7.57 1.20
7 10.96 0.83 10.97 0.83
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
Fourier analyses and short-time Fourier transform techniques. Table III displays some of their results,
where the rst apparent frequency of the building in the EW direction at the beginning and end of ve
earthquakes, along with its minimum value obtained during shaking, is presented. On the basis of these
results, Trifunac et al. concluded that the system frequencies changed from one earthquake to another,
and with the intensity of shaking.
In another study, Alimoradi and Naeim [19] used ground oor responses recorded during the
1992 Big Bear and 1994 Northridge earthquakes as input. They identied the rst (EW) natural
frequency as 0.87 Hz from Big Bear data. They went on to apply their method to three segments of
the Northridge data, and estimated the natural frequencies of the rst three modes along the EW
direction as 0.61, 1.90, 3.41 (rst segment), 0.56, 1.74, 3.11 (second segment), and 0.48, 1.82,
4.20 Hz (third segment). They also found the damping ratios to be approximately 10% for all modes
for segment 1, 15% for all modes for segment 2, and 6.7%, 8.4%, and 15% for three modes for
segment 3 of the Northridge data. However, if SSI is signicant for this building, as reported by
Trifunac et al. [57, 58], then neither the xed-base nor the exible-base natural frequencies were
identied in their study. Indeed, they have extracted pseudo-exible-base parameters, that is, rocking
motion is included, but sway is excluded.
Recently, Todorovska and Trifunac [59] used wave travel times of vertically propagating waves in
the Van Nuys building between the ground oor and roof to estimate the rst natural frequency of the
xed-base system in the EW direction. They found that the xed-base system vibrated at a frequency
of around 1.2 Hz during the Landers and the Big Bear earthquakes (both occurred in 1992). They also
found that the rst frequency varied from 0.85 to 0.55 Hz during the 1994 Northridge earthquake.
As the summaries of previous studies indicate, there appear to be several unresolved issues
regarding the interpretation of data recorded at the Van Nuys Hotel. To wit,
Apparent natural frequencies of the system identied from ambient tests are different from those
expected during earthquakes (we also note that the damping ratios have not been reported in [54]).
In calculating the apparent natural frequencies, rocking response is included, but sway is discarded.
To identify the dynamic characteristics of the system from strong motion data, two approaches
were used: (i) assuming the ground oor response as the input motion [19], and (ii) extracting
information from the relative response spectra [57, 58]. The rst approach is not applicable for
buildings that interact with the surrounding soil (Van Nuys Hotel appears to have this attribute),
and the second approach cannot be used to identify the natural frequencies and damping ratios of
soilstructure systems.
As demonstrated through the simulated (verication) problem earlier, the identication method
proposed herein is able to accurately estimate the modal characteristics from strong motion data
without the need to know Foundation Input Motions. As such, aforementioned problems regarding
the Van Nuys buildings data can be overcome with the proposed method, which is applied here to
data recorded at the four eastwest channels. The ground oor data are excluded because of its
evidently low signal-to-noise ratio. Because of the limited number of sensors along the height of the
building, we seek only the rst three translational modes along the EW direction, and use the data
recorded during four relatively strong earthquakes, which are summarized in Table IV. In addition
to the high level of recorded peak amplitudes (except for the Big Bear earthquake), these
earthquakes are specically chosen so that the identied results can be compared with the previous
studies mentioned earlier. We use the full lengths of recorded signals for the Whittier, Landers, and
Table III. The rst EW apparent natural frequency of Van Nuys Hotel reported by Trifunac et al. [57, 58].
Number Earthquake Date M f
beg
(Hz) f
end
(Hz) f
min
(Hz)
1 San Fernando 02/09/1971 6.6 1.05 0.85 0.70
2 Whittier 10/01/1987 5.9 1.00 0.75 0.80
3 Landers 06/28/1992 7.5 1.00 1.30 0.70
4 Big Bear 06/28/1992 6.5 0.80 0.80 0.80
5 Northridge 01/17/1994 6.4 0.95 0.60 0.45
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
Big Bear earthquakes, and divide the Northridge data into four 15-s segments, because signicant
damages have been reported during this earthquake.
Figure 8 displays the mode shapes identied through the proposed method with E
2
= 0.01 and
E
4
= 0.001. The identied rst modes for all earthquake data (Figure 8(a)) appear nearly the same;
although a small discrepancy is observed at the lower stories during the latter portions of Northridge
earthquake, which is arguably related to the reported damages. This discrepancy is also observed for
the Big Bear earthquake, which may be related to measurement noise, because the level of shaking
is fairly low in this earthquake.
Remark 4
It is useful to note here that the sensors are limited to points shown by markers in Figure 2, and linear
interpolation is used for the mode shape displacements of other stories. Additionally, the contribution
of torsional modes may cause further inaccuracies in estimating the translational modes.
Remark 5
Figures 8(b) and (c) imply that changes are occurring during the nal 45 s of the Northridge earth-
quake: the second and especially the third mode shapes identied from these data windows are differ-
ent from their earlier versions. Because of the limited number of sensors, it is not possible to exactly
identify the location of these changes (a potential extension of the technique proposed here for local-
izing damage is deferred to a future study). Nevertheless, as a comparison, the mode shapes identied
from the last segment of the Northridge earthquake are redrawn in Figure 9, along with mode shapes
identied from ambient tests conducted one month after this earthquake [54]. This gure reveals that
the proposed method can extract the mode shapes of this structure by using only four signals recorded
during the Northridge earthquake (note that the identied mode shapes from ambient data are related to
a pseudo-exible-base system).
Table IV. Earthquake data used in this study.
Number Earthquake Date M Dist. (km) PGA* (g) PSA** (g)
1 Whittier 10/01/1987 5.9 41 0.17 0.20
2 Landers 06/28/1992 7.5 187 0.04 0.19
3 Big Bear 06/28/1992 6.5 152 0.03 0.06
4 Northridge 01/17/1994 6.4 7 0.47 0.59
*PGA, peak ground acceleration.
**PSA, peak structure acceleration.
Figure 8. The identied mode shapes. (a) rst mode, (b) second mode, and (c) third mode.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
In the second step, natural frequencies and damping ratios are extracted from the identied modal
coordinates. The obtained results are displayed in Table V. As observed, natural frequencies, especially
in the rst mode, are decreasing during successive earthquakes, which may be an indication of damage.
The said trend is only violated during the nal segment of the Northridge earthquake. More specically,
the minimum rst natural frequency is observed during the third segment of the Northridge earthquake,
potentially because nonlinearities/damage occurred during that segment, and then it is seen recovering
during the last segment, which incidentally is consistent with the results reported in Figure 10 of Ref. [59].
Remark 6
Damping ratios higher than 20% are not reported here, because they are unusual for civil structures. It
is also useful to note that the natural frequencies and damping ratios identied here may be question-
able because of potential contributions from other (undetected torsional or other higher) modes.
5. CONCLUSIONS
A new system identication method was presented. This method can identify modal properties of civil
structures from their recorded responses when the input motion is unknown. Contrary to currently
available output-only techniques, which are only applicable to ambient vibration data, the proposed
method is able to extract dynamic characteristics of structures using strong motion responses. This is
quite attractive, because the actual input motions exciting the structures are rarely recordable during
earthquakes. This is especially true for the case of soilstructure systems where neither the free-eld
motion nor the recorded response of foundation may be assumed as input to the system. This
Figure 9. Comparison between the mode shapes identied from the last 15 s of the 1994 Northridge
Earthquake data (black), and from ambient test one month after that [54] (gray).
Table V. Identied natural frequencies and damping ratios from earthquake response.
Mode No. Whittier Landers Big Bear Northridge-1 Northridge-2 Northridge-3 Northridge-4
f
n
(Hz) 1 0.87 0.86 0.77 0.63 0.50 0.43 0.50
2 2.92 3.31 2.29 2.00 1.66 1.46 1.13
3 5.41 5.50 4.54 4.24 3.43 3.07 2.81
x(%) 1 8.96 7.51 9.76 2.72 16.67 11.38 12.62
2 7.34 5.01 1.71 9.41 7.25
3 13.74 7.89 3.83 1.67 16.76
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
method comprises two steps. First, a time-frequency blind source separation technique is employed to
decompose the recorded response into a linear combination of modal coordinate signals for which the
combination factors are the mode shapes. In the second step, modal cross-relations are employed to
extract the natural frequencies and damping ratios. To verify the proposed method, two numerical
simulations were presented: (i) identication of a xed-base system without having the input motion
(i.e., the foundation response); and (ii) identication of a exible-base system without having the
foundation input motion (i.e., the result of kinematic interaction between the foundation and the
surrounding soil). Moreover, response signals recorded at the Van Nuys Hotel during four strong
earthquakes were used as a case study. Identication results for both scenarios, that is, simulated
and real-life data, indicate that the proposed method can be successfully employed to identify
dynamic properties of civil structures.
ACKNOWLEDGEMENTS
The work presented in this manuscript was funded, in part, by the NSF Grant CMMI-0755333. Any
opinions, ndings, conclusions or recommendations expressed in this material are those of the authors
and do not necessarily reect the views of the sponsoring agency. The authors also would like to thank
Professor Boashash for providing MATLAB codes for calculating the time-frequency distributions.
REFERENCES
1. Hart GC, Yao JTP. System identication in structural dynamics. Journal of Engineering Mechanics ASCE 1976;
103:10891104.
2. Beck JL, Jennings PC. Structural identication using linear models and earthquake records. Earthquake Engineering
and Structural Dynamics 1980; 8:145160.
3. Beck JL. System identication applied to strong motion records from structures. In: Earthquake Ground Motion and
its Effects on Structures, Datt SK, (ed.), AMD-53, ASME: New York, 1983.
4. Doebling SW, Farrar CR. The state of the art in structural identication of constructed facilities. Report by the ASCE
Committee on Structural Identication of Constructed Facilities, 1999.
5. Soyoz S, Taciroglu E, Orakcal K, Nigbor R, Skolnik D, Lus H, Safak E. Ambient and forced vibration testing of a
reinforced concrete building before and after its seismic retrotting. Journal of Structural Engineering (in print,
http://dx.doi.org/10.1061/(ASCE)ST.1943-541X.0000568).
6. Hazra B, Sadhu A, Lourenco R, Narasimhan S. Re-tuning tuned mass dampers using ambient vibration measurements.
Smart Materials and Structures 2010; 19(11), 115002.
7. Carden EP, Fanning P. Vibration based condition monitoring: a review. Structural Health Monitoring 2004; 3(4):
355377.
8. Caicedo JM, Dyke SJ. Experimental validation of structural health monitoring for exible bridge structures. Structural
Control and Health Monitoring 2005; 12(3-4):425443.
9. Worden K, Friswell MI. Modal-vibration-based damage identication. Encyclopedia of Structural Health Monitoring.
John Wiley & Sons, 2009.
10. Moaveni B, He X, Conte JP, Restrepo JI. Damage identication study of a seven full-scale building slice tested on
the UCSD-NEES shake table. Structural Safety 2010; 32(5):347356.
11. Juang JN, Pappa RS. An eigensystem realization algorithm for modal parameter identication and model reduction.
Journal of Guidance, Control, and Dynamics 1985; 8(5):620627.
12. van Overschee P, Moor BD. Subspace algorithm for the stochastic identication problem. Automatica 1993; 29(3):
649660.
13. Brincker R, Zhang L, Andersen P. Modal identication of output-only systems using frequency domain decomposition.
Smart Mater Struct. 2001; 10:441445.
14. Yuen KV. Bayesian methods for structural dynamics and civil engineering. John Wiley & Sons: Singapore, 2010.
15. Pridham BA, Wilson JC. Identication of based-excited structures using output-only parameter estimation. Earthquake
Engineering and Structural Dynamics 2004; 33:133155.
16. Lin CC, Hong LL, Ueng JM, Wu KC, Wang CE. Parametric identication of asymmetric buildings from earthquake
response records. Smart Materials and Structures 2005; 14:850861.
17. Wolf JP. Dynamic Soil-Structure Interaction. Prentice-Hall, Englewood Cliffs: New Jersey, 1985.
18. Skolnik D, Lei Y, Yu E, Wallace J. Identication, model updating, and response prediction of an instrumented 15-story
steel-frame building. Earthquake Spectra 2006; 22(3):781802.
19. Alimoradi A, Naeim F. Evolutionary modal identication utilizing coupled shear-exural response-implication for
multistory buildings, Part II: application. The Structural Design of Tall and Special Buildings 2006; 15:67103.
20. Boashash B Time-frequency signal analysis and processing: A comprehensive reference. Elsevier Ltd: UK, 2003.
21. Cohen L Time-frequency analysis. Englewood Cliffs, Prentice-Hall: New Jersey, 1995.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
22. Gabor D. Theory of communication. Journal of IEE 1946; 93(III):429457.
23. Barkatm B, Boashash B. Higher order PWVD and Legendre based time-frequency distribution. Proceedings of Sixth
IEEE International Workshop on Intelligent Signal Processing and Communication Systems (ISPACS98); 2, 532-536,
Melbourne, Australia, 1998.
24. Jones DL, Parks TW. A high-resolution data-adaptive time-frequency representation. IEEE Transactions Acoustics,
Speech, and Signal Processing 1990; 38:21272135.
25. Auger F, Flandrin P. Improving the readability of time-frequency and time-scale representations by the reassignment
method. IEEE Transactions on Signal Processing 1995; 43:10681089.
26. Flandrin P. Time-frequency/time-scale analysis. Academic Press: San Diego; 1999.
27. Boashash B. Note on the use of the Wigner distribution for time-frequency signal analysis. IEEE Transactions
Acoustics, Speech, and Signal Processing 1988; 36:151521.
28. Claasen TACM, Mecklenbrauker WFG. The Wigner distribution-A tool for time-frequency signal analysis. Philips
Journal of Research 1980; 35:217250 (part 1): 276300 (part 2): 372389 (part 3).
29. Linh-Trung N, Belouchrani A, Abed-Meriam K, Boashash B. Separating more sources than sensors using time-
frequency distributions. EURASIP Journal on Applied Signal Processing 2005; 17:28282847.
30. Chopra A. Dynamics of structures: Theory and Applications to Earthquake Engineering. Englewood Cliffs, Prentice-
Hall: New Jersey, 1995.
31. Irvine T. An introduction to the shock response spectrum. Available: http://www.vibrationdata.com/tutorials2/
srs_intr.pdf, 2002 09/05/2011.
32. Jutten C, Herault J. Space or time adaptive signal processing by neural network models. Proceedings of AIP Conference,
206-211, Lake Louise, Canada, 1986.
33. Pham DG. Separation of a mixture of independent sources through a maximum likelihood approach. Proceedings of
European Signal Processing Conference, 771774, Brussels, Belgium, 1992.
34. Comon P. Independent component analysis: a new concept?". Signal Processing 1994; 36:287314.
35. Hyvarinen A, Karhunaen J, Oja E. Independent component analysis. John Wiley: New York, 2001.
36. Bousbia-Salah A, Belouchrani A, Bousbia-Salah H. A one step time-frequency blind identication.
Proceedings. of the International Symposium on Signal Processing and its Applications (ISSPA), 1:581584,
Paris, France, 2003.
37. Belouchrani A, Moeness G, Amin MG. Blind source separation based on time-frequency signal representations.
IEEE Transaction on Signal Processing 1998; 46(11):1998.
38. Belouchrani A, Abed-Meriam K, Amin MG, Zoubir AM. Blind separation of nonstationary sources. IEEE Signal
Processing Letters 2004; 11(7):605608.
39. Aissa-El-Bey A, Linh-Trung N, Abed-Meriam K, Belouchrani A, Grenier Y. Underdetermined blind separation of
nondisjoint sources in the time-frequency domain. IEEE Transactions on Signal Processing 2007; 55(3):897907.
40. Aissa-El-Bey A, Abed-Meriam K, Grenier Y. Underdetermined blind audio sources separation using modal decom-
position. EURASIP Journal on Audio, Speech, and Music Processing 2007; Article ID 85438: 115.
41. Belouchrani A, Abed-Meriam K, Cardoso JF, Moulines E. A blind source separation technique using second-order
statistics. IEEE Transactions on Signal Processing 1997; 45(2):434444.
42. Cardoso JF. Perturbation of joint diagonalizers. Telecom Paris, Signal Dept., Technical Report; 94D023, 1994.
43. Cardoso JF, Souloumiac A. Jacobi angles for simultaneous diagonalization. SIAM Journal on Matrix Analysis and
Applications 1996; 17(1):161164.
44. Golub GH, Loan CFV. Matrix computations. John Hopkins University Press: Baltimore, MD, 1989.
45. Belouchrani A, Cichocki A. Robust whitening procedure in blind source separation context. Electronics Letters
2000; 36:20502051.
46. Choi S, Cichocki A, Belouchrani A. Blind separation of second order nonstationary and temporally colored sources.
Proceedings of the 11th IEEE Signal Processing Workshop on Statistical Signal Processing 2001; 444447.
47. Fevotte C, Doncarli C. Two contributions to blind source separation using timefrequency distributions. IEEE Signal
Processing Letters 2004; 11(3):386389.
48. Jury EI. Theory and Application of the Z-Transform method. Krieger Pub. Co, 1973.
49. Xu G, Liu H, Tong L, Kailath T. A least-squares approach to blind channel identication. IEEE Transactions on Signal
Processing 1995; 43:29822993.
50. Pacic Earthquake Engineering Research Center (PEER), http://peer.berkeley.edu/smcat/ 07/01/2011.
51. MATLAB. The language of technical computing. Version 7.0, Mathworks, 2004.
52. Lieven NAJ, Ewins DJ. Spatial correlation of mode shapes the coordinate modal assurance criterion (COMAC).
Proceedings of the 6th Int. Modal Anal. Conf. 1988; 1:690695.
53. Krawinkler H (ed). Van Nuys hotel building tested report: exercising seismic performance assessment. Pacic Earth-
quake Engineering Research Center. Report No. PEER 2005/11. University of California at Berkeley: Berkeley,
California; 2005.
54. Ivanovic SS, Trifunac MD, Novikova EI, Gladkov AA, Todorovska MI. Ambient vibration tests of a seven-story
reinforced concrete building in Van Nuys, California, damaged by the 1994 Northridge earthquake. Soil Dynamics
and Earthquake Engineering 2000; 19:391411.
55. Mulhern MR, Maley RP. Building period measurements before, during and after the San Fernando, California, earth-
quake of February 9, 1971. U.S. Department of Commerce, National Oceanic and Atmospheric Administration:
Washington, DC 1973, I(B): 725733.
RESPONSE-ONLY MODAL IDENTIFICATION OF STRUCTURES USING STRONG MOTION DATA
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
56. Stewart JP, Fenves G. System identication for evaluating soil-structure interaction effects in buildings from strong
motion recordings. Earthquake Engineering and Structural Dynamics 1998; 27: 869885.
57. Trifunac MD, Ivanovic SS, Todorovska MI. Apparent periods of a building, I: Fourier analysis. Journal of Structural
Engineering 2001; 127(5):517526.
58. Trifunac MD, Ivanovic SS, Todorovska MI. Apparent periods of a building, II: time-frequency analysis. Journal of
Structural Engineering 2001; 127(5):527537.
59. Todorovska MI, Trifunac MD. Impulse response analysis of the Van Nuys 7-story hotel during 11 earthquakes and
earthquake damage detection. Structural Control and Health Monitoring 2008; 15:90116.
S. F. GHAHARI ET AL.
Copyright 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe

You might also like