You are on page 1of 18

Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

Contents lists available at SciVerse ScienceDirect

Soil Dynamics and Earthquake Engineering


journal homepage: www.elsevier.com/locate/soildyn

1D system identification of buildings during earthquakes by seismic


interferometry with waveform inversion of impulse responses—method
and application to Millikan library
Mohammadtaghi Rahmani, Maria I. Todorovska n
University of Southern California, Department of Civil Eng., Los Angeles, CA 90089-2531, United States

a r t i c l e i n f o abstract

Article history: Two new algorithms have been introduced as a further development of a robust interferometric
Received 7 November 2011 method for structural health monitoring (SHM) of buildings during earthquakes using data from
Received in revised form seismic sensors. The SHM method is intended to be used in an automatic seismic alert system, to issue a
4 August 2012
warning of significant damage during or immediately after the earthquake, and facilitate decision
Accepted 24 September 2012
Available online 1 November 2012
making on evacuation, to avoid loss of life and injury from possible collapse of the weekend structure
during aftershock shaking. The method identifies a wave velocity profile of the building by fitting an
equivalent layered shear beam model in impulse response functions (virtual source at roof) of the
recorded earthquake response. The structural health is monitored by detecting changes in the identified
velocities in moving time windows, the initial window being used as reference. Because the fit involves
essentially matching phase difference between motion at different floors, the identified velocity profile
is not affected by rigid body rocking, and soil-structure interaction in general, as demonstrated in this
paper. Consequently, detected changes in wave velocity during an earthquake are not affected by
changes in the soil-foundation system, which is a major advantage over SHM by detecting changes in
the observed modal frequencies. Further, the method is robust when applied to real buildings and large
amplitude earthquake response, as demonstrated in previous work. The new fitting algorithms
introduced are the nonlinear least squares (LSQ) fit and the time shift matching (TSM) algorithms.
The former involves waveform inversion of the impulse responses, and the latter - iterative matching of
the pulse time shifts, both markedly reducing the identification error as compared to the previously
used direct ray algorithm, especially for more detailed models, i.e., with fewer floors per layer.
Results are presented of identification of the NS, EW and torsional responses of the densely
instrumented Millikan Library (9-story reinforced concrete building in Pasadena, California) during a
small earthquake.
& 2012 Elsevier Ltd. All rights reserved.

1. Introduction paper introduces two new fitting algorithms, both iterative,


which reduce markedly the error as compared to the previously
This paper presents a significant new development of a used ‘‘direct’’ algorithm [1–3,5–7]. As shown in [1], the accuracy
method for system identification (SI) of buildings from earth- of the ‘‘direct’’ algorithm is poor for fitting more detailed models
quake records by wave travel time analysis, for use in structural (with fewer floors per layer), and inadequate for SHM. The first
health monitoring (SHM), being developed by the authors. The one is a waveform inversion algorithm, which we call ‘‘nonlinear
method involves fitting a layered shear beam model of the least squares (LSQ) fit algorithm’’, and which employs the
building in recorded earthquake response by matching pulse time Levenberg–Marquardt method [8]. The second one is an iterative
shifts in impulse response functions (IRF) with virtual source at ray algorithm, which we call ‘‘time shift matching (TSM) algo-
roof [1–3]. Since their first use for analysis of seismic response of rithm’’, and which fits only the time shifts of the direct pulses,
buildings [4], impulse response functions have shown to be a very recursively from top to bottom. The resolution and accuracy of
useful tool for analysis and system identification of buildings. This the method are discussed in detail. Good accuracy of identifica-
tion of more detailed models is essential for detecting local
damage, which has been a challenge [9–11]. This paper also
n
Corresponding author.
presents an application of the new algorithms to the NS, EW
E-mail addresses: mrahmani@usc.edu (M. Rahmani), and torsional responses of Millikan library (9-story reinforced
mtodorov@usc.edu (M.I. Todorovska). concrete building in Pasadena, California) during Yorba Linda

0267-7261/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.soildyn.2012.09.014
158 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

earthquake of 2002, which was recorded by a dense network of however, can be useful only if local variations can be resolved and
sensors [4]. The fundamental fixed-base frequencies inferred from measured accurately, which depends jointly on the nature of the
the fitted model are compared with published results by an response, model fitted and fitting algorithm [1]. This paper
independent method [12–14]. The SI method can be applied, in directly addresses the accuracy of the identification.
principle, to higher dimensions as well and to any structural The problem of the seismic response of a high-rise building is
model. The SI method is being developed for use in SHM systems similar to that of layered geology to vertically incident shear
to provide early warning (during or immediately after the strong waves. However, there are some major physical differences, and
shaking) that a structure may not be safe, and facilitate decision an identification method successfully applied in geophysics is not
making on evacuation of the weakened structure during shaking necessarily appropriate for buildings. One major difference is that
from aftershocks to avoid loss of life and injuries caused by buildings are slender structures, supported by a foundation, and
potential collapse [15–17]. The basis for such use is discussed in respond to horizontal earthquake shaking also by rocking motion.
the following. Another major difference is that they also deform in bending, to a
The seismic response of structures, commonly viewed as degree that varies considerably from one structure to another
vibration, can also be viewed as a wave propagation problem, [50], which introduces dispersion in the wave propagation. The
the two approaches being equivalent [4,18]. Continuous models, third difference is the stricter demand for accuracy in engineering
e.g., shear beams, are convenient for wave analyses of buildings, applications, especially for SHM. To be practically useful as a
because they can be directly p characterized
ffiffiffiffiffiffiffiffiffi by their wave velocity, decision making tool (evacuate or not), an identification algo-
e.g., shear wave velocity b ¼ m=r, where m is shear modulus and rithm must be sufficiently accurate. The effect of rocking can be
r is mass density [1–7,15–32]. As b is directly related to the eliminated successfully by an appropriate choice of identification
structural rigidity, its decrease indicates loss of stiffness, possibly algorithm [5,6]. The contribution from bending, and the resolu-
caused by damage [2,3,6,22,25]. The time lag t between motions tion and accuracy are addressed in this paper.
recorded at two locations, which is measurable from recorded In the methodology section of this paper, the two new
response, is directly related to b and can be used to identify it. identification algorithms are presented, and measures for testing
According to ray theory, which ignores wave scattering, b ¼ h/t, the goodness of fit and their validity for buildings. The results
where h is the distance. In buildings, t has been measured using section shows their application to the NS, EW and torsional
cross-correlation [25], pulse time shift in normalized input– responses of Millikan library during the Yorba Linda earthquake
output minimization (NIOM) [33–35], and pulse time shift in of 2002, which was recorded at every floor. Results are shown for
impulse response functions computed by deconvolution of the identified velocity profiles, which are the features needed for
recorded response [1–7,15–17,29]. The three approaches are SHM, and also for the modal frequencies of the identified fixed-
closely related and belong to the class of methods referred to in base model, and the inferred rigid body rocking frequencies,
seismology as seismic inteferometry [4,36–44]. Interferometry by which are compared with published results of soil-structure
deconvolution and correlation are closely related, the former system identification of the same building using data from forced
being preferable because the effects of the external source have vibration tests [12–14]. Finally, the main findings of this study are
been removed [4]. The deconvolved response gives the system summarized, implications on SHM of the approximations
impulse response function, representing the system response to a involved are discussed, and research for further development of
virtual source at the location of the reference point, and repre- this interferometric method is identified.
sents also the system Green’s function but for modified boundary
conditions [4,37]. Seismic interferometry by deconvolution, first
applied to buildings in [4], has been gaining increasing interest 2. Methodology
both in analyses of borehole data and buildings [41–48].
Increase in wave travel time t, detected by seismic interfero- 2.1. Model
metry, can be used for SHM as an indicator of possible damage.
Analyses of recorded earthquake response in damaged buildings The building is modeled as an elastic, layered shear beam,
have shown that identified increases in t are consistent with the supported by a half-space, and excited by vertically incident plane
presence, distribution and severity of observed damage shear waves (SV) (Fig. 1a) [1]. The layers may correspond to
[2,3,6,25,34,47,48]. Further, the identified t and its changes are individual floors, or to group of floors. Within each layer, the
not affected by the soil-structure interaction, as shown in [4] medium is assumed to be homogeneous and isotropic, and that
analytically for a model with horizontal response only, and in [5] perfect bond exists between the layers. The building is assumed
on simulated response for a model with coupled horizontal and to move only horizontally, in the y-direction, the foundation
rocking response. This is a major advantage of the wave approach rocking due to soil-structure interaction being ignored. The layers,
to SHM as compared to the prevailing vibrational approach, based numbered from top to bottom, are characterized by thickness hi,
on monitoring changes in observed modal frequencies, which are mass density ri, and shear modulus mi, where i¼1,y,n for the
those of the soil-structure system [5], and are sensitive to changes layers in the
pffiffiffiffiffiffiffiffiffiffiffi
ffi building, which implies shear wave velocities
in the soil [6]. Both soil softening and damage produce the same bi ¼ mi =ri . The displacements at the roof and at the consecutive
effect on the system frequencies, and separating their effects is layer interfaces are u1, u2, yun þ 1. Amplitude attenuation due to
difficult [9,10]. The significance of monitoring changes in wave material friction is introduced via the quality factor Q, which is
velocity instead of changes in the apparent frequency f1,app has equivalent to damping ratio z ¼1/(2Q). The motion of this model
been demonstrated for two damaged RC buildings in [2,3]. is identical to that of a layered half-space excited by vertically
Further, in [6], t helped resolve the degree to which the observed incident SH-waves (Fig. 1b) [36–44].
wandering of the apparent NS frequency of Millikan library [49]
during four earthquakes was caused by softening of the soil as 2.2. Observed impulse response functions
opposed to recoverable and permanent softening of the structure.
Another advantage of the wave travel time approach to SHM is its The observed IRF, hðz,0, omax ; t Þ, at a particular level z, for a
local nature, as localized damage would produce large change in b virtual source at the roof (z ¼0), and in the frequency band
locally but a small change in the frequencies of vibration, which o ¼ 2pf A ð0, omax Þ, is computed from a regularized transfer func-
are global characteristics of response [22,23]. The local nature, tion (TF) between the motions at that level, u(z), and the motion
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 159

Fig. 1. The model (a) Layered shear-beam representing a building. (b) Layered half-space.

8  
at roof, u(0), [4] >
<exp aðmÞ h
1 i omax
SAiðmÞ ðtÞ ¼   aðmÞ
i
cosðomax t Þ
^ uð0Þ
uðzÞ ^ omax >
: aðmÞ þt 2
2
h^ ðz,0; oÞ ¼ 2
ð1Þ i
9uð0Þ9
^ þe 9
>
=
as its inverse Fourier transform aðmÞ
i
þ tsinðomax t Þ  2 ð5Þ
Z omax >
1 aðmÞ þ t2 ;
hðz,0, omax ; t Þ ¼ h^ ðz,0; oÞ eiot do ð2Þ i
2p omax
In Eq. (3), t ðmÞ
i
are time shifts relative to source time, such that
In Eq. (1), e is a regularization parameter chosen to avoid
division by very small numbers [4]. t ð0Þ
1 ¼0
The cut off frequency, omax , determines the width of the ðm1Þ
t ðmÞ
2i1 ¼ t i þ Zm hm ; ðmÞ
t 2i ¼ t ði m1Þ Zm hm ð6Þ
pulses in the IRF, and controls the identification error [1].
For omax o1, the virtual source pulse is a box function in the and the coefficients b aðmÞ and (m)
are functions of the reflection and
i
frequency domain and the sinc function sinomax t=pt in the time transmission coefficients R and T such that
domain. The half-width of the main lobe of the sinc function
  a1ð0Þ ¼ 1
Dt ¼ p=omax ¼ 1= 2f max is a measure of its spread in time.
ðm1Þ ðm1Þ
As omax -1, Dt-0 and the source function approaches the Dirac aðmÞ
2i1 ¼ ai , aðmÞ
2i ¼ ai Rm , i - odd ð7Þ
d-function. As shown in [1], the accuracy with which the central aðmÞ ¼ aði m1Þ Rm , aðmÞ ¼ aði m1Þ , i - even
2i1 2i
time of the pulses in IRFs represents the true pulse arrival time
increases with decreasing pulse width, i.e., increasing omax .
ðmÞ
Y
m  
b ¼2 Tj ð8Þ
2.3. Model impulse response functions j¼2

For the model in Fig. 1, analytical expressions for both TFs and where Rm is the reflection coefficient of waves in m-th layer
band-limited IRFs have been derived using the propagator matrix reflected from the interface with the (m 1)-th layer and Tj is the
for SH waves in layered half-space [51–54]. The expressions for transmission coefficient for waves from the j-th layer into the
the IRFs we use are the same as those in [1,36], presented here for (j 1)-th layer defined as
completeness. Zm mm Zm1 mm1
For virtual source pulse of unit amplitude at roof, the IRF at Rm ¼ ð9Þ
Zm mm þ Zm1 mm1
interface z ¼zm is
m1
2X
aðmÞ h    i 2Zj mj
hðzm ,0, omax ; t Þ ¼ i
SC iðmÞ tt ðmÞ
i
þ SAðmÞ
i
t þt iðmÞ ð3Þ Tj ¼ ð10Þ
i¼1b
ðmÞ
Zj mj þ Zj1 mj1
where SC iðmÞ ðtÞ and SAðmÞ i
ðtÞare inverse Fourier transforms of In Eqs. (4) and (5), aðmÞ
i
are amplitude attenuation factors
attenuated box functions for causal and acausal waves given by
8  
að0Þ
1 ¼0
> ð11Þ
1 <exp aðmÞ i
o max h
ðmÞ
a2i1 ¼ aði m1Þ þ Zm hm zm ; aðmÞ ¼ aði m1Þ Zm hm zm
2i
SC ðmÞ
i
ðt Þ ¼  2 aiðmÞ cosðomax t Þ
omax >
: aðmÞ
i þ t 2
Further, hm is the thickness, Zm ¼1/bm is the vertical slowness,
9 zm ¼1/(2Qm) and mm is the shear modulus, all of these of the
>
=
aðmÞ
i m-th layer.
þtsinðomax t Þ  2 ð4Þ
> As seen from Eq. (3), the impulse response function at a level is
aðmÞ þt2 ;
i an assembly of shifted in time pulses created by transmission and
160 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

reflection at the boundaries (external and internal). There are 2.5. The time shift matching (TSM) algorithm
both causal and acausal pulses, all propagating downward [1,36].
The acausal ones are due to the fact that the virtual source (at top) This iterative algorithm involves progressive adjustment of the
does not coincide with the physical source (at bottom) [1,4,36]. layer velocities until the time shifts of the direct pulses in the
Most prominent are the causal and acausal direct pulses, which model IRFs are close enough to those in the observed IRFs.
are transmitted through each internal boundary, referred to as The time shift is measured from the time of the peak of the
direct pulses, which will be used for the fit [1]. pulse. The b profile is estimated recursively starting from the top
layer, and the average of the peak time for the causal and acausal
direct pulses is matched, which eliminates errors due to poor
2.4. The nonlinear LSQ fit algorithm synchronization of motions recorded by different channels, and
effects of pulse distortion. The starting value for the iteration is
This algorithm involves fitting, in the least squares (LSQ) sense, the estimate by the direct method, and the process is terminated
of pulses in the IRFs, as functions of time, over predefined time when the times of the peaks becomes smaller than the time step
windows. We use the Levenberg–Marquardt method for non- of the data.
linear least squares estimation, which is a fixed regressor, small
residual algorithm, considered to be generally robust, and is 2.6. Frequency band for the fit
implemented in Matlab [8]. Let b ¼(b1,b2,y,bL) be the vector of
the velocity profile, where L is the number of layers. Further, let The choice of the frequency band for the fit is very important.
hobs(zj,tji) (Eqs. (1) and (2)) be the observed IRFs
 in time window j The resolving power of the impulse response functions is limited
mod 
at point i of that window, and let h zj ,t ji ; b be the model IRF at by the effective bandwidth of the data f max , which determines the
the same point in space and time, for velocity profile b (Eq. (3)). width of the source pulse [1]. The resolution is higher and
Here zj is the z-coordinate of that window, and tji is the time of the the error is smaller if the IRFs are more broadband, but limiting
ith point of the jth window. The estimation is based on the the bandwidth is necessary to ensure compatibility of the simple
representation model with the data. The causes of incompatibility include
obs   mod   violation of the assumptions of nondispersive wave propagation
h zj ,t ji ¼ h zj ,t ji ; bn þ eji ð12Þ
and constant Q, and, at high frequencies, ignoring details of the
where bn is the true value of b, and eji is the error, assumed to be structural elements.
identically distributed, statistically independent, zero mean ran- Wave propagation in a pure shear beams is nondispersive
dom variable. Then, the least squares estimate of b, b^ , is such that (except for small dispersion due to material damping), while in a
it minimizes the sum of the squares of the error at all observation pure bending beam it is highly dispersive. Real buildings deform
points both in shear and in bending, the participation of each type of
deformation depending on the structural system. Significant
Nj
J X
X obs   mod   dispersion, i.e., such that there is a large difference in the phase
SðbÞ ¼ ½h zj ,t ji h zj ,t ji ; b 2 ð13Þ
and group velocities over the frequency band of the observation,
j¼1i¼1
would cause ‘‘splitting’’ of the source pulse in the IRFs as it
where Nj is the number of points in the jth window, and J is the propagates further down the building, which can be visually
total number of windows. As a measure of the speed of conver- detected, and fitting a layered shear beam model is such response
gence and error of the fit, the standard deviation of b^ , s, is is not meaningful. Ratios of observed frequencies of vibra-
obtained from the scatter of the successive estimates. tions suggest that the response of many buildings is closer
The value of Q in the model is assumed, as the variation of the to a pure shear beam (1:3:5y.) than to a pure bending beam
pulse amplitudes along the height is mostly governed by the (1:6.27:17.5y) (ignoring soil-structure interaction in the inter-
impedance of the layers [1]. Also, our experience showed that, for pretation of these ratios is in favor of bending beam behavior).
reasonably chosen values of Q, b^ is not sensitive to Q. Conse- Our experience with about 15 buildings shows that, by band-pass
quently, we use either the apparent Q estimated by the direct filtering, the effects of dispersion on the pulse shapes in the IRFs
method [1], or some value that approximately matches the can be reduced, so that the direct pulses closely resemble the
observed apparent damping. We leave detailed study of Q for source pulse. Within such bands, fitting an equivalent shear beam
the future. is possible, with the identified velocities representing the group
The time windows over which IRFs are fitted are chosen so velocity for the band of propagation of a lateral disturbance. In
that they enclose only the main lobes of both the acausal and this paper, we assume that after the filtering, the effects of
causal direct pulses. The reason for this is to maximize the signal dispersion are of second order, affecting the accuracy of the
to noise ratio, as, in practice, the reflected pulses from the layer method. We restrict the frequency band for the fit by visual
interfaces have much smaller amplitudes than the direct pulses, inspection of the pulse shapes in the IRFs, and by checking for
and are difficult to distinguish from the ripples of the sinc significant and systematic differences in pulse arrival times at
functions. different levels for different f max of the band (0, f max ). If possible,
The algorithm requires initial value of b, for which we use the the identification is carried out also in higher disjoint band(s).
estimate by the direct algorithm, obtained from the ratio of Observed transfer-function amplitudes in buildings suggest
distance traveled and pulse time shift (bi ¼hi/ti) [1]. Because the more rapid decay with frequency (due to structural and radiation
Levenberg–Marquardt method is a small residual, iterative damping) than predicted by a constant Q in the model (due to
method, to insure convergence, the initial value has to be reason- structural damping only). This leads to differences in the
ably close to the true value [1]. Our study of several buildings pulse shapes of model and observed IRFs, in the sense that
showed that such initial values are appropriate. beyond some frequency, which is much smaller than the cap-
The two main advantages of the nonlinear LSQ algorithm over ability of the sensors (typically 25 Hz or 50 Hz [55–57]), the
the direct algorithm [1] are that (1) it uses also information about observed transfer-functions have little energy, so that the pulse
the pulse amplitudes, which serves as additional constraint for width in the IRFs does not decrease with increasing frequency, in
the fit, and (2) it is not limited by the assumption of ray theory, contrast with the model [1]. In the experience of the authors,
e.g., that the variation of material properties is smooth [1]. similar pulse shapes of model and data produces better fit, and to
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 161

ensure this, we perform the fit in the effective band (0, f max ), measured from pulse time shifts in the IRF, and also by the
where f max is the maximum frequency beyond which the width of agreement of TFs (between roof and base, within the frequency
the main lobe of pulses in observed IRFs does not decrease. band of the fit, and considering frequency shift due to foundation
rocking), which shows how realistic the model is. The latter is an
2.7. Resolving power of IRFs and accuracy of the identification independent test because the model is identified from phase
delays in the time domain, while the check is in terms of
The resolving power of IRFs to detect variations in the velocity concentrations of energy (large amplitudes) in the frequency
profile, which is important for the detection of localized damage, domain. For the horizontal responses, the observed transfer
can be derived from Gabor’s theory of signals as follows [58,59]. functions correspond to a building that is pinned at the base,
Adopting the pulse half-width as a measure of the uncertainty of and can rotate about a horizontal axis, through which interaction
the time localization of the IRFs, two pulses canbe distinguished with the soil motion takes place. For the torsional response, the
 transfer functions are computed from the ratio of roof and base
as separate processes if shifted in time by t ¼ 1= 2f max . Recalling
that the time delay between an incoming pulse in a layer and its torsion, and correspond to a building on a foundation that cannot
reflection from the other layer boundary (internal or external) is twist but can move horizontally. Therefore, the observed modal
2h/b, the minimum layer width that can be resolved is hmin ¼ frequencies are the apparent frequencies of vibration, which
b=ð4f max Þ ¼ lmin =4, where lmin is the shortest wavelength in the depend on the foundation stiffness, the fundamental mode being
data. The data bandwidth is critical for resolving the top layer in most affected, while those of the model are the fixed-base
which the direct acausal and causal pulses (corresponding to the frequencies. Consequently, the fundamental frequency of the
incoming direct pulse and its reflection from the free surface) fitted model should be higher than that in the observed TF, more
have comparable amplitudes. In the interior layers, finer varia- so for stiffer and heavier structures compared to the soil, while
tions than those afforded by the data bandwidth can be identified the agreement should be closer for the higher modes. The fitting
from a spatial array of observations along the building height, algorithms are ranked by all of the mentioned criteria.
using the direct acausal and causal pulses, which are progres-
sively further apart in the observations closer to the base. The 3. Results and analysis
reflections from internal impedance contrasts are much weaker
than the direct pulses, and can be considered as ‘‘noise’’ in the In this section, the building and data are described,
data, affecting the accuracy of the identification. However, if and identification results and their analysis are presented for
h ohmin ¼ b=ð4f max Þ, the identified velocity profile will be aliased, equivalent uniform models, for which the fit is most robust, and
in the sense that the velocity in a layer will be affected by the for multilayer models by the three fitting algorithms (Direct,
velocity of the neighboring layers, with the radius of influence nonlinear LSQ fit and Time Shift Matching). The analysis includes
and degree depending on the ratio b=f max . This is similar to the (1) comparison of the performance of the three algorithms,
aliasing of frequency in sampling a signal that contains frequen- (2) demonstration that the results are not affected by soil-
cies beyond the Nyqusit frequency [60]. Hence, in the interior structure interaction (including foundation rocking), (3) identifica-
layers, the issue of resolution becomes an issue of accuracy. For a tion of the fixed-base frequencies and mode shapes, and the rigid
given structure, the error can be reduced by: (1) increasing f max , body rocking frequencies, and their comparison with published
which would decrease the radius of the aliasing, (2) increasing the results using forced vibration tests data and independent identi-
number of observation points, which would reduce the error fication method [14].
statistically, and (3) good choice of fitting algorithm, as it will be
shown in the results section of this paper. 3.1. Building description and observations

2.8. Goodness of fit and ranking of fitting algorithms Millikan library (Fig. 2) is a 9-story reinforced concrete build-
ing in Pasadena, California, instrumented over a period of 40
The goodness of fit is examined by analysis of s of the years, and tested extensively, in particular for soil-structure
nonlinear LSQ estimate, the agreement of pulse arrival times as interaction studies [5,6,12–14,61–63]. It is a rare example of a

Fig. 2. Millikan library: (a) photo (courtesy of M. Trifunac); (b) vertical cross-section; and (c) typical floor layout (top) and sensor locations at basement (bottom), redrawn
from [4].
162 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

densely instrumented building for which data is publically avail- The similarities of pulse width for both values, which should have
able and can be used for testing system identification methods. differed by a factor of 1.7 [1], and the very small TF amplitudes
The building is 21  23 m in plan, and vertically extends 43.9 m beyond 15 Hz, suggest effective f max around 15 Hz. Parts (b) also
above grade and 48.2 m above basement level (Fig. 2). Resistance show that the effective data bandwidth is not sufficient to resolve
to lateral forces in the NS direction is provided by RC shear walls on any pulses reflected from the internal boundaries, as well as the
the east and west sides of the building. The RC central core houses direct causal and acausal pulses at the 9th floor.
the elevators and provides resistance to lateral forces in the EW The pulse shapes suggest greater dispersion in the EW
direction. The local soil can be characterized as alluvium, with response over the same frequency interval, which we divided
average shear wave velocity in the top 30 m of about 300 m/s, and into two subbands, the lower one containing the first two modes
depth to ‘‘bedrock’’ of about 275 m [12–14,61–63]. Published work of vibration. Fig. 6 shows the observed IRFs for the whole effective
also suggests uniform mass distribution over the top three, middle band, 0–15 Hz (the solid line is part a), and for the two subbands,
three and bottom three stories [61], and we used r ¼526, 473 and 0–7.5 Hz (part b) and 7–15 Hz (part c), in which we could fit our
490 kg/m3, respectively [1]. model. Table 1 summarizes the observed pulse arrival times t ave i
Yorba Linda earthquake of 3 September, 2002 (M¼4.8, epi- at each level (average of the absolute values for the acausal and
central distance R ¼40 km) was recorded by a dense network of causal pulses), measured from the time of the peak of the pulse
sensors (Fig. 2c) operated by the U.S. Geological Survey [4]. The computed from interpolated data with Dt ¼0.001 s.
building response was small, with maximum rocking angle of
0.012  10  3 rad (NS response). Figs. 3–5 illustrate the observed 3.2. Equivalent uniform model fit
NS (at the West wall), EW and torsional responses. Parts (a) show
the acceleration time histories, parts (b) show the corresponding Table 2 shows the identified equivalent wave velocity beq ¼H/t
IRFs for virtual source at roof, and parts (c) show the transfer for the structure (t ¼pulse travel time from ground floor to roof,
functions between the roof and ground floor acceleration accel- H¼building height), the apparent quality factor Q computed from
erations. (The NS motions at the West wall were used instead of the peak amplitudes of the acausal and causal pulses, SA and SC, as
 
the average of the motions at NS and East walls because they 1=2Q ¼ ln SA=SC =omax t and the corresponding damping ratio
were recorded at all levels.) The acceleration data was made z ¼1/(2Q) all obtained by the direct algorithm [1,36]. It can be
available low pass filtered at 25 Hz and sampled at Dt ¼0.05 s. seen that t is the range 0.1–0.16 s, and beq is in the range
In parts (b) of Figs. 3–5, the solid and dashed lines correspond 240–460 m/s, being the largest for the tosional motions (463 m/s)
to cut off frequencies f max ¼ 2p=omax ¼15 Hz and 25 Hz. and the smallest for the EW motions 1st subband (243 m/s).

Fig. 3. Millikan library NS response during Loma Linda, 2002 earthquake observed at West wall (Fig. 2). (a) Accelerations. (b) Impulse responses for virtual source at roof.
(c) Transfer-function of roof acceleration w.r.t. ground floor.
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 163

Fig. 4. Millikan library EW response during Loma Linda, 2002 earthquake observed at West wall (Fig. 2). (a) Accelerations. (b) Impulse responses for virtual source at roof.
(c) Transfer-function of roof acceleration w.r.t. ground floor.

Comparison of beq for NS and EW motions suggests much stiffer results for ‘‘determined’’ models, that have as many layers as
structure in the NS direction (405 m/s as compared to 243 m/s and there are recordings that can be used for the fit. Because the
347 m/s), which can be attributed to the presence of the external causal and acausal pulses could not be resolved at the 9th floor,
shear walls at the East and West ends of the building (Fig. 2c). In this the top two stories were combined in one layer. For the torsional
study, identifying the structural damping is not of interest, and response, the top three layers were combined in one layer, and
results for Q are included as a byproduct of the analysis. It is the first and second floors were also combined in one layer. The
important to note that the identified Q, is not that of the structure different tables in this group correspond to different components
alone, and that it is an interval value, representative of the band of motion or subbands. The second group (Tables 6 and 7) shows
ð0,f max Þ. Table 2 shows that the apparent Q is in the range 14–19, results for ‘‘underdetermined’’ (9 layer) models, for the NS
implying interval apparent z ¼2.6 to 3.6%. These values are higher component and for the 1st subband of the EW component.
than the modal damping ratios for the fundamental mode of In each table, Column 1 shows the layer number (Fig. 1),
vibration, measured from the transfer-functions (Figs. 3–5). For Column 2 shows the floors contained in the layer, and Column
example, for NS motions, the interval z ¼3.6%, and the modal 3 shows the layer height, hi. The next group of three columns
z ¼1.7–1.8% for the fundamental mode, which is at the low end of (4–6) shows results related to identification of the layer velocity
the band. This suggests that the actual (soil-structure system) by the direct algorithm, which were used as initial values for the
attenuation is larger for higher frequencies than predicted by constant nonlinear LSQ and the TSM algorithms. Column 4 shows the
Q. The difference in the interval z for the two bands in which EW observed travel time through the layer, ti (determined from
motions was analyzed (2.6 Hz in subband 0–7.5 Hz vs. 3.5 Hz in the pulse arrival times in Table 1), Column 5 shows the identified
subband 7–15 Hz) supports this interpretation. The interval z is the velocity bi ¼hi/ti, and Column 6 shows the relative error
largest for the NS motions (z ¼3.6% vs. 2.6% for the 1st subband of EW Dt/t ¼(tpred  tobs)/tobs, where tpred is the travel time through the
motions and for torsion), consistent with the significant foundation layer measured from model IRFs, and tobs is the observed travel
rocking for this component of motion [13–18,63]. times, measured from observed IRFs (same as in column 4). This ratio
is also a measure of Db/b ¼  Dt/t (which follows from b ¼h/t and
3.3. Equivalent multilayer model fit differential calculus). However, it is a rough estimate of Db/b, because
it is obtained from two readings that both have error [1]. The next
The results of the identification based on more detailed models group of four columns (7–10) shows results of identification by the
are summarized in Tables 3–7. The first group (Tables 3–5) shows nonlinear LSQ algorithm. Columns 7, 8 and 9 show the identified b,
164 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

Fig. 5. Millikan library torsional response during Loma Linda, 2002 earthquake observed at West wall (Fig. 2). (a) Difference of NS accelerations recorded at East and West
wall. (b) Impulse responses for virtual source at roof. (c) Transfer-function of roof torsional acceleration w.r.t. ground floor.

Fig. 6. Evidence of dispersion in the observed IRFs of EW response of Millikan library during Loma Linda, 2002 earthquake. (a) 0–15 Hz (solid line) and 0–7.5 Hz (dashed
line). (b) 0–7.5 Hz. (c) 7.5–15 Hz.

the standard deviation of the estimate, sb, the ratio sb/b, and Dt/t. The identified b- profiles for the first group of models (Tables
Finally, Column 11 shows the identified b by the Time Shift Matching 3–5) are shown graphically in Fig. 7. The different columns show
(TSM) algorithm. results by the nonlinear LSQ, TSM and Direct algorithms (left to
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 165

right). Each bar corresponds to a different floor, and when some algorithm, the 2sb interval is shown as a solid line at the right end
floors were merged in a single layer, the bars show the equivalent of the bars. The sketch in the last column shows the layering
value for the floors in the layer. In the plots for the nonlinear LSQ structure. Fig. 8 shows the agreement of the IRFs for the model
(identified by the nonlinear LSQ algorithm) with the observed
Table 1 IRFs for the NS, EW (1st subband) and torsional motions (parts (a),
Observed pulse arrival times t ave
i (average for the causal and acausal pulses) at (b) and (c)), computed for z ¼2%. Further, Figs. 9–11 show,
each floor in Millikan library response during the Yorba Linda earthquake of 2002. respectively for the NS, EW and tosrsional motions, the agreement
of pulse arrival times (part a) and of transfer-functions (part b).
Level zi NS EW Torsion
[m]
The pulse arrival times for the TSM algorithm are indistinguish-
(0–15 Hz) (0–25 Hz) (0–7.5 Hz) (7–15 Hz) (0–25 Hz) able from the observed ones and are not shown. Finally, Fig. 11
t ave
i [s] t ave
i [s] t ave
i [s] t ave
i [s] t ave
i [s] compares the b- profiles for the 8-layer and 9-layer models for NS
and for the 1st subband of EW motions (Tables 6 and 7), all
Roof 0 0 0 0 0 0
identified by the nonlinear LSQ algorithms.
9 4.25 – – – – –
8 8.5 0.04 0.0395 0.0745 0.0415 –
We first discuss the results obtained by the nonlinear LSQ
7 12.75 0.048 0.047 0.0885 0.047 0.039 algorithm for the different components of motion. The b-profiles
6 17 0.0535 0.053 0.095 0.056 0.046 in Fig. 7 (1st column) suggest increasing floor stiffness towards the
5 21.25 0.0615 0.0615 0.101 0.0685 0.053 bottom and a soft first story for NS motions and torsion, which is
4 25.5 0.073 0.0725 0.111 0.083 0.059
consistent with the building design. The soft first story is clearly
3 29.75 0.0785 0.078 0.1305 0.0965 0.064
2 34 0.0835 0.083 0.1485 0.102 – identified despite the expected aliasing errors. The comparison of
Gnd 38.9 0.096 0.096 0.16 0.112 0.084 IRFs in Fig. 8 shows that the agreement of pulses is generally good.
The agreement is worse at the 7th and 8th floors, where the time
Some values are missing where the causal and acausal pulses could not be shift between the acausal and causal pulses is the smallest, and their
resolved or where motion was not recorded.
interference is the strongest [1]. The agreement is the worst for the
1st subband of EW motion, which has the smallest bandwidth, and
Table 2 the identified variation in velocity in the top three stories is not
Identified equivalent uniform model parameters over the building height, where reliable (Fig. 7, row 2 and column 1). Comparison of sb (Fig. 7 and
Qeq is the apparent quality factor.
Tables 3–5) shows that sb/b is the largest (10%) at the 2nd floor for
Component Frequency t [s] beq f 1,eq ¼ 41t Qeq zeq ¼ 1 NS motions, which is above the soft 1st floor, suggesting slower
2Q eq
band [Hz] [m/s] [Hz] [%] convergence of the algorithm in layers with large impedance
contrast relatively to the neighboring layers.
NS 0–15 0.096 405 2.6 13.9 3.6 Next, we compare the agreement of the velocity profiles in
EW 0–7.5 0.16 243 1.56 19.2 2.6 Fig. 7 obtained by the different algorithms. It can be seen that the
EW 7–15 0.112 347 2.23 14.3 3.5
Torsion 0–25 0.084 463 2.98 19.2 2.6
profiles by the TSM algorithm are consistent with those by the
nonlinear LSQ algorithm, except in the top layers for EW motion

Table 3
Comparison of shear wave velocities of Millikan Library NS response during Yorba Linda earthquake of 2002, as identified using
8-layer model by three algorithms: Direct, nonlinear LSQ and Time Shift Matching (TSM) (0–15 Hz).

Layer Floor hi [m] Direct LSQa TSM

ti [s] bi [m/s] 9Dt9/t [%] bi [m/s] sb [m/s] sb/b [%] 9Dt9/t[%] bi [m/s]

1 8–9 8.5 0.04 212.5 5 281 2.3 0.8 11 226


2 7 4.25 0.008 531.3 81 226.6 3.6 1.6 37.5 292
3 6 4.25 0.0055 772.7 64 360.8 11.3 3.1 63.6 599
4 5 4.25 0.008 531.3 137 554.4 27.5 5.0 6 639
5 4 4.25 0.0115 369.6 22 516.6 21.5 4.2 13 554
6 3 4.25 0.0055 772.7 73 638 40.5 6.3 0 657
7 2 4.25 0.005 850.0 80 835 87.3 10.5 20 723
8 Gnd 4.9 0.0125 392.0 100 491.6 18.1 3.7 12 478

a
Normalized_RMS error¼ 28.8%.

Table 4a
Comparison of shear wave velocities of Millikan Library EW response during Yorba Linda earthquake of 2002, as identified using
8-layer model by three algorithms: direct, nonlinear LSQ and Time Shift Matching (TSM) (0–7.5 Hz).

Layer Floor hi [m] Direct LSQa TSM

ti [s] bi [m/s] 9Dt9/t [%] bi [m/s] sb [m/s] sb/b [%] 9Dt9/t[%] bi [m/s]

1 8–9 8.5 0.0745 114 11 173 0.86 0.5 28 132.6


2 7 4.25 0.014 303.6 82 172 2.3 1.3 85 151.8
3 6 4.25 0.0065 654 84 141 1.5 1.1 192 269.7
4 5 4.25 0.0065 654 84 397 27 6.8 23 392.3
5 4 4.25 0.0095 447 31 344 13 3.8 26 461.3
6 3 4.25 0.0195 218 215 328 9.5 2.9 5 408.6
7 2 4.25 0.018 236 69 331 9.5 2.8 2.8 378
8 Gnd 4.9 0.0115 426 78 384 17 4.4 13 362

a
Normalized_RMS error¼24%.
166 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

Table 4b
Comparison of shear wave velocities of Millikan Library EW response during Yorba Linda earthquake of 2002, as identified using
8-layer model by three algorithms: Direct, nonlinear LSQ and Time Shift Matching (TSM) (7–15 Hz).

Layer Floor hi [m] Direct LSQa TSM

ti [s] bI [m/s] 9Dt9/t[%] bi [m/s] sb [m/s] sb/b [%] 9Dt9/t[%] bi [m/s]

1 8–9 8.5 0.0415 205 3.6 247 1.83 0.7 3.6 218
2 7 4.25 0.0055 773 100 294 5.56 1.9 0 333
3 6 4.25 0.009 472 0 393 8.9 2.3 27.8 498
4 5 4.25 0.0125 340 96 457 12.3 2.7 8 510
5 4 4.25 0.0145 293 55 334 6.8 2 10.3 311
6 3 4.25 0.0135 315 30 327 6.5 2 3.7 287
7 2 4.25 0.0055 773 36.5 487 18.7 3.8 9 511
8 Gnd 4.9 0.01 490 15 432 11.6 2.7 5 447

a
Normalized_RMS error¼ 29%.

Table 5
Comparison of torsional wave velocities of Millikan Library EW response during Yorba Linda earthquake of 2002, as identified using
6-layer model by three algorithms: Direct, nonlinear LSQ and Time Shift Matching (TSM) (0–25 Hz).

Layer Floor hi [m] Direct LSQa TSM

ti [s] bi [m/s] 9Dt9/t [%] bi [m/s] sb [m/s] sb/b % 9Dt9/t[%] bi


[m/s]

1 7–9 12.75 0.039 327 5.1 348.6 2.7 0.8 3.8 339
2 6 4.25 0.007 607 57.1 417.4 16.3 3.9 21.4 448
3 5 4.25 0.007 607 21.4 487 19.9 4.1 28.6 567
4 4 4.25 0.006 708.3 33.3 813.4 73.3 9.0 25 797
5 3 4.25 0.005 850 30 739 50.6 6.8 10 850
6 Gnd–2 9.15 0.02 457.5 0 587.4 26 4.4 15 457

a
Normalized_RMS error¼ 37%.

Table 6
Comparison of shear wave velocities of Millikan Library NS response during Yorba Linda earthquake of 2002, as identified using
9-layer model by three algorithms: Direct, nonlinear LSQ and Time Shift Matching (TSM) (0–15 Hz).

Layer Floor hi [m] Direct LSQa

ti [s] bi [m/s] 9Dt9/t [%] bi [m/s] sb [m/s] sb/b [%] 9Dt9/t [%]

1 9 4.25 – 212.5 – 224 4.4 2 –


2 8 4.25 0.04 212.5 5 460 22 4.8 30
3 7 4.25 0.008 531.3 81 290 7 2.4 81
4 6 4.25 0.0055 772.7 64 265 5.3 2 154
5 5 4.25 0.008 531.3 137 410 16 3.9 12
6 4 4.25 0.0115 369.6 22 419 15 3.6 17
7 3 4.25 0.0055 772.7 73 683 59 8.6 18
8 2 4.25 0.005 850.0 80 884 112 12.7 40
9 Gnd 4.9 0.0125 392.0 100 494 18 3.6 32

a
Normalized_RMS error¼ 27.7%.

Table 7
Comparison of shear wave velocities of Millikan Library EW response during Yorba Linda earthquake of 2002, as identified using
9-layer model by three algorithms: Direct, nonlinear LSQ and Time Shift Matching (TSM) (0–7.5 Hz).

Layer Floor hi [m] Direct LSQa

ti [s] bi [m/s] 9Dt9/t [%] bi [m/s] sb [m/s] sb/b [%] 9Dt9/t[%]

1 9 4.25 – 114 – 159 3.7 2.3 –


2 8 4.25 0.0745 114 11 193 2.8 1.5 23
3 7 4.25 0.014 303.6 82 178 2.9 1.6 61
4 6 4.25 0.0065 654 84 142 1.5 1.1 192
5 5 4.25 0.0065 654 84 379 23 6.1 23
6 4 4.25 0.0095 447 31 343 13 3.8 26
7 3 4.25 0.0195 218 215 328 9.3 2.8 2.5
8 2 4.25 0.018 236 69 330 9.3 2.8 8.3
9 Gnd 4.9 0.0115 426 78 380 16.5 4.3 8.7

a
Normalized_RMS error¼ 24%.
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 167

Fig. 7. Identified high resolution velocity profiles, obtained by different algorithms, of Millikan library during the Yorba Linda, 2002, earthquake. (a) NS response (0–15 Hz),
(b) EW response (0–7.7.5 Hz), EW response (7–15 Hz), and (d) torsional response (0–25 Hz). The algorithms used are: LSQ ¼nonlinear least squares fit of the direct pulses,
TSM¼ time shift matching, and Direct¼ the direct algorithm.

1st subband, for the reason mentioned earlier. In contrast, the Comparison of the pulse arrival times, measured from
profiles by the direct algorithm differ considerably, except for the model and observed IRFs (Figs. 9–11), shows remarkably better
torsional model which has fewer layers [1]. fit by the nonlinear LSQ algorithm in comparison to the direct
168 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

Fig. 8. Agreement of IRFs of models identified by the nonlinear LSQ algorithm with the observed ones. (a) NS response, 0–15 Hz. (b) EW response, 0–7.5 Hz. (c) Torsional
response, 0–25 Hz. B ¼2% was assumed.

Fig. 9. Other measures of agreement of the identified models of the NS response of Millikan library with the observations during Yorba Linda earthquake of 2002.
(a) Agreement of pulse arrival times at the different floors. (b) Agreement of transfer-functions. Note: the arrival at the 9th floor could not be resolved.

Fig. 10. Other measures of agreement of the identified models of the EW response of Millikan library with the observations during Yorba Linda earthquake of 2002.
(a) Agreement of pulse arrival times at the different floors. (b) Agreement of transfer-functions. Note: the arrival at the 9th floor could not be resolved.

algorithm. The agreement is very good, except for the 1st error being different for the model and observed IRFs, because of
subband of EW motions near the top for the reasons explained differences in the pulse shapes.
earlier. For the direct algorithm, the agreement, which was The comparison of TFs (second mode only) (Figs. 9–11) is in favor
previously found to be poor for NS motions [1], is much worse of the nonlinear LSQ algorithm. The agreement is very good for both
for the lower frequency subband of EW motions, but is better for algorithms for the NS motions. The agreement is not so good for the
the torsional motions, which is because of the smaller number of EW motions and is slightly better for the nonlinear LSQ algorithm. For
layers. the torsional motions, the agreement is very good for the nonlinear
Figs. 9–11 show better agreement of pulse arrival times for the LSQ algorithm, and not so good for the TSM algorithm. The shift of the
TSM algorithm than for the nonlinear LSQ algorithm. However, first mode we discuss in the next sections.
this does not mean that the TSM algorithm is better, because the Finally, we compare the profiles identified by fitting ‘‘deter-
pulse time shifts are not exactly at the physical arrival times, the mined’’ and ‘‘underdetermined’’ models (Fig. 12). It can be seen
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 169

Fig. 11. Other measures of agreement of the identified models of the Torsional response of Millikan library with the observations during Yorba Linda earthquake of 2002.
(a) Agreement of pulse arrival times at the different floors. (b) Agreement of transfer-functions. Note: the arrival at the 9th floor could not be resolved, and the observed
torsion is not available at the 2nd and 8th floors.

Fig. 12. Results for 9-layer models and comparison with the previous ones (Fig. 7).

that the profiles differ mainly near the top. The comparison of the IRFs and the identified velocities at a particular level are not
transfer-functions (Figs. 9 and 10) is in favor of the determined affected by the structure underneath them, including the soil.
model for the NS motions, and is comparable for the EW motions. Fig. 14 shows IRFs for a simple 2D soil-structure interaction model
with foundation rocking (part a), in which the structure is a uniform
shear beam with properties same as those of the identified equivalent
3.4. Demonstration of insensitivity of SHM to soil-structure uniform model for Millikan library NS response (b ¼405 m/s, Table 2),
interaction and the soil is modeled as a half-space with shear wave velocity of
250 m/s [5]. Parts (b) and (c) compare IRFs (0–15 Hz) and TFs of the
Previous work showed that the identified velocities by this soil-structure interaction model (solid lines), and of the structure
interferometric method are not affected by soil-structure inter- itself (dashed lines), the former representative of observed response,
action, both for a system without and with foundation rocking which in general contains rigid body rocking, and the latter repre-
[4,5]. For completeness of this paper, and due to the importance sentative of the fitted model. It can be seen that the rocking has a
of the issue for SHM, we demonstrate this again. significant effect on the 1st mode frequency in the TF (part c)), but no
Fig. 13, redrawn from [1], shows IRFs for a 3-layer model of effect on the pulse time shifts in the IRFs (part b)). Because all three
Millikan library NS response (no foundation rocking, bi ¼242 m/s, algorithms essentially fit the pulse time shifts in the IRFs, the
569 m/s and 536 m/s from top to bottom), for high f max ( ¼50 Hz) identified wave velocities are not affected by the foundation rocking,
for which the internal reflections are resolved. It can be seen that and represent the true fixed-base properties of the structure. Conse-
no pulses are reflected back to the roof. The internally reflected quently, in SHM application, detected changes in the b-profiles will
pulses that physically propagate upwards become downward not be affected by changes in the soil. In contrast, the pulse
propagating acausal pulses in the IRFs [1,4,36]. Consequently, amplitudes differ, and the identified Q (Table 2) represents the
170 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

data from forced vibration tests conducted in 1975 [63]. The two
identification studies are independent (time domain identification
using broadband excitation in our study and frequency domain
identification using single frequency excitation in [14]). The response
levels during Yorba Linda earthquake of 2002 were comparable to
those typical of forced vibration tests.
Comparison of the apparent frequencies suggests lower values
during the 2002 earthquake than during the test in 1975, by
about 4% for NS response, 7% for EW response and 17% for torsion,
consistent with the long term trend for this building [49],
attributed to structural degradation. Comparison of the fixed-
base and rigid-body rocking frequencies shows good qualitative
agreement, but higher values for f1 identified by our method for
all three degrees of freedom by about 20–30% depending on
which value is used to normalize (3 Hz vs. 2.33 Hz for NS
response, 1.68 Hz vs. 1.39 Hz for the EW response, and 3.5 Hz
vs. 2.94 Hz for torsion). This difference is reflected on the inferred
reduction of the fundamental mode frequency due to soil-
structure interaction, which is larger in our study (42% vs. 22%
for NS motion).
To some degree, the higher f1 identified by our method is due to
difference in the point of fixity, which is at ground level in our study
and at the basement in [14]. Other possible causes for the differences
Fig. 13. Broad-band impulse response functions (0–50 Hz), of a 3-layer model of
are various assumptions made in both methods, e.g., ignoring
Millikan library NS response (three stories per layer), showing that there is no dispersion in our method, and simplifying assumptions about the
constructive interference of the direct pulse and the reflections from the internal foundation deformation in [14], as well as differences in the nature
interfaces (redrawn from [1]). and location of the source of excitation, which is broadband earth-
quake excitation at base in our study, and monochromatic force at
combined effect of structural damping and radiation damping asso- roof in [14]. Based on our analysis of IRFs in a Timoshenko beam,
ciated with rocking motion. dispersion present to some degree in the data also contributes to such
bias. Detailed investigation of this effect will be reported in the future.
It is worth noting that the assumption of uniform properties along the
3.5. Identified fixed-base frequencies and mode shapes
height in the 1-layer model produces a bias in f1 in the opposite
direction, which compensates to some degree the bias caused by
Once the layering structure has been identified, the fixed-base
ignoring dispersion, and may give a closer value to the true one than
modal frequencies and shapes can be computed from the model
the multilayer model for use in studies like those in [4,6]. For the NS
TF. Table 8 shows readings of the apparent and fixed-base
response, the 1-layer model [6] does give closer estimate of f1 to those
frequencies for the first two modes, fi,app and fi, i¼1,2, identified
identified in [14].
from plots of the observed TFs (Figs. 3–5) and the best multi-layer
Once the layered beam model has been identified, it can be used
model TFs (Figs. 9–11), and Fig. 15 shows the first few fixed-base
to predict the response of the structure, by convolving the system
mode shapes. In Table 8, the relative difference between the
function with the base motion, and assess the forces and deforma-
apparent and fixed-base frequencies is also shown, as well as the
tions in the structural members, which can serve as additional
rigid-body rocking frequency, fR, inferred from the relationship
indicators, performance based, of possible damage [64].
1 1 1
¼ þ ð14Þ
2
f 1,app
2
f1
2
fR 4. Summary and conclusions

The identified frequencies, for the NS response are: 4.1. Summary of findings of this study
f1,app ¼ 1.72 Hz, f1 ¼3 Hz and fR ¼ 2.1 Hz, for the EW response are:
f1,app ¼ 1.12 Hz, f1 ¼1.68 Hz and fR ¼1.5 Hz, and for the torsional The most significant findings of this study about the method
response are: f1,app ¼2.4 Hz, f1 ¼3.5 Hz. These results suggest are as follows. (1) The nonlinear LSQ identification algorithm
significant difference for the fundamental mode, and small (waveform inversion), produces much better fit than the other
difference for the second mode. The fundamental fixed-base two algorithms, and should be used for identification of more
frequency is higher than the apparent one by about 40% for the detailed building models. (2) The resolving power of the IRFs for a
NS response, and about 30% for the EW response and torsion. For particular structure is limited by the effective bandwidth of the
the second mode, the fixed-base frequency is slightly lower than data, i.e., the maximum frequency f max for which the observed
the apparent one, by 2.5% for the NS, and 6.6% for the EW transfer function has significant energy, and for which the model
responses, and for torsion the difference is only 1.2%. The trend and building are compatible. The theoretical minimum layer
of the difference for the fundamental mode is consistent with width that can be resolved is hmin ¼ b=ð4f max Þ ¼ lmin =4. If obser-
frequency shift due to soil-structure interaction, which is vations on a dense array of points are available, e.g., there is a
expected to be significant for this stiff and heavy RC building. sensor at every or every other floor, a finer model can be
Next, we compare the identified f1,app, f1 and fR with published identified. However, if h o hmin ¼ b=ð4f max Þ, the identified velocity
results by an independent study [14], also shown in Table 8. Soil- profile will be aliased, i.e., there will be some ‘‘spilling’’ of velocity
structure system identification of buildings has been a very difficult from one layer over into the neighboring ones, or ‘‘bleeding’’ in
problem, and such studies are rare, but do exist for this building. In the identified velocity profile, the radius of influence depending
the following, we compare our results with those of detailed studies on b=f max . Due to aliasing errors, localized severe damage will
of Millikan library published in the late 1980s [12–14], which used produce some change also in the identified velocities of the
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 171

Fig. 14. (a) A simple 2D soil-structure interaction (SSI) model of Millikan library NS response, with foundation rocking [5] simulating the ‘‘observed’’ response.
(b) Comparison of impulse responses of the SSI model (solid lines) and fixed-base structure (dashed lines). (c) Same as part (b) but for the transfer-functions (roof w.r.t.
ground floor). It can be seen that the foundation rocking, which affects significantly the 1st mode frequency, has no effect on the pulse time shifts.

Table 8
Identified frequencies of vibration of Millikan Library (apparent, fixed-base and
rigid body rocking) Yorba Linda, 2002, earthquake and comparison with results by
Luco et al. [14] from forced vibration tests conducted in 1975 [60].

NS EW Torsion

This Luco et al. This Luco et al. This Luco


study [14] study [14] study et al.[14]

f1,app 1.72 Hz 1.79 Hz 1.12 Hz 1.21 Hz 2.4 Hz 2.89 Hz


f1 3.0 Hz 2.33 Hz 1.68 Hz 1.39 Hz 3.5 Hz 2.94 Hz
f 1 f 1,app 42.6% 22.2% 33% 12.3% 31% 1.7%
f1
fR 2.1 Hz 3.01 Hz 1.5 Hz 2.57 Hz Fig. 15. Fixed-base mode shapes of Millikan library NS and EW responses. The
f2,app 7.4 Hz 4.82 Hz  8.7 Hz frequency band of the data used to identify the models is indicated.
f2 7.22 Hz 4.52 Hz 8.6 Hz
f 2 f 2,app  2.5%  6.6% 1.2%
f2

very important for SHM, and a significant advantage over the


modal methods. (4) Once the b- profile has been identified, the
neighboring layers, which should be considered in the inference fixed-base modal frequencies (and shapes) can easily be identified
of damage from detected changes. (3) Wave velocity profiles from the model TFs. Despite biases in their values, the identified
identified by fitting IRFs with virtual source at roof are not fixed-base properties and their changes during an earthquake are
affected by foundation rocking. This insensitivity of the damage not affected by the soil and changes in the soil. It is remarkable
sensitive parameter to the effects of soil-structure interaction is that this is accomplished without the need for a model for the
172 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

soil-structure interaction. (5) The identified model can also be Of interest for SHM is how small and how localized damage
used for prediction of the deformations and forces in the struc- the method can detect. Our experience with damaged RC build-
tural members, in a performance based method for structural ings [2,3,47,48] indicates that earthquake damage in such build-
health monitoring. ings is typically wide spread, and that the correlation between
Significant findings of the identification of Millikan library are change in wave velocity and damage depends on the history of
that (6) an equivalent layered shear beam could be fitted in its NS, prior exposure to strong shaking. In Millikan library, beq dropped
EW and torsional responses in bands that include the first two permanently between 1970 and 2002 by about 22%, most of the
modes of vibration, and for the EW response also in a higher change occurring during the San Fernando earthquake of 1971,
disjoint band. (7) The identified fixed-base frequencies agree which was its first exposure to strong shaking, and which caused
qualitatively with published results by an independent method no visible damage in the structure [6]. In contrast, the subsequent
[14], suggesting significant shift of the fundamental frequency Whittier-Narrows earthquake of 1987 caused 12% or recoverable
due to soil-structure interaction, more so for the NS response. change, interpreted to be due to opening and closing of existing
However, they are systematically higher (by about 20% for the NS cracks in the concrete [6]. Our analysis of a 12-story RC building
and EW motions), the bias likely being due to dispersion, which is in Sherman Oaks, California, showed that minor structural and
ignored in our model. (8) Although motion was recorded at nonstructural damage during the San Fernando earthquake of
every floor, the effective bandwidth ð0,f max Þ for this RC building 1971 produced changes in the equivalent 2-layer model shear
is not sufficient to resolve the top floor, and the identified wave velocities of 23–27% in the top layer and 31–37% in the
velocities in the interior single floor layers are aliased (based on bottom layer (the range describes the variation over NS and EW
the theoretical quarter minimum wavelength rule), more so responses) [47,48]. Similar analyses of more buildings are needed
towards the bottom where the structure is stiffer. The redundancy to further calibrate the method for different types of structures.
of sensors is still useful because it helps reduce the error. Despite
aliasing errors, the identified velocity profile correctly reflects
4.3. Conclusions
decrease in structural stiffness towards the top and a soft first
story. An accurate account of the aliasing errors and their effect
Based on the analysis in this paper, it is concluded that an
on SHM are out of the scope of this paper, and will be reported in
equivalent layered beam model deforming in shear only can be
the future.
fitted in Millikan library’s NS, EW and torsional responses such
that matches vertical pulse propagation within carefully chosen
frequency bands. The waveform inversion algorithm produces the
4.2. Discussion
best fit and should be used for models with thinner layers. Similar
conclusion was reached in our analysis of a 54-story steel
The layered shear beam does not account for details of the
building [48,66] for a band that includes the first five modes of
variations of stiffness in the building, which is constructed from
vibration. Based also on its recent applications to full-scale
many different elements, the SHM method aiming to identify only
buildings, which suffered no visible, minor and severe damage,
the part of the structure (story or group of stories as afforded by
this interferometric SHM method is robust when applied to full-
the data bandwidth and density of sensors) that has been
scale data, sensitive to damage in RC buildings of concern for
damaged and the degree of damage, to facilitate decision making
structural safety, and not sensitive to changes in the soil. Further
on evacuation immediately after the earthquake. The effects of
research is needed to determine its effectiveness for detecting
details in the stiffness distribution are removed by low-pass
damage in steel structures, its limits in detecting localized
filtering of the IRFs before fitting. The remaining longer waves
damage in both RC and steel buildings, and the role of wave
are not sensitive to inhomogeneities of much smaller scale
dispersion on those limits. The layered shear beam model is
[20,54]. Lateral variations of greater extent, e.g., cores and shear
convenient because it does not require a detailed model of the
walls forming vertical waveguides, may have significant effect
structure. A further development of the method by replacing the
even on longer waves, e.g., contributing to wave dispersion [20].
layered shear beam model by more realistic models (e.g., such
Analysis of these effects is left for the future. Insofar, the method
that account for wave dispersion and for finer details in the
is based on the assumption that, within controlled frequency
design), will enable to broaden the frequency band of the fit,
bands, the wave velocity does not vary greatly.
which will increase its resolution and accuracy. Towards this end,
The model in this paper does not considers explicitly the floor
the layered shear beam model has been a useful vehicle, provid-
slabs. Despite their small size, because of the large mass, the slabs
ing valuable insight into this interferometric SHM method.
act as scatterers of the vertically propagating waves, affecting
more the shorter wavelengths, and result in additional phase
delay in the pulse propagation, and in additional damping.
Acknowledgements
The layered building model and the wave inversion algorithm
can account for these effects, the slabs being modeled as thin,
dense layers. We are conducting a detailed study of these effects, This material is based upon work supported by the National
the results of which will soon be reported. As mentioned earlier in Science Foundation under Grant No. MMI-0800399. The strong
this paper, we are also investigating the effects of bending on the motion data used was obtained from the Engineering Center for
identified equivalent layered shear beam velocity profile, using a Strong Motion Data (www.strongmotioncenter.org/). The authors
Timoshenko beam model, and will soon report the results. are also grateful to F. Udwadia, M. Trifunac and T. Heaton for the
Displacement of heavy objects in the building, e.g., toppling of stimulating discussions on this topic, and to the two anonymous
furniture, equipment and books, will affect the high frequencies reviewers for their constructive and insightful comments, which
of response in the vicinity of such events, as higher frequencies significantly improved the manuscript.
attenuate faster with distance. Consequently, they would produce
effects similar to damage only in methods that use the higher References
frequency part of the spectrum (e.g., novelties detection [65]), and
are not expected to have significant effects on detected changes in [1] Todorovska MI, Rahmani MT System identification of buildings by wave
wave velocities from low-pass filtered IRFs. travel time analysis and layered shear beam models—spatial resolution and
M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174 173

accuracy, Structural Control and Health Monitoring http://dx.doi.org/10. [28] Gičev V, Trifunac MD. Energy and power of nonlinear waves in a seven story
1002/stc.1484, preview published online 6 March, 2012, in press. reinforced concrete building. Indian Society of Earthquake Technology
[2] Todorovska MI, Trifunac MD. Earthquake damage detection in the Imperial Journal 2007;44(1):305–23.
County Services Building III: analysis of wave travel times via impulse [29] Kohler MD, Heaton T, Bradford SC. Propagating waves in the steel, moment-
response functions. Soil Dynamics and Earthquake Engineering 2008;28(5): frame factor building recorded during earthquakes. Bulletin of the Seismo-
387–404. logical Society of America 2007;97(4):1334–45.
[3] Todorovska MI, Trifunac MD. Impulse response analysis of the Van Nuys [30] Gičev V, Trifunac MD. Rotations in a shear-beam model of a seven-story
7-story hotel during 11 earthquakes and earthquake damage detection. building caused by nonlinear waves during earthquake excitation. Structural
Structural Control and Health Monitoring 2008;15(1):90–116. Control and Health Monitoring 2009;16(4):460–82.
[4] Snieder R, S- afak E. Extracting the building response using interferometry: [31] Gičev V, Trifunac MD. Transient and permanent shear strains in a building
theory and applications to the Millikan Library in Pasadena, California. excited by strong earthquake pulses. Soil Dynamics and Earthquake Engi-
Bulletin of the Seismological Society of America 2006;96(2):586–98. neering 2009;29(10):1358–66.
[5] Todorovska MI. Seismic interferometry of a soil-structure interaction model [32] Gičev V, Trifunac MD. Predetermined earthquake damage scenarios (PEDS)
with coupled horizontal and rocking response. Bulletin of the Seismological for structural health monitoring, Structural Control and Health Monitoring,
Society of America 2009;99(2A):611–25. published online on 28 June 2011, http://dx.doi.org/10.1002/stc.470 in press.
[6] Todorovska MI. Soil-structure system identification of Millikan Library [33] Kawakami H, Oyunchimeg M. Normalized input–output minimization ana-
North–South response during four earthquakes (1970–2002): what caused lysis of wave propagation in buildings. Engineering Structures 2003;25(11):
the observed wandering of the system frequencies? Bulletin of the Seismo- 1429–42.
logical Society of America 2009;99(2A):626–35. [34] Oyunchimeg M, Kawakami H. A new method for propagation analysis of
[7] Trifunac MD, Todorovska MI, Manić MI. Bulajić BŒ. Variability of the fixed- earthquake waves in damaged buildings: evolutionary normalized input–
base and soil-structure system frequencies of a building – the case of Borik-2 output minimization (NIOM). Journal of Asian Architecture and Building
building. Structural Control and Health Monitoring 2010;17(2):120–51. Engineering 2003;2(1):9–16.
[8] Seber GAF, Wild CJ. Nonlinear regression. Hoboken, New Jersey: John Wiley & [35] Kawakami H, Oyunchimeg M. Wave propagation modeling analysis of
Sons, Inc.; 2003. earthquake records for buildings. Journal of Asian Architecture and Building
[9] Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage identification and Engineering 2004;3(1):33–40.
health monitoring of structural and mechanical systems from changes in [36] Trampert J, Cara M, Frogneux M. SH propagator matrix and Q S estimates from
their vibration characteristics: a literature review, Report LA-13070-MS, Los borehole- and surface-recorded earthquake data. Geophysical Journal Inter-
Alamos National Laboratory, Los Alamos, NM; 1996. national 1993;112:290–9.
[10] Chang PC, Flatau A, Liu SC. Review paper: health monitoring of civil [37] Snieder R, Sheiman J, Calvert R. Equivalence of the virtual-source method and
infrastructure. Structural Health monitoring 2003;2(3):257–67. wave-field deconvolution in seismic interferometry. Physical Review E:
[11] Fan W, Qiao P. Vibration-based damage identification methods: a review and Statistical, Nonlinear, and Soft Matter Physics 2006;73:066620.
comparative study. Structural Health Monitoring 2011;10(1):83–111. [38] Mehta K, Snieder R, Graizer V. Downhole receiver function: a case study.
[12] Luco JE, Trifunac MD, Wong HL. On the apparent change in the dynamic Bulletin of the Seismological Society of America 2007;97(5):1396–403.
behavior of a nine-story reinforced concrete building. Bulletin of the [39] Mehta K, Snieder R, Graizer V. Extraction of near-surface properties for a
Seismological Society of America 1987;77(6):1961–83. lossy layered medium using the propagator matrix. Geophysical Journal
[13] Wong HL, Trifunac MD, Luco JE. A comparison of soil-structure interaction International 2007;169:271–80.
calculations with results of full-scale forced vibration tests. Soil Dynamics [40] Assimaki D, Jamison J, Liu PC. Attenuation and velocity structure for site
response analyses via downhole seismogram inversion. Pure and Applied
and Earthquake Engineering 1988;7(1):22–31.
[14] Luco JE, Trifunac MD, Wong HL. Isolation of soil-structure interaction effects Geophysics 2006;163:81–118.
[41] Filippo Broggini, Snieder Roel. Connection of scattering principles: a visual
by full-scale forced vibration tests. Earthquake Engineering & Structural
and mathematical tour. European Journal of Physics 2012;33:593–613.
Dynamics 1988;16:1–21.
[42] Nakata N, Snieder N. Nearsurface weakening in Japan after the 2011
[15] Todorovska MI. Earthquake damage detection in buildings and early warning
TohokuOki earthquake. Geophysical Research Letters 2011;38:L17302.
based on wave travel times, Proc. 2009 NSF engineering research and
[43] Nakata N, Snieder N. Estimating near-surface shear wave velocities in Japan
innovation conference, Honolulu, Hawaii, June 22–25, 2009, Grant #
by applying seismic interferometry to KiK-net data. Journal of Geophysical
CMMI-0800399, pp.8, National Science Foundation, Washington, DC.
Research 2012;117:B01308.
[16] Todorovska MI, Trifunac MD. Earthquake damage detection in structures and
[44] Pech A, Sánchez-Sesma FJ, Snieder R, Ignacio-Caballero F, Rodrı́guez-Castel-
early warning, Proc. 14th World Conference on Earthquake Engineering,
lanos A, Ortı́z-Alemán JC. Estimate of shear wave velocity, and its time-lapse
October 12–17, 2008, Beijing, China, Paper S05-03-010, pp. 10.
change, from seismic data recorded at the SMNH01 station of KiK-net using
[17] Todorovska MI. Earthquake damage: detection and early warning in man-
seismic interferometry. Soil Dynamics and Earthquake Engineering 2012;39:
made structures, Chapter in ‘‘Encyclopedia of Complexity and System
128–37.
Science’’, RA Meyers (Ed.), Part 5: Complexity in Earthquakes, Tsunamis,
[45] Picozzi M, Parolai S, Mucciarelli M, Milkereit C, Bindi D, Ditommaso R, et al.
and Volcanoes, and Forecasting and Early Warning of their Hazards, W H K
Interferometric analysis of strong ground motion for structural health
Lee (Section Ed.), Springer 2009, 2371-2395; http://dx.doi.org/10.1007/ monitoring; the example of the L’Aquila, Italy, seismic sequence of 2009.
978-0-387-30440-3_146. Bulletin of the Seismological Society of America 2011;101(2):635–51.
[18] Kanai K. Some new problems of seismic vibrations of a structure, Proc. Third [46] Picozzi M. An attempt of real-time structural response assessment by an
World Conf. Earthq. Eng., Auckland and Wellington, New Zealand, January interferometric approach: a tailor-made earthquake early warning for build-
22–February 1, 1965, pp. II-260 to II-275. ings. Soil Dynamics and Earthquake Engineering 2012;38:109–18.
[19] Todorovska MI, Trifunac MD. Antiplane earthquake waves in long structures. [47] Rahmani M, Dehaghani ME, Todorovska MI (2012). Comparative earthquake
Journal of Engineering Mechanics (ASCE) 1989;115(12):2687–708. damage detection in a full scale 12-story RC building using wave travel time,
[20] Todorovska MI, Lee VW. Seismic waves in buildings with shear walls time-frequency and novelties analyses, in preparation.
or central core. Journal of Engineering Mechanics (ASCE) 1989;115(12): [48] Todorovska MI, Rahmani M (2012). Recent advances in wave travel time
2669–86. based methodology for structural health monitoring and early earthquake
[21] Todorovska MI, Trifunac MD. A note on the propagation of earthquake waves damage detection in buildings, The 15th World Conf. on Earthqu. Eng.
in buildings with soft first floor. Journal of Engineering Mechanics (ASCE) (15WCEE), Sept. 24–28, 2012, Lisbon, Portugal, pp. 10.
1990;116(4):892–900. [49] Clinton JF, Bradford SK, Heaton TH, Favela J. The observed wander of the
[22] S- afak E. Detection of seismic damage in multi-story buildings by using wave natural frequencies in a structure. Bulletin of the Seismological Society of
propagation analysis. Proc. Sixth U.S. National Conf. Earthq. Eng. 1998; EERI, America 2006;96(1):237–57.
Oakland, CA, Paper No. 171, pp. 12. [50] Reinoso E, Miranda E. Estimation of floor acceleration demands in high-rise
[23] S- afak E. Wave propagation formulation of seismic response of multi-story buildings during earthquakes. Structural Design of Tall and Special Buildings
buildings. Journal of Structural Engineering (ASCE) 1999;125(4):426–37. 2005;14:107–30.
[24] Todorovska MI, Ivanovic SS, Trifunac MD. Wave propagation in a seven-story [51] Thompson WT. Transmission of elastic waves through a stratified solid.
reinforced concrete building, Part II: Observed wavenumbers. Soil Dynamics Journal of Applied Physics 1950;21:89–93.
and Earthquake Engineering 2001;21(3):224–36. [52] Haskell NA. The dispersion of surface waves in multilayered media. Bulletin
[25] Ivanović SS, Trifunac MD, Todorovska MI. On identification of damage in of the Seismological Society of America 1953;43:17–34.
structures via wave travel times. In: Erdik M, Celebi M, Mihailov V, Apaydin N, [53] Gilbert F, Backus GE. Propagator matrices in elastic wave and vibration
editors. Proc. NATO advanced research workshop on strong-motion instrumenta- problems. Geophysics 1966;2:326–32 XXXI.
tion for civil engineering structures. Istanbul, Turkey: Kluwer Academic Publish- [54] Aki K, Richards P. Quantitative seismology. Freeman and Company; 2002.
ers; 2001. p. 447–68 June 2–5, 1999. [55] Trifunac MD. Zero baseline correction of strong-motion accelerograms.
[26] Gičev V. Investigation of soil-flexible foundation-structure interaction for Bulletin of the Seismological Society of America 1971;61(5):1201–11.
incident plane SH waves. Ph.D. Dissertation, Dept. Civil Eng., Univ. of South- [56] Trifunac MD. A note on correction of strong-motion accelerograms for
ern California, Los Angeles, CA., 2005. instrument response. Bulletin of the Seismological Society of America
[27] Gičev V, Trifunac MD. Permanent deformations and strains in a shear 1972;62(1):401–9.
building excited by a strong motion pulse. Soil Dynamics and Earthquake [57] Trifunac MD, Todorovska MI. Evolution of accelerographs, data processing,
Engineering 2007;27(8):774–92. strong motion arrays and amplitude and spatial resolution in recording
174 M. Rahmani, M.I. Todorovska / Soil Dynamics and Earthquake Engineering 47 (2013) 157–174

strong earthquake motion. Soil Dynamics and Earthquake Engineering [63] Foutch DA, Luco JE, Trifunac MD, Udwadia FE. Full scale three-dimensional
2001;21(6):537–55. tests of structural deformations during forced excitation of a nine-story
[58] Gabor D. Theory of Communications. Journal of the Institute of Electrical reinforced concrete building, Proc. U.S. National Conf. on Earthqu. Eng., Ann
Engineers—Part III: Radio and Communication Engineering 1946;93(26): Arbor, Michigan, 1975:206–215.
429–57. [64] Naeim F, Lee H, Hagie S, Bhatia H, Alimoradi A, Miranda E. Three-dimensional
[59] Gabor D. Communication theory and physics, institute of radio engineers. analysis, real-time visualization, and automated post-earthquake damage
Professional Group on Information Theory 1953;1(1):48–59. assessment of buildings. Structural Design of Tall and Special Buildings
[60] Shannon CE. Communication in the presence of noise. Proceedings of the 2006;15:105–38.
Institute of Radio Engineers 1949;37(1):10–21. [65] Todorovska MI, Trifunac MD. Earthquake damage detection in the Imperial
[61] Jennings PC, Kuroiwa JH. Vibration and soil-structure interaction tests of a County Services Building II: analysis of novelties via wavelets. Structural
nine-story reinforced concrete building. Bulletin of the Seismological Society Control and Health Monitoring 2010;17:895–917.
of America 1968;58(3):891–916. [66] Rahmani M, Todorovska MI (2012). 1D System identification of a 54-story
[62] Udwadia FE, Trifunac MD. Time and amplitude dependent response of steel frame building by seismic interferometry, Earthquake Engineering &
structures. Earthquake Engineering & Structural Dynamics 1974;2:359–78. Structural Dynamics, submitted for publication.

You might also like