TimeFrequency Analysis of Small Frequency

Variations in Civil Engineering Structures

Under Weak and Strong Motions Using a
Reassignment Method

Clotaire Michel1 and Philippe Gueguen2,*

1
Applied Computing and Mechanics Laboratory (IMAC), Ecole Polytechnique
Federale de Lausanne, Lausanne, Switzerland
2
Laboratoire de Geophysique Interne et Tectonophysique, Universite Joseph
Fourier, CNRS, LCPC, France

The analysis of strong and weak motion recordings in existing structures is critical for
understanding both the damaging process during earthquakes and their structural behavior.
The timefrequency representation is one of the existing methods to get information on the
frequency variations in buildings, which may indicate either damage or degradation. In this
article, we use standard timefrequency methods with amplitude normalization and reas-
signment method in order to observe smaller variations in earthquake recordings in build-
ings. The method used in this article is first validated on strong recordings from the R.
Millikan Library (Pasadena, California) and later applied to weak earthquakes recorded in
the Grenoble City Hall (France). This method detects torsion activation and very small
frequency variations under weak motion. We show that these variations are probably due
to variations of the input motion and cannot be attributed to variations in the parameters of
the structural system.

Keywords wignerVille distribution  time-varying stiffness  earthquake engineering 

seismic risk

1 Introduction Recent studies that focused on earthquake engi-

Since Omori [1], certainly the first author who
related the variations of the fundamental frequency
of buildings with damage and structural retrofit-
ting, a large scientific community has dealt
with the building frequency monitoring for struc-
tural and earthquake engineering (e.g., [29]).
neering applications (e.g., [1013]) have shown
how these variations can be monitored with
enhanced methods based on timefrequency repre-
sentation or linear modal analysis. They related the
permanent and transient decrease of the frequency
value during ground shaking to the structural integ-
rity of the building. The damaging process in

The Author(s), 2010. Reprints and permissions:

http://www.sagepub.co.uk/journalsPermissions.nav
*Author to whom correspondence should be addressed. Vol 9(2): 15913
E-mail: philippe.gueguen@obs.ujf-grenoble.fr [1475-9217 (201003) 9:2;15913 10.1177/1475921709352146]

159
160 Structural Health Monitoring 9(2)
buildings during earthquakes produces a perma-
nent loss of structural stiffness and thus a permanent
decrease of the fundamental frequency. Laboratory
dynamic or pseudo-dynamic tests also showed such
a frequency decrease with increasing damage [14].
For this reason, monitoring the frequency of build-
ings may be useful to detect the damage after earth-
quakes, as recently shown in practice by Dunand et
al. [15] after the Boumerdes, Algeria (May 21, 2003)
earthquake. After evaluating the fundamental fre-
quencies for undamaged and damaged buildings
using ambient vibrations, they found less than
20% frequency shift for almost all buildings classi-
fied in slight damage (green classification, i.e.,
people could return to buildings) and less than
40% frequency shift for buildings classified in mod-
erate damage (orange classification, i.e., buildings
could be retrofitted).
Moreover, using the strong motion data col-
lected in buildings, Clinton et al. [11] and
Dunand et al. [12] showed that transient frequency
variations due to the opening and closing process
of pre-existing cracks in the structure occurred
during shaking.
Clinton et al. [11] and Todorovska and Al
Rjoub [16] recently showed that, for long-term
monitoring of buildings, small frequency variations
under weak motion (ambient vibrations) could also
be due to the atmospheric conditions. For exam-
ple, rains can alter the elastic property of soils, and
thus change the response of soil-structure system.
These references show that small variations of fun-
damental frequencies observed in buildings using
ambient vibrations may not be always related to
damaging process but may be related to the
dynamic response of the system composed of soil,
structure, and foundation. Since the monitoring
of building frequency is certainly the easiest way
for building behavior assessment and health
monitoring, the physical meaning of instantaneous
frequency variation is therefore a crucial point that
must be explored in depth. A large set of methods
using the timefrequency representation exists for
monitoring building frequency variations during
earthquakes [1719]. While the decrease is well
and easily observed during the strongest motions,
because of the large decrease in fundamental fre-
quency, the analysis of the smallest frequency var-
iations under strong and weak motions must be
sufficiently precise in order to understand their
physics. In this article, a normalized distribution
is proposed in order to focus on the frequency
variations. Based on the WignerVille distribution
[20,21], this method combines the advanced time
frequency distribution with the reassignment
method [22,23] that enhances the precision of the
energy distribution, well adapted to reveal the
lowest and transitory variations. Based on this
technique, the frequency variations under strong
and weak motions are investigated for two build-
ings in order to analyze the origin of the frequency
shifts observed in buildings under weak motions.
This method is first validated on a well-known
strong motion recorded in the Millikan Library
building (California) and compared to classical
analyzes already published by several authors
(e.g., [11]). It is then applied to weak motions pro-
duced by moderate earthquakes recorded in the
City Hall building of Grenoble (France). The
small variations of the building frequency are stud-
ied and compared to the building behavior
extracted from ambient vibrations. The relevancy
and limits of the observed variations are finally
discussed.
2 The Reassigned and Smoothed
Pseudo WignerVille Distribution
Considering the signal to be analyzed x(t), let
us define xa(t) the analytical signal of x(t) by:
^
xa t xt jxt, 1
^
where xt, the Hilbert transform of x(t), is defined
by:
Z
1 1 x
^
xt d: 2
 1 t  
xa(t) can be rewritten in its polar form:
xa t Et:ejt , 3
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations
with E(t) and u(t) representing the envelope and
the instantaneous phase of the signal, respectively.
The analytical signal xa(t) is generally used instead
of the real signal x(t) for all timefrequency meth-
ods because its Fourier Transform is single-sided,
which is more convenient for computations [17].
During earthquakes, the large amplitudes mask
the information contained in the signal before
and after the strong motion when a classical
timefrequency representation is used. For this
reason, we make the signal stationary by dividing
it by its envelope. This step may distort the phase
of the signal but the results given in this article
show that this effect is minor in this application.
Bradford [18] suggested normalizing each
window of the spectrogram by the maximum in
the spectrum. The method employed here can be
used irrespective of the timefrequency representa-
tion. The studied signal s(t) is given as:
st ejt : 4
Numerous applications of timefrequency meth-
ods have been published, for structural and earth-
quake engineering applications (e.g., [2426,18,19]).
Neild et al. [17] reviewed these methods for structural
vibration analysis. They classified the methods into
three categories: the instantaneous frequency-based
methods, the windowed methods, and the bilinear
distributions-based methods, simultaneous in time
and frequency.
Since the study-system has multiple modes, the
instantaneous frequency could be used only after
the Empirical Mode Decomposition as proposed
by Huang et al. [27] but it is not developed here.
The windowed methods are the most widely used.
The well-known spectrogram, for example, consists
in computing Fourier transforms on short time
windows. The use of such windowed methods
needs a compromise between frequency and time
precision. The length of the window also strongly
influences the smoothing of the results. In order to
avoid this trade-off between frequency and time
precision, timefrequency distributions are used.
They are also called energy distributions because
161
they distribute the energy of the signal in the time
frequency space [23]. Most of them use the Cohens
class distribution [23,17], which keeps the total
energy, the instantaneous energy, and the spectral
density energy of the signal.
The simplest distribution of Cohens class is the
WignerVille distribution (PWV) and is given by:
Z 1 
  
PWV t,f ej2f xa t xa t  d: 5
1 2 2
The integral in time is not only evaluated on
a short-time window like the spectrogram but
also from 1 to +1, so that there is no
limit on the frequency precision. When we com-
pute the product xa t =2xa t  =2, addi-
tional terms are included in the PWV
distribution producing interferences. The use of
the analytic signal limits them because its
Fourier Transform is single-sided and has there-
fore less components. The smoothed
pseudo-WignerVille (SPWV) distribution PspWV
is usually computed. This corresponds to a wind-
owed version in time and frequency of the PWV.
It is equivalent to a time and frequency smooth-
ing [23] and is adapted to finite signals in time.
The smoothing also attenuates interferences. The
PspWV distribution is expressed as follows:
Z1 Z1
PspWV t, f; g, h h gu  ej2f
1
 1 , 6
  
 xa u xa u  dud
2 2
with g and h representing regular windows (for
example, Hamming windows) corresponding to
time and frequency smoothing, respectively. The
energy band of this distribution is, however, gen-
erally thick, making the small frequency variations
difficult to distinguish.
In order to remove these difficulties, Kodera
et al. [22] and Auger and Flandrin [23] developed
the reassignment method. They assume that there
are no physical reasons for the energy distribution
to be symmetric at the vicinity of each time
frequency point. Therefore, they suggest the
162 Structural Health Monitoring 9(2)
reassigning of the value of the local energy distri-
bution to the center of gravity of the domain
around each timefrequency point. The reassigned
distribution is the sum of the distributions
obtained at each point. The mathematical imple-
mentation of the reassignment proposed by
Kodera et al. [22] is developed in Auger and
Flandrin [23] for several classical timefrequency
methods such as the SPWV distribution or the
spectrogram. In addition, reassignment method
reduces the interferences in the WV distribution.
The reassigned SPWV is computed as follows [23]:
Z1
 
PrspWV t0 ,f0 ; g,h PspWV t,f; g,h t0  t^t,f 
1
 
 f0  f^ t,f dtdf,
with
8 has been instrumented since 1968 with two 3C
> PspWV t, f; t  g, h
> ^
<t t, f t  2PspWV t, f; g, h
>
 :
>
> PspWV t, f; t  g, dh
> ^
:f t, f f j dt
2PspWV t, f; g, h
Few applications of the reassignment tech-
nique to natural signals exist in the scientific liter-
ature: for example, Kodera et al. [22] and Pedersen
et al. [28] employed this method for the study of
geomagnetic pulsations and the estimation of dis-
persion curves in geophysics, respectively. For
applications to earthquakes recorded in structures,
the time history of the building motion may have
an initial and fast frequency drop followed by a
more stable part (e.g., [11,19]). The short-term
decrease of the frequency is hardly captured by
classical methods integrating on finite windows
such as spectrogram or even SPWV. This leakage
phenomenon may be corrected by the reassignment
method. The major advantages of this method are
therefore the slenderness of the energy band, the
removal of the majority of interferences, and a
better capture of fast frequency changes [22] that
is well adapted to the building vibrations analyzis.
Once the timefrequency distribution has been cal-
culated, it is also possible to extract the modal fre-
quency values by choosing the maxima in the
required frequency range.
3 Application to Strong Motion
Recorded in the Millikan Library
Building
The Millikan Library on the Caltech campus
(Pasadena, California) is certainly one of the first
instrumented and extensively studied buildings in
the world. It is a 9-story reinforced-concrete (RC)
structure built in 1967 based on RC frames and
some RC shear walls. A harmonic shaker is perma-
nently installed at the top of the structure since the
early seventies to perform forced vibration tests
7 [29]. Thanks to the Californian Strong motion
Instrumentation Program (CSMIP), the building
accelerometers at the top and at the basement.
Immediately after the construction of the building,
forced vibration tests were performed in 1967 and
the measured resonance frequencies under such
weak motion were 1.45 Hz and approximately
1.95 Hz in the EW and NS directions, respec-
tively [29,11].
On February 9, 1971, the San Fernando earth-
quake (local magnitude ML = 6.6 at 31 km), with a
peak ground acceleration (PGA) of 200 cm/s2
(basement, N direction), produced peak top
accelerations of 306 and 341 cm/s2 in the EW
and NS directions, respectively, one of the
greatest recorded in the structure [11]. This earth-
quake induced cracking and spalling of the con-
crete slabs on the ground floor and horizontal
cracks in the core shear walls between the
basement and the second story in the NS
direction [30].
Figure 1 compares the spectrogram distribu-
tion, its reassigned version, the SPWV distribution,
and its reassigned version for normalized record-
ings of this earthquake at the top of the building
in the E direction. In all, 2048 and 464 points
in frequency and time were used, respectively.
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations 163

1.6 5 Log-
1.4 scale
1.2 0
1
0.8 5
1.6
1.4 5
1.2
0
1
0.8 5
1.6 5
Frequency (Hz)

1.4
1.2 0
1
0.8 5
1.6
1.4 5
Transient drop Permanent drop
1.2
0
1
0.8 5
Acc. (m/s2 )

5
0
5
0 10 20 30 40 50 60 70 80 90 100
Time (s)

Figure 1 Timefrequency distributions (from top to bottom: spectrogram, reassigned spectrogram, smoothed
pseudo-WignerVille and reassigned smoothed pseudo-WignerVille) of February 09, 1971 ML = 6.6 San Fernando
Earthquake normalized recording at the roof of the Millikan Library on CalTech campus (California) in E direction. The
black and grey solid lines are the pre- and post-seismic frequencies, respectively [11]. The grey dashed line is the
co-seismic frequency.

The frequency smoothing was done by 512-point
Hamming windows, and the time smoothing for
the SPWV distribution was done by a 46-point
Hamming window. Most of the observations dis-
played in Figure 1 have been reported by Foutch
and Jennings [30] and Clinton et al. [11]. We
observe a rapid decrease in the first frequency
during the first 15 s. The pre-seismic frequency is
greater than 1.3 Hz, but due to the very short
pre-event time window, this frequency cannot be
clearly seen in this figure. The co-seismic fre-
quency, i.e., the minimum value reached during
the earthquake, is 0.94 Hz, i.e., a transitory drop
of 35% with respect to the pre-seismic frequency
value (1.45 Hz) found in Clinton et al. [11].
These authors estimated a co-seismic frequency
of 1 Hz calculated on the whole signal, therefore
a bit larger than the minimum value found here.
This co-seismic frequency occurs 5 s after the peak
acceleration. Once this value is reached, the fre-
quency increases gradually up to the post-seismic
frequency (frequency at the end of the recording)
and equal to 1.15 Hz. This post-seismic frequency
is close to the value obtained after the San
Fernando earthquake by forced and weak vibra-
tion tests performed in 1974 (1.21 Hz) [11]. This
means that during the San Fernando event the
greatest part of the transient stiffness drop is
164 Structural Health Monitoring 9(2)
recovered in 50 s. However, Figure 1 shows that
the recovery is not continuous. For example, a
small decrease of the frequency occurs after
t=70 s, whereas the motion amplitude has
decreased. This small frequency drop is therefore
more likely due to the linear response to the
ground motion than to stiffness reduction. Thus,
during the nonlinear process of stiffness recovery,
the response to the complex input also plays an
important role.
The application of any timefrequency method
to real earthquake recordings in structures faces
some problems. First, even if the proposed normal-
ization method allows to follow the main modes on
the whole signal duration, some modes may not be
activated in some parts of the signal (no energy).
Second, we can observe strange spider-like patterns
in the reassigned SPWV that partly appear as energy
patches in the spectrogram. They are therefore not
mainly due to the WV interferences but due to the
complexity of the input wave field and resonance
frequency beating, e.g., at 45 s. Thus, timefre-
quency distributions of the building vibrations
should be carefully interpreted.
The spectrogram gives the poorest results
because of fuzzy patterns and discontinuities
when variations are too rapid (e.g., at 60 s).
However, its reassigned version gives a more pre-
cise picture of the frequency variation during shak-
ing. The SPWV performs better than the
spectrogram with a clearer, smoother, and thinner
picture but its reassignment is even better. The
reassignment procedure removes most of the inter-
ferences seen with SPWV. This comparison shows
that the reassignment method describes and fol-
lows in more detail the frequency variations and
identifies fast and small frequency drops. The reas-
signment of the spectrogram and the SPWV per-
form equally well in this example but the
reassigned SPWV is preferred in order to avoid
the windowing. Since the proposed method is suc-
cessful for large frequency variations, it is applied
in the following sections to weak earthquakes in
order to assess their capacity to monitor and to
understand small frequency variations.
4 Application to Weak Motions in the
Grenoble City Hall Building
Since November 2004, the City Hall building of
Grenoble (France) has been surveyed by the French
Accelerometric Network (RAP) [31]. The building is
a 13-story RC-structure built in 1967 (Figure 2(a)).
The tower has a 44 m by 13 m plan section and rises
52 m above the ground. Two inner cores, consisting
of RC shear walls, enclose the stairwells and lift
shafts and are located at two opposite sides of the
building. The structural system combines the shear
walls with RC frames and longitudinal beams bear-
ing the RC floors. The accelerometric stations mon-
itoring the building are composed of 3C Episensors
(Kinemetrics) plugged to MiniTitan (Agecodagis)
digitizers, located at the top and at the bottom of
the building. A complete description of the building
and of the network may be found in Michel and
Gueguen [32] and Pequegnat et al. [31].
In addition, ambient vibration tests were per-
formed in the building in order to extract the
full-scale behavior of the structure [32].
Simultaneous recordings of ambient vibrations
were performed in 35 different points of the struc-
ture (Figure 2(b)) and the modal parameters were
extracted using the frequency domain decomposi-
tion (FDD) method [33]. The method comes to cal-
culate the Fourier transforms of the correlation
matrices for each dataset (power spectral density
(PSD) matrices) and then to perform a singular
value decomposition of the PSD matrices at each
frequency. The peak frequencies of the first
singular values give the resonance frequencies
(Figure 2(c)) and the corresponding eigenvectors
give the modal shapes (Figure 2(d)). Only three
modes were accurately determined (Figure 2(c)
and (d)): the first longitudinal bending mode at
1.15 Hz, the first transverse bending mode at
1.22 Hz, and the first torsion mode at 1.44 Hz.
According to the building plans, the structure is
nearly symmetric. This is confirmed by the fact
that these three modes are fairly uncoupled and
show pure motions along the three main degrees
of freedom of the structure. The damping ratios
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations 165

(a) (c)

80
Y1 1.22 Hz
T1 1.45 Hz
60

Eigenvalues of PSD matrices (dB)

X1 1.16 Hz
40

20

0
(b) N OGH5

OGH4 20
OGH6
40 40

60
OGH2
80
OGH1 0.5 1 1.5 2 3 4 5 10
0 40
OGH3 Frequency (Hz)
15
0 0

(d)
50 50
50
40 40
Height (m)

Height (m)
40
Height (m)

30 30
30
20 20
20
10 10
10
44 0 44 0 44
0 13
13 0 13
0 0 0 0
0

Figure 2 (a) City Hall building of Grenoble; (b) location of permanent stations OGH and distribution of ambient vibration
recordings (points); (c) frequency domain decomposition [33] spectrum with main modes; (d) corresponding modal
shapes (after [32]).

under ambient vibrations are close to 1% for each distribution between the bending and the torsion
mode [32]. modes. For the three examples given in Figure 3,
We computed the reassigned SPWV distribu- only the bending mode (at 1.22 Hz) is activated
tion for three earthquakes among the largest during the time window corresponding to the
recorded at the OGH6 roof station in the trans- strongest amplitudes. During the earthquake
verse direction (Figure 3): the Vallorcine motion, the torsion mode (at 1.44 Hz) is not acti-
(ML = 4.9 at 127 km, PGA = 2.3 cm/s2), Lago Di vated, whereas in the pre- and post-event windows
Garda (ML = 5.5 at 340 km, PGA = 0.3 cm/s2), (ambient vibrations), torsion, and bending have
and Laffrey (ML = 3.1 at 15 km, PGA = 0.5 cm/s2) equivalent energy. This observation confirms the
earthquakes. This figure shows the energy prevalence of the first bending mode under
166 Structural Health Monitoring 9(2)

moderate seismic shaking. This phenomenon is 5 Discussion on the Frequency

likely due to the loading by a single remote Decrease Observed under Weak and
source that acts with the same phase on the Strong Motion
whole building foundation, whereas ambient
vibrations are due to multiple sources acting In case of strong motions, the timefrequency
with different phases on the foundation and there- distribution obtained using the reassigned SPWV
fore creating a moment. Nevertheless, some varia- method for the San Fernando earthquake recorded
tions in the building behavior exist under weak in the Millikan Library (Figure 1) clearly shows
motion such as these three study-cases. These var- variations of the resonance frequencies from the
iations are amplified by the fact that the City-Hall pre- to co-seismic values and from the co- to
of Grenoble is founded on a very deep 3D basin post-seismic values. These variations are consistent
introducing variability of seismic ground response with those estimated by other researchers [11,12].
even for weak motion [34]. In order to show the impact of stiffness
We observe also small and transient variations decrease on the recordings in the building, we
of the bending frequency (1.22 Hz), sometimes on used two linear single degree-of-freedom (SDOF)
small windows of ambient vibrations but mainly systems (5% damping), the first one having
during earthquakes (Figure 3). The greatest varia- dynamic characteriztics corresponding to the
tion, a decrease of about 10%, occurs during the post-event building (frequency 1.2 Hz) and the
Vallorcine earthquake that produced a peak accel- second to the co-seismic building (frequency
eration of 11 cm/s2 at the top. 1 Hz). We computed the synthetic time history at

1.6 Log- Log-

Frequency (Hz)

Frequency (Hz)

1.5 5 scale 1.5 5

scale

1.4 1.4
0 1.3 0
1.3
1.2 1.2
5 5
Acc. (m/s )

Acc. (m/s )
2

0.1 0.1
0 0
0.1 0.1
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
1.6 Log- 1.6 Log-
Frequency (Hz)

Frequency (Hz)

1.5 5 scale 1.5

scale
5
1.4 1.4
1.3 0 1.3 0
1.2 1.2
5 5
Acc. (m/s )

Acc. (m/s )
2

0.05 0.05
0 0
0.05 0.05
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
1.6 Log- 1.6 Log-
Frequency (Hz)

Frequency (Hz)

1.5 5 scale 1.5 5

scale

1.4 1.4
1.3 0 1.3 0
1.2 1.2
5 5
0 50 100 150 200 250 300 350
Acc. (m/s )

Acc. (m/s )
2

0.01 0.01
0 0
0.01 0.01
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Time (s) Time (s)

Figure 3 Timefrequency distributions (left: reassigned smoothed pseudo-WignerVille; right: reassigned spectro-
gram) of the normalized recordings in the transverse direction at OGH6 (top) station of Vallorcine (ML = 4.9 at 127 km,
PGA = 2.3 cm/s2 top), Lago Di Garda (ML = 5.5 at 340 km, PGA = 0.3 cm/s2, center), and Laffrey (ML = 3.1 at 15 km,
PGA = 0.5 cm/s2 bottom) earthquakes. The horizontal dashed grey lines correspond to the transverse (1.22 Hz) and
torsion (1.45 Hz) fundamental resonance frequencies obtained under ambient vibrations [32].
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations 167

the roof by convolution of the recorded time his- post-seismic phase, the 1.2 Hz SDOF fits the data
tory at the basement with both SDOF models better and it could explain the decrease at 70 s that
using the Duhamel integral. The 1 Hz SDOF may be due to the history of the seismic input
model reproduces the major part of the time his- motion. However, the greatest part of the varia-
tory as well as the frequency variations (Figure 4). tions, around 20%, can only be explained by non-
Only the part after 10 s of recording is described linearity in the structural stiffness due to damaging
because of a very short pre-event window for time effect.
frequency analyzis. In the initial phase correspond- Under strong motion, the observed drop of the
ing to the strongest motion (i.e., between 10 and building frequency is therefore mainly dominated
40 s), the 1 Hz SDOF reproduces the frequency by the permanent and transient variations of the
variations, but fails to explain the stiffness recov- structural stiffness, due to the opening and closing
ery. Between 65 and 90 s, i.e., corresponding to the process of cracks in concrete. This observation was
Frequency (Hz)

(a)
1.5 Log-
5
scale
0
1
5
Acc. (m/s2)

5
0
5
Frequency (Hz)

1.5 5

0
1
5
Acc. (m/s2)

5
0
5
Frequency (Hz)

1.5 5

0
1
5
Acc. (m/s2)

5
0
5
Frequency (Hz)

(b) 1.5

0 10 20 30 40 50 60 70 80 90
Time (s)

Figure 4 (a) Time history and timefrequency distribution (reassigned smoothed pseudo-WignerVille) of the recorded
motion of the Millikan Library during the San Fernando earthquake (top) and the synthetic obtained using two linear
single degree-of-freedom models having fundamental frequency equal to 1.2 Hz (center) and 1 Hz (bottom); (b) Maximal
amplitude of the timefrequency distributions displayed in (a) extracted from the recordings (black line) and the syn-
thetics (grey lines) corresponding to the 1 Hz single degree-of-freedom (solid grey) and the 1.2 Hz single
degree-of-freedom (dashed grey).
168 Structural Health Monitoring 9(2)

(a) Log-

Frequency (Hz)
5 scale
1.25

1.2 0
1.15
5
Acc. (m/s2)
0.1
0
0.1
Frequency (Hz)

5
1.25

1.2 0
1.15
5
Acc. (m/s2)

0.1
0
0.1
1.3
(b)
Frequency (Hz)

1.2

1.1
0 50 100 150 200 250 300 350
Time (s)

Figure 5 (a) Comparison between timefrequency distributions (reassigned smoothed pseudo-WignerVille) of the
normalized recording at OGH6 during Vallorcine Earthquake and synthetic signal computed using a linear single
degree-of-freedom based on modal parameters determined using ambient vibrations; (b) Maxima of the recording
(black line) and the synthetic (grey line) distributions. The horizontal dashed grey line stands for the frequency obtained
from ambient vibrations experiment.

firstly mentioned by Omori [1], and later concluded
by many studies (e.g., [35,36,10]) for other struc-
tures. The recovery between the co- and
post-seismic frequencies, maybe due to the
re-closing of cracks in the structure, shows that
the stiffness variation is partly transient. Some
others origins, such as soil-structure interaction
variations due to local perturbations of soil foun-
dation properties by atmospheric conditions, have
been recently suggested by Clinton et al. [11] and
Todorovska and Al Rjoub [16] on long time obser-
vations of continuous recordings. Nevertheless, the
low amplitude of the frequency drop observed by
these studies can be neglected in comparison with
those due to stiffness reduction under strong
motion.
Nevertheless, some small variations observed
after the strong motion can be clearly detected by
the reassigned SPWV distribution and may origi-
nate in different factors than structural nonlinear-
ity, such as the input time history.
Concerning the Grenoble City Hall building,
Figure 3 shows how the energy is distributed
between the bending and the torsion mode under
weak motion. As we did for the Millikan Library,
we developed a SDOF model for this building,
fixed using ambient vibration surveys done by
Michel and Gueguen [32], i.e., 1.22 Hz for the
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations
transverse fundamental frequency with 1% damp-
ing. The synthetic time history of the building
motion at the roof was also computed by convolu-
tion of the recorded time history at the basement
(OGH3, Figure 2(c)) with the SDOF model using
the Duhamel integral. This modelling assumes a
constant fundamental frequency value under
weak motion (linear elastic model). The recorded
and synthetic distributions using the reassigned
SPWV method are displayed in Figure 5, corre-
sponding to the strongest earthquake (Vallorcine
earthquake) showing a 10% frequency drop. The
maximum values of these distributions are
extracted (Figure 5(b)). Only points with a value
greater than one-fourth of the maximum value in
the timefrequency plane are considered. The
timefrequency distributions of the experimental
and synthetic time histories show similar frequency
variations (Figure 5), with a correlation coefficient
of 82%. The study gives similar results for other
earthquakes, not displayed in this paper. This
observation allows us to conclude that the
over-simplified and linear model of the building
explains most of the variations under weak
motion. These variations are therefore more
likely due to the input motion (at the basement)
rather than nonlinear stiffness variations in the
building. Moreover, during earthquakes, torsion
motion is much less activated than bending mode
as displayed in Figure 3. In fact, the building
behavior appears to alternate between these two
modes. One explanation could be that ambient
vibrations cause differential motions along the
building foundation produced by the multiple
sources and therefore activating the torsion
mode, whereas the earthquake may be considered
as one single input loading the whole building
foundation. This would imply that the building
eccentricity between the center of mass and the
center of rigidity is very low, which is attested by
the structure regularity and shown by Michel and
Gueguen [32].
Recently, Clinton et al. [11] observed the
changes of building frequency due to the water
content in soil on long time history of the building
169
motion recorded at the Milikan Library building.
However, this phenomena are very different from
the variation observed in our study and would not
be a relevant explanation. The multiple sources of
the seismic noise may explain the complexity of the
ground ambient wavefield producing small varia-
tions in amplitude and frequency.
6 Conclusions
Even if no timefrequency representation is per-
fect for systems with changing frequency content,
this paper shows that the reassigned SPWV follows
precisely the structural frequencies during earth-
quake and weak motions. The reassignment proce-
dure for SPWV and spectrogram distributions
significantly increases the resolution to analyze the
transitory variations of frequencies in buildings for
weak and strong ground motion. The method used
in this paper allows the analysis of earthquake
recordings from structures and from dynamic or
pseudo-dynamic laboratory tests to better relate fre-
quency variations to the amplitude of motion.
The timefrequency distributions of normal-
ized signals, and particularly the reassigned
SPWV method, give a better understanding of
the nonlinear evolution of the resonance frequency
of buildings during earthquake motions. In the
case of the Millikan Library, we reproduce the
well-known fast frequency drop that occurs just
after the peak acceleration until a minimum
value. This drop is followed by a slower increase
in frequency until the post-seismic frequency. This
frequency corresponds to those recorded under
ambient or forced vibrations after the earthquake.
In the case of the San Fernando earthquake
recorded in the Millikan Library, a transient
drop of 35% in E direction was observed in accor-
dance with previous studies whereas the permanent
drop was 16.6% [11]. We showed that even if
the input motion could explain slight variations
in this case, most of them are due to structural
nonlinearity. In case of strong shaking like
those produced by the San Fernando earthquake,
the timefrequency distribution is useful to obtain
170 Structural Health Monitoring 9(2)
a relevant interpretation of the resonance fre-
quency variations during earthquakes and later,
assess the building integrity after a damaging
earthquake.
Under weak motion, other effects such as var-
iations of the input signal may dominate the fre-
quency variations observed on recordings at the
City Hall of Grenoble. In this study, the timefre-
quency analysis shows also clearly the alternation
in time between the bending and the torsion modes
that may also be linked to the nature of the ambi-
ent vibrations. For that reason, the reassigned
SPWV distribution is also well adapted to weak
motion application in order to evaluate smaller fre-
quency variations under moderate earthquake and
ambient vibrations. As shown for the City Hall of
Grenoble, these variations should not be related to
soil-structure interaction effects or transitory
reduction of stiffness but to the input motion.
Acknowledgments
This work is supported by the Rhone-Alps regional auth-
orities through the Thematiques Prioritaires program (pro-
ject VULNERALP) and by the French Research National
Agency (ANR) through the RGCU program (project
ARVISE ANR-06-PRGCU-007-01). The Grenoble City
Hall data are provided by the French Accelerometric
Network (RAP: http://www-rap.obs.ujf-grenoble.fr) oper-
ated by the Laboratoire de Geophysique Interne et
Tectonophysique (LGIT). Data from Millikan Library are
provided by USGS array through the Cosmos web site.
References
1. Omori, F. (1922). The semi-destructive earthquake of
April 26, 1922. Seismological Notes (Imperial
Earthquake Investigation Committee), 3, 130.
2. Carder, D.S. (1936). Observed vibration of buildings.
Bulletin of the Seismological Society of America, 26,
245277.
3. Crawford, R. and Ward, H.S. (1964). Determination of
the natural periods of buildings. Bulletin of the
Seismological Society of America, 54, 17431756.
4. Trifunac, M. (1972). Comparison between ambient and
forced vibration experiments. Earthquake Engineering
and Structural Dynamics, 1, 133150.
5. Celebi, M. (1993). Dynamic characteriztics of five tall
buildings during strong and low-amplitude motions.
The Structural Design of Tall Buildings, 2, 115.
6. Peeters, B. (2000). System identification and damage
detection in civil engineering. PhD Dissertation,
Katholieke Universiteit Leuven, Belgium.
7. Carden, P.E. and Fanning, P. (2004). Vibration based
condition monitoring: A Review. Structural Health
Monitoring, 3, 355377.
8. Hans, S., Boutin, C., Ibraim, E. and Roussillon, P.
(2005). In Situ experiments and seismic analyzis of
existing buildings - Part I: Experimental investigations.
Earthquake Engineering and Structural Dynamics, 34,
15131529.
9. Michel, C., Gueguen, P. and Bard, P.-Y. (2008).
Dynamic parameters of structures extracted from ambi-
ent vibration measurements: An aid for the seismic vul-
nerability assessment of existing buildings in moderate
seismic hazard regions. Soil Dynamics and Earthquake
Engineering, 28, 593604.
10. Mucciarelli, M., Masi, A., Gallipoli, M.R., Harabaglia,
M.R.P., Vona, M., Ponzo, F. and Dolce, M. (2004).
Analysis of RC building dynamic response and
soil-building resonance based on data recorded during
a damaging earthquake (Molise, Italy, 2002). Bulletin of
the Seismological Society of America, 94, 19431953.
11. Clinton, J.F., Bradford, S.C., Heaton, T.H. and Favela,
J. (2006). The observed wander of the natural frequen-
cies in a structure. Bulletin of the Seismological Society
of America, 96, 237257.
12. Dunand, F., Gueguen, P., Bard, P.Y., Rodgers, J. and
Celebi, M. (2006). Comparison of the dynamic para-
meters extracted from weak, moderate and strong
motion recorded in buildings. In: Proceedings of the
First European Conference on Earthquake Engineering
and Seismology, Geneva, Switzerland, Paper 1021.
13. Pai, F.P., Huang, L., Hu, J. and Langewish, D.R.
(2008). Timefrequency method for non-linear system
identification and damage detection. Structural Health
Monitoring, 7, 103127.
14. Calvi, G.M., Pinho, R. and Crowley, H. (2006).
State-of-the-knowledge on the period elongation of
RC buildings during strong ground shaking. In:
Proceedings of the First European Conference of
Earthquake Engineering and Seismology, Geneva,
Switzerland, Paper 1535.
15. Dunand, F., Y. Ait Meziane, P. Gueguen, J.-L.
Chatelain, B. Guillier, R. Ben Salem, M. Hadid, M.
 Michel & Gueguen TimeFrequency Analysis of Small Frequency Variations
Hellel, A. Kiboua, N. Laouami, D. Machane, N.
Mezouer, A. Nour, E.H. Oubaiche et A. Remas. 2004.
Utilisation du bruit de fond pour lanalyse des dommages
des batiments de Boumerdes suite au seisme du 21 mai
2003, Mem Serv. Geol. Alg., 12, 177191.
16. Todorovska, M.I. and Al Rjoub, Y. (2006). Effects of
rainfall on soil-structure system frequency: example
based on poroelasticity and a comparison with
full-scale measurements. Soil Dynamics and
Earthquake Engineering, 26, 708717.
17. Neild, S.A., McFadden, P.D. and Williams, M.S. (2003).
A review of timefrequency methods for structural
vibration analysis. Engineering Structures, 25, 713728.
18. Bradford S.C. (2006). Timefrequency analyzis of sys-
tems with changing dynamic properties. Ph.D. Thesis,
California Institute of Technology, California.
19. Todorovska, M.I. and Trifunac, M.D. (2007).
Earthquake damage detection in the Imperial County
Services Building I: The data and timefrequency ana-
lyzis. Soil Dynamics and Earthquake Engineering, 27,
564576.
20. Wigner, E. (1932). On the quantum correction for ther-
modynamic equilibrium. Physical Review, 40(5), 749759.
21. Ville, J. (1948). Theorie et applications de la notion de
signal analytique. Cables & Transmissions, 2A, 6174.
22. Kodera, K., De Villedary, C. and Gendrin, R. (1976).
A new method for the numerical analysis of
non-stationary signals. Physics of the Earth and
Planetary Interiors, 12, 142150.
23. Auger, F. and Flandrin, P. (1995). Improving the read-
ability of timefrequency and time-scale representations
by the reassignment method. I.E.E.E Transactions on
Signal Processing, 43, 10681089.
24. Trifunac, M.D., Ivanovic, S.S. and Todorovska, M.I.
(2001). Apparent periods of a building. II: Timefre-
quency analyzis. Journal of Structural Engineering,
127, 527537.
25. Hans, S., Ibraim, E., Pernot, S., Boutin, C. and Lamarque,
C.-H. (2000). Damping identification in multi-degree-of-
freedom system via a wavelet-logarithmic decrement - Part
2: Study of a civil engineering building. Journal of Sound
and Vibration, 235, 375403.
171
26. Argoul, P. and Le, T.P. (2004). Wavelet analyzis of
transient signals in civil engineering. Novel Approaches
in Civil Engineering, 14, 311318.
27. Huang, N.E., Long, S.R. and Shen, Z. (1996). The
mechanism for frequency downshift in nonlinear wave
evolution. Advances in Applied Mechanics, 32, 59111.
28. Pedersen, H., Mars, J. and Amblard, P.O. (2003).
Improving surface-wave group velocity measurements
by energy reassignment. Geophysics, 68, 677684.
29. Kuroiwa, J.H. (1967). Vibration test of a multistory
building. Technical Report, Earthquake Engineering
Research Laboratory, California Institute of
Technology, Pasadena, California.
30. Foutch, D.A. and Jennings, P.C. (1978). Foundation
response of a nine-story reinforced concrete building.
Bulletin of the Seismological Society of America, 68,
219229.
31. Pequegnat, C., Gueguen, P., Hatzfeld, D. and Langlais,
M. (2008). The French Accelerometric Network (RAP)
and National Data Centre (RAP-NDC). Seismological
Research Letters, 79, 7989.
32. Michel C., Gueguen P., El Arem S., Mazars J. and
Kotronis P. 2009. Full Scale Dynamic response of a
RC building under weak seismic motions using earth-
quake recordings, ambient vibrations and modelling,
Earthq. Engng and Struct. Dyn., doi 10.1002/eqe.948.
33. Brincker, R., Zhang, L. and Andersen, P. (2001).
Modal identification of output only systems using fre-
quency domain decomposition. Smart Materials and
Structures, 10, 441445.
34. Gueguen, P., Cornou, C., Garambois, S. and Banton, J.
(2007). On the limitation of the H/V spectral ratio using
seismic noise as an exploration tool: application to the
Grenoble valley, a small apex ratio basin. Pure and
Applied Geophysics, 164, 115134.
35. Stubbs, I.R., and McLamore, V.R. (1973). The ambient
vibration survey. In: Proceedings of the Fifth World
Conference on Earthquake Engineering, Rome.
36. Udwadia, F.E. and Trifunac, M.D. (1973). Comparison
of earthquake and microtremor ground motions in El
Centro, California. Bulletin of the Seismological Society
of America, 63, 12271253.