You are on page 1of 32

Journal of Earthquake Engineering

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/ueqe20

Wavelet-Based Analysis for Detection of Isolation


Bearing Malfunction in a Continuous Multi-Span
Girder Bridge

Dionysius M. Siringoringo & Yozo Fujino

To cite this article: Dionysius M. Siringoringo & Yozo Fujino (2021): Wavelet-Based Analysis for
Detection of Isolation Bearing Malfunction in a Continuous Multi-Span Girder Bridge, Journal of
Earthquake Engineering, DOI: 10.1080/13632469.2020.1868363

To link to this article: https://doi.org/10.1080/13632469.2020.1868363

Published online: 15 Feb 2021.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=ueqe20
JOURNAL OF EARTHQUAKE ENGINEERING
https://doi.org/10.1080/13632469.2020.1868363

Wavelet-Based Analysis for Detection of Isolation Bearing


Malfunction in a Continuous Multi-Span Girder Bridge
Dionysius M. Siringoringo and Yozo Fujino
Institute of Advanced Sciences, Yokohama National University, Kanagawa, Japan

ABSTRACT ARTICLE HISTORY


This paper describes techniques for detecting bearing malfunction directly Received 21 July 2020
from the seismic response of the bridge using wavelet transform time- Accepted 20 December 2020
varying identification. Instantaneous frequency of continuous wavelet and KEYWORDS
detail components of discrete wavelet transform are utilized as structural Wavelet analysis; seismically
features, and they characterize isolation bearing condition via statistical isolated bridge; structural
clustering technique. Accuracy and efficacy of the techniques were verified assessment; bearing
in numerical simulations using analytical and finite element models. The malfunction; seismic
techniques were implemented on seismic records from long-term monitor­ monitoring
ing of multi-span continuous-girder isolated bridge. Results demonstrate
that wavelet-based features can effectively characterize the condition of
the isolation bearing directly from seismic responses of girder and piers.

1. Introduction
Seismically isolated bridges are popular in many countries with high risk of earthquakes. In Japan, the use
of seismic isolation for highway bridges has increased significantly after the 1995 Kobe earthquake.
Seismic isolation system increases natural periods of the bridges and reduces the effect of inertia force
from superstructures on the substructures. The use of seismic isolation system on multi-span continuous
bridges has increased significantly in Japan after the 1995 Kobe earthquake. Notable examples are the
Ohito Viaduct, a 29-span continuous prestressed concrete girder bridge with a total length of 725 m and
Tenryu-River Bridge, a 23-span continuous prestressed concrete girder bridge with a total length of
1585 m. The isolation system improves the seismic performance of the bridges and makes multi-span
continuation of superstructures easier because the horizontal force which acts on bridge piers by
temperature change can be reduced using elastic seismic bearings (Unjoh 2014).
Bridges with seismic isolation and elastomeric bearings generally performed well under extreme
ground motions. During the 2011 Tohoku earthquake, for example, only very few of them were
damaged by the earthquake (Kawashima and Buckle 2013; Kitahara, Kajita, and Kitane 2012). Past
observation and monitoring experiences, however, revealed that the behavior of seismically isolated
bridges is significantly influenced by structural details of isolation bearing system such as metallic
bearings, side stoppers and bearing keeper plates (Chaudhary, Abe, and Fujino 2002a; Fujino,
Siringoringo, and Abe 2016). In some cases, such as the Onneto bridge, steel plate girders became
locked with the transverse side stopper on several bearings, and as a result, the deck was not isolated
completely from the pier (Chaudhary, Abé, and Fujino 2002b; Fujino, Siringoringo, and Abe 2016).
The locking of isolation bearings at one or more pier during a large earthquake would result in
unexpected lateral load redistribution to the substructure because significant inertia force of super­
structure will be transmitted to the substructure. This unwanted condition should be identified as early
as possible and mitigated to avoid further malfunction of the isolation bearings at larger earthquake.

CONTACT Dionysius M. Siringoringo dion@ynu.ac.jp Dionysius Siringoringo Institute of Advanced Sciences, Yokohama
National University, 79-1 Tokiwadai Hodogaya-ku Yokohama, Kanagawa 240-8501, Japan.
© 2021 Taylor & Francis Group, LLC
2 D. M. SIRINGORINGO AND Y. FUJINO

With recent advancement in the structural monitoring system, more seismically isolated bridges
are instrumented, and their structural responses subjected to various excitations can be obtained and
analyzed (Fujino et al. 2019). Most vibration-based diagnostic techniques for bridge assessment are
based on a comparison between response of current state and reference state obtained via finite
element model or initial state of the bridge. Past studies show that performance of the seismically
isolated structures can be evaluated based on the comparison of identified global parameters such as
structural or modal parameters using system identification methods (Chaudhary, Abe, and Fujino
2002a; Derkevorkian et al. 2014; Fujino, Siringoringo, and Abe 2016; Fujino et al. 2019; Siringoringo
and Fujino 2015). Most of the methods are based on the modal analysis in time-domain using
piecewise linear approach or in frequency domain by spectral analysis. However, since the isolation
system is expected to behave nonlinearly, characterization of the time-varying property of the
structures must account for both temporal evolution of the frequency and amplitude contents of
vibration. Neither time analysis nor frequency analysis method alone can fully describe these char­
acteristics. In addition, locking of isolation bearing is a local phenomenon and detecting such
phenomenon based on structural global parameter such as modal parameters can be challenging.
Therefore, it is more desirable to have a detection method directly from the analysis of the vibration
signal without the need for performing a complete process of modal analysis.
In this paper, we describe wavelet-based techniques for detecting the malfunction of isolation
bearing directly from seismic records of the structure. The techniques allow us to monitor the behavior
of isolation bearing and detect any changes related to bearing malfunction such as locking phenom­
enon directly from the seismic records of the bridge without the need for structural model or a
reference state. Organization of the paper is such that the condition and model of the isolated bridge
having malfunctioned or locked bearing is firstly described by an analytical model. Next, the paper
describes the theoretical background on methodologies for analyses of the time-varying structural
system by continuous wavelet transform (CWT) and discrete wavelet transform (DWT). Numerical
simulations by MDOF lumped-mass system and three-dimensional finite element model of multi-
span continuous bridge with numerous scenarios of the locked bearing are presented. Finally, we
describe the application on a real case of multi-span isolated bridge using seismic records from
a 3-year continuous structural monitoring system.

2. Seismic-Induced Vibration of Multi-span Continuous Girder Isolated Bridge


2.1. Modeling of Seismic-induced Vibration of Multi-span Isolated Bridge
Seismic response of a pile-supported base-isolated bridge system is quite complex and requires some
simplifying assumptions to facilitate the formulation of a mathematical model. Generally, the axial
stiffness of the bridge deck is much larger than the shear stiffness of the isolation bearing. Therefore,
the bridge deck is commonly regarded as a rigid body. Moreover, previous research works on
continuous multi-span isolation bridge have shown that predominant dynamic characteristics of the
bridge can be adequately represented by several degree-of-freedom lumped-mass systems in the
longitudinal direction (Chaudhary, Abe, and Fujino 2002a). In this type of model, one degree of
freedom represents the base-isolated super-structure system while the other degree of freedom
represents the sub-structure such as pier or abutment. The effect of the incoherence of support motion
to each bridge pier is usually neglected because the distance between piers is not very large.
In this study, we consider a simplified lumped-mass seismic analysis model of the bridge system as
a multi-degree-of-freedom (MDOF) nonlinear, time-variant system subjected to earthquake excitation
as shown schematically in Figure. 1. The equation of motion of the system is described in Equation (1)
where structural vibration is considered only on the bridge longitudinal direction. In this equation, the
isolation bearing is considered to undergo significant deformation under earthquake excitation that
allows its stiffness to enter the nonlinear region, while the stiffness of bridge piers is assumed to remain
in the linear range. Damping force is also considered nonlinear as the consequence of hysteresis
JOURNAL OF EARTHQUAKE ENGINEERING 3

Figure 1. Schematic figure of (a) continuous multi-span seismically isolated bridge and (b) multi-degree-of-freedom lumped mass
system of the bridge.

behavior of isolation bearing. Equation of motion of a continuous multi-span bridge modeled as an


MDOF nonlinear system becomes:
_
M€xðtÞ þ Fc ðtÞ½xðtÞ� þ Fs ðtÞ½xðtÞ� ¼ MΓ€zðtÞ; (1)

M ¼ diagð½Mg Mp1 � � � Mpn �Þ; (2:a)

where
2P
n 3
Cbi ðtÞ Cb1 ðtÞ ��� Cbn ðtÞ
6 i¼1 7
6 � 7
6 7
Fc ðtÞ ¼ 6 Cb1 ðtÞ Cb1 ðtÞ þ Cp1 ��� 0 7 (2:b)
6 .. .. .. .. 7
4 . . . . 5

Cbn ðtÞ 0 ��� Cbn ðtÞ þ Cpn

2P
n 3
Kbi ðtÞ Kb1 ðtÞ ��� Kbn ðtÞ
6 i¼1 7
6 � 7
6 7
Fs ðtÞ ¼ 6 Kb1 ðtÞ Kb1 ðtÞ þ Kp1 ��� 0 7 (2:c)
6 .. .. .. .. 7
4 . . . . 5

Kbn ðtÞ 0 ��� Kbn ðtÞ þ Kpn

� �T
x ¼ xg xp1 � � � xpn (2:d)

In this equation, matrices M, Fc and Fs are the total mass, damping force and stiffness force of the
bridge system, respectively. Matrix Γis a transformation matrix that relates input ground motion to the
mass matrix. Vectors x , x_ and x are acceleration, velocity and displacement of the bridge in the
longitudinal direction (i.e. along bridge axis), respectively, while zðtÞ denotes the input vectors ground
excitation caused by the earthquake. The subscripts g, p, and b under the mass, damping and stiffness
matrices denote the girder, pier and bearing, respectively. Whereas the subscript i denotes the
corresponding number of piers.
4 D. M. SIRINGORINGO AND Y. FUJINO

2.2. Scenario of Bearing Malfunction and Its Modeling


In this study, the elastomeric type of isolation is considered, however the same principle can be extended
to other types of isolation. Figure 2 describes a typical detail of elastomeric bearing used for seismic
isolation at the continuous girder bridge and its side stopper. In thelongitudinal direction, seismic force
displaces the bearing allowing movement of the girder relative to the pier. Meanwhile in the transverse
direction, displacement of elastomer is tolerated only for limited length (i.e. in the order of few
centimeters) and restrained by the side stopper. The side stopper is usually made of two stiffened
angle or welded steel plates located on both sides of the bearing. In design, a gap exists between the
isolation bearing and the two side stoppers. However, due to pounding, misalignment or poor main­
tenance over the time, there are cases where the gap in one or both sides of the isolator is closed, and the
side stopper becomes locked with the edge of the upper steel plate or the elastomer as shown in Figure 2b.
Once the gap is closed, friction from the contacting surfaces prevents movement of the upper steel plate,
and the deck becomes locked with the pier. In such condition, the side stopper’s stiffness will engage
together with the friction force (Fs) resulting in a significant increase of initial stiffness of the deck-pier
connection since the stiffness of the side stopper is larger than that of the elastomeric bearing. As a result,
the pier is fixed to and moves together with the girder. The isolation system does not function properly
until the seismic force is large enough to overcome the friction force and the girder becomes unlocked
from the pier. This process is implemented in the model by changing the stiffness at the pier–girder
connection as shown schematically in Figure 2c and d.

3. Methodologies for Analyses of Time-Varying Structural System by Wavelet


Transform
Wavelet analyses have been implemented in numerous studies for structural health monitoring (SHM)
using seismic response because of its versatility and robustness in dealing with time-varying properties
of the signal and ability in detecting signal discontinuities. Numerous wavelet-based damage detection
methods have been developed and implemented via numerical models or full-scale seismic

Figure 2. (a) Detail of seismic isolation elastomeric bearing with side stopper, (b) schematic figure of condition when side stopper
locked with upper side plate, (c) system with locked bearing, and (d) illustration of the effect of friction force due to locked bearing
on the backbone curve of pier-girder.
JOURNAL OF EARTHQUAKE ENGINEERING 5

experiments. Hera and Hou (2004) and Hou, Hera, and Shinde (2006), for instance, introduced
a wavelet-based methodology to extract the instantaneous modal parameters to assess the structural
health condition of structures subjected to base accelerations. Todorovska and Trifunac (2007)
investigated the inter-story drifts and changes in the first frequencies of a six-story RC structure
severely damaged by the 1979 Imperial Valley earthquake and found a decrease in the system
frequencies which correlated well with the observed damage. Michel and Gueguen (2010) and
Aguirre, Gaviria, and Montejo (2013) found a change in structural stiffness and thus a decrease of
the fundamental frequency can be detected from continuous wavelet analysis. In the abovementioned
studies, the changes in the wavelet transform map are associated with structural damages. Using
a similar logic, we would investigate the possible use of wavelet transform to detect changes related to
bearing malfunction such as locking phenomenon directly from seismic records of the structure.
There are two types of wavelet transforms implemented in this study: CWT and DWT. The CWT is
a linear transformation that decomposes an arbitrary signal x(t) through basis functions; dilations and
translations of a mother wavelet ψ(t) through convolution of the signal and the scaled/shifted mother
wavelet, such that:
ð �t b�
1 1
Wða; bÞ ¼ pffiffi xðtÞψ � dt: (3)
a 1 a
In Equation 3 , W ða; bÞ is the wavelet coefficient that represents a measure of the similarity between
the dilated/shifted mother wavelet and the signal at time t and scale a. The scale a represents dilatation
that is inversely proportional to frequency and contains the periodic or harmonic nature of the signal.
Parameter b indicates the time shifting and * denotes the complex conjugation. The factor p1ffiaffi is
included to ensure energy preservation.
There are various types of mother wavelet available in the literature. In this study, we utilize one of
the most popular and widely used one, the complex Morlet wavelet (Grossmann and Morlet 1984).
The complex Morlet wavelet is described as:
1 fb ðπfc Þ2
t2
ψðtÞ ¼ pffiffiffiffiffiffi ðei2πfc t e Þe fb ; (4)
πfb
where fb denotes the bandwidth parameter, and fc denotes the
pmother
ffiffiffi pwavelet
ffiffi central frequency. As
2
explained by (Yan, Miyamoto, and Brühwiler 2006), when fb fc � 2 the term e fb ðπfc Þ becomes
negligible, and the approximate version of the modified Morlet wavelet can be expressed as:
1 t2
ψðtÞ ¼ pffiffiffiffiffiffi ei2πfc t e fb : (5)
πfb
The dilated version of the Fourier transform of the modified Morlet wavelet becomes:
π2 fb ðaf fc Þ2
F½ψðtÞ� ¼ e : (6)
Equation (6) shows that the relation between scale a and frequency f is given by fc/a that is the localized
Fourier frequency obtained when F½ψðt Þ� reached the maximum. The time and frequency resolution of
wavelet at frequency fi can be determined by the Heisenberg uncertainty principle. Additionally, the
end effects that can significantly influence the quality of wavelet coefficient are estimated by setting the
pffiffi
f f
time interval Δt ¼ β c 2fi b where the value of β is taken as 4 (Yan, Miyamoto, and Brühwiler 2006).

3.1. Extraction of Wavelet Ridge and Instantaneous Frequency


One of the most widely used properties of CWT is the ridge and skeleton. The ridge of the wavelet
transform is related to the instantaneous frequency (IF) of the signal. To provide definition and
6 D. M. SIRINGORINGO AND Y. FUJINO

derivation of wavelet ridges, consider Equation (3) of the CWT represented as the scalar product
(Tchamitchian and Torrésani 1990):
ð
1 1 t b i½ϕx ðtÞ ϕψ ðt a bÞ�
Wða; bÞ ffi Λx ðtÞΛψ ð Þe dt; (7)
2a 1 a
in which Λ and ϕ are the envelope and instantaneous phase of the asymptotic signal x(t) and wavelet
ψðtÞ, respectively. Note that as pointed out by Tchamitchian and Torrésani (1990) and Staszewski and
Wallace (2014), the essential contribution to Equation (8) is provided by a set of stationary points t = ts
of the function:
t b
Φða; bÞðtÞ ¼ ϕx ðtÞ ϕψ ð Þ; (8)
a
where the stationary points can be determined from the equation:
1 ts b
ða; bÞðts Þ ¼ ϕ_ ψ ð Þ: (9)
a a
The stationary point ts is the set of points in time-scale (a,b) domain for which the ts (a,b) is equal to b.
The curve that is a function of b (e.g. a = f(b)) is a ridge of the wavelet transform.
The ridge of wavelet transform can be interpreted as the location where the major energy of
vibration response is localized among the distributed energy density under over-the-time-scale plane
(a,b). Therefore, the ridge can be considered as the governing or dominating vibration characteristics
at a specific time instant. One can obtain the ridge of wavelet from time-scale plane (a,b) by employing
amplitude-based or phase-based algorithms. In this paper, the amplitude-based ridge extraction is
employed. The ridge is obtained by solving the optimization problem of finding the local maxima of
wavelet transform magnitude along the scale axis for individual time. Results of the wavelet ridge are
considered as the IF of the signal and utilized to trace time-varying characteristics of the response.

3.2. Discrete Wavelet Transform (DWT)


Contrary to CWT whose coefficients are calculated at scale values that vary continuously and resulting
in a highly redundant representation, the DWT analysis decomposes the signal into two components,
namely, the approximation (A) and detail (D) components. The signal decomposition is performed
through two complementary low- and high-pass filters in a dyadic manner. The approximations
contain high-scale (low-frequency) components of the signal, while the details contain the low-scale
(high-frequency) components. At each step of decomposition, the approximation Ai in level i can be
further decomposed into Aiþ1 and Diþ1 in level i + 1 at each such that the original signal x(t) can be
expressed as
Xi
xðtÞ ¼ Ai þ D:
j j
(10)

Because of the nature of the detail components that contain the low-scale (high-frequency) compo­
nents of the decomposed signal, they are often used to detect irregularity and discontinuity associated
with the change in signal characteristics. In the case of seismic response of the isolated structure, the
signal discontinuity is expected when a seismic response of the isolator enters the nonlinear range.
Signal discontinuity associated with yielding behavior can be detected from detailed components
obtained via DWT as stated in the previous studies (Aguirre, Gaviria, and Montejo 2013; Pan and Lee
2002).
The same logic can be applied for detecting signal discontinuity associated with the yielding of
isolation layer in a movable (functioning) bearing or the absence of yielding in the case of a locking or
malfunction isolator bearing. The mechanisms behind this phenomenon can be explained as follows.
JOURNAL OF EARTHQUAKE ENGINEERING 7

One can rewrite the equation of motion described in Equation (1) by assuming an undamped case for
simplification. Using the first row of the matrix, the girder acceleration (ag Þis rewritten as:
Xn
ag ¼ €xg ¼ i¼1
ω2bgi ðxpi xg Þ: (11)

Similarly, acceleration of the ith pier (api Þ can be expressed using the corresponding (i + 1)th row of
Equation (1) as:

api ¼ €xpi ¼ ω2bpi ðxg xpi Þ þ ω2pi xpi : (12)

The quantities ωbgi , ωbpi , and ωpi are the circular frequency of the bearing-girder system, the bearing-
qffiffiffiffiffiffiffiffi
pier system and the pier system, respectively, and they are defined, respectively, as ωbgi ¼ KM b i ðt Þ
g
;
qffiffiffiffiffiffiffiffi qffiffiffiffiffi
K
ωbpi ¼ KMb i ðpitÞ and ωpi ¼ Mppii .
Equation (11) shows that girder acceleration depends on the circular frequency of the bearing-
girder system and the relative displacement between girder and piers. Whereas from Equation (12),
one can see that the pier accelerations are a function of circular frequency of the bearing-pier system,
relative displacement between girder and pier, circular frequency of pier system, and displacement � of
the pier itself. Note that both the circular frequency of the bearing-girder system ωbgi and the
bearing-pier system (ωbpi Þ vary with time and with the relative displacement between pier and girder.
Their values depend on the state of bearing stiffness which is bilinear. The values of circular frequency
before yielding are ωebgi and ωebpi , where the superscript e denotes the elastic state; obtained from the
initial (pre-yielding) bearing stiffness. After the yielding, the values of circular frequency become
y y
ωbgi and ωbpi , where the superscript y denotes the post-yielding state with post-yielding bearing
stiffness. Meanwhile, the ωpi is assumed to be constant since the stiffness and mass of the piers are
assumed to be constant and much higher than the stiffness of isolation bearing.
To understand the effect of circular frequency and bearing state on the signal acceleration, one can
differentiate acceleration with time to obtain the rate of change. For girder acceleration, this becomes:
Xn Xn
r r
a_ g ¼ 2 ω ω0 bgi x_ gpi
i¼1 bgi
xgpi þ ω2 x_ r ;
i¼1 bgi gpi
(13)
r r

where xgpi is the relative displacement between girder and pier i, defined as xgpi ¼ xpi xg , and the
r
prime (‘) denotes the differentiation with respect to xgpi . Similarly, one can compute the rate of change
in acceleration of the ith pier as:
r r
a_ pi ¼ 2ωbpi ω0 bpi x_ gpi xgpi ω2bpi x_ gpi
r
þ ω2pi x_ pi : (14)

Both Equations (13) and (14) show that the rate of change in accelerations of girder and pier is
a function of circular natural frequency that varies with time following isolation bearing condition.
Consequently, there would be a discontinuity of acceleration because of the abrupt changes in the ωbgi
and ωbpi due to the change in the stiffness of isolation bearing from the pre-yielding to post-yielding
state and vice versa. Discontinuity in acceleration response can be detected as a time-localization in the
low-scale (high-frequency) components of the decomposed signal using DWT. In this study, the
orthonormal wavelet from the fourth order of the Daubechies family or ‘Db4ʹ for DWT is utilized
(Daubechies 1992). The ‘Db4ʹ wavelet has orthogonal compactly supported wavelet functions that
allow for symmetric and exact reconstruction. The Daubechies’ wavelet family is known to have
worked well in detecting a disturbance in the signal and the Db4 wavelet was found to be useful in
detecting signal discontinuity in its first derivative.
The objective of using wavelet transforms in this study is twofold. The CWT is employed to detect
shifts in the vibration frequency of seismic responses using IF extracted from the ridges of the CWT
time-frequency map. Meanwhile, the detail component (Di) obtained from DWT of acceleration is
8 D. M. SIRINGORINGO AND Y. FUJINO

utilized to detect irregularities in the high frequency associated with the change in stiffness character­
istics of isolation bearing. By combining and comparing the results of CWT and DWT methods,
therefore, one expects to obtain information on both frequency variation and the time of occurrence of
the locking or sliding of the isolation bearing from the recorded seismic response.

3.3. Characterization of Isolator Bearing Condition by K-means Clustering Classification


Results of identification, namely, the IF of CWT and the detail component of DWT inherently contain
statistical variations. This is quite common especially when implemented directly to seismic records
that inevitably contain some measurement noise. To obtain a meaningful decision on the bearing
condition, the results are further analyzed using a statistical approach by comparing the extracted
features using the K-means clustering classification technique (Arthur and Vassilvitskii 2007; Chawla
and Aristides 2013). The k-means clustering algorithm performs a partition of data space into
kclusters. Each cluster is represented by an object, which is the centroid or the mean point, whose
values are estimated in the algorithm.
In this study, two clusters were assigned, namely the isolator bearing with the locked condition and
the normal or movable isolator bearing. The method is implemented as follows. Once the results of
identification are obtained, the Euclidean metric is computed to determine the distance between the
clusters using Equation (15):
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn
DistðU; VÞ ¼ jjU Vjj ¼ i¼1
ðui vi Þ2 : (15)

In this equation, U ¼ ðu1 ; u2 . . . ; un Þ and V ¼ ðv1 ; v2 . . . ; vn Þ are the vectors of vibration features,
namely the IF of CWT and the detail component of DWT of acceleration records obtained from girder
and piers. As shall be seen later, the distance between the two clusters of vibration features at the peak
of seismic excitation are utilized as indictor to determine the condition of the isolator bearing
quantitatively.

4. Numerical Examples by Lumped-Mass MDOF Bridge System


Numerical simulation using MDOF lumped-mass of continuous multi-span base-isolated bridge is
described in this section. The model consists of a continuous girder with five piers illustrated in Figure.
1. Mass of the girder and piers as well as the properties of isolator bearings are listed in Table 1. Note that
since the main objective is to investigate the performance of isolation bearings, piers are assumed to be
undamaged and behave linearly throughout the entire excitation. The input ground acceleration motion
used in the following numerical analyses is the 1995 Kobe earthquake Takatori Station east-west compo­
nent. The ground motion contains 4095 data points at sampling frequency 100 Hz and the peak ground

Table 1. Structural properties used in lumped-mass MDOF isolated bridge


system.
Structural Properties
Mass of Girder (ton) 1000
Mass of Pier (ton) 100
Stiffness of Pier (ton /m) 50000
Bearing Properties
Initial stiffness (ton/m) 2000
Secondary Stiffness (ton/m) 500
Yield Force (ton) 450
Yield Displacement (m) 0.225
Stiffness of Side Stopper
Initial stiffness (ton/m) 20000
Secondary Stiffness (ton/m) 5000
Yield Force (ton) 4000
JOURNAL OF EARTHQUAKE ENGINEERING 9

acceleration is 603.2 cm/s2 at 2:83 s. The step-by-step integration of the Wilson-ϴ method is used to
compute the output acceleration response for a given input ground motion. Bilinear hysteresis behavior of
isolation bearing is modeled by Masing’s rules (Masing 1926). The Masing’s rule is also employed in
nonlinear computation analysis in conjunction with the backbone curve to obtain the response of the
structure.
In the numerical model, all piers are assumed to have the same dimension, materials, and
foundation condition. Therefore, they all have equal stiffness. The pier stiffness accounts for the
stiffness of the pier structure and foundation. All isolation bearings are assumed to have the same
bilinear hysteresis behavior. To simulate the case of locked bearing, one or more bearings is assumed
to be locked while the other piers are movable. This condition is achieved by assigning stiffness of side
stopper as the initial stiffness instead of the original initial stiffness of the isolator as illustrated in
Figure 2c and d

4.1. Observation from Extracted Wavelet Ridge and Instantaneous Frequency of Continuous
Wavelet Transform
Results from two case studies: single locked bearing at pier P3 and multiple locked bearings at pier P2
and P4 are described in this section. In the first case, a single locking of isolator bearing is assumed for
the middle pier (P3) while other bearings remain functioning. Figure 3 shows the results of simulation
of the MDOF system with the scenario explained above. The figure describes the time-frequency map
of the wavelet transform of accelerations and the IF extracted from the ridges of the time-frequency
map by employing the procedure above. The figure shows that the girder acceleration is dominated by
a single low-frequency response at 0.75 Hz, which corresponds to the girder mode. The piers

Figure 3. Results of simulation of MDOF system with one locked bearing. Accelerations, time-frequency map of acceleration and
instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map of accelerations: (a) girder accelera­
tion, (b) pier no.1. acceleration (moveable), (c) pier no.2. acceleration (moveable), (d) pier no.3. acceleration (locked), (e) pier no.4.
acceleration (moveable), and (f) pier no.5. acceleration (moveable).
10 D. M. SIRINGORINGO AND Y. FUJINO

acceleration responses have at least two dominant frequencies, namely, 0.75 Hz and between 3.20 and
3.75 Hz. The higher frequency corresponds to the pier frequency at 3.63 Hz.
Based on the time-frequency map and IF, one can see that the locked pier (P3) has a very different
pattern of frequency distribution compared to other movable piers. At the movable piers (P1, 2, 4 and 5),
accelerations were dominated by the higher frequency components at the time of the largest excitation
between t = 2 s and t = 16 s. During this time interval, the effect of girder frequency at 0.75 Hz is
negligible. This suggests that the piers are isolated from the girder at the time of peak excitation, which is
expected from a functioning isolator bearing. Acceleration of the pier with a locked bearing (P3),
however, shows that the dominating frequency as shown by the IF is generally around 0.75 Hz at the
peak excitation and throughout the response. This indicates that the pier response is dominated by girder
frequency, which could occur only when the pier and girder move together as in the case of fixed or non-
isolated pier-girder connection. Note that in the beginning of the response (t < 2 s) and after the peak
excitation until the end of the response (t = 16–40 s), the accelerations of both locked and unlocked piers
are dominated by girder frequency at 0.75 Hz, which suggests that piers and girder move together in the
beginning and after the occurrence of peak excitation when the response is relatively smaller.
In the second example, two isolation bearings at pier P2 and pier P4 are assumed to be locked while
other bearings function normally. Figure 4 shows the results of simulation. In general, time-frequency map
of the girder and pier acceleration is similar to the case of single isolator locked. The time-frequency map of
girder acceleration is dominated by a single low-frequency component at 0.93 Hz corresponding to the
girder mode. This frequency is slightly higher than the corresponding frequency in the first case due to the
stiffness increase caused by the two locked bearings. Similarly, the pier accelerations have at least two
instantaneous frequencies, namely, 0.93 Hz and between 3.60 and 4.25 Hz. The higher frequency
corresponds to the pier frequency at 4.20 Hz. Note that girder longitudinal frequency is higher in this
case than in the first case because with two locked bearings, the structure becomes stiffer. Figure 4c and

Figure 4. Results of simulation of MDOF system with two locked bearings. Accelerations, time-frequency map of acceleration and
instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map of accelerations: (a) girder accelera­
tion, (b) pier no.1. acceleration (moveable), (c) pier no.2. acceleration (locked), (d) pier no.3. acceleration (moveable), (e) pier no.4.
acceleration (locked), and (f) pier no.5. acceleration (moveable).
JOURNAL OF EARTHQUAKE ENGINEERING 11

e shows that the locked piers (P2 and P4) have a different pattern of frequency distribution compared to the
movable piers. On the locked piers, single low-frequency response at 0.93 Hz dominates the response, while
on the moveable piers, the responses were dominated by higher frequency 4.2 Hz at the time of largest
excitation between t = 2 s and t = 16 s. Figure 4 also shows that in the beginning of the response (t < 2 s) and
after the peak excitation until the end of the response (t = 16–40 s), the accelerations of both locked and
unlocked piers are dominated by the girder frequency 0.93 Hz indicating that piers and girder move
together after the occurrence of peak excitation when the responses.
The above results reveal that bearing condition can be inferred from observation of IF derived from the
wavelet transform girder and pier’s seismic response. For the pier with a functioning isolation bearing, the
acceleration is expected to shift to a higher frequency range at the time of large excitation. Whereas the
pier’s acceleration dominated by low-frequency component corresponding to the girder at the time of large
excitation could be an indication of a locked or malfunction bearing. The results indicate that we can
distinguish the condition of isolator bearing by comparing the time-frequency map and IF obtained directly
from seismic responses recorded at the pier and girder without the need of structural analysis.

4.2. Observation of Detail Components Obtained from Discrete Wavelet Transform


The detail components were obtained by taking DWT of acceleration response using the Daubechies
family fourth-order wavelet or ‘Db4ʹ. Several levels of detail components can be obtained from DWT,
and in this example, we utilize the first level of detail component or D1. The values are normalized at
each time instant (i) with the mean value (μ) and standard deviation (σ) such that:
normðDiÞ ¼ abs½ðDi μÞ=σ�: (16)
Figure 5 exhibits time histories of first-level detail (D1) of the accelerations obtained from MDOF
simulation for the first case of single locking isolator bearing at the middle pier (P3). In general, the D1
values are larger at the time when accelerations are large. As shown by the figure, the D1 values near the
peak of acceleration are larger than those near the end of excitation. In addition to that, the figure
clearly shows several higher and distinct spikes between t = 5 and 10 s for the D1 values of pier 1, 2, 4
and 5. For pier 3, however, these higher and distinct spikes do not appear at this time interval. As
explained previously in Equations (12)–(15), the higher and distinct spikes on the norm of D1 are the
indicator of discontinuity of acceleration response. The discontinuity is caused by abrupt changes in
the ωbgi and ωbpi due to the change in the bearing stiffness from the pre-yielding to post-yielding state
and vice versa.
To describe this phenomenon more clearly, transitions of the state from linear to nonlinear during
loading and unloading are compared with the spikes in the detail functions. Let the state of isolator be
defined as ‘PrY’ for pre-yielding condition, where the stiffness of isolator is the initial stiffness, and as ‘PoY’

Figure 5. First level detail components (D1) obtained from DWT of accelerations and the isolator bearing yield condition from MDOF
system with single bearing locked: (a) pier no.1. (moveable), (b) pier no.2. (moveable), (c) pier no.3. (locked), (d) pier no.4.
(moveable), and (e) pier no.5. (moveable).
12 D. M. SIRINGORINGO AND Y. FUJINO

for the post-yielding where the stiffness of isolator is the secondary stiffness in the bilinear hysteresis curve.
The second row of Figure 5 shows the time history of the yield condition for bearings on all piers. The figure
demonstrates that all bearings have reached the yield state except for the one on pier no.3. When the
bearing condition jumps from elastic (PrY) to yield (PoY), there is a spike or an abrupt change in D1.
Similarly, when the state of isolator changes from PoY to PrY during the unloading, there is also an abrupt
change in D1.
The figure shows clearly that the occurrence of abrupt change in D1 correlates well with the change
in the state of isolator, and the values of detail component D1 also correlate well with the changes in the
slope of acceleration at both the yielding and unloading. Hence, this information can be used as an
indicator of isolator bearing condition. By defining the threshold values of absolute normalized D1 to
be 12 times the standard deviation (σ), the first-level detail component plots can be used as an
indicator of yield condition and hence the associated bearing condition. The premise is that at the
time when a functioning elastomeric isolation bearing enters a nonlinear state during large excitation,
an abrupt change in the normalized detail components (D1) could be detected. On contrary, the
yielding mechanism would not be initiated when the bearing is locked, and as a result, there would not
be any abrupt changes in the detailed components of DWT decomposition of the acceleration.
Similar results were observed for the second example where two isolator bearings at pier P2 and pier P4
were locked. In Figure 6, one can clearly see the presence of higher and distinct spikes between t = 5 and 10 s
on the time-histories of normalized D1 of pier accelerations of the functioning (movable) isolation bearings.
On the other hand, no spikes exceeding the threshold value were detected from the detail components (D1)
of pier accelerations with locked bearings P2 and P4. From the figure, it can be seen that whenever bearing
condition changes from elastic to yielding, higher spikes exceeding the threshold value appear on the
normalized D1. Additionally, the timing of yielding and unloading stated in the figure of yield condition can
be detected from the spikes on the normalized detail components (D1) as shown by the second row of
figures in Figure 6.
The above results demonstrate that although the DWT method cannot give information on the
frequency characteristics of the structure, it is useful in detecting the change in isolator condition when
yielding and unloading occurs. Hence, the observation of detail components from DWT method can be
used as a complementary to the results of IF obtained from the time–frequency map of the CWT
coefficients. Combination of the results from both methods provides a more accurate indicator of isolator
bearing condition.

4.3. Characterization of Isolation Bearing Condition Using k-means Clustering


The above examples show that the condition of bearings can be inferred directly from the results of IF
of CWT and details component of DWT. Two procedures based on k-means clustering of IF and detail
component are explained in this section. First, based on the pattern of IF of piers and girders, one can

Figure 6. First level detail components (D1) obtained from DWT of accelerations and the isolator bearing yield condition from MDOF
system with two bearings locked: (a) pier no.1. (moveable), (b) pier no.2. (locked), (c) pier no.3. (moveable), (d) pier no.4. (locked), and
(e) pier no.5. (moveable).
JOURNAL OF EARTHQUAKE ENGINEERING 13

categorize them into two clusters, the normal or movable bearing and locked bearing. In the first
category, the instantaneous frequencies at the peak of acceleration are close to the pier fundamental
frequency, whereas for the second category they are closer to the girder’s frequency. To provide
a classification of the cluster as locked or movable bearing, the distance between two clusters of
fundamental frequencies is normalized with the absolute difference between the pier and girder
frequency in each case obtained by the model. The normalized distance (ND) between the clusters
is defined as:
� ��
� Uc1 Uc2 �
NDðUc1 ; Uc2 Þ ¼ �� �: (17)
ωp ωg �

In this equation, Uc1 and Uc2 are the centroids of cluster 1 and 2, respectively; whereas ωp and ωg are
the theoretical values of pier and girder frequency, respectively. Note that the clusters can be defined
by employing the procedure explained in Equation (16). In the case of moveable bearing, the ND value
will be close to one since the distance between the centroids of the two clusters will approximately
equal to the difference between pier and girder frequency. On the other hand, the ND value will be
closer to zero for the locked bearing since the distance between the centroids of the two clusters is
much smaller than the difference between pier and girder frequency.
A similar classification of the cluster can be assigned using the results of detail components
obtained from DWT. The distance between the cluster of normalized D1 to the threshold value is
computed for the interval time of peak excitation. The ND between the cluster centroid of detail
components and the threshold value is defined as:

Uc e
U
e ¼
NDðUc ; UÞ ; (18)
e
U
where Uc is the centroid of the cluster of identified detail components of DWT (D1) and U ~ is the
threshold value of the detail component. One can note from Equation (17) that the ND becomes positive
when D1 exceeds the threshold value as in the case of moveable bearing. In the case of locked bearing, the
ND value becomes negative. In addition, the ND amplitudes is a measure of how far the detail
component deviates above or below the threshold value.
To investigate the effectiveness of the k-means clustering method in detecting the bearing condi­
tion, case studies were conducted by considering all possibilities of isolation bearing conditions for
a 5-pier multi-span girder bridge. Overall, there are 32 possible scenarios including no locked bearing
case, 1 pier with locked bearing, 2 piers, 3 piers and 4 piers with locked bearing, and finally all piers
with locked bearings. To simulate a realistic measurement case, Gaussian random noise with a 1%
noise-to-signal ratio of the root-mean-square (RMS) of acceleration amplitude was added to the girder
and pier acceleration records prior to performing the wavelet transform.
Results of cluster classification using instantaneous frequencies obtained via CWT are provided as
tables in Figure 7. The identified ND values shown in the tables are computed in Equation (17). The
abscissa denotes the case with a locked pier number and the ordinate shows the ND values for each
pier. The results clearly demonstrate that piers with locked bearing have smaller ND values close to
zero, while those with normal (movable) bearings have higher ND values close to one. All cases with
locked piers can be accurately identified based on the relative values of ND. The differences in the
values for the same condition of either locked or movable are not very significant and can be attributed
to the noise effect in the acceleration records. In all 32 cases, the simulated bearing condition can be
accurately characterized as a normal (moveable) or as a locked one based on the comparative values of
ND from the table.
Figure 8 describes the results of the classification for the first level of detail component (D1)
computed using Equation (18). The threshold value of absolute normalized D1 was set to be 12
times the standard deviation (σ), which was found through numerical simulation. The tables show the
results from all 32 possible cases of isolation bearing condition. As shown in the figures, the ND values
14 D. M. SIRINGORINGO AND Y. FUJINO

Figure 7. Classification of isolator bearing condition based on k-means clustering of the instantaneous frequencies obtained from
CWT of accelerations. ND denotes the identified normalized distance (ND values closer to one means moveable bearing, closer to
zero means locked bearing). (a) Case of no locked bearing and only one pier with locked bearing, (b) two piers have locked bearings,
(c) three piers have locked bearings, (d) four piers or more have locked bearings. Note: Pij in the abscissa denotes the pier number
with locked bearing.

Figure 8. Classification of isolator bearing condition based on k-means clustering of the first level detail (D1) component obtained
from DWT of accelerations. ND denotes the identified normalized distance (a) Case of no locked bearing and only one pier with
locked bearing, (b) two piers have locked bearings, (c) three piers have locked bearings, (d) four piers or more have locked bearings.
Note: Pij in the abscissa denotes the pier number with locked bearing.

for piers with movable bearings were all classified as a positive value smaller than one, whereas the ND
values for piers with locked bearings were classified as negative with an absolute value smaller than
one. The tables clearly show that bearing condition either locked or movable can be classified correctly
as the intended by simulation in all cases. The amplitudes of ND values may change relatively
indifferent with similar bearing conditions because of the noise effect. However, by comparing the
ND values of all piers in one case, we can clearly classify the bearing condition correctly by referring to
the classification criteria mentioned in Equation (18).

5. Case Study Using Three-dimensional Finite Element Model


The investigation is extended to a more realistic case using a three-dimensional finite element model
of a seismically isolated multi-span continuous bridge. The objectives of the simulation are twofold:
one is to confirm the behavior of seismically isolated bridge with malfunction bearing scenarios and
their seismic responses, and the other is to ascertain that the MDOF model explained above can be
JOURNAL OF EARTHQUAKE ENGINEERING 15

used to capture the essential behavior of the bridge when compared with the three-dimensional finite
element model. The bridge model has four continuous spans of 32 m length and 10 m width each
supported by a 10 m tall pier. The bridge piers and girder are made of reinforced concrete material. All
piers have the same dimension and are assumed to have equal stiffness. Details on cross-sectional
dimension of the pier, girder and bent are provided in Table 1.
Figure 9 illustrates the fishbone model of the bridge built using structural analysis software
SAP2000 (CSI 2013). The bridge consists of three main structural components: girder, pier and bent
that were all modeled by frame element. On each pier, the girder is supported by five rubber isolation
bearings. The elastomeric rubber isolation bearing is modeled by a nonlinear link element that covers
the nonlinear behavior of isolation bearing and allows for nonlinear analyses for seismic response
computation. In the model, all bearings are moveable in longitudinal and lateral directions. They all
have the same structural properties and bilinear hysteresis behavior, except for the bearings at the end
pier which were assumed to have larger stiffness. Seismic analysis was conducted using the 1995 Kobe
earthquake Takatori station ground motions as the ground excitation. The half-scale of the east-west
(EW) and north-south (NS) components of 1995 Takatori ground motions were applied in a direction
parallel and perpendicular to the bridge axis, respectively. Similar to the previous analysis of the
MDOF system, one or more isolation bearings are assumed to be locked while the other piers are
movable to simulate cases of the locked bearing by assigning stiffness of side stopper as the initial
stiffness instead of the original initial stiffness of the isolator. Details on structural information and
isolation bearings are listed in Table 2.
In the analyses, it is assumed that accelerations from the girder and pier caps are recorded and used
to evaluate isolation bearing conditions. Since the bridge girder is a continuous rigid body that has
larger lateral stiffness than the lateral stiffness of isolator bearing, the girder longitudinal accelerations

Table 2. Structural properties used in lumped-mass MDOF isolated bridge system.


Structural Element Area (m2) Moment of Inertia [Ixx, Iyy, Iq] (m4)
Girder 5.80 [14.48, 2.26, 5.9]
Pier 3.00 [0.57,1.00, 1.215]
Bent 1.44 [0.18,0.18,0.292]
Bearing properties Mid pier End Pier
Initial stiffness (kN/m) 3000 5000
Post yield stiffness ratio 0.2 0.2
Yield strength (kN) 312.5 312.5
Ultimate displacement (m) 0.15
Stiffness of Side Stopper
Initial stiffness (ton/m) 30000
Post yield stiffness ratio 0.2
Yield force (ton) 15000

Figure 9. Finite element model of continuous multi-span isolated bridge used in the case study.
16 D. M. SIRINGORINGO AND Y. FUJINO

are equal along the bridge, thus only acceleration from one point in the center of the girder is used for
the analysis. Meanwhile for the piers, accelerations at the pier cap in the bridge axis (longitudinal)
direction are utilized.

5.1. Extracted Wavelet Ridge and Instantaneous Frequency of the Finite Element Results
As the first case, all bearings are assumed to function properly. This means that during the peak
excitation, all isolators have reached the nonlinear stage. This case is investigated to provide
a reference for the other case of the locked bearing. Figure 10 shows the characteristics of time-
histories accelerations, the time-frequency map and corresponding IF. The largest peak in the time-
frequency map occurs during the peak excitation period t = 2–8 s. At this time, the girder response is
dominated by a single low-frequency response at 0.98 Hz, which corresponds to the girder long­
itudinal mode. Based on the IF, it is confirmed that the single-frequency girder longitudinal mode
remains the dominating frequency until the end of excitation. Meanwhile, all piers accelerations have
the same characteristics. One can clearly see nonstationary characteristics of the acceleration time
history which can be verified by the time-frequency map and IF. At the beginning of the response, after
the arrival or the secondary wave, the responses were characterized by higher instantaneous frequen­
cies, between 4.76 Hz and 5.3 Hz. The frequency range of 4.76–5.3 Hz corresponds to the pier
longitudinal frequency at 5.25 Hz. Note that at the time interval t = 2–5 s, which corresponds to the
duration of peak excitation, the IF drops from 5.14 Hz to 4.74 Hz before increasing further to 5.20 Hz
at t = 10 s. After the period of peak excitation, the IF drops much lower than 0.98 Hz until the end of
vibration. The results demonstrate that in a moveable bearing, the pier and girder accelerations during
the peak excitation were dominated by their own respective frequency suggesting the uncoupled

Figure 10. Results of finite element simulation for all piers with moveable bearings; accelerations, time-frequency map of
acceleration and instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map: (a) girder
acceleration, (b) pier no.1. acceleration (moveable), (c) pier no.2. acceleration (moveable), (d) pier no.3. acceleration (moveable),
(e) pier no.4. acceleration (moveable), and (f) pier no.5. acceleration (moveable).
JOURNAL OF EARTHQUAKE ENGINEERING 17

separate systems. In this condition, the piers are isolated from the girder and the high-frequency
content of the pier is not transferred to the girder, which is expected from a functioning isolation
bearing.
In the second simulation, we consider the case of single locked bearing at the middle pier (P3).
Figure 11 reveals the acceleration time histories, time-frequency map and the IF of girder and all piers.
It is evident from the figure that the accelerations of all moveable piers are dominated by two
frequency components: the frequency associated with pier main frequency during the peak excitation
(t = 2–15 s) and the girder frequency at the end of the response. Meanwhile, acceleration of the pier
with locked bearing (P3) is mainly dominated by the girder frequency at around 1.0 Hz. One can
observe the similarity of time-frequency characteristics of accelerations on the girder and the pier with
locked bearing. The similarity indicates that both girder and pier move together as a coupled system.
The same phenomenon was observed when single locked bearing occurs at a different pier, which is
not shown here because of space limitation.
The third case simulates a case of multiple locked bearings at pier P2 and P4. The acceleration time
histories, time-frequency map and the IF of girder and all piers are shown in Figure 12. The figure
shows that the characteristics of the time-frequency map of the piers with locked bearing are very
similar to that of the girder. During the peak acceleration at the time t = 2.5–10 s, the main frequency
content of these pier accelerations is 1.1 Hz which is the frequency of the girder’s longitudinal mode.
Note that this frequency is slightly higher than the corresponding frequency in the first two cases due
to the stiffness increase caused by the two locked bearings. This trend continues until the end of
excitation indicating that girder with pier P2 and girder with P4 move together as coupled system. On
contrary, the piers with movable bearings were characterized by high-frequency component at 4.76–­
5.20 Hz at the peak of excitation between t = 2.5 and 10 s. This is the time when the girder becomes

Figure 11. Results of finite element simulation: accelerations, time-frequency map of acceleration and instantaneous frequency
(noted by red lines) obtained from the ridges of time-frequency map of accelerations: (a) girder acceleration, (b) pier no.1.
acceleration (moveable), (c) pier no.2. acceleration (moveable), (d) pier no.3. acceleration (locked), (e) pier no.4. acceleration
(moveable), and (f) pier no.5. acceleration (moveable).
18 D. M. SIRINGORINGO AND Y. FUJINO

Figure 12. Results of finite element simulation: accelerations, time-frequency map of acceleration and instantaneous frequency
(noted by red lines) obtained from the ridges of time-frequency map of accelerations: (a) girder acceleration, (b) pier no.1.
acceleration (moveable), (c) pier no.2. acceleration (locked), (d) pier no.3. acceleration (moveable), (e) pier no.4. acceleration (locked),
and (f) pier no.5. acceleration (moveable).

isolated from the piers and the isolation bearings start to function. Like in the previous cases, the IF
then drops to the girder frequency 1.1 Hz after the peak excitation until the end of vibration suggesting
that the girder and piers move together as in a non-isolated system at the small amplitude of vibration.
Note that the results of finite element simulations and the characteristics of their time-frequency
maps are consistent with the previous results from the lumped mass simulation. Time-frequency
response characteristics of the moveable and locked bearings are consistent with the MDOF lumped
mass simulation suggesting that the behavior of the bridge can be modeled adequately by the lumped
mass system. Therefore, the behavior of isolation bearing can be observed directly from the IF
evolution of the acceleration via the time-frequency map of CWT.

5.2. Observation of Detail Components Obtained from Discrete Wavelet Transform of the
Finite Element Results
The first level of detail components was computed via DWT of accelerations using the Daubechies
fourth-order wavelet or ‘Db4ʹ. The D1 values were normalized at each time instant by the normalization
procedure described in Equation (16). The time histories of absolute values of the normalized detail
components are plotted for the three cases and compared with the time history of the yield condition for
bearings on all piers. In the first case of all bearings assumed to function properly, Figure 13 shows that
the time histories of D1 are characterized by distinct and higher spikes at the time interval of peak
accelerations. For all piers, the values of absolute normalized D1 exceed the threshold (12 times the
standard deviation (σ)). The second row of the figure illustrates the transitions of the isolator state from
linear to nonlinear during loading and unloading processes. Compared to the spikes in the detail
functions, the transition state figure clearly demonstrates that all bearings have reached the yield state.
JOURNAL OF EARTHQUAKE ENGINEERING 19

Figure 13. First level detail components (D1) obtained from DWT of accelerations and the isolator bearing yield condition from finite
element model with all moveable bearing: (a) pier no.1, (b) pier no.2, (c) pier no.3, (d) pier no.4, and (e) pier no.5.

At the interval time of peak excitation, the bearing condition changes from elastic (PrY) to yield (PoY),
which were followed by the abrupt changes in the acceleration responses as signified by the higher spikes
in the D1 time histories.
In the second case (Figure 14), the first level detail components (D1) of all piers contain
higher values exceeding the threshold value at the time interval t = 2–5 s except for pier P3 in
the middle that has a locked bearing. Comparison between the D1 time histories and the
transitional state of isolator plot reveals a remarkable correlation between the occurrence of
these higher D1 spikes and the transitional state of isolation bearing from elastic (PrY) to yield
(PoY) and vice versa.
Similar results were observed for the third case where two locked isolation bearings at pier P2 and pier P4
(Figure 15). One can clearly see the presence of higher and distinct spikes between t = 2 and 5 s on the time-
histories of normalized detail components (D1) of the functioning (movable) isolator bearings at pier P1, P3
and P5. On the other hand, no spikes exceeding the threshold value were detected from the D1 of pier
accelerations with locked bearings P2 and P4. The figures clearly demonstrate that whenever bearing
condition changes from elastic to yielding, the spikes exceeding threshold values appear on the normalized
D1. The results are consistent with the observation from the D1 component of the MDOF system explained
above.
The results have demonstrated that discontinuity in the acceleration responses caused by the change
in the isolation state at the time of yielding can be detected effectively by the fit level of detail components
obtained from DWT. The results from finite element simulation have confirmed similar remarkable
detection ability observed in the MDOF simulation by the lumped-mass system. The yielding
information contained in the D1 amplitudes provides a good indicator of bearing condition from

Figure 14. First level detail components (D1) obtained from DWT of accelerations and the isolator bearing yield condition from finite
element model for: (a) pier no.1. acceleration (moveable), (b) pier no.2. acceleration (moveable), (c) pier no.3. acceleration (locked),
(d) pier no.4. acceleration (moveable), and (e) pier no.5. acceleration (moveable).
20 D. M. SIRINGORINGO AND Y. FUJINO

Figure 15. First level detail components (D1) obtained from DWT of accelerations and the isolator bearing yield condition from finite
element model for: (a) pier no.1. acceleration (moveable), (b) pier no.2. acceleration (locked), (c) pier no.3. acceleration (moveable),
(d) pier no.4. acceleration (locked), and (e) pier no.5. acceleration (moveable).

which one can infer the possible occurrence of the locked bearing. Therefore, one can infer the bearing
condition from given D1 values at the time of peak excitation and compare their amplitudes with the
threshold value to make a reasonable judgment and informed decision on the bearing condition.

5.3. Implementation of k-means Clustering for Characterization of Isolation Bearing


Condition from Finite Element Simulation
To investigate more comprehensively the effectiveness and accuracy of detection using CWT and
DWT results, the results of simulations by finite element models were summarized in a form of
classification of the bearing condition using the k-means clustering method. The method was
implemented to the CWT and DWT results following the same procedure described in the
previous section for the MDOF simulation. All possible 32 cases of isolation bearing conditions
and combinations were investigated, and for each case, a separate finite element model of the
bridge was constructed. The 32 finite element models include all possible combinations of bearing
scenarios, namely all functioning bearing, locked bearing at one pier, locked bearings at two piers,
three and four piers and locked bearings at all five piers. In all cases, the input excitation and
components of the finite element model were the same except for the pier with simulated locked
bearing. The girder and pier acceleration records were added with random noise with a 1% noise-
to-signal ratio of the RMS of acceleration amplitude before taking the wavelet transform to
simulate measurement noise. For each case, longitudinal accelerations from six locations were
used in the analysis, namely, one node in the center of the girder and five nodes at the respective
pier cap.
Relationship between the centroids of IF values and the locked bearing model shown in Figure 2
can be explained as follows. A locked bearing is a condition where the friction between side stopper
and steel plate above the isolator prevents the isolator to move freely. This condition is modeled by
assigning higher initial stiffness that is of the side stopper instead of the original initial stiffness of the
elastomeric bearing in the backbone curve of the isolator as illustrated in Figure 2c and d. This higher
initial stiffness is attributed to the effect of friction between side stopper and steel plate above the
isolator. As the consequence of the locked bearing condition, the frequency contents of girder and pier
accelerations are dominated by a single low-frequency component that corresponds to girder mode.
This frequency is higher than the frequency of the girder in the case of normal (unlocked) bearing
because of higher initial stiffness. The physical condition of locked and normal (unlocked) bearing can
be inferred from the ND values denoted in Equation (17). The centroids of IF cluster data represent
the dominant IF of the girder and the dominant IF of the pier’s acceleration. The values of Uc1 and Uc2
are closer to the girder’s frequency and pier’s frequency, respectively, when the isolator functions
JOURNAL OF EARTHQUAKE ENGINEERING 21

Figure 16. Classification of isolator bearing condition for the FE simulations based on k-means clustering method of the
instantaneous frequencies obtained from CWT of accelerations. ND denotes the identified normalized distance (ND values closer
to one means moveable bearing, closer to zero means locked bearing). (a) Case of no locked bearing and only one pier with locked
bearing, (b) two piers have locked bearings, (c) three piers have locked bearings, (d) four piers or more have locked bearings. Note: Pij
in the axis label denotes the pier number with locked bearing.

normally. On the other hand, the values of Uc1 and Uc2 become closer to each other when the bearing is
locked because the girder and pier move together as a coupled system.
Figure 16 exhibits the results of cluster classification using instantaneous frequencies obtained
via CWT for the 32 cases of finite element models. The abscissa in each case describes the pier
number with locked bearing and the ordinate describes the identified ND values that are
computed in Equation (17). It is evident from the results that piers with locked bearing can
be identified based on ND values closer to zero, while piers with normal (movable) bearings
were identified with ND values close to one. The method accurately classifies all cases of the pier
with locked bearing from the relative values of ND. Similar to the case studies by the MDOF
system, using the comparative values of ND one can accurately characterize the simulated
bearing condition. There were some differences in the ND values for the same bearing condition,
which can be attributed to the noise effect in the acceleration records. The differences, however,
do not significantly change the classification, since the comparative values of ND in one case still
unmistakably provide an indicator on the bearing condition of that pier.
A similar procedure is implemented for DWT as explained in Equation (18). This time, the cluster
of the identified first level detail components (D1) is obtained from DWT of pier acceleration. By
k-means cluster algorithm, the centroids of the first level detail components noted asUc is defined, as

well as the threshold value of the detail component, U ~ and the ND Uc ; U ~ is computed. The physical
interpretation of this index is explained in Equations (11)–(14). For the case of normal (unlocked)
bearing, the D1 will have higher spikes in the interval of peak acceleration. This will result in higher
value of the centroid of the first level detail components (Uc ). This will result in higher value of the
centroid of the first level detail components (Uc ), and positive numerator in Equation (18) and the ND
value. On the other hand, when the bearing is locked, there will be no high D1 spikes as explained in
Equations (11)–(14). As the result, the centroid of the first level detail components (Uc ) will be smaller
and closer to the threshold value of the detail component U. ~ In this case, the numerator in Equation
(18) and ND value will become negative.
Figure 17 shows the classification results from all 32 possible cases of bearing condition obtained from
the first level detail components (D1). The tables list the ND values with respect to the threshold values set
to be 12 times the standard deviation (σ). The figure clearly shows that ND values for piers with movable
bearings were all classified as positive values smaller than one. Whereas, the piers with locked bearings
were classified as negative with an absolute value smaller than one. Similar patterns were observed and
22 D. M. SIRINGORINGO AND Y. FUJINO

Figure 17. Classification of isolator bearing condition for the FE simulations based on k-means clustering method of the first level
detail component (D1) obtained from DWT of accelerations. (a) Case of no locked bearing and only one pier with locked bearing, (b)
two piers have locked bearings, (c) three piers have locked bearings, (d) four piers or more have locked bearings. Note: Pij in the axis
label denotes the pier number with locked bearing.

Figure 18. Flowchart of analysis for evaluation of bearing condition using CWT, DWT and k-means clustering classification directly
from seismic responses of bridge girder and piers.

confirmed by numerical MDOF simulation. The results demonstrate that isolation bearing condition can
be characterized correctly as either locked or movable based on the comparative value of ND for each
case. This also confirms the classification obtained previously using instantaneous frequencies obtained
via CWT. To provide a better explanation and clarification on the detection using the proposed
technique, a flowchart of analysis for the evaluation of bearing malfunction is added in Figure 18.

6. Implementation on Full-Scale Multi-span Continuous Girder Bridge


6.1. Description of the Case-study Bridge and the Monitoring System
The abovementioned procedure is implemented on a continuous multi-span isolated girder bridge using
seismic response obtained from the long-term seismic monitoring system. The object bridge is Katsuta
JOURNAL OF EARTHQUAKE ENGINEERING 23

Figure 19. (a) Photo of Katsuta Viaduct, (b) Wireless sensor node place on the pier cap, (c) girder sensor node placed on the bottom
flange of the girder beam.

Viaduct located in the suburb of Mito City, Ibaraki Prefecture, east of Japan (Figure 19). The bridge
connects Shin-Nakagawa cable-stayed bridge that crosses the Nakagawa river on the south and the
Katsuta highway on the north. It consists of two separated parts each supporting two traffic lanes. The
upstream part supports the traffic toward Katsuta side, and the downstream part supports the traffic
toward the Shin-Nakagawa bridge. The bridge girder is continuously supported by 12 piers and an
abutment at the north end. The piers are made of reinforced concrete, supported by pile foundation, and
have various heights from 8.9 m to 11.18 m, width 5.5 m and thickness 1.7 m (Figure 20). The total span
length is 381.8 m and width 10.9 m each. The bridge girder is composed of composite steel with an open
cross-section and lower plan bracing. The steel girder beam is connected by shear studs to the reinforced
concrete slab that supports the bridge deck. Elastomeric type of bearing in a form of natural rubber
bearing (NRB) is used on all piers except on the north end and south end piers where the sliding bearings
are used. Each pier uses five units of NRB with dimensions: 52 × 52 × 15.5 cm.

Figure 20. Schematic figure of Katsuta Viaduct and its dimension, (a) layout of the viaduct and the wireless sensor system, (b) layout
and dimension of the pier, (c) description of the natural rubber bearing.
24 D. M. SIRINGORINGO AND Y. FUJINO

The upstream part of the bridge was monitored continuously by a wireless sensor network
consisting of 20 triaxial accelerometers. The sensors were placed on the girder (7 nodes), pier cap
(10 nodes) and on the ground (3 nodes). The pier sensors were placed on the top of pier caps at the
same level as the isolation bearing (Figure 19b). The girder sensors were placed on the side of the
girder right above the isolation bearing (Figure 19c). The sensor arrangement is intended, among
others, to evaluate the performance of isolators during an earthquake by comparing the response of
pier and girder. All sensors are arranged such that the x-axis measuring the bridge transversal
vibration, y-axis measuring the bridge longitudinal vibration and z-axis for the vertical vibration.
Detailed information on the sensors and the wireless monitoring system is available elsewhere (Suzuki
et al. 2016).
The monitoring system was installed in summer 2017 and it has been recording seismic responses
continuously from August 2017 until 2020 (Siringoringo et al. 2019). The sensors are powered by 2D
batteries that need to be replaced every 6 months. Between August 2017 and June 2020, the system has
successfully recorded more than 50 seismic events, comprising near-field and far-field earthquakes
where majority of them are of small and moderate amplitudes. Some of the representative earthquakes
are listed in Table 3. The list covers the relatively large earthquakes with PGA > 100 cm/s2 (the first
four earthquakes), the moderate ones (10 cm/s2 < PGA < 100 cm/s2), and several more frequent small
earthquakes (PGA < 10 cm/s2).

6.2. Analyses of Seismic Records by Continuous Wavelet Transform


Response analysis presented in this paper covers only the upstream part of the bridge because only that
part of the bridge has the monitoring system. Also, despite the fact that sensors and recorded
responses are in three directions, the paper focuses analysis for response in longitudinal (bridge
axis) only, because responses in this direction are significantly affected by the behavior of isolation
bearings as shown in the simulations above. It should be mentioned that in some earthquakes, some
sensors failed to function properly because of the problem in power or signal transmission.
Fortunately, due to the redundant arrangement of two sensors on each pier, acceleration of the pier
can be obtained by at least one of the available sensors. Due to space limitation, only analyses from
three seismic events, two relatively large and one small earthquakes, are presented in detail in this
section, while the rest will be given only in summary.
Figure 21 shows the time-histories and a time-frequency map of acceleration from July 17, 2018
earthquake (Equation 1). On this earthquake, sensor on pier 35 (Ch6) failed to record accelerations,
therefore only accelerations from the other 4 piers are shown. Equation 1 is a near-field earthquake
with an epicenter distance about 15 km from the bridge site, and this is shown by a distinct pulse on

Table 3. Representative examples of recorded seismic events by the wireless sensor network system at Katsuta Bridge from 2018 to
2020 (Note: the earthquakes are listed according to the input amplitude from the largest to smallest).
Earthquake [EQ] No. Date [mm/dd/yy] Epicenter [km] Depth [km] Magnitude [M] Distance [km] Peak Ground Acc. [cm/s2]
EQ1 7/17/2018 36.43N,140.7E 52 M4.8 15 143.70
EQ2 6/17/2019 36.50N,140.6E 80 M5.1 16 137.40
EQ3 3/30/2018 36.44N,140.6E 56 M5.1 11 121.80
EQ4 6/4/2020 36.4N, 140.7E 50 M4.7 14 116.60
EQ5 1/21/2020 36.4N, 140.7E 50 M4.3 14 61.38
EQ6 4/12/2020 36.2N,140.0 E 50 M5.1 53 42.48
EQ7 6/1/2020 36.2N,140.4E 100 M5.3 22 31.64
EQ8 1/14/2020 36.1N,139.9E 50 M5.0 66 26.78
EQ9 9/5/2018 36.4N,141.3E 60 M5.5 72 25.02
EQ10 2/1/2020 36.0N,140.1E 70 M5.3 57 23.57
EQ11 8/4/2019 37.7N,141.7E 50 M6.4 181 16.03
EQ12 5/17/2018 36.35N,140.6E 52 M5.3 75 10.40
EQ13 6/18/2019 38.6N,139.5E 10 M6.8 266 9.51
EQ14 7/7/2018 35.3N,140.6E 70 M6.0 118 7.98
EQ15 4/18/2020 27.2N,140.7E 50 M6.9 1018 2.48
JOURNAL OF EARTHQUAKE ENGINEERING 25

Figure 21. Recorded longitudinal seismic responses of Katsuta viaduct due to EQ1 (July 17th, 2018): accelerations, time-frequency
map of acceleration and instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map of
accelerations: (a) Ch-10 (girder), (b) Ch-13 (girder), (c) Ch-11 pier no.38, (d) Ch-14 pier no.41, (e) Ch-17 pier no.44, and (f) Ch-20
abutment A2.

the accelerations which are commonly observed from near-field seismic records. Note that although
there are seven sensor nodes on the girder, the time and frequency characteristics of the longitudinal
accelerations are very similar among the nodes since the girder is a rigid continuous span, two of them
Ch10 and Ch13 (above P38 and P41) are shown by the Figure 21a and b, respectively. Acceleration
from both sensors exhibits the same characteristics of time-frequency response. The girder long­
itudinal acceleration is dominated by a single frequency component as shown by the IF around
1.56 Hz, which is close to the frequency of girder longitudinal mode. This condition remains the
same without significant variation throughout the entire excitation.
On the other hand, pier accelerations reveal different characteristics. During the largest interval of
excitation between t = 10 s and 25 s, the instantaneous frequencies vary between 4 and 8 Hz on pier 38
(Ch11), pier 41 (Ch14), pier 44 (Ch-17) and 6–10 Hz for abutment A2 (Ch-20). There are two distinct
peaks in the spectra for piers at 4.5 Hz and between 6 and 8 Hz. These are the frequency range of the
pier first and second modes. After t = 25 s, the instantaneous frequencies drop to around 1.56 Hz,
which is close to the frequency of girder longitudinal mode. The results demonstrate that during peak
excitation, the pier and girder responses were dominated by their own respective frequency suggesting
that they exist as uncoupled system. The characteristics of time-frequency maps and instantaneous
frequencies are consistent with the results of simulations by MDOF lumped-mass and finite element
systems for the moveable bearing case. In this condition, the piers are isolated from the girder and the
high-frequency content of the pier is not transferred to the girder, which is expected from
a functioning isolator bearing.
A similar condition was observed on the responses due to three other relatively large earthquakes
(June 17, 2017 (EQ2), March 30, 2018 (EQ3) and April 6, 2020 (EQ4). The girder and piers responses
during EQ3 are shown in detail in Figure 22. One can clearly see from Figure 21a that girder
26 D. M. SIRINGORINGO AND Y. FUJINO

Figure 22. Recorded longitudinal seismic responses of Katsuta viaduct due to EQ3 (March 30th, 2018): accelerations, time-frequency
map of acceleration and instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map of
accelerations: (a) Ch-13 (girder), (b) Ch-6 pier no.35, (c) Ch-11 pier no.38, (d) Ch-14 pier no.41, (e) Ch-17 pier no.44, and (f) Ch-20
abutment A2.

accelerations were dominated by frequency content frequency around 1.56 Hz. The time-frequency
map shows a distinct peak around this frequency at the time t = 10 s, which is the peak of excitation.
The instantaneous frequencies derived from the ridge of the time-frequency map exhibits a relatively
constant value around 1.56 Hz throughout the response. Meanwhile, the piers and abutment exhibit
two frequency contents: the higher frequency range during the period of peak excitation and low-
frequency component near the end of the response. During the interval of largest excitation (t = 10–­
25 s), the instantaneous frequencies vary between 4 and 8 Hz on pier 35 (Ch6), pier 38 (Ch11), pier 41
(Ch14), pier 44 (Ch-17) and 6–10 Hz abutment A2 (Ch-20). There are two distinct peaks in the spectra
at 4.5 Hz and between 6 and 8 Hz.
All the results indicate that isolator bearings at all instrumented piers were functioning properly
during the mentioned large earthquakes. The time-frequency map and IF suggest that the piers are
isolated from the girder at the time of peak excitation, which is an expected behavior from
a functioning isolation bearing. It should be mentioned that although the general trend of values for
the pier at the three earthquakes is similar with the finite element simulations for movable bearing
case, the instantaneous frequencies during the time of peak excitation were not as smooth and at the
constant frequencies as they were in simulations. There were several frequency peaks within the
frequency range of pier’s first and second frequencies. This can be attributed to the fact that although
the piers are designed to be the same, there could be some variations on the exact behavior depending
on the local soil condition. Nevertheless, the general trend of time-frequency and IF characteristics
have indicated that the isolators have functioned properly, and that the girder was fully isolated during
the above large earthquakes.
Next, the typical responses for small earthquake (PGA < 10 cm/s2) are discussed. Figure 23 shows
the response for a small earthquake on April 18, 2020 (EQ15) with PGA 2.48 cm/s2. The figure shows
JOURNAL OF EARTHQUAKE ENGINEERING 27

Figure 23. Recorded longitudinal seismic responses of Katsuta viaduct due to EQ15 (April 18, 2020): accelerations, time-frequency
map of acceleration and instantaneous frequency (noted by red lines) obtained from the ridges of time-frequency map of
accelerations: (a) Ch-13 (girder), (b) Ch-6 pier no.35, (c) Ch-11 pier no.38, (d) Ch-14 pier no.41, (e) Ch-17 pier no.44, and (f) Ch-20
abutment A2.

that similar to the case of a large earthquake, girder longitudinal acceleration is dominated by a single
frequency component throughout the entire excitation. The dominating single frequency, however, is
around 1.8 Hz or slightly higher than the girder longitudinal mode identified from a large earthquake.
A remarkable difference is shown by the time-frequency map of pier accelerations. Contrary to the
responses due to large earthquake, longitudinal accelerations of all instrumented piers were dominated
by a single frequency component around 1.8 Hz, which is the same range of frequency of the girder
response. This condition was observed during the peak acceleration at the time t = 10–30 s until the
end of excitation. There was no change in the dominant frequency of pier acceleration like the one
observed during large earthquakes. The identical dominating frequency of accelerations indicates that
both girder and piers move together as a coupled system as was confirmed in the MDOF simulation
and finite element model of locked bearing. This can be explained as a consequence of small excitation.
As often seen in the case of seismic excitation, an adequate level of excitation amplitude is required for
the isolation system to start engaging in the isolation mechanisms and reached the critical stage of
yielding. Whereas for small level of excitation, the isolation system is still at rest or at the beginning of
the initial state. Therefore, the fact that piers and girder move together as a coupled system should not
be comprehended as an alarming case of locked bearing if it occurs under small excitation amplitude.

6.3. Analyses of Seismic Records by Discrete Wavelet Transform


Figure 24 shows the first level of detail components computed via DWT of accelerations for the three
earthquakes (EQ1, EQ3 and EQ15). The D1 values were computed for all pier accelerations and
normalized by the procedure described in Equation (17). In the first row (EQ1), we can note that
spikes of D1 appear on all pier accelerations at the time interval between 10 and 15 s, which
28 D. M. SIRINGORINGO AND Y. FUJINO

Figure 24. First level detail components (D1) obtained from DWT of accelerations of Katsuta viaduct for Ch-13 (girder), Ch-11 pier
no.38, Ch-14 pier no.41, Ch-17 pier no.44, and Ch-20 abutment A2 due to (a) EQ1 (July 17th,2018), (b) EQ3 (March 30th, 2018) and (c)
EQ15 (April 18, 2020).

corresponds to the time of peak excitation. Similar spikes also appear on the D1 component of pier
accelerations due to EQ3 (second row). In both earthquakes, the normalized D1 values exceed the
threshold (12 times the standard deviation (σ)), indicating the possibility of discontinuity in accelera­
tion responses caused by the transitional condition of isolator from elastic to yielding. Note that the
amplitude of D1 corresponds well with the amplitude of piers and girder accelerations.
For the small earthquakes, EQ15 shown here as the typical example, time histories of normalized D1
values do not consist of obvious spikes within the time interval of peak excitation. Moreover, the
amplitude of normalized D1 values is relatively smaller compared to that of larger earthquakes and do
not exceed the threshold value. This indicates a smaller possibility of discontinuity in acceleration
responses due to the transitional condition of the isolator from elastic to yielding. The results are
consistent with the observation of IF from CWT that show no changes in the frequency component of
pier acceleration (Figure 23). Therefore, it can be inferred from both CWT and DWT results that in
the case of small excitation, the girder and piers move together, and the isolation bearings have not
engaged in the isolation mechanism or reached the critical stage for yielding.

6.4. Implementation of k-means Clustering for Characterization of Isolation Bearing


Condition from Seismic Monitoring Data
The results of CWT and DWT analysis for 15 earthquakes were summarized using the k-means
clustering method to characterize the isolation bearing condition. To provide classification of the
cluster as locked or movable bearing, the distance between clusters of identified instantaneous
frequencies from CWT and the normalized D1 values of DWT are evaluated following the procedure
explained in Equations (18) and (19), respectively. Figure 24 illustrates the results of cluster classifica­
tion using instantaneous frequencies obtained via CWT. The abscissa in each case describes the
earthquake number listed from the largest to smallest PGA, and the ordinate describes the identified
ND values that are computed in Equation (18). The results show that for all large and moderate
earthquakes, the ND values are closer to one indicating piers with normal (movable) bearings. This is
evident by the ND values > 0.5 for all piers in the EQ1 until EQ11. For the relatively smaller earthquake
(EQ12–EQ15) where PGA � 10 cm/s2, the results show ND values less than 0.5 on some piers. In the
two smallest earthquakes, the ND values are even closer to zero on all piers indicating the possibility of
the locked bearing.
JOURNAL OF EARTHQUAKE ENGINEERING 29

Figure 25. Classification of isolator bearing condition based on k-means clustering method of the instantaneous frequencies
obtained from CWT of accelerations on pier P35-P44 and abutment A2 of 15 earthquakes. ND denotes the identified normalized
distance (ND values closer to one means moveable bearing, closer to zero means locked bearing).

Figure 26. Classification of isolator bearing condition based on k-means clustering method of the first level detail component (D1)
obtained from DWT of accelerations on pier P35-P44 and abutment A2 of 15 earthquakes. ND denotes the identified normalized
distance (ND > 0 means moveable bearing, ND < 0 means locked bearing).

Figure 26 describes the results of classification for the first level detail component D1 computed
using Equation (18) for the 15 earthquakes. The tables show that for the large and most of the
moderate earthquakes (EQ1–EQ10), the ND values for piers were classified as positive value smaller
than one indicating the moveable bearing condition. Meanwhile, the ND values for smaller earth­
quakes (EQ11–EQ15) were classified as negative with an absolute value smaller than one indicating the
possibility of locked bearings. These results are consistent with the classification obtained from CWT
(Figure 25). The tables clearly show that for large earthquakes, the isolation bearing has functioned
properly as the movable bearing. For smaller earthquakes, although the results indicate the possibility
of locked bearing, they can be interpreted as the condition where the isolation mechanism has not
been engaged or initiated due to the small amplitude of excitation.

7. Conclusions
This paper describes wavelet-based techniques for detecting a malfunction of isolation bearing directly
from seismic records of the structure. The techniques consist of (1) CWT to identify IF based on which
stiffness changes associated with isolation bearing condition can be traced, and (2) detail components
of DWT based on which discontinuity of acceleration response associated with isolation state can be
identified. To facilitate direct detection based on both CWT and DWT results, classification of bearing
condition was performed using k-means cluster classification method. Results of simulation using
analytical multi-degree of freedom model and three-dimensional finite element model of continuous
30 D. M. SIRINGORINGO AND Y. FUJINO

multi-span isolated bridge have shown the effectiveness and accuracy of the proposed technique to
characterize the behavior of isolation bearing and detect any changes related to bearing malfunction
such as locking phenomenon directly from the seismic record of the bridge without the need of
structural model or reference state. The following are the conclusions of the study.

(1) For a normal (moveable) isolation bearing, the IF of pier acceleration during the peak excita­
tion would have a higher frequency than the girder acceleration. This condition indicates that
girder and pier behave separately as an uncoupled system. In such a condition, when the
isolation has reached the yielding state, high and distinct spikes representing signal disconti­
nuity would appear on the detail components obtained from DWT of the pier acceleration.
(2) On contrary, for a locked bearing, the IF of pier acceleration during the peak excitation would
have a similar frequency with the girder acceleration, indicating that the girder and pier move
together as a coupled system. Because of this condition, the relative displacement between the
girder and pier would be minimum and the isolation would not reach the yielding state.
Consequently, high and distinct spikes representing signal discontinuity would not appear
on the detail components of the DWT of the pier acceleration. These structural features were
used as an indicator to characterize the condition of isolation bearing directly from
accelerations.
(3) The technique was implemented on a continuous multi-span girder bridge using seismic
records from 15 earthquakes with various amplitude levels obtained from the long-term
seismic monitoring system. Results of analyses reveal that isolation bearings have functioned
properly without observed locked bearing during large (PGA > 10 cm/s2) and moderate
(10 cm/s2 < PGA < 100 cm/s2) earthquakes. Possible condition of locked bearing was observed
during small earthquakes (PGA < 10 cm/s2). This was interpreted as a condition where the
isolation mechanism has not been engaged or initiated due to the small amplitude of excitation.

The study has demonstrated that the detection techniques described in this paper can be used
effectively to characterize isolation bearing condition directly from seismic monitoring records. This
can facilitate on-line structural assessment via an SHM system using seismic records.

Acknowledgement
This study is supported by the JSPS Grant-in-Aid Kakenhi C No. 18K04320 and partially by Taisei Foundation for the
first author. These supports are gratefully acknowledged. We also gratefully acknowledge the supports from the Central
Nippon Expressway Co. Ltd. (NEXCO) throughout the bridge instrumentation process. Opinions and findings in this
paper are of the authors and not of the institutions named above.

Funding
This study is supported by the JSPS Grant-in-Aid Kakenhi C No. 18K04320 and partially by Taisei Foundation
Fellowship to the first author.

ORCID
Dionysius M. Siringoringo http://orcid.org/0000-0003-2267-0673
Yozo Fujino http://orcid.org/0000-0001-8993-3779

References
Aguirre, D. A., C. A. Gaviria, and L. A. Montejo. 2013. Wavelet-based damage detection in reinforced concrete
structures subjected to seismic excitations. Journal of Earthquake Engineering 17 (8): 1103–25. doi: 10.1080/
13632469.2013.804467.
JOURNAL OF EARTHQUAKE ENGINEERING 31

Arthur, D., and S. Vassilvitskii. 2007. K-means++: The advantages of careful seeding. Proceedings of the Eighteenth
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ‘07, Society for Industrial and Applied Mathematics,
1027–35. Philadelphia, PA.
Chaudhary, M. T. A., M. Abé, and Y. Fujino. 2002b. Investigation of atypical seismic response of a base-isolated bridge.
Engineering Structures 24 (7): 945–53. doi: 10.1016/S0141-0296(02)00013-5.
Chaudhary, M. T. A., M. Abe, and Y. Fujino. 2002a. Role of structural details in altering the expected seismic response of
base-isolated bridges. Mechanical Systems and Signal Processing 16 (2–3): 413–28. doi: 10.1006/mssp.2002.1480.
Chawla, S., and G. Aristides. 2013. K-means–: A unified approach to clustering and outlier detection.” Proceedings of the
2013 SIAM International Conference on Data Mining, 189–97. Society for Industrial and Applied Mathematics. doi:
10.1137/1.9781611972832.21.
Daubechies, Ingrid. 1992. Ten Lectures on Wavelets, CBMS-NSF regional conference series in applied mathematics. Vol.
61. 33 (1995): 37. Philadelphia, PA: Society for Industrial and Applied Mathematics (SIAM).
Derkevorkian, A., Masri, S. F., Fujino, Y., & Siringoringo, D. M. (2014). Development and validation of nonlinear
computational models of dispersed structures under strong earthquake excitation. Earthquake Engineering &
Structural Dynamics 43 (7): 1089–105. doi:10.1002/eqe.v43.7.
Fujino, Y., D. M. Siringoringo, and M. Abe. 2016. Japan’s experience on long-span bridges monitoring. Structural
Monitoring and Maintenance 3 (3): 233. doi: 10.12989/smm.2016.3.3.233.
Fujino, Y., D. M. Siringoringo, M. Kikuchi, K. Kasai, and T. Kashima. 2019. Seismic monitoring of seismically isolated
bridges and buildings. Seismic Structural Health Monitoring: From Theory to Successful Applications 407–47. doi:
10.1007/978-3-030-13976-6_17.
Grossmann, A., and J. Morlet. 1984. Decomposition of hardy functions into square integrable wavelets of constant shape.
SIAM Journal on Mathematical Analysis 15 (4): 723–36. doi: 10.1137/0515056.
Hera, A., and Z. Hou. 2004. Application of wavelet approach for ASCE structural health monitoring benchmark studies.
ASCE Journal of Engineering Mechanics 130 (1): 96–104. doi: 10.1061/(ASCE)0733-9399(2004)130:1(96).
Hou, Z., A. Hera, and A. Shinde. 2006. Wavelet–based structural health monitoring of earthquake excited structures.
Computer Aided Civil and Infrastructure Engineering 21 (4): 268–79. doi: 10.1111/j.1467-8667.2006.00434.x.
Kawashima, K., and I. Buckle. 2013. Structural performance of bridges in the Tohoku-oki earthquake. Earthquake
Spectra 29 (1_suppl): 315–38. doi: 10.1193/1.4000129.
Kitahara, T., Y. Kajita, and Y. Kitane. 2012. Investigation on the damage cause of the bridge rubber bearings in the 2011
off the Pacific coast of Tohoku Earthquake. Proceedings of 15th World Conference on Earthquake Engineering,
Lisbon.
Masing, G. 1926. Eigenspannungen und verfestigung beim messing (Self stretching and hardening for brass).
Proceedings of the Second International Congress for Applied Mechanics, 332–35. Zurich, Switzerland. (in German).
Michel, C., and P. Gueguen. 2010. Time-frequency analysis of small frequency variations in civil engineering structures
under weak and strong motions using a reassignment method. Structural Health Monitoring 9: 159–71. doi: 10.1177/
1475921709352146.
Pan, T. C., and C. L. Lee. 2002. Application of wavelet theory to identify yielding in seismic response of bi-linear
structures. Earthquake Engineering & Structural Dynamics 31 (2): 379–98. doi: 10.1002/eqe.113.
SI (Computers and Structures Inc). 2013. SAP2000 v16 integrated finite element analysis and design of structures.
Berkeley: CSI.
Siringoringo, D. M., and Y. Fujino. 2015. Long-term seismic monitoring of base-isolated building with emphasis on
serviceability assessment. Earthquake Engineering & Structural Dynamics 44 (4): 637–55. https://doi.org/10.1002/eqe.
2538.
Siringoringo, D. M., Y. Fujino, M. Suzuki, and M. Vaibhav. 2019. Seismic monitoring of cable-stayed bridge using
wireless sensor network. Proceeding of Workshop on Structural Control & Health Monitoring, Stanford, CA. doi:
10.12783/shm2019/32394
Staszewski, W. J., and D. M. Wallace. 2014. Wavelet-based frequency response function for time-variant systems—An
exploratory study. Mechanical Systems and Signal Processing 47 (1–2): 35–49. doi: 10.1016/j.ymssp.2013.03.011.
Suzuki, M., K. Jinno, Y. Tashiro, Y. Katsumata, C. H. Liao, T. Nagayama, and H. Morikawa. 2016. Development and field
experiment of routing-free multi-hop wireless sensor network for structural monitoring. Proceedings of the
International Conference on Smart Infrastructure and Construction, 179–84, Cambridge University, Cambridge,
UK, January 27–June 29.
Tchamitchian, P., and B. Torrésani, B. 1990. Ridge and skeleton extraction from the wavelet transform. Proceeding of
CBMS-NSF Conference on Wavelets and their Applications, Lowell, MA, USA, June 1990.
Todorovska, M. I., and M. D. Trifunac. 2007. Earthquake damage detection in the Imperial County services building I:
The data and time–frequency analysis. Soil Dynamics and Earthquake Engineering 27 (6): 564–76. doi: 10.1016/j.
soildyn.2006.10.005.
Unjoh, S. 2014. Menshin (Seismic isolation) Bridges in Japan. Technical Note of PWRI Japan No. 4288.
Yan, B. F., A. Miyamoto, and E. Brühwiler. 2006. Wavelet transform-based modal parameter identification considering
uncertainty. Journal of Sound and Vibration 291 (1–2): 285–301. doi: 10.1016/j.jsv.2005.06.005.

You might also like