Engineering Structures 29 (2007) 2001–2017

www.elsevier.com/locate/engstruct

Dynamic characteristics of a curved cable-stayed bridge identified from

strong motion records
Dionysius M. Siringoringo ∗ , Yozo Fujino
Department of Civil Engineering, University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan

Received 26 April 2006; received in revised form 5 September 2006; accepted 16 October 2006
Available online 6 December 2006

Abstract

An assessment of dynamic characteristics of the 455 m Katsushika–Harp curved cable-stayed bridge is presented. Dynamics characteristics
such as natural frequencies, mode shapes and modal damping ratios are obtained from seismic response of the bridge by employing a time-domain
multi-input multi-output (MIMO) system identification (SI) technique. The technique makes use of base motions and superstructure accelerations
as pairs of inputs–outputs to realize the coefficients of state-space system matrices. The SI results indicate the occurrence of many closely spaced
modal frequencies with spatially complicated mode shapes. Fourteen global modes in the ranges of 0.45–2.5 Hz were identified, in which the
girder motion dominated most of the modes. The tower modes were associated with girder modes and were characterized by the lowly-damped
motion. Using identification results from six earthquakes, the effects of earthquake amplitude on modal damping ratios were observed.
c 2006 Elsevier Ltd. All rights reserved.

Keywords: Curved cable-stayed bridge; Instrumented bridge; Seismic response; MIMO system identification; Katsushika–Harp bridge

1. Introduction approach using transfer function to minimize the objective

The availability of multi-channel permanent sensors allows
regular full-scale dynamic tests that are essential for continuous
monitoring of a bridge. In a seismically active region,
such as Japan, this instrumentation provides an opportunity
to use system identification (SI) techniques to explain
bridge performance during earthquakes. By employing the SI
technique, it is also possible to monitor any changes in the
bridge behavior without the presence of visually observable
damage. In the context of bridge monitoring, this excellent
opportunity is beneficial to evaluate the adequacy of bridge
seismic design code [1].
The general approach of an earthquake-induced SI is to
use the input–output relation to recreate structural models
that are capable of reproducing the actual responses. In one
early study Beck [2] employed the output-error minimization
method for a linear, time-invariant structural system with
classical damping. McVerry [3] proposed a frequency domain
∗ Corresponding author. Tel.: +81 03 5841 6097; fax: +81 03 5841 7454.
E-mail addresses: dion@bridge.t.u-tokyo.ac.jp (D.M. Siringoringo),
fujino@bridge.t.u-tokyo.ac.jp (Y. Fujino).
function of output error. Chaudhary et al. [4] improved it,
for a more general problem of non-classical damping that
includes the structural model in addition to the modal model.
This method, while powerful and significantly insightful,
requires prior information of structural properties that are
typically unavailable and difficult to obtain, especially for large
and complex structures such as cable-stayed bridges. Most
conventional SI techniques were developed in the frequency
domain due to the common practice of using frequency analyzer
for data acquisition. These approaches offer advantages in
incorporating soil–structure interaction into analysis, but often
suffer from damping estimation especially when closely spaced
modes are present.
Compared to cable-stayed bridges with straight girder, the
curved cable-stayed bridges are relatively few. Examples of
well-known curved cable-stayed bridges are the La Arena
Viaduct in Spain [5], the Safti Link Bridge in Singapore [6],
the Rhine Bridge near Schahausen, Switzerland [7] and the
twin curved cable-stayed bridge at the Malpensa airport in
Milan, Italy [8]. In the analysis of a cable-stayed bridge under
seismic action, the aspects of three dimensionality, multi-modal
contribution, multiple-support excitations and modal coupling

c 2006 Elsevier Ltd. All rights reserved.

0141-0296/$ - see front matter

doi:10.1016/j.engstruct.2006.10.009
2002 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017
Fig. 1. View of the Katsushika–Harp bridge.
are of great importance [9]. Accordingly, the SI techniques
should be developed to provide accurate information that
accommodate observation of such aspects.
Challenges in the SI technique applied to a curved cable-
stayed bridge come from its nature as a large and continu-
ous structure, which makes it difficult to be represented with a
lumped-mass model. Moreover, owing to its complexity, many
modes are closely spaced in frequency. And due to its long span,
spatial variability of ground motions may be significant. For
this latter reason, the SI technique needs to employ a scheme
of multiple-input excitations. More recent works on the SI of
long-span bridges under seismic excitation have attempted to
overcome the problems by employing time-domain algorithms
that include the effect of multi-input [10–12].
The focus of the work described in this paper is twofold: (1)
providing a systematic procedure for SI of the Katsushika–Harp
curved cable-stayed bridge using seismic records, and (2)
evaluating the dynamic characteristics of the bridge based on
the SI results. The SI method adopted here is based on the
input–output mapping technique of the System Realization
using Information Matrix (SRIM) [13,14]. Organization of this
paper is such that a brief account of the bridge is presented first,
followed by explanation of the SI methodology, instrumentation
and seismic records. Afterwards, the results of identification are
explained, including the comparisons with an analytical finite
element model, previous tests using ambient vibration, and
forced vibration. Next the identification of tower modes and the
observation from six earthquakes are presented. Conclusions of
the study are summarized at the end.
2. Description of Katsushika–Harp bridge
The Katsushika–Harp Bridge (Fig. 1) is a curved cable-
stayed bridge located in the Katsushika area of Tokyo
Metropolitan city. The bridge crosses over an estuarial area
between the Arakawa and Nakagawa Rivers. Its girder is made
of steel with the width of 23.5 m and the total length of 455 m,
consisting of 220 m main span, and three side spans of 40.5, 134
and 60.5 m (Figs. 2 and 3). The bridge’s unique asymmetrical
S-shaped girder is composed by two curves with a radius of
334 m and 270 m each. The bridge has two rectangular towers
made of steel; the main tower (65 m in height and 3 m in width)
is located in the middle of the curve and the smaller one (13.8 m
in height and 2.5 m in width) is located at the end of highway
approach to the Hirai Bridge. The bridge stays spring from the
main tower (17 cables) and the right tower (7 cables) to support
the curved girder. Construction of the bridge was completed in
1987, and since then it became the part of the Shutoo Chuo
Loop line in the Tokyo Metropolitan expressway.
For the purpose of monitoring, 32 channels of sensors were
permanently installed on the bridge. The sensors consist of
29 accelerometers deployed at 12 locations, (Fig. 4) and 3
displacement sensors. For system identification, only responses
from accelerometers were utilized. These accelerometers have
a range of frequency between 0.05 and 35 Hz with an accuracy
of 15 microampere per cm/s2 . Among these accelerometers,
six are located on the substructure (on the pile foundation and
pile caps) and the rest were installed on the superstructure
(towers, pier caps and girder). On the girder, sensors were
installed at eight locations along the girder centerline. These
sensors measure accelerations in vertical, transverse and
longitudinal directions, at a data-sampling rate of 100 Hz. It
should be mentioned that all sensors measure the motion in
directions that coincide with the local coordinates of structure
members (i.e. the x-direction of measurement coincides with
the centerline of the bridge).
3. System identification methodology
In this study, the SRIM-based SI is employed to identify the
complex modal frequencies, damping ratios and mode shapes
of the bridge system. The SI consists of two steps, namely,
the realization of system matrices from correlation of input and
output data, and estimation of modal parameters from identified
system matrices. The main advantage of this technique is
its ability to provide a systematic way to identify modal
parameters from multiple-input and multiple-output earthquake
data without prior knowledge of structural properties or models.
Moreover, it allows the identification of a classically as well as
a non-classically damped structural system, and thus eliminates
the need of trial and error in updating the modal parameters to
achieve reasonable agreement with the recorded data.
3.1. Basic formulation of earthquake-induced vibration
The identification starts from a basic equation of motion of
N degree-of-freedom (DOF) linear, time invariant, viscously
damped system subjected to earthquake excitation ü g (t), in the
spatial coordinate {u(t)} and continuous time (t). The equation
is given as
Mü(t) + Cu̇(t) + Ku(t) = −Wüg (t). (1)
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2003

Fig. 2. Dimension and general configuration of the Katsushika–Harp bridge.

(a) Main tower. (b) Right tower.

Fig. 3. Dimensions of girder and tower.

Fig. 4. Sensor positions and measuring directions of the Katsushika Bridge (note: measuring direction is the local coordinate of the structural member).
2004 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017

The quantities M, C and K ∈ R N ×N denote mass, damping

 
D
and stiffness matrices respectively, while the W ∈ R N ×q 
 RB D 

denotes the continuous time input matrix, with q being the [Tp ] = 
 RAB RB D .

(6)
dimension of input vector üg (t). This equation can be expressed .. .. ..
. . .
 
 ··· 
as a finite dimensional, discrete-time, linear, time-invariant,
RAp−2 B RAp−3 B RAp−4 B · · · D
state variable dynamical system:
Matrix [Op ] ∈ R pm×2N is commonly called the
x(k + 1) = Ax(k) + Bz(k) (2a)
observability matrix. The matrix [Tp ] ∈ R pm× pq is a
y(k) = Rx(k) + Dz(k). (2b) generalized Toeplitz matrix composed by the system Markov
parameters (i.e. RAk−1 B). It should be mentioned that time
The vector x(k) = [u(k) u̇(k)]T in the latter expression discretization error that might appear may be made negligible
is a 2n × 1 state vector consisting of displacement and by using a sufficiently small time step. Since most earthquake
velocity. Vector z(k) consists of input acceleration recorded at excitation is sampled at a frequency of 100 Hz, which is
q channels, and y(k) denotes structural acceleration responses sufficiently small as compared to the most natural frequencies
recorded from m output channels. of civil structures (i.e. far below 50 Hz), discretization error in
The integer k = 0, 1, 2, . . . l denotes the time-step number, this case may be considered negligible.
i.e. x(k + 1) = x(k(1t) + 1t), with 1t being the time interval.
The quantities of A ∈ R2N ×2N , B ∈ R2N ×q , R ∈ Rm×2N and 3.2. System realization using information matrix
D ∈ Rm×q are the discrete representation of system matrices.
Their corresponding continuous system matrices are given as The kernel of the SRIM algorithm is identification of
Z 1t observability matrix [Op ]. Let [Op ](m+1 : pm, :) be defined as
a matrix consisting of the last ( p−1)m rows (from the (m +1)th

Ac 1t [Ac ]τ 0
A=e , B= e dτ Bc ,
0
0 row until the ( pm)th row) and all columns of [Op ]. Similarly,
let the [Op ](1 : ( p − 1)m, :) be defined as a matrix consisting
−1 −1
C , D = LM−1 .
 
R = L −M K −M (3)
of the first ( p − 1)m rows and all columns of [Op ]. The above
definitions are rewritten in matrix form as follows:
Matrix [L] ∈ Rm×N is a transformation matrix that connects
RA
 
the position of system degree-of-freedom with the measured
2
output responses. The quantities Ac ∈ R2N ×2N and Bc ∈  RA 
R2N ×q are the system matrix in continuous time and the input 3 
 
[Op ](m + 1 : pm, :) =  RA  and

influence matrix of order 2N × q, respectively, which are both  . 
 .. 
defined as:
RAp−1
    (7)
0 I 0 
R

Ac = Bc = . (4)
−M−1 K −M−1 C −I  RA 
2
 
[Op ](1 : ( p − 1)m, :) =  RA .
 
The system matrices A ∈ R2N ×2N , B ∈ R2N ×q , R ∈  ..
 .

Rm×2N , and D ∈ Rm×q are all unknowns and to be

determined from given earthquake input–output data. With RAp−2
several algebraic manipulations, Eq. (2) can be rewritten for Now, following the equality:
various p time spans in a matrix form as
RA
   
R
yp (k) = [Op ]x(k) + [Tp ]zp (k) (5)  RA2   RA 
3 RA2
   
[Op ](m + 1 : pm, :) =  RA = A
   
where the y p (k), z p (k), [Op ] and [Tp ] are defined as follows  . ..
 .. .
  
  
RAp−2
 
y(k) R RAp−1
 
y(k + 1) RA
= [Op ](1 : ( p − 1)m, :)A
 
  (8)
RA2
 
yp (k) =  y(k + 2)
 
[Op ] = 
  
.. ..

the state matrix A can be identified as:
 
. .
   
 
y(k + p − 1) RAp−1 A = [O∗p ](1 : ( p − 1)m, :)[Op ](m + 1 : pm, :) (9)
 z(k) 
where the asterisk (*) denotes the pseudo-inverse matrix.
 z(k + 1)  Following derivation of Eqs. (8) and (9), the choice of integer
zp (k) =  z(k + 2) 
 
p becomes obvious. The integer p should be selected such that
 .. 

.
 matrix [Op ](m + 1 : pm, :) of dimension ( p − 1)m × 2N has
z(k + p − 1) rank larger than or equal to 2N , hence p ≥ 2N
m + 1.
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2005

The procedure to obtain the observability matrix starts by To obtain the solution for matrix [Op ], Eq. (14a) needs
expanding the vector equation in (5), into a matrix equation, i.e., to be factored into three matrices where the first matrix is
the transpose product of the third matrix. The singular value
[Yp ](k) = [Op ]X(k) + [Tp ][Zp ](k) (10) decomposition is a logical choice for this matrix factorization.
where In this decomposition, only the first size ( pm × ( p − 1)m) row
and column of Rhh matrix is decomposed using singular value
X(k) = [x(k) x(k + 1) ··· x(k + s − 1)], decomposition. Hence, the new form of decomposition is:
[Yp ] = [ y p (k) y p (k + 1) · · · y p (k + s − 1) ]
Rhh (:, 1 : ( p − 1)m) = HΣ 2 VT = H2N H0
 
 y(k) y(k + 1) · · · y(k + s − 1) 
y(k + 1) y(k + 2) · · · y(k + s) " T #
.. .. ..
   2
=  .. 
Σ2N 02N xn o V2N 2

. . . .

× T
= V2N Σ2N VT2N (15)
y(k + p − 1) y(k + p) · · · y(k + p + s − 2) 0m o x2N 0m o xn o Vo
[Zp ] = [ z p (k) z p (k + 1) · · · z p (k + s − 1) ] where the dimension of Rhh (:, 1 : ( p−1)m) is ( pm×( p−1)m).
 z(k) z(k + 1) · · · z(k + s − 1) 
 z(k + 1) z(k + 2) · · · z(k + s) The integer n o indicates the number of zero singular values
.. .. .. .

=  .. (11) and also the number of columns in matrix Vo . The integer m o
. . . .
 
z(k + p − 1) z(k + p) · · · z(k + p + s − 2) denotes the numbers of columns in Ho that are orthogonal to
the column in H2N .
To ensure that the rank of [Yp ] and [Zp ] is at least For noisy data, there are no zero singular values, that is,
equal to that of matrix [Op ], a sufficiently large integer s n o = 0. If no singular values are truncated, m o = m is
must be selected. Thus, rank evaluation must be performed obtained. If some singular values are truncated, m o becomes
before selecting the proper value of integer s. Now let us the sum of m. In both conditions of singular values truncation,
define the following quantities: Ryy , Rzz , Rxx , Ryz , Ryx , and the partial decomposition method guaranties that there are at
Rxz as correlations between input (z), output (y), and state (x) least m columns of Ho that are orthogonal to the columns of
respectively, where the subscript indices denote the product of H2N in Eq. (15). Further factorization of Eq. (15) produces,
correlation in such a way that the correlation of Ap and Bp is
defined as: Rhh [:, 1 : ( p − 1)m] = [Op ]R̂xx [Op ]T
1 × [:, 1 : ( p − 1)m] = H2N Σ2N
2
VT2N . (16)
Rab = Ap (k)BTp (k). (12)
s Finally following the identity in Eq. (16), the observability
If the integer s = l − p is sufficiently large with l being matrix can be obtained:
the total length of data and p the data shifts, the quantities in
[Op ] = H2N and
Eq. (12) approximate the expected values in a statistical sense. (17)
To determine the observability matrix [Op ], one can start by R̂xx [Op ] [:, 1 : ( p − 1)m] = Σ2N
T 2
VT2N .
obtaining the data correlations of input–output matrices: Given the observability matrix, the system matrix A can be
identified from Eq. (9). Subsequently, the modal parameters
Rhh = [Op ]R̂xx [Op ]T (13)
of the structural system can be estimated by solving the
where the quantities Rhh and R̂xx are defined as follows eigenvalues problem of matrix A as:

Rhh = Ryy − Ryz R−1 T A8̂ = 3̃8̂. (18)

zz Ryz (14a)
zz Rxz .
T
R̂xx = Rxx − Rxz R−1 (14b)
The quantity Rhh is determined from the output autocorrela-
tion matrix Ryy , the output–input cross-correlation matrix Ryz
and the inverse of the input autocorrelation matrix, which all
are available from input–output data. It should be mentioned
that the quantity of Rhh exists only if the input autocorrela-
tion matrix Rzz is a non-singular matrix. This requirement will
be satisfied if the input is persistent and rich enough for z(k),
with k = 0, 1, . . . , l or in other words the input block matrix
Zp (k) is of full rank (i.e. qp). If we assume that the input and
the state vectors are two uncorrelated quantities, then the cross-
correlation matrix between them Rxz becomes a 2N × q p zero
matrix, and Eq. (14b) thus reduces to R̂xx = Rxx . This condi-
tion can be achieved if the input is a sufficiently long Gaussian
random signal. With this assumption, Eq. (14a) can be solved
to obtain matrix [Op ], given the matrix Rhh computed from
input–output data.
Matrices 3̃ and 8̂ represent the eigenvalues and eigenvec-
tors of matrix A, respectively. The eigenvalues and eigenvectors
can be real or complex, where in the latter case, they appear as
complex conjugate pairs. The eigenvalues λ̃i are actually ex-
pressed in z-domain, and therefore can be related to the modal
characteristics of dynamical system using the following trans-
formation
λi = ln(λ̃i )/1t. (19)
After transformation, the natural frequency (ωi ) and modal
damping ratio (ξi ) can be estimated:
p
ωi = Re(λi )2 + Im(λi )2 , ξi = −Re(λi )/ωi . (20)
The mode shapes matrix in a coordinate system is obtained
by transforming the eigenvectors in z-domain into a coordinate-
domain using the output-transformation matrix R,
8 = R8̂. (21)
2006 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017

Table 1
List of earthquakes used in analysis

Earthquake Magnitude Input acceleration Total length

(Triggered time) (M j) Max-x (cm/s2 ) Max-y (cm/s2 ) Max-z (cm/s2 ) (s)
Feb 20, 1990 (15:53) 6.5 5.50 5.80 2.00 210.5
Feb 2, 1992 (04:04) 5.9 32.60 43.38 17.26 107.00
Oct 23, 2003 (17:57) 6.8 12.80 11.80 6.10 46.75
Oct 23, 2003 (18:35) 6.8 11.40 9.55 4.36 39.75
Jul 23, 2005 (16:35) 6.0 52.61 29.21 6.93 45.00
Aug 16, 2005 (11:48) 7.2 14.86 15.99 4.80 70.25

In summary, the SI technique involves the following 5.1. Spectra of the acceleration responses
procedures: (1) computation of the correlation matrices (Ryy ,
Ryz and Rzz ), using the p sampled data points of output block Fig. 6 shows the Fourier spectra of input accelerations
matrix [Yp ] and input block matrix [Zp ]. (2) Calculation and recorded at the bottom of the main and the right tower. Inputs at
both towers were characterized by low frequency components
then factorization of the information matrix Rhh using singular-
(0–1.6 Hz) in all three directions. The spectra of vertical
value decomposition. The order of the system is determined by
input at both towers reveal similar trend in all frequencies of
examining the singular values of matrix Rhh . (3) Realization
interest. On contrary, the spectra of both towers in longitudinal
of the observability matrix [Op ], using the results of H2N . (4)
and lateral directions differ from the frequency of 1.0 Hz
Finally, realization of the state matrix A and R, followed by the
onward. The differences here imply non-uniformities of input
estimations of modal parameters by solving the eigenvalues and
excitations at the bottom of each tower. Therefore, these non-
eigenvectors of the realized state matrix.
uniformities should be considered as the case of multi-input
The algorithm has been implemented in a numerical
when implementing the system identification.
simulation using a three-dimensional FEM of cable-stayed
The spectra of accelerations obtained from three channels
bridge to examine its accuracy and efficiency. Results of
at different locations on the girder are shown in Fig. 7. Many
numerical simulation, which was presented elsewhere [15],
peaks were observed within a range of 0.5–2 Hz suggesting the
show that under a reasonable number of output sensors, modal
existence of closely-spaced natural frequencies of the bridge,
parameters of the bridge subjected to simulated filtered-short
the case that is common in a large and flexible structure such
ground motion can be estimated with great accuracy. The errors
as a cable-stayed bridge. Most of the peaks that appear on the
of natural frequency estimates were found less than 2%, while
longitudinal direction also appear on the lateral direction with
the error of damping estimates, which depends on the length of
the almost equal amplitude, indicating the coupling of motions
data, signal-to-noise ratio and non-stationary, reach up to 15%.
between longitudinal and lateral direction. Spectrum of the
Implementation of the algorithm to seismic records from long-
girder’s vertical responses reveals six distinct peaks constantly
span cable-stayed bridge and the evaluation of the results can
appear on the three vertical channels. Some of the peaks in
also be found in Ref. [10].
vertical components have strong corresponding peaks at the
lateral and longitudinal, except for the first and the third peak at
4. Seismic records
the frequency of 0.46 Hz and 0.81 Hz respectively. Later, it was
The present work utilizes ground motions and acceleration confirmed from SI results that these peaks are associated with
responses obtained from six earthquakes (Table 1); among them the pure vertical modes.
are the two sets of main shocks from the October 23, 2004 After observing the spectra of girder responses in three
Chuetsu Niigata Earthquake. The strongest earthquake, in terms directions, it was decided that the range of frequency interest
of response amplitude, is the July 23, 2005 earthquake. This for SI is between 0 and 2.5 Hz, since most dominant
peaks are contained within this range. Furthermore, owing
earthquake had the epicenter 25 km southeast of Tokyo with
to the unique shape of the bridge girder, it was found
the focal depth of 65.6 km [16]. In analysis, the total 210 s of
that the identification of longitudinal and lateral modes is
response from the February 20, 1990 earthquake was divided
difficult to be performed separately. Due to this reason, a
into three frames 60 s each in order to evaluate the effect of
three-dimensional system arrangement is chosen, instead of
earthquake amplitude to modal parameters. All records were
separating them into three in-plane motion systems like in
sampled using the standard sampling frequency of 100 Hz and
the application of modal identification to a straight-girder
baseline-corrected before processing.
cable-stayed bridge [10]. Using this system configuration the
influence of modal displacement and directions in a mode can
5. Data processing and analysis
be visually observed.
Data processing, analysis and the implementation of SI
5.2. Input–output data set
shown in this section utilize the records from the strongest
earthquake in July 23, 2005. Typical input and output responses A system of 6 inputs and 23 outputs (6I–23O) is employed
of the bridge from this largest earthquake are plotted in Fig. 5. for identification. The output channels are obtained from the
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2007

Fig. 5. Typical acceleration time histories recorded at (a) the bottom of main tower, (b) middle of girder recorded during the July 23, 2005 earthquake.

Fig. 6. Fourier amplitude of input acceleration measured on the bottom of the main tower and right tower. (a) Longitudinal direction, (b) lateral direction, (c) vertical
direction, recorded during the July 23, 2005 earthquake.

sensors at the girder, piers and on the top of the towers
(see Fig. 4). The girder sensors consist of 19 accelerometers
mounted at 8 locations along the girder horizontal centerline.
At the towers, four sensors are deployed to identify the
tower modes, consisting of two longitudinal and two lateral
accelerometers at each tower. Six accelerometers on the footing
of the main tower and the right tower are utilized as inputs.
These channels measure response in vertical, lateral and
longitudinal directions at the bottom of the towers and thus can
be viewed as direct inputs to the superstructures.
2008 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017

Fig. 7. Fourier amplitude of girder acceleration in (a) longitudinal direction, (b) lateral direction, (c) vertical direction, recorded during the July 23, 2005 earthquake.

As clearly shown in Eq. (18), the total number of modes

generated by the SI depends on the size of matrix A.
Furthermore, the size of matrix A is determined by the
number of singular values of matrix Rhh . The large singular
values represent the strength of dominating modes, while
the remaining values consist of either noise or fictitious
computational modes. After excluding the entire zero (or small)
singular values that are associated with the noise or fictitious
modes during decomposition of the Rhh matrix, the system
matrix A is realized. To facilitate the selection, a plot of singular
values is utilized, and the singular value cutoff is obtained from
a typical sudden drop of the values in this plot.

5.3. Realization of global modes

As explained in the previous section, every step in realiza-

tion uses an observation of singular value to decide the cutoff
values for the second realization of information matrix Rhh . In
the case of no-noise measurement, the cutoff value can easily
be determined by observing a significant drop in the singular
value plot. In practice, however no such case exists, since there
can hardly be a case of no-noise measurement. The inevitable
presence of noise and small non-linearity can distort the system
to some extent and results in the non-distinguishable singular
value plot. To overcome this limitation, a scheme of singular-
value truncation is implemented. All modes that are associated
with zero or smaller were eliminated (Fig. 8). The second re-
alization was performed by using the retained singular values
and subsequently all corresponding modes were generated. This
will result in many modes comprising the real ones and the
computational modes. To distinguish real modes from fictitious
modes, a modal supervision technique is utilized.
The technique selects certain modes by assigning threshold
values and modal supervisor, by using logic operators, which
Fig. 8. Singular values plot of the information matrix Rhh recorded during the
July 23, 2005 earthquake with p = 100.
are defined by the user. So at first, all modes are generated
then the resulted modes are put into a batch for comparison.
In this batch the resulting modes are passed through threshold
values. The threshold values are (i.e. all modes with the
criteria below were rejected): (1) negative damping ratio or/and
uncharacteristically high damping ratio, (2) Extended Modal
Assurance Criteria (EMAC) [17] less than 90%, (3) modes that
have uncharacteristic shapes or are incomparable with their
FEM counterparts, when compared to the FEM mode shapes
using the Modal Assurance Criterion (MAC) values.
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017
Fig. 9. Consistency of frequency and damping estimates with respect to p-values.
To better illustrate the case, one can refer to the example
given in Fig. 8. This figure shows an example of the singular
value plot for the total bridge system. One can observe that
singular values beyond 100 are almost or less than zero.
Therefore, the value of 100 is selected as the singular value
cutoff. Using this number, the total number of modes obtained
from matrix A is 100/2 = 50, since they all appear in
conjugate frequencies. Upon observing the 50 modes from
the first generation and by adopting the modal supervision
technique explained above, only 14 modes were identified with
confidence. The other modes were perceived as fictitious modes
or noise-generated modes.
5.4. Consistency of modal parameters
The theory states that the minimum value of p ≥ 2N /m + 1
will yield to an exact solution. However, due to measurement
noise that propagates into system, the actual minimum p-
values might be found larger than the ideal minimum, and
thus need to be investigated in order to minimize bias that
might appear in calculation. In order to study the sensitivity
of modal parameters with respect to the length of input–output
in the block matrix ([Yp ] and [Zp ]), several records with
various lengths were evaluated. For this purpose, the number
of columns in both [Yp ] and [Zp ] block matrices were fixed at
3500 for all frames, while the number of rows is incremented
from 10 until all available numbers of rows are utilized. Modal
parameters were calculated separately for these different block
matrices and the modal supervision technique was applied
afterwards.
2009
The results of modal parameters for various p-values
show higher estimates of both frequencies and dampings at
smaller p-values, before they decrease and finally stabilize
at certain values. The stabilized values are thus selected as
the representative estimates. It was found that most of the
modes were identified using p = 50. The p-values that give
the stabilized estimates may vary from one mode to another.
However, the observations show that, generally, the estimates
stabilize around p = 100. Fig. 9 shows typical examples
of the consistency of frequency and damping estimates with
respect to the p-values for the first vertical and the first lateral
mode of Katsushika–Harp Bridge. The figure clearly indicates
that steady results of natural frequency estimates are obtained
around the value of p = 100. To obtain a stabilized estimate
of damping ratio, however, a longer data set and thus larger
p-value is required. Sometimes, even after longer data was
included, a relatively small scatter of damping estimates are
still present. The reasons of the scatter estimates are related
to the facts that the small-valued damping is usually very
sensitive to the non-stationary nature of seismic excitation, and
the presence of measurement noise. Therefore to quantify a
representative damping estimate of one mode, an average value
and its standard deviation near the stabilized values are utilized
(see Table 3, for example).
6. Identification results
6.1. Global modes and natural frequencies
Owing to the double curvature of the bridge girder,
the vibration modes are all three dimensional in principle.
2010 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017

Table 2
Value of the mdp index and the mode type recorded during the July 23, 2005 earthquake

Mode Identified mdp index of the girder Type of mode

no. freq (Hz) X -direction Y -direction Z -direction
1 0.46 0.05 0.04 1.00 1st Vertical (Sym)
2 0.61 0.32 1.00 0.57 1st Lateral + Vertical 1st Sym
3 0.66 0.37 1.00 0.30 2nd Lateral
4 0.81 0.11 0.11 1.00 2nd Vertical (Asym)
5 0.93 0.24 1.00 0.28 3rd Lateral
6 0.99 0.24 1.00 0.21 4th Lateral
7 1.17 0.15 0.82 1.00 3rd Vertical (Sym) + Lateral
8 1.28 1.00 0.43 0.50 Longitudinal dominant mode
9 1.38 0.65 0.95 1.00 1st Torsion
10 1.64 0.27 1.00 0.56 5th Lateral
11 1.68 0.32 0.15 1.00 4th Vertical (Asym)
12 1.81 0.91 1.00 0.98 2nd Torsion
13 2.23 0.34 0.53 1.00 5th Vertical (Sym)
14 2.41 0.19 1.00 0.14 6th Lateral

To determine the type of a mode, a modal displacement
participation index (shortened hereafter as: mdp) is computed
using the following equation:
j
|max{φ(x,y,z) }|
mdp(x,y,z) (φ ) =
j
. (22)
j
|max{φ1:N ,(x,y,z) }|
Eq. (22) implies that for the j-th mode, the index calculates
ratio between the maximum absolute modal displacements of
structural component in one direction divided by the maximum
modal displacement of the structure. By employing the above
index, the participation or contribution of a motion in one
direction to the overall modal displacement can be quantified.
Therefore, the predominant direction and thus type of the mode
is determined. The dominating direction is the direction that
gives the mdp value equal or very close to unity.
Table 2 lists the results of identification with their mdp
index. Since the modes are all girder-dominant, the mdp index
calculates the participation of girder modal displacement in
three principal directions. Accordingly, the modes are classified
into four types of motion, namely: vertical, lateral, longitudinal
and torsion. The mode that is dominated by a motion in one
direction only, (i.e. vertical, lateral and longitudinal) has a
characteristic of mdp value equal to 1 in that corresponding
direction, and accompanied by weak participation (less than
0.5) of motions in other directions. A mode is classified as
torsional mode when mdp values for all directions exceed 0.5.
One can notice from Table 2 that except for the first and the
second vertical modes, all other modes are characterized by
relatively strong participation of modal displacement in other
directions beside the predominant one. This indicates that due
to the effect of the curved girder, most of the modes were highly
coupled.
The SI generates a total of fourteen vibration modes within
the frequency range of 0.4–2.5 Hz. Using the mdp index the
fourteen modes are classified as: five vertical predominant
modes, six lateral modes, one longitudinally predominant
mode and two torsion modes. Although the spectral peaks
indicate closely spaced natural frequencies, all modes were well
separated in the identification and their shapes fit their FEM
counterparts mode shapes.
The fundamental mode of the bridge is the first vertical
bending with a natural frequency of 0.45 Hz. This mode
exhibits a zero node symmetrical pattern of girder between the
main tower and the lower one (main span), while the side span
(Yotsugi approach) moving out-of-phase vertically. The mode
shows a pure vertical bending pattern, in which the participation
of motion in both longitudinal and lateral directions are very
weak as evidenced by the mdp < 0.05. The second vertical
mode was identified at the frequency of 0.81 Hz. The mode has
the characteristic of zero node and symmetrical main span, with
the side span moving in-phase. The participations of motion in
other directions were found stronger than they were in the first
vertical mode. The third mode was recovered at 1.17 Hz, with a
single node asymmetrical pattern of the main span. Unlike the
first two vertical modes, strong participation of lateral motion
appears in this mode (mdp = 0.8).
In the lateral direction, the first mode was found at the
frequency of 0.61 Hz. In this mode both towers move laterally
in-phase with each other as well as with the lateral motion of the
main span. The lateral motion of the main span was coupled
with vertical component (mdp = 0.57). Slightly higher, at
0.66 Hz, the second lateral mode was identified. The mode
has a characteristic of a zero node symmetric shape of the
main span in lateral direction, while the towers move out-of-
phase with each other. The first two lateral modes show the
strong participation of both towers but no participation of end-
piers in motion. More lateral modes were recovered at higher
frequency (0.93, 0.99, 1.64 and 2.41 Hz). They exhibit a single
node asymmetric pattern of the main span with the strong
participation of tower and weak modal displacement of both
end-piers.
As indicated in Table 3, the first torsional mode was
identified at 1.38 Hz. The fact that all sensors were installed
at the centerline of the bridge longitudinal axis makes it rather
difficult to visually observe the torsional nature of this mode.
However, its mdp values suggest the torsional characteristics.
In addition, when compared to the frequency range of the
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2011

Table 3
Comparison of the identified natural frequencies and damping ratios obtained from the seismic records with other tests and FEM

Mode Identified from earthquake Ambient Forced FEM Type of mode

no. 23-Jul-05 vibration vibration
test test
Frequency (Hz) Damping (%)a EMAC (%)
1 0.46 1.58 (0.78) 98.5 0.45 0.45 0.45 1st Vertical (Sym)
2 0.61 1.64 (0.80) 97.0 0.62 0.60 0.60 1st Lateral + Vertical 1st Sym
3 0.66 3.34 (0.26) 99.0 0.70 0.67 0.67 2nd Lateral
4 0.81 1.46 (0.20) 96.8 0.83 0.82 0.82 2nd Vertical (Asym)
5 0.93 2.06 (0.32) 96.2 0.98 0.96 0.92 3rd Lateral
6 0.99 0.65 (0.03) 96.5 1.01 1.01 1.00 4th Lateral
7 1.17 3.64 (0.21) 93.6 1.17 1.18 1.17 3rd Vertical (Sym) + Lateral
8 1.28 1.79 (0.80) 91.8 – – N/A Longitudinal dominant mode
9 1.38 2.61 (0.74) 96.7 1.31 1.38 1.37 1st Torsion
10 1.64 1.33 (0.45) 94.5 1.71 1.63 1.64 5th Lateral
11 1.68 1.09 (0.30) 93.5 1.67 1.67 1.68 4th Vertical (Asym)
12 1.81 2.79 (0.63) 92.9 1.82 1.79 1.81 2nd Torsion
13 2.23 2.62 (0.10) 92.1 2.23 2.22 2.24 5th Vertical (Sym)
14 2.41 1.52 (0.43) 93.4 2.40 2.41 2.41 6th Lateral
Note: –: Not identified, N/A: not available.
a Values in the brackets denote the standard deviation.

first torsion mode identified in ambient vibration test, forced
vibration and also as predicted from FEM, this mode was
indeed identified within the same frequency range of the first
torsional mode.
The two torsion modes were found in relatively high-
frequency modes (i.e. at 1.38 and 1.81 Hz), which imply that the
torsional effect is not so dominant in this curved cable-stayed
bridge. The reason is related to the fact that while the girder
is essentially curved in the horizontal, it is supported at four
locations: both towers and the approach end-piers. Moreover,
the large radii of both girder curves and the position of the main
tower in the transitional point of the curves might have, to some
extent, reduced the effect of torsional motion.
It should be noted that there is one mode that was not
identified either during the forced vibration or ambient test,
but was identified from earthquake records. This mode is the
longitudinal predominant mode identified at the frequency of
1.28 Hz. The mode has a characteristic of longitudinal swing
motion, in which the girder, pier and tower move longitudinally
toward one direction along the centerline of the girder. The mdp
index reveals that the maximum girder modal displacement
occurred in the longitudinal direction. It appears that the
excitation level provided during ambient and forced vibration
tests were insufficient to produce reasonably strong vibration
to excite this mode. The similar fact was observed during
the study of SI applied to seismic response of a large cable-
stayed bridge with a straight girder [10,15]. In both references,
the longitudinal dominant mode was significantly insightful
to assess the performance of isolation devices between the
pier and girder. In this work, however, due to the lack of
sensors in pier-caps, such a performance assessment cannot be
performed.
Fig. 10 illustrates the mode shapes in two types of view:
the front and the plan view: From this figure the relationship
between components (i.e. vertical, lateral and longitudinal) in
the mode can easily be observed. Table 3 lists the frequency
estimates identified from the strongest earthquake (July 23,
2005). It is shown in this table that the natural frequencies
estimated from seismic response are in good agreement with
the results from ambient vibration and forced vibration tests
conducted after the completion of the construction, [18] as
well as the finite element model. Generally, natural frequencies
estimated from the ambient vibration test are slightly higher
than those estimated from the seismic records with the
maximum discrepancy between the two being 5.7%. This
suggests that, during the small level of ambient excitation the
structure is seemingly more rigid than it is during larger levels
of excitation.
6.2. Damping ratio estimates
In the conventional frequency domain SI, the modal
damping ratio can be estimated using the half-power bandwidth
of the Fourier spectrum peaks. In the case of this bridge,
however, estimating damping ratios from half-power bandwidth
are difficult since the spectrum peaks are closely spaced
or overlapped by adjacent peaks, especially in lateral and
longitudinal directions.
One advantage of the SRIM identification algorithm is that
damping estimates are derived directly from the eigenvalues
of the state matrix A. Hence, the values are well separated
since each eigenvalue is associated with a unique mode. In
implementation of the identification algorithm, however, the
initial estimates of damping ratio were rather scattered and
require larger p-values to obtain stabilized results. This fact
was not found during the estimation of natural frequency.
The reasons of the scatter estimates are related to the facts
that the small-valued damping is usually very sensitive to the
non-stationary nature of seismic excitation, the presence of
measurement noise, the selection of p-values and the length
of data. Nevertheless, the representative values of damping for
each mode were obtained by observing the stabilized values
over several ranges of p-values as indicated in the example
2012 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017

Fig. 10. The first eight mode shapes of the Katsushika–Harp Bridge, identified from the July 23, 2005 earthquake records (note: dotted lines represent the stationary
position, i.e. centerline of the girder, towers and piers).

(Fig. 9). The damping estimates were found within the range
of 0.5%–4% with average value and standard deviation of 2±
0.8%.
6.3. Tower modes
While the identification method can simultaneously generate
the global modes, the local modes of structural members can
also be identified by employing the sensors associated only
with those of the members. In this section, the tower modes
and their participation in the global modes are examined.
For this purpose, sensors located on the top of both towers
are utilized as outputs. These sensors measure accelerations
in longitudinal and lateral directions. Fig. 11 shows typical
longitudinal and lateral accelerations recorded on the top of
the main and the right tower. Considering the slender shape
of the tower, especially the main one, and the material they
are made of, a higher level of vibration is expected. In both
towers accelerations in lateral direction were found larger than
the accelerations in longitudinal direction. The identification
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2013

Table 4
Identified modal parameters of the main tower and their comparison with the global modes recorded during the July 23, 2005 earthquake

Mode no. Freq (Hz) Main tower Right tower Associated global mode
Damp (%)a Mode angle (◦) Damp (%)a Mode angle (◦)
1 0.45 1.61 (0.38) 58.7 1.32 (0.19) 148.38 1st Vertical (Sym)
2 0.59 0.94 (0.10) 264.3 0.56 (0.09) 265.27 1st Lateral + Vertical 1st Sym
3 0.65 2.55 (0.17) 264.8 1.67 (0.20) 85.70 2nd Lateral
4 0.81 0.82 (0.09) 83.5 0.99 (0.18) 237.55 2nd Vertical (Asym)
5 0.92 2.23 (0.24) 266.5 2.95 (0.31) 87.13 3rd Lateral
6 1.09 0.85 (0.17) 81.6 2.47 (0.26) 67.5 4th Lateral
7 1.16 0.73 (0.15) 89.2 0.48 (0.17) 98.6 3rd Vertical (Sym) + Lateral
8 1.26 0.41 (0.35) 286.5 1.11 (0.33) 97.1 Longitudinal dominant mode
9 1.38 0.78 (0.18) 271.4 0.66 (0.14) 100.52 1st Torsion
a Values in the brackets denote the standard deviation.

Fig. 11. Typical acceleration responses of the tower recorded during the Feb 2, Fig. 12. Fourier amplitude of the towers’ acceleration responses recorded
1992 earthquake. during the Feb 2, 1992 earthquake.

employs a system of six inputs and two outputs (6I–2O) for
each tower. The inputs were selected from the bottom of the
tower, while the outputs were the two channels on the top of
the main and the right tower. By observing the stabilized plot,
p = 100 was selected. Based on this selection the system
matrix A is realized and its singular values were examined
afterwards. Fig. 8 shows an example of the singular value
plot of the main tower system. One can observe that singular
values beyond 60 are almost or less than zero. Therefore,
the value of 60 is selected as the singular value cutoff.
Using this number, the total number of modes obtained from
matrix A is 60/2 = 30, since they all appear in conjugate
frequencies. Afterwards, upon observing the 30 modes from
the first generation and by adopting the modal supervision
technique explained above, only 9 modes were identified with
confidence.
The results of SI of the main tower (Table 4) indicate that
all tower modes at the frequency range of 0–1.4 Hz were
associated with the girder motion. The interactions between
the girder motion and tower exist in such a way that the
tower modes were resulted from their girder’s predominant
counterpart modes. For example the fundamental tower mode
at 0.45 Hz is the result of the girder–tower interaction mode
in vertical direction. This mode was found as the fundamental
mode in both towers. The alignments of modal displacement
(see mode angle in Table 4) indicate that both towers are leaning
towards the main span. The second and third modes were
dominated by lateral motion, with the difference in the phase.
The second mode exhibits an in-phase motion of both towers,
while the third one shows the out-of-phase motion.
Fig. 12 shows the Fourier spectra of both towers in
longitudinal and lateral directions. Due to larger lateral motion,
peaks in lateral channels dominate the spectra. One can
notice from these spectra that the frequency peaks are indeed
associated with the girder dominant modes. Fig. 13 plots the
alignment of modes identified from the main tower. From the
angle of mode alignment, it can be seen that except for the
first mode (0.45 Hz), all main tower modes were dominated
by the lateral motion. It is interesting to see that even in the
modes that were associated with the girder vertical modes, the
2014 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017
Fig. 13. Mode shape vectors of the main tower. (Note: solid line indicates the principal alignment of the mode, the dotted line indicates the principal axis of the
tower.)
towers’ modal displacement are still lateral-predominant. This
characteristic was not found in the single-plane cable-stayed
bridge with the straight girder, such as the case of the Tsurumi-
Fairway Bridge [19].
In the identification process, damping estimates of tower
modes were stabilized easily and the scatter is confined within
a narrow range of estimates as shown by the small value of
standard deviations (see Table 4). An essential feature observed
here is the small value of damping estimates of both towers.
In the main tower, for instance, except for the modes that are
associated with the 1st vertical, 2nd and 3rd lateral modes,
damping estimates were all identified less than 1%. The mean
value for all modes is 1.21 ± 0.74%. The damping estimates
of the right towers are also small, with the mean value of
1.35 ± 0.86. These results show that damping ratios of both
towers are lower than the damping estimates of the global
modes (mean value: 2.00 ± 0.87%).
The small damping of both towers can be understood as the
effect of the stay cables. Since the flexural rigidity of the cable
is very small, the restraining effect especially in lateral direction
becomes considerably smaller. Therefore, with the slender
shape of both towers, they can be considered as freestanding
cantilever columns. It should be mentioned, however, that due
to the curved shape of the girder, the cables are not necessarily
on one plane. Consequently, there is still some portion of axial
rigidity from the cables acting on the tower in lateral direction.
This axial rigidity may contribute to the tower flexibility in
lateral motion. In a similar case of a single-plane cable-stayed
bridge and similar shape of steel box section tower, but with a
straight girder, an even lower damping ratio of the tower mode
(0%–1%) was observed [20].
7. Comparisons of modal parameters from six earthquakes
The seismic records available for this study cover a range of
accelerations with the maximum amplitude from 2 to 60 cm/s2 .
And as previously mentioned, the identification procedure can
be performed over a relatively narrow time window. Therefore,
by employing several time windows from different levels of
acceleration amplitude, variations of modal parameters can be
observed. In a long record such as the February 20, 1990
earthquake, the responses were divided into three frames of
60 s each, using the time-shifted windows. With this procedure,
the identified modal parameters can be compiled to detect the
possible structural changes during excitation process.
For this purpose, six earthquakes records were divided into
8 frames and the resulting modal parameters were compared.
The maximum input accelerations recorded at the base of
the main tower were found at the range of 2–52 cm/s2 ,
and their corresponding root-mean-squares were between
0.6 and 5.4 cm/s2 . Table 5 shows the results of natural
frequency estimates identified from the eight frames. In this
table, the frequency estimates are listed with the standard
deviation values obtained from p-values around the stabilized
frequencies. In general, constant results of frequency estimates
were obtained in most of the frames. This clearly indicates
that the natural frequency of the structure remains constant
regardless of the level of input amplitude. It should also be
mentioned that in all modes variations of frequency estimates
are very small as noted by the maximum standard deviation that
equals 0.03 Hz.
The results of damping ratios shown in Table 6, on the other
hand, indicate larger variations of estimates. Mean values of
damping ratio of most modes were found within the range of
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2015

Table 5
Natural frequencies of the bridge identified from six earthquakes

Type of mode Identified natural frequency (Hz) and its standard deviation (in bracket)
Feb 20, 1990 Feb 2, 1992 Oct 23, 2004 Jul 23, 2005 Aug 16, 2005
Main Main
Frame Frame Frame shock shock
1 2 3 17:57 18:35

1st Vertical (Sym) 0.45 0.45 0.45 0.45 0.45 0.45 0.46 0.45
(0.00) (0.00) (0.00) (0.01) (0.00) (0.00) (0.00) (0.00)
1st Lateral + Vertical 1st 0.60 0.60 0.60 0.59 0.62 0.59 0.61 0.60
Sym (0.01) (0.00) (0.00) (0.00) (0.00) (0.00) (0.01) (0.01)
2nd Lateral 0.66 0.67 0.67 0.67 0.65 0.66 0.66 0.66
(0.01) (0.00) (0.00) (0.00) (0.01) (0.01) (0.00) (0.00)
2nd Vertical (Asym) 0.81 0.82 0.82 0.84 0.80 0.80 0.81 0.81
(0.00) (0.00) (0.00) (0.01) (0.01) (0.00) (0.01) (0.00)
3rd Lateral 0.96 – – 0.95 0.95 0.93 0.93 –
(0.00) – – (0.01) (0.00) (0.01) (0.01) –
4th Lateral 1.00 1.01 1.00 1.00 1.01 1.00 0.99 1.05
(0.00) (0.00) (0.01) (0.01) (0.02) (0.01) (0.01) (0.01)
3rd Vertical (Sym) + Lateral 1.16 1.18 1.17 1.19 1.16 1.14 1.17 1.16
(0.00) (0.01) (0.01) (0.01) (0.00) (0.01) (0.01) (0.01)
Longitudinal dominant mode 1.28 1.28 – 1.28 1.29 1.27 1.28 1.29
(0.02) (0.01) – (0.01) (0.01) (0.01) (0.01) (0.01)
1st Torsion 1.38 1.38 1.37 1.35 1.36 1.35 1.38 1.37
(0.01) (0.00) (0.01) (0.01) (0.01) (0.01) (0.02) (0.02)
5th Lateral 1.62 1.61 1.64 1.64 – 1.62 1.64 1.64
(0.01) (0.02) (0.01) (0.02) – (0.02) (0.01) (0.01)
4th Vertical (Asym) 1.68 1.67 1.68 1.68 1.68 1.68 1.68 1.68
(0.01) (0.02) (0.01) (0.01) (0.02) (0.02) (0.01) (0.02)
2nd Torsion 1.82 1.83 1.81 1.84 1.86 1.83 1.81 1.86
(0.01) (0.02) (0.03) (0.01) (0.01) (0.02) (0.01) (0.02)
5th Vertical (Sym) 2.22 2.22 – 2.20 2.22 2.22 2.23 2.22
(0.01) (0.01) – (0.02) (0.02) (0.02) (0.02) (0.02)
6th Lateral 2.40 2.41 2.41 2.40 2.41 2.37 2.41 2.44
(0.02) (0.02) (0.01) (0.02) (0.01) (0.03) (0.01) (0.02)

0%–5%. The standard deviation of the estimates, however, was
quite large compared to the mean values, indicating the scatter
estimates.
Fig. 14 attempts to illustrate the trend of damping estimates
with respect to the maximum of input accelerations measured
at the base of the main tower. It is shown in this figure that,
despite the scatter estimates, one can observe a trend of input
amplitude dependence of the mean damping estimates. In the
first four modes, most damping estimates were found to be
within 0%–4% with the mean values increasing as the input
amplitude increased. The dependence of the damping estimates
on the input amplitude is quite apparent for the third and fourth
mode, where the estimates do not show large scatter. For the
first and second mode, however, the large scatter of estimates
makes it more difficult to arrive at conclusive results of
damping-input amplitude relationship. Generally, larger input
excitation will produce larger structural response. And due to
the larger response, greater energy dissipation would take place
on the bridge component such as friction in bearings. Owing to
this mechanism and also by considering the effect of curved
girder and the out-of-plane arrangement of the stays, larger
damping during larger input excitation is anticipated. However,
since damping is small in value and complex in mechanism,
more comprehensive and detailed analysis are required before
arriving at a quantitative relationship between the mechanisms
and their contribution to the overall damping amplification,
especially for such a complex structure during large earthquake
excitation. This is beyond the scope of the current paper and
will be incorporated in future work.
8. Conclusions
The assessment of dynamic characteristics of the
Katsushika–Harp Bridge, a curved cable-stayed bridge using
seismic responses has been presented in this study. Owing to
the dense sensor deployment, excellent quality of measured re-
sponse, and the recent development of algorithms to deal with
the data, the dynamic characteristics of the bridge can be effi-
ciently and accurately estimated.
Fourteen modes were identified within the range of
0–2.5 Hz. Due to the curved shape of the girder; most of them
were closely spaced in frequency and having complicated 3D
motion-coupled mode shapes. Using the modal displacement
ratio, the participation of motion in one direction can be
determined and thus the type of mode is obtained according
to the dominant direction of motion. Based on observation of
these characteristics, the following conclusions are drawn:
2016 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017
Fig. 14. Variation of damping estimates with respect to the maximum input amplitude, recorded from 8 frames of six earthquakes.
1. The characteristics of the modes are somewhat similar to
those of cable-stayed bridge with the straight girder, in
which the girder motion dominated the global mode shapes.
High-coupled motions in three directions, however, were
observed that make it difficult to obtain the pure mode in
one direction only. In all but the first vertical mode, the
effect of lateral motion was found significant. Torsional
effect was found to be relatively small as evidenced by the
high-frequency of torsional dominant modes. This might be
due to the fact that while the girder is essentially curved
in horizontal, it is well supported at four locations: both
towers and the approach end-piers. In addition, the large
radii of both girder curves and the position of the main tower
might have reduced the effect of the torsional effect in the
motion.
2. Tower modes identified within the same frequency range
of the global modes were all associated with girder
modes. Except for the first mode, all tower modes were
characterized by dominant lateral motions. Damping ratios
of tower modes were found lower than the global modes. The
average damping ratios identified at both towers are 1.21%
and 1.3%.
3. Using the response from six earthquakes with the level
of input amplitude between 2 and 52 cm/s2 , the trend of
modal parameters was evaluated. It was found that while the
frequencies remain constant, damping ratios exhibit slight
dependencies on earthquake input amplitude. An increasing
trend of damping ratio was observed with the increase of the
earthquake input amplitude.
The scheme of system realization using information matrix
adopted for SI in this work has been an indispensable
instrument in dealing with voluminous seismic responses
recorded from multi-channels of sensors on the bridge. The
algorithm is straightforward and requires neither a prior
structural model, nor an optimization procedure to obtain
accurate estimates of dynamic characteristics of a complex
structure such as the curved cable-stayed bridge. In the view
of structural health monitoring, an efficient and accurate
identification technique are essential, especially for quick
structural assessment after the occurrence of earthquake
events. In the future, instrumentation of the cables should
be considered. This would be helpful to monitor the degree
of coupling between tower and girder caused by the radial
arrangement of the cables and also to monitor the tension of
cables during extreme shaking such as an earthquake.
Acknowledgement and disclaimer
The authors would like to express their gratitude to
the Tokyo Metropolitan Expressway Public Corporation for
providing the strong motion records and the drawings of
the bridge. Opinions, findings, and conclusions expressed in
this material are those of the authors and do not necessarily
reflect those of the Tokyo Metropolitan Expressway Public
Corporation. The first author gratefully acknowledges the
Japanese Ministry of Education, Culture, Sports, Science and
Technology, for financial support provided during this research
at the University of Tokyo.
 D.M. Siringoringo, Y. Fujino / Engineering Structures 29 (2007) 2001–2017 2017

Table 6
Damping ratio of the bridge identified from six earthquakes

Type of mode Identified damping ratio (%) and its standard deviation (%)
Feb 20, 1990 Feb 2, 1992 Oct 23, 2004 Jul 23, 2005 Aug 16, 2005
Main Main
Frame Frame Frame shock shock
1 2 3 17:57 18:35
1st Vertical (Sym) 1.53 0.52 1.19 2.26 2.22 1.33 1.58 1.26
(0.99) (0.46) (0.30) (1.10) (1.54) (0.34) (0.78) (0.36)
1st Lateral + Vertical 1st 2.16 1.18 – 1.97 2.49 0.74 1.64 2.72
Sym (0.73) (0.13) 0.47(0.31) (0.70) (0.39) (0.26) (0.80) (0.21)
2nd Lateral 1.36 1.65 2.19 1.53 1.98 3.68 3.34 2.23
(0.09) (0.08) (0.28) (0.29) (0.44) (0.24) (0.26) (0.22)
2nd Vertical (Asym) 0.61 0.83 0.66 0.44 0.52 1.04 1.46 1.12
(0.25) (0.10) (0.08) (0.14) (0.23) (0.38) (0.20) (0.09)
3rd Lateral 3.42 – – 2.39 2.89 4.34 2.06 –
(1.79) – – (0.45) (0.09) (0.36) (0.32) –
4th Lateral 2.86 4.36 5.44 2.71 2.15 1.56 0.65 2.67
(1.39) (1.67) (0.72) (1.18) (0.29) (0.34) (0.03) (0.36)
3rd Vertical (Sym) + 1.78 1.95 0.71 0.55 1.21 2.27 3.64 0.79
Lateral (1.06) (0.65) (0.56) (0.41) (0.21) (0.50) (0.21) (0.15)
Longitudinal dominant 2.47 2.33 – 2.58 3.70 2.28 1.79 2.33
mode (0.21) (0.22) – (0.37) (0.43) (0.78) (0.80) (1.30)
1st Torsion 2.91 – 3.41 3.61 3.72 1.23 2.61 0.23
(2.17) 2.56 (-) (1.28) (0.65) (1.03) (0.65) (0.74) (0.17)
5th Lateral 4.79 5.51 1.74 3.30 – 1.72 1.33 2.11
(1.17) (1.07) (1.00) (0.55) – (0.45) (0.45) (0.18)
4th Vertical (Asym) 2.25 – 3.07 0.90 3.06 2.29 1.09 2.38
(0.55) 3.46 (-) (0.30) (0.52) (0.21) (0.85) (0.30) (0.59)
2nd Torsion 8.68 4.01 1.62 1.61 1.36 1.69 2.79 3.57
(0.79) (0.75) (0.61) (1.43) (0.11) (0.72) (0.63) (0.63)
5th Vertical (Sym) 3.63 0.90 2.58 1.62 3.23 1.45 2.62 3.06
(1.00) (0.96) (0.41) (0.91) (0.13) (0.96) (0.10) (0.44)
6th Lateral 1.38 1.76 1.06 1.73 2.47 1.35 1.52 2.88
(0.31) (0.71) (0.65) (0.80) (0.25) (0.35) (0.43) (0.88)

References
[1] Fujino Y, Kikkawa H, Namikawa K, Mizoguchi T. Seismic retrofit
design of long-span bridges on metropolitan expressways in Tokyo.
In: Proceeding of the 6th international bridge engineering conference.
Transportation Research Board; 2005. [CD-ROM].
[2] Beck JL, Jennings PC. Structural identification using linear model and
earthquake records. Earthquake Engineering & Structural Dynamics
1980;8:145–60.
[3] McVerry GM. Structural identification in the frequency domain from
earthquake records. Earthquake Engineering & Structural Dynamics
1980;8:161–80.
[4] Chaudhary MTA, Abe M, Fujino Y, Yoshida J. System identification
of two base-isolated bridges using seismic records. Journal of Structural
Engineering ASCE 2000;126(10):1181–94.
[5] Virlogeux M. Recent evolution of cable-stayed bridges. Engineering
Structures 1999;21:737–55.
[6] Brownjohn JMW, Lee J, Cheong B. Dynamic performance of a curved
cable-stayed bridge. Engineering Structures 1999;21:1015–27.
[7] Deger Y, Cantieni R, deSmet CAM, Felber AJ. Finite element model
optimization of the new Rhine Bridge based on ambient vibration testing.
In: Proceedings of Eurodyn ’96. 1996. p. 817–22.
[8] Gentile C, Martinez y Cabrera F. Dynamic performance of twin curved
cable-stayed bridges. Earthquake Engineering & Structural Dynamics
2004;33:15–34.
[9] Abdel-Ghaffar AM. In: Ito M, editor. Cable-stayed bridges: Recent
developments and their future. Cable-stayed bridges under seismic action.
Elsevier Science Publisher BV; 1991. p. 171–91.
[10] Siringoringo DM, Fujino Y. Observed dynamic performance of the
Yokohama-Bay bridge from system identification using seismic records.
Journal of Structural Control and Health Monitoring 2006;13(1):226–44.
[11] Lus B, Betti R, Longman RW. Identification of linear structural system
using earthquake-induced record. Earthquake Engineering & Structural
Dynamics 1999;28:449–1467.
[12] Smyth AW, Pei J-S, Masri SF. System identification of the Vincent
Thomas suspension bridge using earthquake records. Earthquake
Engineering & Structural Dynamics 2003;32(3):339–67.
[13] Juang JN. System realization using information matrix. Journal of
Guidance, Control, and Dynamics 1997;20(3):492–500.
[14] Juang JN. Identification and control of mechanical systems. Cambridge
University Press; 2001.
[15] Siringoringo DM. System identification of instrumented bridges using
seismic records and their performance evaluation. Doctoral dissertation.
University of Tokyo; 2005.
[16] Japan Meteorological Agency. http://www.jma.go.jp/JMA HP/jma/index.
html.
[17] Pappa RS, Elliot KB. Consistent-mode indicator for the eigensystem
realization algorithm. Journal of Guidance Control and Dynamics 1993;
16(5):825–58.
[18] Sano S, Morimoto C, Ohgaki K. Vibration tests on completed
Katsushika–Harp Bridge. Kawasaki Technical Review 1988;100:121–9
[in Japanese]. http://www.khi.co.jp/tech/index e.htm.
[19] Yamaguchi H, Takano H, Ogasawara M, Shimosato T, Kato M, Okada J.
Identification of dynamic characteristics by field vibration test in Tsurumi
Tsubasa bridge. Journal of Japan Society of Civil Engineering 1996;543:
247–58 [in Japanese].
[20] Kawashima K, Unjoh S. Seismic behavior of cable-stayed bridges.
In: Ito M, editor. Cable-stayed bridges: Recent developments and their
future. Cable-stayed bridges under seismic action. Elsevier Science
Publisher BV; 1991. p. 193–212.