Seismic fragility analysis of a typical two-pylon cable-stayed bridge

in China: Comparison of fragility models

Pang Yutao1, Wang Xiaowei2, Shang Yu3, Wu Li4

1
Postdoctoral Fellow, Department of Civil, Geological, and Mining Engineering, Ecole
Polytechnique, Montréal, Canada. Email: pangyutaoyy@gmail.com
2
PhD Candidate, State Key Laboratory of Disaster Reduction in Civil Engineering, Tongji
University, Shanghai 200092, China. Email: 10_xiaoweiwang@tongji.edu.cn
3
PhD Candidate, State Key Laboratory of Disaster Reduction in Civil Engineering, Tongji
University, Shanghai 200092, China. Email: 0725shangyu@tongji.edu.cn
4
Professor, Faculty of Engineering, China University of Geosciences, Wuhan 430074, China.
Email: lwu@cug.edu.cn (Corresponding author)

ABSTRACT: The necessity of performing a detailed seismic fragility analysis of

cable-stayed bridge is well established in China. This paper focuses on the seismic fragility
analysis of a typical cable-stayed bridge with 400 meters span and H-type pylon. The primary
emphases of the fragility analysis adopted here are the detailed nonlinear modeling of
cable-stayed bridges along with time-history analyses and the generation of fragility curves of
various components (e.g. pylon and pier sections and bearings). Different fragility methods,
namely IDA and regression method, maximum likelihood method and the cloud method using
regression method, are used for fragility analysis and their results are compared. Among these,
a new regression-based method and high-order polynomial are proposed in this work in order
to compare the results with the existing methods to see whether it is better. In general, PGV,
VSI and Sa(T1) gives a better correlation (than PGA) with the demand parameter and should
be used if the fragility is estimated independently. The proposed method of fragility analysis
gives the best estimates of the fragility function. But the cloud and regression method fails to
provide realistic estimates of the fragility function.
KEYWORDS: Fragility analysis; Cable-stayed bridges; Probabilistic safety assessment;
Incremental dynamic analysis; Correlation

INTRODUCTION

In last decades, cable-stayed bridges have been popular not only in China but also
around the world due to appealing aesthetics, full and efficient utilization of structural
materials, increasing stiffness over suspension bridges and efficient and fast mode of
construction (Ren and Obata 1999). Nowadays, China has more than 400 cable-stayed bridges.
Some of these are located in moderate and high seismic zones according to the current
seismological intensity map of China. Cable-stayed bridges are sensitive to earthquake
because of their high flexibility and low inherent damping (Pacheco et al. 1993; Chang et al.
2004). If a cable-stayed bridge collapses in an earthquake, it can cause tremendous loss of life
and property. Therefore, seismic fragility of cable-stayed bridges should be evaluated to
ensure the safety.
Seismic fragility, which defines the probability of physical damage as a function of
seismic intensity measures (IMs), is used to develop vulnerability information for bridges
(Shinozuka et al. 2000a). Three main methods are used for development of fragility functions:
(1) damage data of structures from past events (Basoz and Kiremidjian 1999); (2) expert
opinion; and (3) analytical methods (Hwang and Jaw 1990; Mander and Basoz 1999;
Shinozuka et al. 2000a,b). Nevertheless, when past damage data are absent and appropriate
analytical tools are available, the analytical method is the proper method for fragility
development of engineering and special structures (Padgett and DesRoches 2008). For the
past decade, maximum likelihood method, the cloud method and IDA have become three
main popular approaches for estimating seismic fragility on bridges and other structures, such
as maximum likelihood method (Shinozuka et al. 2000a,b), cloud method (Choi et al. 2004;
Nielson and DesRoches 2007), and incremental dynamic analysis (IDA) (Feng 2008;
Vamvatsikos and Fragiadakis 2010). With regards to the seismic fragility analysis of
cable-stayed bridges, few studies (Casciati et al. 2008; Pang et al. 2014) have been published.
Thus, it needs more studies to evaluate the fragility curves of a cable-stayed bridge within a
probabilistic framework.
The primary objective of this work is to perform seismic fragility analysis of a typical
two-pylon cable-stayed bridge in China using different fragility methods. Few fragility
analyses have so far been conducted for this structure. The highlights of these fragility
analyses are: demand estimations based on nonlinear response-history analysis, the use of
various as-recorded ground motions, and the use of different seismic intensity measures and
the use of different fragility methods. Seismic fragility is estimated from the seismic demands
data using: (a) IDA method and nonlinear regression based method (Ellingwood et al. 2001),
(b) a maximum likelihood based method (Shinozuka et al. 2000a), (c) cloud method and
regression-based method and (d) a modification of the regression-based method proposed in
this work. The emphasis is on a comparative study of the fragility estimates from these four
methods and their effectiveness.

FINITE ELEMENT MODELING

In order to make this research more represent, we collected 55 cable-stayed bridges in

China to investigate the representative span, pylon height and pylon type of cable-stayed
bridge. Fig. 1 shows the statistical results of this investigation. From the results of Fig. 1, we
can see that the dominated main span of cable-stayed bridge ranges from 400 m to 500 m, and
pylon is mainly H shaped or inverted Y shaped while the pylon height is around 150 m to 200
m. Thus, the considered cable-stayed bridge has a length of 70m+130m+400m+130 m+70 m
and H shaped pylon with 156m height as shown in Fig. 2. It consists of two concrete pylons,
144 steel cables and 4 additional RC side piers.

a) main span b) pylon shape

c) pylon height
Fig 1 statistical results of cable-stayed bridge in China

The finite element model of the cable-stayed bridge is built using OpenSees (McKenna
et al. 2006) as shown in Fig. 2. The OpenSees can perform nonlinear time history analysis
considering large geometric effects, local nonlinear geometric effects, and element distributed
plasticity due to material nonlinearity, which have been proved to be crucial for analysis of
cable-stayed bridges by Ren and Obata (1999). A distributed plasticity fiber element is used
for piers and pylons to account for the axial force-moment interaction and the material
nonlinearity. The nonlinear beam-column element is integrated numerically at five integration
points along its length. The cross-section of each beam-column element is divided into
200-400 fibers. The elastic beam-column element is used for the superstructure. The cable
element was modeled as a large-displacement truss element using the Ernst method due to the
easy usage and the capability to account for the sag effect. The cables were assumed to be
straight line element without the distributed mass to neglect the local dynamics (Caetano et al.
2000). The large geometric effect of cable-stayed bridge was considered by a P-Δ geometric
nonlinearity that modified the lateral resistance by the axial load (P) acting through the
displacement (Δ). Rigid links were used to connect the cables to the tower. Since the bridge
was supported by bed-rock, the soil–structure interaction effects were neglected. The
constraints of the supports were set to be fixed. The cable-stayed bridges had low structural
damping, so 3% modal damping matrix was applied in the analysis. The first 10 numerical
natural vibration periods of the evaluation FE model were 6.58s, 2.91s, 2.81s, 2.70s, 2.47s,
2.31s, 1.50s, 1.24s, 1.16s, 1.09s. The ground motion at each support is input uniformly.
 Fig. 2 the overall bridge structural finite element model

SEISMIC FRAGILITY ANALYSIS

In the present paper, the basic assumption is that the uncertainty of ground motions far
supersedes any other uncertainties that affect the fragility of cable-stayed bridge. This
assumption is made due to the good quality and strict control of cable-stayed bridges. Four
methods of fragility, namely IDA and regression method, cloud method, maximum likelihood
method and proposed regression-based method, are used and compared to test their efficiency
to estimate the fragility of components of cable-stayed bridge. 200 far-fault as-recorded
ground motions are selected for nonlinear time-history analysis (NLRHA) to obtain the IDA
curves. Table 1 provides basic information on the 200 records. All ground motions are treated
as independent records and each one of them is used for seismic response analysis. For IDA
analysis of each record, NLRHA of the structure is performed at various scales (e.g. 10 scales)
of the intensity measure (IM). Fig. 3 shows the PGA vs. maximum pylon curvature multi-IDA
plot. These IDA curves are used as reference for fragility estimates. From these IDA curves,
the probability of failure can be calculated in a simple way at each IM level by counting the
number of IDA curves that cross the vertical line corresponding to the limit state under
consideration. The ratio of these IDA curves to the total number of IDA curves (200 in this
case) is the probability of failure at that level of IM, which is also the fragility. This method is
called direct IDA method.

Table 1 basic information of far-fault 200 ground motion records

No. Earthquake event Year Magnitude Mean-Rrup (km) Number of records

1 Loma Prieta 1989 6.93 57.99 45

2 Northridge 1994 6.69 51.45 35
3 Chichi 1999 7.62 54.68 35
4 Whittier Narrows 1987 5.99 42.38 25
5 San Fernando 1971 6.61 60.13 45
6 Imperial Valley 1979 6.53 37.32 15
Total 200
-3
x 10
2
IDA curves
1.8

1.6

1.4
Curvature(1/m)

1.2

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3
PGA(g)
Fig. 3 IDA curves of pylon curvatures using 200 ground motions

Fragility analysis using IDA and regreesion method

This method is proposed by Ellingwood et al. [11] by using IDA and regression for the
fragility evaluation of frames in the United States. They used the response data from IDA
analysis to generate sample responses for a specified intensity measure. The variation in the
response quantity (maximum drift ratio) at any given intensity was then modelled with a
lognormal distribution. The response was expressed as an exponential function of the
intensity adopting a regression approach:
EDP = a(IM)! 𝛿 (1)
in which a and b are regression parameters and 𝛿 is the dispersion of the response EDP
(engineering demand parameters). a(IM)! is the median value of IDA curves as a function
of intensity measure IM. The nonlinear relationship of Eq. (1) is conducted by linear
regression between In(IM) and In(EDP). Then, the failure probability is calculated by the
exceedance of the capacity θc
 !" a(IM)𝑏 !!
Pr = ∅ (2)
!

where ∅(. ) is the standard normal CDF. 𝛽 should include both uncertainties of demand and
capacity. IM could be PGA, PGV and Sa(T1) etc. This method is labelled as ‘IDA and
regression’ method.

Fragility analysis using maximum likelihood estimate

This method is proposed by Shinozuka et al. (2000a) using maximum likelihood
estimate method to estimate the parameters associated with the calculated probability. The
parameters here are the median acceleration capacity (µ! ) and lognormal dispersion of the
uncertainty (β! ), which are adopted to generate fragility functions at the specified IM:
!"(!" !! )
Pr = ∅ (3)
!!

The likelihood function takes Bernoulli distribution:

!
L= !!![ Pr (IM! )]!! [1 − Pr (IM! )]!! (4)
in which Pr(.) represents the fragility functions, n stands for the total number of NLRHA.
When the seismic demand exceeds the capacity, p=1 and p+q=1. The two parameters µ! and
β! are estimated by maximum likelihood:
!!"# !!"#
= =0 (5)
!!! !!!

These two parameters are used to generate the fragility functions. This method can
include the response data of random samples. In this paper, 200 samples using 200 ground
motions are adopted. This method is labelled as “maximum likelihood’ method.

Fragility anylsis using cloud and regression method

The cloud method is proposed by Cornell et al. (2002). The key step of this method is to
build the probabilistic seismic demand models (PSDMs). PSDM is the relationship between
seismic demand and IM using the power law:
𝐼𝑛(𝑆! ) = 𝐼𝑛 𝑐 + 𝑑𝐼𝑛 𝐼𝑀 (6)
where c and d are regression parameters obtained from the NLRHA, n is the total number of
NLRHA. The dispersion 𝛽! of the PSDM can be estimated by:
!
!!![!"(!"#! )!(!" ! !!"# !"! ]!
𝛽! = (7)
!!!

The fragility function can then be generated as:

 !"(!! !! )
Pr = ∅ (8)
! !
!! !!!

in which Sc is the median capacity, Sd is the median demad, 𝛽! is the capacity dispersion. In
this paper, 𝛽! is set to 0.1 to get a small variation of capacity. This method is labelled as
“cloud and regression” method.

Proposed regression-based method

The proposed method here is a modified version of the cloud and regression method
using a high-order regression techniques. The regression format selected for the Sd and 𝛽! is
modified as
𝐼𝑛(𝑆! ′) = 𝑎1 + 𝑎2(𝐼𝑛 𝐼𝑀 )! + ⋯ + 𝑎𝑛(𝐼𝑛 𝐼𝑀 )! (9)
!
[!"(!"#! )!(!" !! )]!
𝛽! ′(𝐼𝑀! ) = !!!
(10)
!!!

where a1, a2,…, an are regression parameters, n is the order value (in this paper, the value is
set to 3), 𝛽! ′(.) is a function of at IM. This method is labelled as ‘proposed method’.

RESULTS AND DISSCUSION

Results of fragility analysis are presented and discussed from three perspective. In
Section 5.1, the selection of proper IM is discussed. In Section 5.2, the fragility curves are
compared in terms of their closeness to the direct fragility method. The effect of sample sizes
on the fragility curves are also evaluated based on the proposed method in Section 5.3.

Selection of proper IM
To conduct the seismic fragility analysis of cable-stayed bridges, the primary work is to
study which seismic parameters can be used to reasonably express the fragility. Thus, in this
paper, a correlation coefficient is firstly applied between the seismic demands and intensity
measures to test the effectiveness of IMs. The Pearson's correlation (Gibbons and Chakraborti,
2003) is applied in the case of the paper. The Pearson's correlation coefficient between the
two variables X and Y is defined as following
!
!!!(!! !!)(!! !!)
ρ= (11)
! ! ! !
!!!(!! !!) !!!(!! !!)

in which, 𝑥 and 𝑦 are the mean values of Xi, and Yi. In this study, Xi represents the seismic
responses of components of the cable-stayed bridge while Yi represents the intensity
measures.
 1 1

0.9 0.9

0.8 0.8

0.7 0.7
Correlation Coefficients

Correlation Coefficients
0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
PGA PGV Ia Ic VSI SMV Sa(T1) HI PGA PGV Ia Ic VSI SMV Sa(T1) HI
Seismic Parameters Seismic Parameters

(a) (b)
1 1

0.9 0.9

0.8 0.8

0.7 0.7
Correlation Coefficients

Correlation Coefficients

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
PGA PGV Ia Ic VSI SMV Sa(T1) HI PGA PGV Ia Ic VSI SMV Sa(T1) HI
Seismic Parameters Seismic Parameters

(c) (d)
Fig 4 Correlation between maximum demands of four components and intensity measures: (a)
curvature at the bottom section; (b) tower displacement; (b) bearing displacement and (d)
cable forces

Eight seismic parameters are selected to test the effectiveness of IMs, namely PGA,
PGV, Arias Intensity (Ia), Characteristic Intensity (Ic), Velocity Spectrum Intensity (VSI),
Sustained Maximum Velocity (SMV), spectral acceleration at the fundamental period T1
(Sa(T1)) and Housner Intensity (HI). Fig 4 shows the correlation coefficients of four different
components. For sectional curvature, it is demonstrated that the strongest correlation are PGV
and VSI. SMV is at the second position. For bearing displacement, Ia is the best and VSI is
the second. For the tower top displacement, the Sa(T1), PGV and VSI show good correlation.
The next two parameters are SMV and HI. For cable forces, the Sa(T1), PGV and VSI ranked
firstly, secondly and thirdly. In general, among the 8 available intensity measures, PGV and
VSI are the proper IMs for the section curvature of pylon, cable forces, bearing displacement
and the top displacement of tower. Thus, due to convenience, PGV is selected as IM to
generate the fragility curves of cable-stayed bridge.

Comparison of different fragility methods

It is noted that the comparison is conducted in terms of the numerical accuracy of the
four fragility methods, namely ‘IDA and regression’, ’maximum likelihood’, ‘cloud and
regression’ and ‘proposed’ methods. The direct IDA method is used as reference. The four
fragility analysis methods are compared in terms of how close each fragility curve is to the
direct IDA fragility discrete data. Moreover, PGV is considered as the IM for all the four
analysis methods. Due to the reason that the limit states associated with cables and top
displacement of tower are hard to define, only the fragility curves of pylon section and
bearings are generated in this paper.

1.0 1.0
Cloud and regression
Maximum Likehoods
0.8 IDA and regression 0.8
Proposed method
IDA direct method
0.6 0.6
Pr

Pr
0.4 0.4
Cloud and regression
Maximum Likehoods
IDA and regression
0.2
Proposed method
0.2
IDA direct method

0.0 0.0
0 100 200 300 400 0 100 200 300 400
PGV(cm/s) PGV(cm/s)

Fig 5 fragility curves of two components using four different method (a) pylon section
curvature; (b) bearing displacement

Fig. 5 shows the fragility curves obtained for two components using four different
method. The fragility plots in the Fig. 5 demonstrate that the ‘proposed’ regression based
method can provide the good accuracy of fragility curves, while the ‘IDA and regression’ and
‘maximum likelihood’ method also show good results overall (very close to ‘proposed’
fragility curves). The ‘cloud and regression’ method results in impractical fragility plots
compared to the other three methods. The fragility curve from ‘cloud and regression’ method
gives higher fragility values due to the calculation of larger dispersion 𝛽! comparing to other
method. A numerical comparison of these curves, considering the direct IDA discrete data as
reference, are performed using three quantities: sum squared error (SSE) averaged over the
total number of intensity levels (Eavg), coefficient of determination (R2), and Pearson’s
correlation coefficient (ρ). This comparison is summarised in Table 2, which shows clearly
that the proposed method provides the good fragility curves. In terms of computation, the IDA
and regression method is found to be the most demanding, followed respectively by the
maximum likelihood method and the proposed method.

Table 2 Closeness of different fragility estimates to the direct IDA fragility

Method Components Eavg R2 ρ
Pylon section 0.021 0.85 0.98
IDA and regression
Bearing 0.023 0.82 0.97
Maximum likelihood Pylon section 0.00106 0.97 0.99
 Bearing 0.00113 0.95 0.99
Pylon section 0.00121 0.96 0.99
Proposed method
Bearing 0.0015 0.97 0.99
Pylon section 0.032 0.75 0.81
Cloud and regression
Bearing 0.036 0.77 0.83

1.0 1.0

0.8 0.8

0.6 0.6
Pr

Pr
0.4 50 samples 0.4 50 samples
120 samples 120 samples
200 samples 200 samples
0.2 300 samples 0.2 300 samples

0.0 0.0
0 100 200 300 400 0 100 200 300 400
PGV(cm/s) PGV(cm/s)

Fig 6 fragility curves of two components at different sample sizes, namely 50, 120, 200, 300:
(a) pylon section curvature; (b) bearing displacement

Effect of sample sizes on fragility

Conventionally, fragility analysis under seismic loading is a computationally intensive
problem due to the reason that this method needs a large number of samples to get an
accuracy result (Lagaros et al. 2009). Moreover, the proposed regression-based method is
based the modification of the cloud method in which the results of fragility are sensitive to the
sample sizes. Thus, in this study, the effect of sample sizes on fragility is investigated to test
whether the 200 samples yield the convergent results. Fig 6 shows fragility curves of two
components of cable-stayed bridges at different sample sizes, namely 50, 120, 200, 300. It can
be seen from Fig. 6 that different sample sizes have different influence on the different
components of cable-stayed bridge, and the increasing sample size can get more accuracy
results of fragility curves. Fig. 6 also shows that the sample size 200 can get convergent
fragility curves of different components of cable-stayed bridge.

CONCLUSIONS

This paper has investigated the seismic fragility of a typical cable-stayed bridge by
comparing different fragility methods, namely IDA and regression method, maximum
likelihood method and the cloud method using regression method. Among these, a new
regression-based method and high-order polynomial are proposed in this work in order to
compare the results with the existing methods to see whether it is better. The significant
conclusions from this study are:
(1) PGV, VSI and Sa(T1) gives a better correlation (than PGA) with the demand
parameter and should be used if the fragility is estimated independently.
(2) The ‘Proposed’ method of fragility analysis gives the best estimates of the fragility
function. But the cloud and regression method fails to provide realistic estimates of the
fragility function.
(3) The ‘maximum likelihood’ method of fragility analysis gives very good estimates of
the fragility values, while the IDA regression based method estimates fragility values quite
well.
(4) The increasing sample size can get more accuracy results of fragility. The sample
size 200 can get convergent fragility of cable-stayed bridge

REFERENCE

Basoz, N., and Kiremidjian, A. (1999). “Development of empirical fragility curves for
bridges.” Proc., 5th U.S. Conf. on Lifeline Earthquake Engineering, ASCE, New York.
Caetano E, Cunha A, Taylor CA. (2000). “Investigation of dynamic cable-deck interaction in
a physical model of a cable-stayed bridge: Part I modal analysis; Part II seismic
response.” Earthquake Eng. Struct. Dyn., 29(4), 481-521.
Casciati, F., Cimellaro, G.P., and Domaneschi, M. (2008). “Seismic reliability of a
cable-stayed bridge retrofitted with hysteretic devices.” Comput. Struct., 8(6),
1769-1781.
Chang, K. C., Mo, Y. L., Chen, C. C., Lai, L. C. and Chou, C. C. (2004) “Lessons learned
from the damaged Chi-Lu cable-stayed bridge,” J. Bridge Eng. 9(4), 343-352.
Choi, E., DesRoches, R., Nielson, B. (2004). “Seismic fragility of typical bridges in moderate
seismic zones.” Eng Struct, 26(2), 187-199.
Cornell, A.C., Jalayer F., Hamburger R.O. (2002). “Probabilistic basis for 2000 SAC federal
emergency management agency steel moment frame guidelines.” J Struct Eng, 128(4),
526–533.
Ellingwood, B.R., Celik, O.C., Kinali, K. (2007). “Fragility assessment of building structural
systems in Mid-America.” Earthquake Eng Struct Dyn, 36(13), 1935–52.
Feng, Q.H. (2008). “Study on seismic vulnerability and risk probability analysis of
super-large bridge.” Doctoral dissertation. Tongji Univ, Shang Hai, China
Gibbons, J. D., Chakraborti, S. (2003) “Nonparametric statistical inference,” New York,
Basel: Marcel Dekker, Inc.
Hwang, H. H. M., and Jaw, J. W. (1990). “Probabilistic damage analysis of structures.” J.
Struct. Eng., 116(7), 1992-2007.
Karim, K.R., Yamazaki, F. (2001). “Effect of earthquake ground motions on fragility curves
of highway bridge piers based on numerical simulation.” Earthquake Eng. Struct. Dyn.,
30(12), 1839-1856.
Mander, J.B., and Basoz, N. (1999). “Seismic fragility curve theory for highway bridges.”
Proc., 5th U.S. Conf. on Earthquake Engineering (TCLEE), No.16, ASCE, Reston, VA,
31-40.
McKenna, F., Fenves, G. L., and Scott, M. H. (2006). OpenSees: Open system for earthquake
engineering simulation, Pacific Earthquake Engineering Center, Univ. of California,
Berkeley, 〈 http://opensees.berkeley.edu/〉.
Nielson, B. G., DesRoches, R. (2007). "Seismic fragility methodology for highway bridges
using a component level approach." Earthquake Eng. Struct. Dyn., 36(6): 823-839
Pacheco, B. M., Fujino Y., Sulekh, A. (1993) “Estimation curves for modal damping in stay
cables with viscous damper,” J Struct. Eng. 119(6), 1961-1979.
Padgett, J. E., Nielson, B.G., DesRoches, R. (2008). “Selection of optimal intensity measures
in probabilistic seismic demand models of highway bridge portfolios.” Earthquake Eng.
Struct. Dyn., 37(5), 711-725.
Pang, Y.T.; Wu, X.; Shen, G.; Yuan, W.C. 2014. Seismic Fragility Analysis of Cable-stayed
Bridges considering Different Sources of Uncertainties, Journal of Bridge Engineering
19(4):0401301
Ren, W. X., Obata, M. (1999) “Elastic-plastic seismic behavior of long span cable-stayed
bridges,” J Bridge Eng. 4(3), 194-203
Shinozuka, M., Feng, M.Q., Lee, J., and Naanuma, T. (2000a). “Statistical analysis of fragility
curves.” J. Eng. Mech., 126(12), 1224-1231.
Shinozuka, M, Feng, M.Q., Kim, H., Kim, S. (2000). “Nonlinear static procedure for fragility
curve development.” J Eng Mech, ASCE, 126(12), 1287-1295.
Vamvatsikos, D., Fragiadakis, M. (2010). “Incremental dynamic analysis for estimating
seismic performance sensitivity and uncertainty.” Earthquake Engng Struct. Dyn. 39(2),
1-16.