You are on page 1of 13

Finite Elements in Analysis and Design 92 (2014) 80–92

Contents lists available at ScienceDirect

Finite Elements in Analysis and Design


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

Calibration of initial cable forces in cable-stayed bridge based on


Kriging approach
J. Zhang a,b, F.T.K. Au a,n
a
Department of Civil Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong, China
b
Department of Civil Engineering, Hefei University of Technology, Hefei, Anhui Province, China

art ic l e i nf o a b s t r a c t

Article history: Although optimization methods are useful in the preliminary design of cable-stayed bridges for sizing of
Received 16 April 2014 components and getting optimal bridge configurations, further fine-tuning is invariably conducted in the
Received in revised form detailed design thus affecting various sensitive properties such as the deck profile and cable forces. The
21 July 2014
zero-displacement method that adjusts the initial cable lengths repeatedly to achieve the design deck
Accepted 12 August 2014
profile is both computationally intensive and prone to convergence difficulty. Therefore a calibration
Available online 15 September 2014
method is proposed based on a Kriging surrogate model built using the uniform design approach. Apart
Keywords: from establishing the relationship between the bridge deck geometry and the initial cable forces by
Baseline model statistical method, the Kriging model also obviates the need for a large number of repeated finite
Cable-stayed bridges
element analyses. A simple cable-stayed bridge is used to verify the feasibility and accuracy of the
Initial cable force adjustment
proposed method. In real-life cable-stayed bridges with many cables, a staged calibration is implemen-
Kriging model
Surrogate model ted so that the Kriging model is used to identify reasonable initial forces in the critical stay cables so that
the zero-displacement method or similar can be used for further adjustments. Verification shows that
this staged calibration can address not only the deck level tolerance but also the control of cable forces.
& 2014 Elsevier B.V. All rights reserved.

1. Introduction between theoretical predictions and field measurements are


brought within certain acceptable tolerances for monitoring and
Cable-stayed bridges have become more and more popular over condition assessment purposes. A cable-stayed bridge typically
the past 50 years because of their desirable structural and aesthetic comprises the deck, the towers and the cables. As the cables are
characteristics. With the continual development in design methodol- effectively the only structural components that allow site control,
ogy and construction technology, cable-stayed bridges have opened up it is essential to study the distribution of cable forces and their
a new era of long-span structures with spans over 1000 m, in which effects on the subsequent bridge behaviour. In view of this, the
suspension bridges have been predominant until recently. Since the development of a finite element model by suitably adjusting the
1990s, a few notable long-span cable-stayed bridges have also been initial cable forces is an important milestone to provide a reference
built in Hong Kong, including the Kap Shui Mun Bridge, Ting Kau for more accurate analyses afterwards. Although adjustments of
Bridge and Stonecutters Bridge, to meet various transportation needs. cable forces affect the geometry of not only the deck but also the
The finite element method is extensively employed to analyse towers, the vertical profile of the deck in particular is often
the behaviour of sophisticated structures under dynamic excita- accorded more importance in the calibration process while the
tions such as earthquake, wind and vehicular loading [1,2]. geometry of the towers is often controlled by the horizontal
However, in the development of finite element models, various displacements at the top. In the calibration process, the number
simplifying assumptions are normally made because of the com- of control points is meticulously chosen to avoid having an over-
plexity of the actual structures. In the determination of the constrained optimization problem that brings along not only
dynamic and static responses, it is inevitable to have discrepancies tedious computations but also lack of solution in many cases.
between the predictions by the structural model and their corre- Deck profile calibration can be carried out iteratively by
sponding measured values. Therefore the initial finite element systematically updating a set of initial cable forces in each cycle
model needs to be properly calibrated so that the discrepancies until certain criteria are satisfied by the final set of cable forces.
Based on the criteria or approach employed, the method of deck
profile calibration can be categorized into several groups. In the
n
Corresponding author. Tel.: þ 852 2859 2650. extensively used zero-displacement method [3], satisfactory deck
E-mail address: francis.au@hku.hk (F.T.K. Au). profile under permanent loading is obtained by adjusting the

http://dx.doi.org/10.1016/j.finel.2014.08.007
0168-874X/& 2014 Elsevier B.V. All rights reserved.
J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 81

initial cable forces to give zero-displacements especially at the response. Using this, the original sophisticated finite element model
cable anchorages. In the optimization method [4,5], the initial can be replaced by a simplified surrogate model constructed by
cable forces are adjusted based on certain objective functions from statistical approximation [16]. The surrogate model, which can be
the perspective of efficiency or economy such as optimizing constructed after determining the relationship between the variables
shapes of post-tensioning functions using the B-spline curves [6], and the responses using design points [13], has the advantages of
maximum stresses in constraints and minimum stresses in stays high efficiency, easy implementation and adequate accuracy. In the
[7], and optimizing the cable sections based on the load combina- simplest surrogate model, the unknown function of interest is
tions [8]. The force equilibrium method [9] can easily account for approximated by a polynomial with the random errors assumed to
the effect of prestressing and the additional bending moments due be normally distributed with zero mean, independent and identically
to the vertical profile of bridge deck, and therefore it can also distributed. Many alternative surrogate models are available, includ-
address concerns of the long-term time-dependent behaviour. The ing the response surface method, Gaussian process meta-model,
unit force method [10] takes into account all relevant effects in the Kriging model, etc.
design of cable-stayed bridges, including construction sequence, To address the drawbacks of various available methods for deck
second-order behaviour, large displacements, sag effect as well as profile calibration, an efficient Kriging surrogate model is proposed to
time-dependent factors. calibrate the initial cable forces. Compared with the conventional
Although various optimization methods are most useful in the response surface method that requires an understanding of the
preliminary design stage for sizing of components and getting an qualitative tendency of the entire design space, the Kriging model
optimal bridge configuration, further fine-tuning is invariably [18] provides better flexibility of modelling response data with multi-
carried out during the detailed design stage to make it practical ple local extreme values. Having originated from geographical space
for actual erection. For example, steel plates are manufactured to statistics, the Kriging algorithm is a data interpolation scheme to
have certain standard thicknesses, stay cables consist of a number predict unknown values from data at known locations and it is a
of strands of standard sizes, reinforcing bars have certain standard virtually unbiased minimum variance estimation model [19]. The local
sizes, and the variations of deck section is often in steps rather estimation characteristics of the model can predict the function value
than being continuous. The stiffness and mass will be slightly distribution satisfactorily by means of a correlation function. With the
varied. While the final design after adjustments is still reasonably development of the Kriging toolbox based on Matlab-DACE [20], the
close to the optimum, various sensitive properties such as the deck Kriging model has been extensively applied in several computer
profile and cable forces are often affected, thus requiring further response models based on approximate simulation [21–24]. This is
calibration. Therefore, whether or not the design of a cable-stayed particularly useful in the application to “computer experiments” such
bridge has gone through the optimization process, subsequent that the sets of input and response are highly dimensional. For
calibration is indispensable for accurate structural analysis of example, in the practical application to the design of an aerospike
service behaviour and long-term monitoring. nozzle, only the constant “global” Kriging model has been shown to be
In view of the complexity of real-life cable-stayed bridges, cable more accurate than the conventional response surface models [25]. As
force calibration is often achieved by iteration using various commer- the Kriging model is much better at prediction capability than many
cial finite element packages with programming facilities [11] that other fitting methods and it can greatly reduce the computation time,
allow optimization to be carried out under various constraints it is extensively applied in the reliability analysis [26].
considered appropriate. After specifying a set of initial cable forces The complicated relationship between the geometry of the
by approximate equilibrium conditions or certain assumed initial cable bridge deck and the initial cable forces of a cable-stayed bridge
strains or stresses, structural analysis is conducted. The redistributed can be obtained by establishing a Kriging model using statistical
cable forces under permanent loading are extracted and the criteria for methodology. Updating can be carried out with the established
iteration are checked. If the deviations are considered unacceptable, Kriging model by minimizing the residuals between the displace-
the cable strains are modified, and then the optimization process is ments of the deck at the control points and the objective values. For
repeated until the specified tolerance is satisfied. Continuing in this those bridges with a large number of cables, the surrogate model
manner, it is theoretically possible to arrive at a set of admissible cable can be used efficiently together with the traditional method. In the
forces eventually. However modern cable-stayed bridges tend to have present study, a cable-stayed bridge with a moderate number of
many closely spaced small-size cables to facilitate erection rather than cables is studied first. A reasonable number of design points are
just a few large-size cables. As these cable-stayed bridges are identified using uniform design to ensure adequate accuracy of the
invariably structures with high degrees of statical indeterminacy and relationship between the cable forces and deck displacements. The
the interaction among cables is not easily predictable, the determina- outputs at these design points calculated by the finite element
tion of suitable initial values of cable forces is often a critical step for model are taken as the real responses of the bridge for this
the subsequent optimization. Actually, an algorithm may fail to numerical study. In practice, bridge loading tests can be carried
converge if there are too many specified objective functions or if the out to measure the displacements and other responses of interest.
tolerances are made too stringent. Moreover, finite element analysis is The quality of the Kriging model constructed is mainly assessed
carried out repeatedly in the iterative process, thereby substantially based on the accuracy of predictions. The uncertainties of the
increasing the amount of computation required and reducing its identified initial cable forces due to modelling errors and measure-
efficiency. It is therefore desirable to develop an efficient and practical ment noise can be quantified separately by the analysis of variance
method for the calibration process. [27,28] so that a full factorial design of experiments can be applied
to the corresponding factors. In addition, the method is also applied
to a long-span cable-stayed bridge in Hong Kong for verification of
2. Use of surrogate models in sophisticated engineering its applicability to real cases.
problems

In many fields of engineering, the surrogate model or meta-model 3. Calibration of initial cable forces by Kriging surrogate
has been promoted as a promising method for finite element model model
updating [12–16] and damage identification [17]. In essence, the
surrogate model examines various design variables and their The construction of the Kriging model for calibration of initial
responses in order to identify the design variables that give the best cable forces consists of several steps as elaborated below.
82 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

3.1. Design of experiment for constructing a surrogate model mean and variance σ 2l . The covariance between any two inter-
polation points xp and xq can be determined as
When “computer experiments” are used involving complicated
Covðzl ðxp Þ; zl ðxq ÞÞ ¼ σ 2l Rðθ; xp ; xq Þ ð5Þ
numerical models and multiple factors to establish an unknown
relationship between the response and inputs, the efficiency and where Rðθ; xp ; xq Þ is the correlation function between x and xq, p

accuracy of the method depends critically on the design of the and θ is a parameter of the correlation function. The correlation
experiment. Classical experimental designs are mostly based on matrix R has the typical element Rpq given by
models of analysis of variance that typically involve the main
Rpq ¼ Rðθ; xp ; xq Þ ð6Þ
effects, interactions among factors and random errors [29]. An
important objective of an experimental design is to provide a good The adopted toolbox for Kriging model [20] provides 7 choices
estimate of all parameters in the resulting regression model with a of correlation model, namely exponential, general exponential,
realistic number of experiments. Unfortunately, with an increasing Gaussian, linear, spherical, cubic and spline. For example, the
number of factors, both the number of parameters in the regres- correlation function of the Gaussian correlation model is given in
sion model and the required number of experiments increase terms of the coordinate components xpj and xqj of points xp and xq
exponentially. Various methods of experimental design are avail- respectively as
able including the full factorial design, central composite design, n
Latin hypercube sampling, etc. Among various experimental Rðθ; xp ; xq Þ ¼ ∏ expð  θj ðxpj  xqj Þ2 Þ ð7Þ
j¼1
designs that spread the design points, the method of uniform
design [29,30] is an efficient fractional factorial design, and hence For an arbitrary point xn and based on condition of the best
uniform design has often been adopted to seek a reasonable linear unbiased estimation, the predicted lth component y^ l of the
number of design points that are uniformly scattered in the response vector y^ can be determined using the predictor model in
domain in order to achieve the necessary accuracy. The method the toolbox as [20]
of uniform design has been applied to various branches of science n n
and technology [21], including civil engineering [31]. y^ l ¼ ðfðxn ÞÞT βl þðrðxn ÞÞT R  1 ðyl  Fβl Þ ð8Þ
n
where βl is the lth column vector of the optimized coefficient
3.2. Theory of Kriging model matrix of regression βn and r(xn) is the correlation function for the
Kriging predictor, which are given by
The essence of Kriging model applied is briefly described here. n
In a problem with s design sites, each design site is defined by n βl ¼ ðFT R  1 FÞ  1 ðFT R  1 yl Þ ð9Þ
coordinate components while each response has m components.
The design variable matrix X of dimensions s  n can be written in rðxn Þ ¼ ½ Rðθ; x1 ; xn Þ; Rðθ; x2 ; xn Þ; …; Rðθ; xm ; xn Þ T ð10Þ
terms of either the ith original n-dimensional column vector xi
containing components to define each design site or the jth
3.3. Procedure for calibration of initial cable forces
s-dimensional column vector xj of X as
X ¼ ½ x1 ⋯ xs T ¼ ½ x1 ⋯ xn  ð1Þ Based on the field measurements from loading tests of the
bridge, an objective function for optimization can be formulated in
Similarly the response matrix Y of dimensions s  m can be
terms of the residuals of the displacements at certain control
written in terms of either the ith original m-dimensional column
points between the field measurements and estimates from the
vector yi containing response components at each design site or
Kriging model, namely
the jth s-dimensional column vector yj of Y as ( )
h
Y ¼ ½ y1 ⋯ y s T ¼ ½ y 1 ⋯ ym  ð2Þ min fOðC 1 ; …; C g Þg ¼ min ∑ ðKM i ðC 1 ; …; C g Þ  DFM;i Þ2 ð11Þ
i¼1
The relationship between the response matrix Y and the design
variable matrix X can be written as where O is the objective function constructed from the initial cable
forces C1–Cg of the g cables, h is the number of control points
Y ¼ FðXÞβ þZðXÞ ð3Þ
chosen on the deck for calibration, KMi is the displacement at the
where the product of design matrix F(X) and coefficient matrix β ith control point estimated from the Kriging model using the
denotes the regression model that approximates the global trend initial cable forces C1–Cg, and DFM,i is the corresponding displace-
of the design space, and Z(X) is the system deviation considered as ment measured by loading test. The procedures of calibration of
a stochastic process that should be independent and identically initial cable forces by the Kriging model are explained below and
distributed, e.g. normally distributed. For the deterministic also compared with the traditional method in the form of
response, polynomials of zero-order (constant), first-order (linear) flowcharts in Fig. 1.
and second-order (quadratic) are provided for the regression
model in the adopted toolbox for Kriging model [20]. Once the (a) The sampling points of the initial cable forces are generated
order is chosen, the lth response yl can be further expressed with k using the techniques of experiment design. With these input
known functions as sampling points, the corresponding selected displacements at
control points of the deck can be obtained from the finite
yl ¼ fðxÞT βl þ zl ðxÞ ð4Þ
element model using any self-developed or commercial
where the k-dimensional column vector f(x) contains the regres- programme.
sion functions, βl is the lth column of the coefficient matrix of (b) The Kriging model is created with the initial cable forces and
regression β of dimensions k  m, which can be estimated based the displacements at control points of the deck associated
on the sampling points, zl(x) is the system deviation, and the limit with the sampling points of the model.
k depends on the order of regression model chosen and the (c) A suitable optimization function is formulated to minimize the
number of coordinate components n. For constant polynomials, residuals of displacements of the deck at the control points
k is 1; for linear cases, k is (nþ 1); and for quadratic cases, k is between the field measurements and estimates from the
(n þ1)(n þ2)/2. The system deviation zl(x) is assumed to have zero Kriging model.
J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 83

Assume initial Experiment design:


cable forces Select initial samples

Optimization by finite
Finite element
element package
calculation

Extract the redistributed Displacements at control points


cable forces

Kriging modelling

Check if error No
is acceptable Optimization analysis

Yes
Initial cable forces
End

Fig. 1. Comparison between the traditional cable adjustment and the Kriging model: (a) traditional cable adjustment method and (b) Kriging model-based method.

(d) Updating procedure is processed in the established Kriging 4.1. Sensitivity analysis of initial cable forces
model. With the calibrated initial cable forces, the finite
element model can be further used for further analysis of Sensitivity analysis of the initial cable forces is then carried out
the bridge. based on a set of estimated initial cable forces. As long as the
estimated initial cable forces are reasonable, the sensitivity results
are not affected much. Each group of cable force is varied while
4. Example 1: A simple cable-stayed bridge keeping all others unchanged for the evaluation of the associated
changes in displacement at the control points. With 10% increase
The initial cable forces of the simple cable-stayed bridge [32,33] and decrease of each group of the initial cable forces, the varia-
shown in Fig. 2 are calibrated by a Kriging model. Table 1 shows tions of the displacements at the control points are calculated and
the structural parameters of the cable-stayed bridge. Each tower is shown in Table 2. Fig. 3 shows that the deck profile is sensitive to
made up of four segments with different sectional properties. The all cable forces and hence all of them should be considered in deck
deck is monolithic with the tower at the intersection. In view of profile calibration. As the sensitivity to the force in Cable 5 is the
the relatively short lengths, the sag effect of the cables is ignored. highest, if its initial value is closer to the correct value, the results
The finite element model of the bridge is constructed using the will converge faster and be more accurate.
commercial package ANSYS Multiphysics 12.0 [34]. As the cable
forces are to be adjusted, the finite element model may comprise 4.2. Traditional calibration method of cable forces
just the deck and the towers, with the stay cables replaced by their
corresponding forces. Symmetry is assumed in this problem and The traditional zero displacement calibration method by the
control points for calibration are taken as the cable anchorages at finite element method is first performed by an iterative optimiza-
deck level and the middle of the main span. For simplicity in tion procedure. To start with, the finite element model is built up
monitoring the subsequent results, the correct forces of Cables 1–6 with the initial cable lengths as calculated geometrically from the
are assumed to be [11 11 10 10 9 9]  103 kN respectively. The deck drawings and the permanent loading is then applied. The vertical
is modelled by 84 beam elements while the towers are modelled displacements at the deck cable anchorages are adopted as the
by 8 beam elements. Each cable is modelled by a single link criteria for iteration. In each cycle, adjustments to the initial cable
element. From finite element analysis, the corresponding vertical strains are worked out based on the non-zero deck displacements
displacements at the control points from Nodes 2–8 under obtained from the previous cycle for further analysis until the deck
permanent loading can be obtained respectively as [ 0.034436 displacements at the cable anchorages come within the tolerance.
0.0094001  0.0091918  0.13097  0.22358  0.29343  0.33057] However, if the tolerance for iterations is set too strict, the cable
m where upward displacement is taken as positive. forces in many cases may stop at the limits of the allowable range
84 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

Fig. 2. Example 1: A simple cable-stayed bridge.

Table 1
Example 1: Structural parameters of cable-stayed bridge.

Parameter Girder Tower Cable

Modulus of elasticity (GPa) 198,201.6 198,201.6 198,201.6


Moment of inertia (m4) 1.1307 0.2106 (top); 0.3452; 0.4315; 0.5179 (bottom) –
Sectional area (m2) 0.3193 0.2025 (top); 0.2276; 0.2694; 0.2973 (bottom) 0.0420 (exterior); 0.0162 (interior)
Dead load (kN/m) 87.5591 – 3.2251 (exterior); 1.2404 (interior)

Table 2
Example 1: Sensitivity analysis of the input of the initial cable forces.

Cable no. Change in cable Change in deck displacement (%)


force (%)
Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8

1 þ10 11.01 34.10 0.06  2.96  4.11  5.01 38.94


 10  11.01  34.10  0.06 2.97 4.12 5.01 53.54
2 þ10 5.95 21.54 0.06  2.25  3.55  4.91 38.94
 10  5.95  21.54  0.06 2.25 3.55 4.91 53.54
3 þ10  21.65  32.65 0.01  2.28  2.41  1.78 43.93
 10 21.65 32.64  0.01 2.28 2.41 1.78 48.55
4 þ10  9.74  24.48  0.02  2.60  3.77  2.14 43.79
 10 9.74 24.47 0.02 2.60 3.77 2.14 48.69
5 þ10  10.51  54.49  0.00  1.03  0.56  0.15 46.14
 10 10.51 54.48 0.00 1.04 0.56 0.15 46.34
6 þ10  5.19  20.26  0.01  3.13  1.20  0.25 46.09
 10 5.19 20.26 0.01 3.13 1.20 0.25 46.39

60%
Node 3
Node 8

40%
Change rate of the deck displacement

Node 2
20%

Node 5 Node 6 Node 7


Node 4
0

Cable 1 10%
Cable 1 10%
Cable 2 10%
-20% Cable 2 10%
Cable 3 10%
Cable 3 10%
Cable 4 10%
-40% Cable 4 10%
Cable 5 10%
Cable 5 10%
Cable 6 10%
Cable 6 10%
-60%
-300 -250 -200 -150 -100 -50 0
Cable location on the deck (m)
Fig. 3. Example 1: Sensitivity of deck displacement to variation of initial cable forces.

and fail to converge. In addition, the method can only cater for construction, which cannot be predicted by the method. In the
control points at cable anchorages, but those elsewhere (e.g. Node present example, the starting values of initial cable forces are
8) have to be adjusted by proper preset of formwork during taken arbitrarily as [8 8 8 8 8 8]  103 kN. After 100 cycles of
J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 85

Table 3
Example 1: Displacements at control points using the traditional zero displacement method.

Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8

Objective value (m)  0.03444  0.00940  0.00919  0.13097  0.22358  0.29343  0.33057
Traditional method (m)  0.023092  0.011350  0.0092069  0.11946  0.18032  0.23777  0.27355
Discrepancy (%)  32.95 20.74 0.18  8.79  19.35  18.97  17.25

Table 4
Example 1: Cable forces using the traditional zero displacement method.

Cable 1 Cable 2 Cable 3 Cable 4 Cable 5 Cable 6

Objective value (103 kN) 11.00 11.00 10.00 10.00 9.00 9.00
Traditional method (103 kN) 11.3899 11.4797 11.2108 10.9181 8.3326 8.3404
Discrepancy (%) 3.54 4.36 12.11 9.18  7.42  7.33

Table 5 Table 6
Uniform design table of U n16 ð1612 Þ. Application table for U n16 ð1612 Þ.

Level Factor s Column number D

1 2 3 4 5 6 7 8 9 10 11 12 2 1 8 0.0908
3 1 4 6 0.1262
1 1 2 4 5 6 8 9 10 13 14 15 16 4 1 4 5 6 0.1705
2 2 4 8 10 12 16 1 3 9 11 13 15 5 1 4 5 6 9 0.2070
3 3 6 12 15 1 7 10 13 5 8 11 14 6 1 3 5 8 10 11 0.2518
4 4 8 16 3 7 15 2 6 1 5 9 13 7 1 2 3 6 9 11 12 0.2769
5 5 10 3 8 13 6 11 16 14 2 7 12
6 6 12 7 13 2 14 3 9 10 16 5 11
7 7 14 11 1 8 5 12 2 6 13 3 10
8 8 16 15 6 14 13 4 12 2 10 1 9
9 9 1 2 11 3 4 13 5 15 7 16 8
10 10 3 6 16 9 12 5 15 11 4 14 7 Table 7
11 11 5 10 4 15 3 14 8 7 1 12 6 Example 1: Experimental design of Kriging surrogate model for the cable-stayed
12 12 7 14 9 4 11 6 1 3 15 10 5 bridge.
13 13 9 1 14 10 2 15 11 16 12 8 4
14 14 11 5 2 16 10 7 4 12 9 6 3 Index no. Cable forces (103 kN)
15 15 13 9 7 5 1 16 14 8 6 4 2
16 16 15 13 12 11 9 8 7 4 3 2 1 Cable 1 Cable 2 Cable 3 Cable 4 Cable 5 Cable 6

1 5.75 8.00 9.50 12.50 15.50 16.25


2 6.50 11.00 14.00 7.25 13.25 14.75
iteration with an allowable tolerance of 0.005 m, the displace-
3 7.25 14.00 5.75 14.75 11.00 13.25
ments at the control points are given in Table 3, with a maximum 4 8.00 17.00 10.25 9.50 8.75 11.75
discrepancy of displacement at Node 2 of above 32.95%. Although 5 8.75 7.25 14.75 17.00 6.50 10.25
the tolerance is set to be 0.005 m, the displacements at the control 6 9.50 10.25 6.50 11.75 17.00 8.75
7 10.25 13.25 11.00 6.50 14.75 7.25
points cannot converge satisfactorily. The corresponding initial
8 11.00 16.25 15.50 14.00 12.50 5.75
cable forces obtained together with the discrepancies with the 9 11.75 6.50 7.25 8.75 10.25 17.00
objective values are shown in Table 4 for subsequent comparison. 10 12.50 9.50 11.75 16.25 8.00 15.50
11 13.25 12.50 16.25 11.00 5.75 14.00
4.3. Kriging model method of calibration 12 14.00 15.50 8.00 5.75 16.25 12.50
13 14.75 5.75 12.50 13.25 14.00 11.00
14 15.50 8.75 17.00 8.00 11.75 9.50
During the experimental design for sampling points, a reason- 15 16.25 11.75 8.75 15.50 9.50 8.00
able range for all the cable forces C1–C6 of Cables 1–6 respectively 16 17.00 14.75 13.25 10.25 7.25 6.50
is chosen to be between 5  103 and 17  103 kN as
C 1 ; C 2 ; C 3 ; C 4 ; C 5 ; C 6 A ½5  103 ; 17  103  ð12Þ
Uniform design is chosen here because of the efficiency and the range of [5,17]  103 kN, Levels ‘1’, ‘2’ and ‘16’, for example,
accuracy. The uniform design tables designated as Un(qs) or U nn ðqs Þ correspond to cable forces of 5.75  103 kN, 6.5  103 kN and
[30] are used, where U means uniform design, n is the number of 17  103 kN respectively. Therefore using Columns 1, 3, 5, 8, 10
trials, q denotes the number of levels for each factor, and s denotes and 11 of Table 5 and the above interpretation of levels, the cable
the number of columns in the table. The tables with an asterisk forces for the 16 cases are worked out as tabulated in Table 7.
have better uniformity and are therefore preferred. In this exam- Finite element analyses of the 16 cases are carried out and
ple, the uniform design table U n16 ð1612 Þ as shown in Table 5 and the displacement results at various control points are shown in
associated application table in Table 6 are used here. U n16 ð1612 Þ has Table 8. Some plots of the Kriging model constructed are shown
16 levels and can well cope with 6 factors for the 6 groups of cable in Fig. 4, which illustrate the complicated relationship and the
forces with a small deviation D of 0.2518. Table 6 shows that challenge of the problem. Assuming the same starting values of
Columns 1, 3, 5, 8, 10 and 11 of Table 5 should be used to form the initial cable forces of [8 8 8 8 8 8]  103 kN as before, iteration by
design for the 6 factors. As the 16 levels are evenly distributed in sequential quadratic programming can be carried out using the
86 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

Table 8
Example 1: Vertical displacements at control points based on the cable forces in Table 7.

Index no. Vertical displacements at control points (m)

Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8

1 0.09378 0.11873  0.00912  0.09863  0.28429  0.52929  0.60943


2 0.08639 0.08225  0.00916  0.10234  0.27536  0.42776  0.48187
3  0.02080 0.02887  0.00916  0.09553  0.17616  0.29405  0.34522
4  0.02819  0.00760  0.00921  0.09925  0.16724  0.19251  0.21765
5 0.09591 0.05113  0.00913  0.09138  0.14355  0.33706  0.41272
6  0.00486 0.05984  0.00917  0.13633  0.25731  0.38118  0.43201
7  0.01225 0.02337  0.00921  0.14005  0.24838  0.27964  0.30445
8 0.06198 0.03754  0.00923  0.05767 0.01087 0.01285  0.01432
9  0.03347 0.02000  0.00916  0.10786  0.29587  0.45914  0.51536
10 0.04076 0.03417  0.00917  0.02548  0.03662  0.16665  0.22523
11 0.03337  0.00231  0.00921  0.02920  0.02769  0.06512  0.09767
12  0.06740 0.00641  0.00925  0.07415  0.14145  0.10923  0.11696
13 0.05670 0.06514  0.00917  0.06628  0.11776  0.25378  0.31203
14 0.04932 0.02866  0.00922  0.07000  0.10884  0.15224  0.18447
15  0.05788  0.02472  0.00922  0.06319  0.00964  0.01853  0.04781
16  0.06527  0.06119  0.00927  0.06690  0.00071 0.08300 0.07975
Displacement of Node 8 (m)

Displacement of Node 8 (m)

3
10 × 103 ×
3
Ca ×10 Ca ×103
ble ble
for for
ces ) ces )
of C ble 1 (kN of C 3 (kN
ble
abl of Ca abl of Ca
e2 fo rces e4 fo rces
(kN
) Cable (kN
) Cable
Displacement of Node 8 (m)

103 ×
Ca
ble ×103
for
ces (kN)
of C ble 5
abl f Ca
e6
orc es o
(kN
) Ca ble f

Fig. 4. (a) Example 1: Displacement at Node 8 predicted by Kriging model from forces in Cables 1 and 2. (b) Example 1: Displacement at Node 8 predicted by Kriging model
from forces in Cables 3 and 4. (c) Example 1: Displacement at Node 8 predicted by Kriging model from forces in Cables 5 and 6.

Kriging model to calibrate the cable forces efficiently. The calcu- compared with the objective values in Fig. 5. With a finite element
lated displacements at the control points are shown in Table 9, model comprising 104 elements, the Kriging model takes 0.6 s for
while the corresponding initial cable forces are given in Table 10. calibration, while the traditional calibration method takes 32 s.
The displacements at the control points obtained by the Kriging Therefore the Kriging model performs better than the traditional
model and the traditional “zero displacement” method are calibration method not only in accuracy but also in efficiency.
J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 87

Table 9
Example 1: Displacements at control points based on Kriging model.

Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8

Objective values (m)  0.03444  0.00940  0.00919  0.13097  0.22358  0.29343  0.33057
Kriging model (m)  0.03377  0.01077  0.00919  0.13226  0.22248  0.29338  0.33105
Discrepancy (%)  1.95 14.57 0.00 0.98  0.49  0.02 0.15

Table 10
Example 1: Calibrated initial cable forces based on Kriging model.

Cable 1 Cable 2 Cable 3 Cable 4 Cable 5 Cable 6

3
Objective values (10 kN) 11.00 11.00 10.00 10.00 9.00 9.00
Kriging model (103 kN) 10.9590 10.9328 10.1017 10.2148 8.7213 8.7686
Discrepancy (%)  0.37  0.61 1.02 2.15  3.10  2.57

0.1
Table 11
Example 1: Uncertainties of key structural parameters.
0.05
Node 4 Parameter Description Coefficient of variation
Displacement of the control points (m)

Node 3 Et Modulus of elasticity of tower 1%, 3%


-0.05
Node 2 Ig Moment of inertia of girder 0.5%, 2%
Eg Modulus of elasticity of girder 1%, 3%
-0.1
wg Dead load of girder 1%, 3%
Node 5 Ag Cross sectional area of girder 0.5%, 2%
-0.15 Ec Modulus of elasticity of cable 1%, 3%
Ac Cross sectional area of cable 0.5%, 2%
-0.2
Node 6
-0.25
Objective values
Objective value Table 12
-0.3 Traditional zero-displacement method Node 7 Example 1: Statistics of identified initial cable forces due to uncertainties of key
Kriging model structural parameters.
Node 8
-0.35
-300 -250 -200 -150 -100 -50 0 Cable Objective value Mean Standard deviation
Cable location on the deck (m) no. (103 kN) (103 kN) (103 kN)

Fig. 5. Example 1: Displacements at control points compared with objective values. 1 11 11.0722 0.2504
2 11 11.0360 0.2615
3 10 10.0961 0.4627
4.4. Evaluation of effects of uncertainties by Kriging model
4 10 10.2290 0.4193
5 9 8.6971 0.2498
The deck profile calibration process is affected by various 6 9 8.7614 0.2189
uncertainties, including errors of field measurements, uncertain-
ties of structural properties, construction errors, etc. and the
Kriging model provides a convenient tool for evaluating the effects provide stronger concrete. For each level of uncertainty, 20
of uncertainties. Systematic errors will definitely affect the relia- samples of random error are introduced. Hence there are totally
bility of calibration result, but this can only be addressed by 2560 calibration runs as described in Section 4.3. The statistics of
improving the measurement methods. In this study, the effects of the identified cable forces due to uncertainties of key structural
random errors on the identified results are further investigated in parameters are given in Table 12.
respect of the modelling errors and measurement errors [35]. The effects of field measurements of deck levels on deck profile
These random errors are assumed to be normally distributed and calibration are analysed. Two levels of uncertainty of deck level
independent of one another. measurement, namely 1% and 5% COV, are assumed for all deck
Only 7 key structural parameters have been chosen for demon- nodes without rigid support. Strictly speaking, all the deck nodes
stration as shown in Table 11 because they govern the amount of without rigid support are subject to independent random error
subsequent computations. For example, the modulus of elasticity and should be considered together. However it is observed that
of the tower is included as contractors often tend to provide interaction between cable forces and deck levels is stronger within
concrete of higher strength in view of the serious consequences either the left or right half, but interaction between the two halves
associated with substandard concrete. The adoption of slip- is less strong. To focus on the former effects and to keep the
forming helps to control the accuracy of the cross sectional problem for demonstration of reasonable size, symmetry is
dimensions and hence the uncertainties of the cross sectional area assumed for simplicity. Considering two levels of uncertainty of
and moment of inertia may be less significant. For each of the key deck level measurements D2–D8 at Nodes 2–8 respectively and 20
parameters, two levels of uncertainty expressed in terms of samples of random error for each level of uncertainty, there are a
coefficient of variation (COV) are considered. Then a full factorial total of 2560 identification runs for full factorial design of experi-
design of experiments is used involving 27 combinations of input ments. However the Kriging model can be used efficiently without
parameters. All parameters are assumed to be unbiased and hence, any need for further finite element analysis. The statistics of the
for example, the mean modulus of elasticity of the tower should be identified cable forces due to uncertainties of deck level measure-
estimated taking into account the tendency for the contractor to ments are given in Table 13.
88 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

To evaluate the contribution from various factors to the their corresponding degrees of freedom. For the key structural
identified cable forces, one may adopt the F-test [36] that is often parameters and the deck level measurements, the F-statistics of
used in the analysis of variance for comparison with factors of the the mean and standard deviation of the chosen factors are
system deviation. The significance FA of factor A is distributed calculated. Fig. 6 shows that the uncertainty of dead load of girder
approximately as F (fA, fe) given by is the most significant among structural parameters, which is
reasonable as the cables are direct load-carrying components.
SA =f A Therefore uncertainties of the dead load, e.g. density of concrete,
FA ¼  Fðf A ; f e Þ ð13Þ
Se =f e accuracy of surface level of in situ concrete slab, etc., will
significantly affect the identification of cable forces. Similarly
where SA is the sum of squares of system deviation, Se is the sum of
Fig. 7 shows the significance of uncertainties in deck level
squares of deviation of experiment of factor A, and fA and fe are
measurements. Measurements at Nodes 7 and 8 are the most
significant for the exterior cables, i.e. Cables 1 and 2. For the
Table 13
Example 1: Statistics of identified initial cable forces due to uncertainties of deck
interior cables, measurements at Nodes 6 and 8 are the most
level measurements. significant. In comparison, measurements at Nodes 3 and 4 are the
least significant for the identification of cable forces. If instruments
Cable Objective value Mean Standard deviation for deck level measurement are limited, one may therefore allo-
no. (103 kN) (103 kN) (103 kN)
cate the more accurate ones to positions of higher significance.
1 11 11.0822 0.2983
2 11 10.9974 0.3021
3 10 9.8661 0.3815 5. Example 2: Ting Kau bridge in Hong Kong
4 10 10.1302 0.3945
5 9 8.5501 0.2668
6 9 8.6482 0.2464
Ting Kau Bridge is a cable-stayed bridge in Hong Kong with a
total length of 1177 m completed in 1998. The four longitudinal

12000 25000
Cable 1 Cable 2
10000 20000

8000
15000
F 6000 Mean F Mean
STD 10000 STD
4000

5000
2000

0 0
Et Ig Eg wg Ag Ec Ac Et Ig Eg wg Ag Ec Ac

3000 4500
Cable 3 4000 Cable 4
2500
3500
2000 3000
2500
F 1500 Mean F Mean
2000
STD STD
1000 1500
1000
500
500
0 0
Et Ig Eg wg Ag Ec Ac Et Ig Eg wg Ag Ec Ac

1000 1600
900 Cable 5 1400 Cable 6
800
1200
700
600 1000
F 500 Mean F 800 Mean
400 STD STD
600
300
400
200
100 200

0 0
Et Ig Eg wg Ag Ec Ac Et Ig Eg wg Ag Ec Ac

Fig. 6. Example 1: Significance of uncertainties of key structural parameters.


J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 89

girders labelled as ‘1st’, ‘2nd’, ‘3rd’ and ‘4th’ are supported from as shown in Fig. 9 from A to D and E to I for those at the Ting Kau
four inclined cable planes emanated from three towers as shown End and Tsing Yi End respectively. The finite element model of
in Fig. 8. At each end of the girders, several back-stay cables are Ting Kau Bridge is constructed using the commercial package
closely spaced as a group to stabilize the towers. They are labelled ANSYS Multiphysics 12.0. The deck, towers and struts are modelled

2000 6000
1800 Cable 1
5000
Cable 2
1600
1400
4000
1200
F 1000 Mean F 3000 Mean
800 STD STD
2000
600
400 1000
200
0 0
D2 D3 D4 D5 D6 D7 D8 D2 D3 D4 D5 D6 D7 D8

500 1800
450 Cable 3 1600 Cable 4
400 1400
350
1200
300
1000
F 250 Mean F Mean
800
200 STD STD
600
150
100 400

50 200
0 0
D2 D3 D4 D5 D6 D7 D8 D2 D3 D4 D5 D6 D7 D8

350 800

300 Cable 5 700 Cable 6


600
250
500
200
F Mean F 400 Mean
150 STD STD
300
100
200
50 100

0 0
D2 D3 D4 D5 D6 D7 D8 D2 D3 D4 D5 D6 D7 D8

Fig. 7. Example 1: Significance of uncertainties of deck level measurements.

North East South


4th
3rd
2nd
1st
West

Ting Kau Tower Main Tower Tsing Yi Tower


127 m 448 m 475 m 127 m

towards Ting Kau towards Tsing Yi

Approach viaduct

Fig. 8. Example 2: Schematic plan and elevation of Ting Kau Bridge in Hong Kong.
90 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

by beam elements while the stay cables are modelled by link give rise to unreasonable cable forces. In addition, the common
elements taking into account the sag effect, as shown in Fig. 10. expediency of modelling a group of closely spaced cables as an
The flexibility method is used first to obtain the initial values of equivalent cable also requires further calibration.
the cable forces [37]. It is a mathematical approach that involves The traditional zero displacement calibration method is first
developing and solving a system of equations that relate cable performed such that the vertical deviations from the design deck
forces and displacement of control points using the concept profile do not exceed 5 mm. However, even with 30 cycles of
of compatibility. However the sole criterion of compatibility may iteration, the vertical displacements at control points still cannot

End back-stay cables End back-stay cables

A B C D E F G H I J

Cable anchorage

Rocker bearing Rocker bearing

Ting Kau End Tsing Yi End


Fig. 9. Example 2: Configuration of end back-stay cables at ends of the bridge deck.

Fig. 10. Example 2: Finite element model of Ting Kau Bridge.

x 10-3 Traditional method x 10-3 Based on Kriging approach


8 2
Ting Kau End

Tsing Yi End
Ting Kau End

Tsing Yi End

1.5
Displacement of control point (m)

Displacement of control point (m)

6
1 Axis 1000

0.5
4
0

2 Axis 1000 -0.5

-1
0
-1.5

-2
-2
-2.5

-4 -3
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Node number Node number
Fig. 11. Example 2: Displacement of 1st girder of Ting Kau Bridge after calibration: (a) based on traditional zero displacement method and (b) after further adjustment based
on Kriging model.
J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92 91

Based on Kriging approach displacement of the 1st girder under permanent load is shown
1000
in Fig. 11(b) and the cable stresses under permanent loading as
shown in Fig. 12 are all acceptable. Therefore the Kriging model is
900 Axis 1000
instrumental in providing a set of reasonable initial cable forces for
subsequent calibration so as to satisfy not only the deck level
800
tolerance but also the allowable cable stresses.
Cable stress (Mpa)

700
6. Conclusions
600
The traditional zero-displacement method to carry out deck
500 profile calibration of cable-stayed bridges suffers from conver-
gence problems, tedious computations and unpredictable cable
400 forces. A calibration method based on a Kriging surrogate model is
therefore proposed. The Kriging model can be built efficiently
300 using the uniform design approach with a prescribed number of
50 100 150 200 250 300 350
finite element analyses. As opposed to the traditional zero-
Cable no. displacement method that requires time-consuming repetitions
Fig. 12. Example 2: Stresses in cables supporting 1st girder of Ting Kau Bridge after of finite element analyses in iteration, the Kriging model provides
calibration. an effective interpolation technique for iteration without any
further finite element analyses. The Kriging model can be utilized
meet the requirement. For example, Fig. 11(a) shows that the effectively in a staged calibration procedure to provide reasonable
displacement of the 1st girder under permanent load has a estimates of critical cable forces such as those in closely spaced
maximum of around 8 mm, where the nodes are numbered group of cables. With the set of reasonable initial cable forces,
sequentially from the Ting Kau End. To comply with the allowable further calibration can be carried out by iteration to satisfy not
tolerance of 5 mm, the cable forces can be further adjusted only the deck level tolerance but also the allowable cable stresses.
manually but this will be extremely time-consuming.
Before conducting calibration using the Kriging model, it is
essential to determine the key factors for the experimental design. Acknowledgements
With totally 384 cables supporting the bridge deck, it is imprac-
tical to consider all cable forces as key factors in view of the huge The authors gratefully acknowledge the Highways Department
amount of workload and difficulties associated with highly con- of the Hong Kong Government for assistance received in producing
strained optimization problems. It is important to choose the key this paper as well as permission of its publication. Any opinions
factors and obtain reasonable initial values for further iteration. As expressed or conclusions reached in the paper are entirely of the
the bridge is asymmetric, the end back-stay cables play an authors.
important role in sharing the unbalanced cable forces and con-
trolling the displacements of side towers. In addition, if one References
focuses only on minimizing the deviations of deck profile in the
calibration, the stresses of some of the end back-stay cables may [1] W.X. Ren, X.L. Peng, Baseline finite element modeling of a large span cable-
be unreasonably high. Considering the actual construction process stayed bridge through field ambient vibration tests, Comput. Struct. 83 (8–9)
on site and the practicality of deck profile calibration, the cable (2005) 536–550.
[2] S. Sritharana, Priestley M.J. Nigel, Seibleb Frieder, Nonlinear finite element
forces in each group of end back-stay cables for a longitudinal analyses of concrete bridge joint systems subjected to seismic actions, Finite
girder are assumed to be the same. This results in 8 factors for the Elem. Anal. Des. 36 (3–4) (2000) 215–233.
experiment design. The cable stress should be above 0 MPa but not [3] P.H. Wang, T.C. Tseng, C.G. Yang, Initial shape of cable-stayed bridges, Comput.
Struct. 46 (6) (1993) 1095–1106.
exceeding 796.5 MPa, which is 0.45 times of guaranteed ultimate
[4] J.H.O. Negrão, L.M.C. Simões, Optimization of cable-stayed bridges with three-
tensile strength of 1770 MPa. Totally 30 levels of the 8 factors are dimensional modelling, Comput. Struct. 64 (1–4) (1997) 741–758.
implemented with the finite element model and the displace- [5] F. Nieto, S. Hernández, J.Á. Jurado, Optimum design of long-span suspension
ments at 42 control points are worked out for the experimental bridges considering aeroelastic and kinematic constraints, J. Struct. Multi-
discip. Optim. 39 (2) (2009) 133–151.
design. The Kriging model is established accordingly. The control [6] M.M. Hassan, A.O. Nassef, A.A. El Damatty, Determination of optimum
displacements include the 40 vertical displacements at the deck posttensioning cable forces of cable-stayed bridges, Eng. Struct. 44 (2012)
anchorages of the end back-stay cables (Fig. 9) on the 4 long- 248–259.
[7] L.M.C. Simões, J.H.J.O. Negrão, Optimization of cable-stayed bridges with box-
itudinal girders, and the 2 longitudinal displacements at the top of girder decks, Adv. Eng. Softw. 31 (6) (2000) 417–423.
side towers. In solving this highly constrained optimization [8] A. Baldomir, S. Hernandez, F. Nieto, J.A. Jurado, Cable optimization of a long
problem, it may be helpful to adopt relaxed tolerances initially at span cable stayed bridge in la Coruña (Spain), Adv. Eng. Softw. 41 (7–8) (2010)
931–938.
certain critical points and then tighten them gradually. Therefore [9] D.W. Chen, F.T.K. Au, L.G. Tham, P.K.K. Lee, Determination of initial cable forces
the initial tolerances at points E and J in Fig. 9 are set at 20 mm, in prestressed concrete cable-stayed bridges for given design deck profiles
and the initial tolerances at points F and I are set at 10 mm; using the force equilibrium method, Comput. Struct. 74 (1) (2000) 1–9.
[10] D. Janjic, M. Pircher, H. Pircher, Optimization of cable tensioning in cable-
elsewhere the tolerance is taken as 5 mm. Iteration by sequential stayed bridges, J. Bridge Eng. 8 (3) (2003) 131–137.
quadratic programming can be carried out using the Kriging model [11] M.X. Ye, Y.Q. Han, M. Zhang, Research on adjusting cable forces of cable-stayed
to obtain a set of more reasonable forces in the end back-stay bridges based on ANSYS, J. China Railway Soc. 28 (4) (2006) 128–131 (in
Chinese).
cables that satisfy the relaxed set of tolerances. Then with these
[12] T. Marwala, Finite element model updating using response surface method.
updated forces in the end back-stay cables and the other cable Collection of technical papers-AIAA/ASME/ASCE/AHS/ASC Structures, Struc-
forces previously calculated by the flexibility method, the zero tural Dynamics and Materials Conference, vol. 7, 2004, pp. 5165–5173.
displacement method by successive adjustments to the initial [13] W.X. Ren, H.B. Chen, Finite element model updating in structural dynamics by
using response surface method, Eng. Struct. 32 (8) (2010) 2455–2465.
cable lengths in each cycle is carried out until the final tolerance [14] L. Deng, C.S. Cai, Bridge model updating using response surface method and
of 5 mm is met. After this staged calibration process, the genetic algorithm, J. Bridge Eng. ASCE 15 (5) (2010) 553–564.
92 J. Zhang, F.T.K. Au / Finite Elements in Analysis and Design 92 (2014) 80–92

[15] W.X. Ren, S.E. Fang, M.Y. Deng, Response surface-based finite-element model [26] J. Zhang, L.M. Zhang, W.H. Tang, Kriging numerical models for geotechnical
updating using structural static responses, J. Eng. Mech. 137 (4) (2011) reliability analysis, Soils Found. 51 (6) (2011) 1169–1177.
248–257. [27] A. Saltelli, K. Chan, E.M. Scott, Sensitivity Analysis, John Wiley & Sons, New
[16] S. Chakraborty, A. Sen, Adaptive response surface based efficient finite York, NY, 2000.
element model updating, Finite Elem. Anal. Des. 80 (2014) 33–40. [28] B. Moaveni, J.P. Conte, Uncertainty and sensitivity analysis of damage identi-
[17] S.E. Fang, R. Perera, A response surface methodology-based damage identifica- fication results obtained using finite element model updating, Comput.-Aided
tion technique, Smart Mater. Struct. 18 (6) (2009) 065009. Civ. Infrastruct. Eng. 24 (5) (2009) 320–324.
[18] H.Y. Gao, X.L. Guo, X.F. Hu, Crack identification based on Kriging surrogate [29] K.T. Fang, D.K.J. Lin, P. Winker, Y. Zhang, Uniform design: theory and
model, Struct. Eng. Mech.: Int. J. 41 (1) (2012) 25–41. applications, Technometrics 42 (3) (2000) 237–248.
[19] N.A.C. Cressie, Statistics for Spatial Data, John Wiley & Sons, New York, 1993. [30] K.T. Fang, Experimental design by uniform distribution, Acta Math. Appl. Sin. 3
[20] S.N. Lophaven, H.B. Nielsen, J. Sondergaard, DACE a Matlab Kriging toolbox (4) (1980) 363–372 (in Chinese).
[EB/OL]. 〈http://www2.imm.dtu.dk/  hbn/dace/〉 (19.02.2005). [31] J.Y. Li, R. Li, Y. Gao, J. Huang, Aerodynamic optimization of wind turbine airfoils
[21] W. Shyy, N. Papila, R. Vaidyanathan, K. Tucker, Global design optimization for
using response surface techniques, Proc. Inst. Mech. Eng. Part A: J. Power
aerodynamics and rocket propulsion components, Prog. Aerosp. Sci. 37 (1)
Energy 224 (6) (2010) 827–838.
(2001) 59–118.
[32] P.H. Wang, T.Y. Tang, H.N. Zheng, Analysis of cable-stayed bridges during
[22] R. Roy, S. Hunduja, R. Teti, Recent advances in engineering design optimiza-
construction by cantilever methods, Comput. Struct. 82 (4–5) (2004) 329–346.
tion: challenges and future trends, CIRP Ann.-Manuf. Technol. 57 (2) (2008)
[33] J. Cheng, Determination of cable forces in cable-stayed bridges constructed
697–715.
[23] H.S. Rahman, K. Alireza, G. Reza, Application of artificial neural network, under parametric uncertainty, Eng. Comput. 27 (3) (2010) 301–321.
Kriging, and Inverse Distance Weighting Models for estimation of scour depth [34] ANSYS, User's Manual, Revision 8.0, Swanson Analysis System, USA, 2003.
around bridge pier with bed sill, J. Softw. Eng. Appl. 3 (10) (2010) 944–964. [35] W.J. Yan, W.X. Ren, T.L. Huang, Statistic structural damage detection based on
[24] P. McLean, P. Lger, R. Tinawi, Post-processing of finite element stress fields the closed-form of element modal strain energy sensitivity, Mech. Syst. Signal
using dual Kriging based methods for structural analysis of concrete dams, Process. 28 (2012) 183–194.
Finite Elem. Anal. Des. 42 (6) (2006) 532–546. [36] D.C. Montgomery, Design and Analysis of Experiments, 3rd ed., John Wiley &
[25] T.W. Simpson, J.J. Korte, T.M. Mauery, F. Mistree, Kriging Models for Global Sons, Inc, New York, 2006.
Approximation in Simulation-Based Multidisciplinary Design Optimization, [37] Zhang J., Au F.T.K. Establishment of baseline models for long-span cable-stayed
vol. 39(12), American Institute of Aeronautics and Astronautics, pp. 2233– bridges, in: Proceedings of the Second International Postgraduate Conference
2241. on Infrastructure and Environment, vol. 2, 2010, pp. 263–270.

You might also like