Tunnelling and Underground Space Technology 41 (2014) 1423

Contents lists available at ScienceDirect

Tunnelling and Underground Space Technology

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

Reliability analysis of tunnel using least square support vector machine

Hongbo Zhao a,, Zhongliang Ru a, Xu Chang a, Shunde Yin b, Shaojun Li c
a
School of Civil Engineering, Henan Polytechnic University, Jiaozuo 454003, Peoples Republic of China
b
Department of Chemical and Petroleum Engineering, University of Wyoming, Laramie, WY 82071, USA
c
State Key Laboratory of Geomechanics and Geotechnical Engineering, Institute of Rock and Soil Mechanics, Chinese Academy of Sciences, Wuhan, Hubei 430071,
Peoples Republic of China

a r t i c l e i n f o a b s t r a c t

Article history: In the reliability analysis of tunnels, the limited state function is implicit and nonlinear, and is difcult to
Received 24 September 2012 apply based on the traditional reliability method, especially for large-scale projects. Least squares support
Received in revised form 28 October 2013 vector machines (LS-SVM) are capable of approximating the limited state function without the need for
Accepted 11 November 2013
additional assumptions regarding the function form, in comparison to traditional polynomial response
Available online 7 December 2013
surfaces. In the present work, the LS-SVM method was adapted to obtain the limited state function. An
LS-SVM-based response surface method (RSM), combined with the rst-order reliability method (FORM),
Keywords:
is proposed for use in tunnel reliability analysis and implementation of the method is described. The reli-
Reliability analysis
Tunnel engineering
ability index obtained from the proposed method applied to particular tunnel congurations under dif-
Response surface method ferent conditions shows excellent agreement with Low and Tangs (2007) method and traditional RSM
Least squares support vector machine results, and indicates that the LS-SVM-based RSM is an efcient and effective approach for reliability
analysis in tunnel engineering.
2013 Elsevier Ltd. All rights reserved.

1. Introduction known implicitly through a numerical procedure. A major dif-

Stability analysis is crucial in tunnel engineering. Numerical
methods such as the nite element method and the discrete
element method as useful tools have been used for stability anal-
ysis (Jing and Hudson, 2002). Limitations of these methods lie in
lacking considerations of uncertainties in variables such as rock
mass strength and the parameters of the support structure. To
overcome above-mentioned limitations, the probabilistic approach
has been developed for stability analysis of tunnels by combining
with other techniques such as nite element (Hoek, 1998; Li and
Low, 2010; Mollon et al., 2009, 2011; Oreste, 2005; Lv and Low,
2011; Su et al., 2011; Deng et al., 2005). The integration of
analytical solutions and Monte Carlo simulation was applied to a
reliability analysis of circular tunnels (Hoek, 1998). Oreste applied
a probabilistic numerical approach to the design of primary tunnel
supports (Oreste, 2005). Li and Low (2010) analyzed the stability
of circular tunnels subjected to hydrostatic stress by combining
a closed-form solution with a semi-probabilistic rst-order
reliability method (FORM). Mollon et al. (2009, 2011) studied
the reliability of rock tunnel stability using a numerical model.
Zhang and Goh (2012) estimated the stability of underground rock
cavern using reliability method. From the practical point of view,
the limit state surface is not known explicitly, whereas it is often
Corresponding author. Tel./fax: +86 3913987584.
E-mail address: bxhbzhao@hotmail.com (H. Zhao).
culty in performing reliability analysis for realistic tunnel engi-
neering problems is that the limited state function is not
available as an analytical, closed-form function. To address this
problem, the response surface method (RSM) has been employed
to estimate the failure surface using an approximate closed-form
expression (Mollon et al., 2009, 2011; Su et al., 2011; Lv and
Low, 2011).
In traditional polynomial-based RSM, the number of samples
required increase in tandem with the order of polynomial used.
This can be time-consuming for practical engineering problems
when a high-order polynomial is desired given a large number of
input variables. Meanwhile, some studies have pointed out that
it is most important to obtain support points for the response sur-
face very close to or exactly at the limit state g(x) = 0 for reliability
analysis (Kim and Na, 1997; Zheng and Das, 2000; Ouypornprasert
et al., 1989). Articial neural networks (ANN) and support vector
machine (SVM) as the alternative approaches can be used to over-
come the above problem (Bucher and Most, 2008; Zhao et al.,
2009).
With the development of articial intelligence, articial neural
network (ANN) model has been applied for reliability analysis in
engineering. ANN-based response surface is able to avoid the prob-
lem of false design points arising from the use o polynomial
response surface (Bauer and Pula, 2000). Deng et al. proposed an
ANN-based second-order reliability method and an ANN-based
Monte Carlo simulation method and applied in civil structure

0886-7798/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.tust.2013.11.004
 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423
Fig. 1. Flowchart of LS-SVM-based response surface method.
engineering (Deng et al., 2005). Elhewy et al. proposed an ANN-
based response surface method to analyze the reliability of struc-
tures (Elhewy et al., 2006). ANN-based response surface is adopted
to approximate the limit state function and calculate the probabil-
ity of failure through the rst- and second-order reliability meth-
ods and a Monte Carlo simulation technique in order to reduce
the number of stability analysis calculations (Cho, 2009). Other
researchers have applied ANN to reliability analysis by combining
it with Monte Carlo simulation, FORM, response surface method,
etc. (Lopes et al., 2010; Cheng and Li, 2008; Cardoso, 2008). The
performance comparison between the ANN-based RSM and the
polynomial-based RSM showed that the ANN-based RSM is more
efcient and accurate than the polynomial-based RSM (Gomes
and Awruch, 2004; Elhewy et al., 2006). Response surface models
in the form of ANN and SVM have the advantage of providing
high-order approximations with smaller pools of samples com-
pared to polynomial functions of comparable order (Gomes and
Awruch, 2004; Deng et al., 2005; Zhao, 2008; Tan et al., 2011).
However, ANN has some inherent drawbacks, such as its slow con-
vergence, a less generalizing performance, arriving at a local min-
imum, and over-tting problems. An alternative machine learning
method, support vector machines (SVMs), is somewhat similar to
ANN but appears to have more merit (Zhao and Yin, 2009).
Li et al. adopted SVM to structural reliability analysis through
combining with FORM and MCSM. The results showed SVMs was
15
Fig. 2. The curve for verication example and its design point.
more accurate and efcient in comparison to the traditional re-
sponse surface method (Li et al., 2006). Zhao have used SVM-based
FOSM and SVM-based Monte Carlo simulation to analyze the reli-
ability of a slope and a tunnel respectively (Zhao, 2008; Zhao et al.,
2009). In the previous study, Sequential minimum optimization
(SMO)-based SVM algorithm was used to build the SVM model.
SMO is an iteration algorithm which is time-consuming. To sim-
plify the implementation of SVM, least squares support vector ma-
chines (LS-SVM) have been proposed as an alternative to SVM
(Suykens and Vandewalle, 1999).
In this paper, a new approach that combines the merits of the
response surface method and the least square support vector ma-
chine for reliability analysis is proposed. LS-SVM is capable of
approximating the limited state function without more assump-
tion of the function form. Then, the reliability index of tunnels
was calculated by combining with the response surface method
(RSM) and FORM. The paper is structured as follows: in Section 2,
the FORM algorithm is reviewed; Section 3 introduces the basic
theory of LS-SVM and RSM, and the LS-SVM-based reliability anal-
ysis method is proposed; and in Section 4, some applications of the
proposed method to the reliability analysis of tunnels are pre-
sented. The Conclusions are given in Section 5.
2. FORM algorithms applying varying dimensionless numbers
Tunnel stability is affected by uncertain variables such as the
elastic modulus of the rock mass, cohesion, friction angle and
in situ stress. The reliability analysis methods quantify the com-
bined uncertainties in a reliability index as a measure of uncer-
tainty. The HasoferLind index, b, is widely used in reliability
analysis (Hasofer and Lind, 1974), for which the matrix formula-
tion for a correlated normal is given by
q
b min X  lT C 1 X  l; 1
X2F
where X is a vector representing the set of random variables xi; l is
the vector of mean values; C is the covariance matrix; and F is the
failure domain. Eq. (1) gives the minimum distance in units of
directional standard deviations from the mean-value point of the
random variables to the boundary of the limit state surface.
16 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423

Table 1
The model of LS-SVM and tentative point at each step.

Step b ai Tentative design point

1 2 3 4 5 x1 x2
1 2826.2990 0.0385 0.0173 0.0303 0.0144 0.1005 10.0000 9.9000
2 809.0826 0.0121 0.0015 0.0224 0.0079 0.0439 3.5559 9.0682
3 312.9748 0.0153 0.0038 0.0070 0.0025 0.0285 6.2823 2.8115
4 55.1386 0.0033 0.0000 0.0087 0.0011 0.0131 0.5131 4.7305
5 21.1187 0.0029 0.0005 0.0013 0.0016 0.0063 1.5632 0.2100
6 24.5037 0.0036 0.0003 0.0026 0.0008 0.0074 1.6876 1.1330
7 21.8072 0.0041 0.0003 0.0033 0.0006 0.0083 1.8344 1.6794
8 21.2915 0.0041 0.0003 0.0035 0.0006 0.0085 1.8063 1.8221
9 21.3858 0.0041 0.0003 0.0036 0.0006 0.0085 1.7686 1.8705

Table 2
Results and comparison with other methods.

LS-SVM-based Simple Lows

RSM RSM method
Reliability index b 2.2990 2.2973 2.2983
Design point x1 1.7490 1.7280 1.6856
x2 1.8966 1.9305 1.9678
Value of performance 0.03 0.12 0.00
function

Fig. 4. Design point with iteration and its comparison with real value.

Fig. 3. Comparison of LS-SVM-based RSM with traditional RSM.

Low and Tang (1997a,b) presented an alternative interpretation

of the HasoferLind index based on the perspective of an expand-
ing ellipsoid in the original space of the basic random variables,
and expressed the index as:
s
 T  
xi  l i xi  l i
b min R1 ; 2
X2F ri ri
where [R] is the correlation matrix, and ri is the standard deviation
of random variable xi. Fig. 5. Variation of reliability index with the number of iterations.
For correlated non-normals, the ellipsoidal perspective remains
valid if Eq. (2) is rewritten as:
Based on the reliability index, the probability of failure can be
s
 T   evaluated from:
xi  lNi 1 xi  li
N
b min R 3
X2F N
ri ri
N pf  1  /b 4
N N
where l and ri are respectively the equivalent normal mean and
i where /(b) is the cumulative distribution function of the standard
equivalent normal standard deviation of random variable xi. normal variable.
 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423 17

Table 3
Comparison of different kernel functions and their parameters.

RBF kernel function (r = 5) Polynomial kernel function Simple RSM Lows method
d=3 d=4 d=5 d=6
Iteration no. 19 14 12 9 24 14
Reliability index b 2.3201 2.2973 2.2987 2.2990 2.2973 2.2983
Relative error (%) 0.95 0.04 0.02 0.03 0.04

X
N
yx ak Kx; xk b 7
k1

where K(x, xk) is the kernel function, where the general kernel func-
tion is the polynomial K(X, Y) = ((XY) + 1)d, d = 1,2,. . .,n; the radial
2
kernel function is KX; Y expf jXYjr2 g and the radial basis func-
tion kernel (RBF) is K(X, Y) = tan h (/(XY) + h) ak are Lagrange multi-
pliers; and b is the scalar threshold. The values of ak and b are
obtained from:
" #   
0 1T b 0
; 8
1 X c1 I a y

where y = [y1,. . .,yN]; 1 = [1,. . .,1], a = [a1,. . .,aN], and Mercers theo-
rem is applied within the X matrix, X = u(xk)Tu(xl) = K(xk, xl), k,
l = 1, . . ., N. c is the tolerance error. The analytical of a and b is then
given by:
   
b 0
U1 9
a y
Fig. 6. A circular tunnel subjected to hydrostatic far eld stress and uniform where
support pressure. " #
0 1T
U : 10
1 X c1 I
The rst and third terms under the square root sign in Eq. (3)
are the equivalent standard normal vectors. Low and Tang
(1997b, 2007) and Low (2004) presented a practical FORM proce- 3.2. Response surface method
dure using constrained optimization based on the above equation.
They also proposed an efcient algorithm for evaluating FORM Closed-form mechanical models are often unable to predict the
through the use of a varying dimensionless number ni by recasting behavior of typically complex structural systems in rock engineer-
Eq. (3) as: ing. Although a reliability analysis of such systems can be carried
q out by Monte Carlo simulation, the large number of structural anal-
b min nT R1 n 5 yses required results in prohibitively high computational costs. The
X2F
number of analyses required can be minimized by using polyno-
where [n] is a column vector of ni. When the value of ni varies with mial approximations of actual limit states in the reliability analy-
constrained optimization, the corresponding value of xi is automat- sis; such approximations are referred to as the response surface.
ically calculated from: The response surface method (RSM) is a simple mathematical form
xi F 1 /ni  6 that avoids lengthy computations in the probabilistic analyses of
complex systems, and an important technique for the study of
how response variables are related to variations in experimental
3. LS-SVM-based response surface method conditions. The conventional RSM uses polynomial functions to
t the actual performance functions based on sample points se-
3.1. Least squares support vector machine (LS-SVM) lected according to some experimental design. Once the response
surface is obtained, a reliability analysis can be readily carried
LS-SVM (Suykens and Vandewalle, 1999) is an alternative form out using the FORM algorithm in Section 2.
of SVM regression. For a given training set of N data points {xk, yk} Many variables have a signicant inuence on the response of
(k = 1, 2, . . ., N) with input data xk 2 RN and output yk 2 r, where RN rock tunnels. Supposing that the response variable y depends on
is an N-dimensional vector space and r is a one-dimensional vector the input variables x1, x2, x3, . . ., experiments using these variables
space. The LS-SVM algorithm describes the model as: are repeated until the desired level of accuracy is obtained to de-
ne the response surface. The basic response surface procedure is
to approximate the response by an nth-order polynomial with
Table 4 undetermined coefcients, and thus to generate a polynomial
Parameters of random variables.
equation using regression analysis and an approximate linear (or
Rock properties Mean value Standard deviation nonlinear) functional relationship between dependent output y
E (MPa) 373 48 and the input variables:
c (MPa) 0.23 0.068
u () 22.85 1.31 y f x1 ; x2 ; x3 ; ::: e 11
18 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423

Table 5
The LS-SVM model for the two performance functions.

Performance function Iteration step b ai Tentative design point

1 2 3 4 5 6 7 E (MPa) c (MPa) u ()
g1(x) 1 0.4223 0.0000 0.0000 210.3310 366.5636 0.0130 0.0151 576.9226 373 0.2300 22.8500
2 0.1241 0.0000 0.0000 284.7769 566.0485 0.0165 0.0194 850.8613 373 0.1871 23.0841
3 0.1011 0.0000 0.0000 290.5850 582.5635 0.0168 0.0198 873.1851 373 0.1851 23.0562
4 0.0991 0.0000 0.0000 291.1212 584.1415 0.0168 0.0198 875.2993 373 0.1849 23.0556
g2(x) 1 0.0034 0.0000 0.0000 0.1733 0.2463 0.0000 0.0000 0.4197 373 0.2300 22.8500
2 0.0003 0.0000 0.0000 0.3291 0.4717 0.0000 0.0000 -0.8009 246.084 0.1902 22.7289
3 0.0002 0.0000 0.0000 0.3230 0.4637 0.0000 0.0000 -0.7868 259.627 0.1855 22.6671
4 0.0002 0.0000 0.0000 0.3179 0.4559 0.0000 0.0000 -0.7739 259.062 0.1894 22.6538

Table 6
Results and comparison with different methods.

Reliability index Low and Tangs method Reliability index Polynimial RSM Reliability index LSSVMRSM
Design point Design point Design point
E (MPa) c (MPa) u () E (MPa) c (MPa) u () E (MPa) c (MPa) u ()
g1(x) 0.6933 373.0000 0.1855 23.0185 0.6938 373.0000 0.1849 23.0555 0.6938 373.0000 0.1849 23.0555
g2(x) 2.5014 259.3402 0.1884 22.6535 2.5016 259.1484 0.1886 22.6679 2.5016 259.1865 0.1882 22.6750

Fig. 8. The inuence of the far eld stress on the reliability index.
Fig. 7. The effect of support pressure on reliability index (p0 = 2.5 MPa).

Table 7
The design point with different support pressures. Table 8
The design point with different far eld stresses.
LS-SVM-RSM Lows method
LS-SVM-RSM Lows method
pi (MPa) E (MPa) c (MPa) u () E (MPa) c (MPa) u ()
p0 (MPa) E (MPa) c (MPa) u () E (MPa) c (MPa) u ()
0 373 0.1849 23.0555 373 0.1855 23.0185
0.05 373 0.1579 23.1675 373 0.1591 23.1009 1.5 373 0.1029 23.8344 373 0.1046 23.6650
0.1 373 0.1309 23.2683 373 0.1328 23.1687 2.0 373 0.1434 23.3825 373 0.1446 23.2949
0.15 373 0.1040 23.3621 373 0.1067 23.2211 2.5 373 0.1849 23.0555 373 0.1855 23.0185
0.2 373 0.0771 23.4512 373 0.0810 23.2569 3.0 373 0.2259 22.8619 373 0.2260 22.8591

A second-order regression model without interaction terms
containing the two input variables x1, x2 is given by:
y a0 a1 x1 a2 x2 a3 x21 a4 x22 e
where a0, a1, . . ., a4 are regression coefcients, and e represents the
error involved in neglecting other sources of uncertainties.
To obtain these constants in RSM analysis, (2n + 1) experiments
are conducted. Such an experimental design is referred to as fully
saturated, implying that the surface ts exactly at the experimen-
tal points. Central samples are used to determine 2n + 1 samples in
12
traditional polynomial response surface using the following
relationship:
xi li  hi ri 13
 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423 19

3.4. Convergence criterion

In the present work, the following criterion was adopted:

jbi  bi1 j 6 e 16
where bi and bi1 are the reliability indices of the current iteration
and previous iteration, and e was taken to be equal to 0.001 in
the present study.

3.5. Computation procedure

The algorithm of LS-SVM-based response surface method is

summarized in the owchart in Fig. 1 and explained in detail as
follows:

Step 1: Generate sample points

Central sampling is adopted for generating 2n + 1 sample points

Fig. 9. Reliability analysis incorporating different correlations between c and u. (see Eq. (13)).

Step 2: Build samples set for LS-SVM

where li is the mean value of the ith random variable, ri is the stan-
dard deviation of the ith random variable; hi is a coefcient which is
assigned by user (in general hi = 1 or 2).
3.3. LS-SVM-based response surface
The displacement/plastic zone values at each sample point,
determined beforehand using FEM or other mathematical model-
ing technique, are used to determine the performance function.
Sample sets consist of the sample points and their corresponding
performance function values.

The limit state function of the reliability analysis is an implicit,
nonlinear, complex relationship between random variables and the
deformation/plastic zone, exemplied by the response surface
method. Based on the LS-SVM, the response surface can be
expressed as:
gX LSSVMX
where g(X) is the performance function, and X represents the
vectors of the random variables. If the LS-SVM model is used, then:
X
N
gX ak KX; X k b 15
k1
The LS-SVM has more powerful regression capabilities than
polynomial-based response surfaces. It is able to reect the nonlin-
earity in the performance function, and is less constrained by the
number of sample points. In order to build the above relationship
LS-SVM(X), a training process based on the known data set is
needed. The necessary training samples were created in the pres-
ent work by using analytical solutions or numerical analyses (e.g.
FEM model results) to obtain the displacements and/or plastic
deformations of the tunnel corresponding to the given random
variables. In each iteration, samples were built based on central
samples adapted to a traditional polynomial response surface. Cen-
tral samples are used to determine 2n + 1 samples (see Eq. (13)).
Step 3: Construct LS-SVM-based response surface
The LS-SVM approach was used in the present study to approx-
imate the performance function. The RBF kernel function was
adopted. The LS-SVM-based response surface was constructed
using sample sets.
14
Step 4: Compute reliability index and design point
The FORM algorithm of Low and Tang (2007) is used to compute
the reliability index and the design point, based on the LS-SVM-
based response surface.
Step 5: Check convergence
Steps 14 are repeated until the reliability index converges. In
the present work, convergence is said to occur when the absolute
difference between the current and the previous tentative reliabil-
ity index is less than 0.001 (see Section 3.4).
3.6. Verication
For a highly nonlinear limit state function, the performance
function is given by:
Gx x31 x21 x2 x32  18 17

Table 9
The design point with different correlation coefcients.

Correlation coefcient LSSVMRSM Lows method

E (MPa) c (MPa) u () E (MPa) c (MPa) u ()
0.9 373 0.1728 23.7622 373 0.1731 23.7408
0.7 373 0.1799 23.3442 373 0.1805 23.3074
0.5 373 0.1849 23.0555 373 0.1855 23.0185
0.3 373 0.1885 22.8472 373 0.1891 22.8150
0.1 373 0.1913 22.6910 373 0.1917 22.6650
0 373 0.1924 22.6271 373 0.1928 22.6041
20 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423
Fig. 10. Comparison of design point with lognormal distribution.
Fig. 11. Cross-section of horseshoe tunnel.
where x1 and x2 are normally distributed with means of 10.0 and
9.9, respectively, and a standard deviation of 5.0 for both. This
Table 10
Parameters of random variables for horseshoe tunnel.
E (MPa) c (MPa) u () r1 (MPa) r3 (MPa)
Mean value 4600 1 35 10 7
Standard deviation 800 0.2 5 2 1.4
Fig. 12. Numerical model and grid.
Fig. 13. Displacements for horseshoe tunnel.
example behaves in a highly nonlinear manner around the design
point (Fig. 2). The LS-SVM method is very capable of presenting
such a nonlinear relationship, as Fig. 2 shows. The reliability index
was computed using the LS-SVM-based response surface method as
described in Section 3.5. The polynomial kernel function (Eq. (12))
was adopted and its parameter d was selected as equal to 5. The val-
ues of b and ak (Eq. (15)) are listed in Table 1. The reliability indices
and design points obtained by the LS-SVM-based RSM are listed in
Table 2. These are consistent with the results from Low and Tangs
(2007) method and traditional RSM, but the method requires less
iteration to arrive at these results (see Figs. 3 and 4). This indicates
that the LS-SVM-based RSM method is both more effective and
more efcient than the traditional RSM method.
The kernel function and its parameters have a signicant inu-
ence on the results and efciency of reliability analysis (see Fig. 5);
comparisons are listed in Table 3, it can be seen from Table 3 that
the polynomial kernel function is superior to the RBF kernel func-
tion. In the former, the best results were obtained when d = 5,
 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423 21

Table 11
Reliability index and design point for each iteration.

Iteration step Reliability index Tentative design point

E (MPa) c (MPa) u () r1 (MPa) r3 (MPa)
1 2.995 3391.2186 0.8667 30.5559 14.0885 7.3733
2 2.405 3049.0307 0.9723 33.3659 12.6422 7.3122
3 2.350 3344.3563 1.0018 31.1629 13.0209 7.0452
4 2.347 3125.1387 0.9904 32.1783 12.5643 6.9150

Table 12
Results and comparison with different methods of horseshoe tunnel.

Reliability index Design point Number of iteration

E (MPa) c (MPa) u () r1 (MPa) r3 (MPa)
Polynimial RSM 2.341 3338.681 0.945 33.072 13.195 7.109 5
LSSVMRSM 2.347 3125.139 0.990 32.178 12.564 6.915 4
Relative error (%) 0.256 6.396 4.757 2.702 4.783 2.733

where E is the elastic modulus and l is Poissons ratio. The values of

pcr, k and s are obtained from the following equations:

2p0  rc
pcr 20
k1

1 sin u
k 21
1  sin u

ck  1
rc 22
tan u

rc
where s 23
k1
and u is the friction angle, and c is the cohesion.
The performance functions of the reliability index for circular
tunnels are given by:
rp
Fig. 14. Comparison of displacement of tunnel wall between FEM and LS-SVM. g 1 x L  24
r0

uip
rather than the other values of d = 3, 4 or 6; for example, when g 2 x eL  25
r0
d = 6 the reliability index did not converge readily. Selection of
the most appropriate kernel function and its parameters are thus where L is the permissible threshold about the plastic zone, and eL is
shown to be important in LS-SVM-based RSM, but SVM theory does the inwards displacement of tunnel wall. In this study, L = 3 and
not indicate any ideal way of determining these parameters. eL = 0.01; p0, pi, r0 and l were regarded as deterministic variables,
and E, c and u are regarded as random variables. To investigate
4. Applications the proposed LS-SVM-based RSM method, reliability analyses were
carried out for variations of the circular tunnel case.
4.1. Circular tunnel subjected to hydrostatic stress
4.1.1. Comparison of polynomial RSM and Low and Tangs (2007)
Suppose that a circular tunnel is excavated in a continuous, method
homogeneous, isotropic, initially elastic rock mass subjected to In this section, a normal distribution was assumed for the ran-
hydrostatic far eld stress p0 and uniform support pressure pi as dom variables. The statistical values assumed by Hoek (1998) in his
shown in Fig. 6. Monte Carlo simulations were adopted; see Table 4. In addition, c
If pi is less than the critical pressure pcr, a plastic zone exists. and u were assumed to be negatively correlated, with a correlation
According to the MohrCoulomb criterion, the plastic zone radius coefcient of 0.5. For the plastic zone performance function
rp and the inward displacement of tunnel wall uip are given by: p0 = 2.5 MPa, pi = 0 MPa and r0 = 1 m. In the LS-SVM-based RSM,
 1=k1 the RBF kernel was adopted, and r = 3. The values of b, ak and
rp 2p0 s
18 the tentative design points for each iteration are listed in Table 5.
r0 k 1pi s
The reliability indices and design points using the method of Low
   and Tang (2007) are listed in Table 6; this shows that the two
uip 1l rp 2
21  lp0  pcr  21  2lp0  pi 19 methods are in close agreement, and are almost identical to the
r0 E r0
results obtained from applying the polynomial RSM for the two
22 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423

performance functions. This indicates that the proposed method Appendix A. VBA code of LS-SVM algorithm
can replace the polynomial RSM for reliability analysis.

4.1.2. Support pressure Option Explicit

The reliability of the design of a circular tunnel was analyzed for The Least Square Support Vector Machine
cases with different support pressures (see Fig. 7). The reliability
Sub LSSVM()
index increases with increasing support pressure. The reliability
Algorithm of the least square support vector machine,
indices and design points obtained by the LS-SVM-based RSM
solution the linear equation for the value of alpha
agree well with Low and Tangs (2007) method for different sup-
N or 6 is the number of samples
port pressures (see Fig. 7 and Table 7).
Dim MatrixX(1 To 6, 1 To 6) As Double
The reliability of the design of a circular tunnel was analyzed for
Dim xx(1 To 2) As Double
different far eld stresses. The results agree well with those of Low
Dim yy(1 To 2) As Double
and Tang (Fig. 8 and Table 8). The results are similar for the cases of
different correlation coefcients, as seen in Fig. 9 and Table 9. Dim N As Integer
The proposed method was also applied to determine the reli- Dim i As Integer
ability of the tunnel design when a lognormal distribution was Dim j As Integer
adopted. The reliability index of 0.6117 is very close to the Dim k As Integer
0.6109 value obtained for normal distribution, and the values of Dim ii As Integer
the design points agree well (see Fig. 10). Dim jj As Integer
Dim l As Integer
4.2. Horseshoe tunnel
N=6

In this application, the reliability of the design of a horseshoe- The value of the rst row and column of MatrixX() is
shaped tunnel was analyzed using the proposed LS-SVM-RSM computed
method. The span of tunnel is about 5 m (Fig. 11). The mean value MatrixX(1, 1) = 0
of the rock mass parameters and in situ stress are shown in Fig. 11.
For i = 2 To N
The elastic modulus E (MPa), friction angle u (), cohesion c (MPa),
MatrixX(1, i) = 1
major principal stress r1 (MPa) and minor principal stress r3 (MPa)
MatrixX(i, 1) = 1
were regarded as random variables; their values are listed in
Table 10. The inward displacement of the tunnel wall was selected Next i
as the criterion of the stability of the tunnel. A two-dimensional -
Compute the value of MatrixX(A), xx,yy are the ith
nite element model of the tunnel tested the reliability analysis
dimension of samples
(Figs. 12 and 13). The RBF kernel function was adopted, with
For j = 2 To N
r = 3. The reliability indices and design points are listed in
For k = 2 To N
Table 11.
For l = 1 To 2
The reliability index obtained by the proposed LS-SVM-RSM
xx(l) = Range(xi).Cells(j - 1, l).Value
agreed very closely with the values obtained by RSM, with a rela-
yy(l) = Range(xi).Cells(k - 1, l).Value
tive error of about 0.26% (seen in Table 12). The performance of the
Next l
response surface is important in reliability analysis; the LS-SVM-
MatrixX(j, k) = Kf(xx, yy, Range(sigma).Cells.Value) + 1 /
RSM and FEM displacements are compared in Fig. 14, which illus-
Range(gama).Cells.Value MatrixX(j, k) = Kf(XX, YY,
trates the close agreement between the two methods.
sigma) + 1 / gama
Next k
5. Conclusions Next j

The proposed LS-SVM-based RSM method combines the LS- compute the alpha using the alpha=A-1Y
SVM and response surface method. The LS-SVM closely represents alpha = Application.MMult(Application.MInverse(MatrixX),
the nonlinear relationship between the random variables and the Range(yi).Value)
deformation/plastic zone of the tunnel without any assumption put the value of alpha into the cell of variant alpha
of the function form being required. LS-SVM based on RSM method Range(alpha).Value = alpha
calls LS-SVM model and avoids calling numerical simulation model
(such as nite element model) each iteration to decrease the time. End Sub
So, the LS-SVM-based RSM method replaces the polynomial RSM Kernel function of RBF
and improves the reliability analysis efciency. The reliability in- Function Kf(xx, yy, sigma2) As Double
dex obtained by LS-SVM-based RSM was compared with Low and Dim temp As Double
Tangs (2007) method and the traditional RSM method under dif- Dim temp1 As Double
ferent sets of conditions and showed excellent agreement, which temp = 0
indicates that the LS-SVM-based RSM is an efcient and effective For i = 1 To 2
method for reliability analysis in tunnel engineering. The proposed temp = temp + (xx(i) - yy(i)) ^ 5
method can also be applied to other rock engineering contexts. Next i

Acknowledgements temp1 = Sqr(temp) / (2 sigma2 ^ 2)

Kf = Exp(-temp1)
Support by the Nation Key Basic Research Program of China (No. Kf = temp
2013CB036405) and Program of National Fund of Science in China
(No. 41072224, 41172244) are gratefully acknowledged.
 H. Zhao et al. / Tunnelling and Underground Space Technology 41 (2014) 1423
End Function
Compoute the performance function value using the LSSVM
Function Perffunc(xxx) As Double
Dim yyy(1 To 2) As Double
Dim alpha1(1 To 6) As Double
Perffunc = Range(alpha).Cells(1, 1).Value
For i = 1 To 5
alpha1(i + 1) = Range(alpha).Cells(i + 1, 1)
For j = 1 To 2
yyy(j) = Range(xi).Cells(i, j).Value
Next j
Perffunc = Perffunc + alpha1(i + 1) Kf(xxx, yyy,
Range(sigma).Cells.Value)
Next i
End Function
References
Bauer, J., Pula, W., 2000. Applications of neural networks as universal approximators
in geotechnical reliability computations. Stud. Geotech Mech. 22 (34), 103
115.
Bucher, C., Most, 2008. A comparison of approximate response functions in
structural reliability analysis. Probab. Eng. Mech. 23, 154163.
Cardoso, J., 2008. Structural reliability analysis using Monte Carlo simulation and
neural networks. Adv. Eng. Softw. 39 (6), 505513.
Cheng, J., Li, Q., 2008. Reliability analysis of structures using articial neural
network based genetic algorithms. Comp. Method Appl. Mech. 197 (4548),
37423750.
Cho, S.E., 2009. Probabilistic stability analyses of slopes using the ANN-based
response surface. Comput. Geotech. 36, 287797.
Deng, J., Gu, D., Li, X., et al., 2005. Structural reliability analysis for implicit
performance functions using articial neural network. Struct. Saf. 27 (1), 2548.
Elhewy, A., Mesbahi, E., Pu, Y., 2006. Reliability analysis of structures using neural
network method. Probab. Eng. Mech. 21 (1), 4453.
Gomes, H.M., Awruch, A.M., 2004. Comparison of response surface and neural
network with other methods for structural reliability analysis. Struct. Safe. 26,
4967.
Hasofer, A.M., Lind, N.C., 1974. Exact and invariant second moment code format. J.
Eng. Mech. 100 (1), 111121.
Hoek, E., 1998. Reliability of HoekBrown estimates of rock mass properties and
their impact on design. Int. J. Rock Mech. Min. Sci. 35, 6368.
23
Jing, L., Hudson, J.A., 2002. Numerical methods in rock mechanics. Int. J. Rock Mech.
Min. Sci. 39, 409427.
Kim, S.-H., Na, S.-W., 1997. Response surface method using vector projected
sampling points. Struct. Safe. 19, 319.
Li, H.S., Lv, Z.Z., Yue, Z.F., 2006. Support vector machine for structural reliability
analysis. Appl. Math. Mech 27 (10), 11351143.
Li, H.Z., Low, B.K., 2010. Reliability analysis of circular tunnel under hydrostatic
stress eld. Comput. Geotech. 37, 5058.
Lopes, P.A.M., Gomes, H.M., Awruch, A.M., 2010. Reliability analysis of laminated
composite structures using nite elements and neural networks. Compos.
Struct. 92 (7), 16031613.
Low, B.K., 2004. Reliability analysis using object-oriented constrained optimization.
Struct. Saf. 26 (1), 6989.
Low, B.K., Tang, W.H., 1997a. Reliability analysis of reinforced embankments on soft
ground. Can. Geotech. J. 34 (5), 672685.
Low, B.K., Tang, W.H., 1997b. Efcient reliability evaluation using spreadsheet. J.
Eng. Mech. 123 (7), 749752.
Low, B.K., Tang, W.H., 2007. Efcient spreadsheet algorithm for rst-order reliability
method. J. Eng. Mech. 133 (12), 13781387.
Lv, Q., Low, B.K., 2011. Probabilistic analysis of underground rock excavations using
response surface method and SORM. Comput. Geotech. 38, 10081021.
Mollon, G., Daniel, D., Abdul, H.S., 2009. Probabilistic analysis of circular tunnels in
homogeneous soil using response surface methodology. J. Geotech. Geoenviron.
Eng. 135, 13141325.
Mollon, G., Daniel, D., Abdul, H.S., 2011. Probabilistic analysis of pressurized tunnels
against face stability using collocation-based stochastic response surface
method. Probab. Eng. Mech. 137 (4), 385397.
Oreste, P., 2005. A probabilistic design approach for tunnel supports. Comput.
Geotech. 32, 520534.
Ouypornprasert, W., Bucher, C., Schueller, G. I., 1989. On the application of
conditional integration in structural reliability analysis. In: Ang, A. H.-S.,
Shinozuka, M., Schueller, G. I. (Eds.) Proc. 5th int. conf. on structural safety and
reliability. pp. 16831689.
Su, Y.H., Li, X., Xie, Z.Y., 2011. Probabilistic evaluation for the implicit limit-state
function of stability of a highway tunnel in China. Tunn. Undergr. Space
Technol. 26 (2), 422434.
Suykens, J.A.K., Vandewalle, J., 1999. Least squares support vector machine
classiers. Neural Process. Lett. 9 (3), 293300.
Tan, X.H., Bi, W.H., Hou, X.L., et al., 2011. Reliability analysis using radial basis
function networks and support vector machines. Comput. Geotech. 38 (2), 178
186.
Zhang, W., Goh, A.T.C., 2012. Reliability assessment on ultimate and serviceability
limit states and determination of critical factor of safety for underground rock
caverns. Tunn. Undergr. Space Technol. 32, 221230.
Zhao, H., 2008. Slope reliability analysis using a support vector machine. Comput.
Geotech. 35 (3), 459467.
Zhao, H., Ru, Z., Zhang, S., 2009. Application of support vector machine to reliability
analysis of underground engineering. J. Rock Soil Mech. 30 (2), 526530 (in
Chinese).
Zhao, H., Yin, S., 2009. Geomechanical parameters identication by particle swarm
optimization and support vector machine. Appl. Math. Model. 33, 39974012.
Zheng, Y., Das, P.K., 2000. Improved response surface method and its application to
stiffened plate reliability analysis. Eng. Struct. 22, 544551.