Safety Science 50 (2012) 629644

Contents lists available at SciVerse ScienceDirect

Safety Science
journal homepage: www.elsevier.com/locate/ssci

Special Issue Article: The First International Symposium on Mine Safety Science and Engineering

Long-term prediction model of rockburst in underground openings using

heuristic algorithms and support vector machines q
Zhou Jian a,, Li Xibing a, Shi Xiuzhi a,b
a
b

School of Resources and Safety Engineering, Central South University, Changsha 410083, China
Postdoctoral Workstation of Daye Nonferrous Metals Company, Huangshi 435005, China

a r t i c l e

i n f o

Article history:
Available online 22 September 2011
Keywords:
Underground openings
Rockburst
Prediction
Classication
Genetic algorithm (GA)
Particle swarm optimization algorithm (PSO)
Support vector machines (SVMs)

a b s t r a c t
Rockburst possibility prediction is an important activity in many underground openings design and
construction as well as mining production. Due to the complex features of rockburst hazard assessment
systems, such as multivariables, strong coupling and strong interference, this study employs support vector
machines (SVMs) for the determination of classication of long-term rockburst for underground openings.
SVMs is rmly based on the theory of statistical learning algorithms, uses classication technique by introducing radial basis function (RBF) kernel function. The inputs of models are buried depth H, rocks maximum tangential stress rh, rocks uniaxial compressive strength rc, rocks uniaxial tensile strength rt,
stress coefcient rh/rc, rock brittleness coefcient rc/rt and elastic energy index Wet. In order to improve
predictive accuracy and generalization ability, the heuristic algorithms of genetic algorithm (GA) and
particle swarm optimization algorithm (PSO) are adopted to automatically determine the optimal
hyper-parameters for SVMs. The performance of hybrid models (GA + SVMs = GA-SVMs) and
(PSO + SVMs = PSO-SVMs) have been compared with the grid search method of support vector machines
(GSM-SVMs) model and the experimental values. It also gives variance of predicted data. A rockburst dataset, which consists of 132 samples, was employed to evaluate the current method for predicting rockburst
grade, and the good results of overall success rate were obtained. The results indicated that the heuristic
algorithms of GA and PSO can speed up SVMs parameter optimization search, the proposed method is
robust model and might hold a high potential to become a useful tool in rockburst prediction research.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
A rockburst (Ortlepp, 1997) is a sudden and violent expulsion of
rock from the surrounding rock mass. In China, with the increase of
mining depth and conditions become more complex, as well as
more and more large-scale underground projects are under

q
The First International Symposium on Mine Safety Science and Engineering
(ISMSSE2011) will be held in Beijing on October 2629, 2011. The symposium is
authorized by the State Administration of Work Safety and is sponsored by China
Academy of Safety Science & Technology (CASST), China University of Mining &
Technology (Beijing) (CUMTB), Datong Coal Mine Group, McGill University (Canada)
and University of Wollongong (Australia) with participation from several other
universities from round the world, research institutes, professional associations and
large enterprises. The topics will focus on mines safety eld: theory on mine safety
science and engineering technology, coal mine safety science & engineering
technology, metal and nonmetal mines safety science & engineering technology,
petroleum and natural gas exploitation safety science & engineering technology,
mine safety management and safety standardization science & technology, occupational health and safety in mine, emergent rescue engineering technology in
mine, etc.
Corresponding author. Tel.: +86 13723887261; fax: +86 731 88879612.
E-mail addresses: csujzhou@126.com (J. Zhou), xbli@mail.csu.edu.cn (X. Li),
sxzcsu@163.com (X. Shi).

0925-7535/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ssci.2011.08.065

construction in the deep zone of intense tectonic activities, such

as water conservancy and hydropower, transportation, defense
and basic physics experiments. These projects provide an important opportunity for the progress of geo-engineering. However,
serious challenges also exist because of the complicated geological
conditions and potential geological hazards (especially highstrength rockburst) during tunnel construction, causing great
losses of life and property. Only from 2001 to 2007, China has deep
engineering metal mine disaster as a result of accidents amounted
to more than 13,000, the death people exceed 16,000 and resulted
in substantial valuable resources can not mine; In recent years, in
the eld of water and trafc engineering, every year due to disaster-induced deep engineering projects have up to thousands of
accidents, the number of casualties have nearly a thousand people,
many engineering schedule delays more than six months or even
1 year, tens of millions or even billions of machinery and equipment obsolescence, huge economic losses (see http://
www.whrsm.ac.cn/zt/emdct973/xmgk/). And with the increase of
mining depth, in situ stress shows a linear or nonlinear increasing
tendency, ground temperature increases, the osmotic pressure of
groundwater will be further raised, rock become hard and brittle,
deteriorating geological conditions, the risk of rock burst will be

630

J. Zhou et al. / Safety Science 50 (2012) 629644

further increased. Therefore, there is a need for the development of

suitable computational methods for the prediction and control of
rockbursts particularly for a safe and economical underground
excavation for construction or mining in burst-prone ground
(Sun and Wang, 2000; Board and Fairhurst, 1983).
For this case, since the rst record of rockburst appeared at a
tin mine in Britain in 1738, numerous related research works,
concerning about the mechanism, characteristics or type, formation condition and disaster control of rock burst have been conducted by many researchers (Deng et al., 2011; Ortlepp and
Stacey, 1994). For example, Cook et al. (1966) provided a theoretical method of predicting rockburst based on the opinion that violent damage of rock occurs when an excess of energy becomes
available during the postpeak deformation stage. Casten and
Fajklewicz (1993) analyzed the rock-burst risk in the case of the
Dickebank seam coal mines (Casten and Fajklewicz, 1993);
Shivakumar et al. (1996) summarized the spatial distribution
characteristics of rockburst at the Kolar Gold Mines. In order to
understand the rockburst mechanism, Zubelewicz and Mroz
(1983) simulated the dynamic instability of rock burst; Linkov
(1996) studied the rockburst in an unstable rock mass and Jiang
et al. (2010) presented the local energy release rate (LERR) to
simulate the conditions causing rockburst. Kaiser et al. (1996)
developed a handbook to guide rockburst support based on energy release approaches. As CRRP (1996) pointed out, rock burst
is a violent failure phenomenon associated with (not caused by)
a seismic event, which can be classied into three types: faultslip burst, strainburst and conbination burst. Currently, based
on the time and scope of rockburst prediction, the methods for
predicting rockburst hazard degrees can be classied into two
categories: long-term forecasts and short-term small-scale regional forecasts (Peng et al., 2010). Short-term prediction of rock
burst is measured by means of some underground rock scene or
the prediction of rock burst phenomena, determining the exact
location of rock burst occurrence and the specic time, the current short-term forecast model is acquired some useful eigenvalue by the appropriate eld measurement method, then
established a mathematical model to analyze the amount of features, and then prediction the rockburst. Several on-site testing
methods commonly used are: the photoelastic method, method
of drilling bits, the resistance method, seismic parameters and
rock mechanics methods, the microseismic monitoring system
method, microgravity method, geophysical prospecting covers
seismic, electromagnetic, geological radar and acoustic emission
method (Jha and Chouhan, 1994; Mansurov, 2001; Frid, 1997;
Alcott et al., 1998; Dowding and Andersson 1986). For example,
Brady and Leighton recorded a seismicity phenomenon before a
moderate rock burst (Brady and Leighton, 1977) while Tang
et al. introduced the feasibility in principle of monitoring and prediction of rockbursts using microseismic monitoring techniques
during tunnel construction of Jinping II hydropower station (Tang
et al., 2010). Tang and Xia presented a seismological method for
prediction of areal rockbursts in deep mine on the basis of the
seismic source mechanism and unstable failure theory (Tang
and Xia, 2010). Long-term prediction of rockburst refers to a preliminary prediction of rockburst trend and qualitative judgments
in project regions during initial project, currently long-term
prediction of rockburst is mainly used by:
 Theoretical prediction method from various perspectives (Tang
et al., 2010; Shi et al., 2010), such as the strength theory, the
rigidity theory, the burst liability theory, the energy theory,
the instability theory, the catastrophe theory, the bifurcation
theory, the theory of dissipative structures and the theory of
chaos have been proposed to study of deformation localization
and stability of the mechanical system in rock.

 Nonlinear science method, such as Pan et al. (2006) proposed a

catastrophe theory to analysis of circular tunnel rockburst, Xie
and Pariseau (1993) investigated the rockburst mechanism
and prediction methods based on fractal geometry, and bifurcation and chaos theory, etc.
 Criterion values method (Tang et al., 2010, Chen et al., 2009),
such as the energy criterion method, impact orientation criterion method, depth prediction critical, and to a variety of rock
burst of factors considered in comprehensive evaluation. Lately,
Zhang and Fu (2008) tried to establish ve factors comprehensive criterion for strain-mode rockburst and its classication.
 Prediction rockburst method based on the priori knowledge,
which can be extracted features samples and accessed knowledge from the case base, achieving to predict future requirements according to prior knowledge (Peng et al., 2010).
Therefore, many theoretical and numerical models and technical means were developed to analyze the occurrence of rockbursts, it will have a better scientic and practical results to
predict the future of rockburst using examples which has
already the occurrence of rock burst in rock engineering.
To this end, Feng and Wang (1994) described a novel approach to
predict probable rock bursts in underground openings based on
learning and adaptive recognition of neural networks. Subsequently
developed a new articial intelligence methods and theories have
been successfully introduced into the prediction of rock burst. To
achieve good results, Wang et al. (1998) proposed a fuzzy comprehensive evaluation method for rockburst prediction based on fuzzy
mathematics. Recently, in terms of rockburst prediction, Chen
et al. (2009) and Zhang et al. (2010) proposed an extenics evaluation
method for rockburst prediction; Zhao (2005) presented a rockburst
classication method based on support vector machines; Gong and
Li (2007) applied the discriminant analysis method to rockburst prediction. Shi et al. (2010) established an unascertained measurement
classifying model to predict the possibility and classication of rockburst. A efcacy coefcient method for predicting the classication
of rockburst was proposed by Wang et al. (2010). A sher discriminant analysis (FDA) model for the prediction of classication of rockburst in deep-buried long tunnel was established by J. Zhou et al.
(2010). Yang et al. (2010) presented a model of predicting the possibility and classication of rock burst based on the combination of
rough set and fuzzy set theory. These studies offered new ideas
and approaches for rockburst prediction. As an important means,
numerical methods and laboratory rock burst test have been
adopted by many researchers gradually in recent years. A nite element model is proposed by Sharan (2004) to predict the potential
occurrence of rockburst in underground openings. Single-face dynamic unloading tests under true-triaxial condition were carried
out by He et al. (2010) for Paleozonic marine sedimentation limestone samples produced by blocks cored from 1140 m depth in Jiahe
coal mine of China. Zhu et al. (2010) proposed a numerical model to
simulate on rockburst of underground opening triggered by dynamic disturbance using the rock failure process analysis (RFPA).
However, due to the complexity of rock mass and the variety of inuencing factors, it is very difcult to predict spacetime distribution
of rockburst exactly. The results of various prediction methods
should be analyzed comprehensively. Moreover, each method has
its own advantages and disadvantages, and understanding, predicting and controlling the rockbursts still pose a considerable challenge
for underground engineering.
The Support vector machines (SVMs) is an efcient machine
learning (ML) technique derived from statistical learning theory
by Vapnik (1995), and has demonstrated its good performance in
classication, regression, and time series forecasting and predition
(Zhao and Yin, 2009; Kovacevic et al., 2010; Khandelwal, 2010;
Jiang et al., 2011) in geotechnical practice and mining science,

631

J. Zhou et al. / Safety Science 50 (2012) 629644

paving the way for numerous practical applications (Guyon and

Christianini, 1999). For this case, in order to establish a more objective and universal nature of the prediction model, this paper presents a long-term rockburst prediction model for underground
openings based on SVMs theory, and takes inuencing factors of
rockburst as input vectors, the grade of rockburst as output vector,
simulating accurately the complicated nonlinear relations between
the rockburst and its inuencing factors. However, SVMs also has
some factors that affect the prediction performance these factors
are usually set by non-heuristics (trial-and-error method; grid
search method (GSM)). For searching optimal parameters, genetic
algorithm (GA) (Goldberg, 1989; Gunyon et al., 2002; Mib et al.,
2006) and particle swarm optimization (PSO) (Kennedy and
Eberhart, 1995; Ren and Bai, 2010; X. Zhou et al., 2010) heuristic
algorithm are introduced to optimization SVMs parameters, and
then predict the rockburst proneness.
In the current study, SVMs optimized by grid search method
(GSM) and heuristic algorithms (genetic algorithm, GA; particle
swarm optimization, PSO) are proposed to forecast rockburst,
GSM, GA and PSO are to nd the optimal settings of parameters
in SVMs. The novel SVMs forecasting method is researched rstly,
then the rockburst forecasting model is constructed by SVMs optimized by grid search method (GSM-SVMs), genetic algorithm (GASVMs) and particle swarm optimization (PSO-SVMs). Finally, the
rockburst forecasting cases are used to testify the forecasting performance of the proposed model.
2. Details of SVMs classier
2.1. The basic of SVMs model
SVMs has originated from the concept of statistical learning theory pioneered by Vapnik (1995). This study uses SVMs as a classication technique. In this section, a brief introduction about how to
construct SVMs for classication problem is presented. More details
can be found in many publications (Vapnik, 1995; Gunn 1998; Chang
and Lin, 2001). The idea of SVMs can be described as follows:
Considering the training set, (xi, yi), i = 1, 2, . . ., m, x e Xd,
y e {1, +1} can be separated by the hyperplane wT x + b = 0, where
w is the weight vector and b is the bias. If this hyperplane maximizes the margin, then the following inequality is valid for all input data:

wT xi byi P 1;

for all xi ; i 1; 2; . . . ; m

The margin of the hyperplane is equal to 2/||w||, any training tuples

that fall on hyperplanes H1 or H2 (i.e., the sides dening the margin)
are support vectors, as shown in Fig. 1. Thus, the problem is the
maximizing of the margin by minimizing of ||w||/2 subject to formula (1). This is a convex quadratic programming (QP) problem.
Lagrange multipliers (ai > 0, i = 1, . . . , m) are used to solve it:
P
2
T
Jp  m
i1 ai w xi byi  1 kwk =2. After minimizing Jp with
respect to both w and b, the optimal weights are given by:
P

w m
i1 ai yi xi : The dual of the problem can be described in the
following way:

J d a 

m X
m
m
X
1X
ai aj yi yj xi xj
ai :
2 i1 j1
i1

The Lagrange multipliers are only non-zero, when (wTxi + b)yi = 1. The
optimal bias for any support vector xi is given by: b = yi  wTxi. Thus,
the linear decision function is created by solving the dual optimization function (DOF), which can be obtained by the following:

f x sgn

m
X

!
ai yi xi xT

i1

where ai s are optimal Lagrange multipliers.

H1
Support
Vector

X1

Optimal Hyperplane
WX+b=0

H2
Separate
Hyperplane

Margin=2/||W||

X2

Fig. 1. Classication of data by SVMs.

For input data with a high noise level, SVMs uses soft margins
that can be expressed as follows with the introduction to the
non-negative slack variables ni, i = 1, . . . , m:

b wT xi yi P 1  ni

for i 1; 2; . . . ; m

4
Pm

To obtain the OSH, it should be minimized the w C i1 nki 12 kwk2

subject to formula (4), where C is the penalty parameter, which controls the tradeoff between the complexity of the decision function
and the number of training examples, misclassied. In the nonlinearly separable cases, the SVMs map the training points, nonlinearly, to a high dimensional feature space using kernel function
K(xi, yi), where linear separation may be possible. Gaussian radial
basis function (RBF) is one of the kernel functions that are given
by K(x, y) = exp(||x y||2/2g2), where g is the width of the RBF kernel. After a kernel function is selected, the QP problem will become:

J d a 

m X
m
m
X
1X
ai aj yi yj Kxi ; xj
ai
2 i1 j1
i1

where
ai
is
derived
by
ai arg max
Ld ;
P
a
0 6 ai 6 C; i 1; . . . ; m: m
a
y

0:
The
decision
function
is
i
i
i1
accordingly modied as

"
#
m
X


f x sgn
ai yi Kxi ; x b

i1

The Gaussian kernel function is used in the SVMs model because of

excellent nonlinear forecasting performance. Thus, the parameters
C and g are the training parameters of SVMs. So far, many techniques have been used for optimization of these parameters. The
most commons of these optimization techniques are grid search
algorithm (GSM) (Chang and Lin, 2001; Zhang et al., 2009), genetic
algorithm (GA) (Gunyon et al., 2002; Mib et al., 2006) and particle
swarm optimization algorithm (PSO) (Kennedy and Eberhart,
1995; Ren and Bai, 2010; X. Zhou et al., 2010). The approach for
the development of the SVMs-based for prediction rockburst can
be described in Fig. 2. Further detailed mathematical description
over SVMs can be referred from Vapnik (1995), Gunn (1998), Chang
and Lin (2001). In the present study, The SVMs program of the classication problem is constructed using Matlab 7.0 software (http://
www.mathworks.com).
2.2. Selecting SVMs parameters via a grid searching method (GSM)
In the present study we describe a SVMs model using the radial
basis kernel function (RBF). The main model parameters are the
penalty-factor, C, and the RBF kernel deviation, g. The values of
these two parameters greatly affect the training and generalization
capability of the SVMs. C and g can be obtained via a grid searching
method (GSM) as described in the bibliography (Chang and Lin,
2001; Zhang et al., 2009; Li, 2009). The main steps are as follows:

632

J. Zhou et al. / Safety Science 50 (2012) 629644

Begin
Collection of data sets
Data preprocessing
Initial (C, g)

Training set

Testing set

Given a population of paticles with

random position and velocities

SVM model training

New popolation

No

Reproduction
Crossover

Evaluate fitness of particle

Fitness
evaluation

5 fold validation
New (C, g)
Grid search

Mutation
No

GA
Optimization

Satisfy
termination?

Average accuracy

Update particle velocity and position

No

Criteria terminating
Yes
Optimal
parameter(C, g)

Update particle best and global best

Yes

GSM
Optimization

SVM model establishment

Is stop condiction satisfied?

Yes

PSO
Optimal
parameter(C, g) Optimization

Model performance evaluation

Finish
Fig. 2. Research architectures for the proposed SVMs-based approach with GSM, GA and PSO optimization method.

(1) Experiment and testing is an effective method to determine

the range of the parameter pairs, (C, g), with the growth indices 22, . . . , 210 and 210, 28, . . ..
(2) A grid search is used to select a pair of the parameters(C, g).
Then cross footing tests (Chang and Lin, 2001; Zhang et al.,
2009) are used to evaluate them. The collected samples, D,
are divided into T sets {G1, G1,. . ., Gt}. Then pick any T1 sets
as the training sets and use the remaining sets as the test
sets. This process is repeated T times. The generalized performance, P, can be described as formula (7):

T X
X

C N  CxN j^xi 2

i1 N2Gi

where Gi is the ith test set, CN is the sample from the test set,
^xi is the parameter vector when DGi is used as the training
sample and CxN j^xi is the output data.
(3) Grid searching is used to calculate the possible pair of
parameters of P. The optimum parameter pair, (C, g), satises
the condition that P is a minimum. If the learning precision
is insufcient, we construct a new 2D network plane centered on the optimum parameters previously obtained. That
is, we select parameter values closer to optimum so that the
model becomes more precise, as shown in Fig. 2.
2.3. Optimizing the SVMs parameters with genetic algorithm (GA)
The free parameters C and g greatly affect the classication
accuracy of SVMs. However, it is not known beforehand what values of the parameters are appropriate. The genetic algorithm (GA)
(Goldberg, 1989) is a subset of evolutionary computing, and inspired by theory of evolution. GA have been considered with
increasing interest in a wide variety of applications (Gunyon
et al., 2002; Mib et al., 2006). Therefore, GA is used to search for
better combinations of the parameters in SVMs. Based on the
Darwinian principle of survival of the ttest, GA can obtain the

optimal solution after a series of iterative computations. Fig. 2 presents the process of optimizing the SVMs parameters with genetic
algorithm, which is described below (Goldberg, 1989; Gunyon
et al., 2002; Mib et al., 2006).
(1) Encoding SVMs parameters and initialization: In this investigation, free parameters C and g are represented by a chromosome which is composed of binary numbers. Each bit of the
chromosome represents whether the corresponding feature
is selected or not. One in each bit means the corresponding
feature is selected, whereas 0 means it is not selected. We
can choose the length of bit strings representing C, g according to the calculation precision required. Randomly generate
an initial population of chromosomes which represent the
values of parameters in SVMs.
(2) Calculating the tness function: Some of experimental data
are used as validation sample in calculating the tness function. The tness function is based on classication accuracy
of SVMs, which is as follows:

Fitness CAvalidation yt =yt yf   100%

where yt and yf represent the number of true and false classications, respectively.
(3) GA operators: Genetic algorithm uses selection, crossover,
and mutation operators to generate the offspring of the
existing population. Selection is performed to select excellent chromosomes to reproduce. Based on tness function,
chromosomes with higher tness values are more likely to
yield offspring in the next generation by means of the roulette wheel. Crossover is performed randomly to exchange
genes between two chromosomes by means of single-point
crossover principle, offspring replaces the old population
and forms a new population in the next generation by the
three operations, the evolutionary process proceeds until
stop conditions are satised.

633

J. Zhou et al. / Safety Science 50 (2012) 629644

Table 1
Criteria and valuation of qualitative indexes in classication of rock burst category.
Reference

Index

Equations

None
rockburst

Wu and Zhang (1997)

Hucka and Das (1974)
Tao (1988)
South African experience
(Singh, 1988)
Ryder (1987)
Neyman et al. (1972)

Failure duration index

Brittleness of rocks

DT
B1 = (rc  rt)/(rc + rt); B2 = sinu
rc/r1
ERR = Uk/U0

>500
>14.5
<3.5%

ESS=|s|rntanud
Wet = Ee/Ep
Ku = U/Ul

Energy release rate (ERR)

Excess shear stress (ESS)

Strain energy storage
index Wet
Brittle deformation
coefcient Ku
Criterion of elastic stain
energy We
Motyczka (Singh, 1988)
Bursting efciency rdtio g
Aubertin et al. (Tang, 2000) Brittleness index modied
(BIM)
Tang (2000)
Burst potential index (BPI)
Hoek and Brown (1997)
Barton (Tang, 2000)

Russenes (1974)
Turchaninov et al. (1972)
Tang and Wang (2002)

Lu (J. Zhou et al., 2010)

Shan Z G (J. Zhou et al., 2010)
Hou (1989)
Pan and Li (2002)
Wang et al. (1998)

Zhang and Fu (2008)

RQD index
Stress coefcient
Rock brittleness index k
Stress index S
Intact coefcient of
rockmass KV
rh/rc P Ks
Lithology criteria
Depth prediction critical
Depth prediction critical
Stress coefcient
Rock brittleness coefeient
Elastic energy index
Five factors
comprehensive criterion

Light
rockburst

Medium
rockburst

Strong
rockburst

50500

650

5.514.5
3.54.2%

2.55.5
4.24.7%

62.5
>4.7%

<5

515

>15

62

26

69

>9

We = rc /2Eu

<40

4040

100200

P200

g = (Et/Es)  100
BIM = F1/F2

3.2%

3.23.8%

3.84.4%

4.4%

rh/rc
rc/rmax
rt/rmax

0.34

<25
60.2
60.3
<20
<0.15
<0.55

0.42
>510
>0.33
2550
0.20.3
0.30.5
2075
0.150.20
0.550.65

0.56
2.55
0.160.25
5070
0.30.55
0.50.8
75130
0.150.25
0.650.75

0.70
<2.5
<0.16
>70
>0.55
>0.8
>130
>0.25
>0.75

Ks

0.25
0.30

0.5
0.40

0.75
0.45

1.00
0.60

rh/rc > Ks
hcr = 0.318rc (1  l)/(3  4l)c
hcr = rc (1  sinu) k [1 + (1 + E/k)1/(1sinu)  E/k]/2Ec
rh/rc
rc/rt
Wet

<0.3
>40
<2.0

0.30.5
4026.7
2.03.5

0.50.7
26.714.5
3.55.0

>0.7
<14.5
>5.0

60.15
60.2
<15
<2.0
60.55

0.150.20
0.20.3
1518
2.03.5
0.550.60

0.200.40
0.30.55
1822
3.55.0
0.600.80

>0.4
>0.55
>22
>5.0
>0.80

rh/rc
(rh + rL)/rc
(rc  ef)/(rt  eb)

rt/rc

r1/rc
rh/rc
rc/rt
Wet
Kv

Note: rh is the maximum tangential stress of surrounding rock, MPa, rc is the axial stress of surrounding rock, MPa, rL is the maximum in situ stress of engineering
area, MPa, r1 is the uniaxial compressive strength of rock, MPa; rt is the uniaxial tensile strength of rock, MPa, Eu is the unloading tangential modulus, MPa; Et is throw
energy of rock fragments after failure of a specimen under uniaxial compression; Es is maximum elastic strain energy, kJm3; Ee is the dissipated energy in the creation
of microfracture and plastic deformation of the rock in one cycle of loading, kJm3; Ep is the elastic energy stored in the rock through loading up to rA and unloading,
kJm3; ef strain before peak; eb strain after peak; l is the poissons ratio of rock; k is the lower modulus, MPa; c is the rock density, KNm3; rn is the normal stress
at the slipping point, MPa; ud is the dynamic friction angle, ; F1 is the area surrounded by strain e axis and before the peak stress axial strain curve; F2 is the area
surrounded by strain e axis and after the peak stress axial strain curve; rmax is the maximum in situ stress, MPa; We is the linear elastic energy, kJ/m3; Es is the
unloading tangential modulus, MPa; u is the internal friction angle of rock, ; Dt is the indicator of dynamic rock failure time, ms; U is the total peak strength of
before rock deformation; Ul is the permanent deformation before peak or plastic deformation; Uk is the rock kinetic energy with destructive ejection, kJm3; U0 is the
maximum elastic strain energy, kJm3; Kv is the rockmass intact coefcient.

2.4. Parameters selection of SVMs based on Particle swarm

optimization (PSO)
PSO (Kennedy and Eberhart, 1995; Ren and Bai, 2010; X. Zhou
et al., 2010) is a populated search method, which derives from
the research for the movement of organisms in a bird ocking or
sh schooling. Similar to genetic algorithms, PSO performs
searches using a population (called swarm) of individuals (called
particles) that are updated from iteration to iteration (Kennedy
and Eberhart, 1995; Ren and Bai, 2010; X. Zhou et al., 2010). The
process of optimizing the SVMs parameters with PSO is presented,
which can be summarized as followings:
(a) Particle initialization and PSO parameters setting: Generate initial particles, the particles are composed of the
parameters C and g. Set the PSO parameters including

number of particles, particle dimension, number of maximal iterations, constraint factor b, and acceleration coefcients c1 and c2, etc.
(b) Fitness denition and evaluation: Compute the tness
function value of each particle. Take current particle as
individual extremum point of every particle and do the
particle with minimal tness value as the global extremum
point.
(c) Stop condition checking: The evolutionary process proceeds
until stopping criteria (maximum iterations predened or
the error accuracy of the tness function) are met. Otherwise, go to step (b).
(d) The proposed approaches for SVMs parameter optimization,
i.e. GA-SVMs and PSO-SVMs, are illustrated in Fig. 2, and
both of them were developed as relatively fast alternatives
for the time-consuming GSM approach.

634

J. Zhou et al. / Safety Science 50 (2012) 629644

Table 2
Basic data for rockburst analysis of some underground projects around the world.
No

Engineering

Rock type

H/m

rh

rc

rt

(Mpa)

(Mpa)

(Mpa)

rh/rc

rc/rt

Wet

Actual
conditions

References
Wang
et al.
(1998)

Diversion Tunnels of Yuzixi

Hydropower Station

Granodiorite

200

90.00

170.00

11.30

0.530

15.040

9.00

Moderate

2# Sub Tunnel of Ertan

Hydropower Station
Underground Cavern of
Taipingyi Hydropower Station
Underground powerhouse of
Laxiwa Hydropower Station
Diversion Tunnels of Tian
shengqiao -II Hydropower
Station
Underground Powerhouse of
Nor- wegian Sima Hydropower
Station
Diversion Tunnels of Swedish
Vietas Hydropower Station
Japanese Guanyuk Tunnel
Diversion Tunnels of Jingping
Hydropower Station
Italian Raibl Lead Zinc Sulde
Working
Soviet Rasvumchorr Workings
Cooling Diversion Tunnels of
Swedish Forsmark Nuclear
Power Station
Norwegian Heggura Road
Tunnel
Norwegian Sewage Road
Tunnel
Underground Cavern of Lijiaxia
Hydropower Station
Underground Cavern of
Pubugou Hydropower Station
Underground Cavern of
Longyangxia Hydropower
Station
Underground Cavern of Lubuge
Hydropower Station
Qinling Tunnel of Xikang
Railway Dyk77 + 176
Qinling tunnel of Xikang
railway T1
Qinling tunnel of Xikang
railway T2
Qinling Tunnel of Xikang
Railway Dyk72 + 440
Qin-ling Tunnel

Syenite

194

90.00

220.00

7.40

0.410

29.730

7.30

Light

Granodiorite

400

62.60

165.00

9.40

0.380

17.530

9.00

Light

Granite

300

55.40

176.00

7.30

0.320

24.110

9.30

Moderate

Dolomitic limestone

400

30.00

88.70

3.70

0.340

23.970

6.60

Moderate

Granite

700

48.75

180.00

8.30

0.270

21.690

5.00

Moderate

Quartzite

250

80.00

180.00

6.70

0.440

26.870

5.50

Light

Quartz diorite
Marble

890
150

89.00
98.60

236.00
120.00

8.30
6.50

0.380
0.820

28.430
18.460

5.00
3.80

Moderate
Moderate

Lead and zinc ore

108.40

140.00

8.00

0.770

17.500

5.00

Strong

Ni nephelineP nepheline
Gneissic granite

57.00
50.00

180.00
130.00

8.30
6.00

0.320
0.380

21.690
21.670

5.00
5.00

Moderate
Moderate

Granitic gneiss

62.50

175.00

7.25

0.360

24.140

5.00

Moderate

Granite

75.00

180.00

8.30

0.420

21.690

5.00

Moderate

Biotite angle, ash plagioclase schist

11.00

115.00

5.00

0.100

23.000

5.70

None

Diorite granite

43.40

123.00

6.00

0.350

20.500

5.00

Moderate

Granite

18.80

178.00

5.70

0.110

31.230

7.40

None

Limestone

34.00

150.00

5.40

0.230

27.780

7.80

None

Granite

56.10

131.99

9.44

0.430

13.980

7.44

Moderate

Granite

54.20

134.00

9.10

0.400

0.147

7.10

Moderate

Granite

70.30

128.30

8.70

0.550

0.148

6.40

Moderate

Granite

60.70

111.50

7.86

0.540

14.190

6.16

Strong

3
4
5

7
a

8
9
10
11
12

13
14
15
16
17a

18
19
20
21
22
23

Migmatite

<1600

54.20

134.00

9.09

0.404

15.000

7.08

Moderate

24
25
26

Qin-ling Tunnel
Kuocang Mountain Tunnel
Riverside Hydropower Station
diversion tunnel

Migmatite
Crystal tuff
Sandstone

<1600
204
203

70.30
35.00
157.30

129.00
133.40
91.23

8.73
9.30
6.92

0.547
0.260
0.580

11.400
14.340
13.180

6.43
2.90
6.27

Moderate
Light
Strong

27

Station

Dolomite

827

148.40

66.77

3.81

0.450

17.530

5.08

Light

Station

Ore

896

132.10

51.50

2.47

0.390

20.860

4.63

Moderate

Station

Red Shale

1117

127.90

35.82

1.24

0.280

28.900

3.67

Light

Station

Sandstone

1124

107.50

21.50

0.60

0.200

36.040

2.29

None

Station

Dolomite

1140

96.41

18.32

0.38

0.190

47.930

1.87

None

Station

Ore

983

167.20

110.30

8.36

0.660

13.200

6.83

Strong

Station

Red shale

853

118.50

26.06

0.77

0.220

33.750

2.89

Light

34
35

Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Riverside Hydropower
diversion tunnel
Huize LeadZinc Mine
Jinchuan 2nd Mine

Sandstone
Granite

920
1000

34.15
60.00

54.20
135.00

12.10
15.04

0.630
0.444

4.480
8.976

3.17
4.86

Light
Light

36
37

Jinchuan 2nd Mine

Jinchuan 2nd Mine

Marble
Migmatite

1000
1000

60.00
60.00

66.49
106.38

9.72
11.20

0.902
0.564

6.841
9.498

2.15
6.11

Light
Light

28
29
30
31
32
33

Su et al.
(2010)

Bai et al.
(2002)

Zhang
et al.
(2010)

Yi et al.
(2010)

635

J. Zhou et al. / Safety Science 50 (2012) 629644

Table 2 (continued)
No

Engineering

38
39
40
41
42
43
44
45a
46
47
48
49
50
51
52
53
54
55
56
57

67
68
69
70
71
72
73
74

75

76

77

78

79

80a

81

82

H/m

rh

rc

rt

(Mpa)

(Mpa)

(Mpa)

rh/rc

rc/rt

Wet

Actual
conditions

Jinchuan 2nd Mine

Jinchuan 2nd Mine
Jinchuan 2nd Mine
Ma Luping mine

Peridotite
Lherzolite
Amphibolite
Sandstone

1000
1000
1000
750

60.00
60.00
60.00
63.80

86.03
149.19
136.79
110.00

7.14
9.30
10.42
4.50

0.697
0.402
0.439
0.580

12.050
16.040
13.130
24.400

2.85
3.50
2.12
6.31

Light
Light
Light
Moderate

Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Ma Luping
Beiminghe

Dolomite
Phosphate rock
Red Shale
Sandstone
Dolomite
Phosphate rock
Red shale
Sandstone
Dolomite
Phosphate rock
Red shale
Limestone

750
750
750
700
700
700
700
600
600
600
600
510

2.60
44.40
13.50
70.40
3.80
57.60
19.50
81.40
4.60
73.20
30.00
15.20

20.00
120.00
30.00
110.00
20.00
120.00
30.00
110.00
20.00
120.00
30.00
53.80

3.00
5.00
2.67
4.50
3.00
5.00
2.67
4.50
3.00
5.00
2.67
5.56

0.130
0.370
0.450
0.640
0.190
0.480
0.650
0.740
0.230
0.610
1.000
0.283

6.670
24.000
11.200
24.400
6.670
24.000
11.200
24.400
6.670
24.000
11.200
9.680

1.39
5.10
2.03
6.31
1.39
5.10
2.03
6.31
1.39
5.10
2.03
1.92

None
Light
Light
Moderate
None
Moderate
Moderate
Strong
None
Moderate
Strong
None

Diorite
Iron ore
Skarn
Dolomitic limestone

510
510
510
225

88.90
59.82
32.30
30.10

142.00
85.80
67.40
88.70

13.20
7.31
6.70
3.70

0.627
0.697
0.479
0.340

10.700
11.700
10.100
23.970

3.62
2.78
1.10
6.60

Strong
Moderate
None
Strong

Granite
Limestone
Clay sandstone
Marble
Limestone
Diorite
Granite
Diastatite anorthose
Breccia marble

375
435
250
100
300
330
223
425
<2520

18.80
34.00
38.20
11.30
92.00
62.40
43.40
11.00

171.50
149.00
53.00
90.00
263.00
235.00
136.50
105.00

6.30
5.90
3.90
4.80
10.70
9.50
7.20
4.90

0.110
0.230
0.720
0.130
0.350
0.270
0.320
0.100
0.620

27.220
25.250
13.590
18.750
24.580
24.740
18.960
21.430
20.000

7.00
7.60
1.60
3.60
8.00
9.00
5.60
4.70
3.10

None
Light
None
None
Light
Strong
Strong
None
Moderate

Grey-white marble

<2520

0.670

26.800

0.85

Light

<2520

0.900

25.700

0.90

Strong

<2520

0.830

28.900

3.20

Strong

<2520

0.930

28.900

3.20

Strong

<2520

0.740

28.900

3.20

Moderate

<2520

1.410

19.200

3.10

Strong

<2520

0.790

22.000

2.00

Moderate

0.559

20.400

2.00

Light

0.464

20.400

2.00

Light

0.237

26.800

0.85

Light

0.290

26.800

0.85

Light

Gray-white marble

0.634

19.700

0.85

Moderate

Granophyric marble

0.488

19.700

2.30

Light

0.436

19.700

2.30

Light

Granophyric marble

0.842

19.700

2.30

Moderate

Granophyric marble

0.417

19.700

2.30

Light

mine
mine
mine
mine
mine
mine
mine
mine
mine
mine
mine
iron mine

Beiminghe iron mine

Beiminghe iron mine
Beiminghe iron mine

58
59
60
61
62
63
64
65
66

Rock type

Section 11 of diversion tunnel

JinpingIIhydropower station
Section 11 of diversion tunnel
JinpingIIhydropower station
Section 22 of diversion tunnel
JinpingIIhydropower station
Section 33 of diversion tunnel
JinpingIIhydropower station
Section 44 of diversion tunnel
JinpingIIhydropower station
Section 55 of diversion tunnel
JinpingIIhydropower station
Section 66 of diversion tunnel
JinpingIIhydropower station
Section 77 of diversion tunnel
JinpingIIhydropower station
Long exploratory tunnel
1 + 693 of Jinping II
hydropower station
Long exploratory tunnel
1 + 731 of Jinping II
hydropower station
Long exploratory tunnel
0 + 568 of Jinping II
hydropower station
Long exploratory tunnel
0 + 600 of Jinping II
hydropower station
Long exploratory tunnel
2 + 215 of Jinping II
hydropower station
Long exploratory tunnel
1 + 560 of Jinping II
hydropower station
Long exploratory tunnel
1 + 640 of Jinping II
hydropower station
Long exploratory tunnel
3 + 390 of Jinping II
hydropower station
Long exploratory tunnel
3 + 580 of Jinping II
hydropower station

Mica marble

Mica marble

46.40

100.00

4.90

Gray-white marble

Gray-white marble

Granophyric marble

23.00

46.20

80.00

105.00

3.00

5.30

References

Yang et al.
(2010)

Zhang and
Li (2009)

Feng et al.
(1994)

Liang
(2004)

(continued on next page)

636

J. Zhou et al. / Safety Science 50 (2012) 629644

Table 2 (continued)
No

Engineering

83

84

85

86

87

Cangling tunnel K98 + 152

K98 + 637
Cangling tunnel K98 + 637
K99 + 638
Cangling tunnel K99 + 638
K100 + 892
Erlang Mountain Tunnel level
guide K261 + 939
Qinling Zhongnanshan
highway tunnel
Fujian Jiuhuashan Tunnel
Kuocangshan tunnel
k155 + 200k156 + 178
Kuocangshan tunnel
k156 + 203k157 + 573
Kuocangshan tunnel
k157 + 573k58 + 078
Kuocangshan tunnel
k157 + 078kl59 + 250
Chengchao iron mine

98
99a
100
101
102
103
104
105

88
89

90
91
92
93

rc/rt

Wet

Actual
conditions

Granophyric marble

0.463

19.700

2.30

Light

Black marble

0.846

27.300

3.10

Moderate

Black marble

0.506

27.300

3.10

Moderate

Weathered and fresh welded tuff,

tuff breccia and the K-feldspar
granite porphyry

0.210

24.300

4.60

Light

0.280

23.600

4.90

Light

0.320

21.300

5.30

Moderate

0.280

23.800

4.80

Light

0.520

21.200

5.50

Moderate

0.650

28.600

6.80

Strong

H/m

rh

rc

rt

(Mpa)

(Mpa)

(Mpa)

<504

13.90

124.00

4.22

0.520
0.112

24.600
29.400

7.30
2.04

Moderate
None

<504

17.40

161.00

3.98

0.139

31.400

2.19

Light

<504

19.00

153.00

4.48

0.151

28.100

2.11

Light

<504

19.70

142.00

4.55

0.155

27.900

2.26

Light

Marble

469

18.70

82.00

10.90

0.230

7.520

1.50

None

Chengchao iron mine

Chengchao iron mine
Chengchao iron mine
Chengchao iron mine
Chengchao iron mine
Chengchao iron mine
Chengchao iron mine
Dongguashan copper mine

Granite porphyry
Diorite
Dioritic porphyrite
Magnetite
Granite
Skarn
Quartz-feldspar porphyry
Siltstone

520
552
583
567
670
670
600
850

28.60
29.80
33.60
26.90
55.90
59.90
68.00
105.50

122.00
132.00
156.00
92.80
128.00
96.60
107.00
187.00

12.00
11.50
10.80
9.47
6.29
11.70
6.10
19.20

0.230
0.230
0.220
0.290
0.440
0.620
0.640
0.560

10.220
11.520
14.450
9.800
20.300
8.260
17.510
9.740

2.50
4.60
5.20
3.70
8.10
1.80
7.20
7.27

Moderate
Moderate
Moderate
Moderate
Strong
Light
Strong
Moderate

106
107
108

Dongguashan copper mine

Dongguashan copper mine
Tongyu Tunnel K21 + 680

Garnet Skarn
Skarn
Limestone

850
790
900

105.50
105.50
47.56

170.00
190.00
58.50

12.10
17.10
3.50

0.620
0.550
0.810

14.050
11.110
16.710

5.76
3.97
5.00

Moderate
Moderate
Light

109

Tongyu Tunnel K21 + 740

Limestone

1030

43.62

78.10

3.20

0.560

24.410

6.00

Light

110

Daxiangling tunnel YK55 + 119

Rhyolite

362

25.7

59.7

1.3

0.43

45.9

1.7

None

111
112
113
114
115a
116a
117
118
119
120a
121
122
123
124
125
126
127
128
129
130
131
132

Daxiangling tunnel ZK55 + 154

Daxiangling tunnel YK55 + 819
Daxiangling tunnel ZK55 + 854
Daxiangling tunnel YK56 + 080
Daxiangling tunnel YK56 + 109
Daxiangling tunnel YK56 + 177
Daxiangling tunnel YK56 + 343
Daxiangling tunnel ZK56 + 374
Daxiangling tunnel YK56 + 421
Daxiangling tunnel YK61 + 305
Daxiangling tunnel YK61 + 382
Daxiangling tunnel YK61 + 400
Daxiangling tunnel ZK61 + 440
Daxiangling tunnel YK61 + 445
Daxiangling tunnel YK61 + 450
Daxiangling tunnel YK61 + 493
Daxiangling tunnel YK61 + 827
Daxiangling tunnel YK61 + 382
Daxiangling tunnel ZK56 + 451
Daxiangling tunnel ZK56 + 479
Daxiangling tunnel ZK61 + 201
Daxiangling tunnel ZK61 + 352

Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite
Rhyolite

374
775
799
811
816
841
959
984
1112
981
808
799
768
764
760
729
724
808
1048
1074
980
839

26.9
40.4
39.4
38.2
45.7
35.8
39.4
40.6
39.0
57.2
55.6
56.9
62.1
29.7
29.1
27.8
30.3
55.6
41.6
40.1
58.2
56.8

62.8
72.1
65.2
71.4
69.1
67.8
69.2
66.6
70.1
80.6
114
123
132
116
94
90
88
114
67.6
72.1
83.6
112

2.1
2.1
2.3
3.4
3.2
3.8
2.7
2.6
2.4
2.5
2.3
2.7
2.4
2.7
2.6
2.1
3.1
2.3
2.7
2.3
2.6
2.2

0.42
0.56
0.60
0.53
0.66
0.52
0.57
0.61
0.56
0.71
0.49
0.46
0.47
0.26
0.31
0.31
0.34
0.49
0.61
0.55
0.69
0.50

29.9
34.3
28.3
21.0
21.5
17.8
25.6
25.6
29.2
32.2
49.5
45.5
55.0
42.9
36.1
42.8
28.3
49.5
25.0
31.3
32.1
50.9

2.4
1.9
3.4
3.6
4.1
4.3
3.8
3.7
4.8
5.5
4.7
5.2
5.0
3.7
3.2
1.8
3.0
4.7
3.7
4.6
5.9
5.2

Light
Light
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Strong
Moderate
Moderate
Moderate
Light
Light
None
Light
Moderate
Moderate
Moderate
Strong
Moderate

94
95
96
97

Long exploratory tunnel

3 + 650 of Jinping II
hydropower station
Long exploratory tunnel
3 + 000 of Jinping II
hydropower station
Long exploratory tunnel
3 + 800 of Jinping II
hydropower station
Cangling tunnel
K97 + 102K98 + 152

rh/rc

Rock type

Denotes testing sample by selecting stochasticly; The total data is 132.

References

Wang
et al.
(2010)

Qin et al.
(2009)

Xu et al.
(2008)

Liu et al.
(2008)

Wang
et al.
(2004)
Kang
(2006)
Zhang
(2007)

637

J. Zhou et al. / Safety Science 50 (2012) 629644

3. Application of SVMs-based model for long-term rockburst

prediction
3.1. Inuence factor
In the driving of tunnels and the construction of other civil engineering works in hard rock, as recognized that the occurrence of
rockbursts is associated with multiple factors such as the geological structures of the rockmass, the geo-stress conditions, rockmass
strength, the excavation method and excavation size, rock blasting
and earthquake (Mansurov, 2001; Wang and Park, 2001; J. Zhou
et al., 2010). At present, different scholars choose different parameters as evaluation index of criterion for rockburst, the classication of rockburst intensity also differed from each scholar (the
criteria of rockburst is listed in Table 1). So it is difcult to use
these criteria in construction of underground engineering (Chen
et al., 2003). In the current study, on the basis of previous researching criteria of rockburst (Table 1), it chooses buried depth H, maximum tangential stress rh, uniaxial compressive strength rc,
uniaxial tensile strength rt, stress coefcient rh/rc, rock brittleness
coefeient rc/rt and elastic energy index Wet as the input parameters. It is generally accepted that the project depth H determines
the depth of the size of the in situ stress, direction and distribution,
thereby affecting the engineering geological and environmental
conditions, so we can use H instead of the in situ stress. The maximum tangential stress rh reects the strata stress character of
rockburst. The main characteristics of rock which impact rockburst
are uniaxial compressive strength rc and uniaxial tensile strength
rt. Elastic energy index Wet reects rocks ability storing elastic energy (Kidybinski, 1981). Research Russenes, Turchaninov, Wang
et al., Zhang et al. and Hoeks experiential criteria synthetically (Table 1), Stress coefcient rh/rc is selected as an input parameter.
The criteria of rockburst presented by Lu, indicate that when rh/
rc is bigger than Ks rockburst will occur, and the value of Ks depends on the rock brittleness coefeient rc/rt (Wang et al., 2004;
J. Zhou et al., 2010). So the grade of rockburst depends on rc/rt.
When the rock brittleness coefeient rc/rt is small, rockburst occurs acutely. When the rock brittleness coefeient rc/rt is big,
rockburst occurs lightly. So rh/rc is selected as an input parameter.
3.2. Datasets
The investigation was carried out in the literatures in recent
years. Field data obtained from Fengs data (Feng and Wang,
1994), Wangs data (Wang et al., 1998), Riverside Hydropower Station diversion tunnel (Zhang et al., 2010), Jinping II Hydropower
Station (Liang, 2004), Qinling tunnel (Su et al., 2010; Bai et al.,
2002), Jinchuan 2nd Mine (Yi et al., 2010), Ma Luping mine (Yang
et al., 2010), Beiminghe iron mine (Zhang and Li, 2009), Cangling
tunnel (Wang et al., 2010), Kuocangshan tunnel (Qin et al., 2009),
Chengchao iron mine (Xu et al., 2008), Dongguashan copper mine
(Liu et al., 2008), Tongyu Tunnel (Wang et al., 2004; Kang, 2006),
Da xianglin tunnel (Zhang, 2007), etc., as shown in Table 2. Table 3
indicates the relevant input parameters used to develop long-term
rockburst prediction models range with their range, mean, standard deviation and skew, respectively. The boxplot of the original
data set is given in Fig. 3. For the most of the data groups, the median is not in the center of the box, which indicates that the distribution of the most of the data groups is not symmetric (Fig. 3). In
addition, all dependent variables do not have any outliers expect
rh, rc, rt and rc/rt.
3.3. SVMs model development and validation
Therefore, SVMs (Vapnik, 1995; Gunn 1998; Chang and Lin,
2001; Li, 2009) model was established using the theory discussed

Table 3
Descriptive statistics of the input parameters with their range, mean, standard
deviation and skew for SVMs modeling.
Parameter

Range

Mean

Standard deviation

Skew

H (m)
rh (Mpa)
rc (Mpa)
rt (Mpa)
rh/rc
rc/rt
Wet

1001140
2.6167.2
18.32263
0.3819.2
0.11.41
0.14755
0.859.3

681.63
54.11
109.92
6.00
0.47
22.23
4.28

273.103
33.024
50.415
3.676
0.219
10.297
2.081

0.357
1.130
0.393
0.956
0.737
0.690
0.317

Fig. 3. Boxplot of the original data set of rockburst.

Table 4
Standard of c1assication for intensities of rockburst.
Rockburst
classication

Failure characteristics

Strong
Rockburst

The surrounding rock is bursted severely, and suddenly

thrown out or ejected into tunnel, accompanied by a
strong burst and the roaring sound, air spray, storms
phenomena with continuity, and rapidly expand to the
deep surrounding rock.
The surrounding rock is deformed and fractured, there is a
considerable number of rock chip ejection, loose and
sudden destruction, accompanied by crisp crackling, and
often presented in the local cavern of surrounding rock.
The surrounding rock is deformed, cracked or rib spalled,
there is a weak sound, and no ejection phenomenon.
No sound of rock burst and rockburst activities.

Moderate
Rockburst

Light Rockburst
None Rockburst

above. For implementing the SVMs, it used the toolbox LIBSVM

for Matlab by Chang and Lin (2001) that can be downloaded from:
http://www.csie.ntu.edu.tw/cjlin/libsvm/. The empirical evaluation was performed on a personal computer with Intel (R) Core
(TM) 2 Duo CPU running at 2.10 GHz and 2 GB RAM. The evaluation
indexes of SVMs model for predicting of classication of long-term
rockburst for underground openings was established based on the
statistical learning theory and the actual characteristics of the project. The evaluation indexes of the proposed model are as follows:
In the present study, H, rh, rc, rt, rh/rc, rc/rt and Wet were dened
as the criterion indices for long-term rockburst prediction of
underground openings in the proposed model (Table 3). In order
to examine the reliability of the model in the SVMs and investigate
different input parameters impact on the predicted results. we
choose H, rh, rc, rt, rh/rc, rc/rt and Wet as the input parameters

638

J. Zhou et al. / Safety Science 50 (2012) 629644

100

Accuracy(%)

90
80
70
60
50
40
30
8
6
4
2

log2(g) -2

0
-4
-6

log2(c)

-5
-8

100

100

90

90

80

80

Accuracy(%)

Accuracy(%)

(1) model 1

70
60
50
40

70
60
50
40

30
8
6

30
8
4
2

log2(g)

-2

-4
-6

-8

-8

-2

-4

-6

log2(g) -2 -4

log2(c)

-6

100

100

90

90

80

80

70
60
50
40
30
8

-8

-8

-2

log2(c)

(3) model 3

Accuracy(%)

Accuracy(%)

(2) model 2

-4

-6

70
60
50
40

30
8
4

log2(g)

-2

-4

-6

-8 -8

-6

-4

-2

log2(c)

(4) model 4

log2(g)

-2

-4

-6

-8 -8

-6

-4

-2

log2(c)

(5) model 5

Fig. 4. The tness curve of selecting best parameters by GSM for different SVMs models.

in model 1, select rh, rc, rt, rh/rc, rc/rt and Wet as the input parameters in model 2, opt H, rh/rc, rc/rt and Wet as the input parameters
in model 3, pick rh, rc, rt and Wet as the input parameters in model

4, make a choice rh/rc, rc/rt and Wet as the input parameters in

model 5. Moreover, according to the extent and intensity of the
characteristics of rokburst phenomenon in the underground

639

J. Zhou et al. / Safety Science 50 (2012) 629644

60

60
58

50

56

40

Fitness

Fitness

54
52
50

Best fitness
Average fitness

20

Best fitness
Average fitness

48

30

10

46
44

0
0

20

40

60

80

100

20

40

60

Evoluationary generations

Evoluationary generations

(1) model 1

(2) model 2

80

100

80

100

56
54

Fitness

52
50
Best fitness
Average fitness

48
46
44
0

20

40

60

80

100

Evoluationary generations

(3) model 3
59

58
56

58
57

Fitness

Fitness

54
52
50
Best fitness
Average fitness

48

55

Best fitness
Average fitness

54

46
44

56

20

40

60

80

100

53

20

40

60

Evoluationary generations

Evoluationary generations

(4) model 4

(5) model 5

Fig. 5. The tness curve of selecting best parameters by GA for different SVMs models.

openings, if the grade of rockburst is divided into four degrees,

namely none rockburst, light rockburst, moderate rockburst,
strong rockburst, respectively. So the SVMs model output is rockburst degree (RD), RD = [none rockburst, light rockburst, moderate
rockburst, strong rockburst]. In addition, the division of rockburst
degree can be described in Table 4 (Wang, et al., 1998; Zhang
et al., 2004). The inputoutput data were scaled to lie between

1 and 1 by using the formula (9) (Li, 2009). In this equation, xnorm
is the normalized value, x is the actual value, xmax is the maximum
value, and xmin is the minimum value.

X norm 2x  xmin =xmax  xmin 1

The election of the parameters plays an important role in the performance of SVMs. Because SVMs generalization performance

640

J. Zhou et al. / Safety Science 50 (2012) 629644

Table 5
Variables and summary of the generated models for SVMs.
Model

Input variables

Method

Best C

Best g

CVA (%)

Type

True (%)

False (%)

H, rh, rc, rt, rh/rc, rc/rt, Wet

GSM-SVMs

45.25

0.250

59.49%

GA-SVMs

4.95

1.958

58.23%

PSO-SVMs

41.04

3.658

54.43%

GSM-SVMs

22.63

0.125

60.61%

GA-SVMs

16.27

3.678

57.58%

PSO-SVMs

29.81

3.693

56.57%

GSM-SVMs

64.00

0.177

63.29%

GA-SVMs

1.58

21.404

55.70%

PSO-SVMs

34.25

21.684

54.43%

GSM-SVMs

22.63

0.500

57.58%

GA-SVMs

7.97

1.151

56.57%

PSO-SVMs

12.23

10.602

51.52%

GSM-SVMs

22.63

0.707

65.57%

GA-SVMs

63.20

10.112

58.20%

PSO-SVMs

99.28

9.312

57.38%

Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set
Train set
Test set

88.61% (70/79)
71.43% (5/7)
94.94% (75/79)
85.71% (6/7)
100.00% (79/79)
85.71% (6/7)
67.68% (67/99)
88.89% (8/9)
97.98% (97/99)
77.78% (7/9)
97.98% (97/99)
77.78% (7/9)
79.75% (63/79)
71.43% (5/7)
100.00% (79/79)
71.43% (5/7)
100.00% (79/79)
71.43% (5/7)
75.76% (75/99)
66.67% (6/9)
77.78% (75/99)
66.67% (6/9)
97.98% (97/99)
66.67% (6/9)
76.23% (75/99)
80.00% (8/10)
97.54% (119/122)
80.00% (8/10)
97.54% (119/122)
90.00% (9/10)

11.39% (9/79)
28.57% (2/7)
5.06% (4/79)
14.29% (1/7)
0.00% (0/79)
14.29% (1/7)
32.32% (32/99)
11.11% (1/9)
2.02% (2/99)
22.22% (2/9)
2.02% (2/99)
22.22% (2/9)
20.25% (16/79)
28.57% (2/7)
0.00% (0/79)
18.57% (5/7)
0.00% (0/79)
18.57% (2/7)
24.24% (4/99)
33.33% (3/9)
22.22% (24/99)
33.33% (3/9)
2.02% (2/99)
33.33% (3/9)
23.73% (24/99)
20.00% (8/10)
2.46% (3/122)
20.00% (8/10)
2.46% (3/122)
10.00% (1/10)

rh, rc, rt, rh/rc, rc/rt, Wet

H, rh/rc, rc/rt, Wet

rh, rc, rt, Wet

rh/rc, rc/rt, Wet

strongly depends on the right setting of hyper-parameters C and the

kernel parameter g, these parameters need to be selected properly.
In this paper three methods are presented for estimating the optimal SVMs parameter settings, and they are GSM (2.2 chapter), GA
(2.3 chapter) and PSO (2.4 chapter). We used a 2-D test function
to evaluate GSM will be used as a representative conventional technique for estimating the optimal SVMs parameter settings, and our
GA/PSO-SVMs method will be used as heuristic algorithms technique for estimating the optimal SVMs parameter settings, and
the experiment results show below.
After the standardization of the data, a training to support vector machine was started using the program gotten through GSM,
GA and PSO. When applying SVMs, the goodness of t is determined by the penalty factor C and insensitive parameter g. Libsvm
(Chang and Lin, 2001; Li, 2009) provides a parameter selection tool
using the RBF kernel: cross validation via parallel GSM. In this
study, the free parameters of SVM were selected followed a 5-fold
cross-validation experiment to control generalization capability of
SVMs, and the RBF kernel is used as the kernel function of the
SVMs because it tends to give better performance. The SVMs uses
the RBF kernel whose parameter, g, must be tuned by means of a
cross-validation process, as well as a parameter C, which establishes a compromise between error minimization and generalization capability in the SVMs, and RBF kernel function is adopted
as the kernel function of the samples training, obtaining best
parameters by GSM. Fig. 4 shows an example of the GSM result,
where the x-axis and the y-axis are log2C and log2g, respectively.
The z-axis is the 5-fold average performance for cross validation
accuracy (CVA). In the initial population, the range for each design
variable is set as follows: The ndings of this experiment were that
SVMs is quite robust against parameter selections. In the present
study, training and testing of SVMs has been carried out by using
Matlab. Take model 1 as an example, the result of the GSM-SVMs
parameter selection by GSM is shown in Fig. 5 (3D view), when C
is 45.25, g is 0.250, the average value of MSE is CVAmse, which is
equal to 59.49%. while 79 sets of training sample data were back

evaluated one by one using the GSM-SVMs model of rockburst

and compared with the actual situation, the correct rate is
88.61%. Therefore, the GSM-SVMs model is feasible and effective
for rockburst forecasting. In order to examine the reliability of
the model in a SVMs without model reliability examination is also
built in the 7 forward samples (Table 2). The prediction accuracy of
GA-SVMs model 1 is found to be 71.43% for testing data. Table 5
shows the evaluation results of GSM-SVMs.
In order to determine reasonable and efcient the parameters
of SVMs, rstly, the appropriate tness function for genetic algorithms (GA) operation is determined, and then optimization
parameters of SVMs model is selected by real coded GA, therefore, the genetic algorithms and support vector machine (GASVMs) model for long-term rockburst prediction is established.
Take model 1 as an example, the parameters research space of
GA are shown below. The searching process of optimal parameters is operated with 100 generations in total, and the initial population is 20. The maximum number of generations is adopted as
the termination criterion for the solution process. C and g are real
coded values. The searching ranges for penalty coefcient C and
the kernel function parameters g are as follows: C 2 [0, 102] and
g 2 [0, 102]. Mutation is performed to alter binary code from 0
to 1 or vice versa, so the rate of mutation is set to 0.05 and
the crossover probability is 0.9. Namely ga_option = struct (maxgen, 100, sizepop, 20, ggap, 0.9, cbound, [0, 100], gbound,
[0, 100], v, 5). The classication success of the test data was used
as the tness criterion in the evaluation function. Fig. 5 illustrates
the correlation curves by GA for the best tness, average tness
versus the generation number for different models. Given the
examination of the reliability of the model and the vast amount
of training and checking to support the vector machine. Finally, it
can obtain the prediction model parameters: penalty coefcient
C = 4.95, kernel function width factor g = 1.958 and the average
value of MSE is 56.57%. The prediction accuracy of GA-SVMs
model 1 is found to be 94.94% and 85.71% for trainging data
and testing data, respectively.

641

J. Zhou et al. / Safety Science 50 (2012) 629644

58

55

56
54

50

52

Fitness

Fitness

Best fitness
Average fitness

50

Best fitness
Average fitness

48
46

45

44
42

40

20

40

60

80

40

100

20

Evoluationary generations
(1)model 1

40

60

80

100

Evoluationary generations
(2)model 2

55

Fitness

50
Best fitness
Average fitness
45

40

35

20

40

60

80

100

Evoluationary generations

(3) model 3
52

58
57

50

Best fitness
Average fitness

56

Best fitness
Average fitness

55

Fitness

Fitness

48
46

54
53
52

44

51

42
50

40

49

20

40

60

80

100

Evoluationary generations
(4)model 4

20

40

60

80

100

Evoluationary generations
(5)model 5

Fig. 6. The tness curve of selecting best parameters by PSO for different SVMs models.

Similarly, then optimization parameters of SVMs model are

selected by Particle swarm optimization (PSO) algorithm, therefore, the PSO-SVMs model for long-term rockburst prediction is
established. The operation procedure was written in Matlab
environment. Take model 4 as an example, a particle contains
optimal C and g. The predened maximum iteration is 100 in order to avoid overtraining during each of the ve runs (5-fold
cross validation). The swarm size is set to 20 particles. The
searching ranges for C and g are as follows: C 2 [101, 102] and

g 2 [102, 103]. The personal factors c1 = 1.5; and social factor

c2 = 1.7, namely pso_option = struct(c1, 1.5, c2, 1.7, maxgen,
100, sizepop, 20, k, 0.6, wV, 1, wP, 1, v, 5, popcmax,
10^2, popcmin, 10^(-1), popgmax, 10^3, popgmin,10^(-2)).
Fig. 6 illustrates the correlation curves by PSO for the best tness, average tness versus the generation number for different
models. Run the computer program to iteration, the search results for the best model parameter are as follows: penalty coefcient C = 41.40, kernel function width factor g = 3.658 and the

642

J. Zhou et al. / Safety Science 50 (2012) 629644

average value of MSE is 54.43%. The prediction accuracy of PSOSVMs model 1 is found to be 100.00% and 85.71% for trainging
data and testing data, respectively. These models optimal parameters are summarized in Table 5.

3.4. Results and discussion

This section is devoted to the presentation and discussion of the
results obtained by the proposed models described in the previous
section. The rockburst database (Table 2) are back evaluated one by
one using the proposed models of long-term rockburst for underground openings and compared with the actual situation, preliminary experiments show a similar behavior of among GSMSVMs, GA-SVMs and PSO-SVMs at a frame classication level and
achieve a high prediction classication accuracies. Therefore, the
SVMs model is feasible and effective for long-term rockburst forecasting for underground openings and can be put into use practice
engineering. In order to compare the difference between our methods, the evaluation results of GSM-SVMs model, GA-SVMs and
PSO-SVMs are also listed in Table 5. We can see from the three
SVMs evaluation results and actual assessment results that the
classication results of both SVM are accordant with the actual situation basically. Table 5 shows the prediction results with different
models and methods, we can see the forecast accuracy is increased
dramatically from GSM-SVMs to GA-SVMs, PSO-SVMs. It shows
that the GA-SVMs has higher prediction accuracy than the GSMSVMs, especially in the model 5, despite changes tremendously,
still has good prediction performance.
The above-mentioned comparisons indicate that all three models (GSM-SVMs, GA-SVMs and PSO-SVMs) are competitive with
each other for long-term rockburst prediction in underground
openings, but the performance of the PSO-SVMs is relatively superior to the others. Moreover, the SVMs has some added advantages.
In particular, the use of the structural risk minimization (SRM)
principle in dening cost function provided good generalization
capacity with the SVMs, which uses the empirical risk minimization (ERM) principle. The number of free parameters is another
advantage of using the SVMs. In the case of the RBF kernel function,
there are only two parameters in the SVM, namely C and g, and the
SVMs is its optimization algorithm, which includes solving a linearly constrained QP function leading to a unique, optimal, and global solution. Moreover, there is a high potential for further
improvement of the SVMs performance. In practical application,
according to the circumstances, project example data can be extensively collected, establishing the corresponding sample database
that the recognition of model can be enhanced, which would be
improved that the prediction results will be more in line with
the facts.
From the Table 5, it is obvious that the accuracy of the PSOSVMs is more than 90% and higher than the GA-SVMs and GSMSVMs; whereas the accuracy of the GA-SVMs is comparable to
the GSM-SVMs. Summarizing, GA-SVMs and PSO-SVMs were
developed as relatively fast alternatives for the time consuming
GSM. Both methods, especially the latter, can efciently solve the
problem of estimating the optimal SVMs parameter settings at a
reasonable computational cost. The various inuencing factors of
rockburst can be as the SVMs input vector, basically unlimited
number of factors. Thus, SVMs can be considered a more comprehensive rockburst proneness and the relationship between various
factors of rockburst. In practical engineering applications, according to the case, widely collect case information, establish the
appropriate sample database, to enhance the models recognition
ability, long-term intelligent SVM model for prediction of rockburst proneness and further improving the prediction is more in
line with actual results.

4. Conclusions
In this work, a prediction model of long-term rockburst is established by SVMs. According to mechanism of rockburst, buried
depth H, rocks maximum tangential stress rh, rocks uniaxial compressive strength rc, rocks uniaxial tensile strength rt, stress coefcient rh/rc, rock brittleness coefeient rc/rt and elastic energy
index Wet are dened as the criterion indices for rockburst prediction in the proposed model. Two hybrid techniques for rockburst
classication of high dimensional data were presented and compared. These techniques are based on different metaheuristic algorithms such as PSO and GA used for parameter selection using the
SVMs classier to identify potentially good gene subsets, and GSM
by an accurate 5-fold cross validation method is also used for SVMs
classier to improve the actual classication. In order to examine
the reliability of the model in the SVMs and investigate different
input parameters impact on the predicted results, which avoids
over-tting or undertting of the SVMs model occurring because
of the improper determination of these parameters. The eld datasets are used to investigate its feasibility in the rockburst prediction for underground openings. The experimental results show
that the classication accuracies of PSO-SVMs is more than 90%
and superior than the GA-SVMs and GSM-SVMs; whereas the accuracy of the GA-SVMs is comparable to the GSM-SVMs, and by comparing the evaluation results among three models, it is indicated
that PSO-SVMs has more excellent prediction performance than
GA-SVMs and GSM-SVMs for long-term rockburst prediction in
underground openings.
Acknowledgments
This research is supported by the National Basic Research Program Project (2010CB732004) of China, the National Natural Science Foundation Project (50934006) of China and the Graduated
Students Research and Innovation Fund Project (CX2011B119) of
Hunan Province of China. The authors would like to express thanks
to these foundations.
References
Alcott, J.M., Kaiser, P.K., Simser, B.P., 1998. Use of microseismic source parameters
for rockburst hazard assessment. Pure and Applied Geophysics 153, 4165.
Bai, M.Z., Wang, L.J., Xu, Z.Y., 2002. Study on a neutral network model and its
application in predicting the risk of rock burst. China Safety Science Journal 12
(4), 6569 (in Chinese).
Board, M., Fairhurst, C., 1983. Rockbursts: Prediction and Control. Institution of
Mining and Metallurgy, London.
Brady, B.T., Leighton, F.W., 1977. Seismicity anomaly prior to a moderate rock burst
case study. International Journal of Rock Mechanics and Mining Sciences and
Geomechanics Abstracts 14 (3), 127132.
Casten, U., Fajklewicz, Z., 1993. Induced gravity-anomalies and rockburst risk in
coalmines a case history. Geophysical Prospecting 41 (1), 113.
Chang, C.C., Lin, C.J., 2001. LIBSVM: a Library for Support Vector Machines. <http://
www.csie.ntu.edu.tw/~cjlin/libsvm/>.
Chen, H.J., Li, N.H., Ni, D.X., Shang, Y.Q., 2003. Prediction of rockburst by articial
neural network. Chinese Journal of Rock Mechanics and Engineering 22 (5),
762768.
Cook, N.G.W., Hoek, E., Pretorius, J.P.G., Ortlepp, W.D., SAlamon, M.D.G., 1966. Rock
mechanics applied to the study of rockbursts. Journal of the South African
Institute of Mining and Metallurgy 66 (3), 435528.
Chen, X., Sun, J.Z., Zhang, J.K., Chen, Q.S., 2009. Judgment indexes and classication
criteria of rock burst with the extension judgment method. China Civil
Engineering Journal 42 (9), 8288 (in Chinese).
Crrp, 1996. A Comprehensive Summary of Five Years of Collaborative Research on
Rockbursting in Hard Rock Mines. CAMIRO Mining Division, Canadian
Rockburst Research Program, Book 1.
Deng, L., Lv, Y., Deng, R.G., 2011. Present situation and consideration of rock burst
hazard. Advanced Materials Research, 136147.
Dowding, C.H., Andersson, C.A., 1986. Potential for rock bursting and slabbing in
deep caverns. Engineering Geology 22, 265279.
Feng, X.T., Wang, L.N., 1994. Rockburst prediction based on neural networks.
Transactions of Nonferrous Metals Society of China 4 (1), 714.
Frid, V., 1997. Rockburst hazard forecast by electromagnetic radiation excited by
rock fracture. Rock Mechanics and Rock Engineering 30 (4), 229236.

J. Zhou et al. / Safety Science 50 (2012) 629644

Goldberg, G.E., 1989. Genetic Algorithms in Search, Optimization and Machine
Learning. Addison-Wesley, NY, USA.
Gong, F.Q., Li, X.B., 2007. A distance discriminant analysis method for prediction of
possibility and classication of rockburst and its application. Chinese Journal of
Rock Mechanics and Engineering 26 (5), 10121018 (in Chinese).
Gunyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene selection for cancer
classication using support vector machines. Machine Learning 46 (1-3), 389422.
Gunn, S.R., 1998. Support Vector Machines for Classication and Regression.
Technical Report, University of Southampton, Department of Electrical and
Computer Science, UK. <http://www.ecs.soton.ac.uk/srg/publications/pdf/
SVM.pdf>.
Guyon, I., Christianini, N., 1999. Survey of support vector machine applications. In:
Proceedings of NIPS99 Special Workshop on Learning with Support Vector.
<http://www.clopinet.com/isabelle/Projects/SVM/applist.html>.
He, M.C., Miao, J.L., Feng, J.L., 2010. Rock burst process of limestone and its acoustic
emission characteristics under true-triaxial unloading conditions. International
Journal of Rock Mechanics & Mining Sciences 47 (2), 286298.
Hoek, E., Brown, E.T., 1997. Practical estimates of rock mass strength. International
Journal of Rock Mechanics and Mining Sciences 34, 11651186.
Hou, F.L., 1989. Criterion and prevention measures on rockburst in circular tunnel.
In: Chinese Society for Rock Mechanics and Engineering. Proceedings of the 2th
National Rock Mechanics and Engineering. Knowledge Press, Beijing, pp. 195
201 (in Chinese).
Hucka, V., Das, B., 1974. Brittleness determination of rocks by different methods.
International Journal of Rock Mechanics and Mining Sciences and
Geomechanics Abstracts 11 (10), 389392.
Jha, P.C., Chouhan, R.K.S., 1994. Long range rockburst prediction: a seismological
approach. International Journal of Rock Mechanics and Mining Sciences &
Geomechanics Abstracts 31 (1), 7177.
Jiang, A.N., Wang, S.Y., Tang, S.L., 2011. Feedback analysis of tunnel construction
using a hybrid arithmetic based on support vector machine and particle swarm
optimisation. Automation in Construction 20 (4), 482489.
Jiang, Q., Feng, X.T., Xiang, T.B., Su, G.S., 2010. Rockburst characteristics and
numerical simulation based on a new energy index: a case study of a tunnel at
2500 m depth. Bulletin of Engineering Geology and the Environment 69, 381
388.
Kang, Y., 2006. Research on Relevant Probloms about Failure Mechnism of
Surrounding Rock in Deep Buried Tunnel. Ph.D. Dissertation. Chongqing
University, Chongqing, pp. 118120 (in Chinese).
Kaiser, P.K., McCreat, D.R., Tannnant, D.D., 1996. In: Canadian Rockburst Support
Handbook. Geomechanics Research Centre.
Kennedy, J., Eberhart, R.C., 1995. Particle swarm optimization. In: Proceedings of the
IEEE International Conference on Neural Network. vol. 4, IEEE Press, New York,
pp, 19421948.
Khandelwal, M., 2010. Evaluation and prediction of blast-induced ground vibration
using support vector machine. International Journal of Rock Mechanics and
Mining Sciences 47 (3), 509516.
Kidybinski, A., 1981. Bursting liability indices of coal. International Journal of Rock
Mechanics and Mining Sciences and Geomechanics Abstracts 18 (4), 295304.
Kovacevic, M., Bajat, B., Gajic, B., 2010. Soil type classication and estimation of soil
properties using support vector machines. Geoderma 154 (3-4), 340347.
Li, Y., 2009. LIBSVM-farutoUltimate Version: a Toolbox with Implements for
Support Vector Machines based on Libsvm. <http://www.ilovematlab.cn>.
Liang, Z.Y., 2004. Study on the Prediction and Prevention of Rockburst in the
Diversion Tunnel of Jinping II Hydropower Station, M.Sc. Thesis. Chengdu
University of Technology, Chendu, pp. 6162 (in Chinese).
Linkov, A.M., 1996. Rockbursts and the instability of rock masses. International
Journal of Rock Mechanics and Mining Sciences and Geomechanics Abstracts 33
(7), 727732.
Liu, Z.J., Yuan, Q.P., Li, J.L., 2008. Application of fuzzy probability model to prediction
of classication of rockburst intensity. Chinese Journal of Rock Mechanics and
Engineering 27 (Supp. 1), 30953103 (in Chinese).
Mansurov, V.A., 2001. Prediction of rockbursts by analysis of induced seismicity
data. International Journal of Rock Mechanics & Mining Sciences 38 (7), 893
901.
Matlab 7.0. Program, Available from The Mathworks Inc., Natick, MA. <http://
www.mathworks.com>.
Mib, S.H., Lee, J.M., Han, I.G., 2006. Hybrid genetic algorithms and support vector
machines for bankruptcy prediction. Expert Systems with Applications 31 (3),
652660.
Neyman, B., Szecowka, Z., Zuberek, W., 1972. Effective methods for ghting rock
burst in Polish collieries. In: Proceedings of the 5th International Strata Control
Conference, vol. 23, pp. 19.
Ortlepp, W.D., 1997. Rock fracture and rockburstsan illustrative study. SAIMM
Monograph Series. Johannesburg, Berlin, M9.
Ortlepp, W.D., Stacey, T.R., 1994. Rockburst mechanisms in tunnels and shafts.
Tunnelling and Underground Space Technology 9 (1), 5965.
Pan, Y.S., Li, Z.H. 2002. The analytic analysis stability of rock structure in mine. In:
Proceedings of National Conference on Solid, Mechanics, Dalian University of
Technology Press, Dalian (in Chinese).
Pan, Y., Zhang, Y., Yu, G.M., 2006. Mechanism and catastrophe theory analysis of
circular tunnel rockburst. Applied Mathematics and Mechanics 27 (6), 841852
(in Chinese).
Peng, Q., Qian, A.G., Xiao, Y., 2010. Research on prediction system for rockburst
based on articial intelligence application methods. Journal of Sichuan
University: Engineering Science Edition 42 (4), 1824 (in Chinese).

643

Qin, S.W., Chen, J.P., Wang, Q., 2009. Research on rockburst prediction with extenics
evaluation based on rough set. In: Tang, C.A. (Ed.), Proceedings of the 13th
International Symposium on Rockburst and Seismicity in Mines. Rinton Press,
Dalian, pp. 937944.
Ren, Y., Bai, G.C., 2010. Determination of optimal SVM parameters by using GA/PSO.
Journal of Computers 5 (8), 11601168.
Russenes, B.F., 1974, Analysis of Rock Spalling for Tunnels in Steep Valley Sides (in
Norwegian), M.Sc. thesis. Norwegian Institute of Technology, Department of
Geology, p. 247.
Ryder, J.A., 1987. Excess shear stress in the assessment of geologically hazardous
situations. Journal of the South African Institute of Mining and Metallurgy 88
(1), 2739.
Sharan, S.K., 2004. A nite element perturbation method for the prediction of
rockburst. Computers and Structures 85, 13041309.
Shi, X.Z., Zhou, J., Dong, L., Hu, H.Y., Wang, H.Y., Chen, S.R., 2010. Application of
unascertained measurement model to prediction of classication of rockburst
intensity. Chinese Journal of Rock Mechanics and Engineering 29 (supp.1),
27202727 (in Chinese).
Shivakumar, K., Rao, M.V.M.S., Srinivasan, C., Kusunose, K., 1996. Multifractal
analysis of the spatial distribution of area rockbursts at Kolar Gold Mines.
International Journal of Rock Mechanics & Mining Sciences 33 (2), 167172.
Singh, S.P., 1988. Burst energy release index. Rock Mechanics and Rock Engineering
21 (1), 149155.
Su, G.S., Zhang, Y., Chen, G.Q., 2010. Identify rockburst grades for jinping
hydropower station using gaussian II process for binary classication. In:
Proceedings of 2010 International Conference on Computer, Mechatronics,
Control and Electronic Engineering (CMCE 2010), vol. 2, Changchun, pp. 364
367.
Sun, J., Wang, S.J., 2000. Rock mechanics and rock engineering in China:
developments and current state-of-the-art. International Journal of Rock
Mechanics and Mining Sciences 37 (3), 447465.
Tao, Z.Y., 1988. Support design of tunnels subjected to rockbursting. In: Romana
(Ed.), ISRM International Symposium, Rock Mechanics and Power Plants, pp.
407411.
Turchaninov, I.A., Markov, G.A., Gzovsky, M.V., Kazikayev, D.M., Frenze, U.K.,
Batugin, S.A., Chabdarova, U.I., 1972. State of stress in the upper part of the
Earths crust based on direct measurements in mines and on tectonophysical
and seismological studies. Physics of the Earth and Planetary Interiors (6), 229
234.
Tang, B.Y., 2000. Rockburst Control using Distress Blasting, Ph.D. Dissertation.
McGill University, Montreal, Canada.
Tang, C.A., Wang, J.M., Zhang, J.J., 2010. Preliminary engineering application of
microseismic monitoring technique. Journal of Rock Mechanics and
Geotechnical Engineering 2 (3), 193208.
Tang, L.Z., Wang, W.X., 2002. New rock burst proneness index. Chinese Journal of
Rock Mechanics and Engineering 21 (6), 874878 (in Chinese).
Tang, L.Z., Xia, K.W., 2010. Seismological method for prediction of areal rockbursts
in deep mine with seismic source mechanism and unstable failure theory.
Journal of central south university of technology: English edition 17, 947953.
Vapnik, V., 1995. The Nature of Statistical Learning Theory. Springer-Verlag,
New York.
Wang, Y.C., Shang, Y.Q., Sun, H.Y., Yan, X.S., 2010. Study of prediction of rockburst
intensity based on efcacy coefcient method. Rock and Soil Mechanics 31 (2),
529534 (in Chinese).
Wang, Y.H., Li, W.D., Li, Q.G., Xu, Y., Tan, G.H., 1998. Method of fuzzy comprehensive
evaluations for rockburst prediction. Chinese Journal of Rock Mechanics and
Engineering 17 (5), 493501 (in Chinese).
Wang, J.A., Park, H.D., 2001. Comprehensive prediction of rockburst based on
analysis of strain energy in rocks. Tunnelling and Underground Space
Technology 16 (1), 4957.
Wang, X.F., Li X.H., Gu, Y.L., Jin, X.G., Kang, Y., Li, D.X., 2004. Application of BP neural
network into prediction of rockburst in tunneling. In: Proceedings of the 2004
International Symposiumon Safety Science and Technology, vol. 4, China
Science Press, Shanghai, pp. 617621.
Wu, Y.K., Zhang, W.B., 1997. Prevention of rockbursts in coal mines in China. In:
Gibowicz S.J., Lasocki, J. (Eds.), Proceedings of the 4th International Symposium
on Rockbursts and Seismicity in Mines, Rotterdam, pp. 361366.
Xie, H.P., Pariseau, W.G., 1993. Fractal character and mechanism of rock bursts.
Chinese Journal of Rock Mechanics and Engineering 12 (1), 2837 (in Chinese).
Xu, M.G., Du, Z.J., Yao, G.H., Liu, Z.P., 2008. Rockburst prediction of chengchao iron
mine during deep mining. Chinese Journal of Rock Mechanics and Engineering
27 (supp.1), 29212928 (in Chinese).
Yang, J.L., Li, X.B., Zhou, Z.L., Lin, Y., 2010. A Fuzzy assessment method of rock-burst
prediction based on rough set theory. Metal Mine (6), 2629 (in Chinese).
Yi, Y.L., Cao, P., Pu, C.Z., 2010. Multi-factorial comprehensive estimation for
jinchuans deep typical rockburst tendency. Science & Technology Review 28
(2), 7680 (in Chinese).
Zhang, H., Wang, Y.J., Li, Y.F., 2009. SVM model for estimating the parameters of the
probability integral method of predicting mining subsidence. Mining Science
and Technology 19 (2), 385388.
Zhang, L.X, Li, C.H., 2009. Study on tendency analysis of rockburst and
comprehensive prediction of different types of surrounding rock. In: Tang,
C.A. (Ed.), Proceedings of the 13th International Symposium on Rockburst and
Seismicity in Mines. Rinton Press, Dalian, pp. 14511456.
Zhang, J.J., Fu, B.J., 2008. Rockburst and its criteria and control. Chinese Journal of
Rock Mechanics and Engineering 27 (10), 20342042 (in Chinese).

644

J. Zhou et al. / Safety Science 50 (2012) 629644

Zhang, J.F., 2007. Study on Prediction by Stages and Control Technology of

Rockburst Hazard of Daxiangling Highway Tunnel. M.Sc. Thesis, Southwest
Jiaotong University, Chendu (in Chinese).
Zhang, L.W., Zhang, D.Y., Qiu, D.H., 2010. Application of extension evaluation
method in rockburst prediction based on rough set theory. Journal of China Coal
Society 35 (9), 14611465 (in Chinese).
Zhang, Z.Y., Song, J.B., Li, P.F., 2004. Rock burst comprehensive forecasting method
for the chamber group of underground power house. Advance in Earth Sciences
19 (3), 451456 (in Chinese).
Zhao, H.B., 2005. Classication of rockburst using support vector machine. Rock and
Soil Mechanics 26 (4), 642644 (in Chinese).
Zhao, H.B., Yin, S.D., 2009. Geomechanical parameters identication by particle
swarm optimization and support vector machine. Applied Mathematical
Modelling 33 (10), 39974012.

Zhou, J., Shi, X.Z., Dong, L., Hu, H.Y., Wang, H.Y., 2010. Fisher discriminant analysis
model and its application for prediction of classication of rockburst in deepburied long tunnel. Journal of Coal Science and Engineering (China) 16 (2), 144
149.
Zhou, X., Li, Z.C., Dai, Z., Zou, X.Y., 2010. QSAR modeling of peptide biological activity
by coupling support vector machine with particle swarm optimization
algorithm and genetic algorithm. Journal of Molecular Graphics and
Modelling 29 (2), 188196.
Zhu, W.C., Li, Z.H., Zhu, L., Tang, C.A., 2010. Numerical simulation on rockburst of
underground opening triggered by dynamic disturbance. Tunnelling and
Underground Space Technology 25, 587599.
Zubelewicz, A., Mroz, Z., 1983. Numerical-simulation of rock burst processes
treaded as problems of dynamic instability. Rock Mechanics and Rock
Engineering 16 (4), 253274.