Rockburst Prediction Using Gaussian Process

Machine Learning

Guo-Shao SU, Ke-Shi ZHANG, Zhi CHEN

School of Civil and Architecture Engineering
Guangxi University
Nanning, China
[email protected]

Abstract— Rockburst is a geological disaster occurred usually in examples, function of degree of membership in fuzzy theory
deep mines. Because of poor understanding of the mechanism cannot escape from effect by man-made choice. ANN is
and influence factors of rockbust, it is very difficult to give limited in solving for small sample problem. SVM cannot
accurate prediction using conventional methods. A new model escape from the blindness which is the common phenomenon
based on Gaussian process (GP), which is a probabilistic kernel in man-made choice of kernel function and its parameter [6].
machine leaning and has become a power tool for solving highly So, it is necessary to find a new effective method for better
nonlinear problems, therefore, is proposed. At first, case histories prediction of rockburst during underground mining.
of rockburst occurrence with the real records of rockburst
intensity and influence factors of rockbust are collected and are Gaussian process (GP) is a new machine learning
taken as prior knowledge to be learned by GP binary technology [7]. In recent years GP model has attracted much
classification machine learning tech, where, maximum tangential attention in the machine learning community, there are a lot of
stress in surround rockmass, uniaxial compressive strength, successful application in the field of solving for nonlinear,
tensile strength of rock, and rockburst tendency index of rock, small samples and high dimensions problems [8]. In this paper,
which can reflect the internal and exterior conditions of we propose a new model of rockburst prediction using GP.
rockburst occurrence nicely are suggested to be main influential
factors of rockburst. Then, the nonlinear mapping relationship
between rockburst intensity and its influence factors can be
established easily by GP model. Finally, prediction for the novel
conditions in deep mines can be obtained using the model. The
new model is applied in prediction for rockburt intensity at
practical projects in China, Norway and USSR. Results of case
study show the model is feasible, effective and simple to
implement for rockburst prediction.
(a) Violent rockburst pit (b) Medium rockburst pit
Keywords-rockburst; prediction; Gaussian process; machine
learning; Figure 1. Photos of rockburst in Jinping tunnel in China.

I. INTRODUCTION II. THEORY OF GP

Nowadays, more and more concerns are given to the A briefly introduction to theory of GP is given in this
rockburst along with the development of underground space section.
excavation (Fig.1). Because occurring suddenly and intensely,
rockburst usually causes injury including death to workers, GP is regarded as one of machine learning tech based on
damage to equipment, and even substantial disruption and gaussian stochastic process and Bayes’ learning theory. In this
economic loss of underground space excavation. Since the section we provide a brief introduction of Gaussian process
pioneering work by Cook [1], many scholars have suggested model for binary classification (GPC). More details about GP
various theories and many prediction methods, and empirical see the works by Kuss and Rasmussen [9].
equations have been established after studying lots of instances
Let y ∈ {−1, 1} denote the class label corresponding to an
[2]. However, because of poor understanding of the mechanism
and influence factors of rockbust, it is very difficult to give input x. The probability of success p(y=1|x) is related to an
accurate prediction using conventional methods. Some scholars unconstrained latent function f (x).
have made some progress on the rockburst prediction with aid In the GPC model Bayesian inference is performed about
of fast developing machine learning method such as fuzzy the latent function f in the light of observed
method [3], artificial neural networks (ANN) [4] and support
data D = {( x i , y i ) | i = 1, … m} . The prior distribution of latent
vector machines (SVM) [5]. The machine learning method
paves a new way in solving the problems mentioned above. function is:
But some disadvantages still exist in the methods. For

978-1-4244-4507-3/09/$25.00 ©2009 IEEE

 p ( f | X,θ ) = N (0, K ) (1) III. CLASSIFICATON EXAMPLE
We generate a simple artificial classification dataset, by
The covariance matrix is defined element-wise, Kij = k(xi, sampling data points from each of two classes from separate
xj,θ). Here, covariance function k is set as squared exponential Gaussian distributions.
function:
First, we use linear logistic model [8] for binary
classification to classify this data. When the predictive
1
(2) probability p>1, the sampling data belong to class “o”,
k y ( x p , xq ) = σ 2f exp( − ( x p − xq ) 2 )
2l 2 otherwise, they belong to the other class “+”. The results are
not so satisfactory (Fig.2). Note, that the ideal predictive
where, hyperparameters θ={ σ f , l }. probabilities depend only on the relative density of the two
classes, and not on the absolute density. We would expect that
The main purpose of classification models is to predict the the structure in the upper right hand corner of the plot may be
class label y* for test inputs x*. The distribution of y* can be very difficult to obtain based on the samples, because the data
computed by computing the expectation: density is very low.
Then, we fit a probabilistic Gaussian process classifier to
this data, using an implementation of the Expectation
p ( y* | D, θ , x* ) = ∫ p(y * | f* ) p ( f* | D, θ , x* ) df* (3) Propagation algorithm. When the predictive probability p>1,
the sampling data belong to class “o”, otherwise, they belong to
If value of p is great than 0.5, y* =1, otherwise, y* =-1. the other class “+”. The results are satisfactory (Fig.3).
4
Unfortunately, predictive distribution Eq. (3) can be
computed analytically, so approximations are needed. The 3

5
approximate posterior is:

0.
0.5
2

p( f | D, θ ) ≈ q( f | D, θ ) = N (m, A) (4) 1

0
The parameters m and A of the posterior approximation can
be found using Laplace’s method or by Expectation -1 0.5
Propagation. Eq.(5) gives the approximate posterior of f*: 5
-2
0.

q ( f* | D, θ , x* ) = N ( μ* , σ *2 ) (5) -3

-4
where mean and variance are： -4 -3 -2 -1 0 1 2 3 4

Figure 2. Classification using linear logistic model

μ* = k*T K −1m (6) 4

0.5

3 5

σ *2 = k ( x* , x* ) − k*T ( K −1 − K −1 AK −1 )k*
0.
0. 7

(7)
2 0. 7
0.9

k* = [k ( x1 , x* ), …, k ( xm , x* )]T (8) 1

where, k* is a vector of prior covariance between x* and the 0

0.9
training inputs X. x* belonging to class 1 can be computed
analytically: -1

-2
0.5

7
μ* 0.

∫
q( y* = 1 | D, θ , x* ) = Φ ( f* ) N ( f* | μ* ,σ *2 ) df* = Φ (
1 + σ *2
) (9) -3

-4
-4 -3 -2 -1 0 1 2 3 4
The optimum hyper-parameters θ can be derived by
maximum likelihood function: Figure 3. Classification using GP model

Finally, we use SVM model to classify this data. The

p( D | θ ) = ∫ p ( y | f ) p ( f | X, θ ) df (10) results are also satisfactory (Fig.4). However, it is worth while
pointing out that GP does not require a predefined structure,
and includes a theoretical framework for obtaining the
optimum hyperparameters self-adaptively. This overcomes the V. CASE STUDY
open questions of SVM model. Moreover, a further advantage Cases from real engineering located in different countries
of GP is that probabilistic results can be provided comparing are collected [10] as training samples (Table Ⅰ) and five cases
SVM model because probabilistic results can provide a great
help for decision makers to obtain different lever of confidence. are collected as prediction samples(Table Ⅱ), where, intensity
of rockburst is divided into four lever: Ⅰ(violent rockburst), Ⅱ
2.5 (medium rockburst), Ⅲ (light rockburst ), Ⅳ (no rockburst).
2 Prediction results by GP model are shown in Table Ⅱ.
1.5 Through comparison analysis for rockburst prediction
1 probability, we can judge the intensity lever of prediction
samples (Table Ⅱ ). For example, at Yuzhixi hydropower
0.5
station project, predictive probability of medium rockburst is
0 0.51, predictive probability of violent rockburst is 0.36, which
-0.5
means probability of occurrence of medium rockburst is
approximate 50% and probability of occurrence of violent
-1
rockburst reaches 36%, we can judge intensity rockburst at the
-1.5 project should be between medium and violent lever (ⅡtoⅠ
-2
lever), which agrees with reality. At Pubugou hydropower
station, prediction probability of medium rockburst reach 0.86,
-2.5
-3 -2 -1 0 1 2 3 we can infer prediction result should be medium rockburst,
Figure 4. Classification using SVM model
which also agrees with reality. Prediction results at other three
projects (Table Ⅱ) totally agree with reality. The results show
GP model can give accurate rockburst prediction.
IV. GP MODEL FOR ROCKBURST PREDICTION
TABLE I. THE TRAINING SAMPLES OF GP MODEL FOR ROCKBURST
Occurrence of rockburst is influenced by various factors. PREDICTION
Maximum tangential stress in surround rockmass σmax, uniaxial
compressive strength σc, tensile strength of rock σt, and Project σc/MPa σt/MPa σmax/MPa Wet Intensity
Tianshengqiao
rockburst tendency index of rock Wet are suggested to be main 88.7 3.7 30.0 6.6 Ⅱ
II hydropower
influential factors of occurrence of rockburst [3]. Ertan
220 7.4 90.0 7.3 Ⅲ
Classification of rockburst intensity is indeed a kind of hydropower
multi classification problem. Multi classification can be Taipingyi
165 9.4 62.6 9.0 Ⅱ
hydropower
realized by loops of further binary classification of result of
Laxiwa
previous binary classification. GP model for rockburst hydropower
176 7.3 55.4 9.3 Ⅱ
prediction is established as following steps: Jinping II
120 6.5 98.6 3.8 Ⅱ
(1) Historical cases (total number is n) are collected as hydropower
Lubuge
training samples for GP machine learning. Import sample 150 5.4 34.0 7.8 Ⅳ
hydropower
couples for GP is noted: ( xi , yik ) , i=1 ， 2 ，… n ， k=1 ， China Lijiaxia
115 5.0 11.0 5.7 Ⅳ
2 ， … m. Where, xi is n-dimension input vector, which hydropower
Norway Sima
represent some quantitative control influential factors of 180 8.3 48.8 5.0 Ⅱ
hydropower
rockburst. Value of yik is 1or -1, while yik =1, rockburst Norway
180 8.3 75.0 5.1 Ⅱ
intensity is classified as kth level；m is total number of classes Sewage tunnel
of rockburst. Sweden
130 6.0 50.0 5.1 Ⅱ
Forsmark
(2) Loop following steps for m-1 times: Sweden
180 6.7 80.0 5.6 Ⅲ
th
Vietas tunnel
a. Classify the rockburst of predictive case is k lever or not. Japan
236 8.3 89.0 5.0 Ⅱ
Kankoshi road
b. Select corresponding training samples (while k lever Italy Raibl
rockburst occur, yi=1, otherwise, yi=-1). Optima hyper- Zinc sulfate
140 8.0 108.0 5.5 Ⅰ
parameters are obtained by maximization of likelihood Yuzhixi
function (Eq.(10)) of training samples. 170 11.3 90 9.0 To be predicted
hydropower
Pubugou
c. Rockburst prediction probability y* is obtained by Eq.(9), hydropower
123 6.0 43 5.0 To be predicted
while y*>0.5, kth lever rockburst occurred. Longyangxia 178 5.7 19 7.3 To be predicted
hydropower
d. k=k+1，return to step b；if k=m，ends loop. Norway
175 7.3 62 5.0 To be predicted
Heggum road
The algorithm with steps mentioned above was
USSR
implemented in Matlab. Rasvumchorr
180 8.3 57 5.5 To be predicted
 TABLE II. COMPARISON OF REAL ROCKBURST INTENSITY AND rockburst prediction. GP model for rockburst processes unique
PREDICTIVE RESULTS
goodness, which can give probabilistic results which can reflect
Predictive probability Predictive Real uncertainty of rockburst prediction. GP processes favorable
Project
Ⅳ Ⅲ Ⅱ Ⅰ result intensity application prospect in geotechnical engineering with many
Yuzhixi complicated nonlinear problems.
hydropower 0.21 0.33 0.51 0.36 ⅡtoⅠ ⅡtoⅠ
station
Pubugou ACKNOWLEDGMENT
hydropower 0.15 0.05 0.82 0.02 Ⅱ Ⅱ
station
This work is supported by National Natural Science
Longyangxia Foundation of China under Grant No. 50809017 and
hydropower 0.87 0.18 0.42 0.03 Ⅳ Ⅳ Postdoctor Foundation of China under Grant No. 20080440812.
station
Norway
Heggum 0.02 0.18 0.61 0.01 Ⅱ Ⅱ
REFERENCES
road tunnel [1] Cook, N.G.W., “A note on rockbursts considered as a problem of
USSR stability,” J South Af Int. Min Metal., 1965, pp. 437–440.
Rasvumchorr 0.02 0.07 0.86 0.02 Ⅱ Ⅱ [2] Xu, L.S., Wang, L.S., et al., “Present situation of rockburst research at
mine home and abroad,” Journal of Yangtze River Scientific Research,
1999,16(4), pp. 24–27, 38.
[3] Wang, Y.H., Li, Q.G., “Method of fuzzy comprehensive evaluations for
In fact, uncertainty characters of influential factors of rockburst prediction,” Chinese Journal of Rock Mechanics and
rockburst occurrence would result in uncertainty of rockburst Engineering, 1998,17(5), pp. 493–501.
prediction result, it is necessary to descript uncertainty using [4] Feng, X. T., Introduction to Intelligent Rock Mechanics, Beijing :
probability value. GP model can give a probability result Science Press, , 2000.
exactly and process unique goodness in the field of uncertainty [5] Feng, X. T., Zhao, H. B., “Prediction of Rockburst Using Support
evaluation for prediction result, whereas conventional machine Vector Machine.” Journal of Northeastern University (Natural Science),
learning tech such as ANN and SVM which can only give 2002, 23(1), 57–59.
fixed value for prediction result are hard to descript uncertainty [6] Vapnik N., The Nature of Statistical Learning Theory, New York:
Springer, 1995.
of prediction result.
[7] Seeger, M., “Gaussian processes for machine learning,” International
In addition, GP model has merits of self-adaptive Journal of Neural System, 2004,14(2), pp. 69–106.
parameters determination and better capacity for solving non- [8] Rasmussen C. E. and Williams C. K. I., Gaussian Processes for Machine
linear small samples problems comparing with some open Learning, Cambridge : The MIT Press, 2006.
questions of ANN and SVM. [9] Kuss, M. and Rasmussen C.E., “Assessing approximate inference for
binary Gaussian process classification,” Journal of machine learning
research, 2005, 6, pp.1679–1704.
VI. CONCLUSIONS [10] Chen, H.J., Li, N.H., et, al., “Prediction of rockburst by artificial neural
network,” Chinese Journal of Rock Mechanics and Engineering, 2003,
A new model of rockburst prediction based on GP machine 22(5), pp. 726–768
learning is proposed in the paper. Results of case studies show
the model is feasible, effective and simple to implement for