Sunteți pe pagina 1din 16

Applied Mathematical Modelling 33 (2009) 39974012

Contents lists available at ScienceDirect

Applied Mathematical Modelling


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

Geomechanical parameters identication by particle swarm


optimization and support vector machine
Hong-bo Zhao a,*, Shunde Yin b
a
b

School of Civil Engineering, Henan Polytechnic University, Jiaozuo 454003, Peoples Republic of China
Department of Chemical and Petroleum Engineering, University of Wyoming, Laramie, WY 82071, USA

a r t i c l e

i n f o

Article history:
Received 7 January 2008
Received in revised form 12 January 2009
Accepted 22 January 2009
Available online 3 February 2009

Keywords:
Back analysis
Geomechanical parameters identication
Support vector machine
Particle swarm optimization

a b s t r a c t
Back analysis is commonly used in identifying geomechanical parameters based on the
monitored displacements. Conventional back analysis method is not capable of recognizing
non-linear relationship involving displacements and mechanical parameters effectively.
The new intelligent displacement back analysis method proposed in this paper is the combination of support vector machine, particle swarm optimization, and numerical analysis
techniques. The non-linear relationship is efciently represented by support vector
machine. Numerical analysis is used to create training and testing samples for recognition
of SVMs. Then, a global optimum search on the obtained SVMs by particle swarm optimization can lead to the geomechanical parameters identication effectively.
2009 Elsevier Inc. All rights reserved.

1. Introduction
Displacement back analysis is commonly used in establishing geomechanical parameters in rock mechanics and engineering [17]. There are mainly three types of displacement back analysis methods: inverse solving method, atlas method
and direct (i.e. optimal) method [6]. Because of the special advantages, the optimal methods are more and more extensively
employed in solving engineering problems [810]. For example, LevenberMarquardt method, GaussNewton method,
Bayesian method, Powell method, Rosenbork method, and genetic algorithm have been proposed to obtain optimal values
of parameters from measured displacement data. Neural network and genetic algorithm were applied to geotechnical engineering back analysis [1115], and later on neural network was replaced by support vector machine [1618]. This has been
a new way for displacement back analysis. However, there are still two problems associated with the optimal method unresolved completely. One is that the relation between the displacement and the mechanical parameters is highly non-linear
and complex, although support vector machine can provide appropriate techniques to learn and represent this non-linear
relation [19,20]. Another is that the search for the estimated parameter values is in a large space and is highly multi-modal.
Some existing techniques such as calculus-based and enumerative techniques are mostly insufcient to handle this
problem.
In this paper, the particle swarm optimization [21] was chosen for its biological and evolutionary appeal in nding the set
of unknown parameters that best matches the modeling prediction with the measured displacement data. Then a new intelligent displacement back analysis method incorporating a support vector machine and a particle swarm optimization is presented in detail. By using this method, the estimation of the geomechanical parameters at the permanent shiplock at the
Three Gorges Project is illustrated.

* Corresponding author.
E-mail address: bxhbzhao@hotmail.com (H.-b. Zhao).
0307-904X/$ - see front matter 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.apm.2009.01.011

3998

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

2. Support vector machine and particle swarm optimization


2.1. Support vector machine
The SVM was proposed by Vapnik [19]. It is used to train non-linear relationships based on the structural risk minimization principle that seeks to minimize an upper bound of the generalization error rather than to minimize the empirical error
implemented in neural networks. This induction principle is based on the fact that the generalization error is bounded by the
sum of the empirical error and a condence interval term that depends on the VapnikChervonenkis (VC) dimension. Using
this principle, the SVM will achieve an optimal model structure by striking the right balance between the empirical error and
the VC-condence interval, eventually giving a better generalization than the neural networks. Another merit of the SVM is
that the training is a uniquely solvable quadratic optimization problem. The SVM uses non-linear mapping based on an internal integral function to transform an input space to a high dimension space and then looks for a non-linear relationship between inputs and outputs in that space. The SVM not only has theoretical support but also can nd global optimum solutions
for problems with small training samples, high dimensions, non-linear and local optima. A wide variety of applications such
as pattern recognition, non-linear regression, etc. have empirically shown the SVMs ability to generalize.
Suppose that we are given a set of observation data (samples) (X1, y1), (X2, y2), . . . , (Xk, yk), Xi 2 Rn, yi 2 R. For the regression
problem based on the SVM, our goal is to nd a linear function f(X) = w  X + b that has at most e deviations from the actually
obtained targets yi for all the training data, and is also as at as possible. We describe the linear function f, in the form

f X w  X b:

1
2

One way to ensure this is to minimize the squared Euclidean norm, i.e. kwk . Formally we can write this problem as a
convex optimization problem by requiring:

minimize
subject to

1
kwk2 ;
2

yi  w  X i  b 6 e;
w  X i b  yi 6 e;

2
i 1; . . . ; n:

According to statistics learning theory, we want to obtain the optimal generalization. Allowing for some errors, we can
introduce the slack variables ni P 0 and ni P 0. Hence we arrive at the following:

minimize
subject to

k
X
1
ni ni ;
kwk2 C
2
i1

yi  w  X i  b 6 e ni ;

w  X i b  yi 6 e; ni ;

3
i 1; . . . ; n:

The constant C > 0 determines the tradeoff between the atness of f and the amount up to which deviations larger than e
are tolerated. Using the optimal method, we can obtain the dual optimization problem

maximize Wa; a 

n
n
n
X
X
1X
ai  ai aj  aj X i  X j
yi ai  ai  e
ai ai ;
2 i;j1
i1
i1

8 n
< Pa  a 0;
i
i
subject to
i1
:

0 6 ai ; ai 6 C;

i 1; 2; . . . ; n:

Solving the above optimal problems, we can obtain the regression function of the SVM

f X

n
X
ai  ai X  X i b:

i1

Based on the KarushKuhnTucker (KKT) conditions for quadratic programming, only a small number of coefcients (ai  ai
will be assumed to have non-zero values, and their data points could be referred to as support vectors.
The non-linear problems can be solved by mapping the data into a high feature dimension space. In high feature dimension space, the inner product can be replaced by the kernel function, i.e. K(Xi, Xj) = / (Xi)/(Xj); we do not need to know the
specic formulation of the non-linear mapping /. So Eqs. (3)(5) can be changed into the following formulation:

maximize Wa; a 

n
n
n
X
X
1X
ai  ai aj  aj KX i  X j
yi ai  ai  e
ai ai ;
2 i;j1
i1
i1

8 n
< Pa  a 0;
i
i
i 1; 2; . . . ; n;
subject to
i1
:

0 6 ai ; ai 6 C;
n
X
f X
ai  ai KX  X i b:
i1

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

3999

The following is three general kernel functions:


(1) Polynomial kernel

KX; Y X  Y 1d ;

d 1; 2; . . . ; n:

10

(2) Radial kernel

(
KX; Y exp 

jX  Yj2

)
:

11

KX; Y tanh/X  Y h:

12

r2

(3) Sigmoid kernel

We can solve the quadratic optimal problem of Eqs. (4), (5), (7), and (8) by a variety of methods, including the interior point
algorithm, the sequential minimal optimization [20], and decompression algorithms, etc.

2.2. Particle swarm optimization


The particle swarm optimization (PSO) was originally designed by Kennedy and Eberhart and has been compared to genetic algorithms for efciently nding optimal or near-optimal solutions in large search spaces [21]. The technique involves
simulating social behavior among individuals (particles) ying through a multi-dimensional search space, each particle
representing a single intersection of all search dimensions. The particles evaluate their positions relative to a goal (tness)
at every iteration. Particles in a local neighborhood share memories of their best positions, and then use those memories to
adjust their own velocities and thus subsequent positions. The original formula developed by Kennedy and Eberhart was improved by Shi and Eberhart with the introduction of an inertia parameter that increases the overall performance of PSO.
The original PSO formulae dene each particle as a potential solution to a problem in D-dimensional space, with particle i
represented Xi = (xi1, xi2, . . . , xiD). Each particle also maintains a memory of its previous best position, Pi = (pi1, pi2, . . ., piD), and a
velocity along each dimension, represented as Vi = (vi1, vi2, . . . , viD). At each iteration, the P vector of the particle with the best
tness in the local neighborhood, designated g, and the P vector of the current particle are combined to adjust the velocity
along each dimension, and that velocity is then used to compute a new position for the particle. The portion of the adjustment to the velocity inuenced by the individuals previous best position (P) is considered the cognitioncomponent, and the
portion inuenced by the best in the neighborhood is the socialcomponent.
Regarding the minimum problem, suppose f(X) is the objection function, Xi = (xi1, xi2, . . . , xin) is the current position of particle, Vi = (vi1, vi2, . . . , vin) is the current speed of particle, Pi = (pi1, pi2, . . . , pin) is the best position which particle ied, then the
best position of particle i can be computed according following formulation:

Pi t 1

Pi t

if f xi t 1 P f P i t;

X i t 1 if f xi t 1 < f P i t:

13

If the population is s, and Pg(t) is the global best position which all particle ied the best position, then

Pg t 2 fP 0 t; P1 t; . . . ; Ps tgjf P g t minff P0 t; f P1 t; . . . ; f P s tg:

14

According to the theory of particle swarm optimization, the following equation presents the process of evolutionary:

v i t 1 wv i t c1 r1 tpij t  xi t c2 r2 tpg t  xi t;

15

xij t 1 xij t v ij t 1;

16

where vi is the velocity for particle i, which represents the distance to be traveled by this particle from its current position; xij
represents the position of particle i; pij represents the best previous position of particle i; pg represents the best position
among all particles in the population; r1 and r2 are two independently uniformly distributed random variables with range
[0, 1]; c1 and c2 are positive constant parameters called acceleration coefcients which control the maximum step size; w
is the inertia weight that controls the impact of previous velocity of particle on its current one. In standard PSO, Eq. (15)
is used to calculate the new velocity according to its previous velocity and to the distance of its current position from both
its own best historical position and its neighbors best position. Generally, the value of each component in vi can be clamped
to the range [vmax, vmax] to control excessive roaming of particles outside the search space. Then the particle ies toward a
new position according to Eq. (16). This process is repeated until a user-dened stopping criterion is reached.
3. An intelligent back analysis algorithm
In this section, an intelligent displacement back analysis algorithm based on the integration of particle swarm optimization (PSO), support vector machine (SVM), and numerical analysis is proposed. The algorithm is described as follows (see
Fig. 1).

4000

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Generate randomly N particles (SVM parameters)

SVM learning

Calculate the fitness of each particle(Eq.18)

Produce N new
particles using
PSO algorithm

No
SVM model is OK?
Recognize the SVM model
Yes

Produce N new
particles (parameters to
be recognized)

SVM model for back analysis

Calculate the fitness of each particle(Eq.19)

Error between monitored


and predicted by SVM is
Minimum?
Recognize rock mass parameter

No

Produce N new
particles using
PSO algorithm

Yes
Get the parameters to be recognized

Fig. 1. Intelligent displacement back analysis based on PSO and SVM.

3.1. Representation of non-linear relationship


The non-linear relationship between displacement and geomechanical parameters can be described using a support vector machine SVM(X) as

SVMX : Rn ! R;

17

Y SVMX;
X x1 ; x2 ; . . . ; xn ;
Y y1 ; y2 ; . . . ; yn ;
where xi(i = 1, 2, . . ., n) is geomechanical parameters, for example, Youngs modulus, friction angle, geo-stress coefcients, etc.
yi(i = 1, 2, . . . , n) is displacements of the key points.
In order to obtain SVM(X), a training process based on the known data set is needed. The training of SVM includes creation of training samples using numerical analysis and determination of training parameters of SVM. The former is performed by using numerical analysis for the given set of tentative geomechanical parameters to obtain the corresponding
displacement of rock mass of key points. Considering inuence of training parameters on generalization performance of
SVM, particle swarm optimization is adopted to search the training parameters in global space. The algorithm is described
as follows for this purpose:
Step 1:

Estimate the valuing ranges of geomechanical parameters to be recognized. A set of tentative geomechanical
parameters is given in their valuing ranges. Numerical analysis is used to calculate the corresponding displacement
of key points for the every set of tentative geomechanical parameters. Each set of geomechanical parameters with
the corresponding displacement of key points is considered as a training sample set. In order to obtain the best

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Step 2:
Step 3:

Step 4:
Step 5:

generalization performance of SVMs both for training samples and new samples having similar conditions, another
set of samples, therefore, should be created to test applicability of SVMs. They are called to be the testing sample
set.
Initialize parameters of particle swarm optimization such as number of evolutionary generation, population size,
inertia weight, acceleration coefcients, range of kernel function and its parameters including C and r.
Select randomly a kernel function from common examples of kernel functions such as polynomials, Gaussian radial
base, and sigmoid. Produce randomly a set of C and r in the given valuing ranges. Each selected kernel function and
its parameters such as C and r is regarded as an individual of SVM.
Use sequential minimal optimization (SMO) algorithm to solve the quadratic programming problems including
each individual to obtain values of Lagrange multipliers and their support vectors.
Use the selected parameters and the obtained support vectors to represent a SVM model. The testing samples is
used to test prediction ability of the SVMs. Applicability of the model is measured by tness as

fitness MaxjY ij  Y 0ij j=Y ij ;

Step 6:
Step 7:

4001

18

where Yij and Y 0ij are the estimated displacement of tentative SVM and calculation of numerical analysis for key
point i of rock mass at the jth testing sample.
If tness is accepted then the training procedure of SVM ends and the best SVMs are found. Otherwise, according to
Eqs. (15) and (16), produce the new particle.
If all new individuals of population size are generated then go to Step 4. Otherwise, go to Step 6.

3.2. Objective function


To formulate the particle swarm optimization to this problem, as in any conventional approach to the displacement back
analysis, objective function must be dened. Here the objective function is dened as

fitness

n
1X
jSVM i X  yi j;
n i1

19

where n is the number of key points, yi is the monitored displacement of the ith key points, SVMi(X) is the predicted displacement of ith key point.
3.3. An intelligent displacement back analysis algorithm based on particle swarm optimization
If support vector machine model representing the non-linear relation between the displacement and a parameter is obtained, the model can be used to recognize parameters. Particle swarm optimization is used to search for the best parameter
system having the minimum error between the predicted displacements as predicted by the model and the actual measured
displacement. This back analysis algorithm can be described as follows:
Step 1:
Step 2:
Step 3:
Step 4:
Step 5:
Step 6:
Step 7:
Step 8:
Step 9:

Determine the particle swarm optimization parameters and the range of parameters to be recognized.
Generate randomly n group of parameters at their given range. Each individual represents an initial solution.
Input a set of rock mass parameters to the model SVM(X) obtained above to calculate the displacement values at
given monitoring points.
Use Eq. (19) to evaluate the tness of the current individuals, i.e. the reasonability of the parameter set.
If all individuals are evaluated, then go to Step 6. Otherwise, go to Step 3.
If the given evolutionary generation is reached, or the best individuals (the parameter to be back recognized) are
obtained, then the evolutionary process ends. Otherwise, go to Step 7.
Update the individuals according to Eqs. (15) and (16).
Repeat Step 7 until all n new individuals are generated. They are used as offspring.
Go to Step 3.

4. Verication of the algorithm


To verify the model, we suppose there is an innite large plate with a hole of radius 1 m in centre (Fig. 2). Suppose there
are theoretical values for Poissons ratio l = 0.25, equably distribution of initial geo-stress with rx = rz = 0.98 MPa, sxz = 0 and
Youngs modulus E = 98 MPa. Displacement values for some key points, indicated by nodes 15 in Fig. 2, are calculated by
elastic nite element method. The suggested algorithm above is used to identify initial geo-stress components rx, rz, sxz.
Before the training of SVMs, totally 25 set of training samples and ve testing samples are created using nite element
analysis (Table 1). The established SVM gave its recognized initial geo-stress rx, rz, sxz of about 0.9787, 0.9762, 0 MPa, respectively, having maximum relative error 0.39% corresponding to the theoretical values (Table 2). The comparison with evolu-

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

4002

Fig. 2. A quarter calculation model for FEM. The number is no. of key points to be compared.

Table 1
Training and testing samples.

rx0 (Mpa)

ry0 (Mpa)

sxy (Mpa)

Displacement (cm)
1

u1x

u2x

u2y

3
u3x

u3y

4
u4x

u4y

u5y

Training samples
1
0.6
2
0.6
3
0.6
4
0.6
5
0.6
6
0.8
7
0.8
8
0.8
9
0.8
10
0.8
11
1
12
1
13
1
14
1
15
1
16
1.2
17
1.2
18
1.2
19
1.2
20
1.2
21
1.4
22
1.4
23
1.4
24
1.4
25
1.4

0.58
0.78
0.98
1.18
1.38
0.58
0.78
0.98
1.18
1.38
0.58
0.78
0.98
1.18
1.38
0.58
0.78
0.98
1.18
1.38
0.58
0.78
0.98
1.18
1.38

0.2
0.4
0
0.2
0.4
0.4
0
0.2
0.4
0.2
0
0.2
0.4
0.2
0.4
0.2
0.4
0.2
0.4
0
0.4
0.2
0.4
0
0.2

0.45136
0.55592
0.10498
0.16394
0.43526
0.80309
0.35680
0.09064
0.17793
0.29721
0.60731
0.34388
0.07804
0.54849
0.64953
0.59583
0.33270
0.79844
0.89848
0.45379
0.58607
1.04710
1.14610
0.70607
0.44256

0.41508
0.53144
0.14153
0.08339
0.31123
0.71490
0.32912
0.10639
0.11925
0.34020
0.51657
0.29602
0.07257
0.52643
0.63941
0.48551
0.26425
0.71250
0.82379
0.44090
0.45578
0.89844
1.00800
0.62926
0.40954

0.17422
0.25745
0.13103
0.07435
0.01762
0.26542
0.13868
0.08172
0.02473
0.24894
0.14672
0.08950
0.03224
0.25648
0.33821
0.09769
0.04017
0.26442
0.34603
0.22157
0.04850
0.27276
0.35425
0.22946
0.17328

0.31696
0.44528
0.19138
0.06289
0.06706
0.50479
0.25213
0.12404
0.00553
0.38353
0.31386
0.18617
0.05700
0.44313
0.56879
0.24929
0.12053
0.50370
0.62810
0.37929
0.18506
0.56523
0.68837
0.44081
0.31483

0.32117
0.51236
0.32242
0.25972
0.19744
0.44891
0.25661
0.19242
0.12864
0.57823
0.19180
0.12615
0.06087
0.51266
0.70056
0.06089
0.00586
0.44809
0.63634
0.45013
0.07155
0.38451
0.57310
0.38454
0.32189

0.17419
0.26666
0.14968
0.10227
0.05474
0.25595
0.13882
0.09122
0.04349
0.27703
0.12839
0.08060
0.03268
0.26589
0.35697
0.07041
0.02231
0.25517
0.34600
0.23111
0.01238
0.24487
0.33544
0.22041
0.17366

0.41845
0.71794
0.52279
0.49379
0.46618
0.53430
0.33510
0.30398
0.27425
0.90586
0.14738
0.11416
0.08232
0.71962
1.01470
0.07569
0.10964
0.53336
0.83021
0.63911
0.30163
0.34708
0.64572
0.45057
0.42130

0.45077
0.79800
0.60842
0.59953
0.59234
0.55363
0.36072
0.35037
0.34174
1.04470
0.11185
0.10006
0.08999
0.79880
1.14190
0.15143
0.16292
0.55170
0.89699
0.71104
0.41704
0.30342
0.65093
0.46167
0.45313

Testing samples
26
0.6
27
0.8
28
1
29
1.2
30
1.4

0.78
1.18
0.58
0.98
1.38

0.2
0
0.4
0.2
0.4

0.00039
0.19435
0.97049
0.79844
0.25997

0.02212
0.22405
0.84716
0.71250
0.24143

0.04618
0.16593
0.28662
0.26442
0.10329

0.05931
0.25474
0.56582
0.50370
0.18678

0.12642
0.38713
0.44890
0.44809
0.19302

0.05562
0.18465
0.26781
0.25517
0.10381

0.21782
0.60749
0.48342
0.53336
0.25441

0.25786
0.69966
0.47531
0.55170
0.27414

tionary support vector machine is shown in Table 2. Comparing displacement at x and y-directions of ve key points using
the theoretical and the identied values of mechanical parameters are shown in Fig. 3. The convergence process of algorithm
is shown in Fig. 4. Variation of the recognized parameters with generations is shown in Fig. 5.

4003

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012


Table 2
Comparison of back analysis results and theoretic solutions.
Parameters to be recognized (MPa)

rx0
ry0
sxy

Theoretical

SVM + PSO

0.98
0.98
0

SVM + GA [18]

Recognized

Absolute error

Relative error(%)

Recognized

Absolute error

Relative error (%)

0.97874
0.97618
0.00000

0.00126
0.00382
0.00000

0.12827
0.39000
0.00560

0.97442
0.96808
0.00560

0.00558
0.01192

0.56908
1.21602

Key points
u1 x

u2 x

u2y

u3 x

u3 y

u4 x

u4 y

-0.05

-0.1

-0.2

-0.25

-0.3

-0.35

-0.4
calculated displacement using theoerical parameters

-0.45

Calculated displacement using recognized parameters


-0.5
Fig. 3. Comparison of displacements from back analysis and nite element calculation.

0.25
0.2

Fitness

displacement (mm)

-0.15

0.15
0.1
0.05
0

Generation
Fig. 4. Variation of tness values with generations.

u5 y

4004

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Generation

a
-0.92

x0(Mpa)

-0.94
-0.96
-0.98
-1
-1.02
-1.04
-1.06

y0(Mpa)

279

237

218

216

203

198

192

173

167

79

58

17

Generatiom

xy(Mpa)

-0.95
-0.955
-0.96
-0.965
-0.97
-0.975
-0.98
-0.985
-0.99
-0.995
-1
-1.005

Generation

0
-0.005
-0.01
-0.015
-0.02
-0.025
-0.03
-0.035
-0.04
-0.045
-0.05

Fig. 5. Variation of the recognized parameters with generations: (a) rx0, (b) ry0, and (c) sxy.

5. Case studies
The permanent shiplock is one of the major components of Three Gorges Project in China. It is one of the largest articial
navigation structures excavated in a rock mass in the world. The permanent shiplock is located on the right side of the Yangtze River. The permanent shiplock is constructed along an azimuth direction of 118 as a double channel with ve stages
(Fig. 6) and a total length of 1617 m. The single shiplock room is excavated in granite and is 280 m in length, 34 m in width,
and 5 m in depth for storing water. Both sides of the shiplock room are high and steep granite slopes. The sidewall of the
shiplock room is vertically cut with a height of 4050 m. The deepest excavation is about 170 m deep. The section 1717
is located at the head of the third shiplock room. Its slope is the highest one of the permanent shiplock area. Also, the geological structures there are the most complex and there is a fault (F215) in this area. The design and stability analysis of this
slope is crucial in the construction of the permanent shiplock. For both analysis and design, proper recognition of the rock
mass mechanical parameters is key problem.
The section 1717 is located at the top of the third shiplock room. The corresponding engineering geological zones are
shown in Fig. 6. They consist mainly of a hard and intact rock mass slightly weathered in some places. It is suitable for deep
excavation of high and steep slopes. The mechanical properties of the rock mass do not vary at the slightly weathered and
fresh zones. The upper portion of the strata is completely weathered thin strata and next to it is a moderately weathered

4005

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Fig. 6. Engineering geology of section 1717 (scale: 1:500).

zone. Therefore, mechanical parameters for these natural rock mass zones, such as the slightly and non-weathered zone, the
completely weathered zone, and the moderately weathered zone are considered to be recognized. The in situ stress is disturbed due to excavation of the slope, which generates unloading and relaxation zones. Because of excavation and blasting, a
damage zone forms at the boundary of the excavation. Therefore, there are two other zones related to engineering activities,
unloading and damaged zones, to be considered in the back analysis of displacement.
Some researchers have suggested regressive equations [12] for calculating stress elds from in situ measurement in the
region. In this study, Fengs assumptions for back analysis were adopted [13]. Therefore, we have the following equations for
slightly weathered and fresh rock mass:

rx ax 0:01168H Mpa;
ry ay 0:03039H Mpa;
rz 4:7152 0:01027H Mpa;
sxy 0; syz 0; szx 0:

20

And the following equations for the strongly weathered rock mass:

rx lcH;
ry cH;
rz cH=1  l;
sxy 0; syz 0; szx 0;

21

where c is the unit weight of the rock mass (106 N/m3), H is depth below ground surface (m), and ax, ay are coefcients to be
back analyzed.

Table 3
Arrangement of displacement monitoring points for section 1717 of the permanent shiplock.
Monitoring point no.

Height above sea level (m)

Location

Date at initial monitoring

TP/BM10GP01
TP/BM11GP01
TP/BM26GP02
TP/BM27GP02
TP/BM28GP02
TP/BM29GP02

230
200
170
200
230
245

North
North
South
South
South
South

June 1995
January 1995
November 1996
November 1995
March 1995
December 1994

slope
slope
slope
slope
slope
slope

4006

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Table 4
The learning and testing samples.
Samples
no.

Parameters to be recognized

Computational displacement for monitoring points (mm)

Youngs modulus E (Gpa)

TP/
BM10GP01

TP/
BM11GP01

TP/
BM26GP02

TP/
BM27GP02

TP/
BM28GP02

TP/
BM29GP02

Coefcient
of
geostress
eld

Dz

Udz

Swf

ax

ay

Learning samples
1
6.0
2
6.0
3
6.0
4
6.0
5
6.0
6
8.0
7
8.0
8
8.0
9
8.0
10
8.0
11
10.0
12
10.0
13
10.0
14
10.0
15
10.0
16
12.0
17
12.0
18
12.0
19
12.0
20
12.0
21
15.0
22
15.0
23
15.0
24
15.0
25
15.0
26
6.0
27
6.0
28
6.0
29
6.0
30
6.0
31
8.0
32
8.0
33
8.0
34
8.0
35
8.0
36
10.0
37
10.0
38
10.0
39
10.0
40
10.0
41
12.0
42
12.0
43
12.0
44
12.0
45
12.0
46
15.0
47
15.0
48
15.0
49
15.0
50
15.0

8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0
8.0
10.0
12.0
15.0
18.0

15.0
18.0
20.0
23.0
25.0
18.0
20.0
23.0
25.0
15.0
20.0
23.0
25.0
15.0
18.0
23.0
25.0
25.0
28.0
20.0
25.0
15.0
18.0
20.0
23.0
15.0
18.0
20.0
23.0
25.0
18.0
20.0
23.0
25.0
15.0
20.0
23.0
25.0
15.0
18.0
23.0
25.0
15.0
18.0
20.0
25.0
15.0
18.0
20.0
23.0

25.0
28.0
30.0
32.0
35.0
30.0
32.0
35.0
25.0
28.0
35.0
25.0
28.0
30.0
32.0
28.0
30.0
32.0
35.0
25.0
32.0
35.0
25.0
28.0
30.0
32.0
35.0
25.0
28.0
30.0
25.0
28.0
30.0
32.0
35.0
30.0
32.0
35.0
25.0
28.0
35.0
25.0
28.0
30.0
32.0
28.0
30.0
32.0
35.0
25.0

3.0
4.0
5.0
6.0
7.0
6.0
7.0
3.0
4.0
5.0
4.0
5.0
6.0
7.0
3.0
7.0
3.0
4.0
5.0
6.0
5.0
6.0
7.0
3.0
4.0
7.0
3.0
4.0
5.0
6.0
5.0
6.0
7.0
3.0
4.0
3.0
4.0
5.0
6.0
7.0
6.0
7.0
3.0
4.0
5.0
4.0
5.0
6.0
7.0
3.0

0.8
1.2
0.6
0.8
2.0
2.0
0.8
1.2
1.6
1.8
1.8
2.0
0.8
1.2
1.6
1.6
1.8
2.0
0.8
1.2
1.2
1.6
1.8
2.0
0.8
1.8
2.0
0.8
1.2
1.6
1.6
1.8
2.0
0.8
1.2
1.2
1.6
1.8
2.0
0.8
0.8
1.2
1.6
1.8
2.0
2.0
0.8
1.2
1.6
1.8

13.11
16.44
19.79
23
24.94
23.58
26.28
9.56
18.7
20.92
12.81
23.15
25.47
27.62
10.38
29.31
10.87
13.8
16.31
28.17
17.49
19.39
32.12
11.45
14.85
26.28
9.56
18.65
21.51
24.75
23
25.5
28.35
10.58
13.13
10.77
14.17
16.69
28.1
29.9
19.95
32.97
11.41
14.92
18.1
15.59
18.26
21.38
23.37
12.86

13.84
16.98
20.21
23.38
25.26
23.92
26.62
10.02
19.17
21.4
13.17
23.45
25.78
28.04
10.89
29.51
11.37
14.26
16.73
28.47
17.88
19.82
32.48
12.05
15.4
26.63
9.98
19.29
21.99
25.16
23.43
25.79
28.54
11.14
13.62
11.32
14.55
16.98
28.49
30.24
20.27
33.19
12.01
15.39
18.42
16.07
18.87
21.77
23.66
13.59

17.87
21.96
26.51
30.51
33.14
31.19
35.16
13.18
23.96
28.45
17.74
30.17
33.86
37.67
14.24
38.13
14.66
19.34
23.62
37.48
24.09
27.55
42.73
15.58
20.58
34.85
13.05
24.48
28.41
32.44
30.34
33.48
36.92
14.62
18.59
14.79
19.12
22.79
37.46
40.13
27.42
42.63
15.71
20.46
24.97
21
26.34
30
32.48
17.29

13.47
15.98
18.97
22.18
24.52
22.56
25.94
9.63
17.79
19.93
12.33
21.79
24.73
26.98
10.15
28.24
10.59
13.39
16.51
26.95
17.17
19.17
30.7
11.22
14.98
25.67
9.19
18.36
20.76
23.81
21.76
24.26
27.24
11.21
13.01
10.65
13.62
16.05
26.54
29.02
19.93
31.6
11.13
14.44
17.36
15.07
18.28
21.02
23.08
12.74

11.6
14.29
17.12
19.79
21.46
20.14
22.43
8.23
16.07
17.89
10.84
19.89
21.94
23.32
8.73
24.94
9.17
11.66
14.11
24.17
15.11
16.42
27.36
9.66
13
22.25
7.97
16.39
18.75
21.39
19.86
21.8
24.13
9.46
11.23
9.2
12.06
14.29
23.92
25.44
17.05
28.19
9.6
12.68
15.41
13.34
15.79
18.23
19.76
11

11.37
14.37
17.37
20.03
21.56
20.6
22.73
8.3
16.26
18.23
11.23
20.22
22.13
23.75
8.94
25.44
9.4
12.02
14.33
24.39
15.48
16.95
27.77
9.85
13
22.69
8.2
16.22
18.8
21.5
20.2
22.19
24.51
9.21
11.41
9.37
12.41
14.72
24.3
25.7
17.42
28.55
9.82
13
15.82
13.66
16.08
18.65
20.26
11.16

Testing samples
51
6.0
52
6.9
53
7.8
54
8.7
55
9.6
56
10.5
57
11.4
58
12.3
59
13.2
60
14.1
61
15.0

9.0
11.0
13.0
15.0
17.0
8.0
10.0
12.0
14.0
16.0
18.0

17.0
20.0
23.0
15.0
18.0
21.0
24.0
16.0
19.0
22.0
25.0

29.0
34.0
28.0
33.0
27.0
32.0
26.0
31.0
25.0
30.0
35.0

5.4
3.8
6.6
5.0
3.4
6.2
4.6
3.0
5.8
4.2
7.0

1.88
1.76
1.64
1.52
1.40
1.28
1.16
1.04
0.92
0.80
2.00

22.00
12.82
28.64
17.55
14.12
22.70
21.10
10.30
26.64
15.68
23.70

22.41
13.20
28.87
17.99
14.70
23.02
20.55
10.89
27.04
16.20
23.89

29.24
17.35
37.25
24.72
18.83
30.45
26.70
14.54
36.09
21.77
32.39

20.90
12.28
27.37
17.00
13.64
22.18
19.38
10.39
25.79
15.71
23.23

18.91
10.85
24.50
15.02
12.01
19.36
17.39
8.79
22.97
13.68
20.13

19.29
11.16
24.78
15.39
12.19
19.79
17.62
8.93
23.16
13.72
20.59

Wwz

4007

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

The parameters to be back analyzed are the coefcients ax and ay (Eq. (20)) for the geostress equation and the deformation
modulus for four kinds of rock mass zones (moderately weathered zone, slightly weathered or fresh zone, unloading deformation zone and damaged zone). Data for deformation modulus, Poissons ratio, and weight for the strongly weathered zone
and the fault F215 were provided by the Yangtze River Water Conservancy Committee. The ranges for these six parameters
to be back analyzed are determined based on the eld monitoring data as follows [13]:
(1)
(2)
(3)
(4)
(5)

Deformation modulus for moderately weathered zone Wwz: 615 GPa.


Deformation modulus for damaged zone Dz: 818 GPa.
Deformation modulus for unloading deformation zone Udz: 1525 GPa.
Deformation modulus for slightly or non-weathered zone Swf: 2535 GPa.
The geostress coefcients ax: 37 and ay: 0.82.0.

The rock masses in all zones are considered to be plastic. Their cohesion c and friction angle f are determined directly from
engineering tests and previous monitoring data [13].

Table 5
Identied the parameters and its comparisons.
Method

Youngs modulus E (Gpa)

SVM + PSO
NN + GA [13]

Coefcient of geostress eld

Wwz

Dz

Udz

Swf

ax

ay

6.000
7.515

9.498
9.683

17.313
18.95

29.253
32.1

4.355
4.793

1.370
1.599

25
Monitored dis placement
Predicted displacement using SVM

Displacement (mm)

20

15

10

TP/BM 1 0 GP0 1 TP/BM 11 GP0 1 TP/BM 2 6 GP0 2 TP/BM 27 GP0 2 TP/BM 2 8 GP0 2 TP/BM 29 GP0 2

Monitored points
Fig. 7. Comparison of displacements from back analysis and monitored displacement.

Table 6
Comparison of measured displacement with predictions from support vector machine and neural network model.
Displacement at monitoring point (mm)

Measurement
Prediction of SVM model
Prediction of NN model [13]

Average absolute error

TP/BM10GP01

TP/BM11GP01

TP/BM26GP02

TP/BM27GP02

TP/BM28GP02

TP/BM29GP02

16.32
17.26
17.36

19.11
17.75
17.79

20.76
23
23.64

16.71
16.71
16.85

19.1
15.1
14.91

16.71
14.93
15.29

1.72
1.83

4008

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

1.75

10
9
8
7
6
5
4
3
2
1
0

1.745
1.74

fitness

Fitness

Feng et al. used the DruckerPrager model to analyze in the FLAC2D and obtained totally 50 sets of training samples and
11 testing samples (Table 4) [13]. The computational model scope was 1000 m in the x-direction, 500 m extended from the
isolated rock mass, and 510 m in the y-direction. There are more than 15,000 nodes and 15,000 quadrilateral elements. Information of displacement monitoring points for section 1717 is listed in Table 3 and the location is shown in Fig. 6. Based on
above algorithm for recognizing rock mass parameters, the value of geostress coefcients and the deformation module for

1.73
1.725
1.72
1.715

20

40

20

60

40

population

population

(a) Initial generation

(b) 100th generation

1.736

1.736

1.734

1.734

1.732

1.732

1.73

60

fitness

1.73

1.728
1.726

1.728
1.726

1.724

1.724

1.722

1.722

1.72

1.72

1.718

1.718

20

40

60

20

population

40

population

(d) 500th generation

(c) 200th generation

Fig. 8. Variation of tness values with generations.

8.5
8

Wwz (GPa)

fitness

1.735

7.5
7
6.5
6
5.5

Generation
Fig. 9. Variation of the recognized parameters Wwz with generations.

60

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

4009

10.2
10
9.8

Dz (GPa)

9.6
9.4
9.2
9
8.8
8.6
8.4
8.2
8

Generation
Fig. 10. Variation of the recognized parameters Dz with generations.

17.35
17.3

Udz (GPa)

17.25
17.2
17.15
17.1
17.05
17
16.95

Generation
Fig. 11. Variation of the recognized parameters Udz with generations.

29.7
29.6

Swf (GPa)

29.5
29.4
29.3
29.2
29.1
29

Generation
Fig. 12. Variation of the recognized parameters Swf with generations.

four kinds of rock mass zones are obtained as 4.3546, 1.3697, 6.0002, 9.4975, 17.3129 and 29.2527 GPa, respectively. Its
comparison with the Fengs results is listed in Table 5. Comparison of displacements at monitoring points between monitored and the identied values of mechanical parameters are shown in Fig. 7 and Table 6. We can see the proposed method
is better than NN-based method from Table 6. The convergence process of algorithm is shown in Fig. 8. Variation of the recognized parameters with generations is shown in Figs. 914.

4010

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

4.65
4.6
4.55
4.5

ax

4.45
4.4
4.35
4.3
4.25
4.2

Generation
Fig. 13. Variation of the recognized parameters ax with generations.

1.38
1.36
1.34
1.32

ay

1.3
1.28
1.26
1.24
1.22
1.2
1.18

Generation
Fig. 14. Variation of the recognized parameters ay with generations.

0.5
0.45

Minimum fitness

0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

363

374

398

417

488

650

840

C
Fig. 15. Relationship between the parameters C of SVM and minimum tness of the tentative SVM.

6. Discussions
Relation between rock mass parameters and displacements is a key element in back analysis. It affects directly the results
of rock mass to be recognized (see Figs. 15 and 16). Support vector machine represents well the relationship between rock
mass parameters and displacements (see Fig. 17). Support vector machine model standing of numerical analysis in back
analysis procedure improved the efciency of back analysis (see Fig. 18). Particle swarm optimization has strong capability
of global searching, and it improves the generalization performance in searching the support vector machine model. This

4011

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

0.5
0.45

Minimum fitness

0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

18

24

25

26

27

27

32

34

Fig. 16. Relationship between the parameters r of kernel function and minimum tness of the tentative SVM.

40

Predicted displacement by SVM (mm)

35

TP/BM10GP01
TP/BM26GP02
TP/BM28GP02

30

TP/BM11GP01
TP/BM27GP02
TP/BM29GP02

25

20

15

10

0
0

10

15

20

25

30

35

40

Monitored displacement (mm)


Fig. 17. Comparison between monitored displacement and predicted displacement by SVM of testing samples.

1.9
1.85

fitness

1.8
1.75
1.7
1.65
1.6

Generation
Fig. 18. Change tendency of the minimum tness of SVMs versus number of evolutionary generation.

4012

H.-b. Zhao, S. Yin / Applied Mathematical Modelling 33 (2009) 39974012

Table 7
The results based on different searching bound.

1
2
3
4
5

Range pf searching

rx0 (Mpa)

ry0 (Mpa)

sxy (Mpa)

Fitness

[1, 0]
[2, 0]
[5, 0]
[10, 0]
[20, 0]

0.978743
0.978743
0.978743
0.978743
0.978743

0.976178
0.976178
0.976178
0.976178
0.976178

0
0
0
0
0

0.01031
0.01031
0.01031
0.01031
0.01031

makes it possible to nd the rock mass parameters in a big global space (see Table 7), which enables the back analysis to be
applied to more complex engineering problems.
7. Conclusions
The paper presents a new intelligent back analysis method for recognizing the geomechanical parameters through combing particle swarm optimization, support vector machine and numerical analysis. Support vector machine is used to build
the non-linear relationship between geomechanical parameters and displacements, and has proved excellent performance
of non-linear representation based on little samples. Particle swarm optimization is used to improve the generalization performance in searching the support vector machine model, and has proved powerful global optimal performance. Overall, the
proposed approach improves the efciency and precision of back analysis, and makes it possible to be applied to more complex engineering problems.
Acknowledgement
The nancial support from Program for New Century Excellent Talents in University (NCET) and Doctoral Fund of Henan
Polytechnic University (No. 648197) are greatly acknowledged.
References
[1] H.A.D. Kirsten, Determination of rock mass elastic modulus by back analysis of deformation measurement, in: Proceedings of the Symposium on
Exploration for Rock Engineering, Johannesburg, 1976.
[2] G. Gioda, L. Jurina, Numerical identication of soil structure interaction pressures, Int. J. Numer. Anal. Meth. Geomech. 5 (1981) 3356.
[3] S. Sakurai, K. Takeuchi, Back analysis of measured displacements of tunnels, Rock Mech. Rock Eng. 16 (1983) 173180.
[4] S. Sakurai, N. Dees Wasmongkol, M. Shinji, Back analysis for determining material characteristics in cut slopes, in: Proceedings of the International
Symposium on ECRF, Beijing, vol. 11, 1986, pp. 770776.
[5] S. Sakurai, Interpretation of the results of displacement measurements in cut slopes, in: Proceedings of the Second International Symposium on
FMGM87, Kobe, vol. 2, 1987, pp. 528540.
[6] Z.L. Feng, R.W. Lewis, Optimal estimation of in-situ ground stress from displacement measurements, Int. J. Numer. Anal. Meth. Geomech. 11 (1987)
397408.
[7] B. Pichler, R. Lackner, H.A. Mang, Back analysis of model parameters in geotechnical engineering by means of soft computing, Int. J. Numer. Meth. Eng.
57 (14) (2003) 19431978.
[8] T. Okabe, Inverse of drilling-induced tensile fracture data obtained from a single inclined borehole, Int. J. Rock Mech. Mining Sci. 35 (6) (1998) 747758.
[9] W.G.Y. William, Aquifer parameter identication with optimum dimension in parameterization, Water Resour. Res. 17 (3) (1981) 664672.
[10] A. Cividini, Parameter estimation of a static geotechnical model using a Bayes approach, Int. J. Rock Mech. Mining Sci. 20 (5) (1983) 215226.
[11] Xia-Ting Feng, Study on genetic-neural network of displacement back analysis, Chin. J. Rock Mech. Eng. 18 (5) (1999) 1 (in Chinese).
[12] Xia-Ting Feng, Introduce of Intelligent Rock Mechanics, Science, Beijing, 2000 (in Chinese).
[13] Xia-Ting Feng, Zhang Zhiqiang, Sheng Qian, Estimating geomechanical parameters relating to the Three Gorges Project permanent shiplock using an
intelligent displacement back analysis method, Int. J. Rock Mech. Mining Sci. 37 (2000) 10391054.
[14] J.H. Deng, C.F. Lee, Displacement back analysis for a steep slope at the Three Gorges Project site, Int. J. Rock Mech. Mining Sci. 38 (2) (2001) 259268.
[15] Yuzhen Yu, Bingyin Zhang, Huina Yuan, An intelligent displacement back-analysis method for earth-rockll dams, Comput. Geotech. 34 (6) (2007)
423434.
[16] Xia-Ting Feng, Hongbo Zhao, Shaojun Li, A new displacement back analysis to identify mechanical geo-material parameters based on hybrid intelligent
methodology, Int. J. Numer. Anal. Meth. Geomech. 28 (2004) 11411165.
[17] Xia-Ting Feng, J.A. Hudson, Shaojun Li, et al, Integrated intelligent methodology for large-scale landslide prevention, Int. J. Rock Mech. Mining Sci. 41
(3) (2004) 750755.
[18] Hongbo Zhao, Xiating Feng, Study on genetic-support vector machine of displacement back analysis, Chin. J. Rock Mech. Eng. 22 (10) (2003) 631635
(in Chinese).
[19] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer, New York, 1995.
[20] J. Smola Alex, Schoelkopf Bernhard, A Tutorial on Support Vector Regression, NeuroCOLT2 Technical Report Series NC2-TR-1998030, 1998.
[21] J. Kennedy, R.C. Eberhart, A discrete binary version of the particle swarm algorithm, in: Proceedings of the 1997 Conference on System, Man, and
Cybernetics, IEEE Service Center, Piscataway, NJ, 1997, pp. 41044109.

S-ar putea să vă placă și