You are on page 1of 7

International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159

Contents lists available at ScienceDirect

International Journal of
Rock Mechanics & Mining Sciences
journal homepage: www.elsevier.com/locate/ijrmms

Artificial neural networks as a basis for new generation of rock failure criteria
Hosein Rafiai n, Ahmad Jafari
School of Mining Engineering, University of Tehran, Tehran 1439957131, Iran

a r t i c l e i n f o abstract

Article history: Conventional methods for prediction of rock strength are based on using classical failure criteria. In this
Received 25 October 2010 study, artificial neural networks were regarded as new tools for considering the strength of intact rock
Accepted 9 June 2011 in a wide range of loading condition from uniaxial tension to triaxial compression. A comprehensive
Available online 25 June 2011
data set of the values of major and minor principal stresses at failure from 1638 laboratory tests on
Keywords: seven rock types was collected. For each rock type, data were randomly divided into two subsets,
Rock strength training and test sets. Neural networks were trained using training sets to predict the value of major
Artificial neural network principal stress at failure from uniaxial compressive stress and minor principal stress. Small
Bieniawski–Yudhbir criterion architecture and regularization method were adopted to avoid over-fitting problems. The same training
Hoek–Brown criterion
sets were used in determining the constants of two popular empirical failure criteria, namely
Ductile failure
Bieniawski–Yudhbir and Hoek–Brown. Then, the test sets were used to examine the accuracy and
generalization of the models in predicting the strength in new situations. Comparison of the results of
the neural network models with those of the empirical criteria showed that the former approach
always lead to less root mean squared error and higher coefficient of determination. On average, for
different rock types, using ANN models led to about 30% decrease in prediction error relative to best
empirical models. These models also showed better flexibility in the prediction of major principal stress
at failure in both brittle and ductile failure regimes.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction increases if these criteria were used for irrelevant rock types
and ranges of stress.
Strength of rock is one of the most important factors influen- In recent years, Artificial Neural Networks (ANNs) have been
cing the design of rock structures, and it has led to many years of frequently used for function approximation in different field of
research in the field of rock mechanics. Several empirical rock science, including geotechnical engineering [15–32]. The attractive-
failure criteria have been developed by many researchers such as ness of ANNs comes from the information processing characteristics
Fairhust [1], Hobbs [2], Murrel [3], Franklin [4], Bieniawski [5], of the system, such as non-linearity, high parallelism, fault tolerance,
Hoek and Brown [6], Ramamurthy et al. [7], Johnston [8], Sheorey learning, and generalization capability [33]. Unlike the classical
et al. [9], and Yoshida et al. [10]. In these criteria, major principal regression methods in which a certain form for the approximation
stress at failure s1 is a function of uniaxial compressive function must be presumed, ANNs do not require any presumption
strength (sc) and minor principal stress (s3) with one or more in this regard. Hence, they give better flexibility and higher accuracy
constant(s) to be determined from regression analysis of experi- in function approximation problems
mental data. Thus, if these constants were estimated, strength of On the subject of rock strength, Meulenkamp and Grima [34]
rock at different values of s3 could be predicted by knowledge of used ANNs to estimate the uniaxial compressive strength for a
sc. Approximate values of these constants for different rock types few different rock types. Singh et al. [35] reported the application
are proposed by several authors [5,6,11–13]. However, empirical of ANNs in predicting both uniaxial compressive and tensile
failure criteria are not usually accurate in all ranges of stress strength of schistose rocks. Canakci and Pala [36] trained ANNs
and for all rock types [14], because they are based on specific to predict tensile strength of basalt. Tiryaki [37] developed an
presumed forms of functions that show good agreement with test ANN-based model for estimating uniaxial compressive strength
results in those particular conditions; the prediction error and modulus of elasticity of some rock types. Another use of
ANNs is described by Zorlu et al. [38] for predicting the uniaxial
compressive strength of sandstones. However, it can be observed
that all of these works have focused on the strength of rock only
n
Corresponding author. Tel.: þ989132230377, fax: þ982188008838. in uniaxial (tensile or compressive) loading condition. In addition,
E-mail addresses: hrafiai@ut.ac.ir, hosein_868@yahoo.com (H. Rafiai). they have usually considered the strength of certain rock types.

1365-1609/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijrmms.2011.06.001
1154 H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159

In this study, ANNs were regarded as a basis for new genera- where pi and wi are the values of ith input and weight, b is the
tion of failure criteria to consider the strength of seven different bias, and f is the activation function of the neuron. Different types
rock types in a wide range of loading condition from uniaxial of activation functions such as, step function, linear function, and
tension to triaxial compression. The networks were planned to sigmoid function can be used in ANNs.
predict the value of s1 from the values of sc and s3 for each rock Two or more neurons can be combined in a layer, and a partic-
type. Results of many laboratory tests were utilized to train the ular network could contain one or more such layers. The layers of a
networks. Subsequently, the accuracy of the trained ANNs and multilayer network play different roles. A layer that produces the
the two well-known failure criteria proposed by Bieniawski and network outputs is called output layer. All other layers are called
Yudhbir [5,11] and Hoek and Brown [6] in prediction of major hidden layers [33]. Based on the configuration of the connections
principal stress at failure were compared. It was observed that between neurons of different layers, various types of ANNs can be
ANN-based models with their flexible formulation can give more obtained. Feed-Forward Neural Networks (FFNNs) are especial type
accurate results. of ANNs in which inputs are received and simply forwarded through
all the next layers to obtain the outputs. There is no feed-back
connection to previous layers in these networks [43].
2. The data set In the learning process, the network is presented with a pair of
training data including input values and corresponding target
In this study, it tried to collect all of available and reliable rock values. The network computes its own outputs using its initial
strength data from worldwide sources [14,39]. Sufficient numbers weights and biases. Then, weights and biases are adjusted based
of data were found for seven rock types, namely sandstone, on a comparison of the output values and the target values, until
quartzite, marble, limestone, granite, dolomite, and coal. The data the network outputs match the targets [42].
include the values of sc, s3, and s1 recorded in uniaxial tension, The typical performance function used as an error index in the
uniaxial compression, and triaxial compression tests, respectively. training phase is the Sum of Squared Errors (SSE). Several algo-
For the results of uniaxial tension tests, the values of s3 were rithms for minimization of the performance function have been
negative and the values of s1 were zero. The results of uniaxial developed. Among them, gradient descent optimization has led to
compression tests were used with zero values of s3 and positive one of the most popular learning algorithms, namely back-
values for s1 while for the triaxial compressive tests, both the propagation. In this algorithm learning iterations consist of two
values of s3 and s1 were positive. phases: feed-forward pass, which simply calculates the output
The brittle failure condition proposed by Mogi [40] (s1 43.4s3) values of the ANN for training data, and backward propagation,
was applied to the strength data. However, coal data that contain which propagates an error signal back from the last layer toward
considerable number of results with s1 o3.4s3 were divided into the first layer. Weights and biases are adjusted as functions of the
two subsets; one consists of results that satisfy the brittle failure back-propagated error signal [43].
condition, and another set that includes the results of both brittle Generalization is a very important aspect of ANN learning. Since
and ductile failure. It was done in order to examine the accuracy and it is a measure of how well the network interpolates to points not
flexibility of different models in prediction of strength in both brittle used during training, the ultimate objective of ANN learning is to
and ductile failure. Number of data and range of values of sc, s3, and produce a learner with low generalization error. Over-fitting of a
s1 for each rock type are shown in Table 1. training set occurs when the ANN memorizes the training patterns,
and consequently loses the ability to generalize. ANNs that over-fit
cannot predict correct output for data patterns not seen during
3. Artificial neural networks training. Over-fitting problem is expected when the ANN architec-
ture is too large, i.e. the network has too many hidden units and
ANNs may be defined as structures comprised of densely irrelevant input units [43]. Optimizing the network architecture is a
interconnected adaptive simple processing elements that are remedy for over-fitting. Unfortunately, it is difficult to know before-
capable of performing massively parallel computations for data hand how large a network should be for a specific application [42].
processing and knowledge representation [41]. These elements There are two other methods for improving generalization
are inspired by biological nervous systems and are called neurons. during training the network [42]. Early stopping is the common
Each neuron with n inputs calculates its output a as [42] technique in which the available data are divided into three subsets
! of training, validation, and test. The method is based on monitoring
Xn
a¼f w i pi þ b ð1Þ the error on the validation set and stopping the training iterations
i¼1 when the validation error is in its minimum level.
Another method for improving generalization is called regu-
Table 1 larization. It is possible to improve generalization if the objective
Properties of the strength data set including the results of uniaxial tension, function is modified by introducing a new term that consists of
uniaxial compression and triaxial compression tests. the sum of squares of the network weights and biases [44]:
Rock type Number of rc (MPa) r3 (MPa) r1 (MPa) F ¼ b SSE þ a SSW ð2Þ
data
Min Max Min Max Min Max where F is the modified objective function, b and a are objective
function parameters, and SSW is the sum of squares of the net-
Sandstone 724 24.3 203.8  14.0 100.0 0.0 565.0
work weights and biases. It is desirable to determine the optimal
Quartzite 53 103.4 303.3  18.6 65.3 0.0 746.9
Marble 168 27.6 137.8  7.5 88.2 0.0 394.5
objective function parameters in an automated fashion. One
Limestone 168 45.0 214.6  9.4 50.0 0.0 432.7 approach to this process is the Bayesian framework [45].
Granite 119 81.6 256.1  12.9 60.0 0.0 771.0 When the data set is small and the network is training for
Dolomite 44 76.2 229.2 0.0 107.0 76.2 735.0 function approximation, Bayesian regularization provides better
Coala (brittle) 318 5.3 51.4  3.4 60.0 0.0 214.3
generalization performance than early stopping. This is because
Coala (brittle and 362 5.3 51.4  3.4 71.4 0.0 228.0
ductile) Bayesian regularization does not require that a validation data set
be separate from the training set. Hence more data can be used
a
Maximum principal stress is applied perpendicular to the bedding plane. for training [42].
H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159 1155

4. Prediction of major principal stress at failure using ANNs of weights and biases) will, on average, gives the best generalization
performance [47]. Determining appropriate number of neurons to
In this study, ANNs were used for the prediction of major include in hidden layer is not an exact science [48]. It is difficult to
principal stress at failure from uniaxial compressive strength and find the optimum network size before training and evaluating the
minor principal stress. FFNNs were adopted due to their appro- results of network. However, some heuristic relations are proposed
priateness for pattern matching problems like strength prediction by researchers for determining the network size. Table 2 sum-
[35]. Pattern matching is basically an input/output mapping marizes some of these heuristics for number of neurons to be used
problem, and the quality of mapping is evaluated based on the in the single hidden layer of a FFNN.
difference between actual and calculated output. In this table, ni and no are the number of input and output
FFNNs can have one or more hidden layers. However, it has been neurons, respectively, nt is the number of training patterns, k is
proved that FFNNs with monotonically increasing differentiable the noise factor (), and y is the over determination constant (y).
functions can approximate any continuous function with one hidden Typical values of k and y frequently used in literature are 4 and
layer, provided that the hidden layer has enough hidden neurons 1.25, respectively [31,34,37]. The same values were adopted in
[46]. Two-layer FFNN, with a hyperbolic tangent sigmoid function in this study. It can be observed that the least number of hidden
the hidden layer and a linear transfer function in the output layer, is neurons proposed by heuristics is three.
a useful structure for function approximation problems [42]. Hence, However, the most popular approach to find the optimal number
this configuration is used for ANNs in this study. of hidden neurons is by trial and error with one of the above rules as
Optimization of network size is a very important task to prevent starting point [52]. Results of many trial and errors also revealed
over-fitting. If several networks fit the training set equally well, then that for all rock types, no significant improvement could be made in
the simplest network (i.e. the network that has the smallest number network performance using more than three neurons in the hidden
layer. Final architecture of the ANNs is shown in Fig. 1.
Levenberg–Marquardt back-propagation algorithm combined
Table 2
with Bayesian regularization was used for training of networks
The heuristics proposed for the number of neuron to be used in hidden layer.
because of fast training and high generalization [42]. It has proved
Heuristic Calculated number of Reference to avoid over-fitting even if the number of hidden neurons were
neurons for this study higher than the optimum value. In other word, implementation of
this algorithm can simply eliminate the guesswork required in
r2ni þ 1 r5 Hecht-Nielsen [41]
nt =yno 6a Seibi and Al-Alawi [48] determining the optimum network size [42]. The algorithm
ni þ no þ 1
generally works best when the network inputs and targets are
2ni 4 Kannellopoulas and Wilkinson [49]
3ni 6 Hush [50] scaled so that they fall approximately in the range [ 1,1] [42].
nt
kðni þ no Þ
3a Aldrich and Reuter [51] The following equation was used for linear scaling of input and
target values to this range
a
The numbers are calculated for dolomite which has the minimum number of
training data with nt ¼ 35.
xs ¼ 2ðxxmin Þ=ðxmax xmin Þ1 ð3Þ
s
where x is the scaled value of parameter with maximum and
minimum values of xmax and xmin, respectively.
Hidden Output In literature, it is proposed to use between 20 and 30% of available
Inputs Output
Layer Layer
data for testing [53–55]. In this study, the data of each rock type were
randomly divided into two sets of training and test set, which include
80% and 20% of the data, respectively. The initial values of weights
and biases adopted in this work were random numbers in the range
[ 1,1], which are frequently used and include the ranges recom-
mended by researchers [25,31,47,54]. MATLAB software was imple-
mented for random selection of training sets, initializing the values of
weights and biases, and training the ANNs. Weights and biases of the
trained networks are reported in Table 3. These values along with the
extremum values of parameters presented in Table 1 can be used to
Fig. 1. Architecture of ANNs used for prediction of major principal stress at failure. calculate the outputs of the ANNs for a given set of inputs.

Table 3
Weights and biases of ANNs for different rock types (wij,k is the weight between jth neuron of ith layer and of k0 th neuron of previous layer, bij is the bias of jth neuron
of ith layer).

Weight/bias Sandstone Quartzite Marble Limestone Granite Dolomite Coal (brittle) Coal (brittle and ductile)

w11,1  0.103 0.990 0.169  0.428 0.855  0.704  0.691  0.419


w11,2  2.700 0.448  0.620 1.208 0.855 0.267  0.032  0.057
w12,1  2.297 0.270  0.122  0.248  0.176  0.333 0.193 0.071
w12,2  0.427 0.862  2.453  0.045  1.200  0.726  0.629  4.373
w13,1  0.586 0.715 0.094 0.658  0.027 0.126  0.058 0.418
w13,2  0.527 0.168 1.057 0.409  0.347  1.555 3.546  0.556
w21,1  0.662 1.055 0.846 0.657 0.855  0.480  0.607  1.714
w21,2 0.409 1.325  1.497  0.327  1.171  0.866  1.230  1.317
w21,3  1.624  0.384 1.157 0.859 0.333  1.065 1.370  1.124
b11  2.649  1.310  0.613 1.164  1.549  0.020  0.418  0.458
b12 0.195 0.795  2.948 0.048  0.984 0.917 0.347  4.943
b13 0.159 0.027  0.183 0.336 0.163  1.872 4.056  0.086
b21  0.362  0.001  0.669  0.633  0.104  0.388  1.021  1.936
1156 H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159

It was observed that the predicted values of s1 using ANNs of intact rock. However, they are usually accurate in a certain
were in excellent agreement with target values included in the range of stresses and are not applicable for all rock types [14].
training data set. However, it is not necessarily an indication of Empirical criteria proposed by Bieniawski and Yudhbir [5,11]
high performance of the networks, because those data have been and Hoek and Brown [6] are very popular due to their good agree-
used in the training phase. Hence, it is possible that the networks ment with test results. Hence, in this study, they were selected as
have memorized the data while their accuracy for new data was qualified representatives of rock failure criteria to compare their
not satisfactory. Therefore, the trained ANNs must be validated results with those of the ANNs. The Bieniawski–Yudhbir (B–Y)
by comparison of the output and target values for test sets. criterion is expressed as
The results of this comparison are discussed in Section 6.  k
s1 s3
¼ 1þA ð4Þ
sc sc
5. Prediction of major principal at failure using empirical where A is a constant that depends on rock type and k ¼0.65 for
failure criteria all rock types. This criterion is used for both brittle and ductile
failure [11].
A failure criterion is an algebraic expression of the stress The Hoek–Brown (H–B) criterion is given by
condition under which a material fails. An empirical criterion sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  ffi
may be adopted having no theoretical basis, selected only to s1 s3
¼ 1þ m þ1 ð5Þ
meet practical requirements of accurate strength prediction and sc sc
simplicity of use. Several empirical failure criteria have been where m is a constant that depends on rock type. This criterion is
proposed over the years in an attempt to estimate the strength only applicable for brittle failure [6].

Table 4 Table 5
Proposed and calculated constants of the B–Y and H–B criteria. Coefficient of determination of different predictive models for test sets.

Rock type A[10,12]


p Ac mp [13,12]
mc Rock type B–Y p B–Y c H–Bp H–B c ANN

Sandstone 4 3.7 17 14.9 Sandstone 0.86 0.87 0.90 0.91 0.94


Quartzite 4 3.4 20 15.7 Quartzite 0.88 0.89 0.91 0.91 0.98
Marble – 2.6 9 8.3 Marble – 0.91 0.94 0.95 0.95
Limestone 2 3.0 10 11.3 Limestone 0.80 0.93 0.93 0.93 0.95
Granite 5 5.0 32 29.0 Granite 0.95 0.95 0.91 0.94 0.96
Dolomite – 3.4 9 13.2 Dolomite – 0.94 0.70 0.89 0.96
Coal (brittle) 3 3.5 10 11.5 Coal (brittle) 0.79 0.86 0.86 0.88 0.96
Coal (brittle and ductile) 3 3.4 – – Coal (brittle and ductile) 0.79 0.82 – – 0.96

50 60 25 50

50 20 40
40

40
30 15 30
30
20 10 20
20
5 10
H-Bp

10
B-Yc

ANN
H-Bc
ANN

ANN

ANN
B-Yp

H-Bc
H-Bp
B-Yp

B-Yp
H-Bc

H-Bp

B-Yc

10
B-Yc
H-Bp

B-Yc

H-Bc

-
0 0 0 0
Sandstone Quartzite Marble Limestone

50 100 20 25

40 80 20
15

30 60 15
10
20 40 10

5 5
10 20
ANN
H-Bp

ANN
B-Yc

B-Yp
H-Bc

B-Yc
H-Bp
ANN

ANN
B-Yp

H-Bp
H-Bc

H-Bc
B-Yp
B-Yc

B-Yc

- - -
0 0
0 0
Granite Dolomite Coal (brittle) Coal (brittle & ductile)

Fig. 2. RMSE values of different predictive models for test sets in MPa.
H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159 1157

Approximate values of the constants, A and m, are proposed by data base, regression analysis was performed on the same data
researchers for different rock types [5,6,11–13]. In addition, in sets that were used for training the ANNs. Table 4 shows the
order to find more accurate values for A and m for the utilized proposed and calculated values of the constants for each rock type

600 700
Sandstone Quartzite
500 600
Predicted value of σ1 (MPa) r2 = 0.94 r2 = 0.98

Predicted value of σ1(MPa)


500
400
400
300
300
200
200
100
100
0
0
0 100 200 300 400 500 600 0 100 200 300 400 500 600 700
-100
Target value of σ1 (MPa) Target value of σ1 (MPa)

400 450
Marble Limestone
350 400
r2 = 0.95
Predicted value of σ1 (MPa)
r2 = 0.95 350
Predicted value of σ1 (MPa)

300
300
250
250
200
200
150 150
100 100
50 50
0 0
0 100 200 300 400
-50 0 100 200 300 400
Target value of σ1 (MPa)
Target value of σ1 (MPa)

700
Granite Dolomite
800 600
Predicted value of σ1 (MPa)

r2 = 0.96
Predicted value of σ1 (MPa)

r2 = 0.96
500
600
400
400 300

200
200
100

0 0
0 200 400 600 800 0 100 200 300 400 500 600 700
Target value of σ1 (MPa) Target value of σ1 (MPa)

250 250
Coal (brittle) Coal (brittle & ductile)
200 r² = 0.96 r2 = 0.96
200
Predicted value of σ1 (MPa)

Predicted value of σ1 (MPa)

150
150

100
100

50
50
0
0 50 100 150 200 250 0
-50 0 50 100 150 200 250
Target value of σ1 (MPa) Target value of σ1 (MPa)

Fig. 3. Predicted values of major principal stress at failure using ANNs vs. target values included in test sets.
1158 H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159

with subscripts ‘‘p’’ and ‘‘c’’, respectively. Two spaces in the Ap Table 6
column are left blank, because no proposed value of A for marble Average values of RMSE and r2 of different predictive models.
and dolomite could be found in literature. In addition, m values
Index parameter B–Y p B–Y c H–B p H–B c ANN
corresponding to brittle and ductile failure modes of coal are not
given, because Hoek–Brown criterion is not applicable for ductile RMSE 38.9 33.1 39.7 32.7 23.7
failure. Hence, these four cases are not considered in further r2 0.86 0.90 0.88 0.92 0.96
analyses. In the following section, results of the empirical failure
criteria and ANNs in prediction of for test sets are compared.
The results show that the trained ANNs can predict the major
principal stress at failure more accurately than empirical failure
6. Results and discussion criteria. ANN models are especially more reliable for the cases
with higher number of data in the utilized data set, e.g. sandstone
Training sets were used in development of predictive models and coal.
based on ANNs and empirical failure criteria. The generalization
of the models can be examined by comparison of their predictions
with corresponding target values included in test data sets. 7. Conclusions
In order to evaluate the accuracy of these models, two indices
of root mean squared error (RMSE) and coefficient of determina- In this study, ANNs were utilized as failure criteria to predict
tion (r2) were used [56]. the value of s1 from the values of sc and s3 over a wide range of
RMSE values for different prediction models are shown in loading condition from uniaxial tension to triaxial compression.
Fig. 2. The results of empirical failure criteria using proposed A comprehensive data set including the results of 1638 laboratory
and calculated constants are presented with subscripts ‘‘p’’ and tests on seven rock types was collected. For each rock type, data
‘‘c’’, respectively. For all rock types, ANN models have lower RMSE were randomly divided into two subsets, training and test sets.
value than the lowest values of empirical models. This contrast is Training sets were used in training ANNs with three neurons in
especially considerable for quartzite and coal data where ANN their single hidden layer. Small architecture and regularization
models give RMSE values of about half of the empirical models method were adopted to avoid over-fitting problems. The same
(Fig. 2b, g, and h). Another advantage of ANNs is revealed in com- training sets were used in determining the constants of
parison of RMSE values of models for prediction of coal strength Bieniawski–Yudhbir and Hoek–Brown failure criteria. Then, test
in only brittle regime and both brittle and ductile regimes. By sets were used to examine the accuracy of these models in the
adding ductile failure data to brittle failure data, RMSE of the ANN prediction of s1 in new situations.
model increased by only 2 MPa while the best empirical criterion Although the same information were used in development of the
(B–Yc with lowest RMSE value) showed more than two times of predictive models, ANN models for all rock types showed lower
this increase in error. This advantage is because of flexibility of values of prediction error. The difference was especially considerable
ANNs in approximation of functions over the ranges that show for quartzite and coal, where error of ANNs was about half of the
different behavior. On the other hand, empirical failure criteria failure criteria. On average, for different rock types, using ANN
presume a specific function form and could be used only over the models led to about 30% decrease in prediction error relative to best
ranges that this form is representative; prediction error increases empirical models. In addition, ANNs showed higher flexibility for
by exceeding this range. prediction of strength in ductile failure. For coal, adding the ductile
Table 5 shows values of coefficient of determination of the failure data to brittle failure data led to only 2 MPa increase in error
prediction models. ANN models are also more accurate in terms of for the ANN model while the best empirical model showed more
the determination coefficients. The flexibility of ANNs is apparent in than two times of this increase in error.
considering the variation of r2 values for coal sets. Adding the data ANNs are powerful tools in predicting the strength of different
for ductile failure mode to brittle failure data of coal led to 0.04 rock types with high accuracy and flexibility. However, it is
decreases in r2 for the best empirical criterion (B–Yc with highest r2 possible to improve their performance using more governing
value) while this value remained unchanged for ANN model. input parameters and more comprehensive data set. Extending
Fig. 3 shows the relation between target values of s1 and those the same study for rock mass can be the subject of future works.
predicted by ANNs for test sets of different rock types. It can be Furthermore, ANN-based failure criteria may be implemented in
observed that test data have a good distribution over the whole numerical codes to analyze the stability of different rock struc-
range of strength. Data points show a linear trend, and are sym- tures provided that the ANNs were trained using rock mass
metrically scattered about the line of best fit. The highest degree strength data.
of scatter is related to sandstone and limestone (Fig. 3a and d),
which is expected, because they are general names for a wide
range of geo-materials with different mechanical properties [57]. Acknowledgements
Negative predicted values of s1 can be observed in Fig. 3a, c, and
g. These values are not physically meaningful and should be The authors would like to thank Evert Hoek for his helps and
regarded as prediction error, because the minimum values of s1 guidance in finding the required strength data set. Valuable
corresponding to the results of uniaxial tension tests are zero. comments of Candan Gokceoglu and Bulent Tiryaki on the earlier
However, this problem has occurred just in three cases out of 328 draft of the paper are also warmly acknowledged.
cases included in the test data sets.
Average RMSE and r2 values of the predictive models for References
different rock types, over their ranges of applicability (i.e. brittle
failure for H–B and both brittle and ductile failure for B–Y and [1] Fairhurst C. On the validity of the ‘‘Brazilian’’ test for brittle materials. Int J
ANN) is shown in Table 6. On average, RMSE value in ANN models Rock Mech Min Sci 1964;1:515–46.
for prediction of strength of different rock types is 28–40% lower [2] Hobbs DW. The strength and stress–strain characteristics of coal in triaxial
compression. J Geol 1964;72:214–31.
than the empirical models. Average r2 value in ANN models is [3] Murrel SAF. The effect of triaxial stress systems on the strength of rock at
0.04–0.1 greater than those of empirical models. atmospheric temperature. Int J Rock Mech Min Sci 1965;3:11–43.
H. Rafiai, A. Jafari / International Journal of Rock Mechanics & Mining Sciences 48 (2011) 1153–1159 1159

[4] Franklin JA. Triaxial strength of rock material. Rock Mech 1971;3:86–98. [30] Gokceoglu C, Zorlu K, Ceryan S, Nefeslioglu HA. A comparative study on
[5] Bieniawski ZT. Estimating the strength of rock materials. J S Afr Inst Min indirect determination of degree of weathering of granites from some
Metall 1974;74:312–20. physical and strength parameters by two soft computing techniques. Mater
[6] Hoek E, Brown ET. Underground excavations in rock. Inst Min Metall, London Charact 2009;60:1317–27.
1980. [31] Majdi A, Beiki M. Evolving neural network using a genetic algorithm for
[7] Ramamurthy T, Rao GV, Rao K. A strength criterion for rocks. In: Proceedings predicting the deformation modulus of rock masses. Int J Rock Mech Min Sci
of the Indian Geotechnical Conference, Roorkee, India; 1985. pp. 59–64. 2010;47:246–53.
[8] Johnston JW. Strength of intact geomechanical materials. J Geotech Eng [32] Tiryaki B, Boland JN, Li XS. Empirical models to predict mean cutting forces
1985;111:730–49. on point-attack pick cutters. Int J Rock Mech Min Sci 2010;47:858–64.
[9] Sheorey PR, Biswas AK, Choubey VD. An empirical failure criterion for rocks [33] Jain AK, Mao J, Mohiddin KM. Artificial neural networks: a tutorial. Comput-
and jointed rock masses. Eng Geol 1989;26:141–59. ing 1996:31–44.
[10] Yoshida N, Morgenstern NR, Chan DH. Failure criterion for stiff soils and [34] Meulenkamp F, Grima MA. Application of neural networks for the prediction
rocks exhibiting softening. Can Geotech J 1990;27(2):195–202. of the hardness of the unconfined compressive strength from Equotip
[11] Yudhbir RK, Lemanza W, Prinzl F. An empirical failure criterion for rock hardness. Int J Rock Mech Min Sci 1999;36:29–39.
masses. In: Proceedings of the Fifth International Congress on Rock [35] Singh VK, Singh D, Singh TN. Prediction of strength properties of some
Mechanics, vol. 1, Melbourne; 1983, pp. B1–8. schistose rocks from petrographic properties using artificial neural networks.
[12] Hossaini SMF. A comparison between Hoek–Brown and Bieniawski criteria Int J Rock Mech Min Sci 2001;38:269–84.
for coal and rocks. In: Proceedings of the Coal Operators Conference, [36] Canakci H, Pala M. Tensile strength of basalt from a neural network. Eng Geol
Wollongong, Australia, 1998, pp. 213–23. 2007;94:10–8.
[13] Marinos P, Hoek E. Estimating the geotechanical properties of heterogeneous [37] Tiryaki B. Predicting intact rock strength for mechanical excavation using
rock masses such as flysh. Bull Eng Geol Env 2001;60:85–92. multivariate statistics, artificial neural networks, and regression trees. Eng
[14] Sheorey PR. Empirical rock failure criteria. Rotterdam: Balkema; 1997. Geol 2008;99:51–60.
[15] Zhang Q, Song J, Nie X. Application of neural network models to rock [38] Zorlu K, Gokceoglu C, Ocakoglu F, Nefeslioglu HA, Acikalin S. Prediction of
mechanics and rock engineering. Int J Rock Mech Min Sci Geomech Abstr uniaxial compressive strength of sandstones using petrography-based models.
1991;28:535–40. Eng Geol 2008;96:141–58.
[16] Lee C, Sterling R. Identifying probable failure modes for underground open- [39] Rocscience Inc. RocData (RocProp database). Toronto, Canada, 2007.
ings using a neural network. Int J Rock Mech Min Sci Geomech Abstr [40] Mogi M. Pressure dependence of rock strength and transition from brittle
1992;29:49–67. fracture to ductile flow. Bull Earthq Res Inst Jpn 1966;44:215–32.
[17] Chan WT, Chow YK, Liu LF. Neural network: an alternative to pile driving [41] Hecht-Nielsen R. Kolmogorov0 s mapping neural network existence theorem.
formulas. Comput Geotech 1995;17:135–56. In: Proceedings of the First IEEE International Conference on Neural Net-
[18] Kanj M, Roegiers JC. A neural approach to sand production prediction in works, San Diego; 1987. pp. 11–4.
wells. Int J Rock Mech Min Sci 1998;35:489. [42] Demuth H, Beal M, Hagan M. Neural network toolbox 5 user0 s guide of the
[19] Huang Y, Wfinstedt S. The introduction of neural network system and its MATLAB software. Natick, Mass; 2007.
applications in rock engineering. Eng Geol 1998;49:253–60. [43] Engelbrecht AP. Computational intelligence. Chichester: Wiley; 2007.
[20] Yang Y, Zhang Q. The application of neural networks to rock engineering [44] Foresee FD, Hagan MT. Gauss–Newton approximation to Bayesian regular-
systems (RES). Int J Rock Mech Min Sci 1998;35:727–45. ization. In: Proceedings of the International Joint Conference on Neural
[21] Lee SJ, Lee SR, Kim YS. An approach to estimate unsaturated shear strength Networks; 1997. pp. 1930–5.
using artificial neural network and hyperbolic formulation. Comput Geotech [45] MacKay DJC. Bayesian interpolation. Neural Comput 1992;4:415–47.
2003;30:489–503. [46] Hornik K. Multilayer feed-forward networks are universal approximators.
[22] Neaupane KM, Achet SH. Use of backpropagation neural network for land- Neural Networks 1997;2:359–66.
slide monitoring: a case study in the higher Himalaya. Eng Geol 2004;74: [47] Sietsma J, Dow RJF. Creating artificial neural networks that generalize. Neural
213–26. Networks 1991;4:67–79.
[23] Wang HB, Xu WY, Xu RC. Slope stability evaluation using back propagation [48] Seibi A, Al-Alawi SM. Prediction of fracture toughness using artificial neural
neural networks. Eng Geol 2005;80:302–15. networks (ANNs). Eng Fract Mech 1997;56:311–9.
[24] Banimahd M, Yasrobi SS, Woodward PK. Artificial neural network for stress– [49] Kanellopoulas I, Wilkinson GG. Strategies and best practice for neural net-
strain behavior of sandy soils: knowledge based verification. Comput work image classification. Int J Remote Sensing 1997;18:711–25.
Geotech 2005;32:377–86. [50] Hush DR. Classification with neural networks: a performance analysis.
[25] Sonmez H, Gokceoglu C, Nefeslioglu HA, Kayabasi A. Estimation of rock In: Proceedings of the IEEE International Conference on Systems Engineering,
modulus: for intact rocks with an artificial neural network and for rock Dayton, Ohio; 1989. pp. 277–80.
masses with a new empirical equation. Int J Rock Mech Min Sci 2006;43: [51] Aldrich C, Reuter MA. The application of neural nets in the metallurgical
224–35. industry. Miner Eng 1994;7:793–809.
[26] Moosavi M, Yazdanpanah MJ, Doostmohammadi R. Modeling the cyclic [52] Basheera IA, Hajmeerb. Artificial neural networks: fundamentals, computing,
swelling pressure of mudrock using artificial neural networks. Eng Geol design, and Application. J Microbiol Methods 2000;43:3–31.
2006;87:178–94. [53] Swingler K. Applying neural networks: a practical guide. New York: Aca-
[27] Wang ZL, Li YC, Shen RF. Correction of soil parameters in calculation of demic Press; 1996.
embankment settlement using a BP network back-analysis model. Eng Geol [54] Looney CG. Advances in feed-forward neural networks: demystifying knowl-
2007;91:168–77. edge acquiring black boxes. IEEE Trans Knowl Data Eng 1996;8(2):211–26.
[28] Monjezi M, Dehghani H. Evaluation of effect of blasting pattern parameters [55] Nelson M, Illingworth WT. A practical guide to neural nets. Reading, Mass:
on back break using neural networks. Int J Rock Mech Min Sci 2008;45: Addison-Wesley; 1990.
1446–53. [56] Everitt BS. The Cambridge dictionary of statistics. New York: Cambridge
[29] Yagiz S, Gokceoglu C, Sezer E, Iplikci S. Application of two non-linear University Press; 2006.
prediction tools to the estimation of tunnel boring machine performance. [57] Brady BGH, Brown ET. Rock mechanics for underground mining. London:
Eng Appl Artif Intell 2009;22:818–24. Chapman & Hall; 1992.

You might also like