You are on page 1of 9

Neural Comput & Applic

DOI 10.1007/s00521-011-0631-5

ORIGINAL ARTICLE

Evaluation of flyrock phenomenon due to blasting operation


by support vector machine
Hasel Amini • Raoof Gholami • Masoud Monjezi •

Seyed Rahman Torabi • Jamal Zadhesh

Received: 27 February 2011 / Accepted: 3 May 2011


 Springer-Verlag London Limited 2011

Abstract Flyrock is an undesirable phenomenon in the are used in the blasting operation. Only a fracture of the
blasting operation of open pit mines. Flyrock danger zone accessible explosive energy is practically employed in rock
should be taken into consideration because it is the major fragmentation, and the rest of the energy is wasted in the
cause of considerable damage on the nearby structures. form of unwanted events such as fly rocks, back breaks,
Even with the best care and competent personnel, flyrock etc. [1, 2]. Although safety has favorably been enhanced in
may not be totally avoided. There are several empirical the mining blasting operation, but various accidents due to
methods for prediction of flyrock phenomenon. Low per- flyrock phenomenon have been reported by the Mine
formance of these models is due to complexity of flyrock Safety and Health Administration (MSHA) [3–6].
analysis. Support vector machine (SVM) is a novel According to Institute of Makers of Explosives (IME),
machine learning technique usually considered as a robust flyrock is defined as the rock propelled beyond the blast
artificial intelligence method in classification and regres- area by the force of an explosion [7]. Many experimental
sion tasks. The aim of this paper is to test the capability of and theoretical researches on flyrock phenomenon have
SVM for the prediction of flyrock in the Soungun copper demonstrated that insufficient delay timing between blast
mine, Iran. Comparing the obtained results of SVM with rows, stemming and burden, geological discontinuities,
that of artificial neural network (ANN), it was concluded excessive charge, and deviation in drilling process is the
that SVM approach is faster and more precise than ANN main reasons of flyrock occurrence [8–11]. The researches
method in predicting the flyrock of Soungun copper mine. by Ladegaard-Pedersen and Holmberg [12] revealed the
relationships between powder factor, density of the
Keywords Flyrock  Blasting  Support vector machine  explosive and flyrock. Lundborg [13] studied the blasting
Artificial neural network  Soungun copper mine operations in granite and proposed a prediction model for
flyrock in hard rock. Also Bajpayee et al. [14] and Raina
et al. [15] have proposed empirical prediction models for
1 Introduction flyrock. These models have focused on the prediction of the
initial velocity and maximum distance of the fragmented
The basic purpose of mining blasting operation is rock rocks from the blasting face. Despite considerable efforts,
fragmentation. In this regard, a large amount of explosives difficult nature of rock engineering problems has caused
previously empirical methods to be not appropriate in
predicting flyrock phenomenon. Most of these models have
been developed based on blasthole diameter. To overcome
H. Amini (&)  R. Gholami  S. R. Torabi  J. Zadhesh
Faculty of Mining, Petroleum and Geophysics, this shortcoming, rather new techniques such as artificial
Shahrood University of Technology, Shahrood, Iran intelligence (e.g., artificial neural network (ANN), fuzzy
e-mail: h.amini@Shahroodut.ac.ir logic (FL), genetic algorithm (GA), and support vector
machine (SVM)) have been implemented in prediction and
M. Monjezi
Faculty of Engineering, Tarbiat Modares University, optimization of mining engineering problems successfully.
Tehran, Iran Tawadrous [16] reported ANNs as a reliable tool for blast

123
Neural Comput & Applic

design. Simultaneous prediction of fragmentation and fly-


rock in blasting operation using ANNs is performed by
Monjezi et al. [17]. A neuro-genetic approach was imple-
mented by Monjezi et al. [18] for prediction flyrock and
backbreak in open pit blasting. Rezaei et al. [19] developed
a fuzzy model to predict flyrock in surface mining. These
studies show that artificial intelligence-based models can
evaluate the flyrock phenomenon with an acceptable
accuracy. Furthermore, statistical learning theory-based
models have been developed in classification and predic-
tion problems recently. SVM is a statistical learning theory,
which has been developed in the reverse order to the
development of neural networks. Relatively easy training,
good generalization in theory and finding the globally best
model, no local optimal, unlike in neural networks, are
some advantageous of SVM [20–24]. A SVM-based
method was used for calculation subsidence coefficient by
Zhi-xiang [25]. Zhang et al. [26] employed the SVM to Fig. 1 Large burden and top priming causes flyrock [29]
determine the parameters for the probability-integral
method of mining subsidence prediction showing SVM
model is much better than the ANN in terms of reliability
and precision. In this study, the flyrock phenomenon due to
blasting operation in Soungun open pit copper mine is
studied by both ANN and SVM-based models. The theory
of neural network has been reported in extensive litera-
tures, thus only the background of SVM is summarized in
the following section.

2 Mechanisms of flyrock

2.1 Rifling

This event occurs when stemming material is insufficient.


Blast gases can stream up the blast hole along the path of
least resistance resulting in stemming ejection and some-
times ejection of the collar rock as harmful flyrock (Fig. 1).
The stemming column should contain individual rocks that
are of disproportionate size to the blasthole diameter, these
can become lethal projectiles. This mechanism of flyrock Fig. 2 Crater effects that could cause flyrock in bench blasting [29]
manifestation is closely related to the stemming release
pulse for airblast [27].
Cratering effect is also happened if blasting rows are
2.2 Cratering incorrectly initiated (initiating back rows earlier than front
rows) [27].
The stemming region of a blast pattern usually contains a
weakened layer due to previous blasting from the bench 2.3 Face bursting
above. In this region, blast gases easily jet into the air and
propagate cracks. The venting gases through this region This phenomenon occurs when explosive charges are in
produce cratering and flyrock events. This is particularly adjacent of the major geological structures or zones of
significant if insufficient stemming is used. Similar effects weakness (Fig. 2). The high-pressure gases of the explo-
can result if short inter-row delays are applied. These gases sives jet along the weakness zones (paths of low resistance)
will produce excessive airblast and flyrock (Fig. 2). and generate noise, airblast, and flyrock [27].

123
Neural Comput & Applic

Table 1 Flyrock distance prediction models [27] jy  f ðxÞje :¼ maxf0; jy  f ðxÞj  eg ð2Þ
Mechanisms of flyrock Prediction model
To estimate a linear regression
pffiffiffi2:6
Rifling 2 m f ðxÞ ¼ ðw  xÞ þ b ð3Þ
Lmax ¼ kg SH sin 2h0
pffiffiffi2:6
Cratering 2 m where w is the weighted matrix, x is the input vector, and
Lmax ¼ kg SH
pffiffiffi2:6
b is the bias term. With precision, one minimizes
Face bursting 2 m
Lmax ¼ kg Xm
B 1
kwk2 þC jy  f ðxÞje ð4Þ
2 i¼1

where C is a parameter usually considered for making a


2.4 Models to calculation of flyrock distance
trade-off between the generalization and the associated
error of the model. Written as a constrained optimization
Some attempts have been made to estimate flyrock dis-
problem, this reads
tance. Lundborg et al. [29] used semi-empirical approach
1 XN  
to estimate flyrock throw distance. Based on conservation Min Lðw; n; n0 Þ ¼ kwk2 þC ni þ n0i ð5Þ
of momentum and the scaling laws of spherical charge, a 2 i¼1
relationship between charge diameter d (inch) and rock
Subject to yi  wT x  b  ni þ e ð6Þ
velocity V (m/s) was obtained. Once rock velocity is known
then the flyrock range Lm (m) is calculated from the Subject to  yi þ wT x þ b  n0i þ e ð7Þ
equation of ballistic trajectories. For bench blasting,
Subject to ni ; n0i ; xi  0 ð8Þ
Lundborg et al. [29] proposed the maximum flyrock
according to (1): For all i = 1,…, m. It should be noted that according to (6)
2=3 and (7), any error smaller than e does not require a nonzero
Lm ¼ 260d ð1Þ
fi or f0i and does not enter the objective function (5).
An alternative approach was attempted by Roth [30] for Generalized kernel-based regression analysis is carried
obtaining the flyrock range. In this approach, the critical out in complete analogy in order to pattern recognition.
variable in all flyrock range calculation was the estimation Introducing Lagrange multipliers (i.e., ai ; a0i ), one thus
of V, the initial flyrock velocity. Richards and Moore [31] arrives at the following optimization problem: for C [ 0,
used a flyrock distance prediction model that is based e [ 0 chosen a priori, maximize
partly on the fundamental laws of projectile motion
1X N X N    
coupled with an empirical formulation that relates face Lða; a0 Þ ¼ ai  a0i xTi xj ai  a0i
2 i¼1 j¼1
velocity to scaled burden. Scaled burden is defined as
burden (or stemming) distance divided by the square root XN    
þ ai  a0i yi  ai þ a0i e ð9Þ
of the charge weight per delay. The model needs to be i¼1
calibrated for each mechanisms of flyrock as shown in
Table 1. Subject to 0  ðai  a0i Þ  C ð10Þ
Where h is drill hole angle, Lmax is flyrock range, m is where xi only appears inside an inner product. To get a
charge weight/m (kg/m), B is burden (m), SH is stemming potentially better representation of the data, the data points
height (m), g is gravitational constant (9.81 m/s2), and k is can be mapped into an alternative space, generally called
a constant parameter. feature space (a pre-Hilbert or inner product space) through
a replacement:
xi  xj ! uðxi Þ  uðxj Þ ð11Þ
3 Support vector machine
The functional form of the mapping u (xi) does not need
In pattern recognition, the SVM algorithm constructs to be known since it is implicitly defined by the choice of
nonlinear decision functions by training a classifier to kernel: k (xi, xj) = u(xi)  u(xj) or inner product in Hilbert
perform a linear separation in some high dimensional space space. With a suitable choice of kernel, the data can
that is nonlinearly related to input space. To generalize the become separable in feature space while the original input
SVM algorithm for regression analysis, an analog of the space is still non-linear. Thus, whereas data for n-parity or
margin is constructed in the space of the target values the two spirals problem are non-separable by a hyper plane
(y) by using Vapnik’s e-insensitive loss function. This in input space, it can be separated in the feature space by
function is shown in Fig. 3. RBF (Gaussian) kernel:

123
Neural Comput & Applic

Fig. 3 Concept of
e-insensitivity. Only the samples
out of the ±e margin will have a
nonzero slack variable, so they
will be the only ones that will be
part of the solution

2
longitudes and 38380 20N latitudes at an altitude of
kðxi ; xj Þ ¼ ekxi xj k =2r2
ð12Þ
2,390 m above sea level. Hydrothermal intrusion is the
Then, the regression estimate takes the form geological dominant phenomenon in the Soungun vicinity.
N X
X N Mineralization of this porphyry deposit is mainly hosted by
yi ¼ ðai  a0i Þuðxi ÞT uðxj Þ þ b altered quartz monzonite rocks. In this deposit, feldspar has
i¼1 j¼1 been altered to sericite and/or kaolinite. The porphyry rock
N X
X N has been changed to a more compact rock with quartz veins
¼ ðai  a0i ÞKðxi ; xj Þ þ b ð13Þ crosscuts. The other minerals observed in the area are
i¼1 j¼1 molybdenite, pyrite, and chalcopyrite. Copper ore is con-
where b is computed using the fact that (6) becomes an sidered as the main product of the mine, whereas molyb-
equality with ni = 0 if 0 \ ai \ C and (7) becomes an denum is a byproduct. The grade of this products increases
equality with n0i ¼ 0 if 0\ai \C. with increasing depth, particularly in the hypo-gene zone.
Several extensions of this algorithm are possible. From The probable reservoir in the Soungun deposit is about
an abstract point of view, it is just needed target function 1,700 million tons, whereas the confirmed ore reserve is
that depends on the vector (w, n). There are multiple degrees about 388 Mt with copper average grade of 0.67%. ANFO
of freedom for constructing this function, including some and detonating cord are applied for blasting operations in
freedom how to penalize, or regularize, different parts of the this mine. One of the undesirable incidents of the blasting
vector, and some freedom how to use the kernel trick. operation is flyrock that a maximum up to 100 m has been
Finally, the algorithm can be modified such that e should observed. A view of Soungun copper mine is shown in
not be specified a priori. Instead, one specifies an upper Fig. 4.
bound 0 \ v \ 1 on the fraction of points allowed to lie
outside the tube asymptotically, and the corresponding e is
computed automatically. This is achieved by using as pri-
mal objective function
!
1 2
X m
kwk þC vme þ jyi  f ðxÞje ð14Þ
2 i¼1

Instead of (4), and treating e C 0 as a parameter that has


been minimized over [32–39].

4 Case study

Soungun copper mine is located in the East Azarbaijan


Province in North West of Iran, between 46450 35E Fig. 4 A view of Soungun copper mine

123
Neural Comput & Applic

5 Dataset The optimal value for e depends on the type of noise


present in the data, which is usually unknown. Even if
A database including 245 datasets collected from the pre- enough knowledge of the noise is available to select an
vious blasting operation was used in the modeling process. optimal value for e, there is the practical consideration of
Table 2 shows input and output parameters. The datasets the number of resulting support vectors. e-insensitivity
were randomly divided into training and testing datasets. In prevents the entire training set meeting boundary condi-
both the methods, 70% of datasets was considered for tions and so allows for the possibility of sparsity in the dual
training and the rest was kept for testing the models. The formulations solution. So, choosing the appropriate value
maximum flyrock induced each blasting round was mea- of e is critical from theory.
sured from the newly created face by a tape meter. Since in this study the nonlinear SVM is applied, it
According to Table 2, blast pattern parameters such as would be necessary to select a suitable kernel function. The
blasthole diameter and length, spacing, burden, stemming, obtained results of previous published researches [42, 43]
powder factor, and specific drilling were considered for indicate that Gaussian radial basis function has a better
prediction of flyrock by SVM and ANN methods. efficiency than other Kernel functions. The form of the
In addition, for increasing model strength in recognizing Gaussian kernel is as follow:
relationship between inputs and corresponding outputs, 2

data should be normalized [40]. Normalization in range kðxi ; xj Þ ¼ ekxi xj k =2r2
ð16Þ
[-1 1] was performed by: where r is a constant parameter of the kernel and can either
xi  xmin control the amplitude of the Gaussian function and the
X ¼2 1 ð15Þ
xmax  xmin generalization ability of SVM. We have to optimize r and
where X is the normalized value, Xi is the actual value in find the optimal one. In order to find the optimum values of
column, and Xmax and Xmin are the maximum and minimum two parameters (r and e) and prohibit the overfitting of the
values of each related column, respectively. model, the datasets were separated into a training set of 170
compounds and a test set of 75 compounds randomly, and
the leave-one-out cross-validation of the whole training set
6 Prediction of flyrock by SVM was performed. The leave-one-out (LOO) procedure con-
sists of removing one example from the training set, con-
Similar with other multivariate statistical models, the per- structing the decision function on the basis only of the
formances of SVM for regression depend on the combi- remaining training data, and then testing on the removed
nation of several parameters. They are capacity parameter example. In this fashion, one tests all examples of the
C, e of e-insensitive loss function, the kernel type K, and its training data and measures the fraction of errors over the
corresponding parameters. C is a regularization parameter total number of training examples. The root mean square
that controls the trade-off between maximizing the margin error (RMSE) was used as an error function to evaluate the
and minimizing the training error. If C is too small then quality of model.
insufficient stress will be placed on fitting the training data. Detailed process of selecting the parameters and the
If C is too large then the algorithm will overfit the training effects of each parameter on generalization performance of
data. But, Wang et al. [41] indicated that prediction error is the corresponding model are shown in Fig. 5. To obtain the
scarcely influenced by C. In order to make the learning optimal r, the support vector learning machines with dif-
process stable, a large value should be set up for C (e.g., ferent r were trained, the r varying from 0.05 to 0.22,
C = 100). every 0.01. We calculated the RMSE on different r,

Table 2 Input and output parameters used for prediction flyrock phenomenon
Type of data Parameter Symbol Max. Ave. Min. Std. dev.

Input parameters Hole diameter (mm) D 152.4 134.87 114 .5 0.28


Hole length (m) L 15 12.07 9.5 1.48
Spacing (m) S 6.5 5.03 3 0.55
Burden (m) B 5 4.17 3 0.42
Stemming (m) T 4.5 4.03 2.8 0.36
Powder factor (kg/ton) Pf 0.93 0.43 0.21 0.1
Specific drilling (m) SD 0.11 0.06 0.01 0.01
Output Flyrock (m) Flyrock 95 69.8 25 19.05

123
Neural Comput & Applic

Fig. 5 Sigma versus RMSE RMSE RMSE


error (left) and Epsilon versus 0.45
RMSE error (right) on LOO
cross-validation 0.40
0.40

0.35
0.35

0.30
Epsilon
Sigma 0.02 0.04 0.06 0.08 0.10 0.12 0.14
0.10 0.15 0.20

Support Vector Machine recognized as the optimum model. This model is shown in
100
Fig. 7.
90
The model performance was evaluated by calculating
80
root mean square error (RMSE) and correlation coefficient
Test Data

70 (R). Figure 8 shows the ANN performance in the testing


60 datasets.
50
40
30 Measured 8 Discussions
Predicted
20
0 10 20 30 40 50 60 70 80 In this paper, we have demonstrated one of the SVM
Samples applications in predicting the flyrock of Soungun copper
mine, Iran. We interrogated the performance of this method
Fig. 6 Performance of SVM in prediction of flyrock for test data set
by comparing the results achieved by the best performance
of ANN Model. Comparing the obtained results presented
according to the generalization ability of the model based overall better performance of SVM model than ANN
on the LOO cross-validation for the training set in order to approach in both training and testing process. Figure 9
determine the optimal one. The optimal r was found as shows the scatter plot of the both methods in testing data
0.13. In order to find the optimal e, the RMSE on different e step.
was calculated. The curve of the RMS versus the sigma and The plots indicated an acceptable correlation coefficient
epsilon is shown in Fig. 5. From this figure, the optimal e (R = 0.97) for prediction the flyrock by SVM modeling.
was found as 0.08. Moreover, as it is quite depicted in Fig. 9, SVM is a more
From the above discussion, the r, e, and C were fixed to applicable model than ANN, particularly in situation where
0.13, 0.08, and 100, respectively, while the support vector the dataset is not distributed uniformly. In addition, for
number of the SVM model was 35. Correlation coefficients proving a good comparison, we have determined the rela-
of both training and testing datasets are 0.99 and 0.97, tive RMSE of the both SVM and ANN. The obtained
respectively. Figure 6 presents the performance of SVM in results showed that the RMSE of SVM is 4.5 while that of
prediction of flyrock providing based on the test step. the ANN is 7.98. In terms of running time, in addition, the
SVM method consumed a considerable less time compared
with the ANN methods in the prediction process. All of this
7 Prediction of flyrock by ANN expression can actually prove the highly performance of
the SVM in making a good prediction. However, ANN can
In order to check the accuracy of SVM in the prediction of be considered as an alternative approach after the SVM for
flyrock, obtained results of SVM were compared with that the prediction of flyrock in this study.
of the artificial neural network. In this way, for optimizing
the best network topology (e.g., number of hidden layers
and their corresponding neurons), trial and error process is 9 Conclusions
usually utilized. By examination of several network
topologies, a feed forward network with structure 7–12–1, Despite the considerable progress made over the last three
tangent sigmoid transfer function in hidden and output decades, significant challenges for wholly omitting of
layers and Levenberg–Marquardt training algorithm was unwanted flyrock phenomenon, injuries, and fatalities still

123
Neural Comput & Applic

Fig. 7 Structure of the Hidden Layer


optimum model

Input Layer Output Layer

S Flyrock

Pf
SD

Bias

Artificial Neural Network exist. Flyrock is caused by a mismatch of the distribution


100
of explosive energy, confinement of the explosive charge,
90
and mechanical strength of the rock. Therefore, exact and
80
accurate prediction of flyrock will be a critical task elim-
Test Data

70
inating some usual problems. Support vector machine
60
(SVM) is a novel machine learning method based on sta-
50 tistical learning theory, which has considerable features
40 including the fact that requirement on kernel and nature of
30 Measured the optimization problem, results in a uniquely global
Predicted
20 optimum, high generalization performance, and prevention
0 10 20 30 40 50 60 70 80
from converging to a local optimal solution. In this
Samples research, comparison of the application of SVM with ANN
Fig. 8 Prediction performance of ANN in testing data set model for prediction of Flyrock phenomenon in the

Fig. 9 Correlation between Artificial Neural Network Support Vector Machine


measured and predicted flyrock 100 110
by two considered artificial R=0.92
90 100 R=0.97
methods
80 90
Predicted Flyrock

Predicted Flyrock

80
70
70
60
60
50
50
40
40
30 Data Points Data Points
30
Linear Fit Linear Fit
20 20
20 40 60 80 100 20 40 60 80 100

Measured Flyrock Measured Flyrock

123
Neural Comput & Applic

Soungun copper mine is demonstrated. Although both 17. Monjezi M, Bahrami A, YazdianVarjani A (2010) Simultaneous
methods are data-driven models, it has been found that prediction of fragmentation and flyrock in blasting operation
using artificial neural networks. Int J Rock Mech Min Sci
SVM makes the running time considerably faster than 47:476–480
ANNs with higher accuracy. In Addition, the SVM tech- 18. Monjezi M, Amini Khoshalan H, Yazdian Varjani A (2010)
nique predicted results are quite reasonably better than Prediction of flyrock and backbreak in open pit blasting opera-
ANN model (Fig. 9). As a result, SVM method is a robust tion: a neuro-genetic approach. Arab J Geosci, doi:
10.1007/s12517-010-0185-3
method of prediction enabling to use in wide variety 19. Rezaei M, Monjezi M, Yazdian Varjani M (2001) Development
application of real world problems. of a fuzzy model to predict flyrock in surface mining. Safety Sci
49:298–305
20. Bishop CM (2006) Pattern recognition and machine learning.
Springer, Berlin
References 21. Steinwart I (2008) Support vector machines. Los Alamos
National Laboratory, information Sciences Group (CCS-3).
1. Singh TN, Singh V (2005) An intelligent approach to prediction Springer, Berlin
and control ground vibration in mines. Geotech Geol Eng 22. Cristianini N, Shawe-Taylor J (2000) An introduction to support
23:249–262 vector machines. Cambridge University Press, UK
2. Monjezi M, Rezaei M, Yazdian A (2010) Prediction of backbreak 23. Wang L (2005) Support vector machines: theory and applica-
in open pit blasting using fuzzy set theory. Expert Syst Appl tions, Nanyang Technological University, School of Electrical
37:2637–2643 and Electronic Engineering. Springer, Berlin
3. MSHA (1994) Accident investigation report: surface nonmetal 24. Martinez-Ramon M, Cristodoulou Ch (2006) Support vector
mine, fatal explosives and breaking agent’s accident. Alton Stone machines for antenna array processing and electromagnetic.
Company Inc., Illinios Universidad Carlos III de Madrid, Spain, p 120
4. MSHA (1999) Report of investigation: fatal explosives accident. 25. Zhi-xiang T, Pei-xian L, Li-li Y, Ka-zhong D (2009) Study of the
Surface nonmetal mine. Compass Quarries Inc, Paradise, Lan- method to calculate subsidence coefficient based on SVM. In:
caster County The 6th international conference on mining science & technology
5. MSHA (1999) Accident investigation report: surface coal mine, procedia earth and planetary science, vol 1, pp 970–976
fatal explosives accident. Appalachian Mining Services, Big 26. Zhang H, Wang HJ, Li YF (2009) SVM model for estimating the
Creek Mining, Inc., Mine no. 2, KY parameters of the probability integral method of predicting
6. Verakis HC, Lobb TE (2001) Blasting accidents in surface mines, mining subsidence. Min Sci Tech 19:0385–0388
a two decade summary. In: Proceedings of the 27th annual con- 27. Little TN (2007) Flyrock risk. In: Proceedings of 30th ISEE
ference on explosives and blasting technique, vol I. International conference on explosives and blasting technique, New Orleans,
society of explosive engineers, Cleveland, pp 145–152 Louisiana, pp 35–43
7. Institute of Makers of Explosives (IME) (1997) Glossary of 28. Bhandari S (1997) Engineering rock blasting operations. A. A.
commercial explosives industry terms. Safety publication, No. Balkema, Rotterdam
12, Washington, DC, p 16 29. Lundborg N, Persson A, Ladegaard-Pedersen A, Holmberg R
8. Holmeberg R, Persson G (1976) The effect of stemming on the (1975) Keeping the lid on flyrock in open-pit blasting. Eng Min J
distance of throw of flyrock in connection with hole diameters. 176:95–100
Swedish Detonic Research Foundation, Report DS 1, Stockholm 30. Roth JA (1979) A model for the determination of flyrock range as
9. Rehak TR, Bajpayee TS, Mowrey GL, Ingram DK (2001) Fly- a function of shot condition. US Department of Commerce, NTIS
rock issues in blasting. In: Proceedings of the 27th annual con- Report No PB81222358
ference on explosives and blasting technique, vol I. International 31. Richards AB, Moore AJ (2002) Flyrock control—by chance or
society of explosives engineers, Cleveland, pp 165–175 design. ISEE conference, New Orleans
10. Shea CW, Clark D (1998) Avoiding tragedy: lessons to be 32. Sanchez DV (2003) Advanced support vector machines and
learned from a flyrock fatality. Coal Age 103(2):51–54 kernel methods. Neurocomputing 55:5–20
11. Siskind DE, Kopp JW (1995) Blasting accidents in mines: a 16 33. Quang-Anh T, Xing L, Haixin D (2005) Efficient performance
year summary. In: Proceedings of the 21st annual conference on estimate for one-class support vector machine. Pattern Recogn
explosives and blasting technique. International society of Lett 26:1174–1182
explosives engineers, Cleveland, pp 224–239 34. Stefano M, Giuseppe J (2006) Terminated ramp-support vector
12. Ladegaard-Pedersen A, Persson A (1973) Flyrock in Blasting II, machines: a nonparametric data dependent kernel. Neural Netw
Experimental Investigation, Swedish Detonic Research Founda- 19:1597–1611
tion, Report DS 13, Stockholm 35. Lia Q, Licheng J, Yingjuan H (2007) Adaptive simplification of
13. Lundborg N (1974) The hazards of flyrock in rock blasting. solution for support vector machine. Pattern Recogn 40:972–
Swedish Detonic Research Foundation, Reports DS 12, 980
Stockholm 36. Agarwala S, Vijaya Saradhib V, Karnick H (2008) Kernel-based
14. Bajpayee TS (2004) Blasting injuries in surface mines with online machine learning and support vector reduction. Neuro-
emphasis on flyrock and blast area security. J Safety Res 35(1): computing 71:1230–1237
47–57 37. Lin HJ, Yeh JP (2009) Optimal reduction of solutions for support
15. Raina AK, Chakraborty AK, Choudhury PB, Sinha A (2011) vector machines. Appl Math Comp 214:329–335
Flyrock danger zone demarcation in opencast mines: a risk based 38. Eryarsoy E, Gary J, Aytug H (2009) Using domain-specific
approach. Bull Eng Geol Environ 70:163–172. doi:0.1007/ knowledge in generalization error bounds for support vector
s10064-010-0298-7 machine learning. Decis Support Syst 46:481–491
16. Tawadrous AS, Katsabanis PD (2007) Prediction of surface 39. Wu ChH, Tzeng GH, Lin RH (2009) A Novel hybrid genetic
crown pillar stability using artificial neural networks. Int J Numer algorithm for kernel function and parameter optimization in
Anal Methods Geomech 31(7):917–931 support vector regression. Exp Syst Appl 36:4725–4735

123
Neural Comput & Applic

40. Boser BE (1992) A training algorithm for optimal margin clas- 42. Dibike YB, Velickov S, Solomatine D, Abbott MB (2001) Model
sifiers. In: Proceedings of the 5th annual workshop on compu- Induction with support vector machines: introduction and appli-
tational learning theory. Pittsburgh, vol 5, pp 144–152 cation. J Comput Civil Eng 15(3):208–216
41. Wang WJ, Xu ZB, Lu WZ, Zhang XY (2003) Determination of 43. Bray M, Han D (2004) Support vector machines identification for
the spread parameter in the Gaussian kernel for classification and runoff modeling. J Hydroinformatics 6:265–280
regression. Neurocomputing 55:643–663

123

You might also like