You are on page 1of 8

ARTICLE IN PRESS

International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516

Contents lists available at ScienceDirect

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

Technical Note

Evaluation and prediction of blast-induced ground vibration using support


vector machine
Manoj Khandelwal 
Department of Mining Engineering, College of Technology and Engineering, Maharana Pratap University of Agriculture and Technology, Udaipur 313 001 India

a r t i c l e in fo

Article history:
Received 18 May 2009
Received in revised form
15 December 2009
Accepted 8 January 2010
Available online 31 January 2010

1. Introduction
management and owners. Thus, it is imperative to predict the
Surface mining operations have increased throughout the vibration level prior to the operations.
world for extraction of minerals from the earth crust. Drilling and A number of researchers have suggested various methods to
blasting combination is still an economical and viable method for minimize the ground vibration level during the blasting. Ground
the excavation and displacement of rockmass in mining as well as vibration is directly related to the quantity of explosive used and
in civil construction works. Whenever an explosive charge distance between blast face to monitoring point as well as
detonates in a blast hole, a gigantic amount of energy in terms geological and geotechnical conditions of the rock units in the
of pressure (50 GPa) and temperature (5000 K) is liberated [1–3]. excavation area [7,8]. There are a number of vibration predictors
Only a fraction of this energy is used for the actual fragmentation available suggested by different researchers [9–15]. All vibration
and displacement of rockmass, and the rest of the energy is predictor equations have their site specific constants, therefore,
wasted and creates a number of nuisances such as blast vibration, they cannot be used in a generalized way.
air blast, flyrock, noise, dust dispersion, back break, etc. [4]. Geological and geotechnical conditions and distance between
The ill effects of blasting are unavoidable and cannot be blast face to monitoring point cannot be altered but the only
completely eliminated but should be minimized to avoid damage factor i.e. quantity of explosive can be estimated based on certain
to the surrounding environment [5]. Among all the ill effects, empirical formulae proposed by the different researchers to make
ground vibration is a major concern to the planners, designers and ground vibrations in a permissible limit. An appropriate and rock
environmentalists. It is very important as compared to other friendly blasting can be only alternative for smooth progress of
blasting nuisances due to involvement of public residing in the the rock removal process.
close vicinity of mining sites, regulating and ground vibration There has always been the need of a simple technique for the
standards setting agencies together with mine owners [6]. Also prediction of blast induced ground vibration by some indirect but
with the emphasis shifting towards eco-friendly, geo-environ- relevant and reliable method with greater accuracy. So, here an
mental activities, the field of ground vibration have become an attempt has been made to predict the ground vibration by support
important parameter for the smooth running of a project. vector machine (SVM) taking into consideration of distance
To avoid socio-economic problems created by induced ground between blast face to monitoring point and maximum explosive
vibrations and to have cost effective blasting operations, charge used in a delay. The prediction capability of SVM is
pre-operational planning becomes essential. The economics, in compared with widely used conventional predictors vis-a -vis
case of small mining projects, may restrict continuous monitoring multi-variate regression analysis.
of ground vibrations during mining operations. By measuring Over the past few years, various artificial intelligent (AI)
vibration data, prior to actual operations and further planning techniques such as Hidden Markov Models (HMM), Artificial
with the help of predictor equations may help the mine Neural Networks (ANN) and Support Vector Machines (SVM) have
been used in various mining, civil and geo-engineering applica-
tions.
 Tel.: + 91 294 2471 379; fax: + 91 294 2471 056. Numerous researchers have used ANN to predict the blast
E-mail address: mkhandelwal1@gmail.com induced ground vibration [6–8,16–21]. ANN is an information

1365-1609/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijrmms.2010.01.007
ARTICLE IN PRESS
510 M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516

processing system simulating the structure and functions of the influence the process and results, when process and results are
human brain. It is a highly interconnected structure that consists not fully understood, and where historical or experimental data
of many simple processing elements (called neurons) capable of are available. The prediction of blast induced ground vibrations is
performing massively parallel computation for data processing also of this type.
and knowledge representation. The neural network is first
trained by processing a large number of input patterns and the
corresponding output. The neural network is able to recognize
2. Mechanism of ground vibration
similarities, when presented with a new input pattern after
proper training and predicting the output pattern. The conven-
When an explosive charge detonates in the blast hole, intense
tional pattern recognition method and ANN requires sufficient
dynamic stresses are set up around it due to sudden acceleration
number of samples, which are sometimes difficult to obtain [22],
of the rockmass by detonating gas pressure on the hole wall.
whereas SVM is based on structural risk minimization principle
The strain waves transmitted to the surrounding rock sets up a
and has very good generalization with few data samples.
wave motion in the ground [33]. The strain energy carried out by
SVM is a new generation learning system based on advances in
these strain waves fragments the rockmass due to different
statistical learning theory, enabling non-linear mapping of an
breakage mechanisms such as crushing, radial cracking and
n-dimensional input space into a higher dimensional feature
reflection breakage in the presence of a free face. The crushed
space, where, for example, a linear classifier can be used. The SVM
zone and radial fracture zone encompasses a volume of
can train non-linear models based on the structural risk
permanently deformed rock. When the stress wave intensity
minimization principle that seeks to minimize an upper bound
diminishes to the level where no permanent deformation occurs
of the generalization error rather than minimize the empirical
in the rockmass (i.e. beyond the fragmentation zone), strain
error as implemented in other neural networks. This induction
waves propagate through the medium as the elastic waves,
principle is based on the fact that the generalization error is
oscillating the particles through which they travel (Fig. 1). These
bounded by the sum of the empirical error and a confidence
waves in the elastic zone are known as ground vibration, which
interval term depending on the Vapnik–Chervonenkis (VC)
closely confirm to the visco-elastic behaviour. The wave motion
dimension. Based on this principle, SVM will achieve an optimal
spreads concentrically from the blast site in all directions and gets
model structure by establishing a proper balance between the
attenuated due to the spreading of fixed energy over a greater
empirical error and the VC-confidence interval, leading eventually
mass of material and away from its origin [34]. Even though, the
to a better generalization performance than other neural network
ground vibration attenuates exponentially with distance but due
models. An additional merit of SVM is that training SVM is a
to large quantity of explosive, it can still be high enough to cause
uniquely solvable quadratic optimization problem, and the
damage to buildings and other man made and natural structures
complexity of the solution in SVM depends only on the complex-
by causing dynamic stresses that exceed material strength [35].
ity of the desired solution, rather than on the dimensionality of
the input space. Thus, SVM use a non-linear mapping, based on a
kernel function, to transform an input space to a high dimension
space and then look for a non-linear relation between inputs and 3. The study area
outputs in the higher dimension space. SVM not only have a
rigorous theoretical background, but also can find global optimal The study was conducted at the Jayant opencast mine of
solutions for problems with small training samples, high dimen- Northern Coalfields Limited (NCL), which is a subsidiary company
sion, non-linearity and local optima. Originally, SVM were of Coal India Limited. It is located at Singrauli, District Sidhi (M.P.),
developed for pattern recognition problems [23–25]. Recently, India. The area of NCL lies geographically between latitudes of 241
SVM has been shown to give good performance for a wide variety 00 –241 120 and longitudes 821 300 –821 450 and belongs to the
of problems, such as non-linear regression. Gondwana super group. The dip of the strata is gentle and varying
Feng et al. [26] modeled non-linear displacement time series of from 21 to 51.
geo-materials using evolutionary support vector machines and The coalfield can be divided into two sub basins, viz. Moher
found very accurate results. Liu et al. [27] used support vector sub-basin (312 km2) and Singrauli Main basin (1890 km2).
machine approach to design the tunnel shotcrete–bolting support. The field is divided into eleven major mining blocks namely
Zhao [28] used SVM for the slope stability analysis and shown Kakri, Bina, Marrack, Khadia, Dhudhichua, Jayant, Nighahi,
that the SVM based first-order second-moment method reliability Amlohri, Moher, Gorbi and Jhingurdah [36].
analysis can be used successfully for slope reliability analysis
based on the limit equilibrium method, such as the Bishop’s
method and Spencer’s method. Zhi-xiang [29] calculated the
subsidence coefficient by SVM.
Khandelwal and Kankar [30] predicted the blast induced air
over pressure incorporating maximum charge per delay and
distance between blast face to monitoring point using support
vector machine and found better results than generalized cube
root predictor equation. Kovačević et al. [31] used SVM for the
estimation of values of soil properties and soil type classification
based on known values of particular chemical and physical
properties in sampled profiles and found very good results.
Khandelwal et al. [32] predicted the blast induced ground
vibration of Dharapani Magnesite Mine, Pithoragarh, India using
SVM and found superior results as compared to vibration
predictor equations.
These applications demonstrate that support vector machine is
capable of solving problems in which many complex parameters Fig. 1. Ground vibration due to blasting.
ARTICLE IN PRESS
M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516 511

Margin

P2 {wT x + b = 1}
P2
P1 P P {wT x + b = 0}
P1 {wT x + b = −1}

Fig. 3. Hyper-plane classifying two classes: (a) small margin and (b) large margin.

where w is an N-dimensional vector. The vector w defines a


direction perpendicular to the hyperplane. The scalar value b
Fig. 2. Blast face of Jayant mine. moves the hyper-plane parallel to itself, this value is sometimes
called the bias (or threshold). A distinct separating hyper-plane
should satisfy the constraints f ðxi Þ Z 1 if yi ¼ 1, and f ðxi Þ r 1 if
The overburden in this area is mostly medium to coarse- yi ¼ 1, or it can be presented as
grained sandstones, carbonaceous shales and shaly sandstones.
yi f ðxi Þ ¼ yi ðwT x þ bÞ Z 1 i ¼ 1; 2; . . . ; N ð2Þ
Fig. 2 shows the blast face of Jayant mine. The mine uses large
dragline (24 m3 bucket size and 96 m boom length) in 40 m Since the resulting geometric margin will be equal to 1/JwJ2, the
benches. The bench is drilled with 311 mm diameter holes. vector w that minimizes JwJ2 under constraint (2) is related to the
Normal blast consists of firing 50–60 holes; consuming 150–200 t vector that forms the optimal hyper-plane. So, the optimal hyper-
of explosive. Each hole of 35–40 m length is charged with 300 kg plane separating the data can be obtained as a solution to the
of explosive and the maximum charge per delay is about 6000 kg. following optimization problem:
Nonel and MS connectors are used for initiation. The inter-hole 1 2
Minimize 2JwJ ð3Þ
delay was 17–25 ms, whereas, inter-row delay was 2–4 times the
inter-hole delay.
Subject to yi ðwT xi þ bÞ Z 1 i ¼ 1; 2; . . . ; N ð4Þ
To find the solution of the above quadratic optimization
problem, the saddle point of the Lagrange function has to be
4. Support vector machine (SVM) determined. The Lagrangian function for optimization problem is
given as
A SVM is a supervised machine learning method based on the
X
N
statistical learning theory. It is a very useful method for Minimize Lðw; b; lÞ ¼ 12JwJ2  li ½yi ðwT xi þ bÞ1 ð5Þ
classification and regression in small-sample cases such as fault i¼1
diagnosis. Pattern recognition and classification using SVM is
where li Z0 are the Lagrange multipliers. To find the saddle point
described here in brief; a more detailed description can be found
one has to minimize this function over w and b and to maximize it
in [37,38].
over nonnegative Lagrange multipliers li Z0. Setting the deriva-
Initially consider a simple case of two classes, which can be
tives of L with respect to w and b to zero. We have:
separated by a linear classifier. Fig. 3 shows triangles and squares
stand for these two classes of sample points, respectively. Hyper- @Lðw; b; lÞ XN XN
¼ w yi li xi ¼ 0 ) w ¼ yi li xi ð6Þ
plane P is one of the separation planes that separate two classes. @w i¼1 i¼1
P1 and P2 (shown by dashed lines) are the planes those are parallel
to P and pass through the sample points closest to P in these two @Lðw; b; lÞ XN
classes. Margin is the distance between P1 and P2. The SVM tries ¼ yi li ¼ 0 ð7Þ
@b i¼1
to place a linear boundary between the two different classes, and
orientate it in such way that the margin is maximized, which Substituting results from Eqs. (6) and (7) into Eq. (5) gives
results in least generalization error. The nearest data points that
X
N
1 X N
used to define the margin are called support vectors. WðlÞ ¼ li  y y l l ðx x Þ ð8Þ
This is implemented by reducing it to a convex optimization i¼1
2 i;j ¼ 1 i j i j i j
problem: minimizing a quadratic function under linear inequality
The notation L(w,b,l) is changed to W(l) to reflect the last
constraints [37]. Consider a training sample set {(xi,yi)}; i= 1 to N,
transformation. To construct the optimal hyper-plane, coefficients
where N is the total number of samples. It is wished to determine,
li are to be determined that maximize the function (8). Thus by
among all linear separation planes that separates input samples
solving the Dual optimization problem, the coefficients ‘lio’ can be
into two classes, which separation plane will have the smallest
obtained which is required to express the ‘w’ from Eq. (6):
generalization error. Let us assume the samples can be classified
into two classes namely triangle and square class. Labels yi = 1 X
N
1 X N
Maximize WðlÞ ¼ li  y y l l ðx x Þ ð9Þ
and yi = + 1 are associated with triangle and square class,
i¼1
2 i;j ¼ 1 i j i j i j
respectively. If data are linearly separable, the hyper-plane
f(x) =0 that separates the given data is given as 8
>
> l Z0
< i
Subject to X
N
X
N ð10Þ
f ðxÞ ¼ wi xi þ b ¼ 0 >
> l y ¼ 0; i ¼ 1; 2; . . . ; N
ð1Þ :i¼1 i i
i¼1
ARTICLE IN PRESS
512 M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516

The value of bo is chosen to maximize margin and calculated as input and output parameters range with their mean and standard
deviation, respectively.
maxyi ¼ 1 ð/w xi SÞ þ minyi ¼ 1 ð/w xi SÞ
o o
bo ¼ ð11Þ
2
This leads to the decision function
6. Testing and validation of SVM
 X
N 
f ðxÞ ¼ sign li yi ðxi xj Þ þb ð12Þ
For training and testing of the data set, the Weka software [40]
i;j ¼ 1
is used, which is a collection of machine learning algorithms
When the training data is not linearly separable in feature for data mining tasks. The algorithms can be applied directly to
space, the optimization problem cannot be solved since no a data set. Weka contains tools for data pre-processing,
feasible solution exists. To allow for the possibility of samples classification, regression, clustering, association rules and visua-
violating constraints, slack variables (xi Z0) are introduced. A lization.
classier which generalizes well is then found by controlling both Sequential Minimal Optimization (SMO) algorithm is used due
the classier capacity (via JwJ) and the number of training errors. to its quickly solving capability. SMO decomposes the overall
The optimal hyper-plane separating the data can be obtained as a quadratic programming problem into sub-problems of quadratic
solution to the following optimization problem: programming by using the Osuna’s theorem to ensure conver-
X
N gence. There are two components in SMO: an analytic method for
1 2
minimize 2JwJ þ C xi \ ð13Þ solving for the two Lagrange multipliers; and a heuristic one for
i¼1 choosing multipliers in optimization. The advantage of SMO lies
in the fact that solving for two Lagrange multipliers can be done
subject to yi ðwT xi þbÞ Z1xi ð14Þ analytically. Thus, numerical quadratic programming optimiza-
and tion is avoided completely.
One-hundred fifty-four blast data sets of the Jayant mine were
xi Z 0; i ¼ 1; 2; . . . ; N ð15Þ used for the training of SVM model, whereas twenty new data sets
where C is a constant representing the error penalty. Rewriting were used for the validation of the SVM model. The results
the above optimization problem in terms of Lagrange multipliers, presented in this section demonstrate the performance of the
leads to the following problem: SVM model. Coefficient of determination between the predicted
and observed values of PPV is taken as a performance measure.
X
N X
N
The prediction was based on the input data sets discussed above.
Maximize WðlÞ ¼ li 12 yi yj li lj ðxi xj Þ ð16Þ
i¼1 i;j ¼ 1
Fig. 4 illustrates the measured and predicted PPV on 1:1 slope
line with their respective coefficient of determination (CoD). CoD
8
>
> 0 r li rC
< Table 2
Subject to X
N
ð17Þ Output parameter with range, mean and standard deviation.
>
> l y ¼ 0; i ¼ 1; 2; . . . ; N
:i¼1 i i
S. No. Output parameter Range Mean Standard
The Sequential Minimal Optimization (SMO) algorithm gives deviation
an efficient way of solving the dual problem arising from the
1. Peak particle velocity 0.31– 15.594 16.808
derivation of the SVM. SMO decomposes the overall QP problem (mm/s) 92.30
into QP sub-problems.

5. Data set

A total of 174 blast vibration records were monitored at


different vulnerable and strategic locations in and around to the
Jayant opencast mine as per the ISRM standards [39]. Out of 174
blast vibration data sets, 154 were used for the training and
testing of SVM model as well as to determine site constants for
the different vibration predictor equations and MVRA equations,
whereas 20 randomly selected data sets were used for the
validation of the SVM model as well as different conventional
vibration predictors and MVRA equations. Explosive charge used
per delay and distance between blast face to monitoring point
was taken as an input, whereas PPV was taken as an output
parameter for the validation and testing of the SVM, MVRA and
widely used conventional predictors. Tables 1 and 2 show the Fig. 4. Measured and predicted PPV by SVM.

Table 1
Input parameter with range, mean and standard deviation.

S. No. Input parameter Range Mean Standard deviation

1. Maximum charge per delay (Qmax) in kilograms 75–6000 1374.339 1691.647


2. Distance of monitoring point from blasting face (D) in meters 35–8400 829.903 1462.66
ARTICLE IN PRESS
M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516 513

between predicted and measured values is as high as 0.960, Table 4


whereas, mean absolute error (MAE) is 0.257 for PPV. Sample testing data set used for validation of the SVM model.

S. No. Qmax (kg) D (m) PPV (mm/s)

7. Prediction by multi-variate regression analysis (MVRA) 1. 5600 1000 6.57


2. 2300 350 69.8
The purpose of multiple regressions is to learn more about the 3. 300 203 8.8
4. 150 105 37.58
relationship between several independent or predictor variables 5. 2487 1898 20.97
and a dependent or criterion variable. The goal of regression
analysis is to determine the values of parameters for a function
that cause the function to best fit a set of data observations
provided. In linear regression, the function is a linear (straight-
line) equation. When there is more than one independent Table 5
Different conventional predictors.
variable, then multivariate regression analysis is used to get
the best-fit equation. Multiple regressions solve the data sets Name Equation
by performing least squares fit. It constructs and solves the
simultaneous equations by forming the regression matrix and USBM (1959) v= K [D/OQmax]  B
Langefors–Kihlstrom (1963) v= K [O (Qmax/D2/3)]B
solving for the co-efficient using the backslash operator. The
General predictor (1964) v= K D  B (Qmax)A
MVRA was done by same data sets, which were used in the SVM Ambraseys–Hendron (1968) v= K [D/ (Qmax)1/3]  B
training. The equation for PPV prediction by MVRA is Bureau of Indian Standard (1973) v= K [(Qmax/D2/3)]B
Ghosh–Daemen predictor (1983) v= K [D/OQmax]  B e  aR
PPV ¼ 15:6755þ 0:0013Qmax ðkgÞ20:0023D ðmÞ ð18Þ CMRI predictor (1993) v= n+ K [D/OQmax]  1
This equation is used for the prediction of PPV for twenty blast
Where v, peak particles velocity (PPV) in mm/s; Qmax, Maximum charge per delay
vibration cases. Fig. 5 illustrates the measured and predicted PPV in kg; D, Distance between blast face to vibration monitoring point in m, and; K, B,
on 1:1 slope line with their respective CoD. The CoD for PPV is a and n, site constants.
0.142. This shows that an MVRA equation is not able to predict the
PPV up to the desired level of accuracy. The MAE for PPV is 2.821.
The high MAE value of PPV shows that prediction by MVRA is Table 6
having high error (Tables 3 and 4). Calculated values of site constants.

Equation Site constants

8. Prediction by conventional predictors K B A a n

USBM 179.3081 1.0904


Table 5 shows the various available conventional vibration
Langefors–Kihlstrom 44.43242 –1.1795
predictor equations proposed by various researchers. The site General predictor 212.27 1.0949 0.5203
constants were determined from the multiple regression analysis Ambraseys–Hendron 329.3063 –0.9702
of earlier used 154 data sets. The calculated values of site constants Bureau of Indian Standard 6.328489 0.2118
for the various predictor equations are shown in Table 6. Ghosh–Daemen predictor 780.36 1.2588 0.0004
CMRI predictor 168.91 1.5669

Fig. 5. Measured and predicted PPV by MVRA.

Fig. 6. Measured and predicted PPV by USBM predictor.

Table 3 Figs. 6–12 illustrate the predicted graph between measured


Sample training data set used for learning the SVM model. and predicted PPV by conventional predictor equations on 1:1
S. No. Qmax (kg) D (m) PPV (mm/s)
slope line with their respective coefficient of correlation.

1. 350 101 15.66


2. 5600 3000 1.64 9. Results and discussion
3. 2400 950 3.43
4. 1800 325 33.27
5. 737 420 13.72
Fig. 13 shows a comparison between predicted PPV by SVM,
MVRA and conventional predictor equations. Here, prediction by
ARTICLE IN PRESS
514 M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516

Fig. 7. Measured and predicted PPV by Langefors–Kihlstrom predictor. Fig. 10. Measured and predicted PPV by Bureau of Indian Standard predictor.

Fig. 8. Measured and predicted PPV by General predictor. Fig. 11. Measured and predicted PPV by Ghosh–Daemen predictor.

Fig. 9. Measured and predicted PPV by Ambraseys–Hendron predictor. Fig. 12. Measured and predicted PPV by CMRI predictor.

SVM is closer to the measured PPV, whereas prediction by


conventional predictors and MVRA has wide variation. The
accuracy of MVRA and conventional predictors deteriorates at PPV protect the surrounding environment and structure. The use
of 10 mm/s and higher. Fig. 13 also revealed that SVM predicted PPV of any conventional vibration predictor without validation may
is very close to the measured PPV line, whereas conventional invite further complication for smooth conduct of mining
predictors show very high level of error. Table 7 shows the CoD and operations. This study indicates that all conventional predictors
MAE of PPV predicted by SVM, MVRA and the various conventional are either over estimating or underestimating the safe explosive
predictors. It can be said that prediction capability of SVM is quite charge to keep the PPV level under the safe limit. Both the
remarkable and compares well to field observations. predictions are not appropriate for the site where populations are
residing very near to the mine.
It was found that coefficient of determination between
10. Conclusions measured and predicted PPV was very high by SVM, whereas it
was very less by different conventional vibration predictors and
Based on the study, it is established that the SVM seems to be MVRA. Coefficient of determination between measured and
the better option for close and appropriate prediction of PPV to predicted PPV was 0.960 by SVM, while it was ranging from
ARTICLE IN PRESS
M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516 515

Fig. 13. Comparison of PPV.

Table 7 Acknowledgments
CoD and MAE of PPV by various models.

Model CoD MAE The financial assistance provided by Council of Scientific and
Industrial Research, New Delhi, India is thankfully acknowledged.
SVM 0.960 0.257 The thanks are also due to the mine management of Northern
MVRA 0.142 2.821
Coalfields Limited, Singrauli for providing necessary assistance
USBM 0.633 1.145
Langefors–Kihlstrom 0.106 2.177 during the field visit.
General predictor 0.615 1.122
Ambraseys–Hendron 0.434 1.071
Bureau of Indian Standard 0.278 2.127 References
Ghosh–Daemen predictor 0.659 1.477
CMRI predictor 0.612 1.491
[1] McKenzie C. Quarry blast monitoring technical and environmental perspec-
tive. Quarry Manage 1990;23–29.
[2] ISRM. Suggested method for blast vibration monitoring. Int J Rock Mech Min
Sci 1992;29:145–6.
[3] Cheng G, Huang SL. Analysis of ground vibration caused by open pit
production blast. In: Holmberg, editor. Explosive and blasting technique.
Rotterdam: Balkema; 2000. p. 63–70.
0.106 to 0.659 by different conventional vibration predictors and [4] Hagan TN. Rock breakage by explosives. In: Proc Natl Symp Rock
0.142 by MVRA. Application of SVM showed exceptional con- Fragmentation, Adelaide, Australia, 1973. p. 1–17.
formity between the measured and predicted PPV as compared to [5] Wiss JF, Linehan PW. Control of vibration and air noise from surface coal
mines—III. US Bur Mines Rep OFR 103(3)–79, 1978.
different conventional vibration predictors and MVRA. [6] Khandelwal M, Singh TN. Evaluation of blast induced ground vibration
Considering the complexity of the relationship among the predictors. Soil Dyn Earthquake Eng 2007;27:116–25.
inputs and outputs, the results obtained by SVM is highly [7] Khandelwal M, Singh TN. Prediction of blast induced ground vibrations and
frequency in opencast mine—a neural network approach. J Sound Vib
encouraging and satisfactory. SVM can learn new patterns that
2006;289:711–25.
are not previously available in the training data set. SVM can also [8] Khandelwal M, Singh TN. Prediction of blast induced ground vibration using
update knowledge over time as long as more training data sets are artificial neural network. Int J Rock Mech Min Sci 2009;46:1214–22.
[9] Duvall WI, Petkof B. Spherical propagation of explosion generated strain
presented. Therefore, the technique results in a greater degree of
pulses in rock. US Bur Mines Rep Invest 5483, 1959. 21pp.
accuracy than any other analysis techniques. [10] Langefors U, Kihlstrom B. The modern technique of rock blasting. New York:
By adopting SVM technique, PPV can be predicted prior to blast Wiley; 1963.
and accordingly blast design can be modified, so that blast [11] Davies B, Farmer IW, Attewell PB. Ground vibrations from shallow sub-
surface blasts. The Engineer (London) 1964;217:553–9.
nuisances can be minimized with greater degree of explosive [12] Ambraseys NR, Hendron AJ. Dynamic behaviour of rockmasses. In: Rock Mech
energy utilization. Eng Pract. London: Wiley, 1968. p. 203–7.
ARTICLE IN PRESS
516 M. Khandelwal / International Journal of Rock Mechanics & Mining Sciences 47 (2010) 509–516

[13] Bureau of Indian Standards. Criteria for safety and design of structures [26] Feng XT, Zhao H, Li S. Modeling non-linear displacement time series of geo-
subjected to underground blast. ISI Bulletin 1973; IS—6922. materials using evolutionary support vector machines. Int J Rock Mech Min
[14] Ghosh A, Daemen JK. A simple new blast vibration predictor. In: Proc 24th US Sci 2004;41(7):1087–107.
Symp Rock Mech, Texas A&M University, 1983. p. 151–61. [27] Liu KY, Qiao CS, Tian SF. Design of tunnel shotcrete–bolting support based on a
[15] Pal Roy P. Putting ground vibration predictors into practice. Colliery Guardian support vector machine approach. Int J Rock Mech Min Sci 2004;41(3):510–1.
1993;241:63–7. [28] Zhao H. Slope reliability analysis using a support vector machine. Comput
[16] Khandelwal M, Singh TN. Prediction of blast induced air overpressure in Geotech 2008;35:459–67.
opencast mine. Noise Vib Worldwide 2005;36(2):7–16. [29] Zhi-xiang T, Pei-xian L, Li-li Y, Ka-zhong D. Study of the method to calculate
[17] Kuzu C. The importance of site-specific characters in prediction models subsidence coefficient based on SVM. Proc Earth Planet Sci 2009;1:
for blast-induced ground vibrations. Soil Dyn Earthquake Eng 2008;28(5): 970–6.
405–14. [30] Khandelwal M, Kankar PK. Prediction of blast-induced air overpressure using
[18] Monjezi M, Dehghani H. Evaluation of effect of blasting pattern parameters support vector machine. Arabian J Geosci 2009 (online).
on back break using neural networks. Int J Rock Mech Min Sci 2008;45(8): [31] Kovačević M, Bajat B, Gajić B. Soil type classification and estimation of soil
1446–1453. properties using support vector machines. Geoderma 2009 (online).
[19] Mohamed MT. Artificial neural network for prediction and control of blasting [32] Khandelwal M, Kankar PK, Harsha SP. Evaluation and prediction of blast
vibrations in Assiut (Egypt) limestone quarry. Int J Rock Mech Min Sci induced ground vibration using support vector machine. Min Sci Tech 2009
2009;46(2):426–31. (in press).
[20] Khandelwal M, Kumar DL, Mohan Y. Application of soft computing to predict [33] Attewell PB. Recording and interpretation of shock effects in rock. Min Miner
blast-induced ground vibration. Engineering with Computers, 2009 (online). Eng 1964:21–8.
[21] Monjezi M, Bahrami A, Varjani AY. Simultaneous prediction of fragmentation [34] Dowding CH. Blast vibration monitoring and control. Englewoods Cliffs:
and flyrock in blasting operation using artificial neural networks. Int J Rock Prentice-Hall; 1985.
Mech Min Sci 2009; doi:10.1016/j.ijrmms.2009.09.008. [35] Siskind DE, Stagg MS, Kopp, JW, Dowding CH. Structure response and damage
[22] Zacksenhouse M, Braun S, Feldman M. Toward helicopter gearbox diagnostics produced by ground vibration from surface mine blasting. US Bur Mines Rep
from a small number of examples. Mech Syst Signal Process 2000;14(4): Invest 8507, 1980. 74pp.
523–43. [36] Singh VK. Northern Coalfields Ltd.: surging ahead with time. J Mines Met
[23] Scholkopf B, Burges C, Vapnik V. Extracting support data for a given task. In: Fuels 2004:51.
Proc 1st Int Conf Knowl Discovery Data Min, Menlo Park, Calif, Am Assoc Artif [37] Vapnik VN. Statistical learning theory. New York: Wiley; 1998.
Intell Press, 1995. [38] Cristianini N, Shawe-Taylor NJ. An introduction to support vector machines.
[24] Schmidt M. Identifying speaker with support vector networks. In: Proc Cambridge: Cambridge University Press; 2000.
Interface ‘96, Sydney, 1996. [39] ISRM. Suggested method for blast vibration monitoring. Int J Rock Mech Min
[25] Muller KR, Smola JA, Scholkopf B. Prediction time series with support Sci 1992;29:145–6.
vector machines. In: Proc Int Conf Artif Neural Networks, Lausanne, 1997. [40] Witten IH, Frank E. Data mining: practical machine learning tools and
p. 999–1004. techniques, 2nd ed.. San Francisco: Morgan Kaufmann; 2005.

You might also like