Evaluation of Safe Explosive Charge in Surface Mines using Artificial
Neural Network
Manoj Khandelwal
Department of Mining Engineering, College of Technology and Engineering,
Maharana Pratap University of Agriculture and Technology, Udaipur 313 001, India
Keywords: Safe Explosive Charge, Blast Vibration Predictors, Artificial Neural Network.
Abstract: The present paper mainly deals with the prediction of maximum explosive charge used per delay (Q
MAX
)
using artificial neural network (ANN) incorporating peak particle velocity (PPV) and distance between blast
face to monitoring point (D). 150 blast vibration data sets were monitored at different vulnerable and
strategic locations in and around major coal producing opencast coal mines in India. 124 blast vibrations
records were used for the training of the ANN model vis-à-vis to determine site constants of various
conventional vibration predictors. Rest 26 new randomly selected data sets were used to test, evaluate and
compare the ANN prediction results with widely used conventional predictors. Results were compared
based on coefficient of correlation (R) and mean absolute error (MAE) between calculated and predicted
values of Q
MAX
.
1 INTRODUCTION
The exploitation of economic minerals from earth
crust is increasing day by day at a faster pace since
last decade to fulfill the increasing demand of
minerals. This has led to the substantial increase in
the consumption of explosive. When an explosive
detonates in a blast hole, a tremendous amount of
energy, in terms of pressure (up to 50 GPa) and
temperature (up to 5000 K), is released (Hino, 1956;
McKenzie, 1990; Cheng and Huang, 2000).
Although, significant developments have taken place
in explosive technology, the explosive energy
utilization has not made much progress due to the
complexity of the various rock parameters. Only a
fraction of explosive energy (20-30%) is used in the
actual breakage and displacement of the rock mass,
and the rest of the energy is spent in undesirable
effects like ground vibrations, fly rocks, noises, back
breaks, over breaks, etc. (Hagan, 1973; Dowding,
1985).
As the ground vibration is the most important
environmental effect of blasting operation some
regulations related to structural damages caused by
ground vibration have been developed (Alipour and
Ashtiani, 2011). The regulations are primarily based
on the peak particle velocity (PPV) resulted from
blasting operations. To come out with proper
amounts of maximum charge per delay which
produces limited ground vibration, several empirical
conventional vibration predictors are available
proposed by different researchers (Duvall and
Petkof, 1959; Langefors and Kihlstrom, 1963;
Ambraseys and Hendron, 1968; Bureau of Indian
Standard, 1973; Pal Roy, 1993). These conventional
predictors are normally used for estimating PPV of
ground vibration by blasting. All the predictors
estimate the PPV mainly based on two parameters
(maximum charge used per delay and distance
between blast face and monitoring point). For the
same excavation site, different predictors give
different values of safe PPV vis-à-vis safe charge
per delay. There is no uniformity in the predicted
result by different predictors (Khandelwal and
Singh, 2009; Khandelwal, 2010). It is well known
that the PPV is influenced by various geological,
geotechnical, blast geometry and explosive
parameters, which have not been incorporated in any
of the available predictors. It seems that there is a
great need to evaluate the efficiency and credibility
of various empirical conventional predictors to
calculate maximum charge per delay.
In the present paper, an attempt has been made to
predict the Q
MAX
using artificial neural network
(ANN) by incorporating peak particle velocity
366
Khandelwal M..
Evaluation of Safe Explosive Charge in Surface Mines using Artificial Neural Network.
DOI: 10.5220/0004761703660371
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 366-371
ISBN: 978-989-758-015-4
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

(PPV) and distance from blast face to monitoring
point (D). Prediction capability of ANN is also
compared by various available conventional
predictors based on coefficient of correlation and
mean absolute error.
2 SITE DESCRIPTION
The field study was conducted at three different
opencast coal mines of Sinagreni Collieries
Company Limited (SCCL), Andhra Pradesh, India.
The SCCL area is mostly covered by limestone of
Pakhals in the western and southern parts and slowly
grades into the sandstone of Gondwana series in
North-easterly direction. The other geological units
found within the project area are Talcher and
Barakars. Kamthis are observed away from the
project area in northern and eastern directions.
The limestone is massive, flaggy and at places
striking in NW-SE direction, dipping towards NE
with dip amount varying from 350 – 400. At the
contact zone between limestone and sandstone,
calaceous beds are observed within grades into
sandstone. The sandstone is soft and coarse grained.
The various units of lower Gondwana are abutting
each other in different directions due to structural
disturbances in that area.
In general, this area consists of soft soil upto 2 m
depth followed by medium to coarse grained grey
sandstone overburden along with shale and thick
coal bands of varying thickness of 17.67 to 49.58 m.
Thickness of top seam is varying from 1.4 to 4.4 m
and the bottom seam thickness is varying from 2.75
to 5.07 m. The partition thickness consisting of
mostly medium grained grey sandstone and it is
varying from 4.87 to 13.0 m.
3 THE PHILOSOPHY OF
ARTIFICIAL NEURAL
NETWORK
An artificial neural network can be considered as a
soft tool to model the brain reflection (solution) to
the given problems. Wide range of capabilities of the
neural networks such as generalization,
classification, noise reduction and prediction have
made the method applicable for solving problems in
various fields of science and technology. In the
neural networks, deduction is performed using
highly interconnected computing cells known as
"neurons", which in fact are mathematical functions
of linear or nonlinear (Haykin, 1994). The
computing cells are usually set in three or more
successive layers, known as network architecture or
topology (Rafiq et al, 2001; Dreyfus, 2004). Number
of the neurons in the extreme outmost layers is equal
to the problem independent and dependent variables.
On the other hand, number of the intermediate
(hidden) layer(s) and number of their respective
neurons is dependent to the problem environment. In
fact hidden layer(s) can be considered as the
computational units of a network.
Feed-forward network with error back
propagation algorithm is the most commonly used in
solving complicated problems. In these networks
that are also called multilayer perceptrons (MLPs),
often one or more hidden layers of sigmoid neurons
followed by output(s) linear neuron(s) give the best
results (Babuska, 2004; Tawadrous, 2006; Demuth
and Beale, 2008).
During training process, the network is given
values of both the independent and actual measured
dependent variables. When the difference between
the model predicted values with that of the real
values reaches to a predefined threshold, the training
process is stopped (Babuska, 2004). Prior to start
network training, both the input and output values
should be normalized (Rafiq et al, 2001; Demuth &
Beale, 2008). Following training, the network
performance is tested applying testing datasets. To
get a more coherent result, the testing datasets
should not be incorporated for learning the network.
For a full coverage of data variability, these datasets
are selected from the sorted original database using a
random mechanism.
4 DATA SET
One of the most important stages in the ANN
technique is data collection. In the present study,
150 blast vibration records were monitored at
different vulnerable and strategic locations in and
around the mines as per ISRM (1992) standards.
Among which, 124 blast vibration data sets were
chosen for the training of the network and rest 26
data sets were used for the testing of the ANN
network. The data was divided into training and
validation datasets using sorting method to maintain
statistical consistency. The range of distance of
monitoring point from blasting face and PPV is 35 –
8400 m and 0.31 – 92.30 mm/s respectively,
whereas range of QMAX is 75 – 6000 kg.
EvaluationofSafeExplosiveChargeinSurfaceMinesusingArtificialNeuralNetwork
367

5 NETWORK ARCHITECTURE
Baheer (2000) and Hecht-Neilsen (1987) indicated
that one hidden layer may be sufficient for most
problems. Two hidden layers may be necessary for a
learning function with discontinuities (Masters,
1994). Lippmann (1987) and Rumelhart et al. (1986)
indicated that there is rarely an advantage in using
more than one hidden layer. Therefore, one hidden
layer was preferred in this study. However, the
number of neurons is the most critical task in the
ANN structure. The heuristics proposed for this
purpose are summarized in Table 1.
Table 1: The heuristics proposed for the number of neuron
to be used in hidden layer(s) (Ni: number of input neuron,
N0: number of output neuron)
Heuristic
Calculated
number of
neuron for
this study
Reference
≤ 2 x Ni+1 5
Hecht-Nielsen
(1987)
3Ni 6 Hush (1989)
(Ni +
N0)/2
2 Ripley (1993)
2Ni / 3 1 Wang (1994)
√(Ni x N0) 2 Masters (1994)
2Ni 4
Kannellopoulas
and Wilkinson
(1997)
As can be seen from Table 1, the number of
neurons that may be used in the hidden layer varies
between 1 and 6, depending on the proposed
heuristics in the literature. The ANN structures were
trained by using number of hidden neurons defined
above. By considering the findings obtained from
different trials, the ANN structure consisting of one
hidden layer with 6 neurons (Fig. 1) was selected for
the given problem. The datasets were normalized
between zero and one considering the maximum
values of input parameters.
Feed forward back propagation neural network
architecture (2-6-1) is adopted due to its
appropriateness to predict the Q
MAX
. Pattern
matching is basically an input/output mapping
problem. The closer the mapping, better the
performance of the network is.
The number of training cycles is important to
obtain proper generalization of the ANN structure.
Theoretically, excessive training, which is also
known as over-learning, can result in near-zero error
on predicting training data. However, this over-
Figure 1: Suggested ANN for the case study.
learning may result in loss of the ability of the ANN
to generalize from the test data, Fig. 2 (Basheer and
Hajmeer, 2000). The increasing point in the error of
the test data or the closest point to the training curve
is considered to represent the optimal number of
cycles for the ANN architecture.
All the input and output parameters were
normalized between 0 and 1. Equation 10 was used
for the scaling of input and output parameters.
Normalized value = (max. value – unnormalized
value) / (max. value – min. value)
Figure 2: A criteria for termination of training and
selection of optimum network architecture (Basheer and
Hajmeer, 2000)
PPV
Distance
Q
MAX
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
368

6 TESTING AND VALIDATION
OF ANN MODEL
To test and validate the ANN model, a data sets
were chosen, which was not used while training the
network, was employed. As Bayesian interpolation
(MacKay, 1992) has been used, there was no danger
of over-fitting or under-fitting problems. Fig. 3
illustrates the calculated and predicted QMAX.
Here, coefficient of correlation is as high as 0.985,
whereas, MAE is 94.36.
Figure 3: Calculated vs. predicted Q
MAX
by ANN.
7 ESTIMATION OF MAXIMUM
CHARGE PER DELAY BY
CONVENTIONAL
PREDICTORS
Table 2 illustrates the various available conventional
vibration predictor equations proposed by different
researchers (Duvall and Petkof, 1959; Langefors and
Kihlstrom, 1963; Ambraseys and Hendron, 1968;
Bureau of Indian Standard, 1973; Pal Roy, 1993).
These predictors have been employed to calculate
the safe quantity of charge that can be blasted per
delay, with minimum abuse in the surrounding rock
mass.
Empirical equations are versions of the following
general form that typically are used by investigators
(Davies et al, 1964)
PPV = K.D
A
.Q
MAX
B
Where,
v = Peak particles velocity (PPV), mm/s,
Q
MAX
= Maximum charge per delay, kg,
D = Distance between blast face to vibration
monitoring point, m, and
K, A, B, and n = Site constants.
Table 2: Different conventional predictors.
Name Equation
USBM (Duvall and
Petkof, 1959)
v = K (D/√Q
MAX
)
-B
Langefors – Kihlstrom
(1963)
v = K [√ (Q
MAX
/D
2/3
])
B
Ambraseys – Hendron
(1968)
v = K [D/ (Q
MAX
)
1/3
]
-B
Bureau of Indian
Standard (1973)
v = K [(Q
MAX
/D
2/3
])
B
CMRI Predictor (Pal
Roy, 1993)
v = n + K (D/√Q
MAX
)
-1
All the conventional vibration predictors have
site specific constants. The value of site constants
also varied as the ground conditions changed.
Moreover, these are derived based on only two
parameters, i.e. QMAX and the distance from
monitoring point to blast face.
These conventional vibration predictors have
been advocated in order to analyse the blast data.
The site constants of these predictors were
determined from the multiple regression analysis of
the 124 blast vibration cases. The calculated values
of site constants as well as their respective
coefficient of correlation (R) for the various
predictor equations are given in Table 3.
Figs. 4-8 demonstrate the prediction capability of
various conventional predictors to predict QMAX.
Here, coefficient of correlation is ranging from
0.752 to 0.316, which is maximum for the CMRI
predictor, whereas minimum for the Langefors-
Khilstrom and Bureau of Indian Standard predictor.
Mean absolute error was ranging from 633.44 to
2910.97. It was maximum for Bureau of Indian
Standard predictor whereas minimum for USBM
predictor.
Figure 4: Calculated vs. predicted Q
MAX
by USBM
predictor.
y=0.997x+11.42
R=0.985
0
1000
2000
3000
4000
5000
0 1000 2000 3000 4000 5000
PredictedQmax(kg)
MeasuredQmax(kg)
ANN
y=0.332x+463.9
R=0.425
0
500
1000
1500
2000
2500
3000
3500
4000
0 1000 2000 3000 4000 5000
PredictedQ
max
(kg)
MeasuredQ
max
(kg)
USBM
EvaluationofSafeExplosiveChargeinSurfaceMinesusingArtificialNeuralNetwork
369

Table 3: Calculated values of site constants.
Name of
Predictor
Site Constants
Final Q
MAX
Equation
K B
n
USBM 166.34 1.291
Q
MAX
=
D
2
(v/166.34)
2/1.29
1
Langefors –
Kihlstrom
0.93 0.857
Q
MAX
=
D
2/3
(v/0.93)
2/0.857
Ambraseys
– Hendron
1093.96 1.424
Q
MAX
=
D
3
(v/1093.96)
3/1.
424
Bureau of
Indian
Standard
0.929 0.428
Q
MAX
=
D
2/3
(v/0.929)
1/0.42
8
CMRI
Predictor
165.9
-
3.28
4
Q
MAX
=
D
2
(v+3.284/165.
9)
2
Figure 5: Calculated vs. predicted Q
MAX
by Langefors –
Kihlstrom predictor.
Figure 6: Calculated vs. predicted Q
MAX
by Ambraseys-
Hendron predictor.
Figure 7: Calculated vs. predicted Q
MAX
by Bureau of
Indian Standard predictor.
Figure 8: Calculated vs. predicted Q
MAX
by CMRI
predictor.
8 CONCLUSIONS
A number of researches have been established to
formulate the PPV and QMAX in the blast-induced
vibrations. All the conventional predictors have site
specific constants and these are not able to predict
the safe charge for even other similar geo-mining
conditions. The predictor equations proposed by
various researchers show good correlation in
calculation of PPV and a low correlation while
calculating maximum safe charge per delay, as these
calculates QMAX by back calculation.
The main aim of this study was to predict
QMAX which is one of the most important factors in
blast pattern designing. ANN method has been
found application on various engineering areas,
particularly where the problem is involved with
complexity and uncertainty. In this study, a three
layer feed forward back propagation neural network
model has been employed to predict the QMAX.
Results were also compared with different available
conventional vibration predictors. The ANN model
predicts QMAX value as an output parameter for a
y=‐0.486x+3292.
R=0.316
0
5000
10000
15000
20000
25000
0 1000 2000 3000 4000 5000
PredictedQ
max
(kg)
MeasuredQ
max
(kg)
LanKihl
y=0.973x+458.2
R=0.505
0
2000
4000
6000
8000
10000
0 1000 2000 3000 4000 5000
PredictedQmax(kg)
MeasuredQmax(kg)
AMHEN
y=‐0.491x+3321.
R=0.316
0
5000
10000
15000
20000
25000
0 1000 2000 3000 4000 5000
PredictedQmax(kg)
MeasuredQmax(kg)
BIS
y=2.397x‐ 536.9
R²=0.752
‐5000
0
5000
10000
15000
20000
0 1000 2000 3000 4000 5000
PredictedQmax(kg)
MeasuredQmax(kg)
CMRI
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
370

given PPV and distance from the blast face. The
comparison shows that results from ANN model are
close to the real ones that are desirable. ANN results
indicate very close agreement for the QMAX with
the field data sets as compared to conventional
predictors.
REFERENCES
Alipour A., Ashtiani M. Fuzzy modeling approaches for
the prediction of maximum charge per delay in surface
mining, Int J Rock Mech Min Sci 2011:48:305–310.
Ambraseys N. R., Hendron A. J., Dynamic behaviour of
rock masses, Rock Mechanics in Engineering
Practices, London: John Wiley and Sons; 1968, p. 203
– 207.
Babuska, R. Fuzzy and neural control. Netherlands, Delft
University of Technology, DISC Course Lecture
Notes, 2004.
Baheer I. (2000) Selection of methodology for modeling
hysteresis behavior of soils using neural networks. J
Comput Aided Civil Infrastruct Eng 5(6):445–63.
Basheer I. A., Hajmeer M. (2000) Artificial neural
networks: fundamentals, computing, design, and
application. J Microbiol Meth; 43:3–31.
Bureau of Indian Standard. Criteria for safety and design
of structures subjected to underground blast. ISI
Bulletin 1973; IS – 6922.
Cheng G., Huang S. L., Analysis of ground vibration
caused by open pit production blast. Explosive and
Blasting Technique, Holmberg (Ed.), Balkema; 2000,
p. 63-70.
Davies B., Farmer I. W., Attewell P. B., Ground vibrations
from shallow sub-surface blasts. The Engineer,
London, 1964; 217, p. 553-559.
Demuth, H., Beale, M. Neural Network Toolbox: For Use
with MATLAB. User’s Guide, Version 4, The
MathWorks Inc, 2008.
Dowding C. H., Blast vibration monitoring and control.
Englewoods Cliffs: Prentice-Hall Inc; 1985, p. 288-
290.
Dreyfus, G. Neural Networks: Methodology and
Applications. (2nd ed.) Paris, Published by Eyrolles,
2004.
Duvall, W. I., Petkof B., Spherical propagation of
explosion generated strain pulses in rock. USBM, RI
5483, 1959; p. 21.
Hagan T. N., Rock breakage by explosives, in: Proc. Nat.
Symp. Rock Fragmentation, Adelaide, 1973. p. 1 – 17.
Haykin S. Neural Networks: A Comprehensive
Foundation. Prentice Hall, 1994.
Hecht-Nielsen R, (1987) Kolmogorov’s mapping neural
network existence theorem, Proceedings of the first
IEEE international conference on neural networks,
San Diego CA, USA, p. 11-14.
Hino K. Fragmentation of rock through blasting. J Ind
Explosive Soc 1956;17(1):1–11.
Hush D. R., (1989) Classification with neural networks: a
performance analysis. Proceedings of the IEEE
international conference on systems Engineering
Dayton Ohia, USA, 277–80.
ISRM. Suggested method for blast vibration monitoring.
Int J Rock Mech Min Sci and Geomech Abst 1992;
29:145-146.
Kanellopoulas I., Wilkinson G. G., (1997) Strategies and
best practice for neural network image classification.
Int J Remote Sensing 18:711–25.
Khandelwal M.. Evaluation and Prediction of Blast
Induced Ground Vibration using Support Vector
Machine, Int J Rock Mech Min Sci 2010:47, 509-516.
Khandelwal M. Singh T. N., Prediction of Blast Induced
Ground Vibration using Artificial Neural Network, Int
J Rock Mech Min Sci 2009,:46, 1214-1222.
Langefors U., Kihlstrom B., The modern technique of rock
blasting. New York: John Wiley and Sons Inc; 1963.
Lippmann R. P., (1987) An introduction to computing
with neural nets. IEEE ASSP Mag; 4:4–22.
MacKay D. J. C., Bayesian interpolation, Neural
Computation 1992; 4:415 – 447.
Masters T., (1994) Practical neural network recipes in
C++. Boston MA: Academic Press.
McKenzie C., Quarry Blast monitoring Technical and
Environmental Perspective, Quarry Management
1990; 23-29.
Pal Roy P., Putting ground vibration predictors into
practice, Colliery Guardian 1993; 241:63-67.
Rafiq M. Y., Bugmann G., Easterbrook D. J., Neural
network design for engineering applications.
Computers & Structures 2001, 79: 1541–1552.
Ripley B. D., (1993) Statistical aspects of neural networks.
In: Barndoff-Neilsen OE, Jensen JL, Kendall WS,
editors. Networks and chaos-statistical and
probabilistic aspects. London: Chapman & Hall, p.
40–123.
Tawadrous A. S., Evaluation of artificial neural networks
as a reliable tool in blast design. International Society
of Explosives Engineers, 2006, 1:1–12.
Wang C. (1994), A theory of generalization in learning
machines with neural application. PhD thesis, The
University of Pennsylvania, USA.
EvaluationofSafeExplosiveChargeinSurfaceMinesusingArtificialNeuralNetwork
371