You are on page 1of 14

Arab J Geosci

DOI 10.1007/s12517-013-1174-0

ORIGINAL PAPER

Blasting-induced flyrock and ground vibration prediction


through an expert artificial neural network based on particle
swarm optimization
D. Jahed Armaghani & M. Hajihassani &
E. Tonnizam Mohamad & A. Marto & S. A. Noorani

Received: 6 April 2013 / Accepted: 4 November 2013


# Saudi Society for Geosciences 2013

Abstract Blasting is a major component of the construction model was selected, and the proposed model was trained and
and mining industries in terms of rock fragmentation and tested using 44 datasets obtained from three granite quarry
concrete demolition. Blast designers are constantly concerned sites in Malaysia. Each dataset involved ten inputs, including
about flyrock and ground vibration induced by blasting as the most influential parameters on flyrock distance and PPV,
adverse and unintended effects of explosive usage on the and two outputs. The results indicate that the proposed method
surrounding areas. In recent years, several researches have is able to predict flyrock distance and PPV induced by blasting
been done to predict flyrock and ground vibration by means with a high degree of accuracy. Sensitivity analysis was also
of conventional backpropagation (BP) artificial neural net- conducted to determine the influence of each parameter on
work (ANN). However, the convergence rate of the flyrock distance and PPV. The results show that the powder
BP-ANN is relatively slow and solutions can be trapped at factor and charge per delay are the most effective parameters
local minima. Since particle swarm optimization (PSO) is a on flyrock distance, whereas sub-drilling and charge per delay
robust global search algorithm, it can be used to improve are the most effective parameters on PPV.
ANNs' performance. In this study, a novel approach of incor-
porating PSO algorithm with ANN has been proposed to Keywords Blasting . Flyrock distance . Ground vibration .
eliminate the limitation of the BP-ANN. This approach was Artificial neural networks . Particle swarm optimization
applied to simulate the flyrock distance and peak particle
velocity (PPV) induced by blasting. PSO parameters and
optimal network architecture were determined using sensitiv-
ity analysis and trial and error method, respectively. Finally, a Introduction

In civil and mining engineering, rock is removed during


construction projects such as tunnels, caverns, road cuts,
D. J. Armaghani : M. Hajihassani (*) : E. T. Mohamad : hydraulic channels, and foundations. Blasting, which is the
A. Marto : S. A. Noorani controlled use of explosives, is considered as the most com-
Faculty of Civil Engineering, Department of Geotechnics and
Transportation, Universiti Teknologi Malaysia (UTM), Skudai,
mon method for rock breakage. Potential environmental im-
81310 Johor Bahru, Malaysia pacts of surface blasting operations include ground vibration,
e-mail: mohsen_hajihassani@yahoo.com airblast, flyrock, dust and fumes, and noise. Flyrock and
D. J. Armaghani ground vibration are the source of the most injuries and
e-mail: danja1364@yahoo.com damage in blasting works. The damage induced by flyrock
E. T. Mohamad and ground vibration in rock blasting has been a serious
e-mail: edy@utm.my problem ever since blasting was introduced (Aghajani
A. Marto Bazzazi et al. 2007). According to Pal Roy (2005), blasting-
e-mail: aminaton@utm.my induced flyrock can travel from a few meters to more than 1,
S. A. Noorani 000 m with poorly controlled blasting. In addition, high
e-mail: s_ahmad_777@yahoo.com ground vibration affects the structural integrity, groundwater,
Arab J Geosci

and ecology of the nearby area (Monjezi et al. 2010). Accurate surrounding areas. Ground vibration is an important blasting
estimation of the safety area boundary is one of the most effect that is extremely difficult to effectively control and may
important parameters in blasting design, especially when be of annoyance to nearby residents. A huge amount of energy
blasting is done adjacent to residential areas. is used in every ground vibration, energy which is supposed to
Artificial neural networks (ANNs) are one of the most be applied to rock fracturing. Different adverse effects, all of
dynamic areas of research in advanced and diverse applica- which basically depend on the intensity of these vibrations,
tions of science and engineering. Backpropagation (BP) is a are the result of these operations. High ground vibration not
common method of training ANNs in order to minimize the only creates problems for the nearby population, but also
objective function. However, it has been proven that BP adversely affects the integrity of the structures, groundwater,
algorithm can easily converge to any local minimum and it and the ecology of the nearby area.
cannot guarantee convergence in learning (Lee et al. 1991; Ground vibration is affected by a number of parameters
Priddy and Keller 2005; Adhikari and Agrawal 2011). To such as blast design, mechanical properties of the rock mass,
overcome these problems, the use of powerful optimization explosive characteristics, distance from the blast face, and
algorithms to optimize ANNs is an advantage. Particle swarm geological conditions (Wiss and Linehan 1978; Khandelwal
optimization (PSO) is a powerful population-based optimiza- and Singh 2005). It is necessary to optimize blasting param-
tion approach for solving continuous and discrete optimiza- eters to reduce ground vibration based on rock mass properties
tion problems. Since PSO is a robust global search algorithm, which include strength, density, longitudinal wave velocity,
it can be used to adjust weights and biases of ANNs in order to impedance, and structural discontinuities (Singh and Sastry
increase the performance of the model. The PSO and ANN 1986). Ground vibration is composed of two components,
employ different approaches in minimizing problems in which PPV and frequency. PPV has been employed as a vibration
the ANN target is to obtain the minimum error between actual index in German DIN Standard 4150 (New 1986; Indian
and predicted values. In this study, ANNs are trained by PSO Standard Institute 1973). It has been proven that PPV is an
algorithm to predict flyrock and peak particle velocity (PPV) important indicator to determine the structure damage
using actual data obtained from three granite quarry sites in criteria (Basu and Sen 2005; Ozer et al. 2011; Karadogan
Malaysia. The obtained results of proposed model were com- et al. 2013).
pared with the empirical approaches. PPV due to ground vibration in surface blasting is a
significant parameter for the prediction of ground vibra-
tion. Basically, PPV depends on two parameters: maxi-
mum charge used per delay as well as distance between
Background blast face and monitoring point. Figure 1 shows the blasting
parameters that influence the magnitude of flyrock and ground
Definitions of terms vibration.

Flyrock refers to the uncontrolled dispersion of rock frag-


ments from blast areas caused by explosive energy (Rezaei
et al. 2011). There is an affective relationship among three
factors, distribution of explosive energy, mechanical strength
of the rock mass, and charge confinement, and any mismatch
between these factors can produce flyrock (Bajpayee et al.
2004). When this happens, much of the explosive energy is
used to throw rock rather than create fragmented rock (Pal
Roy 2005).
In surface blasting, flyrock resulted in numerous cases of
injury and property damage. The main causes of flyrock are
inadequate burden and spacing, inadequate stemming, im-
proper blasthole pattern, inaccurate drilling, overloaded holes,
excessive powder factor, unfavorable geological conditions
(i.e., open joints, weak seams, and cavities), inappropriate
delay timing and sequence, and inaccuracy of delays
(Bhandari 1997; Hemphill 1981; Workman and Calder
1994; Kopp 1994; Kecojevic and Radomsky 2005).
According to Khandelwal and Singh (2009), ground vibra- Fig. 1 Influential parameters on flyrock and ground vibration in blasting
tion is a wave motion, spreading outward from the blast to operation
Arab J Geosci

Table 1 Empirical PPV predictor presented by different researchers

Reference Equation Site constant for granite

USBM by Duvall and Petkof (1959) v =K[R/√Q Max]−B K =179.31, B =1.09


Langefors–Kihlstrom (1963) v =K[√(Q Max/R 2/3)]B K =44.43, B =−1.18
General predictor by Davies et al. (1964) v =KR −B (Q Max)A K =212.27, B =1.09, A =0.52
Bureau of Indian Standard (1973) v =K[(Q Max/R 2/3)]B K =6.33, B =0.22
Ghosh-Daemen predictor (1983) v =K[R/√Q Max]−B e −αR K =780.36, B =1.26, α =0.0004
CMRI by Pal Roy (1993) v =n +K[R/√Q Max]−1 K =168.91, n =1.57

v peak particle velocity (in millimeters per second); Q max maximum charge per delay (in kilogram); R distance between blast face and vibration
monitoring point (in meter); K, B, A, α, and n site constants
USBM United State Bureau of Mines
CRMI Central Mining Research Institute

Current methods to predict blasting-induced flyrock research (Roth 1979; Lundborg 1981; Fletcher and D'Andrea
1986; Adhikari 1999; Rehak et al. 2001; Bajpayee et al. 2004;
Some empirical methods have been suggested by a number Kecojevic and Radomsky 2005). Holmberg and Persson (1976)
of researchers to predict flyrock distance. Lundborg used an empirical approach to estimate flyrock distance using
et al. (1975) have proposed a fundamental empirical equa- high-speed cameras, and predicting flyrock theory in hard rock
tion to predict flyrock distance based on hole and rock diam- such as granite has been developed by Lundborg (1981).
eters as follows: Apart from empirical methods, prediction of flyrock dis-
tance using soft computing techniques has been studied by
Lm ¼ 260Dð2=3Þ ð1Þ many scholars. Monjezi et al. (2011a) employed ANNs to
predict flyrock in blasting operations. They used 192 datasets
in which L m is maximum throw in meters and D is hole to train and evaluate ANN models. Their results showed the
diameter in inches. This formula has become the basis of other high performance of ANN model for predicting flyrock. They

Fig. 2 Location of investigated sites in Johor, Malaysia


Arab J Geosci

Table 2 Description of case studies blasting sites

Site name Distance from Latitude Longitude Rock type RQD Bench Explosive
Johor (km) altitude (m) type

Ulu Tiram 18 1°36′41″ N 103°49′20″ E Granite 65–90 10–15 ANFO


Masai 25 1.5°29′42.16″ N 103°52′27.79″ E Granite 60–95 5–30 ANFO
Pengerang 62 1°22′58″ N 104°7′58″ E Granite 70–90 10–30 ANFO

RQD rock quality designation


ANFO ammonium nitrate and fuel oil

demonstrated that in ascending order, blast ability index, Rezaei et al. (2011) developed a fuzzy model to predict
charge per delay, hole diameter, stemming length, and powder flyrock in the Gol-E-Gohar iron mine of Iran and compared
factor are the most effective parameters on flyrock distance. the fuzzy model results with conventional statistical methods.
In order to estimate flyrock distance, burden, spacing, hole
depth, specific drilling, stemming length, charge per delay,
rock density, and powder factor were used as input parame-
ters. Their findings indicated that the efficiency of the devel-
oped fuzzy model is much better than the statistical models.
They demonstrated that the powder factor and stemming
length are the most effective parameters on flyrock distance,
whereas rock density is the least effective parameter.
Ghasemi et al. (2012) developed two predictive models
based on ANN and fuzzy logic to predict flyrock distance. In
their study, burden, spacing, stemming, hole length, powder
factor, and charge per delay were considered as input parame-
ters to estimate flyrock distance. The results showed that both
models were able to predict flyrock distance, whereas the fuzzy
model exhibited higher performance compared to the ANN
model. Monjezi et al. (2012a) developed a new neurogenetic
model for predicting flyrock and backbreak. They showed that
the most influential parameters on flyrock are stemming and
powder factor, whereas for backbreak, stemming and charge
per delay are the most effective parameters.

Table 3 Input and output parameters and range of them used in hybrid
PSO-based ANN model

Variable category Parameter Unit Symbol Min Max

Input Hole diameter cm A 8.9 15


Hole depth m B 5 30
Charge per delay kg C 36.5 268
Spacing m D 2.65 3.9
Burden m E 1.9 3.15
Stemming m F 1.7 3.6
Powder factor kg/m3 G 0.67 1.05
Rock density g/cm3 H 2.3 2.8
Sub-drilling cm I 24 50
Number of rows – J 3 8
Output Flyrock distance m K 60 405
PPV mm/s L 0.07 5.9
Fig. 3 Flowchart depicting the general PSO algorithm
Arab J Geosci

Fig. 4 Correlation of
determination for different swarm
size in the hybrid models

Current methods to predict blasting-induced ground vibration distances from a blast face were considered as input parame-
ters. They demonstrated that ANN is a more accurate ap-
A number of empirical vibration predictors have been pro- proach compared to regression analysis to predict ground
posed by different researchers for prediction of PPV such as vibration. Khandelwal and Singh (2007) investigated four
Duvall and Petkof (1959), Langefors and Kihlstrom (1963), widely used empirical predictors to estimate the PPV for 150
Davies et al. (1964), Bureau of Indian Standard (1973), Ghosh blast datasets, and computed results were compared with
and Daemen (1983), and Pal Roy (1993). However, different actual field data. Subsequently, they developed an ANN with
PPV values for the same excavation site can be found by these two inputs (maximum charge per delay and distance from
predictors and there is no uniformity in the predicted results. blasting face) and one output (PPV). They indicated that
In addition, there are several parameters which influence ANN results were more accurate compared to empirical pre-
ground vibration such as geological and geotechnical condi- dictors. Iphar et al. (2008) utilized two different methods
tions and blast geometry. It is worth noting that explosive including simple regression and the adaptive neuro fuzzy
parameters have not been considered in any available predic- inference system (ANFIS) to predict PPV induced by blasting.
tors. Table 1 illustrates the current PPV predictor equations They used 44 PPV values obtained from blasting operations in
proposed by different researchers. Turkey. Their results showed that the ANFIS model yields
In addition to the empirical approaches, soft computing better results in comparison to regression analysis.
techniques have been widely used by several researchers to Khandelwal and Singh (2009) used ANN and multivariate
predict ground vibration. Singh and Singh (2005) utilized regression analysis (MVRA) techniques to predict blast-
ANN and regression analysis to predict ground vibration induced ground vibration and frequency by incorporating rock
induced by blasting. In their study, hole diameter, number of properties, blast design, and explosive parameters. A total of
holes, hole depth, burden, spacing, radial, and horizontal 174 blast vibration records were used to predict PPV and

Fig. 5 Mean square error for


different swarm sizes in the
hybrid models
Arab J Geosci

Fig. 6 Correlation of
determination for different values
of C 2/C 1 in the hybrid models

frequency with ten input parameters. ANN results indicated a weight and distance from blasting face were considered as
closer agreement for the PPV and frequency with the field input parameters to predict PPV. They concluded that the
datasets as compared to MVRA prediction. An ANN model predicted PPVs obtained from fuzzy model were much closer
with four input parameters including maximum charge per to the measured values in comparison to those predicted by
delay, distance from blasting face to the monitoring point, statistical model.
stemming, and hole depth was developed by Monjezi et al. Various parameters affect the flyrock and ground vibration
(2011b) to predict PPV. A database consisting of 182 datasets with complicated relationships between them. Empirical
was collected at different strategic and vulnerable locations methods do not consider all these influential parameters on
around the Kandovan tunnel in Iran. They demonstrated that flyrock and ground vibration. Therefore, empirical methods
ANNs are applicable tools to predict blast-induced ground may not be completely suitable for such problems.
vibration. In addition, from the sensitivity analysis, they found
that the distance from the blasting face is the most effective
and stemming is the least effective parameter on the PPV. Case studies
Fisne et al. (2011) utilized fuzzy logic approach and classical
regression analysis to predict PPV using 33 datasets obtained A total of 44 blasting operations have been investigated from
from Akdaglar quarry in Turkey. In their research, charge three granite quarry sites in Malaysia. These sites are located

Fig. 7 Mean square error for


different values of C 2/C 1 in the
hybrid models
Arab J Geosci

near Johor City, the capital of the Johor State, as shown in reduced. This process is repeated until the error is converged
Fig. 2. The weathering grade of the rock in these sites is to a level defined by a cost function such as mean square error
mainly classified into grade II to IV, and the rock strength is (MSE) or root mean squared error (Simpson 1990; Kosko
between 30 and 150 MPa. Granite quarry in the mentioned 1994; Singh et al. 2004). However, an experimental database
sites is blasted using 75-, 115-, and 150-mm-diameter blast including sufficient number of datasets is required to train the
holes. ANFO is used as the main explosive material and ANN models.
dynamite is utilized for initiation. Fine gravels are used as
the stemming material. Description of the case studies blasting
sites is shown in Table 2.
Data collection was conducted over 4 months from Particle swarm optimization
February 2012 to June 2012. During data collection, blasting
parameters including burden, spacing, stemming, hole depth, PSO is a population-based computational method motivated
hole diameter, powder factor, charge per delay, rock density, by the social behavior of organisms such as bird flocking and
and sub-drilling were carefully recorded. In order to measure fish schooling that optimizes a problem by iteratively trying to
flyrock distance, the bench surface was colored and three improve a candidate solution with regards to a given measure
video cameras were placed to measure the projection of of quality. This method was originally developed by Kennedy
flyrock. Following the blasting, the events recorded by the and Eberhart (1995) and Shi and Eberhart (1998). PSO opti-
cameras were reviewed and the locations of the flown rocks mizes a problem by having a population of candidate solutions
were found. Using a measuring tape, the flyrock distances identified as particles. These particles move around in the
were measured. In each blasting operation, PPV was mea- search space according to simple mathematical formula over
sured using a seismograph that was placed near the residential the particles' position and velocity. Each particle movement is
buildings (1,000 m). More than 300 residential buildings exist influenced by its local best known position, called personal
adjacent to the mentioned blasting sites, with the nearest best (p best), and is also guided towards the best known posi-
building located approximately 1,000 m from the blasting tions in the search space, called global best (g best), which is
face. updated as better positions are found by other particles. This is
expected to move the swarm towards the best solutions.
The PSO procedure starts with producing the first particle
Artificial neural networks swarm and its parameters initialize randomly. Consequently,
the initial particle positions are determined. The velocity of a
ANNs are mathematical models inspired by biological
neural networks. As a processing system, ANNs process
the information using simple interconnected elements Table 4 Finding the optimal architecture using trial and error method
known as neurons which are placed in distinct layers of
the network. As the best type of ANNs, multilayer perceptron Model Network architecture Training Testing
(MLP) consists of at least three layers: input layer, output
Hidden Nodes in R2 MSE R2 MSE
layer, and intermediate or hidden layer(s) (Monjezi et al. layers hidden layers
2012b). Each layer consists of one or more nodes. The lines
between the nodes indicate the flow of information from one 1 1 10 0.92 0.050 0.90 0.103
node to the next. 2 1 15 0.94 0.049 0.93 0.037
Prior to interpreting new information, the ANN model has 3 1 20 0.91 0.067 0.91 0.063
to be trained. Various learning algorithms have been suggested 4 1 25 0.93 0.050 0.87 0.124
for training of neural networks, but the BP is the most versatile 5 1 30 0.88 0.098 0.81 0.138
and robust technique and provides the most efficient learning 6 2 10 0.89 0.087 0.85 0.130
procedure for MLP networks (Tawadrous and Katsabanis 7 2 15 0.90 0.074 0.82 0.124
2007; Rezaei et al. 2012). In this technique, the strengths or 8 2 20 0.80 0.154 0.81 0.165
weights of the interneuron connections are adjusted according 9 2 25 0.82 0.130 0.70 0.242
to the difference between the predicted and actual outputs. In a 10 2 30 0.72 0.232 0.77 0.265
feed-forward BP algorithm, the signals flow from input layer 11 3 10 0.89 0.082 0.68 0.308
to the output layer, called forward pass, then the output is 12 3 15 0.87 0.102 0.77 0.156
compared to the actual measured values and the error is 13 3 20 0.89 0.080 0.50 0.368
calculated. The obtained error is propagated back through 14 3 25 0.89 0.082 0.57 0.315
the network to update the individual weights, called backward 15 3 30 0.91 0.092 0.73 0.246
pass. In this process, both the training and the testing errors are
Arab J Geosci

particle is a vector from its former position to the current in which v! ! ! !
new , v , pnew , and p are new velocity, current
position, while the initial velocity of all particles is zero. The velocity, new position and current position, respectively.
next step is generating initial swarm followed by calculating C 1 and C 2 are pre-defined coefficients, p! best is personal
the fitness values for all particles. Fitness function is created best particle position and g!
best is global best position among
and used for this reason. PSO utilizes a simple concept which all particles. Figure 3 illustrates the workflow of PSO
consists of updating movement velocity and particle position. algorithm.
Based on Eqs. (2) and (3), position and velocity of each In training ANNs using PSO, minimization of MSE is
particle are computed and the procedure is continued until considered as the goal function that is used instead of fitness
the termination criteria are met. function in minimization of the conversional problems
    using PSO. The main target of training ANNs using
v! ! ! ! ! !
new ¼ v þ C 1 pbest − p þ C 2 g best − p ð2Þ PSO is to achieve a set of weights and biases that minimize
the MSE. In training ANNs using PSO, the position of each
particle in the swarm corresponds to a set of weights and
p! ! !
new ¼ p þ vnew ð3Þ biases.

Fig. 8 Structure of the selected


hybrid PSO-based ANN model
Arab J Geosci

Ground vibration and flyrock prediction using hybrid


ANN-PSO

In order to apply the PSO-based ANN to predict flyrock


distance and PPV induced by blasting, three granite quarry
sites in Malaysia were investigated. To train and verify the
ability and accuracy of hybrid PSO-based ANN, 44 datasets
were collected from these sites. Each dataset involves ten
inputs, including the most influential parameters on ground
vibration and flyrock as discussed in the background. Apart
from the aforementioned parameters, the distance between
blasting face and the monitoring point is always an effective
parameter on ground vibration measurements. However,
since in the case studies, this distance has a constant value
of 1,000 m, the effect of this parameter was neglected in the
models. In addition, each dataset includes two outputs: values
of flyrock distance and PPVs that were measured during
blasting at mentioned sites. In order to evaluate network
performance, 20 % of datasets were randomly chosen for
testing. Testing datasets have no role in the learning process.
Fig. 9 Correlation of determination for training data in hybrid PSO-ANN
Table 3 shows the investigated parameters and the ranges that model
were used to predict flyrock distance and PPV.
The PSO-based ANN modelling procedure starts with
determining the network parameters. Achieving a set of After significant increases in the network performance for
weights and biases that minimizes the MSE is the objective the swarm size of 300, no significant changes were seen in
of the function. In PSO-based ANN training, the position of the values of R 2 and MSE. Therefore, the candidate value for
each particle in a swarm corresponds to a set of weights and the swarm size is 300.
biases. In other words, the dimension of each particle is equal The next series of sensitivity analysis were conducted by
to the number of weights and biases. In order to minimize setting a fixed value of 300 for swarm size and different values
MSE, the particles move within the weights and bias space, of coefficients C 1 and C 2. The aim of these analyses is to find
updating their positions in each iteration. The overall best
error is known as g best and each particle best error is consid-
ered as p best. New error is calculated using new weights and
biases until satisfactory error is attained. The obtained weights
and biases are used for prediction.
To achieve the desired results, optimized parameters of
hybrid models have to be determined. It should be mentioned
that PSO can only adjust the weights and biases of the network
and the network architecture including the number of hidden
layers, and the number of nodes in each hidden layer should
be determined using trial and error. In order to determine the
optimum PSO parameters including number of particles in the
swarm and coefficients of velocity equation (C 1 and C 2), a
series of sensitivity analysis was conducted on a hybrid PSO-
based ANN model which includes 1 hidden layer, 10 nodes in
hidden layer, and iteration number of 1,000.
The first series of analysis was conducted by setting values
of 1 for both coefficients C 1 and C 2 to determine the optimum
number of particles in the swarm. The results of the analysis
are shown in Figs. 4 and 5. According to the figures, subse-
quent to the increase in swarm size, the correlations increased
and MSEs decreased. It should be mentioned that running Fig. 10 Correlation of determination for testing data in hybrid PSO-
time gradually increased due to swarm size increments. ANN model
Arab J Geosci

Table 5 Investigated parameters to estimate PPV and flyrock by empirical approaches and proposed PSO-based ANN model

Parameter Unit Datasets

1 2 3 4 5 6 7 8 9

Hole diameter cm 15 8.9 11.5 15 11.5 15 8.9 15 11.5


Hole depth m 5 11 25 25 14 25 10 30 20
Charge per delay kg 53 61 269 157 116 144 58 255 185
Spacing m 3 3.6 3.85 2.65 3.3 2.7 3.7 3 3.9
Burden m 2 3 2.95 1.9 2.5 2 3.1 2 2.9
Stemming m 1.9 2.4 3 1.7 2 1.75 2.5 1.8 2.5
Powder factor kg/m3 0.73 1.03 0.84 0.94 1.05 0.85 1.01 0.83 0.72
Rock density g/cm3 2.65 2.8 2.7 2.75 2.6 2.75 2.7 2.4 2.6
Sub-drilling cm 28 44 42 30 35 28 42 34 37
Number of rows – 5 8 4 4 3 7 5 7 8
Flied rock diameter m 0.035 0.049 0.058 0.045 0.04 0.05 0.055 0.038 0.035
Distancea m 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
a
Distance between seismograph and blasting face

the values of coefficients C 1 and C 2 which minimize the previous analysis including 300 particles and values of 1 and
MSE and maximize the correlation between actual and 1.5 for C 1 and C 2 coefficients, respectively. Through these
predictive values. The results of the analysis are shown in processes, the network that yields the best results was selected
Figs. 6 and 7. According to the figures, when the C 2 per C 1 as the optimal network. The results of the analysis are shown
is 1.5, the maximum correlation and minimum error were in Table 4.
obtained for training and testing datasets. Therefore, values As can be seen in the table, the model with 1 hidden layer
of 1 and 1.5 were selected for the C 1 and C 2 coefficients, and 15 nodes in hidden layer presents reasonable correla-
respectively. tions, 94 % for training and 93 % for testing. Finally, a PSO-
With the aim of obtaining optimum network architecture, based ANN model including 1 hidden layer, 15 nodes in
following the trial and error method, 15 hybrid network hidden layer, 1 and 1.5 for C 1 and C 2 coefficients, respec-
models were proposed. The processes were repeated with tively, and a swarm size of 300 was selected. Figure 8
different amounts of hidden layers in the models and different shows the structure of the selected hybrid PSO-based
amounts of nodes in each hidden layer. All models were ANN model to predict flyrock distance and PPV induced
trained with the optimized PSO parameters obtained in by blasting.

Table 6 PPV and flyrock distance predicted by empirical approaches and proposed PSO-based ANN model

Method Datasets

1 2 3 4 5 6 7 8 9

Ground vibration (mm/s) USBM 0.84 0.90 2.02 1.51 1.28 1.44 0.88 1.97 1.65
Langefors–Kihlstrom 62.90 57.90 24.10 34.04 40.70 35.83 61.27 25.57 30.91
General predictor 0.87 0.94 2.02 1.58 1.35 1.51 0.94 2.03 1.72
Bureau of Indian Standard 5.58 5.67 7.76 7.00 6.54 6.85 5.61 7.77 7.24
Ghosh–Daemen predictor 2.37 2.59 6.60 2.10 1.73 1.98 1.12 2.85 2.32
CMRI 2.89 2.88 4.34 3.68 3.39 3.59 2.85 4.26 3.86
PSO-based ANN 0.86 1.67 1.53 3.06 3.53 2.41 1.09 2.88 1.49
Measured value 0.75 1.39 1.65 2.81 3.30 2.70 0.92 3.20 1.55
Flyrock (m) Lundborg predictor 92 127 151 116 104 95 106 95 122
PSO-based ANN 187 117 355 317 281 289 127 275 378
Measured value 210 110 405 350 310 320 120 300 395
Arab J Geosci

Fig. 11 Comparison between


obtained PPV values by different
predictors and measured values

Results and discussion ANN model is highly able to predict flyrock distance, whereas
reasonable error exists between the results obtained by
A MATLAB code was provided to predict flyrock distance Lundborg method and actual values, as can be seen in Table 6.
and PPV using obtained data from three granite quarry sites in According to Table 6, in the case of PPV prediction, the
Malaysia. PSO algorithm was used to weight ANN instead of results of two empirical predictors including USBM and
BP to increase the accuracy and applicability of ANN for the General predictor as well as results obtained by proposed
prediction of flyrock distance and PPV. A number of analyses PSO-based ANN model are proximate to measured values.
were conducted to determine the optimum model by compar- The comparison between obtained PPV values by different
ing the correlation of linear regression between actual data and predictors and measured values is shown in Fig. 11. It should
predicted values obtained by models. Finally, an optimized be mentioned that the predicted values by Langefors–
model with 1 hidden layer and 15 nodes in hidden layer was Kihlstrom predictor (Table 6) were omitted from the figure,
selected. Figures 9 and 10 illustrate the linear regression due to high level of error. As it can be seen in the figure,
between measured data obtained from granite quarry sites predicted PPVs using proposed PSO-based ANN model is
and predicted values obtained by the proposed PSO-based closer to the measured PPVs compared to the obtained PPVs
ANN model. According to the figures, the network performed using USBM and General predictor. Table 7 shows the range
with negligible error and noticeably desirable regression with of error obtained by different predictors. The overall error
correlation of 94 % for training data and 93 % for testing obtained by USBM, General predictor, and proposed PSO-
datasets. based ANN model for nine datasets are 30.31, 29.72, and
In order to evaluate the validity and accuracy of the pro- 10.71 %, respectively. The results indicate that the proposed
posed model, 20 % of the datasets (nine datasets) were select-
ed. Table 5 shows the investigated parameters to estimate PPV
and flyrock by empirical approaches and proposed PSO-based Table 7 The range of error obtained by different predictors
ANN model. Some of these parameters were used in proposed
PSO-based ANN model (according to Table 3) and some Method Error (%)
others were used to predict by empirical approaches. The
Minimum Maximum Overall
results obtained by the proposed model, empirical methods,
and actual data were compared. The predicted results are USBM 4.35 61.21 30.31
tabulated in Table 6 and compared with the measured flyrock General predictor 2.17 59.09 29.72
distance and PPV obtained from blasting operations. It is Bureau of Indian Standard 98.18 509.78 304.77
worth mentioning that the results of empirical formulas were Ghosh–Daemen predictor 10.93 300.01 87.13
obtained by using the site constants for granite, suggested by CMRI 2.73 209.78 112.65
Pal Roy (2005). PSO-based ANN 3.87 20.14 10.71
The comparison between obtained flyrock distance by
proposed PSO-based ANN, Lundborg predictor, and mea- USBM United State Bureau of Mines
sured values demonstrates that the proposed PSO-based CRMI Central Mining Research Institute
Arab J Geosci

Fig. 12 Strengths of the relation


(r ij ) between outputs and input
parameters

PSO-based ANN model is an applicable tool to predict flyrock Figure 12 shows the strengths of the relations (r ij values)
distance and PPV induced by blasting with a high degree of between input and outputs (flyrock distance and PPV) param-
accuracy. eters. Sensitivity analysis results show that powder factor (G)
and charge per delay (C) are the most influential parameters
on flyrock distance, whereas sub-drilling (I) and charge per
Sensitivity analysis delay (C) are the most effective parameters on PPV.

To identify the relative influence of the each input parameter


on the flyrock distance and PPV, sensitivity analysis was Conclusion
performed. For this purpose, the cosine amplitude method
introduced by Yang and Zhang (1997) was utilized. To apply A hybrid PSO-based ANN model has been presented to
this method, all of the data pairs were expressed in common predict flyrock distance and PPV induced by blasting. ANN
X-space. The data pairs used to construct a data array X are has been modified to increase the accuracy and applicability
defined as: of prediction by means of replacing the BP with PSO algo-
rithm. A series of sensitivity analysis was conducted to deter-
X ¼ fx1 ; x2 ; x3 …; xi ; …; xn g mine the optimum parameters of the PSO algorithm. The
optimum architecture of the PSO-based ANN model was
The elements x i in the array X is a vector of lengths of m obtained using trial and error method. Finally, a model with
that is: 1 hidden layer, 15 nodes in hidden layer, and a swarm size of
300 was selected. To generate the proposed hybrid PSO-based
x1 ¼ fxi1 ; xi2 ; xi3 …; xim g ANN model, 44 datasets were collected from three granite
quarry sites in Malaysia. The results show the applicability of
Each of the data pairs can be thought of as a point in m- the proposed model to predict flyrock distance and PPV. It is
dimensional space, where each point requires m-coordinates worth noting that similar to any ANN model, the results of the
for a full description. As such, all of the points in the space, hybrid model are usually reasonable in the range of the data
pair wise, have a relation with the obtained results. The used in the training step. In order to evaluate the applicability
strength of the relation (r ij ) between the dataset X i and X j is of the presented model, nine datasets were selected and the
represented by the following equation: results obtained by presented model and empirical methods
were compared to measured data. It can be concluded that
Xm this network is able to be used as an accurate and applicable
x x
k¼1 ik jk approach to estimate the flyrock distance and PPV induced
rij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xm Xm ð4Þ
by blasting. By sensitivity analysis, it was also found
k¼1
x 2
ik k¼1 jk
x2
that the powder factor and charge per delay are the
Arab J Geosci

most effective parameters on the flyrock distance, whereas Khandelwal M, Singh TN (2007) Evaluation of blast-induced ground
vibration predictors. Soil Dyn Earthq Eng 27:116–125
sub-drilling and charge per delay are the most effective pa-
Khandelwal M, Singh TN (2009) Prediction of blast-induced ground
rameters on the PPV. vibration using artificial neural network. Int J Rock Mech Min 46:
1214–1222
Koop JW (1994) Observation of flyrock of several mines and quarries.
Proceeding of 20th Conference on Explosives and Blasting
Technique Austin Texas, 30 January–3 February. Cleveland
International Society of Explosives Engineers, 75–81
References Kosko B (1994) Neural networks and fuzzy systems: a dynamical sys-
tems approach to machine intelligence. Prentice Hall, New Delhi
Langefors U, Kihlstrom B (1963) The modern technique of rock blasting.
Adhikari GR (1999) Studies on flyrock at limestone quarries. Rock Mech Wiley, New York
Rock Eng 32(4):291–301 Lee Y, Oh SH, Kim MW (1991) The effect of initial weights on premature
Adhikari R, Agrawal RK (2011) Effectiveness of PSO based neural saturation in back-propagation learning. Proceedings of the
network for seasonal time series forecasting. Indian International International Joint Conference on Neural Networks (IJCNN '91),
Conference on Artificial Intelligence (IICAI). Tumkur, India, pp pp 765–770
232–244 Lundborg N, Persson PA, Ladegaard-Pedersen A, Holmberg R (1975)
Aghajani Bazzazi A, Osanloo M, Azimi Y (2007) Flyrock prediction by Keeping the lid on flyrock in opencast blasting. Eng Min J 95–100
multiple regression analysis in Esfordi phosphate mine of Iran. Rock Lundborg N (1981) The probability of flyrock damages. Swedish Detoni
fragmentation by blasting. London Taylor and Francis Group, 649–657 Research Foundation Stockholm DS 5:39
Bajpayee TS, Rehak TR, Mowrey GL, Ingram DK (2004) Blasting Monjezi M, Bahrami A, Yazdani Varjani A (2010) Simultaneous predic-
injuries in surface mining with emphasis on flyrock and blast area tion of fragmentation and flyrock in blasting operation using artifi-
security. J Safety Res 35(1):47–57 cial neural network. Int J Rock Mech Min 47:476–480
Basu D, Sen M (2005) Blast induced ground vibration norms—a critical Monjezi M, Bahrami A, Yazdian Varjani A, Sayadi AR (2011a)
review. In: National Seminar on Policies Statutes and Legislation in Prediction and controlling of flyrock in blasting operation using
Mines Kharagpur India, 30–31 July 2005, 112–113 artificial neural network. Arab J Geosci 4:421–425. doi:10.1007/
Bhandari S (1997) Engineering rock blasting operations. Taylor & s12517-009-0091-8
Francis, Boca Raton Monjezi M, Ghafuri kalajahi M, Bahrami A (2011b) Prediction of blast-
Bureau of Indian Standard (1973) Criteria for safety and design of induced ground vibration using artificial neural networks. Tunn
structures subjected to underground blast ISI Bull IS-6922. Uudergr Sp Tech 26:46–50
Davies B, Farmer IW, Attewell PB (1964) Ground vibrations from Monjezi M, Amini Khoshalan H, Yazdian Varjani A (2012a) Prediction
shallow sub-surface blasts. The Engineer London 553–559 of flyrock and backbreak in open pit blasting operation: a
Duvall WI, Petkof B (1959) Spherical propagation of explosion generated neurogenetic approach. Arab J Geosci 5(3):441-448. doi:10.1007/
strain pulses in rock. USBM Report of Investigation s12517-010-0185-3
Fisne A, Kuzu C, Hudaverdi T (2011) Prediction of environmental Monjezi M, Mohamadi HA, Barati B, Khandelwal M (2012b)
impacts of quarry blasting operation using fuzzy logic. Environ Application of soft computing in predicting rock fragmentation to
Monit Assess 174:461–470 reduce environmental blasting side effects. Arab J Geosci doi:10.
Fletcher LR, D’ Andrea DV (1986) Control of flyrock in blasting. 1007/s12517-012-0770-8
Proceedings of 12th Annual Conference on Explosives and blasting New BM (1986) Ground vibration caused by civil engineering works.
technique Atlanta Georgia, 9–14 February. Cleveland, International Transport and Road Research Laboratory Research Report, 53, 19
Society of Explosives Engineers 167–175 Ozer U, Karadogan A, Kahriman A, Aksoy M (2011) Bench blasting
Ghasemi E, Amini H, Ataei M, Khalokakaei R (2012) Application of design based on site-specific attenuation formula in a quarry. Arab J
artificial intelligence techniques for predicting the flyrock distance Geosci 6:711–721. doi:10.1007/s12517-011-0388-2
caused by blasting operation. Arab J Geosci doi:10.1007/s12517- Pal Roy P (1993) Putting ground vibration predictors into practice.
012-0703-6 Colliery Guardian 241:63–67
Ghosh A, Daemen JK (1983) A simple new blast vibration predictor. Pal Roy P (2005) Rock blasting effects and operations. Taylor & Francis,
Proceedings of the 24th US symposium on rock mechanics, College Boca Raton
Station, Texas 151–161 Priddy KL, Keller PE (2005) Artificial neural networks: an introduction.
Hemphill GB (1981) Blasting operations. McGraw-Hill, New York SPIE, Bellingham
Holmberg R, Persson G (1976) The effect of stemming on the distance of Rehak TR, Bajpayee TS, Mowrey GL, Ingram DK (2001) Flyrock
throw of flyrock in connection with hole diameters. Swedish issues in blasting. Proceeding of 27th Annual Conference on
Detonic Research Foundation Report DS 1, Stockholm Explosives and Blasting Technique Orlando Florida, 28–31
Iphar M, Yavuz M, Ak H (2008) Prediction of ground vibrations resulting January. Cleveland, OH, USA. International Society of Explosives
from the blasting operations in an open-pit mine by adaptive neuro- Engineers, 165–175
fuzzy inference system. Environ Geol 56:97–107 Rezaei M, Monjezi M, Yazdian Varjani A (2011) Development of a fuzzy
Karadogan A, Kahriman A, Ozer U (2013) A new damage criteria norm model to predict flyrock in surface mining. Safety Sci 49:298–305
for blast-induced ground vibrations in Turkey. Arab J Geosci doi:10. Rezaei M, Monjezi M, Moghaddam SG, Farzaneh F (2012) Burden
1007/s12517-013-0830-8 prediction in blasting operation using rock geomechanical proper-
Kecojevic V, Radomsky M (2005) Flyrock phenomena and area security ties. Arab J Geosci 5:1031–1037
in blasting-related accidents. Safety Sci 43:739–750 Roth J.A. (1979) A model for the determination of flyrock range as a
Kennedy J, Eberhart R (1995) Particle swarm optimization. Proceeding of function of shot condition. US Department of Commerce NTIS rep
IEEE International Conference on Neural Networks 1942–1948 no PB81222358, p 61
Khandelwal M, Singh TN (2005) Prediction of blast induced ground Shi Y, Eberhart RC (1998) A modified particle swarm optimizer.
vibration and frequency in opencast mine: a neural network ap- Proceedings of IEEE International Congress on Evolutionary
proach. J Sound Vib 289:711–725 Computation 69–73
Arab J Geosci

Simpson PK (1990) Artificial neural system—foundation, paradigm ap- stability using artificial neural networks. Int J Numer Anal Met
plication and implementations. Pergamon, New York 31(7):917–931
Singh DP, Sastry VR (1986) Rock fragmentation by blasting Wiss JF, Linehan PW (1978) Control of vibration and air noise from
influence of joint filling material. Journal of Explosive surface coal mines. US Bureau of Mines Report OFR 103(3):79
Engineering 18–27 Workman JL, Calder PN (1994) Flyrock prediction and control in
Singh TN, Kanchan R, Saigal K, Verma AK (2004) Prediction of P-wave surface mine blasting. Proceeding of 20th Conference on
velocity and anisotropic properties of rock using artificial neural Explosives and Blasting Technique. Austin, Texas, 30 January–
networks technique. J Sci Ind Res India 63:32–38 3 February, Cleveland, OH: International Society of Explosives
Singh TN, Singh V (2005) An intelligent approach to prediction and Engineers, 59–74
control ground vibration in mines. Geotech Geol Eng 23:249–262 Yang Y, Zang O (1997) A hierarchical analysis for rock engineering using
Tawadrous AS, Katsabanis PD (2007) Prediction of surface crown pillar artificial neural networks. Rock Mech Rock Eng 30:207–222

You might also like