Natural Resources Research, Vol. 29, No.

2, April 2020 (Ó 2019)

https://doi.org/10.1007/s11053-019-09593-3

Original Paper

Prediction of Blast-Induced Rock Movement During Bench

Blasting: Use of Gray Wolf Optimizer and Support Vector
Regression

Zhi Yu ,1 Xiuzhi Shi,1,4 Jian Zhou,1 Xin Chen ,1,4 Xiaohu Miao,2 Bing Teng,2
and Timo Ipangelwa3

Received 3 March 2019; accepted 12 November 2019

Published online: 21 November 2019

A large ore loss and dilution can be expected when using a pre-blast ore boundary for shovel
guidance because of the movement and re-distribution of ore in the muck pile under the
action of explosive energy. Considering the difficulties in collecting measurements at the
post-blast ore boundary, the distribution law and prediction of blast-induced rock movement
were studied in this paper. With the application of a blast movement monitoring system,
which can obtain the most accurate data to determine the ore boundary after the blast, four
blast movement trials were carried out to collect data. Then, statistical analysis, an artificial
neural network (ANN) model, a random forest (RF) model and a gray wolf optimizer
algorithm–support vector regression (GWO-SVR) model were used to analyze the database.
The results of the statistical analysis show that the horizontal, vertical and 3D movements
first increase and then decrease, with the maximum displacement occurring near the top of
the charging section. Furthermore, the horizontal movement exhibited a good linear rela-
tionship with the 3D movement, indicating that the horizontal movement can be considered
instead of the 3D movement to facilitate shovel guidance for reducing ore loss and dilution.
The results also indicate that the GWO-SVR model is more accurate than the ANN and RF
models and that the blast-induced rock movement can be controlled by increasing the
burden 9 spacing and reducing the power factor variables during the mining process.
KEY WORDS: Blast-induced rock movement, Dilution, Prediction, Support vector regression, Gray
wolf optimizer algorithm.

INTRODUCTION

The blast-induced rock movement effect refers

to a rock mass breaking under the action of explo-
1
School of Resources and Safety Engineering, Central South
sive energy, moving toward the free face and then
University (CSU), Yuelu District, Changsha 410083, Hunan, redistributing. The impact of this effect on ore loss
China. and dilution is small when the size of the ore body is
2
Uranium Resource Company Limited, China General Nuclear approximately equal to the size of the blast block.
Power Corporation, Beijing 100029, China. However, for an ore body with characteristics of a
3
Swakop Uranium Proprietary Limited, Swakopmund 999122,
Namibia.
small-scale, thickness and serious traversal by waste,
4
To whom correspondence should be addressed; e-mail: bao- the boundary between the pre-blast and post-blast
[email protected], [email protected] varies greatly. Most mines use the pre-blast bound-

843
1520-7439/20/0400-0843/0 Ó 2019 International Association for Mathematical Geosciences
844 Z. Yu et al.

ary to guide the shovel, as that the blast-induced In contrast, many researchers have used, with
rock movement effect has not been studied suffi- great success, various techniques for measuring
ciently; however, this approach often causes con- blast-induced rock movement. However, all of the
siderable ore loss and dilution and seriously impacts existing measurement techniques are based on the
the economy effect of a mine. According to Gilbride single principle of the technician placing an object at
(1995), Harris (1997), and Taylor and Firth (2003), reference points and interpreting the ore boundary
mineral recovery of an individual blast may be in- by using the position of the reference points before
creased by as much as 25% with only a modest cost and after the blast. For example, Zhang (1994),
increase by accounting for the blast-induced rock Gilbride (1995) and Taylor (1995) attempted to use
movement. Therefore, the study of blast-induced sandbags as reference points to indicate the blast-
rock movement is urgently needed. induced rock movement. Compared to other meth-
Methods for blast-induced rock movement re- ods, using sandbags to measure blast-induced rock
search used worldwide can be divided into theoret- movement is simple, inexpensive and practical, but
ical models, dynamic distinct element simulations approximately 50% of the sandbags will be lost, and
and field test approaches. Theoretical models in- it may take several days for technicians to find
clude the kinematic model of Yang et al. (1989) and sandbags in the muck pile. In this regard, Gilbride
Yang and Kavetsky (1990) and the calculation (1995), Harris et al. (1999, 2001), and Taylor and
model of Lucas and Nies (1990). In addition, four Firth (2003) tested, in some mines in the USA, the
computer codes based on Yang and KavetskyÕs function of a device comprised of a magnetic marker
model were developed: 3DMuck, SABREX, and a cesium vapor gradiometer. While this tech-
BLASTPA and ROCKFRAG. However, the nique is more accurate than using sandbags, the
mathematical model of Yang and Kavetsky (1990) magnetic marker will cause disturbances among
has been verified by only limited data from one case other markers. Therefore, an adequate horizontal
study, and thus, more data are needed to test its distance between neighboring markers is required to
accuracy. The calculation model of Lucas and Nies guarantee data accuracy.
(1990) aims at blast fragmentation and ore grade Meanwhile, Thornton (2009), Yennamani
control, but no field data support this model. (2010), Engmann et al. (2013), and Eshun and
The computer codes UDEC, 3DEC, BUMP, Dzigbordi (2016) used a blast movement monitoring
CAROM, CBLOCK and DMC have been created system (BMM), which was invented at the Univer-
to simulate the blasting process based on the dy- sity of Queensland and commercialized by Blast
namic distinct element model pioneered by Cundall Movement Technologies to achieve high-precision
(1971). In this regard, Gilbride (1995) used UDEC monitoring of the blast-induced rock movement ef-
to predict blast-induced rock movement, but a fect. There are several types of BMMÒ balls, and
complete projectile motion of the rock mass was not several different types of BMMÒ balls can be placed
achieved, and the post-blast grade boundary loca- into one blast hole. The horizontal distance between
tions were not predicted due to the long run times. the same types of BMMÒ balls is smaller than that
Although these computer codes can be used to required using magnetic markers, which means that
simulate blast-induced rock movement, only a few more data can be obtained from a blast trial. This
blasts have been simulated to validate the programÕs technique is easy to operate and is considered the
accuracy, and accurate simulation results for a pro- most accurate method for measuring blast-induced
duction blast have not been obtained. In addition, rock movement.
the distinct motion code (DMC) developed by Pre- In summary, several methods exist for studying
ece and Taylor (1989) has been applied to analyze blast-induced rock movement, and some conclusions
rock motions during blasting. A 3D version of the have been obtained from relevant researches. How-
DMC algorithm, proposed by Preece and Silling ever, theoretical modeling and dynamic distinct ele-
(2016), can be used to simulate blast-induced rock ment simulation cannot obtain accurate blast-induced
movement with different initiation schemes, but its rock movement values, and the field test approaches
performance has not been verified by field tests. are expensive and time-consuming. Hence, a conve-
Therefore, more experimental and production nient and inexpensive method for determining blast-
applications are necessary to ensure that these induced rock movement is of great significance for
computer codes achieve the engineering standard. controlling ore loss and dilution in mines.
Prediction of Blast-Induced Rock Movement During Bench Blasting 845

In recent years, machine learning, the most ac- Rössing Uranium Mine (Fig. 1). Exploration drilling
tive research area in the field of artificial intelligence has identified a mineralized body with two zones,
(AI), has been widely developed and used to solve Zone 1 and Zone 2, each with an apparent strike
many problems in fields such as cancer cell sensi- length of roughly 3 km and a width of roughly
tivity prediction (Menden et al. 2013), rare disease 200 m, to a depth of 400 m. The Husab Uranium
research (Danter 2019), animal agriculture analysis Mine is forecasted to be the second largest producer
(Morota et al. 2018), freight transport prediction of uranium in the world; a minimum of 280,000 t of
(Bakhtyar and Henesey 2014), geochemical anomaly uranium reserves and a mine life of 20 years are
identification (Zou et al. 2017; Zuo 2017), rock expected.
parameter prediction (Armaghani et al. 2015) and The Husab Uranium Mine is situated within the
blast-induced hazard prediction (Ebrahimi et al. central Damara orogenic belt. With the crystalliza-
2016; Dehghani and Shafaghi 2017; Hasanipanah tion of magma, elemental uranium was gradually
et al. 2017a, b; Rad et al. 2018; Armaghani et al. enriched. The magma, which contained alaskite,
2018; Nguyen and Bui 2018; Bui et al. 2019; Nguyen moved along a fault, resulting in the small and nar-
et al. 2019a; Zhou et al. 2019a, b). However, ma- row ore body and major traversal of the waste ore;
chine learning technology has rarely been used to eventually, the alaskite uranium deposit formed
analyze blast-induced rock movement. Hence, sup- (Rong et al. 2016; Zhang and Lu 2018).
port vector regression (SVR), a typical machine Mineral-bearing rock masses are relatively
learning method that has been used successfully for strong, which is why drilling–blasting methods are
solving many geotechnical and mining problems used to break them. At present, the single blasting
(Zhou et al. 2011; Shi et al. 2012; Zhou et al. 2012; N. capacity of a blast block in the Husab Uranium Mine
Li et al. 2017; J. Li et al. 2017; Song et al. 2017; Zhou is approximately 400,000 to 800,000 t, meaning that
et al. 2018), is used to predict the blast-induced rock there is a large volume difference between the ore
movement in this paper, and the optimal parameters body and the blast block. As a result, blast-induced
of the SVR model (C and g) are obtained using the rock movement effect and pre-blast boundary shovel
gray wolf optimizer (GWO) algorithm. guidance cause great ore loss and dilution in the
Based on the above literature review, this paper Husab Uranium Mine.
investigates the law of blast-induced rock movement
and proposes a new hybrid model for predicting
blast-induced rock movement during bench blasting. Collection of Blast-Induced Rock Movement Data
To achieve these aims, the blast-induced rock
movement trial and blast-induced rock movement To analyze blast-induced rock movement effect,
monitoring method were first introduced for data blast-induced rock movement data were collected
collection. Then, a statistical analysis was used to based on four blast movement trials by using 48
study the impact of the initial depth on the blast- BMMÒ balls in August 2017, and 42 BMMÒ balls
induced rock movement and the relationship of the were detected after the blast. The blasting parame-
blast-induced rock movement in different directions. ters of these four blast trials are quite similar (Ta-
Finally, the artificial neural network (ANN) model, ble 1). The procedure of the blast movement trial is
random forest (RF) model and GWO-SVR model illustrated in Figure 2 and elaborated as follows.
were used to predict the horizontal blast-induced Before the blast, the BMMÒ balls were acti-
rock movement, and the modelsÕ predictive perfor- vated by the BMMÒ activator, and these balls
mances are discussed. emitted electromagnetic signals that indicated the
depth of BMMÒ balls and that can be detected by
the BMMÒ detector. Next, the BMMÒ balls were
MATERIALS AND METHODS dropped into the study holes of blast-induced rock
movement, and the initial depth and signal strength
Husab Uranium Mine were recorded by using a BMMÒ detector. Then, the
x, y and z coordinates of the study holes of blast
The Husab Uranium Mine is situated in the movement were recorded by using GPS for the
Erongo region of Namibia. The Husab Uranium calculation of the blast-induced rock movement.
Mine is located 60 km east of Swakopmund, the Subsequently, these holes were back-filled with
third largest city in Namibia, near the existing stemming after installing the BMMÒ balls.
846 Z. Yu et al.

Figure 1. Maps showing the location of the Husab Uranium Mine.

Table 1. Parameters of the four blast trials

Blast ID 1B11D012 1B12D010 2B21D003 2B21D019&11

Bench height (m) 7.5 7.5 7.5 7.5

Sub-drill (m) 1.5 1.8 1.5 1.5
Hole diameter (mm) 165 251 251 251
Total number of holes 436 420 368 226
Burden and spacing (m) 4.6 and 5.0 6.0 and 6.5 6.0 and 6.5 6.0 and 6.5
Stemming (m) 4 4 4 4
Powder factor (kg/m3) 0.69 0.79 0.81 0.86
Number of BMMÒ balls 4 17 11 16

After the blast, the technician held the detector Artificial Neural Network
and located the BMMÒ balls by walking over the
blasted muck pile. Once the technician has found the An ANN can learn and store knowledge from
post-blast location by locating the peak signal from training datasets and use that knowledge to predict
the BMMÒ balls, the surveyor recorded the coordi- new datasets (Schalkoff 1997). A multilayer percep-
nates of that position on the muck pile. Finally, the tron (MLP) is the best type of ANN model, and the
coordinate data of the BMMÒ balls were transferred general structure of an MLP model includes the input
into a personal computer, and the 3D movement layer, the hidden layer and the output layer (Fig. 3)
vectors of BMMÒ balls can be obtained by software (Mohamad et al. 2015). Neurons are placed in these
(BMM Explorer). Then, various software packages layers and can be linked by weights. The construction
such as AutoCAD, MicroStation, Datamine or of an ANN model involves two steps. Firstly, an ANN
Surpac can be used to create digitized shapes of the model should learn from a training dataset. There are
post-blast ore boundary within the blasted muck certain learning algorithms, such as the Levenberg–
pile. Compared to the approach of using the pre- Marquardt algorithm, back-propagation (BP) algo-
blast boundary for shovel guidance, which is imple- rithm and conjugate gradient algorithm, which can be
mented at the Husab Uranium Mine, using the post- used to train an ANN. Among these learning algo-
blast ore boundary obtained by blast movement rithms, the most widely used is the BP algorithm. Its
monitoring can effectively reduce ore loss and dilu- learning process is terminated when the error reaches
tion. the level of the cost function (Marto et al. 2014).
Prediction of Blast-Induced Rock Movement During Bench Blasting 847

Figure 2. Procedure of blast movement trial.

ANNs have been applied successfully to solve

engineering problems such as blast-induced over-
break (Koopialipoor et al. 2019a), blast-induced air
over-pressure (Nguyen and Bui 2018) and blast-in-
duced flyrock (Koopialipoor et al. 2019b). There-
fore, an ANN model was developed in this study to
predict blast-induced rock movement, and its pre-
dictive performance was compared with the perfor-
mance of other AI techniques.

Random Forest

The RF method, a typical machine learning

method, was proposed by Breiman (2001). The RF
algorithm is based on decision-tree algorithms, and
it can combine a large number of decision trees into
a complex model to make the final decision for
solving classification or regression problems. During
Figure 3. An ANN model (7-12-1) for predicting blast- the calculation process, each decision tree in the RF
induced rock movement in this paper. model is trained by randomly selecting variables and
data samples from the initial blast-induced rock
Secondly, after training, a testing dataset can be used movement database. A detailed introduction of the
to analyze the predictive performance of the trained RF model can be found in Breiman (2001) and Liaw
or established ANN model. and Wiener (2002).
848 Z. Yu et al.

The RF model has been used, with satisfactory and the corresponding boundary values. By using ni
predictive performance, to solve mining problems in and ni , Eq. 2 can be rewritten as follows:
blasting operations, such as blast-induced ground
vibration (Dong et al. 2011) and air over-pressure Minimize :
(Nguyen and Bui 2018; Bui et al. 2019). Therefore, 1 XN  
the RF method was used here for blast-induced rock kak2 þC ni þ ni
2 i¼1
movement prediction. 8
< yi  T ðxi ; aÞ  e þ ni ; i ¼ 1; 2; . . . ; N;
>
subject to T ðxi ; aÞ  yi  e þ ni ; i ¼ 1; 2; . . . ; N;
>
:
GWO-SVR Model ni  0; ni  0; i ¼ 1; 2; . . . ; l
ð3Þ
Basic Concept of SVR
where C is a constant called the penalty factor.
The SVM algorithm, an achievement of statis- Equation (3) can be converted into the following
tical learning theory, is a useful method for solving form by using the Lagrangian multipliers, ai and ai ,
classification and regression problems (Vapnik 1995;
and Karush–Kuhn–Tucker conditions (Vapnik 1995;
Collobert and Bengio 2001). The development of
Chang and Lin 2011).
SVMs has fostered the development of support
Maximize:
vector regression (SVR), a machine learning method
for performing linear and nonlinear regression tasks 1X N     
(Drucker et al. 1997; Smola and Schölkopf 2004). H ða; a Þ ¼  ai  ai aj  aj K xi ; xj
2 i;j¼1
Based on the studies of Gunn (1998), Samui et al.
(2008), Faruto and Li (2009), Gopalakrishnan and X
N   X
N  
þ yi ai  ai  e yi ai  ai ð4Þ
Kim (2011), Zhou et al. (2011), Chang and Lin i¼1 i¼1
(2011), Zhou et al. (2012), Shi et al. (2012), Sujay
and Deka (2014), Hasanipanah et al. (2015), Hasa- subject to the constraints,
nipanah et al. (2017b) and Rad et al. (2018), some
key points of the SVR theory are discussed here. X
N  
ai  ai ¼ 0; ai ; ai 2 ½1; C; i 2 ½1; N 
D ¼ fðxi ; yi Þg is a learning sample set, where xi i¼1
and yi are the input values and the corresponding
output values, respectively. The task of an SVR is to After solving Eq. 4, the optimal solution of ai
build a reliable relationship between input variables and ai can be found, and then, Eq. 1 can be trans-
xi and output variables yi . The SVR model used for a formed into the following form (Vapnik 1995; Chang
nonlinear case can be described as (Gunn 1998; and Lin 2011):
Zhou et al. 2011; Shi et al. 2012; Zhou et al. 2012): X
N  
T ð xÞ ¼ ai  ai K ðxi ; xÞ þ b ð5Þ
T ð xÞ ¼ ha; gð xÞi þ b ð1Þ
i¼1

where a, b and T ð xÞ denote the weight coefficient, where K ðxi ; xÞ is the kernel function that is used to
constant coefficient and high-dimensional kernel- map the input parameters to the feature space
induced feature space, respectively. Moreover, a and (Vapnik 1995). A review of the literature shows that
b are optimized when R(C) obtains the minimum the Gaussian RBF (radial primary kernel function)
value. kernel is a superior function compared to other
 2 kernel functions because of its powerful mapping
a  C X N
RðCÞ ¼ þ q ðyi ; T ðxi ÞÞ ð2Þ capability (Vapnik 1995; Li 2009; Chang and Lin
2 P i¼1 e
2011). Hence, the RBF kernel function was chosen
  for application in this study, and the radial primary
where 12 a2  represents the regularization term, and kernel function is defined as:
C is a cost function for assessing the empirical risk. !
Two positive slack variables, ni and ni , are used  kxi ; yi k2
K ðx; yÞ ¼ exp ð6Þ
to describe the distance between the actual values 2r2
Prediction of Blast-Induced Rock Movement During Bench Blasting 849

Figure 4. The social hierarchy of gray wolves (modified from Photophoto).

where r denotes the bandwidth of the RBF.  

~ ¼ C
D ~X ~ ðtÞ
~ P ðt Þ  X ð7Þ
Because the combination of C and g, where C is
the penalty factor and g is the gamma in the RBF
kernel, has an enormous impact on the prediction ~ ð t þ 1Þ ¼ X
X ~D
~ p ðt Þ  A ~ ð8Þ
accuracy, the selection of these parameters is an
extremely important issue. where A~ and D~ are the coefficient vectors; X~p , X
~ are
the current position vectors of the prey and the gray
wolf, respectively (Fig. 5); and t is the current iter-
Gray Wolf Optimizer Algorithm ation.
Second step: hunting behavior. The position of
The GWO was first proposed by Mirjalili et al. omega (x) can be updated by the best solutions of
(2014). This algorithm is inspired by the living habits the positions of the higher-level gray wolves (a, b
of gray wolves in nature. In GWO modeling, the and d) (Bian et al. 2018). In this regard, Eqs. 9–15
GWO algorithm includes a social hierarchy and prey have been proposed by Mirjalili et al. (2014) for
tracking, encircling and attacking prey processes. calculating the new positions of a, b and d; the
According to the social hierarchy, the gray wolf position update process is shown in Figure 6.
population can be divided into four parts: alpha (a),  
beta (b), delta (d) and omega (x) (Fig. 4). The ~ a ¼ C
D ~1 X ~ ðtÞ
~ a ðt Þ  X ð9Þ
leaders of the gray wolves, alpha (a), are responsible
for making decisions such as hunting and sleeping  
~ b ¼ C
D ~2 X ~ ðtÞ
~ b ðt Þ  X ð10Þ
place. Beta (b) denotes the subordinate wolves that
help the alpha group to make decisions. Delta (d),  
the third-ranking group in the gray wolf social ~ d ¼ C
D ~3 X ~ ðtÞ
~ d ðt Þ  X ð11Þ
hierarchy, must submit to all the other more domi-
nant wolves. The reminding gray wolves are the  
~1 ¼ X
D ~1 D
~a  A ~a ð12Þ
lowest-ranking ones and are categorized into the
omega (x) group (Mirjalili et al. 2014).  
The mathematical model of the gray wolf ~2 ¼ X
D ~2 D
~b  A ~b ð13Þ
algorithm includes three steps.
First step: encircling behavior. To hunt prey, the  
gray wolves encircle it, and that behavior can be D ~3 D
~d  A
~3 ¼ X ~d ð14Þ
mathematically modeled as (Mirjalili et al. 2014):
850 Z. Yu et al.

Figure 5. The position vectors of the wolves (modified from Mirjalili et al. (2014) and Photophoto).

 . 1) Establish the blast database, which should

~ ð t þ 1Þ ¼ X
X ~2 þ X
~1 þ X ~3 3 ð15Þ be divided into a training database (here,
70%) and a testing database (here, 30%).
Third step: attacking behavior. The main
2) Pre-process the data, which means linearly
objective of attacking behavior—the final hunting
scaling the model parameter set from 0 to 1.
behavior—is to determine the optimized position of
3) Use the GWO algorithm to find the optimal
the prey. In this regard, the action of the wolves is
C and g for training the SVR model.
determined by the value of j Aj; the wolves can only
4) Establish the final SVR model for predicting
attack the prey while j Aj  1, and the optimization
horizontal movement by using the optimal C
of the GWO will be achieved when the criterion is
and g.
reached (Mirjalili et al. 2014).
5) Evaluate and validate the SVR model by
In practice, the position of the prey corresponds
using testing datasets.
to the optimal parameters (C and g) of the SVR
model. Meanwhile, the search process begins with
A flow diagram that clarifies the calculation
randomly generated gray wolf positions; then, the
process is shown in Figure 7.
positions of a, b and d are obtained by calculating
the corresponding fitness. After finding the best
three positions of a, b and d, the position of the prey
Performance Metrics
(C and g) is estimated by Eq. 15 and used to
establish the optimal SVR model for blast-induced
The root mean square error (RMSE), correla-
rock movement prediction.
tion coefficient (R2), mean absolute error (MAE)
and variance account for (VAF) are widely used
indices for evaluating the accuracy of a prediction
GWO-SVR Model Calculation Process
model (Zhou et al. 2011, 2012; Nguyen et al. 2019a).
The following equations can be used for calculating
The procedure for establishing a GWO-SVR
these four indices:
model involves the following five steps.
Prediction of Blast-Induced Rock Movement During Bench Blasting 851

Figure 6. The position update process of the gray wolf algorithm (modified from Mirjalili et al. (2014) and
Photophoto).

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi RESULTS AND DISCUSSION

1X n
RMSE ¼ ðT ðxi Þ  xi Þ2 ð16Þ
n i¼1 Analysis of Trial Blast Results

n
2 The 3D movement vectors were monitored by
P Pn P n
n T ðxi Þxi  T ðxi Þ xi using BMMÒ balls, and 42 datasets were obtained
i¼1 i¼1 i¼1
R2 ¼ n
2 ! n
2 ! after the trial blasts. In addition, the technician used
P
n
2 P P 2
n P the 3D movement vectors to adjust the ore bound-
n T ð xi Þ  T ð xi Þ n xi  xi
i¼1 i¼1 i¼1 i¼1 ary, and the pre-blast and post-blast ore boundaries
of 2B21D003 are plotted in Figure 8. There is a
ð17Þ considerable difference between the pre-blast
boundary and the post-blast boundary, which means
1X n
that using post-blast ore boundary for shovel guid-
MAE ¼ jxi  T ðxi Þj ð18Þ
n i¼1 ance can effectively reduce ore dilution.

The results obtained from the trial blasts in the
varðxi  Tðxi ÞÞ Husab Uranium Mine, by using the BMMs instru-
VAF ¼ 1  100 ð19Þ
varðxi Þ mentation and software, are demonstrated below.
The horizontal movement under the action of
where xi, T(xi) and n are actual values, predicted explosive energy varied from 1.1 to 14.9 m, and the
values and number of input data, respectively. movement of the lower part of the blast hole was
852 Z. Yu et al.

Figure 7. Flowchart of the GWO-SVR method.

Figure 8. Ore boundary adjustments of 2B21D003. The blue and red polygons represent the pre-blast and
post-blast ore boundaries, respectively.
Prediction of Blast-Induced Rock Movement During Bench Blasting 853

Figure 9, in the same blast hole, the blast-induced

rock movement data obtained from the Husab
Uranium Mine at different depths are quite differ-
ent. To study the influence of the initial depth on the
blast-induced rock movement effect, the depth
coefficient (DC) is defined as:
hl
DC ¼
HB þ HS  l
( ð20Þ
l
DC  HB þH S l
subject to
DC  1

where h, l, HB and HS represent the initial depth of

the BMMÒ ball, the stemming length, the bench
height and the sub-drill length, respectively.
The detonated blast holes can create a
momentary free face for the later detonated blast
holes. Hence, the throwing direction of the rock
mass that is perpendicular to the isochrones is con-
trolled by the natural free face and the centerline of
the blasting network during the blasting process as
shown in Figure 10. To describe the effect of the free
face on the blast-induced rock movement effect, the
number of blast holes between the blast movement
study holes and the natural free face is defined as the
first centerline distance (FCD), and the number of
blast holes between the blast-induced rock move-
ment holes and the centerline is defined as the sec-
ond centerline distance (SCD). As a result, the rock
type (RT), number of free faces (NFF), first cen-
terline distance (FCD), second centerline distance
(SCD), hole diameter (HD), burden 9 spacing
(BS), the depth coefficient (DC), powder factor
(PF), horizontal movement (MH), vertical move-
Figure 9. Scatter diagram of blast-induced rock movement ment (MV) and 3D movement (M3D) are used to
data: a horizontal movement and b vertical movement.
build the database. Note that all the rock types were
assigned according to the gradient of 30 MPa of the
UCS (uniaxial compressive strength) data. Because
larger than that of the upper part. The vertical of the lack of accurate rock parameters (see Ta-
movement displacement varied from  2.4 to 6.6 m, ble 2), the UCS data of the Phoenix Mine were
and the vertical movement showed a different trend obtained from the reference Copper Canyon
as the horizontal movement (Fig. 9). (Phoenix) tectono-stratigraphic column (Yennamani
2010), the UCS estimated from point load testing
results for the rock formation (Pierre Mousset-Jones
Blast-Induced Rock Movement Database 2018, personal communication) and the empirical
data of the physical and mechanical properties of the
To expand the database and achieve high rock (Wang 2004). The quantitative standard and
accuracy for the analysis performance, 95 sets of the database with 137 sets of samples are shown in
samples obtained by using the BMMs in the Phoenix Table 2 and Figure 11, respectively. Moreover, cor-
Mine (Yennamani 2010) were added to the blast- relation coefficients between pairs of input variables
induced rock movement database. As seen in were calculated (Fig. 11), and the HD was removed
854 Z. Yu et al.

Figure 10. Blast design parameters terminology.

Table 2. Rock-type properties

Mine Rock type Description UCS/MPa Valuation

Phoenix Mine Lower battle formation Massive siliceous pebble conglomerate 120–150 5
Cambrian harmony formation Allochthonous micaceous sandstone 90–120 4
Lower battle formation – 90–120 4
Cambrian harmony formation
Virgin fault – 60–90 3
Antler peak limestone formation Limestone 120–150 5
Husab Mine Rossing Formation, Khan Formation, etc. Granite, schist, gneiss, etc. 90–120 4

because the correlation coefficient between HD and blast holes was analyzed using statistical analysis
BS was 0.97, which is greater than the threshold methods, and the results are shown in Figure 12.
value (0.85) suggested by Rovini et al. (2018). The horizontal and 3D movements showed the
trend of increasing first and then decreasing, and the
extreme value was reached near the top of the
Relationship Between Depth Coefficient explosive section. This trend was due to the different
and Blast-Induced Rock Movement driving forces and resistances of the stemming sec-
tion and explosive section. On the one hand, from
The average blast-induced rock movement of the surface to the bottom of the stemming section,
the BMMÒ balls from different initial depths in the the closer to the charging section was, the greater
Prediction of Blast-Induced Rock Movement During Bench Blasting 855

0.0 1.0 2.0 0 10 20 15 25 35 0.60 0.75 -5 5

RT .

4.5
0.24 ** -0.042 0.24 ** -0.14 -0.23
** 0.054 -0.00055
0.18 * 0.042

3.0
NFF
1.5

0.25 ** 0.46 *** -0.35 *** -0.34 *** 0.19 * -0.25

** 0.012 -0.13
0.0

FCD *** *** ***

15
-0.065
-0.34 -0.33 0.13 -0.33 0.074 -0.028

0
SCD
10

-0.013 -0.0025 0.05 0.0036 0.11 -0.11

0

240
HD *** ***
*** 0.72 *
0.97 -0.38 -0.18 -0.084

180
BS ***
*** 0.69 **
30

-0.36 -0.23 -0.13

15

DC ***

0.2
-0.33 -0.085 -0.10

-0.6
PF
0.80

0.025 -0.045
0.60

MH
H
*

15
0.20

5
MV
V
5
-5

3.0 4.0 5.0 0 10 20 180 240 -0.6 0.0 0.6 5 15

Figure 11. Scatter plot matrix of the original data set. Black dots represent data from the Husab Uranium Mine. Yellow dots
represent the data of from the Phoenix Mine. The pairwise relationships, marginal distributions and correlation coefficients are shown
in the lower triangle, diagonal and upper triangle, respectively.

the driving force of the stress waves and explosive Relationships Among Horizontal, Vertical and 3D
gases was. On the other hand, from the top to the Movements
bottom of the explosive section, the horizontal bur-
den increased, which meant that resistance to rock Currently, some mines, such as the Husab
movement in the horizontal direction increased. Uranium Mine, use a horizontal ore boundary to
The curve shape of MV and the position of the develop the shovel area, and the horizontal bound-
extreme value were the same as those of the MH and ary is considered to have a greater influence on
M3D curves. In addition, because of different resis- shovel loading than that of the vertical boundary.
tances in the vertical direction, the rate of decrease Similar relationships between the horizontal, verti-
in the MV curve of the stemming section was lower cal and 3D movements in this study were obtained
than that of the explosive section. From the top to (Fig. 13).
the bottom of the explosive section, resistance of the Figure 13 indicates that horizontal movement
rock movement gradually increased as the horizon- and 3D movement show a good linear relationship,
tal and vertical burden increased. and the fitting coefficient was as high as 0.924.
856 Z. Yu et al.

13
3D Movement Trend of 3D Movement
12 Vertical Movement Trend of Vertical Movement
Horizontal Movement Trend of Horizontal Movement
11

10

Blast-induced Movement/m
9

0
-l/(H B+HS-l) -0.6~-0.4 -0.4~-0.2 -0.2~0.0 0.0~0.2 0.2~0.4 0.4~0.6 0.6~1.0
-1.0~-0.6

Stemming Explosive
Depth Coefficient
Figure 12. Relationship between depth coefficient and blast-induced rock
movement.

However, the relationship between the vertical taset were randomly selected and used as training
movement and 3D movement was not obvious. and testing datasets, respectively, following the re-
Hence, horizontal movement can directly replace searches of Ebrahimi et al. (2016) and Khandelwal
3D movement in the analyses and it can be used to et al. (2017). The respective symbols, range, mean,
guide the shovel. and standard deviation of the blast parameters for
developing the blast-induced movement models are
shown in Table 3.
Prediction of MH Using the ANN Model According to previous studies (Mohamad et al.
2015; Hasanipanah et al. 2016b), one hidden layer
To predict blast-induced rock movement using can decrease the model complexity, and this tech-
the ANN model, a MATLAB code was prepared, nique has been widely applied in solving engineering
and the prepared database (input and output data) problems. Therefore, one hidden layer was used in
was normalized to a range of 0–1 using the following this study to predict blast-induced rock movement.
equation to avoid attributes in greater numeric Furthermore, Hecht-Nielsen (1987) suggested that
ranges dominating those in smaller numeric ranges the number of nodes in the hidden layer (Ni) should
and to reduce computational difficulties (Zhou et al. be not less than 2n + 1, where n is the number of
2013): input variables. Therefore, 15 ANN models with Ni
from 1 to 15 were run five times each to determine
ðymax  ymin Þ  ðx  xmin Þ
ynorm ¼ þ ymin ð21Þ the optimal Ni for predicting blast-induced rock
ðxmax  xmin Þ movement by comparing the best predictive results
among the five runs. From the results of these
where ymax and ymin are the range parameters of the
models (Fig. 14), model 12 with Ni of 12, R2 of 0.700,
map, ynorm is the normalized value, and xmax and
and RMSE of 2.231 for the training datasets and R2
xmin are the maximum and minimum values of the
of 0.658 and RMSE of 2.518 against the testing da-
actual data, respectively. Then, roughly 70% (95/
tasets, was found to be the best predictive ANN
137) and roughly 30% (41/137) of the original da-
model.
Prediction of Blast-Induced Rock Movement During Bench Blasting 857

be determined first. For RF modeling, seven RF

models with mtry ranges of 1–7 and ntree = 2000
were developed, the predictive performance
(checked by RMSE as shown in Fig. 15) of the
training datasets indicates that seven was the opti-
mal value of mtry for predicting blast-induced rock
movement. Subsequently, a grid search technique
was applied with ntree in the range of 1–2000
(Fig. 16) and mtry = 7 to determine the optimal
ntree. There were no significant changes when
ntree = 2000; so, 2000 was selected as the optimal
ntree.

Prediction of MH using the GWO-SVR Model

To obtain the best predictive performance of

the GWO-SVR model, the first step of the GWO-
SVR design is to determine the maximum number of
iterations (Imax) and the number of wolves. Hence,
an investigation was conducted to find the optimal
maximum number of iterations (Imax), and the pre-
dictive performance of the GWO-SVR model is
checked using RMSE. As shown in Figure 17, there
were no significant changes in the fitness curve after
100 iterations. In addition, the number of wolves had
no obvious effect on the predictive performance of
the training datasets. Therefore, 100 and 20 were
chosen to establish the final GWO-SVR model for
predicting blast-induced rock movement.
Figure 13. Relationship between horizontal, vertical and 3D
movements: a horizontal vs. 3D movement; b vertical vs. 3D
movement. Evaluation of Results

Table 3. Descriptive statistics of the input and output parameters Several methods including ANN, RF and
GWO-SVR approaches were used in this study to
Type of data Parameter Range Mean Std. dev.
predict blast-induced rock movement. From a re-
Input RT 3.00–5.00 4.15 0.65 view of previous studies (Mohamad et al. 2015;
NFF 0.00–2.00 1.19 0.66 Ebrahimi et al. 2016; Hasanipanah et al. 2016a;
FCD 0.50–26.50 7.01 4.84
SCD 0.00–20.00 6.46 4.67
Khandelwal et al. 2017; Armaghani et al. 2018;
BS/m2 14.49–39.00 26.10 8.24 Nguyen et al. 2019b), the database was randomly
DC  0.58 to 0.58 0.19 0.27 divided into training and testing datasets, namely
PF/(LB/ton) 0.57–0.80 0.69 0.06 70% (95/137) and 30% (41/137) of the data,
Output MH/m 1.10–19.96 8.22 3.93 respectively. Subsequently, a parameter study of the
ANN model, RF model and GWO-SVR model was
conducted, and the optimal parameters of these
three models were obtained. Finally, an ANN
Prediction of MH Using the RF Model model, an RF model and a GWO-SVR model were
developed, ran 10 times and evaluated using several
In the RF model, two main parameters, performance metrics including RMSE, MAE, R2
including the number of trees (ntree) and the num- and VAF. A model was considered a perfect model
ber of variables used to grow each try (mtry), should when the average RMSE is 0, the average MAE is 0,
858 Z. Yu et al.

Figure 14. Predictive performance of the ANN model with various NH.

Figure 16. Predictive performance of the RF model of training

Figure 15. Predictive performance of the RF model for training datasets with various ntree values.
datasets with various mtry values.

with an average RMSE of 2.847, average MAE of

the average R2 is 1 and the average VAF is 100, and
2.242, average R2 of 0.635 and average VAF of
the values of these performance metrics are given in
38.056. The RF model yielded a predictive perfor-
Tables 4 and 5.
mance using the training data with an average
For systematic comparison of the predictive
RMSE of 2.055, average MAE of 1.691, average R2
performance of the ANN model, RF model and
of 0.871 and average VAF of 73.581 and using the
GWO-SVR model, a simple ranking method devel-
testing data with an average RMSE of 2.199, average
oped by Zorlu et al. (2008) was used here, and the
MAE of 1.828, average R2 of 0.819 and average
ranking results are shown in Table 6. The ANN
VAF of 63.200. Moreover, the results of the GWO-
model yielded a predictive performance using the
SVR model (average RMSE of 1.855, average MAE
training data with an average RMSE of 2.453,
of 1.731, average R2 of 0.809, average VAF of 78.549
average MAE of 1.791, average R2 of 0.636 and
for training and average RMSE of 1.513, average
average VAF of 64.636 and using the testing data
Prediction of Blast-Induced Rock Movement During Bench Blasting 859

0.0420
20 wovles 40 wovles 60 wovles
0.0418 80 wovles 100 wovles 150 wovles
200 wovles 250 wovles 300 wovles
0.0416
0.0416
0.0414 0.0414
0.0412 0.0412
RMSE 0.0410
0.0410

RMSE
0.0408
0.0408
0.0406
0.0406 0.0404
0.0404 0.0402
0.0400
0.0402
0.0398
0.0400 0 25 50 75 100
Iteration
0.0398
0 100 200 300 400 500 600 700 800 900 1000
Iteration
Figure 17. Fitness curves of various GWO-SVR models with different numbers of
wolves and Imax = 1000.

Table 4. Average values of performance metrics of the ANN model, RF model and GWO-SVR model

Method Training dataset Testing dataset

2
RMSE MAE R VAF RMSE MAE R2 VAF

ANN 2.453 1.791 0.636 62.644 2.847 2.242 0.635 38.056

RF 2.055 1.691 0.871 73.581 2.1991 1.828 0.819 63.200
GWO-SVR 1.855 1.731 0.809 78.549 1.513 1.217 0.865 82.906

Table 5. Standard deviations of performance metrics of the ANN model, RF model and GWO-SVR model

Method Training dataset Testing dataset

RMSE MAE R2 VAF RMSE MAE R2 VAF

ANN 0.147 0.190 0.039 4.347 0.319 0.224 0.022 13.434

RF 0.007 0.009 0.002 0.177 0.012 0.013 0.005 0.410
GWO-SVR 0.002 0.002 0.001 0.050 0.005 0.001 0.000 0.119

MAE of 1.217, average R2 of 0.865, average VAF of provided by the ANN and RF models. Figures 18, 19
82.906 for testing) show that the GWO-SVR model and 20 demonstrate the best predictive performance
can yield a predicted horizontal movement that is with the optimal parameters and the fitted scale of
closer to the actual horizontal movement than that the predictive model for predicting the horizontal
860 Z. Yu et al.

Table 6. Ranking values of performance metrics of the ANN model, RF model and GWO-SVR model

Method Description Training dataset Testing dataset Sum ranking

RMSE MAE R2 VAF RMSE MAE R2 VAF

ANN Mean 1 1 1 1 1 1 1 1 16
Std. dev. 1 1 1 1 1 1 1 1
RF Mean 2 3 3 2 2 2 2 2 34
Std. dev. 2 2 2 2 2 2 2 2
GWO-SVR Mean 3 2 2 3 3 3 3 3 46
Std. dev. 3 3 3 3 3 3 3 3

Figure 18. Predicted and actual horizontal movement obtained by the ANN
model.

movement in this study. Consequently, the GWO- each input variable was obtained by comparing the
SVR model can be used to help control ore loss and R2 value of a GWO-SVR model with a sensitivity
dilution caused by the blast-induced rock movement factor and a GWO-SVR model without that factor.
with a high level of prediction accuracy. Based on this principle, three sensitivity analysis
tests were conducted for each of the following
variables: RT, NFF, FCD, SCD, BS, DC and PF.
Sensitivity Analysis The average R2 of the three tests and the R2 of each
test are shown in Figure 21.
The high level of prediction accuracy of the The results of analyses show that the influences
GWO-SVR model was verified in the above analy- of various factors on the blast-induced rock move-
sis, and the sensitivity analysis of the input variables ment prediction are different and these input vari-
of the GWO-SVR model was carried out to measure ables can be listed as FCD, NFF, SCD, RT, PF, BS
the degree of impact of each input variable on the and DC in increasing order of their degree of
horizontal movement prediction. The sensitivity of influence. Among these variables, RT, FCD, DC
Prediction of Blast-Induced Rock Movement During Bench Blasting 861

Figure 19. Predicted and actual horizontal movement obtained by the RF model.

Figure 20. Predicted and actual horizontal movement obtained by the GWO-SVR
model.

and SCD cannot be controlled. Therefore, the To analyze the impact of NFF, BS and PF on
effective blasting parameters that can be used for blast-induced rock movement, the distribution of
blast-induced rock movement control in mining NFF, BS and PF is summarized and plotted in Fig-
engineering are NFF, BS and PF. ure 22. As shown, NFF and average MH first de-
862 Z. Yu et al.

Figure 21. Sensitivity analysis results of the input variables.

creased and then increased. However, this result GWO-SVR model to predict blast-induced rock
does not match the practical experience of blasting movement during bench blasting. The main conclu-
production because, due to the minimum moving sions from this study and analysis are summarized as
space available for the fractured rock, the MH value follows.
for a blast without a free face cannot be greater than
the MH values of the one or two free-face blast 1. The horizontal, vertical, 3D movements first
scenarios. The reason for this result may be that increase and then decrease from the ground
NFF is not the dominant factor when MH is small. to the hole bottom, and the maximum dis-
The average MH gradually decreased as BS in- placement is reached near the top of the
creased. The main reason for this relationship is that charging section. Meanwhile, the horizontal
the volume of the rock mass that is broken around movement and 3D movement show a good
the single blast hole gradually increases as BS in- linear relationship.
creases, leading to the low initial speed and short 2. A new application of the GWO-SVR model
distance of the fractured rock movement. Thus, was proposed for predicting blast-induced
increasing the parameters of the hole pattern can rock movement, and 137 groups of datasets
reduce the MH during the mining process. (70% used for training and 30% used for
In addition, it was found that as PF increases, testing) were collected from the Husab
the average MH increases. With increasing power Uranium Mine and previous literature. Sub-
factor, the explosion energy and the initial speed of sequently, ANN, RF, and GWO-SVR mod-
the fractured rock increase, leadings to the large els were used to predict blast-induced rock
moving distance of fractured rock. Thus, MH can be movement, and their predictive perfor-
controlled by reducing PF. mances were evaluated by RMSE, MAE, R2
and VAF. The performance metrics show
that the GWO-SVR model achieved good
CONCLUSIONS predictive performance compared to that of
the ANN and RF models. Moreover, MH can
This paper investigated the distribution law of be controlled by adjusting BS and PF among
blast-induced rock movement and proposed a the input variables of the GWO-SVR model.
Prediction of Blast-Induced Rock Movement During Bench Blasting 863

collected, and only ANN, RF and GWO-SVR

models were used in this paper. Therefore, future
investigations need to collect more data to expand
the database, and more models must be used to
predict blast-induced rock movement.

ACKNOWLEDGMENTS

The most sincere gratitude is extended for the

financial support from the National Natural Science
Foundation Project of China (Grant No. 51874350;
Grant No. 41807259), the National Key R&D Pro-
gram of China (2017YFC0602902), and the Funda-
mental Research Funds for the Central Universities
of Central South University (2018zzts217). Special
thanks to Blast Movement Technologies for the
high-precision monitoring of the blast-induced rock
movement effect. Additionally, the authors fully
acknowledge the Uranium Resource Company
Limited and the Swakop Uranium Proprietary
Limited for permitting us to use the blast-induced
rock movement data.

REFERENCES

Armaghani, D. J., Hasanipanah, M., Mahdiyar, A., Majid, M. Z.

A., Amnieh, H. B., & Tahir, M. M. (2018). Airblast predic-
tion through a hybrid genetic algorithm-ANN model. Neural
Computing and Applications, 29(9), 619–629.
Armaghani, D. J., Mohamad, E. T., Momeni, E., & Nar-
ayanasamy, M. S. (2015). An adaptive neuro-fuzzy inference
system for predicting unconfined compressive strength and
YoungÕs modulus: A study on Main Range granite. Bulletin of
Engineering Geology and the Environment, 74(4), 1301–1319.
Bakhtyar, S., & Henesey, L. (2014). Freight transport prediction
using electronic waybills and machine learning. In Interna-
tional conference on informative and cybernetics for compu-
tational social systems, 2014 (pp. 128–133).
Bian, X. Q., Zhang, L., Du, Z. M., Chen, J., & Zhang, J. Y. (2018).
Prediction of sulfur solubility in supercritical sour gases using
grey wolf optimizer-based support vector machine. Journal of
Molecular Liquids, 261, 431–438.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–
32.
Bui, X. N., Nguyen, H., Le, H. A., Bui, H. B., & Do, N. H. (2019).
Prediction of blast-induced air over-pressure in open-pit
mine: Assessment of different artificial intelligence tech-
Figure 22. Relationship between horizontal movement and niques. Natural Resources Research, 1–21. https://doi.org/10.
NFF and BS: a horizontal movement vs. NFF; b horizontal 1007/s11053-019-09461-0.
movement vs. BS; c horizontal movement vs. PF. Chang, C. C., & Lin, C. J. (2011). LIBSVM: A library for support
vector machines. ACM Transactions on Intelligent Systems
and Technology (TIST), 2(3), 27.
Although some useful conclusions were ob- Collobert, R., & Bengio, S. (2001). SVMTorch: Support vector
tained, only blast-induced rock movement data of machines for large-scale regression problems. Journal of
the Husab Uranium Mine and Phoenix Mine were Machine Learning Research, 1(2), 143–160.
864 Z. Yu et al.

Cundall, P. A. (1971). A computer model for simulating pro- Hasanipanah, M., Shahnazar, A., Bakhshandeh Amnieh, H., &
gressive, large-scale movement in blocky rock system. In The Jahed Armaghani, D. (2017b). Prediction of air-overpressure
international symposium on rock mechanics. caused by mine blasting using a new hybrid PSO–SVR
Danter, W. R. (2019). DeepNEU: Cellular reprogramming comes model. Engineering with Computers, 33(1), 23–31.
of age–a machine learning platform with application to rare Hecht-Nielsen, R. (1987). KolmogorovÕs mapping neural network
diseases research. Orphanet Journal of Rare Diseases, 14(1), existence theorem. In: Proceedings of the international con-
13. ference on neural networks, 1987 (Vol. 3, pp. 11-14): IEEE
Dehghani, H., & Shafaghi, M. (2017). Prediction of blast-induced Press New York.
flyrock using differential evolution algorithm. Engineering Khandelwal, M., Armaghani, D. J., Faradonbeh, R. S., Yellishetty,
with Computers, 33(1), 149–158. M., Majid, M. Z. A., & Monjezi, M. (2017). Classification and
Dong, L. J., Li, X. B., Xu, M., & Li, Q. Y. (2011). Comparisons of regression tree technique in estimating peak particle velocity
random forest and support vector machine for predicting caused by blasting. Engineering with Computers, 33(1), 45–53.
blasting vibration characteristic parameters. Procedia Engi- Koopialipoor, M., Armaghani, D. J., Haghighi, M., & Ghaleini, E.
neering, 26, 1772–1781. N. (2019a). A neuro-genetic predictive model to approximate
Drucker, H., Burges, C. J., Kaufman, L., Smola, A. J., & Vapnik, overbreak induced by drilling and blasting operation in tun-
V. (1997). Support vector regression machines. Paper pre- nels. Bulletin of Engineering Geology and the Environment,
sented at the advances in neural information processing sys- 78, 981–990.
tems. Koopialipoor, M., Fallah, A., Armaghani, D. J., Azizi, A., &
Ebrahimi, E., Monjezi, M., Khalesi, M. R., & Armaghani, D. J. Mohamad, E. T. (2019b). Three hybrid intelligent models in
(2016). Prediction and optimization of back-break and rock estimating flyrock distance resulting from blasting. Engi-
fragmentation using an artificial neural network and a bee neering with Computers, 35(1), 243–256.
colony algorithm. Bulletin of Engineering Geology and the Li, N., Wang, L. G., & Jia, M. T. (2017). Rockburst prediction
Environment, 75(1), 27–36. based on rough set theory and support vector machine.
Engmann, E., Ako, S., Bisiaux, B., Rogers, W., & Kanchibotla, S. Journal of Central South University, 48(5), 1268–1275.
(2013). Measurement and modelling of blast movement to Li, J., Yue, J. H., Yang, Y., Zhan, X. Z., & Zhao, L. (2017). Multi-
reduce ore losses and dilution at Ahafo Gold Mine in Ghana. resolution feature fusion model for coal rock burst hazard
Ghana Mining Journal, 14, 27–36. recognition based on acoustic emission data. Measurement,
Eshun, P. A., & Dzigbordi, K. A. (2016). Control of ore loss and 100, 329–336.
dilution at Anglogold Ashanti, Iduapriem mine using blast Liaw, A., & Wiener, M. (2002). Classification and regression by
movement monitoring system. Ghana Mining Journal, 16, 49– randomForest. R News, 2(3), 18–22.
59. https://doi.org/10.4314/gmj.v16i1.6. Lucas, R., & Nies, D. (1990). Improving Fragmentation and Ore
Faruto, & Li, Y. (2009). LIBSVM-faruto ultimate version: A toolbox Displacement Control. In: Proceedings of the sixteenth con-
with implements for support vector machines based on LIBSVM. ference on explosives and blasting technique, 1990 (pp. 409–
Software available from https://www.ilovematlab.cn. 422).
Gilbride, L. J. (1995). Blast-induced rock movement modelling for Marto, A., Hajihassani, M., Jahed Armaghani, D., Tonnizam
bench blasting in Nevada open-pit mines. Reno: University of Mohamad, E., & Makhtar, A. M. (2014). A novel approach
Nevada. for blast-induced flyrock prediction based on imperialist
Gopalakrishnan, K., & Kim, S. (2011). Support vector machines competitive algorithm and artificial neural network. The
approach to HMA stiffness prediction. Journal of Engineer- Scientific World Journal, 2014, 643715. https://doi.org/10.115
ing Mechanics, 137(2), 138–146. 5/2014/643715.
Gunn, S. R. (1998). Support vector machines for classification and Menden, M. P., Iorio, F., Garnett, M., Mcdermott, U., Benes, C.
regression. ISIS Technical Report, 14(1), 5–16. H., Ballester, P. J., et al. (2013). Machine learning prediction
Harris, G. W. (1997). Measurement of blast induced rock move- of cancer cell sensitivity to drugs based on genomic and
ment in surface mines using magnetic geophysics. Reno: chemical properties. PLoS ONE, 8(4), e61318.
University of Nevada. Mirjalili, S., Mirjalili, S. M., & Lewis, A. (2014). Grey wolf opti-
Harris, G. W., Mousset-Jones, P., & Daemen, J. (1999). Mea- mizer. Advances in Engineering Software, 69(3), 46–61.
surement of blast-induced rock movement in surface mines Mohamad, E. T., Armaghani, D. J., Momeni, E., & Abad, S. V. A.
by application of magnetic geophysics. Mining Technology N. K. (2015). Prediction of the unconfined compressive
IMM Transactions, 108, A172–A180. strength of soft rocks: A PSO-based ANN approach. Bulletin
Harris, G. W., Mousset-Jones, P., & Daemen, J. (2001). Blast of Engineering Geology and the Environment, 74(3), 745–757.
movement measurement to control dilution in surface mines. Morota, G., Ventura, R. V., Silva, F. F., Koyama, M., & Fer-
CIM Bulletin, 94(1047), 52–55. nando, S. C. (2018). Machine learning and data mining ad-
Hasanipanah, M., Amnieh, H. B., Arab, H., & Zamzam, M. S. vance predictive big data analysis in precision animal
(2016a). Feasibility of PSO–ANFIS model to estimate rock agriculture. Journal of Animal Science, 96(4), 1450–1550. h
fragmentation produced by mine blasting. Neural Computing ttps://doi.org/10.1093/jas/sky014.
and Applications, 30(4), 1015–1024. Nguyen, H., & Bui, X.-N. (2018). Predicting blast-induced air
Hasanipanah, M., Monjezi, M., Shahnazar, A., Armaghani, D. J., overpressure: A robust artificial intelligence system based on
& Farazmand, A. (2015). Feasibility of indirect determina- artificial neural networks and random forest. Natural Re-
tion of blast induced ground vibration based on support sources Research. https://doi.org/10.1007/s11053-018-9424-1.
vector machine. Measurement, 75, 289–297. Nguyen, H., Bui, X.-N., Tran, Q.-H., Le, T.-Q., & Do, N.-H.
Hasanipanah, M., Noorian-Bidgoli, M., Armaghani, D. J., & (2019a). Evaluating and predicting blast-induced ground
Khamesi, H. (2016b). Feasibility of PSO-ANN model for vibration in open-cast mine using ANN: A case study in
predicting surface settlement caused by tunneling. Engi- Vietnam. SN Applied Sciences, 1(1), 125.
neering with Computers, 32(4), 705–715. Nguyen, H., Drebenstedt, C., Bui, X.-N., & Bui, D. T. (2019b).
Hasanipanah, M., Shahnazar, A., Arab, H., Golzar, S. B., & Prediction of blast-induced ground vibration in an open-pit
Amiri, M. (2017a). Developing a new hybrid-AI model to mine by a novel hybrid model based on clustering and arti-
predict blast-induced backbreak. Engineering with Comput- ficial neural network. Natural Resources Research. https://doi.
ers, 33(3), 349–359. org/10.1007/s11053-019-09470-z.
Prediction of Blast-Induced Rock Movement During Bench Blasting 865

Photophoto. http://www.photophoto.cn. Retrieved March 26, Yang, R. L., & Kavetsky, A. (1990). A three dimensional model of
2019. muckpile formation and grade boundary movement in open
Preece, D., & Silling, S. A. (2016). Ore loss and dilution studies of pit blasting. International Journal of Mining and Geological
surface mineral blasting with 3D distinct element heave Engineering, 8, 13–34.
models. Sandia National Lab.(SNL-NM), Albuquerque, NM Yang, R. L., Kavetsky, A., & Mckenzie, C. K. (1989). A two-
(United States). dimensional kinematic model for predicting muckpile shape
Preece, D., & Taylor, L. (1989). Complete computer simulation of in bench blasting. International Journal of Mining and Geo-
crater blasting including fragmentation and rock motion. logical Engineering, 7, 209–226.
Albuquerque, NM (USA): Sandia National Labs. Yennamani, A. L. (2010). Blast induced rock movement mea-
Rad, H. N., Hasanipanah, M., Rezaei, M., & Eghlim, A. L. (2018). surement for grade control at the Phoenix mine. Reno:
Developing a least squares support vector machine for esti- University of Nevada.
mating the blast-induced flyrock. Engineering with Comput- Zhang, S. L. (1994). Rock movement due to blasting and its impact
ers, 34(4), 709–717. on ore grade control in Nevada open pit gold mines. Reno:
Rong, J. F., Lin, Y. Z., & Wang, Z. L. (2016). Geologic features of University of Nevada.
the Husab uranium deposit. Namibia. Acta Geologica Si- Zhang, H. F., & Lu, J. J. (2018). Geological characteristics, ore-
chuan, 36(1), 101–106. https://doi.org/10.3969/j.issn.1006-099 controlling factors and genesis analysis of Husab uranium
5.2016.01.022. deposit in Namibia. Global Geology, 37(1), 105–123.
Rovini, E., Maremmani, C., Moschetti, A., Esposito, D., & Ca- Zhou, J., Aghili, N., Ghaleini, E. N., Bui, D. T., Tahir, M., &
vallo, F. (2018). Comparative motor pre-clinical assessment Koopialipoor, M. (2019a). A Monte Carlo simulation ap-
in ParkinsonÕs disease using supervised machine learning proach for effective assessment of flyrock based on intelligent
approaches. Annals of Biomedical Engineering, 46(12), 2057– system of neural network. Engineering with Computers,.
2068. https://doi.org/10.1007/s00366-019-00726-z.
Samui, P., Sitharam, T. G., & Kurup, P. U. (2008). OCR predic- Zhou, J., Li, X. B., & Shi, X. Z. (2012). Long-term prediction
tion using support vector machine based on piezocone data. model of rockburst in underground openings using heuristic
Journal of Geotechnical & Geoenvironmental Engineering, algorithms and support vector machines. Safety Science,
134(6), 894–898. 50(4), 629–644.
Schalkoff, R. J. (1997). Artificial neural networks (Vol. 1). New Zhou, J., Li, X. B., Shi, X. Z., Wei, W., & Wu, B. B. (2011).
York: McGraw-Hill. Predicting pillar stability for underground mine using Fisher
Shi, X. Z., Zhou, J., Wu, B. B., Huang, D., & Wei, W. (2012). discriminant analysis and SVM methods. Transactions of
Support vector machines approach to mean particle size of Nonferrous Metals Society of China, 21(12), 2734–2743.
rock fragmentation due to bench blasting prediction. Trans- Zhou, J., Li, X. B., Wang, S. M., & Wei, W. (2013). Identification
actions of Nonferrous Metals Society of China, 22(2), 432– of large-scale goaf instability in underground mine using
441. particle swarm optimization and support vector machine.
Smola, A. J., & Schölkopf, B. (2004). A tutorial on support vector International Journal of Mining Science and Technology,
regression. Statistics and Computing, 14(3), 199–222. 23(5), 701–707.
Song, Q. J., Jiang, H. Y., Song, Q. H., Zhao, X. G., & Wu, X. X. Zhou, J., Nekouie, A., Arslan, C. A., Pham, B. T., & Hasani-
(2017). Combination of minimum enclosing balls classifier panah, M. (2019b). Novel approach for forecasting the blast-
with SVM in coal-rock recognition. PLoS ONE, 12(9), induced AOp using a hybrid fuzzy system and firefly algo-
e0184834. rithm. Engineering with Computers, 1–10.
Sujay, R. N., & Deka, P. C. (2014). Support vector machine Zhou, C., Yin, K. L., Cao, Y., Ahmed, B., Li, Y. Y., Catani, F.,
applications in the field of hydrology: A review. Applied Soft et al. (2018). Landslide susceptibility modeling applying
Computing Journal, 19(6), 372–386. machine learning methods: A case study from Longju in the
Taylor, S. L. (1995). Blast induced movement and its effect on Three Gorges Reservoir area, China. Computers & Geo-
grade dilution at the Coeur Rochester Mine. Reno: University sciences, 112, 23–37.
of Nevada. Zorlu, K., Gokceoglu, C., Ocakoglu, F., Nefeslioglu, H., & Aci-
Taylor, D. L., & Firth, I. (2003). Utilization of blast movement kalin, S. (2008). Prediction of uniaxial compressive strength
measurements in grade control. In: F. A. Camisani-Calzolari of sandstones using petrography-based models. Engineering
(Ed.), APCOM 2003. Proceedings of the 31st international Geology, 96(3–4), 141–158.
symposium on application of computers and operations re- Zou, Z. X., Tang, H. M., Xiong, C. R., Su, A. J., & Criss, R. E.
search in the minerals industries, May 14–16, Cape Town, (2017). Kinetic characteristics of debris flows as exemplified
South Africa (pp. 244–247). Johannesburg: The South Afri- by field investigations and discrete element simulation of the
can Institute of Mining and Metallurgy. catastrophic Jiweishan rockslide, China. Geomorphology,
Thornton, D. (2009). The implications of blast-induced movement 295, 1–15.
to grade control. In: Seventh international mining geology Zuo, R. (2017). Machine learning of mineralization-related geo-
conference (pp. 147–154). chemical anomalies: A review of potential methods. Natural
Vapnik, V. (1995). The nature of statistical learning theory. Berlin: Resources Research, 26(4), 1–8.
Springer.
Wang, S. (2004). Practical book for new technologies for
geotechnical engineering. Jilin: Yinsheng Audiovisual Press.