You are on page 1of 21

Natural Resources Research ( 2019)

https://doi.org/10.1007/s11053-019-09461-0

Original Paper

Prediction of Blast-induced Air Over-pressure in Open-Pit


Mine: Assessment of Different Artificial Intelligence
Techniques

Xuan-Nam Bui ,1,2,7 Hoang Nguyen ,1,2,7 Hai-An Le,3 Hoang-Bac Bui,4,5
and Ngoc-Hoan Do1,6

Received 18 November 2018; accepted 19 January 2019

Air over-pressure (AOp) is one of the products of blasting operations for rock fragmentation
in open-pit mines. It can cause structural vibration, smash glass doors, adversely affect the
surrounding environment, and even be fatal to humans. To assess its dangerous effects, seven
artificial intelligence (AI) methods for predicting specific blast-induced AOp have been
applied and compared in this study. The seven methods include random forest, support
vector regression, Gaussian process, Bayesian additive regression trees, boosted regression
trees, k-nearest neighbors, and artificial neural network (ANN). An empirical technique was
also used to compare with AI models. The degree of complexity and the performance of the
models were compared with each other to find the optimal model for predicting blast-
induced AOp. The Deo Nai open-pit coal mine (Vietnam) was selected as a case study
where 113 blasting events have been recorded. Indicators used for evaluating model per-
formances include the root-mean-square error (RMSE), determination coefficient (R2), and
mean absolute error (MAE). The results indicate that AI techniques provide better per-
formance than the empirical method. Although the relevance of the empirical approach was
acceptable (R2 = 0.930) in this study, its error (RMSE = 7.514) is highly significant to
guarantee the safety of the surrounding environment. In contrast, the AI models offer much
higher accuracies. Of the seven AI models, ANN was the most dominant model based on
RMSE, R2, and MAE. This study demonstrated that AI techniques are excellent for pre-
dicting blast-induced AOp in open-pit mines. These techniques are useful for blasters and
managers in controlling undesirable effects of blasting operations on the surrounding
environment.
KEY WORDS: Mining environment, Air over-pressure, Blasting, Artificial intelligence, Open-pit mine.

5
1 Center for Excellence in Analysis and Experiment, Hanoi
Department of Surface Mining, Mining Faculty, Hanoi Univer-
University of Mining and Geology, 18 Vien Street, Duc Thang
sity of Mining and Geology, 18 Vien Street, Duc Thang Ward,
ward, Bac Tu Liem District, Hanoi, Vietnam.
Bac Tu Liem District, Hanoi, Vietnam. 6
2 Faculty of Mining, Saint-Petersburg Mining University, Saint-
Center for Mining, Electro-Mechanical research, Hanoi Univer-
Petersburg, Russia.
sity of Mining and Geology, 18 Vien Street, Duc Thang ward, 7
To whom correspondence should be addressed; e-mail:
Bac Tu Liem District, Hanoi, Vietnam.
3 nguyenhoang@humg.edu.vn
Faculty of Oil and Gas, Hanoi University of Mining and Geol-
ogy, 18 Vien Street, Duc Thang ward, Bac Tu Liem District,
Hanoi, Vietnam.
4
Faculty of Geosciences and Geoengineering, Hanoi University of
Mining and Geology, 18 Vien Street, Duc Thang ward, Bac Tu
Liem District, Hanoi, Vietnam.

 2019 International Association for Mathematical Geosciences


X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

AOp (Segarra et al. 2010; Armaghani et al. 2015b).


INTRODUCTION
Therefore, empirical formulas often provide pre-
dicted results with low accuracy (Sawmliana et al.
Mining projects, especially their blasting oper-
2007; Khandelwal and Kankar 2011; Mohamad et al.
ations, often have significant impacts on the sur-
2012; Hajihassani et al. 2014; Armaghani et al.
rounding environment (Shields 1998; Michieka 2014;
2015a).
Shokri et al. 2016; Hresc et al. 2018). According to
To overcome the disadvantages of the methods
previous researchers, not all the energy released
mentioned above, artificial intelligence (AI) has
from blasting is involved in the fragmentation of
been applied in recent years to reduce the undesir-
rock. Researchers have shown that 80–85% of the
able effects caused by blasting operations (Iphar
total amount of explosive energy is wasted and that
et al. 2008; Nguyen et al. 2018a). In this respect,
it creates specific undesirable environmental effects
Mohamed (2011) successfully developed fuzzy logic
such as ground vibration, flyrock, air over-pressure
(FS) and artificial neural network (ANN) using 162
(AOp) and back-break (Khandelwal and Singh
blasting events for predicting blast-induced AOp in
2005; Verma and Singh 2011; Monjezi et al. 2012;
Assay Cement Company. Their study showed that
Hasanipanah et al. 2018). Among those side effects,
the FS model performed better than the ANN
AOp is considered to be the most harmful as it
model. Using a similar approach, Khandelwal and
causes structural vibration, smashes glass doors, af-
Kankar (2011) employed the support vector ma-
fects the surrounding environment and is even fatal
chine (SVM) algorithm to predict blast-induced
to humans (Bowen et al. 1968; Mayorga 1997; Raina
AOp in three limestone mines with 75 blasting
et al. 2004; Chafi et al. 2010; Armaghani et al.
events. Hajihassani et al. (2014) developed a new
2016b). Therefore, predicting blast-induced AOp is
hybrid model based on ANN and particle swarm
needed for reducing its undesirable effects on the
optimization (PSO) algorithm, i.e., PSO-ANN, for
surrounding environment.
predicting blast-induced AOp in four quartiles of
Specific straightforward methods of minimizing
Malaysia using 62 blasting events and yielded a
the impacts of blast-induced AOp on the surround-
promising result. Hasanipanah et al. (2016) investi-
ing environment increase safety distances and im-
gated and constructed three nonlinear models for
prove the quality and structure of works (Mayor and
predicting blast-induced AOp, including adaptive
Flanders 1990; Army 1998). Using barriers to mini-
neuro-fuzzy inference systems (ANFIS), FS, and
mize the effects of blast-induced AOp and ground
ANN. For their study, 77 blasting events in the
vibration has been the focus of many studies, and
Miduk Copper Mine (Iran) were used. A practical
proposals have been made to predict the damage to
technique was also employed to compare with these
buildings and to protect structures (Al-Hussaini and
models. Eventually, they found that ANFIS was the
Ahmad 1991; Andersen and Nielsen 2005; Remen-
most superior model. Recently, many scientists have
nikov and Rose 2007). However, these methods are
published their studies on predicting blast-induced
only qualitative and inadequate for controlling blast-
AOp with improved accuracy. Hasanipanah et al.
induced AOp.
(2017) successfully built a hybrid model based on
Many scholars have attempted to study and
PSO and support vector regression (SVR) algo-
propose experimental methods for estimating blast-
rithms, i.e., PSO-SVR. A variety of equation forms
induced AOp (Siskind et al. 1980; Loder 1987;
such as quadratic (Q), linear (L), and radial basis
McKenzie 1990; Hustrulid 1999; Rodrı́guez et al.
(RBF) kernels were applied to the PSO algorithm to
2007; Kuzu et al. 2009; Armaghani et al. 2016a).
optimize the SVR model. Multiple linear regression
However, experimental methods are mainly based
(MLR) was also used for comparison with the PSO-
on the relationship between the explosive charge per
SVR models. As a result, Hasanipanah et al. (2017)
delay/explosive charge capacity (W) and monitoring
found that the PSO-SVR-RBF was the best model in
distance (R). In fact, AOp can be influenced by
their study. In another study with the PSO algo-
many factors. According to Khandelwal and Kankar
rithm, AminShokravi et al. (2018) used three forms
(2011), blast geometry, W, R, and vegetation have an
of functions for PSO, including L, Q, and power (P)
impact on AOp. Siskind et al. (1980) concluded that
functions, to predict AOp. An ANN model was also
W and R are the most influential factors on AOp.
adopted to compare with the PSO algorithm. The
However, other factors such as over-charging, weak
authors revealed that the PSO-L model was the best
strata, and atmospheric conditions can also affect
model for predicting blast-induced AOp in their
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Table 1. Several AI techniques and their performance in predicting blast-induced AOp

Reference Technique Performance

Mohamed (2011) FS RMSE = 2.71; VAF = 99.95%


Khandelwal and Kankar (2011) SVM MAPE = 2.10; R2 = 0.855
Hajihassani et al. (2014) PSO-ANN R2 = 0.94
Hasanipanah et al. (2016) ANFIS RMSE = 2.329, R2 = 0.954, and VAF = 95.310
Hasanipanah et al. (2017d) PSO-SVR-RBF RMSE = 0.45; CC = 0.996
AminShokravi et al. (2018) PSO-L MSE = 4.33; R2 = 0.960
Alel et al. (2018) MSO-Rand-Mut RMSE = 2.180; R2 = 0.970; MAPE = 1.73%
Mahdiyar et al. (2018) Monte Carlo RMSE = 3379; R2 = 0.901; VAF = 90.028
Armaghani et al. (2018) GA-ANN RMSE = 0.049; R2 = 0.965; VAF = 96.380

RMSE root-mean-square error, VAF variance account for, MAPE mean absolute percentage error, R2 determination coefficient, CC
coefficient correlation, MSE mean-square error

case study. Alel et al. (2018) also conducted a similar efficient way of finding out the best AI technique for
study using ANNs based on PSO and multi-swarm predicting blast-induced AOp in this study. Seven
optimization (MSO) algorithms for blast-produced different AI techniques were implemented to pre-
AOp predictions in Hulu Langat Mine, Malaysia, dict blast-induced AOp. They include random forest
with 76 blasting events. Their results showed that the (RF) representing decision tree algorithms, SVR
MSO-Rand-Mut model was the best model in their representing nonlinear algorithms, Gaussian process
study. With another approach using AI, Mahdiyar (GP) representing Gaussian group of methods,
et al. (2018) successfully developed a Monte Carlo Bayesian additive regression trees (BART) repre-
simulation technique to predict blast-induced AOp senting Bayesian methods, boosted regression trees
in a quarry in Malaysia. MLR was applied to (BRT) serving boosting algorithms, k-nearest
establish the equation for Monte Carlo simulation. neighbors (KNN) posing for lazy algorithms, and
They found that MLR was quite suitable for Monte ANN rendering for neural networks techniques. An
Carlo simulation in blast-induced AOp. Genetic empirical method was also used to estimate AOp,
algorithm (GA) was also one of the algorithms used and its analysis results were compared with those of
by Armaghani et al. (2018) to optimize the ANN the developed AI models.
model for predicting blast-induced AOp. Accord-
ingly, the number of hidden neurons was optimized
by GA to improve the efficiency of the ANN model, STUDY AREA AND DATA USED
i.e., GA-ANN model. Experimental techniques,
MLR, and the single ANN model were also per- Study Area
formed to predict AOp and compared with the GA-
ANN model. Considering testing datasets, they According to the Vietnam National Coal and
concluded that GA-ANN was the best model with a Mineral Industries Holding Corporation Limited
promising result. Similar works can be found at that (VINACOMIN), there are 30 open-pit coal mines
paper here (Hasanipanah et al. 2017a, b, c). Several and 20 underground coal mines in Quang Ninh. The
results of the techniques as mentioned above are five largest open-pit coal mines produce 1–3 Mt/yr,
presented in Table 1. and the seven largest underground coal mines pro-
Review of literature showed that many re- duce over 2 Mt/yr. (Vinacomin 2010). The Deo Nai
searchers have studied and proposed prediction open-pit coal mine is one of the largest open-pit coal
models for blast-induced AOp. Those models, mines. It is located in Cam Pha city, Quang Ninh
however, are not applicable everywhere because the province, Vietnam (Fig. 1). The mine covers a total
impacts of blasting on the environment are different area of  6 km2 with proven reserves of 42.5 Mt and
from country to country (Nateghi et al. 2009). produces 2.5 Mt/yr.
Moreover, it is challenging to know that which AI The Deo Nai open-pit coal mine is covered by
technique is the best for predicting blast-induced sedimentary rocks of the Hon Gai Formation (T3n-
AOp. Therefore, ‘‘trial-and-error’’ procedure is an rhg) with a Late Triassic age (Fig. 2). The rocks
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 1. Location and view of the Deo Nai open-pit coal mine, Vietnam.

Figure 2. A simplified geological map of the surrounding area.

comprise different types such as conglomerate, Descriptive Data Statistics


gritstone, sandstone, siltstone, claystone, shale,
and coal seams. In general, these rocks are quite In this study, 113 blasting events were recorded
hard with ProtodiakonovÕs strength indexes (f) of from the Deo Nai open-pit coal mine. According to
8–14 (Protodiakonov et al. 1964) and specific Singh et al. (2008), the parameters that affect AOp
weights (c) of 2.62–2.65 t/m3 (Vinacomin 2015). can be divided into two groups, namely controllable
Therefore, an effective and high-performing and uncontrollable parameters. For controllable
blasting method is required to break the rocks in parameters, the settings like W, R, burden (B),
this coal mine. spacing (S), length of stemming (T), powder factor
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Figure 3. Instruments used for data collection in this study: (a) GPS for determining monitoring distance; (b)
Kanomax 2212 air quality meter for measurement of air humidity; (c) Blastmate III instrument for collecting AOp
values.

determined to ensure the floor of the surface is flat. S


Table 2. General specifications of Blastmate III for AOp
monitoring (source http://www.instantel.com) is one of the parameters of the hole pattern. In this
study, the square hole pattern was used for blasting
Features Parameters operations. Then, S was defined as B factor. The
Range 88–148 dB (500 Pa (0.072 PSI) Peak) coefficient of closure can be used to ensure a uni-
Resolution 0.25 Pa (0.0000363 PSI) form level of fragmentation. T is defined by hole
Accuracy ± 10% or ± 1 dB, whichever is larger, depth and length of explosive. It must be determined
between 4 and 125 Hz to be safe and not subject to the ejection of stem-
Frequency Response 2–250 Hz between  3 dB roll-off points
ming or rifling. Eventually, W was calculated based
on the number of bore-hole and explosive charge
capacity per bore-hole. The AOp values were mea-
sured by Blastmate III instrument from Instantel,
(P), blast-hole depth, delay times, and blast-hole Canada (Fig. 3c), which is highly rated for its data
diameter can be tuned by blast engineers. For acquisition and accuracy in regulating AOp (Ta-
uncontrollable parameters, such as geological and ble 2). For monitoring AOp, a microphone was
geophysical conditions, characteristics of rock mass, connected with the Blastmate III device and was
cannot be adjusted by engineers (Khandelwal and located at AOp monitoring locations. The locations
Kankar 2011; Khandelwal and Singh 2013). How- to be monitored may be behind, in front of or on the
ever, due to the lacks of information on geological side of the explosion site depending on the actual
and geophysical conditions as well as characteristics conditions. According to recommendations of
of rock mass, uncontrollable parameters were not Instantel, the microphone should be placed in the
often used to predict blast-induced AOp. Also, direction of the blast site to ensure the accuracy le-
Nguyen et al. (2018a) were recommended that air vel as shown in Figure 4. GPS positioning system
humidity (RH) should be used as an input variable (Fig. 3a) was used to calculate R. RH was measured
for estimating blast-produced AOp. Therefore, se- by JapanÕs Kanomax 2212 air quality meter
ven parameters including W, R, B, S, T, P and RH (Fig. 3b), which is designed with a humidity sensor
were used to predict blast-induced AOp in this that can measure accurately in 15 s (Azuma et al.
study. 2018; Fan et al. 2018). The remaining parameters
Blasting parameters of the blast pattern were were collected from 113 blast patterns of the mine
defined by Hustrulid et al. (2013). P is determined (Fig. 5). Table 3 summarizes the datasets used in
by rock density, rock hardness, hole diameter, and this study.
average size of rock. B depends on explosive and For each blast pattern of the mine, the hardness
borehole parameters such as the mass of charge, of rock mass was determined before calculating the
powder factor, bench height, and hole depth. It is parameters of the blast pattern. As mentioned, the
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 4. Blast-induced AOp monitoring method.

hardness of rock mass in this mine lies in the range


of 8–14 according to ProtodiakonovÕs strength index.
Then, the parameters of blast design were defined,
including bench height, diameter of borehole, depth
of borehole, B, S, number of borehole, T, P, and W.
For this mine, the bench height lies in the range of
12–15 m, the depth of borehole lies in the range of
17–18 m. Also, to ensure safety and slope stability,
the bench angle was chosen equal to 70. Based on
the characteristics of the rock mass, the diameter of
Figure 5. Structure of blast pattern. the borehole of 250 mm was selected for this mine.

OVERVIEW OF ARTIFICIAL
Table 3. Characteristics of the data used
INTELLIGENCE TECHNIQUES USED
W R B S

Min.: 1171 Min.: 84.8 Min.: 7.400 Min.: 7.400


Random Forest
1st Qu.: 9476 1st Qu.: 336.7 1st Qu.: 7.900 1st Qu.: 7.700
Median: 13,019 Median: 453.1 Median: 8.100 Median: 7.800 RF is one of the ensemble machine learning
Mean: 12,815 Mean: 454.6 Mean: 8.088 Mean: 7.835 algorithms proposed by Breiman (2001). It belongs
3rd Qu. :16,489 3rd Qu.: 588.0 3rd Qu.: 8.300 3rd Qu.: 8.000 to the group of decision trees algorithms and can
Max.: 21,771 Max.: 806.3 Max.: 8.900 Max.: 8.300
solve both classification and regression problems.
T P RH AOp RF involves the construction (growing) of multiple
decision trees via bootstrap aggregation (bagging)
Min.: 5.900 Min.: 0.3300 Min.: 75.00 Min.: 91.13
(Effron and Tibshirani 1993; Breiman 1999). It is
1st Qu.: 6.600 1st Qu.: 0.3900 1st Qu.: 82.00 1st Qu.: 113.93
Median: 6.800 Median: 0.4200 Median: 85.00 Median: 121.09 based on the combination of a set of results from
Mean: 6.851 Mean: 0.4195 Mean: 84.99 Mean: 120.77 many decision trees to make a final decision. Each
3rd Qu.: 7.100 3rd Qu.: 0.4400 3rd Qu.: 88.00 3rd Qu.: 129.14 tree is trained by randomly selecting variables and
Max.: 7.900 Max.: 0.5000 Max.: 96.00 Max.: 143.83 data samples from an initial training dataset. For
predicting blast-induced AOp, RF was applied as
follows:
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

 The number of trees was determined to en- Review of literature showed that SVR had been
sure the richness of the forest; successfully implemented for blasting problems in
 Bootstrap with replacement from the original many places (Khandelwal and Kankar 2011; SHI
AOp training dataset. The remaining values, et al. 2012; Amini et al. 2012; Hasanipanah et al.
which were used for validation, are called 2015). In this study, SVR was employed to predict
out-of-bag (OOB) data; blast-induced AOp in the Deo Nai open-pit coal
 For each bootstrap sample, a non-pruning mine using radial primary kernel function (Eq. 3)
regression was developed with modifications and compared with the other AI methods.
at each node;
 At each bootstrap iteration, OOB was used
to predict AOp and averaged from all trees; Gaussian Process
 Performance indices (i.e., RMSE, R2, MAE)
were used to evaluate the performance of GP is one of the nonparametric models in AI. It
predicted AOp values on OOB. is a set of random variables and finite elements with
a Gaussian distribution (Rasmussen 2004). GP is
Longjun et al. (2011) also used RF for predict- determined based on mean function, h(x), and
ing another problem in blasting operation, i.e., covariance function, c(x, x¢). It can be described as:
ground vibration with satisfactory result. However,
f ðxÞ  GPðhðxÞ; cðx; x0 ÞÞ ð5Þ
in this study, RF was used for predicting blast-in-
duced AOp. For regression, the GP proceeds to encrypt the
uncertainty before executing the training. After the
relationship between the function and the data is
Support Vector Regression determined, the Bayes rule is used to update the
beliefs through the function and the posterior dis-
SVM is a machine learning algorithm based on tribution can be computed (Särkkä et al. 2017).
the principle of minimizing structural risk to gener- The literature shows that GP has never been
alize a limited number of samples better and was used to predict blast-induced AOp. Hence, its
proposed by Cortes and Vapnik (1995). Like RF, application for this purpose is investigated and
SVM can solve both classification and regression considered here with the polynomial kernel func-
problems. For regression problems, an SVM is called tion, which is described as:
SVR. SVR relies on a subset of training dataset to
ðxT x0 þ aÞn
build the forecasting model (Basak et al. 2007). cðx; x0 Þ ¼ ; a  0; n2N
For predicting blast-induced AOp, SVR can be ððk xk2 þaÞðkx0 k2 þaÞÞn=2
performed based on one of the following kernel ð6Þ
functions (Prashanth and Nimaje 2018), thus:
Polynomial kernels without a normalizing
 Linear kernel: denominator can be determined to produce a finite-
dimensional feature space of polynomials of the to-
K ðx; yÞ ¼ x  y ð1Þ tal degree  n ðif a[0Þ(Seeger 2004). The regular-
 Polynomial kernel: ization operator for polynomial kernels can be found
in the literature (e.g., Smola et al. 1998; Schreiter
K ðx; yÞ ¼ ½ðx  yÞ þ 1d ; d ¼ ð1; 2; . . .Þ ð2Þ et al. 2016; Chen and Wang 2018).
 Radial primary kernel function:

kx  yk2 Bayesian Additive Regression Trees


K ðx; yÞ ¼ exp½  ð3Þ
r2
BART is an algorithm based on the Bayesian
 Two-layer neural kernel: classification and regression tree algorithm proposed
K ðx; yÞ ¼ tanh½aðx  yÞ  d ð4Þ by Chipman et al. (2010). For regression, BART
performs a random sum of decision trees and is
described as:
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

X
N of supervised learning problems (Altman 1992).
f ðxÞ ¼ DTn ðxÞ; x 2 RP ð7Þ KNN is also called an instance or memory-based
n¼1
learning algorithm. In a regression problem, the
Each regression tree DTn(x) is constructed by output of a data point is equal to that of the nearest
the binary tree structure BTn consisting of topology known data point (in the case of k = 1) or is the
and tree splitting rules, and a vector of parameters weighted average of the production of the closest
related to binary trees BTn (Linero 2018). BART locations, or by a relationship based on the distance
documentation can be found in Chipman et al. to the nearest point. In a nutshell, KNN is an algo-
(2010). BART has been proven to show promising rithm for finding the output of a new data point
results in many fields (Hill 2011; Bleich and Kapel- based solely on the information of k data points in
ner 2014; Bleich et al. 2014). However, it has never the closest training set (k-neighborhood); it is not
been used to predict blast-induced AOp. Thus, a interested in the interferences in the surrounding.
BART model was developed in this study for eval- Details of the KNN algorithm can be found in Song
uating its effectiveness in predicting blast-induced et al. (2017) and Chae et al. (2018).
AOp. KNN has not been used to predict blast-induced
AOp in open-pit mines. Therefore, its application is
investigated in this study to predict blast-induced
Boosted Regression Trees AOp. Determining the number of neighbors and
setting up the KNN model are discussed in the next
BRT is one of the ensemble methods for section.
improving the accuracy of single models (Elith et al.
2008). It is combined with the advantages of two
algorithms including classification and regression Artificial Neural Network
trees (Trevor et al. 2009) and boosting and was first
introduced by Schapire (2003). It was then widely ANN is an information processing model. It is
developed in the statistical community as an ad- based on the activity of the nervous system of an
vanced regression technique (Friedman et al. 2000). organism and includes a large number of intercon-
BRT is capable of combining the outstanding nected neurons that process information. ANN
advantages of tree-based methods, handling various models act like a human brain. They can learn
predictive variables and losing data. BRT processes through experience (through training), store
data without transforming data and ignoring the knowledge and use it to predict unseen data
outliers. Therefore, it seems to be a powerful tool (Schalkoff 1997).
for nonlinear issues. The general architecture of an ANN includes
Although the performance of BRT is evident in three parts, namely an input layer, hidden layer(s),
various areas (Tyree et al. 2011; Müller et al. 2013), and output (Schalkoff 1997) as shown in Figure 6. In
BRT has never been used to predict the undesirable ANN, the number of hidden layers and neurons in
effects of blasting operations in the mining field. each hidden layer depends on many factors such as
Hence, in this study, BRT was used to predict blast- the number of incomes and outcomes of the net-
induced AOp for comparing and evaluating its work, the number of cases in the sample, the noise of
effectiveness with the other models. target data, the complexity of the error function,
network architecture, and the network training
algorithm.
At the input layer, the neurons receive biased
k-Nearest Neighbors
(weighted) input information; they then process and
send the information to neurons in the first hidden
KNN is one of the most straightforward super-
layer through a transfer function. These neurons
vised learning algorithms in AI. It was proposed by
process and compute the data received from the
Altman (1992) and was classified as lazy machine
input layer and then send the results to the second
learning because it does not learn anything from
hidden layer through a transfer function. A thresh-
training datasets. In KNN, all computations are
old value, instead of a transfer function, is some-
conducted to predict the results of new data. KNN
times used to control the output of neurons in a
can be applied for classification and regression types
given layer before the output is moved to the next
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

considered as training dataset in the first part; the


remaining 20% (16 observations) was viewed as a
testing dataset in the second part.
Performance indices including RMSE, R2, and
MAE were employed to evaluate the performance
of the predictive models on the training and testing
datasets. The three indices are computed as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X n
RMSE ¼ ðyi  ^ yi Þ 2 ð8Þ
n i¼1

P
Figure 6. General structure of the ANN model for predicting yi Þ 2
ðyi  ^
blast-induced AOp. 2 i
R ¼1 P ð9Þ
yÞ2
ðyi  
i

layer. The process continues until the results are 1X n

passed to the output layer and the final output is MAE ¼ jyi  ^
yi j ð10Þ
n j¼1
provided (Zerguine et al. 2001).
The outcomes of ANN largely depend on the
where n is the total number of data, yi is measured
training process used, which can be one of two types
value, ^yi is predicted value, and 
y is mean of mea-
namely supervised learning and unsupervised
sured values. To evaluate the stability of the models,
learning (Perez et al. 1994). In this study, the
the different between RMSE and MAE was used
supervised learning method was applied based on
(Willmott and Matsuura 2005). In the most optimal
the input data and the requirements of the output
model, R2 should equal to 1, and both RMSE and
data.
MAE should be equal to 0. Also, the lowest value of
ANN has been successfully applied in various
the difference between RMSE and MAE corre-
aspects of the mining field such as air pollution
sponds to the most stable model.
management (Asif et al. 2018), considering content
Of the seven AI techniques used in the study
distributions of impurities in limestone (Koike and
for predicting blast-induced AOp, the Box-Cox
Matsuda 2003), predicting blasting issues (Mohamed
transformation (Sakia 1992) and 10-fold cross-vali-
2009; Monjezi et al. 2013; Hajihassani et al. 2014;
dation (Kohavi 1995) were applied to avoid over-
Bakhtavar et al. 2017), mineral system analysis
fitting or underfitting for the RF, SVR, GP, BART,
(Tessema 2017), and prediction of slope stability
BRT, KNN models. The min–max normalization
(Chakraborty and Goswami 2017). In this study, an
method (Jayalakshmi and Santhakumaran 2011) was
ANN model was developed for predicting blast-in-
applied to avoid overfitting or underfitting for the
duced AOp in the Deo Nai open-pit coal mine. The
ANN models with the scaling in the intervals
results of this model are compared to those of other
[ 1,1].
AI techniques.

Empirical Model
ARTIFICIAL INTELLIGENCE MODELS
FOR PREDICTING AOP
To evaluate the effectiveness of forecasting
models, an empirical technique proposed by the
For generating AOp predictive models using
United States Bureau of Mines (USBM) was used to
the seven AI methods, the original data were split
predict AOp in open-cast mine based on the rela-
into two parts. Review of literature showed that the
tionship between R and W, and site factors (Haji-
most commonly used train/test ratio is 80:20, which
hassani et al. 2014; Armaghani et al. 2016b). The
is also referred to as the Pareto principle (Nick
relationship between W and R is determined
2008). The 80:20 is indeed a good starting point
through the SD values, which are calculated as:
according to Swingler (1996). Therefore, 80% of the
whole data (approximately 97 observations) was SD = RW 0:33 ð11Þ
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 7. Grid search for optimal RF model in predicting blast-induced AOp.

where R denotes monitoring distance, m; W is 2018). Although the default of ntree was 500 (Brei-
maximum explosive charge capacity, kg; SD is scaled man 2001), this study employed a trial-and-error
distance factor (m kg0.33). procedure with ntree of 500, 1000, 1500, and 2000 to
From the scaled distance, AOp can be calcu- observe the performance of the RF model. A grid
lated as follows: search technique was applied to define the optimal
value of mtry (Ogutu et al. 2011). The training da-
AOp ¼ gðSDÞb ð12Þ taset for RF models is the same as those used in the
empirical technique. The results showed that the
where AOp is measured in dB, g and b are site optimal RF model on training dataset with ntree =
factors and computed by regression analysis method. 2000 and mtry = 5 (Fig. 7).
Depending on the specific conditions of each mine,
site factors are different.
To determine site factors, this study ran a Support Vector Regression
multivariate regression analysis on the training da-
taset by using SPSS version 18.0 (Carver and Nash For SVR, r and cost (C) are two parameters
2011). Note that the training dataset used for the used to control the quality of the AOp predictive
empirical technique is the same as the dataset used model. The training dataset used to develop the
for the seven AI models. The regression analysis SVR model includes 97 blasting events, which are
results showed that the empirical model reached its used in the empirical and RF models. The grid
optimal value with g = 333.384 and b = 0.345. search technique was also utilized to determine the
Hence, the experimental formula for predicting optimal values for hyper-parameters of SVR with r
AOp in this site study is: in the range of 0–1, and C in the range of 0.25–5.
Finally, an optimal SVR model on training dataset
AOp ¼ 333:384ðSDÞ0:345 ð13Þ
was found with r = 0.005 and C = 5. The develop-
ment of the SVR model for predicting blast-induced
AOp in this study is illustrated in Figure 8.
Random Forest

In RF, the two main parameters that determine Gaussian Process


the performance of the model are the number of
trees in the forest (ntree) and the number of vari- With GP, only one parameter, i.e., d, is used to
ables randomly sampled as candidates at each split adjust the modelÕs performance. The same training
(mtry) (Liaw and Wiener 2002; Nguyen and Bui dataset (97 blasting events) was used to develop the
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Figure 8. Determination of optimal SVR model in predicting blast-induced AOp.

GP model like empirical, RF, and SVR models. The defined in the range of 1–5; and nu parameter was
grid search was also employed to find the optimal also found in the range of 1–5 for optimal value. As
value of d in the range of 0–1. The result shows that a result, the optimal values of hyper-parameters
the GP model on training dataset was optimized at were reached at n = 200, p = 4, a = 0.945, x = 5 and
d = 0.02. Figure 9 demonstrates the process of nu = 1 (Fig. 10).
determining the optimal GP model in this study.

Boosted Regression Trees


Bayesian Additive Regression Trees
In the BRT model, three hyper-parameters
For the BART model, five hyper-parameters including boosting iterations (mstop), max tree
including the number of trees (n), prior boundary depth (maxdepth), and shrinkage (sh) determine the
(p), base terminal node hyper-parameter (a), power performance of the predictive model. Using the
terminal node hyper-parameter (x), and degrees of same approach employed in previous models,
freedom (nu) were used to control the performance training datasets and grid search technique were still
of the model. It is difficult to know which BART used for searching the optimal BRT model. 10-fold
model is optimal with the values of hyper-parame- cross-validation with three repeats was also used for
ters. Thus, the grid search technique was again ap- resampling to improve the accuracy of the model. To
plied to identify the optimal values of hyper- find out the optimal BRT model for predicting blast-
parameters in the BART model. The same training induced AOp, mstop was searched in the range of
dataset in the empirical, RF, SVR, and GP models 50–500, maxdepth in the range of 1–10, and nu in the
was also used for the development of the BART range of 0.1–0.5. The search process ends with
model. Accordingly, n parameter was set equal 200 mstop = 50, maxdepth = 2 and nu = 0.1, which were
to ensure the abundance of regression trees; p was the optimal values for the BRT model in this study.
searched in the range of 1–10; a was recommended Figure 11 illustrates the development of the BRT
to default to 0.9, 0.945, or 0.990; x optimal value was model for predicting AOp in this study.
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 9. Search for the optimal value d for the GP model.

Figure 10. Optimal values of hyper-parameters in the BART model.


Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Figure 11. Performance of the BRT on the training datasets with various values of the hyper-parameters.

k-Nearest Neighbors similar regression problems as those in this study


(Nguyen et al. 2018b). More hidden layers will in-
The KNN algorithm only collects results from crease the processing time and the calculation of the
neighbors to make its decision. Accordingly, the ANN (Nguyen et al. 2018a). According to Easley
number of neighbors (k) is the only factor that et al. (2018), a deep neural network with more than
determines the accuracy of the forecasting model. A one hidden layer (multiple hidden layers) is a good
grid search technique was established with k set in candidate for increasing the accuracy of the model.
the range of 1–50 for building the KNN model. The Therefore, we developed the ANN models with one,
same training dataset used in previous models (RF, two, and three hidden layer(s) to test their perfor-
SVR, GP, BART, BRT) was also applied to the mance. Also, the number of hidden neurons in each
development of the KNN model. Finally, a KNN hidden layer should also be designed appropriately.
model with k = 8 was identified as the optimal In some cases, using only a few hidden neurons will
model with the smallest RMSE (Fig. 12). lead to underfitting while using too many hidden
neurons will lead to overfitting (Lu et al. 2017). To
prevent the creation of ANN models that are too
Artificial Neural Network complicated, a trial-and-error procedure with the
number of hidden neurons in each hidden layer was
With ANN, unlike the other AI models (RF, conducted with the hidden neurons lie in the range
SVR, GP, BART, BRT, KNN), the performance of of 5–9. Eight ANN models were established and
the model was determined by the structure of the their performances were computed based on both
ANN and the training algorithm. In the structure of the training and testing datasets (Table 4).
the ANN, the numbers of hidden layers and neurons From Table 4, it can be seen that the ANN
are the decisive factors for the performance of the models with two and three hidden layers, particu-
predictive model (Gurney 2014). In theory, an ANN larly the ANN 7-9-7-5-1 model, seem to offer better
model with one hidden layer is capable of solving performance than the ANN models with one hidden
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 12. Searching the optimal value for the KNN model.

Table 4. ANN models for predicting blast-induced AOp in this RESULTS AND DISCUSSIONS
study

ANN models Training dataset Testing dataset


In AI, it is difficult to know which model is the
best for predicting blast-induced AOp. Therefore, a
RMSE R2 MAE RMSE R2 MAE ‘‘trial-and-error’’ procedure was performed with the
ANN 7-5-1 1.723 0.975 1.080 3.540 0.942 2.375 seven models mentioned above to evaluate their
ANN 7-7-1 1.875 0.971 1.229 3.400 0.945 2.281 effectiveness in predicting blast-induced AOp. To
ANN 7-9-1 1.725 0.976 1.088 3.485 0.951 2.134 assess the performance of the predictive models, the
ANN 7-7-5-1 1.592 0.979 0.944 3.713 0.928 1.886 testing dataset was used as unseen data for each
ANN 7-5-7-1 1.755 0.975 1.049 2.980 0.950 1.609
model through performance indices, including
ANN 7-9-6-1 1.801 0.973 1.131 2.849 0.951 1.795
ANN 7-9-7-5-1 2.006 0.967 1.365 2.571 0.961 1.697 RMSE, R2, and MAE. Note that the same testing
ANN 7-9-5-7-1 1.739 0.975 1.148 2.854 0.948 1.858 dataset was used for all predictive models. Accord-
ingly, the performance indices of the predictive
The best ANN model was shown in bold
models on the testing dataset are computed in
Table 5.
layer. Based on RMSE, R2, and MAE, it can be Based on Table 5, the empirical technique
concluded that although the ANN 7-9-7-5-1 model provided the poorest performance on the testing
yielded the poorest performance on the training dataset. In contrast, the seven AI techniques appear
dataset, it achieved the best performance on the to be much more efficient. However, closer inspec-
testing dataset. The testing dataset as mentioned tion of Table 5 indicates that R2 of the USBM
above was used as unseen data for evaluating the empirical model was better than those of the BRT
performance of the models. Therefore, the perfor- and KNN models. This shows that the relevance of
mance on the testing dataset was considered as the data with empirical methods might be higher than
performance in practical engineering. Based on with AI techniques in some cases. However, its error
RMSE, R2, and MAE, the ANN 7-9-7-5-1 model was was highly significant to guarantee the safety of the
chosen as the best ANN model for predicting blast- surrounding environment with an RMSE of 7.514.
induced AOp in ANN technique in this study. Fig- Figure 14 shows the relationship between measured
ure 13 illustrates the structure of the selected ANN and predicted values for each of the AI predictive
model in the ANN technique for predicting blast- models based on the testing dataset.
induced AOp.
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Figure 13. Construction of the ANN model for predicting blast-induced AOp in this study.

Table 5. Performance indices of predictive models on the training and testing datasets

Model Training datasets Testing datasets

RMSE R2 MAE RMSE R2 MAE RMSE–MAE

Empirical 3.853 0.881 2.501 7.514 0.930 4.585 2.929


RF 1.700 0.972 0.969 2.807 0.943 1.171 1.636
SVR 1.653 0.978 1.090 3.197 0.948 1.847 1.350
GP 2.320 0.973 1.560 2.656 0.949 1.652 1.004
BART 1.796 0.973 1.211 2.704 0.945 1.447 1.257
BRT 1.967 0.969 1.200 3.742 0.898 1.643 2.099
KNN 3.360 0.937 2.578 3.991 0.890 2.596 1.395
ANN 2.006 0.967 1.365 2.571 0.961 1.697 0.874

The best performance was shown in bold type. In which, RMSE representing the root-mean-square error, R2 representing the fitness of the
model, MAE representing mean absolute error, RMSE–MAE representing the stable of the model

Based on Table 5 and Figure 14, it can be seen model (Fig. 9). Therefore, the building process of
that the ANN model performed very well in pre- the GP model was more straightforward than that of
dicting blast-induced AOp with an RMSE of 2.571, the ANN model. In contrast, Figure 10 shows that
R2 of 0.961, and MAE of 1.697 on the testing data- there were five hyper-parameters used to control the
set. Also, the lowest value of RMSE-MAE of 0.874 performance of the BART model for predicting
on the testing dataset showed that the ANN model blast-induced AOp in this study. Thus, the building
was the most stable model among the models used in process of the BART model was more complicated
this study. The GP model also performed well in than those of the ANN and GP models. However,
predicting blast-induced AOp in this study with an the BART model yielded an RMSE of 2.704, R2 of
RMSE of 2.656, R2 of 0.949, and MAE of 1.652. The 0.945, and MAE of 1.447, implying that its perfor-
results indicate that the performance of the GP mance was poorer than those of the ANN and GP
model is slightly poorer than that of the ANN models. The RF model also performed well in pre-
model. However, only one hyper-parameter, i.e., d, dicting blast-induced AOp in this study with an
was used to control the performance of the GP RMSE of 2.807, R2 of 0.943, and MAE of 1.171.
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

Figure 14. Correlation between measured and predicted values of the different AI predictive models based on the testing
dataset.

Accordingly, the performance of the RF model is modelÕs performance (Fig. 7) and ntree recom-
slightly poorer than that of the BART model. The mended for this model was 2000 for higher perfor-
RF model used two hyper-parameters to control the mance in predicting blast-induced AOp. For SVR
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Figure 15. The relative importance of various input parameters to AOp.

model, there were two hyper-parameters used to cant factors, which have a significant influence on
control the modelÕs performance (Fig. 8). The the performance of the forecasting model (Fig. 15).
building process of the SVR model was simpler than The results show that RH, in addition to W and R,
those of the BART and ANN models. However, the was also an influence parameter on blast-induced
results in Table 5 show that the performance of the AOp in practical engineering.
SVR model was poorer than those of ANN, GP,
BART, and RF models based on the RMSE, R2, and
MAE. The BRT model did not perform well in CONCLUSIONS AND REMARKS
predicting blast-induced AOp in this study. Its per-
formance was worse with an RMSE of 3.742, R2 of Blasting is the most effective method to break
0.898, and MAE of 1.643. Furthermore, the building rocks in open-pit mines. However, its undesirable
process of the BRT was complex with three hyper- effects such as ground vibration, air over-pressure
parameters used to control the performance of the (AOp), fly rock, back-break, dust and toxic are
model. Among the seven AI predictive models, the evitable. Among these side effects, AOp is consid-
KNN model was the poorest, which yielded the most ered the most dangerous. Therefore, it needs to be
mediocre performance with an RMSE of 3.991, R2 predicted and tightly controlled to minimize its
of 0.890, and MAE of 2.596. potential adverse impacts on the surrounding envi-
Review of literature showed that W and R were ronment. The results of this study showed that:
the most influential parameters on blast-induced
AOp (Armaghani et al. 2016b; Hasanipanah et al.  AI techniques are advanced methods in pre-
2017d). Many studies used these parameters (W, R) dicting blast-induced AOp. They are able to
as input variables for predicting blast-induced AOp explain very well the nonlinear relationships
(Alel et al. 2018; Armaghani et al. 2018; Faradonbeh between the input variables in predicting
et al. 2018). However, seven input variables namely blast-induced AOp. Based on the predictive
W, R, B, S, T, P, and RH were considered in this models, blast designs can be adjusted to re-
study for predicting blast-induced AOp. Therefore, duce the undesirable effects of AOp on the
a procedure for analyzing the degree of influence of surrounding environment. However, the
independent variables was performed to identify the development of AI models is often involved.
significant factors affecting the quality of the model.  ANN is a powerful AI technique for pre-
Olden algorithm was applied to define the relative dicting blast-induced AOp as well as the
importance of input variables in ANN. It was pro- other blasting issues in open-pit mines. In this
posed by Olden et al. (2004) based on the Garson study, the ANN models with one, two, and
algorithm (Garson 1991; Goh 1995). Olden algo- three hidden layers were employed for AOp
rithm allows evaluating the contributions of input predictions. The results showed that the
variables for the ANN model with multiple hidden ANN 7-9-7-5-1 model yielded the best per-
layers. As a result, W, R, and RH were the signifi-
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

formance. It can be applied in practical AminShokravi, A., Eskandar, H., Derakhsh, A. M., Rad, H. N., &
Ghanadi, A. (2018). The potential application of particle
engineering to predict blast-induced AOp. swarm optimization algorithm for forecasting the air-over-
 The other AI techniques such as RF, SVR, pressure induced by mine blasting. Engineering with Com-
GP, BART, BRT, and KNN are also models puters, 34(2), 277–285.
Andersen, L., & Nielsen, S. R. (2005). Reduction of ground
that should be considered in different condi- vibration by means of barriers or soil improvement along a
tions and areas. Their performance can be railway track. Soil Dynamics and Earthquake Engineering,
optimized more in a specific case or when 25(7–10), 701–716.
Armaghani, D. J., Hajihassani, M., Marto, A., Faradonbeh, R. S.,
they are combined. & Mohamad, E. T. (2015a). Prediction of blast-induced
 W, R, and RH are the significant factors, air overpressure: A hybrid AI-based predictive model.
which have a high influence on the perfor- Environmental Monitoring and Assessment, 187(11),
666.
mance of an AOp predictive model. They Armaghani, D. J., Hajihassani, M., Sohaei, H., Mohamad, E. T.,
should be used for predicting blast-induced Marto, A., Motaghedi, H., et al. (2015b). Neuro-fuzzy tech-
AOp in open-pit mines. Also, other meteo- nique to predict air-overpressure induced by blasting. Ara-
bian Journal of Geosciences, 8(12), 10937–10950.
rological factors such as wind speed and Armaghani, D. J., Hasanipanah, M., Mahdiyar, A., Majid, M. Z.
direction should also be taken into consider- A., Amnieh, H. B., & Tahir, M. M. (2016a). Airblast pre-
ation when anticipating blast-induced AOp in diction through a hybrid genetic algorithm-ANN model.
Neural Computing and Applications, 29, 1–11.
future works. Armaghani, D. J., Hasanipanah, M., Mahdiyar, A., Majid, M. Z.
A., Amnieh, H. B., & Tahir, M. M. (2018). Airblast predic-
Although different AI techniques have been tion through a hybrid genetic algorithm-ANN model. Neural
Computing and Applications, 29(9), 619–629.
reviewed and evaluated for predicting blast-induced Armaghani, D. J., Hasanipanah, M., & Mohamad, E. T. (2016b).
AOp in this study, they should be considered further A combination of the ICA-ANN model to predict air-over-
when applied to other areas. Also, wind speed and pressure resulting from blasting. Engineering with Computers,
32(1), 155–171.
wind direction should be investigated and discussed Army, U. (1998). Technical manual design and analysis of hard-
further in forecasting blast-induced AOp for the ened structures to conventional weapons effects. Army TM5-
future works. 855-1, Washington DC.
Asif, Z., Chen, Z., & Zhu, Z. H. (2018). An integrated life cycle
inventory and artificial neural network model for mining air
pollution management. International Journal of Environ-
mental Science and Technology. https://doi.org/10.1007/s1376
2-018-1813-9.
ACKNOWLEDGMENTS Azuma, K., Ikeda, K., Kagi, N., Yanagi, U., & Osawa, H. (2018).
Physicochemical risk factors for building-related symptoms in
This research was supported by Hanoi Univer- air-conditioned office buildings: Ambient particles and com-
bined exposure to indoor air pollutants. Science of the Total
sity of Mining and Geology (HUMG) and Ministry Environment, 616, 1649–1655.
of Education and Training of Vietnam (MOET). We Bakhtavar, E., Nourizadeh, H., & Sahebi, A. (2017). Toward
also thank the Center for Mining, Electro-Mechan- predicting blast-induced flyrock: a hybrid dimensional anal-
ysis fuzzy inference system. International Journal of Envi-
ical Research of HUMG. ronmental Science and Technology, 14(4), 717–728.
Basak, D., Pal, S., & Patranabis, D. C. (2007). Support vector
regression. Neural Information Processing-Letters and Re-
views, 11(10), 203–224.
REFERENCES Bleich, J., & Kapelner, A. (2014). Bayesian additive regression
trees with parametric models of heteroskedasticity. arXiv
preprint arXiv:1402.5397.
Alel, M. N. A., Upom, M. R. A., Abdullah, R. A., & Abidin, M. Bleich, J., Kapelner, A., George, E. I., & Jensen, S. T. (2014).
H. Z. (2018). Optimizing blastingÕs air overpressure predic- Variable selection for BART: An application to gene regu-
tion model using swarm intelligence. In Journal of Physics: lation. The Annals of Applied Statistics, 8, 1750–1781.
Conference Series (vol. 995, vol. 1, pp. 012046). IOP Pub- Bowen, I. G., Fletcher, E. R., & Richmond, D. R. (1968). Esti-
lishing. mate of manÕs tolerance to the direct effects of air blast.
Al-Hussaini, T. M., & Ahmad, S. (1991). Design of wave barriers Report period. Washington, D.C.: Defense Atomic Support
for reduction of horizontal ground vibration. Journal of Agency, Lovelace Foundation for Medical Education and
Geotechnical Engineering, 117(4), 616–636. Research Albuquerque NM.
Altman, N. S. (1992). An introduction to kernel and nearest- Breiman, L. (1999). Random forests. Technical Report TR567
neighbor nonparametric regression. The American Statisti- (pp. 1–34). University of California-Berkeley, Statistics
cian, 46(3), 175–185. Department.
Amini, H., Gholami, R., Monjezi, M., Torabi, S. R., & Zadhesh, J. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–
(2012). Evaluation of flyrock phenomenon due to blasting 32.
operation by support vector machine. Neural Computing and Carver, R. H., & Nash, J. G. (2011). Doing data analysis with
Applications, 21(8), 2077–2085. SPSS: version 18.0: Cengage Learning.
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Chae, D.-K., Lee, S.-C., Lee, S.-Y., & Kim, S.-W. (2018). On model for prediction of blast-induced flyrock using regression
identifying k-nearest neighbors in neighborhood models for tree technique. Environmental Earth Sciences, 76(1), 27.
efficient and effective collaborative filtering. Neurocomput- Hasanipanah, M., Monjezi, M., Shahnazar, A., Armaghani, D. J.,
ing, 278, 134–143. & Farazmand, A. (2015). Feasibility of indirect determina-
Chafi, M., Karami, G., & Ziejewski, M. (2010). Biomechanical tion of blast induced ground vibration based on support
assessment of brain dynamic responses due to blast pressure vector machine. Measurement, 75, 289–297.
waves. Annals of Biomedical Engineering, 38(2), 490–504. Hasanipanah, M., Naderi, R., Kashir, J., Noorani, S. A., & Qaleh,
Chakraborty, A., & Goswami, D. (2017). Slope stability predic- A. Z. A. (2017c). Prediction of blast-produced ground
tion using artificial neural network (ANN). International vibration using particle swarm optimization. Engineering with
Journal of Engineering and Computer Science, 6(6), 21845– Computers, 33(2), 173–179.
21848. https://doi.org/10.18535/ijecs/v6i6.49. Hasanipanah, M., Shahnazar, A., Amnieh, H. B., & Armaghani,
Chen, Z., & Wang, B. (2018). How priors of initial hyper-pa- D. J. (2017d). Prediction of air-overpressure caused by mine
rameters affect Gaussian process regression models. Neuro- blasting using a new hybrid PSO–SVR model. Engineering
computing, 275, 1702–1710. with Computers, 33(1), 23–31.
Chipman, H. A., George, E. I., & McCulloch, R. E. (2010). Hill, J. L. (2011). Bayesian nonparametric modeling for causal
BART: Bayesian additive regression trees. The Annals of inference. Journal of Computational and Graphical Statistics,
Applied Statistics, 4(1), 266–298. 20(1), 217–240.
Cortes, C., & Vapnik, V. (1995). Support vector machine. Ma- Hresc, J., Riley, E., & Harris, P. (2018). Mining projectÕs economic
chine Learning, 20(3), 273–297. impact on local communities, as a social determinant of
Easley, M., Haney, L., Paul, J., Fowler, K., & Wu, H. (2018). Deep health: A documentary analysis of environmental impact
neural networks for short-term load forecasting in ERCOT statements. Environmental Impact Assessment Review, 72,
system. In Texas Power and Energy Conference (TPEC), 64–70.
2018 IEEE, IEEE (pp. 1–6). Hustrulid, (1999). Blasting principles for open-pit blasting: theo-
Effron, B., & Tibshirani, R. J. (1993). An introduction to the retical foundations. Rotterdam: Balkema.
bootstrap. Monographs on Statistics and Applied Probability, Hustrulid, Kuchta, M., & Martin, R. K. (2013). Open pit mine
57, 436. planning and design, two volume set & CD-ROM pack. Boca
Elith, J., Leathwick, J. R., & Hastie, T. (2008). A working guide to Raton: CRC Press.
boosted regression trees. Journal of Animal Ecology, 77(4), Iphar, M., Yavuz, M., & Ak, H. (2008). Prediction of ground
802–813. vibrations resulting from the blasting operations in an open-
Fan, G., Xie, J., Yoshino, H., Yanagi, U., Hasegawa, K., Kagi, N., pit mine by adaptive neuro-fuzzy inference system. Envi-
et al. (2018). Indoor environmental conditions in urban and ronmental Geology, 56(1), 97–107.
rural homes with older people during heating season: A case Jayalakshmi, T., & Santhakumaran, A. (2011). Statistical nor-
in cold region, China. Energy and Buildings, 167, 334– malization and back propagation for classification. Interna-
346. tional Journal of Computer Theory and Engineering, 3(1),
Faradonbeh, R. S., Hasanipanah, M., Amnieh, H. B., Armaghani, 1793–8201.
D. J., & Monjezi, M. (2018). Development of GP and GEP Khandelwal, M., & Kankar, P. (2011). Prediction of blast-induced
models to estimate an environmental issue induced by air overpressure using support vector machine. Arabian
blasting operation. Environmental Monitoring and Assess- Journal of Geosciences, 4(3–4), 427–433.
ment, 190(6), 351. Khandelwal, M., & Singh, T. (2005). Prediction of blast induced
Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic air overpressure in opencast mine. Noise & Vibration
regression: a statistical view of boosting (with discussion and Worldwide, 36(2), 7–16.
a rejoinder by the authors). The Annals of Statistics, 28(2), Khandelwal, M., & Singh, T. (2013). Application of an expert
337–407. system to predict maximum explosive charge used per delay
Garson, G. D. (1991). Interpreting neural-network connection in surface mining. Rock Mechanics and Rock Engineering,
weights. AI Expert, 6(4), 46–51. 46(6), 1551–1558.
Goh, A. T. (1995). Back-propagation neural networks for mod- Kohavi, R. (1995). A study of cross-validation and bootstrap for
eling complex systems. Artificial Intelligence in Engineering, accuracy estimation and model selection. In Appears in the
9(3), 143–151. international joint conference on artificial intelligence (pp.
Gurney, K. (2014). An introduction to neural networks. Boca 1137–1145). Montreal, Canada.
Raton: CRC Press. Koike, K., & Matsuda, S. (2003). Characterizing content distri-
Hajihassani, M., Armaghani, D. J., Sohaei, H., Mohamad, E. T., & butions of impurities in a limestone mine using a feedforward
Marto, A. (2014). Prediction of airblast-overpressure induced neural network. Natural Resources Research, 12(3), 209–222.
by blasting using a hybrid artificial neural network and par- Kuzu, C., Fisne, A., & Ercelebi, S. (2009). Operational and geo-
ticle swarm optimization. Applied Acoustics, 80, 57–67. logical parameters in the assessing blast induced airblast-
Hasanipanah, M., Amnieh, H. B., Khamesi, H., Armaghani, D. J., overpressure in quarries. Applied Acoustics, 70(3), 404–411.
Golzar, S. B., & Shahnazar, A. (2018). Prediction of an Liaw, A., & Wiener, M. (2002). Classification and regression by
environmental issue of mine blasting: an imperialistic com- randomForest. R News, 2(3), 18–22.
petitive algorithm-based fuzzy system. International Journal Linero, A. R. (2018). Bayesian regression trees for high-dimen-
of Environmental Science and Technology, 15(3), 551–560. sional prediction and variable selection. Journal of the
Hasanipanah, M., Armaghani, D. J., Khamesi, H., Amnieh, H. B., American Statistical Association, 113, 1–11.
& Ghoraba, S. (2016). Several non-linear models in esti- Loder, B. (1987). National Association of Australian State Road
mating air-overpressure resulting from mine blasting. Engi- Authorities. In Australian Workshop for Senior ASEAN
neering with Computers, 32(3), 441–455. Transport Officials, 1985, Canberra..
Hasanipanah, M., Faradonbeh, R. S., Amnieh, H. B., Armaghani, Longjun, D., Xibing, L., Ming, X., & Qiyue, L. (2011). Compar-
D. J., & Monjezi, M. (2017a). Forecasting blast-induced isons of random forest and support vector machine for pre-
ground vibration developing a CART model. Engineering dicting blasting vibration characteristic parameters. Procedia
with Computers, 33(2), 307–316. Engineering, 26, 1772–1781.
Hasanipanah, M., Faradonbeh, R. S., Armaghani, D. J., Amnieh, Lu, S., Qiu, X., Shi, J., Li, N., Lu, Z.-H., Chen, P., et al. (2017). A
H. B., & Khandelwal, M. (2017b). Development of a precise pathological brain detection system based on extreme
X.-N. Bui, Nguyen, Le, H.-A. Bui, and Do

learning machine optimized by bat algorithm. CNS & Neu- in artificial neural networks using simulated data. Ecological
rological Disorders-Drug Targets (Formerly Current Drug Modelling, 178(3–4), 389–397.
Targets-CNS & Neurological Disorders), 16(1), 23–29. Perez, L. G., Flechsig, A. J., Meador, J. L., & Obradovic, Z.
Mahdiyar, A., Marto, A., & Mirhosseinei, S. A. (2018). Proba- (1994). Training an artificial neural network to discriminate
bilistic air-overpressure simulation resulting from blasting between magnetizing inrush and internal faults. IEEE
operations. Environmental Earth Sciences, 77(4), 123. Transactions on Power Delivery, 9(1), 434–441.
Mayor, R., & Flanders, R. (1990). Technical manual simplified Prashanth, R., & Nimaje, D. (2018). Estimation of ambiguous
computer model of air blast effects on building walls. Wash- blast-induced ground vibration using intelligent models: A
ington: US Department of State, Office of Diplomatic case study. Noise & Vibration Worldwide, 49(4), 147–
Security. 157.
Mayorga, M. A. (1997). The pathology of primary blast over- Protodiakonov, M., Koifman, M., Chirkov, S., Kuntish, M., &
pressure injury. Toxicology, 121(1), 17–28. Tedder, R. (1964). Rock strength passports and methods for
McKenzie, C. (1990). Quarry blast monitoring: technical and their determination. Moscow: Nauka.
environmental perspectives. Quarry Management, 17, 23–24. Raina, A., Haldar, A., Chakraborty, A., Choudhury, P., Ramulu,
Michieka, N. M. (2014). Energy and the environment: The rela- M., & Bandyopadhyay, C. (2004). Human response to blast-
tionship between coal production and the environment in induced vibration and air-overpressure: An Indian scenario.
China. Natural Resources Research, 23(2), 285–298. Bulletin of Engineering Geology and the Environment, 63(3),
Mohamad, E. T., Hajihassani, M., Armaghani, D. J., & Marto, A. 209–214.
(2012). Simulation of blasting-induced air overpressure by Rasmussen, C. E. (2004). Gaussian processes in machine learning.
means of artificial neural networks. International Review on Advanced lectures on machine learning: ML summer schools
Modelling and Simulations, 5, 2501–2506. 2003, Canberra, Australia, February 2-14, 2003, Tübingen,
Mohamed, M. T. (2009). Artificial neural network for prediction Germany, August 4-16, 2003, Revised Lectures, 3176, 63.
and control of blasting vibrations in Assiut (Egypt) limestone Remennikov, A. M., & Rose, T. A. (2007). Predicting the effec-
quarry. International Journal of Rock Mechanics and Mining tiveness of blast wall barriers using neural networks. Inter-
Sciences, 46(2), 426–431. national Journal of Impact Engineering, 34(12), 1907–
Mohamed, M. T. (2011). Performance of fuzzy logic and artificial 1923.
neural network in prediction of ground and air vibrations. Rodrı́guez, R., Toraño, J., & Menéndez, M. (2007). Prediction of
International Journal of Rock Mechanics and Mining Sci- the airblast wave effects near a tunnel advanced by drilling
ences, 48(5), 845–851. and blasting. Tunnelling and Underground Space Technology,
Monjezi, M., Hasanipanah, M., & Khandelwal, M. (2013). Eval- 22(3), 241–251.
uation and prediction of blast-induced ground vibration at Sakia, R. (1992). The box-cox transformation technique: A re-
Shur River Dam, Iran, by artificial neural network. Neural view. The Statistician, 41, 169–178.
Computing and Applications, 22(7–8), 1637–1643. Särkkä, S., Álvarez, M. A., & Lawrence, N. D. (2017). Gaussian
Monjezi, M., Khoshalan, H. A., & Varjani, A. Y. (2012). Predic- process latent force models for learning and stochastic con-
tion of flyrock and backbreak in open pit blasting operation: trol of physical systems. arXiv preprint arXiv:1709.05409.
a neuro-genetic approach. Arabian Journal of Geosciences, Sawmliana, C., Roy, P. P., Singh, R., & Singh, T. (2007). Blast
5(3), 441–448. induced air overpressure and its prediction using artificial
Müller, D., Leitão, P. J., & Sikor, T. (2013). Comparing the neural network. Mining Technology, 116(2), 41–48.
determinants of cropland abandonment in Albania and Schalkoff, R. J. (1997). Artificial neural networks (Vol. 1). New
Romania using boosted regression trees. Agricultural Sys- York: McGraw-Hill.
tems, 117, 66–77. Schapire, R. E. (2003). The boosting approach to machine
Nateghi, R., Kiany, M., & Gholipouri, O. (2009). Control negative learning: An overview. In D. D. Denison, M. H. Hansen, C.
effects of blasting waves on concrete of the structures by C. Holmes, B. Mallick, & B. Yu (Eds.), Nonlinear estimation
analyzing of parameters of ground vibration. Tunnelling and and classification (pp. 149–171). New York, NY: Springer.
Underground Space Technology, 24(6), 608–616. Schreiter, J., Nguyen-Tuong, D., & Toussaint, M. (2016). Efficient
Nguyen, H., & Bui, X.-N. (2018). Predicting blast-induced air sparsification for Gaussian process regression. Neurocom-
overpressure: A robust artificial intelligence system based on puting, 192, 29–37.
artificial neural networks and random forest. Natural Re- Seeger, M. (2004). Gaussian processes for machine learning. In-
sources Research, 1–15. https://doi.org/10.1007/s11053-018-94 ternational Journal of Neural Systems, 14(02), 69–106.
24-1. Segarra, P., Domingo, J., López, L., Sanchidrián, J., & Ortega, M.
Nguyen, H., Bui, X.-N., Bui, H.-B., & Mai, N.-L. (2018a). A (2010). Prediction of near field overpressure from quarry
comparative study of artificial neural networks in predicting blasting. Applied Acoustics, 71(12), 1169–1176.
blast-induced air-blast overpressure at Deo Nai open-pit coal Shi, X.-Z., Jian, Z., Wu, B.-B., Huang, D., & Wei, W. (2012).
mine, Vietnam. Neural Computing and Applications, 1–17, h Support vector machines approach to mean particle size of
ttps://doi.org/10.1007/s00521-018-3717-5. rock fragmentation due to bench blasting prediction. Trans-
Nguyen, H., Bui, X.-N., Tran, Q.-H., Le, T.-Q., Do, N.-H., & Hoa, actions of Nonferrous Metals Society of China, 22(2), 432–
L. T. T. (2018b). Evaluating and predicting blast-induced 441.
ground vibration in open-cast mine using ANN: a case study Shields, D. J. (1998). Nonrenewable resources in economic, social,
in Vietnam. [journal article]. SN Applied Sciences, 1(1), 125. and environmental sustainability. Nonrenewable Resources,
https://doi.org/10.1007/s42452-018-0136-2. 7(4), 251–261.
Nick, N. (2008). Joseph Juran, 103, pioneer in quality control, Shokri, B. J., Ardejani, F. D., & Ramazi, H. (2016). Environ-
dies. New York Times, 3, 3. mental geochemistry and acid mine drainage evaluation of an
Ogutu, J. O., Piepho, H.-P., & Schulz-Streeck, T. A. (2011). abandoned coal waste pile at the Alborz-Sharghi coal wash-
Comparison of random forests, boosting and support vector ing plant, NE Iran. Natural Resources Research, 25(3), 347–
machines for genomic selection. In BMC Proceedings, (Vol. 363.
5, Vol. 3 pp. S11). BioMed Central. Singh, T., Dontha, L., & Bhardwaj, V. (2008). Study into blast
Olden, J. D., Joy, M. K., & Death, R. G. (2004). An accurate vibration and frequency using ANFIS and MVRA. Mining
comparison of methods for quantifying variable importance Technology, 117(3), 116–121.
Prediction of Blast-induced Air Over-pressure in Open-Pit Mine

Siskind, D. E., Stachura, V. J., Stagg, M. S., & Kopp, J. W. Tyree, S., Weinberger, K. Q., Agrawal, K., & Paykin, J. (2011).
(1980). Structure response and damage produced Parallel boosted regression trees for web search ranking. In
by airblast from surface mining. Report of investigations Proceedings of the 20th International Conference on World
8485. Washington, DC: United States Bureau of Wide Web, (pp. 387–396). ACM.
Mines. Verma, A., & Singh, T. (2011). Intelligent systems for ground
Smola, A. J., Schölkopf, B., & Müller, K.-R. (1998). The con- vibration measurement: A comparative study. Engineering
nection between regularization operators and support vector with Computers, 27(3), 225–233.
kernels. Neural Networks, 11(4), 637–649. Vinacomin. (2010). Report of coal reserve in Quang Ninh pro-
Song, Y., Liang, J., Lu, J., & Zhao, X. (2017). An efficient instance vince, Vietnam (in Vietnamese-unpublished). Coal Reserve
selection algorithm for k nearest neighbor regression. Neu- in Vietnam. Vietnam: Vinacomin.
rocomputing, 251, 26–34. Vinacomin. (2015). Report on geological exploration of Coc Sau
Swingler, K. (1996). Applying neural networks: A practical guide. open pit coal mine, Quang Ninh, Vietnam (in Vietnamse-
Burlington: Morgan Kaufmann. unpublished). Vietnam: Vinacomin.
Tessema, A. (2017). Mineral systems analysis and artificial neural Willmott, C. J., & Matsuura, K. (2005). Advantages of the mean
network modeling of chromite prospectivity in the Western absolute error (MAE) over the root mean square error
Limb of the Bushveld Complex, South Africa. Natural Re- (RMSE) in assessing average model performance. Climate
sources Research, 26(4), 465–488. Research, 30(1), 79–82.
Trevor, H., Robert, T., & Jh, F. (2009). The elements of statistical Zerguine, A., Shafi, A., & Bettayeb, M. (2001). Multilayer per-
learning: data mining, inference, and prediction. New York: ceptron-based DFE with lattice structure. IEEE Transactions
Springer. on Neural Networks, 12(3), 532–545.

You might also like