You are on page 1of 14

Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

Contents lists available at ScienceDirect

Journal of Rock Mechanics and


Geotechnical Engineering
journal homepage: www.jrmge.cn

Full Length Article

Prediction of flyrock induced by mine blasting using a novel kernel-based


extreme learning machine
Mehdi Jamei a, Mahdi Hasanipanah b, c, *, Masoud Karbasi d, Iman Ahmadianfar e,
Somaye Taherifar f
a
Faculty of Engineering, Shohadaye Hoveizeh Campus of Technology, Shahid Chamran University of Ahvaz, Dashte Azadegan, Iran
b
Department of Mining Engineering, University of Kashan, Kashan, Iran
c
Institute of Research and Development, Duy Tan University, Da Nang, 550000, Vietnam
d
Department of Water Engineering, Faculty of Agriculture, University of Zanjan, Zanjan, Iran
e
Department of Civil Engineering, Behbahan Khatam Alanbia University of Technology, Behbahan, Iran
f
Department of Computer Sciences, Faculty of Mathematics and Computer Sciences, Shahid Chamran University of Ahvaz, Ahvaz, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Blasting is a common method of breaking rock in surface mines. Although the fragmentation with proper
Received 13 March 2021 size is the main purpose, other undesirable effects such as flyrock are inevitable. This study is carried out
Received in revised form to evaluate the capability of a novel kernel-based extreme learning machine algorithm, called kernel
15 July 2021
extreme learning machine (KELM), by which the flyrock distance (FRD) is predicted. Furthermore, the
Accepted 17 July 2021
Available online 20 September 2021
other three data-driven models including local weighted linear regression (LWLR), response surface
methodology (RSM) and boosted regression tree (BRT) are also developed to validate the main model. A
database gathered from three quarry sites in Malaysia is employed to construct the proposed models
Keywords:
Blasting
using 73 sets of spacing, burden, stemming length and powder factor data as inputs and FRD as target.
Flyrock distance Afterwards, the validity of the models is evaluated by comparing the corresponding values of some
Kernel extreme learning machine (KELM) statistical metrics and validation tools. Finally, the results verify that the proposed KELM model on ac-
Local weighted linear regression (LWLR) count of highest correlation coefficient (R) and lowest root mean square error (RMSE) is more compu-
Response surface methodology (RSM) tationally efficient, leading to better predictive capability compared to LWLR, RSM and BRT models for all
data sets.
Ó 2021 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by
Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction damages and hazards reported in the mines (Khandelwal and


Singh, 2005; Khandelwal and Monjezi, 2013; Jahed Armaghani
Drilling and blasting is a common method of extracting minerals et al., 2016; Hasanipanah and Amnieh, 2020; Hasanipanah et al.,
and ores at surface mining operations. Several desirable factors may 2020a). Rock mass resistance, powder factor (PF), and blast en-
be of interest in conducting blasting operations such as improving ergy are key parameters on the flyrock distance (FRD). Based on
the drilling operations efficiency and reducing the total costs. About aforementioned studies, any inconsistency between the above
85% of the energy in blasting operations is dispersed in the earth factors (rock mass resistance, PF and blast energy) may result in the
with adverse effects (Hajihassani et al., 2015; Hasanipanah et al., occurrence of flyrock phenomenon. For example, a proper value of
2015). One of the adverse problems is flyrock which means PF can produce proper blast energy and lead to a good fragmen-
movement or throwing of rock fragments under excessive pressure tation size, while any increase or decrease in PF value can lead to an
due to an unexpected blast of explosives. The previous studies improper blast energy and produce the undesirable effects such as
showed that impermissible throwing of rocks has caused most of flyrock. Note that, the PF is the amount of explosive (kg or g)
needed to fragment 1 m3 or cm3 of rock (Hustrulid, 1999). Ac-
cording to the previous researches, such as Ghasemi et al. (2012),
* Corresponding author. Department of Mining Engineering, University of Hasanipanah and Amnieh (2020) and Jahed Armaghani et al.
Kashan, Kashan, Iran. (2020), there are three types of mechanisms causing flyrock in
E-mail address: Hasanipanah.m@gmail.com (M. Hasanipanah).
surface mines.
Peer review under responsibility of Institute of Rock and Soil Mechanics, Chi-
nese Academy of Sciences.

https://doi.org/10.1016/j.jrmge.2021.07.007
1674-7755 Ó 2021 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-
NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1439

Cratering occurs when the ratio of stemming height (St) to blast- Empirical methods are favorably useful in predicting FRD
hole diameter is very small (Ghasemi et al., 2012; Hasanipanah and dealing with simple methods or similar formulas. The site-
Amnieh, 2020). In the cratering occurrence, the flyrock can be specific nature of empirical techniques is their main drawback
projected in any direction from a crater at the hole collar (Ghasemi due to the applicability of statistical data only for the areas of site
et al., 2012). Using inadequate stemming materials leads to the which are measured. It is worth noting that other models usually
rifling. During rifling, blast gases may create a hole along the low- combine the practical and mechanical components. Several re-
resistance path resulting in the ejection of the collar rock as searchers have recently employed methods on the basis of arti-
harmful flyrock (Ghasemi et al., 2012). The convergence of explo- ficial intelligence (AI) to solve mining and civil engineering
sive charges in major and huge geological structures or their problems (e.g. Alimohammadlou et al., 2014; Sezer et al., 2014;
proximity to brittle plates leads to face bursting. In this situation, Sevgen et al., 2019; Zhou et al., 2020a, b, c, d, 2021a, b; Huang
the burden conditions usually control FRD in front of the face. In et al., 2020, 2021; Bardhan et al., 2021; Can et al., 2021), and
Fig. 1, these three types of mechanisms are shown. We studied the also flyrock which is caused by blasting. Flyrock has inevitable
previous works to categorize flyrock based on two controllable and adverse effects. It is not possible to completely eliminate them
non-controllable factors. Proper design can manage the control- but damage can be somehow prevented by minimizing them to a
lable factors, while natural non-controllable factors cannot be reasonable extent. Flyrock accidents can be controlled and pre-
modified (Bajpayee et al., 2004). The controllable factors include vented by estimating various occurrences of the flyrock phe-
inaccurate drilling, improper burden (B) and spacing (S), and nomenon as one of the most effective ways. Designers and
inadequate St. Note that St is the blast-hole height minus the length technicians intend to minimize flyrock to prepare a safe place for
of the explosive column, B is the distance between the individual equipment and workers in the area surrounded by the mine. For
rows of blast-holes, and S is the distance between blast-holes in any this purpose, different empirical models have been developed for
given row (Hustrulid, 1999). Besides, uncontrollable factors in the the prediction of flyrock considering the effects of certain bur-
flyrock phenomenon include issues such as rock mass properties dens and the diameter of the hole (Roth, 1975). Due to the effect
and unspecified geological conditions (Ghasemi et al., 2012; of other factors on the flyrock, in general, it is not possible to
Faradonbeh et al., 2016). One of the reasons for unspecified strongly predict the performance of empirical models (Ghasemi
geological conditions issue is the uncertainty of geological testing et al., 2012). As mentioned earlier, the AI methods have been
techniques. recently employed in the literature to predict FRD. Nguyen et al.
To predict FRD in surface mines, there are two categories of (2021) offered a hybrid of support vector machine (SVM) and
methods. The first category is the machine method with physical whale optimization algorithm to predict FRD. Then, the perfor-
mechanisms being investigated (Roth, 1975; Little and Blair, 2010). mance of hybrid model was compared with two other data-
On the other words, the machine methods, such as the mechanistic driven models including the artificial neural network (ANN)
Monte Carlo models (Little and Blair, 2010), are based on mecha- and gradient boosting machine. Their results confirmed the
nistic modeling in which the physical mechanisms are clearly effectiveness of the proposed model to accurately predict FRD. In
identified (Ghasemi et al., 2012). The second one is the empirical another study, Ye et al. (2021) investigated the application of
equations not related to the machine methods. In the literature, random forest and genetic programming to predict FRD. Their
some studies, such as McKenzie (2009) and Trivedi et al. (2014), results demonstrated that the performance of genetic program-
have used the empirical equations to predict FRD. The following are ming is better than the random forest, and could be a good tool
the advantages and drawbacks of these methods (Lundborg et al., in this field. Recently, a combination of support vector regression
1975; Monjezi et al., 2012). and gray wolf optimization algorithm was carried out in pre-
The most important advantage of machine methods is their dicting FRD by Jahed Armaghani et al. (2020). They indicated the
entire combination of comprehensive components such as rock fall, integrity and reliability of the combined model to predict FRD.
trajectory and air drag. Besides, the factors mentioned above are The main motivation behind this research is to provide more
not specific to the site. On the other hand, an obvious drawback of accurate models than the previous models in estimating the FRD
this method is its need to several input factors such as partial ve- having appropriate performance. In our analysis, a novel kernel-
locity, mass and throwing angle for measurement of FRD. These based extreme learning machine (ELM) algorithm, called the
factors are mostly specific to the site leading to some undesired kernel ELM (KELM), is proposed to predict FRD based on four fea-
problems (Ghasemi et al., 2012). tures including S, B, St and PF. Moreover, three other data-driven
models including local weighted linear regression (LWLR),
response surface methodology (RSM), and boosted regression tree
(BRT) are also developed for the sake of comparison. The potential
of four models are examined for the first time to predict the FRD
estimation. Furthermore, some efficient validation tools, uncer-
tainty, and sensitivity analysis are conducted to precisely justify the
capability of data-driven models. The results demonstrate that the
KELM, as the best predictive model, has considerable capability to
accurately predict the FRD and regarding to the simplicity of setting
the hyperparameters, it prevents over-fitting.
The rest of this paper is organized as follows: Section 2 devotes
to statistically describing the collected data sets and pre-processing
of the data sets used in simulation. Section 3 explains the meth-
odology of the data-driven models and all validation devices. Sec-
tion 4 includes the application results and discussion, data outlier
detection, uncertainty, and sensitivity analysis. Finally, the con-
Fig. 1. Mechanisms causing flyrock in surface mines (Han et al., 2020). clusions are reflected in Section 5.
1440 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

2. Case studies and statistical analysis In training models, 70% of all data (52 data) were considered and
remaining datasets (21 data) were allocated for the testing stage.
To accomplish the objectives of this study, three granite quarry Data statistics used in the simulation of FRD are tabulated in Table 1.
sites, including the Ulu Tiram, Pengerang and Masai quarry sites, The descriptive statistics specially the skewness and kurtosis
located in Malaysia, are investigated (Fig. 2). The rock quality confirmed Gaussian behavior of the implemented variables except
designation (RQD) and uniaxial compressive strength (UCS) of the St which had kurtosis factor equal to 3.254.
aforementioned sites are varied in range of 45e80 and 30e Fig. 3 displays the dispersion of data points and box plots of
110 MPa, respectively. normalized value of datasets for visually assessment of the dis-
To fragment the rock masses, the drilling and blasting method tribution function. According to this figure, the St on account of
was used in the sites. The drilling process was performed by using maximum skewness (1.682) is far from the normal distribution
the holes with diameters of 75 mm, 115 mm and 150 mm for the whereas rests of variables by less skewness are close to the normal
Ulu Tiram, the Pengerang and the Masai sites, respectively. distribution. In addition, the Anderson-Darling test (Scholz and
Furthermore, in the blasting process, the ANFO was used as the Stephens, 1987) on all datasets ascertains that the FRD sample
main explosive material. More details regarding the studied cases passes the normality test which can be a positive point in
can be found in Hasanipanah et al. (2017). modeling.
The blasting operations produce some undesirable effects such as The degree of linear correlation between the input and target
flyrock, ground vibration and air-overpressure. These unfavorable parameters has a significant effect on determining the type of data
effects should be considered in each blasting site. In the present mining method. The linear dependence of all data sets is illustrated
study, a comprehensive AI based investigation is performed to pre- in a correlation matrix in Fig. 4.
dict FRD in the aforementioned sites. Some blast design parameters It is clear that spacing ðrp ¼ 0:6415Þ on account of the highest
including the S, B, St and PF were measured in the sites, and also the absolute Pearson correlation coefficients is recognized as the most
values of FRD for each blasting round were carefully measured. After (inverse) effective feature on FRD estimation. Regarding the low
each blasting event, the relevant videos were observed to determine values of linear correlation between the rest of input data and FRD,
the locations of the maximum rock projections. Then, the maximum it is not possible to judge definitively on the importance of the
horizontal distances between blasting-point and landed fragments other input variables and it seems reasonable to perform a sensi-
were considered as flyrock value. For this purpose, a hand-held tivity analysis based on the data mining models (Jamei and
global positioning system (GPS) was used. Ahmadianfar, 2020a, b). Eventually, the predictive function of FRD
Totally, 73 sets of data, collected from 73 blasting events, based on four mentioned inputs is defined as
including four input parameters (S, B, St and PF) and one output
parameter (FRD) were prepared and used in the modeling pro-
FRD ¼ f ðS; B; St ; PFÞ (1)
cesses. Note that many researchers have used this range of datasets
to evaluate and predict FRD in their studies. For example, Tonnizam It is noteworthy that all the variables are normalized to restrict
Mohamad et al. (2013), Trivedi et al. (2014), Armaghani et al. (2014), the variables into the range [0, 1] for facilitating the complexity of
Faradonbeh et al. (2016), Zhou et al. (2020d), Fattahi and modeling and computational cost reduction using (Xnor ¼ ðX 
Hasanipanah (2020), and Hasanipanah and Amnieh (2020) have Xmin Þ=ðXmax  Xmin Þ) formulation, where Xnor denotes the
used 39, 95, 44, 76, 65, 80 and 62 datasets in their studies to predict normalized value; X is the original value of variable; and Xmax and
FRD. Therefore, the use of 73 datasets in the present study can be an Xmin are the maximum and minimum values of the variable,
acceptable number of dataset to evaluate FRD. respectively.

Fig. 2. The study area and granite quarry sites in this research.
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1441

Table 1
Statistical indicators associated with the variables used in FRD modeling.

Variable Minimum Maximum Mean Standard Coefficient of Skewness Kurtosis Anderson-Darling test
deviation variation
A2 P-value Passed normality
test (alpha ¼ 0.05)?

Spacing, S 2.65 m 4m 3.303 m 0.4082 m 12.36% 0.102 1.525 2.822 <0.0001 No


Burden, B 1.7 m 3.2 m 2.44 m 0.4591 m 18.82% 0.2497 1.673 5.024 <0.0001 No
Stemming length, St 1.7 m 3.6 m 2.185 m 0.3926 m 17.97% 1.682 3.254 3.482 <0.0001 No
Powder factor, PF 0.67 kg/m3 1.05 kg/m3 0.8852 kg/m3 0.1141 kg/m3 12.89% 0.1342 1.135 1.113 0.0061 No
Flyrock distance, 109 m 334 m 229.4 m 54.54 m 23.78% 0.3556 0.3703 0.4422 0.2807 Yes
FRD

Note: P-value and A2 are the metrics of AndersoneDarling test for identifying normality of data.

Fig. 3. The normalized distribution function of input and output variables.

iteration. The main advantages of this method are high learning


speed and easy implementation. This method is a novel version of
the single layer feed forward network (SLFN) which, unlike the
ANN, is random in nature. The ELM randomly generates input
weights and hidden layer biases, and then maintains them fixed
throughout training process. Recently, a novel version of the ELM,
i.e. KELM, has been introduced by Huang et al. (2011). The KELM
uses the main advantages of the ELM and kernel functions (KFs)
simultaneously. It was demonstrated that it provides a better
prediction efficiency than the ELM with lower computational cost
(Chen et al., 2020a,b). The ELM output with M hidden nodes can be
defined as

X
M   
yk ¼ aj h aj xk þ bj ðk ¼ 1; 2; .; KÞ (2)
j¼1

where xk ¼ ½xk1 ; xk2 ; .; xkn T denotes the input vector with n


nodes, yk ¼ ½yk1 ; yk2 ; .; ykn T represents the output vector with n
nodes, ai stands for the weight vector conecting jth hidden node
Fig. 4. The correlation plot for the input and target variables. and input nodes, aj is the weight vector linking the output nodes
and jth hidden node, bj symbolizes the jth hidden node bias, and h
indicates the nonlinear function for hidden layer.
3. Methodology Eq. (1) can be expressed as

3.1. KELM Z ¼ Ha (3)

ELM was first brought up by Huang et al. (2006). The ELM where Z is the expected output; and H denotes the matrix of hidden
calculates the weights of output in just one step without any layer output, which is defined as
1442 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

2 3 In this research, the radial basis function (RBF) is employed as a KF,


hða1 x1 þ b1 Þ . hðaM x1 þ bM Þ which is described by
H ¼ 4« . « 5 (4) 
 
. 2

hða1 xK þ b1 Þ . hðaM xK þ bM Þ k xi ; xj ¼ exp  xi  xj 2 2r (11)


In order to find the optimal value for the weight vector a, the
ELM needs an objective function (OF) that should be minimized. where r represents the kernel width of RBF.
The OF can be expressed as
8 2 3 92 3.2. LWLR
<XK X
M   =
OF ¼ 4 aj h aj xk þ bj 5  Targetk (5)
: ; In the conventional linear regression methods, all training
k¼1 j¼1
samples are considered by the same weights. This will cause the
problem of under-fitting the data by this type of regression. This
where Targetk represents the target vector.
under-fit negatively impacts the accuracy of the model (Wang et al.,
According to the theory of generalized inverse, the solution of
2016). To resolve this problem, LWLR method, which is an advanced
Eq. (2) is given by
version of the linear regression, has been introduced (Cleveland
and Devlin, 1988). The LWLR method is one of the lazy learning
a ¼ HyZ (6) methods. Assume that the estimated yq ðxÞ output is determined by
regression using the following equation:
where H y denotes the Moore-Penrose inverse matrix (MPIM) of H.
According to the orthogonal projection technique and theory of T
yq ðxÞ ¼ q x (12)
ridge regression (Hoerl and Kennard, 1970), the regularization
factor (F) was considered in the optimization process, thus the where x is the training input data and q stands for the coefficient
solution a can be calculated as factor for the training input. The LWLR method minimizes the
 1 following objective function to achieve the least error:
I
a ¼ HTH þ HTZ (7)
l n o
F X
Minimize wi ½yq ðxi Þ  yi 2 ¼ ðX q  YÞT WðX q  YÞ (13)
where Iis the identity matrix. Therefore, the ELM output function i¼1
can be written as
where yi is the observed data value, W is the diagonal matrix of the
  weights, X denotes the matrix of the input training data, and Y is
I 1 T
gðxÞ ¼ hðxÞa ¼ hðxÞ H T H þ H Z (8) the estimated output vector. In order to minimize the above
F objective function, we must differentiate the above equation with
The ELM achieves high performance, but its random nature is its respect to q and set the result to zero (Jamei and Ahmadianfar,
major drawback. Hence, Huang et al. (2011) introduced the KELM to 2020b):
overcome this shortcoming. Fig. 5 illustrates the structure of the
KELM. The KELM uses the kernel matrix (Kernelðx;xj Þ) instead of the X T WX q  X T WY ¼ 0 (14)
activation function hðxÞ. The kernel matrix based on an orthogonal Therefore, q in the LWLR model can be obtained as follows
projection procedure can be expressed as (Wang et al., 2016; Jamei and Ahmadianfar, 2020b):
    
1
F ¼ H T H : Fj;k ¼ h xj hðxk Þ ¼ Kernel xj ; xk (9)
q ¼ X T WX X T WY (15)
The KELM output function is defined by
The LWLR method uses KFs similar to the SVM model to increase
2 3
Kernelðx; x1 Þ  1 the weight of close points more than the other points. In the pre-
4
gðxÞ ¼ « 5 Fþ I Z (10) sent study, a polynomial kernel with the following equation is
F used:
Kernelðx; xM Þ

m
Therefore, instead of the activation function, KELM needs a KF.
wij ¼ xTi xj þ C (16)

where m is a polynomial degree and C is a constant coefficient. In


this study, LWLR model is developed using Weka 3.8.4 software.

3.3. RSM

In the current study, RSM is used to investigate the effect of


independent variables (S, B, St and PF) on the output (response)
variable (FRD) and also to provide an optimal regression relation-
ship for FRD prediction. The RSM is a statistical tool for modeling
and analyzing the behavior of the process (input) variables on the
response (output) variable (Bucher and Bourgund, 1990). Using
RSM, most information can be obtained with a minimum of
experimental data. The 2nd order RSM model includes linear,
Fig. 5. KELM structure utilized in modeling the FRD values. quadratic and the interaction of input variables terms. The RSM
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1443

Fig. 6. Architecture of BRT approach.

model for the mentioned case can be expressed as follows (Hamid 2010). This process builds better and more accurate models
et al., 2016): compared to a single model. BRT overcomes the largest weakness of
a single decision tree, which is a relatively poor fitting. In BRT, only
X
k k 
X
X k 
k1 X
the first tree of the entire training data is created, subsequent trees
y¼ ðbi Xi Þ þ bii Xi2 þ bij Xi Xj þ ε (17) are grown on the remaining data from the previous tree (Elith et al.,
i¼1 i¼1 i ¼ 1 j ¼ iþ1
2008). Trees are not built on all data and only use some data. At
each stage, each data set is categorized, and this classification is
where ε is a random error vector; and bi ; bii and bij are the
used as a weight to fit the next tree. Boosting operations are per-
regression coefficients which can be calculated by the following
formed to improve the predictive power of the regression tree. This
equation:
operation is similar to the model averaging process, in which the

1 common results of several models are used, except that the
b ¼ XTX XTY (18) boosting operation is a step-by-step process, meaning that in each
iteration step, the models are fitted to the part of a training dataset
In this method, analysis of variance (ANOVA) is used to evaluate (Elith et al., 2008; Naghibi and Dashtpagerdi, 2017). Therefore, two
the significance of regression coefficients (P-value < 0.05). The important parameters, i.e. the shrinkage parameter and the
quality of the obtained regression model is evaluated using the R2adj learning rate parameter, are proposed in the model. The shrinkage
index (Hosseinzadeh et al., 2018). parameter specifies the percentage of training data in each iteration
and is determined by the user. The learning rate parameter in-
3.4. BRT dicates the contribution of each tree in the modeling process (Elith
et al., 2008). This method has several advantages, including the fact
BRT model is a combination of regression trees and boosting that it can analyze large volumes of data at high speeds, is less
technique (Friedman, 2001; Rätsch et al., 2001). This model is one sensitive to over-fitting of models, does not require data distribu-
of the several techniques that help to improve the performance of a tion assumptions, and is also able to determine the most important
single model by using a combination of multiple single models factors in the modeling process (Westreich et al., 2010). The ar-
(Aertsen et al., 2010). BRT uses a combination of two algorithms: (i) chitecture of BRT approach is depicted in Fig. 6.
the classification and regression tree (CART) model and (ii) the
construction and combination of a set of models by boosting 3.5. Theory of leverage approach
technique (Naghibi and Dashtpagerdi, 2017). Boosting is a way to
increase the model’s accuracy, and it works on the basis of building, Checking the validity of the model and detecting outlier data are
combining and averaging a large number of models (Aertsen et al., the most essential parts of the model developing process
1444 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

(Rousseeuw and Leroy, 2005). Among the plenty of numerical and Kling-Gupta (KGE) multi-objective index (Gupta et al., 2009)
graphical methods of outlier data detection, leverage approach is a was examined for reliable selecting the optimal predictive model.
well-known method in the field of model application range and This decision criterion depends on the correlation coefficient,
outlier data detection (Shateri et al., 2015). In this method, the re- variability error and bias of the models as follows:
sidual is defined as the difference between the observed and pre- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dicted data and the following equation is used to construct the H  2  2
StDp FRDp
matrix (Rousseeuw and Leroy, 2005): KGE ¼ 1  ðR  1Þ2 þ 1 þ 1 (25)
StDo FRDo


H ¼ X X T X X 1 (19) where StDp and StDo are the standard deviations in predicted and
observed values of FRD. It should be mentioned that the unit value
where the input matrix X has n rows (number of samples) and m of KGE indicates the perfect consistency between measured and
columns (number of variables used in modeling). The diagonal el- predicted values.
ements of the H matrix are known as the hat index. In the process,
the hat index is plotted against the standardized residual in a dia- 3.7. Model development
gram called Williams. The warning hat index (H* ) value is calcu-
lated from the relation H* ¼ 3ðm þ 1Þ=n. In the Williams diagram, In this research, a comprehensive intelligent data analysis is
the placement of the majority of data in the range of  3 < R < 3 (R conducted for precise prediction of the FRD measurement using a
is the correlation coefficient) and 0 < H < H* indicates that the data novel KELM model. For this purpose, four input variable were
used to develop the model and the values predicted by the model adopted comprised of S, B, St and PF. Three robust data-driven
are in the application domain, and therefore the model is statisti- models including LWLR, RSM and BRT were employed to validate
cally correct and valid (Esfahani et al., 2015). the predictive execution of KELM. The procedure of predicting FRD
using the provided predictive models is presented in Fig. 7. The
KELM model was provided in environment of Matlab Software
3.6. Statistical criteria for evaluation of models
based on a RBF KF, three layers, four input neurons, and two crucial
settings, i.e. the regularization coefficient (F) and width of KF (r)
For the goodness of fit assessment of the provided AI models,
were specified by a trial and error process. The best values of F and r
five efficient performance metrics were taken into account, i.e.
were obtained equal to 5000 and 30, respectively. Besides, in order
correlation coefficient (R), root mean square error (RMSE), mean
to develop the LWLR model, a polynomial kernel was employed to
absolute percentage error (MAPE), and Theil’s inequality co-
provided model. The optimum setting parameters include C ¼ 1.275
efficients comprised of the prediction accuracy or Theil 1 (U1), and
and exponent value of m ¼ 1.07 which were acquired in a trial-
prediction quality or Theil 2 (U2) (Zhang et al., 2019, 2020a, b, c, d,
and-error process and several model executions. The summary of
2021; Wang et al., 2020; Hasanipanah et al., 2020b; Shahrour and
setting parameters for all provided model is listed in Table 2.
Zhang, 2021; Zheng et al., 2021; Zhou et al., 2021c). According to
Moreover, in this research, an ensemble of BRT based on the
Lewis categorization for MAPE indicator, the predictive models for
“fitrensemble” function (Matlab Software) was accomplished using
MAPE less than 10% yields “excellent” results, those for 10% 
a least-squares boosting (LSboost) aggregation algorithm and a 10-
MAPE  20% can be evaluated as the “good” performance, those for
fold cross-validation method. The optimum regularization to gain
20%  MAPE  50% are considered as “acceptable”, and those for
the best results was reported in Table 2. Fig. 8 shows the learn rate
MAPE > 50% lead to “inaccurate” outcomes.
variation in BRT modeling versus R and RMSE in range of 0.01e1.
Correlation coefficient (R), RMSE, MAPE, prediction accuracy
The optimal learn rate equal to 0.66 resulted in the best perfor-
(U1) and prediction quality (U2) are respectively expressed as
mance in the BRT model.
PN   
i¼1 FRDp;i  FRDp FRDo;i  FRDo
R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN   PN   ð0 < R < 1Þ 4. Results and discussion
i¼1 FRDp;i  FRDp i¼1 FRDo;i  FRDo
2 2

(20) 4.1. Application results and analysis

X As mentioned earlier, four data-driven models (i.e. RSM, KELM,


1 N  2 0:5
RMSE ¼ i¼1
FRDo;i  FRDp;i (21) LWLR and BRT) are used to estimate the FRD. According to Section
N
2, a total number of 73 data points have been collected for
  modeling. Fig. 9 illustrates the decision trees obtained by the BRT
100XN FRDo;i  FRDp;i  model. Also, the optimal input variables achieved by the RSM
MAPE ¼ i¼1
(22)
N FRDo;i model using a quadratic approximation are listed in Table 3.
Table 4 presents the values of six performance metrics obtained
RMSE by four models for training, testing and all datasets. As reported in
U1 ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN ffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN (23) Table 4, the BRT model has the highest R (0.983) and lower MAPE
1 2 1 2
N i¼1 FRD o;i þ N i¼1 FRDp;i (4.5102) when compared with the KELM, LWLR and RSM models in
the training phase. In terms of RMSE (12.3984), KGE (0.9514), U1
PN  2 (0.0265) and U2 (0.0565), the KELM models can outperform the
i¼1 FRDo;i  FRDp;i
U2 ¼ PN (24) other models. In the testing phase, according to Table 4, the LWLR
2
i¼1 FRDo;i had the maximum value of R (0.976) and minimum value of RMSE
(13.325), U1 (0.0282) and U2 (0.0561), outperforming the KELM,
where FRDo;i and FRDo are the measured and mean values of RSM and BRT models, while the KELM model had the maximum
measured FRD samples, respectively; FRDp;i and FRDp are the pre- value of KGE (0.9679) and minimum value of MAPE (5.4947)
dicted and mean values of predicted FRD, respectively; and N is the compared to other models. Tables 5 and 6 report the rank of all
number of data. models in the training and testing phases, respectively. Based on
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1445

Fig. 7. Flowchart associated with the process of FRD estimation using four data-driven approaches.

Table 2
could successfully capture the non-linearity of the FRD values. The
The characteristics and setting parameter of proposed AI based approaches. distribution of FRD values achieved by the KELM, LWLR, RSM and
BRT models for all datasets is displayed in the form of box plots in
Model Adjustment parameter
Fig. 12. As shown in this figure, the KELM has the best distribution
KELM F ¼ 5000, r ¼ 30 of FRD values compared with the LWLR, RSM and BRT models. This
LWLR Poly-linear kernel, C ¼ 1.025, exponent value m ¼ 1.07
means that the proposed KELM can present the superior efficiency
BRT “Lsboost” function, optimum learn rate ¼ 0.66, trees number ¼ 500
among all data-driven models.
Fig. 13 exhibits the relative deviation (RD) distribution in the
scatter form for the KELM, RSM, LWLR and BRT models in the
training and testing phases. From this figure, the KELM by RD range
of 15%  RD  13% provides the least under-/over-estimation
compared with LWLR, RSM and BRT models by the RD ranges of
 20:5%  RD  17%,  15:7%  RD  13:6%, and  70%  RD 
15:3%, respectively. According to the obtained results, the KELM
model yields more compression of error distribution as compared
to the LWLR, RSM and BRT models.
Fig. 14 illustrates the pie plots of the RD values for all provided
models. The pie plots indicate that about 62% of the estimated FRD
by the KELM have RD less than 5% while only 56.76%, 49.32% and
54.79% of the estimated FRD by LWLR, RSM and BRT have RD less
Fig. 8. The learn rate variation versus correlation coefficient in testing phase. than 5%, respectively. Also, 5.54% of all datasets have RD more than
10% for the KELM model, while 14.2%, 16% and 17.5% of the pre-
dicted FRD by the LWLR, RSM and BRT models, respectively, provide
the RD more than 10%. Consequently, the results reveal that the
these tables, the KELM can obtain the best rank (rank ¼ 1.33), fol-
KELM model is recognized as a more reliable and robust predictive
lowed by the BRT (rank ¼ 2), RSM (rank ¼ 3) and LWLR
model in comparison with the LWLR, RSM and BRT models for
(rank ¼ 3.67), respectively, at the training stage. In addition, the
estimating the FRD.
best rank belongs to the LWLR model (rank ¼ 1.33), followed by the
It is noteworthy that the extracted relationship obtained from
KELM (rank ¼ 1.67), RSM (rank ¼ 3) and BRT (rank ¼ 4), respec-
the RSM model is written as
tively, in the testing phase. Overall, by averaging the mean ranks
obtained for training and testing phases, it was specified that the
FRD ¼ 1003:2  169:89S  14:374B  242:41St þ 77:832PF
best rank belongs to the KELM (rank ¼ 1.5), followed by the LWLR
(rank ¼ 2.5), RSM and BRT (rank ¼ 3), respectively. þ 37:201SB  101:49SPF þ 41:739St 2
The measured versus predicted FRD values at the training and (26)
testing stages are displayed in Fig. 10. According to this figure, the
error of estimated values obtained by all models, when compared 4.2. Sensitivity analysis
with the measured values, is within the range of 10%. This figure
shows that the RSM, LWLR and KELM have a reasonable precision Sensitivity analysis is one of the most crucial parts of data-
and the estimated FRD values were closer to the 45 line. driven model which gives the reliable and useful information on
Fig. 11 demonstrates the physical behaviors of all models to feature influence on simulation of phenomena. There are different
assess their capability in tackling the nonlinear trend of the FRD sensitivity analysis methods including cosine amplitude method
values for all datasets, where picked intervals for accurate evalua- (CAM) (Khandelwal et al., 2016), partial derivative sensitivity
tion are zoomed beside each plot. Taking a deeper look into the analysis (PDSA), relevancy factor (Chen et al., 2014), and consecu-
calculated results indicates that the KELM, RSM and LWLR models tive excluding input variable from the predictive model
1446 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

Fig. 9. Achieved decision trees from the BRT approach to estimate the FRD.

Table 3
Table 5
The optimum outcome of RSM approach using a quadratic approximation.
Rank of all models for training stage.
Criterion Estimate SE tstat P-value
Metric Rank
(Intercept) 1003.2 186.54 5.378 0
KELM RSM LWLR BRT
S 169.89 55.406 3.066 0.004
B 14.374 66.312 0.217 0.829 R 2 3 4 1
St 242.41 58.944 4.113 0 RMSE 1 3 4 2
PF 77.832 177.86 0.438 0.664 MAPE 2 4 3 1
SB 37.201 18.864 1.972 0.055 KGE 1 2 3 4
SPF 101.49 55.18 1.839 0.073 U1 1 3 4 2
S2t 41.739 11.532 3.62 0.001 U2 1 3 4 2
Average 1.33 3 3.67 2
Note: SE is the standard error, and tstat is the t-statistic.

Table 4
Quantitative evaluation of AI-based approaches for predicting FRD.
Table 6
Model Dataset type R RMSE (m) MAPE KGE U1 U2 Rank of all models for testing stage.

RSM Train 0.9579 14.5284 5.7021 0.9405 0.031 0.0619 Metric Rank
Test 0.9693 15.3446 6.0781 0.9629 0.0323 0.0643
KELM RSM LWLR BRT
All 0.9622 14.7678 5.8103 0.9493 0.0314 0.0627
LWLR Train 0.9567 14.8761 5.3987 0.9137 0.0317 0.0634 R 2 3 1 4
Test 0.976 13.325 5.5344 0.9566 0.0282 0.0561 RMSE 2 3 1 4
All 0.9637 14.4324 5.439 0.9303 0.0307 0.0613 MAPE 1 3 2 4
KELM Train 0.9696 12.3984 4.9367 0.9514 0.0265 0.0529 KGE 1 3 2 4
Test 0.9723 14.5261 5.4947 0.9679 0.0305 0.0609 U1 2 3 1 4
All 0.9706 13.0461 5.0972 0.9582 0.0277 0.0553 U2 2 3 1 4
BRT Train 0.983 13.2532 4.5102 0.8771 0.0288 0.0565 Average 1.67 3 1.33 4
Test 0.9159 29.0483 10.8436 0.6784 0.0621 0.1218
All 0.9547 19.1796 6.3322 0.8084 0.0415 0.0814

metrics of R ¼ 0.8127, RMSE ¼ 31.8143, U1 ¼ 0.0678 and U2 ¼ 0.135


was identified as the second effective parameter followed by the PF
(Ahmadianfar et al., 2019). Sequentially excluding input variables (R ¼ 0.8505, RMSE ¼ 28.5504, U1 ¼ 0.0609 and U2 ¼ 0.1211). Fig. 15
(feature) is one of the most reliable sensitivity analysis approaches depicts the selective metrics (R, KGE, U1 and U2) for sensitivity
which employed in this study on account of more compatibility analysis execution for four combinations and benchmark (original)
with AI models. Therefore, KELM as the main contribution of the model. The parameters S and St by lowest and highest conformity
current paper was employed to handling sensitivity analysis. For metrics (R and KGE) were diagnosed as the significant and insig-
this purpose, four combinations were compared to benchmark nificant features in modeling, respectively.
KELM model (with all features) and given results were reported in
Table 7. One can infer that the parameter S by the most impact on 4.3. Uncertainty analysis
accuracy reduction of original model, in terms of R ¼ 0.3859,
RMSE ¼ 50.1327, U1 ¼ 0.1079 and U2 ¼ 0.2127, was recognized as the To quantitatively evaluate the uncertainty of the models in the
most influential feature. Moreover, the parameter B based on FRD estimation, the uncertainty analysis was performed on the
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1447

Fig. 10. Comparison between the measured and estimated FRD values for all predictive models in training and testing phases: (a) KELM, (b) LWLR, (c) RSM and (d) BRT.

Fig. 11. Capability assessment of the provided data-driven models to capture the non-linearity relationship between the measured and predicted FRD values: (a) KELM, (b) LWLR, (c)
RSM and (d) BRT.
1448 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

Fig. 12. The distribution of FRD values for all datasets in RSM, LWLR, KELM and BRT.

Fig. 13. Variation of RD versus measured FRD for evaluating the accuracy of all predictive models in the training and testing phases: (a) KELM, (b) LWLR, (c) RSM and (d) BRT.
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1449

RD<5% RD<5%
5% RD 10% 5% RD 10%
RD>10% RD>10% 29.04%
32.2%

14.2%
62.26% 5.54%
56.76%

(a) (b)
RD<5% RD<5%
5% RD 10% 5% RD 10%
34.68% 27.71%
RD>10% RD>10%

16%

17.5%

54.79%
49.32%
(c) (d)

Fig. 14. Cumulative absolute RD for (a) KELM, (b) LWLR, (c) RSM and (d) BRT models.

Table 7
The sensitivity analysis results using KELM model.

Metric All-S All-B All-St All-PF All

R 0.3859 0.8127 0.9491 0.8505 0.9711


RMSE 50.1327 31.8143 17.0755 28.5504 12.9468
MAPE 20.0887 12.8613 6.4378 10.4097 5.1115
KGE 0.0406 0.4338 0.7433 0.5129 0.7808
U1 0.1079 0.0678 0.0363 0.0609 0.0275
U2 0.2127 0.135 0.0724 0.1211 0.0549

models. To calculate the uncertainty, first, the error value of each


sample was calculated as ei ¼ FRDo;i  FRDp;i , and then the mean
values of the error e and the deviation of the estimation error Se
were computed using the following equations:
Fig. 15. Qualitative conformity metrics for sensitivity analysis execution.
X
n
e¼ ei (27)
i¼1
Table 8
Uncertainty estimates of FRD for different models.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP
u n Model e Se Width of uncertainty band
u ðe  eÞ2
ti ¼ 1 i KELM 0.551 14.873 29:143
Se ¼ (28) LWLR 0.968 13.761 26:972
n1
RSM 1.552 15.643 30:648
When e is positive, it indicates an overestimation of the model, BRT 5.272 29.269 57:374

and as e is negative, it demonstrates an underestimation of the


model. High values of Se show a high scatter of error values and
thus increase the uncertainty of the model prediction. Using  the KELM model ( 29:143Þ. The highest width of uncertainty band
1:96Se offers a 95% confidence interval (Newcombe, 1998; Sattar, is related to the BRT model (57:374), which indicates more un-
2014). Mean values of error e, the standard deviation of error Se, certainty in the results of this model.
and 95% error confidence band are presented in Table 8 for all
models. 4.4. Applicability domain assessment
According to Table 8, the mean error values for all models used
in this research are positive, indicating the overestimation. The In the present study, leverage approach was used to determine
KELM model has the lowest mean error value (e ¼ 0:551), and the the outlier data and statistical validity of the model. Fig. 16 shows
BRT model has the highest one (e ¼ 5:272). The lowest width of the Williams diagram of the KELM, RSM, LWLR and BRT models.
uncertainty band is related to the LWLR model ( 26:972) and then According to the shape, all data points are in the ranges of
1450 M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451

References

Aertsen, W., Kint, V., Van Orshoven, J., Özkan, K., Muys, B., 2010. Comparison and
ranking of different modelling techniques for prediction of site index in Med-
iterranean mountain forests. Ecol. Model. 221, 1119e1130.
Ahmadianfar, I., Jamei, M., Chu, X., 2019. Prediction of local scour around circular
piles under waves using a novel artificial intelligence approach. Mar. Georesour.
Geotechnol. 39 (1). https://doi.org/10.1080/1064119X.2019.1676335.
Alimohammadlou, Y., Najafi, A., Gokceoglu, C., 2014. Estimation of rainfall-induced
landslides using ANN and fuzzy clustering methods: a case study in Saeen
Slope, Azerbaijan province, Iran. Catena 120, 149e162.
Armaghani, D.J., Hajihassani, M., Mohamad, E.T., Marto, A., Noorani, S.A., 2014.
Blasting-induced flyrock and ground vibration prediction through an expert
artificial neural network based on particle swarm optimization. Arab. J. Geosci.
7, 5383e5396.
Bajpayee, T.S., Rehak, T.R., Mowrey, G.L., Ingram, D.K., 2004. Blasting injuries in
Fig. 16. Williams diagram of applicability domain specifying for all data-driven surface mining with emphasis on flyrock and blast area security. J. Saf. Res. 35,
47e57.
paradigms.
Bardhan, A., Gokceoglu, C., Burman, A., Samui, P., Asteris, P.G., 2021. Efficient
computational techniques for predicting the California bearing ratio of soil in
soaked conditions. Eng. Geol. 291, 106239.
Bucher, C.G., Bourgund, U., 1990. A fast and efficient response surface approach for
3 < R < 3 and 0 < H < H* , and as a result, all developed models structural reliability problems. Struct. Saf. 7, 57e66.
are statistically acceptable and valid. Can, R., Kocaman, S., Gokceoglu, C., 2021. A comprehensive assessment of XGBoost
algorithm for landslide susceptibility mapping in the upper basin of Ataturk
dam, Turkey. Appl. Sci. 11 (11), 4993.
5. Conclusions Chen, G., Fu, K., Liang, Z., Sema, T., Li, C., Tontiwachwuthikul, P., Idem, R., 2014. The
genetic algorithm based back propagation neural network for MMP prediction
in CO2-EOR process. Fuel 126, 202e212.
Any blasting event in surface mines produces a sudden ejection Chen, S., Gu, C., Lin, C., Wang, Y., Amin, M., 2020a. Prediction, monitoring, and
of rock pieces, called flyrock, which may result in human injuries, interpretation of dam leakage flow via adaptative kernel extreme learning
machine. Measurement 166, 108161.
fatalities and property damage. Therefore, a precise prediction of
Chen, Z., Li, H., Goh, A.T.C., Wu, C., Zhang, W., 2020b. Soil liquefaction assessment
FRD is important especially for safety issues. To this view, the using soft computing approaches based on capacity energy concept. Geo-
present research aimed at introducing a novel KELM model to sciences 10 (9), 330.
predict FRD. To make a fair comparison, three other data-driven Cleveland, W.S., Devlin, S.J., 1988. Locally weighted regression: an approach to
regression analysis by local fitting. J. Am. Stat. Assoc. 83, 596e610.
models, i.e. LWLR, RSM and BRT, were also employed. Then, Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression
several statistical and graphical tools were considered to evaluate trees. J. Anim. Ecol. 77, 802e813.
the predictive performance of the provided models. Based on the Esfahani, S., Baselizadeh, S., Hemmati-Sarapardeh, A., 2015. On determination of
natural gas density: least square support vector machine modeling approach.
obtained results, the following conclusions can be summarized: J. Nat. Gas Sci. Eng. 22, 348e358.
Faradonbeh, R.S., Armaghani, D.J., Amnieh, H.B., Mohamad, E.T., 2016. Prediction
(1) Among all considered models, the KELM in terms of R ¼ 0.9706 and minimization of blast-induced flyrock using gene expression programming
and firefly algorithm. Neural Comput. Appl. 29 (6), 269e281.
and RMSE ¼ 13.0461 m and rank scores of 1.33 and 1.67 for Fattahi, H., Hasanipanah, M., 2020. An integrated approach of ANFIS-grasshopper
training and testing, respectively exhibited a higher perfor- optimization algorithm to approximate flyrock distance in mine blasting. Eng.
mance capacity to predict FRD in all datasets. However, the Comput. https://doi.org/10.1007/s00366-020-01231-4.
Friedman, J.H., 2001. Greedy function approximation: a gradient boosting machine.
LWLR method on account of the best rank score (1.33) showed
Ann. Stat. 29 (5), 1189e1232.
the best performance (R ¼ 0.976 and RMSE ¼ 13.325 m) in the Ghasemi, E., Sari, M., Ataei, M., 2012. Development of an empirical model for pre-
testing phase and the KELM method was in the second place dicting the effects of controllable blasting parameters on flyrock distance in
surface mines. Int. J. Rock Mech. Min. Sci. 52, 163e170.
with a slight difference. Accordingly, the KELM and LWLR
Gupta, H.V., Kling, H., Yilmaz, K.K., Martinez, G.F., 2009. Decomposition of the mean
models could be appropriate tools for accurate prediction of squared error and NSE performance criteria: implications for improving hy-
FRD and have the capacity to generalize in other fields, and RSM drological modelling. J. Hydrol. 377, 80e91.
and BRT stood in the next ranks, respectively. Hajihassani, M., Jahed Armaghani, D., Marto, A., Tonnizam Mohamad, E., 2015.
Ground vibration prediction in quarry blasting through an artificial neural
(2) The error analysis demonstrated that the KELM model due to network optimized by imperialist competitive algorithm. Bull. Eng. Geol. En-
lowest RD band (  15%  RD  13%) has the best performance viron. 74, 873e886.
in predicting FRD in comparison with the other approaches. Hamid, H.A., Jenidi, Y., Thielemans, W., Somerfield, C., Gomes, R.L., 2016. Predicting
the capability of carboxylated cellulose nanowhiskers for the remediation of
(3) Based on the sensitivity analysis, it was found that the pa- copper from water using response surface methodology (RSM) and artificial
rameters S and St by lowest and highest conformity metrics neural network (ANN) models. Ind. Crop. Prod. 93, 108e120.
(R and KGE) were diagnosed as the significant and insignif- Han, H., Jahed Armaghani, D., Tarinejad, R., Zhou, J., Tahir, M.M., 2020. Random
forest and Bayesian network techniques for probabilistic prediction of flyrock
icant parameters in modeling, respectively. induced by blasting in quarry sites. Nat. Resour. Res. 29, 655e667.
(4) Based on uncertainty analysis, LWLR and KELM models have Hasanipanah, M., Monjezi, M., Shahnazar, A., Armaghani, D.J., Farazmand, A., 2015.
the least uncertainty while the BRT model has the highest Feasibility of indirect determination of blast induced ground vibration based on
support vector machine. Measurement 75, 289e297.
uncertainty in FRD prediction.
Hasanipanah, M., Armaghani, D.J., Amnieh, H.B., Majid, M.Z.A., Tahir, M.M.D., 2017.
(5) Finally, to detect outlier data and applicability domain of the Application of PSO to develop a powerful equation for prediction of flyrock due
models, the leverage approach was used. The Williams plot to blasting. Neural Comput. Appl. 28 (1), 1043e1050.
showed no outlier data in the models, and all four models Hasanipanah, M., Amnieh, H.B., 2020. A fuzzy rule-based approach to address un-
certainty in risk assessment and prediction of blast-induced flyrock in a quarry.
were statistically valid and correct. Nat. Resour. Res. 29 (2), 669e689.
Hasanipanah, M., Keshtegar, B., Thai, D.K., Troung, N.T., 2020a. An ANN-adaptive
dynamical harmony search algorithm to approximate the flyrock resulting
Declaration of competing interest from blasting. Eng. Comput. https://doi.org/10.1007/s00366-020-01105-9.
Hasanipanah, M., Zhang, W., Armaghani, D.J., Rad, H.N., 2020b. The potential
application of a new intelligent based approach in predicting the tensile
The authors declare that they have no known competing
strength of rock. IEEE Access 8, 57148e57157.
financial interests or personal relationships that could have Hoerl, A.E., Kennard, R.W., 1970. Ridge regression: biased estimation for non-
appeared to influence the work reported in this paper. orthogonal problems. Technometrics 12, 55e67.
M. Jamei et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1438e1451 1451

Hosseinzadeh, A., Najafpoor, A.A., Jafari, A.J., Jazani, R.K., Baziar, M., Bargozin, H., Shateri, M., Ghorbani, S., Hemmati-Sarapardeh, A., Mohammadi, A.H., 2015. Appli-
Piranloo, F.G., 2018. Application of response surface methodology and artificial cation of Wilcoxon generalized radial basis function network for prediction of
neural network modeling to assess non-thermal plasma efficiency in simulta- natural gas compressibility factor. J. Taiwan Inst. Chem. Eng. 50, 131e141.
neous removal of BTEX from waste gases: effect of operating parameters and Tonnizam Mohamad, E., Jahed Armaghani, D., Hajihassani, M., Faizi, K., Marto, A.,
prediction performance. Process Saf. Environ. Prot. 119, 261e270. 2013. A simulation approach to predict blasting-induced flyrock and size of
Huang, G.B., Zhou, H., Ding, X., Zhang, R., 2011. Extreme learning machine for thrown rocks. Electron. J. Geotech. Eng. 18, 365e374.
regression and multiclass classification. IEEE Trans. Syst. Man, Cybern. Part B 42, Trivedi, R., Singh, T.N., Raina, A.K., 2014. Prediction of blast-induced flyrock in Indian
513e529. limestone mines using neural networks. J. Rock Mech. Geotech. Eng. 6 (5), 447e454.
Huang, G.B., Zhu, Q.Y., Siew, C.K., 2006. Extreme learning machine: theory and Wang, J., Yu, L.-C., Lai, K.R., Zhang, X., 2016. Locally weighted linear regression for
applications. Neurocomputing 70, 489e501. cross-lingual valence-arousal prediction of affective words. Neurocomputing
Huang, J., Sun, Y., Zhang, J., 2021. Reduction of computational error by optimizing 194, 271e278.
SVR kernel coefficients to simulate concrete compressive strength through the Wang, L., Wu, C., Gu, X., Liu, H., Mei, G., Zhang, W., 2020. Probabilistic stability
use of a human learning optimization algorithm. Eng. Comput. https://doi.org/ analysis of earth dam slope under transient seepage using multivariate adaptive
10.1007/s00366-021-01305-x. regression splines. Bull. Eng. Geol. Environ. 79, 2763e2775.
Huang, J., Duan, T., Zhang, Y., Liu, J., Zhang, J., Lei, Y., 2020. Predicting the perme- Westreich, D., Lessler, J., Funk, M.J., 2010. Propensity score estimation: neural net-
ability of pervious concrete based on the beetle antennae search algorithm and works, support vector machines, decision trees (CART), and meta-classifiers as
random forest model. Adv. Civ. Eng. 2, 8863181. alternatives to logistic regression. J. Clin. Epidemiol. 63, 826e833.
Hustrulid, W.A., 1999. Blasting Principles for Open Pit Mining: General Design Ye, J., Koopialipoor, M., Zhou, J., Jahed Armaghani, D., He, X., 2021. A novel combi-
Concepts. A.A. Balkema, Rotterdam, the Netherlands. nation of tree-based modeling and Monte Carlo simulation for assessing risk
Jahed Armaghani, D., Koopialipoor, M., Bahri, M., Hasanipanah, M., Tahir, M.M., levels of flyrock induced by mine blasting. Nat. Resour. Res. 30, 225e243.
2020. A SVR-GWO technique to minimize flyrock distance resulting from Zhang, W., Ching, J., Goh, A.T.C., Leung, A.Y.F., 2020a. Big data and machine learning
blasting. Bull. Eng. Geol. Environ. 79, 4369e4385. in geoscience and geoengineering: Introduction. Geosci. Front. 12 (1), 327e329.
Jahed Armaghani, D., Tonnizam Mohamad, E., Hajihassani, M., Alavi Nezhad Khalil Zhang, W., Li, H., Li, Y., Liu, H., Chen, Y., Ding, X., 2021. Application of deep learning
Abad, S.V., Marto, A., Moghaddam, M.R., 2016. Evaluation and prediction of algorithms in geotechnical engineering: a short critical review. Artif. Intell. Rev.
flyrock resulting from blasting operations using empirical and computational https://doi.org/10.1007/s10462-021-09967-1.
methods. Eng. Comput. 32, 109e121. Zhang, W., Li, Y., Wu, C., Li, H., Goh, A.T.C., Lin, H., 2020b. Prediction of lining
Jamei, M., Ahmadianfar, I., 2020a. A rigorous model for prediction of viscosity of oil- response for twin tunnels constructed in anisotropic clay using machine
based hybrid nanofluids. Physica A 556, 124827. learning techniques. Undergr. Space. https://doi.org/10.1016/
Jamei, M., Ahmadianfar, I., 2020b. Prediction of scour depth at piers with debris j.undsp.2020.02.007.
accumulation effects using linear genetic programming. Mar. Georesour. Geo- Zhang, W., Wu, C., Zhong, H., Li, Y., Wang, L., 2020c. Prediction of undrained shear
technol. 38, 468e479. strength using extreme gradient boosting and random forest based on Bayesian
Khandelwal, M., Armaghani, D.J., Faradonbeh, R.S., Ranjith, P.G., Ghoraba, S., 2016. optimization. Geosci. Front. 12 (1), 469e477.
A new model based on gene expression programming to estimate air flow in a Zhang, W., Zhang, R., Wang, W., Zhang, F., Goh, A.T.C., 2019. A multivariate adaptive
single rock joint. Environ. Earth Sci. 75, 739. regression splines model for determining horizontal wall deflection envelope
Khandelwal, M., Monjezi, M., 2013. Prediction of backbreak in open-pit blasting for braced excavations in clays. Tunn. Undergr. Space Technol. 84, 461e471.
operations using the machine learning method. Rock Mech. Rock Eng. 46, 389e Zhang, W., Zhang, R., Wu, C., Goh, A.T.C., Lacasse, S., Liu, Z., Liu, H., 2020d. State-of-
396. the-art review of soft computing applications in underground excavations.
Khandelwal, M., Singh, T.N., 2005. Prediction of blast induced air overpressure in Geosci. Front. 11 (4), 1095e1106.
opencast mine. Noise Vib. Worldw. 36, 7e16. Zheng, G., Zhang, W., Zhang, W., Zhou, H., Yang, P., 2021. Neural network and
Little, T.N., Blair, D.P., 2010. Mechanistic Monte Carlo models for analysis of flyrock support vector machine models for the prediction of the liquefaction-induced
risk. Rock Fragm. Blasting 9, 641e647. uplift displacement of tunnels. Undergr. Space 6 (2), 126e133.
Lundborg, N., Persson, A., Ladegaard-Pedersen, A., Holmberg, R., 1975. Keeping the Zhou, J., Asteris, P.G., Armaghani, D.J., Pham, B.T., 2020a. Prediction of ground vi-
lid on flyrock in open-pit blasting. Eng. Min. J. 176, 95e100. bration induced by blasting operations through the use of the Bayesian
McKenzie, C.K., 2009. Flyrock range and fragment size prediction. In: Proceedings of Network and random forest models. Soil Dyn. Earthq. Eng. 139, 106390.
the 35th Annual Conference on Explosives and Blasting Technique, 2. Denver, Zhou, J., Koopialipoor, M., Li, E., Armaghani, D.J., 2020b. Prediction of rockburst risk
CO, USA. in underground projects developing a neuro-bee intelligent system. Bull. Eng.
Monjezi, M., Khoshalan, H.A., Varjani, A.Y., 2012. Prediction of flyrock and backbreak Geol. Environ. https://doi.org/10.1007/s10064-020-01788-w.
in open pit blasting operation: a neuro-genetic approach. Arab. J. Geosci. 5, Zhou, J., Koopialipoor, M., Murlidhar, B.R., Fatemi, S.A., Tahir, M.M., Armaghani, D.J.,
441e448. Li, C., 2020d. Use of intelligent methods to design effective pattern parameters
Naghibi, S.A., Dashtpagerdi, M.M., 2017. Evaluation of four supervised learning of mine blasting to minimize flyrock distance. Nat. Resour. Res. 29, 625e639.
methods for groundwater spring potential mapping in Khalkhal region (Iran) Zhou, J., Li, C., Koopialipoor, M., Jahed Armaghani, D., Pham, B.T., 2020c. Develop-
using GIS-based features. Hydrogeol. J. 25, 169e189. ment of a new methodology for estimating the amount of PPV in surface mines
Newcombe, R.G., 1998. Two-sided confidence intervals for the single proportion: based on prediction and probabilistic models (GEP-MC). Int. J. Min. Reclam.
comparison of seven methods. Stat. Med. 17, 857e872. Environ. https://doi.org/10.1080/17480930.2020.1734151.
Nguyen, H., Bui, X.N., Choi, Y., Lee, C.W., Jahed Armaghani, D., 2021. A novel com- Zhou, J., Huang, S., Wang, M., Qiu, Y., 2021a. Performance evaluation of hybrid GAe
bination of whale optimization algorithm and support vector machine with SVM and GWOeSVM models to predict earthquake-induced liquefaction po-
different kernel functions for prediction of blasting-induced fly-rock in quarry tential of soil: a multi-dataset investigation. Eng. Comput. https://doi.org/
mines. Nat. Resour. Res. 30, 191e207. 10.1007/s00366-021-01418-3.
Rätsch, G., Onoda, T., Müller, K.R., 2001. Soft margins for AdaBoost. Mach. Learn. 42, Zhou, J., Qiu, Y., Zhu, S., Armaghani, D.J., Li, C., Nguyen, H., Yagiz, S., 2021b. Opti-
287e320. mization of support vector machine through the use of metaheuristic algo-
Roth, J.A., 1975. A model for the determination of flyrock range as a function of shot rithms in forecasting TBM advance rate. Eng. Appl. Artif. Intell. 97, 104015.
condition. In: NTIS Report No PB81222358. US Department of Commerce, Zhou, J., Qiu, Y., Armaghani, D.J., Zhang, W., Li, C., Zhu, S., Tarinejad, R., 2021c.
Washington, DC, USA. Predicting TBM penetration rate in hard rock condition: a comparative study
Rousseeuw, P.J., Leroy, A.M., 2005. Robust Regression and Outlier Detection. John among six XGB-based metaheuristic techniques. Geosci. Front. 12 (3), 101091.
Wiley and Sons.
Sattar, A.M.A., 2014. Gene expression models for the prediction of longitudinal
dispersion coefficients in transitional and turbulent pipe flow. J. Pipeline Syst. Mahdi Hasanipanah obtained his BSc and MSc degrees in
Eng. Pract. 5, 4013011. Mining Engineering from University of Kashan, Iran, and
Scholz, F.W., Stephens, M.A., 1987. K-sample andersoneDarling tests. J. Am. Stat. Islamic Azad University-South Tehran Branch, Iran, in
Assoc. 82, 918e924. 2008 and 2010, respectively, and his PhD in Mining Engi-
Sevgen, E., Kocaman, S., Nefeslioglu, H.A., Gokceoglu, C., 2019. A novel performance neering from University of Kashan, Iran, in 2020. He was
assessment approach using photogrammetric techniques for landslide suscep- affiliated as Institute of Research and Development, Duy
tibility mapping with logistic regression, ANN and random forest. Sensors 19 Tan University, Vietnam, since 2019. His research interests
(18), 3940. include rock blasting, rock mechanics, machine learning
Sezer, E.A., Nefeslioglu, H.A., Gokceoglu, C., 2014. An assessment on producing methods, and optimization algorithms. He has been coop-
synthetic samples by fuzzy C-means for limited number of data in prediction erated in a large number of projects in Iran and Malaysia.
models. Appl. Soft Comput. 24, 126e134.
Shahrour, I., Zhang, W., 2021. Use of soft computing techniques for tunneling
optimization of tunnel boring machines. Undergr. Space 6 (3), 233e239.

You might also like