You are on page 1of 12

Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

Soil Dynamics and Earthquake Engineering 139 (2020) 106390

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering


journal homepage: http://www.elsevier.com/locate/soildyn

Prediction of ground vibration induced by blasting operations through the


use of the Bayesian Network and random forest models
Jian Zhou a, Panagiotis G. Asteris b, *, Danial Jahed Armaghani c, **, Binh Thai Pham d
a
School of Resources and Safety Engineering, Central South University, Changsha, 410083, China
b
Computational Mechanics Laboratory, School of Pedagogical and Technological Education, 14121, Heraklion, Athens, Greece
c
Institute of Research and Development, Duy Tan University, Da Nang 550000, Viet Nam
d
University of Transport Technology, Hanoi, 100000, Viet Nam

A R T I C L E I N F O A B S T R A C T

Keywords: The present study aims to compare the performance of two machine learning techniques that can unveil the
Blasting relationship between the input and target variables and predict the ground vibration (peak particle velocity,
Ground vibration PPV) due to quarry blasting. To this end, a Random Forest (RF) model and a Bayesian Network (BN) model were
Random forest
developed. Before developing these models, and in order to illustrate the necessity of proposing new intelligent
Bayesian network
Feature selection
systems, a new empirical equation is proposed, using maximum charge per delay and distance from the blast-
Machine learning face. The results confirm that there is indeed a need to develop intelligent systems with more input parame­
ters. Thus, a Feature Selection (FS) model is applied to decrease the dimensionality of data and remove the
irrelevant data. The outputs of this technique set five parameters, hole depth, power factor, stemming, maximum
charge per delay and distance from the blast-face, as the most important model inputs necessary to predict PPV.
After constructing FS-BN and FS-RF models and comparing them under different conditions (i.e., computational
cost, accuracy and robustness), it is found that the developed FS-RF model can be introduced as a new model in
the field of blasting environmental issues. The accuracy level of the FS-RF model is quite high; 92.95% and
90.32% for the train and test stages, respectively, while 92.95% and 87.09% accuracy is calculated for train and
test of the FS-BN model. Thus, both developed hybrid models can effectively unveil the relationships between the
input and target variables.

1. Introduction significance to predict the ground vibration accurately, in such a way,


that the blasting environmental effects can be alleviated as much as
One of the most extensively-used techniques for rock excavation possible.
purposes is blasting. On the other hand, the use of this technique is High-pressure gases are generated due to the detonation of the
associated with releasing a huge amount of explosive energy, which can explosive substances and chemical reactions that occur at the time of the
lead to a number of hazards, such as ground vibration, air over-pressure, blasting event. The pressure of the produced gas is able to crush the rock
fly-rock, back break, etc. [1–4]. Conducted studies related to blasting mass situated adjacent to the blast hole and the pressure of detonation
operations and surface mining have shown that the ground vibration is will fade or burst quickly. After that, a wavy motion is produced in the
one of the most destructive results of blasting, which can exert severe ground by the strain waves that are conveyed to the adjoining rocks
damages to neighbouring buildings [5–7]. Ground vibration can nega­ [11]. Strain waves run as the elastic wave, when the stress wave in­
tively affect a wide range of things in the vicinity of the operation site, tensity lowers down to the ground level [12]. Such waves are known as
including adjacent rock masses, slopes, underground works, roads, ground vibration.
railroads, water conduits, while it also has a negative effect on the In general, when recording ground vibrations two factors are taken
ecology of the neighbouring region [8–10]. Thus, it is of a high into account: frequency and peak particle velocity (PPV). PPV is utilized

* Corresponding author.
** Corresponding author.
E-mail addresses: csujzhou@hotmail.com (J. Zhou), panagiotisasteris@gmail.com (P.G. Asteris), danialjahedarmaghani@duytan.edu.vn (D.J. Armaghani),
binhpt@utt.edu.vn (B.T. Pham).

https://doi.org/10.1016/j.soildyn.2020.106390
Received 6 February 2020; Received in revised form 10 May 2020; Accepted 26 August 2020
Available online 17 September 2020
0267-7261/© 2020 Elsevier Ltd. All rights reserved.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

when measuring the ground vibrations since it is a critically significant tree-based, optimization and other, have been utilized for the predic­
index that plays an effective role in controlling the damage criteria [9, tion and optimization of PPV induced by blasting. Among them, the
13,14]. Literature has recently introduced a number of empirical PPV ANN groups have been extensively-used (more than 40 times) in this
predictors [11,15–18]. Two key factors that can be utilized to estimate area of research, while only 8 studies with the use of FIS-based group are
PPV value are the distance from the blast-face and the maximum charge available in literature. However, optimization techniques have been
per delay [19,20]. High degree of PPV prediction is required to deter­ rarely-applied in the field of PPV. Tree-based groups of techniques like
mine the safety area when performing the blast operations; therefore, GP, GEP and classification and regression trees (CART) have been pro­
the above-noted empirical approaches cannot be adequate. The pre­ posed only about 12 times for PPV prediction. It should be noted that the
diction of PPV is also affected by a number of other parameters such as information in Fig. 1, is derived from the available PPV studies for the
powder factor, and stemming. Non-existence of these parameters in the authors and there may exist some other related studies in literature.
PPV prediction process could be the main reason for inadequacy of the To be more specific, some of the most important relevant studies of
above-noted empirical approaches [21,22]. Moreover, some other re­ PPV using ML and SC techniques are reviewed here. Literature contains
searchers introduced semi-empirical approaches to evaluated and a number of studies confirming that ANN outperforms the statistical and
determine PPV resulting from blasting. For example, in order to mini­ empirical techniques (e.g. Ref. [5,6,63]). However, some researchers
mize the damages caused by PPV to the nearby structures, Hakan et al. have reported results of higher accuracy with the use of fuzzy-based
[1] suggested a graphical prediction technique for PPV evaluation, using models like FIS and ANFIS in comparison with the previous ones (e.g.
the distance from the blast-face and the maximum charge per delay Refs. [7,64–66]). In some other studies, the authors attempted to high­
parameters. Considering the same important parameters, similar graphs light the advantage of proposing SVM-based models to solve the PPV
have been proposed to determine and estimate the PPV values for problems (e.g. Ref. [20,67,68]). In addition, hybrid ANN models like
minimizing the possible structural damages (e.g. Ref. [23,24]). PSO-ANN have been selected as the best predictive model of PPV in
In addition to empirical predictors, literature consists of a number of literature (e.g. Ref. [22,69]). Tree-based models such as CART, GP and
statistical techniques introduced by different researchers for PPV pre­ GEP were also introduced as powerful and applicable models in esti­
diction [25–28]. A key advantage of the statistical techniques over the mating the PPV ([27,28,70–72]). Some other techniques such as boosted
empirical ones, is that the former takes into consideration more pre­ generalised additive and eXtreme gradient boosting have also been used
dictors, including the predictors related to explosive substances, rock for PPV estimation [73,74].
mass property, and blasting design [2,24,29]. On the other hand, the In this frame, the formulation of more and more reliable models to
most important disadvantage of the statistical techniques is that their predict PPV induced by blasting is a possible goal. To the best of the
application is limited in cases where new available data are different authors’ knowledge, no research is available that investigates the use of
from the original data [5,30–32]. advanced techniques, such as feature selection (FS) for input selection
In recent years, a number of soft computing (SC) and machine and random forest (RF) and Bayesian Network (BN) for the prediction of
learning (ML) techniques have been extensively utilized to find effective the PPV. To this end, this study developed two hybrid models, including
solutions to civil and mining engineering issues, as well as the ground the FS-RF and FS-BN models, for PPV evaluation. These techniques were
vibration problems resulting from blasting operations [33–62]. The SC selected, as they are well-known for unveiling the relationships between
and ML techniques such as genetic algorithm, adaptive neuro-fuzzy the input variables and out variable, as well as relationships within the
inference system (ANFIS), artificial neural network (ANN), support input variables [75,76]. The next sections explain the techniques
vector machine (SVM), gene expression programming (GEP), fuzzy employed in this present study. In addition, comprehensive information
inference system (FIS), imperialism competitive algorithm (ICA), par­ of the dataset is provided. Thus, a new empirical equation is developed
ticle swarm optimization (PSO), and genetic programming (GP) have to predict PPV values, which is then compared with the most
been applied in the area of PPV prediction. In order to have a better well-known equations. This paper also describes the implementation of
understanding regarding the frequent use of these techniques, Fig. 1 the models developed for evaluating the PPV. Finally, the results of the
shows some of the different technique groups used for PPV prediction developed models are compared together in terms of models’ robust­
and optimization. According to this figure, eight (8) groups of tech­ ness, computational cost, and accuracy.
niques i.e., ANN, hybrid ANN, FIS-based, ANFIS-based, SVM-based,

Fig. 1. Some of the different technique groups used for PPV prediction and optimization.

2
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

2. Methods Table 1
Screening rules for conducting the FS technique.
Both RF and BN methods were utilized in an efficient way aiming at Screening rules Cut-off
predicting many outcomes and also investigating the relationships and
Maximum percentage of missing values 70
causality of the considered variables. Both RF and BN have the capacity Maximum percentage of records in a single category 90
of predicting the target variable with the help of certain rules or con­ Maximum number of categories as a percentage of records 95
ditions. In addition, the predictive outputs of the above-noted methods Minimum coefficient of variation 0.1
can be described and compared quantitatively. The prediction power of Minimum standards deviation 0.0
Value for important variables 0.95
these methods can be improved hybridizing with a well-known input
selection technique, FS. The present study developed two hybrid models
i.e., FS-RF and FS-BN to select the input parameters and predict the PPV DTs are indeed models of high instability and variability; for this reason,
values. This study then compares the accuracy performance of these the trees may show various generalization behaviours with small al­
models. In addition, the computational cost and robustness of these terations to the training data [78]. Moreover, RF adopts the bagging
models were assessed. It is worth mentioning that a new empirical technique in order to form each member of the ensemble from various
equation was also proposed based on the parameters of distance from training datasets. In a random way, the bagging samples generates from
the blast-face and the maximum charge per delay and the proposed the space of DTs virtually similar predictions (i.e., with a low diversity).
empirical equation was evaluated using the available standards/sug­ Table 2 demonstrates the model development rules and thresholds.
gested methods. Fig. 2 shows the steps implemented in this study in
order to evaluate and predict PPV induced by blasting. 2.1.3. Bayesian Network
Bayesian network (BN) is referred to as a graphical model type
2.1. Machine learning techniques representing multivariate probability distributions [77]. Such repre­
sentation can be used to generate a directed acyclic graph (DAG) (see
2.1.1. Feature selection (FS) Fig. 3). The network nodes signify a set of random variables, while the
Feature selection (FS) is capable to effectively enhance the ML and directed arcs represent the causal relationships among the variables. BN
SC performance quality by lowering data dimensionality. FS is mainly makes use of the Tree Augmented Naive Bayes (TAN) method [79] for
aimed at removing the redundant and/or irrelevant features and keep­ the purpose of addressing the interdependencies that exist among the
ing only relevant ones. It should be noted that the elimination of irrel­ variables (Fig. 3). Among all, TAN is the best technique that can be most
evant features has no impact on learning performance. Literature effectively used by BN, as it is able to relax the assumption of naive
comprises various advantages of applying FS to ML and SC techniques. Bayes variable independence by utilizing a tree structure in which each
For instance, FS can enhance the predictive precision of these techniques variable is dependent only upon the output variable and one other
and effectivity of both learning and data gathering [77]. In this paper, variable. Table 3 presents the rules applicable to the BN development.
aiming to recognize the input variables of the highest relevancy and
significance, FS is used to rank the input variables based on intrinsic
properties of the data. As a result, the top variables were selected based Table 2
on thresholds. Table 1 presents the screening rules and thresholds. FS Model development rules for conducting the RF method.
was also utilized to determine the quality of the variables considering Model development rules Cut-off
the correlation that existed between input variables and output one.
Maximum percentage of missing values 70
Then, FS chooses the variables that showed the highest correlations. Exclude fields with a single category majority over (%) 95
Additionally, FS made use of the Pearson chi-square for the purpose of Maximum number of field categories 49
testing the independence level of the output and input variables since all Minimum field variation 0.05
of them were categorical. Number of bins 10
Number of models to build 100
Maximum number of nodes 10000
2.1.2. Random forest Maximum tree depth 10
RF is known as a method of ML, which is able to ensemble many Minimum child node size 5
decision trees (DTs) in order to generate estimations of higher precision.

Fig. 2. Steps taken in this study.

3
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

these parameters included the burden, hole diameter, total charge,


spacing, distance from the blast-face, hole depth, stemming length, sub-
drilling, number of holes, powder factor, and maximum charge per
delay. In addition, to record the PPV values, the Vibra ZEB seismogra­
pher was employed. The distance between the seismograph and the face
of blasting was measured in each blasting which was in the range of
(285–531 m). The significance of this parameter is specified considering
the impacts of this phenomenon. After collection of the data from the
blasting site and analyzing them, while at the same time taking
previously-published articles into consideration, stemming length (ST),
burden to spacing ratio (BS), distance from the blast-face (D), powder
factor (PF), maximum charge per delay (MC), and hole depth (HD) with
rages of (1.9-3.6 m), (0.7-0.92), (285-531 m), (0.23-0.94 kg/m3), (45.8-
305.6 kg), and (5.2-23.2 m), respectively, were selected to apply in the
modelling. Additionally, PPV values with minimum, average and
maximum of 0.13, 5.34, and 11.05 mm/s, respectively were set as the
model output. Fig. 4 shows matrix analysis chart of the input and output
parameters used in this study. In this figure, coefficient of correlation
between the input variables themselves and between the input and
output parameters is presented. It can be seen that some variables have a
meaningful relationships to each other and some others do not have any
relationships. In the following section, and after proposing a new
empirical equation, the model inputs are selected through the use of the
FS technique. Then, the FS-RF and FS-BN models are developed with
their most influential parameters and finally, the best hybrid predictive
model in terms of the PPV prediction will be chosen and introduced.

3. Empirical equation

In this section, a new empirical formula is developed according to the


two important parameters suggested in literature (i.e., MC and D) to
estimate PPV resulting from blasting and then the performance predic­
tion of the proposed equation is compared with the most important
Fig. 3. BN model created by TAN. previously-published equations in literature, described in detail in the
two following sub-sections, “proposing a new empirical equation” and
“performance prediction comparison”.
Table 3
Model development rules for conducting the BN method.
3.1. Proposing a new empirical equation
Model development rules Cut-off

Independence test Likelihood ratio


This section presents the development of a new empirical equation
Significance level 0.01
for the prediction of PPV resulting from blasting. In view of previous
studies related to empirical equations of PPV, it was found that a general
2.2. Comparison of the models’ developed method proposed by Duvall and Petkof [80] is the most frequently-used
and acceptable. Based on their suggested technique [84], there is a need
The proposed RF-BN models were assessed in several ways. First, the to define the parameter of scaled distance (SD) as follows:
FS-RF and FS-BN models were tested for robustness through a sensitivity ( )
D
analysis. Second, the computational cost of the proposed models was SD = √̅̅̅̅̅̅̅̅ (1)
MC
assessed through a twofold comparison between the FS-RF and FS-BN
model with respect to: (1) the time required to run these models and where, D is the distance from the blast-face (m) and MC is the maximum
(2) the achieved accuracy of these two types of models. charge per delay (kg). Then, they suggested that the PPV values can be
estimated through the use of a general form as follows:
2.3. Case study and data
PPV = A (SD)B (2)
In the present study, to accomplish the defined prediction objectives,
a broad dataset was prepared through gathering data from a blasting where, A and B are site constants. After calculations of SD values using D
mine where explosive operations were undertaken. Generally, con­ and MC, the PPV values obtained from the blasting operations were put
ducting explosive operations is very common in this mine and these in Equation (2) and the following equation with A = 81089 and B =
operations are repeated at various intervals. The mine has the capacity − 2.941 was derived:
of producing around 500–700 thousand tons/year, and for the detona­
PPV = 81089 (SD)− 2.941
(3)
tions, a large quantity of explosive substances is used. Typically,
856–9420 kg of explosives are put in blast-holes with diameters of These coefficients i.e., A and B and generally Equation (3) were
76–89 mm. High-explosive operations augment the risk of destruction to developed based on the data collected from the quarry site with rock
the mine. For the purpose of the present study, 102 blasting operations type of granite. It should be noted that Equation (3) can only be used for
were identified and evaluated meticulously taking into account all the same quarry site and its blasting operations. The performance of the
design details. Significant parameters were determined and recorded; newly-proposed empirical equation based on coefficient of

4
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

Fig. 4. Scatterplot matrix of PPV database with correlation.

determination is 0.672 which is considered as acceptable and appli­ equations which are well-known in literature, were selected (Table 4)
cable. It should be mentioned that after the suggested method of Duvall and then, 2 performance indices namely R2, and the a20-index (Equa­
and Petkof [80], some other researchers (e.g. Refs. [16,17]) have tions (4) and (5), respectively) were considered and applied. Table 5
introduced several empirical equations based on MC and D parameters. presents the results of these performance indices for all empirical
Figs. 5 and 6 show the measured PPV values against the amounts of SD in equations.
forms of power and logarithmic, respectively. According to these figures, ∑( )2
the new empirical equation is adequate to predict PPV values. It is R2 = 1 − i ximes − xipred
∑ (4)
important to note that these results are based on the influence of only 2
2
i (ximes − x)
parameters of MC and D to forecast PPV. More details regarding the new
empirical model will be given in the next section. a20 − index =
m20
× 100 (5)
N

3.2. Performance prediction comparison where, ximes and xipred are the measured and predicted values, respec­
tively. x represents the average measured values, N is the total No. of
It is important to evaluate the predictive models/equations using the datasets, and m20 is the No. of samples with values of rate measured/
most significant performance indices. To this end, and in order to show predicted value (range between 0.8 and 1.2). The values of one and
applicability of the newly-proposed empirical model first, 5 empirical

Fig. 5. The power form of SD verses PPV and their derived equation.

5
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

Fig. 6. The logarithmic form of SD verses PPV.

closer to the measured PPV values in the quarry site in comparison with
Table 4
the previously-published equations. This shows the prediction power of
Some of the important empirical equations for PPV estimation.
this equation, however, it seems that by using only 2 input parameters or
Reference Equation Site constant (granite rock variables i.e., MC and D, it is not possible to get an excellent PPV pre­
mass)
diction capacity. Therefore, the use of other important parameters such
Langefors - Kihlstrom PPV = A[√(MC/D2/3 )]B A: 44.43 and B: 1.18 as powder factor and stemming together with MC and D is of interest and
[17]
advantage. In other words, utilizing multi-inputs for prediction of PPV
Davies et al. [16] PPV = QD− B (MC)A Q: 212.27, B: 1.09 and A:
0.52 may lead us to receive a higher level of accuracy compared to proposing
Indian Standard [13] PPV = A[(MC/D2/3 )]B A: 6.33 and B: 0.22 empirical equation. Hence, in the modelling of this study, ST, PF, HD,
Ghosh - Daemen [81] A: 780.36, B: 1.26 and α: and BS, were used together with MC and D to get a higher performance
PPV =
A[D/√MC]− B e− αD 0.0004 prediction of the PPV values.
Roy [18] PPV = n + A: 168.91 and n: 1.57
A[D/√MC]− 1 4. Modelling results

4.1. Feature selection model for input selection

Table 5 The FS technique was used to remove the irrelevant data, as well as
Applying 2 performance indices on the empirical equations.
to reduce the data dimensionality. After constructing FS model, from the
Equation R2 a20-index (%) six input variables, the five most important variables, including D, PF,
Langefors - Kihlstrom [17] 0.275 0 ST, HD and MC, were selected to be used as inputs for developing the RF
Davies et al. [16] 0.654 0 and BN models. The importance of D, PF, ST and HD variables was one
Indian Standard [13] 0.293 0 (1.00), while for the MC parameter it was 0.95. The FS model removed
Ghosh - Daemen [81] 0.635 0
the BS parameters from the inputs’ list because of their low importance.
Roy [18] 0.642 0
This study 0.672 3.1 It should be mentioned that in modelling of FS, the minimum impor­
tance value that we defined for the system was 0.95. The authors
believed that the method of FS should be used in many civil and mining
100% are required for performance indices of R2 , and the a20-index, studies related to computational techniques, as a step prior to modelling.
respectively, if a perfect predictive model is of interest. According to Using this technique or other similar problems, it is possible to control or
results obtained in Table 5, for both indices, the new empirical equation even reduce the model complexity which is an important issue in pre­
in this study showed better prediction performance. In terms of R2, diction and simulation studies.
values of 0.275, 0.654, 0.293, 0.635, 0.642 and 0.672 obtained by
equations of Langefors – Kihlstrom [17], Davies et al. [16], Indian
Standard [13], Ghosh - Daemen [81], Roy [18] and this study, 4.2. Random forest model
confirmed that the newly-proposed equation is able to perform better
than the previous well-known equations. Furthermore, and aiming to The RF model was developed using the five mentioned input vari­
assess the models through an additional alternative method, 10 datasets ables. Before development of the RF models, the data were split into the
were selected randomly from all 102 datasets and the amounts of train and test partitions. Thus, 70% (n = 71) of the data were used for
measured PPV together with the values predicted by the equations of the train phase and 30% (n = 31) were used for the test phase. As
Langefors – Kihlstrom [17], Davies et al. [16], Indian Standard [13], pointed out earlier, RF builds an ensemble model that includes many
Ghosh - Daemen [81], Roy [18] and this study were plotted in one graph DTs (in this study n = 100). It is worth noting that the RF is built on
(Fig. 7). According to Fig. 8, in most of the cases (datasets), the PPV CART methodology. Moreover, several parameters were used to develop
results obtained by the proposed empirical equation in this study are and fine-tune the RF model. The authors quantified the maximum
number of leaf nodes that are permitted in individual trees. If the

6
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

Fig. 7. Comparison of the measured PPV values with the results of empirical equations.

Table 6
Decision rules for PPV derived from the RF model.
No Decision rule Most Rule Forest Interestingness
frequent accuracy accuracy index
category

1 ST > 2.3 AND 2 1.0 1.0 1.0


MC≤223.55 AND
PF > 0.47
2 MC≤147.0 AND 2 1.0 1.0 1.0
D≤395.0 AND
0.47≤PF≤0.85
AND ST > 2.89
3 PF > 0.65 AND 2 1.0 1.0 1.0
308 < D < 406.0
AND MC > 80.55
AND HD≤11.0
4 ST≤2.89 AND 4 1.0 1.0 1.0
MC > 198.21
AND PF > 0.47
5 122.44 < 4 1.0 1.0 1.0
MC≤165.14 AND
ST≤2.4 AND
D≤376.0 AND
HD > 11.0

PPV category 2 = 2.86-5.59 mm/s, and PPV category 4 = 8.32-11.05 mm/s.

parent node is divided, was set. If a child node would comprise fewer
records than the value entered (in this study, the minimum number of
records was 5), the parent node will not be divided. The RF model was
built using the above-mentioned parameters. The accuracy of the best
RF model was 92.95% for the training datasets and 90.32% for the
testing datasets. The FS-RF model identified input of D, as the most
important factor with value of 0.25 followed by PF (0.23), ST (0.20), HD
(0.18), and MC (0.14), respectively.
The RF model identified several decision rules, which are presented
in Table 6. This table shows the details of how the rule is applied and
combined, if the results are in the most frequent category, rule accuracy,
Fig. 8. Graph of BN model to probability evaluation of PPV values. trees accuracy, and interestingness index (INi). The INi was calculated
using the following formula:
maximum number would be exceeded during the next split, the tree ( ( ⃒

))
INi (t) = P(A(t)) × P(B(t)) × P(B(t)|A(t)) + P B(t)⃒A(t) (6)
growth is stopped before the partition occurs. They also quantified the
maximum number of levels under the root node; that is to say, the
number of times the sample is divided recursively. The minimum where, P(A(t)) denotes the trees accuracy; P(B(t)) denotes the rule ac­
number of records that must be kept in check in a child node, after the curacy; P(B(t)|A(t)) signifies correct predictions by both the trees and
the node.

7
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

The RF could not find any rulesets to predict the occurrence of PPV of Table 7
“1” and “3”. In addition, three decision rules resulting in the PPV of “2” Conditional probabilities greater than 0.75 for predicting the PPV using the BN
were identified. It is worth mentioning that the PF and MC were the most model.
frequently appeared factors in these decision rules. Concerning the PPV PPV Hole Stemming Distance Powder Max Probability
of “4”, the RF identified two decision rules. ST and MC were also depth factor charge
frequently appeared factors to predict the occurrence of PPV of “4”. per
delay

4.3. Bayesian Network model 1 15.994- 2.92-3.26 1.00


19.582
1 15.994- 432.6- 1.00
Five input variables including D, PF, HD, ST and MC were used to 19.582 481.8
develop the BN model and predict the occurrence probability of the PPV. 1 15.994- 201.68- 1.00
Similarly to the RF model, several BN parameters were used to develop 19.582 253.64
the PPV model using the BN method. The authors used the maximum 1 15.924- 0.514- 1.00
19.582 0.656
likelihood as the learning method which refers to the conditional 2 ≤ 8.818 ≤ 334.2 1.00
probabilities for each node assumed the values of its parents. To examine 2 ≤ 8.818 ≤ 97.76 1.00
whether paired observations on two variables are independent of each 2 ≤ 8.818 0.514- 1.00
other the likelihood ratio was employed. This ration assessed the inde­ 0.656
3 15.924- 0.83
pendence of target and predictor by calculating a ratio between the
>0.798
19.582
maximum probabilities of a result under two different hypotheses. A 4 8.818- ≤ 2.24 1.00
significance level of 0.01 was set in conjunction with the independence 12.406
test settings. 4 15.994- 2.24-2.58 0.75
Using the abovementioned parameters the BN model was created. 19.582
4 >19.582 ≤ 2.24 0.88
The accuracy of the model was 92.95% for the training datasets and 4 8.818- ≤ 334.2 1.00
87.09% for the testing datasets. The parameter of D was identified as the 12.406
most important factor with value of 0.24 followed by PF (0.23), HD 4 12.406- ≤ 334.2 0.78
(0.20), MC (0.17), and ST (0.15). The modelling graph derived from the 15.994
4 15.994- ≤ 334.2 1.00
BN model is shown in Fig. 8. The BN network encompassed of six nodes,
19.582
one for the target (PPV) and five for the predictors. The BN graph shows 4 >19.582 ≤ 334.2 0.75
that the binary compositions of ST and HD, PF and HD, D and HD, as well 4 8.818- 97.76- 1.00
as MC and HD are desirable predictors for predicting the PPV 12.406 149.72
occurrence. 4 8.818- 0.656- 1.00
12.406 0.798
Besides, several conditional probabilities were identified using the 4 15.924- >0.798 1.00
BN model. This study reported only those probabilities that achieved a 19.582
probability of greater than 0.75 (Table 7). According to the results, four 4 >19.582 >0.798 0.75
conditional probabilities (greater than 0.75) were identified for PPV “1”. PPP category 1 = 0.13-2.86 mm/s, PPV category 2 = 2.86-5.59 mm/s, PPV
It is also shown that the HD values between 15.994 and 19.582 m are the category 3 = 5.59-8.32 mm/s, and PPV category 4 = 8.32-11.05 mm/s.
most influential for this category of PPV. The BN identified three con­
ditional probabilities (greater than 0.75) for predicting the PPV of “2”.
by both models of FS-RF and FS-BN. The results also show that both
As can be seen, the HD of 8.818 m or less was the most influential value.
models predicted the PPV somewhat similarly.
Concerning the PPV of “3”, the BN identified only one conditional
probability. In addition, 11 conditional probabilities were identified for
5.2. Models’ robustness
PPV of “4”. For this category of PPV, several values of HD were influ­
ential, while several other factors, including ST of 2.24 m and less, D of
A sensitivity analysis was conducted to assess the proposed models’
334.2 m and less, MC of 97.76-149.72 kg, PF of 0.656-0.798 kg/m3 and
robustness. Herein, the robustness refers to the changes of accuracy in
0.798 kg/m3 and above were influential on PPV with perfect probability
the models developed against any change in the dataset. A variable-
(1.00).
exclusion approach was used to analyse the models’ sensitivity. The
most important predictor in each model was excluded and then the
5. Comparison of the models developed
models were run. Thus, the parameter of D in FS-RF and FS-BN models
were excluded. Table 9 shows the accuracy achieved after models
5.1. Models’ accuracy
modifications. As can be seen, the FS-RF model was substantially more
robust against the modifications than the FS-BN model in training phase,
The accuracy of the FS-RF and FS-BN models are presented in Figs. 9
while the FS-BN model showed a better robustness in testing phase. In
and 10, respectively. A simple check on these figures illustrates that the
other words, when D which is the most important parameter identified
accuracy of the FS-RF model is slightly higher than FS-BN model for the
by both RF and BN, was removed from the inputs, the changes in per­
testing phase, while both models showed a same accuracy level for
formance prediction of the RF model in training stage is negligible while
training phase. The accuracy performances of the FS-RF model for the
the changes in testing stage is considerable. Moreover, the rate of
training and testing phases was very close to those of the FS-BN model.
changes in training stage of BN is higher than the RF model whereas this
Thus, it can be said that the overall performance of FS-RF model was
rate in testing stage is lower compared to the RF model. It is worth
better than the FS-BN model. It should be noted that in simulation and
noting that the parameter of D which is selected as the most effective one
prediction studies, a higher performance prediction obtained by testing
among all five model inputs, was identified as one of the most important
datasets shows that the specific model was more successful and powerful
factors on PPV in previously published empirical equations as well [11,
in model development or the training stage.
17,18]. Additionally, in many studies of PPV prediction using ML and SC
Moreover, the agreement between these two models is indicated in
techniques [5,6,27,82], this parameter has been chosen as the most
Table 8. According to this table, the agreement between correct pre­
significant one in their developed models.
dictions is 88.73% for the training datasets and 93.54% for testing
datasets. This agreement refers to samples that were similarly predicted

8
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

Fig. 9. Accuracy of FS-RF model.

Fig. 10. Accuracy of FS-BN model.

5.3. Models’ computational cost


Table 8
Agreement between FS-RF and FS-BN models.
The authors compared the time required to run the FS-RF model and
Partition Train % Test % FS-BN model to assess the computational cost of the proposed models.
Agree 63 88.73 29 93.54 The authors observed that the total required time to run the FS-RF model
Disagree 8 11.27 2 6.46 was almost 37 s (2 s for FS+35 s for RF) and run-time for FS-BN model
Total 71 31
was almost 5 s (2s for FS+3s for BN). While considering cost-benefit, the
FS-BN model decreased the time for developing the model by almost
seven-fold, but decreased the testing accuracy by over 3%. However,
Table 9 this model did achieved the same training accuracy achieved by the FS-
Accuracy of the modified models. RF model.
Modified models Number of Training Testing
inputs accuracy (%) accuracy (%)

Main FS-RF model 5 92.95 90.32 5.4. General discussion


Modified FS-RF model 4 91.55 80.65
(variable-exclusion Both FS-RF and FS-BN as new models in the field of PPV prediction,
approach/distance)
were developed using 5 model inputs. Then, these models were evalu­
Main FS-BN model 5 92.95 87.09
Modified FS-BN model 4 80.28 80.65
ated and compared according to 3 different categories: accuracy,
(variable-exclusion robustness and computational cost. Results of accuracy category showed
approach/distance) that the FS-RF is more successful in both training and testing stages
compared to the FS-BN model. They received the same performance

9
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

prediction for training stage, but the FS-RF model received higher pre­ CRediT authorship contribution statement
diction capacity for testing datasets compared to the FS-BN model.
Regarding the robustness category, it was found that the FS-RF model is Jian Zhou: Conceptualization, Formal analysis, Writing - review &
slightly better than the FS-BN model for training datasets when the editing. Panagiotis G. Asteris: Supervision, Writing - review & editing.
parameter of D is removed from the inputs, while the results of testing Danial Jahed Armaghani: Conceptualization, Data curation, Writing -
datasets are inverse. Concerning the last category which is computa­ review & editing. Binh Thai Pham: Formal analysis, Writing - review &
tional cost, the modelling results revealed that the FS-BN model is editing.
significantly better than the FS-RF model. In light of the above discus­
sion and considering the fact that in civil and mining applications, the
Declaration of competing interest
predictive models with higher performance prediction (especially for
testing datasets) are of interested and advantage, the developed FS-RF
The authors declare that they have no known competing financial
model was selected as the best predictive model in this study to esti­
interests or personal relationships that could have appeared to influence
mate PPV induced by blasting and it can be used in further relevant
the work reported in this paper.
studies.
It is important to state that some studies in the field (e.g. Ref. [22,66,
69,83]), considering many model inputs (in the range of 7–11), intro­ Acknowledgment
duced intelligent systems with a higher accuracy level in comparison to
the current study. However, the objective of this study was to reduce the This research was funded by the Innovation-Driven Project of Central
number of input parameters, thus making the developed model easier to South University (2020CX040) and the Shenghua Lieying Program of
utilize. Additionally, through this approach, the level of model Central South University (Principle Investigator: Dr. Jian Zhou).
complexity can be minimized compared to the use models with a higher
number of inut parameters. Moreover, the preparations of the suggested Appendix A. Supplementary data
model inputs are easier for further researchers, designers and engineers.
It can be concluded from this section that proposing a simpler intelligent Supplementary data to this article can be found online at https://doi.
system to predict PPV comes with more advantages than a complex one org/10.1016/j.soildyn.2020.106390.
consisting of many inputs, with only a slightly higher accuracy. Of
course accuracy is of high importance; however setting a balance be­ References
tween model complexity and model accuracy is a better strategy to enjoy
the simultaneous advantages of both important issues. [1] Ak H, Iphar M, Yavuz M, Konuk A. Evaluation of ground vibration effect of blasting
operations in a magnesite mine. Soil Dynam Earthq Eng 2009;29:669–76.
[2] Dindarloo SR. Prediction of blast-induced ground vibrations via genetic
6. Conclusions programming. Int J Min Sci Technol 2015;25. https://doi.org/10.1016/j.
ijmst.2015.09.020.
[3] Armaghani DJ, Mahdiyar A, Hasanipanah M, Faradonbeh RS, Khandelwal M,
This study set out to assess the performance of two advanced ML Amnieh HB. Risk assessment and prediction of flyrock distance by combined
techniques, RF and BN, to predict the PPV. To confirm that a predictive multiple regression analysis and Monte Carlo simulation of quarry blasting. Rock
model with more than 2 mentioned input parameters in the literature is Mech Rock Eng 2016;49:1–11. https://doi.org/10.1007/s00603-016-1015-z.
[4] Armaghani DJ, Hajihassani M, Sohaei H, Mohamad ET, Marto A, Motaghedi H,
needed, a new empirical equation was proposed and compared with the et al. Neuro-fuzzy technique to predict air-overpressure induced by blasting. Arab J
most well-known previously published equations. The results showed Geosci 2015;8:10937–50. https://doi.org/10.1007/s12517-015-1984-3.
that there is indeed a need to develop models with more than 2 model [5] Khandelwal M, Singh TN. Prediction of blast-induced ground vibration using
artificial neural network. Int J Rock Mech Min Sci 2009;46:1214–22.
inputs. After this, and aiming to choose the most significant input var­
[6] Monjezi M, Ghafurikalajahi M, Bahrami A. Prediction of blast-induced ground
iables for proposing ML models, a FS model was constructed and, based vibration using artificial neural networks. Tunn Undergr Space Technol 2011;26:
on the results, from the six initial inputs, the FS model selected the five 46–50. https://doi.org/10.1016/j.tust.2010.05.002.
[7] Armaghani D, Momeni E, Abad S. Feasibility of ANFIS model for prediction of
most important variables which included D, PF, HD, ST and MC.
ground vibrations resulting from quarry blasting. Environ Earth Sci 2015;74:
Consequently, these input variables were used to develop the RF and BN 2845–60.
models. The results of the modelling section of the paper, illustrate that [8] Toraño J, Ramírez-Oyanguren P, Rodríguez R, Diego I. Analysis of the
both FS-RF and FS-BN models are able to provide the same level of ac­ environmental effects of ground vibrations produced by blasting in quarries. Int J
Min Reclamat Environ 2006;20:249–66.
curacy (92.95%) for training datasets, while the accuracy of the FS-RF [9] Singh TN, Singh V. An intelligent approach to prediction and control ground
model (90.32%) is slightly higher than that of the FS-BN model vibration in mines. Geotech Geol Eng 2005;23:249–62.
(87.09%) during the testing phase. Both models were capable to unveil [10] Faramarzi F, Farsangi MAE, Mansouri H. Simultaneous investigation of blast
induced ground vibration and airblast effects on safety level of structures and
the relationships between the input variables and the target variable human in surface blasting. Int J Min Sci Technol 2014;24:663–9.
(PPV). The FS-BN technique represented the PPV model as a directed [11] Duvall W, Petkof B. Spherical propagation of explosion-generated strain pulses in
acyclic graph, which facilitated the understanding of empirical data, as rock. U.S. Department of the Interior, Bureau of Mines; 1959.
[12] Dowding CH. Suggested method for blast vibration monitoring. In: Int. J. Rock
well as the probability concept. On the other hand, the FS-RF model did mech. Min. Sci. Geomech. Abstr. vol. 29. Elsevier; 1992. p. 145–56.
not provide any graphical representation, because the RF technique used [13] Standard I. Criteria for safety and design of structures subjected to under ground
ensemble and bagging approaches to predict the target variable and it blast. ISI; 1973. IS-6922.
[14] Kahriman A. Analysis of ground vibrations caused by bench blasting at can open-
was not easy to build a tree and provide a graphical interface. Thus, the
pit lignite mine in Turkey. Environ Geol 2002;41:653–61.
lack of graphical representation can be a drawback of the RF technique. [15] Ambraseys NN, Hendron AJ. Dynamic behaviour of rock masses. J. Wiley & Sons;
After evaluating these 2 models regarding model accuracy and compu­ 1968.
[16] Davies B, Farmer IW, Attewell PB. Ground vibration from shallow sub-surface
tational cost, it was concluded that the FS-RF model is the best amongst
blasts. Engineer 1964;217.
the two developed models for PPV prediction. However, each of these [17] Langefors U, Kihlström B. The modern technique of rock blasting. New York:
models has its advantages and drawbacks, which should be considered Wiley; 1963.
by researchers, especially in the field of blasting and its environmental [18] Roy P. Putting ground vibration predictions into practice. Colliery Guard 1993;
241:63–7.
issues. Another important output of this study is that a simpler intelli­ [19] Armaghani DJ, Hasanipanah M, Amnieh HB, Mohamad ET. Feasibility of ICA in
gent system, with a slightly lower accuracy level, is better than a com­ approximating ground vibration resulting from mine blasting. Neural Comput Appl
plex intelligent system consisting of many inputs with only a slightly 2018;29:457–65.
[20] Hasanipanah M, Monjezi M, Shahnazar A, Armaghani DJ, Farazmand A. Feasibility
higher level of prediction performance. of indirect determination of blast induced ground vibration based on support
vector machine. Measurement 2015;75:289–97.

10
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

[21] Khandelwal M, Singh TN. Evaluation of blast-induced ground vibration predictors. ground vibration. Eng Comput 2019. https://doi.org/10.1007/s00366-019-00895-
Soil Dynam Earthq Eng 2007;27:116–25. x.
[22] Armaghani DJ, Hajihassani M, Mohamad ET, Marto A, Noorani SA. Blasting- [51] Sarir P, Chen J, Asteris PG, Armaghani DJ, Tahir MM. Developing GEP tree-based,
induced flyrock and ground vibration prediction through an expert artificial neural neuro-swarm, and whale optimization models for evaluation of bearing capacity of
network based on particle swarm optimization. Arab J Geosci 2014;7:5383–96. concrete-filled steel tube columns. Eng Comput 2019. https://doi.org/10.1007/
[23] Nateghi R. Evaluation of blast induced ground vibration for minimizing negative s00366-019-00808-y.
effects on surrounding structures. Soil Dynam Earthq Eng 2012;43:133–8. [52] Asteris PG, Nikoo M. Artificial bee colony-based neural network for the prediction
[24] Hajihassani M, Jahed Armaghani D, Marto A, Tonnizam Mohamad E. Ground of the fundamental period of infilled frame structures. Neural Comput Appl 2019.
vibration prediction in quarry blasting through an artificial neural network https://doi.org/10.1007/s00521-018-03965-1.
optimized by imperialist competitive algorithm. Bull Eng Geol Environ 2014;74: [53] Zhou J, Bejarbaneh BY, Armaghani DJ, Tahir MM. Forecasting of TBM advance
873–86. https://doi.org/10.1007/s10064-014-0657-x. rate in hard rock condition based on artificial neural network and genetic
[25] Hudaverdi T. Application of multivariate analysis for prediction of blast-induced programming techniques. Bull Eng Geol Environ 2019. https://doi.org/10.1007/
ground vibrations. Soil Dynam Earthq Eng 2012;43:300–8. s10064-019-01626-8.
[26] Verma AK, Singh TN. Intelligent systems for ground vibration measurement: a [54] Armaghani DJ, Mohamad ET, Narayanasamy MS, Narita N, Yagiz S. Development
comparative study. Eng Comput 2011;27:225–33. of hybrid intelligent models for predicting TBM penetration rate in hard rock
[27] Shirani Faradonbeh R, Jahed Armaghani D, Abd Majid MZ, Md Tahir M, Ramesh condition. Tunn Undergr Space Technol 2017;63:29–43. https://doi.org/10.1016/
Murlidhar B, Monjezi M, et al. Prediction of ground vibration due to quarry j.tust.2016.12.009.
blasting based on gene expression programming: a new model for peak particle [55] Zhou J, Li C, Koopialipoor M, Jahed Armaghani D, Thai Pham B. Development of a
velocity prediction. Int J Environ Sci Technol 2016;13. https://doi.org/10.1007/ new methodology for estimating the amount of PPV in surface mines based on
s13762-016-0979-2. prediction and probabilistic models (GEP-MC). Int J Min Reclamat Environ 2020.
[28] Hosseini SA, Tavana A, Abdolahi SM, Darvishmaslak S. Prediction of blast-induced https://doi.org/10.1080/17480930.2020.1734151.
ground vibrations in quarry sites: a comparison of GP, RSM and MARS. Soil Dynam [56] Jahed Armaghani D, Asteris PG, Askarian B, Hasanipanah M, Tarinejad R, Huynh V
Earthq Eng 2019;119. https://doi.org/10.1016/j.soildyn.2019.01.011. Van. Examining hybrid and single SVM models with different kernels to predict
[29] Monjezi M, Baghestani M, Shirani Faradonbeh R, Pourghasemi Saghand M, Jahed rock brittleness. Sustainability 2020;12:2229.
Armaghani D. Modification and prediction of blast-induced ground vibrations [57] Armaghani DJ, Asteris PG, Fatemi SA, Hasanipanah M, Tarinejad R, Rashid ASA,
based on both empirical and computational techniques. Eng Comput 2016;32. et al. On the use of neuro-swarm system to forecast the pile settlement. Appl Sci
https://doi.org/10.1007/s00366-016-0448-z. 2020;10:1904.
[30] Mohamed MT. Performance of fuzzy logic and artificial neural network in [58] Zhou W, Shi X, Lu X, Qi C, Luan B, Liu F. The mechanical and microstructural
prediction of ground and air vibrations. Int J Rock Mech Min Sci 2011;48:845. properties of refuse mudstone-GGBS-red mud based geopolymer composites made
[31] Qi C. Big data management in the mining industry. Int J Miner Metall Mater 2020; with sand. Construct Build Mater 2020;253. 119193.
27:131–9. [59] Qi C, Fourie A. Cemented paste backfill for mineral tailings management: review
[32] Qi C, Fourie A, Chen Q. Neural network and particle swarm optimization for and future perspectives. Miner Eng 2019;144. 106025.
predicting the unconfined compressive strength of cemented paste backfill. [60] Qi C, Fourie A, Chen Q, Zhang Q. A strength prediction model using artificial
Construct Build Mater 2018;159:473–8. intelligence for recycling waste tailings as cemented paste backfill. J Clean Prod
[33] Xu H, Zhou J, Asteris P G, Jahed Armaghani D, Tahir MM. Supervised machine 2018;183:566–78.
learning techniques to the prediction of tunnel boring machine penetration rate. [61] Qi C, Fourie A, Chen Q, Tang X, Zhang Q, Gao R. Data-driven modelling of the
Appl Sci 2019;9:3715. flocculation process on mineral processing tailings treatment. J Clean Prod 2018;
[34] Asteris PG, Ashrafian A, Rezaie-Balf M. Prediction of the compressive strength of 196:505–16.
self-compacting concrete using surrogate models. Comput Concr 2019;24:137–50. [62] Qi C, Fourie A, Ma G, Tang X. A hybrid method for improved stability prediction in
[35] Hajihassani M, Abdullah SS, Asteris PG, Armaghani DJ. A gene expression construction projects: a case study of stope hangingwall stability. Appl Soft Comput
programming model for predicting tunnel convergence. Appl Sci 2019;9:4650. 2018;71:649–58.
[36] Asteris PG, Moropoulou A, Skentou AD, Apostolopoulou M, Mohebkhah A, [63] Khandelwal M, Singh TN. Prediction of blast induced ground vibrations and
Cavaleri L, et al. Stochastic vulnerability assessment of masonry structures: frequency in opencast mine: a neural network approach. J Sound Vib 2006;289:
concepts, modeling and restoration aspects. Appl Sci 2019;9:243. 711–25.
[37] Armaghani DJ, Hatzigeorgiou GD, Karamani C, Skentou A, Zoumpoulaki I, [64] Fişne A, Kuzu C, Hüdaverdi T. Prediction of environmental impacts of quarry
Asteris PG. Soft computing-based techniques for concrete beams shear strength. blasting operation using fuzzy logic. Environ Monit Assess 2011;174:461–70.
Procedia Struct Integr 2019;17:924–33. [65] Iphar M, Yavuz M, Ak H. Prediction of ground vibrations resulting from the
[38] Asteris PG, Tsaris AK, Cavaleri L, Repapis CC, Papalou A, Di Trapani F, et al. blasting operations in an open-pit mine by adaptive neuro-fuzzy inference system.
Prediction of the fundamental period of infilled RC frame structures using artificial Environ Geol 2008;56:97–107.
neural networks. Comput Intell Neurosci 2016;2016:20. [66] Ghasemi E, Ataei M, Hashemolhosseini H. Development of a fuzzy model for
[39] Asteris PG, Kolovos KG. Self-compacting concrete strength prediction using predicting ground vibration caused by rock blasting in surface mining. J Vib Contr
surrogate models. Neural Comput Appl 2019;31:409–24. 2013;19:755–70.
[40] Chen H, Asteris PG, Jahed Armaghani D, Gordan B, Pham BT. Assessing dynamic [67] Li DT, Yan JL, Zhang L. Prediction of blast-induced ground vibration using support
conditions of the retaining wall: developing two hybrid intelligent models. Appl Sci vector machine by tunnel excavation. Appl Mech Mater 2012;170:1414–8. Trans
2019;9:1042. Tech Publ.
[41] Asteris PG, Apostolopoulou M, Skentou AD, Moropoulou A. Application of artificial [68] Khandelwal M, Kankar PK. Prediction of blast-induced air overpressure using
neural networks for the prediction of the compressive strength of cement-based support vector machine. Arab J Geosci 2011;4:427–33.
mortars. Comput Concr 2019;24:329–45. [69] Hajihassani M, Jahed Armaghani D, Monjezi M, Mohamad ET, Marto A. Blast-
[42] Zhou J, Li E, Yang S, Wang M, Shi X, Yao S, et al. Slope stability prediction for induced air and ground vibration prediction: a particle swarm optimization-based
circular mode failure using gradient boosting machine approach based on an artificial neural network approach. Environ Earth Sci 2015;74:2799–817. https://
updated database of case histories. Saf Sci 2019;118:505–18. doi.org/10.1007/s12665-015-4274-1.
[43] Jian Z, Shi X, Huang R, Qiu X, Chong C. Feasibility of stochastic gradient boosting [70] Hasanipanah M, Faradonbeh RS, Amnieh HB, Armaghani DJ, Monjezi M.
approach for predicting rockburst damage in burst-prone mines. Trans Nonferrous Forecasting blast-induced ground vibration developing a CART model. Eng Comput
Met Soc China 2016;26:1938–45. 2016;1–10.
[44] Zhou J, Li X, Mitri HS. Classification of rockburst in underground projects: [71] Samareh H, Khoshrou SH, Shahriar K, Ebadzadeh MM, Eslami M. Optimization of a
comparison of ten supervised learning methods. J Comput Civ Eng 2016;30. nonlinear model for predicting the ground vibration using the combinational
4016003. particle swarm optimization-genetic algorithm. J Afr Earth Sci 2017;133:36–45.
[45] Wang M, Shi X, Zhou J, Qiu X. Multi-planar detection optimization algorithm for [72] Fang Q, Nguyen H, Bui X-N, Nguyen-Thoi T. Prediction of blast-induced ground
the interval charging structure of large-diameter longhole blasting design based on vibration in open-pit mines using a new technique based on imperialist competitive
rock fragmentation aspects. Eng Optim 2018;50:2177–91. algorithm and M5Rules. Nat Resour Res 2019. https://doi.org/10.1007/s11053-
[46] Shi X, Jian Z, Wu B, Huang D, Wei WEI. Support vector machines approach to mean 019-09577-3.
particle size of rock fragmentation due to bench blasting prediction. Trans [73] Nguyen H, Bui X-N, Bui H-B, Cuong DT. Developing an XGBoost model to predict
Nonferrous Met Soc China 2012;22:432–41. blast-induced peak particle velocity in an open-pit mine: a case study. Acta
[47] Zhou J, Guo H, Koopialipoor M, Armaghani DJ, Tahir MM. Investigating the Geophys 2019;67:477–90.
effective parameters on the risk levels of rockburst phenomena by developing a [74] Nguyen H, Bui X-N, Tran Q-H, Moayedi H. Predicting blast-induced peak particle
hybrid heuristic algorithm. Eng Comput 2020. https://doi.org/10.1007/s00366- velocity using BGAMs, ANN and SVM: a case study at the Nui Beo open-pit coal
019-00908-9. mine in Vietnam. Environ Earth Sci 2019;78:479.
[48] Mahdiyar A, Jahed Armaghani D, Koopialipoor M, Hedayat A, Abdullah A, [75] Zhou J, Shi X, Du K, Qiu X, Li X, Mitri HS. Feasibility of random-forest approach for
Yahya K. Practical risk assessment of ground vibrations resulting from blasting, prediction of ground settlements induced by the construction of a shield-driven
using gene expression programming and Monte Carlo simulation techniques. Appl tunnel. Int J GeoMech 2016;17. 4016129.
Sci 2020;10:472. [76] Zhou J, Li E, Wei H, Li C, Qiao Q, Armaghani DJ. Random forests and cubist
[49] Chen W, Sarir P, Bui X-N, Nguyen H, Tahir MM, Armaghani DJ. Neuro-genetic, algorithms for predicting shear strengths of rockfill materials. Appl Sci 2019;9:
neuro-imperialism and genetic programing models in predicting ultimate bearing 1621.
capacity of pile. Eng Comput 2019. https://doi.org/10.1007/s00366-019-00752-x. [77] Sammut C, Webb GI. Encyclopedia of machine learning. Springer Science &
[50] Chen W, Hasanipanah M, Rad HN, Armaghani DJ, Tahir MM. A new design of Business Media; 2011.
evolutionary hybrid optimization of SVR model in predicting the blast-induced

11
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

J. Zhou et al. Soil Dynamics and Earthquake Engineering 139 (2020) 106390

[78] Brown G. Ensemble learningvols. 312–20. Encyclopedia of Machine Learning; [82] Monjezi M, Hasanipanah M, Khandelwal M. Evaluation and prediction of blast-
2010. induced ground vibration at Shur River Dam, Iran, by artificial neural network.
[79] Friedman N, Geiger D, Goldszmidt M. Bayesian network classifiers. Mach Learn Neural Comput Appl 2013;22:1637–43.
1997;29:131–63. [83] Dindarloo SR. Peak particle velocity prediction using support vector machines: a
[80] Duvall WI, Petkof B. Spherical propagation of explosion-generated strain pulses in surface blasting case study. J South African Inst Min Metall 2015;115:637–43.
rock. US Department of the Interior, Bureau of Mines; 1959. [84] Zhou J, Li C, Koopialipoor M, Jahed Armaghani D, Thai Pham B. Development of a
[81] Ghosh A, Daemen JJK. A simple new blast vibration predictor (based on wave new methodology for estimating the amount of PPV in surface mines based on
propagation laws). In: 24th US symp. rock mech. American Rock Mechanics prediction and probabilistic models (GEP-MC). Int J Min Reclamat Environ 2020.
Association; 1983. https://doi.org/10.1080/17480930.2020.1734151.

12

You might also like