Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Engineering with Computers (2022) 38:1905–1920

https://doi.org/10.1007/s00366-020-01136-2

ORIGINAL ARTICLE

A new multikernel relevance vector machine based on the HPSOGWO


algorithm for predicting and controlling blast‑induced ground
vibration
Zhi Yu1 · Xiuzhi Shi1 · Jian Zhou1 · Yonggang Gou1 · Xiaofeng Huo1 · Junhui Zhang1,2 · Danial Jahed Armaghani3

Received: 3 June 2020 / Accepted: 28 July 2020 / Published online: 9 August 2020
© Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract
The relevance vector machine (RVM) is considered a robust machine learning method and its superior performance has
been confirmed through many successful engineering applications. To improve the performance of the RVM model, three
single kernel functions, and three multikernel functions, including two newly proposed multikernel functions, tenfold cross-
validation, and the hybrid particle swarm optimization with grey wolf optimizer (HPSOGWO) algorithm were combined to
develop an artificial intelligence (AI) model framework. Afterwards, a new application of the RVM method was used and
introduced for two different datasets of the blast-induced ground vibration. In addition, an artificial neural network (ANN)
model and seven empirical equations were also developed for comparison purposes, and their prediction performances were
evaluated considering three performance metrics, i.e., root mean square error (RMSE), correlation coefficient (R2), and
mean absolute error (MAE). The obtained results showed that the multikernel RVM model can provide better performance
capacity than the single-kernel RVM model. As a result, the AI models were found to be more applicable than the empirical
equations in estimating blast-induced ground vibration. The prediction performance results of these models confirmed that
the selected database has a great impact on the prediction capacity. Therefore, it is a common act to compare the performance
of various models based on the selected database before selecting an optimal predictive model. The proposed model in this
study provides new theoretical and practical support for the prediction of blast-induced ground vibration and can be utilized
by other researchers in similar fields.

Keywords  Relevance vector machine · Kernel function · Hybrid particle swarm optimization with grey wolf optimizer
algorithm · Blast-induced ground vibration

* Xiuzhi Shi Junhui Zhang


sxz_csu@163.com zhangjunhui@xju.edu.cn
* Danial Jahed Armaghani 1
School of Resources and Safety Engineering, Central South
danialjahedarmaghani@duytan.edu.vn
University, Changsha 410083, Hunan, China
Zhi Yu 2
Institute of Geology and Mineral Engineering, Xinjiang
yuzhi_blasting@csu.edu.cn
University, Urumqi 830046, Xinjiang, China
Jian Zhou 3
Institute of Research and Development, Duy Tan University,
csujzhou@hotmail.com
Da Nang 550000, Vietnam
Yonggang Gou
gouyonggang@csu.edu.cn
Xiaofeng Huo
huoxiaofeng@csu.edu.cn

13
Vol.:(0123456789)

1906 Engineering with Computers (2022) 38:1905–1920

1 Introduction to regress short-term wind turbine power. The historical


numerical weather and wind farm power were used as the
Artificial intelligence (AI) is a new technical science training datasets to develop the RVM model, and the future
for simulating, extending, and developing human intel- numerical weather was used to analyze the further wind
ligence. In the process of the development of AI technol- turbine power. As shown, the RVM method is indeed used
ogy, machine learning (ML) was proposed as an important in many areas. However, the RVM method still needs to be
method for implementing AI. The main task of ML meth- further developed in order to obtain higher performances
ods is to build a computer system that learns knowledge in specific applications.
from obtained data and then use that system to predict or To improve the performance of the RVM model, some
recognize problems in many areas. According to a review methods have been proposed by other researchers, and one
of the previous literature, ML methods have been used in of the typical approach is to improve the kernel function.
many areas, such as mining [1–8], mechanical engineer- For example, a multikernel relevance vector machine model
ing [9–11], civil engineering [12–21], and medical science was proposed and utilized in the early fault diagnosis of
[22, 23]. rolling bearings [30], prediction of military vehicle draw-
The relevance vector machine (RVM) is a typical ML bar pull [31], and prediction of performance degradation of
method and can be used to solve nonlinear problems. Since mechanical equipment [32]. The main contributions of these
its development, this algorithm has been widely used to papers are the introduction of a multikernel function that is
solve many engineering problems. Fang [24] combined combined with RBF and a polynomial kernel function and
the wavelet transform (WT) and RVM methods to predict a new application of the multikernel RVM method. Wang
monthly runoff. Based on an analysis of Minjiang River et al. [33] also proposed a multikernel function and com-
database, the qualified rate of the WTRVM model was bined the multikernel function with RBF and a polynomial
found to be 84.33%, which indicated good agreement kernel function. However, there are still some differences
with the actual monthly runoff value. Widodo and Yang between these two multikernel functions, and the main dif-
[25] used machine datasets collected from the condition ference is the combination type. Zhao and Wang [34] devel-
monitoring process to predict the survival probability of oped a multiple scale kernel function that is more suitable
the machine component unit. In their paper, RVM was for a high scale sample space. In this kernel function, only
selected as the intelligent system to train, validate, and the RBF kernel was used, but it can be adjusted according to
test the obtained datasets, and the accuracies of 98% and the collected database. Although some improvements have
95.9% were found in the case of simulation data and exper- been made in the RVM kernel functions, these improve-
imental data, respectively. Agrawal et al. [26] developed ments were mainly focused on the RBF and polynomial ker-
an electricity price prediction model, and two versions nel function, and there has not been not much improvement
of the RVM model, including the radial basis function in the sigmoid kernel function. Meanwhile, performance
(RBF) and polynomial kernel function, were used in the comparisons of single kernel functions and multiple kernel
paper. After training, validation, and testing on the hour functions are still rare.
locational marginal price datasets from 2012 to 2013, the In expanding the application of the RVM method, many
developed model was found to be more accurate than the application cases can be found in the published literature
support vector machine, random forest, and multilayer per- [35–37]. However, the feasibility of using the RVM model
ceptron models. Lima et al. [27] proposed the surface elec- to predict blast-induced ground vibration is unclear. In both
tromyography signal classification method that combined open-pit mines and underground mines, blasting is still the
the RVM method and fractal dimension method. Then, cheapest and most efficient method for breaking rock masses
the proposed method was evaluated by using a publicly [38, 39]. During the blasting process, part of the explosive
available database involving seven distinct types of limb energy can cause some negative effects of blasting [40–42].
motions. The evaluation results showed that the method Blast-induced ground vibration, an important and common
can achieve good classification accuracy and was better negative effect and poses a threat to the safety of the sur-
than other methods. Imani et al. [28] applied the extreme rounding buildings and the residents’ lives [43–46]. There-
learning machine and RVM to predict the daily sea level. fore, it is of great significance to control, weaken, predict,
By using the database collected from the period of January and contain the impacts of blast-induced ground vibration on
2004 to May 2011 at the Dongshi tide gauge station in Tai- buildings, slopes, groundwater, etc. in nearby areas.
wan, the model obtained a minimum RMSE of 34.73 and This paper proposed a new ML model, the RVM model
maximum R2 of 0.93, which showed that the RVM method with single- and multikernel functions optimized by the
is a good method for forecasting the daily sea level. Zhang hybrid particle swarm optimization with grey wolf opti-
et al. [29] tested the feasibility of using the RVM model mizer (HPSOGWO) algorithm, and a new application of the
RVM method was introduced for solving problems related

13
Engineering with Computers (2022) 38:1905–1920 1907

to blast-induced ground vibration. After field monitoring difference between the traditional RVM models and the opti-
and literature referencing, two blast-induced ground vibra- mized RVM model used in this paper is the search method
tion databases, including the Tonglvshan database and the that is used to find hyperparameter combinations, hence the
Buzhaoba database, were collected. Then, the RVM model traditional models were not developed in this paper. Before
was utilized to predict blast-induced ground vibration based calculation, the collected datasets will be scaled to the range
on the above two databases. Finally, the predictive perfor- of  0-1 to reduce the calculation difficulties, and then that
mance of the RVM model was evaluated using three perfor- database will be divided into training datasets and testing
mance metrics and compared with a predeveloped artificial datasets with a ratio of 80–20%. During the calculation
neural network (ANN) model and seven empirical formulas process, a tenfold cross-validation method was selected to
in the field of ground vibration. evaluate the generalization capability of the trained model,
and the prediction performance of the trained model was
verified by three performance metrics, RMSE, R2, and mean
2 Theoretical background absolute error (MAE), and compared with the ANN model
and seven empirical equations (Fig. 1).
In this article, the RVM method was selected as the predic-
tion model. Compared to the traditional RVM algorithm,
three multikernel functions were tested to improve the map- 2.1 Multikernel relevance vector machine
ping performance of the RVM model. In practice, research-
ers usually use the trial-and-error method to search for the 2.1.1 Relevance vector machine (RVM)
best hyperparameter combination for a machine learning
model; however, the search is time-consuming and it dif- An RVM is proposed as a robust ML algorithm to create a
ficult to consider the interaction between each hyperpa- connection between x and y [49–51]. The algorithm is based
rameter. Recently, some meta-heuristic algorithms, such as on the Bayesian framework, and the relationship between
PSO and GWO were proposed, and these algorithms can targets and input vectors can be written as [52]:
provide assistance in solving the above issues. Additionally,

n
( )
compared with the trial-and-error method, many successful yi = 𝜔1 K x, xi + 𝜀i + 𝜔0 . (1)
cases [15, 47, 48] have successfully verified that applying a i=1
meta-heuristic algorithm improves model performance due ( )
to its robust ability to search using hyperparameter com- In the above equation, K x, xi is the kernel function used
binations; hence, an improved meta-heuristic algorithm in the RVM model, and 𝜔0 , 𝜔1 , and 𝜀i are the respectively
(HPSOGWO) was utilized to find the optimal combination deviation, weight of the kernel function, and the additive
of the kernel function parameters in this paper. The only noise, respectively.

Single kernel START


Initial position of wolves PSO
RBF
Data preprocessing
Sigmoid Polynomial
RVM model Attacking behavior
Multi-kernel Collected datasets
RBF+Sigmoid Fitness evaluation Hunting behavior
RBF+Polynomial Polynomial+Sigmoid
Training datasets Testing datasets
Satisfy termination? No
Encircling behavior
Model 1 Fitness 1 Yes Training set X1 Testing set
Model 2 Fitness 2
Optimal hyper-parameter
Model 3 Fitness 3
Training set X2 Testing set
Model 4 Fitness 4
Model 5 Fitness 5 Optimal RVM model Training set X3 Testing set
Model 6 Fitness 6
Model 7 Fitness 7
Prediction performance evaluation Training set X4 Testing set
Model 8 Fitness 8
Model 9 Fitness 9 • • •
• • •
Model 10 Fitness 10 Empirical model Prediction performance • • •
Validating Training
Training set Xn Testing set
1 10 ANN model RMSE MAE R2
Fitness = Σ fitness (i)
10 i =1 Training set Y Testing set
10-folds Cross-validation END

Fig. 1  Proposed framework of the model in this article

13

1908 Engineering with Computers (2022) 38:1905–1920

After assuming the errors are modeled as independent enough iterations, the optimal combination of 𝛼MP and 𝜎MP 2

zero-mean Gaussian variables, the likelihood function can can be obtained and the testing datasets can be predicted by
be expressed as follows [26, 53]: using the weight parameters.
� �
� � � � N
2 −2 ‖t − 𝜙𝜔‖2
2
p t�𝜔, 𝜎 = 2𝜋𝜎 exp − . (2) 2.1.2 Kernel functions
2𝜎 2
For regression purposes, a suitable kernel function can be
Here, 𝜎 2 is the variance of the modeled error in the above
used to map the feature vectors to a higher dimensional
equation, and t is the independent distribution.
space [55–57]. After reviewing the previous literature, three
Based on Bayesian theory, the prior distribution of 𝜔 can
kinds of kernel functions, including polynomial, RBF, and
be expressed as [33]:
sigmoid can be used in the SVM model.

N
( ) Polynomial kernel function [58, 59]:
p( 𝜔|𝛼) = N 𝜔i ||0, 𝛼i−1 . (3) ( )d
i=0 K(x, y) = xT y + 1 . (12)
In the RVM process, conditional probability t of the out- ∗
RBF kernel function [55, 57]:
put vector can be calculated by using the following equation
[32, 54]. � �
‖x − y‖2
K(x, y) = exp − . (13)
( ) ( |
) 2𝜎 2

p( t |t) =

P t∗ |𝜔, 𝛼, 𝜎 2 P 𝜔, 𝛼, 𝜎 2 |t d𝜔d𝛼d𝜎 2 . (4)
|
Sigmoid kernel function [60, 61]:
Subjected to: ( )
K(x, y) = tanh axT y + 1 . (14)
( ) ( ) ( )
| |
p 𝜔, 𝛼, 𝜎 2 |t = p 𝜔|𝛼, 𝜎 2 , t p 𝛼, 𝜎 2 |t , (5) Considering that no single kernel function is suited for all
| |
engineering datasets, it is difficult for a single kernel func-
( ) ( ) tion to match the distribution characteristics of all samples,
p t∗ |𝜔, 𝛼, 𝜎 2 = N t∗ |y(x∗ ;𝜔), 𝜎 2 , (6)
and sigmoid has not been used in the multikernel function,
( ) ( ) / ( ) three multikernel functions including two newly proposed
p 𝜔|𝛼, 𝜎 2 , t = p t|𝜔, 𝜎 2 p( 𝜔|𝛼) p t|𝛼, 𝜎 2 , (7) multikernel functions are utilized in this article to improve
the mapping ability of the kernel functions. The multikernel
( ) functions utilized in this study are presented as follows:
( ) − N2 − 12 tT Ω−1 t
p t|𝛼, 𝜎 2
= (2𝜋) |Ω| exp − , (8) Polynomial + RBF [30, 31]:
2
K(x, y) = 𝜆Kpoly (x, y) + (1 − 𝜆)KRBF (x, y). (15)
( )
( ) (𝜔 − u)T (𝜔 − u)
− N+1 Polynomial + Sigmoid:
1
p 𝜔|𝛼, 𝜎 , t = (2𝜋) 2 |E| 2 exp −
2 −
.
2
(9) K(x, y) = 𝜆Kpoly (x, y) + (1 − 𝜆)KSigmoid (x, y). (16)
( )
In the above equations, some parts such as p t|𝛼, 𝜎 can-
2
RBF + Sigmoid:
not be calculated by means of integration, but these parts
can be solved using the maximum likelihood method [54]: K(x, y) = 𝜆KRBF (x, y) + (1 − 𝜆)KSigmoid (x, y). (17)
( 2
) ( )
𝛼MP , 𝜎MP = arg max p t|𝛼, 𝜎 2 . (10) After developing the above multikernel functions, these
𝛼,𝜎 2
functions will be used in the RVM model to solve the regres-
Subjected to: sion problems.

⎧ 𝛼 new = 𝛾i 2.2 HPSOGWO algorithm


⎪ � i � 𝜇i2
⎪ 𝜎 2 new = ‖t−𝜑𝜇‖2
⎨ ∑N (11) 2.2.1 Particle swarm optimization algorithm (PSO)
N− 𝜇i
⎪ ∑ i=0
⎪ 𝛾i = 1 − 𝛼i i, i.
⎩ PSO was introduced by studying the social behavior of birds
∑ [62, 63]. In this algorithm, a large number of particles are cre-
Here, i, i is the ith element on the diagonal of E. After ated and placed in the search space of an N-dimensional moot
using the above equation, α and σ2 can be calculated. After point, and each particle in a PSO model represents a potential

13
Engineering with Computers (2022) 38:1905–1920 1909

solution of a problem [64, 65]. The particle determines the ⃗ + 1) = X


X(t ⃗ p (t) − A
⃗ ⋅ D.
⃗ (20)
direction and distance of the moving particles and dynami-
cally adjusts its velocity according to its own movement and Here, A⃗ and D⃗ are the coefficient vectors; X⃗ p  , X
⃗ and
the movement of the other particles. Eventually, the swarms t are the current position vector of the prey, the current
move close to the optimum position according to the fitness position vector of the gray wolf, and the current iteration,
function [66]. respectively.
In the PSO model, the position and velocity will updated Hunting:
by Eq. (18) [67]:
{ ( ) ( ) ⃗ 𝛼 = ||C
D ⃗ X ⃗ ||,
⃗ (t) − X(t) (21)
vnew = 𝜔v + c1 p1 Pbest − x + c2 p2 gbest − x | 1 𝛼 |
(18)
xnew = x + vnew .
⃗ 𝛽 = ||C
D ⃗ X ⃗ ||,
⃗ (t) − X(t) (22)
Here, p1 and p2 are random values; c1 and c2 are positive | 2 𝛽 |
acceleration constants; ω is the inertial weight coefficient; x
and xnew are the old and new positions, respectively; v and vnew ⃗ 𝛿 = ||C
D ⃗ X ⃗ ||,
⃗ (t) − X(t) (23)
are the old and new velocity values, respectively; Pbest is the | 3 𝛿 |
best position of a selected particle; and gbest is the best position ( )
of all particles. ⃗1 = X
D ⃗𝛼 − A ⃗𝛼 ,
⃗1 D (24)

2.2.2 Grey wolf optimizer algorithm (GWO) ( )


⃗2 = X
D ⃗𝛽 − A ⃗𝛽 ,
⃗2 D (25)
GWO is a typical nature-inspired computing algorithm that is
invented by referencing the behavior of grey wolves [68, 69]. ( )
Four kinds of wolves, as shown in Fig. 2, can be analyzed, and ⃗3 = X
D ⃗𝛿 − A ⃗𝛿 ,
⃗3 D (26)
corresponding mathematical models can be created. Among
these four kinds of wolfs, alpha (α) is the leader, and beta ( )/
(β) and delta (δ) will help alpha (α) make the decision. Grey ⃗ + 1) = X
X(t ⃗1 + X
⃗2 + X
⃗ 3 3. (27)
wolves will follow the process of encircling prey, hunting prey
and attacking prey, which is similar to the behavior used in Attacking: The final position of the prey will be deter-
nature; hence, three corresponding steps were purposed in mined in this step when the criterion is reached, and the
this algorithm to search the optimal solution for engineering action of the wolves will be determined according to the
problems [64, 68, 70]. value of |A| . If |A| is less than or equal to 1, the wolves will
Encircling: attack the prey.
⃗ = ||C
D ⃗X ⃗ ||,
⃗ (t) − X(t) (19)
| P |

Fig. 2  Four parts of wolves in


the GWO algorithm (Modified
from Photophoto [71])

13

1910 Engineering with Computers (2022) 38:1905–1920

2.2.3 Combination of PSO and GWO General predictor by Davies et al. [87],

The purpose of developing a hybrid PSO and GWO algo- v = aR−b Qcmax . (33)
rithm is to improve the optimization ability of the GWO, so
Bureau of Indian Standard [88],
the exploitation process of the PSO was referenced and applied
to the exploration of the GWO by Singh and Singh [72] in ( )
Qmax b
2017. In this algorithm, Eqs. (21), (22), and (23) are modified v=a . (34)
R2∕3
as follows [72]:
CMRI by Pal Roy [89],
⃗ 𝛼 = ||C
D ⃗ X ⃗ ||,
⃗ (t) − 𝜔 ∗ X(t) (28)
| 1 𝛼 | � �−1
R
v=a+b √ . (35)
⃗ 𝛽 = ||C
D ⃗ X ⃗ ||,
⃗ (t) − 𝜔 ∗ X(t) (29) Qmax
| 2 𝛽 |
Ambraseys–Hendron [90],
⃗ 𝛿 = ||C
D ⃗ X ⃗ ||.
⃗ (t) − 𝜔 ∗ X(t) (30) ( )−b
| 3 𝛿 | R
v=a . (36)
Additionally, the velocity and position updating equations 1∕3
Qmax
are also combined to replace Eq. (27).
Cupta et al. [91],
2.3 Artificial neural network (ANN) ( )b
R cR
v=a e Qmax . (37)
ANNs are is an important branch of ML and have been used Qmax
to solve problems in many areas such as surface settlement
Here, Qmax and R are the max charge per delay and dis-
[73–75], slope instability [76, 77], blast-induced flyrock
tance from the blast block to the monitoring point, respec-
[78, 79], and material strength [80–82]. By using the ANN
tively; a, b and c are constants.
algorithm, knowledge of the training datasets can be learned
and utilized to develop a black-box model, and this model
2.5 K‑fold cross‑validation and performance
will be used to from predictions on the testing datasets. In
metrics
general, the multilayer perceptron is considered to be the
most widely applied ANN model, and the back-propagation
During the multikernel RVM model training process, the
(BP) algorithm is the most famous learning algorithm [83,
hyperparameters should be optimized to increase the gen-
84]. During the learning process, the training function is
eralization capability of the proposed prediction model.
controlled by the cost function, and it will be terminated
For this purpose, K-fold cross-validation is a powerful and
when the error of the predicted results and actual results of
widely used method for evaluating the generalization capa-
the training datasets reach a set level.
bility of a trained model [92–94]. Kohavi [95] suggested
using tenfold cross-validation to optimize the trained model.
2.4 Empirical models
In this procedure, the training datasets will be divided into
ten folds, including nine subsets for training and one sub-
After reviewing the previously published works, seven
set for validating [96, 97]. The process will be repeated 10
empirical equations were found and selected for analyzing
times, and each training dataset will be utilized in training
the blast-induced ground vibration discussed in this article.
and validating the prediction model. The overall perfor-
These equations are listed as follows:
mance of the selected hyperparameter can be calculated by
USBM by Duvall and Petkof [85],
averaging the performances of 10 iterations.
� �−b After training, the predictive power of the developed
R models should be evaluated. For this purpose, RMSE, R2,
v=a √ . (31)
Qmax and MAE were applied in this study [98–106].

Langefors–Kihlstrom [86], √ n
√1 ∑ ( ( ) )2
(√ )b RMSE = √ T xi − xi , (38)
n i=1
Qmax
v=a . (32)
R2∕3

13
Engineering with Computers (2022) 38:1905–1920 1911

� n �2 as outliers. More details regarding our Tonglvshan database


∑ � � ∑
n � �∑n
n T xi xi − T xi xi can be seen in Fig. 4.
i=1 i=1 i=1
R2 = � � �2 �� n � n �2 � , The blast-induced ground vibration database was ran-
∑ � �2
n
∑ � �
n
∑ 2 ∑ domly split with a ratio of 80–20%, and then scaled to the
n T xi − T xi n xi − xi
i=1 i=1 i=1 i=1 range of 0–1 to decrease the calculation difficulties. To
(39) develop the optimal prediction model for the Tonglvshan
open-pit mine, six RVM models with different kernel func-
1 ∑| ( )|
n
MAE = |xi − T xi |, (40) tions (three single kernel functions and three multikernel
n i=1 | |
functions) and various HPSOGWO parameters (number of
wolves of (NOW) 20, 40, 60, 80, 100, 150, 250, and 300, the
where xi is the actual peak particle velocity (PPV) value,
maximum number of iterations (MNI) of 500, a search range
T(xi) is the predicted PPV value, xi is the average actual PPV,
of [0, 1] for λ, and a search range of [0, 10] for other kernel
and n is the number of input datasets. The predictive model
function parameters) were developed, and the optimal RVM
is perfect when R2 is 1 and RMSE and MAE are 0.
model for each kernel function was selected to compare the
prediction performance. As shown in Figs. 5 and 6, model
6 using the combination of RBF and sigmoid obtained the
lowest fitness curve, which means that the model obtained
3 Results and discussion
the best generalization performance in tenfold cross-vali-
dation during the training process. Meanwhile, the gener-
3.1 Case 1: Tonglvshan open‑pit mine
alization capability of these multikernel RVM models was
better than that of the single-kernel RVM models as shown
Tonglvshan open-pit mine is a famous mine in Hubei Prov-
in Fig. 7.
ince, China (Fig. 3). Drilling and blasting is the mining
For comparison purposes, a parameter investigation of
method, and blast-induced ground vibration is a typical
the ANN model was carried out to determine the optimal
side effect and has an impact on the surrounding environ-
number of hidden nodes (NHN) in the hidden layer, and nine
ment and people’s lives. After the monitoring of a large
hidden nodes in the hidden layer were utilized to develop
number of blasting operations in this mine, a total of 137
the final ANN model. Meanwhile, seven empirical mod-
datasets were collected by using the monitoring machine
els were regressed, and the predictive performance of both
EXP3850-3. Seven input parameters, including the Proto-
the AI models and empirical equations were ranked using
dyakonov coefficient (f), the max charge per delay (Qmax),
the method provided by Zorlu et al. [107] and are listed in
total charge (Qtotal), the delay time of detonator (T), burden
Tables 1 and 2.
of the first row (B), the horizontal (DH) and vertical (DV)
According to the ranking values plotted in Fig. 7, it can
distance between blast block and monitoring station, and an
be seen that the predictive performance of the developed
output parameter, PPV, were collected to develop the intel-
AI models with an average ranking value of 61.43 is much
ligence models. After the box plot was plotted, 117 datasets
better than the predictive performance of the developed
were analyzed in this article while 20 datasets were removed
empirical equations with an average ranking value of 28.57.

Fig. 3  A view of Tonglvshan


open-pit mine (Resourced from
Google Earth)

13

1912 Engineering with Computers (2022) 38:1905–1920

Fig. 4  Tonglvshan database

0.24
Polynomial
RBF
RBF+Sigmoid
Sigmoid
0.23 Polynomial+RBF
Polynomial+Sigmoid
RBF+Sigmoid Polynomial+Sigmoid

0.22
Fitness

Polynomial+RBF

0.21
Sigmoid

0.20
RBF

0.19
Polynomial

0.180 0.185 0.190 0.195 0.200 0.205 0.210 0.215


0 100 200 300 400 500
Iterations
Fig. 6  Final fitness values of each prediction model
Fig. 5  Fitness curves of each prediction model

the plotted values of the actual and predicted PPV obtained


Meanwhile, the predictive performance of the multikernel by the single-RBF RVM model.
RVM model with an average ranking value of 65.67 is much
better than that of the single-kernel RVM model with an 3.2 Case 2: Buzhaoba open‑pit mine
average ranking value of 55.67. It is also found that the RBF
kernel function is more suitable for the Tonglvshan database The Buzhaoba open-pit mine is part of the Xiaolongtan Min-
because the single-RBF, RBF + Sigmoid, and RBF + Poly- ing Bureau and is located in Kaiyuan City, Yunnan Province,
nomial RVM models all showed good predictive perfor- China (Fig. 9). That mine is only 16 km away from Kai-
mance with the same ranking value of 78. Figure 8 shows yuan City. The mining area of the Buzhaoba Open-pit mine

13
Engineering with Computers (2022) 38:1905–1920 1913

140 is approximately 11.52 km3 and forms a sag mining area


Training RMSE Testing RMSE
Training R 2 Testing R 2 with a length of 1.8 km, a width of 1.45 km, and a depth of
120 Training MAE Testing MAE 150–180 m. Similar to the Tonglvshan open-pit mine, drill-
Mean=61.43 Mean=28.57
ing and blasting is the mining method, and EXP3850-3 was
100
Mean=55.67 Mean=65.67 also used by Jianguo Wang [108, 109] in this mine to moni-
tor the blast-induced ground vibration. Eventually, 50 data-
Ranking value

78 78 78
80
sets were collected, and 11 parameters including hole depth
66
(HD), subdrill length (SD), stemming length (SL), toe bur-
60
den (TB), spacing (S), burden (B), Qmax, Qtotal, DV, the total
47
42 41 43 43
39 39
distance between blast block and monitoring station (DT),
40
and the PPV values were recorded. More details regarding
20 18
our used Buzhaoba database can be found in Fig. 10.
7
11 Similar to the Tonglvshan open-pit mine database, the
0
Buzhaoba open-pit mine database was also split into train-
ing datasets (40 datasets) and testing datasets (10 datasets)
Polynomial Polynomial+RBF ANN Empirical 3 Empirical 6
RBF Polynomial+RBF Empirical 1 Empirical 4 Empirical 7
Sigmoid RBF+Sigmoid Empirical 2 Empirical 5
and scaled into the range from 0 to 1. Then, six RVM
models (Table 3) with various HPSOGWO parameters
Fig. 7  Performance comparison of 14 prediction models

Table 1  Optimal prediction Model Model 1 Model 2 Model 3 Model 4 Model 5 Model 6


performance of various RVM
models with different kernel Kernel function Polynomial RBF Sigmoid Polynomial + RBF Polyno- RBF + Sigmoid
functions for the dataset of mial + Sig-
case 1 moid
Optimal λ – – – 0.6197 0.4452 0.4239
Optimal width – 0.5394 – 0.3842 – 0.3858
Optimal gamma – – 0.6651 – 1.1784 0.4378
Optimal degree 0.4067 – – 0.5792 0.4782 –
Optimal NOW 20 20 20 40 60 60
Optimal MNI 500 500 500 500 500 500
Training RMSE 0.1970 0.0702 0.1959 0.0301 0.1966 0.0297
Training R2 0.4141 0.9239 0.4214 0.9865 0.4178 0.9868
Training MAE 0.1511 0.0504 0.1486 0.0221 0.1512 0.0216
Testing RMSE 0.1561 0.0754 0.1601 0.0818 0.1595 0.0866
Testing R2 0.5162 0.9179 0.4960 0.8879 0.4878 0.8766
Testing MAE 0.1219 0.0615 0.1198 0.0637 0.1213 0.0684

Table 2  Performance of different ANNs and empirical equations in predicting PPV values for the dataset of case 1
Model Model 7 Model 8 Model 9 Model 10 Model 11 Model 12 Model 13 Model 14
ANN Empirical 1 Empirical 2 Empirical 3 Empirical 4 Empirical 5 Empirical 6 Empirical 7

NHN 9 – – – – – – –
a – 7.578 0.08389 6.604 0.08389 − 0.175 27.39 0.1016
b – 1.199 1.509 1.196 0.7544 6.544 1.177 − 1.16
c – – – 0.6175 – – – 1.693
Training RMSE 0.1234 0.1953 0.2317 0.1953 0.2317 0.1960 0.1994 0.2110
Training R2 0.7816 0.4485 0.2240 0.4487 0.2240 0.4447 0.4249 0.3564
Training MAE 0.0872 0.1431 0.1792 0.1429 0.1792 0.1434 0.1469 0.1574
Testing RMSE 0.1197 0.1891 0.2286 0.1901 0.2285 0.1875 0.1807 0.2283
Testing R2 0.7973 0.4039 0.1587 0.3995 0.1587 0.4121 0.4338 0.2204
Testing MAE 0.0986 0.1480 0.1872 0.1486 0.1872 0.1490 0.1421 0.1744

13

1914 Engineering with Computers (2022) 38:1905–1920

3.5 [107], the performance of these prediction models was


Training datasets
Testing datasets
ranked and compared in Fig. 13.
3.0
Ideal fit Unlike the Tonglvshan database, two empirical models
(empirical 3 and 6) show good prediction performance, and
2.5
their prediction power with ranking values of 80 and 82 is
almost better than that of all the AI models. The reason for
Predicted PPV

2.0
this may be that the DT and PPV in the database have a good
linear relationship, as shown in Fig. 10, which agrees well
1.5
with the empirical model. Meanwhile, the predicted PPV
1.0
calculated by empirical model 6 shown in Fig. 14 agrees
well with the actual PPV values. The performance of the
0.5 AI model with an average ranking value of 45.29 is higher
than the performance of the empirical model with an average
0.0 ranking value of 44.71. In addition, the performance of the
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
multikernel RVM model with a mean ranking value of 42.67
Actual PPV was found to be better than that of the single-kernel RVM
model with a mean ranking value of 39.00, which means that
Fig. 8  Predicted PPV against actual PPV for model 2 the multikernel functions are more suitable than the single
kernel functions. Additionally, polynomial kernel function
(number of wolves of 20, 40, 60, 80, 100, 150, 200, 250, is found to be more suitable for Buzhaoba database due to
300 and maximum number of iterations = 500, a search the good predictive performance of the single-polynomial,
range of [0, 1] for λ, and search a range of [0, 10] for other polynomial + RBF, and polynomial + sigmoid models.
kernel function parameters) and different kernel functions
were developed. The optimal RVM model with each kernel 3.3 Discussion
function was obtained as shown in Table 3. The fitness
curves of these models are plotted in Fig. 11, and it is According to the obtained results shown in the above two
easy to see that the generalization capabilities of model applications, it is easy to conclude that the multikernel
4 and model 5 are No. 1 and No. 2 among these six RVM function indeed provides better generalization ability for
models (see Fig. 12), and the generalization capability of the intelligence model compared with a single kernel. How-
the multikernel RVM model is almost better than that of ever, it can also be found that the intelligence model with the
the single-kernel RVM model. best generalization ability may not be the highest-ranking
After this, a parameter investigation was carried out to value model; for example, model 4 in case 2 shows the best
find the optimal NHN in the ANN model, and four hidden generalization ability but does not have the highest ranking
nodes were selected to develop the best ANN model with value. The reason for that phenomenon may be that the com-
a hidden layer. Finally, the PPV datasets were regressed parison of generalization ability considers all the unknown
by using MATLAB, and the performance is tabulated in datasets, and the comparison of prediction performance in
Table 4. By using the method provided by Zorlu et al. this study only considers a specified dataset combination

Fig. 9  A view of Buzhaoba


open-pit mine (Resourced from
Google Earth and Jianguo Wang
[108, 109])

13
Engineering with Computers (2022) 38:1905–1920 1915

Fig. 10  Buzhaoba database

Table 3  Optimal prediction Model Model 1 Model 2 Model 3 Model 4 Model 5 Model 6


performance of various RVM
models with different kernel Kernel function Polynomial RBF Sigmoid Polynomial + RBF Polyno- RBF + Sigmoid
functions for the dataset of mial + Sig-
case 2 moid
Optimal λ – – – 0.0085 0.0664 0.7914
Optimal width – 2.7287 – 4.1300 – 2.5035
Optimal gamma – – 0.9895 – 1.9712 1.1298
Optimal degree 0.3590 – – 1.9545 1.9628 –
Optimal NOW 20 20 20 100 150 100
Optimal MNI 500 500 500 500 500 500
Training RMSE 0.0670 0.0886 0.0871 0.0767 0.0641 0.0691
Training R2 0.9332 0.8829 0.8860 0.9166 0.9430 0.9289
Training MAE 0.0538 0.0673 0.0574 0.0625 0.0532 0.0544
Testing RMSE 0.0454 0.0529 0.0872 0.0729 0.0756 0.0452
Testing R2 0.9684 0.9453 0.9079 0.9120 0.9269 0.9708
Testing MAE 0.0369 0.0451 0.0652 0.0595 0.0622 0.0383

and cannot represent the performance in all unknown data. polynomial kernel function is more suitable for the Buzha-
Meanwhile, the fitness value was calculated using a tenfold oba database. When a database contains characteristics that
cross-validation method based on training datasets, and the agree well with the characteristics of the empirical model,
total ranking value was obtained using both training datasets the performance of the empirical model may be better than
and testing datasets. that of the intelligence model. Therefore, when developing
The characteristics of the samples have a huge impact a predictive model for solving engineering problems, a com-
on model performance. For example, the RBF kernel func- parison between different models is necessary. However, it
tion is more suitable for the Tonglvshan database, and the can be said that the applicability of the AI model is stronger

13

1916 Engineering with Computers (2022) 38:1905–1920

0.16 compared with the empirical model, and the AI model can
Polynomial
RBF be applied to most of the databases in engineering fields.
0.15 Sigmoid
Polynomial+RBF
Although some works were studied in this paper, some
0.14 Polynomial+Sigmoid work is still worthy of in-depth study. For example, only
RBF+Sigmoid
three single-kernel functions and a simple combination type
0.13
of single-kernel function were analyzed, and more works can
Fitness

0.12
be investigated. Meanwhile, the impact of data characteris-
tics on model performance is also worth studying, which is
0.11 of great significance for future modeling work.
0.10

0.09 4 Conclusions
0.08
0 100 200 300 400 500
In this article, three multikernel relevance vector machine
Iterations models using three multikernel functions were developed
and compared with three single-kernel RVM models, an
Fig. 11  Fitness curves of each prediction model ANN model, and seven empirical equations based on the
Tonglvshan open-pit mine database and Buzhaoba open-pit
mine database. After analysis, the following conclusions can
be extracted:

RBF+Sigmoid
1. A new AI model framework was proposed using the
RVM model, three single kernel functions, three mul-
Polynomial+Sigmoid
tikernel functions, the HPSOGWO algorithm, and the
tenfold cross-validation method. Then that framework
was utilized to predict blast-induced ground vibration
Polynomial+RBF
based on a Tonglvshan open-pit database and a Buzha-
oba open-pit database.
Sigmoid
2. Judging from the generalization ability and average
prediction performance, the multikernel function RVM
RBF
model was found to be better than the single-kernel
RVM model.
Polynomial 3. Affected by the characteristics of the samples, different
0.080 0.085 0.090 0.095 0.100 0.105 0.110 databases are suitable for different prediction models.
Therefore, adequate comparisons of various prediction
Fig. 12  Final fitness values of each prediction model models and different kernel functions are necessary in

Table 4  Performance of different ANNs and empirical equations in predicting PPV values for the dataset of case 2
Model Model 7 Model 8 Model 9 Model 10 Model 11 Model 12 Model 13 Model 14
ANN Empirical 1 Empirical 2 Empirical 3 Empirical 4 Empirical 5 Empirical 6 Empirical 7

NHN 4 – – – – – – –
a – 64.23 0.05666 342.1 0.05666 − 0.6986 342.3 0.4106
b – 1.491 − 2.817 1.545 1.409 28.03 1.545 − 1.674
c – – – 0.515 – – – 0.7468
Training RMSE 0.0363 0.0373 0.1510 0.0173 0.1510 0.0451 0.0173 0.1054
Training R2 0.9877 0.9798 0.6701 0.9956 0.6701 0.9704 0.9956 0.8386
Training MAE 0.0283 0.0286 0.1087 0.0089 0.1088 0.0372 0.0089 0.0713
Testing RMSE 0.0360 0.0574 0.1938 0.0273 0.1937 0.0544 0.0271 0.1440
Testing R2 0.9858 0.9502 0.3950 0.9896 0.3950 0.9516 0.9896 0.6671
Testing MAE 0.0335 0.0364 0.1536 0.0135 0.1537 0.0370 0.0134 0.1035

13
Engineering with Computers (2022) 38:1905–1920 1917

140
Training RMSE Testing RMSE
References
Training R 2 Testing R 2
120 Training MAE Testing MAE 1. Zhou J, Li X, Mitri HS (2018) Evaluation method of rockburst:
state-of-the-art literature review. Tunn Undergr Space Technol
100 Mean=45.29 Mean=44.71 81:632–659. https​://doi.org/10.1016/j.tust.2018.08.029
2. Zhou J, Li C, Koopialipoor M et al (2020) Development of
a new methodology for estimating the amount of PPV in
Ranking value

Mean=39.00 Mean=42.67 80 82
80
72 surface mines based on prediction and probabilistic models
(GEP-MC). Int J Min Reclam Environ 00:1–21. https​://doi.
59
60
54 56 org/10.1080/17480​930.2020.17341​51
51
3. Zhou J, Guo H, Koopialipoor M et al (2020) Investigating the
43
40 35
effective parameters on the risk levels of rockburst phenomena
34
28 by developing a hybrid heuristic algorithm. Eng Comput. https​
://doi.org/10.1007/s0036​6-019-00908​-9
20 18
4. Zhang H, Zhou J, Jahed Armaghani D et al (2020) A combina-
9 9
tion of feature selection and random forest techniques to solve
0
Polynomial Polynomial+RBF ANN Empirical 3 Empirical 6
a problem related to blast-induced ground vibration. Appl Sci
RBF Polynomial+RBF Empirical 1 Empirical 4 Empirical 7 10:869. https​://doi.org/10.3390/app10​03086​9
Empirical 5
Sigmoid RBF+Sigmoid Empirical 2
5. Koopialipoor M, Fallah A, Armaghani DJ et al (2019) Three
hybrid intelligent models in estimating flyrock distance
Fig. 13  Performance comparison of 18 prediction models resulting from blasting. Eng Comput 35:243–256. https​://doi.
org/10.1007/s0036​6-018-0596-4
6. Koopialipoor M, Noorbakhsh A, Noroozi Ghaleini E et  al
3.5 (2019) A new approach for estimation of rock brittleness based
Training datasets on non-destructive tests. Nondestruct Test Eval 34:354–375.
Testing datasets https​://doi.org/10.1080/10589​759.2019.16232​14
3.0 7. Zhou J, Koopialipoor M, Li E, Armaghani DJ (2020) Predic-
Ideal fit
tion of rockburst risk in underground projects developing a
2.5 neuro-bee intelligent system. Bull Eng Geol Environ. https​://
doi.org/10.1007/s1006​4-020-01788​-w
Predicted PPV

2.0
8. Mahdiyar A, Jahed Armaghani D, Koopialipoor M et al (2020)
Practical risk assessment of ground vibrations resulting from
blasting, using gene expression programming and Monte
1.5 Carlo simulation techniques. Appl Sci 10:472. https​: //doi.
org/10.3390/app10​02047​2
1.0 9. Ribeiro Junior RF, de Almeida FA, Gomes GF (2020) Fault
classification in three-phase motors based on vibration signal
analysis and artificial neural networks. Neural Comput Appl.
0.5
https​://doi.org/10.1007/s0052​1-020-04868​-w
10. Maral H, Alpman E, Kavurmacıoğlu L, Camci C (2019) A
0.0 genetic algorithm based aerothermal optimization of tip carv-
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
ing for an axial turbine blade. Int J Heat Mass Transf. https​://
Actual PPV doi.org/10.1016/j.ijhea​tmass​trans​fer.2019.07.069
11. Hemmatian B, Casal J, Planas E et al (2020) Prediction of
Fig. 14  Predicted PPV vs actual PPV for model 13 (Empirical model BLEVE mechanical energy by implementation of artificial
6) neural network. J Loss Prev Process Ind 63:104021. https​://
doi.org/10.1016/j.jlp.2019.10402​1
12. Chen R, Zhang P, Wu H et  al (2019) Prediction of shield
attempts to develop an accurate predictive model for tunneling-induced ground settlement using machine learning
techniques. Front Struct Civ Eng 13:1363–1378. https​://doi.
solving an engineering problem. org/10.1007/s1170​9-019-0561-3
13. Asteris PG, Tsaris AK, Cavaleri L et al (2016) Prediction of
the fundamental period of infilled rc frame structures using
Acknowledgements  The National Natural Science Foundation Project artificial neural networks. Comput Intell Neurosci. https​://doi.
of China (Xiuzhi Shi, 51874350). The National Natural Science Foun- org/10.1155/2016/51049​07
dation Project of China (Jian Zhou, 41807259). The National Key R&D 14. Asteris PG, Nikoo M (2019) Artificial bee colony-based neural
Program of China (Xiuzhi Shi, 2017YFC0602902). The Fundamental network for the prediction of the fundamental period of infilled
Research Funds for the Central Universities of Central South University frame structures. Neural Comput Appl 31:4837–4847. https​://
(Zhi Yu, 2018zzts217). The Innovation-Driven Project of Central South doi.org/10.1007/s0052​1-018-03965​-1
University (Jian Zhou, 2020CX040). 15. Duan J, Asteris PG, Nguyen H et al (2020) A novel artifi-
cial intelligence technique to predict compressive strength of
recycled aggregate concrete using ICA-XGBoost model. Eng
Compliance with ethical standards  Comput. https​://doi.org/10.1007/s0036​6-020-01003​-0

Conflict of interest  The authors declare no conflict of interest.

13

1918 Engineering with Computers (2022) 38:1905–1920

16. Asteris P, Roussis P, Douvika M (2017) Feed-forward neural 33. Wang T, He Y, Shi T et al (2019) Transformer health manage-
network prediction of the mechanical properties of sandcrete ment based on self-powered RFID sensor and multiple ker-
materials. Sensors 17:1344. https​://doi.org/10.3390/s1706​1344 nel RVM. IEEE Trans Instrum Meas 68:818–828. https​://doi.
17. Apostolopoulou M, Armaghani DJ, Bakolas A et al (2019) Com- org/10.1109/TIM.2018.28518​40
pressive strength of natural hydraulic lime mortars using soft 34. Zhao W, Wang L (2019) Multiple-kernel MRVM with LBFO
computing techniques. Proc Struct Integr 17:914–923. https​:// algorithm for fault diagnosis of broken rotor bar in induc-
doi.org/10.1016/j.prost​r.2019.08.122 tion motor. IEEE Access 7:182173–182184. https ​ : //doi.
18. Asteris PG, Apostolopoulou M, Skentou AD, Moropoulou A org/10.1109/ACCES​S.2019.29586​89
(2019) Application of artificial neural networks for the prediction 35. Liu Y, Ye Y, Wang Q et al (2019) Predicting the loose zone
of the compressive strength of cement-based mortars. Comput of roadway surrounding rock usingwavelet relevance vector
Concr 24:329–345. https​://doi.org/10.12989​/cac.2019.24.4.329 machine. Appl Sci. https​://doi.org/10.3390/app91​02064​
19. Xu H, Zhou J, Asteris GP et  al (2019) Supervised machine 36. Chen Y, Zhang T, Zhao W et al (2019) Rotating machinery
learning techniques to the prediction of tunnel boring machine fault diagnosis based on improved multiscale amplitude-aware
penetration rate. Appl Sci 9:3715. https​://doi.org/10.3390/app91​ permutation entropy and multiclass relevance vector machine.
83715​ Sensors. https​://doi.org/10.3390/s1920​4542
20. Chen H, Asteris P, Jahed Armaghani D et al (2019) Assessing 37. An JY, Meng FR, You ZH et al (2016) Improving protein–pro-
dynamic conditions of the retaining wall: developing two hybrid tein interactions prediction accuracy using protein evolutionary
intelligent models. Appl Sci 9:1042. https​://doi.org/10.3390/ information and relevance vector machine model. Protein Sci
app90​61042​ 25:1825–1833. https​://doi.org/10.1002/pro.2991
21. Asteris PG, Nozhati S, Nikoo M et al (2019) Krill herd algo- 38. Gou Y, Shi X, Zhou J et al (2020) Attenuation assessment
rithm-based neural network in structural seismic reliability of blast-induced vibrations derived from an underground
evaluation. Mech Adv Mater Struct 26:1146–1153. https​://doi. mine. Int J Rock Mech Min Sci 127:104220. https​: //doi.
org/10.1080/15376​494.2018.14308​74 org/10.1016/j.ijrmm​s.2020.10422​0
22. Sharma M, Singh G, Singh R (2017) Stark assessment of life- 39. Nguyen H, Bui XN, Moayedi H (2019) A comparison of
style based human disorders using data mining based learning advanced computational models and experimental techniques
techniques. IRBM 38:305–324 in predicting blast-induced ground vibration in open-pit coal
23. Gautam R, Kaur P, Sharma M (2019) A comprehensive review mine. Acta Geophys 67:1025–1037. https​://doi.org/10.1007/
on nature inspired computing algorithms for the diagnosis of s1160​0-019-00304​-3
chronic disorders in human beings. Prog Artif Intell 8:401–424 40. Nguyen H, Bui X-N, Tran Q-H et al (2019) Evaluating and
24. Ruiming F (2019) Wavelet based relevance vector machine model predicting blast-induced ground vibration in open-cast mine
for monthly runoff prediction. Water Qual Res J 54:134–141. using ANN: a case study in Vietnam. SN Appl Sci 1:125. https​
https​://doi.org/10.2166/wcc.2018.196 ://doi.org/10.1007/s4245​2-018-0136-2
25. Widodo A, Yang BS (2011) Application of relevance vector 41. Zhang X, Nguyen H, Bui XN et al (2019) Novel soft computing
machine and survival probability to machine degradation assess- model for predicting blast-induced ground vibration in open-pit
ment. Expert Syst Appl 38:2592–2599. https:​ //doi.org/10.1016/j. mines based on particle swarm optimization and XGBoost. Nat
eswa.2010.08.049 Resour Res. https​://doi.org/10.1007/s1105​3-019-09492​-7
26. Agrawal RK, Muchahary F, Tripathi MM (2019) Ensemble of 42. Hasanipanah M, Monjezi M, Shahnazar A et al (2015) Feasibil-
relevance vector machines and boosted trees for electricity price ity of indirect determination of blast induced ground vibration
forecasting. Appl Energy 250:540–548. https:​ //doi.org/10.1016/j. based on support vector machine. Measurement 75:289–297.
apene​rgy.2019.05.062 https​://doi.org/10.1016/j.measu​remen​t.2015.07.019
27. Lima CAM, Coelho ALV, Madeo RCB, Peres SM (2016) Clas- 43. Yu Z, Shi X, Zhou J et al (2020) Effective assessment of blast-
sification of electromyography signals using relevance vector induced ground vibration using an optimized random forest
machines and fractal dimension. Neural Comput Appl 27:791– model based on a Harris Hawks optimization algorithm. Appl
804. https​://doi.org/10.1007/s0052​1-015-1953-5 Sci 10:1403. https​://doi.org/10.3390/app10​04140​3
28. Imani M, Kao HC, Lan WH, Kuo CY (2018) Daily sea level pre- 44. Nguyen H, Drebenstedt C, Bui XN, Bui DT (2019) Prediction
diction at Chiayi coast, Taiwan using extreme learning machine of blast-induced ground vibration in an open-pit mine by a
and relevance vector machine. Glob Planet Change 161:211–221. novel hybrid model based on clustering and artificial neural
https​://doi.org/10.1016/j.glopl​acha.2017.12.018 network. Nat Resour Res. https​://doi.org/10.1007/s1105​3-019-
29. Zhang J, Yan J, Wu W, Liu Y (2019) Research on short-term 09470​-z
forecasting and uncertainty of wind turbine power based on rel- 45. Khandelwal M, Kumar DL, Yellishetty M (2011) Applica-
evance vector machine. Energy Proc 158:229–236. https​://doi. tion of soft computing to predict blast-induced ground vibra-
org/10.1016/j.egypr​o.2019.01.081 tion. Eng Comput 27:117–125. https​://doi.org/10.1007/s0036​
30. Chen F, Cheng M, Tang B et al (2020) A novel optimized multi- 6-009-0157-y
kernel relevance vector machine with selected sensitive features 46. Khandelwal M, Armaghani DJ, Faradonbeh RS et al (2017) Clas-
and its application in early fault diagnosis for rolling bearings. sification and regression tree technique in estimating peak parti-
Measurement 156:107583. https:​ //doi.org/10.1016/j.measur​ emen​ cle velocity caused by blasting. Eng Comput 33:45–53. https​://
t.2020.10758​3 doi.org/10.1007/s0036​6-016-0455-0
31. Yang F, Sun W, Lin G, Zhang W (2016) Prediction of military 47. Hasanipanah M, Noorian-Bidgoli M, Jahed Armaghani D,
vehicle’s drawbar pull based on an improved relevance vector Khamesi H (2016) Feasibility of PSO-ANN model for predict-
machine and real vehicle tests. Sensors 16:1–20. https​://doi. ing surface settlement caused by tunneling. Eng Comput 32:705–
org/10.3390/s1603​0351 715. https​://doi.org/10.1007/s0036​6-016-0447-0
32. Chen F, Yang Y, Tang B et al (2020) Performance degradation 48. Naghibi SA, Ahmadi K, Daneshi A (2017) Application of sup-
prediction of mechanical equipment based on optimized multi- port vector machine, random forest, and genetic algorithm opti-
kernel relevant vector machine and fuzzy information granula- mized random forest models in groundwater potential mapping.
tion. Measurement 151:107116. https​://doi.org/10.1016/j.measu​ Water Resour Manag 31:2761–2775. https​://doi.org/10.1007/
remen​t.2019.10711​6 s1126​9-017-1660-3

13
Engineering with Computers (2022) 38:1905–1920 1919

49. Tipping ME (2001) Sparse Bayesian learning and the relevance 65. Hasanipanah M, Shahnazar A, Bakhshandeh Amnieh H, Jahed
vector machine. J Mach Learn Res 1:211–244. https​://doi. Armaghani D (2017) Prediction of air-overpressure caused by
org/10.1162/15324​43015​27482​36 mine blasting using a new hybrid PSO–SVR model. Eng Comput
50. Wang F, Gou B, Qin Y (2013) Modeling tunneling-induced 33:23–31. https​://doi.org/10.1007/s0036​6-016-0453-2
ground surface settlement development using a wavelet smooth 66. Shi Y, Eberhart RC (1999) Empirical study of particle swarm
relevance vector machine. Comput Geotech 54:125–132. https​ optimization. In: Proceedings of the 1999 congress on evolution-
://doi.org/10.1016/j.compg​eo.2013.07.004 ary computation, CEC 1999, pp 1945–1950
51. Liu P, Zhu X, Hu X et al (2019) Local tangent space alignment 67. Mohamad ET, Jahed Armaghani D, Momeni E, Abad SV (2015)
and relevance vector machine as nonlinear methods for estimat- Prediction of the unconfined compressive strength of soft rocks: a
ing sensory quality of tea using NIR spectroscopy. Vib Spectrosc PSO-based ANN approach. Bull Eng Geol Environ 74:745–757.
103:102923. https​://doi.org/10.1016/j.vibsp​ec.2019.05.005 https​://doi.org/10.1007/s1006​4-014-0638-0
52. Xia C, Huang M, Qian X et al (2019) Novel intelligent approach 68. Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer.
for peak shear strength assessment of rock joints on the basis Adv Eng Softw 69:46–61. https​://doi.org/10.1016/j.adven​gsoft​
of the relevance vector machine. Math Probl Eng. https​://doi. .2013.12.007
org/10.1155/2019/31827​36 69. Yu Z, Shi X, Zhou J et al (2020) Prediction of blast-induced rock
53. Wang P (2017) Application of Ground Settlement Prediction movement during bench blasting: use of gray wolf optimizer and
Based on EMD-RVM. In: Proceedings of the 2nd International support vector regression. Nat Resour Res 29:843–865. https​://
Conference on Mechatronics Engineering and Information doi.org/10.1007/s1105​3-019-09593​-3
Technology (ICMEIT 2017). Atlantis Press, Paris, France, pp 70. Bian XQ, Zhang L, Du ZM et al (2018) Prediction of sulfur
193–198. https​://doi.org/10.2991/icmei​t-17.2017.35 solubility in supercritical sour gases using grey wolf optimizer-
54. An JY, You ZH, Zhou Y, Wang DF (2019) Sequence-based based support vector machine. J Mol Liq 261:431–438. https​://
prediction of protein-protein interactions using gray wolf opti- doi.org/10.1016/j.molli​q.2018.04.070
mizer–based relevance vector machine. Evol Bioinform. https​:// 71. Photophotp. https​://www.photo​photo​.cn. Accessed 26 Mar 2019
doi.org/10.1177/11769​34319​84452​2 72. Singh N, Singh SB (2017) Hybrid algorithm of particle swarm
55. Ocak I, Seker SE (2013) Calculation of surface settlements optimization and grey wolf optimizer for improving convergence
caused by EPBM tunneling using artificial neural network, SVM, performance. J Appl Math. https​://doi.org/10.1155/2017/20304​
and Gaussian processes. Environ Earth Sci 70:1263–1276. https​ 89
://doi.org/10.1007/s1266​5-012-2214-x 73. Pourtaghi A, Lotfollahi-Yaghin MA (2012) Wavenet ability
56. Shreyas SK, Dey A (2019) Application of soft computing tech- assessment in comparison to ANN for predicting the maximum
niques in tunnelling and underground excavations: state of the surface settlement caused by tunneling. Tunn Undergr Space
art and future prospects. Innov Infrastruct Solut. https​://doi. Technol 28:257–271. https​://doi.org/10.1016/j.tust.2011.11.008
org/10.1007/s4106​2-019-0234-z 74. Moghaddasi MR, Noorian-Bidgoli M (2018) ICA-ANN, ANN
57. Zhang P, Wu HN, Chen RP, Chan THT (2020) Hybrid meta- and multiple regression models for prediction of surface settle-
heuristic and machine learning algorithms for tunneling-induced ment caused by tunneling. Tunn Undergr Space Technol 79:197–
settlement prediction: a comparative study. Tunn Undergr Space 209. https​://doi.org/10.1016/j.tust.2018.04.016
Technol 99:103383. https​://doi.org/10.1016/j.tust.2020.10338​3 75. Zhou J, Sh X, Du K et al (2017) Feasibility of random-forest
58. Nguyen H, Choi Y, Bui X-N, Nguyen-Thoi T (2019) Predicting approach for prediction of ground settlements induced by the
blast-induced ground vibration in open-pit mines using vibra- construction of a shield-driven tunnel. Int J Geomech 17:1–12.
tion sensors and support vector regression-based optimization https​://doi.org/10.1061/(ASCE)GM.1943-5622.00008​17
algorithms. Sensors 20:132. https​://doi.org/10.3390/s2001​0132 76. Verma AK, Singh TN, Chauhan NK, Sarkar K (2016) A hybrid
59. Zhao D, Liu H, Zheng Y et al (2019) Whale optimized mixed FEM–ANN approach for slope instability prediction. J Inst Eng
kernel function of support vector machine for colorectal cancer Ser A 97:171–180. https​://doi.org/10.1007/s4003​0-016-0168-9
diagnosis. J Biomed Inform 92:103124. https:​ //doi.org/10.1016/j. 77. Sevgen K, Nefeslioglu G (2019) A novel performance assess-
jbi.2019.10312​4 ment approach using photogrammetric techniques for landslide
60. Liu X, Yang C (2013) A kernel spectral angle mapper algorithm susceptibility mapping with logistic regression. ANN Random
for remote sensing image classification. In: Proc 2013 6th int For Sensors 19:3940. https​://doi.org/10.3390/s1918​3940
congr image signal process CISP 2013 2:814–818. https​://doi. 78. Zhou J, Aghili N, Ghaleini EN et al (2019) A Monte Carlo sim-
org/10.1109/CISP.2013.67452​77 ulation approach for effective assessment of flyrock based on
61. Herawan T, Ghazali R, Deris MM (2014) Recent advances on intelligent system of neural network. Eng Comput. https​://doi.
soft computing and data mining: proceedings of the first inter- org/10.1007/s0036​6-019-00726​-z
national conference on soft computing and data mining (SCDM- 79. Zhou J, Koopialipoor M, Murlidhar BR et al (2019) Use of intel-
2014) Universiti Tun Hussein Onn Malaysia, Johor, Malaysia ligent methods to design effective pattern parameters of mine
June, 16th-18th, 2014. Adv Intell Syst Comput 287:273–281. blasting to minimize flyrock distance. Nat Resour Res. https​://
https​://doi.org/10.1007/978-3-319-07692​-8 doi.org/10.1007/s1105​3-019-09519​-z
62. Kennedy J, Eberhart R (1995) Particle swarm optimization. In: 80. Majdi A, Rezaei M (2013) Prediction of unconfined compres-
Proceedings of the IEEE international conference on neural net- sive strength of rock surrounding a roadway using artificial
works, Perth, 27 November-01 December 1995, 4, 1942–1948. neural network. Neural Comput Appl 23:381–389. https​://doi.
https​://doi.org/10.1109/ICNN.1995.48896​8 org/10.1007/s0052​1-012-0925-2
63. Zeng N, Zhang H, Liu W et al (2017) A switching delayed PSO 81. Fang Q, Yazdani Bejarbaneh B, Vatandoust M et  al (2019)
optimized extreme learning machine for short-term load fore- Strength evaluation of granite block samples with different
casting. Neurocomputing 240:175–182. https:​ //doi.org/10.1016/j. predictive models. Eng Comput. https​://doi.org/10.1007/s0036​
neuco​m.2017.01.090 6-019-00872​-4
64. Kamboj VK (2016) A novel hybrid PSO–GWO approach for unit 82. Armaghani DJ, Tonnizam Mohamad E, Momeni E et al (2016)
commitment problem. Neural Comput Appl 27:1643–1655. https​ Prediction of the strength and elasticity modulus of granite
://doi.org/10.1007/s0052​1-015-1962-4 through an expert artificial neural network. Arab J Geosci 9:1–
16. https​://doi.org/10.1007/s1251​7-015-2057-3

13

1920 Engineering with Computers (2022) 38:1905–1920

83. Yu Z, Shi X, Zhou J et al (2019) Feasibility of the indirect deter- the strength of fiber-reinforced cemented paste backfill. Eng
mination of blast-induced rock movement based on three new Comput. https​://doi.org/10.1007/s0036​6-020-01014​-x
hybrid intelligent models. Eng Comput. https​://doi.org/10.1007/ 99. Koopialipoor M, Tootoonchi H, Jahed Armaghani D et al (2019)
s0036​6-019-00868​-0 Application of deep neural networks in predicting the penetration
84. Li C, Zhou J, Armaghani DJ, Li X (2020) Stability analysis of rate of tunnel boring machines. Bull Eng Geol Environ 78:6347–
underground mine hard rock pillars via combination of finite dif- 6360. https​://doi.org/10.1007/s1006​4-019-01538​-7
ference methods, neural networks, and Monte Carlo simulation 100. Armaghani DJ, Koopialipoor M, Marto A, Yagiz S (2019)
techniques. Underground Space. https​://doi.org/10.1016/j.undsp​ Application of several optimization techniques for estimating
.2020.05.005 TBM advance rate in granitic rocks. J Rock Mech Geotech Eng
85. Duvall WI, Petkof B (1959) Spherical propagation of explosion- 11:779–789. https​://doi.org/10.1016/j.jrmge​.2019.01.002
generated strain pulses in rock. US Department of the Interior, 101. Jahed Armaghani D, Hasanipanah M, Mahdiyar A et  al
Bureau of Mines (2018) Airblast prediction through a hybrid genetic algorithm-
86. Langefors U, Kihlström B (1963) The modern technique of rock ANN model. Neural Comput Appl 29:619–629. https​://doi.
blasting. Wiley, Hoboken org/10.1007/s0052​1-016-2598-8
87. Davies B, Farmer IW, Attewell PB (1964) Ground vibration from 102. Armaghani DJ, Hasanipanah M, Amnieh HB, Mohamad ET
shallow sub-surface blasts. Engineer 217:553–559 (2018) Feasibility of ICA in approximating ground vibration
88. Indian Standard (1973) Criteria for safety and design of struc- resulting from mine blasting. Neural Comput Appl 29:457–465.
tures subjected to underground blast. ISI Bull I:IS–6922:6 https​://doi.org/10.1007/s0052​1-016-2577-0
89. Pal Roy P (1993) Putting ground vibration predictions into prac- 103. Ly H-B, G.Asteris P, Pham BT (2020) Accuracy assessment
tice. Colliery Guard 241:63–67. https​://doi.org/10.1016/0148- of extreme learning machine in predicting soil compression
9062(93)92499​-g coefficient. Vietnam J Earth Sci 42. https​://doi.org/10.15625​
90. Ambraseys NR, Hendron AJ (1968) Dynamic behavior of rock /0866-7187/42/3/14999​
masses. In Rock Mechanics in Engineering Practice (KG Stagg 104. Armaghani DJ, Asteris PG, Fatemi SA et al (2020) On the use
& OC Zienkievicz, Eds.). London:wiley: 203–207. of neuro-swarm system to forecast the pile settlement. Appl Sci.
91. Gupta RN, Roy PP, Singh B (1988) On a blast induced blast https​://doi.org/10.3390/app10​06190​4
vibration predictor for efficient blasting. In: Proceedings of the 105. Zhou J, Qiu Y, Zhu S, Armaghani DJ, Khandelwal M, Mohamad
22nd international conference on safety in mines, pp 1015–1021 ET (2020) Estimation of the TBM advance rate under hard rock
92. Qi C, Fourie A, Ma G, Tang X (2018) A hybrid method for conditions using XGBoost and Bayesian optimization. Under-
improved stability prediction in construction projects: a case ground Space. https​://doi.org/10.1016/j.undsp​.2020.05.008
study of stope hangingwall stability. Appl Soft Comput J 71:649– 106. Gao J, Amar MN, Motahari MR et al (2020) Two novel combined
658. https​://doi.org/10.1016/j.asoc.2018.07.035 systems for predicting the peak shear strength using RBFNN and
93. Moayedi H, Osouli A, Nguyen H, Rashid ASA (2019) A novel meta-heuristic computing paradigms. Engineering with Comput-
Harris hawks’ optimization and k-fold cross-validation predict- ers. https​://doi.org/10.1007/s0036​6-020-01059​-y
ing slope stability. Eng Comput. https​://doi.org/10.1007/s0036​ 107. Zorlu K, Gokceoglu C, Ocakoglu F et al (2008) Prediction of
6-019-00828​-8 uniaxial compressive strength of sandstones using petrography-
94. Qi C, Fourie A, Chen Q, Zhang Q (2018) A strength prediction based models. Eng Geol 96:141–158. https​://doi.org/10.1016/j.
model using artificial intelligence for recycling waste tailings as engge​o.2007.10.009
cemented paste backfill. J Clean Prod 183:566–578. https​://doi. 108. Jianguo W (2011) Numerical simulation study on the stability of
org/10.1016/j.jclep​ro.2018.02.154 Buzhao Dam West under sudden blasting. Kunming University
95. Kohavi R (1995) A study of cross-validation and bootstrap for of Science and Technology
accuracy estimation and model selection. In Proc 14th Int. Joint 109. Jianguo W, Yonghui H, Jianming Z (2016) BP neural network
Conf. Artif Intell, 14:1137–1145 prediction for blasting vibration in open- pit coal mine. J Henan
96. Zhou J, Li X, Mitri HS (2016) Classification of rockburst in Polytech Univ Nat Sci 35:322–328. https​://doi.org/10.16186​/j.c
underground projects: Comparison of ten supervised learning nki.1673-9787.2016.03.006
methods. J Comput Civ Eng 30:4016003. https:​ //doi.org/10.1061/
(ASCE)CP.1943-5487.00005​53 Publisher’s Note Springer Nature remains neutral with regard to
97. Lin Y, Zhou K, Li J (2018) Prediction of slope stability using four jurisdictional claims in published maps and institutional affiliations.
supervised learning methods. IEEE Access 6:31169–31179. https​
://doi.org/10.1109/ACCES​S.2018.28437​87
98. Li E, Zhou J, Shi X et al (2020) Developing a hybrid model of
salp swarm algorithm-based support vector machine to predict

13

You might also like