Applied Soft Computing 127 (2022) 109343

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Prediction of blasting induced air-overpressure using a radial basis

function network with an additional hidden layer
∗
Ruixuan Zhang a , Yuefeng Li b , Yilin Gui a,c , , Jian Zhou d
a
School of Civil and Environmental Engineering, Queensland University of Technology, Gardens Point, QLD, 4000, Australia
b
School of Computer Science, Queensland University of Technology, Gardens Point, QLD, 4000, Australia
c
Centre for Materials Science, Queensland University of Technology, Gardens Point, QLD, 4000, Australia
d
School of Resources and Safety Engineering, Central South University, Changsha, 410083, China

article info a b s t r a c t

Article history: Blasting operations are the most conventional and frequently used rock breakage approach in the
Received 8 January 2022 field of Civil and Mining Engineering. However, the side effects induced by blasting may cause severe
Received in revised form 30 June 2022 damages to surrounding areas. Air-overpressure (AOp) is one of the side effects induced by blasting
Accepted 13 July 2022
operations, which is defined as the air pressure wave generated by blasting operation that exceeds
Available online 25 July 2022
normal atmospheric pressure. It can result in potential structural damage and glass breaking and
Keywords: therefore needs to be well predicted and subsequently minimized. In this study, 76 sets of blasting
Air-overpressure data were collected to develop a predictive model to estimate AOp value. However, due to the small
Rock blasting size of dataset, it is hard to determine the complexity of the model. Therefore, for the purpose of
Machine learning developing a machine learning model with appropriate complexity, a radial basis function network
Radial basis function network with an additional second hidden layer (RBF-2) is proposed, which is trained by incremental design
principle and modified Levenberg–Marquardt algorithm. The performance of the proposed RBF-2 is
compared with those of five other machine learning techniques, i.e., multilayer perceptron (MLP),
RBF, MLP optimized by genetic algorithm (GA-MLP), multi adaptive regression spline (MARS) and
random forest (RF). The results demonstrate that the proposed RBF-2 network outperforms other
models with RMSE of 2.02/1.98, MAPE of 1.32%/1.40%, and R of 0.9828/0.9735 in training/testing stage.
Findings revealed that the proposed RBF-2 network emerged as the most efficient, powerful and robust
technique in predicting blast induced AOp compared with other machine learning models.
© 2022 Elsevier B.V. All rights reserved.

1. Introduction equations have two disadvantages according to Singh [6]: one is

that the empirical equations can only be used in a specific site and
Drilling and blasting techniques have been widely employed in cannot be applied to other sites, and the other one is that apart
the excavations of railway, highway, mining and tunnel construc- from the two blasting design parameters, i.e., maximum charge
tion, for its advantages of low cost and high flexibility. The con- per delay and distance from blasting location, other effective
sequences of blasting can be categorized into desirable rock frag- parameters are not included, which means that the accuracy
mentation and undesirable side effects. The typical undesirable of the empirical equations are limited. With the development
side effects include ground vibration, flyrock, and air overpres- of the artificial intelligence, machine learning techniques have
sure (AOp). These side effects can result in structural damages, been widely explored in this field of study and demonstrated to
injuries, and property damages [1,2]. Thus, one main research have the superiority over the traditional empirical methods. To
topic in the field of drilling and blasting is how to predict and facilitate the readability, some recent machine learning studies
minimize the undesirable side effects. in predicting blast induced AOp are reviewed here. Three ar-
Many studies have been carried out by researchers to study tificial neural network (ANN) models were developed by [7–9]
the relationship between blast induced AOp and its influencing using 56, 110 and 38 sets of data respectively. The results of
factors. Traditional studies use empirical equations to predict these studies demonstrated that ANN is superior over traditional
blasting induced side effects [2–5]. However, these empirical empirical equations in predicting blast induced AOp. In these
models, only the blast design parameters are selected as input of
∗ Corresponding author at: School of Civil and Environmental Engineering, models. To further improve the accuracy of model, [10] consid-
Queensland University of Technology, Gardens Point, QLD, 4000, Australia. ered the impact brought by atmosphere and introduced several
E-mail address: [email protected] (Y. Gui). atmospheric influencing factors as input variables, i.e., Azimuth,

https://doi.org/10.1016/j.asoc.2022.109343
1568-4946/© 2022 Elsevier B.V. All rights reserved.
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

air temperature, wind velocity and wind direction. [11] compared require large size of initialized population to search for any possi-
the performances of different ANN approaches including multi- ble solution, and each individual is trained during each iteration.
layer perceptron (MLP), Bayesian regularized neural network and Therefore, in this study, an attempt has been made to develop an
hybrid neural fuzzy inference system, and the results showed efficient and powerful model to predict blast induced AOp. Based
that MLP is the more suitable in predicting AOp. Other than MLP on the findings provided by [20] and [15], the performance of
approach, two novel ANN approaches, brain inspired emotional ANFIS is better than traditional ANN, and the performance of ANN
neural network (BI-ENN) and cascaded forward neural network is then better than other approaches like SVM and tree-based
(CFNN), were used by [12] and [13] to study the relationship be-
models. As a result, a network use combination of ANFIS and
tween blast induced AOp and its influencing factors. Furthermore,
radial basis function network has been proposed, which is called
other machine learning techniques were also used to develop AOp
RBF-2. Different from traditional RBF network, which has only one
predictive models. Adaptive neuro fuzzy inference system (ANFIS)
was used by [14] and [15]. Both of these studies compared the hidden layer, the proposed RBF-2 network has a second hidden
performances between developed ANFIS models and MLP model, layer. The second hidden layer added is not only connected to the
and proved that ANFIS models had better performances. [16] previous layer, but also connected to the input layer. The purpose
applied support vector machine (SVM) to estimate AOp induced of this layer is to calculate the product linear combination of input
by blasting. Predictive models based on decision trees were also vectors and the output from the previous layer, which can be
implemented by many studies. In [17], two tree-based mod- regarded as the fuzzy rule part in ANFIS. The proposed RBF-2
els were implemented, i.e., M5 tree and genetic programming network is trained by a modified Levenberg–Marquardt algorithm
(GP). [18] compared the performances of three ensemble tree- and incremental design principle. The predictive capability of
based models, Cubist, random forest (RF) and gradient boosting RBF-2 is compared with 5 widely used predictors, i.e., MLP, RBF,
machine (GBM). The results showed that Cubist algorithm had GA-MLP, MARS and RF.
better performance over other two techniques. Another widely This paper is organized as following, Section 2 described the
used boosted tree model, extreme gradient boosting (XGBoost) theoretical concept of the applied methods. In Section 3, the
is used by [19] to predict AOp induced by blasting. In order to data collection and evaluation indices used will be given. Then,
determine the most suitable technique in predicting AOp, [20]
in Sections 4 and 5, the results of predictive models developed
developed and compared several predictive models based on
are reported and a discussion based on the results obtained is
various machine learning techniques, including ANN, SVM, RF, GP,
provided. In the last section of this study, conclusions are drawn
Bayesian regression trees, boosted regression trees and k-nearest
neighbors (KNN). They proved ANN had better performance. which contain the main results, contributions, and limitations of
These are the traditional implementation of machine learning the proposed model.
models. Many other studies used hybrid training mode to im-
prove the performance of the model. Normally, machine learning
2. Methodology and data collection
models like ANN and ANFIS are trained using gradient-based al-
gorithm such as Levenberg–Marquardt algorithm. However, these
gradient-based algorithm have the problem of local minima. This 2.1. Basic of the Traditional RBF network
is because these algorithms optimize the objective function by
calculating the derivatives of objective function. Thus, the opti- Fig. 1 shows a standard three-layer RBF network. Different
mization would stop when it reaches to a local minimum where from ordinary ANN, the activate function used in RBF’s hidden
the derivatives equal to zero. Therefore, to solve this problem, layer is radial basis function. Traditional RBF approach sets the
many studies combined meta-heuristic algorithm with gradient-
weight between input layer and hidden layer as 1. Thus, when
based algorithm to further optimize the model parameters. Ge-
given a set of input, only those neurons whose center are close
netic algorithm (GA) was used by [21] and [22] to optimize the
to the input can be activated. RBF has the advantage of fast con-
performance of ANN. [23] proposed a hybrid ANFIS-Polynomial
neural network (ANFIS-PNN) and used GA to optimize the model. vergence and is able to implement approximation of any function.
Imperialist competitive algorithm (ICA) was used by [24] and
[25] to optimize the performance of ANN. In [26] and [27], par- [ When given a I dimensional sample p, represented as
xp,1 , xp,2 , . . . , xp,i , . . . , xp,I the input of the hth hidden neuron
]
ticle swarm optimization (PSO) was selected to optimize the
performance of ANN. Firefly algorithm (FA) was used by [28] can be calculated as:
to develop a hybrid FA-fuzzy system (FA-FS). In [29], the per-
formance of ANFIS model was improved by stochastic fractal Xp,h,i = xp,i ∗ ui,h (1)
search algorithm (SFS). Other than optimizing model parameters,
where up,i represents the weight between the ith input nodes and
meta-heuristic algorithm can also be used to search for optimum
the hth hidden neuron. Normally, the value of ui,h is set as 1 in
hyperparameters. GA was used by [30] and [31] to optimize the
ordinary RBF network.
hyperparameters of cubist and boosted smoothing spline (BSTSM)
The kernel function for hidden layer is given as:
respectively. The performance of SVM was improved by [32],
where PSO algorithm was used to optimize the hyperparameters Xp,h − ch 2
 
ϕh Xp,h,i = exp(−
( )
of SVM. ) (2)
These applications demonstrate that machine learning tech- σh
niques have superiority in solving problems, where the results Notice that the kernel function used here is not a standard gaus-
are influenced by various factors and the relationship between sian function. The output layer is a summation layer, where the
results and its influencing factors is unclear. The prediction of output of the model can be calculated as:
blast induced AOp is obvious of this type. According to the studies ∑
wh ϕh Xp,h,i + w0
( )
reviewed, it is obvious that the use of hybrid learning mechanism op = (3)
can improve the predictive capability of the model. However,
training of hybrid model is time consuming, especially when the where wh is the weight between output and the hth hidden
searching space is huge. This is because meta-heuristic algorithms neuron, and w0 is the bias.
2
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 1. Structure of Traditional RBF.

Fig. 2. Structure of RBF-2.

2.2. Proposed RBF-2 network hidden layer 2, the other connections between layers are set as 1
as shown in Fig. 2. Therefore, in RBF-2, the adjustable parameters
The proposed RBF-2 network consists of 4 layers, an input are the center and width of the RBF layer, and the weights and
layer, an output layer, and two hidden layers, the structure of the biases between input layer and hidden layer 2. The output of
proposed RBF-2 network is shown in Fig. 2. the proposed network can be calculated following the procedure
Different from the traditional connection of ANN, where the described below.
layers are only connected to the previous and subsequent layers, [ 1. Given a set of input ] variables, represented as Xp =
in the proposed network, the second hidden layer is connected xp,1 , xp,2 , . . . , xp,i , . . . , xp,I . The output of the first hidden layer
to both input layer, and hidden layer 1. The activate function can be calculated as:
used in first hidden layer is radial basis function. The two hidden Xp − ch 2
 
o1,h = ϕh Xp,h,i = exp(−
( )
layers must have the same number of hidden neurons. Signals ) (4)
propagated from hidden layer 1 is multiplied with that from input σh
layer. To reduce the complexity of the network in order to avoid where ϕh is the activate function for the hth hidden neuron, and
overfitting, excluding the connections between input layer and ∥·∥ represents the Euclidean distance.
3
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

2. As the second hidden layer is connected to the input layer, where P is the number of training samples, and N is the number
the output of the hidden layer 2 can be represented as: of network parameters.
I
∑ The whole Jacobian matrix is calculated and stored in the
o2,h = o1,h ∗ ( xp,i wh,i + bh ) (5) implementation of LM algorithm, which requires a lot of memory.
i=1 Thus, traditional LM algorithm is not capable for large training
where wh,i and bh are the weights and bias for hidden neuron h. dataset.
3. The output layer acts as a summation layer, and thus the The training of RBF adopts a modified version of LM Algo-
output of the network can be calculated as: rithm, which calculates the quasi-Hessian matrix rather than
Jacobian matrix. The update rule of the modified algorithm can
Xp − ch 2
H H
(   ) I
be expressed as:
∑ ∑ ∑
Op = o2,h = [exp − ∗( xp,i wh,i + bh )] (6)
σh
h=1 h=1 i=1
∆k+1 = ∆k − (Qk + µk I)−1 gk (9)
Overall, network parameters need to be optimized in the pro-
posed RBF network are the center ch and width σh of hidden where µk is the combination coefficient and I is a unit matrix. Qk
neurons in hidden layer 1, and weight wh,i and bias bh for the con- is quasi Hessian matrix, calculated by:
nection between input layer and hidden layer 2. According to the ∑
Eq. (6), it is obvious that the output of RBF-2 consists of different Q = qp qp = jTp jp (10)
linear combinations of input features. As the simplest regression
method, the linear combination is very robust. Therefore, by using gk is gradient vector calculated by:
the combination of different linear relationships between input ∑
features, the robustness of RBF-2 is strengthened which ensure gk = ηp ηp = jTp ep (11)
the RBF-2 is able to avoid potential overfitting.
where ep is the difference between the expected output and the
2.3. Training of RBF-2 network output, represented as follows:

e p = y p − op (12)
The training of the proposed RBF-2 uses an incremental design
principle. The advantage of this principle is that it can eliminate jp is the row of Jacobian matrix calculated by
the randomness brought by the initialization of the network
∂ ep ∂ ep ∂ ep ∂ ep
[ ]
parameters, where it is possible to obtain a desirable solution
in just one try. Different from traditional training approach of jp = , ,..., ,..., (13)
∂ ∆1 ∂ ∆2 ∂ ∆n ∂ ∆N
RBF, the proposed RBF-2 use a modified Levenberg–Marquardt
(LM) algorithm to train the network parameters, where the cen- ∆n denotes the adjustable variables in the RBF-2 network. In this
ter, width, connection weight and bias are updated during the study, four kinds of variables in RBF-2 network can be adjusted,
training process. i.e., center, width, input weight and output weight. As a result, jp
The basis of incremental design is that the neurons are added can be represented as:
one by one each time. In each time, two neurons in hidden layers
∂ ep ∂ ep ∂ ep ∂ ep
[ ]
1 and 2 are added together to compensate for the training sample
jp = ,..., ,..., ,..., (14)
with the largest error. The procedure of the incremental design is ∂wh ∂ ch ∂σh ∂ ui,h
described as follows:
Each partial derivative in Eqs. (14) and (15) can be calculated
1. The network is initialized with zero hidden neuron. using Eq. (4) to (6).
2. Evaluate the training dataset using the current network.
3. Find the training sample with the largest error.
4. Add two hidden neurons in hidden layer 1 and 2 respec- 2.5. Parameter update procedure control
tively.
5. Update the network parameters using the modified LM The process of updating parameters is described below:
algorithm.
6. Repeat step 2 to 5 until the desired error or the number of 1. The training dataset is evaluated using the previous net-
hidden neurons reached. work denoted as ∆k , and the root mean squared error
The pseudo code of the proposed RBF-2 training procedure is (RMSE) Ek is calculated as:
described in Fig. 3 and the update of parameters are described √
in details in the following two sections, and an example of pre- e21,1 + e22,1 + · · · + e2p,1 + · · · + e2P ,1
diction of AOp using the proposed RBF-2 network is in the third Ek = (15)
P
following section.
2. The parameters in previous network are updated using the
2.4. Update of parameters improved LM algorithm, and a new network is obtained
denoted as ∆k+1 .
In traditional LM algorithm, the update rule is:
3. The training dataset is evaluated again using the updated
∆k+1 = ∆k + (JkT Jk + µk I)−1 JkT ek (7) network ∆k+1 , and a new RMSE value Ek+1 is obtained.
4. Comparing Ek+1 with Ek , if Ek+1 > Ek , the updated network
where J is Jacobian matrix, represented as:
⎡ ∂e ∆k+1 is recovered to the previous ∆k , and the combination
1 ∂ e1 ⎤ coefficient µk is increased. If Ek+1 ≤ Ek , the updated
···
⎢ ∂ ∆1 ∂ ∆N ⎥ network ∆k+1 is kept, and µk is reduced. The adjustment
. .. .. ⎥
⎢ ..
⎢
J =⎢ . . ⎥
⎥ (8) of µ follows the LM scheme described in [33].
⎣ ∂ eP ∂e ⎦ P 5. Repeat step 1 to step 4 until the desired error or maximum
··· iteration times is reached.
∂ ∆1 ∂ ∆N
4
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 3. Pseudo code for RBF-2 training.

2.6. Implementation of RBF-2

In this section, a proposed RBF-2 network with the structure of

7-2-2-1 is first implemented using as an example to explain how
the proposed approach works. Prior to the training of network,
the preprocessing of training dataset is essential, the details of
which will be explained in the following section.
After the preprocessing of data, the network is initialized with
Fig. 4. Initial network structure.
zero hidden neuron in hidden layers. The initial network is shown
in Fig. 4.
The output of the network can be expressed as the sum of
is shown in Fig. 5, where the added neurons are shown in yellow,
input variables. After going through the training dataset, the one
and the connections can be adjusted are marked in red.
with the largest error can be found, denoted as [XA , YA ], where XA
The network is trained follow the procedure described in the
is the input vector and YA is the desired output. Thus, two neurons previous two sections. The updated network is shown in Fig. 6,
are added to the two hidden layers. The center of the added RBF where XA′ and τ1 represent the updated center and width, w1 and
unit is set as XA , and the width is set as 1. The weights and bias of b1 are the weights and bias after adjustment.
the connection between input layer and hidden layer 2 are also The training dataset is evaluated again using the network
set as 1. After adding two neurons, the structure of the network shown in Fig. 6, and another sample which has the largest error
5
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 5. RBF-2 with a structure of 7-1-1-1.

Fig. 6. Updated RBF-2 with a structure of 7-1-1-1.

is found, denoted as [XB , YB ]. Then, two new neurons are added and fuel oil (ANFO), and dynamite was used for initiation. All
to the network, the new network is shown in Fig. 7. blast-holes has a diameter of 89 mm and were stemmed by fine
Similarly, the center of the new RBF unit is set as XB and width gravels.
is set as 1, and the connection weights and bias are all set as AOp generated by blasting operations was monitored using a
1. The whole network is updated again using the modified LM VibraZEB seismograph, the values of which were recorded us-
algorithm, and the network parameters are updated, including ing linear L type microphones. Microphones used operates at a
the parameters of the first added two neurons. After training of frequency response of 2 to 250 Hz, and thus are capable for
the network, the updated network is shown in Fig. 8. Note that all measuring AOp in the frequency range where there is potential
the adjustable parameters are updated in this process, including risk for damaging human hearing and structures. The monitoring
those connected to the first neuron. point of AOp is located in front of the quarry bench, and has a
Proposed RBF network with more hidden neurons can be distance of 150–720 m from blast face. The range of recorded AOp
obtained by repeating the described procedure. In order to obtain values are given in Table 1.
the best performance of the proposed RBF-2 approach, several In the 76 blasting operations, various combinations of blasting
different structures of the network are experimented, and the design parameters were implemented and collected, including
details are given in Table 2. hole depth, maximum charge per delay, burden, spacing, stem-
ming length, powder factor, and distance from the blast-face. The
3. Case study and parameter selection details of each parameter are given in Table 1.

3.1. Data collection 3.2. Descriptive analysis

76 sets of blasting data are collected from Hulu Langat quarry In the previous section, 76 sets of AOp values along with its 7
site in Selangor state, Malaysia. This quarry lies in the south of influencing factors were collected in Hulu Langat quarry site. The
Selangor, which is located in a latitude of 3◦ 7′ 0′′ N and a longitude quartiles, probability densities and outliers of each influencing
of 101◦ 49′ 1′′ E. Blasting operations are executed at a frequency factor and AOp are shown in Fig. 9 using violin plots, in which
of 8 to 10 times a month depending on the weather. Weather the data structure is displayed by the combination of density
conditions are classified into grade III to V using the method estimates and box plots. The top and bottom of the plot depicts
proposed by [34], and the rock mass strength of the quarry is the maximum and minimum values of data. The thick line in the
between 50 to 70 MPa. In blasting operations carried out in this middle of the box represents the median value of the data, and
quarry, main explosive material used were ammonium nitrate the upper and lower boundaries of which represent the first and
6
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 7. RBF-2 with a structure of 7-2-2-1.

Fig. 8. Updated RBF-2 with a structure of 7-2-2-1.

third quartile. The two lines at the top and bottom of the box for AOp as an example, the minimum and maximum values of
are outlier truncation lines, and the points above the top line which are approximately 80 and 140 respectively, and the median
or under the bottom line are regarded as potential outliers. The of which is about 110 dB. The probability density of AOp is
color shaded area around the box depicts the probability density close to a Gaussian distribution. It can be seen clearly that the
of data. Therefore, the violin plot shown in Fig. 9 provides the dataset has a reasonable distribution where all the distributions
range, density and outliers of the relevant influencing factors of variables are close to a Gaussian distribution and there is no
used to develop AOp prediction models. Taking the violin plot obvious outlier in the dataset. Therefore, no outlier detection or
7
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 9. Violin plot of input and output variables.

Table 1
Description of dataset.
Symbol Unit Category Min Max Mean Standard deviation
Hole depth HD m Input 5.2 23.2 13.4871 3.7399
Stemming St m Input 2.0 3.5 2.5743 0.3737
Burden B m Input 1.4 3.2 2.3757 0.4592
Spacing S m Input 2.5 4.1 3.2993 0.4510
Maximum charge per delay MC kg Input 50.8 82.7 71.3700 9.4738
Powder factor PF kg/m3 Input 0.5 0.9 0.6697 0.1156
Distance D m Input 150.0 710.0 390.3571 144.8953
AOp AOp dB Output 89.2 134.1 109.7786 16.0834

transformation of variable is needed. In addition, the correlation Table 2

Comprehensive results of RBF-2 with different structure.
matrix plot in Fig. 10 demonstrates the relationship between AOp
Structure Training performance
and its influencing factors, where the correlation coefficients be-
RMSE Score MAPE Score R Score Total score Rank
tween different variables are given. It is can be seen clearly from
the matrix that each influencing factor collected has a relative 7-1-1-1 3.65 2 2.76% 1 0.9423 1 4 5
7-2-2-1 3.35 3 2.55% 2 0.9515 2 7 3
relationship with AOp. In addition, it is also obvious that the 7-3-3-1 2.56 4 1.82% 4 0.9724 4 12 2
AOp is highly affected by D and MC. As demonstrated in Fig. 9 7-4-4-1 2.02 5 1.32% 5 0.9828 5 15 1
and Table 1, different influencing factors have different valuing 7-5-5-1 3.68 1 2.21% 3 0.9675 3 7 3

range, which can have impact on prediction models. Therefore, Test performance

all variables were normalized into [0, 1]. The normalization of RMSE Score MAPE Score R Score Total score Rank
variables follows the equation given below: 7-1-1-1 2.99 3 2.35% 2 0.9159 2 7 5
7-2-2-1 2.36 4 1.81% 3 0.9544 4 11 2
xi − xmin 7-3-3-1 3.01 2 1.60% 4 0.9290 3 9 3
xi = (16) 7-4-4-1 1.98 5 1.40% 5 0.9735 5 15 1
xmax − xmin
7-5-5-1 4.00 1 2.85% 1 0.9070 1 3 4
In this study, Hole depth (HD), Stemming (St), Burden (B), Spacing
(S), Maximum charge per delay (MC), Powder factor (PF), Distance
(D) and measure AOp data were selected to develop the RBF-2
model to predict blasting induced AOp. The pairwise relationship 3.3. Evaluation of model
between different variables are shown in Fig. 10, which proves
Three indices root mean squared error (RMSE), mean abso-
that each of the 7 influencing factors affect AOp individually.
lute percentage error (MAPE) and correlation coefficient (R) are
Therefore, to estimate AOp, 7 blast design parameters, i.e., HD,
selected to evaluate the performance of the models. These evalu-
St, B, S, MC, PF, D were used as input variables of RBF-2 network, ation indices are used to quantitatively describe the relationship
and the collected AOp data were adopted as output of the model. between the predicted and measured values, and are used by
The whole dataset was split into training and testing dataset, with many machine learning related studies for evaluation purposes
a proportion of 80% and 20% respectively. [7,23,35–39]. The formulas used to calculate these indices are
8
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 10. Correlation matrix for input and output variables.

given as follows: Table 3

√ Comprehensive results of RBF-2 and other models.
∑n Structure Training performance
1 (ỹi − yi )2
RMSE = (17) RMSE Score MAPE Score R Score Total score Rank
n
⏐n ⏐ RBF-2 2.02 6 1.32% 6 0.9828 5 17 2
1 ∑ ⏐ ỹi − yi ⏐ MLP 4.16 1 2.30% 3 0.9248 1 5 7
MAPE = ⏐ yi ⏐ ∗ 100%
⏐ ⏐ (18) RBF 2.71 4 2.10% 4 0.9697 4 12 4
n
1 GA-MLP 0.83 7 0.39% 7 0.9971 7 21 1
∑n MARS 3.84 3 2.69% 1 0.9316 2 6 6
i (yi − y(actual) )(ỹi − y(predict) ) RF 2.10 5 1.56% 5 0.9846 6 16 3
R = ∑n ∑n (19)
2 2
i (yi − y(actual) ) i (ỹi − y(predict) )
SVM 3.95 2 2.66% 2 0.9328 3 7 5
Test performance
where yi represents the measure AOp, ỹi is the predicted AOp,
y(actual) and y(predict) means the average of the measured and pre- RMSE Score MAPE Score R Score Total score Rank

dicted AOp values. When the value of RMSE equals to 0, MAPE = RBF-2 1.98 7 1.40% 7 0.9735 7 21 1
MLP 3.83 2 2.80% 2 0.8893 3 7 6
0, and R2 = 1 means the model can provide a perfect prediction.
RBF 3.69 3 2.34% 4 0.8988 4 11 4
GA-MLP 6.86 1 4.58% 1 0.6599 1 3 7
4. Predictive model results MARS 2.88 6 2.18% 6 0.9223 6 18 2
RF 3.45 4 2.69% 3 0.8875 2 9 5
In this section, the results obtained by the proposed RBF-2 SVM 3.19 5 2.31% 5 0.9065 5 15 3
model are described. For the purpose of testing the reliability of
the proposed RBF-2 network, in addition to a single RBF network,
four other machine learning models MLP, MARS, RF as well as been mentioned that many studies use meta-heuristic techniques
a hybrid machine learning model GA-MLP were implemented as like GA and PSO to optimize the network parameters of ANN.
comparison. Prior to the training of models, the whole dataset Therefore, a hybrid GA-MLP model was also implemented to test
with 76 sets of data is divided into training dataset which consists the reliability of the proposed model. The developed GA-MLP
of 61 sets of data (80%), and the remaining 15 sets of data (20%) model is a classical hybrid model used by many studies [21,22,
are used as testing. 40,41]. In GA, the mutation rate and crossover rate in GA was set
In order to obtain the optimum network structure of RBF-2, as 0.25 and 0.7 according to [21], and the population size and it-
several models are selected for testing with numbers of hidden eration times were 150 and 200 respectively. The MLP optimized
neurons of 1, 2, 3, 4 and 5. The results of these networks are by GA had 8 hidden neurons which was the previously obtained
summarized in Table 2. Based on the results given in Table 2, optimum structure. Apart from ANN approach, MARS,RF and SVM
the models are ranked according to the performance score of the were also implemented as comparison. The hyperparameters of
three evaluation indices. RF and SVM were tuned by Bayesian optimization. The summary
Furthermore, the bar chart given in Fig. 11 shows the over- of the performance index results and ranking results of these
all performance score of each model. According to Table 2 and models and the proposed RBF-2 model are given in Table 3 and
Fig. 11, the best overall structure of RBF-2 is 7-4-4-1, which has the plot of the predictive result of each model is given in Fig. 12,
the best performance in both training testing dataset. and the comparison of evaluation indices for models is given
Next, predictive models based on RBF and MLP were also built, Fig. 13.
the best structures of which are determined using trial and error Also, a bar chart is given in Fig. 14 to show the overall perfor-
technique. Finally, the RBF with 10 hidden neurons and MLP with mance of each model. It can be clearly seen that the proposed
8 hidden neurons were determined as the optimum models. It has RBF-2 model outperforms other models by achieving the best
9
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 11. Performance comparison for RBF-2 with different structure.

Fig. 12. Testing performances of models.

performance in testing dataset and the second best performance Seven influencing factors including HD, St, B, S, MC, PF and D were
in training dataset. To further compare and analysis the reliability used as the input variables of models. For each machine learning
of RBF-2, the testing performances of all models are evaluated technique, the model with the best predictive performance was
using a Taylor graph shown in Fig. 15. Different from scatter obtained using trial and error technique. Table 3 and Fig. 14
plot which only use two indexes to represent the precision of summarize the performance of each technique in both training
model, Taylor chart can show the accuracy of the model using and testing stage. According to the bar chart given in Fig. 14, it can
three indices, i.e., standard deviation, correlation coefficient and be clearly seen that the RBF-2 performs very well in both training
centered root mean squared error. The black point is the reference and testing dataset and there is a significant difference between
point representing the measured AOp data. the performance of the proposed RBF-2 model and other models.
The evaluation indices given in Table 3 also demonstrate the
5. Discussion reliability of the proposed RBF-2 technique in comparison with
other models, which means the proposed RBF-2 is more suitable
Seven machine learning models, i.e., RBF-2, MLP, RBF, GA-MLP, than other methods in predicting AOp induced by blasting. It
MARS, RF and SVM were developed to predict blast induced AOp. is obvious that the proposed RBF-2 method is able to provide
10
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 13. Testing performance comparison of models.

Fig. 14. Performance comparison for models.

the best overall predictive result by achieving higher value of This means that the GA-MLP model is obviously overfitting, and
R (0.9828 in training stage, 0.9735 in testing stage) and lower therefore the GA-MLP model is not reliable in predicting blast
value of RMSE (2.02 in training stage, 1.98 in testing stage) and induced AOp, where the precision of model is of importance.
MAPE (1.32% in training stage, 1.40% in testing stage). This can Therefore, it is obvious that the RBF-2 network has stronger
also be seen from the Taylor plot given in Fig. 15 which shows robustness. Besides, development of hybrid model like GA-MLP is
the comparison of the performances of models in testing stage very time consuming. In this study, the training of GA-MLP with
where a significant difference can be seen from the performance a population size of 150 and iteration time of 200 takes nearly
of RBF-2 and those of other models. 20 min, while that of other models only use a few seconds. This
Although the hybrid GA-MLP model achieved better result in is the situation when only 61 sets of data are used for training and
training stage (RMSE = 0.83, MAPE = 0.39%, R = 0.9971), its the network only consist of 8 hidden neurons, the running time
performance in testing stage was poor (RMSE = 6.86, MAPE = of hybrid models will be much longer when the dataset is larger
4.58%, R = 0.6599). The testing result R means the relationship and the network has more adjustable parameters. In addition, it
between the predicted result of GA-MLP and measured AOp is is nearly impossible to obtain an optimum hybrid model in one
far from linear, which can also be seen from Fig. 12. The Taylor try. Many times of training are necessary to find the suitable
plot in Fig. 15 also confirms the unreliability of GA-MLP, as the controlling parameters of meta-heuristic algorithm and hyper-
point representing GA-MLP is far away from the reference point. parameters of ANN. Therefore, the cost of obtain an optimum
11
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

Fig. 15. Taylor plot of results of models.

hybrid model is too high, and the obtained model also has the approximation ability and robustness of RBF-2 is stronger. When
risk of overfitting. As a result, compared with the proposed RBF- compared with hybrid model like GA-MLP, the RBF-2 network has
2 method, the hybrid GA-MLP approach is less robust and thus is much higher robustness and requires less training time, which
not suitable for predicting blast induced AOp. means our RBF-2 network is more practical in solving engineering
According to the performances indices given in Table 3 and the problems. Based on the above evaluation results and discussion, it
bar chart shown in Fig. 14, it can be seen clearly that MLP has the is obvious that the RBF-2 model has better prediction ability than
worst performance, with RMSE of 4.17 and 3.83, MAPE of 2.30% the previously used models, and therefore it can be introduced as
and 2.8%, R of 0.9248 and 0.8893 in training and testing stage a robust, practical, efficient and powerful method in predicting
respectively. Compared with MLP, the performances of RBF were blast induced AOp. In addition, this technique can be used before
much better in both training and testing dataset, which demon- conducting blasting operations to predict and further control the
strated that RBF network is more suitable than MLP in predicting AOp induced.
blast induced AOp. Therefore, as an upgraded version of RBF
network, the proposed RBF-2 network considerably improve the 6. Conclusion
performance of RBF network. Furthermore, in addition to MLP and
RBF, the performances of other three models, i.e., MARS, RF and In this study, an idea of developing an efficient machine learn-
SVM are close. MARS and SVM had balanced performances in both ing model with appropriate complexity is applied to the field of
training and testing dataset. As for RF, its overall scoring in Fig. 14 prediction of blast induced AOp. For this purpose, an upgraded
is slightly higher than the other two models, for its better training version of RBF network, called RBF-2 network is proposed, which
performance. Although RF can achieve a close performance as has an additional hidden layer and connections. The proposed
RBF-2 network in training dataset, the testing performance of RBF-2 network is trained by a modified Levenberg–Marquardt
which is not ideal. This is probably because the decision trees algorithm and incremental design principle. To examine the per-
in RF are slightly overfitting due to the setting of tree depth. formance of the proposed network, a dataset consisting of 76
Although the increase of tree depth is able to improve the training sets of blast data is collected from Hulu Langat granite quarry
performance of decision tree, it can lead to overfitting. Even site, which includes 7 blast design parameters (i.e., HD, St, B,
though RF uses the Bagging method to strengthen the model S, MC, PF and D). During the modeling process, 80% (61 sets)
robustness, the performance of which in testing stage is still non- of data are used as training and 20% (15 sets) of data are used
ideal. Compared with RF, the proposed RBF-2 network is able to as testing. Several RBF-2 networks with different structures are
reach higher accuracy with stronger robustness. This is because developed, and the performances of which are evaluated using
the RBF-2 network has more complicated structure which endue three evaluation indices, RMSE, MAPE and R. After evaluating the
it with stronger approximation ability. Meanwhile, RBF-2 uses results of these networks, the RBF-2 model with a structure of 7-
combination of different linear relationship of input features to 4-4-1 is selected as the optimum model. Subsequently, 6 other
strengthen its robustness. machine learning models, i.e., MLP, RBF, GA-MLP, MARS, RF and
As a result, in comparison with traditional implementation of SVM are also implemented for comparison. Finally, after evaluat-
machine learning approach, such as RBF, MLP, MARS and SVM, ing the performance of all models, it was found that the proposed
the proposed RBF-2 network can easily achieve higher accuracy. RBF-2 network with RMSE (1.88/2.49), MAPE (1.20%/1.71%) and
Compared with RF which has a close training performance, the R(0.9852/0.9455) in training and testing stages outperforms other
12
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

models. This confirmed that the RBF-2 model is superior over [10] U. Ozer, A. Karadogan, M.C. Ozyurt, et al., Modelling of blasting-induced
other models in predicting blast induced AOp. Among other six air overpressure wave propagation under atmospheric conditions by using
ANN model, Arab. J. Geosci. 13 (16) (2020) 1–11.
models, RF was found more suitable in predicting blast induced
[11] H. Nguyen, X.N. Bui, H.B. Bui, et al., A comparative study of artificial
AOp, with RMSE values of (2.10 and 3.45), MAPE values of (1.56% neural networks in predicting blast-induced air-blast overpressure at Deo
and 2.69%), and R values of (0.9846 and 0.8875). However, its Nai open-pit coal mine, Vietnam, Neural Comput. Appl. 32 (8) (2020)
testing performance proved that it cannot handle the trade-off 3939–3955.
between robustness and approximation ability as well as the [12] V.A. Temeng, Y.Y. Ziggah, C.K. Arthur, A novel artificial intelligent model for
proposed RBF-2 network. Although hybrid model like GA-MLP predicting air overpressure using brain inspired emotional neural network,
Int. J. Mining Sci. Technol. 30 (5) (2020) 683–689.
achieved relatively good result in training phase, its performance [13] J. Zeng, M. Jamei, M. Nait Amar, et al., A novel solution for simulating air
in testing stage was poor, and thus can be regarded as unreliable. overpressure resulting from blasting using an efficient cascaded forward
In addition, it needs to be noticed that the cost of developing an neural network, Eng. Comput. (2021) 1–13.
optimum hybrid model is too high for its long training period and [14] D. Jahed Armaghani, M. Hajihassani, H. Sohaei, et al., Neuro-fuzzy tech-
the risk of overfitting. Therefore, compared with other models, nique to predict air-overpressure induced by blasting, Arab. J. Geosci. 8
(12) (2015) 10937–10950.
the RBF-2 model has the advantage of higher robustness, higher
[15] M. Hasanipanah, D. Jahed Armaghani, H. Khamesi, et al., Several non-linear
efficiency and stronger predictive ability. In terms of practical models in estimating air-overpressure resulting from mine blasting, Eng.
application, the RBF-2 model can be used by designers to pre- Comput. 32 (3) (2016) 441–455.
dict and subsequently control the AOp to avoid damages to the [16] M. Khandelwal, P.K. Kankar, Prediction of blast-induced air overpressure
surroundings. using support vector machine, Arab. J. Geosci. 4 (3) (2011) 427–433.
[17] B. Ramesh Murlidhar, B. Yazdani Bejarbaneh, D. Jahed Armaghani, et al.,
The main limitation of this study is that the proposed RBF-
Application of tree-based predictive models to forecast air overpressure
2 network still suffers from local minima problem because it induced by mine blasting, Nat. Resour. Res. 30 (2) (2021) 1865–1887.
is trained by gradient based algorithm. Further studies can be [18] H. Nguyen, X.N. Bui, Q.H. Tran, et al., A comparative study of empirical
conducted with combination of optimization algorithms like Jaya and ensemble machine learning algorithms in predicting air over-pressure
algorithm and efficient training method to further improve the in open-pit coal mine, Acta Geophys. 68 (2) (2020) 325–336.
performance of RBF-2 and maintain the efficiency of the model [19] Z. He, D.J. Armaghani, M. Masoumnezhad, et al., A combination of
expert-based system and advanced decision-tree algorithms to predict air-
at the same time. overpressure resulting from quarry blasting, Nat. Resour. Res. 30 (2) (2021)
1889–1903.
CRediT authorship contribution statement [20] X.N. Bui, H. Nguyen, H.A. Le, et al., Prediction of blast-induced air over-
pressure in open-pit mine: assessment of different artificial intelligence
Ruixuan Zhang: Conceptualization, Methodology, Software, techniques, Nat. Resour. Res. 29 (2) (2020) 571–591.
[21] E. Tonnizam Mohamad, D. Jahed Armaghani, M. Hasanipanah, et al.,
Validation, Formal analysis, Writing – original draft, Visualisation.
Estimation of air-overpressure produced by blasting operation through a
Yuefeng Li: Methodology, Writing – review & editing. Yilin Gui: neuro-genetic technique, Environ. Earth Sci. 75 (2) (2016) 1–15.
Conceptualisation, Methodology, Supervision, Writing – review & [22] D. Jahed Armaghani, M. Hasanipanah, A. Mahdiyar, et al., Airblast predic-
editing, Visualisation, Project administration. Jian Zhou: Writing tion through a hybrid genetic algorithm-ANN model, Neural Comput. Appl.
– review & editing. 29 (9) (2018) 619–629.
[23] H. Harandizadeh, D.J. Armaghani, Prediction of air-overpressure induced by
blasting using an ANFIS-PNN model optimized by GA, Appl. Soft Comput.
Declaration of competing interest 99 (2021) 106904.
[24] D. Jahed Armaghani, M. Hasanipanah, E. Tonnizam Mohamad, A combi-
The authors declare that they have no known competing finan- nation of the ICA-ANN model to predict air-overpressure resulting from
cial interests or personal relationships that could have appeared blasting, Eng. Comput. 32 (1) (2016) 155–171.
to influence the work reported in this paper. [25] D. Jahed Armaghani, M. Hajihassani, A. Marto, et al., Prediction of blast-
induced air overpressure: a hybrid AI-based predictive model, Environ.
Monit. Assess. 187 (11) (2015) 1–13.
Data availability [26] M. Hajihassani, D. Jahed Armaghani, M. Monjezi, et al., Blast-induced air
and ground vibration prediction: a particle swarm optimization-based
Data will be made available on request. artificial neural network approach, Environ. Earth Sci. 74 (4) (2015)
2799–2817.
[27] A.I. Lawal, S. Kwon, G.Y. Kim, Prediction of an environmental impact of
References
tunnel blasting using ordinary artificial neural network, particle swarm and
Dragonfly optimized artificial neural networks, Appl. Acoust. 181 (2021)
[1] H. Bakhshandeh Amnieh, M. Hakimiyan Bidgoli, H. Mokhtari, et al., Appli-
108122.
cation of simulated annealing for optimization of blasting costs due to air
[28] J. Zhou, A. Nekouie, C.A. Arslan, et al., Novel approach for forecasting the
overpressure constraints in open-pit mines, J. Min. Environ. 10 (4) (2019)
blast-induced AOp using a hybrid fuzzy system and firefly algorithm, Eng.
903–916.
Comput. 36 (2) (2020) 703–712.
[2] D.E. Siskind, Structure Response and Damage Produced by Ground Vibra-
[29] J. Ye, J. Dalle, R. Nezami, et al., Stochastic fractal search-tuned ANFIS model
tion from Surface Mine Blasting, US Department of the Interior, Bureau of
to predict blast-induced air overpressure, Eng. Comput. (2020) 1–15.
Mines, 1980.
[30] Q. Fang, H. Nguyen, X.N. Bui, et al., Estimation of blast-induced air
[3] W.A. Hustrulid, Blasting Principles for Open Pit Mining: General Design
overpressure in quarry mines using cubist-based genetic algorithm, Nat.
Concepts, Balkema, 1999.
[4] C. Kuzu, A. Fisne, S.G. Ercelebi, Operational and geological parameters in Resour. Res. 29 (2) (2020) 593–607.
the assessing blast induced airblast-overpressure in quarries, Appl. Acoust. [31] H. Nguyen, X.N. Bui, Soft computing models for predicting blast-induced air
70 (3) (2009) 404–411. over-pressure: A novel artificial intelligence approach, Appl. Soft Comput.
[5] Prediction of airblast-overpressure induced by blasting using a hybrid 92 (2020) 106292.
artificial neural network and particle swarm optimization. [32] M. Hasanipanah, A. Shahnazar, H. Bakhshandeh Amnieh, et al., Prediction
[6] T.N. Singh, Artificial neural network approach for prediction and control of air-overpressure caused by mine blasting using a new hybrid PSO–SVR
of ground vibrations in mines, Min. Technol. 113 (4) (2004) 251–256. model, Eng. Comput. 33 (1) (2017) 23–31.
[7] M. Khandelwal, T.N. Singh, Prediction of blast induced air overpressure in [33] M.T. Hagan, M.B. Menhaj, Training feedforward networks with the
opencast mine, Noise Vib. Worldwide 36 (2) (2005) 7–16. Marquardt algorithm, IEEE Trans. Neural Netw. 5 (6) (1994) 989–993.
[8] C. Sawmliana, P.P. Roy, R.K. Singh, et al., Blast induced air overpressure [34] R. Ulusay, J.A. Hudson, The Blue Book–The Complete ISRM Suggested
and its prediction using artificial neural network, Min. Technol. 116 (2) Methods for Rock Characterization, Testing and Monitoring: 1974–2006,
(2007) 41–48. ISRM and Turkish National Group of ISRM, Ankara, 2007.
[9] E.T. Mohamad, M. Hajihassani, D.J. Armaghani, et al., Simulation of [35] C.K. Arthur, V.A. Temeng, Y.Y. Ziggah, Novel approach to predicting blast-
blasting-induced air overpressure by means of artificial neural networks, induced ground vibration using Gaussian process regression, Eng. Comput.
Int. Rev. Model. Simul. 5 (6) (2012) 2501–2506. 36 (1) (2020) 29–42.

13
R. Zhang, Y. Li, Y. Gui et al. Applied Soft Computing 127 (2022) 109343

[36] Y. Qiu, J. Zhou, M. Khandelwal, et al., Performance evaluation of hy- [39] R. Zhang, C. Zhang, M. Yu, A similar day based short term load forecasting
brid WOA-XGBoost, GWO-XGBoost and BO-XGBoost models to predict method using wavelet transform and LSTM, IEEJ Trans. Electr. Electron.
blast-induced ground vibration, Eng. Comput. (2021) 1–18. Eng. 17 (4) (2022) 506–513.
[37] J. Zhou, Y. Qiu, M. Khandelwal, et al., Developing a hybrid model of [40] M. Khandelwal, D.J. Armaghani, Prediction of drillability of rocks with
Jaya algorithm-based extreme gradient boosting machine to estimate strength properties using a hybrid GA-ANN technique, Geotech. Geol. Eng.
blast-induced ground vibrations, Int. J. Rock Mech. Min. Sci. 145 (2021) 34 (2) (2016) 605–620.
104856. [41] Y. Azimi, S.H. Khoshrou, M. Osanloo, Prediction of blast induced ground vi-
[38] S.R. Dindarloo, Prediction of blast-induced ground vibrations via genetic bration (BIGV) of quarry mining using hybrid genetic algorithm optimized
programming, Int. J. Mining Sci. Technol. 25 (6) (2015) 1011–1015. artificial neural network, Measurement 147 (2019) 106874.

14