Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Transportation Geotechnics 36 (2022) 100806

Contents lists available at ScienceDirect

Transportation Geotechnics
journal homepage: www.elsevier.com/locate/trgeo

COSMA-RF: New intelligent model based on chaos optimized slime mould


algorithm and random forest for estimating the peak cutting force of
conical picks
Jian Zhou a, Yong Dai b, Kun Du c, *, Manoj Khandelwal d, Chuanqi Li e, Yingui Qiu f
a
School of Resources and Safety Engineering, Central South University, Changsha 410083, China
b
School of Resources and Safety Engineering, Central South University, Changsha 410083, China
c
School of Resources and Safety Engineering, Central South University, Changsha 410083, China
d
Institute of Innovation, Science and Sustainability, Federation University Australia, Ballarat, VIC 3350, Australia
e
Laboratory 3SR, CNRS UMR 5521, Grenoble Alpes University, Grenoble 38000, France
f
School of Resources and Safety Engineering, Central South University, Changsha 410083, China

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

Keywords: Since conical pick cutting is a complex process of multi-factor coupling effects, theoretical model construction for
Conical picks cutting force prediction is a quite difficult task. In this paper, various novel intelligent models based on chaos-
Cutting force optimized slime mould algorithm (COSMA) and random forest (RF) are proposed for this task. In the proposed
Chaos optimized slime mould algorithm
COSMA-RF methods, the chaos algorithms with the ergodicity and randomness are introduced to chaotically
Random forest
determine the initial position to form a COSMA, and the SMA and COSMA are used to tune the hyperparameters
of RF and mean square error are assigned as a fitness function. Consequently, 205 data samples having seven
variables (tensile strength of the rock σt, compressive strength of the rock σ c, cone angle θ, cutting depth d, attack
angle γ, rake angle α and back-clearance angle β) and one output parameter peak cutting force (PCF) are collected
from previous literature. Additionally, the performance of optimal COSMA-RF models is comprehensively
compared with the existing theoretical formulae and four common machine algorithms, namely RF, extreme
gradient boosting, extreme learning machine and back propagation neural network. The results indicate that
Logistic map optimized SMA (LSMA), Sine map optimized SMA (SINSMA) and Sinusoidal map optimized SMA
(SSMA) have better convergence ability and accuracy compared with original SMA. LSMA-RF, SINSMA-RF and
SSMA-RF models yield better PCF prediction performance compared with theoretical formulae and common
machine algorithms. Furthermore, sensitive analysis shows σt, σ c, d and β are significantly sensitive to PCF.

Introduction force (NF), cutting force (CF) and sideway force (SF). CF is an energy-
consuming component promoting the picking movement, through
Partial-face tunnelling machine (e.g. roadheader) is becoming which the breaking torque and performance of roadheader can be
increasingly popular in civil and mining industries due to their high directly evaluated. At present, linear-cutting machines (LCM) are per­
efficiency and safety [54,9,7,8]. The design of the cutting head and the formed for CF measurement, which is considered to be one of the most
arrangement of picks are the key components to determine the perfor­ effective test methods of pick force [34]. However, LCM equipment is
mance of roadheader. Compared with the chisel pick and radial pick, the expensive, and not available in most research institutions. Therefore,
conical pick is known for its energy-saving and long service life pick many researchers have discussed the influence of rock physical and
type, so the research in this study is aimed at the conical pick [53]. mechanical properties, pick parameters and cutting parameters on the
Specific energy (SE) and the force acting on the conical pick are the most CF, and established a series of CF prediction models [10,14,40,13].
popular and effective parameters in evaluating rock-breaking efficiency Evans [10] considered that the rock eventually breaks in the form of
[50,44]. The forces acting on the pick are generally divided into normal tensile failure, and firstly established the theoretical model for conical

* Corresponding author.
E-mail addresses: j.zhou@csu.edu.cn, csujzhou@hotmail.com (J. Zhou), 205511027@csu.edu.cn (Y. Dai), dukuncsu@csu.edu.cn (K. Du), m.khandelwal@
federation.edu.au, mkhandelwal1@gmail.com (M. Khandelwal), Chuanqi.Li@univ-grenoble-alpes.fr (C. Li), 195512085@csu.edu.cn (Y. Qiu).

https://doi.org/10.1016/j.trgeo.2022.100806
Received 26 May 2022; Received in revised form 23 June 2022; Accepted 2 July 2022
Available online 8 July 2022
2214-3912/© 2022 Published by Elsevier Ltd.
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

pick CF estimation. Based on Evans’ study, a large number of modified foundation for global search and improve the optimization ability.
models have been established and further studies have been conducted Finally, a sensitivity analysis is proposed for calculating the importance
by numerous scholars. Goktan [14] and Roxborough and Liu [40] found of input parameters.
that the CF of a conical pick is partly affected by the friction angle be­
tween the pick and rock, so conical pick CF prediction models consid­ Methodology
ering the effect of friction angle between pick and rock were established.
In addition to Goktan [14], and Roxborough and Liu’s [40] consider­ One-dimensional chaotic mapping function
ations, Goktan and Gunes [13] proposed a semi-empirical model for CF
prediction, and the results demonstrate the superiority of their model. As a dynamic behavior of complex nonlinear systems, chaos theory
Bilgin et al. [2] used a conical pick to test 22 kinds of rocks and minerals has received widespread attention in the past few decades, and chaos
with strength from 10 MPa to 170 MPa on an LCM, finding that the theory has been widely used in different scientific fields such as mete­
cutting force is proportional to the compressive strength and tensile orology, economics, and chemistry [37,26,12]. Chaotic sequences are
strength of rock. Based on the data samples conducted by Bilgin et al. traversal, random, and sensitive dependent on initial conditions. In
[2], Tiryaki [45] predicted CF by considering the physical and me­ recent years, chaos theory has been integrated into various meta-heu­
chanical properties of rock such as tensile strength, compressive ristic optimization algorithms [19,22]. Past studies have shown that if
strength and density. chaotic sequences are used to replace randomly generate initial pop­
Su and Ali [42] used Particle Flow Code in 3 Dimensions (PFC3D) to ulations in the original algorithm, the diversity of the initial individuals
simulate the rock-breaking process of conical picks. The peak cutting is increased and a promising search area can be achieved [22]. For
force (PCF) obtained by numerical simulation was compared with the optimizing the SMA in this paper, six chaotic maps are introduced to
model established by Evans [10], Roxborough and Liu [40], and Goktan generate the variables. The definition and equations of chaotic maps are
[14]. The results showed that the peak cutting force obtained by nu­ determined as follows:
merical simulation is well correlated with the empirical model estab­
lished by predecessors, demonstrating the potential prediction ability of Chebyshev map
numerical simulation. Based on the severe abrasive wear of cutting Chebyshev maps, also known as Chebyshev polynomials, originated
tools, Shao et al. [41] conducted an LCM test based on Taguchi’s L25 from the expansion of cosine function and sine function of multiple
orthogonal array and built an artificial neural network (ANN) and angles [30]. The Chebyshev map of the n order is expressed as:
multiple linear regression (MLR) models for cutting force prediction. ( )
xn+1 = cos ncos− 1 xn k = 4 xn ∈ [ − 1, 1] (1)
The result indicated that the ANN model showed excellent prediction
performance and cutting depth affects cutting force most. Through
varying the cutting depth and line spacing of the LCM test, Wang et al. Sine map
[48] constructed several strong empirical models for cutting force pre­ Sine map has a relatively simple structure and fast generating chaotic
diction. With basic geometric parameters, fracture parameters of rock sequence. Due to its special range, xn+1 ∈ [-1, 1], the sine function can
and installation parameters of a conical pick as influencing factors, fix both independent variable and range in the interval [-1, 1], as shown
Wang et al. [46] proposed a conical cutting force prediction model based in Eq. (2) [15]:
on Coulomb-Mohr Criterion which is validated by the orthogonal test. a
xn+1 = sin(πxn ) 0 < a⩽4 (2)
The proposed model is potential for cutting force prediction and the 4
sensitivity analysis showed cutting depth affects the cutting force most,
but the cutting angle is the opposite. Yasar [56] reported that theoretical ICMIC map
models are difficult to predict the conical cutting force, and built semi- ICMIC map has the advantages of uniform traversal, determined by
theoretical models based on 165 test samples of 47 types of rock which Eq. (3) [31]:
were collected from previous studies. However, in the above research, ( )
a
few data and a small number of rock physical, mechanical parameters xn+1 = sin a ∈ (0, ∞) (3)
xn
and pick parameters are considered. In addition, Roxborough and Liu
[40], and Yasar [56] pointed out that PCF prediction of rock-cutting
Tent map
interaction is a complex nonlinear problem. The established models
A tent map, in mathematics, refers to a segmented linear map, named
lacked generalization ability and are often difficult to be applied in other
because its function image resembles a tent, with uniform probability
scenarios. Therefore, it is necessary to establish a conical pick cutting
density and ideal correlation characteristics [29]. Its mathematical
force prediction model with strong generalization ability considering
expression is shown in Eq. (4):
multi parameters and multi-samples.
{
The machine learning (ML) algorithm is a key component in the field 2xn xn < 0.7
of data science. The theories and methods of ML have been applied to xn+1 = (4)
2(1 − xn ) xn ⩾0.7
solve complex problems in engineering applications
[20,5,1,21,47,23,32;62,64,66,67,68]. Compared with the empirical and Logistic map
theoretical models mentioned above, ML has a better ability to deal with Logistic mapping chaotic sequence is a very simple chaotic map in
multi-dimensional relations problems and shows high accuracy in terms of mathematical form. Its distribution is characterized by a rela­
regression prediction. When the data increases, the training data in the tively uniform probability of intermediate values and dense values at
ML model can be adjusted immediately without changing the complex both ends of the interval [33]. Its mathematical expression is determined
formula, and a lot of computing time can be saved. As a part of ML, by Eq. (5):
random forest (RF) has significant advantages over other traditional
machine learning methods in anti over fitting, fast operation speed and xn+1 = λxn (1 − xn ) 0 < λ⩽4 (5)
auto feature selection [63]. Therefore, this paper attempts to use the RF
to predict peak cutting force and use the slime mould algorithm (SMA) Sinusoidal map
to optimize the hyper-parameters of the RF to improve the prediction As can be seen from the following equation, the Sinusoidal map is
performance. Based on the lack of diversity in population initialization more nonlinear than the several chaotic maps described above [58]:
for SMA, six chaotic mapping algorithms with randomness, ergodicity
xk+1 = axk2 sin(πxk ) a = 2.3 (6)
and regularity are proposed to initialize the population, to lay the

2
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 1. Structure of a random forest.

Slime mould algorithm


Random forest
The slime mould algorithm uses a mathematical model to simulate
Breiman [3] combined decision trees into a random forest. Fig. 1
the slime mould foraging behavior and morphological changes [27]. The
shows the structure of a random forest. The idea behind RF is to average
slime mould foraging includes three stages: approaching the food stage,
multiple decision trees suffering from high variance, which builds a
wrapping food stage and oscillating stage, and finally adjusting the
more robust model with better generalization performance and is less
position of the whole slime mould to the optimal food place.
prone to over-fitting. In addition to using bagging technology to provide
“sample perturbation”, RF also adds “attribute perturbation”, which
Approach food
further increases the degree of difference between the final decision
The following equation simulates the process of slime mould
trees, thus further improving the generalization performance of the final
approaching food based on odour in the air:
integration model. The construction of an RF generally includes the
{ ̅̅→ → (→ ̅̅ ̅→)
following steps: ̅̅̅̅̅→
̅→ ̅̅
Xb (t) + vb⋅ W ⋅XA (t) − XB (t) , r < p
{ (
(a). A data set T = (x, y1 ), (x2 , y2 ), … xm , ym
)}
with a size of K is X(t + 1) = ̅̅→ (9)

vc⋅X(t), r⩾p
randomly selected for N times with sampling without replacement, and
one sample is selected each time. Finally, the selected N samples are ̅̅̅→ ̅̅̅→
where, XA (t) and XB (t) denote two slime moulds stochastically
used to train a decision tree. ̅̅→
(b). If each sample has M attributes when each node of the decision selected from the population in the t iteration. X(t) is the current posi­
̅̅̅→
tree needs to be split, m attributes are randomly selected from these M tion, Xb (t) represents the position with the highest concentration of food
attributes, and the condition m ≪ M are met. Then, Eq. (7) is used to →
in the t iteration. The range of vb is in the range of [-a, a], a = arctanh(1-
select one of the m attributes as the split attribute of the node. (t/T)). t refers to the current iteration, and T represents the maximum
(c). During the formation of the decision tree, each node should be iteration. →vc decreases linearly from one to zero, r is a random number
split according to Step (b) until it can’t be split. Note that there is no between zero and one. S(i) represents the fitness value of the ith slime
pruning during the formation of the decision tree. mould individual, N represents the population size of slime moulds, and
(d). Establish a large number of decision trees according to Steps (a) DF represents the optimal fitness value in all iterations.
~(c), so as to constitute a random forest, and the final result is deter­ p = tanh|S(i) − DF| (i = 1, 2, …, N), S(i) represents the fitness value of the
mined by averaging the results of all decision trees. ith slime mould individual, N represents the population size of slime
[ ∑ ∑ ] →
min min (yi − v1 )2 + min (yi − v2 )2 (7) moulds, and DF represents the optimal fitness value in all iterations. W
represents the weight of slime mould, whose formula is as follows:
where, for the partition feature A, the corresponding arbitrary ⎧ ( )
partition point s two sides are divided into data sets V1 and V2, and the ⎪ 1 + r⋅log bF − S(i) + 1 , conditions


mean–variance of the respective sets of V1 and V2 is the smallest, and the ̅̅̅̅̅̅̅̅̅̅̅̅→ ⎨ bF − wF
W(Smellindex(i)) = ( ) (10)
sum of the mean variances of V1 and V2 is the minimum of the feature ⎪

⎪ bF − S(i)
⎩ 1 − r⋅log + 1 , others
and eigenvalue division points, where v1 is the mean of V1 and v2 is the bF − wF
mean of V2.
Smellindex = sort(S) (11)
1∑
T= (T1 + T2 + ⋯ + TN ) (8)
N where bF and wF refer to the optimal and worst fitness values in the
current iteration, respectively. SmellIndex indicates the sorted fitness
Where T refers to the final result of the random forest, T1, T2 … TN
value sequence, and conditions denote the individuals whose fitness
denote the results of the N decision tree.
value ranks in the top half of the population.
→ → ̅̅̅→
Changes invb, →vc and W can adjust the optimal position X (t) to up
b

3
­
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 2. Geometrical features of conical pick cutting.

date individual positions. The role of r is to form a search vector at any


angle, that is, to search the solution space in any direction, thus 1∑ n ⃒ ⃒
MAE = ⃒PCFi − PCF ′ ⃒
i (13)
increasing the possibility of finding the optimal solution. n i=1

n ⃒ ⃒
Wrap food and oscillation 1∑ ⃒PCFi − PCFi′ ⃒
MAPE = ⃒

⃒ × 100%
⃒ (14)
Equation (12) simulates the feedback obtained by slime moulds n i=1 PCFi
based on the food concentration in the search area. If the food concen­
∑n ( ′ )2
tration in the location is high, the slime moulds in the location will be PCFi − PCFi
given positive feedback, otherwise, the slime moulds in the location will R2 = 1 − ∑i=1
n (15)
i=1 (PCFi − PCFi )
be given negative feedback. In this way, the fitness of slime moulds in
different locations will be used to adjust their weights of them and √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n
( ′ )2
further optimize the search mode of the whole population. RMSE = PCFi − PCF i (16)
n i=1

⎪ rand⋅(UB − LB) + LB, rand <z
⎨ ̅̅→ → ( ̅̅ ̅→)
̅→ ̅̅ where,PCF, PCF and PCF refer to the actual PCF values, the pre­

̅→
X ∗ = Xb (t) + vb⋅ W⋅XA (t) − XB (t) , r < p (12)

⎩ ̅̅→ dicted PCF values and the average of actual PCF values, respectively.

vc⋅X(t), r⩾p
where UB and LB represent the upper and lower boundaries of the Global performance indicator
search space, and rand and r represent random numbers between [0,1].
The conversion probability Z is set to 0.03, as recommended by Li et al. The comprehensive ranking of different models is cumbersome and
[27]. space-consuming. Therefore, a comprehensive evaluation index called
→ →
In conclusion,vb, →
vc and W were used to simulate the variations of the global performance indicator was introduced in this study [6]. In
venous width and the oscillation frequency of slime moulds, so that order to balance the weight of each index (MAE, RMSE, MAPE and R2),
slime moulds could approach the food more slowly when the food each index should be normalized in the interval [0,1]. The formula of
concentration is low, and approach the food faster when a high con­ GPI is as follows:
centration of food is found. ( )

4
GPIi = ̃ j − Indexij
ϕj Index (17)
j=1

Evaluation indicator
where, ϕj equals 1 for MAE, MAPE and RMSE, and equals − 1 for
Four pervasive and effective indicators, mean absolute error (MAE), R2.Index
̃ j represents the median of scaled values of index j, Indexij rep­
mean absolute percentage error (MAPE), determination coefficient (R2) resents the scaled value of index j for model i. It can be seen from the
and root mean square error (RMSE), is introduced to evaluate the ac­ formula that the more each indicator is less (higher for R2) than the
curacy of prediction models. MAE and RMSE can indicate the prediction median value, the higher the value of GPI will be. Therefore, a better
error directly, and relatively low values of MAE and MAPE can reflect model has a higher value of GPI.
the good performance of the model. MAPE depends on both the size of
predicted values and error, so the scale of the data does not affect MAPE. Proposed intelligent model
R2 reflects the correlation between predicted and actual values, which is
in the range of [0, 1]. Meanwhile, a small value of MAPE and a large Data Sources and Data Description
value of R2 indicates the excellent performance of the model. These
indicators can be obtained by [23–25,52,57,59,61,65]: The data set used in this study is collected and compiled from

4
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 3. Distributions of all variables in the data set.

published literature and composed of 63 rock types (Granite, limestone, previously developed PCF prediction models only considered a few rock
sandstone, andesite, tuff, etc.) from around the world types, and the prediction models that comprehensively considered
[4,39,43,2,49,51,28,55,42,36,54]. It is worth mentioning that various rock types in this study have stronger generalization ability. The

Fig. 4. Flowchart for constructing PCF prediction models.

5
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Table 1
Unimodal and multimodal test functions.
Unimodal functions (Dim = 30) Initial range Fmin

F1 (x) = ni=1 x2i [− 100,100] 0
∑ n
∏ [− 10,10] 0
F2 (x) = ni=1 |xi | + |xi |
i=1
( )2
∑n ∑i [− 100,100] 0
F3 (x) = i=1 j=1 xj

F4 (x) = max{|xi |, 1⩽i⩽n } [− 100,100] 0


i
∑n− 1 [ ( )
2 2
]
[− 30,30] 0
F5 (x) = i=1 100 xi+1 − xi + (xi − 1)2
∑n
F6 (x) = i=1 ix4i + random[0, 1) [− 1.28,1.28] 0
Multimodal functions (Dim = 30) Initial range Fmin
∑ ( √̅̅̅̅̅̅̅ )
F7 (x) = ni=1 − xi sin |xi | [− 500,500] − 418.9
829n
∑n [ ]
F8 (x) = x2i − 10cos(2πxi ) + 10
i=1
[− 5.12,5.12] 0
( √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ) )
1 ∑ n
( ∑
1 n
[–32,32] 0
2
F9 (x) = − 20exp − 0.2 x i − exp cos(2 π xi ) + 20 + e
n i=1 n i=1
( )
1 ∑n 2 ∏ n
xi [− 600,600] 0
F10 (x) = x −
i=1 i
cos √̅ + 1
4000 i=1 i

geometrical features of conical pick cutting are shown in Fig. 2. Based on parameter adjustment, grid search, random search and algorithm opti­
the fact that tensile strength of the rock σt, compressive strength of the mization [25]. However, the practice of a large number of scholars has
rock σ c, cone angle θ, cutting depth d and rake angle α have been widely proved that manual parameter adjustment is time-consuming and inef­
used to form empirical and theoretical PCF prediction models. It should ficient, the parameter selected by the grid search is always fixed and
be emphasized that σ t and σ c are correlated, while Evans [10] and Yasar easy to fall into the local optimality, and although the random search
[56] pointed out rock cuttability is correlated with the brittleness (σ c/ can find the global optimal, it is difficult to ensure that the best com­
σ t). Thus, σ t and σc are comprehensively considered in this study. bination of parameters is given. Therefore, this paper proposes a meta-
Additionally, attack angle γ, and back-clearance angle β have been heuristic algorithm called the slime mould algorithm to optimize the
extensively discussed in many literatures and have been shown to exert random forest, and the optimization algorithm can continuously iterate
significant influence on PCF. Therefore, the seven parameters σ t (MPa), to find the optimal parameter according to the historical parameter re­
σ c (MPa), θ (◦ ), d (mm), α (◦ ), γ (◦ ), and β (◦ ) are selected as input var­ cord, which not only reduces the operation time but also can find the
iables to develop PCF prediction models. The distribution of all variables optimal solution more accurately.
in the data set is shown in Fig. 3. It can be observed that the distribution Fig. 4 presents the flowchart for constructing and determining
of the pick operational parameters d and s, and the physical and me­ optimal PCF prediction models. In the beginning, a data set consisting of
chanical parameters of rock σ c and σt are relatively dispersed, which 205 samples is established and seven parameters that influence PCF are
indicate these parameters are broadly studied during experiments. Other set as input variables. For building robust ML models, data split ratio of
parameters are mainly concentrated on certain values, revealing that 4:1 is applied to divide the training set (80%) and testing set (20%)
these parameters are fixed at different levels. The relationship between which is a widely acknowledged scheme in the community. The role of
all input parameters and PCF is not simply linear. the training set is to select the model that is most appropriate/reason­
able in terms of predictive accuracy, generalization ability and model
complexity. The effectiveness of this method is verified by the testing set
Model framework
[35,67,16,17]. The dataset is collected in published literature, and
existing no outliers, duplicate and missing values, so the data cleaning
Tuning parameters is a crucial step in solving a problem using ma­
process is unnecessary. To enhance the prediction accuracy and
chine learning, also known as hyperparameter optimization. At present,
computational efficiency of the model, the dataset is normalized into
methods of hyperparameter optimization mainly include manual

Table 2
Results of testing functions with SMA and COSMA during 200 iterations.
Algorithm F1 F2 F3 F4 F5

AVG STD AVG STD AVG STD AVG STD AVG STD

CSMA 0.00620 0.02757 0.02228 0.05692 62.20433 113.16697 0.00251 0.01063 23.82984 22.77274
ISMA 2.31308 8.69293 0.05844 0.21916 26.08964 43.00787 0.09013 0.26068 17.28039 12.60216
LSMA 0.32751 1.66533 0.01294 0.05651 162.21995 778.23008 0.01903 0.04824 13.87558 11.86159
SINSMA 0.00568 0.02811 0.00073 0.02170 0.57227 2.79963 0.00212 0.00578 17.33603 11.63935
SSMA 0.95470 5.00572 0.01533 0.04520 72.29952 213.75903 0.01094 0.05591 16.16779 13.01360
TSMA 0.00970 0.03592 0.06840 0.30178 23.82801 51.64368 0.01861 0.06033 19.37806 11.97432
SMA 0.65012 2.18890 0.00981 0.03680 191.29460 503.43076 9.05862 27.02720 17.68214 12.42263
Algorithm F6 F7 F8 F9 F10
AVG STD AVG STD AVG STD AVG STD AVG STD
CSMA 0.02768 0.02542 − 12545.89790 38.48191 0.11902 0.38413 0.06023 0.30018 0.11439 0.31471
ISMA 0.01657 0.01217 − 12550.44689 35.67603 0.00663 0.02935 0.05198 0.22181 0.18494 0.37687
LSMA 0.02607 0.02668 ¡12555.48672 17.85786 0.00050 0.00254 0.01711 0.07259 0.05151 0.21412
SINSMA 0.01639 0.01572 − 12544.31225 39.37316 0.02519 0.13199 0.00813 0.03610 0.25559 0.28710
SSMA 0.02259 0.01980 − 12555.18408 18.97580 0.02393 0.08440 0.03577 0.18155 0.02204 0.11848
TSMA 0.02447 0.02052 − 12550.81365 28.86591 0.00112 0.00549 0.07228 0.24149 0.03850 0.18470
SMA 0.02100 0.01454 − 12531.94957 109.92645 0.01108 0.05961 0.06672 0.23259 0.03884 0.15361

6
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 5. Comparisons between SMA and COSMA.

[− 1,1]. In the next step, for enhancing the ability of SMA to get rid of Propose chaos optimized slime mould algorithm
local optimum parameters, six chaos mapping functions are proposed to
optimize SMA. Afterwards, two hyperparameters of RF, namely the Benchmark function validation
number of features used to grow each tree (mtry) and the number of trees
(ntree) are optimized by SMA and COSMA in the training stage and the To verify and compare the feasibility of COSMA, namely, Chebyshev
mean square error (MSE) is assigned as a fitness function. In this way, map optimized SMA (CSMA), Sine map optimized SMA (SINSMA),
the best hyperparameters of each model can be determined with fitness ICMIC map optimized SMA (ISMA), Tent map optimized SMA (TSMA),
convergence and thus the optimal models can be determined. During the Logistic map optimized SMA (LSMA), Sinusoidal map optimized SMA
last stage, in order to prove the superiority of the optimal RF-based (SSMA), ten different types of benchmark functions are introduced in
models, four commonly used theoretical formulae and four classical this study. As shown in Table 1, there are six unimodal functions and
ML models are introduced in this study. The models exhibiting lower GPI four multimodal functions. The optimization capability of COSMA can
value will be regarded as the optimal PCF prediction model and rec­ be fully investigated through a variety of classical benchmark functions.
ommended in engineering practice. As shown in Table 1, there are six unimodal functions and four

7
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

(0.00788). The prediction performance in the training and testing set


of SMA-RF and six COSMA-RF are shown in Fig. 7 and Fig. 8. Confidence
and prediction boundaries define the upper and lower bounds of the
correlation interval and define the width of the interval. The width of the
interval indicates how uncertain the fitting coefficients and predicted
observations are. This paper sets a 95% prediction interval. This interval
indicates that there is a 95% chance that the mean of observations and
observations are actually contained within the lower and upper limits of
predictions. As can be seen from Fig. 7 and Fig. 8, the confidence band
and prediction band are narrow, especially for the training set, which
conforms to the principle of machine learning, indicating that the pre­
diction performance of each model is relatively stable. In order to
compare the prediction performance of each model in a more compre­
hensive way, GPI is used to comprehensively evaluate all the evaluation
index values of the training set and testing set. The results are shown in
Table 3. The results show that SSMA-RF (GPI = 2.281), SINSMA-RF (GPI
= 1.636) and LSMA-RF (GPI = 1.419) are optimal RF models, and SMA-
RF (GPI = -4.422) has the comparatively worst prediction performance.
It further indicates that the prediction performance of COSMA-RF is
Fig. 6. Evolution of fitness values for SMA-RF and COSMA-RF.
better. It should be emphasized that the proposed models can test new
data not exceeding the range of variables in the training set to guide the
multimodal functions to investigate the optimization ability of COSMA. design of the cutting head [11,60,18].
To ensure the fairness of the comparison experiment, all algorithms
are carried out under the same conditions. Where the dimension was set Comparison with theoretical formulae
as 30, the population was set as 50, and the iteration time was set as 200.
In order to reduce the influence of random factors in the algorithm, the For a few decades, some researchers had proposed some commonly
results of each benchmark function ran independently 30 times. For the used theoretical formulae for PCF prediction [10,14,40,13], as shown in
ten benchmark functions, the average value (AVG) and standard devi­ Table 4. For demonstrating the superiority of the proposed models, the
ation (STD) of each algorithm are taken as the final evaluation index, as same testing set is substituted into formulae to test their predictive
shown in Table 2. The value in the bold format in Table 2 represents the performance on PCF. Fig. 9 shows the scatter plots of the predicted PCF
optimal value of each index of each function. using four theoretical formulae. The predicted PCF values in Fig. 9 (a)
and (b) are approximately identical and when the measured PCF values
Exploitation competence analysis are large, the predicted PCF values are smaller than the measured PCF
values, indicating that the prediction ability of these theoretical
The experimental results in Table 2 show that LSMA performed best formulae is limited and they have a weak ability for generalization. In
in F7-8, the AVG was the smallest in F5. and the AVG of SINSMA was the Fig. 9 (c) and (d), the distribution characteristics of scattered points are
smallest in F1-4, F6 and F9, the STD was the smallest in F2-5 and F9. similar, but the deviation is greater in Fig. 9 (d). To sum up, values of
SSMA obtains the smallest AVG and STD in F10, and CSMA and ISMA predicted PCF vary dramatically, losing fidelity at most points, demon­
have the smallest STD in F1 and F6 respectively. It can be seen that the strating that these formulae y do not provide satisfactory accuracy.
optimization ability of COSMA was improved compared with the orig­ According to the scatter distribution and prediction error index in
inal SMA, especially LSMA, SINSMA and SSMA. Figs. 7-9, it is obvious that the proposed hybrid ML models have better
In order to reflect the dynamic convergence performance of the prediction performance for PCF prediction. Therefore, the proposed
optimization algorithm, the convergence curves of 10 benchmark hybrid ML models are recommended for practical application.
functions under COSMA and SMA are given in Fig. 5. It is obvious that
SINSMA has the smallest initial fitness, demonstrating there is an Comparison with common ML model
excellent population in benchmark functions such as F1-F6 and F8-10. It
is worth mentioning that COSMA has an obvious initial excellent pop­ To further compare and evaluate the superiority of the proposed
ulation compared with the original SMA in F3, which shows that the models, four traditional and common ML models are introduced in this
proposed COSMA has the potential to improve the quality of the initial study, namely RF, extreme gradient boosting (XGBoost), extreme
population and global optimization ability and reduce the convergence learning machine (ELM) and back propagation neural network (BPNN).
time. The same training set and testing set are used to train and test the
aforementioned ML models, respectively. Table 5 summarizes the
Results and Discussion evaluation indicator values of the common ML models and Table 6
shows the GPI values compared with three optimal COSMA-RF models.
Comparison study of SMA-RF and COSMA-RF models Three optimal COSMA-RF models clearly outperform the common ML
models with lower values of MAE, MAPE and RMSE. Significantly, the
Fig. 6 presents the evolution of fitness in SMA-RF and six COSMA-RF. value of R2 of the XGBoost model on the training set is larger than three
It can be obviously seen from Fig. 6 that different models have different optimal COSMA-RF models, but the R2 value of the XGBoost model on
initial fitness and iteration processes, and eventually converge to various the testing set is obviously small and the remaining evaluation index is
values. The initial fitness of LSMA-RF is the lowest, and the iterative relatively large, existing over-fitting phenomenon in XGBoost model.
process remains stable. SMA-RF has the largest initial fitness, indicating The GPI values for three optimal COSMA-RF models and four common
that an inferior initial population appears in SMA-RF, and it reaches its models indicate that SINSMA-RF (GPI = 1.5584), SSMA-RF (GPI =
optimal fitness (0.00806) in 124 generations, which is the worst 1.5733) and LSMA-RF (GPI = 1.4908) clearly outperform common
compared with COSMA-RF. In addition, SINSMA-RF has the lowest models and RF (GPI = -1.8301) performs best among common models,
convergence value, followed by SSMA-RF (0.00746), TSMA-RF followed by XGBoost (GPI = -2.6073), BPNN (GPI = -3.6526) and ELM
(0.00756), CSMA-RF (0.00756), ISMA-RF (0.00775) and LSMA-RF (GPI = -3.1976). To further testify the superiority of the proposed

8
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 7. Performance comparison on the training set: (a) CSMA-RF (b) ISMA-RF (c) LSMA-RF (d) SINSMA-RF (e) SSMA-RF (f) TSMA-RF (g) SMA-RF.

9
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Fig. 8. Performance comparison on the testing set: (a) CSMA-RF (b) ISMA-RF (c) LSMA-RF (d) SINSMA-RF (e) SSMA-RF (f) TSMA-RF (g) SMA-RF.

10
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Table 3 predicted value and RMSE between actual and predicted values. As can
Ranking of different SMA-RF and COSMA-RF models according to GPI. be seen from Fig. 10, the prediction performance of three optimal
CSMA ISMA LSMA SINSMA SSMA TSMA SMA COSMA-RF models is roughly identical and clearly outperform the
common models.
GPI 0.026 − 0.599 1.419 1.636 2.281 0.026 − 4.422
Overall, the three recommended models not only have better accu­
racy and robustness compared with the previous published theoretical
formulae but also have a great improvement in the single RF model.
Table 4 Therefore, it is recommended to use SINSMA-RF, SSMA-RF and LSMA-
Theoretical formulae for PCF prediction.
RF models to predict PCF.
References Formulae

Evans [10] 16π σt 2


PCF = ) σd Sensitivity of variables
( θ σc t
cos2
2
Roxborough and Liu [40]
PCF =
16πd2 σc σ2t
) ) Sensitivity analysis refers to the calculation and analysis of the in­
fluence of changes of an uncertain factor on the prediction accuracy
[ (θ ( ( θ )]2
2σt + σc cos2 / 1 + tan(φ)/tan
2 2
Goktan [14] (θ
) under the assumption that other uncertain factors remain unchanged, so
4πd2 σt sin2 + φ as to determine the sensitivity value and select influential factors.
2
PCF =
Admittedly, excessive mechanical forces acting on conical pick would
)

cos + φ
2 affect the machine stability and increase energy consumption [28], thus
]
Goktan and Gunes [13] [1
12⋅π⋅σt ⋅d2 ⋅sin2 (90 − α) + φ it is reasonable to understand what parameters influence the PCF more
2
FC =
[1
] and then tune them.
cos (90 − α) + φ
2 To compare and analyze the sensitivity of selected input parameters,
the internal program of the RF is used to predict out-of-bag (OOB) data
through random perturbation, and the mean square error (MSE) of the
model, the Taylor diagram [38] of the testing set is adopted to visually two predictions after and before random perturbation of all decision
shows the following information: Standard deviation of the actual and trees is taken as a measure of this characteristic variable. The calculation
predicted values, the correlation coefficient between the actual and formula is as follows:

Fig. 9. Theoretical formulae for PCF prediction.

11
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

Table 5
Common models for PCF prediction.
Model Training set Testing set
2
MAE MAPE (%) R RMSE MAE MAPE (%) R2 RMSE

RF 2.3694 37.97 0.8562 4.1748 3.2792 57.16 0.8043 5.1349


XGBoost 3.1758 97.24 0.9452 3.7125 3.9806 144.05 0.8637 4.3062
ELM 3.0142 53.82 0.8332 4.0799 4.2509 82.63 0.7027 5.2872
BPNN 2.1130 40.95 0.8893 3.4144 5.0088 98.28 0.6481 7.9749

Table 6
Ranking of optimal COSMA-RF and different common ML models according to GPI.
Model SINSMA-RF SSMA-RF LSMA-RF RF XGBoost ELM BPNN

GPI 1.5584 1.5733 1.4908 − 1.8301 − 2.6073 − 3.6526 − 3.1976

1 ∑N
(
Scorei = msei1n − msei2n ) (18)
N n=1

where N denotes the number of decision trees, msei1n and msei2n refer
to the MSE of OOB after and before perturbation. A large value of Scorei
means there is a strong relationship between input variables and output
variables, indicating the input variable is significant for output variable
prediction. As shown in Fig. 11, σ t, σc, d and β are clearly more sensitive
to PCF than to θ, γ and α, which explains why σt, σ c, and d are common
parameters in theoretical formulae. It is further concluded that β also has
a significant effect on PCF, which is consistent with Yasar’s [56]
conclusion, suggesting the addition of β in the PCF prediction model is
effective and the adjustment of β in practice for decreasing PCF is rec­
ommended. Additionally, θ, γ, and α also have an effect on PCF.

Conclusions

This study comprehensively compared and investigated the predic­


tion performance of the hybrid RF-based optimization techniques in PCF
prediction. A data set having seven input variables - tensile strength of
the rock σ t, compressive strength of the rock σc, cone angle θ, cutting
depth d, attack angle γ, rake angle α and back-clearance angle β and one
Fig. 10. Taylor graph of optimal COSMA-RF and different common ML models. output variable PCF were first established. Whereas 80% and 20% of the
data were used to train and test the model, respectively.
For avoiding the search being trapped in a local optimum, chaos
optimization algorithm i.e., Chebyshev map, Sine map, ICMIC map, Tent
map, Logistic map and Sinusoidal map are combined with slime mould
algorithm. Benchmark function validation indicated the optimization
ability of LSMA, SINSMA and SSMA is greatly improved than SMA. To
obtain the optimal hyperparameters in the RF algorithm, the original
meta-heuristic SMA and proposed COSMA were combined with the RF
algorithm and mean square error was assigned as the fitness function.
The global performance index (GPI) values for the training and testing
set revealed that SSMA-RF (GPI = 2.281), SINSMA-RF (GPI = 1.636) and
LSMA-RF (GPI = 1.419) are clearly outperformed the other hybrid
models. The three optimal COSMA-RF models obviously outperformed
the theoretical formulae of predicting PCF and yielded lower MAE,
MAPE, RMSE and R2. In addition, four other predictive models i.e., RF,
XGBoost, ELM and BP were built to predict PCF for comparison pur­
poses. The results indicated that COSMA-RF models showed signifi­
cantly prediction performance enhancement compared with the four
non-optimized models. The sensitive analysis in the RF algorithm
showed that the PCF prediction is much more sensitive to σc, σt, d and β
than to θ, γ and α. To sum up, SSMA-RF, SINSMA-RF and LSMA-RF are
recommended for PCF prediction in practice. It should be noted that the
Fig. 11. Sensitive scores of seven input parameters. COSMA-RF model can guide the design of the cutting head. However, it
can’t substitute the necessary and practical tools – such as numerical
models, laboratory tests and field tests, etc. – that are normally needed
for a cutting head design.

12
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

CRediT authorship contribution statement [21] Khandelwal M, Monjezi M. Prediction of backbreak in open-pit blasting operations
using the machine learning method. Rock Mech Rock Eng 2013;46(2):389–96.
[22] Kohli M, Arora S. Chaotic grey wolf optimization algorithm for constrained
Jian Zhou: Conceptualization, Methodology, Visualization, Super­ optimization problems. J Comput Des Eng 2018;5:458–72.
vision, Writing – review & editing, Funding acquisition. Yong Dai: [23] Li CQ, Zhou J, Khandelwal M, Zhang XL, Monjezi M, Yu Z, et al. Six Novel Hybrid
Resources, Methodology, Visualization, Writing – original draft. Kun Extreme Learning Machine—Swarm Intelligence Optimization (ELM–SIO) Models
for Predicting Backbreak in Open-Pit Blasting. Nat Resour Res 2022. https://doi.
Du: Formal analysis, Investigation, Writing – review & editing. Manoj org/10.1007/s11053-022-10082-3.
Khandelwal: Visualization, Writing – review & editing. Chuanqi Li: [24] Li E, Yang F, Ren M, Zhang X, Zhou J, Khandelwal M. Prediction of blasting mean
Formal analysis, Writing – review & editing. Yingui Qiu: Writing – re­ fragment size using support vector regression combined with five optimization
algorithms. J Rock Mech Geotech Eng 2021;13:1380–97.
view & editing. [25] Li E, Zhou J, Shi X, Armaghani DJ, Zhi Yu, Chen X, et al. Developing a hybrid
model of salp swarm algorithm-based support vector machine to predict the
strength of fiber-reinforced cemented paste backfill. Eng Comput 2021;37:
Declaration of Competing Interest 3519–40.
[26] Li P, Duo Xu, Zhou Z, Lee W-J, Zhao Bo. Stochastic Optimal Operation of Microgrid
Based on Chaotic Binary Particle Swarm Optimization. IEEE Trans Smart Grid
The authors declare that they have no known competing financial 2016;7:66–73.
interests or personal relationships that could have appeared to influence [27] Li S, Chen H, Wang M, Heidari AA, Mirjalili S. Slime mould algorithm: A new
the work reported in this paper. method for stochastic optimization. Future Gener Comput Syst-the Int J Escience
2020;111:300–23.
[28] Li X, Wang S, Ge S, Malekian R, Li Z. A Theoretical Model for Estimating the Peak
Acknowledgements Cutting Force of Conical Picks. Exp Mech 2018;58:709–20.
[29] Lian SG, Sun JS, Wang ZQ. A block cipher based on a suitable use of the chaotic
standard map. Chaos, Solitons Fractals 2005;26:117–29.
This research was funded by the National Science Foundation of [30] Liu H, Wang X, Kadir A. Image encryption using DNA complementary rule and
China (42177164), the Innovation-Driven Project of Central South chaotic maps. Appl Soft Comput 2012;12:1457–66.
[31] Liu W, Sun K, Zhu C. A fast image encryption algorithm based on chaotic map. Opt
University (No. 2020CX040), the Fundamental Research Funds for the
Lasers Eng 2016;84:26–36.
Central Universities of Central South University (2022ZZTS0480) and [32] Mei X, Li C, Sheng Q, Cui Z, Zhou J, Dias D. Development of a hybrid artificial
Postgraduate Innovative Project of Central South University intelligence model to predict the uniaxial compressive strength of a new aseismic
(1053320213104). layer made of rubber-sand concrete. Mech Adv Mater Struct 2022:1–18.
[33] Miyazaki T, Araki S, Uehara S. Some Properties of Logistic Maps over Integers.
Ieice Trans Fundam Electron Commun Comput Sci 2010;E93A:2258–65.
References [34] Pan Y, Liu Q, Kong X, Liu J, Peng X, Liu Qi. Full-scale linear cutting test in
Chongqing Sandstone and the comparison with field TBM excavation performance.
Acta Geotech 2019;14:1249–68.
[1] Armaghani DJ, Harandizadeh H, Momeni E, Maizir H, Zhou J. An optimized system
[35] Phoon K-K, Ching J, Shuku T. Challenges in data-driven site characterization.
of GMDH-ANFIS predictive model by ICA for estimating pile bearing capacity. Artif
Georisk: Assess. Manage. Risk. Eng Syst 2021;Geohazards:1–13.
Intell Rev 2022;55(3):2313–50.
[36] Polat C. Roadheader Performance Prediction Using Portable Linear Cutting
[2] Bilgin N, Demircin MA, Copur H, Balci C, Tuncdemir H, Akcin N. Dominant rock
Machine. Fen Bilimleri Enstitüsü 2015.
properties affecting the performance of conical picks and the comparison of some
[37] Poon CS, Barahona M. Titration of chaos with added noise. PNAS 2001;98:
experimental and theoretical results. Int J Rock Mech Min Sci 2006;43:139–56.
7107–12.
[3] Breiman L. Random forests. Mach Learn 2001;45:5–32.
[38] Qiu Y, Zhou J, Khandelwal M, Yang H, Yang P, Li C. Performance evaluation of
[4] Copur H, Bilgin N, Tuncdemir H, Balci C. A set of indices based on indentation tests
hybrid WOA-XGBoost, GWO-XGBoost and BO-XGBoost models to predict blast-
for assessment of rock cutting performance and rock properties. J South Afr Inst
induced ground vibration. Eng Comput 2021:1–18. https://doi.org/10.1007/
Min Metall 2003;103:589–99.
s00366-021-01393-9.
[5] Dai, Yong, Manoj Khandelwal, Yingui Qiu, Jian Zhou, M. Monjezi, and Peixi Yang.
[39] Roepke WW, Voltz JI. “Coal-cutting forces and primary dust generation using
2022. A hybrid metaheuristic approach using random forest and particle swarm
radial gage cutters. Report of investigations/1983.” In.: Bureau of Mines, Twin
optimization to study and evaluate backbreak in open-pit blasting, Neural
Cities, MN (USA). Twin Cities Research Center; 1983.
Computing & Applications, 34: 6273-88.
[40] Roxborough FF, Liu ZC. Theoretical considerations on pick shape in rock and coal
[6] Despotovic M, Nedic V, Despotovic D, Cvetanovic S. Review and statistical analysis
cutting; 1995.
of different global solar radiation sunshine models. Renew Sustain Energy Rev
[41] Shao W, Li X, Sun Y, Huang H. Parametric study of rock cutting with SMART(*)
2015;52:1869–80.
CUT picks. Tunn Undergr Space Technol 2017;61:134–44.
[7] Du K, Li X, Su R, Tao M, Lv S, Luo J, et al. Shape ratio effects on the mechanical
[42] Su O, Akcin NA. Numerical simulation of rock cutting using the discrete element
characteristics of rectangular prism rocks and isolated pillars under uniaxial
method. Int J Rock Mech Min Sci 2011;48:434–42.
compression. Int J Min Sci Technol 2022;32(2):347–62.
[43] Sundae LS, Myren TA. In Situ Comparison of Radical and Point Attack Bits. Bureau
[8] Du K, Sun Y, Zhou J, Wang SF, Tao M, Yang C, et al. Low amplitude fatigue
of Mines: US Department of the Interior; 1987.
performance of sandstone, marble, and granite under high static stress. Geomech
[44] Tiryaki B, Boland JN, Li XS. Empirical models to predict mean cutting forces on
Geophys Geo-Energy Geo-Resources 2021;7(3):1–21.
point-attack pick cutters. Int J Rock Mech Min Sci 2010;47:858–64.
[9] Du K, Li XF, Yang CZ, Zhou J, Chen SJ, Manoj K. Experimental investigations on
[45] Tiryaki B. Estimating Rock Cuttability using Regression Trees and Artificial Neural
mechanical performance of rocks under fatigue loads and biaxial confinements.
Networks. Rock Mech Rock Eng 2009;42:939–46.
J Central South Univ 2020;27(10):2985–98.
[46] Wang SJ, Zong XM, He B, Kang KX. “Research on Prediction Model of Conical Pick
[10] Evans Ivor. Theory of the cutting force for point-attack picks, Int J Min Eng; 1984,
Cutting Force Based on Coulomb-Mohr Criterion.” In: 8th Asia Conference on
2.
Mechanical and Materials Engineering (ACMME). Natl Univ Singapore, ELECTR
[11] Feng X, Jimenez R. Predicting tunnel squeezing with incomplete data using
NETWORK; 2020.
Bayesian networks. Eng Geol 2015;195:214–24.
[47] Wang SM, Zhou J, Li CQ, Armaghani DJ, Li XB, Mitri HS. Rockburst prediction in
[12] Georgeot B, Shepelyansky DL. Exponential gain in quantum computing of quantum
hard rock mines developing bagging and boosting tree-based ensemble techniques.
chaos and localization. Phys Rev Lett 2001;86:2890–3.
J Central South Univ 2021;28(2):527–42.
[13] Goktan RM, Gunes N. A semi-empirical approach to cutting force prediction for
[48] Wang X, Okan Su, Wang Q-F, Liang Y-P. Effect of cutting depth and line spacing on
point-attack picks. J S Afr Inst Min Metall 2005;105:257–63.
the cuttability behavior of sandstones by conical picks. Arabian J Geosci 2017;10.
[14] Goktan RM. A suggested improvement on Evans’ cutting theory for conical bits.
[49] Wang X, Su O, Wang QF, Liang YP. Effect of cutting depth and line spacing on the
Proc Fourth Symp Mine Mech Autom 1997:57–61.
cuttability behavior of sandstones by conical picks. Arabian J Geosci 2017;10.
[15] Hua Z, Zhou Y, Pun C-M, Philip Chen CL. 2D Sine Logistic modulation map for
[50] Wang X, Wang Q-F, Liang Y-P, Okan Su, Yang L. Dominant Cutting Parameters
image encryption. Inf Sci 2015;297:80–94.
Affecting the Specific Energy of Selected Sandstones when Using Conical Picks and
[16] Jin YF, Yin Z-Y, Zhou WH, Shao JF. Bayesian model selection for sand with
the Development of Empirical Prediction Models. Rock Mech Rock Eng 2018;51:
generalization ability evaluation. Int J Numer Anal Methods Geomech 2019;43
3111–28.
(14):2305–27.
[51] Wang X, Wang Q-F, Liang Y-P. Effects of cutting parameters affecting on specific
[17] Jin YF, Yin Z-Y, Zhou WH, Horpibulsuk S. Identifying parameters of advanced soil
cutting energy of conical picks. J Coal Sci Eng (China) 2018;43(02):563–70.
models using an enhanced Transitional Markov chain Monte Carlo method. Acta
[52] Xie C, Nguyen H, Bui XN, Nguyen VT, Zhou J. Predicting roof displacement of
Geotech 2019;14(6):1925–47.
roadways in underground coal mines using adaptive neuro-fuzzy inference system
[18] Jin YF, Yin Z-Y. Enhancement of backtracking search algorithm for identifying soil
optimized by various physics-based optimization algorithms. J Rock Mech Geotech
parameters. Int J Numer Anal Methods Geomech 2020;44(9):1239–61.
Eng 2021;13(6):1452–65.
[19] Kaur G, Arora S. Chaotic whale optimization algorithm. J Comput Des Eng 2018;5:
[53] Yang D, Li J, Wang L, Gao K, Tang Y, Wang Y. Experimental and theoretical design
275–84.
for decreasing wear in conical picks in rotation-drilling cutting process. Int J Adv
[20] Khandelwal M, Singh TN. Prediction of blast-induced ground vibration using
Manuf Technol 2015;77:1571–9.
artificial neural network. Int J Rock Mech Min Sci 2009;46(7):1214–22.

13
J. Zhou et al. Transportation Geotechnics 36 (2022) 100806

[54] Yasar S. Determination of optimum rock cutting data through single pick cutting [62] Zhou J, Qiu Y, Khandelwal M, Zhu S, Zhang X. Developing a hybrid model of Jaya
tests. Geotech Lett 2019;9:8–14. algorithm-based extreme gradient boosting machine to estimate blast-induced
[55] Yasar S, Yilmaz AO. Vertical rock cutting rig (VRCR) suggested for performance ground vibrations. Int J Rock Mech Min Sci 2021;145.
prediction of roadheaders. Int J Min Reclam Environ 2019;33:149–68. [63] Zhou J, Dai Y, Khandelwal M, Monjezi M, Yu Z, Qiu Y. Performance of Hybrid SCA-
[56] Yasar S. A General Semi-Theoretical Model for Conical Picks. Rock Mech Rock Eng RF and HHO-RF Models for Predicting Backbreak in Open-Pit Mine Blasting
2020;53:2557–79. Operations. Nat Resour Res 2021;30:4753–71.
[57] Yong W, Zhang W, Nguyen H, Bui XN, Choi Y, Nguyen-Thoi T, et al. Analysis and [64] Zhou J, Qiu Y, Armaghani DJ, Zhang W, Li C, Zhu S, et al. Predicting TBM
prediction of diaphragm wall deflection induced by deep braced excavations using penetration rate in hard rock condition: a comparative study among six XGB-based
finite element method and artificial neural network optimized by metaheuristic metaheuristic techniques. Geosci Front 2021;12(3):101091.
algorithms. Reliab Eng Syst Saf 2022;221:108335. [65] Zhou J, Qiu Y, Zhu S, Armaghani DJ, Li C, Nguyen H, et al. Optimization of support
[58] Yu M, Sun K, Liu W, He S. A hyperchaotic map with grid sinusoidal cavity. Chaos, vector machine through the use of metaheuristic algorithms in forecasting TBM
Solitons Fractals 2018;106:107–17. advance rate. Eng Appl Artif Intell 2021;97:104015.
[59] Zhang P, Yin Z-Y, Jin Y-F, Chan THT. A novel hybrid surrogate intelligent model [66] Zhou J, Huang S, Qiu Y. Optimization of random forest through the use of MVO,
for creep index prediction based on particle swarm optimization and random GWO and MFO in evaluating the stability of underground entry-type excavations.
forest. Eng Geol 2020;265. Tunn Undergr Space Technol 2022;124:104494.
[60] Zhang P, Yin Z-Y, Jin Y-F. Machine learning-based modelling of soil properties for [67] Zhou J, Huang S, Zhou T, Armaghani DJ, Qiu Y. Employing a genetic algorithm and
geotechnical design: review, tool development and comparison. Arch Comput grey wolf optimizer for optimizing RF models to evaluate soil liquefaction
Method Eng 2022;29:1229–45. potential. Artif Intell Rev 2022:1–33.
[61] Zhang P, Yin ZY, Jin YF. Bayesian neural network-based uncertainty modelling: [68] Zhou J, Shen X, Qiu Y, Shi X, Khandelwal M. Cross-correlation stacking-based
application to soil compressibility and undrained shear strength prediction. Can microseismic source location using three metaheuristic optimization algorithms.
Geotech J 2022;59:546–57. Tunn Undergr Space Technol 2022;126:104570.

14

You might also like