Geotech Geol Eng (2022) 40:3191–3209

https://doi.org/10.1007/s10706-022-02086-w

ORIGINAL PAPER

A Predictive Model for Estimating the TBM Penetration

Rate Based on Hybrid ICA‑ANN and DEA‑AHP Algorithms
Mosleh Eftekhari · Nasrollah Eftekhari

Received: 5 August 2021 / Accepted: 15 February 2022 / Published online: 4 March 2022
© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2022

Abstract Predicting the penetration rate of tun- performance. The determination coefficient ­(R2) and
nel boring machine (TBM) is a complex and chal- the mean square error (MSE) indices of the train-
lenging task that plays a crucial role in the schedule ing and testing datasets are 0.901, 0.887, 0.04 and
planning and cost estimation of tunneling projects. 0.0645, respectively. This study shows that the afore-
In this study, Mutual Information (MI) is utilized to mentioned methods could be utilized for enhancing
select the most effective parameters on penetration the prediction of the penetration rate of TBMs.
rate among various rock mass and machine operating
parameters. Moreover, Data Envelopment Analysis- Keywords Penetration rate · Tunnel boring machine
Analytic Hierarchy Process (DEA-AHP) approach is (TBM) · Mutual information (MI) · Imperialist
implemented in Imperialist Competitive Algorithm- competitive algorithm (ICA) · Data envelopment
Artificial Neural Network (ICA-ANN) to improve analysis (DEA) · Analytic hierarchy process (AHP)
its performance as well as to achieve the best possi-
ble architecture of the network. The aforementioned
methods are applied on a database consists of 430 1 Introduction
data collected from the Lot 2 of Zagros tunnel pro-
ject. The results indicated that the network with Evaluating the performance of Tunnel Boring
6 input variables, including uniaxial compressive Machine (TBM) is one of the main challenges and
strength (UCS), Brazilian tensile strength (BTS), dis- issues of tunneling projects, and an acceptable pre-
tance between the planes of weakness (DPW), rock diction of the performance can lead to a reliable esti-
quality designation (RQD), thrust and revolution per mate of the time and budget required for the project.
minute (RPM), 9 neurons in the hidden layer and In recent decades, various theoretical and empirical
the penetration rate as the output provides the best methods have been developed to predict the TBM
performance parameters, especially penetration rate.
Numerous parameters affect the penetration rate that
M. Eftekhari (*) can be divided into two general categories: rock mass,
Department of Mining Engineering, Faculty
including the properties of intact rock and disconti-
of Engineering, Tarbiat Modares University, Jalal
AleAhmad, Nasr Street, 14115‑14 Tehran, Iran nuities, and machine operating parameters. Theoreti-
e-mail: [email protected] cal methods, which are based on the breakage mecha-
nism analysis, focus on the theoretical relationship
N. Eftekhari
between the cutting forces acting on cutters and its
Faculty of Technology and Mining, Yasouj University,
Choram, Iran penetration and also, in some cases, the parameters

Vol.: (0123456789)
13
 3192 Geotech Geol Eng (2022) 40:3191–3209

of rock in order to estimate the penetration rate the rock cutting behaviors and mechanisms under the
(Ozdemir et al. 1978; Roxborough and Phillips 1975; TBM disc cutter, including the finite element method
Sanio 1985; Sato 1991; Snowdon et al. 1982). Rox- (Bilgin et al. 2000; Li et al. 2018; Liu et al. 2002;
borough and Phillips (1975) developed theoretical Qi et al. 2016; Yang et al. 2016), the displacement
equations for estimating cutting forces based on the discontinuity method (Haeri and Marji 2016; Marji
rock compressive strength along with the disc diam- et al. 2009; Marji 1997, 2015), and the distinct ele-
eter and the tip edge angle. A set of predictor equa- ment method (Eftekhari et al. 2014; Gong et al. 2005,
tions was developed by Ozdemir et al. (1978) to esti- 2006a, b; Moon and Oh 2012).
mate the normal and rolling forces using the cutting In recent years, artificial intelligence techniques
geometry (i.e., spacing and penetration) along with have become one of the approaches to solve vari-
uniaxial compressive and tensile strength. Howarth ous geotechnical problems, including penetration
and Roxborough (1982) conducted a research study rate prediction (Afradi and Ebrahimabadi 2020; Eft-
to incorporate the effects of joints. Sanio (1985) ekhari et al. 2010; Khandelwal and Armaghani 2016;
offered a tensile failure model for the chip formation Singh et al. 2004; Verma and Singh 2011). Intelli-
and introduced some equations for estimating the cut- gent methods make it possible to incorporate more
ting forces. Rostami and Ozdemir (1993) and Ros- parameters than empirical and theoretical approaches.
tami (1997) proposed some formulas for estimation Many researchers have developed a variety of mod-
of the cutting forces acting on constant cross section els with different input parameters to predict TBM
(CCS) disc cutters. These equations are more com- performance. Some of these approaches have limita-
monly referred as the CSM model which is one of tions, including low learning rates and being caught
the well-known prediction models in TBM tunneling up in local optima (Ghaleini et al. 2019; Lee et al.
projects (Rostami 2013). The empirical methods are 1991; Wang et al. 2004). In order to overcome these
based on the relationships between the measured rock obstacles, optimization algorithms such as Genetic
mass properties, machine operating parameters and Algorithm (GA), Particle Swarm Optimization (PSO)
the performance parameters. Many researchers have and Imperialist Competitive Algorithm (ICA) have
predicted the penetration rate based on the param- been developed to facilitate the optimization of the
eters of the rock mass or its combination with the weights and biases of the networks in network train-
operational parameters of the machine (Farmer and ing. The combination of these optimization algo-
Glossop 1980; Graham 1976; Tarkoy 1974). Some rithms and artificial neural network have been used
investigators have introduced the penetration rate as to predict TBM performance (Afradi and Ebrahima-
a function of rock mass classification systems such badi 2021; Afradi et al. 2021; Armaghani et al. 2019,
as Rock Mass Rating (RMR), Rock Structure Rating 2017; Mohamad et al. 2017). In the ICA, after creat-
(RSR), Rock Quality Designation (RQD), Geological ing a random number of countries, the value of Mean
Strength Index (GSI) and Rock Mass Quality Index Square Error (MSE) of training data is determined as
(Q) (Innaurato et al. 1991; Park et al. 2001; Ribac- the cost function (Armaghani et al. 2019). The coun-
chi and Fazio 2005; Sapigni et al. 2002; Sundaram tries with the lowest costs are selected as the impe-
et al. 1998). The NTH method, which is based on the rialists. However, the MSE of the training data set
empirical relationship between the mechanical and alone may not be the most appropriate criterion for
geological parameters of the rock mass and the per- selecting the countries because, in some cases, the
formance of TBM (Bruland 1998), and the QTBM error of training data may decline with increasing the
method, which is presented by Barton (2000) with number of hidden neurons, while the error associated
modifications to the Q system, are the two empirical with the testing data increases (Kurt et al. 2009). For
models widely used in practice. Although the pro- this reason, many scholars (Armaghani et al. 2017;
posed methods naturally take into account the inter- Eftekhari et al. 2010; Kurt et al. 2009) employ differ-
action between the rock mass and the machine, they ent performance indices such as MSE and coefficient
are site-specific and cannot be easily generalized to of determination to assess the networks and choose
other rock properties or other areas (Armetti et al. the optimal one.
2018; Brino et al. 2015). Furthermore, numerical In this study, instead of selecting imperialists (best
modeling methods have been adopted to investigate countries) based on the countries’ cost (training set

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3193

error), the countries are ranked in terms of their effi- (14 km), the middle section (9 km) and lot 2 (26 km)
ciency and the top-ranked countries are determined as the southwest section. The maximum depth of the
as the imperialists. The efficiency is simply defined tunnel is 1000 m, with an average of about 400 m.
as the ratio of output(s) to input(s). Anything whose The lot 2 of the tunnel was excavated using a dou-
increase leads to an improvement in the efficiency ble shield TBM. The main specifications of the TBM
can be considered as an output. Likewise, an input are given in Table 1.
can be defined as anything whose decrease leads to According to the studies of geology and engineer-
a lower efficiency (Shayeghi et al. 2017). Therefore, ing geology carried out in different stages and consid-
the MSE of both the training and testing datasets are ering the repetition of units in different parts, 21 units
regarded as the inputs, while their determination coef- of engineering geology was distinguished in the study
ficients are regarded as the outputs. Since it is difficult area. The longitudinal geological profile of Zagros
to determine the weighting of these inputs and out- long tunnel (lot 2) is presented in Fig. 2.
puts indices to rank the countries, the hybrid of Data The obtained 430 datasets from the Zagros tun-
Envelopment Analysis and Analytic Hierarchy Pro- nel, which are composed of PR (m/h), six intact
cess (DEA-AHP), which has been widely used in the rock properties (i.e., Uniaxial Compressive Strength
field of decision-making, is incorporated in the ICA- (UCS), Elastic modulus, Porosity, Density, Brazilian
ANN to determine the best weights and biases for Tensile Strength (BTS) and Quartz content), five rock
any network architecture. Moreover, the DEA-AHP mass properties (i.e., Distance between the Planes of
is applied to find the best architecture among the net- Weakness (DPW), RQD, RMR, GSI and Q system),
works with different number of hidden neurons. and four operational parameters of TBM (i.e., RPM,
It is worth noting that the selection of the input Thrust force, Torque and power), are used in this
variables plays a crucial role in artificial intelligence, study. The main physical and mechanical parameters
and can have a significant impact on its performance. of the rock in the route of tunnel are summarized in
So far, different effective parameters have been con- Table 2. This table also includes description of opera-
sidered as the input parameters in the literature. tional parameters and penetration rate.
Hence, in this study, Mutual Information (MI) method
is used to select the predictors for the hybrid proposed 2.2 Mutual Information (MI)
approach. The MI is an information theoretic measure
that describes the dependency or association between Selection of the input variables is an important pre-
two random variables, especially in nonlinear systems processing step for model development. Generally, a
(Wang et al. 2016). If the MI between a given vari- complex prediction model with many input variables
able and the output was negligible, the variable would may suffer from some disadvantages such as difficulty
be excluded from the predictors. in their use and the long calculation times. Moreo-
This paper aims to predict the penetration rate of ver, more complex models typically lead to increased
TBM using the ICA-ANN by incorporating the DEA- parametric uncertainty due to difficulty of parameter
AHP approach to find the optimal weights and biases estimation. Therefore, it is useful to select a subset
of the network as well as its architecture, based on of relevant and indispensable variables for develop-
collected data from the Lot 2 of Zagros water convey- ing effective prediction models. Accordingly, if the
ance tunnel project. dependency between a given variable and the output
of model is negligible, then the subset of the input
variables does not include that variable.
2 Materials and Methods A large number of variable selection algorithms
have been developed in recent years [see (Sutha and
2.1 Study Area and Data Sets Tamilselvi 2015) for more information]. Among
them, the MI is widely used to measure relevance
Zagros water conveyance tunnel, with a total length between features and decision (Yu et al. 2011). One
of 49 km and diameter of 6.73 m, is located in in advantage of this criterion is to capture the nonlin-
western part of Iran (Fig. 1). The Zagros tunnel was ear dependency between variables while some other
divided into three sections: the northeast section classical criteria (e.g., Pearson’s correlation filter and

Vol.: (0123456789)
13
 3194 Geotech Geol Eng (2022) 40:3191–3209

Fig. 1  Location of study

area in Iran (red line is the
tunnel route) (Engineers
2010)

Table 1  Main specifications of TBM and Y with probability mass functions pXi and pY are
Parameter Value defined respectively as:

∫
Machine diameter 6.73 m H(Y) = − pY (y)log(pY (y))dy (1)
Cutter diameter 432 mm
Number of disc cutters 42
Disc nominal spacing 90 mm
∫ (2)
Max. operating cutterhead thrust 28,134 kN at 350 bar
H(Xi ) = − pXi (x)log(pXi (x))dx
Cutterhead power 2100 kW
Cutterhead speed 0–11 This quantity measures the uncertainty on these
Cutterhead torque (nominal) 4450 kN.m at 9 rpm variables. If the entropy of a variable is large, then
thrust cylinder stroke 1700 mm its uncertainty is high, and vice versa. The concept of
Conveyor capacity (approx.) 690 t/h entropy can be extended to a pair of random varia-
Total TBM weight (approx.) 573 ton bles. Conditional entropy, H(Y|Xi), defines how much
uncertainty exists in the output Y given knowledge of
one of the input parameters Xi:
principal component analysis) are essentially lim-
∫ ∫
ited to linear dependencies. Also, the MI can capture H(Y ||Xi ) = − pXi (x) pY (y|x )log(pY (y|x ))dydx
dependence arising from moments higher than the (3)
conditional mean (Frénay et al. 2013).
where pY (y| x) denote the conditional density of
Let Y denotes the target output and X = {X1, X2,…,
y given x. Building on this definition, the mutual
Xn} is a subset of variables. Then, the entropy of Xi

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3195

Fig. 2  Profile of geology and geological map of Zagros long tunnel (Engineers 2010)

Table 2  Statistical Parameter Unit Symbol Min Max Average

information regarding the
established database Uniaxial compressive strength MPa UCS 15 150 42.5
Elastic modulus GPa E 4 30 6.25
Porosity % n 2 15 7.27
Density gr/cm3 ρ 2.3 2.6 2.44
Brazilian tensile strength MPa BTS 1 6 3.16
Quartz content % q 0 5 1.27
Distance between the planes of weakness cm DPW 10 200 44
Rock quality designation % RQD 25 90 67
Rock mass rating RMR 28 60 46
Geological strength index GSI 18 60 35
Quality system Q 0.32 4.86 2.3
Thrust force kN F 2000 9940 7096
Revolution per minute RPM 3 7.38 5.2
Torque kN.m T 92 1263 805
Power % P 23 60 39
Penetration rate m/h PR 1.06 4.86 2.44

( )
information,I Xi , Y , is defined as the difference in This quantity measures the amount of informa-
the uncertainty of Y with and without knowledge of tion that one variable contains about the other (Huoh
Xi, and characterizes the influence of Xi exerts on Y: 2013). In other words, the MI quantifies the statisti-
( ) cal dependence between two random variables, and
I Xi , Y = H(Y) − H(Y ||Xi ) (4) it provides a better criterion than the autocorrelation

Vol.: (0123456789)
13
 3196 Geotech Geol Eng (2022) 40:3191–3209

function, which only measures linear dependence with actual values of the predicted variable, until the
(Fraser and Swinney 1986). error measure is at an acceptable value that can be
In practice, if there was a database with unknown defined as the mean square error (MSE) (PK 1990)
densities, it would be impossible to directly compute which is defined as:
the MI. In such a condition, using the nearest neigh- N
bor approach for estimating this quantity can be very 1 ∑( )2
MSE = yreal − ynet (6)
cumbersome (Kraskov et al. 2004). Generally, this N i=1
numerical approach for estimating MI is more attrac-
tive than other methods such as Bayesian density esti- where N indicates samples number, ­yreal is the real
mators (Steuer et al. 2002). In this paper, the k-near- output and ­ynet is the output of the network.
est neighbor based estimator is used because it has Back propagation (BP) technique is one of the
proven to be successful in estimation of MI (Rossi most popular optimization methods among the MLP
et al. 2006). In this method, MI is estimated from the learning algorithms (Basheer and Hajmeer 2000;
distribution of the nearest-neighbour-distance of the Gordan et al. 2019), but it is likely to get stuck in
data points, as (Kraskov et al. 2004): local minima. Hence, optimization approaches have
n n n
( ) d ∑ dy ∑ d(x,y) ∑
̂I Xi , Y = −𝜓(k) + 𝜓(n) + x log{𝜀x i,k } − 𝜓(k) + 𝜓(n) + log{𝜀y i,k } + 𝜓(k) − 𝜓(n) + log{𝜀(x,y) i,k } (5)
n i=1 n i=1 n i=1

where ψ denotes the digamma function, and εxi,k, εyi,k been proposed to improve ANN training process.
and ε(x, y)i,k are the distance between xi and its k­ th near- Imperialist Competitive Algorithm (ICA) has been
est neighbor, distance between yi and its ­kth nearest recently developed by (Atashpaz-Gargari and Lucas
neighbor and the distance between (xi, yi) and its ­kth 2007) that has inspired from a socio-political phe-
nearest neighbor, respectively. Also, n denotes the nomenon. One of the advantages of this algorithm is
number of samples, and d is the dimension of the its high ability to search the global optimization even
input space. in nonlinear optimization problems (Ahmadi et al.
2013). This algorithm is population-based algorithm
2.3 Imperialist Competitive Algorithm‑Artificial and, like other evolutionary algorithms, begins with
Neural Networks (ICA‑ANN) initial populations called countries which are in fact
the candidate solution of the problem. The countries
Artificial Neural Network (ANN) has been inspired are divided into two groups, imperialists and colo-
from the brain architecture and its parallel func- nies, which together form empires. In order to iden-
tionality and can identify the complicated relations tify the imperialists and their colonies, the cost of all
among variables. Indeed, it experiences the same as countries is calculated and some of the best countries
human and then generalizes results of these experi- are chosen as the imperialists and the rest form the
ences. Multilayer perceptron (MLP) neural network colonies of these imperialists based on their power
is one of the most frequently used feedforward net- (see Fig. 3).
works in estimation and prediction, which consists Afterward, these colonies start moving toward
of three layers of neurons with specific mathematical their relevant imperialist country. While moving
relationships. The input and output layers comprise toward the imperialist, some of these colonies may
the parameters affecting on the target variable(s) and reach to a better position than that of the imperial-
the parameter(s) supposed to be predicted, respec- ist. In this case, the imperialist exchanges its posi-
tively. The hidden layer(s) functions as the connec- tion with the colony and the algorithm continues with
tion between input and output layers and the number the imperialist in its new position. The competition
of neurons in hidden layers has an important effect among the empires forms the core of this algorithm,
on the performance of the network. During the train- which will gradually result in decrease in the power of
ing stage, the internal weights of the neural network weak empires and, finally, their collapse (see Fig. 4).
are changed and the network responses are compared The process continues until all weak empires collapse

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3197

Fig. 3  Generating the

initial empires (Atashpaz-
Gargari and Lucas 2007)

Fig. 4  Elimination of the

weakest empire (Atashpaz-
Gargari and Lucas 2007)

or the system meets certain termination criteria, e.g. 2016; Ahmadi et al. 2013; Moayedi and Armaghani
maximum number of decade (Armaghani et al. 2019). 2018). The MSE of the training set is considered to
More information about ICA technique could be be the cost function in this algorithm, while it would
found in (Atashpaz-Gargari and Lucas 2007). be better to take into account both training and test-
This optimization algorithm has been used to opti- ing results, because the network with the least train-
mize the weights and biases of Artificial Neural Net- ing error is not necessarily the best network. Hence,
works in different engineering fields (Aghajani et al. in order to select the best network among different

Vol.: (0123456789)
13
 3198 Geotech Geol Eng (2022) 40:3191–3209

networks, different criteria, such as MSE and coef- vi ≥ 0 i = 1, … , m

ficient of determination (­R2), for training and testing
datasets are usually used. In some cases, the obtained where Ek is the relative efficiency of a DMU “k”,
results of the criteria for different networks are very ur and vi are the weights given to output r and input
close to each other; therefore, it seems not easy to fig- i, respectively, yrk is the amount of output r from
ure out the best network (Armaghani et al. 2019). A DMU “k”, xik is the amount of input i from DMU
method of ranking could be a useful way to solve this “k”, yrj is the amount of output r from DMU “j”,
problem. In this study, Data Envelopment Analysis- xij is the amount of input i from DMU “j” and N, n
Analytic Hierarchy Process (DEA-AHP) approach and m are the number of DMUs, outputs and inputs,
is proposed to incorporate in ICA-ANN method to respectively.
assess the countries based on their efficiency instead However, DEA is not a complete ranking method,
of cost. Moreover, it could facilitate finding the best it merely provides the classification of the DMUs into
network among different architectures. two groups: efficient and inefficient (Esmaeilzadeh
and Hadi-Vencheh 2015; Jahanshahloo and Afzaline-
2.4 Data Envelopment Analysis‑Analytic Hierarchy jad 2006; Kritikos 2017; Sinuany‐Stern et al. 2000).
Process (DEA‑AHP) Nevertheless, several approaches have been proposed
to fully rank DMUs in the context of DEA (Cook and
Data envelopment analysis (DEA) is one of the prom- Kress 1990; Cook et al. 1992; Friedman and Sinu-
inent nonparametric techniques to measure and evalu- any-Stern 1998; Green et al. 1996; Sinuany‐Stern
ate the performance of a set of comparable entities, et al. 2000). One of these approaches is to utilize one
called decision-making units (DMUs), in terms of of the more popular Multi-Criteria Decision Mak-
their relative efficiencies. These DMUs use multiple ing (MCDM) methods, Analytic Hierarchy Process
inputs to produce multiple outputs (Emrouznejad and (AHP). AHP is a subjective method originally pro-
Yang 2018). DEA, as a linear programming-based posed by Saaty (1980) for assessing qualitative crite-
method, was originally presented by (Charnes et al. ria to assign weights to the operating units. AHP uti-
1978). They introduced CCR model to produce the lizes pairwise comparisons to establish priorities for
efficiency frontier based on the concept of Pareto both criteria and options based on the experts’ opin-
optimum. DEA has been widely applied in a broad ions (Lin et al. 2011).
variety of fields such as banking, transportation, It is worth noting that DEA-AHP approach not
energy, communication, mining, agriculture, finance, only eliminates the difficulties of extraction of
health care, education, sport and insurance (Liu et al. weights based on pairwise comparison between
2013). DEA does not require any particular func- DMUs and decision maker’s preferences in AHP
tional form to relate inputs to outputs when determin- but also avoids overestimating the DMUs’ efficiency
ing efficiency, but determines the relative efficiency by only self-assessment in DEA. This model does
by using the data of each DMU (Budeba et al. 2015). not require the experts’ opinions (Gholipour et al.
The efficiency ­Ek of the kth DMU is calculated by 2014). Due to the advantages of this approach, sev-
the following CCR model: eral researchers have applied these two techniques
∑n together to rank or evaluate DMUs (An et al. 2018;
MaxEk = uy
r=1 r rk
(7) Giokas and Pentzaropoulos 2008; Mohammadi et al.
2017; Rezaei et al. 2020; Tseng and Lee 2009). DEA-
S.t. AHP method used in this study consists of two steps.
∑m In the first step, DEA is utilized to create the relative
vi xik = 1
i=1 efficiency of each pair of DMUs. For any pair of units
A and B, the following DEA runs without consider-
≤0
∑n ∑m
ur yrj − vx j = 1, … , N ing the other DMUs.
r=1 i=1 i ij
∑n
(8)
ur ≥ 0
MaxEAA = uy
r=1 r rA
r = 1, … , n
S.t.

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3199

∑m
vx =1 Ejj + Ejk
i=1 i iA ajk = andajj = 1 (10)
Ekk + Ekj

≤1
∑n
uy
r=1 r rA
In the second step, the pairwise comparison matrix
calculated in the previous step is utilized for prioritiz-

≤0
∑n ∑m ing the DMUs.
uy − vx
r=1 r rB i=1 i iB

ur ≥ 0 r = 1, … , n 3 Results and Discussion

vi ≥ 0 i = 1, … , m In this study, the dependency between the penetration

rate of TBM and rock mass and machine operational
∑n parameters are evaluated using the MI variable selec-
MaxEBA = uy
r=1 r rB
(9) tion approach. The values of MI between three groups
of parameters, intact rock properties (i.e., UCS, Elas-
S.t. tic Modulus, Porosity, Density, BTS, Quartz content),
∑m rock mass properties (i.e., DPW, RQD, RMR, GSI, Q)
vi xiB = 1 and TBM operational parameters (i.e., Thrust, RPM,
i=1
Torque, Power), and the PR are shown in Fig. 5. It
≤1
∑n can be seen that, in overall, rock mass properties have
uy
r=1 r rB stronger correlations with the PR in comparison with
∑n ∑m the intact rock properties. Moreover, the PR has the
ur yrA − EAA vx =0 highest dependency on the RPM which is one of the
r=1 i=1 i iA
important parameters that is controlled by the opera-
ur ≥ 𝜀
tor, which directly affects the rate of penetration (Jing
r = 1, … , n
et al. 2019). The thrust is another machine operational

vi ≥ 𝜀
parameter that has a fairly good correlation with the
i = 1, … , m PR. These two variables can be found as the effective
parameters on the PR in the literature (Armaghani
EBB and EAB are calculated by the same method.
et al. 2019; Eftekhari et al. 2010; Grima et al. 2000;
Finally, the relative efficiency ratio for every pair of
Yagiz and Karahan 2011). In addition, all examined
units j and k needed for AHP is calculated as follows:
rock mass properties indicate a significant impact on

Fig. 5  Mutual informa- 0.8

tion between different
0.7 TBM operational parameters
intact rock, rock mass and
machine operating param- intact rock properes
eters and the penetration 0.6 rock mass properties
rate
0.5
MI (bit)s

0.4

0.3

0.2

0.1

Variables

Vol.: (0123456789)
13
 3200 Geotech Geol Eng (2022) 40:3191–3209

the PR which is consistent with the previous studies Table 4  The multicollinearity diagnosis statistics for the vari-
(Benardos and Kaliampakos 2004; Eftekhari et al. ables
2010; Salimi et al. 2016), while only the UCS and Variable Collinearity statistics
BTS can be regarded as the intact rock variables cor-
Before elimination After elimination
related with the PR (Eftekhari et al. 2010; Rostami
and Ozdemir 1993; Salimi and Esmaeili 2013; Salimi Tolerance VIF Tolerance VIF
et al. 2016). Therefore, these nine variables were cho- Thrust 0.688 1.453 0.779 1.283
sen as the inputs for the ICA-ANN employed in this RPM 0.678 1.474 0.685 1.459
study. UCS 0.252 3.974 0.271 3.684
Since there are likely high level of correlations BTS 0.269 3.717 0.305 3.278
between intact rock properties as well as rock mass DPW 0.338 2.960 0.394 2.535
properties, it may lead to multicollinearity problems. RQD 0.035 28.173 0.529 1.890
Multicollinearity occurs when two or more than two RMR 0.028 35.863 – –
predictor variables are inter-correlated (Xiong et al. GSI 0.119 8.437 – –
2019). Although some scholars have argued that Q 0.093 10.746 – –
multicollinearity does not affect the ANN’s perfor-
The excluded variable is shown in bold
mance as it does on the performance of linear regres-
sion because they integrate at least one hidden layer
between independent and dependent variables (Bejou the collinearity between variables (Liao and Valliant
et al. 1996; Moosmayer et al. 2013), some others 2012). As shown in Table 4, the values of VIF and
have proposed that when two predictors are signifi- tolerances for the rock mass classification systems are
cantly correlated, the one that has lower correlation more than 5 and less than 0.2, respectively. Since the
with the dependent variable would be excluded from RMR, GSI and Q have a strong correlation with the
the inputs before training the network (Ling and Liu RQD and also a lower correlation than the RQD with
2004; Zhang et al. 2019). Thus, first, Pearson’s cor- the PR, these variables were eliminated from the vari-
relation coefficients are analyzed to evaluate the cor- ables and the new values of VIF and tolerances have
relation between the independent variables (Table 3). also been shown in Table 4.
It is found that there is a strong inter-correlation Therefore, in this study, two types of networks
among the rock mass classification systems (i.e., were examined: (1) Type 1_ the networks with nine
RQD, RMR, GSI and Q), and also between the intact parameters (i.e., Thrust, RPM, UCS, BTS, DPW,
rock parameters. Then, the variance inflation factors RQD, RMR, GSI, Q) in the input layer, and (2) Type
(VIF) and tolerances are utilized to check for mul- 2_ the networks with six parameters (i.e., Thrust,
ticollinearity among predictor variables (Table 4). RPM, UCS, BTS, DPW, RQD) as the inputs.
Critical values of VIF > 5 and tolerance < 0.2 indicate

Table 3  Correlation matrix

Thrust RPM UCS BTS DPW RQD RMR GSI Q
for the variables
Thrust 1

RPM 0.331 1

UCS 0.037 0.015 1

BTS 0.029 0.069 0.814 1

DPW -0.248 -0.372 0.337 0.12 1

RQD 0.165 0.114 0.112 -0.049 0.53 1

RMR 0.239 0.142 0.2 0.065 0.48 0.97 1

GSI 0.16 0.162 0.246 0.143 0.402 0.891 0.919 1

Q 0.184 0.104 0.31 0.133 0.532 0.899 0.934 0.9 1

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3201

In order to predict the penetration rate of TBM, the both MSE and determination coefficient, so that the
structure of the network should be optimized. Deter- network with lower MSE or higher ­R2 has the more
mination of a network structure includes the selection intense color.
of the number of hidden layers as well as the num- According to this Table, the obtained results of
ber of neurons per layer. Many researchers have men- all the performance indicators are very close to each
tioned that the constructed ANN with a hidden layer other; therefore, it seems not easy to choose the best
can approximate any complicated nonlinear function model. Hence, in addition to optimizing the weights
with adequate accuracy (e.g. Ashtiani et al. 2020; and biases of the networks with different architecture
Balestrassi et al. 2009; Hornik et al. 1989; Khandel- using the aforementioned method, the DEA-AHP
wal and Armaghani 2016). Thus, in this study, all net- method was also used to choose the best network
works were created using one hidden layer. among the networks presented in Table 5. The results
When it comes to the number of hidden neurons, of evaluating and ranking different networks have
many investigators have stated that there is no heu- also been summarized in Table 5. As can be seen in
ristic method for determining the number of hidden this table, the architecture of the best network con-
neurons (Chong 2013a, b; Sexton et al. 2002; Svozil sists of 6 input neurons, 9 neurons in the hidden layer
et al. 1997). Thus, like many research studies (Arma- and one output neuron. Figure 6 shows the flow dia-
ghani et al. 2019; Basheer and Hajmeer 2000; Haykin gram of the study from input selection to the best net-
1994; Kurt et al. 2009), the optimum number of hid- work architecture. The determination coefficient (­R2)
den neurons was determined using a trial-and-error and the mean square error (MSE) indices of the train-
procedure. Various networks with different number of ing and testing datasets are 0.901, 0.887, 0.04 and
neurons in the hidden layer, with ranging from 1 to 0.0645, respectively. The best relationships between
20, were constructed and trained with the ICA-ANN. the actual and predicted penetration rate of TBM
In the ICA optimization, three factors (i.e., number of using the ICA-ANN model developed in this study is
countries, number of imperialists and number of dec- illustrated in Fig. 7.
ades) are the most effective parameters on its perfor- In addition, although there is no significant differ-
mance. They can be determined through a trial-and- ence between the results of the networks with 6 and
error procedure (Tashayo et al. 2019). For selecting 9 input variables, the values of R ­ 2 and MSE of net-
the best values of these factors, the different mod- works with 6 and 9 inputs with different number of
els were examined. Based on obtained results, the neurons in hidden layer are shown for training and
number of countries, imperialists and decades were testing datasets in Figs. 8 and 9, respectively. As
selected equal to 300, 20 and 900, respectively, for shown in Fig. 8, in the training data, networks with 9
predicting the penetration rate of TBM. inputs have a higher coefficient of determination and
In this study, the DEA-AHP method is used to rank lower MSE than networks with 6 inputs, while in the
and compare countries in the ICA-ANN method for testing set (Fig. 9), networks with 6 inputs have bet-
each architecture of the networks. For this purpose, ter performance indicators. It can be concluded that
the MSE and the coefficient of determination of both increasing the number of inputs leads to over-training
the training and testing datasets are used for evaluat- and causes the network to lose its power of generali-
ing and ranking of countries instead of merely com- zation. Therefore, multicollinearity could affect the
paring the MSE of the training dataset. Each country performance of the neural networks to some extent,
is taken into account as a DMU that has several inputs as reported in some research studies (Ling and Liu
and outputs. The MSE of the training and testing 2004; Zhang et al. 2019).
datasets are considered as the inputs and the determi-
nation coefficients of these datasets are considered as
the outputs for each DMU. 4 Conclusions
The obtained results of the MSE and the coeffi-
cient of determination of both the training and testing This paper aims to predict the penetration rate of
datasets for each number of neurons in hidden layer TBM using the hybrid of ICA-ANN and DEA-AHP
for both types of the network are given in Table 5. approaches. The DEA-AHP method has been used
In this table, a color intensity rating was used for to prioritize the countries created in the ICA-ANN

Vol.: (0123456789)
13
 3202 Geotech Geol Eng (2022) 40:3191–3209

Table 5  The results of performance indices obtained by the ICA-ANN

MSE R2
Model Number of Number of DEA-AHP
No. Inputs hidden Neurons Rank
Train Test Train Test
1 6 1 0.0501 0.076 0.81 0.79 39
2 6 2 0.0475 0.07 0.854 0.822 35
3 6 3 0.0438 0.0676 0.879 0.852 31
4 6 4 0.0415 0.0685 0.87 0.873 22
5 6 5 0.0417 0.0682 0.871 0.871 25
6 6 6 0.0434 0.0666 0.887 0.856 29
7 6 7 0.0411 0.0649 0.902 0.877 5
8 6 8 0.0413 0.066 0.89 0.875 11
9 6 9 0.04 0.0645 0.901 0.887 1
10 6 10 0.0407 0.0633 0.903 0.882 2
11 6 11 0.0409 0.0663 0.887 0.879 7
12 6 12 0.0411 0.0663 0.889 0.877 9
13 6 13 0.0415 0.0658 0.888 0.874 12
14 6 14 0.0425 0.0653 0.897 0.864 13
15 6 15 0.0409 0.0653 0.898 0.879 4
16 6 16 0.0414 0.0667 0.88 0.876 14
17 6 17 0.0414 0.0643 0.9 0.874 6
18 6 18 0.0406 0.065 0.894 0.882 3
19 6 19 0.0447 0.0736 0.825 0.847 36
20 6 20 0.0437 0.067 0.878 0.856 30
21 9 1 0.0487 0.0818 0.802 0.768 40
22 9 2 0.0442 0.0814 0.846 0.772 37
23 9 3 0.0404 0.0785 0.882 0.798 32
24 9 4 0.0388 0.0801 0.895 0.785 33
25 9 5 0.0373 0.0772 0.908 0.809 24
26 9 6 0.0371 0.0746 0.91 0.829 10
27 9 7 0.0371 0.0763 0.911 0.821 16
28 9 8 0.0373 0.0739 0.908 0.835 8
29 9 9 0.0372 0.0776 0.909 0.806 26
30 9 10 0.0355 0.0768 0.923 0.812 17
31 9 11 0.0359 0.0768 0.92 0.813 18
32 9 12 0.0373 0.0778 0.908 0.806 27
33 9 13 0.037 0.0773 0.911 0.808 23
34 9 14 0.0377 0.0763 0.905 0.818 21
35 9 15 0.0354 0.0792 0.925 0.792 28
36 9 16 0.041 0.0811 0.881 0.786 34
37 9 17 0.0381 0.0746 0.902 0.833 15
38 9 18 0.0382 0.0759 0.905 0.826 19
39 9 19 0.0448 0.0843 0.856 0.763 38
40 9 20 0.0385 0.0755 0.9 0.827 20

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3203

Thrust
TBM operational parameters

RPM

Torque

Power Thrust n1

UCS RPM n2

Thrust

E UCS n3
intact rock properties

RPM

q BTS n4
Mutual
UCS
Information
Multicollinearity
Density analysis DPW n5 PR
analysis
(MI) BTS

n RQD n6

DPW

BTS RMR n7

RQD

DPW GSI n8
rock mass properties

RQD Q n9

RMR Input Layer Hidden Layer Output Layer

GSI

Fig. 6  Flow diagram of the study from input selection to the best network architecture

method as well as to select the network with the best (thrust, RPM) parameters, were selected as the first
performance among the networks with different hid- set of input parameters for the network. Among the
den neurons. In order to identify the most effective selected variables, the PR has the highest dependency
variables on the penetration rate of TBM, the MI on the RPM which is controlled by the operator, and
method was employed and 9 parameters, including directly affects the rate of penetration. In addition, in
the intact rock (UCS and BTS), rock mass (DPW, overall, rock mass properties have stronger correla-
RQD, RMR, Q and GSI) and machine operating tions with the PR in comparison with the intact rock

Vol.: (0123456789)
13
 3204 Geotech Geol Eng (2022) 40:3191–3209

4 properties. Three of these parameters, i.e., RMR, Q

R² = 0.901 and GSI, were excluded from the inputs to eliminate
3.5 the effect of multicollinearity among variables. The
remaining parameters were considered as the second
set of the input parameters for the network. Various
3
networks with different number of neurons in the
Predicted PR (m/h)

hidden layer, with ranging from 1 to 20, were con-

2.5 structed and trained with the ICA-ANN. The MSE
and the coefficient of determination of both the train-
2 ing and testing datasets were used for evaluating and
ranking of countries instead of merely comparing the
1.5
MSE of the training dataset. Each country was con-
sidered as a DMU with two inputs (i.e., the MSE of
the training and testing datasets) and two outputs (i.e.,
1
1 1.5 2 2.5 3 3.5 4 the determination coefficients of the training and test-
Measured PR (m/h) ing datasets). The results of the networks with two
(a) Training dataset types of input parameters showed that in spite of the
fact that the results of all the performance indicators
4 are very close to each other, in the training data, net-
R² = 0.887 works with 9 inputs have a higher coefficient of deter-
3.5 mination and lower MSE than networks with 6 inputs,
while in the testing set, networks with 6 inputs have
better performance indicators. Therefore, increasing
3
the number of inputs leads to over-training and causes
Predicted PR (m/h)

the network to lose its power of generalization, and

2.5 also eliminating the effect of multicollinearity could
slightly improve the performance of the network.
2 The values of ­R2 and MSE of the training and test-
ing datasets for the best architecture (i.e., 6 × 9 × 1)
1.5
are 0.901, 0.887, 0.04 and 0.0645, respectively. The
results demonstrate that the developed methodology
can successfully be used for predicting the penetra-
1
1 1.5 2 2.5 3 3.5 4 tion rate of TBM.
Measured PR (m/h)
(b) Testing dataset

Fig. 7  The best results of ICA-ANN model for prediction of

TBM Penetration Rate a training dataset b testing dataset

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3205

Fig. 8  Performance indica- 6 Inputs 9 Inputs

tors of networks with 6 1
and 9 inputs with different
number of hidden neurons 0.9
for training dataset a ­R2 b
MSE
0.8

0.7

R2
0.6

0.5

0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of hidden neurons
(a) R2

6 Inputs 9 Inputs
0.06

0.055

0.05

0.045
MSE

0.04

0.035

0.03

0.025

0.02
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of hidden neurons
(b) MSE

Vol.: (0123456789)
13
 3206 Geotech Geol Eng (2022) 40:3191–3209

Fig. 9  Performance indica- 6 Inputs 9 Inputs

tors of networks with 6 1
and 9 inputs with different
number of hidden neurons 0.9
for testing dataset a ­R2 b
MSE
0.8

0.7

R2
0.6

0.5

0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of hidden neurons
(a) R2

6 Inputs 9 Inputs
0.09

0.08

0.07

0.06
MSE

0.05

0.04

0.03

0.02
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of hidden neurons
(b) MSE

Funding The authors declare that no funds, grants, or References

other support were received during the preparation of this
manuscript. Afradi A, Ebrahimabadi A (2020) Comparison of artificial neu-
ral networks (ANN), support vector machine (SVM) and
Data availability Enquiries about data availability should be gene expression programming (GEP) approaches for pre-
directed to the authors. dicting TBM penetration rate. SN Appl Sci 2(12):1–16
Afradi A, Ebrahimabadi A (2021) Prediction of TBM pen-
Declarations etration rate using the imperialist competitive algorithm
(ICA) and quantum fuzzy logic. Innovative Infrastructure
Conflict of interest The authors declare that they have no Solutions 6(2):1–17
known competing financial interests or personal relationships Afradi A, Ebrahimabadi A, Hallajian T (2021) Prediction of
that could have appeared to influence the work reported in this TBM penetration rate using fuzzy logic, particle swarm
paper. optimization and harmony search algorithm. Geotech
Geol Eng. https://​doi.​org/​10.​1007/​s10706-​021-​01982-x
Aghajani A, Kazemzadeh R, Ebrahimi A (2016) A novel
hybrid approach for predicting wind farm power produc-
tion based on wavelet transform, hybrid neural networks

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3207

and imperialist competitive algorithm. Energy Convers Charnes A, Cooper WW, Rhodes E (1978) Measuring the
Manage 121:232–240 efficiency of decision making units. Eur J Oper Res
Ahmadi MA, Ebadi M, Shokrollahi A, Majidi SMJ (2013) 2(6):429–444
Evolving artificial neural network and imperialist com- Chong AY-L (2013a) Predicting m-commerce adoption deter-
petitive algorithm for prediction oil flow rate of the reser- minants: a neural network approach. Expert Syst Appl
voir. Appl Soft Comput 13(2):1085–1098 40(2):523–530
An Q, Meng F, Xiong B (2018) Interval cross efficiency for Chong AY-L (2013b) A two-staged SEM-neural network
fully ranking decision making units using DEA/AHP approach for understanding and predicting the deter-
approach. Ann Oper Res 271(2):297–317 minants of m-commerce adoption. Expert Syst Appl
Armaghani DJ, Mohamad ET, Narayanasamy MS, Narita N, 40(4):1240–1247
Yagiz S (2017) Development of hybrid intelligent models Cook WD, Kress M (1990) A data envelopment model
for predicting TBM penetration rate in hard rock condi- for aggregating preference rankings. Manage Sci
tion. Tunn Undergr Space Technol 63:29–43 36(11):1302–1310
Armaghani DJ, Koopialipoor M, Marto A, Yagiz S (2019) Cook WD, Kress M, Seiford LM (1992) Prioritization models
Application of several optimization techniques for esti- for frontier decision making units in DEA. Eur J Oper
mating TBM advance rate in granitic rocks. J Rock Mech Res 59(2):319–323
Geotech Eng 11(4):779–789 Eftekhari M, Baghbanan A, Bagherpour R (2014) The effect of
Armetti G, Migliazza MR, Ferrari F, Berti A, Padovese P fracture patterns on penetration rate of TBM in fractured
(2018) Geological and mechanical rock mass condi- rock mass using probabilistic numerical approach. Arab J
tions for TBM performance prediction. The case of “La Geosci 7(12):5321–5331
Maddalena” exploratory tunnel, Chiomonte (Italy). Tunn Eftekhari M, Baghbanan A, Bayati M (2010) Predicting pen-
Undergr Space Technol 77:115–126 etration rate of a tunnel boring machine using artificial
Ashtiani S-HM, Rohani A, Aghkhani MH (2020) Soft comput- neural network. In: ISRM International Symposium-6th
ing-based method for estimation of almond kernel mass Asian Rock Mechanics Symposium. International Soci-
from its shell features. Sci Hortic 262:109071 ety for Rock Mechanics and Rock Engineering
Atashpaz-Gargari E, Lucas C (2007) Imperialist competitive Emrouznejad A, Yang G-l (2018) A survey and analysis of the
algorithm: an algorithm for optimization inspired by first 40 years of scholarly literature in DEA: 1978–2016.
imperialistic competition. In: 2007 IEEE congress on Socioecon Plann Sci 61:4–8
evolutionary computation. IEEE, p 4661–4667 Engineers SC (2010) Report of geological study and engineer-
Balestrassi PP, Popova E, Paiva Ad, Lima JM (2009) Design ing geotechnical of Nowsoud (Zagros) tunnel. Unpub-
of experiments on neural network’s training for non- lished report
linear time series forecasting. Neurocomputing Esmaeilzadeh A, Hadi-Vencheh A (2015) A new method
72(4–6):1160–1178 for complete ranking of DMUs. Optimization
Barton NR (2000) TBM tunnelling in jointed and faulted rock. 64(5):1177–1193
CRC Press, Boca Raton Farmer I, Glossop N (1980) Mechanics of disc cutter penetra-
Basheer IA, Hajmeer M (2000) Artificial neural networks: fun- tion. Tunnels and Tunnelling 12(6):22–25
damentals, computing, design, and application. J Micro- Fraser AM, Swinney HL (1986) Independent coordinates for
biol Methods 43(1):3–31 strange attractors from mutual information. Phys Rev A
Bejou D, Wray B, Ingram TN (1996) Determinants of relation- 33(2):1134
ship quality: an artificial neural network analysis. J Bus Frénay B, Doquire G, Verleysen M (2013) Is mutual informa-
Res 36(2):137–143 tion adequate for feature selection in regression? Neural
Benardos A, Kaliampakos D (2004) Modelling TBM perfor- Netw 48:1–7
mance with artificial neural networks. Tunn Undergr Friedman L, Sinuany-Stern Z (1998) Combining ranking scales
Space Technol 19(6):597–605 and selecting variables in the DEA context: the case of
Bilgin N, Tuncdemir H, Balci C, Copur H, Eskikaya S (2000) A industrial branches. Comput Oper Res 25(9):781–791
model to predict the performance of tunneling machines Ghaleini EN, Koopialipoor M, Momenzadeh M, Sarafraz
under stressed conditions. In: Proceedings of the world ME, Mohamad ET, Gordan B (2019) A combination
tunnel congress of AITES-ITA, Durban, pp 47–53 of artificial bee colony and neural network for approxi-
Brino G, Peila D, Steidl A, Fasching F (2015) Prediction of mating the safety factor of retaining walls. Eng Comput
performance and cutter wear in rock TBM: application to 35(2):647–658
Koralm tunnel project. In: Proceedings of Geoingegneria Gholipour R, Jandaghi G, Rajaei R (2014) Contractor selection
Ambientale e Mineraria, Turin, pp 41–54 in MCDM context using fuzzy AHP. Iran J Manage Stud
Bruland A (1998) Hard rock tunnel boring: vol 1–10. Ph. D. 7(1):151–173
Thesis, Norwegian University of Science and Technol- Giokas DI, Pentzaropoulos G (2008) Efficiency ranking of
ogy (NTNU) the OECD member states in the area of telecommunica-
Budeba M, Joubert JW, Webber-Youngman R (2015) A pro- tions: a composite AHP/DEA study. Telecommun Policy
posed approach for modelling competitiveness of 32(9–10):672–685
new surface coal mines. J South Afr Inst Min Metall Gong Q-M, Zhao J, Jiao Y-Y (2005) Numerical modeling
115(11):1057–1064 of the effects of joint orientation on rock fragmenta-
tion by TBM cutters. Tunn Undergr Space Technol
20(2):183–191

Vol.: (0123456789)
13
 3208 Geotech Geol Eng (2022) 40:3191–3209

Gong Q-M, Zhao J, Hefny A (2006a) Numerical simulation of by TBM cutting tools. Tunn Undergr Space Technol
rock fragmentation process induced by two TBM cutters 81:428–437
and cutter spacing optimization. Tunnelling and Under- Liao D, Valliant R (2012) Variance inflation factors in the anal-
ground Space Technology 21(3–4) ysis of complex survey data. Surv Methodol 38(1):53–62
Gong Q, Jiao Y, Zhao J (2006b) Numerical modelling of the Lin M-I, Lee Y-D, Ho T-N (2011) Applying integrated DEA/
effects of joint spacing on rock fragmentation by TBM AHP to evaluate the economic performance of local gov-
cutters. Tunn Undergr Space Technol 21(1):46–55 ernments in China. Eur J Oper Res 209(2):129–140
Gordan B, Koopialipoor M, Clementking A, Tootoonchi H, Ling FYY, Liu M (2004) Using neural network to predict per-
Mohamad ET (2019) Estimating and optimizing safety formance of design-build projects in Singapore. Build
factors of retaining wall through neural network and bee Environ 39(10):1263–1274
colony techniques. Eng Comput 35(3):945–954 Liu H, Kou S, Lindqvist P-A, Tang C (2002) Numerical sim-
Graham PC (1976) Rock exploration for machine manufactur- ulation of the rock fragmentation process induced by
ers. In: Proceedings of Exploration for Rock Engineer- indenters. Int J Rock Mech Min Sci 39(4):491–505
ing, Johannesburg, pp 173–180 Liu JS, Lu LY, Lu W-M, Lin BJ (2013) A survey of DEA
Green RH, Doyle JR, Cook WD (1996) Preference voting and applications. Omega 41(5):893–902
project ranking using DEA and cross-evaluation. Eur J Marji MF (1997) Modeling of cracks in rock fragmentation
Oper Res 90(3):461–472 with a higher order displacement discontinuity method.
Grima MA, Bruines P, Verhoef P (2000) Modeling tunnel bor- Dissertation, Middle East Technical University, Ankara,
ing machine performance by neuro-fuzzy methods. Tunn Turkey
Undergr Space Technol 15(3):259–269 Marji MF (2015) Simulation of crack coalescence mechanism
Haeri H, Marji MF (2016) Simulating the crack propagation underneath single and double disc cutters by higher order
and cracks coalescence underneath TBM disc cutters. displacement discontinuity method. J Cent South Univ
Arab J Geosci 9(2):124 22(3):1045–1054
Haykin S (1994) Neural networks: a comprehensive founda- Marji M, Hosseini Nasab H, Hossein Morshedi A (2009)
tion. MacMillan College Publishing Co., New York Numerical modeling of crack propagation in rocks under
Hornik K, Stinchcombe M, White H (1989) Multilayer feedfor- TBM disc cutters. J Mech Mater Struct 4(3):605–627
ward networks are universal approximators. Neural Netw Moayedi H, Armaghani DJ (2018) Optimizing an ANN model
2(5):359–366 with ICA for estimating bearing capacity of driven pile in
Howarth D, Roxborough F (1982) Some fundamental aspects cohesionless soil. Eng Comput 34(2):347–356
of the use of disc cutters in hard-rock excavation. J South Mohamad ET, Faradonbeh RS, Armaghani DJ, Monjezi M,
Afr Inst Min Metall 82(11):309–315 Abd Majid MZ (2017) An optimized ANN model based
Huoh Y-J (2013) Sensitivity analysis of stochastic simulators on genetic algorithm for predicting ripping production.
with information theory. UC Berkeley Neural Comput Appl 28(1):393–406
Innaurato N, Mancini A, Rondena E, Zaninetti A (1991) Fore- Mohammadi A, Sadeghi M, Shojaei P, Rezaei A (2017)
casting and effective TBM performances in a rapid exca- Evaluating the efficiency of Iran’s provincial tax offices
vation of a tunnel in Italy. In: Proceeding of the 7th Inter- and ranking them by DEA/AHP. Iran J Manage Stud
national Congress ISRM, Aachen, pp 1009–1014 10(3):729–750
Jahanshahloo G, Afzalinejad M (2006) A ranking method Moon T, Oh J (2012) A study of optimal rock-cutting condi-
based on a full-inefficient frontier. Appl Math Model tions for hard rock TBM using the discrete element
30(3):248–260 method. Rock Mech Rock Eng 45(5):837–849
Jing L-j, Li J-b, Yang C, Chen S, Zhang N, Peng X-x (2019) Moosmayer DC, Chong AY-L, Liu MJ, Schuppar B (2013) A
A case study of TBM performance prediction using field neural network approach to predicting price negotiation
tunnelling tests in limestone strata. Tunn Undergr Space outcomes in business-to-business contexts. Expert Syst
Technol 83:364–372 Appl 40(8):3028–3035
Khandelwal M, Armaghani DJ (2016) Prediction of drillabil- Ozdemir L, Miller R, Wang F (1978) Mechanical tunnel boring
ity of rocks with strength properties using a hybrid GA- prediction and machine design, Report to National Sci-
ANN technique. Geotech Geol Eng 34(2):605–620 ence Foundation, Colorado School of Mines, Golden
Kraskov A, Stögbauer H, Grassberger P (2004) Estimating Park C, Park C, Synn J, Sunwoo C, Chung S (2001) TBM pen-
mutual information. Phys Rev E 69(6):066138 etration rate with rock mass properties in hard rock. In:
Kritikos MN (2017) A full ranking methodology in data envel- Proceedings of the AITES-ITA 2001 World Tunnel Con-
opment analysis based on a set of dummy decision mak- gress, Milan, pp 413–419
ing units. Expert Syst Appl 77:211–225 Qi G, Zhengying W, Hao M, Qiao C (2016) Numerical and
Kurt MB, Sezgin N, Akin M, Kirbas G, Bayram M (2009) experimental research on the rock-breaking process of
The ANN-based computing of drowsy level. Expert Syst tunnel boring machine normal disc cutters. J Mech Sci
Appl 36(2):2534–2542 Technol 30(4):1733–1745
Lee Y, Oh S-H, Kim MW (1991) The effect of initial weights Rezaei A, Hassani H, Tziritis E, Mousavi SBF, Jabbari N
on premature saturation in back-propagation learning. In: (2020) Hydrochemical characterization and evaluation of
IJCNN-91-Seattle international joint conference on neu- groundwater quality in Dalgan basin, SE Iran. Ground-
ral networks. vol 1. IEEE, p 765–770 water Sustain Develop. https://​doi.​org/​10.​1016/j.​gsd.​
Li G, Wang W, Jing Z, Zuo L, Wang F, Wei Z (2018) Mecha- 2020.​100353
nism and numerical analysis of cutting rock and soil

Vol:. (1234567890)
13
Geotech Geol Eng (2022) 40:3191–3209 3209

Ribacchi R, Fazio AL (2005) Influence of rock mass parame- rock mechanics and mining sciences & geomechanics
ters on the performance of a TBM in a gneissic formation abstracts, vol 19. Elsevier, Amsterdam, pp 107–121
(Varzo Tunnel). Rock Mech Rock Eng 38(2):105–127 Steuer R, Kurths J, Daub CO, Weise J, Selbig J (2002)
Rossi F, Lendasse A, François D, Wertz V, Verleysen M (2006) The mutual information: detecting and evaluat-
Mutual information for the selection of relevant variables ing dependencies between variables. Bioinformatics
in spectrometric nonlinear modelling. Chemom Intell 18(suppl_2):S231–S240
Lab Syst 80(2):215–226 Sundaram N, Rafek A, Komoo I (1998) The influence of rock
Rostami J (1997) Development of a force estimation model for mass properties in the assessment of TBM performance.
rock fragmentation with disc cutters through theoretical In: Engineering geology: a global view from the Pacific
modeling and physical measurement of crushed zone Rim. p 3553–3559
pressure. Dissertation, Colorado School of Mines, Colo- Sutha K, Tamilselvi JJ (2015) A review of feature selection
rado, USA algorithms for data mining techniques. Int J Comput Sci
Rostami J (2013) Study of pressure distribution within the Eng 7(6):63
crushed zone in the contact area between rock and disc Svozil D, Kvasnicka V, Pospichal J (1997) Introduction to
cutters. Int J Rock Mech Min Sci 57:172–186 multi-layer feed-forward neural networks. Chemom Intell
Rostami J, Ozdemir L (1993) A new model for performance Lab Syst 39(1):43–62
prediction of hard rock TBMs. In: Proceedings of rapid Tarkoy PJ (1974) Predicting tunnel boring machine (TBM)
excavation and tunneling conference, Boston, Massachu- penetration rates and cutter costs in selected rock types.
setts, pp 793–809 In: Proceedings of 9th Canadian Rock Mechanics Sym-
Roxborough FF, Phillips HR (1975) Rock excavation by disc posium, Montreal, pp 263–274
cutter. International journal of rock mechanics and min- Tashayo B, Behzadafshar K, Tehrani MS, Banayem HA,
ing sciences & geomechanics abstracts, vol 12. Elsevier, Hashemi MH, Nezhad SST (2019) Feasibility of impe-
Amsterdam, pp 361–366 rialist competitive algorithm to predict the surface settle-
Saaty T (1980) The Analytic Hierarchy Process Mcgraw Hill, ment induced by tunneling. Eng Comput 35(3):917–923
New York. Agricultural Economics Review 70 Tseng Y-F, Lee T-Z (2009) Comparing appropriate decision
Salimi A, Esmaeili M (2013) Utilising of linear and non-linear support of human resource practices on organizational
prediction tools for evaluation of penetration rate of tun- performance with DEA/AHP model. Expert Syst Appl
nel boring machine in hard rock condition. Int J Mining 36(3):6548–6558
Mineral Eng 4(3):249–264 Verma AK, Singh TN (2011) Intelligent systems for ground
Salimi A, Rostami J, Moormann C, Delisio A (2016) Applica- vibration measurement: a comparative study. Eng Com-
tion of non-linear regression analysis and artificial intel- put 27(3):225–233
ligence algorithms for performance prediction of hard Wang X, Tang Z, Tamura H, Ishii M, Sun W (2004) An
rock TBMs. Tunn Undergr Space Technol 58:236–246 improved backpropagation algorithm to avoid the local
Sanio H (1985) Prediction of the performance of disc cutters in minima problem. Neurocomputing 56:455–460
anisotropic rock. International journal of rock mechanics Wang W, Ying L, Zhang J (2016) On the relation between
and mining sciences & geomechanics abstracts, vol 22. identifiability, differential privacy, and mutual-informa-
Elsevier, Amsterdam, pp 153–161 tion privacy. IEEE Trans Inf Theory 62(9):5018–5029
Sapigni M, Berti M, Bethaz E, Busillo A, Cardone G (2002) Xiong B, Newton S, Li V, Skitmore M, Xia B (2019) Hybrid
TBM performance estimation using rock mass classifica- approach to reducing estimating overfitting and collinear-
tions. Int J Rock Mech Min Sci 39(6):771–788 ity. Eng Constr Archit Manage. https://​doi.​org/​10.​1108/​
Sato K (1991) Prediction of disc cutter performance using a ECAM-​08-​2018-​0353
circular rock cutting rig. In: Proceedings of the First Yagiz S, Karahan H (2011) Prediction of hard rock TBM pene-
International symposium on Mine Mechanization, vol 1. tration rate using particle swarm optimization. Int J Rock
Golden, Colorado, pp 31–40 Mech Min Sci 48(3):427–433
Sexton RS, Johnson RA, Hignite MA (2002) Predicting Yang H, Wang H, Zhou X (2016) Analysis on the rock–cutter
internet/e‐commerce use. Internet research interaction mechanism during the TBM tunneling pro-
Shayeghi H, Ahmadpour A, Heiran MMHK (2017) Optimal cess. Rock Mech Rock Eng 49(3):1073–1090
operation of wind farm in presence of pumped-storage Yu D, An S, Hu Q (2011) Fuzzy mutual information based
station as smart infrastructure and load estimation using min-redundancy and max-relevance heterogeneous fea-
artificial neural networks. In: 2017 Smart Grid Confer- ture selection. Int J Comput Intell Syst 4(4):619–633
ence (SGC). IEEE, p 1–7 Zhang G, Wang M, Liu K (2019) Forest fire susceptibility mod-
Singh T, Kanchan R, Saigal K, Verma A (2004) Prediction of eling using a convolutional neural network for Yunnan
p-wave velocity and anisotropic property of rock using province of China. Int J Disaster Risk Sci 10(3):386–403
artificial neural network technique
Sinuany-Stern Z, Mehrez A, Hadad Y (2000) An AHP/DEA Publisher’s Note Springer Nature remains neutral with regard
methodology for ranking decision making units. Int to jurisdictional claims in published maps and institutional
Trans Oper Res 7(2):109–124 affiliations.
Snowdon R, Ryley M, Temporal J (1982) A study of disc cut-
ting in selected British rocks. International journal of

Vol.: (0123456789)
13