ARMA 24–0036

Application of supervised learning algorithms to predict engineering

characteristics of soft to strong rock masses using actual TBM
performance data
Hanan Samadi
MSc., School of Geology, College of Science, University of Tehran, Tehran, Iran
Jafar Hassanpour
PhD., Associate Professor, College of Science, School of Geology, University of Tehran, Tehran, Iran
Jamal Rostami
PhD., Professor, Earth Mechanic Institute, Department of Mining Engineering, Colorado School of Mines, USA
Jitendra Khatti
PhD., Department of Civil Engineering, Rajasthan Technical University, Kota, Rajasthan, India

Copyright 2024 ARMA, American Rock Mechanics Association

This paper was prepared for presentation at the 58th US Rock Mechanics/Geomechanics Symposium held in Golden, Colorado, USA, 23-26 June 2024. This paper
was selected for presentation at the symposium by an ARMA Technical Program Committee based on a technical and critical review of the paper by a minimum of
two technical reviewers. The material, as presented, does not necessarily reflect any position of ARMA, its officers, or members. Electronic reproduction, distribution,
or storage of any part of this paper for commercial purposes without the written consent of ARMA is prohibited. Permission to reproduce in print is restricted to an
abstract of not more than 200 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgement of where and by whom the paper was
presented.

ABSTRACT: Knowing accurate values of rock engineering parameters in hard rock condition is the key in estimating and adjusting
TBM performance. Therefore, this paper has tried to predict the rock engineering factors using some indicators related to machine
performance, mechanical specification, and TBM operational parameters. In this study, TBM penetration per revolution (Pr) as
driving parameter, TBM operational parameters including cutterhead torque (CT), thrust force (TF), cutter load (Fn), cutterhead
rotational speed (RPM), and also cutterhead diameter as a mechanical machine specification are used to predict geomechanical
parameters based on reverse analysis in different types of soft to strong rocks (carbonate, volcanic, metamorphic). Soft computing
methods are trained using operating and driving data as well as rock engineering parameters gathered from several tunnelling projects
in Iran which were excavated in hard and soft rock (UCS = 0-200 MPa) condition using TBM. Among the data collected during the
pre-construction and construction phases of the project, the dataset was classified into several zones (107 sections), and the reasonable
distribution intervals of the main tunneling factors corresponding to each zone were defined (80% for training and 20% for testing).
The results of the proposed models indicate an acceptable and reliable accuracy.

1. INTRODUCTION Hassanpour et al. (2011) explored the relationship

between the TBM field penetration index (FPI) and the
The tunnel boring machine (TBM) is widely utilized for
rock quality designation (RQD) and unconfined
the construction of long tunnels due to its high efficiency compressive strength (UCS). The NTNU model,
and cost-effectiveness (Liu et al. 2016, Wu et al., 2019). proposed by the Norwegian University of Science and
The performance of the TBM during tunnel excavation is Technology, was developed for calculating the TBM
significantly influenced by the rock mass parameters.
penetration rate based on rock mass parameters (Bruland
However, obtaining accurate values of these parameters
1998). Rostami (2008) introduced the CSM model, which
through direct observation and in-situ testing of the
relies on full-scale laboratory cutting tests and multiple
working face is limited in tunnel boring machines regression analysis for assessing the TBM penetration
(Yamamoto et al. 2003, Farrokh and Rostami 2009, rate and the normal thrust applied on disc cutters. Benato
Hassanpour et al. 2015, Hassanpour et al. 2019). and Oreste (2015) formulated a new empirical approach
In recent years, researchers have sought to establish to estimate the penetration per revolution (Pr) for TBM
correlations between TBM operational parameters and tunneling based on data from TBM monitoring of alpine
rock mass specifications. These studies have predicted tunnels in the North-West of Italy. Various rock
rock mass parameters based on monitoring data of TBM classification systems, such as rock mass rating (RMR),
operation using back analysis methods. Farrokh et al. QTBM based on Q system, and rock structure rating
(2010) investigated the influence of ground conditions on (RSR), have been utilized to estimate TBM performance
machine performance based on information obtained (Bieniawski et al. 2007, Hassanpour et al. 2010, Rostami
from the Ghomroud tunnel, which is 36 km in length. 2016).
In recent decades, there has been a significant increase in argillaceous, pyroclastic and volcanic, and low-medium
the application of soft computing techniques, such as grade metamorphic rocks. The investigation included
artificial intelligence (AI) methods, in studying the geomechanical and machine data from the Zagros water
relationship between rock mass properties and the conveyance tunnel lot 2, Ghomrood water conveyance
performance parameters of tunnel boring machines and tunnel lots 2, 3 and 4, and Karaj water conveyance tunnel
other aspects of mechanized excavation. Several studies lot 1, totaling a combined length of 63.4 km. Table 1
(Jung et al. 2019, Gao et al. 2020, Mahmoodzadeh et al. presents a summary of the primary characteristics of the
2022) have utilized machine learning (ML) and deep TBM tunnelling projects. Additionally, Table 2 displays
learning (DL) models developed with field data, including the geomechanical parameters of the tunneling projects.
geomechanical parameters and TBM operational Table 1. Main characteristics of tunnelling projects
parameters, across various geotechnical ground
conditions. Tunnel TBM
Available TBM
Projects length diameter
data (km) type
(km) (m)
The significance of rock mass parameters, including
Zagross WCT Double
uniaxial compressive strength, rock quality designation, (Lot 2)
26 5
shield
6.73
brittleness index (BI), and distance between plane of Ghomrood WCT
21.5 15.7
Double
4.52
(Lots 2, 3 & 4) shield
weakness (DPW), on the performance of tunnel boring
Karaj WCT (Lot Double
machines in hard rock conditions has been discussed in 15.9 8.2 4.66
1) shield
previous research (Gong and Zhao 2009, Yagiz and
Karahan 2015, Samadi et al. 2022). Table 2. Geomechanical parameters of tunneling projects
The present research develops, trains, tests, and analyzes Projects
UCS
RQD RMCI RMR
(MPa)
the capabilities of machine learning approaches. Zagross WCT
Considering the gap identified in the literature survey, the 15-150 45-95 12-120 39-65
(Lot 2)
present research contains the following novelty. Ghomrood WCT (Lots 2,
25-150 10-100 4-170 43-76
3 & 4)
 The present research checks the capabilities of soft Karaj WCT (Lot 1) 30-150 10-100 6-150 27-73
computing approaches in the reliability analysis of
geomechanical properties of hard and soft rocks. 3. DATABASE DEVELOPMENT
 This study aims to use supervised predictor models to To construct the database for this study, data on rock
predict geomechanical parameters based on TBM engineering characteristics and TBM operational
operating and driving factors. A database of parameters were gathered. The input parameters include
necessary data from different mechanized tunnelling penetration per revolution (Pr) as a TBM performance
projects was applied to predict and calculate the target parameter, operational parameters such as cutterhead
geomechanical parameters based on input torque (CT), thrust force (TF), cutterhead rotation speed
parameters, including operating and driving (RPM), field penetration index (FPI), and thrust per cutter
parameters of boring machine. The obtained results of (Fn), as well as machine specifications like cutterhead
models were evaluated using the superior loss diameters (D). Additionally, combined parameters known
functions of machine learning. as thrust index (THI) and torque index (TQI), which are
 This study also demonstrates the impact of the associated with cutter load, cutterhead torque, cutterhead
database multicollinearity on the performance, diameter, and penetration per revolution, were also
accuracy, and overfitting of the developed models. included.

 This work draws the hypothesis statement for Previous work by Dullmann (2014) indicated that the TQI
reliability analysis and performs the ANOVA and can be determined using the Eq. (1). In this index, the
PCA tests to select the research hypothesis. geometric characteristics, as well as other performance
parameters like penetration per revolution and cutting
 The score analysis and regression error characteristics surface (A), are taken into account. Furthermore, the THI
curve are carried out to find the optimum performance was calculated using Eq. (2), considering specific contact
soft computing model for reliability analysis of cutting force in kN/m2.
geomechanical rock properties. MSP Tq
Torque index (TQI) = , MSP = (1)
A Pr
Pr Th∗1000
2. PROJECT DESCRIPTION Thrust index (THI) = , SCF = (2)
SCF A
The research utilized data obtained from a specialized The unit of TQI and THI are MN.m/(mm/rev)/m2 and
database containing multiple tunnel sections from various (mm/rev)/(kN/m2); respectively. It is important to note
projects, involving a diverse range of lithological units that constant parameters such as cutter spacing, disc cutter
with varying compressive strength of carbonate and
diameter, number of disc cutters, and tip width were not Fig. 1 displays box plot diagrams for three distinct
included in the machine learning analysis. categories: the total dataset, the training subset, and the
testing subset of the monitored TBM performance and
operational parameters as input variables, in addition to
geological parameters as output variables. Moreover,
Table 3 presents the statistical indices of the developed
dataset.
Table 3. The summary of statistical analysis of defined
parameters in database
Parameter Unit Min. Max. Ave.
Pr mm/rev 2 21 9.1
(a) (b) Fn kN 57.7 500 140
RPM rev/min 3.04 11.6 6.7
CT kN/m 220 1210 696
TF kN 2510 7790 5093
TQI MN.m/(mm/rev)/m2 3.8 42.6 16.4
THI (mm/rev)/(kN/m2) 0.17 3.7 1.04
FPI kN/mm/rev 2.7 70.7 17.8
UCS MPa 20 170 68
RQD % 10 100 59

The dataset was subjected to principal component

(c) (d)
analysis (PCA), which is a widely used multivariate
statistical technique for identifying patterns, trends, and
reducing the complexity of datasets by reducing the
number of dimensions, such as independent variables. In
PCA, linear combinations of original variables known as
principal components (PCs) are utilized to effectively
capture the differences across datasets. The eigenvalues
and percentage of variation of the PCs are depicted in Fig.
2 (a), providing insight into which PCs best explain the
variability in the dataset. Furthermore, the component
(e) (f) plot in rotated space presented in Fig. 2 (b) indicates that
the dataset does not exhibit clear natural grouping and is
an outlier for this measure.

(g) (h)
(a)

(i) (j) (b)

Fig. 1. The boxplot diagrams of defined parameters for three Fig. 2. (a) Scree plot for the first 10 PCs, (b) The component
categories including total dataset, training and testing sub plot in rotated space
dataset
4. SOFT COMPUTING TECHNIQUES Deep neural network (DNN) is a type of neural network
with multiple hidden layers, capable of learning complex
Support vector machine (SVM) is a powerful and patterns in data. Multi-layer perceptron (MLP) is a type
versatile classification algorithm that finds the best
of feedforward neural network that can learn non-linear
hyperplane that separates data into different classes. K- relationships in data through multiple layers of
nearest neighbors (KNN) is a simple and effective interconnected neurons. The Takagi-Sugeno fuzzy model
algorithm for classification and regression tasks, using the (TSF) is a type of fuzzy logic system that uses if-then
majority class or average value of the k-nearest data
rules and fuzzy sets to represent uncertain information.
points to predict the class or value of a new data point.
Gene expression programming (GEP) is an evolutionary
Radial basis function (RBF) is a type of kernel function algorithm used for symbolic regression. The detailed
used in machine learning algorithms, such as SVM, that description and background of these methods can be
measures the similarity between data points in a high- found in the literature and will not be repeated here.
dimensional space.

(a) (b) (a)

(a) (b)
(b)

(c) (d) (c)

(c) (d)
(d)

(e) (f) (e)

(e) (f)
(f)

(g) (g)
Fig. 3. The correlation between measured and predicted values Fig. 4. The correlation between measured and predicted values
of UCS of RQD
To deploy ML models, the computational capacity of (LCL = 2.755) and the GEP model achieving the lowest
Google Colab, which is compatible with Python 3, was accuracy (LCL = 7.144) for RQD. Additional statistical
utilized. Google Colab is a prominent platform for measures are used to further assess the performance of the
scientific computing in Python, offering a cost-effective ML models alongside the LCL metric. Furthermore, the
solution and streamlining package management and lower-dimensional projections of the predictor space
execution processes. The computational tasks were based on the input factors using the KNN network are
performed on a high-performance Intel (R) Core (TM) i5- depicted in Fig. 5.
1155G7 CPU running at 2.5 GHz, with a 16 GB RAM
Table 4 presents the statistical metrics for each model,
capacity, ensuring efficient and precise performance. The
providing a comprehensive assessment of their
ML models were optimized using training datasets to
performance for UCS. The measures include root mean
achieve superior performance in estimating rock
square error (RMSE), mean square error (MSE),
geomechanical properties based on boring machine data.
normalized root mean square error (NRMSE), root
Optimization involved setting hyperparameters to optimal
relative squared error (RRSE), relative standard error
values, with the trial-and-error method employed to fine-
(RSE), mean absolute deviation (MAD), mean absolute
tune the hyperparameters of the models.
percentage error (MAPE), relative root mean square error
(RRMSE), mean bias error (MBE), relative absolute error
(RAE), root mean squared logarithmic error (RMSLE),
log cash loss (LCL), and coefficient of determination
(R2). Similarly, Table 5 presents the statistical metrics for
each model, providing a comprehensive assessment of
their performance for RQD.
Table 4. The results of loss functions in the development of ML
models for UCS
MLP GEP DNN KNN RBF TSF SVM
RMSE 1.319 0.569 0.802 0.962 0.206 0.888 0.207
MSE 1.748 0.328 0.643 5.040 0.733 1.499 0.737
MAPE 0.000 0.001 0.002 0.003 0.001 0.000 0.001
MAD 0.245 0.324 0.284 0.245 0.223 0.218 0.190
RRMSE 0.047 0.176 0.210 0.176 0.271 0.249 0.271
(a)
NRMSE 0.009 0.004 0.005 0.006 0.001 0.005 0.001
RSE 0.268 0.853 0.040 7.891 1.484 3.473 1.493
RRSE 0.518 0.923 0.201 2.809 1.218 1.864 1.222
R2 0.836 0.853 0.905 0.790 0.805 0.844 0.611
MBE 0.116 0.045 0.068 0.000 0.015 0.007 0.015
RAE 1.869 1.010 0.253 0.385 1.427 2.458 1.419
RMSLE 0.057 0.140 0.113 0.095 0.145 0.133 0.145
LCL 6.739 10.116 9.108 6.939 7.200 5.748 7.190

Table 5. The results of loss functions in the development of ML

models for RQD
MLP GEP DNN KNN RBF TSF SVM
RMSE 0.08 0.20 1.15 0.80 0.297 0.3636 0.1169
MSE 0.42 1.16 0.12 0.12 1.157 1.402 0.690
MAPE 0.005 0.004 0.004 0.006 0.003 0.005 0.000
MAD 0.197 0.262 0.092 0.084 0.283 0.243 0.179
RRMSE 0.489 0.535 0.497 0.458 0.013 0.244 0.053
(b) NRMSE 0.000 0.002 0.012 0.008 0.002 0.003 0.001
RSE 0.11 0.02 0.00 0.79 0.041 0.218 0.030
Fig. 5. The lower-dimensional projection of the predictor space RRSE 0.34 0.15 0.07 0.89 0.203 0.467 0.173
with K = 3 based on KNN for rock mechanical properties R2 0.81 0.82 0.80 0.70 0.803 0.857 0.913
prediction MBE -0.29 -0.00 -0.08 -0.23 -0.21 -0.11 -0.03
RAE 0.00 0.18 0.59 0.91 0.052 0.701 0.129
The distribution histograms of results and correlations RMSLE 0.18 0.01 0.02 0.0 0.038 0.090 0.051
between the estimated results of the ML models and those LCL 6.48 7.14 5.62 4.76 6.710 4.669 2.755
obtained from the actual conditions for UCS are presented
in Fig. 3, and for RQD in Fig. 4. Notably, all models The ranking of each model is presented numerically, with
demonstrate high accuracy in predicting rock a visual representation in Fig. 6. These results indicate
geomechanical properties, with the TSF model achieving
that the SVM model (with scores of 62 and 68 for UCS
the highest accuracy (LCL = 5.748) and the GEP model and RQD, respectively) demonstrated the highest level of
achieving the lowest accuracy (LCL = 10.116) for UCS.
accuracy among the models. Conversely, the TSF model
Furthermore, SVM model achieving the highest accuracy showed lower accuracy, with scores of 27 and 22 for UCS
 statistical evaluation metrics using a novel thermal map
matrix referred to as the accuracy matrix, which is utilized
and RQD, respectively. The ranking order based on
to evaluate the efficiency of the generated models. The
prediction accuracy for UCS and RQD is as follows:
accuracy of the ML models is calculated for each class
UCS: SVM → DNN → RBF → GEP → MLP → KNN
using the ideal value as a reference. For Class I, the ideal
→ TSF value for UCS would be 50, and for Class II, III, and IV,
RQD: SVM → DNN → RBF → GEP → MLP → KNN the ideal values would be 36, 16, and 5, respectively. The
KNN model demonstrated the highest accuracy among
→ TSF the UCS class divisions, with values approximately at 45,
39, 21, and 2. Hence, the figure indicates that KNN, TSF,
DNN, and MLP are the most precise predictors for each
UCS class, while SVM and RBF are the least accurate for
predicting UCS in terms of class strength division.

(a)

(a)

(b)
Fig. 6. Ranking score of the applied ML models for the rock
geomechanical properties prediction, (a) UCS), (b) RQD
A thorough analysis was performed to evaluate the
prediction outcomes of various soft computing methods
for rock geomechanical properties (specifically UCS and
RQD) through the utilization of normalized error analysis,
as illustrated in Fig. 7. The performance of the employed
algorithms was assessed using Taylor diagrams, which
considered the method prediction results, normalized
standard deviation, normalized root mean square error, (b)
and variance. This comprehensive evaluation aimed to Fig. 7. Taylor chart for comparison of multiple supervised
provide a comprehensive understanding of the models to predict rock geomechanical properties, (a) UCS, (b)
effectiveness and reliability of the different soft RQD
computing techniques in predicting rock geomechanical Similarly, the RQD values (ranging from 0 to 100) are
properties. The results of this evaluation offer valuable divided into four classes to illustrate the thermal map
insights into the strengths and limitations of each method, matrix, with Class I representing 0-25, Class II
which can be utilized to inform future research and representing 25-50, Class III representing 50-75, and
application of soft computing methods in the field of Class IV representing 75-100. Fig. 8 (b) exhibits the
geomechanics. statistical evaluation indicators through a newly
The UCS values (ranging from 0 to 200 MPa) are developed thermal map matrix known as the accuracy
categorized into four classes to illustrate the thermal map matrix, which is used to evaluate the performance of the
matrix, with Class I representing 0-50, Class II generated models. The accuracy of the ML models is
representing 50-100, Class III representing 100-150, and calculated for each class using the ideal value as a
Class IV representing 150-200. Fig. 8 (a) presents the reference. For Class I, the ideal value for RQD would be
17, and for Class II, III, and IV, the ideal values would be
18, 43, and 29, respectively. The GEP model increasing UCS, while other models exhibit varying
demonstrated the highest accuracy among the RQD class predictions when TQI is increased. Subsequently, the
divisions, with values approximately at 15, 17, 48, and 27. variation of UCS with respect to the parameter FPI for
Thus, the figure illustrates that most of the applied each ML model was examined, as depicted in Fig. 10.
machine learning algorithms are the most accurate
predictors for each RQD class, while the SVM is the least
accurate for predicting RQD in terms of class strength
division.

(a) (b)

(a)

(c) (d)
(b)
Fig. 8. Accuracy matrix to analyze the performance of the ML
models in the prediction of rock geomechanical properties, (a)
UCS, (b) RQD
Several machine learning models were developed that
demonstrate exceptional predictive accuracy in
estimating rock geomechanical properties. It is, therefore,
necessary to analyze input trends and output patterns of
rock geomechanical properties to ensure the
generalization capabilities of these models rather than (e) (f)
solely relying on cross-correlation. To accomplish this,
certain input variables were maintained constant while
varying others within a specified range.
Initially, the efficacy of the models in predicting the UCS
by varying the input parameter TQI within its range of
values (4-43 kN/(mm/rev)/m2) was assessed, while
holding FPI constant at 18 kN/mm/rev. Additionally, in
subsequent stages, TQI was held constant (at 17
kN/(mm/rev)/m2) while FPI was varied across its range
of values (2-70 kN/mm/rev). It is worth noting that both (g)
TQI and FPI increase with an increase in UCS. Fig. 9. Behavior of the ML models about the parameter TQI for
Leveraging machine learning models, the variation of UCS analysis (FPI is held constant)
UCS with respect to the parameters TQI and FPI was
Similar to the previous analysis, the models display
examined.
declining, constant, increasing, and erratic trends in UCS
Fig. 9 illustrates the variations in UCS with respect to as FPI is increased. The KNN, RBF, and TSF models
changes in the parameter TQI for each of the trained ML consistently predict an increasing UCS, while other
models. The results indicate that all ML models models yield varying results. Thus, the KNN, RBF, and
demonstrate declining, constant, increasing, and erratic TSF models have proven to be accurate in this scenario.
trends in UCS as a function of increasing TQI. Notably, Also, the models' efficacy in predicting the RQD was
the RBF and TSF models consistently predict an evaluated by varying the input parameter TQI within its
range of values, while holding FPI and THI fixed at 18
kN/mm/rev and 1.046 (mm/rev)/(kN/m2), respectively. In
this case, TQI increases with an increase in RQD.
Therefore, we examined how RQD shifts as a function of
the parameter TQI for each ML model, leading to Fig. 10
displaying the variations in RQD. The results show that
the RBF and KNN models consistently predict an
increasing RQD, while other models demonstrate varying
predictions as TQI is increased. Notably, THI and FPI
cannot be considered singularly as changeable or fixed (a) (b)
factors for RQD analysis due to their inherent
relationship, which results in changes to both parameters
if one is altered. Therefore, TQI was exclusively
considered as a changeable factor in analyses of RQD.

(c) (d)

(a) (b)

(e) (f)

(c) (d)

(g)
Fig. 11. Behavior of the ML models about the parameter TQI
for RQD analysis (FPI and THI are held constant)
(e) (f)

5. SENSITIVITY ANALYSIS
In order to contextualize the findings, a sensitivity
analysis may be conducted to evaluate the most influential
model parameters. A comprehensive understanding of
this subject is imperative for reducing data complexity
and preventing model overfitting. In this study, the mutual
information (MI) technique was employed for the
sensitivity analysis. MI, a feature selection approach
(g) rooted in information theory, leverages information gain
Fig. 10. Behavior of the ML models about the parameter FPI in the construction of decision trees. The MI between two
for UCS analysis (TQI is held constant) variables quantifies the extent to which one variable can
 be inferred from observations of the other. Utilizing MI of the construction process, potentially leading to issues
is straightforward with categorical input and output data, such as jamming, clogging, and collapse accidents. To
and though originally designed for textual data, it may be address these challenges, the development of a
adapted for numerical data. The efficacy of MI may be comprehensive model for predicting rock engineering
assessed by the degree to which it reduces entropy. As parameters must carefully consider the selection of input
illustrated by Eq. (3), the MI score should be a positive factors, accurate principle component analysis,
integer between zero and infinity. A high MI value multivariable regressions, hyper-parameter optimization
indicates the importance of the feature for model training of machine learning models, and the choice of appropriate
and thus warrants greater consideration in the final score. machine learning algorithms.
Conversely, a low MI score, such as 0, suggests limited
Furthermore, the interpretation of the contribution of each
or no correlation between the attribute and the objective.
input variable is crucial for the integrity of the developed
MI (feature; target) = Entropy (feature) – Entropy (feature model. Despite this, there is a lack of comprehensive
| target) (3) evaluations that consider the relationship between TBM
tunneling data and rock engineering parameters based on
The calculated MI scores for all inputs for both UCS and
specific driving and operating factors of a boring machine
RQD are depicted in Fig. 11, highlighting the link
using supervised and unsupervised learning algorithms.
between the RPM and CT parameters and the most and
Additionally, there is limited research that evaluates
least significant impacts, respectively, on the UCS. Also,
geomechanical parameters with boring machine data
this figure presents the link between the CT and Pr
based on reverse analysis.
parameters and the most and least significant impacts,
respectively, on the RQD. The influence of other This study has yielded several conclusions as follows:
parameters on the rock geomechanical properties is also
(1) To validate the developed model, geomechanical
noteworthy.
properties and boring machine data were collected and
categorized into 107 zones based on the geotechnical
conditions along tunnel routes. The performance and
geomechanical data were gathered from various
tunnelling projects in Iran, resulting in a dataset from
different engineering geological units, 80% of which was
randomly selected for training and 20% for final testing.
(2) Soft computing techniques were applied to estimate
geomechanical parameters in different types of rocks
(carbonate, low-medium grade metamorphic, volcanic)
based on collected data with varying ranges of values.
Notably, these models were used from data with a UCS
(a) range of 0-200 MPa.
(3) Machine learning-based models were found to be a
practical alternative to other costly and time-consuming
methods for determining rock geomechanical factors
based on boring machine data using reverse analysis. The
prediction accuracy of the models was ranked as follows:
SVM → DNN → RBF → GEP → MLP → KNN → TSF
(4) For RQD prediction, the RBF and KNN models were
found to be more accurate than others when considering
the performance of the ML models in comparison to
practical mode. Also, for UCS prediction, the RBF, TSF,
and KNN models were found to be more accurate than
(b) others when considering the performance of the ML
Fig. 12. Sensitivity score of input parameters, (a) UCS, (b) models in comparison to practical mode
RQD
(5) Sensitivity analysis indicated that the RPM and CT
parameters had the most and least significant impacts,
6. CONCLUSIONS respectively, on UCS, while CT and Pr had the most and
In the context of TBM tunneling construction, the least significant impacts, respectively, on RQD.
geological conditions in front of the excavation face can (6) This study is significant in that it provides tunnel
have a significant impact on both the progress and safety engineers with a means to accurately estimate rock
geomechanical properties based on TBM data without 14. Liu, Q., Huang, X., Gong, Q., Du, L., Pan, Y., and Liu. J.
requiring extensive time and financial resources. The (2016). Application and development of hard rock TBM
findings of this paper can be applied to upcoming projects and its prospect in China. Tunnel. Under. Space. Tech. 57,
with similar geomechanical properties and TBM 33-46. https://doi.org/10.1016/j.tust.2016.01.034
15. Mahmoodzadeh, M., Taghizadeh, M., Mohammed, A.H.,
specifications.
Ibrahim, H.H., Samadi, H., Mohammadi, M., and Rashidi,
S. (2022). Tunnel wall convergence prediction using
REFERENCES optimized LSTM deep neural network. Geomech. Eng.
31(6), 545-556.
1. Benato, A. and Oreste. P. (2015). Prediction of penetration https://doi.org/10.12989/gae.2022.31.6.545
per revolution in TBM tunneling as a function of intact 16. Rostami, J. (2008). Hard rock TBM cutterhead modeling
rock and rock mass characteristics. Inter. J. Rock. Mech. for design and performance prediction. Geomech. Tunnel.
Min. Sci. 74, 119-127. (Aust. J. Geotech. Eng.). 1, 18-28.
https://doi.org/10.1016/j.ijrmms.2014.12.007 https://doi.org/10.1002/geot.200800002
2. Bieniawski, Z.T., Celada, B., and Galera, J.M. (2007). 17. Rostami, J. (2016). Performance prediction of hard rock
TBM excavability: prediction and machine-rock tunnel boring machines (TBMs) in difficult ground.
interaction. Rapid. Excav. Tunnel. Conf. 1118-1130. Tunnel. Under. Space. Tech. 57, 173-182.
3. Bruland, A. (1998). Hard rock tunnel boring. PhD Thesis, https://doi.org/10.1016/j.tust.2016.01.009
Norwegian University of Science and Technology. 18. Samadi, H., Hassanpour, J., and Tarigh Azali, S. (2022).
4. Dullmann, J. (2014). Performance optimization and wear Developing GEP technique for prediction of EPB-TBM
forecast in hydro-shield tunneling in soft ground. PhD performance in limestone strata. TBMDigs 2022. Austria,
Thesis, Ruhr University. 158-164.
5. Farrokh, E., Rostami, J., and Laughton, C. (2010). Study of 19. Wu, Z.J., Zhang, P.L., Fan, L.F., and Liu, Q.S. (2019).
various models for estimation of penetration rate of hard Numerical study of the effect of confining pressure on the
rock TBMs. Tunnel. Under. Space. Tech. 30, 110-123. rock breakage efficiency and fragment size distribution of
https://doi.org/10.1016/j.tust.2012.02.012 a TBM cutter using a coupled FEM-DEM method. Tunnel.
6. Farrokh, E. and Rostami, J. (2009). Effect of adverse Under. Space. Tech. 88, 260-275.
geological condition on TBM operation in Ghomroud https://doi.org/10.1016/j.tust.2019.03.012
tunnel conveyance project. Tunnel. Under. Space. Tech. 20. Yagiz, S., and Karahan, H. (2015). Application of various
24(4), 436-446. https://doi.org/10.1016/j.tust.2008.12.006 optimization techniques and comparison of their
7. Gao, X., Shi, M., Song, X., Zhang, C., and Zhang, H. performances for predicting TBM penetration rate in rock
(2020). Recurrent neural networks for real-time prediction mass. Inter. J. Rock. Mech. Min. Sci. 80, 308-315.
of TBM operating parameters. Auto. Const. 98, 225-235. https://doi.org/10.1016/j.ijrmms.2015.09.019
https://doi.org/10.1016/j.autcon.2018.11.013 21. Yamamoto, T., Shirasagi, S., Yamamoto, S., Mito, Y., and
8. Gong, Q.M. and Zhao, J. (2009). Development of a rock Aoki, K. (2003). Evaluation of the geological condition
mass characteristics model for TBM penetration rate ahead of the tunnel face by geo-statistical techniques using
prediction. Inter. J. Rock. Mech. Min. Sci. 46(1), 8-18. TBM operational data. Tunnel. Under. Space. Tech. 18(2),
https://doi.org/10.1016/j.ijrmms.2008.03.003 213-221. https://doi.org/10.1016/S0886-7798(03)00030-0
9. Hassanpour, J., Rostami, J., Khamehchiyan, M., Bruland,
A., and Tavakoli, H.R. (2010). TBM performance analysis
in pyroclastic rocks, a case history of Karaj Water
Conveyance Tunnel (KWCT). Rock. Mech. Rock. Eng. 43,
427-445. https://doi.org/10.1007/s00603-009-0060-2
10. Hassanpour, J., Rostami, J., and Zhao, J. (2011). A new
hard rock TBM performance prediction model for project
planning. Tunnel. Under. Space. Tech. 29, 595-603.
https://doi.org/10.1016/j.tust.2011.04.004
11. Hassanpour, J., Rostami, J., Zhao, J., and Tarigh Azali, S.
(2015). TBM performance and disc cutter wear prediction
based on ten years’ experience of TBM tunnelling in Iran.
Geomech. Tunnel. 8(3), 239-247.
12. Hassanpour, J., Esmaeili Vardanjani, S., Rostami, J., and
Cheshomi, A. (2019). Engineering geological studies used
for redesigning and employing a hard rock TBM in soft
rock formations of Chamshir water conveyance tunnel.
Geopersia. 9(1), 1-20.
13. Jung, J.H., Chung, H., Kwon, Y.S., and Lee, I.M. (2019).
An ANN to Predict Ground Condition ahead of Tunnel
Face using TBM Operational Data. KSCE J. Civil. Eng. 23,
3200-3206, https://doi.org/10.1007/s12205-019-1460-9