This article was downloaded by: [North Dakota State University]

On: 04 March 2015, At: 06:24

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Geomechanics and Geoengineering: An International

Journal
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tgeo20

Prediction model of tunnel boring machine

performance by ensemble neural networks
a b c a b
Zhiye Zhao , Qiuming Gong , Yun Zhang & Jian Zhao
a
School of Civil and Environmental Engineering , Nanyang Technological University ,
Nanyang Avenue, Singapore
b
Ecole Polytechnique Fédérales de Lausanne (EPFL) , Rock Mechanics Laboratory , CH-1015,
Lausanne, Switzerland
c
College of Architecture and Civil Engineering , Beijing University of Technology , Beijing,
China
Published online: 05 Jun 2007.

To cite this article: Zhiye Zhao , Qiuming Gong , Yun Zhang & Jian Zhao (2007) Prediction model of tunnel boring machine
performance by ensemble neural networks, Geomechanics and Geoengineering: An International Journal, 2:2, 123-128, DOI:
10.1080/17486020701377140

To link to this article: http://dx.doi.org/10.1080/17486020701377140

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 Geomechanics and Geoengineering: An International Journal
Vol. 2, No. 2, June 2007, 123--128

Prediction model of tunnel boring machine performance by ensemble neural networks

ZHIYE ZHAO*†, QIUMING GONG‡§, YUN ZHANG† and JIAN ZHAO‡
†School of Civil and Environmental Engineering, Nanyang Technological University, Nanyang Avenue, Singapore
‡Ecole Polytechnique Federales de Lausanne (EPFL), Rock Mechanics Laboratory, CH-1015, Lausanne, Switzerland
§College of Architecture and Civil Engineering, Beijing University of Technology, Beijing, China

(Received 9 December 2006; in final form 22 March 2007)

The penetration rate of a tunnel boring machine (TBM) depends on many factors ranging from the machine design to the geological properties.
Therefore it may not be possible to capture this complex relationship in an explicit mathematical expression. In this paper, we propose an ensemble
Downloaded by [North Dakota State University] at 06:24 04 March 2015

neural network (ENN) to predict TBM performance. Based on site data, a four-parameter ENN model for the prediction of the specific rock mass
boreability index is constructed. Such a neural-network-based model has the advantages of taking into account the uncertainties embedded in the site
data and making appropriate inferences using very limited data via the re-sampling technique. The ENN-based prediction model is compared with a
non-linear regression model derived from the same four parameters. The ENN model outperforms the non-linear regression model.

Keywords: Tunnel boring machine performance; Specific rock mass boreability index; Ensemble neural network

1. Introduction prediction model based on the rock compressive strength

Tunnel boring machines (TBM) are now widely used to exca-
vate rock tunnels. Extensive studies have been carried out to
establish the relationship between the rock mass quality and
TBM performance. In general, the penetration rate primarily
reflects the machine efficiency in rock breakage. Four
parameters, i.e. rock strength, rock brittleness index, joint
spacing, and orientation, are the main influencing factors on
the TBM performance. The penetration rate primarily reflects
the machine efficiency in rock breakage. The advance rate and
utilization relate to the suitability of the the machines, the rock
mass, and the project management. The cutter wear depends not
only on the rock mineral composition, rock texture, and tunnel
face variation such as mixed face, but also on the cutter char-
acteristics and the operating factors.
The planning of tunnel projects and the selection of construc-
tion methods require effective prediction of TBM performance.
Therefore the prediction of TBM performance in a given rock
mass has been the subject of much research in the last decade.
Because of the complexity of the interaction between rock mass
and TBM, it is not possible to predict TBM performance theo-
retically. Various prediction models have been developed dur-
ing the last 30 years, including single-factor models (Graham
1976, Farmer and Glossop 1980, Nelson 1983, Hughes 1986,
O’Rourke et al. 1994), multiple-factor models (Rostami 1997,
Bruland 1998, Nelson et al. 1999, Barton 2000, Grima et al.
2000, Bruines 2001), simple parameter models such as the
*Corresponding author. Email: CZZHAO@ntu.edu.sg
(Graham 1976), complex probability models (Nelson et al.
1999), and neural-fuzzy models (Grima et al. 2000, Bruines
2001).
Another approach to predicting TBM performance is based
on the rock mass classification system (Sapigni et al. 2002,
Ribacchi and Lembo-Fazio 2005). Morimoto and Hori (1986)
assessed the effect of rock mass on TBM excavation rate for the
headrace tunnel of the Hayakido Hydro Power Station.
Innaurato et al. (1991) utilized the uniaxial compressive
strength c and the rock structure rating (RSR) to predict
TBM penetration rate. The analysis was based on 112 tunnel
sections, which can be considered homogeneous with regard to
TBM excavation. Laughton and Nelson (1996) classified rock
masses into 36 classes, representing combinations of block size
rank (four classes), degree of alteration or weathering (three
classes), and level of water inflow at the heading (three classes).
Alber (1996, 2000) used rock mass strength scm to evaluate the
specific penetration rate (mm/revolution/MN/cutter). The inte-
grated mass system (IMS) has been extensively used to study
the correlation between TBM performance and rock mass
classes in Hong Kong (McFeat-Smith and Askilsrud 1993,
McFeat-Smith 1999, Grandori et al. 1995). In all these rock
mass classification systems, rock strength, joint spacing or
frequency, and joint conditions were used as the key influen-
cing parameters.
The artificial neural network (NN) is a computational model
for information processing based on biological neural net-
works. It has been successfully applied to a wide range of
civil engineering applications, such as fault detection

Geomechanics and Geoengineering: An International Journal

ISSN 1748-6025 print=ISSN 1748-6033 online Ó 2007 Taylor & Francis
http:==www.tandf.co.uk=journals
DOI: 10.1080=17486020701377140
 124 Z. Zhao et al.

(Jakubek and Strasser 2004), concrete strength prediction
(Jiang et al, 2003), and modelling the behaviour of steel at
elevated temperatures (Zhao 2006). An ensemble neural net-
work (ENN) is a collection of a finite number of neural net-
works that are trained for the same task. Compared with a single
NN, the ENN can provide better generalization, especially
when the training data are relatively sparse.
In this paper, an ENN based on site information is used to
model the boreability of the rock mass. As the data set is
limited, the bootstrap sampling technique is used to training
each component NN in the ENN structure. The ENN-based
rock mass boreability prediction model is compared with a
non-linear regression model which was derived from the same
data set. The results demonstrate superior accuracy of the ENN
model compared with the regression model.
Downloaded by [North Dakota State University] at 06:24 04 March 2015
masses were mainly composed of granite rock masses with
joint spacing ranging from very wide to very narrow and with
different weathering grades (Zhao et al. 2006). The correspond-
ing TBM performance data and TBM specifications were col-
lected and analysed to calculate SRMBI. Although many tunnel
face mappings have been performed during tunnel construction,
not all of them were used to establish the database because core
rock samples were not obtained in some locations owing to the
time required to carry out the coring procedure. In addition, the
results of some rock experiments were excluded because the
samples failed along structural weak planes. In total, 47 sets of
data were used for neural network analysis. The general range
of the data corresponding to the five parameters is shown in
table 1.
The input--output mapping for both the regression analysis
and the neural network analysis is aimed to determine the

following relationship:
2. Database properties and regression analysis
SRMBI ¼ f ðUCS; Jv; Bi; Þ ð1Þ
Four rock mass parameters are identified as the factors influen-
A non-linear regression analysis based on the 47 sets of data
cing the TBM penetration process: uniaxial compressive
was carried out (Gong 2006). The aim of non-linear regression
strength (UCS), brittleness (Bi), joint spacing (Jv), and joint
is to fit a model to these given data. The program finds the best-
orientation () (Gong 2006). The uniaxial compressive strength
fit values of the variables in the model which can be interpreted
of the rock affects the formation of the crushed zone during
scientifically. The non-linear regression proceeds as follows.
TBM penetration. The higher the rock strength, the more load is
required on the cutter. The brittleness of the rock affects the size
(a) Start with an initial estimated value for each variable in
of the crushed zone and crack initiation and propagation during
the equation.
rock indentation and chip formation. A higher brittleness results
(b) Generate the curve defined by the initial values and cal-
in easier rock fragmentation. Because of the existence of joints,
culate the sum of the squared residuals.
the principal stress field is changed which affects the rock
(c) Adjust the variables to make the curve come closer to the
chipping process. An increase in joint spacing results in a
data points.
decrease in the penetration rate. The preferred joint angle
(d) Stop the calculations when the adjustments make virtually
between the tunnel axis and the joint plane for optimum pene-
no difference to the sum of squares.
tration rate is 60 . The penetration rate decreases for angles
(e) Report the best-fit results.
greater or less than 60 . The rock mass boreability is expressed
by the boreability index, defined as the ratio of the applied
The non-linear regression analysis based on NLREG
thrust per cutter to the penetration per revolution (Hamilton
(Sherrod 1996) gives
and Dollinger 1979, Wanner and Aeberli 1979, Sundin and
Wanstedt 1994). It should be noted that the boreability index
is a cutter force normalized by the number of revolutions per SRMBI ¼ 37:06UCS0:26 Bi0:10
 
minute (r.p.m.) and the penetration rate. Thus it facilitates 0:84e0:06Jv þ e0:09 sinðþ30Þ ð2Þ
comparison of the performance of different TBMs. Because
of a change in the efficiency of the cutting action at the cutter
head for different cutter forces, the boreability index is not a
constant (Hamilton and Dollinger 1979, Gong et al. 2007).
Table 1. Range of parameters
Therefore the boreability index cannot accurately represent
the rock mass boreability. The specific rock mass boreability Minimum Maximum Data distribution
index (SRMBI), defined as the boreability index at the critical
Orientation angle () 10 80 Evenly distributed
point of penetration of 1 mm/rev and based on in situ tests, is a Joint spacing (Jv) 0 30 Joint count Jv ,16
better representation of rock mass boreability (Gong et al. for most data
2007). The SRMBI eliminates the influence of the cutter force Compressive strength 100 260 Most data between
(UCS) 160 and 200 MPa
variation on rock mass boreability. Rock brittleness (Bi) 8 22 Most data between
The parameters characterizing rock mass boreability were 14 and 18
determined from data obtained from extensive rock material Specific rock mass 110 240 Most data between
boreability index 130 and 210
strength experiments and tunnel face mappings in two deep (SRMBI) kN/cutter/mm/rev
tunnel sewerage systems (DTSSs) in Singapore. The rock
 Tunnel boring machine performance model 125

The coefficient of determination (R2) obtained is 74.94%. It
indicates that the regression model can explain 75% of the total
variation in the 47 data sets.
3. Ensemble neural network modelling
The networks in an ENN are usually trained independently and
their predictions are combined (Sollich and Krogh 1996). In
other words, any of the component networks in the ENN could
provide a solution to the task by itself, but better results might
be obtained by the ENN because of its better generalization.
Different methods can be employed to combine the solutions
achieved by the component networks. A typical ENN architec-
ture is shown in figure 1.
In general, an ENN is constructed in two steps: creating
Downloaded by [North Dakota State University] at 06:24 04 March 2015
best weight configuration of the individual component network
is the one with the smallest training mean square error (MSE).
Since good regression ensemble members must be both accu-
rate and diverse, the training of each component network
should also have high accuracy and diversity. Based on a
trial-and-error procedure, it is found that the BPNN with two
hidden nodes generally provides the best generalization, and so
all component networks have the same two hidden nodes.
The ENN algorithm (figure 2) is as follows.
1. The data set is split into two parts: the training data set and
the test data set. The training data set is used for learning
the various component neural network models. The test
data set is not used in the network training, but for testing
the performance of the trained ENN model.
2. Three component networks are used in the ENN. Each

component network has one input layer with four input

component networks, and combining these networks into an
ENN. In order to create component networks, good regression
or classification component networks must be both accurate
and diverse. a number of training parameters, including the start
initial conditions, the training data, the topology of the nets,
and the training algorithm, can be manipulated to find networks
with different generalization abilities (Sharkey 1999).
In this study, back-propagation neural networks (BPNNs) Randomly select a few sets of
were adopted as the component networks. BPNNs, which are training data from the total training
data, based on the re-sampling
multilayer feed-forward networks combined with a back-pro- procedure.
pagation learning algorithm, were proposed by Rumelhart et al.
(1986). The main idea of a BPNN is that the error in the units of
the hidden layer is determined by back-propagating the error in
the units of the output layer; this is known as the back-propaga-
tion learning rule. The back-propagation algorithm or general- Create an ENN,
determine the no. of the component
ized delta rule is based on an error-correction learning rule. It is NNs and no. of the hidden nodes in
a gradient descent method which minimizes the total squared each component NN.
error of the output computed by the network. Because of the
limited data available, the bootstrap sampling method is
employed to create component neural networks (Efron and
Tibshirani 1993), where ‘bootstrap’ means that one available For each component NN, randomly
sample gives rise to many others by resampling. It adds a new run several times, then choose the
training set containing N data by sampling data N times uni- best structure, which has the
formly at random from the total training data. During the minimum training-MSE.
sampling, the repeated data randomly picked can be used in
the new training set. The resampled sets are often called boot-
strap replicates.
In the proposed ENN, each component network is created Combine the ENN with each best
component NN using the
several times, and the best weight configuration of this compo- averaging weight method.
nent network is kept in the ENN. The criterion to choose the

NN 1 Task solution 1
Using testing data to
calculate the test-MSE.
Input NN 2 Task solution 2 Better solution

end
NN m Task solution m

Figure 1. Architecture of the ensemble neural network. Figure 2. Flowchart of the proposed ensemble neural network.
 126 Z. Zhao et al.

nodes, one hidden layer with two hidden nodes, and one 220

Predicted Specific Rock Mass Boreability Index

output layer with one output node. 210
3. Run each component network several times randomly. For 200
each run, calculate the MSE of the component network by 190

(kN / cutter / mm / rev)

using the training data set. Then perform a comparison to 180
select the best weight configuration of the individual com- 170
ponent network that has the smallest training MSE. 160
4. Combine the ENN with each best component network 150
using the simple average output of the component 140
networks. 130
5. The MSE of the test data set will be used as the perfor- 120
mance measurement of the ENN. 110
100
100 110 120 130 140 150 160 170 180 190 200 210 220
Measured Specific Rock Mass Boreability Index (kN / cutter / mm / rev)
4. Rock mass boreability analysis by ensemble neural
network Figure 3. Comparison of the measured and predicted SRMBI (test data).
Downloaded by [North Dakota State University] at 06:24 04 March 2015

Only three of the 47 data sets collected had an SRMBI value in

the range 110--130. A preliminary ENN analysis showed a 220

Predicted Specific Rock Mass Boreability Index

relatively poor correlation over that range because the sparse 210
data and so these three data sets were excluded from the ENN 200
analysis. Therefore the ENN training and testing carried out to 190
(kN / cutter / mm / rev)

predict the specific rock mass boreability index was based on 44 180
sets of site data. Thirty-six sets were randomly selected as the 170
training data, and the remaining eight sets as the test data. The 160
training data for each component neural network were ran- 150
domly selected based on the bootstrap sampling technique. 140
For the purposes of comparison, three types of neural network 130
were used: a single NN with two hidden nodes, an ENN with 120
four component networks, and an ENN with 10 component 110
networks. 100
Twenty runs were carried out for each type of neural net- 100 120 140 160 180 200 220
work, and the statistical results are summarized in table 2. It can Measured Specific Rock Mass Boreability Index (kN / cutter / mm / rev)
be seen that both ENNs are more accurate than the single NN. Figure 4. Comparison of the measured and the predicted SRMBI (training
The single NN is quite sensitive to the initial weights assigned data).
for each run, which explains its large standard deviation. When
the number of component networks is increased from four to
ten, both the training and the test results are improved, indicat-
ing better generalization. The values of R2 for Ensemble-4 are Four parametric studies were performed to study the beha-
0.76 for training and 0.78 for testing, and for Ensemble-10 they viour of the parameters governing SRMBI. When the effect of
are 0.78 for training and 0.81 for testing. Both ENN results (R2) one parameter on the SRMBI was under study, the other three
are better than that for the non-linear regression analysis (R2 = parameters were kept constant, so that the correlation between
0.75. Figures 3 and 4 show typical neural network testing and the studied parameter and the SRMBI could be obtained.
training comparisons between the measured data and the neural Figure 5 shows the relationship between UCS and SRMBI. In
network prediction. general, SRMBI increases with increasing UCS. The analysis
shows that UCS has a significant effect on SRMBI. The rela-
tionship can be divided into three sections. The first section is
for UCS ,100 MPa; SRMBI increases slowly within this
Table 2. TBM boreability modeling (20 runs)
range. In the second section (100 MPa , UCS , 200 MPa)
Single Ensemble 4 Ensemble 10 there is a rapid increase in SRMBI. In the third section (UCS .
200 MPa), the increase in SRMBI becomes slow again.
Test MSE Minimum 123.12 71.68 71.63
Mean 256.07 175.00 127.47
The relationship between Bi and SRMBI is shown in figure 6.
SD 155.58 92.44 50.70 SRMBI decreases with increasing Bi. The analysis shows that
Train MSE Minimum 112.58 111.35 120.09 Bi has a greater effect on SRMBI at higher values of UCS.
Mean 246.43 167.67 140.43
SD 190.08 28.43 8.32
Figure 7 shows the relationship between Jv and SRMBI.
When Jv is between 0 and 45, SRMBI varies over a large
 Tunnel boring machine performance model 127

200 170

190
UCS = 100, Bi = 10, Jv = 5
180 UCS = 100, Bi = 10, Jv = 10
SRMBI (kN/cutter/mm/rev)

SRMBI (kN/cutter/mm/rev)
160
170 UCS = 100, Bi = 10, Jv = 15

160
150 150
140
130
angle = 60, Jv = 5, Bi = 10 140
120
angle = 60, Jv = 10, Bi = 10
110 angle = 60, Jv = 15, Bi = 10
100 130
0 30 60 90 120 150 180 210 240 270 300 330 360 0 10 20 30 40 50 60 70 80 90
UCS (MPa) Angle (degree)

Figure 5. The effect of rock strength on specific rock mass boreability index. Figure 8. The effect of joint orientation on specific rock mass boreability
index.
Downloaded by [North Dakota State University] at 06:24 04 March 2015

200 other three parameters. As  increases from 0 to 90 , SRMBI

190 first decreases and then increases. It reaches a minimum value
SRMBI (kN/cutter/mm/rev)

180 at approximately 60 .

170 Based on the ENN simulation results (figures 5--8), the
160 influence of the rock mass parameters on SRMBI can be sum-
150
marized as follows.
140
1. UCS and Jv have the greatest effect on SRMBI.
130
angle = 80, Jv = 15, UCS = 100 2. The effect of Bi on SRMBI is more significant at higher
120 angle = 80, Jv = 15, UCS = 200
values of UCS (i.e. UCS = 300).
110 angle = 80, Jv = 15, UCS = 300
3. The angle  has the least effect on SRMBI.
100 4. A small SRMBI (i.e. ,140) can be achieved with the
0 5 10 15 20 25 30
following parameters: UCS # 200, Jv $ 25, Bi $ 15,
Brittleness Index
and 30 #  # 90 .
Figure 6. The effect of rock brittleness index on specific rock mass boreability index.

5. Conclusions
200

190
Many engineering problems have complex relationship
180
between various parameters, and so it is difficult to construct
SRMBI (kN/cutter/mm/rev)

170
a pure mathematical model. Neural networks offer an alterna-
160 tive tool for establishing the relationship between such para-
150 meters. In this study, an ensemble neural network was
140 developed to establish the complex non-linear relationship
130 between the specific rock mass boreability index (SRMBI)
angle = 20, Bi = 20, UCS = 100
120 and four influential rock mass properties: rock compressive
angle = 20, Bi = 20, UCS = 200
110
strength, rock brittleness, joint spacing, and joint orientation.
angle = 20, Bi = 20, UCS = 300
100
Of these four rock mass parameters, rock unaxial compressive
0 5 10 15 20 25 30 35 40 45 strength and volumetric joint count have the greatest effect on
Jv the SRMBI. This result agrees with the theoretical analysis, i.e.
the formation of rock chips is mainly caused by the tensile
Figure 7. The effectof volumetric jointcount onspecific rockmassboreabilityindex.
failure of rock which is directly related to the rock strength
and joint properties. Parametric studies based on the ensemble
neural network are consistent with in situ measurements by
range. However, when Jv is greater than 30, its effect on SRMBI Bruland (1998).
becomes less significant. It should be noted that the statistical data were collected
It can be seen from figure 8 that the joint orientation angle  only from granitic rock masses in T05 and T06, Singapore.
has a relatively small influence on SRMBI compared with the The machine configurations are fixed in the TBMs used in
 128 Z. Zhao et al.
these two tunnels. Therefore care must be taken when the
model is used to predict the penetration rate in other rock
masses or using different types of machine. In this model,
the effect of the in situ stress on the penetration rate is not
taken into account.
References
Alber, M., Prediction of penetration and utilization for hard rock TBMs. In
Eurock’96, edited by G. Barla, pp. 721--725, 1996 (Balkema:
Rotterdam).
Alber, M., Advance rates of hard rock tbms and their effects on project econom-
ics. Tunnelling Underground Space Technol., 2000, 15(1), 55--64.
Barton, N., TBM Tunnelling in Jointed and Faulted Rock, 2000 (Balkema:
Rotterdam).
Bruines, P. A., The use of neurofuzzy modeling for performance prediction of
tunnel boring machines. In Modern Tunnelling Science and Technology,
Downloaded by [North Dakota State University] at 06:24 04 March 2015
edited by M. Kimura, T. Adachi, and K. Tateyama, pp. 583--588, 2001
(Swets & Zeitlinger: Lisse, The Netherlands).
Bruland, A., Hard rock tunnel boring. Doctoral thesis, Norwegian University of
Science and Technology, Trondheim, 1998.
Efron, B. and Tibshirani, R., An Introduction to the Bootstrap, 1993. (Chapman
& Hall: New York).
Farmer, I. W. and Glossop, N. H., Mechanics of disc cutter penetration. Tunnels
Tunnelling Int., 1980, 12(6), 22--25.
Gong, Q. M., Development of a rock mass characteristics model for TBM
penetration rate prediction. PhD thesis, Nanyang Technological
University, Singapore, 2006.
Gong, Q. M., Zhao, J., and Jiang, Y. S., In situ TBM penetration tests and rock
mass boreability analysis in hard rock tunnels. Tunnelling Underground
Space Technol., 2007, 22(3), 303--316.
Graham, P. C., Rock exploration for machine manufacturers. In Exploration for
Rock Engineering, edited by Z. T. Bieniawski, Vol. 1, pp. 173--180, 1976
(Balkema: Rotterdam).
Grandori, R., Sem, M., Lembo-Fazio, A., and Ribacchi, R., Tunnelling by
double shield TBM in the Hong Kong granite, in 8th International
Congress for Rock Mechanics, 1995, Vol. 1, pp. 569--574.
Grima, M. A., Bruines, P. A., and Verhoef, P. N. W., Modeling tunnel boring
machine performance by neuro-fuzzy methods. Tunnelling Underground
Space Technol., 2000, 15(3), 259--269.
Hamilton, W. H. and Dollinger G. L., Optimizing tunnel boring machine and
cutter design for greater boreability, in RETC Proceedings, 1979, Vol. 1,
pp. 280--296.
Hughes, H. M., The relative cuttability of coal measures rock. Min. Sci.
Technol., 1986, 3, 95--109.
Innaurato, N., Mancini, R., Rondena, E., and Zaninetti, A., Forecasting and
effective TBM performance in a rapid excavation of a tunnel in Italy. In
Proceedings of the 7th International Congress on Rock Mechanics, edited
by W. Wittke, pp. 1009--1014, 1991 (Balkema: Rotterdam).
Jakubek, S. M. and Strasser, T. I., Artificial neural networks for fault detection
in large-scale data acquisition systems. Eng. Appl. Artif. Intell., 2004, 17,
233--248.
Jiang, N., Zhao, Z. Y. and Ren, L. Q., Design of structural modular neural
networks with genetic algorithm. Adv. Eng. Software, 2003, 34, 17--24.
Laughton, C. and Nelson, P. P., The development of rock mass parameters for
use in the prediction of tunnel boring machine performance. In Eurock’96,
edited by G. Barla, pp. 727--733, 1996 (Balkema: Rotterdam).
McFeat-Smith, I., Mechanised tunnelling for Asia. Workshop Manual, IMS
Tunnel Consultancy Ltd, 1999.
McFeat-Smith, I. and Askilsrud, O. G., Tunnel boring machines in Hong Kong,
in RECT Proceedings, 1993, pp. 401--413.
Morimoto, T. and Hori M., Performance Characteristics of a Tunnel Boring
Machine from the Geomechanical Viewpoint. International Journal of
Rock Mechanics and Mining Sciences & Geomechanics Abstracts, 1986,
23(1), 55--66.
Nelson, P. P., Tunnel boring machine performance in sedimentary rock.
Doctoral dissertation, the Graduate School of Cornelll University, USA,
1983, p. 438.
Nelson, P. P., Yousof A. Al-Jalil and Laughton C., Improved strategies for TBM
performance prediction and project management. In RETC Proceedings,
1999, pp. 963--979.
O’Rourke, J. E., Spring, J. E., and Coudray, S. V., Geotechnical parameters and
tunnel boring machine performance at Goodwill Tunnel, California. In
Rock Mechanics Models and Measurements: Challenges from Industry,
edited by P. Nelson and S. E. Laubach, 1994 (Balkema: Rotterdam).
Ribacchi, R. and Lembo-Fazio, A., Influence of rock mass parameters on the
performance of a TBM in a gneissic formation (Varzo Tunnel). Rock
Mech. Rock Eng., 2005, 38(2), 105--127.
Rostami, J., Development of a force estimation model for rock fragmentation
with disc cutters through theoretical modeling and physical measurement
of crushed zone pressure. Doctoral dissertation, Department of Mining
Engineering, Colorado School of Mines, Golden, CO, 1997, p. 382.
Rumelhart, D. E., Hinton, G. E., and Williams, R. J., Learning internal repre-
sentations by error propagation. In Parallel Distributed Processing:
Foundation, edited by D. E. Rumelhart and J. L. McClelland, Vol. 1,
pp. 318--362, 1986 (MIT Press: Cambridge, MA).
Sapigni, M., Berti, M., Bethaz, E., Busillo, A., and Cardone, G., TBM perfor-
mance estimation using rock mass classifications. Int. J. Rock Mech. Min.
Sci., 2002, 39, 771--788.
Sharkey, A. (Ed.), Combining Artificial Neural Nets: Ensemble and Modular
Multi-Net Systems, 1999 (Springer: London).
Sherrod, H. P., Nonlinear regression analysis program. NLERG Manual, 1996,
p. 70.
Sollich, P. and Krogh, A., Learning with ensembles: how over-fitting can be
useful. In Advances in Neural Information Processing Systems 8, edited by
D. S. Touretzky, M. C. Mozer, and M. E. Hasselmo, pp. 190--196, 1996
(MIT Press: Cambridge, MA).
Sundin, N. O. and Wanstedt, S., In Rock Mechanics Models and Measurements:
Challenges from Industry, edited by P. Nelson and S. E. Laubach, 1994,
p. 8 (Balkema: Rotterdam).
Wanner, H. and Aeberli, U., Tunnelling machine performance in jointed rock, in
4th Congress of the International Society for Rock Mechanics, 1979, Vol. 1,
pp. 573--580.
Zhao, J., Gong, Q. M. and Eisensten, Z., Tunnelling through a frequently
changing and mixed ground: a case history in Singapore. Tunnelling
Underground Space Technol., 2007, 22(4), 388--400.
Zhao, Z. Y., Steel column under fire: a neural network based strength model.
Adv. Eng. Software, 2006, 37(2), 97--105.