You are on page 1of 21

- 1275 -

A Fuzzy Model for High-Speed Railway


Tunnel Convergence Prediction in Weak
Rock

Amoussou Coffi ADOKO

PhD Student, ASCE Student Member, Geotechnical Engineering Department,
China University of Geosciences (Wuhan)
e-mail: aadoko2004@yahoo.fr

Qing-Jun ZUO
PhD Student, Geotechnical Engineering Department, China University of
Geosciences (Wuhan)
e-mail: qjzcug@sina.com

Li WU
*

Professor, Faculty of Engineering, China University of Geosciences (Wuhan)
e-mail: lwu@cgu.edu.cn; tel: +8613871159076
*Corresponding author

ABSTRACT
Early prediction of tunnel diameter convergence is very crucial for a tunnel construction with
the New Austrian Tunnelling Method (NATM) because it can help in any quick required
adjustment of the design and consequently deadly hazards can be avoided. In this paper, a
fuzzy model for the convergence prediction of high-speed railway tunnel was formulated
based on Mamdani algorithm, triangular and trapezoidal membership functions using 135
datasets collected from a construction site in Hunan province (China). The inputs parameters
included surrounding rock mass class index (SRM index), ground engineering conditions
rating index (GEC index), tunnel overburden (H), rock density (d), distance between
monitoring station and working face (D), and the elapsed time (T). The model performance
was assessed by the variance account for (VAF), root mean square error (RMSE) indices and
the coefficient of determination (R
2
). In addition, a statistical regression analysis was
performed for comparison purpose with the proposed model. The results showed overall good
prediction accuracy. Finally a sensitivity analysis indicated that the SRM index and the GEC
index were the most effective parameters controlling the convergence, while the density had
least effect on the tunnel convergence.
KEYWORDS: Tunnel convergence, New Austrian Tunnelling Method (NATM),
Mamdani Fuzzy Systems.
I NTRODUCTI ON
The convergence of a tunnel refers to the amount of closure on the tunnel diameter and is
closely related to the tunnel deformation and its stability (Ozsan & Karakus, 2006). This is due to
the loss of stress-strain equilibrium state of the rock mass caused by the tunnel driving, resulting
in a redistribution of stress around the excavation and rock deformations (AFTES, 2001). The
Vol. 16 [ 2011] , Bund. Q 1276

amount of convergence is a key factor in tunnel serviceability and stability. For example, during
the construction stage of a tunnel, knowing the convergence trends is important for assessing the
tunnel stability and consequently the construction and design parameters can be appropriately
adjusted.
Especially, when the tunnel is designed under the New Austrian Tunnelling Method (NATM)
consideration, a systematic convergence measurement through instrumentation and monitoring
program, is required (Galler et al., 2009)
An accurate convergence monitoring needs an appropriate instrumentation of the tunnel.
Usually, the deformations are measured with geodetic surveying (total stations), laser scanners
(profilometers) and tape extensometers. The instruments are placed at regular intervals between 5
m and 50 m depending on the rock mass properties. The data are processed manually or
automatically as digital data and transmitted (e.g. by remote optical fiber sensors) to the data
processing unit or simply recorded to be used later (Kavvadas, 2005; Kolymbas, 2008; Simeoni &
Zanei, 2009). In addition, Automatic Tunnel Monitoring Systems (ATMSs) are gaining more
popularity mainly in complex urban conditions such as proximity of the foundation of super-
structures, bridges, subway lines, riverbed and soft ground (Chung et al., 2006). Special
monitoring methods for old tunnel rehabilitation are used as well (Torres et al., 2011). In china,
the use of tape extensometers is recommended for railway projects (P.R. China Professional
Standard, 2007).
The literature highlights that the measurement of tunnel convergence allows a better and
practical judgment of safety during and after tunnels construction (Asano et al., 2003;
Kontogianni & Stiros, 2002 & 2005; Liu et al., 2006; Bernard et al., 2006; Bezuijen & Talmon,
2006; Unlutepe & Ozener, 2008; Kontogianni et al., 2008 Jiang & Shen, 2010). Therefore the
prediction of the amount of convergence can help in optimizing the design by saving cost and
time. Also, an adequate interpretation of the monitoring data of the displacements is needed in
order to get a realistic prediction of the convergence (Schubert et al., 2004). In engineering
routine in China, statistical model (least square regression method for example) is systematically
implemented to estimate the tunnel convergence (P.R. China Professional Standard, 2007).
Although this method is well accepted, one of its weaknesses is related to the large amount of
data needed to make an accurate prediction. This is the main reason that motivated recently some
researchers to investigate alternative methods based on artificial neural network and expert
systems (Wang et al., 2004; Yoo & Kim, 2007; Lee & Akutagawa, 2009; Rafiai & Moosavi,
2011).
An accurate prediction of a tunnel deformation at the designing stage is very difficult even
though computational methods are often used. This is due, for instance, to complex geological
condition, unpredictability of rock material behavior and lack of knowledge about the real ground
condition resulting obviously into a kind of uncertain prediction (Lee & Akutagawa, 2009).
Fortunately, fuzzy set theory provides the theoretical framework to process such epistemic
uncertainties associated with the rock mass deformation. It has been suggested that fuzzy
inference systems (FISs) can be used as predictive model for complex non linear behavior of rock
mass. Many models based on FISs have been implemented successfully in the field of rock
engineering and geomechanics (Azimi et al., 2010; Monjezi & Rezaei, 2011; Jalalifar et al., 2011;
Yazig & Gogceoglu, 2010; Monjezi et al., 2010; Rezaei et al., 2011; Adoko & Wu, 2011).
Vol. 16 [ 2011] , Bund. Q 1277

In this paper, we are exploring the practical use of fuzzy inference system to predict the
amount of tunnel convergence and its velocity at the earlier stages of construction.
OVERVI EW OF FUZZY SET THEORY AND FUZZY
SYSTEMS
Fuzzy set theory was introduced by Zadeh (1965) and it provides the means for representing
epistemic uncertainty using set theory and describes the concept of gradualness and bipolarity
(Dubois & Prade, 2010). A fuzzy set is an extension of the concept of a crisp set. While a crisp
set only allows full membership or no membership to every element of the universe of discourse,
a fuzzy set allows for partial membership. Basically fuzzy set theory includes fuzzy set (type-1
fuzzy set or simply fuzzy set, type-2 fuzzy set and higher order fuzzy set), fuzzy functions, fuzzy
logic, fuzzy inference system, fuzzy probability, and hybrid fuzzy set (combination of other soft
computing techniques such as neural network and genetic algorithm).
Fuzzy set
A fuzzy set differs from an ordinary set on their membership or characteristic function. On
one hand, in the classical set theory, the characteristic function of a crisp set A, takes value 0 or 1
according to whether a given element of the universal set belong to A or not, respectively. On the
other, a fuzzy set has membership function with value varying from 0 to 1. Let X be the universal
set and x a generic element of X, for a given crisp set A, the characteristic function can be defined
as:
( )
1 if
0 if
A
x A
x
x A

(1)
On the contrary for a given fuzzy set A, the membership function is defined as:
( ) [ ] , 0,1
A
x A x (2)
In Eq. (2) [0,1] denotes the interval of real numbers from 0 to 1 inclusive of the end points.
Fuzzy set theory offers mathematical framework to represent gradual and partial membership to a
set. Therefore the nearer the membership function is to 1, the higher the degree of membership of
x in A. Fuzzy set allows many operations (union, intersection, products, relation etc) similar to
the case of classical set theory.
The determination of membership function is vital for any application of this theory. There
are different types and shapes of membership functions such as triangular, trapezoidal (linear),
Gaussian, bell-shaped etcThe most used membership functions are triangular and trapezoidal
for easy computation (Huang & Siller, 1997; Dixit & Dixit, 2008; Adoko & Wu, 2011) and can
be calculated by Eq. (3) and (4) respectively:
( ) max min , , 0
A
x a c x
x
b a c b

( (
=
( (


(3)
Vol. 16 [ 2011] , Bund. Q 1278

( ) max min ,1, , 0
A
x a d x
x
b a d c

( (
=
( (


(4)
Where ( )
A
x denotes the membership function and a, b, c and d are constant such as
a b c d < < <

Fuzzy I nference Syst ems ( FI Ss)
Fuzzy system models define relationships between input and output variables of a system by
using linguistic labels (i.e., fuzzy sets) in a collection of IF-THEN rules (Celikyilmaz & Trksen,
2009). To infer in a rule based on fuzzy model, the fuzzy proposition needs to be represented by
an implication function. The implication function is called fuzzy IF-THEN rule. A fuzzy IF-
THEN rule assumes the form if x is A then y is B (A is called the antecedent or premise and B
is the consequent; they are linguistic variables defined by fuzzy sets on universes of discourse X
and Y respectively). Generally, a fuzzy system consists of many rules. Each rule in the fuzzy
model is a relation such as [ ] : 0,1
i
R X Y and can be calculated by Eq. (5) (Grima, 2000):

( ) ( ) ( ) ( )
, , , 1, 2,...,
i i i
R A B
x y I x y i n = = (5)
Many FISs have been proposed such as Zadeh, Mamdani, Takagi-Seguno-Kang (TSK),
Mizumoto Tsukamoto and Singleton models (Celikyilmaz & Trksen, 2009). The most
commonly used in geotechnical engineering and related fields are the Mamdani-type and Sugeno-
type (Adoko & Wu, 2011).

Figure 1: General Illustration of a FIS structure
Vol. 16 [ 2011] , Bund. Q 1279

A generalized fuzzy inference structure has been explained by Celik Yilmaz and Trksen
(2009), illustrated in Fig. 1 and it includes the steps provided in the passage below: 1) the
fuzzification which consists in assigning a membership value to each new data vector; 2) the
aggregation of antecedents using MIN connective usually; 3) the implication step where the
model output fuzzy sets are identified. In Mamdani type inference (Mandani & Assilian, 1974)
AND connective is used as implication connective, where MIN operation is used to obtain the
membership function of the model output fuzzy set; 4) the aggregation of consequents using
usually MAX operator in Mamdani inference; 5) the defuzzification in which the fuzzy output is
converted into crisp (numeric) value using methods such as centroid of the area (COA) or center
of gravity, height, mean of the maximum, bisector of the area (BOA), smallest of maximum
(SOM) and largest of maximum (LOM) methods.
The method of centroid of the area (COA) is the most prominent and physically appealing of
all the defuzzification methods. The crisp value is obtained by taking the position of center of
gravity in a single axis. Thus, the crisp value is given by Eq. (6):

( )
( )

A
z
COA
A
z
z zdz
z
z dz

=
}
}
(6)
Where
COA
z is the crisp value of the output z and ( )
A
z is the aggregated output membership
function (Dixit & Dixit, 2008; Celikyilmaz & Trksen, 2009)
Mamdani Fuzzy Model
Proposed in 1975 by Mamdani and Assilian, this method was first applied to control a steam
engine and boiler combination by synthesizing a set of linguistic control rules obtained from
experienced human operators (Mamdani & Assilian, 1975).
In recent years, Mamdani algorithm has been used to solve complex problems in the field of
rock engineering (Adoko & Wu, 2011) and can be represented by Eq. (7) as below (Hamidi et al.,
2010):

1 1 2 2
if is and is and ... is then is B (for 1, 2,..., )
i i r ir i
x A x A x A y i k = (7)
where k is the number of rules,
1
,...,
r
x x are the antecedents (input) variables,
1
,...,
i ir
A A and
i
B are
linguistic terms (fuzzy set) and y is the output variable or consequence variable. In this model,
max-min composition is commonly used (Ross, 2004) and is represented by Eq. (8); it is
illustrated with a two-rule in Figure 2

( ) ( ) ( ) ( ) ( )
max min , (for 1, 2,..., )
k k k
C A B
z input x input y k r
( ( = =

(8)
where
k
C
,
k
A
and
k
B
are the membership functions of output z for the kth rule, input x and
input y respectively (Ross, 2004; Monjezi & Rezaei, 2011)


Vol. 16 [ 2011] , Bund. Q 1280


Figure 2: The Mamdani FIS (after, Mamdani & Assilian, 1975)
DEVELOPMENT OF THE PROPOSED MODEL
Overview of t he geo- engineering condit ions of t he
const ruct ion sit e
The construction site where data were collected is located in Hunan province and is a part of
a construction project of 15 tunnels and auxiliary excavations for high speed railway (350km/h)
which are supposed to respond to the increasing demand for fast transportation between
Changsha and Kunming in southern China.
The tunnel of this study, is a two track railway tunnel, 912 m long located in a straight line;
the overburden varies between 25 m and 141.53 m. The section to be driven has 1262 cm of
height and 1320 cm of span making sure that the minimum cross section for serviceability is 100
m
2
according to the specifications of the Chinese code for design of high speed railway tunnel
(P.R. China Professional Standard TB, 2009).
The lithology of the area includes mainly a succession of weathered calcareous slate and
tuffaceous slate rocks of the Quaternary Holocene, covered by silty clay and gravel soil. The rock
occurrence orientation (dip direction/dip) is 296/39 for the calcareous slate layers while the
tuffaceous 210/17 with joint strikes parallel to the tunnel axis affecting fair favorably the tunnel
excavation (Bieniawski, 1989). Geological explorations revealed from weak to strong weathered
rock composed of clay minerals, sericite, feldspar quartz and other minerals; non filled cracks are
developed while rock texture and structure are unclear. The rock of the area is considered as of

Vol. 16 [ 2011] , Bund. Q 1281

poor quality and belongs to class III, IV and V according to the Chinese code GB 50218-94 for
rock classification (P.R. China National Standard, 1995 & 2008).
Consequently, the selected construction method is based on the design principle of the New
Austrian Tunnelling Method (NATM) with partial excavations supported temporarily by
shotcrete, rock bolts wire mesh and steel ribs as first lining.
Dat a collect ion and met hods
In this paper, the data that has been used to develop the fuzzy model include the tunnel
convergence (crown and sides), rock mass geomechanical properties as well as ground conditions.
They were collected from the convergence monitoring program and geotechnical investigation
(field and laboratory test). We recorded the tunnel deformation data after the first lining was
posed, using digital extensometers (SGS-1) and total station which have been processed
according to the technical regulation for such work in China (P.R. China Professional Standard,
2007).
Firstly, we used 475 datasets (each dataset consists of two parameters, the measured
cumulative convergence and the elapsed time) for the statistical regression prediction as a routine
data processing method for convergence monitoring. During the preparation of the database, we
sorted the datasets in accordance with each tunnel segment where the rock mass of each
monitoring sections showed similar characteristics, otherwise excellent prediction performance
will not be guaranteed (P.R. China Professional Standard, 2007). Secondly, we employed 135
datasets (each dataset consists of seven input and output parameters) obtained from field and
laboratories measurements to form the fuzzy model database which has been divided into two:
122 datasets (90%) for the model construction and the remaining for the model testing. For the
purpose of selecting the testing datasets, we utilized a sorting method to make sure that every
class of data is covered. Next, we considered the aspect of the statistical consistency of the testing
dataset by choosing data based on a regular interval. The testing dataset was used for the model
performance evaluation. Also, we performed a sensitivity analysis in order to identify the most
effective input parameters controlling the tunnel convergence using the original datasets.
Finally, the proposed model which is intended to predict the deformation of a tunnel driven in
weak rock condition employs Mamdani fuzzy inference systems with triangular and trapezoidal
membership functions. This model has been found appropriate to our research based on the
results of similar research carried out by other authors such as Monjezi and Rezaei (2011). The
model has been computed using the fuzzy toolbox of MATLAB R2010a software.
I nput s and out put paramet ers of t he fuzzy model
As far as input and output parameters were concerned, factors that influence a NATM tunnel
behavior such as tunnel geometry, ground excavation and support conditions were used to
identify six inputs of the proposed FIS.
Firstly, class indices were assigned to the surrounding rock mass based on the rock
classification of the Chinese code GB 50218-94 in which the classification depends on
parameters such as uniaxial compressive strength, intactness factor, p-wave velocity (very similar
to the well known Rock Quality Designation) as well as groundwater condition and initial stress
Vol. 16 [ 2011] , Bund. Q 1282

state (Lee et al., 1996; P.R. China National Standard, 1994 & 2008). The class rating indices of
the surrounding rock mass (SRM index) are shown in table 1. Secondly, the ground engineering
(excavation and support) conditions were also represented by a class rating system. The rating
value is on a scale of 100% to 0% (from certain to highly uncertain conditions). It had been
obtained from a quick survey on the basis of expert knowledge and experience on the
construction site (designers and contractors). This index described the difficulties on the choice
and the implementation of the excavation and support methods which depend mainly on the rock
class. It should be noticed that this class rating is not intended to make any ground classification,
even though it does. Rather it takes into consideration the inherent uncertainties associated with
the difficult conditions (unpredicted faults, joints and groundwater etc.) in which the required and
optimal design of support is quasi impossible and consequently the deformations of the temporary
lining cannot be properly foreseen. Fuzzy inference systems allow quantifying effectively such
uncertainties in rock engineering (Adoko & Wu, 2011). Table 2 gives details on the ground
engineering conditions rating index (GEC index). The other inputs include, rock density (d), the
tunnel overburden (H), the distance between the monitoring station and the tunnel heading face
(D) and the elapsed time (T) after the working face has passed the monitoring station.
The output parameters are: the amount of the total convergence (C
CUM
) and the velocity of the
convergence (C
VEL.
) for the crown and the sides (top and bottom heading). Table 3 provides a
sample of the datasets used while table 4 shows the variation of the model parameters. It should
be emphasized that the proposed fuzzy model do not have two outputs; rather we developed two
models (one for the convergence and the other for the convergence speed) but with same structure
and input parameters.
Table 1: Surrounding rock mass index (SRM Index) and tunnel segmentation

Table 2: Ground engineering conditions rating index (GEC Index)


Chainage
000m-
065m
065m-
140m
140m-
260m
260m-
370m
370m-
460m
460m-
630m
630m-
822m
Tunnel Segment No1 No2 No3 No4 No5 No6 No7
Rock classification V IVb IIIa IVa IIIb IVb- IVa V
Class Index 5 4 1 3 2 4-3 5
Ground
Engineering
Conditions
Very good Good Acceptable Fairly difficult Very difficult
Class Rating
Index
100-91 90-71 70-61 60-51 50-00
Effect on the
foreseen
deformation
(convergence)
deformation
fits within
expected
range
deformation
fits within
expected
range after
few
adjustment
Expected
deformation
are known
fairly
Expected
deformation are
known fairly with
special measure for
support
Expected
deformation are
unknown despite
of special
measures
Vol. 16 [ 2011] , Bund. Q 1283

Table 3: Sample of data used in the model formulation
Table 4: Statistical description of the model parameters
Type of data Parameters Symbol Unit Minimum Maximum
Inputs

Surrounding rock class
index
(SRM index) logic - -
Ground engineering
condition rating index
(GEC index) logic - -
Density d g/cm
3
2.3 2.8
Overburden H m 25 143
Time
C
CUM
T day 20 73
C
VEL.
T day 1 73
Distance to working face D m 25 65
Output Cumulative Convergence C
CUM
mm 2.5 37
Convergence velocity C
VEL.
mm/day -0.5 3.20

St at ist ical regression analysis
In reference to the convergence monitoring data processing, the Chinese code recommends
the use of simple least square regression analysis with exponential, logarithmic or hyperbolic
fitting functions (P.R. China Professional Standard, 2007).
In this study, we computed a nonlinear simple regression model of the tunnel convergence
based on the least square method (LSM) in Excel, where the fitting curves were exponential and
logarithmic functions according to the rock geomechanical properties of the tunnel segments and
the monitoring stations for the purpose of achieving the best performances.
A logarithmic regression of the convergence (crown) is provided in Fig. 3. The results of the
statistical regression analysis of the cumulative convergence C
CUM
as well as the performances
(coefficient of determination R
2
) in which the best fitting functions have been considered, are
summarized in the table 5. The estimated velocity of convergence (C
VEL.
) can be obtained by the
first derivative of the computed regression functions.

No SRM
index
GEC
index
Density
(g/cm
3
)
H
(m)
D
(m)
T
(day)
C
CUM
(mm) C
VEL
(mm/day)
Crown Sides Crown Sides
1 1 88 2.70 126 55 20 16.8 5.6 0.65 0.8
2 3 70 1.98 80 25 60 34.65 9.02 0.2 0.10
3 3 70 2.51 70 30 70 19.6 6.4 0.10 0.15
4 2 80 2.20 50 42 35 25.6 12.8 0.7 1.07
5 1 88 2.56 90 35 25 22.5 8.12 0.19 0.06
6 2 85 2.30 90 60 60 35.4 15.5 0.24 0.13
7 4 60 2.10 40 50 35 23.8 19.4 1.3 0.7
8 5 50 2.10 25 65 25 22.1 16.7 2.19 1.43
9 5 45 1.95 35 60 45 28.7 14.8 1.3 0.72
10 1 95 2.8 82 45 45 12.5 7.6 0.02 0.09
11 4 55 2.20 50 25 30 16.6 18.32 1.4 2.01
12 2 80 2.50 135 30 73 32.0 22.45 0.15 0.3
Vol. 16 [ 2011] , Bund. Q 1284


Figure 3: Crown convergence prediction with logarithmic fitting function
Table 5: Summary of the statistical regression analysis for the convergence prediction
Segment Crown Side (Top heading) Side (Bottom heading)
No1
( ) 6.76ln 1.72
0.97
y x
R
= +
=

0.068
2.09
0.90
x
y e
R
=
=

0.065
0.38
0.84
x
y e
R
=
=

No2
( ) 6.37ln 1.53
0.99
y x
R
= +
=

( ) 2.96ln 0.85
0.99
y x
R
= +
=

( ) 2.77ln 2.02
0.96
y x
R
= +
=

No3
( ) 4.2ln 2.32
0.98
y x
R
= +
=

( ) 2.21ln 0.76
0.97
y x
R
= +
=

( ) 1.98ln 1.04
0.98
y x
R
= +
=

No4
( ) 5.51ln 1.45
0.93
y x
R
= +
=

0.052
1.96
0.92
x
y e
R
=
=

0.134
0.32
0.85
x
y e
R
=
=

No5
( ) 5.22ln 1.09
0.96
y x
R
= +
=
( ) 5.28ln 0.78
0.98
y x
R
= +
=
( ) 2.01ln 0.84
0.96
y x
R
= +
=

No6
( ) 5.85ln 1.62
0.96
y x
R
= +
=
0.057
2.01
0.91
x
y e
R
=
=

0.071
0.28
0.88
x
y e
R
=
=

No7
( ) 6.8ln 1.05
0.98
y x
R
= +
=

0.094
0.93
0.92
x
y e
R
=
=

0.081
0.37
0.89
x
y e
R
=
=

The results suggested that the logarithmic regression yielded to the best performance with
values of R
2
higher than 0.93
Fuzzy Model formulat ion
In this study, a new model based on Mamdani FIS is introduced for the prediction of tunnel
convergence after posing the first lining (temporary lining). For that matter, the SRM index, GEC
index, density (d), overburden (H), distance between the monitoring station and the tunnel
heading face (D) and the elapsed time (T) are used as input. Fig. 4 illustrates the structure of the
proposed fuzzy model with the parameters for the convergence prediction; the structure for the
convergence velocity is the same, as well.
As far as membership functions were concerned, we chose triangular and trapezoidal
membership functions for the input and output parameters due to their construction simplicity and
Vol. 16 [ 2011] , Bund. Q 1285

the excellent results that had been achieved by employing them in the area of rock engineering
and geotechnical engineering (Adoko & Wu, 2011). Fig. 5-12 show the membership function
where five levels of classification are used: from very low (VL) to very high (VH) with
low (L), medium (M) and high (H) in between.
For the construction of the if-then rules, the theoretical number of rules is 15625 (6 inputs
with 5 levels is 5
6
) which can be computer consuming (Azimi et al., 2010). Usually, a rule
reduction method is required in order to reduce the amount of rules by eliminating those which
are unlike to happen. For example the rule: if SRM index is VH and GEC index is VL and
density is VH and overburden is VH and distance is VH and time is VH then convergence is VL
is impossible. The literature reveals different methodology of rule reduction such as effective
rule method (Azimi et al., 2010), design of experience (DOE) methods (Rafiai et al., 2011) or
methods based on experts experience (Monjezi & Rezaei, 2011). In this paper, we developed in
total 118 rules using the experts experience method because appropriate to our case and practical
results showed that there were many rules which contradicted the principles and design of
NATM. For example, if rock engineering quality is good, then the expected convergence cannot
be higher. So the rules were obtained by combining one by one each level (5 in total) of the input
parameters and the rules having realistic consequents, were selected. Also the fuzzy operator
NOT (complement of fuzzy set) was used at time and helped in reducing considerably the amount
of rules. Finally the Mamdani algorithm was applied appropriately to map the inputs to the output
on the basis of the constructed fuzzy rules of the model.
The last stage of the model corresponds to the defuzzification where the numeric output
values were obtained from the fuzzy output by using centroid of area (COA) defuzzier. It has
been shown that the COA gives very acceptable results for a wide range of problems in
geotechnical engineering field (Adoko & Wu, 2011).
Finally, the results of the model are illustrated by the fuzzy reasoning mechanism (If-then
rules editor) and rule viewer in the MATLAB R2010a environment as shown in Fig. 13-15
(pieces of shoot screen of the displayed results). The proposed fuzzy model suggests that the
convergence can be predicted when input data is entered. For instance, in Fig. 14, the output is
predicted as 8.29 mm if SRM index is 1, GEC index is 88, density (d) is 2.56 g/cm
3
, overburden
(H) is 90m, the distance from the monitoring station to the working face (D) is 35m and the
elapsed time (T) is 25 days.
Similarly the convergence velocity is predicted as 0.24 mm/day if SRM index = 1, GEC
index =88, d = 2.56 g/cm
3
, H = 90m, D = 35m and T = 25 days (Fig. 15)


Vol. 16 [ 2011] , Bund. Q 1286


Figure 4: Structure of the proposed fuzzy model
Figure 5: Membership
function of SRM Index
Figure 6: Membership
function of GEC Index
Figure 7: Membership
function of Density
Figure 8: Membership
function of Overburden
1 2 3 4 5
0
0.2
0.4
0.6
0.8
1
SRM Index
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
40 50 60 70 80 90
0
0.2
0.4
0.6
0.8
1
GEC Index
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
2.3 2.4 2.5 2.6 2.7 2.8
0
0.2
0.4
0.6
0.8
1
Density
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
20 40 60 80 100 120 140
0
0.2
0.4
0.6
0.8
1
Overburden
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
Vol. 16 [ 2011] , Bund. Q 1287

Figure 9: Membership
function of Distance (D)
Figure 10: Membership
function of Time (T)
Figure 11: Membership
function of Convergence
Figure 12: Membership
function of Convergence
velocity

Figure 13: If-then rules editor for the model (convergence)
20 30 40 50 60 70
0
0.2
0.4
0.6
0.8
1
Distance (D)
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
20 30 40 50 60 70
0
0.2
0.4
0.6
0.8
1
Time (T)
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
5 10 15 20 25 30 35 40
0
0.2
0.4
0.6
0.8
1
Convergence
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
-0.5 0 0.5 1 1.5 2 2.5 3 3.5
0
0.2
0.4
0.6
0.8
1
Convergence velocity
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
VL L M H VH
Vol. 16 [ 2011] , Bund. Q 1288


Figure 14: Rule viewer showing the results of the fuzzy reasoning

Figure 15: Rule viewer showing the results of the fuzzy reasoning (convergence
velocity)
Predict ion performance of t he model
With the purpose of evaluating the prediction performance of the model, the remaining
datasets (testing datasets) were used to run the proposed model. The indicators that were selected
to assess the performance included the variance account for (VAF) and root mean square error
(RMSE) for being some of the most common used for fuzzy predictive models (Cabalar et al.,
2010; Monjezi et al., 2010; Monjezi & Rezaei, 2011). While the VAF index represents the
percentage ratio of the difference between the variances of the measured and predicted outputs,
and the variance of the predicted outputs, the RMSE index measures the deviation between the
measured and predicted data. On one hand, the closer the value of the VAF is to 100%, the
smaller the variability is; and therefore the better the model prediction capabilities. On the other,
lower value of the RMSE suggests better performance (Alvarez Grima & Babuska, 1999). The
values of these indexes are calculated using Eqs. (9) and (10):

( )
( )
var
100 1
var
i i
i
y y
VAR
y
| |
=
|
|
\ .

(9)
( )
2
1
1
N
i i
i
RMSE y y
N
=
=


(10)
where VAR stands for the variance,
i
y is the i
th
measures element,
i
y

is the i
th
predicted element
and N is the number of dataset. As far as the crown convergence of the tunnel segment No1 was
concerned, the calculated VAFs were 89.95% and 75.20% for the convergence (C
CUM
) and its
Vol. 16 [ 2011] , Bund. Q 1289

velocity (C
VEL.
), respectively; while the calculated RMSEs were 0.20 mm and 0.85 mm/day,
respectively. The results are provided in table 6 and table 7.
Table 6: Variation of the calculated indices of performance


The indices of performance indicate very high prediction accuracy for the amount of total
convergence (C
CUM
). However, as far as the convergence velocity is concerned, the model is not
very accurate showing fairly good and acceptable the performance.
In addition, the coefficient of determination (R
2
) between the measured and predicted values
(by the fuzzy model), obtained from a regression fitting, comes to corroborate the model
performance. For example, the calculated (R
2
) was 0.96 for the crown convergence of the tunnel
segment No1, indicating high fitting between the output of the model and the field measurements
(Fig. 16). The values of (R
2
) are summarized in table 8, which vary from 0.88 to 0.97 and 0.65 to
0.80 for the cumulative convergence and its velocity, respectively.
Finally the performances of the fuzzy model (tables 6-8) were compared to those of the
statistical model (table 5). It is observed that the performances of the fuzzy and the statistical
model are very close to each other, suggesting that the fuzzy model can be used to predict the
tunnel convergence.

Figure 16: Comparison between measured and predicted convergence
from the fuzzy model.

Table 7: Performance indices VAR (%) and RMSE of the fuzzy model
Crown Side (Top heading) Side (Bottom heading)
y = 1.0737x - 0.2222
R = 0.9621
0
5
10
15
20
25
0 5 10 15 20 25
P
r
e
d
i
c
t
e
d

c
o
n
v
e
r
g
e
n
c
e

(
m
m
)
Measured convergence (mm)
C
CUM
C
VEL

VAF 78.54 % - 95.08% 59.23 % -78.40%
RMSE 0.09 mm- 0.32mm 0.40 mm/day- 2.05mm/day
Vol. 16 [ 2011] , Bund. Q 1290

C
CUM
C
VEL.
C
CUM
C
VEL.
C
CUM
C
VEL.

VAR RMSE VAR RMSE VAR RMSE VAR RMSE VAR RMSE VAR RMSE
No1 89.95 0.2 75.20 0.85 84.15 0.21 69.6 1.51 84.15 0.27 59.23 2.05
No2 91.04 0.22 72.10 0.41 78.54 0.32 70.1 1.62 90.61 0.19 61.3 1.5
No3 95.08 0.09 74.33 0.88 80.30 0.29 65.8 1.03 91.32 0.15 70.23 1.13
No4 90.14 0.13 70.12 0.71 81.16 0.28 70.1 1.09 88.75 0.26 65.8 1.2
No5 91.05 0.11 77.45 0.76 82.33 0.25 72.4 1.28 86.88 0.27 78.4 1.51
No6 89.33 0.21 65.89 0.40 86.51 0.18 69.3 1.31 79.50 0.3 75.3 1.45
No7 90.21 0.25 76.77 1.02 83.52 0.22 68.2 1.45 92.07 0.14 69.8 1.6
Table 8: Values of R
2
indicating a comparison between measured and predicted convergence
using simple linear regression
Crown Side (Top heading) Side (Bottom heading)
C
CUM
C
VEL.
C
CUM
C
VEL.
C
CUM
C
VEL.

No1 0.96 0.80 0.94 0.69 0.95 0.65
No2 0.91 0.79 0.95 0.62 0.90 0.68
No3 0.98 0.75 0.97 0.65 0.92 0.69
No4 0.90 0.76 0.93 0.70 0.90 0.61
No5 0.92 0.81 0.92 0.68 0.92 0.60
No6 0.95 0.75 0.95 0.71 0.95 0.75
No7 0.88 0.79 0.90 0.73 0.91 0.72
PARAMETERS SENSI TI VI TY ANALYSI S
Sensitivity analysis refers to the study of how the uncertainty in the output of a statistical
model can be attributed to different variations in the inputs of the model (Saltelli et al., 2008). It
can be used to determine the most effective input parameters on the output (objective) parameter
by applying methods such as Cosine Amplitude Method (CAM), Relative Strength Effect (RSE),
Fractional Factorial Design (FFD) as well as Analysis of Variance (ANOVA) technique: Fourier
Amplitude Sensitive Test or the Solobs indices (Cukier et al., 1978; Sobol 1993; Saltelli et al.,
1999). In this study, the sensitivity analysis was performed using (CAM) for being one the most
used methods in the field of rock engineering (Jong & Lee, 2004; Monjezi et al., 2010; Bahrami
et al., 2011).
This method is based on a pairwise comparison of two data pairs represented by the output
and each of the inputs which are expressed in common X-space resulting in a data array X
defined as:
{ }
1 2 3
, , ,...,
n
X X X X X = (11)
n, being the number of the inputs and output parameters. Each of the elements, X
i
, in the data
array X is a vector of m coordinates (m samples of data) defined as:
{ }
1 2 3
, , ,..., 1, 2,...,
i i i i im
X x x x x i n = = (12)
Therefore the strengths of relations (r
ij
) between the output and each of the input parameters,
can be computing using Eq. (13):
V


E
simila
crown
distan
densit
F
the G
F
distan
T
T
conve
F
T
Vol. 16 [ 20
Eq. (13) sugge
ar), the more
n convergenc
nce (D) are t
ty (d) is the le
or the amoun
EC index and
or the conver
nce (D).
The least effec
These results
ergence, espec
Figure 17: Str
Table 9: Valu
r
ij

SRM Index
GEC Index
Density (d
Overburden
Distance (D
Time (T)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SR
r
i
j
11] , Bund.
ests that the m
the input is e
ce computed
the most eff
east. The com
nt of cumulati
d the distance
rgence speed
ctive input par
show good c
cially with th
rength of the r
ues of r
ij
betw
C
CUM
x 0.98
x 0.94
d) 0.70
(H) 0.90
D) 0.93
0.92
0.98
RM Index GE
. Q
1
m
ij ik
k
r x x
=
=

more (r
ij
) is cl
effective on th
in Excel. It
fective inputs
mputed values
ve convergen
(D).
, the most eff
rameters on th
conformity wi
e SRM index
relation (r
ij
) b
ween the conve
of the
Crown
M
C
VEL.
8 0.95
4 0.75
0 0.69
0 0.91
3 0.96
2 0.98
0.94
0
C Index Den
Mode
2
1
m
jk ik
k
x x
=

lose to unity
he output. Fi
is shown tha
s parameters
s of r
ij
are pro
nce, the most
ffective param
he convergen
ith the princi
x and the dista
between the co
ergence, conv
e fuzzy mode
Side (Top
C
CUM

0.97
0.91
0.72
0.93
0.90
0.90
0.7
0.9
nsity Overbu
(H)
el input param
2
1
m
k jk
k
x
=

(the data vec


g. 17 illustrat
at the SRM
on the amou
vides in table
effective para
meters are: SR
nce, is the den
iple of tunnel
ance (D) (Sch
onvergence a
vergence spee
el
p heading)
C
VEL.

0.94
0.70
0.72
0.92
0.95
0.98
9
0.93
urden
)
Distance
eters
ctors are almo
tes the results
index, GEC
unt of conve
e 9. The resul
ameters are th
RM index, tim
nsity (d).
l deformation
hubert et al., 2
and the input p
ed and the inp
Side (Bottom h
C
CUM

0.96
0.95
0.75
0.80
0.80
0.89
0.92
(D) Time (T)
129
(13
ost collinear o
s of
ij
r for th
index and th
ergence, whil
lts indicate tha
he SRM index
me (T) and th
n and diamete
2004).

parameters
put parameter
heading)
C
VEL.

0.96
0.78
0.79
0.69
0.93
0.98
1
3)
or
he
he
le
at:
x,
he
er
rs
Vol. 16 [ 2011] , Bund. Q 1292

CONCLUSI ON
On the basis of the results of this study, it is concluded that the proposed fuzzy model is
capable of predicting the tunnel convergence with good accuracy. This model employed
Mamdani fuzzy inference technique and the parameters affecting a tunnel deformation such as
tunnel geometry, rock mass geomechanical properties and ground engineering conditions. The
performance indices used to assess the model performance, were the variance account for (VAF),
the root mean square error (RMSE) and determination coefficient (R
2
). They showed very good
results indicating that the fuzzy model can be used as convergence prediction model. A
comparison with the statistical model corroborated the fuzzy model capability as well.
However, as far as the convergence velocity is concerned, the performance indices suggested
that the fuzzy model didnt perform excellently, but still can predict the convergence velocity
with fair accuracy. Further research should be done to improve the fuzzy model capacity in
predicting the convergence velocity by for instance, increasing the datasets used and determining
the optimum parameters (memberships function, effective rules, etc.) of the fuzzy system.
Finally, the results of the sensitivity analysis based on the cosine amplitude method, indicated
that the surrounding rock class, the distance between the monitoring station and the working face,
the elapsed time and the ground engineering condition are the most important parameters
controlling the tunnel convergence.
ACKNOWLEDGEMENT
The authors appreciate the support of China University of Geosciences (Wuhan)
REFERENCES
1. Adoko, A. C. and L. Wu (2011) Fuzzy inference systems-based approaches in
geotechnical engineering- a review, Electronic Journal of Geotechnical Engineering,
Vol. 16 M, pp 1543-1558.
2. AFTES (2002) The convergence-confinement method: Recommendation GT7R6A1, No.
174.
3. Alvarez Grima, M. and R. Babuska (1999) Fuzzy model for the prediction of unconned
compressive strength of rock samples, International Journal of Rock Mechanics and
Mining Sciences, Vol. 36, No.3, pp 339-349.
4. Asano, T, M. Ishihara, Y. Kiyota, H. Kurosawa and S. Ebisu (2003) An observational
excavation control method for adjacent mountain tunnels, Tunnelling and Underground
Space Technology, Vol. 18, No. 2-3, pp 291-301.
5. Azimi, Y., M. Osanloo, M. Aakbarpour-Shirazi and B.A. Aghajani (2010) Prediction of
the blastability designation of rock masses using fuzzy sets, Int J Rock Mech Min, Vol.
47, No.7, pp 1126-1140.
6. Bahrami A., M. Monjezi, K. Goshtasbi and A. Ghazvinian (2011) Prediction of rock
fragmentation due to blasting using artificial neural network, Engineering with
Computers, Vol. 27, No. 2, pp 177-181.
Vol. 16 [ 2011] , Bund. Q 1293

7. Bernard, R, J. Pacovsk and I. Zemnek (2006) Geo-monitoring performed during the
construction of the Valk highway tunnels, Tunnelling and Underground Space
Technology, Vol. 21, No. 3-4, pp 226-227.
8. Bezuijen, A and A.M. Talmon (2006) Calculation models based on monitoring during
tunnel construction, Tunnelling and Underground Space Technology, Vol. 21, No. 3-4,
pp 380-381.
9. Bieniawski, Z.T. (1989) Engineering rock mass classification, Wiley, New York.
10. Cabalar, A.F., A. Cevik, C. Gokceoglu and G. Baykal (2010) Neuro-fuzzy based
constitutive modeling of undrained response of Leighton Buzzard Sand mixtures, Expert
Systems with Application, Vol. 37, No.1, pp 842-851.
11. Celikyilmaz, A., I.B. Turksen (2009) Modeling uncertainty with fuzzy logic with recent
theory and applications, Springer-Verlag, Berlin Heidelberg.
12. Chung, H-S., B-S. Chun, B-H. Kim and Y-J. Lee (2006) Measurement and analysis of
long-term behavior of Seoul metro tunnels using the automatic tunnel monitoring
systems, Tunnelling and Underground Space Technology, Vol. 21, No. 3-4, pp 316-317.
13. Cukier, R.I., H.B. Levine and K.E. Shuler (1978) Nonlinear sensitivity analysis of
multiparameter model systems, Journal of Computational Physics, Vol. 26, No. 1, pp 1-
42.
14. Dixit, P.M. and U.S. Dixit (2008) Modeling od metal forming and machining processes
by finite element and soft computing methods, Springer-Verlag, London.
15. Galler, R, E. Schneider, P. Bonapace, B. Moritz and M. Eder (2009) The new guideline
NATM - the Austrian practice of conventional tunnelling, BHM Berg-und
Httenmnnische Monatshefte, Vol. 154, No. 10, pp 441-449.
16. Grima, M. A. (2000) Neuro-fuzzy modelling in engineering geology, A.A.Balkem,
Rotterdam.
17. Hamidi, J.K. et al.(2010) Application of fuzzy set theory to rock engineering
classification systems: an illustration of the rock mass excavability index, Rock Mech
Rock Eng, Vol. 43, No.3, pp 335-350.
18. Huang, Y. T. and T.J. Siller (1997) Fuzzy representation and reasoning in geotechnical
site characterization, Computers and Geotechnics, Vol. 21, No. 1, pp 65-86.
19. Jalalifar, H. et al. (2011) Application of the adaptive neuro-fuzzy inference system for
prediction of a rock engineering classification system, Computers and Geotechnics, Vol.
38, No. 6, pp 783-790
20. Jiang, B.S. and C.R. Shen (2010) Study of Surrounding Rock's Stability Based on
Monitoring Measurement of Highway Tunnel, Proceedings of the 2010 International
Conference on Electrical and Control Engineering (ICECE 2010), pp 2059-2062.
21. Jong, Y-H., C-I. Lee (2004) Influence of geological conditions on the powder factor for
tunnel blasting, International Journal of Rock Mechanics and Mining Sciences, Vol. 41,
Supplement 1, pp 533-538.
22. Kavvadas, M.J. (2005) Monitoring ground deformation in tunnelling: Current practice in
transportation tunnels, Engineering Geology, Vol. 79, No. 1, pp 93-113.
Vol. 16 [ 2011] , Bund. Q 1294

23. Kolymbas, D. 2008 Tunnelling and Tunnel Mechanics A Rational Approach to
Tunnelling, Springer-Verlag, Berlin Heidelberg.
24. Kontogianni, V.A. and S. C. Stiros (2002) Predictions and observations of convergence
in shallow tunnels: case histories in Greece, Engineering Geology, Vol. 63, No. 3-4, pp
333-345
25. Kontogianni, V.A. and S. C. Stiros (2005) Induced deformation during tunnel
excavation: Evidence from geodetic monitoring, Engineering Geology, Vol. 79, No. 1-2,
pp 115-126.
26. Kontogianni,V., C. Papantonopoulos and S. Stiros (2008) Delayed failure at the
Messochora tunnel, Greece, Tunnelling and Underground Space Technology, Vol. 23,
No. 3, pp 232-240.
27. Lee, C.F., Wang Sijing and Yang Zhifu (2006) Geotechnical aspects of rock tunnelling
in China, Tunnelling and Underground Space Technology, Vol. 11, No. 4, pp 445-454.
28. Lee, J-H., S. Akutagawa (2009) Quick prediction of tunnel displacements using
Artificial Neural Network and field measurement results, International Journal of the
JCRM, Vol. 5, No. 2, pp 53-62.
29. Liu, Z-C., W-J. Li, S-M. Zhang, Y-Q. Zhu (2006) Synthetical analysis on monitoring of
Wushaoling railway tunnel, Tunnelling and Underground Space Technology, Vol. 21,
No. 3-4, pp 363-364.
30. Mamdani, E. H. and S. Assilian (1975). An experiment in linguistic synthesis with a
fuzzy logic controller. International Journal of Man-Machine Studies, Vol. 7, No. 1, pp 1-
13.
31. Monjezi, M. and M. Rezaei (2011) Developing a new fuzzy model to predict burden
from rock geomechanical properties, Expert Systems with Application, Vol. 38, No. 8,
pp 9266-9273.
32. Monjezi, M., M. Rezaei and A. Yazdian (2010) Prediction of backbreak in open-pit
blasting using fuzzy set theory, Expert Systems with Applications, Vol. 37, No. 3, pp
2637-2643
33. Navarro Torres, V.F et al. (2011) Application of a new convergence measurement
technique for the rehabilitation of old Rossio railway tunnel, Lisbon, Geomechnanics
and Geoengineering, Vol. 6, No. 2, pp 109-118.
34. Ozsan, A. and M. Karakus (2006) Site investigations and convergence measurements for
a twin metro tunnel driven in Ankara clay, Turkey, Proceedings of the 10th IAEG
International Congress (IAEG2006), Paper No. 504.
35. P.R. China National Standard (1995), GB 50218-94 Standard for Engineering
Classification of Rock Masses. (in Chinese)
36. P.R. China National Standard (2008), Technical Guidelines for Geological Exploration
for Railway Tunnel, No. 105. (in Chinese)
37. P.R. China Professional Standard (2007), TB 10121-2007 Technical code for monitoring
measurement of railway tunnel. (in Chinese)
38. P.R. China Professional Standard (2009), TB 10621-2009 Code for Design of High Speed
Railway. (in Chinese)
Vol. 16 [ 2011] , Bund. Q 1295

39. Rafiai, H. and M. Moosavi (2011) An approximate ANN-based solution for
convergence of lined circular tunnels in elasto-plastic rock masses with anisotropic
stresses , Tunnelling and Underground Space Technology, (In Press, Corrected Proof)
40. Rezaei, M., M. Monjezi and V.A. Yazdian (2011) Development of a fuzzy model to
predict flyrock in surface mining, Safety Science, Vol. 49, No.2, pp 298-305.
41. Ross, T.J. (2004) Fuzzy Logic with Engineering Applications, John Wiley and Sons.
42. Saltelli, A. et al. (2008) Global Sensitivity Analysis- The Primer, John Wiley and Sons.
43. Saltelli, A., S. Tarantola and K.P.-S. Chan (1999) Quantitative model-independent
method for global sensitivity analysis of model output, Technometrics, Vol. 41, No. 1,
pp 39-56.
44. Schubert, W., K. Grossauer and E.A. Button (2004) Interpretation of displacement
monitoring data for tunnels in heterogeneous rock masses, Rock Mechanics and Mining
Sciences, Vol. 41, No. SUPPL. 1, paper 3B 19.
45. Simeoni, L. and L. Zanei (2009) A method for estimating the accuracy of tunnel
convergence measurements using tape distometers, International Journal of Rock
Mechanics and Mining Sciences, Vol. 46, No. 4, pp 796-802.
46. Sobol, I.M. (1993) Sensitivity estimates for nonlinear mathematical models,
Mathematical Modeling and Computational Experiment, Vol. 1, pp 407-414.
47. Unlutepe, A. and H. Ozener (2008) Geotechnical measurements at Izmir LRT system
tunnels, Tunnelling and Underground Space Technology, Vol. 23, No. 6, pp 734-741.
48. Wang, X.R., Y.H. Wang, and S.H. Zhang (2004) Study on judge system of fuzzy
inference to classification of tunnel surrounding rock, Proceedings of the International
Young Scholars' Symposium on Rock Mechanics - Boundaries of Rock Mechanics
Recent Advances and Challenges for the 21st Century, pp 203-206.
49. Yagiz, S. and C. Gokceoglu (2010) Application of fuzzy inference system and nonlinear
regression models for predicting rock brittleness, Expert Systems with Applications,
Vol. 37, No. 3, pp 2265-2272.
50. Yoo, C.S. and J-M. Kim (2007) Tunneling performance prediction using an integrated
GIS and neural network, Computers and Geotechnics, Vol. 34, No. 1, pp 19-30.
51. Zadeh, L. A. (1965) Fuzzy sets, Inf. Control Vol. 8, No. 3, pp 338-353.
2011 ejge

You might also like