You are on page 1of 19

Engineering Geology 74 (2004) 227 – 245

www.elsevier.com/locate/enggeo

Fuzzy set approaches to classification of rock masses


A. Aydin *
Department of Earth Sciences, James Lee Science Building, The University of Hong Kong, Pokfulam Road, Hong Kong, China
Received 4 November 2003; accepted 30 March 2004
Available online

Abstract

Rock mass classification is analogous to multi-feature pattern recognition problem. The objective is to assign a rock mass
to one of the pre-defined classes using a given set of criteria. This process involves a number of subjective uncertainties
stemming from: (a) qualitative (linguistic) criteria; (b) sharp class boundaries; (c) fixed rating (or weight) scales; and (d)
variable input reliability. Fuzzy set theory enables a soft approach to account for these uncertainties by allowing the expert to
participate in this process in several ways. Hence, this study was designed to investigate the earlier fuzzy rock mass
classification attempts and to devise improved methodologies to utilize the theory more accurately and efficiently. As in the
earlier studies, the Rock Mass Rating (RMR) system was adopted as a reference conventional classification system because
of its simple linear aggregation.
The proposed classification approach is based on the concept of partial fuzzy sets representing the variable importance or
recognition power of each criterion in the universal domain of rock mass quality. The method enables one to evaluate rock
mass quality using any set of criteria, and it is easy to implement. To reduce uncertainties due to project- and lithology-
dependent variations, partial membership functions were formulated considering shallow ( < 200 m) tunneling in granitic rock
masses. This facilitated a detailed expression of the variations in the classification power of each criterion along the
corresponding universal domains. The binary relationship tables generated using these functions were processed not to derive
a single class but rather to plot criterion contribution trends (stacked area graphs) and belief surface contours, which proved
to be very satisfactory in difficult decision situations. Four input scenarios were selected to demonstrate the efficiency of the
proposed approach in different situations and with reference to the earlier approaches.
D 2004 Elsevier B.V. All rights reserved.

Keywords: Fuzzy set theory; Multi-criteria classification; RMR; Rock mass classification; Subjective uncertainty; Partial fuzzy sets

1. Introduction terms to facilitate communication of information,


guide detailed investigations, predict their properties
The main goal of classification systems is to and behavior or establish relationships. Rock mass
group objects or events using common scales and classifications serve two purposes: characterization
and empirical design (Fig. 1). Useful summaries of
these classifications and their extensions can be
* Fax: +852-2517-6912. found in Bieniawski (1989) and Singh and Goel
E-mail address: aaydin@hku.hk (A. Aydin). (1999). Of the existing classification systems, Rock

0013-7952/$ - see front matter D 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.enggeo.2004.03.011
228 A. Aydin / Engineering Geology 74 (2004) 227–245

Fig. 1. Classification of rock mass classification systems according to their functions (purposes) and contents (modified from Russo et al., 1998).

Mass Rating (RMR) (Bieniawski, 1973) and Q and communicating the experience and underlying
(Barton et al., 1974) are the most commonly used. assumptions of experts in expressing various states
These are multi-parameter classification systems or levels of a classification criterion in linguistic
based on the analyses of tunneling case histories. terms. What do experts exactly mean when they use
They employ similar sets of linguistic and/or numer- these terms that relate to approximation and possi-
ical criteria recognized to influence behavior of rock bility? How can we ensure that these terms mean
masses and stability of rock structures. The relative the same thing in every case and to every practi-
contributions of each criterion to the engineering tioner or expert? How can we utilize this fuzziness
quality of rock masses are expressed in the form of in expert opinion and the decision process? Fuzzy
rating factors, over arithmetic (RMR) or logarithmic set approach reveals not only the basis for experts’
scales (Q). evaluations but also makes them more objective,
Because the dominance and interaction of each particularly through the process of setting functional
criterion (or parameter) vary with specific applica- expressions (i.e., construction of membership func-
tion (characterized by purpose, dimensions, depth tions) of their perception, belief, support or confi-
and loading mechanisms), engineering geological dence in the values assigned.
environment (characterized by litho-structural and In rock mass classification processes, subjective
geotechnical properties, and hydrogeological, geo- (or non-random) uncertainties result from: (a) use of
morphological and geodynamic settings) and rock qualitative (linguistic) terms (as input value of some
mass quality (or class), devising a universally valid criteria) whose meanings vary from one expert to
classification system is not a realistic goal. The the other; (b) predetermined and sharp class bound-
growing number of rock mass classifications mod- aries, whereas the rock mass quality is gradational
ified from existing ones or newly formulated for in nature; (c) prescribed rating (or weighting) scales
different applications (such as slope stability, foun- representing contribution (importance) of each crite-
dation bearing capacity, excavatibility and grouting) rion to the overall quality (assuming the overall
and specific rock types shows universal recognition quality is characterized perfectly by the used crite-
of the variations in the type and importance (rating) ria, and in all geological settings and applications);
of classification parameters with project/purpose and and (d) reliability of input value (quantitative and
geology. linguistic) of each criterion. Fuzzy set theory ena-
Like many other classification systems in use in bles a soft classification approach to account for
engineering geological practice, rock mass classifi- these uncertainties by allowing the expert (of a rock
cations often involve criteria whose values are engineering project and/or a particular engineering
assigned in linguistic terms. Practitioners and edu- geological environment) to participate in this pro-
cators are familiar with the difficulty of capturing cess in several ways.
A. Aydin / Engineering Geology 74 (2004) 227–245 229

This paper was designed to investigate the earlier set is uniquely characterized by a membership
attempts (at engineering classification of rock function expressing this support between 0 and 1
masses) based on the fuzzy set theory and to devise (Fig. 3).
improved methodologies to utilize the theory more A fuzzy subset A of a universal set X (= [x]) can be
accurately and efficiently. As in the earlier studies, represented as a set of ordered pairs of discrete values
the RMR system was adopted also in this study as a of the domain variable x and the corresponding
reference conventional classification system because membership values lA(x):
of its simple linear aggregation operator.
A ¼ ½x; lA ðxÞ x a A; AoX ð1Þ

2. Rock Mass Rating system In fuzzy set grammar, linguistic variables (e.g.,
descriptive qualifiers commonly used in engineering
In this study, the latest version of the Rock Mass geology such as ‘‘extremely strong’’, ‘‘highly weath-
Rating system (Bieniawski, 1989) was used as the ered’’, ‘‘widely spaced’’, ‘‘very tight’’ and ‘‘very
reference classification structure (Table 1). RMR wet’’) are replaced by fuzzy sets. Assuming symmet-
system is based on five main parameters: The arith- ric triangular membership distribution (as in Fig. 3)
metic sum of the ratings corresponding to the input within UCS (MPa) interval [200, 400], the fuzzy set
values of these parameters is referred to as ‘‘the basic expressing ‘extremely strong rock’ can be represented
RMR’’ (Fig. 2). The engineering quality or class of in vector form as:
the rock mass is represented by ‘‘the total RMR’’,
obtained by adjusting ‘‘the basic RMR’’ for the A ¼ ½0=200; 0:5=250; 1:0=300; 0:5=350; 0=400 ð2Þ
influence of joint orientation for a specific excavation
face (Fig. 2). where 0, 0.5 and 1.0 are the membership grades
corresponding to the selected UCS values and ‘‘/’’ is
a delineator.
3. Basics of fuzzy set theory Various segments of a membership function rep-
resent the limits of our expectation that an object (or
Fuzzy set theory is a generalization of ordinary or variable) belongs to the corresponding fuzzy set. The
classical set theory; it consists of mathematical tools relevant terms are shown in Fig. 4 for a trapezoidal
developed to model and process incomplete and/or membership function. The fuzzy sets presented
gradual information, ranging from interval-valued above are normal and convex sets that: (a) satisfy
numerical data to symbolic and linguistic expressions h(A) = max{lA(x)} = 1; and (b) display strictly mono-
(Dubois and Prade, 2000). The concepts and defini- tonic increase and/or decrease of its membership
tions of the fuzzy set theory (Dubois and Prade, 1980; values lA(x) with increasing values of its elements
Ross, 1995; Klir et al., 1997) are briefly summarized x, respectively.
below to facilitate the discussion on fuzzy rock mass The definition of a fuzzy set depends on the
classifications. situation (or context) that modifies our perception or
expectation of the meaning of the linguistic vari-
3.1. Fuzzy set and its membership function able. For example, the intact rock weathering clas-
sification by GSL (1995) suggests that ‘‘large pieces
Unlike crisp (or ordinary) sets, fuzzy sets have of highly weathered (Grade IV) rock can be broken
no sharp or precise boundaries (Fig. 3). As a result, by hand’’: Fuzzy sets constructed using this criteri-
the degree of membership lA(x) of elements x to a on to represent Grade IV rock may differ for
fuzzy set A depends on their positions (or compat- different people and rock types (Fig. 5a). Similarly,
ibility) with respect to the centre (or the central quantitative data may sometimes be imprecisely
concept). An element’s position within a fuzzy set described due to non-random uncertainties: UCS
is determined by the degree of belief (or support) of ‘‘about 40 MPa’’ implies different magnitude of
that the element belongs to that set. Thus, a fuzzy variations if it is determined indirectly by point load
230 A. Aydin / Engineering Geology 74 (2004) 227–245

Table 1
The rock mass rating system (after Bieniawski, 1989)
(A) Classification parameters and their ratings
Parameters Range of values
1 Intact rock PLI (MPa) >10 4 – 10 2–4 1–2 PLI is not preferred
strength UCS (MPa) >250 100 – 250 50 – 100 25 – 50 5 – 25 1 – 5 < 1
Rating 15 12 7 4 2 1 0
2 Drill core 90 – 100 75 – 90 50 – 75 25 – 50 < 25
quality in
RQD (%)
Rating 20 17 13 8 3
3 Spacing of >2 m 0.6 – 2 m 200 – 600 60 – 200 < 60
discontinuities mm mm mm
Rating 20 15 10 8 5
4 Condition of Very Slightly Slightly Slickensided or
discontinuities rough rough rough gouge < 5 mm
(see E for details) or separation
< 1 – 5 mm
Not Soft gouge>5 mm
continuous or separation>5 mm
No Separation Separation
separation < 1 mm < 1 mm
Unweathered Slightly Highly Continuous Continuous
weathered weathered
Rating 30 25 20 10 0
5 Groundwater Inflow per None < 10 10 – 25 25 – 125 >125
condition 10 m tunnel
length (l/m)
Joint water 0 < 0.1 0.1 – 0.2 0.2 – 0.5 >0.5
pressure/major
principal
stress
General Completely Damp Wet Dripping Flowing
conditions dry
Rating 15 10 7 4 0

(B) Rating adjustment for discontinuity orientations


Strike and dip orientations Very favorable Favorable Fair Unfavorable Very unfavorable
Rating
Tunnels and mines 0 2 5  10  12
Foundations 0 2 7  15  25
Slopes 0 5  25  50  60

(C) Rock mass classes determined from total ratings


Total rating 100 p 81 80 p 61 60 p 41 40 p 21 < 21
Class number I II III IV V
Description Very good rock Good rock Fair rock Poor rock Very poor rock

(D) Meaning of rock classes


Class number I II III IV V
Average stand-up time 20 years for 15 m span 1 year for 10 m span 1 week for 5 m span 10 h for 2.5 m span 30 min for 1 m span
Cohesion of rock >400 300 – 400 200 – 300 100 – 200 < 100
mass (kPa)
A. Aydin / Engineering Geology 74 (2004) 227–245 231

Table 1 (continued)
(D) Meaning of rock classes
Class number I II III IV V
Friction angle of >45 35 – 45 25 – 35 15 – 25 < 15
rock mass (j)

(E) Guidelines for classification of discontinuity conditions


Rating
Discontinuity <1 m 6 1–3 m 4 3 – 10 m 2 10 – 20 m 1 >20 m 0
length
(persistence)
Separation None 6 < 0.1 mm 5 0.1 – 1.0 4 1 – 5 mm 1 >5 mm 0
(aperture) mm
Roughness Very 6 Rough 5 Slightly 3 Smooth 1 Slickensided 0
rough rough
Infilling None 6 Hard 4 Hard filling 2 Soft 2 Soft filling 0
(gouge) filling >5 mm filling >5 mm
< 5 mm < 5 mm
Weathering Unweathered 6 Slightly 5 Moderately 3 Highly 1 Decomposed 0
weathered weathered weathered

index apparatus or directly from laboratory testing gradational classes of objects, which makes it im-
(Fig. 5b). possible to represent all the elements x with their
membership functions lA(x) in any other form. In
3.2. Analytical representation of membership analytical method trapezoidal membership functions
functions can be characterized by the height and four distinct
elements of the fuzzy set interval. The generic
Membership functions can be represented by
different methods. Of these methods, analytical
representation is used when a fuzzy set is defined
over a continuous domain, such as real numbers or

Fig. 2. RMR classification for characterization and design purposes. Fig. 3. Pictorial definitions of crisp and fuzzy sets.
232 A. Aydin / Engineering Geology 74 (2004) 227–245

The fuzzy relation R between two fuzzy sets A = [x/


lA(x)] and B = [ y/lB( y)] is a fuzzy set with dual
membership in both sets:
Z
R¼ ½ðx; yÞ=lR ðx; yÞ ð5Þ
A B

where the degree of dual membership is given by:

Fig. 4. Pictorial definition of the segments of a (trapezoidal)


lR ðx; yÞ ¼ ½lA ðxÞ ^ lB ðyÞ ¼ min½lA ðxÞ; lB ðyÞ ð6Þ
membership function, encapsulating variations in our subjective
judgment over the fuzzy set domain. (B: boundary of support zone). 3.4. Fuzzy arithmetic operations

Of the four basic fuzzy arithmetic operations, fuzzy


analytical expressions and the corresponding graph- addition and subtraction (Fig. 8) are calculated
ical representation are given Fig. 6. Taking b = c, according to (Kaufmann and Gupta, 1991):
these expressions reduce to those of triangular
membership functions. lAðþÞB ðzÞ ¼ _flA ðxÞ ^ lB ðyÞg

3.3. Fuzzy set operations z ¼ x þ y@lA ðxÞ ¼ lB ðyÞ ð7Þ

Three basic fuzzy set operations mentioned in this


paper are summarized below: lAðÞB ðzÞ ¼ _flA ðxÞ ^ lB ðyÞg
The union (A[B) and intersection (A\B) of two z ¼ x  y@lA ðxÞ ¼ lB ðyÞ ð8Þ
fuzzy sets A and B are given, respectively by (Fig. 7):
where A and B are fuzzy numbers (i.e., normal fuzzy
lA[B ðxÞ ¼ ½lA ðxÞ _ lB ðxÞ ¼ max½lA ðxÞ; lB ðxÞ ð3Þ sets) of the same universe, i.e., A, BoX and x, y,
z a X. The same rules apply for multiplication
(z = x y) and division (z = x/y).
lA\B ðxÞ ¼ ½lA ðxÞ ^ lB ðxÞ ¼ min½lA ðxÞ; lB ðxÞ ð4Þ
3.5. a-cuts of fuzzy sets

The symbols _ and ^ represent fuzzy union and In fuzzy set theory, the notion of a-cuts represents
intersection operations, respectively. the intervals (support range) corresponding to selected

Fig. 5. Context dependency in constructing fuzzy sets: (a) changing definitions of Grade IV rock by two people according to the breakability
criterion; (b) influence of the reliability of test procedures on the expected range of deviations implied by the qualifier ‘‘about’’.
A. Aydin / Engineering Geology 74 (2004) 227–245 233

Fig. 6. Five elements of a trapezoidal membership function used for its analytical representation.

membership values in [0,1]. The a-cut of a fuzzy set A two fuzzy sets (N = 2). As the procedure involves
(denoted as aA) is a crisp set: calculation of FWA for 2N permutations at each
a
A ¼ fxAAðxÞ z ag ð9Þ selected a-cut, it becomes a bulky task for more
than a few fuzzy sets.
Accordingly, the a-cuts of the trapezoidal function
in Fig. 6 can be written in generic form as: 3.7. Defuzzification methods

a Defuzzification is a process of reducing an aggre-


A ¼ ½ða þ ðb  aÞaÞ; ðd  ðd  cÞaÞ ð10Þ
gated (or clipped) fuzzy set into a crisp number,
At a = 0.5, the fuzzy set represented by Eq. (2) will presumably the most representative value of that
have the following interval: fuzzy set interval. The two methods mentioned in this
paper are described below.
0:5
A ¼ ½250; 350 ¼ ½0=200; 1=250; 1=350; 0=400
ð11Þ 3.7.1. Centre of area method
The centre of area (or centroid) method is the most
commonly used defuzzification method, despite its
3.6. Fuzzy weighted average (FWA) algorithm
complexity. The method aims to determine the point
zc at which the area under the clipped membership
This computational algorithm (Dong and Wong,
function lC is equally divided:
1987) is an aggregation procedure developed for
cases where both the meaning and importance (or
weights) of the criteria are fuzzy. The algorithm is zc ¼ ½jzlC ðzÞdz=½jlC ðzÞdz ð12Þ
based on the a-cut representation of fuzzy sets and
interval analysis. Unlike other numerical aggregation 3.7.2. Ranking index method
procedures, the membership values rather than sup- The ranking index (Juang et al., 1987) evaluates
port domain are discretized. Fig. 9 presents a the position of a fuzzy set in the universal domain, i.e.,
summary of the method for the simple case of the interval including the entire spectrum of classes of
objects or qualities. The procedure is explained in Fig.
10. This method is limited to cases where h(A) = 1 and
cannot be used for clipped fuzzy sets.

4. Earlier fuzzy rock mass classification


approaches

The use of fuzzy set theory in multi-criteria


Fig. 7. Union and intersection of two fuzzy sets A and B. classifications is uniquely effective, as it enables
234 A. Aydin / Engineering Geology 74 (2004) 227–245

Fig. 8. Fuzzy arithmetic operations: (a) addition; and (b) subtraction (after Kaufmann and Gupta, 1991).

incorporation of expert’s knowledge as well as is discussed, earlier fuzzy rock mass classification
quantification of qualitative information while retain- methods are briefly introduced. Particular note is
ing the imprecise nature of the criteria. Before the taken of their consistency with the theory and their
fuzzy classification approach proposed in this paper mutual differences.

Fig. 9. Application of FWA algorithm (based on Dong and Wong, 1987).


A. Aydin / Engineering Geology 74 (2004) 227–245 235

Fig. 10. Ranking index defuzzification method (based on Juang and Lee, 1990).

4.1. Nguyen’s (1985) approach on this table reveals the most likely class and member-
ship degree of the evaluated rock mass to that class.
Nguyen (1985) and Nguyen and Ashworth (1985) The type of fuzzy membership function proposed
proposed the first fuzzy rock mass classification ap- by Nguyen implies that if the input value of a criterion
proach based on min – max aggregation operation pro- is at the centre of any of its RMR class ranges, then
posed by Bellman and Zadeh (1970) for multi-criteria the rock mass fully belongs to that class. Subsequent-
decision modeling. This approach aims at the selection ly, this expert opinion of strong support is discarded
of most likely rock mass class as summarized in Fig. during the aggregation process as the criterion with
11. Membership functions lij for each criterion j at each the weakest support is retained for final selection.
rock class i are constructed based on subjective judg- Moreover Nguyen suggested that the degree of mem-
ment. With this type of membership function, a rock bership of a rock mass with RQD = 77 to RMR Class
mass usually belongs to all five classes with different II, i.e., lII,RQD, can also be calculated by dividing the
degrees of belief or support. Membership values input value (77) to the central value of that class range
corresponding to input are listed in the form of a fuzzy (82.5). This method is not consistent with the graph-
binary relation (Aj Bi) table. The min – max operation ical example, and does not produce the values given

Fig. 11. Nguyen’s (1985) rock mass classification approach.


236 A. Aydin / Engineering Geology 74 (2004) 227–245

in the binary relation table shown in Fig. 11. Another weighting (W) scales are both represented by five
shortcoming of this approach is that the fuzzy min – triangular membership functions (labeled A to E) in
max operation: (a) is appropriate for systems with both the universal rock mass quality domain (includ-
non-interactive criteria (Dubois and Prade, 1980), ing all possible fuzzy sub-sets of rock quality classes)
whereas criteria in rock mass classification systems and in the universal domain of importance (or
are inevitably interactive and have different impor- weights). Both universal domains are discretized into
tance or weights; (b) is conservative in class selection, 17 elements (1– 17) for simpler implementation of the
which can be avoided by mean – max operation; and proposed methodology. Then a fuzzy rating subset is
(c) fails when two or more similar values emerge from assigned to each criterion according to its input value
the min operation. relative to corresponding RMR class intervals. After
this, an FWA algorithm is utilized to obtain overall
4.2. Juang and Lee’s (1990) approach ratings, first for JCS and finally for the rock mass.
This final fuzzy set is defuzzified into a rock mass
This fuzzy classification method is based on Dong classification index (RMCI) using Juang et al.’s (1987)
and Wong’s (1987) FWA algorithm (Fig. 9) and ranking index method.
aggregates the individual (fuzzy) ratings of different It is not clear how the criteria weights were
criteria into an overall classification rating. The se- assigned in this approach. An important shortcoming
lected classification structure is a modified version of of the approach is the assignment of a single fuzzy set
RMR system (Fig. 12), but lumps JS and JC together to each numerical criterion regardless of its input
as JCS, presumably to limit the number of criteria to a value position on the corresponding class interval
manageable five (FWA becomes too tedious with range. Dual membership is not considered, with an
more). The fuzzy subsets for the rating (Rt) and inevitable loss of quantitative information.

Fig. 12. Juang and Lee’s (1990) rock mass classification approach (input and RMR ratings from Hoek and Brown, 1982).
A. Aydin / Engineering Geology 74 (2004) 227–245 237

4.3. Habibagahi and Katebi’s (1996) approach gradational boundaries. In constructing these func-
tions, the support base of each interval (of quantita-
In this approach, each of the numerical RMR tive criteria) extends into the neighboring class
criteria (UCS, RQD and JS) is fuzzified by five intervals such that the intersections of sets are all
trapezoidal membership functions defined over the at l = 0.5. This ensures that dual membership is
universal domain of the criterion in question. Fuzzi- obtained, for a given input value of a criterion.
fication of numerical criteria could be useful only to The largest product of RMR rating (corresponding
account for uncertainty in assigned ratings due to to each fuzzy set) and membership grade is taken as

Fig. 13. Habibagahi and Katebi’s (1996) rock mass classification approach. Refer to Fig. 14 for the evaluation of reliability parameters.
238 A. Aydin / Engineering Geology 74 (2004) 227–245

the fuzzy rating. This process of obtaining so-called (or importance) [0,1] of that criterion (Fig. 14). The
‘‘refined ratings’’ obviously alters the ordered struc- data are assigned reliability levels based on the
ture of ratings as a lower rating may be obtained for confidence in factors such as method of measure-
a higher input value. For example, for UCS = 100, ments, field experience, and total number of measure-
fuzzy rating = lA RMR rating = 6.0; whereas for ments, while the weight of each criterion for reliability
UCS = 75, fuzzy rating = 7.0. evaluation is assigned from subjective judgment.
The class intervals for each of the remaining RMR
criteria (JC, GW, JO) (input given mostly in descriptive
terms) are expressed by five triangular membership 5. Proposed fuzzy rock mass classification
functions plotted in the domain of corresponding RMR approach
ratings. However, fuzzy ratings for these criteria were
selected at the maximum membership level (i.e., l = 1). The proposed approach is based on the concept of
In other words, fuzzy sets for linguistic criteria were partial fuzzy sets representing the variable importance
used only symbolically. Because all the so-called fuzzy or recognition power of each criterion in the universal
ratings are in crisp numbers, the aggregation process to domain of rock mass quality. The method enables
obtain the total rating is exactly the same as in RMR evaluation of rock mass quality using any set of
system. The total rating is mapped onto the fuzzy sets criteria, and is easy to implement.
of rock mass classes represented by trapezoidal mem-
bership functions (Fig. 13) constructed using the same 5.1. Construction of partial fuzzy sets
universal domain and the central values of each class
interval as in RMR. The centroid method of defuzzi- In a multi-criteria classification system in which
fication is employed to find a final rock mass rating. criteria have variable and imprecise importance in the
Habibagahi and Katebi also deduced a rating universal domain of rock mass quality, fuzzy sets to
scheme for the reliability of classification using the represent their importance are necessarily partial, in
arithmetic weighted average of the product of reliabil- that a single criterion does not have the power to fully
ity [0,1] of each criterion and corresponding weight classify the object in the entire domain. This is
different from the partial membership concept where
the total membership is 1 at any point in the universal
domain. In previous fuzzy rock mass classification
attempts, the class memberships corresponding to
input values of each parameter are represented by
fuzzy sets with full support for at least one value on
the interval, whereas the importance of parameters is
taken into account in the form of crisp weighting
coefficients. This ignores fuzziness in the importance
of the parameters, which is further complicated by
their dependency on the specific application, engi-
neering geological environment and rock mass class.
In accordance with these observations, the partial
fuzzy sets (Fig. 15) were formulated to reflect each
parameter’s classification power of each class, by
combining fuzzy importance and multi-class member-
ship concepts. The aggregation of these partial sets
corresponding to input directly produces the level of
support for each class. The sixth parameter of the
Fig. 14. An evaluation of reliability of fuzzy classification of rock RMR system (joint orientation) was not considered in
masses (based on Habibagahi and Katebi, 1996). Refer to Fig. 13 this method, as it is not a characterization criterion,
for Rei, rri and Wi values. but can be easily incorporated into the proposed
A. Aydin / Engineering Geology 74 (2004) 227–245 239

method. The partial membership functions answer the


question ‘‘what are the possibilities (expressed in
terms of degree of belief in recognition power of each
criterion) for a rock mass to belong to each quality
class if it is characterized only by a single criterion?’’
Fig. 15 suggests that, except for extremely low values
of UCS, RQD and JS, a rock mass cannot be fully
characterized by a single criterion, and thus the rock
mass belongs to multiple classes with variable degrees
of memberships.
To reduce uncertainties due to project- and lithol-
ogy-dependent variations, the partial membership
functions were formulated considering shallow
( < 200 m) tunneling in granitic rock masses. Such
a restricted definition of the applicability range of the
classification facilitated a detailed expression of the
variations in the classification power of each param-
eter. For example, when UCS is very low, then the
granitic rock mass is likely either highly or com-
pletely weathered or involves a fault gouge, and
consequently is very likely of Class V quality. On
the other hand, a very high UCS does not guarantee
Class I quality as the structural parameters become
more decisive. Similarly, a low RQD value provides
more specific information than a high RQD value
because by definition RQD considers all core pieces
whose axial lengths are longer than 10 cm. Average
JS is more informative, especially for large values,
but it is insufficiently trustworthy to be used alone.
JC characterizes shear strength and permeability of
joints. Because the number of joint sets is not
explicitly included as a parameter in RMR, assess-
ment of block size and removability, which signifi-
cantly influences the structural quality of rock mass,
is not accounted for, and this increases the uncer-
tainty interval (i.e., the spread of the fuzzy sets).
Though GW condition is not an intrinsic quality
parameter, it was included in the original definition
of basic RMR. In shallow tunneling operations,
adverse GW conditions severely degrade the quality
of rock masses as construction media, but workabil-
ity and stability are not fully determined by GW
alone.
It is worth noting that universal domains of de-
scriptive terms can be numerically expressed by an
equally divided unit interval [0,1]. Similarly, partial
Fig. 15. Partial fuzzy sets representing the recognition power of fuzzy sub-sets defined over a descriptive domain can
RMR criteria. easily be translated onto the corresponding numerical
240 A. Aydin / Engineering Geology 74 (2004) 227–245

domain, where numerical input is available. Because The alternative was to use graphical tools as
descriptive terms do not in general correspond to visual identification or decision aids. Plotting param-
equal numerical ranges, each interval should be scaled eter contribution trends and belief surface contours
independently. proved to be very satisfactory in difficult decision
situations, particularly when used together. To dem-
5.2. Classification and visual interpretation onstrate the efficiency of this approach, four input
scenarios were selected: the input and graphical
In general, the classification criteria do not uni- results are presented in Fig. 16. For each input data
formly support the same rock class, and any given set, a matrix of ratings was prepared (Fig. 16) by
rock mass volume belongs to multiple classes with reading the corresponding membership values or
different degrees of memberships. Moreover, stron- support directly from the partial fuzzy sets (Fig.
gest support levels are not always available to con- 15). The key advantage of this system is that it is
tiguous classes, resulting in multiple isolated peaks. able to preserve all the information used to evaluate
Numerous trials with different aggregation operators the rock mass quality, which is usually lost due to
(e.g., min – max, min –mean, etc.) for different input the application of different aggregation and/or defuz-
scenarios showed that a universal operator could not zification processes.
be formulated for this evaluation structure. The ag- To maintain the clarity and consistency of the
gregation process itself was leading to the loss of discussion, the results are presented according to:
valuable information about the rock mass quality (a) degree of cumulative support; (b) degree of
variations and the responsible criteria. difference d between the supports for other classes;

Fig. 16. Classification performance of the proposed approach for the selected input scenarios (Cases 1 – 4).
A. Aydin / Engineering Geology 74 (2004) 227–245 241

and (c) degree of uniformity of support. Cumulative Decision: Class III –IV (lIII – IV = 2.68/5); dual mem-
support may imply simple addition of independent bership; high possibility of mixed Class III and Class
parameters. The trend surface graphs could be re- IV conditions.
duced to a crisp value using centroid defuzzification
method, if necessary. Case 4 (from the same drillhole; depth interval:
On the other hand, an acceptability threshold for 81.0– 84.5 m)
dual membership should be assigned to distinguish
cases where, despite high degree of support, no single (a) highest cumulative support: 2.70 for Class IV
class can be declared to represent the rock mass (b) difference from the nearest highly supported class:
quality according to the graphs. This threshold was d(2.70 –2.30)>0.2 from Class III
assumed to be d = 0.2 to account for fuzziness in (c) degree of support uniformity: Class III and IV both
assigned interval boundaries at all a-cut levels (i.e., non – -uniformly supported. Class III is not
type 2 fuzzy sets). Lastly, the relative uniformity of supported by UCS
support or contribution of each parameter was used to
help visual identification of the most likely class. Each Decision: Class IV (lIV = 2.7/5); high possibility of
of the investigated cases (Fig. 16) is evaluated in the encountering Class III condition.
following paragraphs.
A study of these cases confirms that this approach
Case 1 (from Nguyen, 1985) avoids the pitfalls and shortcomings of the earlier
fuzzy classification approaches, and also enhances the
(a) highest cumulative support: 2.6 for Class III reliability and usefulness of classification process. It
(b) difference from the nearest highly supported class: provides a simple and straightforward procedure for
d(2.6 –1.79)>0.2 from Class IV evaluating rock mass quality.
(c) degree of support uniformity: Class III more
regularly supported than all the other classes 5.3. Comparison with earlier fuzzy rock mass
classifications
Decision: Class III (lIII = 2.6/5); only Class IV may be
encountered locally. To compare the existing and proposed fuzzy rock
mass classification approaches, the structures of all
Case 2 (from Nguyen, 1985) classification approaches were standardized. For ex-
ample, JS was taken as one of the main parameters in
(a) highest cumulative support: 2.02 for Class II Juang and Lee’s method, instead of as a sub-parameter
(b) difference from the nearest highly supported class: of Joint Condition and Spacing (JCS).
d(2.02 –1.85) < 0.2 from Class III Because the membership functions of each param-
(c) degree of support uniformity: highly irregular; eter for each rock class in Nguyen’s (1985) approach
Class III not supported by JS are not known, the min – max method could be direct-
ly compared only for the two examples given in his
Decision: Class II (lII = 2.1/5); high degree of possi- paper. For the other cases, the partial membership
bility of encountering other classes. values as used in the proposed approach were utilized
to establish the binary relation table of min – max
Case 3 (from a drillhole in a granitic rock mass; aggregation method. Similarly, the criteria weights
depth interval: 39.5– 43.0 m) in Juang and Lee’s (1990) approach were assigned
subjectively. The overall FWA was therefore calculat-
(a) highest cumulative support: 2.68 for Class III ed assuming equal importance of the criteria.
(b) difference from the nearest highly supported class: To establish the distribution of cumulative mem-
d(2.68 –2.55) < 0.2 from Class IV berships of each criterion over five rock classes,
(c) degree of support uniformity: Class III more defuzzification methods were avoided in this compar-
uniformly supported than Class IV ison process. The distribution of classes for Nguyen’s
242
A. Aydin / Engineering Geology 74 (2004) 227–245
Fig. 17. Comparison of existing rock mass classification approaches for the cases evaluated in Fig. 16.
A. Aydin / Engineering Geology 74 (2004) 227–245 243

(1985) approach was presented as a histogram taking tion structure cannot produce strong evidence to
the minimum membership values along each of the support a particular class. Therefore, the proposed
five columns of rock classes and without evaluating approach successfully retains all the inherent uncer-
the maximum of the minimum values, which is tainty in the system, while enabling at the same time
obvious from the histogram representation. The class derivation of a crisp number (by using the centroid
distribution for Juang and Lee’s approach was shown method) if required for numerical analysis or design
as a fuzzy set in which the final FWA was converted purposes. Finally, it should be remembered that the
from a range of 1– 17 to a range of 0 –100. On the results are strongly controlled by the membership
other hand, to compare the fuzzy approaches with the functions designed by the subjective judgment of the
RMR, the class distribution was illustrated as stacked expert and for a specific geological setting.
columns taking equal height (or equal membership)
for each parameter’s contribution to the corresponding 5.4. Reliability of input values of criteria
rock class.
A comparison of the results for the same cases Reliability in fuzzy classification context is a
evaluated in Fig. 16 are summarized below (Fig. 17): function of belief in the accuracy and/or representa-
tiveness of input values, and as such should have an
Case 1. The rock mass is generally agreed to be of
effect on fuzziness in classification. A low degree of
Class III quality, but Habibagahi and Katebi’s system
input reliability should result in upper and lower
strongly suggested Class II. This general agreement is
bounds of the overall quality. Reliability in this
not unexpected, because of RMR’s inclination
context depends on site investigation team, methods
towards Class III.
and conditions, density and type of data, geological
Case 2. The rock mass is generally agreed to belong heterogeneities and other factors. For instance, deter-
to Class II, but Nguyen (with his own values) mining RQD and JS only by vertical drillholes results
suggested Class III quality with a degree of support in low reliability in input values of these criteria. On
of 0.60. When the values of the binary relation table the other hand, the classifier’s sense of reliability also
are derived from the proposed partial fuzzy sets, the varies, particularly with experience. Therefore, reli-
min – max aggregation yields Class II with a very low ability of input is a source of uncertainty that should
support of 0.12. This observation confirms the over- be directly included in the classification process,
conservatism of this method. unlike in Habibagahi and Katebi’s (1996) independent
Case 3. The proposed approach predicts mixed Class assessment approach (Fig. 14). It is suggested that this
III – IV conditions, whereas all other approaches can be readily incorporated into the proposed ap-
indicate Class III quality rock mass. Despite the basic proach by using type-2 fuzzy sets (Mendel, 2001).
RMR value of >60, its stacked column representation
agrees with the other fuzzy approaches.
6. Discussion and conclusions
Case 4. The proposed approach indicates Class IV
with a strong possibility of better (Class III) 6.1. Conventional classifications
conditions, a result distinct from the other approaches.
Interestingly, stacked column representation of the Bieniawski (1989) introduced a series of rating
RMR suggests mixed Class II – IV conditions, while curves for the numerical criteria to refine the rating
the basic RMR indicates Class II – III boundary. selection procedure in RMR. Ratings for the input
In general, the support ranges in the results of the values were no longer taken as the central value or
proposed approach usually coincide with the entire assigned from linear interpolation. The use of these
universal domain of rock mass quality. This is a rating curves significantly reduces uncertainty due to
natural consequence of the interpretation of the inter- the prescribed class boundaries of the numerical
actions between the selected criteria in constructing criteria. These curves can easily be translated into
the partial fuzzy sets, and the fact that moderate input fuzzy membership functions that enable a site-specific
values of any criterion used in the RMR’s classifica- formulation and incorporation of other aspects of
244 A. Aydin / Engineering Geology 74 (2004) 227–245

fuzziness as proposed by the partial fuzzy set concept. be capable of expressing variability of rock quality
Obviously, where the data is in the form of linguistic conditions due to all types of non-random uncertain-
variables, subjective uncertainty is still valid. ties or fuzziness in a rock mass classification process.
Conventional classification approaches reduce var- In conclusion, the proposed partial fuzzy set concept
iability in ‘‘relatively uniform’’ rock mass domains to is a practical and generic fuzzy rock mass classifica-
a single (crisp) value. Perhaps the weakest link in the tion structure capable of being adapted to different
classification process is the subjectivity involved in geological and engineering environments and of
the selection of such domains over natural exposures tackling inherent uncertainties in the classification
and excavation faces, and along boreholes. Levels and process.
aspects of required uniformity may change with the Membership functions can be viewed as concise
project and engineering geological environment, and quantitative expressions of our degree of expec-
while subtle litho-structural changes between adjacent tation (or belief) in respect of a great number of ‘‘IF –
areas or intervals often become a decisive factor. THEN’’ statements, and this degree of expectation
Consequently, relevant details are overlooked and forms the basis of fuzzy reasoning. For example, the
practical applications of these classifications (e.g., statement ‘‘IF RQD is about 35, THEN rock mass
support selection) are based on a single number. quality is very likely poor’’ is presented by the Class
Furthermore, the required support system usually IV membership function on Fig. 14. In this statement,
differs from that predicted by conventional classifica- the possibility qualifier ‘‘very likely’’ can be replaced
tions using pre-excavation data, and may involve by a truth (‘‘very true’’) or a certainty (‘‘almost
many different combinations of rock mass re-qualifi- certainly’’) qualifier. Aggregation of these statements
cation, reinforcement and external support options. for a specific set of conditions (i.e., input values) is
Thus, presenting variable degrees of belief for the the fuzzy inference process.
possibility of encountering different rock mass classes Fuzzy set theory also offers various means of
during excavation within the classified domain is quantifying the efficiency of classification structure
necessary to help reduce significant departures from based on fuzzy measures. However, as these require
the predicted (single-class) conditions. It is recom- discussion of a rather different theoretical back-
mended that ‘‘uniformity’’ of rock mass domains be ground, the use of fuzzy measures and fuzzy integrals
defined on the basis of site-specific criteria, and be in assessment of classification structure and rock mass
presented in the form of fuzzy sets to clearly commu- classification will be presented in another paper.
nicate the rules and assumptions involved.

6.2. Fuzzy approaches Acknowledgements

The earlier fuzzy classification approaches demon- The study is funded by the CRCG of The
strated the potential of the soft modeling tools of the University of Hong Kong. The author would like to
fuzzy set theory. These approaches are based on thank Ms. K. Battacharrya for her assistance in
different modifications of RMR classification struc- carrying out the tedious calculations involved in the
ture, universal domains and class intervals, member- comparison of the existing fuzzy approaches.
ship functions, and aggregation and defuzzification
procedures. However, these procedures are cumber- References
some to put into practice, and are not always based on
sound assumptions. This probably explains their fail- Barton, N., Lien, R., Lunde, J., 1974. Engineering classification of
ure to gain support as practical tools. rock masses for the design of tunnel support. Rock Mechanics 6,
Unlike the earlier approaches, the proposed ap- 189 – 236.
proach addresses the issue of relevance and sufficien- Bellman, R.E., Zadeh, L.A., 1970. Decision-making in a fuzzy
environment. Management Science 17 (4), B141 – B164.
cy of RMR criteria by constructing partial fuzzy sets Bieniawski, Z.T., 1973. Engineering classification of jointed rock
specifically for granitic rock masses at shallow masses. Transactions of the South African Institution of Civil
depths. The partial fuzzy set concept was shown to Engineers 15, 335 – 344.
A. Aydin / Engineering Geology 74 (2004) 227–245 245

Bieniawski, Z.T., 1989. Engineering Rock Mass Classifications. Kaufmann, A., Gupta, M.M., 1991. Introduction to Fuzzy Arithme-
Wiley, New York, p. 251. tic: Theory and Applications. Van Nostrand Reinhold, New
Dong, W.M., Wong, F.S., 1987. Fuzzy weighted averages and im- York, p. 361.
plementation of the extension principle. Fuzzy Sets and Systems Klir, G.J., Clair, U.S., Yuan, B., 1997. Fuzzy Set Theory: Founda-
21, 183 – 199. tions and Applications. Prentice Hall PTR, New Jersey, p. 245.
Dubois, D., Prade, H., 1980. Fuzzy Sets and Systems: Theory and Mendel, J.M., 2001. Uncertain Rule-Based Fuzzy Logic Systems:
Applications. Academic Press, New York, p. 393. Introduction and New Directions. Prentice Hall PTR, New
Dubois, D., Prade, H. (Eds.), 2000. Fundamentals of Fuzzy Sets. Jersey, p. 555.
Kluwer, Norwell, MA, p. 647. Nguyen, V.U., 1985. Some fuzzy set applications in mining geo-
GSL, 1995. The description and classification of weathered rocks mechanics. International Journal of Rock Mechanics and Min-
for engineering purposes. Geological Society Engineering ing Sciences & Geomechanics Abstracts 22 (6), 369 – 379.
Group Working Party Report. QJEG, 28, 207 – 242. Nguyen, V.U., Ashworth, E.A., 1985. Rock mass classification by
Habibagahi, G., Katebi, S., 1996. Rock mass classification using fuzzy sets. In: 26th US Symposium on Rock Mechanics, Rapid
fuzzy sets. Iranian Journal of Science and Technology 20 (3), City, SD, pp. 937 – 945.
273 – 284. Ross, T.J., 1995. Fuzzy Logic with Engineering Applications.
Hoek, E., Brown, E.T., 1982. Underground Excavations in Rock. McGraw-Hill, New York, p. 600.
E & FN Spon, London, p. 527. Russo, G., Kalamaras, G.S., Grasso, P., 1998. A discussion on the
Juang, C.H., Lee, D.H., 1990. Rock mass classification using concepts of geomechanical classes, behavior categories, and
fuzzy sets. Tenth Southeast Asian Geotechnical Conference. technical classes for an underground project. Gallerie e Grandi
Chinese Institute of Civil and Hydraulic Engineering, Taipei, Opere Sotterranee 54, 41 – 51.
pp. 309 – 314. Singh, B., Goel, R.K., 1999. Rock Mass Classification—A Practical
Juang, C.H., Burati, J.L., Kalidindi, S.N., 1987. A fuzzy system for Approach in Civil Engineering. Elsevier, Amsterdam, p. 267.
bid proposal evaluation using microcomputers. Civil Engineer-
ing Systems 4, 124 – 130.

You might also like