Rock Mech Rock Eng (2014) 47:717–732

DOI 10.1007/s00603-013-0426-3

ORIGINAL PAPER

Prediction of the Rock Mass Diggability Index by Using Fuzzy

Clustering-Based, ANN and Multiple Regression Methods
Omid Saeidi • Seyed Rahman Torabi •

Mohammad Ataei

Received: 3 January 2013 / Accepted: 20 April 2013 / Published online: 9 May 2013
 Springer-Verlag Wien 2013

Abstract Rock mass classification systems are one of the Keywords Rock mass diggability index  Intelligence
most common ways of determining rock mass excavat- system  ANN  ANFIS  Clustering approach  FCM 
ability and related equipment assessment. However, the SCM  Multiple regression
strength and weak points of such rating-based classifica-
tions have always been questionable. Such classification
systems assign quantifiable values to predefined classified 1 Introduction
geotechnical parameters of rock mass. This causes partic-
ular ambiguities, leading to the misuse of such classifica- The selection of a suitable overburden loosening method
tions in practical applications. Recently, intelligence has crucial importance in geotechnical and mining engi-
system approaches such as artificial neural networks neering design projects. On many projects, the ease of rock
(ANNs) and neuro-fuzzy methods, along with multiple excavation is mainly based on whether the rock can be
regression models, have been used successfully to over- economically excavated by mechanical excavators or if
come such uncertainties. The purpose of the present study drilling and blasting is necessary. The initial selection of
is the construction of several models by using an adaptive such equipment should optimise productivity and avoid
neuro-fuzzy inference system (ANFIS) method with two any subsequent revisions to excavation practice, with the
data clustering approaches, including fuzzy c-means resulting additional expenditure, production delays and
(FCM) clustering and subtractive clustering, an ANN and losses. There are a number of methods available for
non-linear multiple regression to estimate the basic rock assessing the excavatability of rock and related equipment
mass diggability index. A set of data from several case assessment. Rock mass classifications, as an example, have
studies was used to obtain the real rock mass diggability recently been quite popular and are mostly being used for
index and compared to the predicted values by the con- the preliminary design and planning purposes of a project.
structed models. In conclusion, it was observed that ANFIS According to Bieniawski (1989), ‘‘a rock mass classifica-
based on the FCM model shows higher accuracy and cor- tion scheme is intended to classify the rock masses, provide
relation with actual data compared to that of the ANN and a basis for estimating deformation and strength properties,
multiple regression. As a result, one can use the assimila- supply quantitative data for support estimation, and present
tion of ANNs with fuzzy clustering-based models to con- a platform for communication between the exploration,
struct such rigorous predictor tools. design, and construction groups.’’
In surface mining and civil engineering projects, terms
such as diggability, rippability, excavatability and bore-
ability are widely used to assess rock mass excavation
methods and required equipment. The rippability index was
defined for assessing overburden loosening by rippers
O. Saeidi (&)  S. R. Torabi  M. Ataei
(Caterpillar Inc. 1988). The seismic wave velocity of the
Faculty of Mining Engineering, Petroleum and Geophysics,
Shahrood University of Technology, Shahrood, Iran rock mass measured in the field is one of the widely used
e-mail: [email protected] indices for the prediction of rippability and the production

123
718 O. Saeidi et al.

rate of bulldozers (Caterpillar Inc. 1980; State Department different positive degrees of membership. In that case, it is
of Highways, Colorado 1969; Bailey 1975). Goktan and represented by the fuzzy clustering instead of the classical
Eskikaya (1991) developed a ‘‘rock mass rippability index’’ clustering (Dumitrescu et al. 2000).
applicable to sedimentary rocks of surface lignite mines. There are two prominent fuzzy clusterings, namely,
The index, which is a combination of the rock uniaxial fuzzy c-means (FCM) clustering and subtractive clustering.
compressive strength and the coefficient of relative rock FCM clustering was first introduced by Dunn (1973) and
mass weakness, was found to correlate well with the rates improved by Bezdek (1981). This algorithm is based on
at which the rock can be ripped and dozed for loading. minimising an objective function that represents the dis-
Scoble and Muftuoglu (1984) proposed an index which tance from any given data point to a cluster centre-
defines five rock classes based on four geotechnical weighted by that data point’s membership value.
parameters, namely; uniaxial compressive strength, bed- When we lack a clear idea of how many clusters there
ding spacing, joint spacing and weathering. The index, should be for a given set of data, subtractive clustering can
derived by the summation of the rated values of these input be used. Subtractive clustering (Chiu 1994), which is an
parameters, considers both geotechnical factors and exca- extension of the mountain clustering method (Yager and
vating equipment capabilities. Kirsten (1982) developed Filev 1994), is a fast one-pass algorithm for estimating the
the ‘‘excavatability index’’ based on the Q rock mass number of clusters and the cluster centres in a set of data.
classification system originally created for tunnelling. The The subtractive clustering algorithm supposes that each
input parameters of the system are uniaxial compressive data point is a potential cluster centre and calculates a
strength, number of joint sets, RQD, joint roughness, joint measure of the possibility that each data point would define
alteration, joint orientation and joint spacing. Howarth the cluster centre, based on the density of surrounding data
(1987) defined boreability as prediction of the penetration points (The MathWorks Inc. 2009)
rate of a rock cutting machine in a rock mass. The bore- Multiple regression analysis has been successfully
ability is a comprehensive variable related to the rock mass applied for developing predictive models in different rock
properties and machine specifications, as well as machine engineering practices (Ulusay et al. 2001; Gunes Yilmaz
operation parameters. et al. 2007; Gong and Zhao 2009; Yagiz and Gokceoglu
In rock mass classification processes, subjective uncer- 2010). The use of multiple regression analysis has advan-
tainties result in the use of linguistic terms (as the input tages in that it involves the use of categorical variables to
value of some criteria) that have different meanings among fit the regression model and the modelling of a non-linear
researchers. These can create predetermined and sharp relationship between dependent and independent variables
class boundaries that are improper for describing the nat- (Chun et al. 2009).
ural gradations of rock quality. However, linguistic terms The purpose of the present study is to construct models
are generally preferred by engineers and engineering using an adaptive neuro-fuzzy inference system (ANFIS)
geologists. How can we make the most of these linguistic method, an artificial neural network (ANN) and non-linear
terms in an expert opinion and decision process? The regression analyses for predicting the rock mass diggability
uncertainties of these parameters can be expressed as fuzzy index and to make a comparison of the prediction levels
sets when there is a reason to believe that not all values in between the developed models by using the related pre-
the interval have the same degree of confidence (Juang diction values and results.
et al. 1998). Since fuzzy systems can tolerate a wide range
of uncertainties and are capable of describing complex and
non-linear multivariable problems in a transparent way, the 2 Rock Mass Diggability Index
past few years have witnessed a rapid growth in the number
and variety of their applications in the areas of rock Diggability prediction has traditionally been based on
mechanics and engineering geology (Alvarez Grima and intact rock strength and discontinuity intensity. Geotech-
Babuška 1999; Alvarez Grima 2000; Sonmez et al. 2003; nical core logging offers an obvious means of rock mass
Gokceoglu et al. 2004; Gokceoglu and Zorlu 2004; Sonmez appraisal for diggability prediction. The principle ground
et al. 2004; Li et al. 2006; Tutmez and Tercan 2007). conditions, which affect excavating equipment perfor-
Clustering is the most consequential approach in modern mance during digging and loading, are (Scoble and Muf-
data mining technology and is used in processing dat- tuoglu 1984):
abases. The general philosophy of clustering is to divide
– Block size and shape, as defined by discontinuity
the initial set into homogenous groups based on the simi-
spacing and orientation, modified by any ground
larity of properties. In classical clustering, each pattern is
preparation conducted;
assigned to a unique class, while in real life, some of the
– Intact rock and discontinuity strength;
patterns could belong to more than one cluster, with

123
Prediction of the Rock Mass Diggability Index 719

– Bulk density of rock mass (bank and loose); metamorphic limestones. The effect of folding is reduced
– Rock abrasivity, determined by mineralogy; beyond the carbonaceous part of the district toward the
– Moisture content and degree of weathering (physical west and metabasic rocks. Figure 2a, b shows the major
and chemical) in modifying the nature of the above joint sets planes on a rock wall at this mine and a view of
factors. the mine’s slopes, respectively.
The rock mass diggability index proposed by Scoble and
3.2 Aghdarreh Gold Mine
Muftuoglu (1984) is one of the most advanced conven-
tional diggability classification systems and has been
The Aghdarreh gold mine is located in the Azarbayjan-e
adopted as the reference diggability classification system in
Gharbi province northwest of Takab city (Fig. 3). Gold
this study. The considered ground parameters in the system
mineralisation and other elements are created in micritic
are weathering (W), rock strength (UCS), joint spacing
limestones, which belong to the Ghom formation. The
(J) and bedding spacing (B).
oldest rock type in the area is the arkoses sandstone, related
The diggability index can be obtained by the arithmetic
to the Laloon formation.
sum of the ratings corresponding to the input values of
Figure 4a, b shows the jointed rock masses at the Ag-
these parameters, as shown in Table 1, that forms the basis
hdarreh gold mine surveyed during this study. As can be
for the proposed diggability classification chart (Table 2).
seen, many joint sets have intersected the rock mass.

3 Case Studies 3.3 Structure of the Database

3.1 Anguran Zinc and Lead Mine For the aim of this study, two sets of databases were
obtained from the literature (Scoble and Muftuoglu 1984;
The Anguran zinc and lead mine is located in the southwest Iphar and Goktan 2006). In addition, rock mass structural
of Zanjan province in northwest Iran (Fig. 1). The mine is a and strength data from the two presented case studies, i.e.
part of the Oroumieh-Poldokhtar zone, where it extends Anguran and Aghdarreh mines, were collected. Geological
between the Zagros belt and the central Iranian district. The structures, as needed in the rock mass diggability index,
mine is a central core of a turned anticline and dominant were surveyed along scan lines in all slopes in the mines.
rock at the mine area, composed of graphite-bearing Rock blocks were gathered from the mine sites and

Table 1 Diggability index

Parameter Rock class
rating method (Scoble and
Muftuoglu 1984) I II III IV V

Weathering Completely Highly Moderately Slightly Unweathered

Rating (W) 0 5 15 20 25
Strength, UCSa (MPa) \20 20–40 40–60 60–100 [100
b
Strength, Is(50) \0.5 0.5–1.5 1.5–2 2–3.5 [3.5
Rating (S) 0 10 15 20 25
Joint spacing (m) \0.3 0.3–0.6 0.6–1.5 1.5–2 [2
Rating (J) 5 15 30 45 50
a Bedding spacing (m) \0.1 0.1–0.3 0.3–0.6 0.6–1.5 [1.5
Uniaxial compressive strength
b Rating (B) 0 5 10 20 30
Point load strength index

Table 2 Diggability classification scheme (Scoble and Muftuoglu 1984)

Excavation class I II III IV V VI VII

Index \40 40–50 50–60 60–70 70–95 95–100 [100

(W ? UCS ? J ? B)
Ease of digging Very Easy Moderately Difficult Very Extremely Marginal without
easy difficult difficult difficult blasting

123
720 O. Saeidi et al.

Aghdarreh gold
mine

Fig. 1 Anguran zinc and lead mine (http://maps.google.com/maps) Fig. 3 Aghdarreh gold mine (http://maps.google.com/maps)

Jset 2 planes
Jset 1 planes

Bedding planes

(a) Jointed rock mass

(a) Joint sets and bedding planes

(b) Joint set and beddings

(b)
Fig. 4 Occurrence of joint and bedding planes in the slopes of the
Fig. 2 a Rock joints and beddings outcrops in a slope in the Anguran Aghdarreh gold mine
mine. b View of Anguran mine slopes and excavation and haulage
operations

transferred to the laboratory for determining the strength 4 The Adaptive Neuro-Fuzzy Inference System
parameters. Uniaxial compressive strength, UCS, and point
load strength, Is(50), tests were carried out according to the 4.1 The Takagi–Sugeno Fuzzy Inference System
procedure recommended in the International Society for
Rock Mechanics (ISRM) standard (ISRM 1981). Table 3 Fuzzy logic theory is a superset of traditional logic that has
presents a summary of the collected data from the case been developed to carry the concept of partial truth, i.e.
studies. truth values between ‘‘absolutely true’’ and ‘‘absolutely

123
Prediction of the Rock Mass Diggability Index 721

Table 3 Laboratory tests and field database

Source Parameter Number of data Minimum Maximum Mean Standard deviation

Scoble and Muftuoglu (1984) W 10 15 25 20.5 3.69

UCS (MPa) 10 40.9 89.9 67.6 15.5
J (m) 10 0.50 4 1.87 0.92
B (m) 10 0.1 15 2.95 4.33
DI 10 55 125 103 20.31
Iphar and Goktan (2006) W 46 15 25 20.5 3.69
UCS (Mpa) 46 1.5 174.1 34.3 41.82
J (m) 46 0.25 2 0.97 0.54
B (m) 46 0.15 2 0.86 0.52
DI 46 25 125 70 27.0
Anguran zinc and lead mine (this study) W 16 5 20 15 2.56
UCS (Mpa) 16 15 150 85 45.2
J (m) 16 0.75 3.5 1.35 0.85
B (m) 16 0.2 2.5 1.25 0.9
DI 16 45 145 123 34
Aghdarreh gold mine (this study) W 23 15 25 18.3 3.25
UCS (Mpa) 23 12 135 78 25
J (m) 23 0.3 5 3.4 2.2
B (m) 23 0.5 4 1.5 0.87
DI 23 35 95 58 17
W weathering, UCS uniaxial compressive strength, J joint spacing, B bedding spacing, DI rock mass diggability index; for the weathering, data
are related to their rates

false’’. Fuzzy conditioning statements are linguistic Fuzzy logic and neural networks are widely used arti-
expressions that can store the knowledge acquired using an ficial inference systems. Each approach has its merits and
input variable x and the corresponding output variable y. drawbacks. To take advantage of both approaches, inte-
Their general form is: gration of these systems has been proposed by many
if x is A then y is B researchers in recent years. The acronym ANFIS derives its
name from adaptive neuro-fuzzy inference system. Using a
where x and y are variables in the universes of X and Y; A and given input–output data set ANFIS, an FIS whose mem-
B are fuzzy sets based on X and Y, respectively. Generally, bership function parameters are tuned (adjusted) using
the ‘if’ part of the rule is called the antecedent, while the either a back-propagation algorithm alone or in combina-
‘then’ part is called the consequent. A general FIS structure tion with the least squares method is constructed.
comprises five parts: (1) a fuzzification interface that trans- This adjustment allows the fuzzy systems to learn from
forms crisp input data into a fuzzy value; (2) a rule base that the data they are modelling. An FIS and functionally
contains a number of fuzzy if–then rules; (3) a database that equivalent adaptive network are shown in Fig. 6a, b,
defines the membership functions of fuzzy sets used in fuzzy respectively.
rules; (4) a decision-making unit that performs inference In our experiment, a neural fuzzy model is used (Zhou
operations on the rules; and (5) a defuzzification interface, et al. 1997) that consists of five layers:
which transforms fuzzy results into a crisp output (Kasabov Layer 1: Each node, i in this layer generates a mem-
1998) (Fig. 5). The so-called Takagi–Sugeno (TS) method bership grade of a linguistic label. For instance, the node
of fuzzy inference was introduced by Takagi and Sugeno function of the ith node might be:
(1985) and consists of a set of fuzzy rules, each describing a (  2bi )1
local linear input–output relationship. x  v i
Q1i ¼ lAi ðxÞ ¼ 1 þ ð2Þ
The TS model has the following logic structure: r1
if x is A1 and y is B1 then f1 ¼ p1 x þ q1 y þ r1 ð1Þ
where x denotes the input to node i, Ai denotes the lin-
If f1 is a constant instead of a linear equation, then we have guistic label (low, high etc.) associated with this node and
a zero-order TS fuzzy model. {r1 ; ti ; bi} is the parameter set that changes the shape of

123
722 O. Saeidi et al.

Fig. 5 Fuzzy inference system

(FIS) with crisp output (Jang
Knowledge Base
1999) OUTPUT
INPUT
Database Rulebase
Fuzzification Defuzzification

Decision
System

Fig. 6 An adaptive network

and a fuzzy inference system
(FIS) (after Jang 1999)

the membership function. Parameters in this layer are summation of all incoming signals. This node is repre-
referred to as the ‘‘antecedent parameters’’. sented by Eq. (5):
Layer 2 (rule layer): Each node in this layer calculates P
X wi fi
the ‘‘firing strength’’ of each rule via multiplication: 5
Qi ¼ 
Wi fi ¼ P ð5Þ
wi
Q2i ¼ wi ¼ lAi ðxÞ  lBi ðyÞ i ¼ 1; 2 ð3Þ
4.2 Establishment of an ANFIS Structure
Layer 3: The ith node of this layer calculates the ratio of
the ith rule’s firing strength to the sum of all rules’ firing For a given data set, different ANFIS models can be con-
strengths: structed using different identification methods. The sub-
tractive clustering method (SCM) and FCM are the
 i ¼ P wi
Q3i ¼ W ; i ¼ 1; 2 methods used in this paper to identify the antecedent
2
j¼1 wj membership functions. The purpose of clustering is to
The outputs of this layer are also called normalised firing identify natural groupings of data from a large data set to
strengths. produce a concise representation of a system’s behaviour
Layer 4: The defuzzification layer, where each node (The MathWorks Inc. 2009).
function is given by:
 i fi ¼ W
 i ðpi x þ qi y þ ri Þ 4.2.1 ANFIS Modelling Based on the Subtractive
Q4i ¼ W ð4Þ
Clustering Method (SCM)
where W  i is the output of layer 3. Parameters in this layer
are referred to as the ‘‘consequent parameters’’. Subtractive clustering (Chiu 1994) is an algorithm for
Layer 5: The single node in this layer is a circle node, estimating the number of clusters and the cluster centres in
labelled R, which computes the ‘‘overall output’’ as the a set of data. The subtractive clustering algorithm assumes

123
Prediction of the Rock Mass Diggability Index 723

" #1
that each data point is a potential cluster centre and esti- c 
X 
dij ð2=ðm1ÞÞ
mates a measure of the possibility that each data point lij ¼ ð8Þ
would define the cluster centre, based on the density of k¼1
dkj
surrounding data points (The MathWorks Inc. 2009). For n where dij = ||ci - xj|| is the Euclidean distance between the
data points, a density measure (Di) is defined for each data ith cluster centre and the jth data point, and m is the
point, xi, as:
fuzziness index. Third, the cost function is computed
Xn
according to the next equation shown below. Stop the
Di ¼ expððjxi  xj j2 Þ=ðra =2ÞÞ2 ð6Þ
process if it is below a certain threshold.
j¼1
X
c c X
X n
where ra denotes a positive constant which specifies the JðU; c1 ; . . .; cÞ ¼ Ji ¼ lm 2
ij dij ð9Þ
radius of the neighbourhood (cluster radius) and can take i¼1 i¼1 j¼1
values in the range [0, 1]. It can be observed that the
For the final step, new c fuzzy cluster centres ci,
density measure for a data point is a function of its distance
i = 1, 2,…, c are computed using Eq. (10):
to all other data points. Hence, a data point that has many Pn m
neighbouring points will have a high potential of being a j¼1 lij xj
Ci ¼ Pn m ð10Þ
cluster centre. j¼1 lij
The first cluster centre is chosen to be the data point that
has the highest density measure. Then, the density measure The fuzzy rule base for the Sugeno model based on
for each data point, xi, is reduced as follows: subtractive clustering and FCM were conducted using
MATLAB software (The MathWorks Inc. 2009). The
proposed architecture of the ANFIS model is shown in
Di ¼ Di  Dc1 exp½ððjxi  xc1 j2 Þ=ðrb =2ÞÞ2  ð7Þ Fig. 7.
where xc1 denotes the first cluster centre and Dc1 denotes It can be seen from Fig. 7 that W, UCS, J and B were
its density measure. The constant rb denotes the radius of applied as input parameters into the ANFIS models and
the neighbourhood within which the reduction in density the rock mass diggability index was obtained as the
will be measurable. The constant rb is usually greater output parameter. The proposed models (SCM and FCM)
than ra in order to prevent possessing closely spaced were trained with 71 data sets (Table 3) for the training
centres. Clearly, data points close to the cluster centre phase and the remaining 24 data sets were used for the
will have significantly decreased density measures, so checking phase. In Fig. 8, the membership functions
that they have a smaller chance of being chosen as the (Gaussian type) for the inputs is shown based on the
next cluster centre. The next cluster centre is selected FCM method.
and the density measure is reduced again. This process is The most efficient cluster radius, ra in Eq. (6), was
iterated until a sufficient number of clusters is achieved. determined utilising a set of values which varied from 0 to
?1, with step size of 0.1. Then, each of the TS fuzzy
models was separately trained on the training data and its
4.2.2 ANFIS Modelling Based on the Fuzzy c-Means performance was measured for the test data. It was
Clustering Method (FCM) observed that, by specifying the value of 0.5 for the clus-
tering radius, the associated TS model yielded the lowest
FCM clustering is the most prominent fuzzy clustering root mean square error (RMSE). Also, using this value for
algorithm, which is based on minimising an objective the ra, four Gaussian-type membership functions were
function that represents the distance from any given data extracted for the inputs, which are symbolised as C-1 to
point to a cluster centre weighted by that data point’s C-4 (Fig. 9).
membership value. FCM partitions a collection of n vectors An illustration of the processing stages of the con-
xi, i = 1, 2,…, n into fuzzy groups, and finds a cluster structed TS model obtained from SCM for the rock mass
centre in each group such that the cost function of the diggability index prediction from a given input is shown in
dissimilarity measure is minimised. Fig. 10. As can be seen, by passing a row of inputs
i = 1, 2,…, c are randomly selected from the n points. including W = 20, UCS = 89 MPa, J = 2.15 m and
The stages of the FCM algorithm are, therefore, described B = 7.55 m from the TS model, the membership functions
in brief. Firstly, the cluster centres ci, i = 1, 2,…, c are related to each cluster are affected in rules. The antecedent
selected randomly from the n points {x1, x2, x3,…, xn}. memberships of rules have four parts and, so, the fuzzy
Second, the membership matrix U, using the following ‘‘and’’ operator is applied. This operator simply selects the
equation, is computed: minimum of the three values that represents the result of

123
724 O. Saeidi et al.

Fig. 7 ANFIS structure based

on SCM and FCM W

Rock mass
UCS (MPa) ANFIS
Diggability Index

J (m)

B (m)

Fig. 8 The antecedent Gaussian membership functions of the rock mass diggability index modelling using the FCM clustering method

the antecedent for that rule. The output membership gravity,’’ which is the most prevalent of all the defuzzifi-
function is then truncated by this value. Afterwards, the cation methods.
outputs of the rules are combined into a single fuzzy set via
any of the aggregation methods. The TS fuzzy systems 4.3 Prediction of the Rock Mass Diggability Index
finally take the result of the aggregation process and pro- Using the ANFIS Models
duce a single real-world output by means of the defuzz-
ification method, which is the last stage of the FIS. The Using the MATLAB functions genfis2 and genfis3, one can
defuzzification method used in this study is the ‘‘centre of execute fuzzy processes introducing input and output data

123
Prediction of the Rock Mass Diggability Index 725

Fig. 9 The antecedent membership functions a weathering, b UCS, c joint spacing, d bedding spacing of the rock mass diggability index
modelling using SCM

Fig. 10 Fuzzy if–then rules generated by subtractive clustering; W weathering, UCS uniaxial compressive strength, CS joint spacing, B bedding
spacing, DI rock mass diggability index

in the subtractive and FCM clustering-based models, Once the fuzzy clustering is completed and the com-
respectively. By the determination of antecedent member- ponents of the TS fuzzy system are specified, the adaptive
ship functions and the number of fuzzy if–then rules, the learning procedure begins. The parameters of the fuzzy
linear least squares estimation was used to determine the system membership and consequent equations change
consequent parameters of each rule. through the learning process. The adjustment of these

123
726 O. Saeidi et al.

parameters is facilitated by a gradient vector. This gradient where var symbolizes the variance, yi is the measured
vector supplies a measure of how well the FIS is modelling value, ^yi is the predicted value and N is the number of
the input–output data for a given set of parameters. When samples.
the gradient vector is obtained, any of several optimisation The interpretation of the above performance indices are
routines can be applied in order to search for the best set of as follows: the higher the VAF, the better the model per-
parameters by minimising the sum of the squared differ- forms. For example, a VAF of 100 % means that the
ence between actual and desired outputs (The MathWorks measured output has been predicted exactly. A VAF of 0
Inc. 2009). % means that the model performs as poorly as a predictor
The ANFIS model was trained on training samples and a using simply the mean value of the data. The lower the
testing set was used as checking data to prevent the training RMSE and MAPE, the better the model performs (Alvarez
data from over-fitting. The antecedent membership func- Grima and Babuška 1999; Jalalifar et al. 2011; Bashari
tions and fuzzy rule parameters for both clustering models et al. 2011). Contrary to the VAF, the RMSE also accounts
were justified through 100 training epochs. The checking for a bias in the model, i.e. an offset between the measured
error decreases up to a certain point in the training, and and the predicted data.
then it increases. This increase represents the point of When the VAF, RMSE and MAPE performance
model over-fitting. The ANFIS method chooses the model parameters are considered for each predictive model
parameters associated with the minimum checking error (Table 4), it is clear that the developed ANFIS model based
just prior to this turning point. on FCM gives a better prediction performance than that of
In Fig. 11a, b, a comparison between the outputs of the the SCM-based model.
fuzzy-based subtractive clustering model trained by ANFIS
with actual training and testing data is shown. As can be
seen, good agreement is observed between actual data and 5 Multiple Regression Analysis
outputs of the SCM model in both the training and the test
data. Multiple regression is a method of finding a non-linear
Figure 12 demonstrates the error reduction trend by model of the relationship between the dependent variable
increasing training epochs using the SCM method. The and a set of independent variables. Unlike traditional
result is reasonable, because the test set error does not show linear regression, which is restricted to estimating linear
any significant over-fitting in the model. models, non-linear regression can estimate a model with
Figure 13a, b shows plots of the actual and predicted
rock mass diggability index using the FCM main model
and FCM trained by the ANFIS model, respectively.
According to the plots, it is seen that the FCM output data
was improved after training by the ANFIS model. Cross-
correlation between actual data and SCM and FCM trained
by ANFIS is presented in Fig. 14a, b, respectively. It is
seen that ANFIS with the FCM clustering-based model
predicted the rock mass diggability index more accurately
than ANFIS with subtractive clustering-based model.
In order to check and compare the prediction perfor-
mances of the ANFIS clustering-based models, in addition
to the correlation coefficient, R2, the variance accounted for
(VAF), the RMSE and the mean absolute percentage error
(MAPE) performance indices were used:

 
ðvarðyi  ^yi ÞÞ
VAF ¼ 1  100 % ð11Þ
varðyi Þ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
uX
RMSE ¼ t ðyi  ^yi Þ2 =N ð12Þ
i¼1

N  
1X  ðyi  ^yi Þ
MAPE ¼  100 ð13Þ
N i¼1  yi  Fig. 11 Comparison of the SCM model outputs with a training data
and b testing data

123
Prediction of the Rock Mass Diggability Index 727

Fig. 12 Reduction of training error by epochs

Fig. 14 Cross-correlation between actual and predicted data of the

rock mass diggability index using a SCM and b FCM ANFIS-trained
models

Fig. 13 Comparison of the actual and predicted data of the rock mass
multiple regression analysis was carried out using the sta-
diggability index for a FCM and b FCM trained by the ANFIS
method tistical software package, SPSS (2007) according to the
data gathered in Table 3. Two common regressions, i.e.
arbitrary relationships between independent and dependent linear and non-linear methods, were attempted in this
variables (SPSS 2007). In Fig. 15, the relationships study, as in Eqs. (14) and (15):
between the rock mass diggability index and its subclasses
is shown. DI ¼ 7:63 þ 1:22W þ 0:19UCS þ 18:77J þ 16:9B ð14Þ
It can be seen that most of the effective parameters are 0:32 0:088 0:3 0:2
DI ¼ 3:08  W  UCS J B ð15Þ
tend to have a power law relation with the rock mass
diggability index, except for the weathering factor, which where DI denotes the rock mass diggability index,
shows a linear relation. Since the rock weathering was W denotes weathering, UCS is the uniaxial compressive
implemented in this classification qualitatively, therefore, strength in MPa, J is the joint spacing in metres and B is the
the rate values have been used in the calculations. Lastly, bedding spacing in metres.

123
728 O. Saeidi et al.

Table 4 Performance parameters of the neuro-fuzzy models

Performance parameters Main model Trained by ANFIS
FCM SCM FCM SCM
Training data Test data Training data Test data Training data Test data Training data Test data

RMSE 7.58 11.76 85 165 3.63 10.46 0.1621 157.09

VAF 89.1 89.2 54.4 53.2 97.5 92.01 99.99 75.6
MAPE 9.46 14.89 25 34 4.84 11.89 0.15 92.65
FCM fuzzy c-means clustering-based method, SCM subtractive clustering-based method, RMSE root mean square error, VAF variance accounted
for, MAPE mean absolute percentage error

Table 5 presents the comparison of the significance of Since all the independent variables in Eqs. (14), and (15)
each model from Eqs. (14) and (15). As can be seen, the have expected coefficient signs regarding the rock mass
non-linear model, Eq. (15), has better performance and diggability index, then there is no need to check for any
accuracy in predicting the rock mass diggability index. type of multicollinearity in the regression models.
However, the linear model, Eq. (14), provided a high
correlation coefficient (R2 = 0.9) but it gives a P-value of
more than 0.05 for the intercept value of the model (in 6 Artificial Neural Network (ANN) Analysis
bold). Statistically, for a regression model to be acceptable
at the 95 % level of confidence, the obtained P-values must Artificial neural networks (ANNs) characterise knowledge
be smaller than 0.05. In addition, the standard errors of the that is embedded in many disciplines: neurosciences, math-
regression analysis for all variables in the non-linear model ematics, statistics, physics, computer science and engineer-
are smaller than that of the linear model. ing (Haykin 1999). The network allows for self-learning,

Fig. 15 The relationships between the rock mass diggability index and a weathering, b UCS, c joint spacing, J and d bedding, B

123
Prediction of the Rock Mass Diggability Index 729

Table 5 Comparison between the significance of the regression models

Model Variables Coefficients Standard error P value F Significance F Multiple R R2

Non-linear model, Eq. (15) Intercept 3.08407 0.42310 6.46 9 10-9 131.499 8.22 9 10-23 0.963 0.928
W 0.32405 0.13965 0.026
UCS 0.08766 0.01697 6.57 9 10-6
J 0.29694 0.05730 6.22 9 10-6
B 0.19452 0.05145 0.0005
Linear model, Eq. (14) Intercept 7.63274 8.84989 0.39345
W 1.21526 0.52846 0.02664 95.442 3.29 9 10-20 0.950 0.903
UCS 0.18591 0.03952 0.00003
J 18.77248 4.92816 0.00046
B 16.89801 4.88119 0.00127

self-organisation and parallel processing, and is well sui- Feed forward activation
ted to problems involving matching an input pattern to a
set of output patterns where deep reasoning is not
W
required. A network needs to be trained at the second step
before any prediction can be done. Out of the different
UCS
algorithms available, back-propagation neural networks
(BPNs) are the most popular ANN technique and have J DI
been widely used in problems of pattern feature (pf)
classifications, prediction, adaptation control, system Target layer
B
identification and so on. The mathematical equations used
Input layer
in BPNs can be found in Rumelhart and McClelland
(1986). Due to its popularity and its flexibility and
adaptability in modelling a wide spectrum of problems in
many application areas, in this study, a feed-forward ANN Back propagation (hidden layer with 8 neurons)
architecture with the Levenberg–Marquardt back-propa-
Fig. 16 The architecture of the artificial neural network (ANN)
gation algorithm (trainlm) has been employed in order to model used in the study; W weathering, UCS uniaxial compressive
obtain an accurate model using MATLAB version strength,J joint spacing, B bedding spacing, DI rock mass diggability
7.8.0.347 (Demuth et al. 2005). index
The same data sets employed in the ANFIS and non-
linear multivariable regression analyses were used for When using the function ‘‘train’’ in the MATLAB software,
developing the ANN model. Since the hidden layer is the by default, the data are randomly divided so that 70 % of the
most crucial element of the network, the number of neu- samples are assigned to the training set, 15 % to the validation
rons in the hidden layer has been selected carefully set and 15 % to the test set (The MathWorks Inc. 2009).
according to suggested heuristics by different authors Figure 17 shows cross-correlations between measured
(Hecht-Nielsen 1987; Alvarez Grima 2000). values of the rock mass diggability index and predicted
Running the ANN program several times, it was deter- values using ANN. It was also observed that RMSE values
mined that the architecture (4 9 8 9 1) provides the best for training, testing and validation phases decreased with
prediction model with high correlation. It is evident that the time to the lowest level, gradually. For training, validation
performance of the ANN architecture increases when the and testing data correlation coefficients of 0.99, 0.98 and
number of neurons used in hidden layer(s) increases. 0.98 was obtained, respectively. In general high correlation
However, a minimal number of hidden layer(s) and a coefficient, R = 0.97 was obtained which shows the rig-
minimal number of neurons as possible should be selected orousness of the model for predicting the index.
by considering acceptable performance criterion (Iphar
2012). This particular structure type means that the ANN
has a total of three layers, with four neurons, representing 7 Performance Evaluation of the Models
the W, UCS, J and B in the input layer, one hidden layer
with eight neurons, followed by one neuron in the output In this study, ANFIS fuzzy clustering-based, multiple
layer that is eventually the estimated DI (Fig. 16). regression and ANN models were constructed to predict

123
730 O. Saeidi et al.

Fig. 17 Cross-correlations between the predicted (outputs) and measured (target) data of the rock mass diggability index using the ANN:
a training set, b validation set, c testing set and d all data

Table 6 Performance parameters of the proposed models and RMSE were calculated. The results are given in
Parameter ANFIS Non-linear regression ANN Table 6.
When the VAF and RMSE performance parameters are
FCM SCM
considered for each predictive model (Table 6), it can be
R2 0.9823 0.5648 0.93 0.9745 clearly seen that the developed ANFIS model with FCM
RMSE 10.46 157.09 18.35 12.34 method gives a better prediction performance than that of
VAF 92.01 75.6 85.3 88.32 the ANFIS with SCM, ANN and classical multiple
regression model. In addition, the ANFIS with FCM-based
the rock mass diggability index by using rock mass model has a lower RMSE and a higher correlation coeffi-
properties including weathering, W, uniaxial compressive cient than the other models proposed. Therefore, it can be
strength, S, joint spacing, J, and bedding, B, as subclasses said that the ANFIS with FCM-based model is the best
of the index. prediction tool for the estimation of the rock mass digga-
To assess the performance of the models, the perfor- bility index from its subclasses, rock mass weathering,
mance parameters defined in Eqs. (11) and (12) for VAF UCS, joint spacing and bedding spacing.

123
Prediction of the Rock Mass Diggability Index 731

8 Conclusions Alvarez Grima M, Babuška R (1999) Fuzzy model for the prediction
of unconfined compressive strength of rock samples. Int J Rock
Mech Min Sci 36:339–349
Rock engineering classifications have posed themselves as Bailey AD (1975) Rock types and seismic velocities versus rippa-
an integral part of empirical approaches for design purposes bility. In: Proceedings of the 26th Annual Highway Geology
in rock engineering. They provide quantitative information Symposium, Coeur d’Alene, Idaho, August 1975, pp 135–142
during the design stage of a project. Over the last 40 years, Bashari A, Beiki M, Talebinejad A (2011) Estimation of deformation
modulus of rock masses by using fuzzy clustering-based
there have been a large number of classification systems modeling. Int J Rock Mech Min Sci 48:1224–1234
developed by different authors. The adopted procedure in Bezdek JC (1981) Pattern recognition with fuzzy objective function
most of these classification systems is the quantification of algorithms. Plenum Press, New York
geotechnical parameters that are related to machine perfor- Bieniawski ZT (1989) Engineering rock mass classifications. Wiley,
New York, 251 pp
mance, leading to a single final rating or index. The final Caterpillar Inc. (1980) Caterpillar performance handbook, 11th edn.
ratings obtained in this way are then related to the ease of Caterpillar, Peoria
rock excavation classes and machine types to be used. Caterpillar Inc. (1988) Caterpillar performance handbook, 19th edn.
Despite their widespread use, these systems have some Caterpillar, Peoria
Chiu SL (1994) Fuzzy model identification based on cluster
limitations in practical applications, which may result in estimation. J Intell Fuzzy Syst 2(3):267–278
their misuse. The most important deficiencies of such rat- Chun BS, Ryu WR, Sagong M, Do JN (2009) Indirect estimation of
ing-based classification systems are the existence of sharp the rock deformation modulus based on polynomial and multiple
transition boundaries between two adjacent rock classes regression analyses of the RMR system. Int J Rock Mech Min
Sci 46(3):649–658
contrary to the gradational variation nature of rock and the Demuth H, Beale M, Hagan M (2005) MATLAB version 7.8.0.347.
inclusion of subjective uncertainties, which necessitate an Neural network toolbox for use with MATLAB. The MathWorks
expert judgment. Inc., p 348
In this study, an artificial neural network (ANN), the Dumitrescu D, Lazzerini B, Jain LC (2000) Fuzzy sets and their
application to clustering and training. CRC Press, Boca Raton
adaptive neuro-fuzzy inference system (ANFIS) combined Dunn JC (1973) A fuzzy relative of the ISODATA process and its use
with fuzzy c-means (FCM) and the subtractive clustering in detecting compact well-separated clusters. J Cybern 3:32–57
method (SCM), and multiple regression models were Gokceoglu C, Zorlu K (2004) A fuzzy model to predict the uniaxial
developed for one of the conventional rock excavation compressive strength and the modulus of elasticity of a
problematic rock. Eng Appl Artif Intell 17:61–72
classification systems, the rock mass diggability index. It Gokceoglu C, Yesilnacar E, Sonmez H, Kayabasi A (2004) A neuro-
was concluded that the proposed ANFIS FCM-based model fuzzy model for modulus of deformation of jointed rock masses.
performed best when the variance accounted for (VAF, Comput Geotech 31:375–383
92.01), root mean square error (RMSE, 10.46) and correla- Goktan RM, Eskikaya S (1991) Prediction of ripping machine
performance in terms of rock mass properties. Civil Eng South
tion coefficient (R2, 0.982) are considered. However, since Afr 31(1):13–24
the ANN model performance was better than that of multiple Gong QM, Zhao J (2009) Development of a rock mass characteristics
regression, it can be considered as a strong predictor for rock model for TBM penetration rate prediction. Int J Rock Mech Min
engineering systems in assimilation with fuzzy clustering- Sci 46(1):8–18
Gunes Yilmaz N, Yurdakul M, Goktan RM (2007) Prediction of
based methods, as in the ANFIS models. radial bit cutting force in high-strength rocks using multiple
In addition, in this study, basic classification of the rock linear regression analysis. Int J Rock Mech Min Sci
mass diggability index was investigated; however, there are 44(6):962–970
many other parameters of concern, such as random joints and Haykin S (1999) Neural networks: a comprehensive foundation, 2nd
edn. Prentice Hall, Upper Saddle River, p 841
aperture and fillings in a jointed rock mass, as well as Hecht-Nielsen R (1987) Kolmogorov’s mapping neural network
machine parameters, that could affect the rock mass digga- existence theorem. In: Proceedings of IEEE First International
bility index. Further study in future works is needed in order Conference on Neural Networks, San Diego, California, June
to modify this index to comprise those effective parameters. 1987, pp 11–14
Howarth DF (1987) Mechanical rock excavation—assessment of
cuttability and boreability. Proc RETC 1:145–164
Acknowledgments The authors wish to thank the staff of the An- Iphar M, Goktan RM (2006) An application of fuzzy sets to the
guran and Aghdarreh mines for facilitating the data gathering during Diggability Index Rating Method for surface mine equipment
this study. The authors who allowed us to use their data in this selection. Int J Rock Mech Min Sci 43:253–266
manuscript are gratefully thanked. Iphar M (2012) ANN and ANFIS performance prediction models for
hydraulic impact hammers. Tunn Undergr Space Technol
27:23–29
ISRM (1981) Rock characterization testing and monitoring. In:
References Brown TT (ed) ISRM suggested methods. Pergamon Press,
Oxford p 211
Alvarez Grima M (2000) Neuro-fuzzy modeling in engineering Jalalifar H, Mojedifar S, Sahebi AA, Nezamabadi-pour H (2011)
geology. Balkema, Rotterdam Application of the adaptive neuro-fuzzy inference system for

123
732 O. Saeidi et al.

prediction of a rock engineering classification system. Comput SPSS Inc. (2007) SPSS Regression Models (Version 16). http://www.
Geotech 38:783–790 winwrap.com
Jang JSR (1999) ANFIS: adaptive-network-based fuzzy inference State Department of Highways, Colorado (1969) Rock rippability
system. IEEE Trans Syst Man Cybern 23(3):665–685 study (final report). State Department of Highways, Colorado,
Juang CH, Jhi YY, Lee DH (1998) Stability analysis of existing Rep Pb 189.842
slopes considering uncertainty. Eng Geol 49:111–122 Takagi T, Sugeno M (1985) Fuzzy identification of systems and its
Kasabov NK (1998) Foundations of neural networks, fuzzy systems, applications to modeling and control. IEEE Trans Syst Man
and knowledge engineering. MIT Press, Cambridge Cybern 15:116–132
Kirsten HAD (1982) A classification system for excavation in natural The MathWorks Inc. (2009) MATLAB. Fuzzy logic toolbox, user’s
materials. Civil Eng South Afr 24:293–308 guide
Li W, Mei S, Zai S, Zhao S, Liang X (2006) Fuzzy models for Tutmez B, Tercan AE (2007) Spatial estimation of some mechanical
analysis of rock mass displacements due to underground mining properties of rocks by fuzzy modelling. Comput Geotech
in mountainous areas. Int J Rock Mech Min Sci 43:503–511 34:10–18
Rumelhart DE, McClelland JL (1986) Parallel distributed processing: Ulusay R, Gokceoglu C, Sulukcu S (2001) Draft ISRM suggested
explorations in the microstructure of cognition, vols 1–2. MIT method for determining block punch strength index (BPI). Int J
Press, Cambridge Rock Mech Min Sci 38:1113–1119
Scoble MJ, Muftuoglu YV (1984) Derivation of a diggability index Yager R, Filev D (1994) Generation of fuzzy rules by mountain
for surface mine equipment selection. Min Sci Technol clustering. J Intell Fuzzy Syst 2(3):209–219
1:305–322 Yagiz S, Gokceoglu C (2010) Application of fuzzy inference system
Sonmez H, Gokceoglu C, Ulusay R (2003) An application of fuzzy and nonlinear regression models for predicting rock brittleness.
sets to the geological strength index (GSI) system used in rock Exp Syst Appl 37:2265–2272
engineering. Eng Appl Artif Intell 16:251–269 Zhou Q, Purvis M, Kasabov N (1997) A membership function
Sonmez H, Tuncay E, Gokceoglu C (2004) Models to predict the selection method for fuzzy neural networks. In: Proc ICONIP,
uniaxial compressive strength and the modulus of elasticity for Dunedin, Springer Verlag Singapore, pp 785–788
Ankara agglomerate. Int J Rock Mech Min Sci 41:717–729

123