Safety Science 55 (2013) 26–33

Contents lists available at SciVerse ScienceDirect

Safety Science
journal homepage: www.elsevier.com/locate/ssci

A novel fuzzy inference system for predicting roof fall rate in underground
coal mines
Mojtaba Razani a, Abdolreza Yazdani-Chamzini a,⇑, Siamak Haji Yakhchali b
a
Young Researchers Club, South Tehran Branch, Islamic Azad University, Tehran, Iran
b
Department of Industrial Engineering, College of Engineering, University of Tehran, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Roof fall is one of the most important hazards connected with underground coal mines, because signif-
Received 20 July 2012 icantly affects financial and human losses. On the other hand, according to different parameters make
Received in revised form 22 August 2012 a significant impact on roof fall and these factors are ill-defined or even immeasurable, this problem is
Accepted 24 November 2012
an uncertain and complex issue. Fuzzy logic is a useful tool to handle the existing uncertainty to be more
Available online 31 January 2013
adapted to the real world problems. In this paper, fuzzy inference system (FIS) is applied to predict roof
fall rate more accurate, precise, and sure for controlling, mitigating, and/or even eliminating the risk of
Keywords:
roof fall. The model utilizes subtractive clustering method to generate fuzzy rules based on 109 data of
Roof fall rate
FIS
roof performance from US coal mines. The established model is evaluated by testing dataset based on
ANN three indices, including coefficient of determination (R2), mean absolute error (MAE), and root mean
MVR square error (RMSE). The results derived from the FIS model in comparison with artificial neural network
Underground coal mines (ANN) and multivariate regression (MVR) model demonstrate that prediction of roof fall rate by the FIS
model is more accurate and satisfied.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction different aspects of science and engineering real world problems

(Bakhtyar et al., 2008; Sylaios et al., 2009; Khatibi and Montazer,
Underground coal mining has been recognized as one of the 2010; Yagiz and Gokceoglu, 2010; Arabacioglu, 2010; Tay and
most hazardous activities in all around the world, because under- Lim, 2011; Malinowska, 2011; Alidoosti et al., 2012; Amindoust
ground coal miners deal with different occurrences that may lead et al., 2012; Fouladgar et al., 2012; Pouzols et al., 2012). Some
to fatal or non-fatal injuries. Roof fall is one of the most important applications of FIS to nonlinear systems can be found in (Du and
hazards that the miners face with. It accounted for 18.18% of all Zhang, 2008). Mamdani systems, Takagi–Sugeno (TS), and Tsukam-
fatal accidents in coal mines, contributing about 35.29% of all fatal oto models are three common methods of deductive inference for
accidents in below-ground operations in 2005 (Palei and Das, fuzzy systems (Ross, 2010). The TS fuzzy systems are widely used
2008). According to the importance of the problem, different re- in modeling complex systems because of their simplicity in the
searches are conducted to find the relationship between the roof inference procedure, the possibility to incorporate a general condi-
fall and the effective parameters on occurrences of the roof fall in tion on the physical structure of the system into the fuzzy system,
underground coal mines. Table 1 lists several studies conducted flexibility, and user friendly in producing fuzzy model (Sylaios
on finding the relation between the roof fall and other effective et al., 2008, 2009). In this study, the TS fuzzy inference system is
factors. applied for finding the relation between roof fall rate and other
However, the complexity of geological deposit and variability of effective parameters.
mining parameters leads to the occurrences of unwanted roof falls Likewise, it cannot be determined exactly which occurrences
(Palei and Das, 2009). Therefore, applying a proper technique that recorded in associated with particular types of roof falls; so that,
can take simultaneously both the complexity and inherent uncer- the roof fall data from US underground coal mines is unlabeled.
tainty connected with roof fall problem into account help design- The merit of using the clustering methods is to determine the
ers to analyze the problem more accurate and precise. Fuzzy members of such unlabeled data sets to fuzzy classes when the gi-
logic is a useful mathematical tool for modeling the existing uncer- ven data set does not have any known label values (Bertoncini and
tainty and complexity. This method is successfully applied in the Hinders, 2010).
The main aim of the study is to develop an FIS model to predict
⇑ Corresponding author. Tel./fax: +98 9132820820. the roof fall rate in order to obtain a more accurate, precise, and
E-mail addresses: [email protected] (A. Yazdani-Chamzini), yakhchali@ sure model. To show the capability and effectiveness of the
yahoo.com (S.H. Yakhchali). constructed model, the results are compared with artificial neural

0925-7535/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ssci.2012.11.008
 M. Razani et al. / Safety Science 55 (2013) 26–33 27

Table 1
A list of several studies on modeling roof fall.

Proposed by Parameters involved in modeling Application

Kidybinski (1977) Average roof rock strength Investigations, estimation and classification of roofs in mines in
the United States of America for the selection of suitable
mechanized support for longwalls
Unrug and Szwilski (1980) Roof quality index (RQI) Influence of strata control parameters on longwall mining design
Newman and Bieniawski Strata weatherability, High horizontal stresses, and the roof Entry and roof support design in underground room-and-pillar
(1986) support reinforcement factor coal mines
Molinda and Mark (1994) Groundwater, surcharge, rock strength, strong bed, Quantifying descriptive geological information for use in coal
discontinuities, spacing, cohesion, roughness, persistence, mine design and roof support selection
bedding contact, and moisture sensitivity
Mark (1999) Entry with, coal mine roof rating (CMRR), cut depth, and cover Stability of extended cut
Molinda et al. (2001) CMRR Roof stability
Mark et al. (2002) Rock quality designation (RQD), uniaxial compressive strength Longwall tailgate design, roof bolt selection, longwall mining
(UCS), and diametral point load testing through open entries and recovery rooms, and roof fall
evaluations
Deb (2003) Intersection diagonal span (IS), CMRR, Primary roof support Analysis of coal mine roof fall rate
(PRSUP), and Depth of mine
Duzgun and Einstein (2004) Injury, equipment damage, interruption and delay in operation, Assessment and management of roof fall risks in underground
and clean up coal mines
Duzgun (2005) Distribution function of the annual roof fall and cost model Risk assessment and management methodology for roof fall
hazards in underground mines of the Zonguldak coal basin
Torabi et al. (2008) UCS Prediction of the roof rock strength in underground coal mines
Palei and Das (2008) Number of bolts per row, anchorage strength of bolt, spacing Predicting the effects of contributing parameters on roof falls in
between the bolts, width of gallery, mean rock density, and rock underground coal mines
mass rating (RMR)
Javanshir et al. (2009) UCS, tensile strength, moisture sensitivity, joints, bedding Modeling and classification of roof behavior in coal mines
thickness, orientation and dips of discontinuities relevant to face,
and surface of working area
Palei and Das (2009) Width of gallery, mining height (MH), depth of cover (DOF), seam Prediction of roof fall risks in bord and pillar workings in coal
thickness, roof support status, immediate roof, face and specific mines by using Logistic regression model
mine
Ghasemi and Ataei (2012) CMRR, PRSUP, IS, and DOF Predicting roof fall rate in coal mines

.
network (ANN) and multivariate regression (MVR) models based such as: pattern reorganization, classification, data processing, and
on three indices, including coefficient of determination (R2), mean process control. ANN technique has some unique futures which
absolute error (MAE), and root mean square error (RMSE). distinguish them from other data processing systems include abil-
The rest of the paper is organized as follows: multivariate ity to work successfully even when they are partly damaged, par-
regression is briefly illustrated in the ext section. Section 3 de- allel processing, ability to make generalization, and little
scribes artificial neural network (ANN). The basic concepts of fuzzy susceptibility to errors in data sets (Malinowski and Ziembicki,
inference system (FIS) are presented in Section 4. Section 5 ex- 2006). An artificial neural network simulates the human brain
plains the indices applied in model performance. Data set is de- mechanism to implement computing behavior (He and Xu, 2007).
scribed in Section 6. In Section 7, the FIS system is applied to ANN is developed based on biological neural networks, which
predict the roof fall rate and its performance is investigated in neurons are the basic building blocks ones. An artificial neuron is
comparison with ANN and MVR models. In Section 8, results and a model of a biological neuron. An artificial neuron receives signals
discussion are discussed. Section 9 includes the conclusions of from other neurons, gathers these signals, and when fired, trans-
the study. mits a signal to all connected neurons (Engelbrecht, 2002). An arti-
ficial neuron model is depicted in Fig. 1.
2. Multivariate regression analysis As seen in Fig. 1, xi (i = 1, 2, ..., n) is the input signal of n other
neurons to a neuron j; wij is the connection weight between the
Regression analysis is widely used for modeling relationship be- ith neuron and the neuron j; hj is the activation threshold of the
tween inputs and output variables. Based on basic concepts of mul- neuron j; f is the transfer function, and yj is the output of the neu-
tivariate regression (MVR), relationship between inputs variables ron. yj Is calculated through the following equation:
and the output variable is generally explored by the function that !
X
n
is fitted to a dataset. Standard form of the MVR model in statistical yi ¼ f wij xi  hj ð2Þ
analysis can be defined as: i¼1

y ¼ b1 x1 þ b2 x2 þ . . . þ bn xn þ c ð1Þ f is generally linear, step, threshold, logarithmic sigmoid (logsig),

where y is dependent variable, xn (n = 1, 2, . . . , n) are independent hyperbolic tangent sigmoid (tansig) functions.
variables, b1 ; b2 ; . . . ; bn are the regression coefficients, which repre- A neural network contains of three layers, including one input
sent the amount of the dependent variable changes when the corre- layer, several middle layers (hidden layers) and one output layer.
sponding independent variables change 1 unit. c Is constant term,
which represent the amount the dependent variable will be when w1 j
x1
all the independent variables are 0.
x2 w2 j ∑ θi f (∑ wij xi − θ j ) yj
3. Artificial neural networks (ANNs)
xn wnj
ANN technique has emerged as a powerful modeling tool which
can be applied for many scientific and/or engineering applications, Fig. 1. Artificial neuron model.
28 M. Razani et al. / Safety Science 55 (2013) 26–33

4. Fuzzy inference system

Fuzzy approach is a methodology that is capable of taking into

account uncertain and imprecise reasoning. The fuzzy inference
systems are recently more popular tools for solving engineering
problems because of their unique features in predicting complex
phenomena. A fuzzy system is a nonlinear mapping between in-
puts and outputs, which the mapping of inputs to outputs is in part
characterized by a set of ‘‘IF-THEN’’ rules (Sylaios et al., 2009).
A typical fuzzy-logic based approach involves three stages: (1)
fuzzifier, (2) interference engine, (3) defuzzifier as shown in
Fig. 3. In first stage, linguistic variables and MFs are mapped by
transferring the input vector into fuzzy If–Then rules. In other
words, input vector (crisp values) may be translated into linguistic
variables such as very high, high, medium, low, and very low. The
experts’ knowledge on control targets forms the rule base. In this
Fig. 2. A typical artificial neural network architecture. paper, TS inference model, depicted in Fig. 4 for two input vari-
ables, is applied for modeling roof fall rate (RFR); so that the con-
sequence of a rule is derived from a function of input linguistic
It should be noted that there is no theoretical limit on the number variables as presented in the following part:
of hidden layers but typically there is just one or two (Sumathi and
ð1Þ
Paneerselvam, 2010). Fig. 2 depicts an artificial neural network r1;...;n : IF x1 is A1 and . . . and xp is AðpÞ
n THEN y1;...;n
architecture employed in this study. ¼ f1;...;n ðx1 ; . . . ; xp Þ ð3Þ
In this study, a multi-layer feed-forward perceptron (MLP) with
a back propagation learning algorithm is employed to model gold where r 1;...;n represents the fuzzy rules; p is the number of input
price. During the modeling stage, coefficients are adjusted through data; and n is the number of rules. AnðpÞ are input linguistic values;
comparing the model outputs with actual outputs. A five-step pro- xn is the nth input variable; and f1;...;n are deterministic functions
cedure can be described to present the learning process of an ANN corresponding to the output of the fuzzy model (y1;...;n ).
as follows: In second stage, the fuzzy inference engine uses these fuzzy If-
Then rules to assign a map from fuzzy inputs to fuzzy outputs
1. Input–output vectors are randomly selected as training based on fuzzy composition rules (Li, 2006). Finally, the defuzzifier
datasets, is used to transfer fuzzy sets into crisp value. The output of the fuz-
2. The structure of network is formed, zy model using the weighted averaging is as follows:
3. Network outputs are computed for the selected inputs, X
n
4. Connection weights are adjusted according to performance wi yi
measure, i¼1
y¼ ð4Þ
5. The process of adjusting the weights is continued until perfor- Xn

mance measures are satisfied.

wi
i¼1

where wi is the firing strength for each rule.

5. Model performance

In order to evaluate the established model, several measures of

accuracy are employed to see how far the model is capable of
forecasting the rate of roof fall. For this reason, the models are eval-
uated by three estimators contain of the coefficient of determina-
tion (R2), the root mean square error (RMSE), and the mean
absolute error (MAE). The aforementioned estimators are obtained
Fig. 3. Structure of FIS. by:

Fig. 4. Graphical TS inference method for two input variables.

 M. Razani et al. / Safety Science 55 (2013) 26–33 29

X
N
ðAi  Pi Þ2
i¼1
R2 ¼ 1  ð5Þ
X
N
 i Þ2
ðAi  A
i¼1

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
uX
u ðA  P Þ2
u i i
t i¼1
RMSE ¼ ð6Þ
N

X
N
Fig. 5. Method of measuring intersection diagonals (Molinda et al., 2000).
jAi  Pi j
i¼1
MAE ¼ ð7Þ
N
coal mining. As a result, its failure is the main cause of fatalities
 i is the average
where Pi is predicted values, Ai is observed values, A and injuries. In many cases, increasing the roof bolt density is
of observed set, and N is the number of datasets. the simplest way in order to reduce the level of roof fall risk.
R2 shows how much of the variability in dependent variable can PRSUP is a roof bolt density indicator that is calculated by the
be explained by independent variable(s). R2 is a positive number following equation (Molinda et al., 2000; Ghasemi and Ataei,
that can only take values between zero and one. A value for R2 2012):
close to one shows a good fit of forecasting model and a value close
to zero presents a poor fit. Lb  N b  C
PRSUP ¼ ð8Þ
MAE would reflect if the results suffer from a bias between the 14:5  Sb  W e
actual and modeled datasets (Khatibi et al., 2011). RMSE is a used where Lb is length of the bolt in m, Nb is the number of bolts per
measure in order to calculate the differences between values pre- row, C is the bolt capacity in KN, Sb is the spacing between rows
dicted by a model and the values observed from the thing being of bolts in m, and We is entry width in m.
modeled. RMSE and MAE are non-negative numbers that for an IS: Roof fall reports in underground coal mines show that most
ideal model can be zero and have no upper bound. of the falls occur in intersections. The intersection fall rate
proves that intersections are much more likely to fall than entry
6. Data set or crosscut segments between intersections (Molinda et al.,
2000). Because unlike of entries and crosscuts, rock load applied
The data set involved in this study includes 109 observations of on roof in intersections is proportional to the cube of the span
the roof fall rate against its affecting parameters (Molinda et al., (Molinda et al., 1998). As seen in Fig. 5, the IS can be defined
2000). In order to construct the ANFIS model for the roof fall rate, as the sum of the two intersection diagonals.
the existing data set is separated into training and test sets. For DOF: In underground coal mines, deep of cover is one of the root
achieving the aim, 82 observations (75%) are randomly selected causes of roof fall. Because a rise in depth leads to a vertically
to formulate the model and the rest of the observations (25%) are and horizontally increase in virgin stress levels in the rock mass
used to reflect the performance of the different constructed mod- (Ghasemi and Ataei, 2012). Hence, it is harder to achieve
els. Base on the ARIMA model, the affecting parameters on the roof enough stability at high depth, and a set of measurements
fall rate are extracted as described in the following. should be conducted to ensure roof stability.
One of the most important steps in developing a successful pre- MH: Mining height is one of the most important factors in the
dicting model is the selection of the input variables, which deter- roof fall accidents. Based on the study accomplished by Fotta
mines the architecture of the model. Based on the literature and Mallett (1997), regardless of the employment size, as min-
review and a lot of face-to-face interviews, five input parameters ing height increases, miners are at increasingly higher risk of
for predicting the roof fall rate (RFR) are recognized, including injury from accidents due to slips or falls and accidents involv-
CMRR, PRSUP, IS, DOF, and MH. Each of these variables and their ing shuttle cars and falls of ground. Basic descriptive statistics
impact on RFR are illustrated in the following part. on the dataset involved in modeling are presented in Table 2.

RFR: The roof fall rate is considered as the output parameter

that is a function of total number of roof falls and drivage. Dri- 7. Prediction model
vage is estimated by covering annual production (excluding
longwall production) into linear feet of advance (Dolinar et al., Based on classical assumptions of the _linear model, the best
2001). RFR is calculated by dividing the total number of roof forecast of Y (variable Y regressed on X) is Y can be calculated by:
falls by the drivage. _
CMRR: Quality of roof rock plays a key role in the rate of roof Y ¼ XðX T XÞ1 X T ð9Þ
fall. Roof fall reports in coal mines demonstrated that the weak
Here, Y is RFR, and X = [1, CMRR, PRSUP, IS, DOF, MH] is the set of
roof was the main reason of fatal incidents (Ghasemi and Ataei,
independent variables defined, respectively, as the constant term,
2012). CMRR, developed by Molinda and Mark (1994), is an
Coal Mine Roof Rating, Primary roof support, Intersection diagonal
indicator for representing the quality of roof rock in coal mines.
span, Depth of cover, and Mining height, so that the model is re-
This indicator takes a value between 0 and 100, which 0
sulted as follows:
expresses a weaker roof and 100 shows a stronger roof. CMRR
classification is one of the most common methods because of RFR ¼ 5:26  0:1½CMRR þ 0:482½PRSUP þ 0:0114½IS
considering natural causes of roof fall such as groundwater,
 9:854E  5½DOF  0:136½MH ð10Þ
bedding and other discontinuities, and strength of roof rock.
PRSUP: Roof bolts are usually applied as the only primary sup- The performance measures of the MVR model for data set are
port systems overhead protection of miners in underground listed in Table 3. From Table 3, it can be seen that the MVR model
30 M. Razani et al. / Safety Science 55 (2013) 26–33

Table 2
Statistical parameters of each data set.

Type of data Mean Median Maximum Minimum Std. Dev. Skewness Kurtosis
CMRR Input 47.72477 45 78 28 11.10461 0.854792 3.667385
DOF Input 445.5505 400 1100 40 226.57 1.182488 4.064498
MH Input 6.284404 6.5 10 3 1.940226 0.093543 2.704866
IS Input 63.44771 63 78.4 50 5.607594 0.373796 3.418295
PRSUP Input 5.711009 5.32 14.67 2.46 2.286372 1.186847 4.747508
FALL Output 2.750826 0.52 31.82 0 5.251759 2.955404 12.82796

Table 3
Testing dataset applied for performance measurement.

No. CMRR PRSUP IS (ft) DOF (ft) MH (ft) RFR

Actual Predicted
MVR ANN FIS
1 59 5.32 65.9 150 6 0 2.049 4.75E04 1.76E17
2 75 3.93 60 400 7 0 0.158 2.26E02 6.54E16
3 45 4.55 59 350 5.5 0.23 2.361 1.27E02 1.34E+00
4 55 5.47 75.2 1050 10 4.05 1.501 1.53E01 4.59E69
5 55 7.89 75.2 1050 10 0.35 2.867 2.01E01 7.22E76
6 37 5.52 58 400 10 2.25 2.981 8.05E02 2.23E+00
7 58 3.1 64.8 800 6 0 0.376 4.97E02 7.26E02
8 50 4.92 63 350 7 0 2.150 2.03E02 2.01E03
9 39 6.98 60 400 4.3 0.36 4.221 4.19E03 1.79E+00
10 49 4.64 57 150 3 0 2.493 4.17E02 2.10E10
11 47 5.3 62.9 500 3 0 2.823 5.72E02 2.01E01
12 47 3.98 66.3 300 3 0 2.268 6.64E02 2.22E+00
13 30 9.14 56 300 6 2.8 5.906 2.44E02 4.00E+00
14 40 5.9 63 400 7 0 1.421 9.44E04 5.89E16
15 52 3.93 63 400 7 3.17 3.840 6.63E02 5.87E01
16 35 6 64 200 7.5 0 3.949 1.01E01 1.22E16
17 50 7.95 60 300 6 20 3.336 5.87E01 2.38E+01
18 40 5.9 63 400 7 0 3.336 9.44E04 5.89E16
19 39 5.89 69 500 7.5 0.16 3.357 1.19E01 3.51E+00
20 55 3.71 62.7 1000 6.5 0.28 0.702 1.08E02 1.80E04
21 55 8.21 75.2 1050 10 0.79 3.047 2.57E01 7.02E76
22 38 4.32 60 600 4 0.65 2.669 1.44E01 1.78E+00
23 45 4.55 59 350 5.5 0.29 2.361 1.27E02 1.34E+00
24 44 4.36 62 350 3.5 0 2.544 3.31E02 9.94E01
25 58 3.1 75.7 800 6 3.33 0.519 6.03E03 1.42E09
R2 0.039 0.687 0.872
MAE 2.834 1.711 1.119
RMSE 4.033 2.54 1.72

is not satisfied. The major conceptual limitation of all regression Therefore, an FIS model based on TS inference system is con-
techniques is that one can only ascertain relationships, but never structed to recognize the complex patterns involved in the data
be sure about underlying causal mechanism (Yilmaz and Kaynar, set; so that, ANN is not able to properly identify them. Similar to
2011). the ANN model, using established algorithms in MATLAB 7.11
Then, according to the basic concepts of ANN design and using package software, the best-fitted FIS model is selected. In this
productive algorithm in MATLAB 7.11 package software in order to study, there are five input variables in the problem of RFR predic-
obtain the optimum network architecture; several network archi- tion, including CMRR, PRSUP, IS, DOF, and MH. Therefore, fuzzy ‘‘IF-
tectures are established to select the best-fitted ANN model. Before THEN’’ rules can be defined as follows:
constructing the ANN model, all variables were normalized to the
interval of 0 and 1 to provide standardization using Eq. (11): IF ðCMRR is . . .Þ; AND ðPRSUP is . . .Þ; AND ðIS is . . .Þ; AND
ðDOF is . . .Þ; AND ðMH is . . .Þ; THEN
X norm ¼ ðX  X min Þ=ðX max  X min Þ ð11Þ ðRFR ¼ a  CMRR þ b  PRSUP þ c  IS
þ d  DOF þ e  MH þ gÞ: ð12Þ
The best fitted network based on the best forecasting accuracy
with the test data is contained of five inputs, eighteen hidden The parameters a, b, c, d, e, and g are estimated from the training
and one output neurons (in abbreviated form, N(5-18-1)). This con- dataset using linear least-squares estimation. The structure of the
firms that simple network structure that has a small number of FIS model is depicted in Fig. 6. In order to generate the optimum
hidden nodes often works well in out-of-sample forecasting number of ‘‘IF-THEN’’ rules and determine the MFs, the subtractive
(Areekul et al., 2010; Khashei and Bijari, 2011). This can be due clustering method is applied. This method is utilized to granule
to the overfitting problem in neural network modeling process that conditional attributes (technical indicators) and it reduces proba-
allows the established network to fit the training data well, but bly the amount of FIS rules. The method uses the given search ra-
poor generalization may happen. The effectiveness of the ANN dius to measure the density of data points in the feature space
model for predicting the RFR is presented in Table 3. It can be seen (Chiu, 1994). The algorithm employed in this method selects the
that the output of the established model is not fitted well. data point with the highest potential to be the first cluster centre,
 M. Razani et al. / Safety Science 55 (2013) 26–33 31

Fig. 6. Input–output structure of established FIS model.

removes all data points in the vicinity of the first cluster centre in by using control surface as depicted in Fig. 8. As seen in the figure,
order to determine the next data cluster and its centre location and Fig. 8a shows the inter dependency of RFR on CMRR and PRSUP,
iterates on this process until all of the data is within radii of a clus- Fig. 8b depicts inter dependency of RFR on CMRR and IS, Fig. 8c
ter centre (Yetilmezsoy et al., 2011). The optimum cluster radius shows inter dependency of RFR on CMRR and DOF, and Fig. 8d de-
can be systematically resulted by changing in the values of four picts inter dependency of RFR on CMRR and MH.
parameters range of influence (ROI), squash factor (SF), accepted
ratio (AR), and rejected ratio (RR) around their default (ROI = 0.5, 8. Results and discussion
SF = 1.25, AR = 0.5, and RR = 0.15).
The optimum clustering parameters are found by a trial and er- The results derived from the developed fuzzy model agree well
ror approach as follows: range of influence: 0.1, squash factor: 1.2, with the recorded data and outperform MVR and ANN models.
accept ratio: 0.2 and reject ratio: 0.15. It produces the highest R2 of Three performance indices applied on the dataset demonstrate
0.872 with lowest MAE and RMSE of 1.119 and 1.72, respectively. the effectiveness and capability of the fuzzy model based on sub-
The optimal number of rules for best-fitting model is 84. A sample tractive clustering to predict the RFR values in coal mines. The per-
of rules generated by the ANFIS model is presented in Fig. 7. formance measures of the FIS, ANN, and MVR models for testing
After forming the initial FIS structure, the training stage is data set are listed in the last three row of Table 3. It can be shown
accomplished with the help of the training dataset. Then, the that the R2 value for the FIS model is 0.872, which is significantly
established model is checked by the testing data set. For achieving greater than that obtained in ANN and MVR, which are 0.687 and
the aim, the input vectors from the test data set are presented to 0.039 (R2 less than 80% is unsatisfactory (Shamseldin et al.,
the constructed network and the predicted outputs are compared 1997)), respectively.
with the actual ones for the performance measurement as listed The MAE value for the FIS model is 1.119, which is considerably
in the last column of Table 3. smaller than that resulted by ANN and MVR, which are 1.711 and
The interdependency of input and output parameters derived 2.834, respectively. The RMSE value for the FIS model is 1.72,
from the rules generated by subtractive clustering can be shown which is dramatically smaller than that yield by ANN and MVR,

1. If (CMRR is in1cluster1) and (PRSUP is in2cluster1) and (IS is in3cluster1) and (DOF is in4cluster1) and (MH is in5cluster1) then (RFR is out1cluster1) (1)
2. If (CMRR is in1cluster2) and (PRSUP is in2cluster2) and (IS is in3cluster2) and (DOF is in4cluster2) and (MH is in5cluster2) then (RFR is out1cluster2) (1)
3. If (CMRR is in1cluster3) and (PRSUP is in2cluster3) and (IS is in3cluster3) and (DOF is in4cluster3) and (MH is in5cluster3) then (RFR is out1cluster3) (1)
4. If (CMRR is in1cluster4) and (PRSUP is in2cluster4) and (IS is in3cluster4) and (DOF is in4cluster4) and (MH is in5cluster4) then (RFR is out1cluster4) (1)
5. If (CMRR is in1cluster5) and (PRSUP is in2cluster5) and (IS is in3cluster5) and (DOF is in4cluster5) and (MH is in5cluster5) then (RFR is out1cluster5) (1)
6. If (CMRR is in1cluster6) and (PRSUP is in2cluster6) and (IS is in3cluster6) and (DOF is in4cluster6) and (MH is in5cluster6) then (RFR is out1cluster6) (1)
7. If (CMRR is in1cluster7) and (PRSUP is in2cluster7) and (IS is in3cluster7) and (DOF is in4cluster7) and (MH is in5cluster7) then (RFR is out1cluster7) (1)
8. If (CMRR is in1cluster8) and (PRSUP is in2cluster8) and (IS is in3cluster8) and (DOF is in4cluster8) and (MH is in5cluster8) then (RFR is out1cluster8) (1)
9. If (CMRR is in1cluster9) and (PRSUP is in2cluster9) and (IS is in3cluster9) and (DOF is in4cluster9) and (MH is in5cluster9) then (RFR is out1cluster9) (1)
10. If (CMRR is in1cluster10) and (PRSUP is in2cluster10) and (IS is in3cluster10) and (DOF is in4cluster10) and (MH is in5cluster10) then (RFR is out1cluster10) (1)
11. If (CMRR is in1cluster11) and (PRSUP is in2cluster11) and (IS is in3cluster11) and (DOF is in4cluster11) and (MH is in5cluster11) then (RFR is out1cluster11) (1)
12. If (CMRR is in1cluster12) and (PRSUP is in2cluster12) and (IS is in3cluster12) and (DOF is in4cluster12) and (MH is in5cluster12) then (RFR is out1cluster12) (1)
13. If (CMRR is in1cluster13) and (PRSUP is in2cluster13) and (IS is in3cluster13) and (DOF is in4cluster13) and (MH is in5cluster13) then (RFR is out1cluster13) (1)
14. If (CMRR is in1cluster14) and (PRSUP is in2cluster14) and (IS is in3cluster14) and (DOF is in4cluster14) and (MH is in5cluster14) then (RFR is out1cluster14) (1)
15. If (CMRR is in1cluster15) and (PRSUP is in2cluster15) and (IS is in3cluster15) and (DOF is in4cluster15) and (MH is in5cluster15) then (RFR is out1cluster15) (1)
16. If (CMRR is in1cluster16) and (PRSUP is in2cluster16) and (IS is in3cluster16) and (DOF is in4cluster16) and (MH is in5cluster16) then (RFR is out1cluster16) (1)
17. If (CMRR is in1cluster17) and (PRSUP is in2cluster17) and (IS is in3cluster17) and (DOF is in4cluster17) and (MH is in5cluster17) then (RFR is out1cluster17) (1)
18. If (CMRR is in1cluster18) and (PRSUP is in2cluster18) and (IS is in3cluster18) and (DOF is in4cluster18) and (MH is in5cluster18) then (RFR is out1cluster18) (1)
19. If (CMRR is in1cluster19) and (PRSUP is in2cluster19) and (IS is in3cluster19) and (DOF is in4cluster19) and (MH is in5cluster19) then (RFR is out1cluster19) (1)
20. If (CMRR is in1cluster20) and (PRSUP is in2cluster20) and (IS is in3cluster20) and (DOF is in4cluster20) and (MH is in5cluster20) then (RFR is out1cluster20) (1)
21. If (CMRR is in1cluster21) and (PRSUP is in2cluster21) and (IS is in3cluster21) and (DOF is in4cluster21) and (MH is in5cluster21) then (RFR is out1cluster21) (1)
22. If (CMRR is in1cluster22) and (PRSUP is in2cluster22) and (IS is in3cluster22) and (DOF is in4cluster22) and (MH is in5cluster22) then (RFR is out1cluster22) (1)
23. If (CMRR is in1cluster23) and (PRSUP is in2cluster23) and (IS is in3cluster23) and (DOF is in4cluster23) and (MH is in5cluster23) then (RFR is out1cluster23) (1)
24. If (CMRR is in1cluster24) and (PRSUP is in2cluster24) and (IS is in3cluster24) and (DOF is in4cluster24) and (MH is in5cluster24) then (RFR is out1cluster24) (1)
25. If (CMRR is in1cluster25) and (PRSUP is in2cluster25) and (IS is in3cluster25) and (DOF is in4cluster25) and (MH is in5cluster25) then (RFR is out1cluster25) (1)
26.

Fig. 7. A sample of rules generated for the FIS model.

32 M. Razani et al. / Safety Science 55 (2013) 26–33

Fig. 8. Control surface of RFR on (a) CMRR and PRSUP, (b) CMRR and IS, (c) CMRR and DOF and (d) CMRR and MH.

which are 2.54 and 4.033, respectively. According to the yield re- References
sults, the capability and efficiency of the FIS model for predicting
RFR is more accurate, precise, and sure than two other models. Alidoosti, A., Yazdani, M., Fouladgar, M.M., Basiri, M.H., 2012. Risk assessment of
critical asset using fuzzy inference system. Risk Management 14, 77–91.
Therefore, designers and engineers of underground mines can take Amindoust, A., Ahmed, Sh., Saghafinia, A., Bahreininejad, A., 2012. Sustainable
into account this model as a reference in order to implement pre- supplier selection: a ranking model based on fuzzy inference system. Applied
ventive measures for controlling and minimizing roof fall. Since Soft Computing 12 (6), 1668–1677.
Arabacioglu, B.C., 2010. Using fuzzy inference system for architectural space
CMRR and DOF depend on ground conditions, these two parame- analysis. Applied Soft Computing 10, 926–937.
ters are uncontrollable. While the remaining of parameters (i.e. Areekul, P., Senjyu, T., Toyama, H., Yona, A., 2010. A hybrid ARIMA and neural
PRSUP, IS, and MH) depend on mining conditions and mine design; network model for short-term price forecasting in deregulated market. IEEE
Transactions on Power Systems 25 (1), 524–530.
so that, these three parameters are controllable. Hence, controlling
Bakhtyar, R., Ghaheri, A., Yeganeh-Bakhtiary, A., Baldock, T.E., 2008. Longshore
measures should be focused on mining conditions and mine sediment transport estimation using a fuzzy inference system. Applied Ocean
design. Research 30, 273–286.
Bertoncini, C.A., Hinders, M.K., 2010. Fuzzy classification of roof fall predictors in
microseismic monitoring. Measurement 43, 1690–1701.
Chiu, S.L., 1994. Fuzzy model identification based on cluster estimation. Journal of
9. Conclusion Intelligent Information Systems 2, 267–278.
Deb, D., 2003. Analysis of coal mine roof fall rate using fuzzy reasoning techniques.
International Journal of Rock Mechanics and Mining Sciences 40, 251–257.
This study focuses on developing a TS fuzzy inference system Dolinar, D., Mark, C., Molinda, G., 2001. Design of primary roof support system in
for predicting the values of RFR. Model application is performed U.S. coal mines based on the analysis of roof fall rates. In: Proceedings of Fourth
on 109 datasets of roof performance from US coal mines on the ba- International Symposium on Roof Bolting in Mining, Aachen, Germany,
Martens-PN ed., Druckerei Mainz GmbH, pp. 235–252.
sis of five input variables, including CMRR, PRSUP, IS, DOF, and MH.
Du, H., Zhang, N., 2008. Application of evolving Takagi-Sugeno fuzzy model to
The results of the formed FIS model demonstrate a high correlation nonlinear systems. Applied Soft Computing 8, 676–686.
between recorded and predicted values. The satisfactory accuracy Duzgun, H.S.B., 2005. Analysis of roof fall hazards and risk assessment for
Zonguldak coal basin underground mines. International Journal of Coal
of the prediction model obtained by FIS in comparison with two
Geology 64, 104–115.
popular methods ANN and MVR confirms that FIS model based Duzgun, H.S.B., Einstein, H.H., 2004. Assessment and management of roof fall risks
on subtractive clustering method can be successfully applied for in underground coal mines. Safety Science 42, 23–41.
the problem of the RFR prediction. The results show that the FIS Engelbrecht, A.P., 2002. Computational Intelligence. an Introduction. John Wiley &
Sons, Ltd.
model can serve as a powerful and useful tool for the prediction Fotta, B., Mallett, L.G., 1997. Effects of Mining height on Injury Rates in U.S.
of RFR in US underground coal mines. The output of the model Underground Nonlongwall Bituminous Coal Mines. NIOSH, IC 9447.
can be used for selecting proper reaction strategy in order to re- Fouladgar, M.M., Yazdani-Chamzini, A., Zavadskas, E.K., 2012. Risk evaluation of
tunneling projects. Archives of Civil and Mechanical Engineering 12 (1), 1–
duce or eliminate the risk of the roof fall. 12.
Ghasemi, E., Ataei, M., 2012. Application of fuzzy logic for predicting roof fall rate in
coal mines. Neural Computing and Applications. http://dx.doi.org/10.1007/
Acknowledgement s00521-012-0819-3.
He, X., Xu, Sh., 2007. Process Neural Networks Theory and Applications. Springer.
Javanshir, M., Ataei, M., Torabi, S.R., 2009. Modeling and classification of roof
The authors would like to acknowledge the financial support of behavior in coal mines. International Journal of Industrial Engineering and
University of Tehran for this research. Production Management 19 (9), 45–53.
 M. Razani et al. / Safety Science 55 (2013) 26–33 33

Khashei, M., Bijari, M., 2011. A novel hybridization of artificial neural networks and Newman, D.A., Bieniawski, Z.T., 1986. Modified version of the Geomechanics
ARIMA models for time series forecasting. Applied Soft Computing 11 (2), Classification for Entry Design in Underground Coal Mines. Transactions of
2664–2675. Society of Mining Engineering AIME 280, 2134–2138.
Khatibi, V., Montazer, G.A., 2010. A fuzzy-evidential hybrid inference engine for Palei, S.K., Das, S.K., 2008. Sensitivity analysis of support safety factor for predicting
coronary heart disease risk assessment. Expert System with Application 37, the effects of contributing parameters on roof falls in underground coal mines.
8536–8542. International Journal of Coal Geology 75, 241–247.
Khatibi, R., Ghorbani, M.A., Kashani, M.H., Kisi, O., 2011. Comparison of three Palei, S.K., Das, S.K., 2009. Logistic regression model for prediction of roof fall risks in
artificial intelligence techniques for discharge routing. Journal of Hydrology bord and pillar workings in coal mines: an approach. Safety Science 47, 88–96.
403, 201–212. Pouzols, F.M., Barros, A.B., Lopez, D.R., Sánchez-Solano, S., 2012. Enabling fuzzy
Kidybinski, A., 1977. Roof Rock Stability Tests and Powerd Support Selection technologies in high performance networking via an open FPGA-based
Systems for Longwalls; State-of-the-Art Report, USDI/CMI Joint Research Project development platform. Applied Soft Computing 12, 1440–1450.
No. 14-01-0001-1450, Rep. No. 2. Ross, T.J., 2010. Fuzzy logic with engineering applications, third ed. John Wiley &
Li, Z., 2006. Fuzzy chaotic systems: modeling, control, and applications. Springer- Sons, Ltd.
Verlag, Berlin, Heidelberg. Shamseldin, A.Y., O’Connor, K.M., Liang, G.C., 1997. Methods for combining the
Malinowska, A., 2011. A fuzzy inference-based approach for building damage output of different rainfall-runoff models. Journal of Hydrology 197, 203–229.
risk assessment on mining terrains. Engineering Structures 33, 163– Sumathi, S., Paneerselvam, S., 2010. Computational Intelligence Paradigms: Theory
170. and Applica-tions Using MATLAB. Taylor and Francis Group, LLC.
Malinowski, P., Ziembicki, P., 2006. Analysis of district heating network monitoring Sylaios, G., Gitsakis, N., Koutroumanidis, T., Tsihrintzis, V., 2008. CHLfuzzy: a
by neural networks classification. Journal of Civil Engineering and Management spreadsheet tool for the fuzzy modelling of chlorophyll concentrations in
12 (1), 21–28. coastal lagoons. Hydrobiologia 610, 99–112.
Mark, C., 1999. Application of the Coal Mine Roof Rating (CMRR) to extended cuts. Sylaios, G., Bouchette, F., Tsihrintzis, V.A., Denamiel, C., 2009. A fuzzy inference
Mining Engineering, 52–56. system for wind-wave modeling. Ocean Engineering 36, 1358–1365.
Mark, C., Molinda, G.M., Barton, 2002. New Developments with the Coal Mine Roof Tay, K.M., Lim, Ch.P., 2011. A fuzzy inference system-based criterion-referenced
Rating. In: Proceedings of 21st International Conference on Ground Control in assessment model. Expert System with Application 38, 11129–11136.
Mining. West Virginia University, Morgantoen, WV, pp. 294–301. Torabi, S.R., Sereshki, F., Zare, M., Javanshir, M., 2008. An empirical approach in
Molinda, G., Mark, C., 1994. The Coal Mine Roof Rating (CMRR)-A Practical Rock prediction of the roof rock strength in underground coal mines. Coal Operators’
Mass Classification for coal mines. USBM IC 9387, 83pp. Conference, The AusIMM Illawarra Branch, pp. 132–136.
Molinda, G.M., Mark, C., Bauer, E.R., Babich, D.R., Pappas, D.M., 1998. Factors Unrug, K., Szwilski, A., 1980. Influence of Strata Control Parameters on Longwall
influencing intersection stability in US coal mines. In: Proceedings of the 17th Mining Design. In: Proceedings of the 21st U.S. Symposium on Rock Mechanics,
International Conference on Ground Control in Mining, Morgantown. West Rolla, Mo., 1980, pp. 720–728.
Virginia University, USA, pp 267–275. Yagiz, S., Gokceoglu, C., 2010. Application of fuzzy inference system and nonlinear
Molinda, G.M., Mark, C., Dolinar, D., 2000. Assessing coal mine roof stability through regression models for predicting rock brittleness. Expert System with
roof fall analysis. In: Proceedings of the New Technology for Coal Mine Roof Application 37, 2265–2272.
Support. US Department of Health and Human Services, Centers for Disease Yetilmezsoy, K., Fingas, M., Fieldhouse, B., 2011. An adaptive neuro-fuzzy approach
Control and Prevention, National Institute for Occupational Safety and Health, for modeling of water-in-oil emulsion formation. Colloids and Surfaces A:
NIOSH Publication No. 9453, pp 53–72. Physicochemical and Engineering Aspects 389, 50–62.
Molinda, G.M., Mark, C., Debasis, D., 2001. Using the Coal Mine Roof Rating (CMRR) Yilmaz, I., Kaynar, O., 2011. Multiple regression, ANN (RBF, MLP) and ANFIS models
to assess roof stability in US coal mines. Journal of Mines Metals and Fuels for prediction of swell potential of clayey soils. Expert Systems with
(India), 314–321. Applications 38, 5958–5966.