RESEARCH ARTICLES

Assessment of rib spalling hazard degree in

mining face based on background subtraction
algorithm and support vector machine
Lei Si1,2, Zhongbin Wang1,3,*, Xinhua Liu1, Chao Tan1 and Rongxin Xu1
1
School of Mechatronic Engineering, China University of Mining and Technology, Xuzhou 221116, China
2
Jiangsu Engineering Technology Research Centre on Intelligent Equipment for Fully Mining and Excavating, and
3
Jiangsu Key Laboratory of Mine Mechanical and Electrical Equipment, China University of Mining and Technology, No. 1, Daxue Road,
Xuzhou 221116, China

degree of rib spalling, avoid unnecessary production

Rib spalling is one of the common hazards in a fully
mechanized mining face. In order to accurately assess accidents and improve the safety factor of fully mecha-
the hazard degree, this study proposes a new method nized coal face3,4.
based on background subtraction algorithm and sup- Studies on rib spalling have mainly focused on its pre-
port vector machine (SVM). First, the architecture vention and control technology5–10. In addition, many
diagram of rib spalling feature analysis is constructed, studies reveal the influence regularity of confining pres-
and the rib spalling feature indices are determined, sure on the mechanical properties and spalling effect of
including the duration area, height and the centre of coal wall11–14. The above-mentioned studies analysed
ribs spalling height. Then, the specific feature analysis rarely time, area and quantity of rib spalling, although the
process of rib spalling is performed using the back- mechanism and prevention of rib spalling have been
ground subtraction algorithm. Furthermore, some considered. This cannot meet the requirements of safe
virtual 3D rib spalling animations are generated using
video surveillance of rib spalling for intelligent fully me-
3D Studio Max (3Ds Max) software to verify the rea-
sonability of extracted features. Thereafter, the chanized coal mining.
assessment model of rib spalling hazard degree is In view of the above problems, an assessment method
established based on SVM. Three assessment models of rib spalling hazard degree based on machine vision is
based on SVM, back propagation neural network (BP- proposed here. In the field of machine vision, background
NN) and artificial immune (AI) algorithm have been subtraction algorithm is a commonly used method for
developed. The assessment accuracy of SVM (reaching moving object detection in fixed scenes and has been
85%) is obviously higher than that of BP-NN (75%) widely used in many situations15–18. Hence, a background
and AI (70%) algorithm. The results indicate the fea- subtraction method using double learning rate Gaussian
sibility and superiority of the proposed method in the mixture model is introduced to capture the position and
assessment of rib spalling hazard degree. region of rib spalling, and the feature analysis system is
developed to achieve the feature extraction of rib spalling
Keywords: Background subtraction algorithm, hazard in mining face.
degree assessment, mining face, rib spalling, support In this study, the hazard degree assessment of rib
vector machine. spalling is actually a pattern classification problem.
Therefore, an intelligent classifier should be established
IN coal mines of China, it has been observed that, as the to realize the efficient hazard assessment. Various artifi-
shearer cutting height increases in longwall face, coal cial intelligence pattern recognition techniques have been
wall failure and spalling often occur, especially for brittle widely used for fault detection or condition monitoring,
coal seam. Rib spalling will make hydraulic support such as artificial neural networks19, adaptive neuro-fuzzy
under uneven roof pressure, and the falling coal is easy to inference system20, support vector machine (SVM)21 and
smash into the support area, threatening the safety of so on. SVM has good generalization ability and robust
mine workers. In addition, a large number of coal blocks performance under small sample conditions. It can effec-
will produce a greater impact, resulting in abrupt load tively avoid the problem of over learning, local mini-
change of scraper conveyor, thus affecting the stability of mum, nonlinear and dimension disaster in traditional
the mining area power grid. This kind of problem affects statistical methods22. In this study, SVM classifier is con-
the safety of the entire fully mechanized coal mining structed based on the extracted features to automatically
face1,2. Therefore, it is necessary to evaluate the hazard evaluate the hazard degree of rib spalling, and experi-
ments and simulations are performed to verify the appli-
*For correspondence. (e-mail: wangzbpaper@[Link]) cability and effectiveness of the proposed method.

CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2001

RESEARCH ARTICLES

Figure 1. Flowchart of double learning rate Gaussian mixture background model.

Basic theory Gaussian distribution that best matches It,x. Thus a new
Gaussian distribution is needed to model it and the
Background subtraction algorithm matching result of It,x can be marked by a matching indi-
cator Mt,x, l(t, x):
In the background subtraction algorithm, the commonly
used background models are Gaussian background model ⎧1, if n = l (t , x)
M t , x, n = ⎨ (1)
and Gaussian mixture background model23,24. Lin et al.25 ⎩0, otherwise
presented an improved Gaussian mixture background
modelling method with double learning rate25. Figure 1 ρt ,x,l (t ,x ) (α ) ∈ R denotes the learning rate of parameters
shows the process flow of double learning rate Gaussian μ and σ2, and ηt ,x ( β ) ∈ R denotes the learning rate of
mixture background model, which mainly consists of five weight ω. ηt ,x ( β ) is calculated based on the results of the
steps. previous pixel classification, as shown below.
(1) Maintenance of background model: The double
learning rate Gaussian mixture background modelling ⎧(1 − βb )ηt −1,x + ηb βb , if Ot -1,x =0
method can be represented by pseudo code as follows: ⎪ 2
⎪ β N ( I t , x |μt −1, x,b (t , x ) , σ t −1,x,b (t ,x ) ), if Ot −1,x =1
ηt , x ( β ) = ⎨ d , (2)
Parameters: Tσ = 2.5, σ 02 = 100, ω0 = 0.01 // σ 02 is the ini- ⎪ β s , if Ot −1,x =2
tial variance and ω0 is the weight. ⎪ β , if O
⎩ m t −1, x =3
//Match model
M t , x,n = 0, ∀n = 1, 2,..., N where ηb is a preset constant, β = [βb β d β s β m ]T
dt ,x,n = inf, ∀n = 1, 2,..., N // inf represents a larger ∈ R 4 are the adjustment parameters, b(t , x) =
initial value. arg max n =1,2, , N ωt ,x,n .
(2) Preliminary detection of foreground: The fore-
For n = 1, 2, ... , N do ground and background can be preliminarily distin-
guished using the following equation.
if I t ,x − μt −1, x, n |≤ Tσ σ t −1,x, n then dt ,x, n = −ωt −1,x,n
l (t , x) = arg min n =1,2, , N dt , x, n ⎧0, if ωt , x, n ≥ Tw
if dt ,x,l (t , x ) ≠ inf then M t , x,l (t ,x ) = 1 else l (t , x) = 0 Ft ,x,n = ⎨ . (3)
//Update model ⎩1, otherwise

ωt ,x, n = (1 − ηt , x ( β ))ωt −1,x,n + ηt ,x ( β ) M t ,x,n , ∀n = 1, 2, ,N Tw ∈ R is a preset threshold, and a binary foreground

if M t , x,l (t ,x ) = 1 then mapping can be expressed as Ft = {Ft , x,l (t ,x ) ∀x}. How-
//Update parameters ever, rapid illumination changes can be easily mistakenly
ρt ,x,l (t , x ) (α ) = α N ( I t , x |μt −1, x,l (t ,x ) , σ t2−1,x,l (t , x ) ) detected as a foreground. To solve this problem, Lin et
μt , x,l (t , x ) = (1 − ρt ,x,l (t ,x ) (α )) μt −1,x,l (t , x ) + ρt , x,l (t ,x ) (α ) I t ,x al.25 analysed the difference of pixel values between two
ωt ,x, n = (1 − ηt , x ( β ))ωt −1,x,n + ηt ,x ( β ) M t ,x,n , frames and the interference of illumination changes can
∀n = 1, 2,..., N , be effectively identified and removed, as shown below.
else
//Replacemodel ⎧0, if |I t ,x − I t −1,x | ≤ Td
Dt ,x = ⎨ . (4)
⎩1, otherwise
k = arg min n =1,2, ,N ωt −1,x,n , μt , x, k = I t , x , σ t2,x,k = σ 02 ,

ωt ,x,k = ω0 , ωt ,x,n = ωt , x, n / The foreground mapping Ft can be replaced by

ωt ,x,k = ω0 , N
Ft′ = (Ft ∩ (Ft −1′ ) ∪ (Ft ∩ Dt ).
∑ ωt ,x,n∀n = 1, 2,..., N , (5)
n =1
Thus, the foreground detected by the above operation
where It,x denotes the single channel greyscale image pix- may include shadow, static foreground and motion fore-
el value of position x at time t. l (t , x) is used to mark the ground, which need further classification.

2002 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019

 RESEARCH ARTICLES

Figure 2. Architecture diagram of feature analysis of rib spalling.

(3) Shadow detection: This can be achieved through ⎧0, if Ft , x,l (t , x ) = 0, (background)
a shadow recognition method based on colour space ⎪
proposed by Horprasert et al.26. Suppose that the ⎪1, if Ft , x,l (t , x ) = 1 and St , x =1, (shadow)
Ot , x =⎨ .
background image is Bt = {μt ,x,b (t ,x ) |∀x} at time t, ⎪2, if Ft , x,l (t , x ) = 0 and type(I t , x ) = static foreground
b(t , x) = arg max n =1,2, , N ωt , x,n , and the current image is ⎪3, otherwise (motion foreground)
⎩
I t , x , then the luminance difference can be calculated
(9)
from the following equation
By consulting the literature25,26, the relevant parameters
2
φ (α x ) = arg minα x ( I t ,x − α x Bt , x ) , (6) are set as follows: α = 0.025, ηb = 0.01, βd = 0.01,
βs = 1/900, βm = 1/6000, Tw = 0.24, Td = 10 and τ = 0.7.
where αx denotes the luminance intensity ratio, which
indicates the difference between the luminance of the Support vector machine
current pixel and the background image. Thus the lumin-
ance change can also be represented as follows SVM is a machine learning method proposed by Vapnik
et al.30. It can transform low-dimensional nonlinear pro-
CDt ,x = ||I t , x - α x Bt , x ||. (7) blems into high-dimensional linear problems for solving,
and has good classification effect in the case of few
KaewTraKulPong and Bowden27 proposed the shadow samples.
judgment condition In the construction process of SVM classifier, the
selection of parameters (kernel parameter σ and penalty
⎧⎪1, if φ (α x ) < 2.5σ t ,x,b (t ,x ) and τ < CDt ,x < 1 factor C) has a major impact on the performance of the
St , x = ⎨ , (8) classifier. Reasonable parameter values can import SVM
⎪⎩0, otherwise with higher accuracy and stronger generalization ability.
At present, there are three kinds of parameter selection
where τ is a preset threshold, and the binary shadow methods in practical applications, including grid search
mapping can be expressed as St = {St ,x |∀x}. method, cross-validation method and VC dimensional
boundary method31,32. Among them, the k-fold cross-
(4) Target extraction: In fact, Ft and St may include validation method is the most commonly used method.
some noises and interferences which can be removed by The k-fold cross validation method randomly puts n
morphological processing28. First, the square structure training samples into k containers, that is, k-fold: S1,
element of 5 × 5 is used to corrode Ft and St separately S2, ..., Sk. Then, k experiments are carried out and set Si is
to remove less foreground interference, and then an selected as the testing set for each experiment. The rest of
expansion operation is performed to restore the shape of the subset is used as the training samples, so that the
the remaining foreground. number (li) of samples misclassified in each experiment
can be obtained. The k-fold cross validation error is
(5) Target type classification: Through above steps, defined as
the background, foreground and shadow can be effective-
k
ly classified. In order to further classify the foreground l
E = ∑ i. (10)
into static foreground and motion foreground, we i =1 l
have used the formula of Chen et al.29. Therefore,
the type of I t ,x can be expressed by the following In practical applications, there are usually two ways to
formula: select k: (1) k = 10. So ten-fold cross-validation error can
CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2003
RESEARCH ARTICLES

Figure 3. Flowchart of feature analysis of rib spalling.

be obtained. (2) k = l. Only one sample is taken for Determination of rib spalling feature indexes
testing in one verification process, and the rest is used as
training samples. So it can also be called leave-one-out, The occurrence of rib spalling is mainly manifested by
(LOO for short). Here, the l-fold cross-validation error the slip failure of coal wall. In this process, larger coal
becomes the LOO error. blocks fall down to the floor, forming obvious depression
When choosing kernel parameter and penalty factor, area on the coal wall. In order to accurately describe the
different parameter values can be regarded as different characteristics of rib spalling, it is necessary to reasona-
schemes; thus the ten-fold cross-validation error or LOO bly select the feature indices, which can reflect the scale
error of each scheme can be calculated. The optimal and hazard degree of rib spalling. Due to poor production
parameter combination is the scheme corresponding to environment of fully mechanized coal face, we cannot
the minimum error. acquire good quality surveillance photographs. Therefore,
it is necessary to avoid the characteristic index which is
difficult to extract from these photographs, such as the
Feature analysis of rib spalling based on
depth information of coal-wall strip area. Based on the
background subtraction algorithm
viewpoints of experts and field operators, four main
indices for analysing the characteristics of rib spalling are
Architecture diagram of rib spalling feature
selected here, viz. duration, area, height and centre-height
analysis
of rib spalling.
Feature analysis of rib spalling is used to extract the
characteristics of rib spalling from the process of occur- Feature analysis process of rib spalling in mining
rence to the end and then express them with data. This dace
process needs to be described by a reasonable analysis
system (Figure 2). Feature analysis of rib spalling is the Based on the background subtraction method with double
basis of hazard degree evaluation and the feature analysis learning rate Gaussian mixture background modelling,
system mainly consists of two parts, viz. feature indices the characteristics of rib spalling are analysed; Figure 3
of rib spalling and feature analysis process. is a flowchart of feature analysis of rib spalling. The

2004 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019

 RESEARCH ARTICLES

Figure 4. Some screenshots of rib spalling animation.

Table 1. Ten groups of characteristic data of simulated rib spalling

Rib spalling Area of rib Height of rib Centre height of

Number duration (s) spalling (m2) spalling region (m) rib spalling (m)

1 3.500 1.671 1.067 1.673

2 3.667 1.853 1.078 1.592
3 3.833 1.732 1.110 1.691
4 4.333 1.558 1.087 1.621
5 4.667 1.915 1.059 1.598
6 4.833 1.649 1.101 1.601
7 5.167 1.873 1.093 1.671
8 5.667 1.657 1.086 1.664
9 5.833 1.702 1.063 1.607
10 6.167 1.815 1.102 1.673

foreground detection operation can judge the falling of P (a binary image), the number of pixels whose values
rib spalling, and at this time the occurrence time is rec- are not zero in P as N1 the number of total pixels in It+n as
orded until the moving foreground target cannot be de- N2, and the estimated area of coal wall in It+n as ST, the
tected. Then the rib spalling duration is obtained. In rib spalling area S can be calculated by the following
addition, by extracting the rib spalling region from the formula
last frame, the other three features are further analysed.
(1) Rib spalling duration: According to Figure 3, we N1
S= ST . (12)
can get the start frame It and end frame It+n of rib spal- N2
ling, where n is the interval between the two frames.
Then the duration of rib spalling can be obtained using
(3) Height of rib spalling region: By traversing the rib
the following formula
spalling region P column by column, the column with the
largest number of non-zero pixels can be found, and the
n number of non-zero pixels in the column is counted as N3.
T= , (11)
Fr Assuming that the number of rows in It+n is R and the
height of coal wall is estimated to be HT, the height of rib
where T is the rib spalling duration and Fr is the frame spalling region H can be expressed as
rate.
(2) Area of rib spalling: In the process of rib spalling, N3
H= HT . (13)
the falling coal block can be detected as a foreground, R
and the depression area after rib spalling can be detected
by shadow detection algorithm. By analysing the lumi- (4) Centre-height of rib spalling: By dividing the
nance and colour changes of the coal wall, the depression first-order matrix of the image by the zero-order matrix,
area can be identified as the rib spalling region. Denoting the gravity centre of P can be obtained, where the origin
the rib spalling region detected from the end frame It+n as is the first point in the upper left corner of the image.
CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2005
RESEARCH ARTICLES

Figure 5. Detection results of foreground in the first group of simulated rib spalling animation.

Then the centre-height of rib spalling HC can be calcu- Figure 5 shows the detection results of the foreground
lated as and rib spalling region. In this figure, the white portions
represent the foreground region, and the yellow portions
R − Cy is the rib spalling region. The detection results of the strip
HC = HT . (14) area are not accurate because of the influence of caving
R
coal blocks. With the process of rib spalling, the strip
area is gradually stable and the detection results are more
Simulation analysis satisfactory. Seen from the 110th frame, the detection re-
sults of the rib spalling region are consistent with the
In order to verify the performance of the feature analysis shape in the simulation scene.
method for rib spalling, we used the 3D Studio Max (3ds Figure 6 shows the detection results of the other nine
Max) software to establish several groups of virtual 3D groups of simulated animation. It can be seen that the
rib spalling animation, and some screenshots are shown results are approximately the same as those in the anima-
in Figure 4. The frame rate of the simulation animation is tion, indicating the feasibility and effectiveness of the
30 frames/s, frame width is 720 pixels and frame height algorithm. On this basis, four features indices, viz. rib
is 480 pixels. The parameters of the simulation coal wall spalling duration, area, height and centre-height were
are 3 m length, 2 m height, 6 m2 area, and the pixels are further analysed.
586 × 312. Under the same conditions, ten groups of In order to intuitively reflect the analysis results, the
animation with different characteristics are formed to estimated values of rib spalling animations are compared
quantitatively test the feasibility of the feature analysis with the actual values (Figure 7). It can be seen from Fig-
method (Table 1). ure 7 that the average error of rib spalling duration for the
Taking the first group of simulated rib spalling anima- ten groups of animations is 0.203s, the error rate is
tion as an example, the rib spalling image is greyed. 4.26%; the average error of rib spalling area is 0.100 m2,
2006 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019
 RESEARCH ARTICLES

Figure 6. Detection results of ten groups of simulated rib spalling animation.

the error rate is 5.43%; the average error of rib spalling directly evaluates the damage degree. Hazard assessment
height is 0.058 m, the error rate is 5.35% and the average is actually a classification of hazard level caused by rib
error of rib spalling centre-height is 0.092 m, the error spalling, so it can be regarded as a problem of pattern
rate is 5.61%. Therefore, the features of rib spalling recognition. By means of field investigation, the hazard
obtained by this method can be used as the basis for eva- degree of rib spalling can be classified into four grades:
luating the hazard degree. safe, slight, medium and serious. Through prior sample
data, a multi-class classifier based on SVM is trained, and
Assessment model of rib spalling hazard the evaluation process of rib spalling damage degree
degree-based on support vector machine is carried out according to the extracted characteristic
data.
The hazard degree is mainly determined by the amount of There are many ways to solve the problem of multi-
falling coal blocks during rib spalling, but it is difficult to class classification. Here, the ‘one-to-one’ classification
accurately acquire the coal blocks in the actual mining method is adopted. In this way, a total of six classifiers
process. Therefore, on the basis of the extracted features (C1, C2, ..., C6) need to be trained and the category with
from characteristic analysis of rib spalling, this study in- the largest number of votes is the class of the sample.
CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2007
RESEARCH ARTICLES

Figure 7. Comparison of estimated feature values and actual values of rib spalling.

Figure 8. Assessment model for rib spalling hazard degree.

Figure 8 shows the the assessment model for rib spalling bracket and switching power supply and rib spalling
hazard degree. assessment system. Artificial coal wall is composed of
coal powder, pigment and a small amount of cement so as
Experimental validation to ensure that its surface characteristics are similar to the
actual coal wall in the grey image mode; the length,
Construction of experimental platform width and height are 3, 2 and 1 m respectively. The
assessment system of rib spalling includes a webcam, PC
Due to harsh coal mining conditions, safety cannot be and assessment software. The webcam is fixed on the
guaranteed and the costing is quite high; so an experi- bracket to collect videos of rib spalling and the image
mental platform similar to the field situation was con- data are transmitted to the PC through Ethernet. The
structed (Figure 9). The platform mainly consists on three recognition software of rib spalling in PC is used to ana-
parts: artificial simulated coal wall, position adjustment lyse the acquired images and display the final assessment
2008 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019
 RESEARCH ARTICLES

Figure 9. Structure of experimental platform.

Figure 10. Sketch of simulative rib spalling trigger.

results. In this experimental system, the occurrence of The purpose of training phase is to establish a multi-
simulated rib spalling is triggered by an industrial PE class classifier with excellent performance; Figure 11
plastic mesh (Figure 10). The duration time can be shows the flow chart of the training phase. First, N
changed by controlling the speed of the plastic mesh. The groups of parameters σ and C are set and γ = 1/σ 2, γ ∈
rib spalling area, height and centre-height are set in [–10, 10], C ∈ [–10, 10]. σ and C will take discrete
advance to construct different shapes of rib spalling. values in this given range. Through ten-fold cross valida-
tion and l-fold cross validation, the classification error En
Analysis of results can be calculated, where n = 1, 2, ..., N. The multi-class
classifier corresponding to the minimum value En is
Using the above experimental schemes, we conducted 30 selected as the final solution.
tests of each rib spalling type and a total of 120 samples Figure 12 is a plot of the error changes with different
were acquired. Due to the difference of the feature parameter selections. It can be observed from Figure 12 a
indexed and data level, the first step is to normalize the that the variation of γ has little effect on the error of ten-
characteristic data of rib spalling: the duration time T, rib fold cross-validation method, while the influence of C on
spalling area S, height of rib spalling region H and centre- the error is larger. The error value is the smallest at
height of rib spalling HC. The sample datasets were log2C = 5. Figure 12 b shows that the variation of C has
divided into four categories according to the hazard little effect on the error of l-fold cross-validation method,
degree of rib spalling; numbers 1 to 4 represent safe, slight, while the effect of γ on the error is larger. The minimum
medium and serious respectively. Twenty samples were error value occurs at log2γ = 4. Overall, the parameters of
randomly selected from each category as training SVM can be set as: γ = 16, C = 32.
samples and the remaining 40 samples were used for test- After the training stage, the other 40 samples were
ing. used to test the assessment performance of the SVM
CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2009
RESEARCH ARTICLES

Figure 11. Flowchart of the training phase.

Figure 12. Error changes with different parameter selection methods: (a) 10-fold cross-validation method and
(b) one-fold cross-validation method.

classifier; Figure 13 shows the classification results. It six samples, ensuring the safety of coal mine production,
can be observed that six samples are misclassified and the although the assessment accuracy is not very high.
classification accuracy reaches 85%. The reason for the In order to more objectively analyse the effect of the
error in assessment results is that the eigenvalues of some proposed method, two widely applied algorithms, BP
samples are relatively close. It should be noted that the neural network (BP–NN) and artificial immune (AI) algo-
safety degrees are misclassified into unsafe grades in the rithm, have been compared in this study. According to
2010 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019
 RESEARCH ARTICLES
some reports33,34, the parameters of the two algorithms It can be observed that the classification accuracy
are selected as follows. In BP–NN, the input layer node based on BP-NN and AI algorithm only 75% and 70%
n = 4, hidden layer node l = 2, and output layer node respectively, which is significantly lower than that of
m = 4. The input of the input layer is four eigenvalues of proposed method. The comparison result indicates the
each group of samples and the output of the output layer superiority of SVM in dealing with the classification
[1, 0, 0, 0] represents the first category, the output [0, 1, problem of small sample size, and demonstrates the
0, 0] represents the second category, and so on. The feasibility and effectiveness of the proposed method in
learning rate is η = 0.1, the maximum number of itera- the assessment of rib spalling hazard degree.
tions is 100. In AI algorithm, the number of detectors is
10 and the selection threshold is 0.5. The training and
Conclusion
testing samples are the same as in the proposed method,
and Figures 14 and 15 show the assessment results for
Rib spalling is one of the common hazards in fully
two algorithms.
mechanized mining face. Its occurrence will seriously
affect the safe and reliable operation of mining equip-
ment, and even threaten the lives of workers. Here, a new
method based on background subtraction algorithm and
SVM is proposed to evaluate the hazard degree of rib
spalling. The feature analysis system of rib spalling is
established and analysed using the background subtrac-
tion method. Four indices are reasonably selected, includ-
ing the duration of rib spalling, the area height and
centre-height of rib spalling. Furthermore, experiments
were carried out and three assessment models were
developed. The results indicate the feasibility and supe-
riority of the proposed method in the assessment of rib
spalling hazard degree.
Figure 13. Assessment results based on the proposed method. Thus, this study makes a significant contribution in the
assessment of rib spalling in coal mining face and will
enable safe and efficient production of coal mines. How-
ever, there are some limitations, such as the idealistic
experimental environment. In a future work, we will con-
sider different lighting illumination to reflect the real
working environment in coal mining face.
Conflict of interest: The authors declare that they have
no competing interests.
1. Yao, Q. L. et al., Numerical investigation of the effects of coal
seam dip angle on coal wall stability. Int. J. Rock Mech. Min.,
2017, 100, 298–309.
2. Wang, Z. H., Yang, J. H. and Meng, H., Mechanism and
Figure 14. Assessment results based on back propagation neural net- controlling technology of rib spalling in mining face with large
work. cutting height passing through fault. J. China Coal Soc., 2015, 40,
42–49.
3. Peng, R. et al., Experimental research on the structural instability
mechanism and the effect of multi-echelon support of deep
roadways in a kilometre-deep well. PLoS ONE, 2018, 13,
e0192470.
4. Zhang, G. C. et al., Analysis of gateroad stability in relation to
yield pillar size: a case study. Rock Mech. Rock Eng., 2017, 50,
1263–1278.
5. Likar, J. et al., Analysis of geomechanical changes in hanging
wall caused by longwall multi top caving in coal mining. J. Min.
Sci., 2012, 48, 135–145.
6. Bhaskaran, S. et al., Experimental studies on spalling
characteristics of Indian lignite coal in context of underground
coal gasification. Fuel, 2015, 154, 326–337.
7. Bai, Q. S. et al., Numerical modeling on brittle failure of coal wall
in longwall face – a case study. Arab. J. Geosci., 2014, 7, 5067–
Figure 15. Assessment results based on artificial immune algorithm. 5080.

CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019 2011

RESEARCH ARTICLES
8. Wang, J. C., Mechanism of the rib spalling and the controlling in International Conference on Intelligent Data Acquisition and
the very soft coal seam. J. China Coal Soc., 2007, 32, 785–788. Advanced Computing Systems: Technology and Applications,
9. Fang, X. Q., He, J. and Li, H. C., A study of the rib fall Berlin, Germany, 2013, vol. 12, pp. 425–428.
mechanism in soft coal and its control at a full-mechanized top- 25. Lin, H. H., Chuang, J. H. and Liu, T. L., Regularized background
coal caving face. J. China Univ. Min. Technol., 2009, 38, 641– adaptation: a novel learning rate control scheme for Gaussian
645. mixture modeling. IEEE Trans. Image Process., 2011, 20, 822–
10. Zhang, Y. L. et al., Effect analysis of rib spalling prevention 836.
system in hydraulic support. J. China Coal Soc., 2011, 36, 691– 26. Horprasert, T., Harwood, D. and Davis, L. S., A statistical
695. approach for real-time robust background subtraction and shadow
11. Zhang, H. L., Wang, L. G. and Qin, H., Study of spalling detection. In Proceedings of IEEE ICCV'99 Frame-Rate
mechanism and control techniques of mining roadway. Rock Soil Workshop, 1999, pp. 1–19.
Mech., 2012, 33, 1462–1466. 27. KaewTraKulPong, P. and Bowden, R., An improved adaptive
12. Zhang, S. Q., Technology research of excavating tunnel temporary background mixture model for real-time tracking with shadow
support and prevent sloughing. Saf. Coal Mines, 2013, 22, 19–20. detection. In Proceedings of the 2nd European Workshop on
13. Liu, G. F. et al., Failure characteristics, laws and mechanisms of Advanced Video Based Surveillance Systems, Kingston upon
rock spalling in excavation of large-scale underground Thames, UK, AVBS01, September 2001, pp. 1–5.
powerhouse caverns in Baihetan. Chin. J. Rock Mech. Eng., 2016, 28. An, J., Zhang, G. C. and Liu, Y. N., Application of mathematical
35, 865–878. morphology in processing of medical eye images. Mech. Eng.
14. Jiang, Q. et al., Evaluation method of general geostress based on Autom., 2016, 1, 15–17.
spalling features of wall rock. Rock Soil Mech., 2011, 32, 1452– 29. Chen, H. T., Lin, H. H. and Liu, T. L., Multi-object tracking using
1459. dynamical graph matching. In Proceedings of the 2001 IEEE
15. Jeyabharathi, D. and Dejey, D., Efficient background subtraction Computer Society Conference on Computer Vision and Pattern
for thermal images using reflectional symmetry pattern (RSP). Recognition, Madison, Wisconsin, 2003, p. 210.
Multimedia Tools Appl., 2018, 77, 22567–22586. 30. Vapnik, V. N., The nature of statistical learning theory. IEEE
16. Sugimura, D., Teshima, F. and Hamamoto, T., Online background Trans. Neural Network, 1995, 8, 988–999.
subtraction with freely moving cameras using different motion 31. Karimi, M., A new approach to history matching based on feature
boundaries. Image Vision Comput., 2018, 76, 76–92. selection and optimized least square support vector machine.
17. Liu, X. et al., Background subtraction using spatio-temporal group J. Geophys. Eng., 2018, 15, 2378–2387.
sparsity recovery. IEEE T. Circuits Syst. Video Technol., 2018, 28, 32. Arenas, M. P. et al., Novel austenitic steel ageing classification
1737–1751. method using eddy current testing and a support vector machine.
18. Goyal, K. and Singhai, J., Review of background subtraction Measurement, 2018, 127, 98–103.
methods using Gaussian mixture model for video surveillance 33. Chen, Y. G., Prediction algorithm of PM2.5 mass concentration
systems. Artif. Intell. Rev., 2018, 50, 241–259. based on adaptive BP neural network. Computing, 2018, 100,
19. Wu, J. D. and Tsai, Y. J., Speaker identification system using 825–838.
empirical mode decomposition and an artificial neural network. 34. Sadeghi, M., Maghooli, K. and Moin, M. S., Using artificial
Exp. Syst. Appl., 2011, 38, 6112–6117. immunity network for face verification. Int. Arab. J. Inf. Technol.,
20. Polat, K. and Güneş, S., An expert system approach based on 2014, 11, 354–361.
principal component analysis and adaptive neuro-fuzzy inference
system to diagnosis of diabetes disease. Digit. Signal Process.,
2007, 17, 702–710. ACKNOWLEDGEMENTS. The supports of the Fundamental Re-
21. Si, L. et al., Identification of shearer cutting patterns using search Funds for the Central Universities (No. 2017QNA12) and the
vibration signals based on a least squares support vector machine Priority Academic Program Development (PAPD) of Jiangsu Higher
with an improved fruit fly optimization algorithm. Sensors, 2016, Education Institutions in carrying out this research are acknowledged.
16, 90. We thank the reviewers for their useful suggestions on the original
22. Ukil, A., Support vector machine. Comput. Sci., 2002, 1, 1–28. manuscript.
23. Shahrizat, S. M. and Nooritawati, M. T., Background modelling
and background subtraction performance for object detection. Pro-
ceedings of the 6th International Colloquium on Signal Processing Received 22 October 2018; revised accepted 9 February 2019
and its Applications, Melaka, Malaysia, 2010, pp. 236–241.
24. Sadykhov, R. K. and Kuchuk, S., Background substraction in
grayscale images algorithm. In Proceedings of the 7th IEEE doi: 10.18520/cs/v116/i12/2000-2012

2012 CURRENT SCIENCE, VOL. 116, NO. 12, 25 JUNE 2019