Engineering Failure Analysis 101 (2019) 407–417

Contents lists available at ScienceDirect

Engineering Failure Analysis

journal homepage: www.elsevier.com/locate/engfailanal

Failure analysis of the rubber-textile conveyor belts using

T
classification models
Miriam Andrejiovaa, Anna Grincovab, , Daniela Marasovac
⁎

a
Faculty of Mechanical Engineering, Technical University of Kosice, Letna 9, Kosice 042 00, Slovak Republic
b
Faculty of Electrical Engineering and Informatics, Technical University of Kosice, Letna 9, Kosice 042 00, Slovak Republic
c
Faculty of Mining, Ecology, Process Control and Geotechnology, Technical University of Kosice, Park Komenskeho 14, Kosice 042 00, Slovak Republic

ARTICLE INFO ABSTRACT

Keywords: In real operations, the requirements on conveyor belts are very high and are constantly in-
Rubber-textile conveyor belt creasing. They mainly regard high wear resistance and resistance to the effects of the transported
Damage material, long service life, high longitudinal strength, low weight, etc.
Classification model The main purpose of the present paper is the classification of the damage to rubber-textile
Decision tree
conveyor belts that occurs due to the impact of the transported material (at the impact of the
Regression analysis
material onto a conveyor belt at the chute). The experimental research was focused on the three
types of conveyor belts of various internal structures, on the impact height, and the type of the
material falling onto a conveyor belt and effects thereof on the incurred damage. The damage
classification was carried out using two different models. The models were created using the
classification tree and the multiple regression analysis. The models were created using the
training set of 48 conveyor belt samples. The accuracy and predictive ability of the models were
verified using the testing set of 16 conveyor belt samples.

1. Introduction

A conveyor belt is a carrying and traction element of a belt conveyor. From the structural point of view, we distinguish between
two types of conveyor belts, the rubber-textile and the steel-cord conveyor belts. Damage to or defects of conveyor belts of both types
are discussed by many authors who apply various methods of the damage degree identification. Such research is carried out in
laboratory conditions but also in a real operation. Bajda assumes that the resistance of conveyor belts to puncture and to lengthwise
slits belongs to the most important criteria applied to the service life evaluation. In his papers [1,2], he presents the results of
laboratory tests facilitating the definition of the impact of the conveyor belt structure on the belt wear, in particular on the puncture.
In the same laboratory, Bajda [3] also uses a magnetic diagnostic equipment to evaluate the condition of steel cords in the steel-cord
conveyor belts without the need to remove the belt cover layers. Fedorko describes the laboratory method of analysing the failures of
rubber-textile conveyor belts subjected to the tensile force [4]. Blazej [5], however, describes a diagnostic device used directly in the
operation. The device facilitates the objective magnetic evaluation of the belt core condition, ensuring thus a higher service life and a
lower proportion of scrap belts that must be renovated. Qiao [6] introduces, within the online monitoring, a novel method of the
Integrative Binocular Vision Detection (IBVD) intended for the detection of longitudinal tear of conveyor belts. Based on the infrared
and visible fusion, the IBVD sensor device collects the fusion images of the belt for the period of time shorter than 18 ms. In article
[7], Chen describes new types of the Support Vector Data Description (SVDD) method used to detect the image-based conveyor belt

⁎
Corresponding author.
E-mail address: anna.grincova@tuke.sk (A. Grincova).

https://doi.org/10.1016/j.engfailanal.2019.04.001
Received 7 June 2018; Received in revised form 26 March 2019; Accepted 1 April 2019
Available online 01 April 2019
1350-6307/ © 2019 Elsevier Ltd. All rights reserved.
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

faults.
A smooth operation of belt conveyors is also influenced by the problem of the belt splice control [8,9]. Blazej [10] examine,
within his testing, the load and elongation of adhesive bond in splices. The test results facilitate the optimal selection of adhesive
mixtures used to join the rubber-textile belts. The cover layers of conveyor belts are also made of the adhesive mixture - rubber.
Hakami [11] examines the effects of the load, advancing speed, hardness, and friction on the rubber wear rate. In paper [12],
Mazurkiewicz presents the expert system intended for estimating a technical condition of the conveyor belt splice condition within
his efforts aimed at increasing the service life and reliability of the entire conveyance system.
A site where the material falls onto the belt is critical; it is the place where the belt damage largely occurs. The analysis of the belt
conveyor failures related to such impact site, where the material is located, is described in articles [13,14]. The analysis is based on
the mathematical modelling of a conveyor belt. Komander [15] presents the results of the research of the effects of the rubber cover
thickness, rubber cover strength, and the type and tensile strength of the crosswise belt reinforcements (breakers) on the belt
puncture resistance by means of two coefficients: the Critical Energy and the Average Impact Energy.
For the purpose of adjusting and monitoring the behaviour of the internal structure of a conveyor belt loaded with the material,
some authors use the metrotomography analysis [16]. In his paper [17], Fedorko uses the industrial tomography to examine the belt
carcass failures while identifying the distances between the individual fibres in the warp and the weft of the carcass in the laboratory
samples and to analyse thus the observed defects. Bocko et al. [18] test the mechanical properties of the rubber conveyor belts
through the impact load. The internal damage to conveyor belts was evaluated using the computer tomography.
The creation of regression models aimed at the identification of the impact forces at the site where the material impacts the belt is
used in papers [19–21]. A different approach to the evaluation of the damage severity, using the logistic regression, is used in the
papers [22,23]. Ambrisko et al. [24] used the DOE method to evaluate the effects of the selected factors on the conveyor belt wear.
The efficient data utilisation is facilitated by the data mining techniques. According to [25,26], the data mining is an analytical
methodology of obtaining non-trivial hidden and potentially useful information from the data. The data mining techniques also
include the classification trees. According to Rokach and Maimon [27], the classification trees are used to classify an object into a
predefined set of classes based on their attribute values. The decision trees and the decision rules are applied in many areas as a
powerful solution to classification problems [26].
Ma et al. [28] performed the classification of the damage to the steel-cord conveyor belts using the algorithm combined with the
information entropy and the multi-class SVM (Support Vector Machine), based on the binary tree. The decision tree and the relative
risk were also used in the paper [29].
Khan et al. [30] compare the classification algorithms (e.g., Naïve Bayes Classifier, Decision Tree, Multiclass-Classifier, etc.) for
the purpose of damage classification in smart composite laminates.
Patel et al. [31] describe the ore classification model on the basis of the machine vision using the SVM algorithm. The SVM model
was compared with other methods, such as the K-Nearest Neighbour, Classification Discriminant Analysis, Naïve Bayes, Classification
Tree, and the Probabilistic Neural Network. Rahmadani et al. [32] compare the classification models created using the Naïve Bayes
Classifier and the decision tree in a set of data obtained using the genetic algorithm. Malinowska [33] applies the classification and
regression trees to the assessment of the building damage caused by the surface deformation. Muralidharan et al. [34] describe the
condition monitoring of the self-aligning carrying idler (SAI) in a belt conveyor using the statistical functions and the decision tree
algorithm.
The present paper deals with the analysis and the classification of the conveyor belt damage at the impact load of a conveyor belt.
The damage classification model will be created using the classification tree and the regression analysis.

2. Material and methods

2.1. Problem formulation

The conveyor belt resistance to any wear and damage that eventually shorten the belt operational service life is very important for
the practice. The experimental research should therefore be focused on the monitoring and analysis of the conveyor belt damage with
different input parameters. The factors significantly affecting the belt operating service life include mainly the structural parameters
and the properties of the conveyor belt itself, the type of the transported material, and the chute design. In the laboratory conditions
it is not possible to monitor all the factors, that is why we chose only the selected factors (conveyor belt types with various internal
structures, types of the falling material, and impact height). The objective of the experimental research was to monitor the occurrence
of conveyor belt damage, depending on the above listed selected factors.

2.2. Experiment execution

The experimental measurements were carried out using the testing device (Fig. 1) intended for the simulation of the material
impact onto a conveyor belt at the chute and for the related monitoring of the puncture resistance of conveyor belts. Within the
experiment, the tested conveyor belt samples (sized 1.4 × 0.2 m) were fixed in the hydraulically controlled jaws (4) and stretched to
achieve the operating tension recommended by the manufacturer (in this case 1/10 of the belt strength). A drop hammer with an
impactor (3) (sphere, pyramid) is dropped in a free fall from the determined impact heights (maximum range of 0–2.6 m) onto the
belt sample. The simulation of the impact of materials of various weights is ensured using the additional weights (2) (weight range of
50–100 kg).

408
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Fig. 1. Testing device scheme (1 – skeleton, 2 –drop hammer with a weight, 3 – impactor, 4 – hydraulic jaws).

Within our experiment, the damage type and the magnitude of the impact force at the impact of the material of two different types
were monitored. The impact of the brittle incohesive material was simulated by a spherical impactor and the impact of the sharp-
edged material was simulated by a pyramidal impactor (Fig. 2). Based on the experience from the previous research, various impact
height ranges were chosen for individual impactors. For the spherical impactor, the height range was from 1.2 to 2.6 m with 0.2 m
increments; for the pyramidal impactor, the height range was from 0.8 to 2.2 m with 0.2 m increments. The previous experimental
research indicated that no visible damage occurs below the minimum determined value and complete damage to the conveyor belt
samples incurs above the maximum determined value. The total weight of the drop hammer was determined as 90 kg in both cases.
The testing was carried out using three different types of conveyor belts with different carcass structures (Table 1).

2.3. Conveyor belt type

Rubber-textile conveyor belts (Fig. 3) offer a wide range of applications across all industries. They are used in the transportation
of bulk and piece materials in various operating conditions. The carcass consists of several textile plies coated with the layer of a
special rubber mixture providing the adhesive properties between the plies. In basically consists of polyamide fibres P or polyester

Fig. 2. Impactor (a detail the spherical impactor and the pyramidal impactor attachment).

409
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Table 1
Conveyor belts used in the experiment.
Conveyor belt type Designation Characteristics

EP2000/4 CB1 Conveyor belt with the strength of 2000 N.m−1 and four plies (polyester in the warp and polyamide in the weft)
P2000/4 CB2 Conveyor belt with the strength of 2000 N.m−1 and four plies (polyamide)
P2000/4 + breakers CB3 Conveyor belt with the strength of 2000 N.m−1 and four plies (polyamide) with breakers

Fig. 3. Rubber-textile conveyor belt.

Fig. 4. Some examples of the caused damaged.

fibres E. Polyamide (P) is a fully synthetic fibre resistant to humidity and chemicals. Polyester (E) is a fully synthetic fibre with the
high strength and good tensile properties. It is resistant to acids, bases, and humidity. It is frequently processed together with
polyamide. The carcass of such conveyor belt consists of polyester fibres in the warp and polyamide fibres in the transversal direction
(in the weft, EP).
Cover layers facilitate the contact between the belt and the transported material, or between the conveyor rollers and drums. They
protect the carcass against the mechanical damage caused by the transported material, against the effects of humidity, and chemical
and thermal impacts that affect the conveyor belt. Their thickness depends on the properties of the transported material. Some
conveyor belts have the carcass equipped with breakers to increase their puncture resistance.
The research experiment was carried out using the samples of three rubber-textile conveyor belts of the same strengths but with
different internal carcass structures. The used conveyor belt types are listed in Table 1.

2.4. Damage classification

The tested conveyor belt samples are divided, in terms of the resulting damage degree, into 4 basic groups (Table 2). A detailed
damage classification is presented in article [23].

Table 2
Sample classification by the damage degree.
Damage 0 Damage 1 Damage 2 Damage 3

Unserious damage Serious damage

No visible damage Visible damage to the top or the bottom cover Visible damage to the top or the Puncture (concurrent complete damage to the
layers (cracks, incisions) without any visible bottom cover layers and visible top cover layer, the carcass, and the bottom
damage to the carcass damage to the carcass cover layer)

Damage 0 and Damage 1 (unserious damage) are defined by the damage degree that does not cause the instant belt decommissioning when the
conveyor belt is used in the real operation. On the other hand, Damage 2 and Damage 3 (serious damage) are defined as the damage causing the
conveyor belt decommissioning. Examples of the possible damage to the test pieces are presented in Fig. 4.

410
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Fig. 5. Research procedure and used evaluation methods.

3. Theory/calculation

3.1. Procedure and methods of the experimental research

For the purpose of the assessment of the dependence between the two categorical variables, the Pearson's Chi-Square Test of
Independence or the Fisher's Exact Test were used.
The damage was classified using the classification tree and the multiple regression analysis. The purpose of the classification tree
is to create a classification rule (a classification model), on the basis of the training set of samples, to be applied to the decision
making, based on the selected variables, on which classification damage degrees will be assigned to individual conveyor belt samples.
Another option of the damage classification is also the classification model obtained by the multiple regression analysis. The pre-
dictive ability of both classification models (using the classification tree and using the multiple regressions) is verified using the
testing set of conveyor belt samples.
The course and the methods of the experimental research evaluation are presented in Fig. 5.

3.2. Regression model

The underlying concept of the regression analysis theory is the determination of the relationship between the dependent variable
Y and k-independent variables Xi, i = 1, ⋯, k. The regression model may be written as follows
k
Y= 0 + i Xi +
i=1 (1)
where β0 and βi for =1, ⋯, k are the regression model parameters and ε is the random error.
The statistical significance of the model is verified using the F-test of the model's statistical significance. The statistical sig-
nificance of individual regression model parameters is verified using the test of statistical significance of the regression parameter. To
verify the quality of the regression model, the coefficient of determination R2 is used.

3.3. Classification tree and creation thereof

Decision trees belong to the fundamental principles of the symbolic methods of machinery learning. They represent strong tools
used for the purpose of classification and prediction. Their attractiveness causes that in comparison with the neural networks the
decision trees represent the classification rules. The decision trees may also be defined as a structure used for the prediction of the
target attribute while applying simple decision-making rules.
The application of the classification trees is limited to the discrete data classification. In the case of continuous variables, they
must be converted into discrete variables. In order to perform the first step of the algorithm, i.e., the root selection and splitting, a
more appropriate attribute must be found for such splitting. This is carried out applying various criteria, depending on the character
of particular attributes. The decision-making on an appropriate attribute for the splitting is most frequently carried out using the
following criteria: Entropy, Informational Gain and Informational Gain Ratio, Gini Index, and Chi-Square Test. At present, there are
several algorithms that search, on the basis of a particular criterion, the most appropriate attribute for the splitting. In the paper, we
will use the C5.0 algorithm that is suitable for all types of attributes; however, the output variable is categorical. A splitting criterion
is the Information Gain and Entropy.
Let S be a node with n training data. Let Y be the target attribute that may acquire m different values, of classes Yj (j = 1, 2, ⋯, m).
Then the formula for the calculation of the entropy of node S is as follows
m
H (S ) = p (Yj ) log2 p (Yj ),
j=1 (2)

411
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

where p(Yj) is the probability that any instance belonging to the node S will be classified as the class Yj.
Let A be the attribute that may acquire the values Ai (i = 1, 2, ⋯, k). Then the formula for the calculation of entropy of node S
with the testing attribute A will be as follows
k
H (S , A ) = p (Ai ) H (Ai ),
i=1 (3)
where p(Ai) is the probability that any instance belonging to the node S is of the value of the attribute A = Ai while H(Ai) represents
the value of entropy in the group of cases that have the value of attribute A = Ai. It applies that
m
H (Ai ) = p (Ail ) log2 p (Ail ),
j=1 (4)
where
nl (Ai )
p (Ail ) = ,
mi (5)
whereas nl(Ai) represents the number of training data belonging to the class of the target attribute Yl with the value of attribute
A = Ai.
The information gain IG(S, A) of the attribute A represents the expected entropy decrease caused by dividing the cases on the basis
of the values of the given variable. The information gain compares the entropies before and after the division. It measures how much
information we obtained by the division depending on the selected attribute. The information gain for the node S and the attribute A
can be summed using the formula
IG (S, A) = H (S ) H (S , A). (6)
To fulfil the goal of creating a minimum decision tree, for each node the splitting is carried out while choosing the attribute with
the lowest entropy, i.e., with the maximum possible information gain.

3.4. Classification tree accuracy

The overall accuracy A of the classification of the output variable Y according to the classification model and the real division of
the cases into classes is calculated using the formula
C E
A= ,
C (7)
where E is the number of incorrectly classified cases and C is the total number of cases in the set.
The evaluation of the classification accuracy is carried out using the Cohen's Kappa κ.
po pe
=
1 pe (8)
where po is the probability of the classification success (accuracy A) and pe is the probability of success due the chance.
To assess the agreement of the classification using the Cohen's Kappa, we will use the following scale: < 0% is no agreement,
0–20% is poor, 20–40% is fair, 40–60% is moderate, 60–80% is good, 80% or higher is very good, excellent agreement.

4. Result and discussion

The experimental research was carried out with the objective to:

• analyse the input variables, the occurrence of damage, and the impact of the selected variables on the damage degree,
• create a classification model using the training set, applying the decision tree, and verify it with the testing set of measurements,
• create a classification model using the training set, applying the regression analysis, and verify it with the testing set of mea-
surements.

4.1. Characteristics of variables

Within the analysis and evaluation of the experimental research, we monitor the following variables: Drop Hammer Impactor
(IMPACTOR), Impact Height (HEIGHT), Conveyor Belt Type (CB), and Conveyor Belt Damage Degree (DAMAGE).
The IMPACTOR variable is characterised by the drop hammer impactor and has two categories (sphere, pyramid). The pyramidal
impactor simulates the impact of the sharp-edged material. The impact of the brittle incohesive material is simulated by the spherical
impactor.
The HEIGHT continuous variable (m) represents the impact height from which the drop hammer with the impactor is dropped
onto the conveyor belt sample.

412
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Table 3
Summary of variables.
Variables Description

Dependent variables
DAMAGE Damage type: Damage 0, Damage 1, Damage 2, Damage 3

Independent variables
HEIGHT Impact height (meter)
IMPACTOR Impactor type: sphere, pyramid
CB Conveyor belt type: CB1, CB2, CB3 (Table 1)

Input variables (HEIGHT, IMPACTOR, CB) as well as the DAMAGE output variable will be
regarded as the categorical variables.

The CB variable is characterised by the conveyor belt type and has three categories (CB1, CB2, CB3). In all the cases, a rubber-
textile conveyor belt was used (Table 1).
The damage degree identification consists in the visual inspection of the conveyor belt after the test is completed. The DAMAGE
output variable represents the severity of the damage to the composite at the site where the material falls onto the belt and has four
categories. The variable characteristics are listed in Table 3.
The experimental research was carried out using 48 conveyor belt samples forming the training set. The spherical impactor was
used in 24 experiments (50%) and the pyramidal impactor in 24 experiments (50%). The samples were extracted from the conveyor
belts of three types. For each type, the same numbers of samples were used, i.e., 16 samples. The impact height was ranging from
0.8 m to 2.6 m with the constant increment of 0.2 m. For the analysis purposes, this variable was divided into three classes: the impact
height below 1.4 m (Height 1), from 1.6 m to 2.0 m (Height 2), and above 2.2 m (Height 3). In each interval, the tests were carried out
in equal amounts.
Out of the total number of the tested samples, 10 samples were categorised as Damage 0 (20.8%) and 17 samples as Damage 1
(35.5%). Damage 2 (20.8%) was identified in 10 samples and Damage 3 (22.9%) in 11 samples.
The statistically significant impact of the falling impactor on the conveyor belt damage degree was confirmed by the Pearson's
Chi-Square Test of Independence (p-value < 0.001). The Pearson's Chi-Square Test of Independence also confirmed the statistically
significant impact of the impact height on the conveyor belt damage degree (p-value = 0.002). In both cases, the association is
moderate, Cp = 0.54. On the other hand, no statistically significant impact of the conveyor belt type on the damage degree was
confirmed (p-value = 0.938). The occurrence of different degrees of the damage to conveyor belts, depending on the conveyor belt
type, is shown in Fig. 6.
With the spherical impactor, no damage occurred at the impact heights below 1.6 m. At the impact heights of 1.8 m and higher,
the cases of damage (bulge) to the bottom cover layer were observed, without any damage to the top cover layers. In the case of the
CB1 conveyor belt, the complete damage to the top or the bottom cover layers was observed at the impact height of 2.4 m. In the case
of two other types (CB2, CB3), no damage to the carcass and no puncture were observed even at the maximum impact height of 2.6 m.

Fig. 6. Occurrence of damage (conveyor belt type).

413
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Fig. 7. Classification tree.

4.2. Classification model using the decision tree

The classification model of the conveyor belt damage is created using the training set, applying the decision tree. The training set
consists of 48 conveyor belt samples. The classification criterion is the DAMAGE variable with four classification classes: Damage 0,
(D0), Damage 1 (D1), Damage 2 (D2), and Damage 3 (D3). The result of the decision tree is presented in Fig. 7.
All three input variables were used for some of the tree levels. The first splitting occurs depending on the type of the falling drop
hammer. The next splitting of both branches depends on the impact height.
With the spherical impactor, another splitting criterion is the impact height. The conveyor belt damage is not visible (Damage 0)
at the impact heights less than or equal to 1.6 m. Serious damage occurs at the impact heights above 2.2 m (regardless of the conveyor
belt type).
With the pyramidal impactor, serious damage (Damage 2, Damage 3) occurs at the impact heights above 1.4 m, regardless of the
conveyor belt type. The situation is only better (Damage 1) when the impact height is reduced below 1.4 m.
The decision tree may be converted into the form of the decision-making rules. Each route from the root to the leaf of the tree
corresponds to one rule. The rules in our resulting decision tree are as follows:

IF (IMPACTOR=Sphere) AND (HEIGHT < =1.6m) THEN (DAMAGE=DAMAGE 0)

IF (IMPACTOR=Sphere) AND 1.6 m < =HEIGHT < =2.2m) THEN (DAMAGE=DAMAGE 1)
IF (IMPACTOR=Sphere) AND (HEIGHT > =2.2m) THEN (DAMAGE=DAMAGE 2)
IF (IMPACTOR=Pyramid) AND (HEIGHT < =1.4m) THEN (DAMAGE=DAMAGE 1)
IF (IMPACTOR=Pyramid) AND (HEIGHT > =1.4m) AND (CB=CB1) THEN (DAMAGE=DAMAGE 3)
IF (IMPACTOR=Pyramid) AND (HEIGHT > =1.4m) AND (CB=CB2) THEN (DAMAGE=DAMAGE 3)
IF (IMPACTOR=Pyramid) AND (1.4 m < =HEIGHT < =1.8m) AND (CB=CB3) THEN
(DAMAGE=DAMAGE 2)
IF (IMPACTOR=Pyramid) AND (HEIGHT > =1.8m) AND (CB=CB3) THEN (DAMAGE=DAMAGE 3)

The description of the agreement between the classification obtained from the experiment and the classification determined by
the decision tree, we will use the confusion matrix (Table 4). The comparison indicates that the classification model determined by
the decision tree incorrectly classified one sample as Damage 1 instead of Damage 0; three conveyor belt samples were classified as
Damage 1 instead of Damage 2; and one sample was classified as Damage 2 instead of Damage 3.

4.3. Classification model using the regression model

The monitoring of the effects of the selected independent variables on the damage degree will be based on the model:

DAMAGE = f (HEIGHT , IMPACTOR, CB ).

The regression model equation is as follows:

414
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Table 4
Confusion matrix (training set).
Observed damage Classification determined by the decision tree

Damage 0 Damage 1 Damage 2 Damage 3

Damage 0 9 1 0 0
Damage 1 0 17 0 0
Damage 2 0 3 7 0
Damage 3 0 0 1 10

The confusion matrix indicates that out of the total number of 48 samples, 43 were classified correctly. The overall accuracy, determining the
probability that the conveyor belt samples are classified correctly, is 89.6%. The Cohen's Kappa is κ = 0.86.

Table 5
Confusion matrix (testing set).
Observed damage Classification determined by the decision tree

Damage 0 Damage 1 Damage 2 Damage 3

Damage 0 1 0 0 0
Damage 1 1 3 0 0
Damage 2 0 0 7 1
Damage 3 0 0 1 2

The model's predictive ability was verified using the testing set of 16 conveyor belt samples. The comparison of the damage degrees identified by the
visual inspection and using the classification model indicates that out of 16 samples, 13 were classified successfully (Table 5). The overall accuracy
of the model is 81.5% and the Cohen's Kappa is κ = 0.716.
The aforesaid indicates that the obtained model has good classification ability and good prediction ability.

DAMAGE = 0 + 1 X1 + 2 X2 + 3 X3 + , (9)
where β0, βi, i = 1, ⋯3 are the model parameters, X1 is the impact height (HEIGHT, m), X2 is the drop hammer impactor type
(IMPACTOR), and X3 is the conveyor belt type (CB). The DAMAGE, IMPACTOR, and CB variables are categorical variables with four,
two, or three categories. If the damage degree is determined as 0 (or 1, 2, 3, respectively), then DAMAGE = 0 (or DAMAGE = 1,
DAMAGE = 2, DAMAGE = 3, respectively). With the pyramidal impactor X2 = 0 and with the spherical impactor X2 = 1. If we
consider the conveyor belt type CB1 (or CB2, CB3, respectively), then X3 = 1 (or X3 = 2, X3 = 3, respectively).
The point estimate of the regression model parameters is
DAMAGE = 0.333 + 1.806X1 1.806X2 0.188X3 . (10)

All the regression model parameters, as well as the regression model, are statistically significant (p-value < alpha). The value of
the multiple correlation coefficient is 0.891.
The resulting values of the DAMAGE dependent variable, with regard to the discrete nature thereof, will be adjusted by rounding
it to the nearest integer. The comparison of the damage results obtained from the regression model and the classification obtained
from the experiment indicates that the regression model incorrectly classified 7 samples out of the total number of 48 samples
(Table 6). The overall accuracy of the classification model is 85.4% and the Cohen's Kappa is 0.802.

5. Conclusion

The operational service life of conveyor belts is significantly influenced by the belt wear and damage. In order to assess conveyor
belts correctly, in terms of the belt wear and damage, it is necessary to know the patterns occurring within the damage process.

Table 6
Confusion matrix (training set).
Observed damage Classification determined by the regression model

Damage 0 Damage 1 Damage 2 Damage 3

Damage 0 8 2 0 0
Damage 1 0 15 2 0
Damage 2 0 0 10 0
Damage 3 0 0 3 8

The model's predictive ability was again verified using the testing set of 16 conveyor belt samples. The comparison of the damage degrees observed
by the visual inspection and by the classification model indicates that 13 out of 16 samples were classified successfully (Table 7). The overall
accuracy of the model is 81.5% and the Cohen's Kappa is 0.716.

415
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

Table 7
Confusion matrix (testing set).
Observed damage Classification determined by the regression model

Damage 0 Damage 1 Damage 2 Damage 3

Damage 0 1 0 0 0
Damage 1 0 4 0 0
Damage 2 0 2 6 0
Damage 3 0 0 1 2

The aforesaid indicates that the classification and the prediction ability of this model is good too.

Within the experimental research, three input variables were assessed: the conveyor belt type (the same strengths but different
carcass structures), impact height, and impactor type. The resulting finding of the evaluation of the impact of these variables on the
conveyor belt damage degree, using the Pearson's Chi-Square Test of Independence, was the statistically significant effects of two
variables: the impactor type and the impact height. The experimental research also provided the following information:

• In the case of the pyramidal impactor simulating the impact of the sharp-edged material, some damage occurs already at the
heights below 1.4 m. At the impact heights above 1.4 m, serious damage occurs (Damage 2, Damage 3).
• In the case of the spherical impactor, serious damage occurs at the impact heights above 2.2 m.
• The impact height has a significant impact on the occurrence of serious damage to a rubber-textile belt in the case of the drop
hammer with the pyramidal impactor (impact of the sharp-edged material) as well as in the case of the drop hammer with the
spherical impactor (impact of the brittle material).
• The carcass structure has a significant effect on the occurrence of serious damage to conveyor belts in the case of the drop hammer
with the pyramidal impactor at the impact heights above 1.4 m.
• The CB3 conveyor belt (P2000/4 + breakers), as compared to CB1 (EP 2000/4) and CB2 (P2000/4) belts, has a little better
properties, in terms of the resistance to serious damage.

The comparison of the classification models indicates that both created models have good classification and prediction properties.
Apparently, the classification model created using the classification tree is a little better than the regression model. One of the reasons
is the classification output that is more detailed and better arranged for the practical applications thereof.

Acknowledgements

This article is the result of the Project VEGA 1/0577/17 and the Project KEGA 009TUKE-4/2016.

References

[1] M. Bajda, Comparison of puncture resistance test results for different conveyor belt designs, International Multidisciplinary Scientific GeoConference Surveying
Geology and Mining Ecology Management, vol. 17, 2017, pp. 125–132.
[2] M. Bajda, Laboratory tests of conveyor belt parameters affecting its lifetime, International Multidisciplinary Scientific GeoConference Surveying Geology and
Mining Ecology Management, vol. 17, 2017, pp. 495–502.
[3] M. Bajda, R. Blazej, L. Jurdziak, A new tool in belts resistance to puncture research, Mining Sci. 23 (2016) 173–182.
[4] G. Fedorko, V. Molnar, P. Michalik, M. Dovica, T. Toth, T. Kelemenova, Extension of inner structures of textile rubber conveyor belt – failure analysis, Eng. Fail.
Anal. 70 (2016) 22–30.
[5] R. Blazej, L. Jurdziak, Condition-based Conveyor Belt replacement strategy in lignite mines with random belt deterioration, IOP Conf. Ser. 95 (2017) 042051.
[6] T. Qiao, L. Chen, Y. Pang, G. Yan, C. Miao, Integrative binocular vision detection method based on infrared and visible light fusion for conveyor belts longitudinal
tear, Measurement 110 (2017) 192–201.
[7] G. Chen, X. Zhang, Z.J. Wang, F. Li, Robust support vector data description for outlier detection with noise or uncertain data, Knowl.-Based Syst. 90 (2015)
129–137.
[8] A. Temerzhanov, I. Stolpovskikh, A. Sładkowski, Analysis of reliability parameters of conveyor belt joints, Transp. Prob. 7 (2012) 107–112.
[9] H. Yang, G. Xu, X. Fan, A reliability analysis method of cloud theory – Monte Carlo based on performance degradation data, Maint. Reliab. 17 (2015) 435–442.
[10] R. Blazej, M. Bajda, M. Hardygóra, Monitoring creep and stress relaxation in splices on multiply textile rubber conveyor belts, Acta Montan. Slovaca 22 (2017)
116–125.
[11] F. Hakami, A. Pramanik, N. Ridgway, A.K. Basak, Developments of rubber material wear in conveyer belt system, Tribol. Int. 111 (2017) 148–158.
[12] D. Mazurkiewicz, Maintenance of belt conveyors using an expert system based on fuzzy logic, Arch. Civil Mech. Eng. 15 (2015) 412–418.
[13] S. Honus, P. Bocko, T. Bouda, I. Ristovic, M. Vulic, The effect of the number of conveyor belt carrying idlers on the failure of an impact place: a failure analysis,
Eng. Fail. Anal. 77 (2017) 93–101.
[14] H. Gondek, J. Neruda, J. Pokorný, The dynamics of impacts tools the loading boom bucket wheel excavators, Appl. Mech. Mater. 683 (2014) 213–218.
[15] H. Komander, M. Bajda, G. Komander, G. Paszkowska, Effect of strength parameters and the structure of steel cord onveyor belts on belt puncture resistance,
Appl. Mech. Mater. 683 (2014) 119–124.
[16] G. Fedorko, V. Molnar, P. Michalik, M. Dovica, T. Kelemenova, T. Toth, Failure analysis of conveyor belt samples under tensile load, J. Ind. Text. 48 (2019)
1364–1383.
[17] G. Fedorko, V. Molnar, M. Dovica, T. Toth, J. Fabianova, J. Strohmandl, H. Neradilova, M. Belusko, Analysis of defects in carcass of rubber–textile conveyor belts
using metrotomography, J. Ind. Text. 47 (2018) 1812–1829.
[18] P. Bocko, O. Marada, T. Bouda, Material and Component Testing of Conveyor Belts and its Numerical Analyses, Modern Methods of Construction Design, vol. 16,
(2014), pp. 565–571.
[19] G. Fedorko, V. Molnar, D. Marasova, A. Grincova, M. Dovica, J. Zivcak, T. Toth, N. Mikusova, Failure analysis of belt conveyor damage caused by the falling

416
M. Andrejiova, et al. Engineering Failure Analysis 101 (2019) 407–417

material Part 1: experimental measurements and regression models, Eng. Fail. Anal. 36 (2014) 30–38.
[20] A. Grincova, D. Marasova, Experimental research and mathematical modelling as an effective tool of assessing failure of conveyor belts, Maint. Reliab. 16 (2014)
229–235.
[21] A. Grincova, M. Andrejiova, D. Marasova, Measuring and comparative analysis of the interaction between the dynamic impact loading of the conveyor belt and
the supporting system, Measurement 59 (2015) 184–191.
[22] M. Andrejiova, A. Grincova, D. Marasova, Failure analysis of rubber composites under dynamic impact loading by logistic regression, Eng. Fail. Anal. 84 (2018)
311–319.
[23] M. Andrejiova, A. Grincova, D. Marasova, G. Fedorko, V. Molnar, Using logistic regression in tracing the significance of rubber-textile conveyor belt damage,
Wear 318 (2014) 145–152.
[24] L. Ambrisko, D. Marasova, Evaluation the Quality of Rubber Composites Using the DOE Method, Quality - Access to Success, vol. 18, (2017), pp. 60–63.
[25] I.H. Witten, E. Frank, H. Ma, Data Mining: Practical Machine Learning Tools and Techniques, 3rd ed., Morgan Kaufmann, Burlington, MA, 2011.
[26] M. Kantardzic, M. Data Mining, Concepts, Models, Methods, and Algorithms, 2nd ed., John Wiley & Sons, New York, 2002.
[27] L. Rokach, O. Maimon, Data Mining with Decision Trees: Theory and Applications, 2nd ed, World Scientific Pub Co., Singapore, 2014.
[28] H.W. Ma, Q. Mao, X. Zhang, D. Zhang, H. Chen, Defects classification of steel cord conveyor belt based on rough set and multi-class v-SVM, Adv. Mater. Res. 328-
330 (2011) 1814–1819.
[29] M. Andrejiova, A. Grincova, D. Marasova, Measurement and simulation of impact wear damage to industrial conveyor belts, Wear 368 (2016) 400–407.
[30] A. Khan, Ch.K. Ryoo, H.S. Kim, Comparative study of classification algorithms for damage classification in smart composite laminates, Proc. SPIE 10167,
Nanosensors, Biosensors, Info-Tech Sensors and 3D Systems, 2017.
[31] A.K. Patel, S. Chatterjee, A.K. Gorai, Development of machine vision-based ore classification model using support vector machine (SVM) algorithm, Arab. J.
Geosci. 10 (2017) 107.
[32] S. Rahmadani, A. Dongoran, M. Zarlis, Comparison of naive Bayes and decision tree on feature selection. Using genetic algorithm for classification problem, J.
Phys. Conf. Ser. 978 (2018) 012087.
[33] A. Malinowska, Classification and regression tree theory application for assessment of building damage caused by surface deformation, Nat. Hazards 73 (2014)
317–334.
[34] V. Muralidharan, S. Ravikumar, H. Kangasabapathy, Condition monitoring of self aligning carrying idler (SAI) in belt-conveyor system using statistical features
and decision tree algorithm, Measurement 58 (2014) 274–279.

417