You are on page 1of 14

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/297673218

Failure rate analysis of jaw crusher using Weibull model

Article  in  ARCHIVE Proceedings of the Institution of Mechanical Engineers Part E Journal of Process Mechanical Engineering 1989-1996 (vols · March 2016
DOI: 10.1177/0954408916636922

CITATIONS READS

2 359

2 authors, including:

Ravi Sinha
Indian Institute of Technology (ISM) Dhanbad
4 PUBLICATIONS   18 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Ravi Sinha on 26 November 2018.

The user has requested enhancement of the downloaded file.


Original Article

Proc IMechE Part E:


J Process Mechanical Engineering
Failure rate analysis of jaw crusher 0(0) 1–13
! IMechE 2016

using Weibull model Reprints and permissions:


sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0954408916636922
uk.sagepub.com/jpme

RS Sinha and AK Mukhopadhyay

Abstract
The primary crusher is essential equipment employed for comminuting the mineral in processing plants. Any kind of
failure of its components will accordingly hinder the performance of the plant. Therefore, to minimize sudden failures,
analysis should be undertaken to improve performance and operational reliability of the crushers and its components.
This paper considers the methods for analyzing failure rates of a jaw crusher and its critical components application of a
two-parameter Weibull distribution in a mineral processing plant fitted using statistical tests such as goodness of fit and
maximum likelihood estimation. Monte Carlo simulation, analysis of variance, and artificial neural network are also
applied. Two-parameter Weibull distribution is found to be the best fit distribution using Kolmogorov–Smirnov test.
Maximum likelihood estimation method is used to find out the shape and scale parameter of two-parameter Weibull
distribution. Monte Carlo simulation generates 40 numbers of shape parameters, scale parameters, and time. Further, 40
numbers of Weibull distribution parameters are evaluated to examine the failure rate, significant difference, and regres-
sion coefficient using ANOVA. Artificial neural network with back-propagation algorithm is used to determine R2 and is
compared with analysis of variance.

Keywords
Software reliability, Monte Carlo simulation, analysis of variance, artificial neural networks, failure data analysis

Date received: 7 July 2015; accepted: 8 February 2016

distribution from the pool of TBFs data.3,4 The K-S


Introduction test is often employed for testing the null hypothesis
The reliability of a unit comprising subsystems in where a data set has a distribution described by a fully
series is evaluated by multiplying the reliabilities of determined distribution function.5,6 Failure data are
individual subsystem.1 The major problem of the min- examined by exponential, normal, lognormal, gamma,
eral processing plant equipment is how long a unit logistic, and two-parameter Weibull distributions.
may operate without failure. The failure of capital- Maximum likelihood (ML) estimation method is
intensive machine components cannot be ignored in used to determine the shape and scale parameter of
the competitive environment. The sudden and fre- two-parameter Weibull distribution.7–9 The estima-
quent failure of equipment/components leads to pro- tion theory is a cornerstone in statistical analyses,
duction loss. As the number of components increase, and several techniques are used to assess the param-
the reliability of the unit decreases. Operational reli- eters, of which ML estimation, graphical procedure,
ability investigation is desirable for reducing the moments method, and weighted least square method
maintenance cost and improving the performance of are some of the most used ones.10–14 ML estimation is
the plant.2 Crushers are designed in such a way that asymptotically unbiased with the minimum variance
they reduce rock in a manner where all pieces are less and maximum likelihood estimation, and is the tool
than a stated size. But no crusher has been devised used for parameter estimation.15 In this two-
which produces only fragments greater than a speci-
fied size. A crusher always produces various sizes of
particles with substantial amount of fines. The time- Department of Mining Machinery Engineering Indian School of Mines,
between-failures (TBFs) data of jaw crusher of 2-year Dhanbad, India
duration from the plant maintenance record book
Corresponding author:
have been considered for analysis. The RS Sinha, Department of Mining Machinery Engineering Indian School of
Kolmogorov–Smirnov (K-S) test performs multiple Mines, Dhanbad 826004, India.
goodness-of-fit (GOF) tests determining the best Email: ravi.dr0091@mme.ism.ac.in

Downloaded from pie.sagepub.com by guest on March 9, 2016


2 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

parameter Weibull distribution, density function is of ANNs. The ANN model having the best prediction
given in equation (1)16 and failure rate (h) is given performance is detected by trying various networks.
in equation (2) The ANN method has determined R2 values, which
indicates the GOF of jaw crusher and its components.
  
 t   t  The R2 values of ANN are compared with ANOVA. It
fðtÞ ¼ e ,  4 0,  4 0 ð1Þ is shown that the R2 value of ANN model was approxi-
 
mately same as that compared to ANOVA. The results
  also indicated that the designed model is useful, reli-
 t 1
hðtÞ ¼ ð2Þ able, and quite effective for analyzing the effects of
 
failure rate of jaw crusher.51 The goal of this paper is
to find out failure rates and R2 using ANOVA and
where h(t) is the failure rate, b is the shape parameter, validate the result by ANN through R2 values.
Z is the scale parameter, and t is the time. No sparse The actual performance of plant machineries is pri-
data lead to large error limit for two-parameter marily a function of failure rates of its components
Weibull fitting.17 From equation (2), Weibull failure and maintenance philosophies applied by the manage-
model helped to find out failure rate, significant dif- ment of the specific plant. Prior to taking up the study
ferences, and R2 (regression coefficient) by ANOVA and during general survey of the plant, it was assumed
(analysis of variance). In this study, the two-para- that the reliability of the plant and its equipment
meter Weibull distribution has two parameters as could be improved after identifying the frequency of
shape and scale, and time is also used as a third par- failure of critical components and adopting suitable
ameter. Shape parameter (b) and scale parameter (Z) is maintenance scheme, which finally determines the
a function of time (t) of two-parameter Weibull distri- asset utilization.
bution. Taking these three parameters (shape, scale,
and time), we proceed to analyze the failure rate by
ANOVA. Most of the research works carried out so
Mineral processing plant
far uses ANOVA found in the areas of wear and coat- The equipment in a mineral processing plant works in
ing.18–24 By Monte Carlo simulation (MCS), 40 num- series. The plants include four general types of oper-
bers of shape parameters, scale parameters, and time ations which are sizing, comminuting, mineral recov-
data are generated from two-parameter Weibull distri- ery, and dewatering. A jaw crusher is used as the
bution. These data are used for 23 factorial design primary type size reduction equipment. The plant
using ANOVA to minimize the overall number of equipment layout is shown in Figure 1. There are no
examinations. The random variability is explained by standby equipments and all equipments work in
the second factor, so it is more easy to find out the series. If one particular unit in the circuit is down,
significant differences. The significant interaction of the operation of entire plant stops. Efficient mainten-
one variable changes depending on the level of the ance program of different units for improving oper-
other factor. ANOVA also determined the relation ational reliability thus becomes essential.
between observed and predicted value by R2, which
indicated GOF of jaw crusher and its components.
Jaw crusher
Using ANOVA, 40 numbers of shape parameter,
scale parameter, and time determined the failure Crushing is the first stage of size reduction of rocks
rates and R2 of jaw crusher and its components. received from mines. Crushers are heavy-duty
Artificial neural network (ANN) models have the machines, used to reduce the run-of-mine ore down
capability of detecting complex underlying nonlinear to a predetermined size. A line diagram of a double
relationships between 40 numbers of input and output toggle jaw crusher and its main components is shown
data, and trained using several effective training algo- in Figure 2. The jaws are set at an angle to each other.
rithms.25,26 The input data was considered as shape The moving jaw moves forward towards the fixed jaw
parameter, scale parameter, and time, and output as causing fragmentation of rocks inside the crushing
failure rate of two-parameter Weibull distribution chamber and then moves backward in the next half
from equation (2). However, in the last few years, of its working cycle. The crushed materials get dis-
ANNs with regression analysis have been used to pre- charged by gravity from the bottom of the crusher
dict uniaxial compressive strength (UCS) from texture during the return stroke.
characteristics of rocks,27–29 air quality forecasting
community,30–35 electricity consumption,36–40 field of
wood science,41–50 etc. This paper applied ANN Methodology
method for jaw crusher and its components. ANNs
Data collection and treatment of data
are designed according to their connection architec-
ture, learning algorithm, number of hidden layers, The time-between-failure (TBF) data of jaw crusher
number of nodes in a hidden layer, and transfer func- of 2-year duration are collected from a plant mainten-
tion. Also, these design criteria affect the performance ance record book. The components of the crusher

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 3

Figure 1. Block diagram mineral processing plant operational unit.

where supx is the supremum of the set of distances.


For this analysis, the TBFs are treated as complete
samples. The GOF describes how well it fits a set of
observations in a particular distribution.52–55 Forty
numbers of failure data of Table 1 are tested in dif-
ferent distributions and it is observed that these data
in two-parameter Weibull distribution fits best. TBFs’
distribution and the parameters are estimated by the
ML method using WEIBULLþþ, a computer pack-
age for statistical analysis. Table 2 presents the K-S
test of components of jaw crusher for the distributions
considered in this study from equation (4). It is
observed that in two-parameter Weibull distribution,
the data fit more closely to the input field data com-
pared to other distributions.
Figure 2. A double toggle jaw crusher.

which are critical and having major impact on plant’s


Maximum likelihood estimation
operation have been considered for analysis. TBFs of The ML estimation is a method of estimating the par-
these components are presented in Table 1. TBFs’ ameters of a statistical model. When applied to a fail-
data are considered for those components only ure data and given a statistical model, the ML
whose failure caused stoppage of functioning of the estimation provides estimates for the model’s param-
whole equipment. Forty number of such component eters. The shape and scale parameters of the two-
failures have been recorded in 2 years. parameter Weibull distribution are found out by the
ML estimation method. The ML estimation requires
minimal distributional assumptions, and is valuable
Goodness of fit
for obtaining a descriptive measure for the purpose
Life data analysis (LDA) is a method to evaluate of summarizing observed data, but it has no basis for
GOF and ML estimation of jaw crusher and its com- testing hypotheses or constructing confidence inter-
ponents. GOF uses Kolmogorov–Smirnov (K-S) test vals. It is a standard approach to parameter estima-
to identify the best distribution among normal, log- tion and inference in statistics. Let x1, x2. . .xn be a
normal, Weibull, exponential, gamma, generalized random sample size n drawn from a probability dens-
gamma, logistic, and log-logistic distributions from ity function fx (x; ) where  is an unknown param-
the TBF data. The K-S test is used as a measure of eter. The likelihood function of this random sample is
GOF of different theoretical distributions. In statis- the joint density of the n random variables and is a
tics, the K-S test is a nonparametric test and it is function of the unknown parameter. Thus
used for GOF. The empirical distribution function
Fn for n independent and identically distributed obser- Y
n

vations Xi is given in the following equation L¼ fxi ðxi , Þ ð5Þ


i¼1

1X n
is the likelihood function. The ML estimation of  is a
Fn ðxÞ ¼ I½1, x ðxi Þ ð3Þ solution of
n i¼1

where I½1, x ðxi Þ the indicator function, equal to 1 if d log L


¼0 ð6Þ
Xi 4 x and equal to 0 otherwise. The K-S statistic for d
a given cumulative distribution function F(x) is
Now, we are going to apply the ML estimation to
  estimate Weibull parameter, namely the shape and the
Dmax ¼ supx Fn ðxÞ  FðxÞ ð4Þ
scale parameters. Consider the Weibull pdf given in

Downloaded from pie.sagepub.com by guest on March 9, 2016


4 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

Table 1. TBFs of the jaw crusher.

System/Subsystem No. of failures Time between failure (TBF)

Jaw crusher unit 40 12, 60, 70, 108, 110, 132, 135, 165, 180, 185, 200, 216, 225, 245, 250, 273, 300, 324, 330,
340, 350, 360, 360, 360, 360, 360, 365, 370, 380, 396, 400, 405, 410, 420,425, 430, 430,
431, 432, 432
Back toggle plate 11 396, 455, 567, 600, 650, 711, 767, 790, 800, 800, 1800
Liner plate 10 492, 500, 560, 600, 610, 620, 636, 670, 730, 888
Jaw plate 9 60, 100, 210, 400, 556, 708, 810, 1100, 1248
Tie rod 10 708, 720, 750, 790, 810, 828, 900, 950, 1100, 1320

Table 2. Goodness-of-fit test of jaw crusher components.

K-S test (Dmax)

System/Subsystem Exponential Normal Lognormal Gamma Weibull Logistic Best fit model

Jaw crusher 43.98 1.08 1.28 3.83 0.33 6.53 Weibull


Back toggle plate 1.2  1010 3.91  108 1.2  1010 2.24  107 1.0  1010 1.15  108 Weibull
chute liner plate 8.52  108 0.010 1.73  105 1.34  109 1.0  1010 0.012 Weibull
Jaw plate 1.0  1010 3.03  107 5.5  109 4.42  108 1.0  1010 2.45  107 Weibull
Tie rod 5.36  108 2.34  107 7.0  107 1.0  1010 1.0  1010 5.79  106 Weibull
K-S: Kolmogorov–Smirnov.

equation (1), then likelihood function will be as given


in equation (7). These parameters are presented in Table 3. ML estimation of jaw crusher.
Table 3.
System/Subsystem b Z
Yn   1 x 
 xi i Jaw crusher 1.17 261
Lðx1 , . . . , xn ; , Þ ¼ e  ð7Þ
  Back toggle plate 1.28 1186
i¼1
Chute liner plate 4.33 714
Jaw plate 0.62 788
Monte Carlo simulation Tie rod 3.18 1060

The MCS is a process to run simulation with small


number of data to obtain a particular distribution of regression on Y. In Figure 3, results clearly indicate
with an aim to generate large number of data for ana- that median line and true parameter line of RRX pro-
lysis. The MCS-based observations are better than the vides the least deviation from the real failure data of
man-made observations for predicting a particular the jaw crusher and its components. Forty numbers of
distribution.56–58 The MCS considered 40 numbers shape parameters, scale parameters, and time of two-
of randomly selected data using two-parameter parameter Weibull distribution are shown in Table 4.
Weibull distribution. The MCS method uses Time (t) data of two-parameter Weibull distribution
WEIBULLþþ software. When the simulation is com- are generated by the MCS. Shape parameter (b) and
pleted, the results are used to describe the likelihood scale parameter (Z) is a function of time (t) of two-
distribution. It allows estimation of parameter likes parameter Weibull distribution. Where b is the shape
rank regression on X (RRX) and rank regression on parameter, also known as the Weibull slope, Z is the
Y (RRY). Rank regression method for parameter esti- scale parameter which indicates the time between fail-
mation is also known as the least squares method. The ures, and t is the time of equipment/components during
regression is on the X-axis, and then the line is fitted operation. These 40 numbers of data are used for the
so that the horizontal deviations from the points to analysis of failure rate, effect (and interactions) of com-
the line are minimized. If the regression is on the ponent parameters, and R2 through ANOVA.
Y-axis, then this means that the distance of the verti-
cal deviations from the points to the line is minimized.
Analysis of the results of MCS has shown that the
Analysis of variance
rank regression on X results is closer to the actual Factorial design is one of the most effective techniques
distribution used to generate the data than the results using ANOVA to minimize the overall number of

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 5

Figure 3. Monte Carlo simulation plotting for jaw crusher of (a) RRX and (b) RRY analysis of 40 simulated data.

(interaction between parameter). In the present study,


Table 4. Shape parameter, scale parameter, and time shape parameter, scale parameter, and time have been
(input data) of jaw crusher. considered for their influence on jaw crusher and its
component’s failure rates using 23 full factorial
No. b Z t No. b Z t
design. A total number of eight runs are conducted
1 2.7 564 381.5 21 0.4 190.2 127.6 to complete the design. The ANOVA is used to deter-
2 0.9 128.4 41.2 22 1.7 490.3 265.6 mine the failure rate, effect (and interactions) of com-
3 2.0 398.2 232.8 23 0.8 249.6 62.3 ponent parameters, and R2 value of the jaw crusher
4 0.4 332.9 30.0 24 0.5 120.5 152.9 and its components through DOE þþ tool. The vari-
5 4.4 248.4 195.5 25 0.7 462.7 95.7 ance of a set of n equally likely values can be written
6 1.7 510.7 270.3 26 0.4 224.2 375.8
in equation (8) as
7 0.7 648.9 141.5 27 0.8 143.4 36.7
1X n
8 3.4 670.0 650 28 1.6 194.9 98.7 S2 ¼ ðxi  Þ2 ð8Þ
n 1¼1
9 1.4 127.0 60.8 29 0.7 385.2 79.0
10 0.9 139.5 41.6 30 3.1 228.0 162.1 where xi is the number of sample, and m is the
11 0.7 214.8 47.2 31 2.7 377.5 256.4 expected value as shown in the following equation
12 3.7 405.2 305.1 32 2.1 300.6 179.8
13 2.5 270.5 178.3 33 1.3 195.7 84.6 1X n
¼ xi ð9Þ
14 2.2 289.0 274.9 34 1.7 236.4 126.5 n i¼1
15 1.8 444.8 250.5 35 0.4 128.8 225.9
16 3.1 142.0 100.5 36 1.0 245.6 87.3
17 0.8 242.8 69.1 37 1.3 195.0 86.7 Artificial neural networks
18 0.4 225.5 116.4 38 0.3 290.9 44.5
The ANNs have been trained to overcome the restric-
19 1.3 330.4 143.6 39 0.5 180.1 225.9
tion of the traditional methods in solving multifaceted
20 0.6 206.7 36.5 40 1.3 280.7 126.5
problems.68 The ANN algorithm simulates human
learning processes through establishment and
reinforcement of linkages between the input and
examinations.59–61 ANOVA is used to understand output data. The linkages then connect input and
statistical significance of failure data and to examine output data in the absence of training data.69
the jaw crusher failure rates.62 The MCS 40 numbers Numerous ANN algorithms have been suggested,
of data uses ANOVA to determine the influence of such as radial basis function,70 Elman recurrent,71
failure rates on two-parameter Weibull distribution. and Hopfield neural networks;72 however, back-pro-
The ANOVA gives failure rate equations, which high- pagation algorithm73–75 is the most popular. In this
lights the influence of different factors and their study, the back propagation algorithm has been used,
importance.63–67 Using failure rate equations, failure which is more powerful, accurate, and faster. A neural
rates of crusher and its components are calculated. network has an input layer, an output layer, and a
The ANOVA method is used at confidence level of number of hidden layers connected to each
90%, i.e. with p 4 0.1 indicating significant difference other.76–81 The input layer has three neurons

Downloaded from pie.sagepub.com by guest on March 9, 2016


6 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

Table 5. Output data of jaw crusher.

Failure Failure Failure Failure


No. rate No. rate No. rate No. rate

1 0.0001 11 0.0013 21 0.0063 31 0.0137


2 0.0017 12 0.0025 22 0.0019 32 0.0018
3 0.0056 13 0.0055 23 0.0015 33 0.0027
4 0.0013 14 0.0157 24 0.0018 34 0.0017
5 0.0029 15 0.0073 25 0.0011 35 0.0017
6 0.0092 16 0.004 26 0.0015 36 0.0018
7 0.0025 17 0.0015 27 0.0021 37 0.0009
8 0.0551 18 0.0012 28 0.0016 38 0.0013
9 0.0027 19 0.0029 29 0.0049 39 0.0033
10 0.0016 20 0.0012 30 0.0293 40 0.0006

Figure 4. Configuration of artificial neural network


Table 6. Parameter and data selected for ANN process.
architecture.
System/components Data Training Testing Validation

Jaw crusher 40 28 6 6
corresponding to the shape parameter, scale param- Back toggle plate 40 28 6 6
eter, and time, and an output is the failure rate as seen Chute liner plate 40 28 6 6
in equation (2). The input layer, hidden layer, and an Jaw plate 40 28 6 6
output layer are shown in Figure 4. The failure rates/ Tie rod 40 28 6 6
hour as an output of jaw crusher is mentioned in
Table 5. The network having six hidden layers and
the lowest error is selected. MATLAB neural network
tool is used for the configuration, training, and opti- In statistics, the p-value is a function of the observed
mization of the ANN model. Forty numbers of data data that is used for testing a statistical hypothesis.
are randomly divided into three groups for training, Before performing a test, a threshold value is chosen,
testing, and validation of the model. Among 40 num- called the significance level of the test, traditionally 5%
bers of data, 28 (70% of whole data) data were or 10%. p-Value of jaw crusher and its components are
selected for the ANN training process, 6 (15% of given in Table 9. The ANOVA shows p-value to be 0.1,
whole data) data selected for the ANN validation pro- i.e. with a confidence level of 90% or more. p-Values
cess, and the remaining 6 (15% of whole data) data 40.1 are regarded to have a statistically significant
were used for the ANN testing process and clearly contribution for the jaw crusher and its components.
mentioned in Table 6. The ANN is used to determine p-Value is affected by the variables which are explained
R2, which indicates GOF of the jaw crusher and its later. In Table 9, it can be observed that p-value of jaw
components. Advantages and disadvantages of the crusher and back toggle plate has one significant vari-
statistical method in this paper are presented in able as a AC (shape parameter and time parameter)
Table 7. interaction, chute liner plate and tie rod has no signifi-
cant variables, and jaw plate has three significant vari-
ables as a shape parameter (A), time (C), interaction of
Results and discussion
shape parameter, and time parameter, i.e. AC.
The shape parameter, scale parameter, and time data
of different components are estimated by the MCS
method. The factorial design is accomplished from
Failure rate analysis by mathematical modeling
analysis of variance. Factorial design is used to deter- The mathematical model developed consisting main
mine failure rates using shape parameter, scale par- variables A, B, C and their interaction for a 23 factorial
ameter, and time data. The failure rate of a system design using equation (10) to obtain failure rates h(t).
usually depends on time and its unit is failure rate/
hour. Failure rates of the crusher and its components hðtÞ ¼ X0 þ X1 A þ X2 B þ X3 C þ X4 AB
are examined using 23 full factorial designs consider- þ X5 AC þ X6 BC ð10Þ
ing two levels (high level and low level) and three
variables. High level is indicated as þ1 and low level where X0, X1, X2, X3, X4, X5, and X6 are regression
as 1. The level and ranges of the variables affecting coefficients, A is the shape parameter, B is a scale
the failure rate is shown in Table 8. parameter, B is a scale parameter, and C is the time.

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 7

Table 7. Advantages and disadvantages of statistical methods.

Method Advantage Disadvantage

Kolmogorov–Smirnov test It only applies to continuous distributions. The critical region of the K-S test is no longer
valid.
ML Estimation Maximum likelihood provides a consistent The likelihood equations need to be specifically
approach to parameter estimation problems. worked out for a given distribution and esti-
mation problem.
Analysis of variance The main advantage of the within subjects design Data not completely independent (assumption of
is that it controls for individual differences ANOVA).
between participants. The variability due to the
participants is assumed not to vary across the
treatment levels.
Monte Carlo simulation It generates very accurate distribution parameter. It is necessary to make some possibly unwar-
ranted parametric distributional assumptions
at the data simulation stage.
Artificial neural network Neural networks are quite simple to implement. Handling of time series data in neural networks is
a very complicated topic.

Table 8. Ranges of the factors used in factorial design for jaw crusher.

Level of range

Variables Symbol Jaw crusher Back toggle plate Chute liner plate Jaw plate Tie rod

1 þ1 1 þ1 1 þ1 1 þ1 1 þ1
b A 0.4 4.4 0.5 4.8 1.6 16.4 0.2 2.4 1.2 15.4
Z B 170 647 560 2470 579 890 170 2800 750 940
t C 30 650 140 2475 380 770 10 2600 440 950

Table 9. p-Value of jaw crusher and its components. Table 11. Epoch and validation result of jaw crusher and its
components.
p-Value
Best
Jaw Back toggle Chute liner Jaw System Best epoch validation
Term crusher plate plate plate Tie rod
Jaw crusher 23 0.00000012
A 0.1715 0.1718 0.6604 0.069 0.7199
Back toggle plate 237 0.000000017
B 0.6126 0.5458 0.4988 0.3832 0.4889
Chute liner plate 16 0.00000037
C 0.1817 0.2066 0.5323 0.0664 0.5788
Jaw plate 10 0.000000096
AB 0.2297 0.1878 0.4885 0.3605 0.5501
Tie rod 17 0.000000011
AC 0.0936 0.0887 0.4592 0.0656 0.4101
BC 0.1979 0.1879 0.4854 0.2895 0.4174

Table 12. R2, % error and variance of jaw crusher and its
components.
Table 10. Failure rate value of jaw crusher and its
components. System/ S2
Component ANOVA ANN % Error (Variance)
System/Subsystem Failure rate [h(t)]/hours
Jaw crusher 98.98% 98.99% 0.01 1.7e-8
Jaw crusher 1.14  103 Back toggle plate 98.87% 98.95% 0.08 7.56e-5
Back toggle plate 1.19  103 Chute liner plate 92.83% 94.37% 1.65 1.0e-8
Chute liner plate 3.29  103 Jaw plate 99.64%, 99.14%, 0.50 6.4e-9
Jaw plate 5.61  104 Tie rod 93.12% 95.19% 2.22 1.91e-7
Tie rod 2.56  103
ANOVA: analysis of variance; ANN: artificial neural network.

Downloaded from pie.sagepub.com by guest on March 9, 2016


8 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

Figure 5. Fitted vs. actual value of (a) jaw crusher, (b) back toggle plate, (c) chute liner plate, (d) jaw plate, and (e) and tie rod.

The regression equations are developed using p-values For chute liner plate
from Table 6.
For jaw crusher
hðtÞ ¼ 0:0010 þ 0:0004A þ 0:0007B þ 0:0006C
hðtÞ ¼ 0:0013  0:0006A þ 0:0001B  0:0006C þ 0:0007AB þ 0:0008AC þ 0:0008BC ð13Þ
þ 0:0002AB þ 0:0011AC þ 0:0001BC ð11Þ

For back toggle plate For jaw plate

hðtÞ ¼ 0:0012  0:0005A þ 0:0001B  0:0006C hðtÞ ¼ 0:0005  0:0003A  0:0002B  0:0004C
þ 0:0001AB þ 0:0013AC þ 0:0001BC ð12Þ þ 0:0002AB þ 0:0004AC þ 0:0003BC ð14Þ

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 9

Figure 6. Observed vs. predicted value of (a) jaw crusher, (b) back toggle plate, (c) chute liner plate, (d) jaw plate, and (e) and tie rod.

For tie rod plate has low failure rate and back toggle plate, chute
liner plate, and tie rod have high failure rate. The
failure was determined through ML estimation and
hðtÞ ¼ 0:0007  0:0002A þ 0:0008B þ 0:0005C
ANOVA.
þ 0:0007AB þ 0:0009AC þ 0:0008BC ð15Þ

ANOVA regression coefficient (R2) analysis


From equations (11) to (15), the calculated failure
rates are shown in Table 10. From the table, jaw The R2 indicates GOFs of a mathematical model. The
crusher, chute liner plate, tie rod, and back toggle value of R2 of crusher and its components are pre-
plate exhibit high failure rate, and jaw plate has low sented in Table 12. According to Figure 5, the R2
failure rates. Among of all components, jaw plate has values predicted by ANOVA are 98.98% for jaw
the minimum failure rate value. As per Table 3, jaw crusher, 98.87% for back toggle plate, 92.83% for

Downloaded from pie.sagepub.com by guest on March 9, 2016


10 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

chute liner plate, 99.64% for jaw plate, and 93.12% for distribution has best fit from TBFs data of jaw
tie rod. This result means that the designed model is crusher. Using of regression equation confirmed that
capable of explaining at least 98.98%, 98.87%, the failure rate of jaw crusher is 1.14  103, back
92.83%, 99.64%, and 93.12% of R2 value. Failure toggle plate is 1.19  103, chute liner plate is
rate and R2 value indicate that jaw plate has low failure 3.29  103, jaw plate is 5.61  104, and tie rod is
rate and are shown in Tables 10 and 12. These values 2.56  103. R2 values were predicted by the
also support the applicability of using ANOVA. ANOVA and they are found to be 98.98% for jaw
crusher, 98.87% for back toggle plate, 92.83% for
chute liner plate, 99.64% for jaw plate, and 93.12%
ANN regression coefficient (R2) analysis for tie rod. Both failure rate value and R2 value indi-
In this study, feed-forward and back-propagation cated that the jaw plate has low failure rate.
multilayer ANN is chosen for solution of the prob- The ANN has 23 epochs for jaw crusher, 237
lem. The relationship between observed values and epochs for back toggle plate, 16 epochs for chute
predicted values by the ANN models for all data is liner plate, 10 epochs for jaw plate, and 17 epochs
illustrated in Figure 6. The figure shows the graphic of for tie rod. Jaw plate has minimum epochs with min-
error variation depending on iteration of the ANN imum error of about 0.000000096. Using the ANN,
model chosen for the bonding strength. About 1000 R2 value was found to be 98.99% for jaw crusher,
epochs have been done in this ANN process. The 98.95% for back toggle plate, 94.37% for chute
ANN has 23 epochs for jaw crusher, 237 epochs for liner plate, 99.14% for jaw plate, and 95.19 % for
back toggle plate, 16 epochs for chute liner plate, 10 tie rod. The results showed that the ANOVA and
epochs for jaw plate, and 17 epochs for tie rod. It is ANN model has a very high R2 value between pre-
decided that the 0.000000096 targeted error values for dicted and observed values. The R2 value is high
jaw plate would be sufficient for training of the ANN. which agrees that the results are strongly correlated.
The training of the ANN was stopped after 10 epochs The results displayed that the R2 value of ANN and
because the targeted value 0.000000096 was reached ANOVA is approximately the same. Such analysis is
and is shown in Table 11. Among all the components, important for better utilization of the equipment as
jaw plate has the minimum epoch and also minimum well as its components.
error. The R2 values are found to be 98.99% for jaw
crusher, 98.95% for back toggle plate, 94.37% for Declaration of conflicting interests
chute liner plate, 99.14% for jaw plate, and 95.19% The author(s) declared no potential conflicts of interest with
for tie rod using the ANN model. The epoch and R2 respect to the research, authorship, and/or publication of
values have indicated that jaw plate has minimum this article.
error amongst other components of jaw crusher and
are shown in Tables 11 and 12. Funding
The author(s) received no financial support for the research,
Comparitive analysis of ANOVA and ANN authorship, and/or publication of this article.

On the other hand, the ANN model is approximately


References
same as prediction R2 value when compared to the
ANOVA model and are shown in Table 12. It has 1. Castro D and Cavalca KL. Maintenance resources opti-
been observed that the percentage error for both mization applied to a manufacturing system. Reliab Eng
Syst Safety 2006; 91: 413–420.
regressions is less than 2%. This is indicative of
2. Kuo W and Zuo MJ. Optimal reliability modelling -
close agreement between the observed and predicted Principles and applications. Hoboken, NJ: New Jersey:
results. The results displayed that ANOVA and ANN John Wiley & Sons, 2003.
prediction models can be used for accurate prediction 3. Kumar U, Klefsjo B and Granholm S. Reliability inves-
because they have a high degree of explanatory. From tigation for a fleet of load haul dump machines in a
the observation of variance, it is concluded that the Swedish mine. Reliab Eng Syst Safety 1989; 26: 341–361.
variance is zero for the jaw crusher and its components 4. Crutcher HL. A note on the possible misuse of the
are as shown in Table 12. Variance is zero indicating Kolmogorov-Smirnov test. J Appl Meteorol 1975; 14:
that the data points tend to be very close to the mean. 1600–1603.
5. Steinskog DJ, Tjøstheim DB and Kvamsto NG. A cau-
tionary note on the use of the Kolmogorov-Smirnov
Conclusion test for normality. Monthly Weather Rev 2007; 135:
1151–1157.
Considering the relationship between inputs and 6. Abbasi B, Jahromi AHE, Arkat J, et al. Estimation the
output, the results obtained by the prediction parameter of Weibull distribution using simulated algo-
models are highly encouraging. The K-S test, ML esti- rithm. Appl Math Comput 2006; 183: 85–93.
mation, ANOVA, and ANN are found to be effective 7. Al-Fawzan MA. Methods for estimation the parameters
statistical methods for failure rate analysis. The K-S of the Weibull distribution. King Abdulaziz City for
test confirmed that the two-parameter Weibull Science and Technology, Saudi Arabia, October 2000.

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 11

8. Pasha GB, Khan MS and Pasha AH. Empirical analysis 29. Corani G. Air quality prediction in Milan: Feed-for-
of the Weibull distribution for failure data. J Statist ward neural networks, pruned neural networks and
2013; 13: 33–45. lazy learning. Ecol Model 2005; 185: 513–529.
9. Ross R. Graphical method for plotting and evaluating 30. Robles LAD, Ortega JC and Fu JS. A hybrid ARIMA
Weibull distribution data. In: Proceedings of interna- and artificial neural networks model to forecast particu-
tional conference on properties and applications of dielec- late matter in urban areas: The case of Temuco. Chile.
tric materials Brisbane Australia, 1994, pp.250–253. Atmos Environ 2008; 42: 8331–8340.
10. Lehman EH. Shapes, moments and estimators of the 31. Fernando HJ, Mammarella MC and Grandoni G.
Weibull distribution. Trans IEEE Reliab 1962; 11: Forecasting PM10 in metropolitan areas: Efficacy of
32–38. neural networks. Environ Pollut 2012; 163: 62–67.
11. White JS. The moments of log-Weibull order statistics. 32. Ordieres J, Vergara E, Capuz R, et al. Neural network
Technimetrics 1962; 11: 374–386. prediction model for fine particulate matter (PM2.5) on
12. Usher JS. Weibull component reliability prediction in the US–Mexico border in El Paso (Texas) and Ciudad
the presence of masked data. IEEE Trans Reliab 1996; Juarez (Chihuahua). Environ Model Software 2005; 20:
45: 229–232. 547–559.
13. Brain LJ and Antle CE. Estimation of parameter in the 33. Bárcenas OP, Olivas ES and Guerrero JDM. Unbiased
Weibull distribution. Technimetrics 1967; 9: 621–627. sensitivity analysis and pruning techniques in neural
14. Dubey SD. Asymptotic properties of several estima- networks for surface ozone modelling. Ecol Model
tion of Weibull parameter. Technimetrics 1965; 7: 2005; 182: 149–158.
423–434. 34. Perez P, Trier A and Reyes J. Prediction of PM2.5 con-
15. Mann and Schafer NR. Method for statistical analysis centrations several hours in advance using neural net-
of reliability and life data. New York: John Wiley and works in Santiago, Chile. Atmos Environ 2000; 34:
Sons, 1974. 1189–1196.
16. Liu C-C. A comparison between the Weibull and lognor- 35. Ekonomou L. Greek long-term energy consumption
mal models used to analyse reliability data. PhD Thesis, prediction using artificial neural network. Energy
University of Nottingham, UK, 1997. 2010; 35: 512–517.
17. Kumar A, Mahapatra MM and Jha PK. Modeling the 36. Geem ZW and Roper WE. Energy demand estimation
abrasive wear characteristics of in-situ synthesized Al– of South Korea using artificial neural network. Energy
4.5%Cu/TiC composites. Wear 2013; 306: 170–178. Policy 2009; 37: 4049–4054.
18. Tjong SC and Ma ZY. Microstructural and mechanical 37. Sozen A, Arcaklioglu E and Ozkaymak M. Turkey’s net
characteristics of in situ metal matrix composites. Mater energy consumption. Appl Energy 2005; 81: 209–2021.
Sci Eng 2000; 29: 49–113. 38. Sozen A, Akcayol MA and Arcaklioglu E. Forecasting
19. Chu MG and Premkumar MK. Mechanism of TiC for- net energy consumption using artificial neural network.
mation in Al/TiC in-situ metal matrix composites. Energy Sources Part B 2006; 1: 147–155.
Metall Trans 1993; 24: 2803–2805. 39. Kankal M, Akpinar A, Komurcu MI, et al. Modeling
20. Wahab MN, Daud AR and Ghazali MJ. Preparation and forecasting of Turkey’s energy consumption using
and characterization of stir cast-aluminium nitride rein- socio-economic and demographic variables. Appl
forced aluminium metal matrix composites. Int J Mech Energy 2010; 88: 1927–1939.
Mater Eng 2009; 4: 115–117. 40. Cook DF and Whittaker AD. Neural network process
21. Surappa MK and Rohatgi PK. Preparation and proper- modeling of a continuous manufacturing operation.
ties of cast aluminium-ceramic particle composites. Eng Appl Artif Intell 1993; 6: 559–564.
J Mater Sci 1981; 16: 883–993. 41. Wu H and Avramidis S. Prediction of timber kiln
22. Sahoo P and Koczak M. Analysis of in situ formation drying rates by neural networks. Drying Technol 2006;
of titanium carbide in aluminium alloys. Mater Sci Eng 24: 1541–1545.
1991; 144: 37–44. 42. Ceylan I. Determination of drying characteristics of
23. Miracle DB. Metal matrix composites from science to timber by using artificial neural network sand mathe-
technological significance. Compos Sci Technol 2005; matic almodels. Drying Technol 2008; 26: 1469–1476.
65: 2526–2540. 43. Mansfield SD, Iliadis L and Avramidis S. Neural net-
24. Hornik K. Some new results on neural network work prediction of bending strength and stiffness in
approximation. Neural Networks 1993; 6: 1069–1072. western hemlock (Tsuga heterophylla Raf.). Holz-for-
25. Lippmann R. An introduction to computing with schung 2007; 61: 707–716.
neural nets. IEEE ASSP Mag 1987; 4: 4–22. 44. Fernandez FG, Palacios P and Esteban LG. Prediction
26. Singh VK, Singh D and Singh TN. Prediction of of MOR and MOE of structural plywood board using
strength properties of some schistose rocks from petro- an artificial neural network and comparison with a
graphic properties using artificial neural networks. Int J multivariate regression model. Compos Part B: Eng
Rock Mech Mining Sci 2001; 38: 269–284. 2012; 43: 3528–3533.
27. Singh R, Vishal V, Singh TN, et al. A comparative 45. Packianather MS and Drake PR. Comparison of neural
study of generalized regression neural network and minimum distance classifiers in wood veneer defect
approach and adaptive neuro-fuzzy inference systems identification. Proc IMechE, Part B: J Engineering
for prediction of unconfined compressive strength of Manufacture 2005; 219: 831–841.
rocks. Neural Comput Appl 2012; 23: 499–506. 46. Castellani M and Rowlands H. Evolutionary feature
28. Singh R, Vishal V and Singh TN. Soft computing selection applied to artificial neural networks for
method for assessment of compressional wave velocity. wood veneer classification. Int J Prod Res 2008; 46:
J Sci Iran Trans Civil Eng 2012; 19: 1018–1024. 3085–3105.

Downloaded from pie.sagepub.com by guest on March 9, 2016


12 Proc IMechE Part E: J Process Mechanical Engineering 0(0)

47. Avramidis S and Iliadis L. Predicting wood thermal from plant precursors by a two-level full factorial
conductivity using artificial neural networks. Wood design. Chem Eng J 2010; 160: 99–107.
Fiber Sci 2005; 37: 682–690. 65. Rajesha S, Krishnab AG, Rajuc PRM, et al. Statistical
48. Avramidis S and Wu H. Artificial neural network and analysis of dry sliding wear behaviour of graphite rein-
mathematical modeling comparative analysis of non- forced aluminium MMCs. Procedia Mater Sci 2014; 6:
isothermal diffusion of moisture in wood. Holz Roh 1110–1120.
Werkst 2007; 65: 89–93. 66. Vettivel SC, Selvakumar N, Leema N, et al. Electrical
49. Zhang J, Cao J and Zhang D. ANN based data fusion resistivity, wear map and modelling of extruded tung-
for lumber moisture content sensors. Trans Inst Meas sten reinforced copper composite. Mater Des 2014; 56:
Control 2006; 28: 69–79. 977–996.
50. Tiryaki S, Ozsahin S and Yıldırım I. Comparison of 67. Kaytez F, Taplamacioglu MC, Cam E, et al.
artificial neural network and multiple linear regres- Forecasting electricity consumption: A comparison of
sion models to predict optimum bonding strength of regression analysis, neural networks and least squares
heat treated woods. Int J Adhes Adhes 2014; 55: support vector machines. Electric Power Energy Syst
29–36. 2015; 67: 431–438.
51. Olivares AM and Forero CG. Goodness of fit testing. 68. Campbell JB. Introduction to remote sensing. London:
Int Encyclop Educ 2010; 7: 190–196. Taylor & Francis, 2002.
52. Qiao H and Tsokos CP. Estimation of three parameter 69. Vojislav K. Learning and soft computing: Support vector
Weibull probability distribution. Math Comput Simul machines, neural networks, and fuzzy logic models.
1995; 39: 173–185. Complex Adaptive Systems. Massachusetts: MIT
53. Sarhan AM. Reliability estimations of components Press, 2001.
from masked system life data. Reliab Eng Syst Safety 70. Rakkiyappan R and Balasubramaniam P. Delay-depen-
2001; 74: 107–113. dent asymptotic stability for stochastic delayed recur-
54. Zhou JC and Wu DF. Estimation of Weibull param- rent neural networks with time varying delays. Appl
eters with linear regression method. Mater Sci Technol Math Comput 2008; 198: 526–533.
2010; 26: 1008–1011. 71. Nguyen MQ, Atkinson PM and Lewis HG. Super-reso-
55. Aydogdu H, Senoglu B and Kara M. Parameter estima- lution mapping using Hopfield neural network with
tion in geometric process with Weibull distribution. fused images. IEEE Trans Geosci Remote Sens 2006;
Appl Math Comput 2010; 217: 2657–2665. 44: 736–749.
56. Elmahdy EE and Aboutahoun AW. A new approach 72. Werea K, Buic DT, Dicka B, et al. A comparative
for parameter estimation of finite Weibull mixture dis- assessment of support vector regression, artificial
tributions for reliability modelling. Appl Math Model neural networks, and random forests for predicting
2013; 37: 1800–1810. and mapping soil organic carbon stocks across an
57. Reliability synthesis element user guide. Weibull þþ 9 Afromontane landscape. Ecolog Ind 2015; 52: 394–403.
ReliaSoft Corporation, USA, 2014. 73. Murray AI. Applications of neural networks. Dordrecht:
58. Negrete CC. Optimization of cutting parameters for Kluwer, 1995.
minimizing energy consumption in turning of AISI 74. Ali E, Guang W and Ibrahim A. Empirical relations
6061 T6 using Taguchi methodology and ANOVA. between compressive strength and micro-fabric proper-
J Clean Prod 2013; 53: 195–203. ties of amphibolites using multivariate regression, fuzzy
59. Hajjaji N, Renaudina V, Houasb A, et al. Factorial inference and neural networks: A comparative study.
design of experiment (DOE) for parametric energetic Eng Geol 2014; 183: 230–240.
investigation of a steam methane reforming process 75. Lee S and Evangelista DG. Earthquake-induced
for hydrogen production. Chem Eng Process 2010; 49: landslide susceptibility mapping using an artificial
500–507. neural network. Nat Hazards Earth Syst Sci 2006;
60. Baradeswaran A, Vettivel SC, Perumal AE, et al. 6: 687–695.
Experimental investigation on mechanical behaviour, 76. Conforti M, Pascal S, Robustelli G, et al. Evaluation of
modelling and optimization of wear parameters of prediction capability of the artificial neural networks
B4C and graphite reinforced aluminium hybrid com- for mapping landslide susceptibility in the Turbolo
posites. Mater Des 2014; 63: 620–632. River catchment (northern Calabria, Italy). Catena
61. Mandal N, Doloi B and Mondal B. Development of 2014; 113: 236–250.
flank wear prediction model of Zirconia Toughened 77. Wu W, Dandy GC and Maier HR. Protocol for
Alumina (ZTA) cutting tool using response surface developing ANN models and its application to the
methodology. Int J Refract Metals Hard Mater 2011; assessment of the quality of the ANN model develop-
29: 273–280. ment process in drinking water quality modelling.
62. Sinha RS and Mukhopadhyay AK. Failure analysis of Environ Model Software 2014; 54: 108–127.
jaw crusher and its components using ANOVA. J Brazil 78. Parinet J, Julie M and Nunet P. Predicting equilibrium
Soc Mech Sci Eng 2015; DOI: 10.1007/s40430-015- vapor pressure isotope effects by using artificial neural
0393-6. networks or multi-linear regression-A quantitative
63. Chang SH, Teng TT and Ismail N. Screening of factors structure property relationship approach.
influencing Cu(II) extraction by soybean oil-based Chemosphere 2015.
organic solvents using fractional factorial design. 79. Shadizadeh SR, Karimi F and Zoveidavianpoor M.
J Environ Manage 2011; 92: 2580–2585. Drilling stuck pipe prediction in iranian oil fields: An
64. Gottipati R and Mishra S. Process optimization of artificial neural network approach. Iran J Chem Eng
adsorption of Cr (VI) on activated carbons prepared 2010; 7.

Downloaded from pie.sagepub.com by guest on March 9, 2016


Sinha and Mukhopadhyay 13

80. Zoveidavianpoor M. A comparative study of artificial 81. Zoveidavianpoor M, Samsuri A and Shadizadeh SR.
neural network and adaptive neurofuzzy inference Prediction of compressional wave velocity by artificial
system for prediction of compressional wave velocity. neural network using some conventional well logs in a
Neural Comput Appl 2014; 25: 1169–1176. carbonate reservoir. J Geophys Eng 2013; 10(4): 045014.

Downloaded from pie.sagepub.com by guest on March 9, 2016

View publication stats

You might also like