You are on page 1of 21

J. of Mult.-Valued Logic & Soft Computing, Vol. 25, pp. 535–554 ©2015 Old City Publishing, Inc.

Reprints available directly from the publisher Published by license under the OCP Science imprint,
Photocopying permitted by license only a member of the Old City Publishing Group.

An ANFIS Algorithm for Forecasting Overall


Equipment Effectiveness Parameter in Total
Productive Maintenance

E BRU T URANOGLU B EKAR1 , M EHMET C AKMAKCI2,∗


AND C ENGIZ K AHRAMAN3

1
Department of Industrial Engineering, Faculty of Engineering,
Izmir University, Karabaglar, Izmir 35350, Turkey
2
Department of Industrial Engineering, Faculty of Engineering,
Dokuz Eylul University, Buca, Izmir 35160, Turkey
3
Department of Industrial Engineering, Faculty of Engineering,
Istanbul Technical University, Macka, Istanbul 34367, Turkey
E-mail: kahramanc@itu.edu.tr

Received: July 24, 2013. Accepted: March 18, 2014.

Total Productive Maintenance (TPM) is a successful technique used for


corrective, preventive and predictive maintenance policies. It is impor-
tant in identifying the success and overall effectiveness of the manu-
facturing process for long term economic viability of business. Overall
equipment effectiveness (OEE) is commonly used and well-accepted
metric for TPM implementation in many manufacturing industries. In
this study, Adaptive Neuro-Fuzzy Inference System (ANFIS) is used to
obtain forecasted results for OEE parameter in TPM through some pre-
determined inputs such as availability, performance efficiency and rate
of quality. Triangular type of membership functions was determined as
low, medium, and high for each input parameter in the ANFIS model.
Fuzzy c-means clustering algorithm was used for determining of the
membership degrees of membership functions for each input parame-
ter. This study is important to forecast the risk by OEE in the TPM.
With the predicted results of OEE performance an appropriate mainte-
nance strategy can be developed and the production can be improved.
This can also help reducing the risk level of breakdowns or failures at
any critical equipment

Keywords: Adaptive neuro-fuzzy inference system, overall equipment effective-


ness, total productive maintenance, performance improvement of overall equip-
ment effectiveness.


Corresponding author: E-mail: mehmet.cakmakci@deu.edu.tr

535
536 E BRU T URANOGLU B EKAR et al.

1 INTRODUCTION

Nowadays, the elements of Q (quality), C (cost) and (D) delivery are becom-
ing a vital business strategy leading to success and growth in most organi-
zations. Lean production leads to many operational benefits to support the
Q, C, D element in organization by focusing on wastes elimination such as
(1) waste of rework, (2) waste of overproduction, (3) waste of wait time (4)
waste in delivery (production lead time), (5) waste in processing, (6) waste of
inventory and (7) waste of motion. The wastes generated in production have
a strong relationship with the availability of production equipment. The mal-
function and breakdown of equipment would results in poor quality products
and as a consequence delay deliveries. Thus, a systematic and strategic main-
tenance management, such as TPM, in maintaining the production’s equip-
ment is really significance to support the successful of lean production. A
typical example of applying TPM methodology to support the lean produc-
tion is on poorly maintained machines or equipment. Through a strategic
maintenance management, the defects and variations could be eliminated at
their source [2]. TPM seeks to maximize equipment effectiveness through-
out the lifetime of the equipment. The significance role of TPM element as
an important complementary to support the success of lean production was
not been comprehensively defined in quite a number of researches related to
lean production [34]. It strives to maintain the equipment in optimum con-
dition in order to prevent unexpected breakdown, speed losses, and quality
defects occurring from process activities [24]. The unexpected failures, the
down time associated with such failures, line stoppage, the loss of production
and, the higher maintenance costs are major problems in any process plant.
Measuring risk in maintenance helps in designing an alternative strategy to
minimize the risk resulting from breakdowns or failures. Therefore making
decisions concerning a selection of a maintenance strategy using a risk-based
approach is essential to develop cost effective maintenance polices for mech-
anized and automated systems [49]. TPM can be defined in terms of OEE
which in turn can be considered a combination of the operation maintenance,
equipment management, and available resources. The goal of TPM is to max-
imize equipment effectiveness, and OEE is the core metric for measuring the
success of TPM implementation program [51].
Metrics of TPM achieving or exceeding the predetermined targets presents
an improved production whereas any instability of TPM metrics is presenting
a risk of low performance in production. Hence accurate forecasting of these
metrics is significantly important for reducing the variation in the production
performance. Various approaches and models have been applied to describe
and forecast some systems. Among those, two distinct approaches are known
F ORECASTING E QUIPMENT E FFECTIVENESS 537

to be more applicable namely statistical approach, including time series mod-


els and econometric models and soft computing techniques (including as its
main elements fuzzy logic, various probabilistic-type reasoning, neural net-
works, evolutionary computation, etc.), designed to tackle imprecision and
uncertainty involved in a complex nonlinear system [19]. The evolution of
soft computing techniques has helped in understanding the various aspects
of nonlinear systems and thereby making it possible to model them as well
as predict their future response. These make their analysis and control easier
[37]. Recent reviews on soft computing around the world indicate that the
number of soft computing based engineering applications is increasing. One
popular soft computing method is neuro-fuzzy technique which is a hybrid
combination of artificial neural networks (ANN) and fuzzy inference sys-
tem [50]. Neuro-fuzzy systems combine the semantic transparency of rule-
based fuzzy systems with the learning capability of neural networks. Fur-
thermore, neuro-fuzzy modeling has been recognized as a powerful tool that
can facilitate the effective development of models by combining informa-
tion from various sources, such as empirical models, heuristics and data.
Hence, in most cases neuro-fuzzy models can be better used to explain solu-
tions to users than completely black box models such as neural networks
[40, 44]. In the recent past, a variety of neuro-fuzzy model architectures
were proposed and described in studies belonging to some researches [5, 18,
27, 31].
One of the most important architectures of neuro-fuzzy model is Adaptive
Neuro-Fuzzy Inference System (ANFIS) which has shown significant results
in modeling nonlinear functions is a new improved tool and a data driven
modeling approach for determining the behavior of imprecisely defined com-
plex dynamical systems [25]. It can construct an input–output mapping based
on both human knowledge in the form of fuzzy if-then rules with appropri-
ate membership functions and stipulated input–output data pairs. It applies a
neural network in determination of the shape of membership functions and
rule extraction [48]. It has increasingly founding diverse applications and has
gained special attention in forecasting.
ANFIS analysis is quite flexible in regards to the number and form of the
experimental data, which makes it possible to use more informal experimen-
tal designs than with statistical approaches. Table 1 shows the comparison of
ANFIS algorithm with some of the conceptual methods [1].
As seen in Table 1, ANFIS approach has fuzzy data modeling, which
requires a method (1) for transforming human knowledge or experience into
the rule base and database of a fuzzy inference system and (2) for tuning the
membership function to minimize the output error measure or maximize per-
formance index, with high precision and reliability. Also it is widely used in
538 E BRU T URANOGLU B EKAR et al.

modeling and
nonlinearity

uncertainty

forecasting
complexity

Intelligent
non-crisp
Multiple
Method

data set
Inputs

Data

Data
and

and
√ √ √ √
ANFIS Approach
√ √ √
ANN
√ √
Fuzzy Regression
√ √
Fuzzy Inference System
√ √ √
Genetic Algorithm
√ √
Decision Tree
Fuzzy data

processing

processing
Reliability

ambiguity

Data pre-

and post-
modeling

Precision
Method

Dealing
High

√ and √ √ √
ANFIS Approach

ANN
√ √ √
Fuzzy Regression
√ √
Fuzzy Inference System
√ √
Genetic Algorithm
√ √ √
Decision Tree

TABLE 1
The features of the ANFIS approach versus other methods

forecasting problems and provides strong computational power using by its


software. With this motivation, ANFIS was employed in this study.
In this study, it is aimed to forecast a risk by OEE parameter through some
predetermined inputs using an ANFIS model. The rest of the paper is orga-
nized as follows. Section 2 explains the information about the OEE and the
literature on measuring risk in TPM. Section 3 introduces the fundamentals
of ANFIS and presents a literature review on ANFIS. In Section 4, an ANFIS
model to forecast risk by OEE parameter in TPM is presented. In the last
section, results and conclusion are given.

2 PROBLEM DEFINITION

The efficient maintenance of the production and other plant machinery is cru-
cial in determining the success and overall effectiveness of the manufacturing
process. OEE can be said to be a measure of progress of TPM in an organi-
zation. Definition of OEE includes six big losses, including downtime and
other production losses that reduce output/machine hour or capacity utiliza-
tion and does not include factors that reduce capacity utilization, e.g. planned
F ORECASTING E QUIPMENT E FFECTIVENESS 539

FIGURE 1
Publication frequencies of OEE

downtime, lack of material input, lack of labor, etc. The six large losses are
given below [3]:

Downtime losses: 1) Breakdown losses categorized as time losses and quan-


tity losses caused by equipment failure or breakdown. 2) Set up and adjust-
ment losses occur when production is changing over from requirement of one
item to another.
Speed losses: 3) Idling and minor stoppage losses occur when production
is interrupted by temporary malfunction or when a machine is idling. 4)
Reduced speed losses refer to the difference between equipment design speed
and actual operating speed.
Quality losses: 5) Quality defects and rework are losses in quality caused by
malfunctioning production equipment. 6) Reduced yield during start-up are
yield losses that occur from machine start-up to stabilization.

According to literature review a lot of studies concentrated on OEE, a


well-known efficiency metric that allows evaluation of the impact of sev-
eral hidden losses, by comparing the actual performance of equipment with
respect to its theoretical potential. Figure 1 shows the publication frequen-
cies of OEE according to years between 2004 (including 2004 and earlier)
and 2013. Some of these publications are journal articles, books/e-books,
and so on. Figure 2 shows the distribution of these publications according
to publication categories. According this figure, most of the studies on OEE
are published in articles (journal, magazine and newspaper) with a rate of
83 percent. For example, Ljungberg [38] argued that it should be beneficial
to change focus and use a comprehensive model for losses and proposes a
540 E BRU T URANOGLU B EKAR et al.

FIGURE 2
Percentage of publication categories of OEE

TPM model with eight equipment losses. The author also explained that the
data collection problem has not been sufficiently treated in the literature and
has suggested a method for collecting disturbance data where computerized
systems are combined with manual recording.
Jeong and Phillip [30] explained that accurate estimation of equipment
utilization is very essential. They presented a new loss classification scheme
for computing OEE for a capital-intensive industry and provide justification
for this scheme. They also presented the methodology for designing the nec-
essary data collection system that can serve as a template for any industry.
Bamber et al. [14] explored the purpose of the OEE concept in modern oper-
ations. This paper discussed that in order to effectively address all six big
losses and hence improve OEE, cross-functional team is necessary. Cross-
functional team accordingly has the combined necessary skills and knowl-
edge of entire system of manufacture to identify correctly the practices and
activities that relate to the six big losses. Additionally, through the use of
cross-functional team, it is more likely that the responsibility and authority
to carry out improvements is gained from management. Holmgren [35] iden-
tified maintenance related losses, and their causes, in order to describe dif-
ferent deviations in the maintenance process that contributes to incidents and
accidents at the Swedish Railway. Wang [20] used Data Envelopment Analy-
sis to evaluate the efficiency score for when the utility function considers its
many attributes. A multiple linear regression model was used to formulate the
regression equation for predicting the efficiency score in terms of the TPM
award months, the number of employees, the ratio of spending TPM to sales,
the sales growth and OEE. Finally, the proposed methodology can identify
F ORECASTING E QUIPMENT E FFECTIVENESS 541

a peer group of efficient factories against which to benchmark. The actual


improvement process may involve identifying the operating practices and
procedures of the benchmark factories and engaging in re-engineering pro-
grams. Nachiappan and Anantharaman [45] expressed the importance of the
quantitative OEE analysis for the whole factory. Sarkar [11] pointed out how
six sigma methodology has been applied for process improvement consider-
ing OEE as a parameter. Raja and Kannan [42] optimized the OEE param-
eters using Evolutionary Programming using MATLAB. They analyzed the
impact of yield in calculation of OEE in casting industry and suggested to
the practicing industry to run the process with the given optimum condi-
tions. Braglia et al. [33] developed an alternative losses classification struc-
ture to divide the losses that can be directly ascribed to equipment, from the
ones that are spread in the line for engine basements production. Starting
from this losses classification structure, an approach based on OEE is devel-
oped to evaluate the criticalities and the effectiveness of the line. Results
show that it successfully highlights the progressive degradation of the ideal
cycle time, explaining it in terms of: bottleneck inefficiency, quality rate,
and synchronisation-transportation problems. Garza-Reyes et al. [26] demon-
strated the relationship between OEE and Process Capability (PC) and sug-
gested the existence of a “cut-off point” beyond which improvements in PC
have little impact on OEE. They developed a discrete-event simulation model
of a bottling line. Labib et al. [8] developed an integrated model of main-
tenance decision making by utilizing analytic hierarchy process and Fuzzy
Logic. Shahin and Attarpour [7] estimated a linear relationship between OEE
and mean time between failures in order to develop decision making grid for
making more accurate maintenance policies. Zammori et al. [21] taken into
account the stochastic nature of the OEE, an approximated procedure based
on the application of the Central Limit Theorem was presented. The results
obtained demonstrate that the stochastic OEE can help in battling variation,
for it allows one to identify the hidden losses that account for most of the vari-
ability and to estimate the impacts of potential corrective actions in terms of
both efficiency and efficacy. Relkar and Nandurkar [6] simulated a manufac-
turing scenario by using WITNESS software to identify bottleneck machine
with different combinations of mean time between failures and repair time
results into variation in output. They used MiniTab15 software for regression
analysis which establishes relation between OEE and time between failures,
repair time. The process of OEE was optimized by using response surface
methodology to identify optimized zone for maximizing output. Puvanas-
varan et al. [43] explained the inclusion of customer demand in obtaining
OEE value of any particular equipment. Besides that, the equipment without
ideal cycle time, which means those processes carried out in constant cycle
time are possibly to be evaluated with performance ratio. As a consequence,
542 E BRU T URANOGLU B EKAR et al.

the machine utilization and capacity used could be quantified and visualized
using the performance ratio data of the OEE proposed. Tsarouhas [41] car-
ried out the analysis of failure and repair data of the limoncello production
line over a period of 8 months. Descriptive statistics at machine and at line
level were computed, and also all the components of the OEE were calcu-
lated. As a conclusion in this study, the statistical analysis provided a useful
perspective and helped managers make better decisions about how to run and
improve their processes more efficiently and effectively.
It is seen that a considerable amount of literature has been published in
relation to the definition of OEE and its various applications. It is clear that
to the best knowledge of the authors, this is the first study to forecast risk by
OEE parameter in TPM using ANFIS.
The OEE was first proposed by Nakajima [47] as the key performance
indicator to support TPM and is calculated by obtaining the product of avail-
ability of the equipment, performance efficiency of the process and rate of
quality products as follows [10, 38]:

OEE = Availability (A) × Performance efficiency (E) × Rate of quality (R)


(1)
where

Loading time − Down time


Availability (A) = × 100 (2)
Loading time

Process Amaount ∗ Theoretical cycle time


Perfomance efficiency (P) = × 100
Operating time
(3)

Process Amaount − Defect amount


Rate of quality (R) = × 100 (4)
Process amount

This metric has become widely accepted as a quantitative tool essential for
measurement of productivity in manufacturing operations [23]. The OEE
measure is central to the formulation and execution of a TPM improvement
strategy [38].
Nowadays, the OEE is a widely accepted method to monitor the actual
performance of an equipment in relation to its nominal capabilities under
optimal operating conditions. Nevertheless, the computation of the OEE is
not always straightforward and many drawbacks and difficulties can still be
found in many industrial applications. Problems typically arise while collect-
ing the data required for the analysis and/or when it is necessary to adapt the
losses classification structure of the OEE to fit the requirements of a specific
F ORECASTING E QUIPMENT E FFECTIVENESS 543

industry. Due to the day-today variability of the manufacturing performances,


the daily value of the OEE tends to be excessively uneven and it is hard to
evaluate the real efficiency of equipment and to define sensible improvement
targets [21]. In this way one can easily evaluate the OEE parameter under
fuzziness. In this study among other intelligent technique, we focus on the
application of ANFIS to forecast OEE parameter in TPM.

3 METHOD
3.1 Adaptive Neuro-Fuzzy Inference System (ANFIS)
A neural-fuzzy system is a hybrid of neural networks and fuzzy systems in
such a way that neural networks or neural networks algorithms are used to
determine parameters of fuzzy system. The main intention of neural-fuzzy
approach is to improve a fuzzy system automatically by means of neural net-
work method.
In this section, it has been constructed an ANFIS model that is functionally
equivalent to a fuzzy control system with TSK (Takagi–Sugeno–Kang) fuzzy
rules whose consequents are linear combinations of their preconditions [16,
37]. For simplicity, assume that the fuzzy control system under consideration
has two inputs x1 and x2 and one output y and that rule base contains two
TSK fuzzy rules as follows:

R 1 : IF x1 is A11 AND x2 is A12 , THEN y = f 1 = a01 + a11 x1 + a21 x2 + . . . + a21 x2 ,


(5)

R 1 : IF x1 is A21 AND x2 is A22 , THEN y = f 2 = a02 + a12 x1 + a21 x2 + . . . + a22 x2 .


(6)
j
where xi an input variable, y is the output variable, Ai are linguistics terms
j
of the precondition part with membership functions μ Aij (xi ), ai ∈ R are
coefficients of linear equations f j (x1 , x2 , . . . , xn ), and j = 1, 2, . . . , M, i =
1, 2, . . . , n. For given input values x1 and x2 , the inferred output y ∗ is calcu-
lated by

μ1 f 1 + μ2 f 2
y∗ = (7)
μ1 + μ2

where μ j are firing strengths of R j , j = 1, 2, and are given by

μ j = μ A j (x1 ) × μ A j (x2 ) , j = 1, 2, (8)


1 2

The neural network structure includes six layers.


544 E BRU T URANOGLU B EKAR et al.

Layer 1: Every node in this layer is an input node that just passes external
signals to the next layer. Layer 2: Every node in this layer acts as membership
function μ Aij (xi ), and its output that specifies the degree to which the given
j
xi satisfies the quantifier Ai . Usually it is chosen μ Aij (xi ) to be bell-shaped
with a maximum equal to 1 and a minimum equal to 0, such as

1
μ Aij (xi ) =  , (9)
 2 bij
j j
1+ xi − mi /σi

⎧ ⎡   ⎞2 ⎤bij ⎫
⎪ ⎛ ⎪

⎨ xi − m i
j ⎪

⎢⎝ ⎠ ⎦⎥
μ Aij (xi ) = exp − ⎣ , (10)

⎪ σi
j ⎪

⎩ ⎭

 
j j j
where m i , σi , bi is the parameter set to be tuned. Parameters in this layer
are referred as precondition parameters. Layer 3: Every node in this layer is
labeled as  and multiples the incoming signals μ j = μ A j (x1 ) × μ A j (x2 ),
1 2
and sends the product output. Each node output represents the firing strength
of a rule. Layer 4: Every node in this layer is labeled as N and calculates the
normalized firing strength of a rule. That is, the jth node calculates the ratio
of the jth rule’s firing strength to the sum of all the rules’ firing strengths,

μ j = μ j /μ A j (x1 ) + μ A j (x2 ) . (11)


1 2

Layer
 5: Every j node in this layer calculates the weighted consequent
 value
j j j j j j
μ j a0 + a1 x1 + a2 x2 , where μ j is the output of layer 4 and a0 , a1 , a2
is the parameter set to be tuned. Parameters in this layer are referred as con-
sequent parameters. Layer 6: The only node in this layer is labeled as , and
it sums all incoming signals to obtain the final inferred result for the whole
system.

3.2 Literature Review on ANFIS


In the literature, a series of work can be explored that proposed different
ANFIS based models for forecasting. ANFIS is generally used forecasting
of energy [9, 13, 29], and stock market [22]. In maintenance, Garcı́a et al.
[15] designed a real-time monitoring system for Condition Based Mainte-
nance using a forecasting model based in ANFIS combined with a Gray-
Scale Health Index (HI) was implemented to evaluate a conveyor belt system.
Kaur et al. [4] aimed to evaluate and compare the application of different soft
computing techniques-ANN and ANFIS to construct models for prediction
F ORECASTING E QUIPMENT E FFECTIVENESS 545

of Software Maintenance Effort. The maintenance effort data of two com-


mercial software products was used in this study, and they illustrated that
ANFIS technique gives out the most accurate model. According to the lit-
erature review, there are a few studies in maintenance using ANFIS. To the
best knowledge of the authors, this is the first study to forecast risk by OEE
parameter in TPM using ANFIS.

4 ANFIS MODEL FORECASTING RISK BY OEE PARAMETER


IN TPM

In this section, a case study is presented to illustrate the detailed applica-


tions of the proposed ANFIS model. ANFIS modeling and prediction of OEE
parameter in TPM starts by obtaining a data set (input-output data points).
The data set shown in Figure 3 and 4 is taken from a single production line
from a company operating in the automotive industry. In this study, ANFIS
is used to obtain forecasting results for the OEE parameter in TPM through
some predetermined inputs, which are availability, performance efficiency
and rate of quality. Accurate forecasting of OEE reduces uncertainty of equip-
ment efficiency in the production. According to predetermined OEE perfor-
mance, an appropriate maintenance strategy can be designed and this can help
reducing the risk level of performance at any critical equipment. The overall
framework of forecasting OEE in TPM is shown as Figure 5.
ANFIS Toolbox of MATLAB 7.11 was used for model simulations. For
improved performance, all variables were normalized into a smaller range
through Min-max normalization method. This method scales numbers in a
data set to improve accuracy of subsequent numeric computations. Let X old ,
X max and X min be main value, maximum and minimum of each input variable
data, respectively, and X max , X min be maximum and minimum of normalized

FIGURE 3
The input data set obtained from a company operating in the automotive industry
546 E BRU T URANOGLU B EKAR et al.

FIGURE 4
The output data obtained from a company operating in the automotive industry

data, respectively, then normalization of X old called X new can be obtained as


 
 xold − xmin   
xnew = (xmax − xmin ) + xmin (12)
xmax − xmin

In this system, the Takagi–Sugeno fuzzy inference system (TS-FIS) pro-


posed by Takagi and Sugeno was built with inputs - availability, performance

FIGURE 5
The overall framework of forecasting OEE in TPM
F ORECASTING E QUIPMENT E FFECTIVENESS 547

efficiency, and quality rate. Triangular type of membership functions was


determined for each input parameter in the ANFIS system. Fuzzy c-
means (FCM) clustering algorithm was used for determining the member-
ship degrees of membership functions of inputs. They are introduced and
expressed by fuzzy linguistic values such as ‘low’, ‘medium’ and ‘high’, and
also 9 fuzzy rules were used to build the TS-FIS for modeling the forecasting
OEE parameter in TPM.
TS-FIS model can be seen as a combination of linguistic and mathematical
regression modelling in the sense that the antecedents describe fuzzy regions
in the input space in which the consequent functions are valid. It also can be
regarded as a quasi-linear system (i.e., a linear system with input-dependent
parameters). In this sense, a TS-FIS model can be seen as a mapping from
the antecedent (input) space to a convex region (polytope) in the space of
the parameters of a quasi-linear system [44]. Therefore, in this study, as per
requirement of TS- FIS based ANFIS method, only input space is defined as a
triangular membership function with low, medium and high fuzzy linguistic
variables whereas the output space is defined as a linear equation for each
fuzzy if-then rules. So this definition is a result of the method preferred to be
used. And, it facilitates the analysis of TS-FIS model in a framework similar
to that of linear system and also provides a numerical simplicity.
After constitution of TS-FIS model, the data belonging to the period
between first week and last week of the year 2012 was used as the train-
ing and testing part of the ANFIS model and the structure of ANFIS model
is given in Figure 6. Also Table 1 gives the parameters for ANFIS model for
OEE parameter in TPM.

FIGURE 6
The structure of ANFIS model for OEE parameter in TPM
548 E BRU T URANOGLU B EKAR et al.

Architecture of Takagi–Sugeno fuzzy inference system


Membership Functions Type for Input Triangular Membership Function (Trimf)
Membership Functions Type for Output Linear
Fuzzy rules 9
Training of Takagi–Sugeno fuzzy inference system
ANFIS parameter optimization method Hybrid (a mixture of backpropagation and
least squares)
Error tolerance 0
Epochs 40

TABLE 2
Parameters of ANFIS model for OEE parameter in TPM

In Table 2, it is seen that ANFIS uses a hybrid learning algorithm that


combines the gradient method and the least squares estimate (LSE) to identify
the membership function parameters of TS-FIS. If the adaptive network under
consideration has only one output, then

output = F(I, S) (13)

where I is the set of input variables and S is the set of parameters. If there
exists a function H such that the composite function H ◦ F is linear in some
of the elements of S, then these elements can be identified by the least squares
method. More formally, if the parameter set S can be decomposed into two
sets

S = S1 ⊕ S2 (14)

where ⊕ represents sum such that H ◦ F is linear in the elements of S2 then


upon applying H to Equation (13), we have;

H (output) = H ◦ F( I , S) (15)

which is linear in the elements of S2 . Now given values of elements of S1


training data which can be denoted by P, can be input into Equation (15) and
a matrix equation obtained:

AX = B (16)

where X is an unknown vector whose elements are parameters in S2 . Let


|S2 | = M, then the dimensions of A, X and B are P × M, M × 1 and P ×
1, respectively. Since P, the number of training data pairs is usually greater
F ORECASTING E QUIPMENT E FFECTIVENESS 549

than M, the number of linear parameters, this is an overdetermined problem,


and generally there is no exact solution to Equation (16). Instead, LSE of X
and X ∗ is sought to minimize the squared error AX − B 2 .The most well-
known formula for X ∗ uses the pseudo-inverse of X [39].

X * = (A T A)−1 A T B (17)

While Equation (17) is concise in notation, it is expensive in computation


when dealing with the matrix inverse and, moreover, it becomes ill defined if
A T A is singular. As a result, sequential formulas are employed to compute
the LSE of X. This sequential method of LSE is more efficient, especially
when M is small and can be easily modified to an on-line version i.e., for
systems with changing characteristics. Specifically, let the ith row vector of
matrix A defined in Equation (16) be aiT and the ith element of B be biT , then
X can be calculated iteratively using the following sequential formulas in [12,
39];
 T 
X i+t = X i + S i+1 ai+1 bi+1 − ai+1
T
Xi (18)

Si ai + 1ai+1
T
Xi
Si+1 = Si − (19)
1 + a i+a
T
Si ai+1

for i = 0, 1, ..., P − 1, where Si is the covariance matrix, and the least


squares estimate X ∗ is equal to X p . The initial conditions to bootstrap Equa-
tions (18) and (19) are X 0 = 0 and S0 = γ I, where γ is a positive large
number and I is the identity matrix of dimension M × M. When dealing with
multi-output adaptive networks, output in Equation (13) is a column vector
and Equations (18) and (18) still apply except that biT is the i th row matrix
of B.
The gradient method and the least squares estimate can be combined to
update the parameters in an adaptive network. Each epoch of this hybrid
learning procedure is composed of a forward pass and a backward pass. In
the forward pass, input data is supplied and functional signals go forward to
calculate each node’s output until the matrices A and B in Equation (16) are
obtained. The parameters in S2 are identified by the sequential least squares
formulas in Equations (18) and (19). After identifying parameters in S2 the
functional signals keep going forward until the error measure is obtained.
In the backward pass, the error rates or the differential error measure with
respect to each node output, propagate from the output end toward the input
end, and the parameters in S1 are updated by the gradient descent method.
For given fixed values of parameters in S1 , the parameters in thus found are
550 E BRU T URANOGLU B EKAR et al.

Weeks Actual OEE Predicted OEE


39 92.3695 91.5057
40 91.5661 90.5931
41 71.8543 69.0095
42 89.5616 88.6891
43 75.4250 75.6843
44 87.8045 87.4252
45 57.0996 56.5158
46 76.6395 75.3279
47 100.6252 100.2157
48 109.4822 106.2256
49 98.0915 97.0566
50 90.6835 89.7724
51 93.0372 92.4274
52 96.7736 96.5044

TABLE 3
Actual and predicted OEE parameters for the weeks between 39 and 52

guaranteed to be the global optimum point in the S2 parameter space due to


the choice of the squared error measure. Not only does this hybrid learning
rule decrease the dimension of the search space in the gradient method, but,
in general, it also cuts down substantially the convergence time.
In this study the employed training errors are the mean squared error
(MSE) of the training data set at each epoch. If Yt is the actual observation for
time period t and Ft is the forecast for the same period, then MSE is defined
as in Equation (20) [46].

1
n
MSE = (Yt − Ft )2 (20)
n t=1

The ANFIS model for OEE parameter in TPM was trained by selecting a
certain number of data points. For the whole data set, the MSE was found
0.6336%.
After training, the trained network is used to test with the rest of 13 sam-
ples. Table 3 gives the actual and the predicted OEE for the weeks between
39 and 52.
According Table 3, the performance of the ANFIS model for the test set,
which is the absolute mean square error, was found to be 1.8526%. It is sig-
nificantly seen that ANFIS structure gives closer forecasted values than the
actual values. This value demonstrates that ANFIS model can be used to fore-
cast OEE parameter with an efficient performance in TPM. Accurate fore-
casting of OEE parameter in TPM and achieving the target value of OEE are
F ORECASTING E QUIPMENT E FFECTIVENESS 551

important to maintain the equipment availability, the quality of the product


and the maintenance strategy in a systematic way and to reduce level of the
risk in production process.

5 CONCLUSION

The wastes generated in production have a strong relationship with the avail-
ability of production equipment. The malfunction and breakdown of equip-
ment would results in poor quality products and as a consequence delay deliv-
eries. Thus, a systematic and strategic maintenance management, such as
TPM, in maintaining the production’s equipment is really significance to sup-
port the successful of lean production. TPM implementation could eliminate
the defects and variations at their source and increase productivity of manu-
facturing and production lines. TPM employs OEE as a quantitative metric
for measuring the performance of a productive system. The OEE measure-
ment applies equally well in both batch/discrete and process plants, and pro-
vides good guidance as to which problem represents the greatest opportunity
for improvement and the elimination of potential waste and reducing of risk
level. In this study, it is illustrated that forecasting risk by OEE parameter in
TPM using ANFIS which takes the advantage of fuzzy inference capacity and
learning function. To the best knowledge of the authors, there is not a recent
work on the integration of current TPM approach with Artificial Intelligence
techniques. The contribution of this study is to provide new approaches to
TPM by using ANFIS. Future research should focus on the development of
new parameters in TPM taking into consideration results of the study.

REFERENCES

[1] A. Azadeh, M. Saberi and S.M. Asadzadeh, An adaptive network based fuzzy inference
system-auto regression-analysis of variance algorithm for improvement of oil consumption
estimation and policy making: the cases of Canada, United Kingdom, and South Korea,
Applied Mathematical Modelling, 35(2) (2011), 581–593.
[2] A. H. Bakri, A.R.A. Rahim, N.M. Yusof and R. Ahmad, Boosting Lean Production via
TPM, Procedia - Social and Behavioral Sciences, 65(2012), 485–491.
[3] A.J. De Ron and J.E. Rooda, OEE and equipment effectiveness: an evaluation, Interna-
tional Journal of Production Research, 44(23) (2006), 4987–5003.
[4] A. Kaur, K. Kaur and R. Malhotra, Soft Computing Approaches for Prediction of Software
Maintenance Effort, International Journal of Computer Applications, 1(15) (2010), 80–86.
[5] A. K. Palit and D. Popovic, Nonlinear combination of forecasts using artificial neural net-
work, fuzzy logic and neuro-fuzzy approaches, In The ninth IEEE international conference
on fuzzy systems, (2000), 566–571.
552 E BRU T URANOGLU B EKAR et al.

[6] A.S. Relkar and K.N. Nandurkar, Optimizing Overall equipment effectiveness through
simulation analysis, 2nd Annual International Conference on Operations Research and
Statistics (ORS 2012) (2012).
[7] A. Shahin and M R. Attarpour, Developing Decision Making Grid for Maintenance Policy
Making Based on Estimated Range of Overall Equipment Effectiveness, Modern Applied
Science, 5(6) (2011).
[8] A.W. Labib, G.B. Williams and R.F. O’Connor, An intelligent decision analysis mainte-
nance system: application of AHP and fuzzy logic, Proceedings of The Fourth International
Symposium on AHP, Vancouver, Canada, 12–16 July (1996).
[9] B. Akdemir and N. Çetinkaya, Long-term load forecasting based on adaptive neural fuzzy
inference system using real energy data, Energy Procedia, 14 (2012), 794–799.
[10] B. Dal, P. Tugwell and R. Greatbanks, Overall equipment effectiveness as a measure for
operational improvement: a practical analysis, International Journal of Operations & Pro-
duction Management, 20(11) (2000), 1488–502.
[11] B.N. Sarkar, Capability enhancement of a casting process in a small steel foundry through
six sigma: A case study, International Journal of Six Sigma and Competitive Advantage, 3
(2007), 56–71.
[12] C.G. Graham, Adaptive filtering prediction and control, Prentice-Hall (Englewood Cliffs,
N.J), 1984.
[13] C.H. Cheng and L.Y. Wei, One step-ahead ANFIS time series model for forecasting elec-
tricity loads, Optim. Eng., 11(2) (2010), 303–317.
[14] C.J. Bamber, P. Castka, J.M. Sharp and Y. Motara, Cross-functional team working for
overall equipment effectiveness (OEE), Journal of Quality in Maintenance Engineering,
9(3) (2003), 223–238.
[15] C.M. Garcı́a, T. Escobet and J. Quevedo, PHM Techniques for Condition-Based Main-
tenance Based on Hybrid System Model Representation, Annual Conference of the Prog-
nostics and Health Management Society (2010).
[16] C-T. Lin and C.S. George Lee, Neural fuzzy systems: A Neuro-Fuzzy Synergism to Intel-
ligent Systems, Prentice Hall PTR, 1996.
[17] D. Nauck, F. Klawonn and R. Kruse, Foundations of Neuro-Fuzzy Systems, John Wiley
& Sons Ltd, England, 1997.
[18] D. Nauck and R. Kruse, Neuro-fuzzy systems for function approximation, Fuzzy Sets and
Systems, 101 (1999), 261–271.
[19] D. Ruan, J. Kacpryzk and M. Fedrizzi, Soft computing for risk evaluation and man-
agement applications in technology, environment and finance, Physica–Verlag Heidelberg
(2001), New York.
[20] F.K. Wang, Evaluating the efficiency of implementing total productive maintenance, Total
Quality Management & Business Excellence, 17(5) (2006), 655–667.
[21] F. Zammori, M. Braglia and M. Frosolini, Stochastic overall equipment effectiveness,
International Journal of Production Research, 49(21) (2011), 6469–6490.
[22] G.S. Atsalakis and K.P. Valavanis, Forecasting stock market short-term trends using a
neuro-fuzzy based methodology, Expert Systems with Applications, 36 (2009), 10696–
10707.
[23] H.H. Samuel, P.D. John J. Shi and S. Qi, Manufacturing system modeling for productivity
improvement, Journal of Manufacturing Systems, 21(4) (2002), 249–60.
[24] I.P.S. Ahuja and J.S. Khamba, Total productive maintenance: Literature review and direc-
tions, International Journal of Quality and Reliability Management, 25(7) (2008), 709–
756.
F ORECASTING E QUIPMENT E FFECTIVENESS 553

[25] J. Kim and N. Kasabov, ANFIS: Adaptive neuro-fuzzy inference systems and their appli-
cation to nonlinear dynamical systems, Neural Networks, 12 (1999), 1301–1319.
[26] J.A. Garza-Reyes, S. Eldridge, K.D. Barber and H. Soriano-Meier, Overall equipment
effectiveness (OEE) and process capability (PC) measures A relationship analysis, Inter-
national Journal of Quality & Reliability Management, 27(1) (2010), 48–62.
[27] J. S. R. Jang, ANFIS: Adaptive network based fuzzy inference system, IEEE Transactions
on Systems, Man, and Cybernetics, 23(3) (1993), 665–684.
[28] J. S. R. Roger, Fuzzy modeling using generalized neural networks and Kalman filter algo-
rithm, In Proceedings of the Ninth National Coference on Artificial Intelligence, 1991.
[29] K. Lia, H. Sua and J. Chua, Forecasting building energy consumption using neural net-
works and hybrid neuro-fuzzy system: A comparative study, Energy and Buildings, 43
(2011), 2893–2899.
[30] K.Y. Jeong and D.T. Phillips, Operational efficiency and effectiveness measurement, Inter-
national Journal of Operations & Production Management, 21(10) (2001), 1404–1416.
[31] L. X. Wang, Adaptive fuzzy systems and control: Design and stability analysis, Engle-
wood Cliffs, NJ: Prentice Hall, 1994.
[32] M. Bertolini, M. Bevilacqua, F.E. Ciarapica and G. Giacchetta, Development of Risk-
Based Inspection and Maintenance procedures for an oil refinery, Journal of Loss Preven-
tion in the Process Industries, 22 (2009), 244–253.
[33] M. Braglia, M. Frosolini and F. Zammori, Overall equipment effectiveness of a manu-
facturing line (OEEL), an integrated approach to assess systems performance, Journal of
Manufacturing Technology Management, 20(1) (2008), 8–29.
[34] M. Hedelind and M. Jackson, How to improve the use of industrial robots in lean manu-
facturing systems, Journal of Manufacturing Technology Management, 22(7) (2011), 891–
905.
[35] M. Holmgren, Maintenance-related losses at the Swedish Rail, Journal of Quality in
Maintenance Engineering, 11(1) (2005), 5–18.
[36] M. Sugeno and G. T. Kang, Structure Identification of Fuzzy Model, Fuzzy Sets and Sys-
tems, 28 (1988), 15–30.
[37] O. Kayak and L.A Zadeh et al. (Eds.), Fuzzy Inference Systems: A Critical Review, Com-
putatioanl Intelligence: Soft Computing and Fuzzy-Neuro Integration with Applications,
Springer, Berlin, 1998.
[38] O. Ljungberg, Measurement of overall equipment effectiveness as a basic for TPM activi-
ties, International Journal of Operations and Production Management, 18(5) (1998), 495–
507.
[39] P.B. Piero, Adaptive neural fuzzy inference systems (ANFIS): Analysis and applications,
Lecture notes, 2000.
[40] P. C. Panchariya, A.K. Palit, D. Popovic and A. L. Sharma, Nonlinear system identifica-
tion using Takagi–Sugeno type neuro-fuzzy model, In Second international IEEE confer-
ence on intelligent systems, (2004), 76–81.
[41] P.H. Tsarouhas, Evaluation of overall equipment effectiveness in the beverage industry: a
case study, International Journal of Production Research, 51(2) (2013), 515–523.
[42] P.N. Raja and S.M. Kannan, Evolutionary programming to improve yield and overall
equipment effectiveness of casting industry, Journal of Engineering and Applied Sciences,
2(11) (2007), 1735–1742.
[43] P. Puvanasvaran, Y.S. Teoh and C.C. Tay, Consideration of demand rate in Overall Equip-
ment Effectiveness (OEE) on equipment with constant process time, Journal of Industrial
Engineering and Management, 6(2) (2013), 507–524.
554 E BRU T URANOGLU B EKAR et al.

[44] R. Babuška and H. Verbruggen, Neuro-fuzzy methods for linear system identification,
Annual Reviews in Control, 27 (2003), 73–85.
[45] R.M. Nachiappan and N. Anantharaman, Evaluation of overall line effectiveness (OLE)
in a continuous product line manufacturing system. Journal of Manufacturing Technology
Management, 17(7) (2006), 987–1008.
[46] S. Makridakis, S. C. Wheelwright and R. J. Hyndman, Forecasting: Methods and appli-
cations (3rd ed.), New York: John Wiley & Sons, 1998.
[47] S. Nakajima, Introduction to TPM, Productivity Press, Cambridge, MA (1988).
[48] T. Efendigil, S. Önüt and C. Kahraman, A decision support system for demand forecasting
with artificial neural networks and neuro-fuzzy models: A comparative analysis, Experts
Systems with Applications, 36 (2009), 6697–6707.
[49] U. Kumar, Maintenance strategies for mechanized and automated mining systems; a reli-
ability and risk analysis based approach, Journal of Mines, Metals and Fuels, Annual
Review (1998), 343–347.
[50] Y. Dote and S.J. Ovaska, Industrial applications of soft computing: a review, Proc. IEEE,
(2001), 1243–1265.
[51] Z. Juric, A.I. Sanchez and A. Goti, Money-based overall equipment effectiveness, Hydro-
carbon Processing, 85(5) (2006), 43–50.
Copyright of Journal of Multiple-Valued Logic & Soft Computing is the property of Old City
Publishing, Inc. and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.

You might also like