Journal of Industrial and Systems Engineering

Vol. 1, No. 3, pp 235-250

Fall 2007

A One-Stage Two-Machine Replacement Strategy Based on the Bayesian

Inference Method
Mohammad Saber Fallah Nezhad1, Seyed Taghi Akhavan Niaki2*, Abdolhamid Eshragh-
Jahromi3

Department of Industrial Engineering, Sharif University of Technology

P.O. Box 11365-9414, Azadi Ave., Tehran, Iran
1
[email protected]

2
[email protected]

3
[email protected]

ABSTRACT

In this research, we consider an application of the Bayesian Inferences in machine replacement

problem. The application is concerned with the time to replace two machines producing a
specific product; each machine doing a special operation on the product when there are
manufacturing defects because of failures. A common practice for this kind of problem is to fit
a single distribution to the combined defect data, usually a distribution with an increasing
hazard rate. While this may be convenient, it does not adequately capture the fact that there are
two different underlying causes of failures. A better approach is to view the defect as arising
from a mixture population: one due to the first machine failures and the other due to the second
one. This allows one to estimate the various parameters of interest including the mixture
proportion and the distribution of time between productions of defective products for each
machine, separately. To do this, first we briefly introduce the data augmentation method for
Bayesian inferences in the context of the finite mixture models. Then, we discuss the analysis of
time-to-failure data and propose an optimal decision-making procedure for machine
replacement strategy. In order to demonstrate the application of the proposed method we
provide a numerical example.

Keywords: Reliability; Weibull distribution; Preventive maintenance; Bayesian inference

1. INTRODUCTION AND LITERATURE REVIEW

Preventive maintenance (PM) involves the repair, replacement, and maintenance of equipments in
order to avoid unexpected failure during use. The objective of any PM program is the minimization
of the total cost of inspection, repair, and equipment downtime (measured in terms of lost
production capacity or reduced product quality) (Mann et al. 1995).

*
Corresponding Author
236 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

In order to perform PM two approaches have evolved in the literature. The traditional approach is
based on the use of statistical and reliability analysis of equipment failure. Under statistical-
reliability (S-R)-based PM, the objective of achieving the minimum total cost is pursued by
establishing fixed and statistically optimal PM intervals, at which to replace or overhaul equipment
or components. The second approach involves the use of sensor-based monitoring of equipment
condition in order to predict the time of machine failure. Under condition-based (C-B) PM, intervals
between PM work are no longer fixed, but are performed only “when needed” (Mann et al. 1995).

The primary disadvantage of (S-R)-based PM is that the results of the calculations are based on the
use of the mean value as the measure of central tendency. If the standard deviations of these means
are large, then the probability of ascertaining the maintenance interval with accuracy is small. In
many of these cases, the plant is over-maintained. Other disadvantages include more emergency
maintenance, more overtime, and less equipment utilization (Mann et al. 1995).

With the development of a condition-based maintenance (CBM) technique, a more dynamic

preventive maintenance practice could be applied. By integrating prediction tools, CBM can
determine the required maintenance action prior to any predicted failure based on the conditions
observed prior to a previous failure. From this aspect, this technique can be called condition-based
predictive maintenance (CBPM) (Zhou et al. 2006). It has been proven that CBPM is an effective
way to minimize maintenance costs, improve operational safety, and reduce the frequency and
severity of in-service system failures (Zhou et al. 2006 and Mobley 1989).

CBM is carried out in response to significant deterioration in a unit’s condition or performance as

indicated by a change in a monitored parameter. PM allows the machine to be taken off-line at a
predetermined time, which allows production loss to be minimized by scheduling production
around the down time (Saranga 2002).

CBM techniques can be classified according to the type of symptoms they are designed to detect.
According to Moubray (1990) the classifications are:

• Dynamic effects, such as vibration and noise levels;

• Particles released into the environment;
• Chemicals released into the environment;
• Physical effects, such as cracks, fractures, wear and deformation;
• Temperature rise in the equipment;
• Electrical effects, such as resistance, conductivity, dielectric strength, etc.

CBM has been widely accepted in practice in the past few years since it enables maintenance
decisions to be made based on the current state of the equipment, thus avoiding unnecessary
maintenance (replacement) and hence making timely maintenance actions when there is a strong
indication of impending failure (Jardine et al. 1997).

The available literature on discrete time maintenance models predominantly treats an equipment
deterioration process as a Markov chain. Sherwin and Al-Najjar (1999) presented a Markov model
to determine the inspection intervals for a phased deterioration monitored complex components in a
system with severe down time costs. An example involved roller bearing in paper mills with three
phases; no defect, possible defect and final deterioration towards failure. In the last phase,
continuous monitoring was used. The output of the model was an optimum inspection rate for each
phase given a switching rule for going over to continuous monitoring. Wang and Hwang (2004)
A One-Stage Two Machines Replacement Strategy … 237

presented a Markov model that could be applied to construct the relationships among maintenance
cycle, maintenance personnel allocation, human recovery factor, and system’s tolerance time. Zhou
et al. (2006) presented a dynamic opportunistic condition-based predictive maintenance policy for a
continuously monitored multi-unit series system that was proposed based on short-term
optimization with the integration of imperfect effect into maintenance actions. In their research, it
was assumed that a unit’s hazard rate distribution in the current maintenance cycle could be directly
derived through CBPM. Whenever one of the units fails or reaches its reliability threshold, the
whole system has to stop and PM opportunities arise for the system units. Jardine et al. (1997)
presented an optimal replacement policy based on Markov stochastic process. Gupta and Lawsirirat
(2006) presented a simulation based optimization method for strategically optimum maintenance of
monitoring-enabled multi-component systems using continuous-time jump deterioration models.
Sherwin (1999) with the concept of opportunity maintenance suggests new ways to construct and
update preventive schedules for a complex system by making better use of system failure down
time to do preventive work. Sinuany-Stern et al. (1997) concentrated on the 2-action version of this
preventive schedules problem. They suggested an extremely practicable decision rule in partial
observability, and proved empirically that this rule more than satisfactory competes with the state-
of-the-art generic algorithm when implemented with its recommended grid usage. Sinuany-Stern
(1993) considered a production system (machine) which deteriorates over time and the system
deterioration over time was assumed to be Markovian. Moreover, the time scale assumed discrete
and the ‘true’ state of the system (excellent, medium and bad) was not directly observable. What is
observed was the performance of the system measured in terms of ‘number of defectives’ per time
period. At the end of each period, a decision was to be made: whether to replace the system or not
and the objective was to minimize the total cost in the long run.

Some authors applied Bayesian inference approach in machine replacement strategy. Mazzuchi and
Soyer (1996) proposed a Bayesian approach to machine replacement problem in which they tried to
minimize the total maintenance costs per unit time, TC (T ) , including the cost of doing a planned
preventive maintenance and the cost of in-service replacement. The objective function of their
research was defined as:

T
Cf ∫ m (x )dx + C p

TC (T ) = 0
T

Where C p was the cost of performing a planned preventive maintenance, C f was the cost of in-
service replacement, and m (x ) was the hazard density function of the time to failure. Merrick et al.
(2003) presented a Bayesian semi-parametric proportional hazard model to describe the failure
behavior of machine tools. The semi-parametric setup was introduced using a mixture of Dirichlet
process prior.

The amount of literature on single item maintenance is enormous. In one hand, when the parameters
of the failure distribution are unknown but constant, Valdez-Florez and Feldman (1989) performed
an extensive review. On the other hand, Wilson and Popova (1998) considered situations in which
the parameters are random variables and Bayesian parametric analysis was performed.

While in most of the published research an infinite time horizon is assumed (see Chen and Popova
(2000) for Bayesian policies over a finite time horizon), Damien et al. (2007) considered the
problem of a finite horizon single item maintenance optimization structured as a combination of
238 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

preventive and corrective maintenance in a nuclear power plant environment. They presented
Bayesian semi-parametric models to estimate the failure time distribution and costs involved. The
objective function of the optimization problem was the expected total cost of maintenance over the
pre-defined finite time horizon. Typically, the mathematical modeling of failure times was based on
parametric models.

Sethi et al. (2000) considered a single machine, multiproduct manufacturing system, operating at
finitely many quality levels, in which the quality of the machine deteriorates according to a
continuous-time Markov process. They assumed that the only way to improve the quality is to
replace the machine by a new one and derived some conditions on the stability of the system under
a simple class of scheduling-replacement policy named the recurrence of the total work backlog.

Hamada et al. (2004) presented a fully Bayesian approach that simultaneously combined non-
overlapping (in time) basic event and higher-level event failure data in fault tree quantification.
Such higher-level data often correspond to train, subsystem, or system failure events. The fully
Bayesian approach also automatically propagates the highest-level data to lower levels in the fault
tree.

Childress and Durango-Cohen (2005) formulated a stochastic version of the parallel machine
replacement problem and analyzed the structure of optimal policies under general classes of
replacement cost functions.

Hritonenko et al. (2007) combined known continuous- and discrete-time models of equipment
replacement. They showed that the optimal equipment lifetime was shorter when the embodied
technological change was more intense.

2. PROBLEM STATEMENT

In this research, we employ the concept of condition monitoring in machine replacement problem.
Similar to Sinuany-Stern (1993) research, we observe the true state of the machines indirectly by
measuring their performance in terms of ‘number of defective items’ per time period assuming that
the machines are close to “wear-out” period such that the hazard rate is increasing. Furthermore,
while in machine replacement strategies, a usual practice is to assume that the state space of the
decision system is a discrete one, in this paper we relax this assumption and consider continuous
state spaces with Weibull probability distribution. By this assumption since the complexity of the
decision-making problem increases, we will consider only one stage. Moreover, in order to define a
continuous state space, we employ the idea of the finite mixture models. Finite mixture models arise
in a variety of industrial and scientific applications. In theses applications, we are concerned with
the time to replace two machines that produce a specific product and each machine doing a specific
operation when there failures due to manufacturing defects. A common practice to model this
problem is to fit a single distribution, usually a distribution with an increasing hazard rate, to the
combined defect data. Increasing hazard rate is due to the assumption of the machines being close to
wear-out period according to “bathtub model.” In the bathtub model, the failure rate, m (.) ,
decreases with age or usage when the equipment is new – the time during which this applies is
known as the infant mortality period – and it is usually relatively short. This is followed by a period
of constant failure rate (the useful life period), and then a “wear-out” period. The latter is
characterized by a failure rate, which is an increasing function of age or usage (Tsang, 1995). While
this approach may be convenient, it does not adequately capture the fact that there are two different
underlying causes of defect. A better approach is to view the defect as arising from a mixture
population: one due to the first machine failures and the other due to the second machine failures.
A One-Stage Two Machines Replacement Strategy … 239

This allows one to estimate the various parameters of interest including the mixture proportion and
the machine reliability distribution, separately.

The main difference between this research and the previous ones is the choice of the state variable.
By analyzing the outcome of the machines (products) and based upon the quality of the products,
we try to determine the optimal strategy using Bayesian inference. The optimal strategy in previous
researches was obtained based on the machine failure data.

In section three, we present the assumptions and some of derivations required in the new approach.
We discuss the ways to analyze time-to-failure data in section four. Notations are defined in section
five. The application of the Bayesian inference in decision-making method is given in section six.
In order to demonstrate the application of the proposed method we provide a numerical example in
section seven. A sensitivity analysis comes in section eight, and finally, the conclusion comes in
section nine.

3. ASSUMPTION AND DERIVATIONS

The assumptions involved in the proposed methodology are:

1. The horizon of the decision-making process is one stage.

2. The time between producing defective products, ( t ), follows a Weibull probability distribution
with parameters of α and β as m t ( α , β ) = αβ t β −1 .

3. An appropriate prior for α , Q (α ) , is assumed to be a Gamma distribution with parameters of

ba a−1 −bα
a and b given by Q(α) = α e .
Γ(a)

4. For the prior distribution of the shape parameter, β , it is convenient to define a discrete
distribution by discretization of the beta density on ( β L , βU ) (Mazzuchi and Soyer, 1996). This
allows for great flexibility in representing prior uncertainty. The beta density is given by

Γ (c + d ) ( β − βL ) ( βU − β )
c −1 d −1

g (β ) = for 0 ≤ βL ≤ β ≤ βU
Γ (c ) Γ ( d ) ( βU − βL )
c +d −1

where βU , β L , c , d > 0 are the specified parameters. Then, we define the prior distribution for β as

β l +δ 2
PL = Pr {β = β l } = ∫ g ( β )d β
β l −δ 2

2l − 1 β − βL
where β l = β L + δ and δ = U for l = 1, 2,..., k
2 k
240 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

5. If t i represents the time between production of defective products then, using likelihood method,
at the current time t B we obtain the posterior state of variable α and β l , denoted by α and β 1 , as
(Mazzuchi and Soyer, 1996):

(
f α β l , t1 , t 2 ,..., t n = Gamma ( a * , b * ) )
βl
n
⎛ n
⎞
Where a * = a + n and b * = b + ∑
i =1
t i
βl
+ ⎜ B ∑t i ⎟
⎝
t −
i =1 ⎠

And

{∏ t }
β l −1
(b )
*
* a
β n n

{ } j =1 j
l
Pr β l t1 , t2 ,..., tn = Pl
∑ P β {∏ t }
β i −1
(b )
k *
n n * a
i i j =1 j
i =1

6. We assume that when the time between productions of defective products is less than a threshold
like T * , then the machine has failed and should be replaced. This has practical applications in many
industries like the press shops when the evaluation of the machines is based on the quality of the
products.

7. We consider three types of costs in the model:

a. Replacement cost

b. Cost of defective products in a stage

c. Cost of system failure (production break down)

8. Assuming k(t) to be the distribution function of time between production of defective products,
then, according to assumption 5, the probability of machine failure using Bayesian inference is:

f ( t ) dt = ∫0
*

∫ f ( t α , β ) g (α , β ) dα d β dt
∞

∫
T T
Probability of machine failure =
0 0

k
where g (α , β ) = ∑ Q (α β ) Pr {β } . Substituting for the terms and using integration we have:
i =1
i i

* *
k
*a
T* b *a
∑ Pr {β } ∫
k

∑ αβ t β −1e
b βi
* *

∫ f (t ) dt = ∫ Pr {β i } d αdt = αβ i t βi −1e −αt α a −1e −b α d α dt

T T βi * * * *
−α t a −1 −b α
α
( )
i

( )
e i
Γ a
i *
0 0
i =1 Γ a i =1
0 *
(1)
T*
⎛ ⎞
k
−1 k
* ⎜ ⎟
∑ ∑
1 1
Pr {β i } a *b *a Pr {βi } a *b *a ⎜ * −
*
= = ⎟
(t ) ( )
a* *a a*
βi β
i =1 +b* i =1 ⎜ b T * i
+b* ⎟
0 ⎝ ⎠
A One-Stage Two Machines Replacement Strategy … 241

9. Assuming an overall time in decision making stage to be H, according to failure function, the
expected number of defective products in the stage can be calculated by Bayesian inference as:

The expected number of defective products = ∫ m (t ) dt = ∫ m (t α , β ) g (α , β ) d α d β dt

∞

∫
H H

0 0 0

Hence we have
∞ k
∫ ∫ ∑αβ t k (α β i ) Pr {β i } d α dt
H
β
∫ m (t ) dt =
H
i −1
i
0 0 0
i =1

k ⎛ ∞ b *a
*
⎞
∑ ⎜⎜ Pr {β } ∫ ∫
H
αβi t βi −1 α a −1e −b α d α dt ⎟
* *
=
i =1
⎝
i
0 0
( )
Γ a* ⎟
⎠
k
a*
∑∫
H
= β i t βi −1 dt
i =1
0 b*
k
⎛ a* ⎞
= ∑ ⎜⎝ Pr {β } b
i =1
i *
H βi ⎟
⎠

10. After machine replacement, the time between productions of defective products for new
machine follows a negative exponential with constant failure rate λ , where λ follows Gamma
(1, λ ' ).

4. ANALYSIS OF TIME-TO-FAILURE DATA

Instead of fitting a single distribution to the time to repair or replace of two machines that produce a
specific product, it is much more informative to separately estimate the parameters of the mixture
model. In other words, we ideally model the time as a mixture of data from the two distributions as
in equation (2).

f (t , p ) = pf 1 (t ) + (1 − p )f 2 (t ) (2)

Where f 1 (t ) and f 2 (t ) are the probability distributions of the first and the second machine failures,
respectively, and the proportion p measures the extent of the first machine failure. The mixing
proportion p and the parameters of f 1 (t ) and f 2 (t ) have important physical interpretations in the
decision-making process. However, before applying this model, first we employ the Bayesian
inference to the mixture model using data augmentation technique (Sinuany-Stern 1993). To do
this, let zi = 1 if the defective product belongs to the first machine, and zi = 2 if the second one has
produced it. To choose the prior probability distribution of the mixing proportion, f ( p ) , in the
absence of specific prior knowledge, it is common to use non-informative priors. Referring to
Jeffrey’s non-informative prior (Tsang 1995), for the mixing proportion p, we take a Beta prior
1
distribution with parameters a = b = .
2

It is possible to express the unknown parameter p, in terms of a metric f ( p ) , so that the

corresponding likelihood is data translated. This means that the likelihood curve for f ( p ) is
completely determined a priori except for its location, which depends on the data yet to be
observed. Then to say that we know little a priori relative to what the data is going to tell us, may be
242 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

expressed by saying that we are almost equally willing to accept one value of f ( p ) as another. This
state of indifference may be expressed by taking f ( p ) to locally uniform, and the resulting priori
distribution is called non-informative for f ( p ) with respect to the data.

⎡ ∂ 2 log p ( x p ) ⎤
According to Jeffreys' Rule, assuming φ ( p ) = − E ⎢ ⎥ , the informative for p should be
x p
⎣⎢ ∂p 2 ⎥⎦
chosen so that locally f ( p ) = φ ( p ) . Specifically for the Beta distribution case, the information
0.5

measure is φ ( p ) ≈ p −1 (1 − p ) . Hence f ( p ) ≈ p −0.5 (1 − p ) . In other words, we conclude that

−1 −0.5

f ( p ) as a non-informative prior distribution is Beta (0.5, 0.5) (See Box and Tiao 1992).

With the aforementioned preliminaries, we are now ready to specify the data augmentation
algorithm for the problem at hand. Let t = (t 1 ,t 2 ,...,t n ) be the exact failure times, z = ( z 1 , z 2 ,..., z n )
be the latent variables corresponding to the defect modes, and R to be the number of times zi ’s
equal to one. Then, we can obtain the distribution of p t , z as:

f (t , z p )f ( p )
f (p t ,z ) =
∫ f (t , z p )f ( p )dp
Γ(a + b ) p a (1 − p )b
p R (1 − p ) n − R
Γ(a )Γ(b )
=
Γ(a + b ) p a (1 − p )b
∫p (1 − p ) n − R
R
dp
Γ(a )Γ(b )
Γ(a + R + b + n − R ) p a + R (1 − p )b + n − R
=
Γ(a + R )Γ(b + n − R )

which is a Beta distribution with parameters of R + a and n − R + b (Tanner and Wong 1987; Nair
et al. 2001).

5. NOTATIONS

We will use the following notations and definitions in the rest of the paper:

C cost of producing one unsatisfactory product

Ci cost of replacing machine i
C if cost of in-service replacement for machine i
p measures the extent of the of the first machine to produce defective products
Pirf Probability of in-service replacement for machine i after replacement
Pif Probability of in-service replacement for machine i before replacement
1
Standard expected mean of failure rate for machine i after replacement (new machine failure rate)
λi'
H overall time in period
r( p) distribution function of p
A One-Stage Two Machines Replacement Strategy … 243

The probability of an in-service replacement for machine one in the case of replacement, assuming
f (t ) = λ1'e − λ1t is:
'

∞
∫ ∫ f ( t λ ) h ( λ ) d λ dt
* *

k ( t ) dt =
T T
∫
0 0 0
T *
∞
λe ( − λt ) λ1'e
( − λλ )d λ dt
'

= ∫ ∫
0 0
T *
∞ (
− λ λ ' +t )d λdt
= ∫ ∫
0 0
λλ1'e
T *
1 T*
= ∫ λ' dt =
(λ ) λ ' +T
2 *
+t
0 '

In other words,

*
T*
∫ f (t ) dt =
T
P1rf =
0
λ1' + T *

The probability of an in-service replacement for machine one in the case of not replacing machine
one, assuming increasing failure rate, will be calculated using equation (1) as:

⎛ ⎛ ⎞
a* ⎞

∑ Pr {β }
k *
⎜1 − ⎜ b ⎟
P1f =
i =1
i
⎜ ⎝ T *βi + b * ⎠⎟ ⎟
⎝ ⎠

6. THE DECISION MAKING APPROACH

In order to apply Bayesian inference, assume n defective products has been produced. According to
what we state in section 4, for pre-replacement situation we have:

1. p t , z is distributed as B e ta ( R + a , n − R + b )

βl
⎛ n ⎛ n ⎞ ⎞
2. α t , z is distributed as Gamma denoted by Γ ⎜ a + R ,b +
1
⎜
t iβl + ⎜t B −
⎜
ti ∑ ∑ ⎟
⎟
⎟
⎟
⎝
i :z i =1 ⎝ i :z i =1 ⎠ ⎠

3. discrete distribution of β 1 t , z is

{∏ t }
βl −1
(b )
*
* a
β ln
n

{ = β t ,z } =
j , z j =1 j
Pr β 1 Pl (3)
∑ P β {∏ t }
l βi −1
(b )
k *
n n * a
i i j , z j =1 j
i =1

The distributions of α 2 t , z and β 2 t , z are obtained in a similar manner.

244 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

For machines’ post-replacement situation, to determine p, the extent of a machine to produce

defective products, for machine one and two we let the stochastic numbers e1 and e2 to denote the
numbers of defective products that the new machine one and two produces in the decision-making
stage, respectively. According to Bayesian inference, if we decide to replace any machines, we have
the following results:

4. p after machine one replacment is distributed as Beta(a + e1 , n − R + b )

5. p after machine two replacment is distributed as Beta(R + a,b + e 2 )

6. p after replacing both machines is distributed as Beta (a + e1 ,b + e 2 )

Then by applying the expected cost for each alternative, we get:

⎧C 1 + P1rf c 1f + P2f c 2 f + E ( cost of defective products ) , ⎫

⎪ ⎪
⎪C 2 + P1f c 1f + P2rf c 2 f + E ( cost of defective products ) , ⎪
V = M in ⎨ ⎬
⎪ P1f c 1f + P2f c 2 f + E ( cost of defective products ) , ⎪ (4)
⎪ ⎪
⎩C 1 + C 2 + P1rf c 1f + P2rf c 2 f + E ( cost of defective products ) ⎭

The first and the second term in the minimization program given in (4) indicate the expected cost
when we replace the first and the second machine, respectively. The third term refers to the
expected cost when none of the machines are replaced, and the fourth term is the expected cost
when both machines are replaced.

An important feature of this approach is that we use the historical data of defective products to
update the distribution function of failure rates and with considering appropriate assumptions, we
determine the optimal policy at the start of the current stage.

Applying equation (4) and conditional expectation, we have the following result:

E ( cost of defective products ) = C

1
⎛p ∞
m1 (t ) dt + (1 − p )
∞
m 2 (t ) dt ⎞⎟r ( p ) d
∫ ∫
⎜
0⎝ 0 ∫
0 ⎠
(5)

In order to calculate the expected costs of defective products in equation (4) we define the following
events:

FMR: The event of the First Machine Replacement

SMR: The event of the Second Machine Replacement
NMR: The event of No Machine Replacement
BMR: The event of Both Machine Replacements

Then, the costs associated with different decisions are:

A One-Stage Two Machines Replacement Strategy … 245

⎛ k ⎛ a * ⎞⎞
⎜1 − ⎛ b2 ⎞ 2 ⎟⎟ ⎛H⎞ ⎛ ⎛ k ⎛ ⎞⎞⎞
∑ { } ( ) ∑ { }
*
T* ⎜ a*
Cost of FMR :C1 + C + ⎜ Pr β ⎜ ⎟
* ⎟ ⎟⎟ 2 f
C + E ( p FMR ) C ⎜ ' ⎟ + ⎜ 1 − E ( p FMR ) C ⎜ ⎜ Pr β i 2* H βi ⎟ ⎟ ⎟
i ⎜ ⎜ *βi ⎜ ⎟ ⎜ ⎜ ⎟⎟
⎝ λ1 ⎠ ⎜⎝ ⎠ ⎠ ⎟⎠
* 1f
λ1 + T
'
⎜ i=1 ⎜ ⎝ T + b2 ⎠ ⎟ ⎟ ⎝ i=1 ⎝ b2
⎝ ⎝ ⎠⎠

⎛ k ⎛ a * ⎞⎞
⎛ T* ⎞ ⎛ b1* ⎞ 1 ⎟ ⎟ ⎛ k ⎛ a1* βi ⎞ ⎞ ⎛ ⎛ H ⎞⎞
Cost of SMR :C2 + ⎜ ' ⎟
⎜ λ +T* ⎟
⎝ 2 ⎠
C2 f +
⎜
⎜ ∑ { }
⎜ i=1
Pr β i
⎜1 − ⎜ ⎟
⎜ ⎜ T *βi + b * ⎟ ⎟ ⎟
⎜ ⎝ 1 ⎠ ⎟⎟
C1 f + E ( p SMR ) C ⎜
⎜
⎜
⎜
Pr β i∑ { }
* ⎟ ⎟ ⎜ (
H ⎟ ⎟ + ⎜ 1− E ( p SMR) C ⎜ ' ⎟ ⎟
⎜ λ ⎟⎟ )
⎝ ⎠⎠ ⎝ i=1 ⎝ b1 ⎠⎠ ⎝ ⎝ 2 ⎠⎠
⎝

⎛ k ⎛ a * ⎞⎞ ⎛ k ⎛ a * ⎞⎞
⎛ b* ⎞ 2 ⎟ ⎛ b* ⎞ 1 ⎟ ⎛ k ⎛ ⎞⎞ ⎛ ⎛ k ⎛ ⎞⎞ ⎞
∑{ }
⎜
Cost of NMR: ⎜ Pr βi ⎜1−⎜ *β 2 * ⎟ ⎟⎟C2 f
⎜ ⎜ T i +b ⎟ ⎟
⎜ ⎝
⎜
∑{ } ⎜ ⎜ T i +b ⎟ ⎟ ⎜ ⎜
a*
⎟ ⎟ ∑ { }
⎜
(⎜ ⎜
a*
) ∑ { }
+ ⎜ Pr βi ⎜1−⎜ *β 1 * ⎟ ⎟⎟C1f + E( p NMR) C⎜ ⎜ Pr βi 1* Hβi ⎟⎟ + ⎜ 1− E( p NMR) C⎜ ⎜ Pr βi 2* Hβi ⎟⎟⎟
⎟⎟ ⎟
⎜ i=1 2 ⎠ ⎟⎟ ⎜ i=1 ⎜ ⎝ 1 ⎠ ⎟⎟ ⎝ i=1 ⎝ b1 ⎠⎠ ⎝ ⎝ i=1 ⎝ b2 ⎠⎠ ⎠
⎝ ⎝ ⎠⎠ ⎝ ⎝ ⎠⎠

⎛ T* ⎞ ⎛H⎞ ⎛ ⎛H ⎞⎞
Cost of BMR:C1 + C2 +
T*
λ1' + T *
C1 f + ⎜
⎜ λ' + T* ⎟ 2 f
(⎜ λ' ⎟ ⎜
) (
⎟ C + E p BMR C ⎜ ⎟ + ⎜ 1 − E p BMR C ⎜
⎜ λ'
( )) ⎟⎟
⎟⎟
⎝ 2 ⎠ ⎝ 1⎠ ⎝ ⎝ 2 ⎠⎠

Since p machine one replacment is distributed as Beta (a + e1 , n − R + b ) , we have:

(
E ( p FMR ) = E E ( p FMR e1 ) )
⎛ a + e1 ⎞
=E⎜ ⎟
⎝ n − R + b + a + e1 ⎠
∞ ⎛ ∞
exp ( −λ1H )( λ1H ) ( a + i ) ⎞
i

= ⎜ ∑∫⎜ 0 Γ ( i + 1)( n − R + b + a + i )
i =1 ⎝
f ( λ1 ) d λ1 ⎟
⎟
⎠
∞ ⎛ ∞
exp ( −λ1H )( λ1H ) ( a + i ) ' − λ1λ1' ⎞
i

= ⎜ ∑∫⎜ 0 Γ ( i + 1)( n − R + b + a + i )
i =1 ⎝
λ1e d λ1 ⎟
⎟
⎠
∞
H i ( a + i ) λ1'
= ∑
(λ + H )
i +1
i =1
1
'
(n − R + b + a + i )

We note that in deriving the last equation we assumed a Poisson distribution for the number of
defective products in stage (with mean of λ1 H ).
By similar reasoning, we have:

∞
H i ( a + R ) λ2'
E ( p SMR ) = ∑
(λ )
i +1
i =1
'
2 +H (R + a + b + i )

Based on the expectation of beta distribution we have:

a+R
E ( p NMR ) =
n +a +b
246 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

⎛
( −(λ ) )(λ H ) (λ H ) ( a + i ) ' −λ λ ' −λ λ ⎞
i j
∞ ∞
⎜ ∞ ∞ exp + λ2 H ⎟
E ( p BMR ) = ∑∑ ∫ ∫
1 1 2
λ1e λ2e d λ1d λ2 ⎟
' '
1 1 2 2
⎜ Γ ( i + 1) Γ ( j + 1)( j + b + a + i )
j =1 i =1 ⎜ ⎟
0 0
⎝ ⎠
∞ ∞
H i+j
(a + i ) λ λ
' '
= ∑∑ 1 2

(λ + H ) ( λ2' + H ) (b + a + i + j )
i +1 j +1
'
j =1 i =1
1

which can be approximated by

H
λ1' λ2'
E ( p BMR ) = =
H
+
H λ2' + λ1'
λ1' λ2'

7. A NUMERICAL EXAMPLE

Consider a manufacturing process in which there are two serial machines producing a specific
product. For increasing hazard rate assumption, we have generated values of β from standard beta
distribution with parameter c=2 and d=1 and values of α have been generated from an exponential
distribution with mean 1. Generated data are shown in Table 1.

Assuming C1 = 1000, C2 = 1300, T * = 0.6, λ1' = 50, λ2' = 50, C1 f = 2000, C2 f =2500, H =100, C =5 , we have:

Cost of FM R : 2667.61
Cost of SMR : 2098.15
Cost of BM R : 2357.43
Cost of NM R : 2242.53

Table 1: Time between productions of defective products

Machine Two Machine One

0.63308 0.527716
0.822558 0.025934
0.140781 0.402137
0.21674 0.072486
0.023657 2.215276
0.6156 11.37043
0.280686 1.93114
0.091386 4.389833
0.741723 1.209809
0.34208 0.016308

Hence, the decision regarding the replacement of machine two is the best.
A One-Stage Two Machines Replacement Strategy … 247

8. A SENSITIVITY ANALYSIS

To analyze the effect of different values of the model parameters, in this section we present a
sensitivity study of the numerical example given in section 7. This study is based on different
values of the in-service replacement costs and the defective product cost.

If the in-service replacement cost is unknown, then based on different combinations of C1f and C2f
we can estimate the cost associated with each decision and determine the best strategy. Table 2
shows the results of this study.

The results of Table 2 show that the decision FMR is never the best. For large values of both C1f and
C2f replacing both machines is the best decision. Small values of C1f and C2f leads NMR to be the
best and moderate values of C1f and C2f results in replacing machine two.

The results of the sensitivity analysis based on different defective product costs are shown in Table
3. From the results of Table 3, we see that small values of C leads to replacing machine two and for
the large values of C, replacing both machines is the best decision. The other two decisions are not

Table2: The estimated cost of each decision for different combinations of C1f and C2f

C1f C2f NMR BMR SMR FMR

1000 1000 1114.773 2331.344 1756.604 1855.283
1200 1200 1286.962 2335.613 1823.966 1964.38
1400 1400 1459.151 2339.881 1891.327 2073.476
1600 1600 1631.341 2344.15 1958.688 2182.573
1800 1800 1803.53 2348.419 2026.05 2291.67
2000 2000 1975.719 2352.688 2093.411 2400.766
2200 2200 2147.908 2356.957 2160.773 2509.863
2400 2400 2320.098 2361.225 2228.134 2618.959
2600 2600 2492.287 2365.494 2295.495 2728.056
2800 2800 2664.476 2369.763 2362.857 2837.153
3000 3000 2836.665 2374.032 2430.218 2946.249
3200 3200 3008.855 2378.3 2497.58 3055.346
3400 3400 3181.044 2382.569 2564.941 3164.442
3600 3600 3353.233 2386.838 2632.302 3273.539
3800 3800 3525.422 2391.107 2699.664 3382.636
4000 4000 3697.612 2395.375 2767.025 3491.732
4200 4200 3869.801 2399.644 2834.387 3600.829
4400 4400 4041.99 2403.439 2901.748 3709.926
4600 4600 4214.179 2408.182 2969.109 3819.022
4800 4800 4386.369 2412.451 3036.471 3928.119
5000 5000 4558.558 2416.719 3103.832 4037.215
5200 5200 4730.747 2420.988 3171.193 4146.312
5400 5400 4902.936 2425.257 3238.555 4255.409
248 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

the best in none of the cases. The pictorial representation of the results of Table 3 is also given in
Figure 1.

9. CONCLUSIONS AND RECOMMENDATIONS FOR FUTURE RESEARCH

In this paper, first we briefly introduced the data augmentation method for Bayesian inferences in
the context of the finite mixture models. Then, we discussed the analysis of time-to-failure data and
the application of the methods of decision-making for the machine replacement strategy. In order to
demonstrate the application of the proposed method we provided a numerical example.

The proposed approach may be applied to any decision-making problems in which we need to
update the probability distribution function of its states by Bayesian inference approach.

For further researches in the subject of this research, we recommend the following:

1. In general, we can consider the case in which there are n machines instead of two.

2. We can consider other distribution functions for the time between producing of defective
products.

3. Failure rates, costs, and the probabilities in the mixture model all may be considered as fuzzy
parameters.

Table3: The estimated cost of each decision for different values of C

C FMR SMR BMR NMR

1 2418.892 2001.958 2349 2038.72
2 2480.784 2025.915 2351 2089.44
3 2542.676 2049.873 2353 2140.16
4 2604.568 2073.83 2355 2190.88
5 2666.46 2097.788 2357 2241.6
6 2728.352 2121.746 2359 2292.32
7 2790.244 2145.703 2361 2343.04
8 2852.136 2169.661 2363 2393.76
9 2914.028 2193.618 2365 2444.48
10 2975.92 2217.576 2367 2495.2
11 3037.812 2241.534 2369 2545.92
12 3099.704 2265.491 2371 2596.64
13 3161.596 2289.449 2373 2647.36
14 3223.488 2313.406 2375 2698.08
15 3285.38 2337.364 2377 2748.8
16 3347.272 2361.322 2379 2799.52
17 3409.164 2385.279 2381 2850.24
18 3471.056 2409.237 2383 2900.96
19 3532.948 2433.194 2385 2951.68
20 3594.84 2457.152 2387 3002.4
A One-Stage Two Machines Replacement Strategy … 249

Figure 1: The cost diagrams of different decisions

REFERENCES

[1] Box G.E.P., Tiao G.C. (1992), Bayesian inferences in statistical analysis; Wiley Classic Library
Edition; New York.

[2] Chen T.M., Popova E. (2000), Bayesian maintenance policies during a warranty period;
Communications in Statistics –Stochastic Models 16; 121-142.

[3] Childress S., Durango-Cohen P. (2005), On parallel machine replacement problems with general
replacement cost functions and stochastic deterioration; Naval Research Logistics 52; 409-419.

[4] Damien P., Galenko A., Popova E., Hanson T. (2007), Bayesian semi-parametric analysis for a single
item maintenance optimization; European Journal of Operational Research 182; 794-805.

[5] Gupta A., Lawsirirat C. (2006), Strategically optimum maintenance of monitoring-enabled multi-
component systems using continuous-time jump deterioration model; Journal of Quality in
Maintenance Engineering 12; 306-329.

[6] Hamada M., Martz H.F., Reese C.S., Graves T., Johnson V., Wilson A.G. (2004), A fully Bayesian
approach for combining multilevel failure information in fault tree quantification and optimal follow-
on resource allocation; Reliability Engineering and System Safety 86; 297-305.

[7] Hritonenko N., Yatsenko Y. (2007), Optimal equipment replacement without paradoxes: A continuous
analysis; Operations Research Letters 35; 245-250.

[8] Jardine A.K.S., Banjevic D., Makis V. (1997), Optimal replacement policy and the structure of
software for condition-based maintenance; Journal of Quality in Maintenance Engineering 3; 109-119.

[9] Mann L., Saxena A., Knapp G. (1995), Statistical-based or condition-based preventive maintenance;
Journal of Quality in Maintenance Engineering 1; 46-59.

[10] Mazzuchi T.A., Soyer R. (1996), A Bayesian perspective on some replacement strategies; Reliability
Engineering and System Safety 51; 295-303.

[11] Merrick J.R.W., Soyer R., Mazzuchi T.A. (2003), A Bayesian semi-parametric analysis of the
reliability and maintenance of machine tools; Technometrics 45; 58-69.
250 Fallah Nezhad, Akhavan Niaki and Eshragh-Jahromi

[12] Mobley K.R (1989), An introduction to predictive maintenance; Butterworth-Heinemann, New York.

[13] Moubray J. (1990), Reliability centered maintenance; Butterworth-Heinemann, Oxford.

[14] Nair V.N., Tang B., Xu L. (2001), Bayesian inference for some mixture problems in quality and
reliability; Journal of Quality Technology 33; 16-28.

[15] Saranga H. (2002), A dynamic opportunistic maintenance policy for continuously monitored systems;
Journal of Quality in Maintenance Engineering 8; 92-105.

[16] Sethi S.P., Sorger G., Zhou X.Y. (2000), Stability of real-time lot-scheduling and machine replacement
policies with quality levels; IEEE Transactions on Automatic Control 45; 2193-2196.

[17] Sherwin J.D. (1999), Age-based opportunity maintenance; Journal of Quality in Maintenance
Engineering 5; 221-235.

[18] Sherwin J.D., Al-Najjar B. (1999), Practical models for condition-based monitoring inspection
intervals; Journal of Quality in Maintenance Engineering 5; 203-209.

[19] Sinuany-Stern Z.S. (1993), Replacement policy under partially observed Markov process;
International Journal of Production Economics 29; 159-166.

[20] Sinuany-Stern Z.S., David I., Biran S. (1997), An efficient heuristic for a partially observable Markov
decision process of machine replacement; Journal of Computers in Operations Research 24; 117-126.

[21] Tanner M.A., Wong W.H. (1987), The calculation of posterior distributions by data augmentation;
Journal of the American Statistical Association 82; 528-540.

[22] Tsang A. (1995), Condition-based maintenance: tools and decision making; Journal of Quality in
Maintenance Engineering 1; 3-17.

[23] Valdez-Florez C., Feldman R. (1989), A survey of preventive maintenance models for stochastically
deteriorating single unit systems; Naval Research Logistics 36; 419-446.

[24] Wang C.H., Hwang S.L. (2004), A stochastic maintenance management model with recovery factor;
Journal of Quality in Maintenance Engineering 10; 154-164.

[25] Wilson J.G., Popova E. (1998), Bayesian approaches to maintenance intervention; In Proceedings of
the section on Bayesian Science of the American Statistical Association, 278-284.

[26] Zhou X., Xi L., Lee J. (2006), A dynamic opportunistic maintenance policy for continuously
monitored systems; Journal of Quality in Maintenance Engineering 12; 294-305.