Sequential Imperfect Preventive Maintenance Models with

Two Categories of Failure Modes

Daming Lin,1 Ming J. Zuo,2 Richard C. M. Yam3
1
Department of Mechanical and Industrial Engineering, University of Toronto,
Toronto, Ontario, Canada M5S3G8
2
Department of Mechanical Engineering, University of Alberta, Edmonton, Alberta,
Canada T6G 2G8
3
Department of Manufacturing Engineering and Engineering Management, City
University of Hong Kong, Hong Kong, People’s Republic of China

Received February 2000; revised July 2000; accepted September 12, 2000

Abstract: In this paper, we introduce the concept of two categories of failure modes: maintain-
able failure modes and nonmaintainable failure modes into the modeling of preventive maintenance
(PM) activities. The hybrid model proposed by Lin et al. is used to model the effect of a PM activ-
ity on the failure rate function of maintainable failure modes. However, PM does not change the
hazard rate function of nonmaintainable failure modes. PM is performed at a sequence of intervals.
The objective is to determine the optimal PM schedules to minimize the mean cost rate. Numerical
examples for Weibull hazard rates are given. c 2001 John Wiley & Sons, Inc. Naval Research Logistics
48: 172–183, 2001

Keywords: preventive maintenance (PM); Age Reduction Model; Hazard Rate Model; Hybrid
Model; maintainable failure modes; nonmaintainable failure modes

1. INTRODUCTION
Determination of a PM policy is an important issue in optimal maintenance planning. A PM
policy specifies how PM activities should be scheduled. PM policies can be divided into two main
categories: periodic and sequential. Periodic PM defines that a system is maintained at integer
multiples of some fixed period and undergoes only minimal repair at failures between these PMs.
Minimal repair only restores the function of the system when it is failed, but does not change the
general health condition of the system. Sequential PM is the same as periodic PM except that the
system is maintained at a sequence of intervals which may have unequal lengths. Periodic PM is
more convenient to schedule, whereas sequential PM is more realistic when the system requires
more frequent maintenance as it ages.
The general health condition of a piece of equipment has been measured with hazard rate func-
tion and effective age. The higher the hazard rate function value, the worse the equipment health

Correspondence to: M.J. Zuo

c 2001 John Wiley & Sons, Inc.

 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 173

condition. For equipment whose health condition deteriorates with its age, the term ‘‘effective
age’’ has been used to measure its health condition. For deteriorating equipment, the higher the
effective age, the worse the health condition of the equipment.
A PM model is used to measure the effects of a PM activity. Many publications have reported
PM models. Traditional PM models assume that the system after PM is either ‘‘as good as new’’
or ‘‘as bad as old.’’ The more realistic assumption is that the system after PM lies somewhere
between as good as new and as bad as old. This kind of PM is called imperfect PM. There are
two main alternatives for modeling an imperfect PM activity. The first one assumes that PM is
equivalent to minimal repair with probability p and equivalent to replacement with probability
1−p [1–3, 10, 11, 16]. The second one models the PM’s effects directly in terms of how the hazard
rate function or the effective age of the equipment is changed by the PM [4–6, 8, 9, 13–15]. Using
the concept of effective age, we may say that a certain PM has restored the health condition of a
piece of equipment with a calendar age of 5 years to ‘‘as good as a 4-year old.’’ Such a statement
indicates that the PM has made the equipment younger and thus healthier.
Lie and Chun [6] and Nakagawa [12] introduce the concept of adjustment factors in hazard rate
function and effective age to model the effects of PM. Nakagawa [15] proposes two PM models
assuming that the required time for PM is negligible. Let’s use ti to represent the point of time
when the ith PM is performed on the equipment (i ≥ 1). We will assume that a new piece of
equipment is put in use at time t0 = 0. These two PM models are described below.
1. Hazard Rate PM Model: The hazard rate function after the ith PM becomes
ai hi−1 (x) for x ∈ (0, ti+1 − ti ) when it was hi−1 (x) for x ∈ (0, ti − ti−1 ),
where ai > 1 is the adjustment factor for the hazard rate function due to the
ith PM. Nakagawa [15] also assumes that the hazard rate function is zero for
a new piece of equipment. Based on these model assumptions, we can say that
the hazard rate PM model represents the situation wherein the equipment’s
hazard rate function is an increasing function of time when there are no PM
interventions, each PM resets the hazard rate function value to zero, and the
rate of increase of the hazard rate function gets higher after each additional PM.
2. Age Reduction PM Model: The effective age after the ith PM reduces to bi Ei
if the equipment’s effective age was Ei just prior to this PM, where bi < 1 is the
improvement factor in the effective age of the equipment due to the ith PM. The
effective age of equipment is the same as its actual age before the first PM is
performed. In other words, at time t1 − (just prior to time t1 ), the equipment’s
effective age and the actual age are both t1 . At time t1 + (just after the first
PM is completed), the effective age becomes b1 t1 while the actual age of the
equipment remains the same. The equipment’s health condition right after the
first PM is represented by its hazard rate function value which is equal to the
same value as that when the equipment’s actual age was b1 t1 . If we use h0 (x)
to represent the failure rate function prior to the first PM for x ∈ (0, t1 ), then
h1 (x) = h0 (b1 t1 + x) for x ∈ (0, t2 − t1 ) represents the failure rate function
of the equipment in the time interval of (t1 , t2 ). The hazard rate function of
the equipment is then a function of its effective age and each PM reduces the
effective age of the equipment to a certain extent.
Based on the descriptions of these two models, we can say that the age reduction PM model
has the advantage of determining the instantaneous failure rate function value right after a PM.
In other words, we can say h(t1 +) = h(b1 t1 ) if we use h(t) to denote the hazard rate function of
the equipment as a function its actual age. On the other hand, the hazard rate PM model has the
174 Naval Research Logistics, Vol. 48 (2001)

advantage of allowing the rate of increase of the hazard function to be higher after each additional
PM is performed on the equipment. Based on these observations, Lin, Zuo, and Yam [7] propose
the Hybrid PM Model that combines the advantages of the Age Reduction PM Model and the
Hazard Rate PM Model. In the Hybrid PM Model, it is assumed that effects of each PM are
modeled by two aspects: one for its immediate effects after the PM is completed and the other
for the lasting effects when the equipment is put to use again. This hybrid model captures both
by how much the effective age is reduced the instant the PM is completed and how much faster
the hazard rate function will increase after the equipment is put into use again. With this Hybrid
PM Model, the failure rate function after the first PM may be written as

h(t1 + x) = a1 h(b1 t1 + x), for x > 0.

Using this Hybrid PM Model, we can still use the concept of effective age by saying that the
effective age of the equipment becomes b1 t1 right after the first PM. However, the failure rate at
this moment is not the same as the failure rate when the equipment’s age is b1 t1 . Instead, it is
equal to a1 h(b1 t1 ). The adjustment factor a1 that is used in the Hazard Rate PM Model is also
used here to find the failure rate function value right after the PM is completed. Figure 1 is given
for comparison of these three models before and after the first PM is performed assuming a linear
hazard rate function when there is no PM intervention.
Two popular approaches used to determine the maintenance intervals for a sequential PM are the
reliability-based method and the optimization method. For the former method, PM is performed
whenever the system reliability (or the hazard rate of the system) reaches a predetermined level

Figure 1. Comparison of the reported PM models.

 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 175

[4–6, 8, 9]. The latter one selects the optimal intervals to minimize the mean cost rate or total life
cycle cost [13–15].
It is reasonable to assume that system failure modes can be grouped into two categories:
maintainable failure modes and nonmaintainable failure modes. PM can reduce the hazard rate of
the maintainable failure modes but cannot change the hazard rate of the nonmaintainable failure
modes. For example, PM of a machine may include oiling, tightening screws, and cleaning, etc. But
this kind of PM activities can only improve the failure modes which are affected by such working
environment of the machine. Those failure modes which are affected only by the inherent design
of the machine, however, cannot be improved by the PM activities. In this paper, we introduce the
concept of these two categories of failure modes to model PM activities. We also use the Hybrid
PM Model to model the effects of PM on the maintainable failure modes. We develop optimal
PM intervals through two alternatives: (1) leaving the PM intervals as decision variables and (2)
allowing the PM intervals to be determined by the hazard rate or reliability limit. The objective
functions in both cases are to minimize the mean cost rate. Numerical examples for the case of
Weibull hazard rates will be given for illustration purposes.

Notation:

ha (t) = hazard rate of the nonmaintainable failure modes of the system,

hb (t) = hazard rate of the maintainable failure modes of the system,
Ha (t) = cumulative hazard rate of the nonmaintainable failure modes of the system,
Hb (t) = cumulative hazard rate of the maintainable failure modes of the system,
hk (t) = hazard rate of the system between the (k − 1)th PM and the kth PM,
Hk (t) = cumulative hazard rate of the system between the (k − 1)th and the kth PMs,
λ = predetermined acceptable level of hazard rate,
xk = scheduled PM intervals, k = 1, 2, . . . , N ,
tk = x1 + x2 + · · · + xk , k = 1, 2, . . . , N ,
yk = effective age of the system just before the kth PM, k = 1, 2, . . . , N ,
N = scheduled number of PM intervals (the N th PM is actually a replacement),
ak = adjustment factor in hazard rate of the maintainable failure modes after the kth PM,
1 = a0 ≤ a1 ≤ a2 ≤ · · · ≤ aN −1 ,
Qk−1
Ak = i=0 ai , k = 1, 2, . . . , N
bk = adjustment factor in effective age due to the kth PM, 0 = b0 ≤ b1 ≤ b2 ≤ · · · ≤
bN −1 < 1,
cm = cost of minimal repair,
cp = cost of PM,
cr = cost of replacement,
C = mean cost rate of the system,
(1−bk )α 1/(α−1)
dk = [ β1 +Ak β2 −(β1 +Ak+1 β2 )b
α] .
k

Assumptions:

1. The planning horizon is infinite.

2. The hazard rate functions of the nonmaintainable and maintainable failure
modes of the system, ha (t) and hb (t), are continuous and strictly increasing if
there are no PM interventions.
3. The times for PM, minimal repair and replacement are negligible.
176 Naval Research Logistics, Vol. 48 (2001)

4. PM is performed at t1 , t2 , . . . , tN −1 , and the system is replaced at tN .

5. The system is restored to as good as new state at replacement.
6. Maintainable failure modes and non-maintainable failure modes are indepen-
dent and they are competing to cause system failure.
7. The Hybrid PM Model proposed by Lin, Zuo, and Yam [7] will be used to
model the effects of PM on maintainable failure modes.

2. MODEL FORMULATIONS AND OPTIMAL SOLUTIONS

We consider the situation where a system is preventively maintained at t1 , t2 , . . . , tN −1 and
replaced at tN . Minimal repair is performed at failures between PMs. Replacement of the system
restores the system to the as good as new state. The system has the hazard rate due to maintainable
failure rate Ak hb (t) between the (k − 1)th and the kth PMs, i.e., in time interval (tk−1 , tk ). The
hazard rate of the nonmaintainable failure rate in this time interval is ha (t). Thus, the hazard
rate of the system in time interval (tk−1 , tk ) is hk (t) = ha (t) + Ak hb (t). Based on the model
descriptions, we have

yk = xk + bk−1 xk−1 + · · · + bk−1 bk−2 · · · b2 b1 x1 ,

yk = xk + bk−1 yk−1 ,

xk = yk − bk−1 yk−1 .
Hence the mean cost rate is
PN
cr + cp (N − 1) + cm k=1 [Hk (yk ) − Hk (bk−1 yk−1 )]
C= PN −1 . (1)
[ k=1 (1 − bk )yk + yN ]

There are two main alternatives to determine the PM intervals in the literature. One is to select
optimal PM intervals to minimize the mean cost rate, i.e., to leave PM intervals as decision
variables in the optimization problem [4–6, 8, 9]. The other is to determine PM intervals by the
hazard rate or reliability limit [13–15]. They will be used respectively in the following model
formulations.

2. 1. Model 1
In this model, the PM times t1 , t2 , . . . , tN −1 and the replacement time tN are all treated as inde-
pendent decision variables. Alternatively, the decision variables are N and yk (k = 1, 2, . . . , N ).
That is, the objective in this model is to determine the optimal values of N and yk (k = 1, 2, . . . , N )
to minimize the mean cost rate expressed in Eq. (1).
Setting the differentiation of Eq. (1) with respect to yk to zero leads to

hk (yk ) − bk hk+1 (bk yk ) = (1 − bk )[ha (yN ) + AN hb (yN )], k = 1, 2, . . . , N − 1, (2)

and

cm [ha (yN ) + AN hb (yN )] = C. (3)

 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 177

THEOREM 1: For a fixed yN (0 < yN < ∞), the solution to Eq. (2) with respect to yk
(yk > 0) exists if 1 − ak bk > 0, k = 1, 2, . . . , N − 1. Further, the solution is unique if ha (t) and
hb (t) are differentiable and h′a (t) and h′b (t) are strictly increasing.

PROOF: If 1 − ak bk > 0, we can show that the lhs of Eq. (2) is less than the rhs when yk = 0:

hk (0) − bk hk+1 (0) = (1 − bk )ha (0) + Ak (1 − ak bk )hb (0) < (1 − bk )[ha (yN ) + AN hb (yN )].

With 1 − ak bk > 0, the lhs of Eq. (2) is greater than the rhs when yk → +∞:

hk (yk ) − bk hk+1 (bk yk ) ≥ (1 − bk )ha (yk ) + Ak (1 − ak bk )hb (yk ) → +∞ as yk → +∞.

It should be noted that the rhs is a constant for a fixed yN . Therefore, there exists some yk (yk > 0)
which satisfies Eq. (2). Moreover, when h′a (t) and h′b (t) are strictly increasing, the derivative of
the lhs of Eq. (2) with respect to yk is

h′k (yk ) − b2k h′k+1 (bk yk ) > (1 − b2k )h′a (yk ) + Ak (1 − ak bk )h′b (yk ) > 0.

That is, the lhs of Eq. (2) is a strictly increasing function of yk . Therefore, the solution to Eq. (2)
with respect to yk is unique.
Note that the condition 1 − ak bk > 0 means that the hazard rate adjustment factor ak (> 1)
should be smaller than the reciprocal of the age reduction factor bk . This is a reasonable condition
since PM tends to improve the system’s condition. The extent of system improvement by PM
is determined by the values of ak and bk : The smaller the values of ak and bk , the better the
improvement. If the condition 1 − ak bk > 0 does not hold, the solution to Eq. (2) may not exist.
Then the optimal yk may be attained at boundary points or may not exist. In the following, we
assume that the condition 1 − ak bk > 0 holds.
Substituting each solution to Eq. (2), yk (k = 1, 2, . . . , N − 1), into Eq. (3), we have
"N −1 # N
X X cr + cp (N − 1)
hN (yN ) (1 − bk )yk + yN − [Hk (yk ) − Hk (bk−1 yk−1 )] = , (4)
cm
k=1 k=1

where each yk (k = 1, 2, . . . , N − 1) is some function of yN . Then, the lhs of Eq. (4) is a function
only of yN .
THEOREM 2: If 1 − ak bk > 0, k = 1, 2, . . . , N − 1, ha (t) and hb (t) are differentiable,
ha (0) = 0, hb (0) = 0, and h′a (t) and h′b (t) are strictly increasing, the solution to Eq. (4) with
respect to yN (yN > 0) exists and is unique.

PROOF: Since

hk (yk ) − bk hk+1 (bk yk ) ≥ (1 − bk )ha (yk ) + Ak (1 − ak bk )hb (yk ) ≥ 0, k = 1, 2, . . . , N − 1,

the solution to Eq. (2) with respect to yk is zero when yN = 0. Thus, if yN = 0, the lhs of Eq.
(4) equals zero and is smaller than the rhs of Eq. (4). On the other hand, the derivative of the lhs
of Eq. (4) with respect to yN is found to be
"N −1 #
X
[h′a (yN ) + AN h′b (yN )] (1 − bk )yk + yN > 0
k=1
178 Naval Research Logistics, Vol. 48 (2001)

if yN > 0. Hence, the lhs of Eq. (4) is a strictly increasing function of yN (yN ≥ 0) and equals
zero at yN = 0. Therefore, the solution to Eq. (4) with respect to yN (yN > 0) exists and is
unique.
Based on the above results, we obtain the following algorithm for finding the optimal PM
schedule:
Step 1. Solve Eq. (2) with respect to yk (k = 1, 2, . . . , N − 1). The solutions are functions
of yN .
Step 2. Substitute the solutions in Step 1 into Eq. (4) and solve it with respect to yN .
Step 3. Choose N to minimize [ha (yN ) + AN hb (yN )], where yN is the solution obtained in
Step 2.
Step 4. Compute yk , k = 1, 2, . . . , N from expressions obtained in Steps 1 and 2 for the value
of N obtained in Step 3.
Step 5. Compute xk from xk = yk − bk−1 yk−1 , k = 1, 2, . . . , N .
When the condition ha (0) = 0 and hb (0) = 0 in Theorem 2 does not hold, the solution to Eq.
(4) with respect to yN (yN > 0) may or may not exist. But the above algorithm still works if the
solution to Eq. (4) with respect to yN (yN > 0) exists.
It should be noted that Model 1 reduces to a pure hazard rate reduction model when bk =
0, k = 1, 2, . . . , N , and a pure age reduction model when ak = 1, k = 1, 2, . . . , N .

2. 2. Model 2
In this model, PM is performed whenever the hazard rate of the system reaches the prede-
termined level λ and the decision variable is N . This implies that the hazard rate at time ti
(i = 1, 2, . . . , N ) must equal λ. That is,

hk (yk ) = ha (yk ) + Ak hb (yk ) = λ, k = 1, 2, . . . , N. (5)

Solving Eq. (5) with respect to yk (k = 1, 2, . . . , N ), we get the solutions as expressions of λ.

Substitute these expressions of λ. Substitute these expressions into Eq. (1). Then, C becomes a
function of λ. Differentiating C with respect to λ and setting it to zero leads to
PN −1 hk (yk )−bk hk+1 (bk yk )
k=1 ′ (yk )
hk
+ hhN (yN )
′ (yN ) C
PN −1 1−bk
N
= , (6)
k=1 h′ (yk ) + h′ (yN )
1 cm
k N

where yk (k = 1, 2, . . . , N ) is a function of λ. Solving Eq. (6) with respect to λ, we obtain λ as

a function of N . Then, we can find N which minimizes
PN −1 hk (yk )−bk hk+1 (bk yk )
k=1 h′k (yk )
+ hhN (yN )
′ (yN )
PN −1 1−bk 1
N
, (7)
k=1 h′ (yk ) + h′ (yN )
k N

where yk (k = 1, 2, . . . , N ) is a function of λ and thus a function of N .

Based on the above discussions, we derive the algorithm for obtaining the optimal PM schedule
as follows.
Step 1. Solve Eq. (5) with respect to yk (k = 1, 2, . . . , N − 1). The solutions are functions
of λ.
 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 179

Step 2. Substitute the solutions in Step 1 into Eq. (6) and solve it with respect to λ.
Step 3. Choose N to minimize the function in expression (7) where yk (k = 1, 2, . . . , N ) is
the expression obtained in Steps 1 and 2.
Step 4. Compute yk , k = 1, 2, . . . , N from expressions obtained in Steps 1 and 2 for the value
of N obtained in Step 3.
Step 5. Compute xk from xk = yk − bk−1 yk−1 , k = 1, 2, . . . , N .
It can be shown that Eqs. (5) and (6) are equivalent to Eqs. (2) and (3), respectively, when
bk = 0, k = 1, 2, . . . , N . This means that the optimal PM schedules in Model 1 and Model 2 are
the same for a pure hazard rate reduction policy.

3. NUMERICAL EXAMPLES
We consider the case when the hazard rate functions ha (t) and hb (t) of the system are Weibull
and have the same shape parameter α, i.e.,

ha (t) = β1 tα−1 , hb (t) = β2 tα−1 , α > 1, β1 > 0, β2 > 0.

For Model 1, Eq. (2) becomes

β1 ykα−1 +Ak β2 ykα−1 −bk [β1 (bk yk )α−1 +Ak+1 β2 (bk yk )α−1 ] = (1−bk )(β1 yN
α−1 α−1
+AN β2 yN ),

for k = 1, 2, . . . , N − 1. Solving this equation we get

 1/(α−1)
(1 − bk )(β1 + AN β2 )
yk = yN , k = 1, 2, . . . , N − 1. (8)
β1 + Ak β2 − (β1 + Ak+1 β2 )bα
k

Substituting equation (8) into equation (3) leads to

[cr + cp (N − 1)]1/α
yN = PN −1 . (9)
{cm (1 − 1/α)[β1 + AN β2 + (β1 + AN β2 )α/(α−1) k=1 dk ]}1/α

Then, ha (yN ) + AN hb (yN ) becomes

(β1 + AN β2 )[cr + cp (N − 1)](α−1)/α

PN −1 .
{cm (1 − 1/α)[β1 + AN β2 + (β1 + AN β2 )α/(α−1) k=1 dk ]}(α−1)/α

Minimization of ha (yN ) + AN hb (yN ) is equivalent to minimization of function

cr + cp (N − 1)
B(N ) = PN −1 , N = 1, 2, . . . .
(β1 + AN β2 )−1/(α−1) + k=1 dk

The optimal value of N can be found by solving inequalities B(N + 1) ≥ B(N ) and B(N ) <
B(N − 1). These two inequalities imply that

cr cr
D(N ) ≥ and D(N − 1) < , (10)
cp cp
180 Naval Research Logistics, Vol. 48 (2001)

where
PN −1
(β1 + AN β2 )−1/(α−1) + k=1 dk
D(N ) = − (N − 1).
(β1 + AN +1 β2 )−1/(α−1) − (β1 + AN β2 )−1/(α−1) + dN
Thus, the optimal value N ∗ must satisfy inequalities (10). The optimal PM intervals are computed
from xk = yk − bk−1 yk−1 (k = 1, 2, . . . , N ), where yk (k = 1, 2, . . . , N ) are given in Eqs. (8)
and (9).
To compute the optimal PM schedule, we need to know the cost parameters cm , cp , and cr , the
Weibull parameters α, β1 , and β2 , and the adjustment factors ak and bk . For the cost parameters,
actually we only need to know the cost ratios cr /cp and cm /cp . One has to obtain these parameters
from data analysis using the failure and maintenance database. For example, a goodness-of-fit test
needs to be conducted to see if the Weibull distribution can be used and if yes we need to estimate
the parameters of the Weibull distribution. As far as the adjustment parameters are concerned, the
users need to observe the failure rate function before and after each PM. The failure rate function
values after a fixed amount of time has elapsed from every PM can be compared to fit the trend
of the failure rate as a function of the number of PM activities. In our numerical example in
this section, however, we will arbitrarily select these parameter values to illustrate the model and
results presented earlier in this paper.
Table 1 shows the optimal PM schedules, N ∗ , x1 , x2 , . . . , xN ∗ , in Model 1 with and with-
out nonmaintainable failure modes considered, respectively, for cr /cp = 2, 5, 10, 20, 50, where
cm /cp = 4, α = 2, β1 = 2, β2 = 3, and

6k + 1 k
ak = , bk = , k = 0, 1, 2, . . . .
5k + 1 2k + 1
From Table 1, we can see that more PM intervals are needed and the PM intervals at the beginning
tend to be larger for the case with nonmaintainable failure modes considered than when the
nonmaintainable failure modes are ignored. The PM interval is becoming shorter and shorter
except the last one for both cases. Besides, the shortening speed at the beginning is much faster
when non-maintainable failure modes are considered than when they are ignored.
For Model 2, solving Eq. (5) with respect to yk (k = 1, 2, . . . , N ), we have
 1/(α−1)
λ
yk = , k = 1, 2, . . . , N. (11)
β1 + Ak β2

Substituting Eq. (11) into Eq. (6) and solving it with respect to λ, we get
 (α−1)/α
cr + cp (N − 1)
λ= , (12)
(1 − 1/α)cm E(N )
where
N −1  
X β1 + Ak+1 β2 α
E(N ) = 1− b (β1 + Ak β2 )−1/(α−1) + (β1 + AN β2 )−1/(α−1) .
β1 + Ak β2 k
k=1

Then expression (7) becomes

(α−1)/α
[E(N )]1/α

cr + cp (N − 1)
, (13)
(1 − 1/α)cm F (N )
 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 181

Table 1. Optimal PM schedules for Weibull case in Model 1.

With Nonmaintainable Failure Modes Without Nonmaintainable Failure Modes

cr /cp 2 5 10 20 50 2 5 10 20 50

N∗ 1 3 6 9 13 1 3 5 7 11
x1 0.447 0.485 0.609 0.775 1.100 0.447 0.504 0.648 0.838 1.207
x2 0.262 0.329 0.419 0.595 0.249 0.321 0.415 0.597
x3 0.350 0.258 0.328 0.466 0.310 0.234 0.303 0.436
x4 0.214 0.272 0.386 0.183 0.237 0.341
x5 0.180 0.229 0.326 0.267 0.191 0.274
x6 0.281 0.194 0.276 0.155 0.224
x7 0.165 0.235 0.238 0.184
x8 0.140 0.199 0.151
x9 0.224 0.169 0.125
x10 0.143 0.104
x11 0.120 0.164
x12 0.101
x13 0.164

where
N
X −1
F (N ) = (1 − bk )(β1 + Ak β2 )−1/(α−1) + (β1 + AN β2 )−1/(α−1) .
k=1

Minimization of the function in (13) is equivalent to minimization of function

[cr + cp (N − 1)](α−1)/α [E(N )]1/α

Q(N ) = .
F (N )

Inequalities Q(N + 1) ≥ Q(N ) and Q(N ) < Q(N − 1) imply that

cr cr
W (N ) ≥ and W (N − 1) < , (14)
cp cp

where

[E(N + 1)]1/(α−1) [F (N )]α/(α−1)

W (N ) = − (N − 1).
[E(N )]1/(α−1) [F (N + 1)]α/(α−1) − [E(N + 1)]1/(α−1) [F (N )]α/(α−1)

Thus, the optimal value N ∗ must satisfy inequalities (14). The optimal PM intervals are computed
from xk = yk − bk−1 yk−1 (k = 1, 2, . . . , N ), where yk (k = 1, 2, . . . , N ) and λ are given in
Eqs. (11) and (12).
Table 2 gives the optimal PM schedules in Model 2 with and without nonmaintainable failure
modes considered, respectively, for the same parameters adopted in the example of Model 1.
From Table 2, we also observe that more PM intervals are needed and the PM intervals at the
beginning are larger for the case with nonmaintainable failure modes considered than when the
nonmaintainable failure modes are ignored. The PM interval is becoming shorter and shorter for
182 Naval Research Logistics, Vol. 48 (2001)

Table 2. Optimal PM schedules for Weibull case in Model 2.

With Nonmaintainable Failure Modes Without Nonmaintainable Failure Modes

cr /cp 2 5 10 20 50 2 5 10 20 50

N∗ 1 4 6 9 13 1 3 5 8 11
x1 0.447 0.517 0.622 0.766 1.067 0.447 0.553 0.671 0.835 1.180
x2 0.298 0.358 0.441 0.614 0.290 0.351 0.437 0.618
x3 0.233 0.281 0.346 0.481 0.211 0.257 0.319 0.451
x4 0.193 0.233 0.287 0.399 0.201 0.250 0.354
x5 0.196 0.242 0.337 0.162 0.202 0.285
x6 0.167 0.205 0.286 0.165 0.233
x7 0.174 0.242 0.135 0.191
x8 0.148 0.206 0.112 0.158
x9 0.125 0.174 0.130
x10 0.147 0.108
x11 0.124 0.090
x12 0.105
x13 0.088

both cases. Besides, the shortening speed at the beginning is much faster when nonmaintainable
failure modes are considered than when they are ignored.

4. CONCLUDING REMARKS
In this paper we have presented a general PM policy which assumes that PM only reduces the
hazard rate of maintainable failure modes of the system, but does not affect the hazard rate of
nonmaintainable failure modes of the system. Two models for obtaining the optimal PM schedules
have been studied. Generally, the optimal solutions in the two models are not the same. When the
Hybrid PM Model reduces to the Hazard Rate PM Model (bk = 0), however, it can be shown that
the optimal PM schedules obtained from Model 1 and Model 2 are exactly the same. From the
examples, we see that the optimal solution to yk is analytically found in the special Weibull case.
For the general case, however, the optimal yk may not be analytically obtainable. The algorithms
in Section 2 may become numerical procedures and standard numerical approaches for solving
an equation are required in Step 1 and Step 2.
From Table 1, we can see that xk decreases for all k values except for the last one. This coincides
with Nakagawa’s [15, p. 297] comment: ‘‘it would be reasonable to do frequent PM with age, but
it would be better to do the last PM as late as possible because the system should be replaced at
next PM.’’ From Table 2, however, we observe that xk decreases for all k values. Therefore, we
may conclude that Model 1 would be more practical than Model 2.
In Model 2, the hazard rate at each PM remains the same (equals λ). The predetermined level
λ is selected to minimize the mean cost rate. An alternative and reasonable way is to determine
λ to meet practical requirement of the system reliability. That is, λ is set to be a specific level λ0
(λ0 is known). Then yk (k = 1, 2, . . . , N ) can be computed from Eq. (5). The mean cost rate C
is a function only of N . Thus, we can find N which minimizes C and finally obtain the optimal
PM schedule.
Some issues deserving further research include how to separate failure modes into maintainable
and nonmaintainable failure modes, how to use maintenance and failure database to obtain life
 Lin, Zuo, and Yam: Sequential Imperfect Preventive Maintenance 183

distribution parameters and PM adjustment factors, and when the models presented in this paper
and other literatures can be applied effectively.

ACKNOWLEDGMENTS
This research was partially supported by the City University of Hong Kong and the Natural
Sciences and Engineering Research Council of Canada. Comments from referees and the Editor
are very much appreciated.

REFERENCES
[1] M. Brown and F. Proschan, Imperfect repair, J Appl Probab 20 (December 1983), 851–859.
[2] P.K.W. Chan and T. Downs, Two criteria for preventive maintenance, IEEE Trans Reliab R-27 (October
1978), 272–273.
[3] R.A. Fontenot and F. Proschan, ‘‘Some imperfect maintenance models,’’ Reliability theory and models,
Academic, New York, 1984, pp. 83–101.
[4] V. Jayabalan and D. Chaudhuri, Optimal maintenance and replacement policy for a deteriorating system
with increased mean downtime, Nav Res Logistics 39 (1992), 67–78.
[5] V. Jayabalan and D. Chaudhuri, Cost optimization of maintenance scheduling for a system with assured
reliability, IEEE Trans Reliab 41 (March 1992), 21–25.
[6] C.H. Lie and Y.H. Chun, An algorithm for preventive maintenance policy, IEEE Trans Reliab R-35
(April 1986), 71–75.
[7] D. Lin, M.J. Zuo, and R.C.M. Yam, Sequential imperfect preventive maintenance models, Int J Reliab
Qual Safety Eng 2000, to appear.
[8] M.A.K. Malik, Reliable preventive maintenance scheduling, AIIE Trans 11 (September 1979), 221–
228.
[9] A. Monga, M.J. Zuo, and R.W. Toogood, Reliability-based design of systems considering preventive
maintenance and minimal repair, Int J Reliab Qual Safety Eng 4 (March 1997), 55–71.
[10] D.N.P. Murthy and D.G. Nguyen, Optimal age-policy with imperfect preventive maintenance, IEEE
Trans Reliab R-30 (April 1981), 80–81.
[11] T. Nakagawa, Optimum policies when preventive maintenance is imperfect, IEEE Trans Reliab R-28
(October 1979), 331–332.
[12] T. Nakagawa, A summary of imperfect preventive maintenance policies with minimal repair, RAIRO
Oper Res 14 (August 1980), 249–255.
[13] D.G. Nguyen and D.N.P. Murthy, Optimal preventive maintenance policies for repairable systems,
Oper Res 29 (November–December 1981), 1181–1194.
[14] T. Nakagawa, Periodic and sequential preventive maintenance policies, J Appl Probab 23 (June 1986),
536–542.
[15] T. Nakagawa, Sequential imperfect preventive maintenance policies, IEEE Trans Reliab R-37 (August
1988), 295–298.
[16] S.-H. Shen and C.-T. Liou, A generalized sequential preventive maintenance policy for repairable
systems with general random minimal repair costs, Int J Syst Sci 26 (March 1995), 681–690.