Failure and Maintenance 2005
description
Transcript of Failure and Maintenance 2005
European Journal of Operational Research 161 (2005) 183–190
www.elsevier.com/locate/dsw
Production, Manufacturing and Logistics
Failure replacement and preventive maintenancespare parts ordering policy
Timothy S. Vaughan *
Department of Management and Marketing, University of Wisconsin-Eau-Claire, Eau Claire, WI 54702-4004, USA
Received 11 September 2002; accepted 2 June 2003
Available online 14 November 2003
Abstract
This paper addresses inventory policy for spare parts, when demand for the spare parts arises due to regularly
scheduled preventive maintenance, as well as random failure of units in service. A stochastic dynamic programming
model is used to characterize an ordering policy which addresses both sources of demand in a unified manner. The
optimal policy has the form ðsðkÞ; SðkÞÞ, where k is the number of periods until the next scheduled preventive main-
tenance operation. The nature of the ðsðkÞ; SðkÞÞ policy is characterized through numeric evaluation. The efficiency of
the optimal policy is evaluated, relative to a simpler policy which addresses the failure replacement and preventive
maintenance demands with separate ordering policies.
� 2003 Elsevier B.V. All rights reserved.
Keywords: Inventory; Failure replacement; Preventive maintenance; Dynamic programming
1. Introduction
The problem of providing an adequate yet effi-
cient supply of spare parts, in support of mainte-
nance and repair of plant and equipment, is an
especially vexing inventory management scenario.
Spare parts for plant and equipment may be veryexpensive, and thus costly to keep in inventory.
Nonetheless, spares must be on hand when needed,
in order to avoid costly plant shutdown or
equipment unavailability. In the face of this com-
bination of high cost and high criticality, random
* Tel.: +1-715-836-4408.
E-mail address: [email protected] (T.S. Vaughan).
0377-2217/$ - see front matter � 2003 Elsevier B.V. All rights reserv
doi:10.1016/j.ejor.2003.06.026
failure of units in service typically generates a low-
volume, intermittent demand process.
In addition to random failure replacement, de-
mand for spare parts may also arise under a policy
of regularly scheduled shutdown and preventive
maintenance for the larger system in which the
parts are used. In contrast to the low volumefailure replacement demand, preventive mainte-
nance may present a ‘‘lumpy demand’’ scenario in
which a larger number of units are required at a
known point in time.
This paper is directed toward an efficient
ordering policy recognizing both preventive
maintenance requirements, and requirements due
to random failure of units in service. The premiseis that greater efficiency will be realized by
ed.
184 T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190
addressing these two sources of demand in a uni-fied manner, relative to the use of separate order-
ing policies for the two demand streams.
This scenario is closely related to a system of
‘‘age replacement’’ and associated inventory
ordering policies. Under age replacement, a ran-
domly failing item is replaced upon reaching some
specified age T , or upon failure, whichever occurs
first. Barlow and Proschan (1965) originally stud-ied the problem in terms of finding the best
replacement age T . Subsequently, Kaio and Osaki
(1978, 1981) examined the situation where spares�inventory ordering policy and replacement age are
jointly optimized. Park and Park (1986) extended
this model to the case of random lead time. Kabir
and Al-Olayan (1994, 1996) further extended the
analysis to the case of multiple units in service, andthe possibility of holding more than a single unit of
inventory.
Whereas the age replacement literature assumes
units are replaced upon reaching a certain number
of periods in service, the current paper is addressed
toward a scenario where the system in question is
shut down for preventive maintenance or overhaul
at regularly scheduled intervals. The timing of thepreventive replacement is not necessarily dictated
by the age of any specific unit, but rather, on the
requirements of the larger system in which the unit
is used. As such, we treat the time between pre-
ventive maintenance operations as a given
parameter, the determination of which is beyond
the scope of the present investigation.
We assume a system having n identical parts inservice (see Walker, 1997). Random failure of
units in service generates intermittent, single-unit
demands between preventive maintenance periods.
At preventive maintenance, all units in service are
inspected, and a decision is made as to which ones
should be proactively replaced. Providing suffi-
cient inventory in support of preventive mainte-
nance is of paramount importance, so as not todelay a schedule of inter-related preventive main-
tenance activities.
We should note that the present investigation is
directed toward the case of spare parts in support
of plant maintenance and equipment, rather than
spare parts in support of equipment at a number
of geographically dispersed customer sites, or in
support of product service or warranty repair. Thispaper is also directed toward the case where the
parts in question are not repairable. Kennedy et al.
(2002) provide a recent survey of literature on
spare parts inventories.
The remainder of the paper is organized as fol-
lows. In Section 2 a stochastic dynamic program-
ming model is formulated, implicitly defining the
optimal policy for the scenario at hand. The char-acteristics of the optimal policy, examined through
numeric evaluation of the dynamic programming
model, are discussed in Section 3. Section 4 pre-
sents a comparison between the cost of the optimal
policy vs. a that of a simpler policy which treats the
two sources of demand separately. Summary and
possible extensions are discussed in Section 5.
2. Model development
Assume a system containing n identical com-
ponents. Each component fails according to a
constant failure rate k, such that the time to failurefor each component is exponentially distributed
with mean 1=k periods, independent of theremaining n� 1 units. The total number of part
failures during any given period then has Poisson
distribution with mean kn. Failed units must be
replaced from existing inventory, or expedited.
Scheduled preventive maintenance occurs every
T periods, during which the n units in service are
inspected, and some or all are replaced in a pro-
active manner. We assume the number of unitsreplaced is binomial with parameters n and p. Aswith the failure replacement, demands due to
preventive maintenance must be satisfied through
existing inventory or expedited.
We desire an optimal ordering policy which
accounts for both sources of demand as the pre-
ventive maintenance period approaches. Toward
this end, we will develop a stochastic dynamicprogramming characterization of the optimal
ordering policy.
For the derivations to follow it will be conve-
nient to adopt a common notation to represent
both the random failures and preventive mainte-
nance demands. Define DðtÞ as the number of units
required during period t. Without loss of general-
T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190 185
ity, assume period 0 is a preventive maintenanceperiod. Following dynamic programming conven-
tion, we will number the periods backwards in
time, thus period 1 is the period immediately prior
to the scheduled preventive maintenance in period
0, etc. Following this approach, periods 0; T ;2T ; . . . represent successively earlier preventive
maintenance periods, generating binomial ðn; pÞdemands on the inventory system. Any period t forwhich t mod T 6¼ 0 is an operational period, during
which the random failure process generates Pois-
son-distributed demand. In general, then, we have:
PtðdÞ ¼ P ðDðtÞ ¼ dÞ ¼ e�knðknÞd=d!;d ¼ 0; 1; 2; . . . ; t ¼ 1; 2; 3; . . . ; tmodT 6¼ 0;
ð1Þ
PtðdÞ ¼ P ðDðtÞ ¼ dÞ ¼n
d
� �pdð1� pÞðn�dÞ
;
d ¼ 0; 1; 2; . . . ; n; t ¼ 0; T ; 2T ; . . . ð2Þ
Let CH equal the carrying cost per unit, per
period, and CO the fixed cost of placing an order for
any quantity greater than zero units. Let B1 rep-
resent the penalty cost associated with inventory
shortage in the case of random part failure between
preventive maintenance periods. Similarly, let B2
represent the penalty cost associated with inven-
tory shortage during the preventive maintenance
operation. In both cases, the penalty includes the
cost of emergency ordering and expediting, as well
as a penalty associated with process downtime or
costly delay of the preventive maintenance project.
As above, it will be convenient to adopt a common
notation for the shortage penalty:
BðtÞ ¼ B1; t ¼ 1; 2; 3; . . . ; tmodT 6¼ 0; ð3Þ
BðtÞ ¼ B2; t ¼ 0; T ; 2T ; . . . ð4ÞLet xðtÞ represent the inventory on hand at the
beginning of period t. Let QðtÞ represent the
number of units ordered at the beginning of period
t, in response to observing xðtÞ units on hand.
These units will not be received into inventory
until the end of period t, i.e., after a one-period
lead time. The length of an operational period is
then implicitly assumed to be equal to the order
lead time. The preventive maintenance period isnot necessarily of equal length.
Inventory shortage results if period t demandDðtÞ is greater than xðtÞ. The shortage in any given
period is then ½DðtÞ � xðtÞ�þ, where ½y�þ ¼maxðy; 0Þ. Any shortages that occur during periodt are expedited, and any outstanding order quan-
tity QðtÞ is reduced by the number of units expe-
dited. Ending inventory is then ½xðtÞ � DðtÞþQðtÞ�þ. Under the expediting assumption, period
t � 1 begins with n units in operation.
Following the dynamic programming approach
of Bellman (1957), we will define the optimal pol-
icy using a recursive equation. Define TCtðxðtÞÞ asthe minimum total expected cost of the inventory
system over periods t; t � 1; t � 2; . . . ; 0, if xðtÞunits are on hand at the beginning of period t. Wecan recursively express TCtðxðtÞÞ as follows:TCtðxðtÞÞ
¼ minQðtÞ
COdðQðtÞÞ(
þX1d¼0
½BðtÞ½d � xðtÞ�þ
þ CH½xðtÞ � d þ QðtÞ�þ
þ TCt�1ð½xðtÞ � d þ QðtÞ�þÞ�PtðdÞ);
t ¼ 1; 2; 3; . . . ð5Þ
The function dðqÞ ¼ 0 if q ¼ 0, while dðqÞ ¼ 1 if
q > 0. Thus, the term COdðQðtÞÞ applies the
ordering cost CO if an order is placed for any
quantity greater than zero. The summation term
represents an expected cost over all possible DðtÞ,using the appropriate PtðdÞ ¼ P ðDðtÞ ¼ dÞ for theperiod at hand. Here the appropriate shortage
penalty is applied to the shortage quantity½d � xðtÞ�þ, carrying costs are levied on ending
inventory ½xðtÞ � d þ QðtÞ�þ, and period t � 1 is
begun with ½xðtÞ � d þ QðtÞ�þ units on hand.
The boundary condition can be written as
TC0ðxð0ÞÞ ¼Xn
d¼xð0Þþ1Bð0Þðd � xð0ÞÞP0ðdÞ; ð6Þ
i.e., the expected penalty cost if the preventive
maintenance operation scheduled for period 0 is
begun with xð0Þ units on hand.
0
1
2
3
4
5
6
1112131415161718191k = # of Periods Until Preventive Maintenance
s(k)S(k)
Fig. 1. Optimal ðsðktÞ; SðktÞÞ ordering policy n ¼ 5, k ¼ 0:001,
CO ¼ 25, CH ¼ 0:5, B1 ¼ 1000.
186 T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190
3. Policy characterization
Analytical evaluation of the optimal policy is
difficult, given the non-stationary nature of the
problem at hand. We will therefore examine the
results of numerical characterization of the opti-
mal policy, derived through simple backward
evaluation of the recursive relationship given in (5)and (6). The optimal policy was evaluated over all
32 combinations of the following parameter levels:
n ¼ 5, 20; k ¼ 0:001, 0.01; CO ¼ 25, 50; CH ¼ 0:5,2; B1 ¼ 300, 1000. These parameters are varied
relative to a constant B2 ¼ 1000. The preventive
maintenance requirements have binomial ðn; pÞdistribution, with p ¼ 0:7. While not required by
the model, the motivation for B2 PB1 applied hereis that shortage during preventive maintenance
upsets the entire maintenance schedule, possibly
idling a number of related preventive maintenance
activities.
At each t, TCðxðtÞÞ was evaluated over xðtÞ ¼0; 1; 2; . . . ; 2n. The time between preventive main-
tenance periods is T ¼ 100 in all cases. The
recursive evaluation was extended backwardsthrough t ¼ 499, thus the evaluation represents
five successive preventive maintenance intervals.
In every case, the optimal policies covering
t ¼ ½299–200�, [399–300], and [499–400] were
identical. Thus, once the analysis extends far en-
ough back to escape end-of-horizon effects, the
optimal policy becomes ‘‘cyclical-stationary’’ in
the sense that the policy depends on t only throughtmodT , i.e. the number of periods remaining untilthe next preventive maintenance period. The pol-
icy covering periods (99–0), and occasionally
periods (199–100), displayed minimal differences
from the ‘‘cyclical-stationary’’ policy. These end-
of-horizon effects are primarily characterized by
some nervousness in the SðktÞ parameter discussedbelow, in the periods immediately prior to pre-ventive maintenance.
The ‘‘cyclical-stationary’’ optimal policy has the
form ðsðktÞ; SðktÞÞ, where kt ¼ tmodT is the num-
ber of periods at time t until the next scheduled
preventive maintenance. Thus QðtÞ ¼ 0 if xðtÞ >sðktÞ, while QðtÞ ¼ SðktÞ � xðtÞ if xðtÞ6 sðktÞ.
The optimal policy is obviously non-stationary
in the narrower sense that the ordering parameters
depend on kt. As the preventive maintenance de-mand becomes sufficiently distant, however, the
optimal policy becomes concerned only with
the failure replacement demand process. The
ðsðktÞ; SðktÞÞ policy thus converges to a stationary
ðs; SÞ policy as kt becomes large. For clarity in
subsequent discussion, we will refer to this policy
as ðsFR; SFRÞ, where the FR subscript refers to
failure replacement. As the stationary ðs; SÞ policyhas been widely studied and is well understood, we
will focus here on the deviation from the ðsFR; SFRÞpolicy as kt approaches 0.
Fig. 1 (n ¼ 5, k ¼ 0:001, CO ¼ 25, CH ¼ 0:5,B1 ¼ 1000) displays one of the most predominant
and intuitive behaviors of the optimal ðsðktÞ; SðktÞÞpolicy. (The shaded regions in Figs. 1–5 are
bounded above by SðktÞ and below by sðktÞ.) Inthis case sðktÞ ¼ 0 until kt ¼ 1, at which point sð1Þjumps to 4. This essentially guarantees that an
order will be placed in period 1 to cover the pre-
ventive maintenance demand, if such an order has
not been placed already. This increase in sðktÞnever occurs prior to period 1 in any of the cases
evaluated. Intuitively, the model suggests there is
no point in forcing an order to cover preventivemaintenance demand any sooner than necessary.
SðktÞ, on the other hand, increases from SFR ¼ 1
to SðktÞ ¼ 5 at kt ¼ 13. The intuition here is that if
an order is to be placed (as governed by the sðktÞparameter discussed above) the SðktÞ parameter
should at some point recognize the preventive
0
1
2
3
4
5
6
1112131415161718191
k = # of Periods Until Preventive Maintenances(k)S(k)
Fig. 2. Optimal ðsðktÞ; SðktÞÞ ordering policy n ¼ 5, k ¼ 0:001,
CO ¼ 50, CH ¼ 0:5, B1 ¼ 1000.
0
1
2
3
4
5
6
7
1112131415161718191k = # of Periods Until Preventive Maintenance
S(k) s(k)
Fig. 3. Optimal ðsðktÞ; SðktÞÞ ordering policy n ¼ 5, k ¼ 0:01,
CO ¼ 25, CH ¼ 2, B1 ¼ 1000.
T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190 187
maintenance demand. In Fig. 1, the order-up-to
level accounts for preventive maintenance demand
as early as 13 periods prior to the preventivemaintenance itself.
Fig. 2 (n ¼ 5, k ¼ 0:001, CO ¼ 50, CH ¼ 0:5,B ¼ 1000) provides a useful contrast to Fig. 1, in
that CO has increased from 25 to 50 with all other
parameters held constant. The distinguishing
characteristic, again common to all scenarios, is
that the order-up-to level begins to account for
preventive maintenance demand at larger kt as CO
increases.
Fig. 3 (n ¼ 5, k ¼ 0:01, CO ¼ 25, CH ¼ 2,
B ¼ 1000) demonstrates that when sFR > 0, the
optimal ðsðktÞ; SðktÞÞ policy may actually implydecreasing sðktÞ for some number of periods priorto the eventual spike at kt ¼ 1. This has the effect
of forgoing an order that would have been placed
under a larger sðktÞ. In Fig. 3, as the inevitable
order to cover preventive maintenance demand
draws near, the policy responds to the probability
that a current inventory xðtÞ ¼ 1 will suffice until
receipt of that order. In short, the number ofperiods of exposure to the lower inventory levels
does not merit the cost of a replenishment
order, given that a subsequent replenishment order
will soon follow. At larger kt the ‘‘window of
exposure’’ to stock-out is large enough that an
order is justified when xðtÞ ¼ 1.
Fig. 4 (n ¼ 5, k ¼ 0:01, CO ¼ 50, CH ¼ 0:5,B ¼ 1000) displays another common characteristicof the optimal policy, in that SðktÞ decreases fromSFR for some interval prior to the eventual in-
crease. As in Fig. 3, the policy is responding to the
fact that a subsequent order must surely be placed
no later than kt ¼ 1. As such, the order-up-to level
need not cover the same number of periods� de-mand than it would if there were no impending
order to cover preventive maintenance require-ments. The SðktÞ parameter is essentially attempt-
ing to avoid carrying large inventory into a period
that will experience an order driven by preventive
maintenance requirements. Interestingly, this
behavior reflects logic similar to that underlying
the Wagner–Whitin algorithm for deterministic
process lot-sizing (Wagner and Whitin, 1958).
Finally, Fig. 5 (n ¼ 20, k ¼ 0:01, CO ¼ 25,CH ¼ 0:5, B ¼ 1000) displays a policy in which all
these effects are present simultaneously. The gen-
eral impression is that of a step-wise ‘‘draw down’’
of the ordering parameters, prior to the eventual
increase to cover preventive maintenance demand.
First intuition may suggest sðktÞ and Sðkt) willboth steadily increase as kt decreases, representinga gradual build up of inventory in anticipation ofthe upcoming preventive maintenance demand.
The basic counter-argument to this logic is that
that timing of the preventive maintenance demand
is known. As such, there is no need for a larger sparameter until the latest possible moment. In-
deed, the s parameter may actually decrease for
reasons discussed within the context of Fig. 3. The
0
1
2
3
4
5
6
7
8
9
1112131415161718191
k = # of Periods Until Preventive Maintenances(k)S(k)
Fig. 4. Optimal ðsðktÞ; SðktÞÞ ordering policy n ¼ 5, k ¼ 0:01,
CO ¼ 50, CH ¼ 0:5, B1 ¼ 1000.
0
5
10
15
20
25
1112131415161718191k = # of Periods Until Preventive Maintenance
s(k)S(k)
Fig. 5. Optimal ðsðktÞ; SðktÞÞ ordering policy n ¼ 20, k ¼ 0:01,
CO ¼ 25, CH ¼ 0:5, B1 ¼ 1000.
188 T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190
S parameter may decrease as well, for reasons
discussed within the context of Fig. 4. Eventually,
of course, SðktÞ must increase to cover the pre-
ventive maintenance demand, while sðktÞ increasesat kt ¼ 1 to force such an order if one has not yet
occurred ‘‘naturally’’.
4. Cost comparison
To characterize the cost savings afforded by the
optimal ðsðktÞ; SðktÞÞ policy, we will compare the
cost to that of a simpler policy which addresses
the failure replacement and preventive mainte-
nance demands separately. Under this policy, theðsFR; SFRÞ policy is used to satisfy failure replace-
ment demand through all periods t in which kt 6¼ 1.
An optimal single period ordering policy is used at
kt ¼ 1 to provide for the preventive maintenance
demand. This policy is given by the optimal ðs; SÞat t ¼ 1 of the backward evaluation of (5) and (6),
and will be denoted as ðsPM; SPMÞ. In general, then,we have:
ðs0ðktÞ; S0ðktÞÞ ¼ ðsFR; SFRÞ; kt 6¼ 1; ð7Þ
ðs0ðktÞ; S0ðktÞÞ ¼ ðsPM; SPMÞ; kt ¼ 1: ð8Þ
Let TC0tðxðtÞÞ represent the expected cost of fol-
lowing this policy through periods t; t � 1;t � 2; . . . ; 0. We then have:
TC00ðxð0ÞÞ ¼ TC0ðxð0ÞÞ ð9Þ
and
TC01ðxð1ÞÞ ¼ TC1ðxð1ÞÞ; ð10Þ
where TC0ðxð0ÞÞ and TC1ðxð1ÞÞ are given by (6) and(5), respectively.
Let Q0ðtÞ ¼ S0ðktÞ � xðtÞ for xðtÞ6 s0ðktÞ, and
Q0ðtÞ ¼ 0 for xðtÞ > s0ðktÞ. For tP 2 the cost of the
simpler policy is then given by
TC0ðxðtÞÞ
¼ COdðQ0ðtÞÞ þX1d¼0
½BðtÞ½d � xðtÞ�þ
þ CH½xðtÞ � d þ Q0ðtÞ�þ
þ TC0t�1ð½xðtÞ � d þ Q0ðtÞ�þÞ�PtðdÞ: ð11Þ
As kt ! 1 the ratio TC0tðxðtÞÞ=TCtðxðtÞÞ ! 1, as
the portion of total cost affected by ðsðktÞ;SðktÞÞ 6¼ ðsFR; SFRÞ is diminished. Similarly, any
long-run comparison of the two policies is con-
founded by the number of periods comprising the
preventive maintenance interval. To eliminate this
effect from the comparison, we will evaluate the
relative performance of the two policies condi-tionally, under various ðt; xðtÞÞ in the interval
where ðsðktÞ; SðktÞÞ 6¼ ðsFR; SFRÞ.Table 1 displays the maximum value over all xðtÞ
of the ratio TC0tðxðtÞÞ=TCtðxðtÞÞ for the 32 scenarios
evaluated, at t ¼ 2; 3, and 10. As can be seen, the
optimal policy may generate considerable cost
Table 1
Cost comparison of optimal ðsðktÞ; SðktÞÞ policyScenario n k CO CH B1 Max. TC0
10ðxð10ÞÞ=TC10ðxð10ÞÞ
Max. TC03ðxð3ÞÞ
=TC3ðxð3ÞÞMax. TC0
2ðxð2ÞÞ=TC2ðxð2ÞÞ
1 5 0.001 25 0.5 300 1.3971 1.7121 1.7745
2 5 0.001 25 0.5 1000 1.1231 1.5343 1.6276
3 5 0.001 25 2 300 1.0000 1.0000 1.0000
4 5 0.001 25 2 1000 1.0086 1.3773 1.4806
5 5 0.001 50 0.5 300 1.6334 1.8394 1.8752
6 5 0.001 50 0.5 1000 1.3847 1.7127 1.7793
7 5 0.001 50 2 300 1.0000 1.0000 1.0000
8 5 0.001 50 2 1000 1.2277 1.5842 1.6639
9 5 0.01 25 0.5 300 1.1149 1.4233 1.4956
10 5 0.01 25 0.5 1000 1.1556 1.5855 1.7036
11 5 0.01 25 2 300 1.0153 1.1104 1.2594
12 5 0.01 25 2 1000 1.1732 1.4618 1.5365
13 5 0.01 50 0.5 300 1.3538 1.6264 1.6870
14 5 0.01 50 0.5 1000 1.415 1.7671 1.8394
15 5 0.01 50 2 300 1.0469 1.3282 1.4661
16 5 0.01 50 2 1000 1.0583 1.2721 1.3507
17 20 0.001 25 0.5 300 1.0205 1.2827 1.4253
18 20 0.001 25 0.5 1000 1.0084 1.0974 1.2461
19 20 0.001 25 2 300 1.0165 1.2169 1.2784
20 20 0.001 25 2 1000 1.0016 1.0021 1.0727
21 20 0.001 50 0.5 300 1.0574 1.5126 1.6248
22 20 0.001 50 0.5 1000 1.0280 1.3235 1.4555
23 20 0.001 50 2 300 1.1555 1.4047 1.4572
24 20 0.001 50 2 1000 1.0080 1.1081 1.2525
25 20 0.01 25 0.5 300 1.0437 1.1554 1.3380
26 20 0.01 25 0.5 1000 1.0650 1.2681 1.4441
27 20 0.01 25 2 300 1.0208 1.0324 1.1769
28 20 0.01 25 2 1000 1.0303 1.1581 1.2782
29 20 0.01 50 0.5 300 1.1207 1.4210 1.5672
30 20 0.01 50 0.5 1000 1.1469 1.5131 1.6501
31 20 0.01 50 2 300 1.0364 1.1987 1.3595
32 20 0.01 50 2 1000 1.0197 1.0300 1.1321
T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190 189
savings, in the event that the inventory trajectoryprovides an ordering opportunity shortly prior to
period 1. (The minimal end-of-horizon effects in-
cluded in this comparison are unavoidable, given
the structural difference between the optimal vs.
sub-optimal policies, coupled with our interest in a
conditional rather than a long-run cost compari-
son. These end-of-horizon effects tend to work in
favor of the myopic sub-optimal policy.)In general, the cost savings tend to be greatest
at small CH and large CO. Given the source of
the optimal policy�s advantage, this is under-
standable. The advantage of the optimalðsðktÞ; SðktÞÞ policy lies in properly recognizing
the preventive maintenance demand, and includ-
ing those units in an order placed within some
proper interval prior to the preventive mainte-
nance. The savings associated with this ability to
economize on ordering cost are naturally greatest
when CO is large. Conversely, since the optimal
policy implies ordering the preventive mainte-nance units earlier than under the simpler policy,
the advantage of doing so is greatest when CH is
small.
190 T.S. Vaughan / European Journal of Operational Research 161 (2005) 183–190
Finally, note that scenarios 3 and 7 in Table 1represent a special case (small kn, large CH, small
B1) in which no inventory is carried for the pur-
pose of failure replacement under the ðsFR; SFRÞpolicy. In these cases, the optimal ðsðktÞ; SðktÞÞpolicy is identical to the simpler policy.
5. Conclusion
This paper has presented a dynamic program-
ming characterization for a spare parts ordering
policy. The situation investigated is that in which
demand for spare parts arises from two sources,
i.e. random failure of units in service, and bulk
replacement at regularly scheduled preventive
maintenance intervals. The optimal policy isshown to have the form ðsðktÞ; SðktÞÞ, where kt isthe number of periods until the next preventive
maintenance operation.
The optimal ðsðktÞ; SðktÞÞ policy is generally
characterized by SðktÞ increasing to cover the
preventive maintenance demand some number of
periods prior to the preventive maintenance
period. This increase in SðktÞ occurs farther inadvance of the preventive maintenance period, the
larger is the fixed cost per order, and the smaller is
the carrying cost per unit. The sðktÞ parameter
increases in the period prior to the preventive
maintenance operation, to a level which forces an
order to cover the preventive maintenance de-
mand.
Prior to their respective increases, sðktÞ and/orSðktÞ may decrease from their values under the
stationary failure replacement policy. This draw-
down of the two ordering parameters is in recog-
nition of the impending order that must be placed
prior to the preventive maintenance operation.
The appropriate magnitude and duration of the
alternative ðsðktÞ; SðktÞÞ depend highly on the de-
mand and cost parameters. Nonetheless, a simplebackward evaluation of the recursive relationship
presented in this paper provides an optimal policy
for a given parameter set.
Extensions of this paper include the case of lead
time greater than one period, and more impor-
tantly, the case of a random lead time. Intuitively,
this would cause the increase in sðktÞ to occur prior
to kt ¼ 1, and may impact the policy in other waysas well. Another possible extension would be to
assume an increasing failure rate for units in ser-
vice, rather than the constant failure rate implied
by the exponential distribution. This could have
the effect of prescribing lower inventory levels
immediately after preventive maintenance, when
presumably a larger number of new units are in
operation. Unfortunately, these extensions willprobably render an intractable model under the
dynamic programming approach, and will require
some other means of evaluation.
Acknowledgements
The author would like to thank the two anon-ymous reviewers for their constructive comments,
which helped to greatly improve the original
manuscript.
References
Barlow, R.E., Proschan, F., 1965. Mathematical Theory of
Reliability. Wiley, New York.
Bellman, R.E., 1957. Dynamic Programming. Princeton Uni-
versity Press, Princeton, NJ.
Kabir, A.B.M.Z., Al-Olayan, A.S., 1994. Joint optimization of
age replacement and continuous review spare provisioning
policy. International Journal of Operations and Production
Management 14 (7), 53–69.
Kabir, A.B.M.Z., Al-Olayan, A.S., 1996. A stocking policy for
spare part provisioning under age based preventive replace-
ment. European Journal of Operational Research 90 (1),
171–181.
Kaio, N., Osaki, S., 1978. Optimum ordering policies with lead
time for an operating unit in preventive maintenance. IEEE
Transactions on Reliability 27, 270–271.
Kaio, N., Osaki, S., 1981. Optimum planned maintenance
policies with lead time. IEEE Transactions on Reliability 30,
79.
Kennedy, W.J., Patterson, J.W., Fredendall, L.D., 2002. An
overview of recent literature on spare parts inventories.
International Journal of Production Economics 76, 201–
215.
Park, Y.T., Park, K.S., 1986. Generalized spare ordering
policies with random lead time. European Journal of
Operational Research 23 (3), 320–330.
Wagner, H.M., Whitin, T.M., 1958. Dynamic version of the
economic lot size model. Management Science 5, 89–96.
Walker, J., 1997. Base stock level determination for ‘‘insurance
type’’ spares. International Journal of Quality and Reliabil-
ity Management 14 (6), 569–574.