Information Sciences 178 (2008) 3382–3392

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

Integrated production planning and preventive maintenance in

deteriorating production systems
El-Houssaine Aghezzaf a,*, Najib M. Najid b
a
Department of Industrial Management, Ghent University, Technologiepark 903, B-9052 Zwijnaarde, Belgium
b
Department GMP, IUT de Nantes, 2 Av. du Prof. Jean Rouxel, PB. 595-44475 Carquefou Cedex, France

a r t i c l e i n f o a b s t r a c t

Article history: This paper discusses the issue of integrating production planning and preventive mainte-
Received 20 March 2007 nance in manufacturing production systems. In particular, it tackles the problem of inte-
Received in revised form 8 March 2008 grating production and preventive maintenance in a system composed of parallel
Accepted 4 May 2008
failure-prone production lines. It is assumed that when a production line fails, a minimal
repair is carried out to restore it to an ‘as-bad-as-old’ status. Preventive maintenance is car-
ried out, periodically at the discretion of the decision maker, to restore the production line
to an ‘as-good-as-new’ status. It is also assumed that any maintenance action, performed
Keywords:
Production planning
on a production line in a given period, reduces the available production capacity on the line
Preventive maintenance during that period. The resulting integrated production and maintenance planning
Failure-prone manufacturing systems problem is modeled as a nonlinear mixed-integer program when each production line
implements a cyclic preventive maintenance policy. When noncyclical preventive mainte-
nance policies are allowed, the problem is modeled as a linear mixed-integer program. A
Lagrangian-based heuristic procedure for the solution of the first planning model is pro-
posed and discussed. Computational experiments are carried out to analyze the perfor-
mance of the method for different failure rate distributions, and the obtained results are
discussed in detail.
Ó 2008 Elsevier Inc. All rights reserved.

1. Introduction

To cope with the current tough competition many manufacturing companies have invested in highly automated produc-
tion systems with sophisticated equipments. To be economically sustainable, these costly equipments should be exploited to
the last instant of their maximum possible productive time. When an unplanned downtime, caused by a production line fail-
ure, occurs it often trims down the system’s productivity and renders the current production plan obsolete. Revising the pro-
duction plan in an emergency situation is usually very expensive and often causes increased variability in product quality
and in service level. It is, therefore, essential that production planning and preventive maintenance activities be carried
out in an integrated way to hedge against these often avoidable failures and re-planning occurrences. This paper proposes
and discusses models to generate such integrated production and maintenance plans which aim at achieving an optimal
trade-off between the various production and maintenance costs.
The issue of integrating production planning and preventive maintenance, in failure-prone manufacturing systems, is
becoming an active area of research due to its importance in the current highly competitive environment. However, it is fre-
quently tackled at the operational (the scheduling) level. In particular, scheduling problems on unreliable machines have re-
ceived a considerable attention in the literature [1,6,11,12,17,18]. A large number of these scheduling problems assume that

* Corresponding author.
E-mail addresses: [email protected] (E.-H. Aghezzaf), [email protected] (N.M. Najid).

0020-0255/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.ins.2008.05.007
 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392 3383

maintenance periods are known in advance, and thus reduce to scheduling problems with machine availability constraints
[1,15,19]. A recent survey on scheduling problems with limited machine availability can be found in [20]. A general result is
that these scheduling problems, with availability constraints, under different machine configurations and various objective
functions are NP-hard [13,16]. Various heuristics as well as exact methods for their solution are proposed [3,8].
At the tactical level, there are only few papers discussing this issue. Wienstein and Chung [23] presented a three-part
model to resolve the conflicting objectives of system reliability and profit maximization. An aggregate production plan is first
generated, and then a master production schedule is developed to minimize the weighted deviations from the specified
aggregate production goals. Finally, work center loading requirements, determined through rough cut capacity planning,
are used to simulate equipment failures during the aggregate planning horizon. Several experiments are used to test the sig-
nificance of various factors for maintenance policy selection. These factors include the category of maintenance activity,
maintenance activity frequency, failure significance, maintenance activity cost, and aggregate production policy. Aghezzaf
et al. [4] presented an integrated production and preventive maintenance planning model for a single-line production sys-
tems which can be minimal repaired at failure. They assumed that maintenance actions carried out on the production line
reduce its capacity, and proposed a mathematical programming model to establish an optimal integrated production and
maintenance plan for the single-line production systems.
In Section 2, a mathematical programming model for the integrated production and maintenance problem in production
systems implementing cyclic preventive maintenance policies is proposed. In Section 3, a heuristic solution algorithm for the
solution of the proposed model is proposed and discussed. Section 4 presents an integrated production and maintenance
planning model for systems implementing general preventive maintenance policies. The last section presents some compu-
tational experiment to analyze the performances of the proposed models and their algorithms. Finally, some concluding re-
marks are presented in Section 6.

2. Integrated production planning and cyclic preventive maintenance

Consider a planning horizon H of length T ¼ N s covering N periods of fixed length s, and a set of items i 2 P to be pro-
duced on a set of capacitated parallel production lines j 2 L during this planning horizon. Each production line j 2 L of the
system produces each item i 2 P at a known production rate, expressed in product units per unit of time. During each period
t 2 H, a demand dit of item i 2 P should be satisfied. It is assumed that each production line j 2 L has a known nominal capac-
ity (given in time units) and is denoted by jj . Each preventive or corrective maintenance action performed on the production
line consumes a certain percentage of this capacity. Thus, each preventive maintenance action and each corrective mainte-
nance action on production line j 2 L consume, respectively hpj ¼ ajj and hrj ¼ bjj capacity units (where 0 6 a 6 1 and
0 6 b 6 1). It is also assumed that the failure distribution of each production line j 2 L is known. Let fj ðtÞ and F j ðtÞ denote
its corresponding probability density and cumulative distribution functions, respectively. Let rj ðtÞ denotes the failure rate
function of the production line j 2 L at time t. It is well known that rj ðtÞ is given by
fj ðtÞ
rj ðtÞ ¼ :
1  F j ðtÞ
Finally, assume that preventive maintenance tasks are performed periodically on each production line j 2 L, in the beginning
of periods t ¼ 1; ðnj þ 1Þ; ð2nj þ 1Þ; ð3nj þ 1Þ; . . . ; N, that corrective maintenance tasks are carried out on the production line
when a failure occurs, and that any maintenance action is completed within the period in which it started. Fig. 1 shows an
example of an integrated production and cyclic preventive maintenance plan and also the preventive maintenance cycle

Fig. 1. An integrated production and cyclic preventive maintenance plan.

3384 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392

for each production line. The parameters T 1 ; T 2 , and T m shown in Fig. 1 are maintenance cycles of the corresponding
lines.

2.1. A production and cyclic preventive maintenance planning model

The proposed mathematical program here models the problem of determining optimal integrated production and cyclic
preventive maintenance plans in a multi-line production system. It assumes that when a production line fails, a minimal
repair is carried out to restore it to an ‘as-bad-as-old’ status. When a preventive maintenance is carried out, the production
line is restored to an ‘as-good-as-new’ status. It also assumes that the expected failures increase with elapsed time since the
last preventive maintenance.
To define the parameters of the model, let cpj be the cost of preventively maintaining production line j 2 L in the beginning
of each maintenance cycle, and crj be the cost of performing a corrective maintenance action on this line when a failure oc-
curs. Also, let fitj be the setup cost of producing item i 2 P in period t 2 H on production line j 2 L; pjit be the variable cost of
j
producing item i 2 P in period t 2 H on production line j 2 L; hit be the inventory holding cost of item i 2 P in period t 2 H, and
qij be the processing time of item i 2 P on production line j 2 L (expressed in time units per item). For the variables of the
model, let xjit be the quantity of item i 2 P produced on line j 2 L during period t 2 H; Iit be the inventory of item i 2 P at
the end of period t 2 H; T j ¼ nj s be the length of the preventive maintenance cycle for production line j 2 L, and yjit be a binary
decision variable set to 1 if item i 2 P is produced on line j 2 L during period t 2 H and 0 otherwise.

2.1.1. The mathematical model

! Z !
XX X X Tj
ðPCPMÞ Minimize Z PM ¼ ðfitj yjit þ pjit xjit Þ þ hit Iit þ dT=T j e cpj þ crj r j ðtÞdt
t2H i2P j2L j2L 0
X
subject to xjit þ Ii;t1  Iit ¼ dit for all i 2 P; t 2 H; ð1Þ
j2L
X
qij xjit 6 jj ðtÞ for all j 2 L; t 2 H; ð2Þ
i2P
i
xjit  dtN yjit 6 0 for all i 2 P; j 2 L; t 2 H; ð3Þ
X j
yit 6 1 for all j 2 L; t 2 H; ð4Þ
i2P

xjit ; Iit P 0; yjit 2 f0; 1g; T j 2 fs; 2s; . . . ; Nsg for all i 2 P; j 2 L; t 2 H;
i P
where dtN ¼ Ns¼t dis , and the function jj ðtÞ defines the available capacity of the production line j 2 L in period t 2 H as is ex-
plained below.
Constraints (1) are the flow conservation constraints defined for any item i 2 P and any period t 2 H. They guarantee that
the available inventory of an item augmented with the quantity produced, on each production line, of that item is sufficient
to satisfy its demand in period t. The remainder is stocked for subsequent periods. Constraints (2) are the capacity restric-
tions defined for each production line j 2 L in each period t 2 H. They guarantee that the quantity produced on a production
line does not exceed the available capacity of that line in period t. Constraints (3) force payments of the fixed costs, in any
period t 2 H, for any item i 2 P that is produced during that period. Constraints (4) ensure that at any moment no more than
one product is scheduled on a production line.
It is time now to define exactly how the available capacity function of each production line in each period is determined.
Consider a production line j 2 L, and assume that its preventive maintenance cycle is given by T j ¼ nj s. Assume also that a
maintenance action on this line took place in the beginning of a period s 2 H, and that it covers the periods
fs; s þ 1; . . . ; ðs þ nj  1Þg. The available capacity of the production line j 2 L in period s is then given by equation:
Z s
jj ðsÞ ¼ jj  hpj  hrj rj ðuÞdu: ð5Þ
0

The capacity jj ðtÞ for the periods t 2 fs þ 1; . . . ; ðs þ nj  1Þg, during which only the necessary corrective maintenance actions
take place, is given by equation:
Z s
jj ðtÞ ¼ jj  hrj rj ðu þ ðt  sÞsÞdu: ð6Þ
0

Notice that when preventive maintenance periods are known in advance, the available capacity of each production line,
during each period, can be determined, and that the expected maintenance cost can be computed in advance. In this case, the
problem (PCPM) reduces to the usual multi-item multi-line capacitated lot-sizing problem. Therefore, one can conclude that
the problem (PCPM) belongs to the class of NP-hard problems since it contains, as a special case, the single-line capacitated
lot-sizing problem known to be NP-hard [7].
 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392 3385

2.1.2. An exact algorithm for problem (PCPM)

The variables ðT j Þj2L defining the preventive maintenance cycles, for the production lines of the system, make the objec-
tive function of the problem (PCPM) to become a nonlinear function. If these preventive maintenance cycles are known in
advanced the problem reduces to a multi-item capacitated multi-line lot-sizing problem. This latter problem can be solved
with any usual linear mixed integer solver like Cplex. Alternatively, many branch and bound, branch and cut, and branch and
price exact methods as well as many heuristic methods were proposed to solve the single-line version of this problem
[2,5,9,10,14,22]. A recent survey on the multi-item capacitated lot-sizing problem can be found in [24]. The following steps
describe an algorithm to solve the problem (PCPM):

 For each vector of preventive maintenance cycles ðn1 ; . . . ; nj ; . . . ; njLj Þ of the production lines j 2 L, with nj 2 f1; 2; . . . ; Ng,
perform the following two steps:
1. Compute the expected maintenance cost corresponding to the selected vector of maintenance cycles, and determine for
each production line j 2 L the available capacity jj ðtÞ in each period t 2 H.
2. Solve the resulting pure production planning problem, with the above computed available capacities ðjj ðtÞt2H Þj2L , and
add together the resulting expected maintenance and production costs.
 Rank these total costs and select the plan resulting in the least cost, this is the optimal integrated production and cyclical
preventive maintenance plan.

The number of possible combinations of the preventive maintenance cycle sizes is exponential of order N jLj , where N is the
number of periods in the planning horizon as defined above. Now, observe that when the number of production lines is very
small (2 or 3), then the complexity of the above procedure depends actually on the complexity of the algorithm used to solve
the pure capacitated lot-sizing problems.
The following section discusses a heuristic solution method, based on Lagrangian relaxation, for the problem (PCPM).
Since the single-line production and cyclic preventive maintenance planning problem plays a central role in this procedure,
a detailed procedure for solving this latter problem efficiently is given below.

2.2. Single-line integrated production and cyclic preventive maintenance planning model

Consider the problem of planning production and cyclic preventive maintenance on a single production line j 2 L, denoted by
ðPCPMjSL Þ. Assume that the demand fractions of items i 2 P, which are to be produced on the production line j during each period
t 2 H, are known in advanced and are denoted by djit . Let Ijit be new variables defining the proportion of inventory of item i 2 P at
the end of period t 2 H resulting from the production on line j 2 L. Assume as before that T j ¼ nj s denotes the length of the pre-
ventive maintenance cycle for the production line j 2 L. The problem ðPCPMjSL Þ can now be rewritten as follows:
XX j j
PCPMjSL Minimize Z jPM ¼ ðfit yit þ pjit xjit þ hit Ijit Þ þ UjPM ðnj Þ
t2H i2P

subject to xjit þ Iji;t1  Ijit ¼ djit for all i 2 P; j 2 L; t 2 H; ð7Þ

X
qij xjit 6 jj ðtÞ; for all j 2 L; t 2 H; ð8Þ
i2P
i
xjit  dtN yjit 6 0 for all i 2 P; j 2 L; t 2 H; ð9Þ
X j
yit 6 1 for all j 2 L; t 2 H; ð10Þ
i2P

xjit ; Ijit P 0; yjit 2 f0; 1g; nj 2 f1; 2; . . . ; Ng for all i 2 P; t 2 H;

R
where the
R
available capacities are given by jj ðtÞ ¼ jj  hpj  hrj 0s rj ðuÞdu if t ¼ ðn  1Þnj þ 1, and
r s
jj ðtÞ ¼ jj  hj 0 rj ðu þ ðt  ððn  1Þnj þ 1ÞÞsÞdu if ðn  1Þnj þ 1 6 t 6 nnj , for 1 6 n 6 dN=nj e and ðn  1Þnj þ 1 6 t 6 nnj with
t 6 N. The maintenance cost UjPM ðnj Þ is given by
0 1
dN=nj e
X nnj
X  Z s 
U j @c p r
cj rj ðu þ ðt  ðn  1Þnj  1ÞsÞdu A:
PM ðnj Þ ¼ j þ
n¼1 t¼ðn1Þnj þ1;t6N 0

It must be clear by now that the decision variables in this last model are the usual production variables xjit ; yjit , Ijit , and the
variable nj which defines the optimal size of the preventive maintenance cycle. To determine the optimal values of produc-
tion plan and the size of the maintenance period, one may proceed as follows.

2.2.1. A single production line algorithm

For each preventive maintenance cycle nj = 1 to N, corresponding respectively to the cases of one preventive maintenance
in the beginning of each production period, and to the case of one preventive maintenance in beginning of the whole plan-
ning horizon.
3386 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392

1. For each preventive maintenance cycle of size nj , determine the corresponding maintenance cost function UjPM ðnj Þ, then
n
determine for each period t 2 H the available capacity jj j ðtÞ.
2. Solve the resulting pure production planning problem ðPCPMjSL ðnj ÞÞ, for this fixed cycle of size nj , to obtain the optimal
b j ðnj Þ using any efficient specialized algorithm, or a usual linear mixed integer solver.
value Z PM
3. Compare the resulting values Z b j ðnj Þ for nj ¼ 1 to N. The value of n b j ðnj Þ is chosen as
^ j that results in the least value of Z
PM PM
the optimal preventive maintenance cycle size for the production line, and the optimal production plan resulting from the
j
solution of ðPCPMSL ðn^ j ÞÞ, is chosen as the optimal production plan for the line.

To take advantage of this single-line production and cyclic preventive maintenance algorithm, a heuristic solution meth-
od for the problem (PCPM), based Lagrangian relaxation decomposition, is proposed in the following section.

3. A Lagrangian relaxation based approximation procedure for (PCPM)

The mathematical program (PCPM) models the integrated production and preventive maintenance planning problem
on parallel production lines implementing cyclic preventive maintenance policies. To take advantage of the properties
of the single-line version of the problem, the problem (PCPM) is reformulated to make these single-line versions appear
as sub-problems. Two new sets of variables djit and gjit are introduced, both defining the fraction of demand of item i 2 P
produced on production line j 2 L in period t 2 H, (i.e. djit ¼ gjit ). The new variables Ijit , defining the proportion of inventory
of item i 2 P at the end of period t 2 H resulting from production on line j 2 L, are also introduced. The resulting model
becomes then.

3.1. A reformulated mathematical model ðPCPMR Þ

X XX j j  Z nj s !
ðPCPMR Þ Minimize Z PM ¼ ðfit yit þ pjit xjit þ hit Ijit Þ þ dN=nj e cpj þ crj r j ðtÞdt
j2L t2H i2P 0

subject to xjit þ Iji;t1  Ijit ¼ djit for all i 2 P; j 2 L; t 2 H; ð11Þ

X
qij xjit 6 jj ðtÞ for all j 2 L; t 2 H; ð12Þ
i2P
i
xjit  dtN yjit 6 0 for all i 2 P; j 2 L; t 2 H; ð13Þ
X j
yit 6 1 for all j 2 L; t 2 H; ð14Þ
i2P

djit  gjit ¼ 0 for all i 2 P; j 2 L; t 2 H; ð15Þ

X j
git ¼ dit for all i 2 P; t 2 H; ð16Þ
j2L

xjit ; Ijit ; djit ; gjit P 0; yjit 2 f0; 1g; nj 2 f1; 2; . . . ; Ng for all i 2 P; j 2 L; t 2 H:
Constraints (11) are the new flow conservation constraints defined for each production line j 2 L. Constraints (15) and (16)
guarantee, for each item i 2 P and each period t 2 H, that the sum of the quantities shipped from inventories accumulated by
the production lines j 2 L are sufficient to cover the demand of item i 2 P in period t 2 H.
Examining carefully the problem ðPCPMR Þ, one can observe that if constraints (15) are relaxed, the problem splits up into
ðjLj þ 1Þ sub-problems. The first jLj sub-problems consist of planning production and cyclical preventive maintenance on
each single production line j 2 L for a given demand vector dj . The last sub-problem is a simple linear program, which assigns
the demands to be produced to the production lines with the objective of minimizing the marginal production costs of each
production line.
Now, let kjit be the Lagrangian multipliers associated with constraints (15). The resulting sub-problem for each production
line j 2 L is given by
XX
PCPMjLR ðkÞ : Minimize V jPM ðkÞ ¼ Z jPM  kjit djit
t2H i2P

subject to xjit þ Iji;t1  Ijit ¼ djit for all i 2 P; j 2 L; t 2 H; ð17Þ

X
qij xjit 6 jj ðtÞ for all j 2 L; t 2 H; ð18Þ
i2P
i
xjit  dtN yjit 6 0 for all i 2 P; j 2 L; t 2 H; ð19Þ
X j
yit 6 1 for all j 2 L; t 2 H; ð20Þ
i2P
 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392 3387

xjit ; Ijit ; djit P 0; yjit 2 f0; 1g; nj 2 f1; 2; . . . ; Ng for all i 2 P; j 2 L; t 2 H;

where jj ðtÞ is the available capacity of the production line j 2 L in period t 2 H, defined by Eqs. (5) and (6), and Z jPM given by
XX j j  Z nj s 
Z jPM ¼ ðfit yit þ pjit xjit þ hit Ijit Þ þ dN=nj e cpj þ crj r j ðtÞdt :
t2H i2P 0
P P
As already mentioned, observe that if the vector d is given, then the quantity t2H i2P kjit djit in the objective function of
j
j
ðPCPMR ðkÞÞ becomes a constant. The resulting problem reduces then to the problem of planning production and cyclical pre-
ventive maintenance on a single production line, in which djit are the demands for each item i 2 P in each period t 2 H.
The last sub-problem ðDASPðkÞÞ determines, for each item i 2 P in each period t 2 H, the best assignment of the demand dit
to the production lines j 2 L, and is given bt
XXX
DASPðkÞ : Minimize V 0PM ðkÞ ¼ kjit gjit
j2L t2H i2P
X
subject to gjit ¼ dit for all i 2 P; t 2 H;
j2L

gjit P 0 for all i 2 P; j 2 L; t 2 H: ð21Þ

b j ðkÞ and V
If V b 0 ðkÞ are, respectively, the optimal values of the problems ðPCPMj ðkÞÞ and ðDASPðkÞÞ for a given k 2 R,
PM PM LR
then the Lagrangian dual associated with the above relaxed problem is given by
!
X
W LD ¼ Maximize b j b 0
V PM ðkÞ þ V PM ðkÞ :
k2R
j2L

In the following sections, a heuristic procedure to solve the problem (PCPM) using the sub-problems ðPCPMjLR ðkÞÞ and
ðDASPðkÞÞ is presented and discussed. The value obtained with this heuristic procedure is then compared to a very good low-
er bound on the optimal solution of the problem to evaluate and analyze the optimality gap.

3.2. A Lagrangian-based solution procedure for the problem (PCPM)

To launch the solution process explained below, one needs to start with a first assignment of demands to the production
lines. One may start with an assignment of demands to the production lines in proportions of their processing rates (fair
share):

qij
djit ¼ P dit for all i 2 I; t 2 H; j 2 L:
j0 2L qij0

Step 1: Solve the linear programming relaxations of the problems ðPCPMjLR Þ obtained from the problems ðPCPMjLR ðkÞÞ by let-
ting k ¼ 0 and dj as defined above. Use the single production line algorithm as described above but where the pure
production planning problems ðPCPMjSL ðnj ÞÞ, for fixed values of nj , are solved as linear programs (LP-relaxation).
Step 2: Get the shadow prices kjit associated with the constraints (17) for the optimal problem ðPCPMjSL Þ, that is, the optimal
ðPCPMjSL ðn
^ j ÞÞ. To assure convergence, after a given fixed number of iterations, multipliers obtained by the usual sub-
gradient procedure are added to these shadow prices, component by component.
Step 3: Using these shadow prices kjit solve the problem ðDASPðkÞÞ and get the new values for dj . These values are used in
step 1 to get the new shadow prices as explained in step 2, and the process is continued until the values of the prob-
lems ðPCPMjLR Þ are not changing.
Step 4: Once the final values of dj are obtained, then for each line one may use the single production line algorithm
described above, but now the problems ðPCPMjSL Þ are solved as integer linear programs. The collection of optimal
plans for each production line j 2 L is taken as the best approximate to the optimal solution for the problem PCPM.

3.3. An illustrative example

Consider a production system composed of two identical production lines each with a maximal capacity j1 ¼ j2 ¼ 15.
Two items are to be produced in lots on these two lines during a planning horizon of eight periods. The setup, production
and holding costs on each line of each item in each period are respectively 25, 5, and 2. It is also assumed that qij ¼ 1 for
all i 2 P and j 2 L. Table 1 shows the periodic demands of each item in each period.
Assume that the cost of any preventive maintenance action and the costs of any corrective maintenance action on any of
the two production lines are given, respectively by cp1 ¼ cp2 ¼ 40 and cr1 ¼ cr2 ¼ 35. Also, assume that the two production lines
3388 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392

Table 1
Products’ periodic demands

Periods Demand of product 1 Demand of product 2

1 7 6
2 6 4
3 7 4
4 6 4
5 4 6
6 6 4
7 7 4
8 6 4

Table 2
Expected number of failures

Age Expected number of failures

½0; 1s½ 0.901
½1s; 2s½ 1.489
½2s; 3s½ 1.664
½3s; 4s½ 1.749
½4s; 5s½ 1.799
½5s; 6s½ 1.833
½6s; 7s½ 1.857
½7s; 8s½ 1.875

are subject to random failures according to a Gamma distribution Cðm ¼ 2; m ¼ 2Þ with a shape parameter m ¼ 2 and a rate
parameter m ¼ 2. Table 2 shows the expected number of failures as a function of the production line’s age: Recall that the
density function of a Gamma distribution Cðm; mÞ with a shape parameter m and a rate parameter m is given by
(
mm tm1 expðmtÞ
CðmÞ if t; m P 0; m > 0;
f ðtÞ ¼ ð22Þ
0 if t < 0:
Finally, let hpj ¼ 1 and hrj ¼ 5 be respectively the amounts of capacity lost when a preventive maintenance action and an
unplanned corrective maintenance action take place, respectively. Table 3 shows the available capacities in each period
given as a function of the size of the preventive maintenance cycle T ¼ ks.
The expected maintenance cost for each production line can now be obtained as indicated in Table 4, if the size of the
preventive maintenance cycle is given.
To get the optimal solution of this problem 64 pure production planning problems are solved with a linear mixed integer
solver. The problem is then solved with the Lagrangian relaxation based heuristic procedure as described above. The following
are the summary results. The optimal value of the problem (PCPM) is given by Z b PM ¼ 1735:89 where the first production line is
maintained after each three periods and the second production line maintained after each four periods. The Lagrangian-based

Table 3
Available capacities in each period as a function of the size of the preventive maintenance cycle T ¼ ks

Periods Available capacities as a function of T ¼ ks

k¼1 k¼2 k¼3 k¼4 k¼5 k¼6 k¼7 k¼8
1 9.49 9.49 9.49 9.49 9.49 9.49 9.49 9.49
2 9.49 7.55 7.55 7.55 7.55 7.55 7.55 7.55
3 9.49 9.49 6.68 6.68 6.68 6.68 6.68 6.68
4 9.49 7.55 9.49 6.26 6.26 6.26 6.26 6.26
5 9.49 9.49 7.55 9.49 6.00 6.00 6.00 6.00
6 9.49 7.55 6.68 7.55 9.49 5.84 5.84 5.84
7 9.49 9.49 9.49 6.68 7.55 9.49 5.72 5.72
8 9.49 7.55 7.55 6.26 6.68 7.55 9.49 5.63

Table 4
Expected maintenance costs as a function of the size of the preventive maintenance cycle T ¼ ks

Period k¼1 k¼2 k¼3 k¼4 k¼5 k¼6 k¼7 k¼8

Cost 572.39 494.68 487.46 486.19 487.97 493.90 506.77 500.84
 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392 3389

approximation procedure provided the value Zb LR ¼ 1770:09 where the first production line is maintained after each four peri-
ods and the second line maintained after each three periods.

4. Production and general preventive maintenance planning

In many situations additional constraints, such as limited size and number of maintenance crews combined with the pro-
ductivity requirement, can make it difficult to implement cyclical preventive maintenance strategies for all production lines
of the systems. In this section, a model is presented that is more general in the sense that it relaxes the cyclic restriction, and
can handle additional constraints. Before developing the model, some necessary additional parameters and variables should
be defined.
Let jjs ðtÞ be the expected loss in production capacity of line j, in period t, when the last preventive maintenance action
before time t on the line j has taken place in period s, s 6 t. Thus, this expected loss is then given by
( Rs
j
hp þ hr 0 r j ðuÞdu if t ¼ s;
j s ðtÞ ¼ Rs ð23Þ
hr 0 r j ðu þ ðt  sÞsÞdu if t > s:

Let cjs ðtÞ be the expected maintenance cost of line j, in period t, when the last preventive maintenance action before time t
on the line j has taken place in period s; s 6 t. Thus, this expected cost is then given by
( Rs
cp þ cr 0 rj ðuÞdu if t ¼ s;
cjs ðtÞ ¼ Rs ð24Þ
cr 0 r j ðu þ ðt  sÞsÞdu if t > s:

Finally, let zjst be a binary decision variable set to 1 if the preventive maintenance cycle on line j covering period t has
started at period s, and 0 otherwise. The general integrated production and maintenance planning model (PGMP) is given by
!
XX X j j X X
PGMP : Minimize Z GPM ¼ ðfit yit þ pjit xjit Þ þ hit Iit þ cjs ðtÞzjst
t2H i2I j2L t2H s2H;s6t
X
subject to xjit þ Ii;t1  Iit ¼ dit for all i 2 P; t 2 T; ð25Þ
j2L
X X
qij xjit þ jjs ðtÞzjst 6 jj for all j 2 L; t 2 T; ð26Þ
i2I s2H;s6t
i
xjit  dtN yjit 6 0 for all i 2 P; j 2 L; t 2 T; ð27Þ
X j
yit 6 1 for all j 2 L; t 2 T ð28Þ
i2I
X
zjst ¼ 1 for all j 2 L; t 2 T; ð29Þ
s2H;s6t

zjst  zjst1 6 0 for all j 2 L; s 2 T; t 2 T; s < t; ð30Þ

xtij ; Iit P 0; ytij ; zjst 2 f0; 1g for all i 2 P; j 2 L; and s; t 2 H:

Constraints (26) are the new capacity constraints. These constraints define for each production line j 2 L the available
capacity during each period. Constraints (29) determine the preventive maintenance periods for each production line. Con-
straints (30) guarantee, that if a period s is the last preventive maintenance period before t > s, then s is also the last pre-
ventive maintenance period before t  1.
The advantage of this problem is that the objective function in not any more nonlinear, one can then solve it using any
solver for linear mixed integer programs. In the following section, this model will be used to generate lower bounds for the
test problems (PCPM) to assess the optimality gap for the heuristic procedure.

5. Design of experiment and computational results’ analysis

To assess the quality of the solutions of the problem (PCPM) obtained with the Lagrangian relaxation based heuristic, a
design of experiment based on some critical planning parameters which are thought likely to have a significant impact on
these heuristic solutions is developed. Seven parameters are considered for the analysis: number of products, number of pro-
duction lines, number of periods in the planning horizon, cost structure, demand structure, capacity tightness, and failure
distribution of each production line. The design of experiment is structured in the same way as in [21].
The number of product jPj is selected from {10, 25}, the number of periods N is selected from {12, 24}, and the number of
production lines jLj is selected randomly from {2, 4}. Failure distribution of each production line j 2 L is selected from Gamma
distributions with a shape parameter 2 and rate parameters randomly selected from the set {1, 2}, and Weibull distributions
with a shape parameter 2 and scale parameters randomly selected from the set {3, 4}.
3390 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392

The cost structure is assumed to have the following properties: the setup cost fitj for each product is selected randomly
from two different intervals ½10; 50 for low setup cost problem and ½75; 100 for high setup cost problems, the production
cost pjit is selected randomly from ½5; 10, and the holding cost is then defined as hit ¼ a%maxj ðpjit Þ, where a is uniformly dis-
tributed between 5 and 20. The demand for each product i 2 P is assumed to have a stationary mean d  . The mean demands
i
are randomly sampled from the interval ½75; 100, and the period-by-period demands for each product i 2 P are generated
from a truncated normal distribution having mean d  and a standard deviation randomly sampled from the interval
i
½0:25di ; 0:5d
i .
The capacity is the last parameter set for each production line in each test problem. A target average utilization of capacity
factor b is first introduced to define the capacity tightness. The factor b is set to 0.75, 0.85, and 0.95 corresponding respec-
tively to situations with loose, moderately loose, and tight capacity constraints. The capacity required, in each period, is then
computed as if a lot-for-lot solutions were implemented. These values are averaged over all periods and the result is divided
by the number of production lines. The capacity of each production line in each period is then obtained by dividing the later
result by the target average utilization of capacity factor b.
For each combination of number of products, number of production lines, number of periods, setup cost structure, and
capacity tightness 15 test problems are generated. These 15 test problems are divided in three groups. A first group of five
test problems is generated with Gamma failure distributions of all production lines, then a second group of five other test
problems is generated with Weibull failure distributions of all production lines are Weibull distributions, and finally, a last
group of five test problems is generated with failure distributions selected randomly from Gamma or a Weibull distributions
(mixed). In this last group, test problems in which the failure distributions are all the same are discarded. This makes a total
of 720 test problems. Each test problem is solved with the proposed Lagrangian-based solution procedure to obtain a very
good upper bound Z UB PM on the optimal value the problem (PCPM). The model (PGMP) is then solved for the same instance to
get a lower bound on the optimal value the problem (PCPM). Finally, the gap is computed as follows:

Z UB G
PM  Z PM
Gap ¼ 100 %:
Z UB
PM

Table 5 shows the summary of the results of the averages of above ratios computer for each test group, and Table 6 pre-
sents the computational times required for the Lagrangian-based solution procedure. Notice that G-*, W-*, and M-* in the two
tables mean, respectively, that the line has a Gamma failure distribution, a Weibull, or mixed (some have Gamma failure
distributions and some others Weibull distribution).
The average solution gap ratios shown in Table 5 indicate that the tightness of the capacity constraint has a significant
effect on the solution gap. Test problems with tight capacity constraints have much larger average gaps than those with loose

Table 5
Summary of solution gaps (mean)

Lines Items Periods Capacity tightness

Loose Moderately loose Tight
Setup costs
Low High Low High Low High
G-2 10 12 2.034 6.842 3.840 13.273 15.609 16.653
G-2 10 24 2.132 7.171 4.025 13.910 16.358 17.452
G-2 25 12 2.294 7.719 4.332 14.973 17.608 18.786
G-2 25 24 1.553 5.226 2.933 10.137 11.921 12.718
G-4 10 12 2.250 7.570 4.249 14.685 17.270 18.425
G-4 10 24 1.590 5.348 3.002 10.375 12.201 13.017
G-4 25 12 1.658 5.577 3.130 10.818 12.722 13.573
G-4 25 24 1.924 6.474 3.634 12.558 14.768 15.756
W-2 10 12 1.272 4.279 2.402 8.301 9.762 10.415
W-2 10 24 1.752 5.893 3.308 11.432 13.444 14.343
W-2 25 12 1.215 4.089 2.295 7.932 9.328 9.952
W-2 25 24 1.561 5.250 2.947 10.185 11.977 12.779
W-4 10 12 1.933 6.502 3.650 12.613 14.833 15.825
W-4 10 24 1.701 5.721 3.211 11.098 13.050 13.923
W-4 25 12 1.663 5.593 3.140 10.851 12.760 13.614
W-4 25 24 1.368 4.603 2.583 8.929 10.500 11.202
M-2 10 12 2.039 6.858 3.850 13.304 15.646 16.692
M-2 10 24 1.681 5.656 3.175 10.972 12.903 13.766
M-2 25 12 1.506 5.066 2.844 9.828 11.557 12.330
M-2 25 24 1.331 4.476 2.512 8.683 10.211 10.894
M-4 10 12 1.862 6.263 3.515 12.149 14.287 15.243
M-4 10 24 1.773 5.965 3.348 11.571 13.607 14.518
M-4 25 12 1.781 5.991 3.363 11.621 13.666 14.581
M-4 25 24 1.425 4.793 2.690 9.298 10.934 11.665
 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392 3391

Table 6
Summary of the heuristic computational times (min)

Lines Items Periods Capacity tightness

Loose Moderately loose Tight
Setup costs
Low High Low High Low High
G-2 10 12 1.29 1.62 1.76 2.05 2.04 2.49
G-2 10 24 1.30 1.58 1.61 1.98 2.32 2.23
G-2 25 12 3.13 3.45 3.73 4.12 4.52 4.66
G-2 25 24 2.60 3.17 3.31 4.35 4.47 4.79
G-4 10 12 4.03 4.80 4.86 5.83 5.90 7.48
G-4 10 24 4.47 4.64 4.90 5.77 6.34 7.29
G-4 25 12 15.69 19.10 20.56 25.64 24.42 30.84
G-4 25 24 17.17 18.99 19.23 24.44 26.75 29.46
W-2 10 12 1.36 1.74 1.83 1.90 2.15 2.39
W-2 10 24 1.44 1.54 1.85 1.88 2.12 2.52
W-2 25 12 2.71 3.26 3.44 3.95 4.40 4.90
W-2 25 24 3.02 3.44 3.63 4.21 4.03 4.75
W-4 10 12 3.99 5.09 5.76 6.43 5.74 7.03
W-4 10 24 4.32 5.20 4.89 6.41 5.97 7.79
W-4 25 12 15.96 20.68 21.34 24.92 27.13 27.32
W-4 25 24 15.48 19.35 22.76 25.33 24.66 27.90
M-2 10 12 1.37 1.59 1.89 2.18 1.99 2.26
M-2 10 24 1.49 1.56 1.69 1.90 2.11 2.61
M-2 25 12 2.95 3.47 3.69 4.32 4.36 5.03
M-2 25 24 2.81 3.26 3.71 4.37 4.32 4.94
M-4 10 12 3.86 4.51 4.88 6.35 6.39 7.81
M-4 10 24 3.82 5.07 5.20 5.97 6.34 6.90
M-4 25 12 16.37 19.99 22.21 23.55 23.18 27.09
M-4 25 24 16.44 19.78 19.43 22.61 23.17 26.95

capacity. This gap varies from 2% for loose capacity problems up to 16% when capacity is tight. Actually, this result was ex-
pected since, for test problems with loose capacity, the two independently obtained optimal production and maintenance
plans usually produce feasible near-optimal integrated plans for the problem (PCPM). These ratios show also that the effect
of setup cost is significant. When comparing the low and high setup cost columns in Table 5, one can see that the differences
vary between 4% and 8%. However, when capacity is tight these differences are not important. This is probably due to the fact
that when the capacity is tight, production takes place more frequently making the impact of the setup cost less significant.
Table 5 shows also that when Weibull distributions with scale parameters of 3 or 4 are used, the solution gaps become smal-
ler than when Gamma distributions are used. This is also expected to some extent, since the expected capacity loss, given in
function of the time elapsed since the last preventive maintenance, is smaller in case of a Weibull distribution with scale
parameters of 3 or 4. Overall, the results obtained are fairly good in terms of optimality gaps.
Table 6 shows the computational times of the heuristic procedure. These times vary from 2 to 30 min for hard and large
cases. This can be explained by the fact that Cplex is used a certain number of times during each iteration to solve pure pro-
duction problems. Most probably, a more careful implementation of the procedure would reduce significantly these compu-
tational times.
Finally, notice that in the proposed Lagrangian-based solution procedure, the assignment of demands to the production
lines is carried out on the basis of marginal costs obtained through the solutions of LP-relaxations of the single-line linear
mixed integer sub-problems and corrected with subgradient-based multipliers. When no improvement is realized the last
assignment of demands is kept and the sub-problem is solved as a mixed integer problem. The test problems indicate that
a large part of the gaps resulted from non-optimal assignments of the demands to the lines. To improve this gap the Lagrang-
ian dual of the problem ðPCPMR Þ in which the variables g are eliminated and constraints (16) are relaxed is currently
investigated.

6. Conclusion

This paper investigated the problem of integrating production planning and preventive maintenance in multi-line pro-
duction systems subject to failures. It is assumed that the capacities of these production lines progressively decline as they
produce due deterioration. Two mathematical programming models for the problem are proposed. The first model assumes
that each production line of the system implements a cyclic preventive maintenance policy with a fixed cycle size. The sec-
ond model relaxes this last restriction and establishes for each production line specific maintenance periods, which do not
necessarily fall at equally distant epochs. The first model is an NP-hard nonlinear mixed integer program, for which a heu-
ristic method based on Lagrangian relaxation is developed. The sub-problems, used to iteratively construct an approximate
3392 E.-H. Aghezzaf, N.M. Najid / Information Sciences 178 (2008) 3382–3392

optimal solution, are single-line versions of the problem, one for each production line, and then a last sub-problem is a linear
program used to assign demands to the production lines. The LP-relaxations of these single-line sub-problems are solved to
provide marginal production costs on each line, and these marginal costs combined with subgradient multipliers are used to
assign the demands to the production lines. A series of test problems were generated and solved with the Lagrangian-based
solution procedure. The optimality gaps were assessed using lower bounds obtained with the second model, which is a linear
mixed integer program. These results show that procedure performs fairly well in terms of solution gap. Computational
times are still high for some cases. Improving the implementation of the procedure to reduce these times is now carried out.

References

[1] I. Adiri, J. Bruno, E. Frostig, A.H.G. Rinnooy Kan, Single machine flow-time scheduling with a single breakdown, Acta Inform. 26 (1989) 679–696.
[2] A. Aggarwal, J.K. Park, Improved algorithms for economic lot size problems, Oper. Res. 41 (1993) 549–571.
[3] R. Aggoune, M.-C. Portmann, Flow shop scheduling problem with limited machine availability: a heuristic approach, Int. J. Prod. Econ. 99 (1–2) (2006)
4–15.
[4] E.H. Aghezzaf, A. Jamali, D. Ait-Kadi, A production and maintenance planning model for production systems subject to preventive maintenance with
minimal repair at failure, in: Proceedings of the Fifth International Industrial Engineering Conference, Québec, Canada, 2003.
[5] G. Belvaux, L.A. Wolsey, BC-PROD: a specialized branch-and-cut system for lot-sizing problems, Manage. Sci. 46 (2000) 724–738.
[6] J. Birge, J.B.G. Frenk, J. Mittenthal, A.H.G. Rinnooy Kan, Single-machine scheduling subject to stochastic breakdowns, Naval Res. Logist. 37 (1990) 661–
677.
[7] G.R. Bitran, H.H. Yanasse, Computational complexity of the capacitated lot size problem, Manage. Sci. 28 (1982) 1174–1186.
[8] J. Blazewicz, J. Breit, P. Formanowicz, W. Kubiak, G. Schmidt, Heuristic algorithms for the two-machine flowshop problem with limited machine
availability, Omega J. 29 (2001) 599–608.
[9] G.D. Eppen, R.K. Martin, Solving multi-item lot-sizing problems using variable redefinition, Oper. Res. 35 (1987) 832–848.
[10] M. Florian, M. Klein, Deterministic production planning with concave costs and capacity constraints, Manage. Sci. 26 (1971) 669–679.
[11] K.D. Glazebrook, Evaluating the effects of machine breakdowns in stochastic scheduling problems, Naval Res. Logist. 34 (1987) 319–335.
[12] G.H. Graves, C.Y. Lee, Scheduling maintenance and semiresumable jobs on single machine, Naval Res. Logist. 46 (1999) 845–862.
[13] W. Kubiak, J. Blazewicz, P. Formanowicz, J. Breit, G. Schmidt, Two-machine flow shops with limited machine availability, Eur. J. Oper. Res. 136 (2002)
528–540.
[14] A. Lambert, H. Luss, Production planning with time-dependent capacity bounds, Eur. J. Oper. Res. 9 (1982) 275–280.
[15] C.Y. Lee, Machine scheduling with an availability constraint, J. Global Opt. 9 (1996) 395–416.
[16] C.Y. Lee, Minimizing the makespan in the two machines flowshop scheduling problem with an availability constraint, Oper. Res. Lett. 20 (1997) 129–
139.
[17] C.Y. Lee, Z.L. Chen, Scheduling jobs and maintenance activities on parallel machines, Naval Res. Logist. 47 (2000) 145–165.
[18] W. Li, J. Cao, Stochastic scheduling on a single machine subject to multiple breakdowns according to different probabilities, Oper. Res. Lett. 18 (1995)
81–91.
[19] X. Qi, T. Chen, F. Tu, Scheduling the maintenance on single machine, J. Oper. Res. Soc. 50 (10) (1999) 1071–1078.
[20] G. Schmidt, Scheduling with limited machine availability, Eur. J. Oper. Res. 121 (2000) 1–15.
[21] W.W. Trigeiro, L. Joseph Thomas, J.O. McClain, Capacitated lot sizing with setup times, Manage. Sci. 35 (3) (1989) 353–366.
[22] C.P.M. Van Hoesel, A.P.M. Wagelmans, An O(T3) algorithm for the economic lot-sizing problem with constant capacities, Manage. Sci. 42 (1996) 142–
150.
[23] L. Wienstein, C.H. Chung, Integrated maintenance and production decisions in a hierarchical production planning environment, Comput. Oper. Res. 26
(1999) 1059–1074.
[24] L.A. Wolsey, Solving multi-item lot-sizing problems with an MIP solver using classification and reformulation, Manage. Sci. 48 (2002) 1587–1602.