You are on page 1of 5

Proceedings of the 13th IFAC Symposium on

Information Control Problems in Manufacturing


Moscow, Russia, June 3-5, 2009

Coupling Simulation and Optimization for an Integrated Production and


Maintenance Planning in Manufacturing Systems

D. Bergeron*. M. A. Jamali.*
D. Ait-Kadi**

*Department of Industrial Engineering, University of Quebec at Trois-Rivieres,


Trois-Rivieres (Qc), Canada, (e-mail: Daniel.Bergeron1@uqtr.ca, Jamali@uqtr.ca).
**Department of Mechanical Engineering, Pavillon Adrien-Pouliot, Laval University,
Quebec (Qc), Canada, (e-mail: Daoud.Ait-Kadi@gmc.ulaval.ca)

Abstract: This paper discusses the issue of integrating production and maintenance in manufacturing
production systems. The objective is to find an integrated lot-sizing and preventive maintenance strategy
of the system that satisfies the demand for all product types without backlogging, and which minimizes
the sum of production and maintenance costs. Optimal buffer sizing is also considered in this model. The
procedure used to reach the objective is valid for other maintenance policies as well. A simulation model
was developed to evaluate the total cost of operating the production system when a finite-capacity buffer
stock is introduced between each pair of machine. The simulation model gives the opportunity to
evaluate the performance of a production system under different integrated production and maintenance
plans as well as under different system configurations.
Keywords: Manufacturing systems, production, preventive maintenance, simulation, optimization,
integrated strategies.

machines produce until they are starved (upstream buffer is


1. INTRODUCTION empty). Each maintenance action, preventive or corrective, is
performed at fixed costs. The cost incurred for a correction
Production and maintenance are two interrelated aspects of action is higher than the cost incurred for a preventive action.
any manufacturing process. While minor maintenance tasks Furthermore, at the end of each period of the planning
are more integrated during the production time than ever horizon, a holding cost is applied. The demand is known in
before, the other maintenance tasks must be planned with the advance for each product type since the material is “pulled”
objective of, on one hand, minimizing as much as possible by demand. A setup cost is incurred each time the product
the impact on the production planning and, on the other hand, type being manufactured is different from the last product
respecting the best frequency equipment maintenance. type that has been manufactured. Furthermore, to minimize
However, while this area of research has grown considerably backlogging, a very high penalty cost is incurred.
recently, models integrating production and maintenance are
still rare in the literature. Most models consider either The model proposed in this paper has been developed with
maintenance planning models or production planning models. Arena simulation software. OptQuest, the Arena-embedded
Maintenance planning models found in the literature do not optimization software, is used to find the optimal solution
explicitly consider their impact on production planning. that minimizes maintenance and production costs. An optimal
Likewise, most production planning models do not take into solution is one that contains the optimal PM periodicity of
account the impact of corrective maintenance (CM) actions, each machine, the optimal size of each buffer, and the
or even preventive maintenance (PM) actions, in their optimal production plan of a given production line.
planning. If the optimal production or maintenance strategies
are valid when taken apart, they may not be optimal anymore The remainder of this paper contains a brief review of the
when taken at once. This constitutes the basis of our research. literature in Section 2. The combined simulation and
optimization model is presented in Section 3. Section 4 gives
The model proposed in this paper is a multi-product results for production lines counting several machines.
integrated production and maintenance manufacturing system Conclusion is made in Section 5.
over a planning horizon that includes a buffer between each
pair of machines to minimize the impact of failures and 2. LITERATURE REVIEW
preventive actions. Each machine is subject to random
failures. When a failure occurs, a minimal repair is Most papers found in the literature seem to overlook the
performed. A PM action on a given machine leads to a inter-dependency between maintenance and production
renewal of this machine. Upstream machines keep producing planning. Since available production capacity is affected by
until they are blocked by a full buffer, while downstream maintenance actions and that elapsed production time affects

978-3-902661-43-2/09/$20.00 © 2009 IFAC 1091 10.3182/20090603-3-RU-2001.0549


13th IFAC INCOM (INCOM'09)
Moscow, Russia, June 3-5, 2009

machine failure probabilities, this integration should be In opposition to Aghezzaf, Jamali, and Ait-Kadi’s model,
considered. Adiri, Bruno, Frostig et al. (1989), Cassady and considered as the basic model, the planning horizon () in
Kutanoglu (2003), Graves and Lee (1999), Iravani and our model is not limited to 1 cycle of 8 equal periods (), but
Duenyas (2002) showed in their paper that the single rather counts 625 cycles (or 5000 periods) in order to get
machine-scheduling problem subject to random failures is production and maintenance costs based on an average per
NP-hard even if the breakdowns are known in advance. Lee cycle. This number seemed high enough to obtain
(1996), Lee and Chen (2000), Lee, Lei and Pinedo (1997) reproducible results within a 2% error margin.
also found that the problem is NP-hard when they considered
various versions of the problem. In addition, Allahverdi, For each period τ, a predefined demand dit has to be satisfied.
Gupta and Aldowaisan (1999), in their review, state that The production system has a nominal capacity  fixed for
setup time if often neglected in majority of scheduling each machine during each period, and each maintenance
research papers, and that this omission may conduct to poor action consumes a certain amount of this capacity. In our
quality solutions. example containing 4 machines,  is set to 20. For all
other examples,  is set to 15. Thus, in this last case,
Weinstein and Chung (1999) proposed a model counting processing time is deterministic and equals 1/15 of a period
three part types to evaluate an organization’s maintenance length. Table 1 illustrates setup cost, production cost and
policy. However, the reliability parameters are only holding cost used in the basic model for each product type.
considered during the second stage of the planning process. We used same values to facilitate results’ comparisons. In our
Iravani and Duenyas (2002) suggested a model integrating model, we added the waiting customer cost to take delays
maintenance and production control for a single machine / into account. Such delays could not be taken into account in
single product system. They also demonstrate that policies the basic model since it was an analytic model. A setup
considering production and maintenance control apart operation is carried before processing the first part and
perform badly. Ashayeri, Teelen and Selen (1996) came to between each producing lot. Lot-sizing is determined by the
the same conclusion, but with N-machine production lines. optimization module while it is seeking for an optimal
Aghezzaf, Jamali and Ait-Kadi (2007) proposed a model production plan.
counting one machine and two product types. The model
presented in this paper is based on the latter. Beside the fact Table 1. Cost data for products
that Aghezzaf, Jamali, and Ait-Kadi’s model contains only
one machine, the main drawback of this model concerns the Waiting
Holding
fact that it has been developed analytically. Hence, no delay Production Customer
Product Setup Cost
can occur in their model since the analytical model is based Cost Cost
Type Cost (unit cost /
on an average. However, a simulation model is much closer (unit cost) (unit cost
period)
to reality; even with the best production plan, no company is / period)
immunized to an abnormally long corrective maintenance, to A 25 5 2 100
a machine that suddenly keeps on breaking, or to a two- B 25 5 2 100
machine-break at same time. These events might inevitably
lead to delays. When the analytical model cannot take these The data contained in Table 2 show demand for products A
cases into account easily, a simulation model can. and B during 1 cycle (i.e. 8 periods). Since it is impossible to
avoid backlogging completely, a high penalty cost of 100 is
In order to avoid this major drawback and to be able to incurred in case such an event occurs. Therefore, the
expand Aghezzaf, Jamali, and Ait-Kadi’s model, we opted OptQuest optimization module would eliminate results where
for a combination of simulation and optimization instead of there are several backlogs. This optimization module is based
relying on an analytic model. This option is also preferred for on a combination of Tabu Search and Scatter Search
future extensibility of the model. From all the papers that we algorithms. The Tabu Search algorithm acts as the memory of
are aware of, our model would be the only one integrating the optimization module; it avoids re-evaluation of solutions
production and PM planning, for a given set of products, on a already evaluated, and also gauges the diversification and
serial production line where machines are separated by intensification of solutions to come. On the other side, the
buffers. Furthermore, in our model, not only production and Scatter Search algorithm looks similar to the Genetic
maintenance actions are optimized to minimize costs, buffer Algorithm; it is based on a population and creates new
sizing is also optimized in order to reduce blockage and solutions based on a combination of previous solutions.
starvation of machines on the production line. However, this combination is not randomized, but rather
based on a probabilistic strategy.
3. SIMULATION AND OPTIMIZATION MODEL
Table 2. Periodic demand for products
The model proposed by Aghezzaf, Jamali, and Ait-Kadi Period 1 2 3 4 5 6 7 8
(2007) contains only one machine () and two product Demand for
types (A and B). Our model can contain much more. But, in 2 3 2 3 2 3 2 3
Product A
order to validate our model, we have started by comparing Demand for
our results with theirs, for such a small model. These results 3 2 3 2 3 2 3 2
Product B
will be discussed later in this section.

1092
13th IFAC INCOM (INCOM'09)
Moscow, Russia, June 3-5, 2009

The production cost ( ) is calculated in the following way:


2>?@#- -A:
∑∈ ∑∈         
 = +∙ ∑!
" ,
 ∑+ ∈ C 7*+ = ∑+ ∈ C D*+ )AE , ∈ (2)
whereas the maintenance cost () is calculated as follows: 7*+ + 8*+F ≥ D*+ )AE - ∈ 2 H<D , ∈ (3)
# ∙ ∑! !
" $ %  + #& ∙ ∑" ' %  ∑* ∈ I 7*+ ≤  - )AE - ∈ 2 H<D : ∈  (4)
 =
(
7*+ , 8*+ , 9*+, , % ≥ 0 )AE - ∈  H<D , ∈ (5)
The model parameters are:
5*+ ∈ {0, 1} )AE - ∈  H<D , ∈ (6)
)*+ fixed setup cost for producing item , in period -
.*+ variable cost of producing one unit of item , in period , ≥ 1 )AE : ∈  (7)
-
ℎ*+ variable cost of holding one unit of item , by the end The constraint (2) forces to produce exactly the number of
of period - product type , required by the demand D over a cycle S. The
0*+ variable cost of delaying one unit of item , by the end constraint (3) states that the demand must be meet every
of period - period -. The constraint (4) limits the production to the
 duration of each basic planning period capacity of the machine  - for a period -. Constraints (5,
1 number of periods of fixed length  in a cycle 2 6) are for non-negativity. Constraint (7) stands because each
( number of periods of fixed length cycle 2 over the machine is considered as having exactly one buffer slot.
horizon H
ℎ*+ variable cost of holding one unit of item , by the end When a PM action occurs during period -, the remaining
of period -
available capacity : - for each period can be calculated
0*+ variable cost of delaying one unit of item , by the end
using the following:
of period -
 fixed cost of allowing one buffer slot on any buffer O
# cost of each PM action  - =  − 4 − 4& NP E > + - − 1 D>.
#& cost of each CM action (# ≤ #& 
4 machine capacity consumed during a PM action Since a PM action occurs only once every PM cycle % on a
given machine :, for all other periods -,  - can be
4& machine capacity consumed during a CM action
calculated using the following:
 - function that defines the available capacity for a
period - O
 - =  − 4& NP E > + - − 1 D>.
 maximum production capacity of each single machine
during 
 number of machines Notice that E - is the known failure rate of a given machine
 horizon (equals (1) : at time - and is defined by:
set of products (A = 1, B = 2, …)
) -
E - =
and the model variables are: 1 − Q -

5*+ binary variable (5*+ equals 1 if one unit of item , is where ) - and Q - are, respectively, the failure
produced after a unit of any other item ,, 5*+ equals 0 probability density function and the cumulative distribution
otherwise) function of the machine :. As an example, in the case the
7*+ quantity of item , produced in period - failure distribution is given by a gamma distribution R 2, 1,
8*+ inventory of item , by the end of period - we would get:
9*+ quantity of item , delayed by the end of period -
, size of the buffer placed between machines : and -
E - =
: + 1 (including one slot on the machine :) -+1
% PM cycle on machine :
$ %  number of PM actions over the cycle 2 considering a In order to validate our model, we first compared our optimal
PM cycle % for machine : production planning results with those of Aghezzaf, Jamali,
' %  number of CM actions over the cycle 2 considering and Ait-Kadi in the case the machine is reliable. We obtained
a PM cycle % for machine : the exact same results, illustrating that our optimization
module is efficient for such a small problem. Notice that in
The objective function is then defined by: our model, during a given period, product A is always being
produced before product B. Also, notice that since the
,< = =  +  (1)

1093
13th IFAC INCOM (INCOM'09)
Moscow, Russia, June 3-5, 2009

machine is reliable, there is no buffer,  - =  ,  = 0, backlog), is reflected in our average production cost. This
and 9*+ = 0. may explain the differences.

In the case the machine is unreliable, both production costs When the failure distribution used is R 2, 1, the calculated
and maintenance costs should get higher since failures cause minimal maintenance cost is obtained for %* = 4 and equals
unavailability of the machine, and that delays are likely to 223.34. Our coupled simulation and optimization system
occur. Batch size should then be smaller and more setup gave us a value of 225.14. Similarly, when the failure
operations should be required. It is assumed that the machine distribution is given by R 2, 2, we obtained a maintenance
has known failure probability density function ) - and cost of 449.81 for %*= 2, which is slightly higher than
cumulative distribution function Q -. In our examples, the theoretical cost being 446.68.
failure probability density function ) - is a gamma
distribution function. The available remaining capacities for 4. GENERALIZATION FOR MULTIPLE MACHINES
each period provided in the basic model contained some
mistakes; it seemed their results were calculated using This section contains two examples, one having 2 machines
R 2, 2 set of parameters for gamma distribution function and one having 4 machines. These machines are identical to
instead of R 2, 1. Furthermore, their optimal maintenance the one presented in Section 3. However, the failure
costs were calculated for a horizon of only one cycle, making distribution of a given machine can either be R 2, 1 or
some of them erroneous on a longer horizon (for T ∈ {3, 5, 6, R 2, 2. Every pair of machines is separated by a buffer. A
7}). The results we obtained by simulation were within 2% of cost of 2 is incurred for each buffer slot used in the final
the corrected results of Aghezzaf, Jamali and Ait-Kadi (2007) design. No cost is considered for the buffer slot contained on
for a gamma distribution function having, respectively, a machine.
parameters R 2, 1 and R 2, 2. The cost of a PM is set to #
= 28, and the cost of a minimal repair action at failure is #& = To avoid running the optimization module for a long period
35. Furthermore, it is assumed that all maintenance actions of time, we implemented a procedure which allows the
are perfectly performed. PM actions are performed every %, optimization module to start with a near-optimal initial

where % = T, 2T, 3T... This consumes 4 = H solution min = = {Y% Z∗ , \, ] , Y Z∗}. During the first
capacity units. Corrective maintenance actions are performed stage of this procedure, the optimization module seeks for the
every time machines fail and it consumes 4& =  optimal PM periodicity matrix, Y% Z∗. The optimal matrix
capacity units. Notice that 0 ≤ H ≤  ≤ 1, since more capacity found in the first stage, is used during the second stage where
units are consumed to carry a CM action than to carry a PM the size of buffers matrix \, ] is optimized. During the
action. In all of our examples, H = 0.067, and  = 0.333. third stage, the production plan Y Z is optimized in order to
minimize production costs. Finally, at the last stage, the
Concerning the remaining available capacities in each period matrixes obtained during the first three stages are used as the
as a function of maintenance period length % = T, compared initial solution for the combined optimization of PM
to theoretical results, all obtained results had an error margin periodicity, buffer size and production plan. There are three
below 2%. However, this statement does not apply to replications per simulation. The best 25 solutions obtained
production costs. Since the optimal production plan obtained are refined with seven more replications. The optimization
by Aghezzaf, Jamali, and Ait-Kadi for each value of % was module is stopped when there has been no improvement of
not available, we were unable to reproduce their results. the best solution over the last 100 simulations.
Furthermore, their results are valid only for a horizon  = 1
cycle. In addition, even if one wants to avoid backlogs as Table 3 shows results for a production line containing 2
much as possible, this event is likely to happen. Even if it machines having respectively R 2, 1 and R 2, 2 failure
happens rarely during 625 cycles, this backlog cost (100 per distributions. The machines are separated by a buffer.

Table 3. Optimal production plan, buffer size and PM periodicity for each machine in a 2-machine production line
^_`a`bc Period 1 2 3 4 5 6 7 8 Cost
Product A 2 3 5 2 0 3 2 3
Y Z∗
Product B 8 0 0 6 6 0 0 0

\, ] [5]
Y% Z∗ [4, 2]
Total cost 1105.42
d`_bc Period 1 2 3 4 5 6 7 8 Cost
Production Product A 2 3 7 0 0 3 2 3
Plan Product B 8 0 0 6 6 0 0 0

\, ] [5]
Y% Z∗ [3, 2]
Total cost 1092.37

1094
13th IFAC INCOM (INCOM'09)
Moscow, Russia, June 3-5, 2009

The upper part of Table 3 contains results obtained after the 5. CONCLUSION
first three stages, while the lower part of the table contains
final results. The differences between both parts of the table In this paper, we suggested a new model integrating
are greyed. Final results demonstrate that the first machine production and maintenance planning that is closer to reality,
should be preventively maintained every 3 periods, while this and may therefore be interesting for manufacturers. This
maintenance should be made every 2 periods for the second model is interesting since it aims to determine the optimal
machine. The buffer size is equal to 5. production plan considering both preventive and corrective
maintenance requirements. The production plan obtained is
Table 4 illustrates final results for a production line having 4 then more robust in case a machine failure occurs. Since both
machines separated by 3 buffers. Machines 1, 2 and 4 have preventive and corrective maintenance actions are taken into
G 2, 1 failure distributions while machine 3 has a G 2, 2 account before creating the production plan, the production
failure distribution. In this case, Cghi is equal to 20 instead of plan is more optimal in case one or more failures occur
15. This is because the production of one part necessitated 4 during the production.
units of capacity (one unit per machine). The number of
backlogs became way too important. Final results Future research will focus on complex production systems
demonstrate that the first, second and fourth machines should (CPS) instead of production lines. Furthermore, this model
be preventively maintained every 4 periods, while this could be extended to consider the impact of a limited
maintenance should be made every 2 periods for the third maintenance staff, of an unknown demand, or of imperfectly
machine. The buffer sizes are equal, respectively, to 2, 1 and carried maintenance actions. This could lead us to the
4. The optimization module, however, took about an hour to integration of a third important aspect of the manufacturing
give us the results on a 2 GHz processor. An optimization process to our model: the quality.
module with a dedicated algorithm would certainly provide
faster results.

Table 4. Optimal production plan, buffer sizes and PM periodicity for each machine in a 4-machine production line
d`_bc Period 1 2 3 4 5 6 7 8 Cost
Product A 2 4 6 0 0 3 2 3
Y Z∗
Product B 8 0 0 2 10 0 0 0

\, ] [2, 1, 4]
Y% Z∗ [4, 4, 2, 4]
Total cost 1508.54

REFERENCES Graves, G. H., and C.-Y. Lee (1999). Scheduling


maintenance and semiresumable jobs on a single
Adiri, I., J. Bruno, et al. (1989). Single machine flow-time machine. Naval Research Logistics, 46 (7): 845-863.
scheduling with a single breakdown. Acta
Informatica, 26 (7): 679-696. Iravani, S. M. R., and I. Duenyas (2002). Integrated
maintenance and production control of a
Aghezzaf, E. H., M. A. Jamali, et al. (2007). An integrated deteriorating production system. IIE Transactions,
production and preventive maintenance planning 34 (5): 423-435.
model. European Journal of Operational Research,
181 (2): 679-685. Lee, C.-Y. (1996). Machine scheduling with an availability
constraint. Journal of Global Optimization, 9 (3):
Allahverdi, A., J. N. D. Gupta, et al. (1999). A review of 395-416.
scheduling research involving setup considerations.
Omega, 27 (2): 219-239. Lee, C.-Y., and Z.-L. Chen (2000). Scheduling jobs and
maintenance activities on parallel machines. Naval
Ashayeri, J., A. Teelen, et al. (1996). A production and Research Logistics, 47 (2): 145-165.
maintenance planning model for the process
industry. International Journal of Production Lee, C.-Y., L. Lei, et al. (1997). Current trends in
Research, 34 (12): 16. deterministic scheduling. Annals of Operations
Research, 70 (0): 1-41.
Cassady, C. R., and E. Kutanoglu (2003). Minimizing Job
Tardiness Using Integrated Preventive Maintenance Weinstein, L., and C.-H. Chung (1999). Integrating
Planning and Production Scheduling. IIE maintenance and production decisions in a
Transactions, 35 (6): 503 - 513. hierarchical production planning environment.
Computers & Operations Research, 26 (10-11):
1059-1074.

1095

You might also like