Vous êtes sur la page 1sur 17

Journal of Quality in Maintenance Engineering

Emerald Article: Production control of unreliable manufacturing systems


producing defective items
F. Mhada, A. Hajji, R. Malham, A. Gharbi, R. Pellerin
Article information:
To cite this document: F. Mhada, A. Hajji, R. Malham, A. Gharbi, R. Pellerin, (2011),"Production control of unreliable
manufacturing systems producing defective items", Journal of Quality in Maintenance Engineering, Vol. 17 Iss: 3 pp. 238 - 253
Permanent link to this document:
http://dx.doi.org/10.1108/13552511111157362
Downloaded on: 18-07-2012
References: This document contains references to 19 other documents
To copy this document: permissions@emeraldinsight.com
This document has been downloaded 573 times since 2011. *
Users who downloaded this Article also downloaded: *
Laura C. Engel, John Holford, Helena Pimlott-Wilson, (2010),"Effectiveness, inequality and ethos in three English schools",
International Journal of Sociology and Social Policy, Vol. 30 Iss: 3 pp. 140 - 154
http://dx.doi.org/10.1108/01443331011033337
Aryati Bakri, Peter Willett, (2011),"Computer science research in Malaysia: a bibliometric analysis", Aslib Proceedings, Vol. 63
Iss: 2 pp. 321 - 335
http://dx.doi.org/10.1108/00012531111135727
Sandrine Roginsky, Sally Shortall, (2009),"Civil society as a contested field of meanings", International Journal of Sociology
and Social Policy, Vol. 29 Iss: 9 pp. 473 - 487
http://dx.doi.org/10.1108/01443330910986261
Access to this document was granted through an Emerald subscription provided by OXFORD BROOKES UNIVERSITY

For Authors:
If you would like to write for this, or any other Emerald publication, then please use our Emerald for Authors service.
Information about how to choose which publication to write for and submission guidelines are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
With over forty years' experience, Emerald Group Publishing is a leading independent publisher of global research with impact in
business, society, public policy and education. In total, Emerald publishes over 275 journals and more than 130 book series, as
well as an extensive range of online products and services. Emerald is both COUNTER 3 and TRANSFER compliant. The organization is
a partner of the Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.
*Related content and download information correct at time of download.
Production control of unreliable
manufacturing systems producing
defective items
F. Mhada
Department of Electrical Engineering, Ecole Polytechnique de Montreal,
Montreal, Canada
A. Hajji
Department of Operations and Decision Systems, Laval University,
Quebec, Canada
R. Malhame
Department of Electrical Engineering, Ecole Polytechnique de Montreal,
Montreal, Canada
A. Gharbi
Department of Automated Production Engineering, LCCSP Laboratory,
Ecole de Technologie Superieure, Quebec, Canada, and
R. Pellerin
Department of Mathematics and Industrial Engineering,
Ecole Polytechnique de Montreal, Montreal, Canada
Abstract
Purpose This paper seeks to address the production control problem of a failure-prone
manufacturing system producing a random fraction of defective items.
Design/methodology/approach A uid model with perfectly mixed good and defective parts has
been proposed. This approach combines the descriptive capacities of continuous/discrete event
simulation models with analytical models, experimental design, and regression analysis. The main
objective of the paper is to extend the Bielecki and Kumar theory, appearing under the title Optimality
of zero-inventory policies for unreliable manufacturing systems, under which the machine considered
produced only good quality items, to the case where the items produced are systematically a mixture
of good as well as defective items.
Findings The paper rst shows that for constant demand rates and exponential failure and repair
time distributions of the machine, the Bielecki-Kumar theory, adequately revisited, provides new and
coherent results. For the more complex situation where the machine exhibits non-exponential failure
and repair time distributions, a simulation-based approach is then considered. The usefulness of the
proposed models is illustrated through numerical examples and sensitivity analysis.
Originality/value Although the decisions taken in response to demands for productivity have a
direct impact on product quality, management quality and production management have been
traditionally treated as independent research elds. In response to this need, this paper is considered
as a preliminary work in the intersection between quality control and production control issues.
Keywords Manufacturing systems, Unreliable, Defective, Production control
Paper type Research paper
The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1355-2511.htm
JQME
17,3
238
Journal of Quality in Maintenance
Engineering
Vol. 17 No. 3, 2011
pp. 238-253
qEmerald Group Publishing Limited
1355-2511
DOI 10.1108/13552511111157362
1. Introduction
In the 1980s Akella and Kumar (1986), and subsequently Bielecki and Kumar (1988)
developed an optimal stochastic control analysis of a model of a failure prone machine
(with two states: up and down), subjected to a constant rate of demand of a single part
type, and with both storage and backlog costs. The main conclusion of their analysis,
with the objective of minimizing long-run average cost, was that there exists an
optimum, typically positive level of nished parts inventory one should aim at
maintaining whenever possible in meeting the demand, and because of the simple
structure of their problem, they were able to obtain an explicit solution. This type of
production policy became known as a hedging policy, which stipulates to produce at
full capacity if the stock is smaller than a threshold level, produce nothing if the stock
is higher than the threshold level, and produce at the demand rate when the stock is at
the threshold level. The hedging policy was also shown to be susceptible to
generalizations in multi part Kim and Gershwin (1985), and multi state failure prone
manufacturing systems (Sharifnia, 1988). In hedging policies, the critical inventory one
aims at maintaining is associated with a cost per unit time, and this cost acts much like
an insurance policy one pays upfront so as to hedge against future machine failures
which may induce shortages and ensuing large backlog costs. The
Akella-Bielecki-Kumar (AKB) theory has come to be regarded as an after the fact
analytically founded justication of the ideas behind Kanban and related
manufacturing production disciplines.
One of the main conclusions of the mathematical analysis of manufacturing systems
under uncertainty is that it is optimal for production to get sufciently ahead of
demand, or equivalently, that there is a need for storing inventory or work in process at
a level consistent with the level of uncertainty to meet product demand. While the
uncertainty alluded to in the AKB theory was rst and foremost a consequence of
unpredictable machine failures and repair times, in an effort to integrate quality issues
in the design of production policies, we consider here the impact of defective
production on inventory sizing. Note that quality is not merely an additional source of
uncertainty in the process. Its specicity is that it is only partially observable in
general, and thus decisions based on quality observations (such as preventive
maintenance for example) will unavoidably be tainted in part with error. In control
theoretic terms, the situation is such that control actions can affect system
observability properties, with the latter in turn affecting the quality of the control. At
this stage of our analysis, we circumvent the resulting difcult dual control problems
by assuming that a xed fraction of the production is always defective and that quality
observations of the inventory of nished parts are continuous and exact. The
Bielecki-Kumar theory is revisited under this modied model.
The rest of the paper is organized as follows: a brief review of the relevant literature
is presented in section 2. In section 3, our machine model and optimal control
formulation are presented; the solution to the optimal control problem is obtained by
displaying equivalence with a Bielecki-Kumar problem in section 4. Numerical results
are presented in section 5, while section 6 is our conclusion.
2. Literature review
To the best of our knowledge, only a limited number of papers have thus far considered
the intersection between quality control and production control issues, although as
Unreliable
manufacturing
systems
239
early as in 2003, Inman et al. (2003) have convincingly argued that there are several
classes of decisions simultaneously affecting quality and manufacturing productivity.
As we see it, the most comprehensive contribution has been a series of works by Kim
and Gershwin (2005, 2008; and Schick et al., 2005), who introduce dynamic models of
machines including in addition to operational and failure states, various (partially
observable) defective quality states. They also extend the decomposition methods of
Gershwin (1987) to incorporate the effects of quality states in the approximate
analytical assessment of buffer sizing/quality related decisions in transfer lines.
Finally, a systematic taxonomy of quality/quantity issues is presented in Schick and
Gershwin (2007). An important feature of the Kim-Gershwin models is that they
correspond to a uid view of the production process, although quality attributes
remain attached to discrete parts. At times, this can become a source of contradictions,
and in our modeling framework, attempts are made to embed quality also within a
uid paradigm.
By contrast with Kim and Gershwin, Colledani and Tolio (2005, 2006, 2011; and
Colledani, 2008) tackle simultaneous quality and production issues within an entirely
discrete framework: they consider a production system with manufacturing and
inspection machines. Machines are unreliable (they can fail in different modes). In
addition the behaviour of the machines can be monitored by control charts, which are
used to generate information about their state. A discrete Markov chains based
analysis is carried out for small size systems.
There are some works addressing the economic models of integrating production,
maintenance and quality: Chakraborty et al. (2008) presents a production system which
can move from an in control state to an out of control state, and is also subject to
random breakdowns. The system is periodically inspected and corrective actions are
taken, according to the state of the system; the objective is to characterize an optimal
inspection policy. Instead, in our model, we recognize that any manufacturing system,
even in an in control state, will produce on average a fraction of defective parts. In
such a model, transition towards an out of control state (not modeled in our paper)
will correspond to an increase in the fraction of defective parts produced on average.
3. Formulation of an optimal control problem
3.1 Manufacturing system model
We assume the manufacturing system consists of a single unreliable machine with two
states, or equivalently modes: an operational mode in which systematically, a constant
fraction b of the total number of parts produced is defective, and a failure mode with
zero production. The two modes evolve according to a Markov chain, with rates failure
rate p and repair rate r as shown in Figure 1. Furthermore, the following assumptions
are made:
(1) The machine M produces a single type of parts.
(2) The rate of demand for parts is a constant d.
(3) The maximum machine production rate is k.
(4) The production of parts is considered to be a uid process.
(5) There is an innite supply of raw parts.
(6) The nished parts storage space has a nite capacity to be specied.
JQME
17,3
240
(7) Good and defective nished part ows are perfectly mixed.
(8) Finished parts are tested as they are drawn to satisfy demand.
(9) Demand is satised from the uid nished parts mix, and the fraction of
detected defective parts is discarded.
(10) Demand is satised from the inventory if available and otherwise backlogged.
(11) Unit inventory storage cost is c
p
while backlog storage cost is c
n
.
(12) Inspection and handling costs additional relative to the perfect quality case are
assumed to be K per defective nished part pulled. So the additional cost
long-term unit time cost caused by the existence of defective nished part will
be proportional to the number of nished parts pulled per unit time, more
specically K
bd
12b
.
Since our model is based on the Bielecki-Kumar theory while the majority of
assumptions cited previously are derived from this model. The remaining assumptions
are part of our general three-state machine model (submitted to the 49th IEEE
Conference on Decision and Control, 2010).
We dene production rate u(t) as the number of parts produced by M per unit of
time: 0 # u(t) # k.
3.2 Inventory dynamics and optimal control formulation
As was the case for the Bielecki-Kumar model, machine M has a discrete state g with
two possible values: g = 1 (M operational), and g = 0 (M has failed). The difference
here though is that when M is operational it produces a mixed parts ow including
both good parts (a goodput) and defective parts. It is assumed that the ratio of
instantaneous defective parts ow to instantaneous total parts throughput is a
constant b. We designate by x
1
(t) the inventory of good parts when non negative, and
the backlog of good parts otherwise, while x
2
(t) and x(t) respectively designate the
inventory of defective parts, and the total inventory of parts. Note that x
2
(t) cannot
become negative; also, according to our model, x
1
(t) and x
2
(t) must reach zero at the
same time. Finally, when x(t) is negative, it is equal to x
1
(t) (there is no interest to know
the defective parts shortage since this quantity does not generate any cost). The system
dynamics can be written as follows:
dx
1
dt
= u(t)(1 2b)I [g(t) = 1] 2d (1)
Figure 1.
Modes of Machine M
Unreliable
manufacturing
systems
241
x
2
(t) =
b
1 2b
x
1
(t)I [x
1
(t) $ 0] (2)
x(t) = x
1
(t) - x
2
(t) (3)
where in the previous, I[.] is the indicator function.
Parallel to the Bielecki-Kumar analysis, our objective is to determine the optimal
feedback production control policy {u
f
} minimizing the following long-term measure
of combined storage and backlog costs:
J
T u
f
=
T!1
lim
1
T
E
_
T
0
c
p
x
-
(t) - c
n
x
2
t ( ) - K
bd
1 2b
_ _
dt
_ _
= K
bd
12b
- J
u
f
(4)
with x
-
(t) = max(x(t); 0); x
2
(t) = max(2x(t); 0) and
J
u
f
=
T!1
lim
1
T
E
_
T
0
c
p
x
-
(t) - c
n
x
2
t ( )
_ _
dt
_ _
(5)
So to minimizeJ
T u
f
, we have to minimizeJ
u
f
.
4. Analytical approach: equivalence to a Bielecki-Kumar problem
In the following, we display a reformulation of the optimal control problem equations
(1)-(5), which indicates that it is equivalent to two versions of Bielecki-Kumar problems
provided adequate machine/cost parameter modications are carried out.
4.1 First equivalence
Based on equations (2)-(3), and the assumption that when x(t) is negative, it is equal to
x
1
(t), it is possible to write the following relations:
x
-
(t) = x
-
1
(t) - x
-
2
(t)
= x
-
1
(t) -
b
(12b)
x
-
1
(t)
i.e.:
x
-
(t) =
1
(1 2b)
x
-
1
(t) (6)
x
2
(t) = x
2
1
(t) (7)
The cost to be minimized in equation (5) can now be rewritten in terms of the x
1
(t)
variable as:
J
u
f
=
T!1
lim
1
T
E
_
T
0
c
p
1 2b
_ _ x
-
1
(t) - c
n
x
2
1
t ( )
_ _
dt
_ _
JQME
17,3
242
With the x
1
(t) dynamics:
dx
1
dt
= ~ u(t)I [g(t) = 1] 2d
and : ~ u(t) = u(t)(1 2b):
As such, it becomes a Bielecki-Kumar problem with a unit storage cost increased to
c
*
p
= c
p
(1b)
21
and a maximum production rate decreased to k* = k(1b). This
allows us to conclude that the optimal policy is again a so-called hedging policy. It
involves producing whenever possible at the maximum rate k
*
in the ctitious
modied machine (with k the real production rate), until a critical inventory level z
*
of
good parts is attained, or equivalently a level (1 2b)
21
z* of total nished parts is
attained in the real system.
In the following, we detail the precise values of z
*
and the optimal cost J
*
according
to Bielecki and Kumar (1988).
Optimal (good parts) hedging point z
*
:
z
*
= 0 if
k(12b)p(c
n
-
c
p
(12b)
)
c
p
(12b)
(k(12b)2d )( p-r)
# 1 and
k(12b)2d
p
.
d
r
z
*
= 1if
k(12b)2d
p
,
d
r
z
*
=
1
r
d
2
p
(k(12b)2d )
log(
k(12b)p(c
n
-
c
p
(12b)
)
c
p
(12b)
(k(12b)2d )( p-r)
) otherwise
(10)
Optimal cost J
*
:
J
*
=
c
n
pk(12b)d
( p-r)(rk(12b)2rd2pd )
if z
*
= 0
J
*
=
c
p
(12b)
d
r-p
-
c
p
(12b)
r
d
2
p
k(12b)2d
log(
k(12b)p(c
n
-
c
p
(12b)
)
c
p
(12b)
(k(12b)2d )( p-r)
) if z
*
. 0
(11)
4.2 Second equivalence
We dene the effective variable ~ x(t) by the following relations:
~ x
-
(t) = x
-
(t) =
x
-
1
(t)
(1 2b)
(12)
~ x
2
(t) =
x
2
1
(t)
(1 2b)
(13)
The cost to be minimized in equation (5) can now be rewritten in terms of the ~ x(t)
variable as:
J
u
f
=
T!1
lim
1
T
E
_
T
0
c
p
~ x
-
(t) - c
n
(1 2b) ~ x
2
t ( )
_ _
dt
_ _
(14)
Unreliable
manufacturing
systems
243
with the ~ x(t) dynamics:
d~ x
dt
= u(t)I g(t) = 1
_
2
d
(1 2b)
(15)
As such, it becomes a Bielecki-Kumar problem with a unit backlog storage cost
decreased to c
n
* = c
n
(1b) and a constant demand rate increased to d* = d(1b)
21
.
Again, this allows us to conclude that the optimal policy is again a hedging policy with
values of z
*
and optimal cost J
*
specied in the following, and consistent with the
values in subsection 4.1.
Optimal hedging point (total nished parts) z
*
:
(16)
Optimal cost J
*
:
J (z
*
) =
c
n
(12b)pk
d
(12b)
( p-r)(rk2r
d
(12b)
2p
d
(12b)
)
if z
*
= 0
J (z
*
) =
c
p
d
(12b)
r-p
-
c
p
r
d
(12b)
2
p
k2
d
(12b)
log(
kp(c
n
(12b)-c
p
)
c
p
(k2
d
(12b)
)( p-r)
) if z
*
. 0
(17)
4.3 The boundaries between regions dened by distinct cases of z
*
:
The limit that denes the systems ability to meet the demand is given
by k (1 2b)
r
r-p
= d
_ _
. So if k (1 2b)
r
r-p
. d
_ _
= b , 1 2
d
k
r
r-p
_ _
, then the
system has enough capacity to meet the demand rate and in this case z* will be
nite. Otherwise we can conclude that the value of z
*
will be equal to 1. z* , 1
denes a security stock against future failure; this case includes two possibilities:
z* = 0 and z* . 0. The previous section development gives us a boundary condition
to differentiate between these two possibilities; this boundary condition is:
k (1 2b) (c
n
-
c
p
(12b)
) p
c
p
(12b)
( k(1 2b) 2d ) ( p - r)
= 1 (18)
Equation (18) denes a second order boundary condition with respect to (1 2b), which
can be written as:
a (1 2b)
2
- b (1 2b) - c = 0
where:
a ; k p c
n
; b ; 2k r c
p
and c ; d ( p - r) c
p
In this case a and c are positive and b is negative. The function f (b) =
a (X(b) 2r
1
) (X(b) 2r
2
) is a quadratic equation and r
1
and r
2
are its roots. The
discriminant of this quadratic equation is: D = b
2
24 a c
We consider the following cases:
(i) D , 0: i.e.: c
n
.
k c
p
r
2
4 d p ( p-r)
JQME
17,3
244
In this case f (b) = a ( X -
b
2a
)
2
2
b
2
24ac
4a
2
_ _
. 0 ; ;0 , b , 1 2
d (r-p)
k r
.
(ii) D . 0
In this case f (b) = a (X 2r
1
) (X 2r
2
) where r
1
=
2b2

D
_
2a
, r
2
=
2b-

D
_
2a
and
0 , r
1
, r
2
. We recall that X = 1 2b and 0 , b , 1 2
d (r-p)
k r
. So:
If max 0; (1 2r
2
)
_
, b , min (1 2r
1
) ; 1 2
d (r-p)
k r
_ _
then f (b) , 0 i.e. z* = 0.
If 0 , b , max 0; (1 2r
2
)
_
or min (1 2r
1
) ; 1 2
d (r-p)
k r
_ _
, b , 1 2
d (r-p)
k r
_ _
then f (b) . 0 i.e. z* . 0.
An interesting case arises from this study, namely the case when the values of c
n
verify C
1
n
;
k r 2d ( p-r)
k p
c
p
, c
n
,
k r
2
4 d p ( p-r)
c
p
; C
2
n
i.e.: (1 2r
2
) . 0 and D . 0.
For all c
n
satisfying the previously mentioned inequality, the optimal hedging point
z* displays the unusual behavior that:
z* . 0 for 0 , b , 1 2r
2
z* = 0 for 1 2r
2
, b , 1 2r
1
z* . 0 for 1 2r
1
, b , 1 2
d (r-p)
k r
_

_
(19)
This behavior can be justied by the fact that to minimize the total cost we have to nd
the hedging point that ensures a balance between two contradictory effects; the rst
effect is that the cost of storage increases as b increases and this tends to drive the
hedging point towards zero. The second effect is that a fraction of parts produced is
defective and this fraction increases as b increases; this in effect decreases machine
productivity thus justifying the need for more hedging.
Figure 2 illustrates the regions where z
*
is equal to zero or strictly positive for a
machine considered subject to a demand rate d = 1, with a failure rate p = 0:1, a repair
Figure 2.
Regions of z* . 0, z* = 0,
and z* = -1 in the
b 2c
n
plane for d = 1,
p = 0:1, r = 0:5, k = 4
c
p
= 1
Unreliable
manufacturing
systems
245
rate r = 0.5, a maximum production rate k = 4 and a storage cost c
p
= 1. We notice
that for some values of c
n
, the optimal hedging point z
*
can successively go through
z* . 0, z* = 0, z* . 0, and z* = -1 as b increases.
5. Numerical results
The analytical results are shown in Table I. The machine considered is subject to a
demand d = 1, with a failure rate p = 0.01, a repair rate r = 0.1, a maximum
production rate k = 1.5, a storage cost c
p
= 2, and a backlog cost c
n
= 15.
To illustrate the effect of the fraction of defective parts b variation on the optimal
hedging point z
*
and the optimal average cost J (z
*
), a sensitivity analysis was
conducted. Table I details this variation, and presents the optimal hedging point and
the incurred optimal cost for the sensitivity analysis cases. It clearly appears that the
results obtained make sense and conrm our expectations in the sense that when the
fraction of defective parts increases, the optimal hedging point and cost increase.
While the solution for a Markovian machine was obtained analytically, in the next
section, a simulation-based approach is proposed to deal with non-Markovian
situations.
6. Extension to non-exponential failure and repair times distributions
Within the framework of the classical control theory of the last 30 years, no
satisfactory method has been developed for the stochastic optimal control of
manufacturing systems subject to non-exponential machine up and down times. With
non-exponential failure and repair time distributions, analytical approaches such as
the one detailed in section 4 are very difcult to develop. However, that is the situation
that is usually encountered in real manufacturing systems. We refer the reader to Law
and Kelton (2000), chapter 6, for details on commonly used failure and repair
probability time distributions.
To deal with such situations we will adopt an approach, which has been
successfully used to control production, setup and preventive maintenance activities in
several manufacturing congurations (see Gharbi and Kenne (2003) and the references
therein). This approach combines the descriptive capacities of continuous/discrete
event simulation models with analytical models, experimental design, and regression
analysis. A block diagram of the resulting control approach is depicted in Figure 2.
The structure of the proposed control approach presented in Figure 3 consists of the
following sequential steps:
Analytical results
B z
*
0 10.509 J(z
*
)
0.05 12.483 39.201
0.1 15.551 44.105
0.15 20.967 51.303
0.2 33.494 63.324
Table I.
Analytical results
JQME
17,3
246
(1) The control problem statement of the manufacturing system, as shown in
sections 3 and 4, consists of an extension of the Bielecki-Kumar model. The
problem of the optimal ow control for the unreliable manufacturing system
producing a fraction of defective items is described in this rst step.
(2) The control factor z, for production rates control, describes the optimal control
policy.
(3) If the manufacturing system is facing exponential failure and repair time
distributions the analytical approach shown in sections 3, 4 and 5 provide the
optimal control policy. Otherwise, go to step 4.
(4) The simulation model uses the control policy dened in the previous step 2 as
the input for conducting experiments in order to evaluate the performances of
the non Markovian manufacturing system. Hence, for given values of the
control factor, the long-run average cost is obtained from the simulation model.
The simulation model was developed using Visual SLAM language (Pritsker
and OReilly, 1999).
(5) The experimental design step denes a one-factor experimentation plan with a
given number of replications (see Montgomery, 2001, for more details).
Figure 3.
Proposed approach
Unreliable
manufacturing
systems
247
(6) The regression analysis is then used to obtain the relationship between the
long-run average cost and the control factor. The obtained model is then
optimized in order to determine the best value of the control factor called here z
*
and the optimal incurred cost J (z
*
).
6.1 Simulation model
The simulation model consists of several networks, each of which describes a specic
task in the system (i.e. demand generation, control policy, states of the machines,
inventory control [. . .], etc.). The diagram of the proposed simulation model is shown in
Figure 4 with the following block notation descriptions:
(1) The INITIALIZATION block initializes the problem variables.
(2) The Demand Arrival block performs the arrival of a demand at each d
21
unit of
time. A verication test is then performed on the inventory level of good items,
and the inventory or the backorder is updated.
(3) The CONTROL POLICY segment block is dened in section 3 and governed by
a hedging point policy. The feedback control policy is dened by the output of
the FLAG block. This block is used to permanently verify the variation in the
stock level in order to specify the best action to carry out (production rate).
(4) The PARTS PRODUCTION block performs the production of nished goods
according to the policy dened by the CONTROL POLICY block.
Figure 4.
Diagram of simulation
model
JQME
17,3
248
(5) The update inventory blocks perform the variation of the inventory level of
defective and good items when a part is produced or when a demand arrival
occurs. Off-line runs of the simulation model using the control policy for z = 10
and a defective items rate b = 0 and 0.2, is illustrated in Figure 5. This
illustration makes it possible to observe the evolution of the inventory levels.
For example, when a failure occurs (arrow 1) a drop in the inventory levels is
observed. In such a situation the machine is under repair and the defective items
level is equal to zero (arrow 2). When the machine is repaired, arrows 3 and 4
show clearly the difference between the two cases. In fact, with a positive
defective items rate the machine needs more time to raise the inventory level
from the backlog zone.
(6) The failure-repair block performs two functions: it denes the time-to-failure of
the machine, and repairs a broken one.
(7) The update the incurred cost block calculates the cost of inventory, backlogs
and setup actions.
6.2 Validation of the analytical solution with the simulation-based approach
To crosscheck the robustness of the proposed approach and the validity of the
simulation results we carried on several experimentations and analysis detailed in the
following steps:
(1) Consider the analytical optimal hedging points z
*
(for the different cases of b)
as input and estimate the incurred cost J (z
*
) with the simulation model.
(2) Apply the simulation based approach to develop a regression model and to nd
a near optimal z
*
and the corresponding incurred cost for the different cases of b.
(3) Compare the estimated control policy to the optimal one through the Students
t-test of the following H0 hypothesis.
Figure 5.
Stock trajectory
Unreliable
manufacturing
systems
249
H0. The incurred cost for the estimated control policy Y
1
( )is different from
that given by the optimal control policy Y
2
( ).
The simulation model is used to generate two data samples related to the estimated
and optimal control policies. For each control policy, N replications are performed to
obtain N incurred cost observations. For a 1 2a condence level, H0 is rejected if the
value of the t distribution with y degrees of freedom and a condence level 1 2a
(i.e. t
y;12a=2
) is greater than the t given by:
t 2

Y
1
2

Y
2

S
2
1
-S
2
2
_ _
N21 ( )
2N22

2
_
N
_ (20)
where y = 2N 22,

Y
i
= 1=N
_ _

N
j=1
Y
ij
; S
2
i
= 1=N
_ _

N
j=1
Y
ij
2

Y
i
_ _
2
; i = 1; 2:
For N = 26 and a = 5%, the students t-test table gives t
50;0:975
= 2:145.
The results presented in Table II clearly show that the students t-test table value is
always greater than the t given by equation (20). Hence, the estimated control policy
performs as well the optimal control policy in the Markovian case.
6.3 Sensitivity analysis for non-Markovian cases
The hypothesis test validates the proposed approach and states that simulation
experiments could be combined with statistical analysis to obtain near-optimal
hedging point policies. For non-Markovian processes (i.e. non-exponential failure and
repair probability distributions), the hedging point policy is used as near-optimal
control policy with estimated stock threshold values. As illustrative examples, we
consider the following numerical values. To compare the Markovian versus the
Exponential Gamma Lognormal
b z
*
Cost J R
2
z
*
Cost J R
2
z
*
Cost J R
2
0 10.4927 39.2767 0.96 11.1402 40.0139 0.99 8.7766 36.8669 0.99
0.05 13.1277 43.9008 0.95 13.1425 44.8360 0.99 10.5418 41.2920 0.99
0.1 16.0390 51.7775 0.96 16.0892 52.0250 0.99 13.0591 47.8941 0.99
0.15 21.0839 63.5431 0.94 21.1492 64.6178 0.98 17.2181 59.3363 0.99
0.2 34.0644 90.0561 0.96 34.2881 93.5784 0.99 27.6148 85.8751 0.98
Table III.
Results for
non-Markovian cases
Analytical results
Simulation results
of the analytical z
*
Simulation results of
the near-optimal z
*
Students t-test
b z
*
J(z
*
) z
*
Y
2
z
*
Y
1
t
0 10.509 39.201 10.509 39.2710 10.4927 39.2767 0.07431234
0.05 12.483 44.105 12.483 43.8701 13.1277 43.9008 0.10464877
0.1 15.551 51.303 15.551 51.7690 16.0390 51.7775 0.02129824
0.15 20.967 63.324 20.967 63.5428 21.0839 63.5431 0.00048124
0.2 33.494 89.715 33.494 90.0558 34.0644 90.0561 0.00038697
Table II.
Validation
JQME
17,3
250
non-Markovian experimental results we have xed the standard deviation equal to the
mean for the Lognormal and the Gamma distributions.
.
Failure times: Gamma, Lognormal, with mean values equal to 100 and the same
standard deviation s = 100.
.
Repair times: Gamma, Lognormal, with mean values equal to 10 and the same
standard deviation s = 10.
The obtained results for the considered distribution and the different values of b are
given in Table III, where R
2
gives the proportion of the total variation in the estimated
cost attributable to the variability of the process. Hence, well-t regression models are
characterized by large R
2
values.
Given that the values of R
2
for the considered distributions are greater than that of
the exponential one, we can conclude that the obtained hedging point policy is
consistent with the behaviour of the considered manufacturing system.
7. Conclusion
We have presented an extension of the Bielecki-Kumar single part, single unreliable
machine subjected to a constant demand rate, with storage and backlog costs optimal
production control problem to the case where the machine, even when fully operational,
systematically produces a fraction of defective parts. A uid model with perfectly
mixed good and defective parts has been proposed. It has been established that the
optimal production policy is still of the hedging type with an associated critical
inventory of total parts, or equivalently, a critical inventory of good parts. The
presence of defective parts is seen to produce a combined increase in storage costs
relative to the perfect production case, and a reduction in the effective maximum
production rate of the machine. While the solution for a Markovian machine was
obtained analytically, the optimal hedging policy for the non-Markovian case was
obtained via simulation optimization techniques with the Markovian solution possibly
helping to center the parameter search space. We hope that these results can play the
role of building blocks in recent efforts aimed at integrating quality considerations
when designing production control policies in manufacturing systems.
References
Akella, R. and Kumar, P.R. (1986), Optimal control of production rate in a failure-prone
manufacturing system, IEEE Transactions on Automatic Control, Vol. AC-31, pp. 116-26.
Bielecki, T. and Kumar, P.R. (1988), Optimality of zero-inventory policies for unreliable
manufacturing systems, Operations Research, Vol. 36, pp. 532-41.
Chakraborty, T., Giri, B.C. and Chaudhuri, K.S. (2008), Production lot sizing with process
deterioration and machine breakdown, European Journal of Operational Research,
Vol. 185 No. 2, pp. 606-18.
Colledani, M. (2008), Integrated Analysis of Quality and Production Logistics Performance in
Asynchronous Manufacturing Lines, The International Federation of Automatic Control,
Seoul.
Colledani, M. and Tolio, T. (2005), Impact of statistical process control (SPC) on the performance
of production systems: part one (small systems), Proceedings of the 5th International
Conference on Analysis of Manufacturing Systems: Production Management, Zakynthos,
pp. 77-84.
Unreliable
manufacturing
systems
251
Colledani, M. and Tolio, T. (2006), Impact of quality control on production system performance,
Annals of the CIRP, Vol. 55, pp. 453-6.
Colledani, M. and Tolio, T. (2011), Integrated analysis of quality and production logistics
performance in manufacturing lines, International Journal of Production Research, Vol. 49
No. 2, pp. 485-518.
Gershwin, S.B. (1987), An efcient decomposition method for the approximate evaluation of
tandem queues with nite storage space and blocking, Operations Research, Vol. 35,
pp. 291-305.
Gharbi, A. and Kenne, J.P. (2003), Optimal production control problem in stochastic
multiple-product multiple-machine manufacturing systems, IIE Transactions, Vol. 35,
pp. 941-52.
Inman, R.R., Blumenfeld, D.E., Huang, N. and Li, J. (2003), Designing production systems for
quality: research opportunities from an automobile industry perspective, International
Journal of production Research, Vol. 41, pp. 1953-71.
Kim, J. and Gershwin, S.B. (1985), Flow optimization in exible manufacturing systems,
International Journal of Production Research, Vol. 23, pp. 81-96.
Kim, J. and Gershwin, S.B. (2005), Integrated quality and quantity modeling of a production
line, OR Spectrum, Vol. 27, pp. 287-314.
Kim, J. and Gershwin, S.B. (2008), Analysis of long ow lines with quality and operational
failures, IIE Transactions, Vol. 40, pp. 284-96.
Law, A.M. and Kelton, W.D. (2000), Simulation Modeling and Analysis, 3rd ed., McGraw-Hill,
New York, NY.
Montgomery, D.C. (2001), Design and Analysis of Experiments, 5th ed., John Wiley & Sons,
New York, NY.
Pritsker, A.A.B. and OReilly, J.J. (1999), Simulation with Visual SLAM and AWESIM, John
Wiley & Sons, New York, NY.
Schick, C. and Gershwin, B. (2007), A taxonomy of quality/quantity issues in manufacturing
systems, The Analysis of Manufacturing Systems: AMS, pp. 1-6.
Schick, I.C., Gershwin, S.B. and Kim, J. (2005), Quality/Quantity Modeling and Analysis of
Production Lines Subject to Uncertainty, nal report presented to General Motors Research
and Development Center, May 10.
Sharifnia, A. (1988), Production control of a manufacturing system with multiple machine
states, IEEE Transactions on Automatic Control, Vol. 33, pp. 620-5.
About the authors
F. Mhada has an engineering degree (2000) from Ecole Mohammedia dIngenieurs (Morocco) in
automatic. She is preparing for a PHD degree in E

cole Polytechnique of Montreal supervised by


Roland P. Malhame and Robert Pellerin. Her main research interest is manufacturing systems
analysis, modeling, and optimization.
A. Hajji received his undergraduate degree in Mechanical Engineering from Ecole Nationale
dIngenieurs de Tunis, Tunisia (1999) and his MEng (2003) and PhD (2007) in Automated
Production Engineering, both from Ecole de Technologie Superieure, Montreal. After two years
as Postdoctoral Researcher at Ecole Polytechnique de Montreal, in 2010 he joined Laval
University, where he is Professeur Adjoint. He gained practical and research experience through
several collaborations and applied research projects. His main research interest is manufacturing
systems modelling, and control, simulation as well as integrated reactive models in ERP
systems.
JQME
17,3
252
R. Malhame received the Bachelors, Masters and PhD degrees in Electrical Engineering from
the American University of Beirut, the University of Houston, and the Georgia Institute of
Technology in 1976, 1978 and 1983 respectively. After a single years stay at the University of
Quebec (Chicoutimi), and CAE Electronics Ltd (Montreal), he joined Ecole Polytechnique de
Montreal in 1985, where he is Professor of Electrical Engineering. In 1994 and 2004, he spent
sabbatical periods with LSS CNRS, and E

cole Centrale de Paris, France respectively. His current


research interests are in stochastic control, and the analysis and optimization of complex
networks, in particular manufacturing and communication networks. Past contributions include
a statistical mechanics-based approach to electric load modeling. He is Director of Groupe
dEtudes et de Recherche en Analyse des Decisions (GERAD).
A. Gharbi is a Professor in the Department of Automated Production Engineering. He is also
the Director of C2SP laboratory at Ecole de Technologie Superieure in Montreal. He received his
Engineering Degree in Industrial Engineering from the Universite du Quebec a` Trois-Rivie`res in
1986 and his MScA and PhD in Industrial Engineering, both from Ecole Polytechnique de
Montreal in 1988 and 1992, respectively. His main research interests are in manufacturing
system design and control using simulation, experimental design and OR modelling. Other
research and teaching interests include multiple criteria decision making, maintenance, project
planning, and control of overhaul manufacturing systems and quality control.
R. Pellerin is Full Professor in the Department of Mathematics and Industrial Engineering at
Ecole Polytechnique de Montreal. He holds degrees in Engineering Management (BEng) and
Industrial Engineering (PhD). He has practised for more than 12 years in production
management and enterprise resource planning (ERP) systems implementation. He is also a
Certied Professional in Operations Management (CPIM) and Project Management (PMP). His
current research interests include project planning, manufacturing execution, and enterprise
system implementation and integration. He is a member of the CIRRELT research group and
he is the current Chairman of the Jarislowsky/SNC-Lavalin Research Chair in international
project management. R. Pellerin is the corresponding author and can be contacted at:
robert.pellerin@polymtl.ca
Unreliable
manufacturing
systems
253
To purchase reprints of this article please e-mail: reprints@emeraldinsight.com
Or visit our web site for further details: www.emeraldinsight.com/reprints