Professional Documents
Culture Documents
a r t i c l e in f o abstract
Available online 22 December 2010 This paper aims at presenting a method to optimize the maintenance planning for a commercial heavy
Keywords: vehicle. Such a vehicle may be considered as a multi-components system. Grouping maintenance operations
Condition-based maintenance related to each component reduces the global maintenance cost of the system. Classically, the optimization
Multi-component maintenance model problem is solved using a priori reliability characteristics of components. Two types of methods may be used,
Maintenance grouping i.e. static or dynamic methods. Static methods provide a xed maintenance planning, whereas dynamic
Rolling horizon methods redene the groups of maintenance operations at each decision time. Dynamic procedures can
Gamma process incorporate component information such as component states or detected failures. For deteriorating systems,
reliability characteristics of each component may be estimated thanks to deterioration models and may be
updated when a degradation measure is available. This additional information on degradation features allows
to better follow the real state of each component and to improve the maintenance planning.
& 2010 Elsevier Ltd. All rights reserved.
0951-8320/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2010.11.009
602 K. Bouvard et al. / Reliability Engineering and System Safety 96 (2011) 601610
manual. Maintenance operations are scheduled at xed time methods with rolling horizon: They optimize repeatedly the
intervals. These maintenance time intervals depend on the vehicle maintenance cost on a nite horizon at each decision time. They
conguration and the specied vehicle usage. This usage is dened are based on a long-term plan which can be adapted according
by the customer expectations but sometimes the current vehicle to the available short-term information.
usage does not correspond to the actual one:
A commercial heavy vehicle is used during several years. It is then
If the vehicle operates in more severe environment than the justied to choose a rolling horizon method for such a system.
specied one, some components will wear more rapidly and Dynamic maintenance optimization method uses a priori
could cause a vehicle immobilization. Maintenance cost will reliability characteristics of components to update the mainte-
then increase with the immobilization and corrective costs, nance planning at each decision time. Information about compo-
which are very expensive. nents like the component age, the component use or a detected
If the actual vehicle usage is less severe than the specied one, failure can be incorporated.
some components will be changed before their actual end of
useful lifetime. In this case, the maintenance cost could be 1.4. Main contribution of the paper
reduced if the early replacements were avoided.
Scheduling maintenance operations on the long-term means to use
The idea is to develop a maintenance model which is able to an innite horizon or a rolling horizon. Although the computations are
optimize dynamically the maintenance decisions on a multi- exible on an innite horizon, short-term information like component
component system. This method should be based on monitoring failure or degradation level cannot be included into the maintenance
component information and should group dynamically the main- scheduling. Using a rolling horizon brings the advantage to adapt the
tenance operations. maintenance decision when short-term information is available.
Maintenance cost optimization is performed on a nite horizon
1.3. Literature overview which is specied at the current inspection date. Optimization is then
repeated on a new dened horizon at the next inspection date and so
Maintenance models were rst built to manage the mainte- on. This process reduces the system maintenance cost on a long-term
nance operations of a single component. Then, researches focus on horizon.
the modelling and optimization of maintenance of multi-compo- Our method aims at scheduling system maintenance operations
nent systems [6]. and at decreasing the system maintenance cost. Maintenance cost
A maintenance model is composed of a model of system optimization is performed on a rolling horizon.
reliability behavior and of a maintenance policy model. The main A rst idea is to group several maintenance operations at each date.
aim of a maintenance model is to evaluate the overall maintenance The grouping should reduce the maintenance stop number on the
cost incurred by the maintenance decisions. scheduling horizon and then should reduce the logistic cost on the
Many overviews of possible maintenance strategies for multi- system maintenance cost. Even if executing j maintenance operations
components systems are given by [19,10,12]. Multi-components at the same date should save on the logistic cost comparing to j
maintenance optimization methods are based on [15]: maintenance operations which are executed at j different dates, the
system maintenance cost could be indirectly penalized:
economic aspect: grouping N maintenance operations saves
cost compared with N individual maintenance actions;
with the reduction of the component useful life if the main-
tenance dates are advanced,
failure interaction: the failure of a component induces the
failure of other components;
with the increasing of component failure probability which
could imply a system immobilization if the maintenance date is
structural relationship: if some components form a part and one
executed too late.
of them fails, the others should be maintained at the same time.
To incorporate the current component degradation level into with G the Gamma function where
maintenance decisions, we introduce deterioration models. Z 1
Gq uq1 expu du 2
0
2.1. Deterioration models The failure probability is dened by
Fi t PZi t Z Zi,failure
Suppose that component degradation can be characterized by a
physical variable Zi with i{1,y,n}. For each component, one
Gat, bZi,failure Zi 0
failure mode and one maintenance operation are considered. Fi t 3
Gat
From an initial degradation level Zi(0), the deterioration model
can estimate the future degradation level Zi(t) of a component in a with
dynamic environment [14]. The model is based on a stochastic Z 1
degradation process. A threshold failure set Zi,failure is dened from Gu,v xu1 expx dx 4
v
technical specications to determine the component failure
(Fig. 1). When the degradation level Zi reaches Zi,failure, the
component i is considered as failed. By iteration, the model can Fig. 2 illustrates three degradation paths Zi(t) and the failure
compute the failure probability function Fi(t) (Fig. 2). probability function Fi(t) of the component i.
Assume that at each inspection date Tk, the component degra- Note E and Var, respectively, the mean and the variance of the
dation levels are automatically retrieved and the degradation Gamma process.
processes are modelled by Gamma processes. This type of process
describes a gradual and monotone degradation [1,18].
Zi t,t 4 0 is a Gamma process with shape parameter a and scale 2.2. Maintenance assumptions & operation grouping
parameter b:
Suppose that each component follows a classical age-replace-
Zi(0) 0, ment policy, [3]. A component is replaced when it reaches a
Zi(t) has independent increments, specied operational age or when it fails.
For t 4 0 and h4 0, Zi t hZi t follows a Gamma distribution: Each maintenance stop cost is considered as a combination of a
xed set-up cost S and a specic maintenance cost si . The xed set-
bah ah1
f z z expb z 1 up cost S represents the downtime cost and includes the production
Gah
loss during maintenance operation. The cost S is independent on
the considered operation i. The specic cost si of operation i
depends on the time-to-failure Ti,failure:
si t CP , 8t oTi,failure
i
200
Degradation level Z
150
Zi,failure
100
50
0
0 10 20 30 40 50 60 70 80 90 100
Time
1
Failure probability
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70 80 90 100
Time
Fig. 2. Different degradation paths and associated failure probability functions. a 0:4; b 0:2; Zi 0 0; Zi,failure 100.
604 K. Bouvard et al. / Reliability Engineering and System Safety 96 (2011) 601610
We assume that the time to replace a component is negligible. cost J2 on HP will be decomposed of u set-up costs S and the specic
Considering the following notations: costs si .
u X
X
HP is the scheduling horizon and the optimization maintenance J2 si tG j u S 6
horizon. On this horizon HP, m maintenance operations are to be j 1 i A Gj
scheduled. It is possible to have more than one occurrence of a with tG j the optimal maintenance date of the group Gj.
maintenance operation on this horizon HP. For each group Gj of Z components, we save a cost CGj :
Gu is a maintenance group, it is composed of maintenance X
operations which are mutually different and executed at the CGj Z1 S si tG j si yi 7
i A Gj
same date tG u . Gu is a nonempty subset of {1,y,m} and is the uth
group of the grouping structure SGk,
where si tG j si yi is the additional cost when the operation i is
SGk is a grouping structure which is built at each inspection date executed at tG j instead of yi .
Tk. SGk is a collection of mutually exclusive groups {G1,y,Gu} Maintenance optimization aims at nding the grouping struc-
with u the number of groups on the scheduling horizon. ture with a minimal maintenance cost J2 on a nite horizon HP.
Maintenance groups of the grouping structure SGk cover all
maintenance operations, that is
3 Gj \ Gl |, 8j,l A f1, . . . ,ug2 , 3. Algorithm of the adaptive maintenance policy
3 [l A f1,...,ug Gl f1, . . . ,mg.
In this section, we propose a method to provide an optimized,
Maintenance scheduling has to consider several system dynamic and well-suited maintenance planning for a multi-com-
constraints: ponents system like as for instance commercial heavy vehicle. This
method uses degradation models with a maintenance optimization
A commercial heavy vehicle must guarantee a minimal avail- using a rolling horizon.
At each decision time Tk , the procedure is running through
ability between two maintenance stops. This constraint sup-
poses that a minimal time d must be respected between two 4 steps, see (Fig. 3):
consecutive planned stops and that the time between two
occurrences of the same operation has to be greater than d. estimation of failure probability function;
Maintenance preparation and organization requires a minimal individual maintenance optimization;
time dmin . This constraint supposes that the time between the grouping maintenance optimization;
current date and the next maintenance stop has to be superior decision step and maintenance planning update.
than dmin .
these functions, the optimal dates of maintenance operations can previously computed (see Fig. 5):
be computed according to the chosen maintenance policy and the
additional costs can be evaluated if the maintenance rules are not HP 0, max yi 11
i A f1,...,ng
respected.
The vehicle maintenance plan is updated at each inspection date With this denition, the n components of the system are taken into
Tk. The time origin is re-scheduled at t 0 to simplify the notations. account for the maintenance decisions. Moreover, the complexity
time of grouping computation is reduced with this nite horizon.
3.2. Step 2: individual optimization Maintenance cost optimization is then performed on this
horizon HP.
In this step, the aims are at:
3.2.3. Incorporating other maintenance operation occurrences
identifying the optimal maintenance date of each operation, In practice, a system is composed of components which wear
dening the scheduling horizon HP, with different degradation speeds. For example, engine oil has to be
determining the maintenance operation to schedule on the changed more frequently than driving belt on a commercial heavy
horizon HP, vehicle. However, the scheduling horizon HP covers maintenance
building the additional cost function of each maintenance operations of the n system components. Several occurrences of
operation. maintenance operations have to be scheduled on this horizon for
components with fast degradation speed.
For the rth occurrence of operation i, the maintenance date
3.2.1. Computing the optimal maintenance dates yri ,r Z2 is dened using the optimal maintenance date yi of operation
Let us dene yi the optimal maintenance date of operation i. On i and the nominal replacement age xni of the associated component.
the scheduling horizon, it is supposed that maintenance decisions Since cost optimization is performed on the horizon HP, we
are stationary and are updated when short-term information is consider the occurrences of operations which are scheduled before
available. yi represents the date at which the mean maintenance maxi A f1,...,ng yi .
cost of operation i per time unit on a innite horizon, noted Ci , is yri r Z2 veries (see Fig. 6):
minimal (see Fig. 4). For an age-replacement policy [3,17], this yri yi r1 xi 12
asymptotic cost is given by
with
Q (
S CPi bi 1 t1 0
l 0 Fi ljzi rZ2
0
Ci tjzi P Q , 8t w e,w A N 8 r
1 tj 2 lj2
F tjz0i
0 i
maxyi rmaxi A f1,...,ng yi
with bi CCi CPi The date yi is deduced from the following
equation: 3.2.4. Building the penalty cost function
To nd the best balance between reducing the useful life of some
dCi 0 components and increasing the failure probability of the others, we
y jz 0 9
dt i i introduce the penalty functions (Fig. 7).
n
Note xi the nominal and optimal replacement age of component A penalty function hi evaluates the expected additional cost of
when it is new. Optimal asymptotic maintenance cost of operation i shifting the maintenance time from the optimal maintenance date
is dened by yi . The detailed procedure to compute these penalty functions is
given in [7,21].
Ci Ci xi jz0i 0 10
The penalty function hi associated to operation i depends on:
3.2.2. Dening the scheduling horizon the optimal maintenance date yi ,
Scheduling horizon HP is dened according to the current the preventive and corrective replacement cost of the operation
inspection date and the optimal maintenance dates yi A f1,...,ng i respectively CPi and CCi ,
Fig. 7. Penalty cost of a component whose deterioration process follows a Gamma process Ga(2,1) with maintenance costs S 20, CPi 10 and CCi 90.
the failure probability of the component Fi tjz0i , We suppose that the optimal maintenance dates are sorted in
the survival probability of the component Rtjz0i 1Fi tjz0i . increasing order and noted tni with i A f1, . . . ,mg where m represents
the maintenance operation number to schedule on the horizon HP.
The penalty function hi of the rst occurrence of a maintenance We dene HGu ti the penalty function of the group Gu with u A N
operation i is dened as [7,21] when the operations i A Gu are executed at the date ti ti Di t
8P
0
h
Qq1
i instead of the date tni . Di t is the delay or advance time from the
>
< q k 1 e Ci Fi yi qejz0i :bi p k 1 Ri yi pejz0i , k o0
hi yi ke Pk h i optimal maintenance date tni . For a maintenance group Gu, the
> Q
: q 1 Fi yi q ejz0i bi e Ci q1
0
p 1 Ri yi p ejzi , k Z0 optimal replacement date tG u is dened by
!
13 X
HGu tGu HGu min hi t 16
t
with the empty product equal to 1. i A Gu
We suppose that the previous maintenance operation is exe-
cuted at its optimal maintenance date. According to this assump- If the maintenance operations i A Gu are executed at the date tG u , the
tion, we build the penalty cost function of the rth occurrence which saving equals the difference between the saved set-up costs thanks
veries: to the maintenance grouping and the total penalty cost due to the
8 early and delayed replacements. For the group Gu, the saving QGu
> 1, 8kr n
>
> P0
>
h Qq1 i veries at the date tG u the following property:
>
< q k 1 e Ci Fi xi q e bi p k 1 Ri xi p:e , 8n o ko 0
r
hri yi ke
> 0, k0
QGu tG u cardGu 1 SHG u 17
>
> h i
>
> P Q
: kq 1 Fi xi q e bi e Ci q1
p 1 Ri xi p e ,
8k4 1 If the saving is positive then the group Gu is considered as cost-efcient.
14 The aim is to nd the grouping structure SGk, which is cost-
efcient and which minimizes the system maintenance cost on the
with n the oor of r1 xi =e. scheduling horizon HP at each inspection date Tk.
In failure case, the penalty function hi of the failed component is Dekker et al. [7] developed an algorithm to deliver the grouping
dened as structure with only one occurrence of operation on the scheduling
(
0, si t 0 horizon. This structure groups operations whose optimal main-
hi t 15 tenance dates are consecutive. Wildeman [21] enhanced this
1, 8t 40
algorithm in reducing the computation complexity.
With this assumption, opportunistic maintenance operation could We adapt this algorithm to nd heuristically the optimal
be scheduled during the system immobilization time. grouping structure on the scheduling horizon HP. This adapted
algorithm veries the following constraints, see Fig. 8:
3.3. Step 3: grouping maintenance optimization
between two consecutive maintenance stops, the system has to
Step 3 aims at grouping the maintenance operations from the guarantee a minimal availability time d,
previous step to minimize the system maintenance cost on the grouping several occurrences of the same operation is
scheduling horizon HP. forbidden,
K. Bouvard et al. / Reliability Engineering and System Safety 96 (2011) 601610 607
cost optimization is frozen when the preparation maintenance maintenance group Gn1 are xed. The maintenance planning is then
time dmin is reached. updated at the date D1 tG1 .
When Gu veries these assumptions, the maximal saving QGu is with u the number of maintenance stops on the cycle length Tcycle
X
evaluated. The same computation is performed for the mainte- sDp si Dp 21
nance groups Gu1 fi 1, . . . ,jg and Gu2 fi, . . . ,j1g. i A Gp
n
If QG is smaller than QG or QG , then the maintenance group Gu with Gp the maintenance group, which executed at the date Dp and
u u1 u2
(
cannot be part of the optimal structure because there are CPi in case of preventive replacement
subsets of the group Gu whose saving is higher than the saving si Dp C in case of corrective replacement 22
Ci
QGu , [7].
Else Gu can be part of the optimal grouping structure. Fig. 10 presents several simulated paths of the mean maintenance
cost per time unit for the considered system on which the adaptive
This algorithm delivers the list of maintenance groups which policy is applied with an inspection period t equal to 6. The mean
could be part of the optimal grouping structure. Maintenance maintenance cost per time unit tends to an asymptotic value, which
groups are dened with their optimal maintenance date and the is noted C2 t. Note C1 the asymptotic mean maintenance cost per
associated saving. The optimal grouping structure SGk is dened time unit of the studied system with the proposed scheduling
with the maintenance groups which maximize the total saving Q
and veries the operational constraint. The total saving Q is the sum Table 1
of the group savings QGu on the scheduling horizon HP: Component deterioration parameters.
X
Q QGu 19 Component i ai bi Ei Vari Zi,failure CPi CC i
Gu A SGk
System 1 1 1 1 1 1 100 80 980
Note tG p , the pth planned maintenance date computed by the 2 1.5 1.5 1.5 1 100 80 980
algorithm and Gp, the maintenance group which is executed at this 3 2 2 2 1 100 80 980
date (see Fig. 8). System 2 1 2E 1 4E 2 1 5 100 80 980
2 3E 1 6E 2 1.5 5 100 80 980
3.4. Step 4: update and decision 3 4E 1 5E 2 2 5 100 80 980
Fig. 10. Maintenance cost per time unit with an adaptive maintenance policy applied on the studied system with an inspection period t which is equal to 6.
Fig. 11. Inspection period inuence on the asymptotic maintenance cost per time Fig. 12. Percentage of the maximal estimated saving costs with the adaptive
unitC1 the asymptotic mean maintenance cost per time unit with the classical maintenance policy compared to the classical dynamic policy with respect to the
dynamic policyC2 the asymptotic mean maintenance cost per time unit with the variance of deterioration process increments with the inspection period equal to 1.
adaptive maintenance policy.
increase with the multiplication of specic instrumentation and 4.2.2. Condition-based operation
commercial heavy vehicle will be more expensive. It is then not For monitored components, we suppose that the penalty cost
realistic both from a technological and an economical point of view function is expressed according to the Eq. (14) and depends on:
to consider that all vehicle components are monitored.
Accordingly, the main practical issue is that the commercial the optimal maintenance date yi ,
heavy vehicle has to be maintained according to different policies: the preventive and corrective replacement costs of the opera-
tion i, respectively, CPi and CCi ,
the failure probability of the component Fi tjz0i ,
Some vehicle components are replaced when they reach their the survival probability of the component Ri tjz0i 1Fi tjz0i ,
optimal operational age yi . This age is specied by suppliers and computed from step 1 (Section 3.1).
maintained by an age-based operation.
Others vehicle components, which can be monitored, are Thanks to these assumptions, it could be possible to bring forward
maintained according to their degradation condition. and delay the maintenance execution from the optimal mainte-
nance date.
In this part, we propose an adaptation of the adaptive policy which However, the corrective replacement cost CCi is very difcult to
was previously presented: Step 1 of the adaptive method considers evaluate in practice. This cost represents the cost when the
only the second class of components. Steps 3 and 4 are running in component which is maintained by the operation i fails and it is
the same way as presented in Sections 3.3 and 3.4. To take account unthinkable to leave a failed component on the vehicle in order to
of the applicative constraints, we need to introduce additional compute its corrective cost! The idea is to link the component
assumptions to build the penalty cost function of operations in step corrective cost with the vehicle immobilization cost Cimmo. A pre-
2. These assumptions on the penalty cost function are developed liminary work aims at classifying vehicle components, which are
according to the type of component in the following sections. based according to their degradation level, according to their
impact level on the vehicle immobilization. If a component fails
and brings immediately vehicle to a standstill, its impact level is the
highest. On the contrary, the components which do not provoke
4.2.1. Age-based operation immediately vehicle immobilization have a lower impact level. The
Only the optimal operational age yi is available for components more critical the component is for the risk of vehicle immobiliza-
which follow an age-based policy. No lifetime law is available. In tion, the higher the corrective replacement cost is. Then a relation-
this case, we suppose that the penalty cost function of this age- ship between the preventive component replacement cost CPi , the
based operation i decreases linearly from the preventive replace- component impact level on vehicle immobilization and the vehicle
ment cost CPi to 0 between the last replacement time and the time immobilization cost Cimmo is used to compute the component
when the maintained component reach yi . No penalty cost is added corrective replacement cost CCi .
when the component is maintained at its optimal operational age Another issue is to evaluate the vehicle immobilization cost.
yi . Thanks to these assumptions, the age-based operations can be This cost depends on several operational parameters like the
brought forward since suppliers guarantee that the components customer activity, the market where the vehicle runsyand all
run properly before their optimal operational age yi . However, no these contributions are often not easy to quantify in practice.
information about reliability component is available after the age
yi . It is then better to avoid a maintenance delay for this type of
components. The penalty cost function of this age-based operation i 5. Conclusions
is assumed to be increasing from 0 to the immobilization vehicle
cost, noted Cimmo. This cost Cimmo represents all costs which are In this paper, we propose an adaptive and dynamic method to
imputed for an unscheduled vehicle stop. The penalty function hi of schedule maintenance operations of a multi-component system
an age-based operation i veries: using online condition monitoring information in the maintenance
8 grouping decision rule. Dynamic maintenance grouping is per-
< S CPi ke,
>
8k A N=yi r ke r 0 formed to optimize the maintenance costs on a rolling horizon
e
hi yi k xi 23 and it is shown on numerical examples that this approach allows
>
: SC
immo, 8k A N=ke 4 0 for cost savings which can justies the decision of investing in
condition monitoring devices and system, and to switch to more
Fig. 13 illustrates the basic form of the penalty function of an age- dynamic maintenance plans.
based maintenance operation. In this perspective, a European FP7 project Maintenance On
Demand started in September 2009, which deals with the main-
tenance issues for commercial heavy vehicles [2]. The aim is to
schedule maintenance operation using this kind of adaptive
approach, while further incorporating other operational and busi-
ness constraints. Moreover, the component deterioration models
will be performed using collected wear data and covariates.
References
[4] Budai G, Huisman D, Dekker R. Scheduling preventive railway maintenance [13] Rifkin J. The age of access: the new culture of hypercapitalism, where all of life
activities. Journal of the Operational Research Society 2006;57(9):103544. is a paid-for experience. Putnam Publishing Group; 2000.
[5] Castanier B, Grall A, Berenguer C. A condition-based maintenance policy with [14] Singpurwalla ND. Survival in dynamic environment. Statistical Science
non-periodic inspections for a two-series system. Reliability Engineering and 1995;10(1):86103.
System Safety 2005;87(1):10920. [15] Thomas LC. A survey of maintenance and replacement models for maintain-
[6] Cho D, Parlar M. A survey of maintenance models for multi-unit systems. ability and reliability of multi-item systems. Reliability Engineering
European Journal of Operational Research 1991;51(1):123. 1986;16(4):297309.
[7] Dekker R, Smit A, Losekoot J. Combining maintenance activities in an opera- [16] Van der Duyn Schouten FA, van Vlijmen B, Vos de Wael S. Replacement policies
tional planning phase: a set-partitioning approach. IMA Journal of Mathe- for trafc control signals. IMA Journal of Mathematics Applied in Business &
matics Applied in Business and Industry 1992;3:31531.
Industry 1998;9:32546.
[8] Dekker R, Scarf PA. On the impact of optimisation models in maintenance
[17] van der Duyn Schouten FA, Vanneste SG. Analysis and computation of (n,N)-
decision making: the state of the art. Reliability Engineering and System Safety
strategies for maintenance of a two-component system. European Journal of
1998;60(2):1119.
Operational Research 1990;48:26074.
[9] Dekker R, Wildeman RE, van der Duyn Schouten FA. A review of multi-
component maintenance models with economic dependence. Mathematical [18] Van Noortwijk JM. A survey of the application of gamma processes in
Methods of Operations Research 1997;45:41135. maintenance. Reliability Engineering and System Safety 2009;94(1):221.
[10] Deloux E, Castanier B, Berenguer C. Predictive maintenance policy for a [19] Wang H. A survey of maintenance policies of deteriorating systems. EJOR
gradually deteriorating system subject to stress. Reliability Engineering & 2002;139:46989.
System Safety 2008;94(2):41831. [20] R.E. Wildeman, The art of grouping maintenance. PhD thesis, Erasmus
[11] Marquez AC, Gupta JND. Contemporary maintenance management: process, Universiteit Rotterdam, 1996.
framework and supporting pillars. Omega 2006;34(3). [21] Wildeman RE, Dekker R, Smit ACJM. A dynamic policy for grouping main-
[12] R.P. Nicolai, Maintenance models for systems subject to measurable deteriora- tenance activities. European Journal of Operational Research 1997;99(3):
tion. PhD thesis, Erasmus Universiteit Rotterdam, 2008. 53051.