You are on page 1of 8

Available online at www.sciencedirect.

com

European Journal of Operational Research 192 (2009) 479–486


www.elsevier.com/locate/ejor

Stochastics and Statistics

A facility location and installation of resources model


for level of repair analysis
Eduardo Siqueira Brick *, Eduardo Uchoa
Departamento de Engenharia de Produção, Universidade Federal Fluminense, Rua Passo da Pátria 156, sala 306,
Bloco D, Niterói, RJ 24210-240, Brazil

Received 25 October 2006; accepted 30 August 2007


Available online 1 October 2007

Abstract

The Level of Repair Analysis – LORA – is an analytic methodology aimed at determining: (i) the optimal location of facilities that
compose a maintenance structure; (ii) the quantity of required resources in each facility; and (iii) the best repair policies, i.e., rules that
determine if a given component should be discarded or repaired, and where those actions should take place. This work presents a mixed-
integer programming model for LORA that is more comprehensive than others in the literature, being suitable to many practical situ-
ations. The model was applied to 15 substantial real world problems, and considering distinct maintenance policies to some of them,
resulted in 22 different solutions, all of which could be achieved by a commercial Mixed-Integer Programming (MIP) solver in reasonable
times.
 2007 Elsevier B.V. All rights reserved.

Keywords: Maintenance; Level of Repair Analysis; Capacitated Location Problem; Integer Programming

1. Introduction A technical system is composed by a large number of


components that can be modeled by tree structures,
Level of Repair Analysis (LORA) is an analysis method- according to various criteria. These structures may encom-
ology used to determine the optimal location of facilities pass many levels, known as indentation levels. For the sake
and equipments for the maintenance of technical systems of LORA, the suitable criterion to place a component in a
components and the more appropriated repair policies. particular structure indentation level is the relation con-
Such systems may be used to provide services, to extract tain–contained (father–son), associated with the possibility
raw materials, or to manufacture products. LORA is fre- that it can be removed from its father and substituted by a
quently applied to systems deployed in different locations, spare. Depending on the maintenance resources available
spread over large geographic regions, possibly covering at the places where the systems are deployed, only a subset
more than one continent. Examples of such systems include of the components can be removed and substituted at these
a surveillance system composed of radars extended along a locations. These are called Line Replaceable Units (LRU).
border, a communications system composed by a set of The components that can be removed and replaced only at
switching equipments located in different cities, or a places where special facilities exist are called Shop Replace-
defense system that relies on aircrafts distributed among able Units (SRU). We assume that a SRU may contain
various airbases. other SRU, but not a LRU.
Fig. 1 depicts part of the tree structure of a complex sys-
* tem. It represents an aircraft, illustrating the fact that these
Corresponding author.
E-mail addresses: brick@producao.uff.br (E.S. Brick), uchoa@produ- components may be located at any indentation level in the
cao.uff.br (E. Uchoa). structure.

0377-2217/$ - see front matter  2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2007.08.043
480 E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486

AIRCRAFT A1

NAVIGATION MOTOR LRU 1 COMMUNICATION


SUBSYSTEM SUBSYSTEM SUBSYSTEM

COMPRESSOR TURBINE LRU 2 SRU 11 SRU 12

LRU 3 LRU 4 LRU 5 SRU 111 SRU 112

SRU 31 LRU 31

SRU 311 LRU 311

SRU 3111 SRU 3112

Fig. 1. Example of a system tree structure.

Nodes that do not represent physical components with an extreme situation with respect to geographical dis-
replaceable at the operation sites are not considered in tribution of the assets. It recommends a maintenance struc-
LORA. In the above example, LRU 2 and 4 could be ture in three echelons: Operational (in the same locations
linked directly to the system AIRCRAFT A1. This means where the systems are deployed); Intermediate (geographi-
that the LORA model discussed in this paper deals only cally close to places where the technical systems operate)
with the maintenance of components that can be removed and Depot (specialized and centralized maintenance, per-
from the system. When an organization owns various sim- haps thousands of miles away from the operation sites).
ilar systems, it is frequent to find common components Usually the first two echelons are organic, whereas third
among them. Thus, a particular LRU or SRU may be part parties typically operate the Depot maintenance level.
of the structure of more than one system. Ideally, a LORA model should be integrated with the
It is assumed that it is possible to associate the failure of overall problem of selecting the ‘‘best’’ support system
a system with a specific LRU. When this occurs, that LRU for the technical systems owned by an organization, con-
is replaced by a spare. The defective LRU, removed from sidering adequate effectiveness and efficiency measures
the system, generates a demand for maintenance actions and a set of possible support system alternatives. This
by the support system. This demand originates at the places should be the main objective of product support logistics,
were the technical systems are deployed. There are many as defined by Barros et al. (2001). Usually, system availabil-
ways in which a LRU may fail. These are called failure ity is the effectiveness measure and life cycle cost (or total
modes. The first decision in LORA is between LRU repair ownership cost) is the efficiency measure. The solution of
and disposal, depending on the failure mode. In case of this broader problem — maximize availability subject to
repair, LORA also prescribes to which maintenance facility cost constraints or minimize costs subject to availability
this LRU should be sent for service, also depending on the constraints — would also include non-LORA issues like
failure mode. If the failure mode occurs in a SRU con- the determination of the spare inventory levels at each sup-
tained in the LRU, the repair means the substitution of port and operational sites and could be further complicated
the defective SRU. The faulty SRU may be repaired or dis- by considering the stochastic nature of some inputs, like
carded. Many types of resources are required to perform failure and maintenance times. Alfredsson (1997) has pro-
each of the possible maintenance actions, namely: test posed a model to address this overall problem, using mar-
instruments, tools, spares, technical documentation, sup- ginal analysis, but acknowledged that he had to make
port software, special facilities, consumables and special- many simplifying assumptions in order to circumvent con-
ized human resources. Some repairs may be performed vexity problems. He has also assumed that the maintenance
on premises belonging to third parties, in which case it is structure is taken as pre-defined and must consist of two
not necessary to install resources, but the unit cost of repair echelons; one central site, supporting various regional lev-
is usually higher. els which, in turn, support a number of operational sites.
The American Department of Defense, where the Sleptchenko et al. (2003) have addressed the same problem.
LORA was initially formulated (MIL-STD-1390 D), deals They have used a modification of the VARI-METRIC
E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486 481

procedure (Sherbrooke, 1992), which also relies on mar- j costs frj and each installed unit provides the capacity
ginal analysis, to obtain near-optimal solutions and had Crj. The maximum number of units of the resource r that
to make several approximations to avoid problems with can be installed in the site j is Mrj.
non-convex constraints in the model. Define xtkaj as a continuous variable that indicates the
Barros and Riley (2001) have proposed an integer pro- percentage of the demand t, generated in k, that should
gramming model for LORA, as well as branch and bound be attended by means of the execution of the service action
algorithms to find exact solutions. Nevertheless, Barros & a in site j. Define yrj as a integer variable representing the
Riley have assumed a number of simplifying hypotheses, number of units of the resource r to be installed in site j.
resulting in a formulation that does not fit very well in real The model is presented below.
situations. For instance, the issues of the geographical dis- XXXX XX
Min d tk ctkaj xtkaj þ frj y rj ð1Þ
tribution of the operation sites or the capacity of the main- t2T j2J r2R j2J
k2K a2A
tenance facilities are not considered. In addition, it is XX
assumed that, if a facility is capable of performing the s:t: xtkaj ¼ 1 8t 2 T ; 8k 2 K; ð2Þ
a2A j2J
repair of a component belonging to a certain indentation
level, then it will be capable to repair all other components xtkaj 6 y rj 8t 2 T ; 8k 2 K; 8a 2 A; 8j 2 J ;
belonging to that same level. Gutin et al. (2005) proved 8r 2 Rjurta > 0; ð3Þ
recently that this model may be solved in polynomial time, XXX
using an algorithm for network flow over a bipartite graph. d kt urta xtkaj 6 C rj y rj 8r 2 R; 8j 2 J ; ð4Þ
t2T k2K a2A
This is an indication of the limitations of the Barros &
Riley model, because it shows that even one of the most 0 6 xtkaj 6 1 8t 2 T ; 8k 2 K; 8a 2 A; 8j 2 J ;
basic facility location problems, namely the Simple Plant ð5Þ
Location Problem, which is NP-hard, cannot be repre-
0 6 y rj 6 M rj ; y rj 2 Z 8r 2 R; 8j 2 J : ð6Þ
sented in that model. Another integer programming model
for LORA, also less general than the one proposed in this The objective function is given by the sum of the costs to
paper, has been presented by Bousseta (1995). execute the service actions, which are proportional to the
There are commercial software packages available for demand, and are called variable costs, and the costs to
LORA. Examples in North America and Europe are, install the resources at the service sites, which are called
respectively, Opus 10 (Systecon AB, 2007) and EDCAS fixed costs (although they may also depend on the demand
(Tools for Decision Group, 2007). It is not possible to in the capacitated problem). Restrictions (2) guarantee that
determine exactly what models and resolution techniques all generated demand is met by service actions. Restrictions
are implemented. Probably, the combinatorial part of the (3) condition the execution of each possible service action,
problem, concerning the location of maintenance facilities in each site, to the installation of at least one unit of the
and resources, is treated by heuristics. required resources. Restrictions (4) control the resources
capacities consumption. The following should be noted
2. A generic location problem model regarding the above formulation:

This section presents an integer programming formula- • The case where there is only one kind of resource, one
tion for the discrete location of facilities and installation kind of service action and all Mrj = 1, is the traditional
of resources with limited capacities problem, broader than formulation for the Capacitated Facility Location.
the ones already published in the literature, according to • The case where there is only one type of resource and
the recent surveys of ReVelle and Eiselt (2005), Klose one kind of service action, but there are sites with
and Drexl (2005), ReVelle et al. (2008) and works of Ghi- Mrj > 1, is the problem treated by Ghiani et al. (2002).
ani et al. (2002) and Wu et al. (2006). The next section • Restrictions (3) already follow logically from the others,
describes the LORA problem and show how it can be but are known to be very useful to improve the formu-
reduced to this more general problem. lation linear relaxation.
Define K as the set of demand generating sites and T as a • The model supports situations where some resources
set of tasks, or specific demand types. The data dtk is the may be considered with unlimited capacity, by just
number of tasks t 2 T, demanded at site k 2 K. Define J choosing Crj sufficiently high.
as a set of sites potentially capable to accomplish tasks. • Depending on the situation, it may be desirable to force
A task may be accomplished by the execution of one of the x variables to be binary too. This variant will be
the possible service actions defined in set A. The data ctkaj called Uniform Service Action per Origin and per Task
is the cost to accomplish task t, generated at site k, per- Policy (USAOTP). This policy requires that all demand
forming the service action a 2 A, in the site j 2 J. The exe- of the kind t, originated at site k, be always serviced at
cution of a service action requires resources belonging to a the same site j, with the same kind of service action a.
set R. The data urta is the capacity of the resource r 2 R, • Depending on the situation it may be desirable to
used to accomplish task t, by the execution of the service choose a unique service action for each t 2 T. This does
action a. The installation of one unit of resource r in site not imply that demand generated at different sites k
482 E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486

must be attended at the same site j. This variant will be where the systems are deployed, there are conditions (such
called Uniform Service Action per Task Policy as human, material and information resources) that make
(USATP) and requires the addition of new variables it possible to identify the LRU with the lowest indentation
xta, t 2 T, a 2 A, and of the following restrictions: level where the failure has occurred and to remove it. Nev-
X ertheless, we assume that there are no technical conditions
xta ¼ 18t 2 T ð7Þ to repair LRU or to substitute and repair SRU in these
a2A
X sites.
xtkaj 6 xta 8t 2 T ; 8a 2 A; 8k 2 K; ð8Þ Let T be the set of all failure modes in S. There is a univ-
j2J
ocal relation between T and Q so that each failure mode t is
0 6 xta 6 1; xta 2 Z 8t 2 T ; 8a 2 A: ð9Þ associated with only one q 2 Q. Let T(q)  T be the set of
failure modes t, associated with the component (LRU or
• Another possibility is to choose a unique service site for SRU) q 2 Q. Let kt be the failure rate for the failure mode t.
each t 2 T. This variant will be called Uniform Service Each occurrence of failure mode t requires an answer
Site per Task Policy (USSTP) and requires the addition from the support system. The immediate answer is the sub-
of new variables xtj, t 2 T, j 2 J, and of the following stitution of the LRU with the lowest indentation level, in
restrictions: which the failure has occurred. That is, the LRU q such
X that t 2 T(q), or such that t 2 T(q2) and q2 2 L(q). The next
xtj ¼ 1 8t 2 T ; ð10Þ step is to dispatch the defective LRU to a site where there
j2J
X are facilities and resources required to perform the neces-
xtkaj 6 xtj 8t 2 T ; 8j 2 J ; 8k 2 K; ð11Þ sary maintenance. It is assumed that failure modes do
a2A not occur simultaneously. There are four possible mainte-
0 6 xtj 6 1; xtj 2 Z 8t 2 T ; 8j 2 J : ð12Þ nance actions to be considered:

In the following section it will be shown that the LORA • aD = discard the LRU q;
problem, considering its variants, may be formulated using • aC = repair the LRU q if the failure mode t has occurred
the above model. To represent the situation where some in the LRU q and not in one of its SRU;
decisions are not valid, such as where it is not possible to • ad = repair the LRU q by substituting the defective SRU
execute a certain service action a in a specific site j, it q2, followed by the disposal of SRU q2;
will be necessary to force some variable groups xtkaj to be • ac = repair the LRU q by substituting the defective SRU
zero or to be excluded from the formulation. Similarly, q2, followed by the repair of SRU q2.
some restrictions (3) and (4) may be removed to represent
the fact that some service sites j already have unlimited These maintenance actions require the use of different
resources. kinds of resources belonging to the set R (set of required
resources). Each maintenance action a 2 A, in response to
3. The LORA model failure mode t 2 T, uses the quantity urta of the resource
r 2 R. Naturally, when a given action for a failure mode
Let S be the set of all systems to be supported by the does not need a specific resource, the correspondent quan-
maintenance structure object of the analysis. Let Q(s) be tity will be zero.
the set of all repairable components belonging to s 2 S. The broad productive system, including both technical
The unit cost of component q 2 Q(s) is c(q). These compo- and support systems, may be distributed along various geo-
nents are organized according to a tree structure, using the graphical places g 2 G (set of geographical places covered
criteria contain–contained. Let Q1(s)  Q(s) be the set of by the productive process). These places may receive facil-
all LRU belonging to s 2 S and Q2(s)  Q(s) be the set of ities where the technical systems will be installed and oper-
all SRU belonging to s 2 S. The quantity of LRU q 2 Q1(s) ated or its components may be repaired or discarded, upon
in the system s 2 S is n(q, s). Q, Q1 and Q2 are, respectively, failure. The facilities for repair or disposal may be organic
the sets of all repairable items, LRU and SRU existing in or owned by third parties.
the set S. The subset K of G represents the places where technical
It is assumed that in a system structure a LRU may con- system instances will be installed and operated. The other
tain other LRU and SRU, while a SRU may contain only sites belonging to G will be able to host facilities for repair
other SRU. Given LRU q 2 Q1, let L(q)  Q2 be the set of and disposal. They are represented by the set J. In order to
all SRU contained in q, and not contained in another LRU represent the situation where a facility for repair and dis-
also belonging to q. The quantity of the SRU q2 2 L(q) posal is located in the same place of an operation facility,
contained in LRU q is equal to n(q2, q). sites belonging to subsets K and J may have the same geo-
When the system fails, this failure will be due to a LRU graphical coordinates.
q. This failure may have occurred in SRU q2 2 L(q), Subset J comprehends sites where organic or third par-
belonging to LRU q, or in one part of this LRU that does ties maintenance facilities may be located. Organic mainte-
not belong to any SRU. It is assumed that in the sites nance facilities may be located on sites subset JO  J and
E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486 483

will require the availability of resources to perform the X


n
i1
repair and disposal maintenance actions. Outsourced third F ðn; dÞ ¼ 1=ð1 þ dÞ : ð15Þ
i¼1
party maintenance facilities may be located on sites subset
JT  J. By hypotheses, these sites already have all the nec- Therefore, the average unit cost of the maintenance
essary resources to perform the maintenance actions they action a, related to failure mode t, in LRU q, such that
are requested. The subset JD  JO comprehends all sites t 2 T(q) or t 2 T(q2) and q2 2 L(q), generated at site k
where only organic maintenance facilities capable of exe- and executed in site j, for a period of n years, discounted
cuting exclusively LRU disposal maintenance actions can at an annual rate of d, is shown in Table 1.
be located. Meeting the demands at the possible maintenance sites
The quantity of units of the technical system s 2 S imply the existence of resources in these places. The costs
installed in site k 2 K is represented by n(k, s). These sys- of these resources are not directly proportional to the
tems will operate according to different regimes, such as demand therein serviced.
continuous operation (24 h/7 d), 8 h per business day, Some of these are fixed costs, which do not depend at all
1000 h per annum, etc. Thus, each operation regime defines (within reasonable limits) on the demand size, such as civil
an annual quantity of operation hours h(s, k). The demand construction cost of facilities and fixed maintenance costs.
for maintenance actions of type t, to be performed in LRU Since these costs depend on site j, these resources r have
q of system s, generated at site k, will be a function of the been modeled with fixed cost Frj, unlimited capacity Crj,
parameters n(k, s), h(k, s), n(q, s) and kt associated with and maximum quantity per site Mrj = 1. When these
components q 2 Q1(s) and of the parameter n(q, q2) associ- resources are required by some maintenance action, the
ated with components q2 2 Q2(s)jq2 2 L(q). The annual parameter urta is arbitrarily set to 1.
average demand for maintenance actions generated at site Another type of cost, also proportional to the demand
k for the failure mode t, in LRU q, such that t 2 T(q2) size, but in a different way (because it depends only on
and q2 2 L(q), or such that t 2 T(q), are given, respectively, lumped amounts of the demand), corresponds to resources
by (13) e (14): with limited capacity, which must be installed in discrete
X
d tk ¼ nðs; kÞ:hðs; kÞ:nðq; sÞ:nðq2 ; qÞ:kt quantities. As an example, suppose that a circuit-board test-
8s2S ing equipment may be used in a certain place (due to avail-
8k 2 K; 8tjt 2 T ðq2 Þ and q2 2 LðqÞ; ð13Þ ability or man-power limitations) for 3000 h per annum. If
X the total demand requires the use of 8000 h of this resource,
d tk ¼ nðs; kÞ:hðs; kÞ:nðq; sÞ:kt it will be necessary to buy 3 units. In these cases, these
8s2S
resources may be modeled with unit cost Frj, a limited
8k 2 K; 8tjt 2 T ðqÞ: ð14Þ capacity Crj, and maximum quantity per site Mrj P 1. The
capacity of resource r, used to perform maintenance action
Supposing that adequate resources to perform mainte- a, in response to failure mode t, is given by urta.
nance actions in a certain place exist, meeting the demand The installation of resources will be considered only in
in this location will imply a variable cost proportional to sites j 2 JO (those able to receive organic facilities capable
the size of this demand. This cost will be composed of of performing maintenance actions). By hypothesis, sites
two parts: one is the cost to transport the failed LRU j 2 JT, where there are third parties able to execute mainte-
between the operation and maintenance sites and the other, nance actions have all the required resources. It is usual
the direct costs to perform the maintenance actions (LRU that unit costs to perform maintenance actions by third
disposal, LRU repair, or LRU repair with SRU disposal or parties are higher.The generic LORA model may be formu-
repair). lated as: Minimize (1), Subject to (2)–(6), where dtk is given
The transportation cost, in general, depends on the loca- by (13) and (14), ctkaj is given by (16)-(19) and Frj, Crj, Mrj,
tion of the two sites involved, the transportation modal and
on some characteristics of the LRU to be transported
(among which weight, volume, value and dangerousness Table 1
Maintenance actions costs
are the most important). The cost of the two-way transpor-
tation of LRU q between k and j sites is ctqkj. Action Description Average discounted unit cost
(ctkaj)
The unit cost of a repair action on LRU q, such that
t 2 T(q), is a function of the failure mode t and of the site ac LRU repair, with SRU F ðn; dÞ:ðctqkj þ cstj þ crtj Þ ð16Þ
substitution and repair
j where the maintenance will be executed. This cost is rep-
resented by crtj. The unit cost to substitute the SRU q2 such ad LRU repair, with SRU F ðn; dÞ:ðctqkj þ cstj þ cðq2 ÞÞ ð17Þ
that t 2 T(q2) and q2 2 L(q) is cstj. Finally, the unit cost to substitution and disposal
dispose of a LRU or a SRU is equal to the unit cost of the
discarded component c(q). aC LRU repair F ðn; dÞ:ðctqkj þ crtj Þ ð18Þ
All unit costs considered are at present value, using an
annual discount rate d, for n years of system operation. aD LRU disposal F ðn; dÞ:ððctqkj =2Þ þ cðqÞÞ ð19Þ
The discount factor F(n, d) is given by:
484 E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486

urta are defined hereinabove. In addition, the variables yrj Table 2


and the restrictions (3) and (4) corresponding to places in Problem sizes
JT, are omitted. Since ac, ad and aC type actions cannot Problem jSj jKj jJOj jJTj jQ1j jQ2j jTj jRj
be executed in sites j 2 JD, the corresponding xtkaj variables 1 1 15 4 2 10 14 36 9
must be eliminated. Restrictions (4), corresponding to 2 1 15 8 2 10 14 36 9
resources r with unlimited capacity Crj, must also be 3 2 16 8 2 16 32 74 9
4 3 16 8 2 26 45 113 9
eliminated. 5 4 16 10 3 32 55 126 9
For the Uniform Service Action per Task Policy 6 5 20 15 3 53 100 206 9
(USATP) option, variables and restrictions (7)–(9) must 7 1 23 12 3 53 100 206 9
be added; for the Uniform Service Site per Task Policy 8 6 23 15 3 56 106 215 14
(USSTP) option, variables and restrictions (10)–(12) must 9 6 23 12 3 56 106 215 14
10 7 23 6 1 59 116 256 14
be added; and for the Uniform Service Action per Origin 11 8 23 10 1 66 132 296 14
and per Task Policy (USAOTP), the restriction that all 12 8 23 12 3 66 132 296 20
variables xtkaj must be binary must be added. 13 9 23 12 3 70 140 310 20
14 9 23 3 3 70 140 310 20
4. Computational experiments 15 2 23 20 6 14 22 58 20

The proposed models were tested on 15 problems gener-


ated to be statistically similar to those found in certain Bra- For each such problem we selected a combination of
zilian military and civilian groups of systems sharing the maintenance policies (USATP, USSTP, USAOTP, or free
same maintenance structure. Location decisions are essen- policy) and generated the corresponding LORA MIP mod-
tial when performing LORA for these systems, since they els. The lifecycle and the annual discount rate used were,
are spread over a large continental area, including remote respectively, 15 years and 6%. The resulting 22 integer pro-
operation sites in the Amazon jungle. There is always the gramming instances were run in a machine with Pentium
question of whether is it better to minimize transportation IV 3.2 GHz processor and 1 GB of RAM, using CPLEX
costs by opening maintenance facilities closer to the opera- 9.1 as the mixed-integer programming solver. The results
tion sites or to minimize repair costs by opening only a few obtained are shown in Table 3.
facilities, typically close to industrial centers. There is also Columns CONs, VARs, and NON ZEROS indicate the
the need of considering possible savings by outsourcing number of constraints, variables and non-zero coefficients
part of the maintenance, in which case potential outsour- in each MIP instance after CPLEX preprocessing. Column
ced maintenance sites may be located in North America, ROOT TIME is the total time spent in root node, before
Europe or Asia. The sizes of the test problems are shown starting the tree enumeration. This time includes the solu-
in Table 2. tion of the linear relaxation of the model, the addition of

Table 3
Test results
Problem Policy CONs VARs NON ROOT TIME ROOT B&C TOTAL TIMES FINAL
ZEROS (second) GAP (%) NODES (seconds) GAP (%)
1 FREE 18861 2869 57724 5 0.03 107 8 0
2 USAOTP 43748 8468 147728 10 0.04 804 105 0
3 USATP 31740 4851 97359 12 0.05 171 29 0
4 USSTP 101613 17930 316066 109 0.01 3 118 0
5 USATP + USSTP 157084 28081 492512 456 0.02 306 4286 0
6 USATP + USAOTP + USSTP 824998 154180 2814300 3476 0.00 1 3476 0
6 FREE 384582 52189 1193194 1243 0.04 321 2428 0
7 USATP 215712 36229 674990 907 0.01 3 959 0
7 USAOTP 764385 141280 2583225 678 0.05 1088 10000 0.02
8 USAOTP 800316 150079 2717221 1190 0.03 2109 10000 0.02
9 USAOTP 434996 84952 1512508 382 0.03 10550 10000 0.01
10 USAOTP 69553 358164 1205880 415 0.02 2000 3602 0
11 USAOTP 605136 118388 2038324 890 0.02 2400 6934 0
12 USAOTP 644382 128775 2191754 1685 0.03 2000 9585 0
13 USAOTP 817717 163948 2787644 1270 0.04 1748 10000 0.02
14 USAOTP 210612 47109 732018 180 0.00 1 180 0
14 USAOTP + USSTP 210658 47133 732271 112 0.00 1 112 0
14 USAOTP + USATP 242610 59475 833697 514 0.00 1 514 0
14 USATP + USSTP 277207 59545 868708 652 0.00 1 652 0
14 USAOPT + USATP + USSTP 242656 59545 834157 614 0.00 1 614 0
14 FREE 62979 11602 189599 12 0.00 1 12 0
15 USAOPT 202399 46683 711627 556 0.04 9500 8102 0
E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486 485

cuts to strengthen this relaxation and the application of pri- maintenance site is potentially able to service any failure
mal heuristics to find good solutions. Column ROOT GAP mode, using any maintenance action. Consequently, the
gives the percentage difference between the best solution number of possible alternatives is larger and this will
found while still at the root and the lower bound given lead naturally to cheaper solutions.
by the strengthened linear relaxation. Column B&C 7. It can be adapted to handle situations where failure rates
NODES is the size of the enumeration tree used by CPLEX vary with time. For systems with mechanical compo-
to achieve optimality. Finally, TOTAL TIME is the total nents, operating for long periods of time, the aging effect
time spent. We set a time limit of 10,000 seconds for each may increase the failure rate, with the result that the
run. If a proven optimal could not be found within that demand will also increase with time. Therefore, it may
time, column FINAL GAP contains the percentage differ- be desirable to delay the installation of a portion of
ence between the best solution ever found and the current the resources, according to the demand growth. It is
lower bound. When the instance can be solved to optimal- possible to adapt the model to this situation, at the cost
ity within the time limit the final gap is zero. of an increase in the number of variables and con-
We remark that, although solving those instances to straints. The solution would be to divide the life cycle
proven optimality may require hours to explore several time in P periods, each one with its failure rates. The
hundreds of nodes in the branch-and-cut tree, the CPLEX decision variables x and y would have a new index
solver always found solutions guaranteed to be very close p 2 P. Resources quantities may then vary with the time
to the optimal in a few minutes, right after solving the first periods, depending on the demand generated during
linear relaxation. In all cases those solutions had a gap of each one.
less than 0.10%, good enough for most practical purposes.
Nevertheless, the proposed model still has some limita-
5. Conclusions tions that may be considered in future works:

This paper proposed a MIP model for the discrete loca- • It does not take explicitly into consideration that mainte-
tion of facilities and installation of capacitated resources nance decisions affect spares stock levels required to guar-
and showed its applicability to LORA problems. Compu- antee the desired system availability. For instance, it is
tational experiments have shown that quite large problems possible that the repair of a LRU by its third part man-
can be solved in reasonable times, attesting that the overall ufacturer may take several months, while the same
approach is practical. We believe that the proposed model repair by an organic installation would take only a few
is more general and realistic than others published previ- days. In this case, the decision to repair at the manufac-
ously. Its use to solve LORA problems presents the follow- turer premises will imply in the increase of the number
ing desirable characteristics: of spare items to be stocked, in order to keep the same
system availability level. This additional cost is not con-
1. The input data can be obtained with the use of well sidered explicitly in the proposed model. One possibility
known and widely used logistic support analysis meth- would be to estimate those costs using an auxiliary
ods, such as Maintenance Task Analysis (MTA) and model to capture the effect of the repair time of each ser-
Failure Mode, Effect and Criticality Analysis (FMECA). vice site in the spare stock levels. This solution has been
2. The input data is very detailed, thus adaptable to many used by Blanchard and Fabrycky (1998). It implies that
real situations. For instance: (i) transportation costs repair times must be defined in advance, before running
may be defined for each component and for each route, the LORA model. However, this does not take into
(ii) variable maintenance costs may be dependent on account the fact that repair times may depend signifi-
each failure mode, maintenance action and maintenance cantly on decisions made by the LORA model. For
site, (iii) the costs for the installation of resources in dif- example, if the total site demand is close to its capacity,
ferent sites may be different. there will be a queue of components waiting for repair,
3. There is no limitation on the quantity of resources types. leading to a repair time larger than assumed.
4. It is a true capacitated model. Resources capacities can • It does not consider the stochastic aspects of LORA. The
be set for each maintenance site. This feature is very use- model relies on average failure rates and does not con-
ful, for instance, to model situations where different sup- sider how failures are actually randomly distributed on
port sites may work different amounts of hours per day. time. Similar considerations may be done with respect
5. There are several maintenance service policies to choose to estimating costs by averages. If those (and many
from. other possible) stochastic aspects are considered to be
6. For historical or practical reasons many support systems critical in a certain application, one would like to have
have been organized in multi-echelon structures. This a model leading to a robust solution over a large set
kind of structure is arbitrary and constitutes, in fact, a of scenarios.
simplification. When a multi-echelon structure is used,
an ‘‘a priori’’ choice is made about the arrangement of Finally, as mentioned in Section 1, it is desirable to
the support sites, whereas in the proposed model any integrate a LORA model within the overall problem of
486 E.S. Brick, E. Uchoa / European Journal of Operational Research 192 (2009) 479–486

selecting the ‘‘best’’ support system for the technical Bousseta, A., 1995. Mathematical tools in cost effectiveness optimization
systems owned by an organization, considering availabil- process. In: Proceedings of the International Conference on Industrial
Logistics (ICIL-95), Ouro Preto (Brazil).
ity as the effectiveness measure and life cycle cost as the Gutin, G., Rafiey, A., Yeo, A., Tso, M., 2005. Level of repair analysis and
efficiency measure. This is a very complex and difficult minimum cost homomorphisms of graphs. Lecture Notes in Computer
problem indeed, as can be attested by the difficulties Science 3521, 427–439.
pointed out by Alfredsson (1997) and Sleptchenko Ghiani, G., Guerriero, F., Musmanno, R., 2002. The capacitated plant
et al. (2003). The development of models to solve the full location problem with multiple facilities in the same site. Computers
and Operations Research 29, 1903–1912.
problem of life cycle cost minimization, subject to Klose, A., Drexl, A., 2005. Facility location models for distribution system
availability restrictions, and over a large set of forecasted design. European Journal of Operational Research 162, 4–29.
scenarios, may be considered one of the greatest chal- MIL-STD-1390 D. Military Standard: Level of Repair Analysis, US Navy.
lenges for researchers of the field in the forthcoming ReVelle, C.S., Eiselt, H.A., 2005. Location analysis: A synthesis and
years. survey. European Journal of Operational Research 165, 1–19.
ReVelle, C.S., Eiselt, H.A., Daskin, M.S., 2008. A bibliography for some
fundamental problem categories in discrete location science. European
References Journal of Operational Research 184, 817–848.
Sherbrooke, C.C., 1992. Optimal Inventory Modeling of Systems: Multi-
Alfredsson, P., 1997. Optimization of multi-echelon repairable item Echelon Techniques. John Wiley& Sons, New York.
inventory systems with simultaneous location of repair facilities. Sleptchenko, A., van der Heijden, M.C., van Harten, A., 2003. Trade-off
European Journal of Operational Research 99, 584–595. between inventory and repair capacity in spare part networks. Journal
Barros, L., Riley, M., 2001. A combinatorial approach to level of of the Operations Research Society 54, 263–272.
repair analysis. European Journal of Operational Research 129, Systecon AB. Introducing OPUS 10. Stockolm, <http://www.systecon.se/
242–251. documents/Introducing_OPUS10.pdf> (accessed in May, 2007).
Barros, L., Riley, M., Brown, D., 2001. Special millennium issue of EJOR: Tools for Decision Group, 2001. EDCAS.<http://www.tfdg.com/Prod-
A global view of industrial logistics. European Journal of Operational ucts/EDCAS/> (accessed in May, 2007).
Research 129, 231–234. Wu, L.Y., Zhang, X.S., Zhang, J.L., 2006. Capacitated facility location
Blanchard, B., Fabrycky, W., 1998. Systems Engineering and Analysis. problem with general setup cost. Computers and Operations Research
Prentice-Hall, Upper Saddle River, NJ. 33, 1226–1241.

You might also like