International Journal of Production Economics 208 (2019) 329–342

Contents lists available at ScienceDirect

International Journal of Production Economics

journal homepage: www.elsevier.com/locate/ijpe

A simulation-based optimization approach evaluating maintenance and T

spare parts demand interaction effects
James M. Wakirua,∗, Liliane Pintelona, Peter N. Muchirib, Peter K. Chemwenoc
a
Centre of Industrial Management/Traffic and Infrastructure, KU Leuven, Celestijnenlaan 300, 3001 Heverlee, Belgium
b
Department of Mechanical Engineering, Dedan Kimathi University of Technology, Nyeri, Kenya
c
Department of Design, Production and Management, University of Twente, Drienerlohaan 5, 7522 NB Enschede, Netherlands

ABSTRACT

Industrial facilities frequently experience significant production losses due to unanticipated failures, sub-optimal maintenance, operational and spare parts logistics
challenges. These among other factors directly affect the plant's performance measures such as availability, repair time and costs. Consequently, optimization
addresses such challenges. However, a fundamental problem presented here relates to the need for a framework that assists in the determination of critical system to
be optimized, variables that significantly impact the performance of such systems, and subsequently undertake optimization. To realistically model such com-
plexities, a framework that applies the discrete simulation model of critical repairable subsystems, undergoing deterioration is proposed. The study utilises empirical
maintenance data, where Pareto analysis is employed to identify critical subsystems, while expert input is incorporated to derive model variables. A full factorial
Design of Experiment (DOE), is employed to establish the variables with significant main and interaction effects on the total repair time and subsequently employed
as decision variables for a simulation-based optimization. The proposed framework is demonstrated in a case study of a thermal power plant. Simulation results
highlight the turbocharger as the critical subsystem, while spares availability, the time between overhaul (TBO) and reliance on different maintenance strategies
exhibit most significant main and interaction effects. The optimization results obtained demonstrate that TBO, spares availability and reliance on various main-
tenance strategies, provide a significant impact on the reduction of the repair time. The framework enhances maintenance decision making by optimizing the plants'
operational and maintenance related factors identified.

1. Introduction carried out by Jardine and Tsang (2013) reports that maintenance
budgets on average are 20.8% of the total plant operating budget. In-
1.1. Background tegrating such measures would ultimately drive the facility towards
achieving the expected plant economics due to savings in maintenance
It is essential industrial facilities such as power plants can supply and operational costs. Optimal maintenance leads to maximized effi-
electric power to the population consistently without interruptions. ciency and productivity as well as reduced waste in both equipment and
This requirement is amplified when a plant supporting critical in- personnel usage, which leads to the significance of joint optimization
stallations, such as in the health and security sectors, where an inter- modelling of maintenance, inventory, and manpower also suggested by
ruption, could lead to significant risks along with both social and eco- Van Horenbeek et al. (2013). Distinct from maintenance strategies
nomic losses. Moreover, most facilities supplying utility grids maintain optimization, operational and other aspects such as inventory man-
contracts specifying hefty penalties in the event of deviation from the agement, manpower planning, procedures and outsourced main-
supply agreement, which makes it imperative for such power plants to tenance, likewise require to be optimized to ensure successful plant
address factors likely to cause interruption of power production and operations. While addressing these aspects, consideration of main-
supply. Besides natural disasters and short-supply of consumables like tenance quality is vital, because it is influenced by various aspects,
fuel, downtime retains a pivotal role in contributing to power produc- while it impacts the equipment reliability and consequently reduced
tion interruption. A growing body of literature recognizes the sig- productivity.
nificant role of maintenance-related downtime, contributing to power
plant downtime and subsequently, reduced profitability (Alabdulkarim
et al., 2011; Bouslah et al., 2018). It follows that addressing main-
1.2. Notations
tenance-related downtime will considerably reduce power generation
interruptions and Operational and Maintenance (O&M) cost. A survey
Throughout this paper, the following notations are adopted.

∗
Corresponding author.
E-mail address: [email protected] (J.M. Wakiru).

https://doi.org/10.1016/j.ijpe.2018.12.014
Received 12 April 2018; Received in revised form 11 October 2018; Accepted 11 December 2018
Available online 13 December 2018
0925-5273/ © 2018 Elsevier B.V. All rights reserved.
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

results, with a brief evaluation, whereas a summary discussion is pre-

Decision variables
sented in Section 5. Finally, Section 6 contains concluding remarks and
Ao Engine operational availability
Tr Total repair time (hrs. per year) the proposed future work.
Model parameters
i Number of maintenance actions 2. Relevant literature review
n Subsystems modelled n = {1,2,3,4,5}
FSj Failure severity for jth severity level The British Standards Institute defines maintenance as, “a combi-
n Random time to next failure for nth subsystem nation of all technical, administrative and managerial actions during
Tdi Random diagnosis time for ith maintenance action (hrs.) the life cycle of an item destined to preserve it in or restore it to, a state
Tdn Total diagnosis time for n th subsystem (hrs.)
in which it can fulfil the required function” (EN13306, 2010). This
Tln Total lead-time for nth subsystem (hrs.)
Oc Number of Outsourced contractors
definition enlarges the sphere of maintenance, to cover various tasks
t Simulation time (hrs.) encompassed in Corrective Maintenance (CM) and Preventive Main-
tstn Initial time to failure for nth subsystem tenance (PM) that incorporate administrative and managerial actions,
j Severity levels as per ISO 14224:2016; j = {1,2,3,4} such as manpower management, spares sourcing and logistics. System
i Impact factor for ith maintenance action reliability, influenced by deterioration, and maintenance strategies
I Engine operating time/Running hours performed owing to equipment failure, are critical factors to the reali-
γ, β, η Location, scale and shape parameters for WEIB (λn)
zation of a plant's objectives, and overcoming various operational
f Spares availability or spares fill-rate
Ri Maintenance actions ( i ; i = {1,2,3,4,5})
challenges (Jardine and Tsang, 2013). Stochastic models such as
Trn Total repair time for nth subsystem (hrs.) Markov models (MM) have proven more robust, in modelling dete-
Tri Random repair time for ith maintenance action rioration and failure of equipment owing to their probabilistic nature
tbo Time between overhauls (TBO) (Nguyen et al., 2013). Semi-Markov Decision Process (SMDP), contrary
tn Simulation time (years) to MM which is memoryless, calls for the memory of the process to be
Mtech Mechanical technicians renewed each time it attains a state (Dehayem Nodem et al., 2011).
DT Downtime (hrs.)
Other Markov models, such as time-based Markov models are limited to
dealing with time as a deterioration factor. Both the prior severity and
the particular maintenance action carried out regarding its quality and
1.3. Study aim and motivation for the research harshness, can influence the deterioration of various equipment. An age
reduction coefficient can be used to represent the actual influence of
Power plant system reliability represents an essential factor in the maintenance and deterioration on the equipment, (e.g., Duan et al.,
success of a power generation project. A decline in reliability due to 2018), or a hazard rate adjustment factor (e.g., Xia et al., 2017). To
critical equipment downtime directly affects both, the availability to abate the impact of deterioration on equipment, inspections, CM, PM
generate power and consequently its profitability. Increased downtime and Overhaul actions are carried out (Bouslah et al., 2018; Rivera-
can be attributed to several reasons influencing O&M aspects such as Gómez et al., 2016). However, due to diverse factors, some organisa-
increased failure frequency, protracted lead times in sourcing critical tions outsource these maintenance actions, either, to focus on their
spares, imperfect maintenance, human errors in diagnosis and main- competencies, or due to technological advances or lack of the required
tenance, and lack of enough maintenance personnel. These factors de- competencies (de Almeida et al., 2015). In an assessment of PM con-
monstrate a significant impact on the installation's economics, tactical tract maintenance, Wu (2012) proposes two types of outsourcing,
and strategic planning. The factors above, both individually as well as where type I, both CM and PM are outsourced and type II, only CM is
interactively affect maintenance, where optimizing them would im- outsourced.
prove the availability of a given power plant and reduce repair and Another significant aspect of maintenance is the often utilization of
subsequent maintenance time, hence improving the economics and spares in both PM and CM actions hence the importance of joint or
performance of the plant. Moreover, in the plants that retain main- integrated maintenance and spare policies optimization. Integrated
tenance records, deriving maintenance optimization programs utilizing studies considering joint optimization of maintenance strategies and
knowledge embedded in this data is not straightforward. The primary spares/inventory policies, predominantly have focused on specific as-
challenge faced by many practitioners is the establishment and selec- pects, such as, evaluating optimal PM optimization (e.g., Lei et al.,
tion of significant optimization variables affecting equipment perfor- 2010), optimal production plan (e.g., Ba et al., 2016), optimal in-
mance, from the empirical data, and further implementing an optimi- ventory holding (e.g., Cai et al., 2017), optimal maintenance policy
zation of the variables. The challenge of undertaking maintenance (e.g., Nguyen et al., 2014). Nevertheless, some studies address diverse
optimization based on significant decision variables derived from em- aspects, for instance, Rezg et al. (2005), presented integrated modelling
pirical data, with the objective of offering maintenance decision sup- and optimization for a randomly failing production unit. Similarly, Lei
port, presents the motivation to this study. et al. (2010) presented a modularized Stochastic Timed Petri-Nets
This study proposes a discrete event simulation (DES) model fra- (STPNs) model, for system level PM optimization and dynamic main-
mework that considers the utilization of empirical maintenance data to tenance decision making. Gharbi and Kenné (2005) advanced simula-
derive critical subsystems, modelling parameters, and variables. We tion and experimental design to determine the optimal control policies
model a repairable system, whose deterioration is assumed to follow a and values of input factors while minimizing total cost of inventory,
Semi-Markov Decision Process (SMDP). Design of Experiment (DOE) repair, and PM. Zohrul Kabir and Farrash (1996), presented a simula-
approach is utilized to quantify the effects and interactions of the tion model seeking for a joint determination of optimal age replacement
variables on equipment availability and total repair time. Markedly, and spare part provisioning policy, while Sarker and Haque (2000),
optimization is performed applying the identified significant variables, developed a simulation model for a system operating with block re-
with an objective of minimizing the total repair time. DES is utilized placement and continuous review inventory policy. However, much of
due to its robustness in dealing with complex, highly variable and un- the research addressing PM (e.g., Roux et al., 2008; Van Horenbeek;
certain events as a decision support tool also corroborated by (Alrabghi Pintelon, 2013), disregard the effect of CM-related failure on the
and Tiwari, 2015). system, which undoubtedly impacts on the equipment reliability, hence
The remaining part of the paper proceeds as follows: Section 2 affects PM actions and outcome. Moreover, failures addressed using
presents a brief literature review of relevant studies, while Section 3 CM, are considered to incorporate spare replacement and following two
describes the methodology used for this study. Section 4 presents the states “as good as new” (AGAN) and “as bad as old” (ABAO), (e.g.,

330
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Dijoux et al., 2016; Rivera-Gómez et al., 2016), which does not offer a consider the interactive effects, where several statistical methods such
comprehensive representation of CM actions that may have repairs with as full factorial analysis and analysis of variance (ANOVA) offer the
or without spare replacement, with random reliability outcome. means to address this aspect. For instance, Bouslah et al. (2018) em-
Another significant challenge found in considerable studies is the ployed ANOVA, to illustrate the interactions of variables, while estab-
employment of one-factor-at-a-time (OFAT) analysis while analyzing lishing variables that were significant in an integrated production,
the impact of various factors on the plant performance measures as quality and maintenance control policy.
previously indicated. Subsequently, these results are applied in main- The first contribution of this paper is to fill the gap in the literature,
tenance decision support, while disregarding their interactive char- by developing an integrated framework, derived from empirical data,
acteristics. As Antony et al. (2003) alludes, despite its popularity, ease for the joint optimization of maintenance and spares policy.
of use and precision on the effect of one variable, OFAT offers less Specifically, we consider maintenance on different engine subsystems,
precise estimates, it does not consider interrelationships (i.e. interac- subject to imperfect maintenance and reliability degradation. The cri-
tions) and cannot be used in optimization, challenges overcome by the teria used to select the subsystems, represent the impact to the system
design of experiments (DOE) technique. Similarly, Bouslah et al. (2018) performance, in this case, downtime, an aspect corroborated by
suggest the interaction effects play an essential role to obtain an op- (Alrabghi and Tiwari, 2015). Moreover, most studies have dwelt on
timal trade-off solution while considering an integrated control policy. multiple systems, for example, several engines in a plant, whereas our
While considering previous optimization studies in maintenance, it is study deals with subsystems within one single engine. The distinctive
becoming exceptionally challenging to overlook the interactive effects effect of respective subsystems, such as failure and deterioration, have
of various O&M aspects, on the plant's performance. been overlooked in such models, which may lead to generalized and
As regards evaluating the plant's performance, several factors, for uncomprehensive optimization solutions. In this case, subsystems are
instance, repair time taken, availability of manpower resources and inter-linked to form a system, for instance, a power plant engine, has
logistical challenges involved in spare sourcing, and others, adversely the cylinder, governor, fuel and turbocharger subsystems, which have
affect the plant's availability (Nguyen et al., 2013). Due to the complex unique operational and failure characteristics.
dynamic characteristics of these variables and their interrelationship, Secondly, this paper considers the reliability degradation of distinct
the use of classical analytical models in optimization is unsustainable. subsystems, which typically require the prescription of clear CM stra-
Hence, the use of simulation techniques, an approach which is also tegies. This distinction may not be forthcoming due to imperfect
corroborated by various studies (Alrabghi and Tiwari, 2016; maintenance (imperfect diagnosis and repair) and alternative options
Nowakowski; Werbińka, 2009; Rezg et al., 2005), while, Sharma et al. available. Hence, the need to explore the value of several maintenance
(2011) accentuate that the use of simulation in maintenance optimi- strategies simultaneously, that can be prescribed to the failed sub-
zation has been increasing steadily. Several studies reviewed in this systems, which is seldom done in research while modelling imperfect
field employing simulation are highlighted in Table 1. maintenance of deteriorating subsystems (Lei et al., 2010).
Surprisingly, the lack of carefully examining the effects of CM is The third contribution of this paper aims to make the selection of
noted, where most studies consider minimal repair towards bringing a significant variables affecting the maintenance performance measures
component back to the previous operating state. This aspect is seen more judicious, by involving expert assessment and employing statis-
perhaps not a complete representation of both maintenance and spare tical techniques to quantify their effects and interactions. Notably, this
policies, that potentially address component reliability and degradation study will address the aspect of systematic selection of significant
challenges, both under PM and CM. Additionally, CM failures which are variables subsequently employed in a maintenance optimization pro-
addressed by various maintenance and spare strategies consequently gram. Key plant performance indicators, such as, spares availability,
demonstrate an impact on the PM strategy, due to fundamental and maintenance strategy reliance, and others derived from the empirical
observable reliability changes on the system impacted by the CM ac- data and agreed upon with the plant maintenance team, are subjected
tions. to this process.
Furthermore, the studies assumed perfect maintenance and con- Lastly, we employ the significant variables as decision variables in a
sidered only one CM maintenance action, i.e. replacement of the failed simulation-based optimization with an objective of minimizing the total
item. This aspect may not offer a realistic maintenance analysis, be- repair time. A case study approach is adopted to obtain in-depth in-
cause repair actions may not always justify spare replacement, which formation and capture the complexities brought by the different vari-
restricts the research to perfect maintenance. Despite the importance of ables, derived from plant's empirical data.
imperfect maintenance, which mimics real maintenance scenarios,
there remains a paucity of evidence on the application of aspects im-
3. Methodology
plying imperfectness such as imperfect maintenance, imperfect re-
placement, and failure diagnosis time, also alluded by Van Horenbeek
The methodology consists of several steps. Data collection, pre-
et al. (2013).
processing and exploration represent the first and second step respec-
While performing maintenance optimization, most of the studies
tively. The third step incorporates identification of the model output
lack a systematic framework for selecting variables that significantly
parameters, while key input variables for the simulation model are
affect the objective measure, which is subsequently considered as
extracted in the fourth step. The simulation model is developed and
control variables in the optimization program. Such frameworks should
subsequent experiments performed in the fifth step. The simulation

Table 1
Sample summary of simulation-based articles-joint maintenance and spares optimization.
Article Variables investigated/affecting performance

Li et al. (2013) Manpower planning

Alabdulkarim et al. (2011) Manpower availability, equipment usage, spares fill rate. Accurate fault diagnosis
Savsar (2015) Maintenance policies - age-based (ABP) and block based (BBP)
de Smidt-Destombes et al. (2007) Maintenance interval, the spare part inventory level and the repair capacity
Scarf and Cavalcante (2012), Maintenance quality (inspection induced defects, quality of PM and replaced component)
Duan et al. (2018) Maintenance quality
Bouslah et al. (2018) Maintenance, quality, and Production-inventory policies

331
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

experiment results are evaluated and interpreted in the sixth step. In Fig. 1 illustrates an example where ideally a subsystem retaining FS2 ,
this step, significant variables to be employed in the optimization are should be diagnosed to undergo R2 , but due to imperfect diagnosis,
determined. Finally, the seventh step addresses the implementation of a there is the possibility of undergoing one of the other Ri apart from R5
simulation-based optimization. (mandatory for all subsystems during overhaul). For the purpose of
analysis, the assumption is considered that due to imperfect main-
3.1. Data collection tenance, FS2 undergoing R2 , transitions to FS3, whereas, if perfect
maintenance was considered (not in this case), it could transition to FS1.
In this study, the analysis uses maintenance data describing failures Following this process, a specific FSj stochastically transitions to a
recorded from thermal power plant engines, over a five-year period. posterior state between low FS1 to high FS4 . The scheduled overhaul
Each equipment or engine (fuel oil-driven) consists of several inter- maintenance reduces the failure severity to FS1 with some probability
linked subsystems such as a governor, exhaust, cylinder and turbo- of FS2 which mimics a renewal state (near AGAN) with stochasticity.
charger. The data contains, subsystem failure occurrences, separate This process is modelled for all the FSj with possibilities of subjection to
maintenance actions carried out, and components details, such as the all the Ri .
date and time of both the failure occurrence and repair finalization. To The operating severities. FSj ; j = {1,2,3,4}
address inconsistencies, the International Standards Organization - ISO The finite maintenance actions. Ri ; i = {1,2,3,4,5}
14224 steps, were followed during pre-processing. In this step, various
components were linked to their respective subsystem, while expert 3.3. Model output parameters
consultations were employed where clarity was needed.
The model considers two performance measures of interest as the
3.2. Data exploration outputs. Firstly, the engine operational availability Ao , is the proportion
of time the engine is running compared to total time including down-
Following the maintenance data exploration, several aspects are times due to failures and overhaul (PM). We use the running hours I
considered towards enabling modelling of the study. against the total running length (includes running hours and down-
time). Secondly, total repair time Tr , is the total value of the repair time,
3.2.1. Critical engine subsystems selection accrued by the engine (five subsystems) on purely CM repair processes.
The power plant engine was decomposed into subsystems, from This time value excludes the spare sourcing lead-time Tl and diagnosis
which the four critical subsystems were selected based on failure fre- time Td delays incurred during the CM activities. The assumption of PM
quency and the individual contribution to the power lost in megawatts- actions having pre-planning therefore, no lead times involved.
hours (MWhr) using Pareto analysis. The fifth critical subsystem item- Equations (3)–(5) provide an annualized average time value based on tn
ized as ‘others’ incorporated summation of all the remaining sub- , which is more intuitive compared to the planning horizon of the plant.
systems. I
Ao =
I + DT (1)
3.2.2. Maintenance actions – Ri and subsystem state modelling
Five maintenance actions employed in data exploration were clas- Tm = Td + Tr + Tl (2)
sified following the ISO 14224:2016 classification and categorised on n
T
estimated time incurred by the action as corroborated by the main- Tr = k=1 rk

tenance team. See Table 2. tn (3)

The state of each subsystem is depicted by two variables, a discrete n
Tdk
stochastic state FSj which take the values {1,2,3,4} denoting the op- Td = k=1
tn (4)
erational state, in terms of severity at time t, and impact factor i
(hazard rate adjustment factor) which takes a value between 0 and 1, n
T
k = 1 lk
and controls the changes of n at time t. In this case, we assume that the Tl =
tn (5)
stochastic state of the subsystem is dependent on the stochastic prior
state FSj and the maintenance action Ri carried out as discussed in the
next section. 3.4. Model parameter extraction

3.2.3. SMDP modelling of failure severity states Several model parameters derived for the critical subsystem's ana-
Failure severity FSj , is the impact due to Ri on a subsystem, de- lysis include; the time to the initial failure generation tstn , which re-
picting the seriousness or harshness of a failure. This was used to de- present the time the first failure of a respective nth subsystem occurred.
termine the reliability measure attained by a subsystem following a The reliance factor/utilization probability i , for the different Ri , was
respective Ri and subsequently used for diagnosis whilst selecting ap- computed from the subsystem failure frequency, while the mean time to
propriate Ri by the maintenance team. While employing SMDP, FS1 is repair Tri , was derived from the maintenance action time classifications
depicted as causing minor production stop, FS2 , moderate, FS3 , severe using a uniform distribution (due to the classification entailing
and FS4 , an extensive stop in production/operation, as adopted from minimum and maximum time). A probability distribution was fitted for
(ISO 14224:2016, 2016). Low FSj is assumed to have low Tri , while high the better random forecast of maintenance interventions that utilized
FSj to have high Tri due to the intense maintenance action Ri required. more than 13 h, to represent R 4 . The tbo and Tr5 , that characterize R5,

Table 2
Maintenance actions classification based on ISO14224:2016.
Ri Maintenance action Description

i=1 Do almost nothing Incorporate minor adjustments, may not cause a stoppage, e.g. tightening
i=2 Minor Incorporate spare replacement bringing the subsystem back to an operational state
i=3 Moderate Both moderate and major repair actions incorporate logistical lead times during spares sourcing and repair actions
i=4 Major
i=5 Overhaul PM action for whole engine overhaul as per the maintenance schedule

332
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Fig. 1. Sample illustration of FS2 modelling using SMDP.

were derived from the PM planning manual of the engine. The main- the subsystem.
tenance team identified Tdn , the estimated time a failed subsystem will 4. The availability of maintenance technicians is not a constraint. This
incur while being diagnosed to indicate the type of FSj , hence identify is because the expert analysis and interviews revealed that the
the appropriate intervention or Ri . The time to next failure n , for each technicians were shared from a pool, hence their availability was
subsystem, fitted to a probability distribution, was established by not viewed as a constraint.
computing the time between repairing the subsystem to operable state 5. PM incorporates a modularized design for the subsystem's main-
up until its next failure. Spares availability f, also referred to as fill-rate tenance characterized by spare replacement. Thus, PM is assumed
or instantaneous reliability of spares (Louit et al., 2011), was provided not to incur repair time.
by the plant supply chain, as the estimated probability of the plant
holding spares stocks on hand to deal with the maintenance require- 3.6. Analysis, evaluation, and interpretation
ment. Sourcing lead-time Tl for both local and imported spares, was
derived similarly from the plant supply chain. In the instances requiring This section encompasses two parts wherein the first part, considers
spare sourcing, the maintenance data included information on both the model results following a one-factor-at-a-time (OFAT) analysis,
actual repair and spare sourcing lead times amalgamated, without while in the second part, a set DOE, following a full 2k factorial design
differentiating the two. Hence, to derive the actual repair time Tr in the (k as the number of variables/controls to evaluate their effect to the
model, we subtracted the estimated spare sourcing lead time Tl from the engine Ao and Tr ) is conducted and computations of the controls' (here
combined actual repair time plus spare parts sourcing lead times de- also termed as variables') average effects and interactions generated.
rived from empirical data. Therefore, due to the derivation and mod- The process will follow the 3-factor full factorial experiment with re-
elling of this aspect, it is expected that Tl will have an impact on Tr solution V, as described by NIST/SEMATECH (2012). This resolution or
values. An outsourced contractor Oc , is partially involved in the PM ability to separate main effects and low-order interactions from one
(specialised activities), is constrained to working solely during the another, has the capacity to estimate all main effects and two-way in-
daytime shift. teractions (Sanchez et al., 2006). The significant variables selected
during this step meet three conditions; firstly they have both main ef-
3.5. Modelling fects and the corresponding interaction effects, an aspect corroborated
by (Mccullagh, 2002). Secondly, they are statistically significant at a
A discrete event simulation modelling framework is developed significant level of 0.5. Finally, using expert assessment, they are con-
which mimics the system normal running I until failure occurrence and sidered significant based on the size and other techniques such as
subsequent CM repair action undertaken, while PM is carried out after a graphical methods.
time interval tbo . A i ranging from 0 to 1, was introduced for esti-
mating the subsystem hazard rate adjustment factor (impact of the
3.7. Simulation-based optimization
maintenance action on the n ) in each Ri . The extreme values ρ = 0
depict ABAN, while ρ = 1 construe AGAN. SMDP was employed to
This section outlines the 7-step simulation-based optimization ap-
model the stochastic deterioration process of the subsystems, based on
proach based on the conceptual framework proposed by Alrabghi and
the specific maintenance actions Ri , where initial FSj values were ran-
Tiwari (2016), further adapted for the optimization phase discussed in
domly assigned to the subsystems and posterior FSj generated as dis-
this study, as illustrated in Fig. 2.
cussed in Section 3.2.3. During the modelling step, the following as-
The first step considers defining the scope of the optimization,
sumptions are made:
which addresses the establishment of the maintenance policies to be
employed in the process. The second step involves identifying the
1. Maintenance actions are imperfectly performed, hence, the equip-
maintenance policies and the respective decision variables to be in-
ment state does not attain “as good as new” (AGAN) state after
corporated in the optimization, while the third step, constitutes the
maintenance intervention.
formulation of the objective function. Depending on the outcome of the
2. We consider that two or more subsystems cannot fail simulta-
preceding step, the fourth step formulates the constraints. The selection
neously, because, a single subsystem failure, causes the engine
of the optimization algorithm and setting up the required algorithm
stoppage.
parameters is performed in the fifth and sixth steps respectively.
3. Concerning the principle of FSj transition, we consider only the
Finally, the optimization results are evaluated and interpreted while
imperfect maintenance action and the stochastic failure severity of
considering the plant/asset's current context. To sum up, the 7-step

Fig. 2. Schematic framework for simulation-based optimization approach.

333
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Table 3 Table 4
Sample data standardisation using ISO 14224. Various subsystem time to next failure distributions.
Equipment class Equipment type Sub-units/Subsystems Subsystem tstn (Hrs) n distribution Parameters Corresponding p-value

Description Code Description Code Description Code X2 Test K-S Test

Combustion engines CE Diesel engine DE Cylinder CYL Governor 2200 EXPO (2.82e+003) 0.341 > 0.15
Turbocharger TC Turbocharger 1998 16 + WEIB (3.55e+003, 0.55 > 0.15
Fuel pumps FS 0.854)
Lubrication system LO Others 640 2 + EXPO (723) 0.316 > 0.15
Exhaust EXH Lubrication 1260 13 + EXPO (1.28e+003) < 0.005 0.0753
Cylinder 1800 40 + EXPO (582) 0.477 > 0.15

process is scalable and will offer insights when followed step by step.
4.3. Model parameter extraction

4. Results Table 4 summarises the tstn and n for each of the critical subsystems
selected, where the tstn was computed based on the assumption of the
This section presents the results, brief evaluation and discussion, analysis commencement of January 2011. It indicates the governor has
while a wide-ranging discussion will be dealt with more extensively in the highest tstn inferring its less susceptibility to failure during the early
Section 5. running hours, after commissioning compared to other subsystems. The
Weibull distribution estimate represented as WEIB (α, β, γ) with the
shape parameter β and scalar parameter α and exponential distribution
4.1. Data collection and pre-processing (EXPO) has the mean i.e. EXPO (mean). The Weibull and Exponential
distributions were characterized with a third parameter distribution (γ),
The maintenance data from the power plant, recorded in a free text also known as location parameter or failure free time, which indicates
structure was standardized to meet the analysis requirement following that failures start at a finite time and not at t = 0, for instance, turbo-
the ISO 14224:2016. The classification entailed the equipment class and charger failures (γ = 16). The governor, lubrication system, cylinder
types Combustion engine (CE) and Diesel engine (DE) respectively, and others fit an exponential distribution with a random mean. This
while subsystem classification as illustrated in Table 3. signifies failure occurrences that are independent of each other and
The data classification enabled the data to be categorised using the randomly distributed and could be attributed to high replacement
various subsystems utilized for the data exploration discussed in the strategy hence tending to near constant or steady state as also corro-
next section. borated by (Louit et al., 2011).
The turbocharger subsystem fitted a Weibull distribution, exhibiting
a shape parameter β < 1, which indicates that the failure rate de-
4.2. Data exploration
creases over time. The components have their hazard rate decreasing
due to less severing strategies for instance replacement that have a
Fig. 3 illustrates a Pareto chart prioritising the engine subsystems in
lower impact on the n . This is contrary to an intensive regenerative
the plant using the individual contribution to the power lost in mega-
strategy which has a high negative impact on the n due to the strategy
watt-hours (MWHrs). What stands out in the chart is the first four
characteristics, where the component renewal is near ABAO, hence,
subsystems, i.e. cylinder, governor, turbocharger and lubrication
lowering n depicts a shorter life experience for the subsystem.
system, which cumulatively contribute 86% of total power lost hence
Table 5 provides the different CM actions alongside extracted
were selected as critical subsystems to be modelled. In this study, the
parameters like the probability of utilization i , repair time classes Tri
four critical subsystems are modelled with an extra one “others” which
and Tdi . Tri followed the Ri classification as discussed in Section 3.2.
is a summation of the remaining subsystems. As regards to maintenance
Mean time to preventive maintenance (overhaul)Tr5 for R5 had a uni-
decision support context, explicit strategic focus on the four critical
form distribution of minimum 192 h and maximum 224 h as depicted
subsystems would potentially improve and enhance the engine perfor-
from the preventive maintenance schedule. The impact factor i , acts
mance.
as a multiplier to the time to the next failure n of an individual sub-
system. The i , was derived from proportionating the respective total
n following the approach utilized by Wakiru et al. (2018), which
compares and aligns with major repair 4 as 0.80 under the assumption
that attaining AGAN status is difficult in deteriorating systems, due to
the introduction of errors such as diagnosis, tooling and human related
errors. This is based on an empirical analysis of the estimated percen-
tage changes in n of the various Ri , where 0.8 was seen as optimal for
R 4 . Respective Tdi for the various CM actions were estimated by the
plant maintenance team.
Table 6 presents the various parameters used while addressing the

Table 5
Repair time for various repair actions.
Repair Action Tdi (hrs.) Tri (hrs.) i (%) i ISO 14224:2016
classification
Do Nothing 0.15 0–1.0 16% 1 = 0.65 Minor
Minor Repair 0.50 1.0–7.0 57% 2 = 0.75 Moderate
Moderate Repair 0.7 7.0–13.0 15% 3 = 0.85 Severe
Major Repair 2.0 Over 13.0 12% 4 = 0.80 Catastrophic
Fig. 3. Pareto analysis for subsystem/components using Power lost.

334
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Table 6
Spares availability and sourcing lead times.
Maintenance action Spares needed (%) f (%) Spares sourcing

Local (%) Lead-time (Hrs) Import (%) Lead-time (Hrs)

Moderate repair 80.19 90 10 1–5 90 36–120

Major repair 100.00 90 5 4–24 95 120–1080

spare inventory and logistics. The percentage requirement of local and Hence, an objective and realistic analysis show the CM related repairs
import sourcing, as well as the respective lead times, were provided by
4
incur more time (Tr + Tl ) and more activity i = 1 Ri , which is logical and
the experts from the plant maintenance and supply chain departments. corroborated by various studies, for instance, (Alrabghi and Tiwari,
2015). Further review demonstrates an inconsistent variation of Ao
indicating the possible effect of other variables. However, a general
4.4. Model
increase is noted from 82.79% to 87.56%. A similar analysis indicates
that increasing f improves Ao , while Tr increases and Tl decreases re-
The developed model mimics the operation of the engine, where the
spectively. To explicitly understand the level of dependency between
performance measurements were Ao and Tr . Fig. 4 illustrates the con-
the variables while impacting the performance measures, a full factorial
ceptual representation of the model.
orthogonal design experiment is employed as discussed in the following
The model was run with a replication length of 105,120 h equiva-
section.
lent of 12 years, reflecting the age and the first planning horizon based
on the planned useful life. The industry average design engine lifespan
is estimated to be 20–25years. Hence, this correlates to the lifetime of
4.5.2. Full factorial effects and interactions experiment results
similar engines. For each simulation, 55 replications were made which
A 3-factor complete factorial design experiment was carried out, to
allowed a large sample size to be considered, hence reducing the 95%
determine significant variables. The respective variable ranges illu-
confidence interval half width of the performance measures Tr (from
strated in Table 9, were prepared after expert interviews and according
±45.00 to ±5.00) There was no significant variation in the percentage
to the procedure (save for tbo) used by (Sanchez et al., 2006; Zhu et al.,
change of the half width beyond 55 replications.
2017). An analysis of the effects and the respective p-value for different
ranges was done in addition to expert consultation while assessing the
4.5. Analysis of results appropriateness of the ranges as seen in Table 11. The ranges were
computed using the model parameters derived as described in Section
4.5.1. Model results 4.3. We utilize the ranges adopted in both the full factorial experiment
The model generated Ao of 90.001%, and Tr of 1526 h per year re- (in this Section 4.5.2) and the simulation-based optimization in Section
spectively. The achieved Ao and Tr using the model were close to the 4.6.
empirical values as analysed from the data of 92% and 1621 h respec- The criteria employed to identify the significant variables, con-
tively. This may be attributable to the stochasticity in the modelling sidered, first, that the variable has both, the main effect and corre-
parameters. However, the generated Ao is very close to the empirical sponding interaction effect, secondly, the P-Value confirms the variable
value of 92%. Thus, we can assume that our model is a valid re- (s) is statistically significant at the significance level of 0.05. Also, to
presentation of the modelled processes in the power plant. Further visualise the error size of an effect, graphical methods were used, while,
evaluating the individual contribution to Tr , the cylinder, turbocharger plots of residuals were used to check for model adequacy. The simu-
and governor subsystems, each contributed 23.2%, 19.9%, 18.1% re- lation was carried out using 55 replications, and results depict average
spectively of the total as depicted in Table 7. values, using resolution V to estimate the average effects and interac-
It should be noted; the values indicated are average values over the tions. Table 10 provides a sample of generated average main and in-
simulation time tn . The consolidative characteristic of “other” sub- teraction effects on the Ao and Tr from the designed experiment, while a
systems as seen in Table 7 is attributed for the high Tr and Tl , hence not detailed list is shown in Table A1, in the appendix.
incorporated in the analysis for individual subsystems in this section. As Table 10 indicates, an increase in tbo , will averagely increase the
The turbocharger incurs the highest Tm compared to the other critical Ao by 5.87% while decrease theTr by 612.325 h. The improvement of Ao
subsystems. Both the turbocharger and governor subsystem indicate is because of compensated repair time by lengthened running or op-
high Tl , implying substantial failures characterized by spares sourcing eration time I due to a decline of PM activities as also seen in Section
requirement. This could offer a pointer towards, the need of a more 4.5.1 and Table 8. The availability of spares potentially reduces the
plausible strategy, that could incorporate spares inventory, reuse, re- subsystem downtime due to spares sourcing lead times Tl which posi-
condition, or cannibalization strategies, if possible, to overcome the tively impact I, hence improve the Ao . An increase in 2 will averagely
challenge. Cylinder subsystem has high Td , possibly attributed to the increase Ao (11.99%) and increase Tr (712.62 h). This indicates, that
complexity of the assembly, that often require disassembly to diagnose increased reliance of R2 , offers a negative impact on Tr attributable to
internal components failure in the subsystem. the fact that R2 incorporates solely repair without spares requirement,
It can be seen from Table 8, obtained by carrying out an OFAT hence the modelled Tr retains a higher value.
analysis, that increasing the tbo gradually decreases Tr in a non-linear Considering interaction effects, for instance, the interaction of tbo
manner. As a result of extended tbo , the frequency of R5 utilization and f causes a significant change in Ao and a modest decrease of Tr by a
(renewal instances) declines and more failures occur, requiring CM factor of 12.52% and 92.07 h respectively. This infers that the effect on
interventions, which are represented in the last column of Table 8. f by tbo positively affects the performance of the engine Ao and Tr .
Additionally, this aspect is confirmed by the marked increase in the Similar interaction to some extent is seen between 2 and 3 with a
utilization of various Ri under CM, for instance, R2 (818–991), R3 modest improvement on Tr and a modest change on Ao . This implies
(221–264) and R 4 (152–187). The observed increase in Tl , can therefore that the tbo effect on Ao does not depend on the effect of f , signifying
be related to the increased utilization of both R3 (moderate repair) and that tbo has a negative effect at high f but a positive effect at low f .
R 4 (major repair), which subsequently demand replacement of spares While evaluating the impact on Tr , the effect of tbo , to some extent de-
that incur high lead-times, an aspect demonstrated by the increase of Tl . pends on the effect of f , where f has a negative effect at high tbo but a

335
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Fig. 4. Conceptual representation of a discrete event simulation model.

positive effect at low tbo . The main and interaction effects are presented variables. A sample of the evaluation is shown in Table 11.
in Table A1. To evaluate the variable ranges employed in this exercise In summary, it can be concluded from the DOE performed, that the
along with expert interviews, we employed three sets of ranges and variables show a marked interaction impact on Tr compared to Ao ,
confirmed they gave same results in terms of significance of the hence Tr is selected as the performance measurement to be employed in

336
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Table 7 Table 9
Summary of subsystem time variables. Variables ranges used in the optimisations.
Subsystem Trn Tln Tdn Tmn Variable Values Ranges

Turbo charger 303.38 128.58 9.99 441.95 tbo 9000 7000 12000
Governor 275.89 122.40 8.46 406.75 2 57 45.60 68.40
Cylinder 354.24 10.28 20.10 384.62 3 15 12.00 18.00
Lubrication 254.00 13.04 7.76 274.80 Td1 0.15 0.07 0.23
Others 338.61 520.64 34.30 893.55 Td2 0.5 0.25 0.75
Td3 0.7 0.35 1.05
Td 4 2 1.00 3.00
the optimization. Subsequently, the significant variables established f 90 72 99
include, tbo , 2 , 3, and f. Due to the interactive characteristics of the Oc 1 1 2
significant variables, an optimization considering them jointly and
several constraints, is advanced employing a simulation-based optimi-
zation as discussed in the next section. 4. Constraints definition

Two constraints are defined, the first considers 2 and 3 , as illustrated

4.6. Simulation-based optimization
in Table 5 in Section 4.3. Both constitute a combined 72% (57% and 15%
respectively), hence we formulate the constraint as 2 + 3 = 72%.
The simulation-based optimization framework described in Section
Secondly, to ensure the optimization considered a better scenario than the
3.7 is followed here step by step:
base, we define Ao to be more than the base value of 90%, as depicted in
Section 4.5.1, hence, the problem can be formulated as follows:
1. Defining the scope of the optimization
n
T
k=1 rk
The simulation-based optimization addresses the subsystems iden- Minimize: Tr =
tn
tified and modelled as per Section 4.3. In this case study, the optimi-
zation scope will include the total repair time and fill rate which ad- Subject to:
dresses the spares policy.
Ao 90%
2. Applicable maintenance policies and decision variables for optimi-
+ = 72%
zation 2 3

The decision variables established as significant in Section 4.5.2 are 5. Optimization algorithm
considered for the maintenance optimization. The variables include the
time between overhaul tbo addressing PM policy, fill rate f that deals We employ OptQuest for the optimization problem at hand, which
with the inventory policy. Other variables addressing the CM policy in- offers the possibilities to optimize simulation experiments employing
clude the maintenance strategy reliance factors 2 and 3. However, Oc , neural network filter and apply heuristics known as tabu search and
the number of outsourced contractors will be included despite not being scatter search (Kelton et al., 2010). Moreover, it offers both the option
in the DOE performed in Section 4.5.2, because we view it as an essential of setting up goals for explicit handling of bounds enforced on output
aspect, that could influence downtime caused by PM strategy and sub- values and the graphical output presentation that enhances inter-
sequently influence CM policy in terms of reliability. Table 12 presents pretation. Furthermore, it is employed, to gain insights, while con-
the summary of the decision variables (possible and revised choices), and sidering the consistency the variable bounds, as previously employed in
the possible number of solutions in both cases. The simulation optimi- Section 4.5.2.
zation process utilises the revised choices to reduce the solution space.
6. Simulation optimization set-up
3. Objective function formulation
The optimization using OptQuest was done with varying the
The critical policies to consider based on the variables selected in number of replications from 5 to 55 for each simulation, allowing
Step 2 include the repair times, maintenance strategy reliance and spare OptQuest to test for the statistical significance between the average
parts availability. The effect of spares availability is incorporated in Tr , objective function of our best previous and current simulation, whose
since it introduces lead time and affects the repair time due to multiple intention is to rule out inferior solutions. The optimization process was
handling. In the same way, the maintenance strategy reliance effect only stopped and results retrieved when no change is noted on the
would be noticed in generating Tr . Hence, minimizing the total repair optimal solution after 200 simulations. The Tr baseline before the op-
time will be the only objective. timization is 1526.1 h as seen in Section 4.5.1.

Table 8
Annualized times and frequency of utilization for Ri while varying tbo .
tbo (hrs.) A o (%) Annualized time over tn (hrs.) The frequency of utilization over tn

Tr Tl Td Tr + Tl R1 R2 R3 R4 R5 4
i = 1 Ri

7000 82.79 1763.51 415.23 83.39 2178.74 236 818 221 152 18 1427
8000 91.60 1735.11 429.22 84.44 2164.33 280 824 225 154 17 1483
9000 90.01 1526.13 685.23 80.88 2211.36 256 833 232 165 15 1486
10000 87.00 1376.3 929.83 79.51 2306.13 230 857 246 165 13 1498
11000 87.60 1187.11 1222.13 70.57 2409.24 193 883 254 175 11 1505
12000 87.56 1175.89 1442.13 76.69 2618.02 205 991 264 187 11 1647

337
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Table 10
A sample of significant main and interaction effects.
tbo f 2 3 tbo + f 2 + 3 tbo + 2 f+ 2 tbo + 3

Δ Ao (%) 5.87 19.4 11.99 5.59 1.52 1.5 −1.95 −0.66 −1.89
Prob > |t| < 0.0001 < 0.0001 < 0.0001 0.0004 < 0.0001 0.0446 0.0225 0.0134 0.0267
Δ Tr (hrs.) −612.33 548.14 716.62 −209.83 −92.07 −42.38 −94.96 −379.9 78.58
Prob > |t| < 0.0001 < 0.0001 < 0.0001 0.0001 0.01850 0.0038 0.0160 < 0.0001 < 0.0001

Table 11
Evaluation of different variable ranges scenarios.
Variable 50%:100%:150% 80%:100%:120% 90%; 100%:110% Significant

Effect p-value Effect p-value Effect p-value

f 1464.18 0.0000 548.14 0.0000 718.52 0.0000 Yes

3 708.7 0.0007 −209.83 0.0001 88.79 0.0000 Yes
tbo f −421.5 0.0000 −92.07 0.0185 −68.72 0.0003 Yes
f 3 −282.18 0.0002 −112.27 0.0350 −31.9 0.0244 Yes
Td2 70.10 0.8840 160.22 0.6092 26.26 0.7426 No
f Td2 −25.20 0.9581 −18.20 0.9534 −2.52 0.9748 No

7. Results and discussion as also corroborated by Alabdulkarim et al. (2011). Furthermore, due
consideration should be taken on probable risks that accompany en-
Fig. 5 illustrates the optimization results, which indicates the best hanced f such as increased risk due to spare costs, insurance costs,
value being attained at the 186th simulation and no further change pilferages and storage space requirement.
seen.
The optimization resulted in more than 30 solutions where the Tr is
more than 1% below the base line value. Table 13 presents the top ten 5. Discussion
optimal solutions. From these results, we can see that optimality is at-
tained while enhancing PM, spares availability, while, a modest change The initial model analysis results are significant in at least two
in the reliance of minor and moderate repair strategies. Scenario I significant respects. In the first place, they guide the practitioner while
generated the highest Ao and power generated in MWHrs as production selecting the critical subsystem, where in our case, the turbocharger is
output for the engine, while scenario A, returned the lowest Tr . picked based on total maintenance time. In the second place, they offer
The values of 2 and 3 in the optimized scenarios, indicate more insights on the bottlenecks suffered by the plant using various main-
reliance on minor repair actions ( 2 ). This was expected, because of the tenance policies. Among the subsystems' identified through the ap-
inherent characteristics of the minor repair actions, which exhibit lower proach as critical regarding portended time characteristics, such as
repair times and no spare sourcing lead time. Oc in this case depicts a repair time, lead-time to procure spare parts and diagnosis includes the
higher value of 2, which can be attributed to the fact that the out- turbocharger, cylinder and governor. Hence, such subsystems necessi-
sourced capacity is only available during the daytime shift, hence the tate further investigations with a view of identifying robust main-
increase by 1 to ensure availability of the resource. tenance strategies, which seeks to optimize operational availability. The
As far as reviewing the optimization results is concerned, there is proposed simulation is demonstrated as useful for decision support as
strong concurrence that enhancement of tbo , f , leads to optimization of observed from the results illustrated in Section 4. As an example, as
the plant's performance measures. An implication of this is, firstly, the described in Section 4.5.1 (Table 7), the turbocharger and governor,
possibility of re-evaluating the tbo to optimize the planned maintenance exhibit lengthier lead-time delays based on the modelled stochastic
effectively, based on the lifecycle of the subsystems. Secondly, spares spare part sourcing lead times. An implication of this is the need to
inventory challenge if overcome, shows a potential positive optimiza- investigate further the specific spare part inventory that frequently
tion solution, an outcome also corroborated by Kennedy et al. (2002). requires to be sourced, either locally or imported from the original
In this case, an introduction of different inventory policies such as equipment manufacturer (OEM). In real-life, determining the stocking
consignment stocking, employing other recovery actions such as re- policy is not straightforward in the absence of a decision support fra-
conditioning and adoption of industrial symbiosis, would probably mework, owing to the stochasticity of time characteristics like repair,
ensure spares are available throughout and could positively impact on diagnosis, and sourcing lead times. Moreover, the interaction effects
the optimization. A caution here would involve carrying out a cost- between the stochastic spare part availability and other variables, in-
benefit analysis to balance the cost and improvement on either Ao or Tr fluence the stocking strategy, yet modelling this aspect and evaluating
its influence on system availability is somewhat challenging. However,

Table 12
Variable ranges and the possible number of solutions.
Variable Ranges Possible choices Revised choices Remarks

tbo 7000 12,000 5000 5000 No change

2 45 68 230 23 Changed from step 0.1 to step 1.0
3 12 18 60 6 Changed from step 0.1 to step 1.0
f 72 99 27 27 No change
Oc 1 2 2 2 No change
Possible solutions 3,726,000,000 37,260,000

338
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Fig. 5. Graphical representation of the optimization results.

from the simulation model, it is observed that it is possible to derive plausible framework. It lays emphasis that the analysts must explore
insights on such interaction effects through the simulation modelling factor effects concurrently to understand how their simulation model
approach. For instance, the turbocharger and governor which exhibit behaves when its factors are changed. Despite the use of p-values, in-
longer sourcing delays could benefit through initiatives like prioritised formation about the size of an effect and its possible error must be al-
stocking, Just-in-Time (JIT) or consignment stocking from the OEM/ lowed to interact with expert knowledge. Graphical methods, ad-
local agent. ditionally provide a valuable means of allowing information in the data
Moreover, for the above, utilization of condition monitoring tech- and the mind of the expert to interact appropriately, an aspect also
niques such as oil analysis or vibration analysis on such subsystems, corroborated by (Kleijnen et al., 2005).
potentially can provide early warnings that assist the plant in advanced Taken collectively, these results suggest it is essential to optimize
planning of spare parts sourcing, a strategy similarly suggested by variables jointly since the decision variables or controls can interact
Eruguz et al. (2018). While addressing the diagnosis delay and accuracy with each other and yield a sub-optimal solution. For instance, as seen
for critical subsystems such as the cylinder and turbocharger, the plant in the results in Table 10, variable such as 2 and f negatively affected
could re-train the technicians and bring in better diagnostic technolo- Tr individually, based on their individual main effects on Tr . However,
gies, which could enhance better detection of failure severities as this when their interactions were considered, their combined beneficial
aspect in most cases is influenced by technician knowledge and ex- influence on Tr was positive. As an example, the combined effects of 2
perience, an observation also corroborated by Wang (2012). This im- and f generated a decrease in Tr despite both having negative effects
provement would also yield better repair strategies, for instance, a individually. Hence, this suggests that performing optimization based
component replacement for high severity failures. on a single variable (or the OFAT approach) could eventually lead to
Importantly, modelling stochastic failure diagnosis Tdi times, despite sub-optimal maintenance optimization, and by extension, considerable
its insignificant impact on Tr in this case, illustrate the significance of modelling time owing to repetitive OFAT experiments. This observation
this seldom utilized maintenance downtime constituent, as a modelling corroborates to a greater extent the findings by (Sarker and Haque,
variable which influences the quality of the maintenance action, an 2000) who suggested that this type of model using interactions is ap-
aspect also corroborated by Van Horenbeek et al. (2013). Further in- plicable in maintenance strategies of multiple components in max-
vestigations concerning Td would lead to unearthing aspects such as the imizing service levels, similarly considered in this study.
suitability of a technician skills, their response time, maintenance With regard to the optimization results in Section 4.6, the results
quality, availability of tools, and could further examine automation of indicate that the total repair time would be optimized by considering
decisions as concerns the use of sensor or historical data, aspects also firstly, the CM related factors (i.e., the maintenance strategies to rely on
corroborated by Bousdekis et al. (2015). This aspect could be addressed and the spares availability), and secondly the PM-related factors (TBO
by employing appropriate fault diagnostic systems, for example uti- and Outsourced maintenance). Concerning the reliance on different
lizing thresholds such as RUL and hazard rates to trigger decisions of maintenance strategies, the results show that a slight increase in minor
appropriate interventions under CBM, an area the authors view for repair with a small decrease in reliance on moderate repair actions
future work. would contribute to an optimal solution. This result may be explained
While reviewing the second phase where main and interaction ef- by the fact that minor repair retains a lower repair time strategy with no
fects are determined, the results doubtlessly, despite dependent on the spare requirement. As for the spares availability, the results are as ex-
case study characteristics, has some reliable conclusions. The sys- pected, where to reduce repair time, spares availability should be
tematic determination of significant variables, employing various guaranteed. However, this aspect requires further investigation con-
techniques, based on the main and interaction effects demonstrates a sidering the operational context of different plants, where a trade-off

Table 13
Optimization parameters while minimizing Tr .
Scenario A B C D E F G H I J

Tr 1432.30 1432.74 1437.65 1437.70 1438.48 1440.06 1440.44 1441.84 1442.37 1444.35
tbo 12000 12000 12000 12000 11976 11976 12000 12000 12000 12000
2 58 58 58 58 58 58 58 58 58 58
3 14 14 14 14 14 14 14 14 14 14
f 99 99 99 99 99 99 99 99 99 99
Oc 2 2 2 2 2 2 2 2 2 2
Ao 91.21 91.01 91.43 91.29 91.42 91.32 91.14 91.47 91.44 91.23
Power Generated 87135 86940 87340 87203 87336 87232 87061 87382 87348 87149

339
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

may be required between stocks level, holding costs, pilferage and for the simulation model. Expected differences would emanate while
spares shelf-life. Concerning TBO, the results indicate that for this case, modelling due to the differences in the operational context. For in-
the increased interval would offer lower repair time. The observed stance, some plant set-ups constitute redundant or buffer subsystems to
decrease in Tr could be attributed to the modelling approach of the time guarantee continuous operations, an aspect that should be considered
variants Tr , Td and Tl as indicated in Section 4.5.1. Moreover, the (Rezg et al., 2005).
possible interference of protracted Tl , thereby reducing respective Trn
cannot be ruled out. However, an introduction of other performance
measure such as total maintenance cost or time that amalgamate the 6. Conclusion
time incurred (Td, Tr and Tl ), will undoubtedly cause this observation to
change. The present study was designed to develop a methodology that
From the optimization results, contract or outsourced maintenance commences with empirical data, to analyse the effects and interactions
during equipment overhaul as currently implemented at the power of various maintenance and operational factors and their impact on the
plant negatively influenced the total repair time for the power plant. optimization of the engine total repair time using a simulation model.
This influence was due to limitations such as availability timelines of The case study of a thermal power plant was advanced with con-
the outsourced resource (often only during one shift), or the type of sideration of repairable subsystems with both PM and CM actions. The
outsourced maintenance service. For the latter, two types of main- subsystem deterioration problem was formulated stochastically in-
tenance services are outsourced, wherein the type I, both corrective and tegrating the previous subsystem severity influenced by maintenance
preventive maintenance services are outsourced, and in type II, only, actions. Further to identifying the turbocharger as critical among the
CM is outsourced. The availability limitation is observed as negatively subsystems using total maintenance and repair time, the study has
impacting the overhaul time since repairs would otherwise have been shown that the different subsystems have different repair, diagnosis and
performed during both the day and night shifts. Secondly, outsourcing lead time's characteristics offering different impacts to their life cycle
only PM services limits the maintenance repair processes, since the times. Spares availability, reliance on minor and moderate repair stra-
outsourced resource are deprived of valuable insights which could be tegies and TBO were demonstrated to have the most potent effect on
derived from performing CM actions. As an example, root causes of total repair time. The analysis results while evaluating the model,
chronic subsystem failures which are correctively repaired could be conceivably support the hypothesis that interactions of the variables
missed during PM actions, yet such root causes could be appropriately play a role in influencing the performance measures. The optimization
addressed during the CM actions by the outsourced resource, who are exercise which openly supports the relevance of interactions, indicates
more specialised in maintaining the critical subsystems. This finding is that tbo , spares availability and maintenance strategies reliance as es-
also demonstrated by Wu (2012), where he suggests that attention sential ingredients. These findings have a significant implication in
should be given while selecting the type of outsourced resource for understanding parameters that require further investigation while car-
plant maintenance. He argues that for better maintenance processes, rying out maintenance decision making, whereby, if enhanced, would
organisations should outsource type I services (CM and PM). Further, greatly improve the maintenance strategies, resource allocations and
the ageing process of the subsystem could be a factor to consider during ensure priorities are set right to improve the availability of the engine
outsourcing, where a balance between in-house and outsourced main- and eventually the plant economics. This combination of findings pro-
tenance should be considered. As an example, ageing components may vides some support for the conceptual premise that while carrying out
require more outsourced maintenance services. This observation cor- maintenance optimization, a balance of decision variables used, need to
roborates findings by Bazargan (2016), who recommended a combi- be struck by considering their effects, interactions and expert knowl-
nation of in-house and outsourced in such circumstances. edge. The research lays a groundwork for future studies into other
The proposed methodology is generalizable and can be scaled to fit maintenance strategies with their possible interactions towards an in-
different applications that constitute multiple subsystems. Table 14 il- depth optimization model. Moreover, the simulation-based experiments
lustrates several potential applications, indicating the respective sector, and optimization on real power plant data verify the validity and ro-
system, and possible subsystems. It is essential to bear in mind that the bustness of the developed technique.
application of this proposed methodology to various applications, for Several limitations of the present study should be acknowledged,
instance, the manufacturing sector would require some adaptations to which should be dealt with in detail in future studies. For example,
suit the operational context. Firstly, the identification of the system to modelling shared resources like technicians amongst several systems,
be investigated, for instance, the cement mill/grinder in a cement plant. which could offer decision support concerning the same. Another lim-
Secondly, for the identified system, a precise segmentation of the sub- itation of the study is based on the fact that it considered only PM and
systems constituting it (indicated in the “sample subsystems” column in CM policies, whereas additional maintenance and restorative strategies
Table 14). The third aspect will involve the adoption of the metho- identified like condition monitoring, spares reconditioning, and reuse
dology as highlighted in Section 3, where maintenance data is collected, of spares could be incorporated. An optimization extension utilizing the
preprocessed and explored. From the data exploration, plants exhibiting failure diagnosis remains a possible aspect interesting to be investigated
numerous subsystems, prioritisation based on the performance mea- about the human factor effect on maintenance. Other studies will be
sures employed, for instance, downtime, may be employed to establish needed to investigate the impact and balancing of in-house and out-
critical subsystems. For the identified subsystems, the different model sourced maintenance, to develop a full picture of outsourced main-
parameters are extracted from empirical data and expert assessment tenance.
while the performance measures and objective functions are formulated

Table 14
A sample of different possible applications of the proposed methodology.
Sector/Industry Sample System Sample subsystems

Manufacturing Cement mill Weigh feeder, conveyor system, mill/grinder, elevator, water system, air, and blower system
Exploration A drilling rig (fossil/geothermal/ Air system, a hoisting system, power supply, top-drive/rotary table, well monitoring system.
offshore)
Aviation Turbofan engine Air intake/compressor, combustion chamber, Turbine, exhaust system, ignition system, engine fuel system,
lubrication, exhaust system

340
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Disclosure statement Funding

No potential conflict of interest was reported by the authors. This research did not receive any specific grant from funding
agencies in the public, commercial, or not-for-profit sectors.

Appendix A. Design of Experiment full factorial analysis

Table A1
Main and interaction effects considering Tr .

Effect Estimate D.f Sum of Squares F Ratio p-value Significant

tbo −612.325 1 1499767.6 384.4216 0.00000 Yes

f 548.138 1 1201818.9 308.0512 0.00000 Yes
2 716.62 1 2054176.9 526.5282 0.00000 Yes
f 2 −379.9 1 577303.6 147.9749 0.00000 Yes
tbo 3 78.58 1 49404 31.1416 0.00000 Yes
3 −209.828 1 176110.32 46.3816 0.00010 Yes
2* 3 −42.38 1 3591.71 36.6776 0.00380 Yes
tbo 2 −94.96 1 36069.6 9.2454 0.01600 Yes
tbo f −92.073 1 33909.4 8.6917 0.01850 Yes
f* 3 −112.273 1 25210.2 9.8302 0.03500 Yes
tbo Td2 54.12 1 11715.9 3.0856 0.11710 No
3 Td2 35.02 1 4907.7 1.2925 0.28850 No
Td2 70.10 1 2543.4 0.0222 0.8840 No
Td3 37.35 1 5580.09 0.0478 0.83250 No
Td 4 12.62 1 2549 0.0062 0.9373 No
f Td 4 8.39 1 1128.00 0.0035 0.9530 No
Td2 Td3 −29.24 1 3418.74 0.0293 0.86840 No
3 Td3 15.71 1 987.22 0.0085 0.92900 No
tbo Td3 28.3 1 3204.13 0.1784 0.68020 No

Table A1, illustrates the main and interaction effects of the decision variables, where D.f is the degrees of freedom. Main and Interaction effects
that are significant, means that the variables are statistically significant (P-Value 5%), in relation to Tr . The significant interaction effects indicate
the validity of determining a balanced trade-off solution while integrating the decision variables.

References systems. Int. J. Prod. Econ. 134, 271–282. https://doi.org/10.1016/j.ijpe.2011.07.

011.
Dijoux, Y., Fouladirad, M., Nguyen, D.T., 2016. Statistical inference for imperfect main-
Alabdulkarim, A.A., Ball, P.D., Tiwari, A., 2011. Rapid modeling of field maintenance tenance models with missing data. Reliab. Eng. Syst. Saf. 154, 84–96. https://doi.
using discrete event simulation. In: Proceedings of the 2011 Winter Simulation org/10.1016/J.RESS.2016.05.017.
Conference (WSC), pp. 637–646. https://doi.org/10.1109/WSC.2011.6147792. Duan, C., Deng, C., Gharaei, A., Wu, J., Wang, B., 2018. Selective maintenance scheduling
Alrabghi, A., Tiwari, A., 2016. A novel approach for modelling complex maintenance under stochastic maintenance quality with multiple maintenance actions. Int. J. Prod.
systems using discrete event simulation. Reliab. Eng. Syst. Saf. 154, 160–170. https:// Res. 7543, 1–19. https://doi.org/10.1080/00207543.2018.1436789.
doi.org/10.1016/j.ress.2016.06.003. EN13306, 2010. Maintenance-maintenance Terminology. British Standards Institution BS
Alrabghi, A., Tiwari, A., 2015. State of the art in simulation-based optimisation for EN 1330.
maintenance systems. Comput. Ind. Eng. 82, 167–182. https://doi.org/10.1016/j.cie. Eruguz, A.S., Tan, T., van Houtum, G.J., 2018. Integrated maintenance and spare part
2014.12.022. optimization for moving assets. IISE Trans. 50, 230–245. https://doi.org/10.1080/
Antony, J., Chou, T., Ghosh, S., 2003. Training for design of experiments. Work. Stud. 52, 24725854.2017.1312037.
341–346. https://doi.org/10.1108/00438020310502642. Gharbi, A., Kenné, J.P., 2005. Maintenance scheduling and production control of mul-
Ba, K., Dellagi, S., Rezg, N., Erray, W., 2016. Joint optimization of preventive main- tiple-machine manufacturing systems. Comput. Ind. Eng. 48, 693–707. https://doi.
tenance and spare parts inventory for an optimal production plan with consideration org/10.1016/j.cie.2004.12.007.
of CO2emission. Reliab. Eng. Syst. Saf. 149, 172–186. https://doi.org/10.1016/j.ress. ISO 14224:2016, 2016. Petroleum, Petrochemical and Natural Gas Industries —
2016.01.006. Collection and Exchange of Reliability and Maintenance Data for Equipment. (ISO
Bazargan, M., 2016. Airline maintenance strategies – in-house vs. outsourced – an opti- 14224:2016), British Stanadards.
mization approach. J. Qual. Mainten. Eng. 22, 114–129. https://doi.org/10.1108/ Jardine, A.K., Tsang, A.H., 2013. Maintenance, Replacement, and Reliability, second ed.
JQME-08-2015-0038. CRC Press, Taylor & Francis Group, Boca Ranton, FL.
Bousdekis, A., Magoutas, B., Apostolou, D., Mentzas, G., 2015. A proactive decision Kelton, D., Sadowski, R., Swets, N., 2010. Simulation with Arena, Fifth. ed. McGraw-Hill,
making framework for condition-based maintenance. Ind. Manag. Data Syst. 115, Inc., New York, NY, USA ©2010.
1225–1250. https://doi.org/10.1108/IMDS-03-2015-0071. Kennedy, W.J., Wayne Patterson, J., Fredendall, L.D., 2002. An overview of recent lit-
Bouslah, B., Gharbi, A., Pellerin, R., 2018. Joint production, quality and maintenance erature on spare parts inventories. Int. J. Prod. Econ. 76, 201–215. https://doi.org/
control of a two-machine line subject to operation-dependent and quality-dependent 10.1016/S0925-5273(01)00174-8.
failures. Int. J. Prod. Econ. 195, 210–226. https://doi.org/10.1016/j.ijpe.2017.10. Kleijnen, J.P.C., Sanchez, S.M., Lucas, T.W., Cioppa, T.M., 2005. State-of-the-Art review: a
016. user's guide to the brave new world of designing simulation experiments. Inf. J.
Cai, J., Yin, Y., Zhang, L., Chen, X., 2017. Joint optimization of preventive maintenance Comput. 17, 263–289. https://doi.org/10.1287/ijoc.1050.0136.
and spare parts inventory with appointment policy. Math. Probl Eng. 2017. https:// Lei, Y., Liu, J., Ni, J., Lee, J., 2010. Production line simulation using STPN for main-
doi.org/10.1155/2017/3493687. tenance scheduling. J. Intell. Manuf. 21, 213–221. https://doi.org/10.1007/s10845-
de Almeida, A.T., Cavalcante, C.A.V., Alencar, M.H., Ferreira, R.J.P., de Almeida-Filho, 008-0194-1.
A.T., Garcez, T.V., 2015. Multicriteria and multiobjective models for risk, reliability Li, H.Z., Jing, T., Zhang, H., 2013. Arena-based modeling of the maintenance operation
and maintenance decision analysis. Int. Ser. Oper. Res. Manag. Sci. https://doi.org/ for a wind farm. Appl. Mech. Mater. 401–403, 2205–2208. https://doi.org/10.4028/
10.1007/978-3-319-17969-8. www.scientific.net/AMM.401-403.2205.
de Smidt-Destombes, K.S., van der Heijden, M.C., van Harten, A., 2007. Availability of k- Louit, D., Pascual, R., Banjevic, D., Jardine, A.K.S., 2011. Optimization models for critical
out-of-N systems under block replacement sharing limited spares and repair capacity. spare parts inventories-a reliability approach. J. Oper. Res. Soc. 62, 992–1004.
Int. J. Prod. Econ. 107, 404–421. https://doi.org/10.1016/j.ijpe.2006.08.013. https://doi.org/10.1057/jors.2010.49.
Dehayem Nodem, F.I., Kenné, J.P., Gharbi, A., 2011. Simultaneous control of production, Mccullagh, P., 2002. What is a statistical model? 1. Ann. Stat. 30, 1225–1310. https://doi.
repair/replacement and preventive maintenance of deteriorating manufacturing org/10.1214/aos/1035844977.

341
J.M. Wakiru et al. International Journal of Production Economics 208 (2019) 329–342

Nguyen, K., Van, P. Do, Grall, A., 2014. A joint predictive maintenance and spare parts Scarf, P.A., Cavalcante, C.A.V., 2012. Modelling quality in replacement and inspection
provisioning policy for multi-component systems using RUL prediction and im- maintenance. Int. J. Prod. Econ. 135, 372–381. https://doi.org/10.1016/j.ijpe.2011.
portance measure. In: Proceedings of the 2nd …, pp. 1–8. 08.011.
Nguyen, T.P.K., Yeung, T.G., Castanier, B., 2013. Optimal maintenance and replacement Sharma, A., Yadava, G.S., Deshmukh, S.G., 2011. A literature review and future per-
decisions under technological change with consideration of spare parts inventories. spectives on maintenance optimization. J. Qual. Mainten. Eng. 17, 5–25. https://doi.
Int. J. Prod. Econ. 143, 472–477. https://doi.org/10.1016/j.ijpe.2012.12.003. org/10.1108/13552511111116222.
NIST/SEMATECH, 2012. e-Handbook of Statistical Methods [WWW Document]. URL. Van Horenbeek, A., Buré, J., Cattrysse, D., Pintelon, L., Vansteenwegen, P., 2013. Joint
http://www.itl.nist.gov/div898/handbook/. maintenance and inventory optimization systems: a review. Int. J. Prod. Econ. 143,
Nowakowski, T., Werbińka, S., 2009. On problems of multicomponent system main- 499–508. https://doi.org/10.1016/j.ijpe.2012.04.001.
tenance modelling. Int. J. Autom. Comput. 6, 364–378. https://doi.org/10.1007/ Van Horenbeek, A., Pintelon, L., 2013. A dynamic predictive maintenance policy for
s11633-009-0364-4. complex multi-component systems. Reliab. Eng. Syst. Saf. 120, 39–50. https://doi.
Rezg, N., Chelbi, A., Xie, X., 2005. Modeling and optimizing a joint inventory control and org/10.1016/j.ress.2013.02.029.
preventive maintenance strategy for a randomly failing production unit: analytical Wakiru, J., Pintelon, L., Muchiri, P., Chemweno, P., 2018. Maintenance optimization:
and simulation approaches. Int. J. Comput. Integrated Manuf. 18, 225–235. https:// application of remanufacturing and repair strategies. Procedia CIRP 69, 899–904.
doi.org/10.1080/0951192052000288152. https://doi.org/10.1016/j.procir.2017.11.008.
Rivera-Gómez, H., Gharbi, A., Kenné, J.P., Montaño-Arango, O., Hernandez-Gress, E.S., Wang, W., 2012. An overview of the recent advances in delay-time-based maintenance
2016. Production control problem integrating overhaul and subcontracting strategies modelling. Reliab. Eng. Syst. Saf. 106, 165–178. https://doi.org/10.1016/j.ress.
for a quality deteriorating manufacturing system. Int. J. Prod. Econ. 171, 134–150. 2012.04.004.
https://doi.org/10.1016/j.ijpe.2015.10.008. Wu, S., 2012. Assessing maintenance contracts when preventive maintenance is out-
Roux, O., Jamali, M.A., Kadi, D.A., Châtelet, E., 2008. Development of simulation and sourced. Reliab. Eng. Syst. Saf. 98, 66–72. https://doi.org/10.1016/j.ress.2011.10.
optimization platform to analyse maintenance policies performances for manu- 004.
facturing systems. Int. J. Comput. Integrated Manuf. 21, 407–414. https://doi.org/ Xia, T., Xi, L., Pan, E., Fang, X., Gebraeel, N., 2017. Lease-oriented opportunistic main-
10.1080/09511920701574214. tenance for multi-unit leased systems under product-service paradigm. J. Manuf. Sci.
Sanchez, S.M., Moeeni, F., Sanchez, P.J., 2006. So many factors, so little timeySimulation Eng. 139, 071005. https://doi.org/10.1115/1.4035962.
experiments in the frequency domain. Int. J. Prod. Econ. 103, 149–165. https://doi. Zhu, Q., Peng, H., Timmermans, B., van Houtum, G.J., 2017. A condition-based main-
org/10.1016/j.ijpe.2005.06.007. tenance model for a single component in a system with scheduled and unscheduled
Sarker, R., Haque, A., 2000. Optimization of maintenance and spare provisioning policy downs. Int. J. Prod. Econ. 193, 365–380. https://doi.org/10.1016/j.ijpe.2017.07.
using simulation. Appl. Math. Model. 24, 751–760. https://doi.org/10.1016/S0307- 014.
904X(00)00011-1. Zohrul Kabir, A.B.M., Farrash, S.H.A., 1996. Simulation of an integrated age replacement
Savsar, M., 2015. Availability analysis of a power plant by computer simulation. Int. J. and spare provisioning policy using SLAM. Reliab. Eng. Syst. Saf. 52, 129–138.
Electr. Comput. Energ. Electron. Commun. Eng. 9 (9), 495–498. https://doi.org/10.1016/0951-8320(96)00013-0.

342