Chapter 3

Spare Parts Inventory

Control based on
Maintenance Planning

3.1 Introduction
Spare parts demand forecasting is essential to controlling spare parts inventories
and avoiding high spare part shortage and holding costs. Time series methods estim-
ate demand based on history (see e.g. Syntetos and Boylan, 2005), and as such they
may work well when the historical situation is comparable with the future. They
respond reactively to unprecedented factors and cannot predict the timing of sudden
peaks in demand. This is especially problematic for spare parts demand because of
its intermittency and lumpiness (Petropoulos and Kourentzes, 2015).
Advance demand information (ADI) is information on demand, either perfect or
imperfect, that is available ahead of the actual demand occurrence (see also Tan
et al., 2007). This concept has been widely used in various industrial settings out-
side of the spare parts context, e.g. the demand forecasting of e-commerce (Ozer,
2003), customized products (Gallego and Ozer, 2001; Johnson and Whang, 2002),
construction industry (van Donselaar et al., 2001). To overcome the limitations of
time series methods in dealing with spare part demand intermittency and lumpiness,
in this chapter we propose to use planned maintenance tasks as a source of ADI for
spare parts inventory control.

45
3.1. Introduction 46

We focus on maintenance tasks that prescribe to inspect a part of the asset, and
depending on the condition of the part, it is then either immediately replaced by
a spare part, or it may remain in the asset. Such on-condition maintenance tasks
are a cost-effective tool for ensuring that parts continue to meet their functional
and safety requirements, and they therefore constitute an important part of modern
maintenance policies for aircraft, trains, and other capital assets.
The resources that enable maintenance, e.g. mechanics and a maintenance hangar,
need to be planned ahead of the actual maintenance. To enable this maintenance
logistics planning, companies specify which on-condition maintenance tasks will be
performed some time periods into the future (see Section 3.3 for a detailed discus-
sion). It is this maintenance logistical plan that we propose to use as a source of
ADI in this paper. We need to overcome two complications when using this form of
ADI to control spare parts inventories. First, on-condition maintenance tasks are a
form of imperfect demand information: Only upon inspection does it become clear
whether an on-condition maintenance task constitutes a spare part demand. So using
the maintenance plan involves dealing with this inherent uncertainty. Second, while
the need for logistics planning forces companies to plan on-condition maintenance
tasks ahead of time, this plan is only available and reliable a few months into the
future.
We contribute to literature by proposing a new approach for joint forecasting
and inventory control based on the maintenance plan. The approach endogenously
links maintenance tasks to parts usage based on maintenance data. Moreover, our
approach integrates the demand forecasting model with an inventory control model.
The natural approach for this would be to somehow extrapolate the demand fore-
cast for the first few months, but this leads to myopically ordering too many parts
when demand is forecasted to be high, which is very costly for slow moving parts.
We contribute a method to overcome this: We extend the demand forecast using a
hybrid forecasting approach, and we propose a forward-looking inventory procedure
that explicitly takes into consideration the risk that parts stay in inventory for a
long time after procurement. The hybrid forecasting approach combines an ADI-
based approach for the first few periods with a regular forecasting approach for the
remainder of the horizon. The forward-looking inventory policy solves in each period
a stochastic dynamic program based on the latest demand forecasts over the horizon.
The forecasting element of our approach is conceptually related to spare parts
forecasting models that use the delay-time model (e.g. Wang and Syntetos, 2011),
but unlike those approaches our approach does not need the delay-time distribution
47 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

of each component as input, but it estimates from data all parameters needed to
arrive at forecasted demand distributions. Other approaches improve forecast accur-
acy at the expense of increasing the difficulty, the effort and the operational cost to
predict demand, e.g. investing in a condition monitoring system (Topan et al., 2018),
see also Driessen et al. (2010). In particular, while our method needs more data than
typical time-series methods, we need considerably less data and information than ap-
proaches based on the delay-time model (e.g. Poppe et al., 2017; Wang and Syntetos,
2011) or on system monitoring systems (Lin et al., 2017; Topan et al., 2018). This
reduction in the data requirements constitutes an important step towards enabling
the application of ADI-based approaches in practice, and in particular towards rolling
out such approaches over entire spare parts assortments.
We illustrate this by assessing the potential value of implementing our approach
for inventory control based on data for thousands of parts of two companies, and
compare it to a state-of-the-art time series forecasting approach (viz. Syntetos and
Boylan, 2001). We find that optimizing inventory using the maintenance plan yields
a very substantial cost reduction of 23-51%, compared to the benchmarks. To our
knowledge, this is the first proof that an ADI-based approach yields value in a prac-
tical setting, where all parameters need to be estimated from data.
The remainder of this chapter is organized as follows. The next section gives
an overview of the relevant literature. In Section 3.3 we discuss the availability of
planned on-condition maintenance tasks in practice. In Section 3.4 we discuss our
approach. Section 3.5 gives the setup and results of our experiments using two sets
of real company data. The final conclusions are presented in Section 3.6.

3.2 Literature Review

Spare parts demand can either be forecasted based on historical data, advance
demand information or a combination of both (Driessen et al., 2010). Since our ADI-
based method represents an alternative to time-series forecasting methods for inter-
mittent spare parts demand, we first briefly review literature those latter methods
(for a more detailed review see Van Wingerden et al., 2014). Time series forecasting
methods for intermittent demand include parametric and nonparametric methods.
Croston’s method (Croston, 1972) is an early example of a parametric method. Re-
cent works analyze and improve this method (Syntetos and Boylan, 2005; Syntetos
and Boylan, 2001; Teunter et al., 2011; Syntetos, 2001; Shale et al., 2006) and pro-
pose alternative parametric methods (Willemain et al., 1994; Ghobbar and Friend,
3.2. Literature Review 48

2003; Eaves and Kingsman, 2004; Romeijnders et al., 2012). For insightful discus-
sions regarding parametric methods refer to Boylan and Syntetos, 2010 and Prak
and Teunter, 2019. Non-parametric methods construct empirical distributions of de-
mand, see e.g. Van Wingerden et al. (2014), Willemain et al. (2004), Porras and
Dekker (2008), and Zhu et al. (2017).
We next review contributions that apply ADI in demand forecasting and invent-
ory control. ADI may take different forms (e.g. service contract, sensor data from
machines, part age, etc) and topics in literature include how to derive demand from
different forms of ADI and how to respond to ADI. ADI literature can be divided into
two streams: perfect ADI refers to situations where the quantity and timing of de-
mand is known in advance, and imperfect ADI refers to cases where some information
regarding demand is known, but not the exact quantity and timing.
Hariharan and Zipkin (1995) study where customers place their orders in advance
(i.e. a demand leadtime), and show that this form of perfect ADI is mathematically
equivalent to a reduction in the supply leadtime. Gallego and Ozer (2001) also study
perfect ADI and show that state-dependent (s, S) policies are optimal in the periodic
review model for positive set-up cost. The study gives a lower bound to its extension
to a distribution system, which is studied in Ozer (2003). Gallego and Ozer (2003)
consider perfect ADI in a multi-echelon system. They find that the value of ADI on
each echelon is influenced by the lead time of that echelon, and prove the optimality
of state-dependent, echelon base-stock policies. The key finding for perfect ADI is
that the inventory position should include outstanding demand.
We now discuss papers with imperfect ADI. Tan et al. (2007) study an inventory
problem with ADI that might either be realized as demand, wait in the system one
more period or leave the system without demand realization with given probabilities.
They show that the optimal policy is of order-up-to type, with the order-up-to level
depending on the ADI information. Tan et al. (2009) consider imperfect ADI in an
ordering and rationing problem with two demand classes. ADI is used to make a
better rejection decision to lower class demand.
In addition to the above papers utilizing ADI to obtain (optimal) inventory
policies, Abuizam and Thomopoulos (2005) and Tan (2008) apply ADI in demand
forecasting. Abuizam and Thomopoulos (2005) propose a Bayesian method to up-
date the expected amount of orders. However, Bayesian updates might fail in some
problems as they rely on the distribution assumption, give one sided updates and fail
to consider customer patterns (Tan, 2008). Therefore, Tan (2008) combine expert
judgmental prediction and demand estimation from ADI. ADI are subject to change
49 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

in time and orders are partially materialized. Historic record is used to model the
order changing behavior. Van der Auweraer et al., 2018 review the forecasting meth-
ods in which ADI takes the form of installed base information. Forecasting methods
are evaluated either in forecasting accuracy or in the integrated inventory system.
Syntetos et al. (2010) and Boylan and Syntetos (2006) stress on the importance of
considering inventory metrics rather than standard forecasting accuracy measures in
evaluating the forecasting method. Simulation is the most widely used in evaluating
the forecasting method and inventory performance, e.g. Wang and Syntetos (2011),
Zahedi-Hosseini et al. (2017) and Poppe et al. (2017).

Using the imperfect ADI concept for improving spare parts inventory control could
potentially aid practitioners to overcome the difficulties posed by spare part demand
intermittency and lumpiness, and a few researchers have developed approaches in
this general direction. Deshpande et al. (2006) track the age of parts in aircraft and
use this information as ADI to improve spare parts inventory control. Pince et al.
(2015) consider a manufacturer with contractual obligations to provide parts to its
customers, and study the drop in demand rate resulting from contract expiration.
Their proposed policy reduces the base stock level ahead of actual contract expiry.
Basten and Ryan (2015) consider a single stocking point that satisfies demands res-
ulting from corrective and preventive maintenance, and assume that perfect ADI is
available for preventive maintenance. They propose heuristics for order and invent-
ory allocation decisions, and find that the joint inventory requirement will be reduced
due to the effect of risk pooling. Romeijnders et al. (2012) develop a two-step method
that makes use of the component repairs in spare parts demand forecasting. They fo-
cus on comparing various time series forecasting methods without ADI, but mention
that ADI in this setting can considerably increase forecast accuracy.

Wang and Syntetos (2011), Topan et al. (2018), Lin et al. (2017) and Poppe et
al. (2017) pioneer new approaches towards spare parts inventory control driven by
planned/foreseen maintenance, and as such they are arguably most closely related
to our present work. These works are conceptually based on delay-time degradation
models. To our knowledge, this key idea was introduced in Wang and Syntetos (2011),
in the context of a block-based inspection policy. They assume that distributions for
both the initial and the delay-time of the delay-time model are known. They focus
solely on forecasting, and develop a forecasting model by computing the conditional
probabilities of parts needing replacement during inspection and between the inspec-
tion intervals. They find that this model can reduce the forecast error substantially.
Topan et al. (2018) consider an asset monitored by a real-time condition monitoring
3.3. On-condition maintenance tasks as advance demand information 50

system that generates imperfect warnings that may indicate that a part is failing.
They develop effective spare parts inventory control policies in this situation. Poppe
et al. (2017) consider an asset monitored by a real-time condition monitoring system,
and investigate the impact of adopting a condition based maintenance policy on in-
ventory control, where they use corrective maintenance and periodic maintenance as
benchmarks. In contrast, Lin et al. (2017) investigate the value of condition mon-
itoring systems for spare parts inventory control without changing the maintenance
policy, and find that it may be substantial.
All these approaches rely heavily on component degradation information, in the
form of real-time condition monitoring and/or complete distributional information
on the degradation process. In contrast, our approach needs no such information.
Instead, we propose to directly estimate the probability that a part needs replacement
in an on-condition maintenance task from data. This greatly simplifies applying our
approach in practice. In addition, we are the first to use real data to test the value
of ADI approaches for spare parts inventory control. Note that our tests involve
assessing the value of the model for inventory control, and as such they contribute
to an understanding of the potential value of our approach in practice.

3.3 On-condition maintenance tasks as advance de-

mand information
We first discuss the on-condition maintenance concept versus other maintenance
concepts, and explain how planned on-condition maintenance tasks constitute ADI.
We then discuss in more detail how information regarding planned on-condition main-
tenance tasks arises in practice. The discussion is based on the experience of the
authors working closely with two maintenance organizations (cf. Section 3.5). The
second author has been working with these two companies for over 10 years. The first
author conducted site visits to these and other maintenance organizations to verify
the validity of the idea using maintenance plan for spare parts demand forecasting.
The third author has worked with many industrial partners and accumulated deep
insights to the maintenance industry. Finally, we note that Driessen et al. (2010)
bring up a similar discussion based on in-depth interviews with a wide range of
maintenance organizations. (Unlike us, they do not develop forecasting methods. )
The simplest form of maintenance is perhaps break-down maintenance, i.e. run-
to-failure. In contrast, preventive maintenance encompasses a wide range of main-
tenance strategies aimed at preventing failures, and we discuss such strategies next.
51 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

In time-based maintenance a part or component is replaced periodically, e.g. after a

fixed amount of time (e.g. every 6 months) or usage (e.g. every 20000 landings of
an aircraft). Time-based maintenance can be planned ahead easily, and no condition
information is needed to apply it, but it has the disadvantage that the useful life of
the replaced parts may be poorly used. Therefore, when it is economically feasible
to do so, companies inspect parts of the asset before deciding upon replacement;
the part is then only replaced if degradation is above some threshold, hence the
term on-condition maintenance task. This approach is typically motivated using a
delay model for part degradation (Wang, 2011; Wang, 2012). Arguably, on-condition
maintenance tasks are an example of condition-based maintenance, but most scholars
reserve this latter term for situations where the condition is real-time monitored. In
practice, the prevalence of real-time condition monitoring systems is still low because
of their high associated cost (Topan et al., 2018).

In this chapter, we propose to use planned (but not necessarily periodic) on-
condition maintenance tasks as spare parts ADI. This generic idea is broadly applic-
able across a wide range of maintenance organizations. The key and only require-
ments of the proposed approach are that on-condition maintenance tasks are known
beforehand (e.g. 1 month into the future), and that information on past on-condition
maintenance tasks and resulting spare parts usage is stored. In typical high-tech as-
set maintenance settings, the first requirement is satisfied in the sense that a broad
range of on-condition maintenance tasks can be accurately foreseen beforehand, and
data collection is often compulsory because of traceability requirements. Indeed, as-
set maintenance must be planned ahead of time to organize availability of the asset,
qualified mechanics, tools, maintenance hangar, etc. The scope of asset maintenance
is typically also known beforehand. Therefore, for those assets which have a big
ratio of planned maintenance, using information on the on-condition maintenance
task is very beneficial from cost perspective. Moreover, increasing adoption of main-
tenance management software has made data on the maintenance plan available in
formats useful for automatic decision making, which has increased the potential of
and need for the approach we propose. In the following, we explain in more detail
the applicability of the approach for modularly designed assets. Modularly designed
assets contain many line-replaceable units (LRUs) that can be removed during asset
maintenance. Examples include aircraft, trains, trams, and many other high-tech
machines. LRUs are typically removed periodically in order to inspect them. In par-
ticular, each LRU has an associated inspection interval and inspection scope. This
scope consists in the on-condition maintenance tasks on parts of the LRU that to-
3.3. On-condition maintenance tasks as advance demand information 52

gether constitute the inspection. Both inspection intervals and degradation limits for
on-condition tasks are typically prescribed by the manufacturer of the LRU, which
bases its prescriptions on quantitative analysis in so-called reliability-centered/risk-
based maintenance studies, see e.g. Khan and Haddara (2003) and Moubray (1997).
The inspection of LRUs after removal from the asset is typically carried out in spe-
cialized repair shops. Therefore, parts of the LRU that are replaced depending on
their condition are typically referred to as shop-replaceable units (SRUs). Example:
the manufacturer may specify that the rear servomotor (=LRU) of a certain type of
aircraft must be removed for inspection every 8000 flight hours. Moreover, the man-
ufacturer specifies that if, during inspection, it is found that the coil (=SRU) of the
servomotor shows any signs of corrosion, it must be replaced.

Asset maintenance is clustered in order to efficiently satisfy the component safety

requirements prescribed by manufacturers. E.g. in aviation it is common practice to
define maintenance with several depth levels, e.g. a, b, c and d level maintenance.
In a-level maintenance the scope is small, while d-level maintenance encompasses
the removal and inspection of a wide range of LRUs. The various levels of asset
maintenance and their frequency are designed such that inspection intervals of indi-
vidual LRUs are guaranteed. As a consequence of this careful design of the various
checks, the work scope of such checks is typically specified beforehand, i.e. it is known
which LRUs will be removed for inspection in which check. Moreover, maintenance
organizations make detailed plannings of the maintenance of their fleet, in order to
align availability of bottleneck resources such as maintenance hangars, mechanics and
tooling, and moreover to ensure that the operational capabilities of the fleet remain
at a sufficiently high level. Example (continued): The fleet maintenance plan of
an operator specifies that in the upcoming four weeks, each week two aircrafts of a
specific type will undergo c-level maintenance, which includes removal and inspection
of the rear servomotor installed in those aircraft.

The main idea of this chapter is then to use the maintenance plan, and in particu-
lar the on-condition maintenance tasks that can be derived from the plan, as input to
forecast spare parts demand. Maintenance plans may be made years in advance, but
the plan is not reliable far into the future. This may for example be a consequence of
cumulative forecasted usage (e.g. flight hours/kms) deviating from actual cumulative
usage, or changes in the plan, etc. However, on the short term (e.g. a few months
into the future instead of years) cumulative usage can be more accurately forecasted,
making such deviations rare. Another reason for the plan to be more reliable on the
short term is that deviations on the short term cause operational disruptions as well
53 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

as unavailability and/or idle time of bottleneck resources. We propose to base ADI

on this reliable time horizon, and we develop a method that reverts to time-series
forecasting for periods beyond this horizon. Example (continued): The coil in
the rear servomotor has historically been replaced in one out of three repairs. Based
on two planned removals of the rear servomotor per week, an expected demand of
2/3 = 0.67 coils per week is forecasted for the next four weeks.
To some extent, the approach can help to predict peaks in demand. Example
(continued): Suppose that an operator decides to inspect all servomotors in the
fleet in April, to avoid any problems in the busy summer months. As a consequence,
there will be 18 inspections per week for the month April. Based on this, expected
demand of coils suddenly grows to 6 per week, or about 24 per month, and using the
maintenance plan would enable to predict the sudden demand hike. However, note
that there would still be considerable remaining variability in coil demand because of
variations in actual coil replacement rates.
We emphasize that while maintenance plans and tasks are nowadays increasingly
available in maintenance management systems, such systems are not ubiquitous, and
even if an investment in such systems is made, it is not necessarily trivial to extract
information from such systems in a format usable for spare parts decision making.
Apart from developing methods to use on-condition maintenance plans for spare parts
inventory control, one of the goals of this chapter is testing their potential value in
practice using data from companies in aerospace and train maintenance that were
able to extract the necessary data from their system. We believe this assessment may
help in driving business cases for the proposed approach.

3.4 Methods
Our approach is applicable in general for maintenance organizations that perform
on-condition maintenance tasks. For concreteness, in what follows we adopt termin-
ology of a repair shop. We focus on one specific LRU/component that is regularly
inspected by the repair shop, which is an establishment specialized in repairs of line-
replaceable units (cf. Section 3.3). Inspection consists of determining the condition
of parts of the component: if a part is degraded beyond some acceptable level, then
it must be removed and discarded, and replaced by a spare part. So for the repair
shop, each component repair corresponds to a number of on-condition maintenance
tasks: one for each part in the component.
To complete repairs quickly, the repair shop keeps a local inventory of spare
3.4. Methods 54

parts, and our focus is on inventory control of one specific spare part that may be
used in the component. In case the part is needed but out of stock, an emergency
order is placed, and after the emergency leadtime the repair continues. Emergency
orders are a common way to avoid very long and costly delays of maintenance, and
the emergency order leadtime is understood to be much shorter than the regular
leadtime. Note that placing an emergency order is (mathematically) equivalent to
a lost sale, and we take this latter perspective. We consider the penalty costs for
the lost sale to be ce , where ce includes emergency order cost and costs of delaying
the repair. Note that delaying a component repair may be costly as it requires the
mechanic to store the inspected component, and to later retrieve it, which is time-
consuming. As is customary in industry (cf. Romeijnders et al., 2012), inventory is
reviewed periodically, resulting in a periodic-review, single-item, lost-sales inventory
system. We denote periods by t ∈ {1, . . . , T }. Here, T is the last period before the
end of the horizon.

We consider a constant lead time L for a regular order. Parts ordered at the start
of period t arrive at the start of period t + L. We let L = 1, which corresponds to
a situation where parts ordered one period are available in the next period. This
is reasonable because repair shop inventory is replenished from a central warehouse
every period. Moreover, since many SRUs are relatively inexpensive, it is affordable
to avoid stock-outs in this central warehouse. More importantly, focusing on this
assumption avoids very technical inventory models and allows us to focus on the
exposition of the key ideas regarding the integration of ADI and inventory control.
For the same reasons, we assume no economies of scale in ordering. Inventory has
holding cost h per part per time unit, and since we work with a finite horizon, leftover
inventory at the end of period T is penalized with cost s per part. This penalty may
either reflect the cost of scrapping the inventory (in case this is really the end of the
horizon in which the part is used), or it may reflect the cash tied up in inventory at
the end of the horizon, which causes a potential loss of opportunity, cf. Section 3.4.2.

As described in Section 3.3, we focus on cases where the repair shop knows the
number of on-condition maintenance tasks (component repairs) that will be carried
out some periods in advance. In particular, at the start of period t, the repair shop
knows the number of on-condition maintenance tasks for periods t, t + 1, . . . , t + Tm ,
where Tm corresponds to the number of periods that tasks are known in advance.
Note that the spare part demand for period t is only revealed during period t: only
upon inspection does it become clear whether a part needs replacement. Also, we
assume that the repair shop keeps track of past on-condition maintenance tasks and
55 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

the resulting spare parts demand. (This is often required for quality assurance reasons
anyhow.)

3.4.1 Forecasting

The goal of this section is to arrive at a demand forecast for upcoming periods that
can serve as a basis for inventory control. Note that for this latter purpose, a demand
distribution forecast rather than a point estimate is needed. Let dt denote the actual
spare part demand in period t, and let At denote the number of maintenance tasks
in period t. At the start of period t, we know the values dt+i and At+i for i < 0,
because those periods are in the past. We also know At+i for 0 ≤ i ≤ Tm : this is the
ADI.
Conceptually, an on-condition maintenance task results in a spare part demand
with some (failure) probability p. In practice, such a probability needs not not be
stationary; it may be subject to change as the components age, and as their usage
pattern changes, etc. Moreover, the precise value of this probability is unknown. We
therefore suggest to estimate the value of this unknown probability from data, by
updating the forecasted failure probability pbt in every period t as follows:

dt−1

 (1 − α)b
p +α if At−1 > 0 (3.1)
t−1
pbt = At−1
if At−1 = 0 (3.2)

pbt−1

Here, α is a smoothing factor. pb could be initiated as 0, or using the first few months
of the demand history.
To forecast demand more than Tm periods in advance (that is, dt+i where i > Tm )
in our proposed method, we revert to standard time-series methods to forecast the
average demand per period, which will be denoted by λt . We opt for the well-studied
Syntetos-Boylan approximation (SBA) (Syntetos and Boylan, 2001), which is an
improvement to Croston’s method (Croston, 1972):

0 b
bt = (1 − α ) St
λ (3.3)
2 b kt

Here, α0 is a smoothing parameter. The estimated demand size Sbt is updated by

(
(1 − α0 )Sbt−1 + α0 dt−1 for dt−1 > 0 (3.4)
Sbt =
Sbt−1 for dt−1 = 0 (3.5)
3.4. Methods 56

and the estimated demand interval b

kt is updated by
(
(1 − α0 )b
kt−1 + α0 kt−1 for dt−1 > 0 (3.6)
kt =
b
kt−1
b for dt−1 = 0 (3.7)

At the start of period t, to arrive at a forecasted demand distribution D

b t,t+i for
some upcoming period t + i, we distinguish between the cases i > Tm and i ≤ Tm .
For i ≤ Tm , the spare parts demand has a binomial distribution with At+i trials
and success probability p. Since p is unknown, we substitute the estimated value pbt
to obtain for 0 < i ≤ Tm : D b t,t+i ∼ B(At+i , pbt ), where B(n, p) denotes a binomial
distribution with parameters n and p. For i > Tm , At+i is not available. Because
the Poisson distribution is a good fit on spare part demand in general, for i > Tm we
forecast the demand distribution as D b t,t+i ∼ poisson(λ
bt ).
We note that there are substantial difference between using B(At+i , pbt ) for fore-
casting, versus using poisson(λ
bt ). Most importantly, B(At+i , pbt ) reacts immediately
to a large number of planned maintenance tasks, while poisson(λ
bt ) only changes after
the maintenance tasks have been executed. Secondly, using the binomial distribu-
tion B(At+i , pbt ) has the advantage that we explicitly know an upper bound on the
number of replacements, which may help reduce the stock in certain situations.

3.4.2 Inventory optimization

We develop an approach for determining the amount xt of spare parts to order in
some arbitrary period t. Since demand is non-stationary, xt should not myopically
depend on the forecasted demand during leadtime alone, but it should be forward
looking. This is easily seen based on an example: Let Tm = 1 and consider two
situations at time t: 1) pbt = 0.2, λt = 0.01, At = 10, At+1 = 10; 2) pbt = 0.2, λt = 2,
At = 10, At+1 = 10. In both cases, demand on the short term is likely around 2
since pˆt At+1 = 2, but in the first case, demand is expected to go down to 0.01 in
subsequent periods, while in the second case, demand is expected to remain around
2. That means that any items remaining at the end of period t + 1 will likely stay in
stock longer in the first case than in the second case, which should be reflected in the
order decision.
To arrive at a forward-looking policy, in each period t we solve a stochastic dy-
namic program (SDP) over periods t+i ∈ {t, . . . , T }. This SDP uses the demand dis-
tributions over said periods, but the exact demand distribution is unknown. Instead,
it is natural to use the forecasts constructed in period t: D
b t,t+i . In the following, we
57 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

briefly summarize the steps that occur in each period, and we subsequently give the
SDP used to determine the order quantity xt in each period t. For summarizing the
steps, we introduce yt and yt0 to denote the on hand inventory at the beginning and
the end of period t, respectively. Here, yt is understood to include the items that
arrive in period t.
In each period, first the order placed in the previous period arrives. Thus yt =
0
yt−1 + xt−1 . Then the order amount xt is decided. Next, spare part demand Dt
happens. Demand is satisfied by on hand inventory yt . Thus at the end of period t
we have on hand inventory yt0 = (yt − Dt )+ = (yt−1
0
+ xt−1 − Dt )+ . The holding cost
h · yt0 and emergency ordering costs ce · (Dt − yt )+ are incurred. Subsequently, the
next period starts.
To arrive at an SDP equation for deciding xt , let ft,t+i (yt+i ) denote the optimal
total discounted cost from period t + i until the end of the time horizon T , when
the starting inventory in that period is yt+i , and based on the forecasts obtained in
period t. Then ft,t+i satisfies the recursive equation:

∞
P (D̂t,t+i = d)(yt+i − d)+
X
ft,t+i (yt+i ) = min h (3.8)
xt+i ∈{0,1,...}
d=0
∞
P (D̂t,t+i = d)(d − yt+i )+
X
+ ce
d=0
∞
X
+ P (D̂t,t+i = d)ft,t+i+1 (yt+i+1 ) (3.9)
d=0

= min gt,t+i (xt+i ) (3.10)

xt+i ∈{0,1,...}

where yt+i+1 = (yt+i −d)+ +xt , and with the boundary condition ft,T +1 = −s·yT +1 .
This boundary condition reflects that, in our approach, we assume that at the end
of the horizon, inventory must be scrapped. This is in line with assumptions in our
numerical experiment. However, even in situations where the end of the horizon does
not necessarily correspond to part obsolescence, it may be wise to add a penalty to
ending the last period with inventory, since this may avoid large procurements near
the end of the horizon because of the end-of-horizon effect. So we believe that having
this penalty is wise, even if not all inventory would be scrapped at the end of the
horizon. Note that (3.10) implicitly defines gt,t+i (xt+i ). We then obtain the amount
to order in period t as xt ∈ arg minx gt,t (x). Note that ft,t+i (yt+i ) corresponds to
estimated costs based on the forecast constructed at the start of period t. Hence,
after updating the forecast, at the start of each period a new SDP is constructed to
3.5. Assessment of potential value of the method 58

arrive at xt .
We next discuss the computational effort required to solve the SDP defined in (3.9)
using backward induction. A practical upper bound for the maximal planned optimal
on-hand inventory (i.e. yt + xt ) can be obtained by using a very high percentile of
the demand distribution, because it cannot be optimal to order an amount that will
only be used with negligible probability. In particular, a Poisson random variable
√
with mean λ̄ exceeds U (λ̄) = λ̄ + 10 + 10 λ̄ with almost vanishing probability
for the Poisson case; for the Binomial case we could use the natural bound At .
yt can then be bounded by 2U (λ̄). Then backward induction for each inventory
decision involves computations over T − t + 1 periods, with 2U (λ̄) states for each
period. Computing the value for each state involves at most a constant times U (λ̄)
computations, so computation time for each decision (i.e. solving the SDP) can be
bounded by some constant times U (λ̄)2 T . For the experiments that we report in
Section 3.5.2, computation time was (much) less than a second for each decision.

3.5 Assessment of potential value of the method

In this section we use data from two maintenance companies in order to quantify
the potential value of the ideas and methods developed in this paper. In Section 3.5.1
we discuss the data and the setup of the experiments. Section 3.5.2 gives the results.

3.5.1 Data and experimental setup

We first briefly describe the two companies and their data, and then discuss the
experimental setup. The first company is Fokker Services (FS). FS provides com-
prehensive in-house component maintenance, repair and overhaul support to aircraft
operators in dedicated repair shops. Components are typically delivered to FS ac-
cording to the aircraft operator maintenance plan. FS subsequently determines the
condition of parts during initial inspection. Failed parts generate demand for spare
parts, which are delivered from a warehouse next to the repair shop. (The warehouse
is replenished from a central warehouse in the Netherlands, but this replenishment
is left out of the scope of this study. )
Repair data over a period of 134 months are available. Data are cleaned by
removing some parts which are not applicable in our model, e.g. if the bill of material
coefficients are larger than one for the part/if the part is used in a quantity larger
than one in a single repair. Each type of component and each type of spare part
has a unique serial number. In our analysis, we use the following information that
59 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

is gathered about component repairs: Period in which the component arrived at the
repair shop, component serial number, which spare parts (serial numbers) were used
in the repair operation. As a component might generate demand of different spare
parts types, we call the component and each corresponding type of spare part as
one spare part-component pair. The data set includes 24,455 different spare part-
component pairs.
We designate the first 84 months as the training set. Within this training set,
we use the average of the first 48 months to arrive at an initial estimate for the
model parameters. We then use the approach discussed in Section 3.4.1 to update
the parameters (e.g. parts failure probability) for the remaining 36 months in the
training set. The test set contains the last 50 months of data, i.e. months 85-134. In
the test set, we keep updating the parameters, and we record performance statistics
such as holding costs and emergency ordering costs.
Real SRU prices are not available. In the experiment, we consider the following
4 parameters as the experimental factors: (i) holding cost per item per time unit
h, (ii) emergency shipping cost ce , (iii) scrapping cost at the end of horizon s, (iv)
maintenance plan lead time Tm .
For the base case, we use the following parameters. At 24% holding costs per
year, a typical low part price of 5 euros amounts to h = 0.1 euros/month. Costs for
scrapping parts at the end of the horizon are set at s = 5 euros, because the costs
of scrapping are dominated by the lost investment. Regarding the emergency costs,
we found in discussions at various repair shops that delaying repairs is inconvenient
because it typically requires the mechanic to temporarily store the component, and to
later retrieve it. Additionally, even relatively short repair delays may harm customer
satisfaction. As a consequence, we set the penalty costs for an emergency shipment
as ce = 20 euro. Finally, we set Tm = 3. We design our experiments around these
base case parameters, and a sensitivity analysis is conducted to explore the effect of
changes to the base case.
Since each component constitutes an on-condition maintenance task that may
result in usage of the part, we can directly apply our methods for each spare part-
component combination. Our method is used to determine the replenishment quant-
ity in each period. Subsequently, we simulate the dynamics of the system using the
real demand and maintenance data (cf. Section 3.4.2), and obtain holding, emer-
gency shipping, and scrapping costs for all parts. To assess the value of ADI, we will
use as a benchmark a method that does not use ADI. Like the method proposed in
this chapter, the benchmark uses the recursive approach (3.9-3.10) to set spare parts
3.5. Assessment of potential value of the method 60

orders. However, the benchmark uses the time series forecast D

b t,t+i ∼ pois(λ
bt ) for
all future periods, including those with 0 < i ≤ Tm . Note that the value of At is
not needed for the benchmark, and note that the Syntetos-Boylan approximation is
used to determine λt . So the benchmark represents the state-of-the-art time-series
method.
Note that our experiments compare holding, emergency shipping, and scrapping
costs for both the proposed approach and the benchmark. Thus costs for performing
the actual on-condition maintenance tasks (i.e. labor costs & downtime of equipment)
are not considered in either approach. This is reasonable, since the latter costs are
exogenous to the model, and should be considered as sunk costs.
We also test our approach at another company: the Netherlands Railways (NS).
NS is by far the largest operator of passenger railway transport in the Netherlands.
The maintenance department of NS tracks the repair actions of main components of
trains over 35 months, and we obtained that data. The history covers information
over 138,347 repair actions on main components. At NS components may either be
replaced as part of the maintenance plan, or upon unplanned failure. The former
covers 2,727 types of components and 749 types of parts, and the latter covers 3,935
types of components and 1,485 types of spare parts. Ideas in this chapter are applic-
able to the former case, and we only use that data. We designate the first 25 months
of demand as training data, and the last 10 months as test data. Out of the training
set the first 20 periods are used for initialization of forecast parameters, and the last
5 for updating those parameters. The other settings are the same as in the FS case.
We mainly evaluate the proposed approach and benchmarks using inventory con-
trol metrics, but to provide a broader perspective, we also assess the performance of
the proposed forecast approach in isolation using the root-mean-square-error (RMSE)
and mean-absolute-deviation (MAD). In evaluating the forecasting accuracy we use
the point forecast of demand rather than the demand distribution. Therefore, we con-
sider the mean of the binomial distribution estimated in 3.4.1 and the benchmarks
include moving average (MA), exponential smoothing forecast (ES), Croston’s fore-
cast method (CR), Syntetos-Boylan approximation (SBA) and forecasting method of
Teunter et al. (Teunter et al., 2011) (TSB). The length of moving periods in the MA
method is set to be 12. The smoothing constant in ES is 0.2. We give 0.2 to both the
smoothing constant of the demand size and that of the demand interval in CR and
its modification SBA. The smoothing factors of the demand size and the probability
in TSB is set to be 0.2 and 0.1 respectively. The initial forecast in the benchmarks
is made over the first 48 months in the Fokker case and 20 months in the NS case.
61 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

3.5.2 Results

We compare the total cost of all spare parts of our proposed approach to the costs
of the benchmark that only uses time-series forecasts. Figure 3.1 shows the relative
cost reduction of all the spare parts at Fokker Services, in the Total Costs (TC),
Holding Costs (HC), Emergency Costs (EC), and Scrapping Costs (SC). Figure 3.2
does the same for NS. Each column in Figure 3.1 and 3.2 represents a single setting of
parameters: For each case the base values of parameters are h = 0.1, p = 0, ce = 20,
s = 5, Tm = 3. Figure 3.4 and 3.5 show the forecasting performance measures for
the Fokker Service and the NS respectively. Average over all types of spare parts are
given.
As the FS case has a relatively long demand history, we can make a relatively
accurate categorization of spare parts based on the number of months with positive
demand during the training period (84 months) to explore the value of the mainten-
ance plan on each category. The three categories are very-slow moving (1-5 months
with positive demand), slow-moving (6-20 months), and fast moving (21-84 months).
We have 24,455 types of part - component combinations in total. Very-slow moving
includes 21,011 combinations, slow moving covers 2,846 combinations and relatively
fast moving has 598 combinations. Figure 3.3 shows the cost reduction in each cat-
egory. We have the following observations.

• We observe that our approach reduces the total cost compared to the bench-
mark by 48% and 23% in average for Fokker Services and NS, respectively.
This illustrates that the value of the maintenance plan is very high in invent-
ory control. In eight out of ten instances in the FS case and in nine out of ten
instances in the NS case, our approach outperforms the benchmark with regard
to all three cost components. Cost reductions are mainly driven by reductions
in emergency shipping cost, followed by the holding cost and scrapping cost.
The emergency shipping cost contributes 89% to the total saving in the Fok-
ker case and 68% in the NS case. Note that since many spare parts are very
reliable, and since components have a life cycle of 5-20 years, scrapping costs
may be a substantial part of total costs because even with low stocks there is
always a risk of leftovers (cf. van Jaarsveld and Dekker, 2011). This explains
the substantial costs of scrapping for the cases. However, note that the scrap-
ping costs difference between the ADI method and the benchmark is minimal,
which implies that the assumption of scrap costs is not essential for our results.
Finally, we note that the cost reduction for Fokker Services is higher than the
3.5. Assessment of potential value of the method 62

·105
7
holding cost (ADI)
penalty cost (ADI)
6 scrapping cost (ADI)

holding cost (BM)

5 penalty cost (BM)
scrapping cost (BM)

0
1 2 3 4 5 6 7 8 9 10

Figure 3.1: The effect of parameters on the value of maintenance planning (Fokker Service).
Numbers 1-10 on the horizontal axis represent experiments with h = 0.1, h = 0.5, h = 1,
ce = 15, ce = 20, ce = 30, Tm = 1, Tm = 3, Tm = 5, Tm = 38 respectively

cost reduction for NS. This is mainly due to the fact that the maintenance plan
is more stable for NS, which reduces the value of ADI, cf. Section 3.4.1.

• The holding cost rate h has more effect on our method than the benchmark
while the emergency shipping cost ce has larger impact on the benchmark.
When h is increased from 0.1 to 0.5, the cost of our approach is increased
by 59% while 34% for the benchmark in the Fokker case. For h from 0.5 to
1, it’s 28% for our approach and 14% for the benchmark. The effect of h is
monotonic in general. For cm from 15 to 20 and to 30, the cost of our approach
is increased by 13% and 17% respectively, while for the benchmark it’s 16% and
24%. Therefore, we can conclude that our approach on average orders more
than the benchmark as to have less penalty and holding cost. The method
apparently orders at the right moment. When h/cm is large enough, the value
of the maintenance plan vanishes since the optimal policy under both methods
63 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

·105
1
holding cost (ADI)
0.9 penalty cost (ADI)
scrapping cost (ADI)
0.8
holding cost (BM)
penalty cost (BM)
0.7
scrapping cost (BM)

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9 10

Figure 3.2: The effect of parameters on the value of maintenance planning (NS). Numbers
1-10 on the horizontal axis represent experiments with h = 0.1, h = 0.5, h = 1, ce = 15,
ce = 20, ce = 30, Tm = 1, Tm = 3, Tm = 5, Tm = 38 respectively

does not place any order. When cm /h is large, both of the methods have larger
stocks. However, using the maintenance plan yields an upper bound for the
spare parts demand while the benchmark does not have access to such an upper
bound. In addition, the benchmark might place an order in the period when
there is no component arrivals as it only uses the history demand in forecasting.
Again, the ADI method places timely orders. This is verified by the observation
of the increase in the holding cost reduction with the increase of cm in both
Figure 3.1 and 3.2.

• The value of maintenance plan is not sensitive to Tm . In the Fokker case,

obtaining the maintenance plan one lead time ahead achieves 46% in cost re-
duction, while obtaining it 5 lead times ahead achieves a 52% cost reduction.
In the NS case, the cost reduction increases from 20% to 24% by increasing Tm
from 1 to 10. The marginal benefit decreases dramatically with the increase of
3.5. Assessment of potential value of the method 64

·105
7
VSM ADI
SM ADI
6 FM ADI

VSM BM
5 SM BM
FM BM

0
1 2 3 4 5 6 7 8 9 10

Figure 3.3: The effect of parameters on the value of maintenance planning, for the part
categories very slow moving (VSM), slow moving (SM), and fast moving (FM) (Fokker
Service). Numbers 1-10 on the horizontal axis represent experiments with h = 0.1, h = 0.5,
h = 1, ce = 15, ce = 20, ce = 30, Tm = 1, Tm = 3, Tm = 5, Tm = 38 respectively
65 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

0.2
RMSE

MAD
0.15

0.1

5 · 10−2

0
ADI ES MA CR SBA TSB

Figure 3.4: RMSE and MAD for the Fokker Service

0.4
RMSE

MAD
0.3

0.2

0.1

0
ADI ES MA CR SBA TSB

Figure 3.5: RMSE and MAD for the NS

3.5. Assessment of potential value of the method 66

Tm . Therefore, we conclude that it is necessary to obtain the maintenance plan

in advance as it brings substantial cost saving. However, obtaining the main-
tenance plan one lead time ahead contributes the most to the inventory cost.
It is not very cost effective to invest more in order to obtain the maintenance
plan much earlier.

• Figure 3.3 shows that costs reductions are substantial over all categories, and
relative cost reductions increase only slightly for faster moving categories. The
biggest absolute contribution to the total cost savings is made by the very
slow moving parts: 53% (averaged over the various cost parameter settings).
This is because the very-slow moving category accounts for a large proportion
of the parts. It is interesting that our approach performs well for the very-
slow moving category, because that category is very hard to forecast using
traditional techniques. Note furthermore that the very-slow moving category
is less sensitive to the holding cost rate h than the other two categories in both
the ADI method and the benchmark. This is because holding cost accounts for
a smaller proportion of the total cost for very-slow moving items.

• In addition to the inventory metrics, our ADI-based forecasting approach also

performs better comparing to the time series methods in the evaluation of
forecasting accuracy. The method ADI has an advantage over all other time
series approaches in both RMSE and MAD in each case, shown in Figure 3.4
and 3.5. Note that as we update λt for Poisson distribution by SBA in Section
3.4.1, the point forecast by SBA method in evaluating forecasting accuracy
is exactly the mean value of the Poisson distribution used in the inventory
control system. Therefore, our approach which considers ADI leads to less cost
in the inventory system and a better forecasting accuracy than the benchmark
in which demand is estimated to be Poisson distributed and its corresponding
point forecast.

• The results are consistent with an intuitive interpretation. The proposed

method can better take into account the time interval between positive de-
mands. Therefore, it prevents the system from keeping redundant stocks. This
leads to less inventory holding cost, as we observe in both cases. The proposed
approach differentiates spare part demand forecasts in different time periods by
building the dependence between spare part demand and its origin, compon-
ent arrivals. By responding to the maintenance plan, our approach makes the
spare part demand forecasting more accurate and the inventory decisions more
67 Chapter 3. Spare Parts Inventory Control based on Maintenance Planning

appropriate. As a result, we have less penalty cost for emergency shipment and
less scrapping for leftover stocks at the end of time horizon. In this way, our
approach can better achieve the goal of having the right amount of stocks at the
right moment. We expect the value of our approach in practice to be highest
for the very slow moving items, because especially such items are notoriously
difficult to control for human decision makers.

3.6 Conclusions and future research

Spare parts demand forecasting is essential to spare parts inventory control but
difficult as the demand has the feature of irregularity and lumpiness. We proposed
and tested ideas to apply ADI, in the form of planned on-condition maintenance tasks,
to improve spare parts inventory control under these circumstances. Incorporating
this form of ADI into forecasting makes the demand forecast nonhomogeneous over
the forecast horizon. Accordingly, as argued in this chapter, the inventory control
must become forward looking, and we propose an inventory optimization approach
to reflect this. We determined the potential value of the combined forecasting and
inventory optimization approach using industry data, and found that potential sav-
ings are very substantial: 51% for the aerospace maintenance case, and 23% for the
train maintenance case.
Some comments are needed to put these figures into perspective. First of all,
while aircraft component maintenance typically results from checks planned by the
operator, this information is currently only shared on an ad hoc basis, e.g. mainten-
ance organizations inform the repair shop in advance when they expect a substantial
number of removals of a specific component in a short time interval. This is mainly
in their own interest: If the repair shop is prepared, then spare part shortages are
rare. The present research solidifies these findings and underlines the economical
value of such information sharing in the supply chain. Moreover, the study provides
compelling evidence that investing in a more structured sharing of information, e.g.
in the form of a data platform, can simultaneously reduce inventory and increase
part availability. Train maintenance organizations likewise inform the repair shop
typically on an ad hoc basis. The present study shows that it would be better to
more structurally organize this, as structural sharing of information would allow for
a substantial reduction in the mismatch between spare parts demand and supply.
Finally, it is interesting to note the rather marked difference in cost reduction for
Fokker Services and NS, though in both cases cost reductions are substantial. We
3.6. Conclusions and future research 68

believe that this is caused by a more uneven maintenance pattern in the aircraft
industry compared to train maintenance.
Our approach focuses on the most common case: If parts are used in an on-
condition maintenance task, they are used in quantity 1. While this holds for the
vast majority of parts, one could generalize it to situations where multiple parts
of the same type may be replaced in a single maintenance task (van Jaarsveld et
al., 2015). Another direction for future research is related to our forecast method:
Suppose many on-condition maintenance tasks are upcoming, i.e. At+i is high. This
will likely increase future demand, so future demand is likely to be higher than λ̂t .
It would perhaps be interesting to adapt the forecasting method for λ̂t such that it
already responds to expected changes in demand. Finally, any experience on a broad
implementation of the ideas pursued in this chapter would likely teach us valuable
lessons that cannot be learned from this preliminary study alone.