Condition-Based Maintenance using the Inverse

Gaussian Degradation Model

Nan Chen∗† , Zhi-Sheng Ye† , Yisha Xiang§ , and Linmiao Zhang†

†
Department of Industrial and Systems Engineering, National University of Singapore
§ Department of Management Science, Sun Yat-Sen University

November 17, 2014

Abstract
Condition-based maintenance has been proven effective in reducing unexpected failures with
minimum operational costs. This study considers an optimal condition-based replacement
policy with optimal inspection interval when the degradation conforms to an inverse Gaus-
sian process with random effects. The random effects parameter is used to account for
heterogeneities commonly observed among a product population. Its distribution is updated
when more degradation observations are available. The observed degradation level together
with the unit’s age are used for the replacement decision. The structure of the optimal re-
placement policy is investigated in depth. We prove that the monotone control limit policy
is optimal. We also provide numerical studies to validate our results and conduct sensitivity
analysis of the model parameters on the optimal policy.
Keywords: Optimal replacement, inverse Gaussian process, Heterogeneity

1 Introduction
This paper considers the condition-based maintenance (CBM) of a unit subject to a continuous-
time stochastic degradation, and monitored through periodic inspections. CBM suggests
regular inspections using modern sensor technology. Maintenance actions are then made
∗
Corresponding author: isecn@nus.edu.sg

1
based on the inspection of working conditions. Because CBM makes use of the in-situ infor-
mation, it most often outperforms the traditional age- and block-based maintenance policies.
Therefore, the use of condition monitoring techniques and CBM has increased rapidly over
recent years (see Wang 2002; Jardine, Lin, and Banjevic 2006, for reviews and examples).
The choice of the stochastic process that characterizes the physical deterioration, the so
called prognostic models, will obviously affect the prediction of remaining useful life, and
therefore influence the decision of the maintenance strategy and its economic performance. In
the literature, many prognostic models have been proposed. In particular, when the health
states are directly observable (Jardine et al. 2006) as in our study, they can be classified
based on whether the degradation states are discrete or continuous. In the scenario with
discrete states, Markov chain models (e.g. Chen et al. 2003; Xiang et al. 2012; Yeh 1997;
Bloch-Mercier 2002) are often adopted. After specifying the probability transition matrix
among all states, these models can be used to calculate the time-to-failure distribution from
any state. The Markov chain models are particularly useful when the degradation states of
the unit cannot be precisely measured, and a rough category (e.g., good, moderate, weak)
has to be adopted instead.
In recent decades, the fast development of sensing technologies enables accurate online
measurements of the degradation levels. Under such circumstance, the degradation states
are considered continuous as opposite to discrete. Consequently, stochastic models with con-
tinuous states are commonly used (e.g. Albin and Chao 1992) to characterize the degradation
process. Notably among them, the Wiener process, Gamma process, and the inverse Gaus-
sian (IG) process, including their variants, attract significant attention because of their nice
mathematical properties and clear physical interpretations (see Ye and Xie 2014, for review).
An attractive feature of these models is that they possess independent degradation incre-
ments. This property significantly simplifies the prediction of the remaining useful life given
the historical degradation observations. As a result, it becomes much easier to determine
the optimal maintenance policies. For example, optimal CBM policies based on the Wiener
process degradation models have been extensively investigated (e.g. Hontelez, Burger, and
Wijnmalen 1996; Elwany, Gebraeel, and Maillart 2011; Guo, Wang, Guo, and Si 2013). How-
ever, since the Wiener process is not always monotone, it fails to model many degradations
well such as crack growth and wears. For such monotone degradations, Gamma processes
and IG processes are more suitable as they inherently impose the monotone constraints on
the sample paths. CBM policies based on Gamma processes can be found in the literature
as well. See Dieulle et al. (2003); Grall et al. (2002a); Liao et al. (2006) for examples and
van Noortwijk (2009) for an excellent review. Nevertheless, different from Wiener process
or Gamma process, which have been studied for a long time, IG process draws significant

2
attention only recently. IG process was introduced by Wang and Xu (2010) to the reliability
literature, and further investigated by Ye and Chen (2014). They had demonstrated signifi-
cant advantages of IG process, including but not limited to, i) clear physical interpretation
for degradation caused by accumulated small damages; ii) dual relationship with Wiener
process, providing intuitive understanding; iii) flexible ways to incorporate random effects
and covariates to account for different types of heterogeneities; iv) explicit analytical for-
mulas for important quantities, such as PDF and CDF of remaining useful life distribution.
However, despite the increasing attention on IG process (see e.g. Wang and Xu 2010; Ye and
Chen 2014; Ye et al. 2014; Peng et al. 2014; Liu et al. 2014), there is scarce literature on
CBM policies based on IG processes despite their useful applications.
While most of existing CBM policies assume that the prognostic models are fixed for all
units across the population, this assumption is increasingly challenged. Because of diverse
usage and environmental differences, the degradation characteristics of units from the same
population are often different. As reported in many studies (e.g., Lawless and Crowder 2004;
Gebraeel et al. 2005; Chen and Tsui 2013; Liao and Tian 2013), heterogeneities often exist
across the population, causing different degradation patterns. Many degradation models have
been proposed to integrate the heterogeneities. Generally speaking, there are two approaches
to take into account the heterogeneities. The first one is to use random effects models while
the second is to impose a prior distribution on one or some of the model parameters. The crux
of these two approaches is essentially the same, that is, some parameters are unit-specific
and different across unit. The rest are common parameters shared by the entire population.
As more degradation observations are collected, unit-specific parameter(s) can be estimated
more precisely and can characterize the degradation process better. Unfortunately, while
accounting for heterogeneity becomes common and has been well-recognized in degradation
modeling and analysis, it also makes the optimal CBM policy difficult to find because the
degradation process becomes nonstationary and age dependent.
In this paper, we develop an optimal inspection/replacement policy for units whose
degradation can be modeled by an IG process. We incorporate the random effects in the
IG process to model the unit-specific heterogeneity, and investigate the effects of such a
heterogeneity on maintenance planning. The unit is inspected periodically. Upon each
inspection, we make the decisions on whether a preventive replacement is needed. Failure is
not self-announcing and can only be revealed through inspections. Operating in the failure
state, the unit incurs a downtime cost. Our goal is to find the optimal inspection interval
and replacement policy to minimize the total operational costs which include inspection cost,
preventive/corrective replacement cost and downtime cost. We formulate the problem as a
Markov decision process, and analyze the structural properties of the optimal policies. Our

3
model shares some common features with the one proposed in Elwany et al. (2011), but
it also differs from it in many aspects. The main differences are summarized as follows.
First, the failure mechanism in our model is different. Elwany et al. (2011) used a Wiener
process with linear drift to model the degradation, whose degradation paths are not strictly
increasing. However, in many applications, the degradation is monotone. We consider the
IG process with heterogeneity which are more suitable in such scenarios. Second, in Elwany
et al. (2011) zero inspection cost is assumed, and the inspection interval becomes irrelevant to
total operational cost. In practice, inspection costs might not be negligible, and it becomes
crucial to determine the optimal inspection schedule to minimize the total cost. In our
model, we explicitly consider the inspection costs and seek for optimal inspection interval in
addition to the replacement policy. This formulation contains their model as a special case,
and is expected to be more widely applicable. Third, we consider the downtime cost in our
model. Since the failure is not self-announcing, the unit often incurs additional cost due to
efficiency or quality loss when operating in the failure state. The downtime cost is especially
alarming if the inspection interval is long, and consequently it becomes one deciding factor on
the optimal inspection interval. Last but not least, we introduce a proof framework based on
likelihood ratio ordering and associated properties. This technique is more generic, and can
be applied more widely to other degradation models (e.g., Gamma processes with random
shape parameters).
The remainder of the paper is organized as follows. Section 2 reviews the IG process
and identifies important stochastic properties related to CBM. Section 3 formulates the
CBM problem as a Markov decision process, and analyzed the structural information of the
optimal policy in depth. Section 4 presents numerical studies to validate the theoretical
results. It also provides sensitivity analysis to examine the effects of different parameters on
the optimal maintenance decisions. Section 5 concludes the article with discussions. Some
technical details and supplementary information are provided in the Appendix.

2 Inverse Gaussian Process with Heterogeneity

The inverse Gaussian (IG) process is recently proposed as a degradation model because it
has many nice properties and it is flexible in dealing with covariates and random effects
(Ye and Chen 2014). Similar to the Gamma process, the IG process possesses monotone
degradation paths. Mathematically, the IG process {Yt , t ≥ 0} satisfies many important
properties. It has independent increments, i.e., Yt1 − Ys1 is independent of Yt2 − Ys2 for
t2 > s2 ≥ t1 > s1 ≥ 0. In addition, the increment Yt − Ys follows an inverse Gaussian
distribution IG (µ[Λ(t) − Λ(s)], η[Λ(t) − Λ(s)]2 ), where the transformed time scale function

4
Λ(t) is monotone increasing with Λ(0) = 0, and IG(a, b), a, b > 0 denotes the IG distribu-
tion. In addition to Λ(t), the process parameters µ, η > 0 control the degradation rate and
degradation volatility.
In practice, different units often exhibit different degradation patterns due to a number
of reasons, e.g., heterogeneous working conditions (Ye et al. 2013) or variations in the raw
materials. To accurately model and predict the degradation for each individual unit, it is
important to account for such heterogeneity. In this article, we consider an IG process model
with random effects where the degradation rate µ is assumed to be random. Different units
have different realizations of µ, causing heterogeneities in their degradation rates. This model
is called random-drift model throughout the paper. Prior to any degradation observations,
µ−1 is typically assumed to follow a truncated normal distribution T N (ω, κ−2 ), κ > 0 to
exclude negative values, which has probability density function (PDF)

κ · φ[κ(µ−1 − ω)]
f (µ−1 ; ω, κ−2 ) = , µ > 0, (1)
1 − Φ(−κω)

where φ(·) and Φ(·) are PDF and cumulative distribution function (CDF) of the standard
normal distribution, respectively.
Suppose the degradation of a unit is observed at times tj , j = 1, 2, · · · , n with degradation
levels Yj = Ytj . Let Λj = Λ(tj ). Given the observations Yn = [Y1 , Y2 , · · · , Yn ], µ−1 is still
p
truncated normal. It is readily shown that hµ−1 |Yn i ∼ T N (ω̃n , κ̃−2
n ) with κ̃n = ηYn + κ2
and ω̃n = (ηΛn + ωκ2 )/κ̃2n . The updated parameters of the truncated normal distribution
only depend on Yn and Λn . In particular, κ̃n increases as Yn . When Yn is large, κ̃−2 n
becomes small and thus hµ−1 |Yn i tends to degenerate to the true (but unobservable) value.
Compared with the unconditional distribution of µ−1 , the conditional distribution hµ−1 |Yn i
is more accurate to predict the future degradation level. Using the updated distribution
hµ−1 |Yn i, the distribution of the degradation level at time tn+1 can be obtained by first
conditioning on µ and then marginalizing over hµ−1 |Yn i. Following Ye and Chen (2014), we
have
s  2 2
ω̌ κ̌ − ω̃n2 κ̃2n η(Λn+1 − Λn )2

1 − Φ [−ω̌κ̌] κ̃n η(Λn+1 − Λn )2
f (Yn+1 = y|Yn ) = · ·exp − ,
1 − Φ[−ω̃n κ̃n ] κ̌ 2π(y − Yn )3 2 2(y − Yn )
p
where y > Yn , κ̌ = ηy + κ2 and ω̌ = (ηΛn+1 + ωκ2 )/κ̌2 are functions of y. It is noted that
with heterogeneity, {Yt } still has the Markov property as hYn+1 |Yn i only depends on Yn .
Although the conditional distribution of future degradation has explicit density function,
it does not belong to any familiar distribution class. Nevertheless, we can establish some
interesting stochastic orders which are extremely useful in characterizing the IG process with

5
 µ1 ≺LR µ2 µ1 ≺LR µ2
Y Unit 2 Y

y2 Unit 2
Unit 1
Unit 1

y
y1 t2 t2 + ∆ t1 t1 + ∆

t t+∆ Age t Age t

Figure 1: Illustration of the stochastic degradation with

heterogeneity

heterogeneity. To begin with, we first review two important definitions of stochastic orders.

Definition A random variable X is said to be smaller than a random variable Y in the

stochastic order, denoted by X ≺ Y , if P(X ≤ t) ≥ P(Y ≤ t) ∀t.

Definition A random variable X is said to be smaller than a random variable Y in the

likelihood ratio order, denoted by X ≺LR Y , if the ratio of their respective PDFs fY (t)/fX (t)
is non-decreasing in t over the unions of the supports of X and Y .

From the definition, we can find that the stochastic order between two random variables
is solely determined by their CDFs. Therefore, we also use FX (·) ≺ FY (·) to denote the
relationship X ≺ Y , where FX (·), FY (·) are the CDF of X and Y respectively. Some useful
properties regarding the stochastic orders are placed in Appendix A-I for reference.
The stochastic orders provide a concise framework in studying the degradation behaviors
when heterogeneity exists. Because the degradation rate µ is unknown for any particular
unit, it needs to be estimated based on historical degradation data. Intuitively, at age t, a
unit with a higher degradation level (larger Yt ) means it is likely to degrade faster in the
stochastic sense. As a result, it is also expected to have larger degradation increments in the
future. In contrast, if a unit reaches a given degradation level y at a later stage, it is likely
to degrade slower in the stochastic sense, and we expect its future degradation increments
to be smaller. This intuition can be graphically illustrated in Figure 1, and formally stated
in Lemma 1 and Lemma 2.

Lemma 1 hYt+∆ |Yt i is stochastically non-decreasing in Yt , i.e., hYt+∆ |Yt = y1 i ≺ hYt+∆ |Yt =
y2 i provided y1 ≤ y2 .

6
Lemma 2 hYt+∆ |Yt i is stochastically non-increasing in t, i.e., hYt1 +∆1 |Yt1 = yi ≺ hYt2 +∆2 |Yt2 =
yi provided t1 > t2 and Λ(t1 + ∆1 ) − Λ(t1 ) ≤ Λ(t2 + ∆2 ) − Λ(t2 ).

Their proofs can be sketched as follows. We first establish the likelihood ratio order hYt+∆ −
Yt |µ1 i ≺LR hYt+∆ −Yt |µ2 i when µ1 < µ2 . We further showed that hµ|Yt = y1 i ≺LR hµ|Yt = y2 i
when y1 < y2 . Combining two orders and using Proposition A.2, Lemma 1 can be proved.
Similar proof can be constructed for Lemma 2. The details are provided in Appendix A-II
and A-III respectively.

Remark The stochastic orderings of the stochastic process Y (t) in terms of age and degra-
dation level are essential to ensure the structural properties of the replacement policy devel-
oped subsequently. Therefore, we expect that the proposed framework is applicable to other
stochastic processes with similar properties, such as Wiener process with random drift rates
(Wang 2010), Gamma process with random shape parameters (Lawless and Crowder 2004).
Even though details of the proofs might vary according to the processes under study, similar
statements to Lemma 1 and Lemma 2 are expected to hold.

3 Condition-based Maintenance for Heterogeneous Degra-

dation
3.1 Problem formulation
In this section, we consider the CBM policy when the periodic inspection is employed to
check the unit’s degradation. The degradation follows the IG process with random drift
parameter µ. For notation simplicity, we focus on the stationary IG process with Λ(t) = t,
and we use “replacement” and “maintenance” interchangeably throughout the paper.
The unit is considered failed when its cumulative deterioration exceeds the failure thresh-
old D. After failure, the unit works non-properly and the failure can only be revealed through
inspections. However, before the failure is detected, downtime costs are incurred to repre-
sent the efficiency or quality loss. Inspections are equally spaced in the planning horizon.
Upon each inspection, if the unit is already failed, corrective maintenance (CM) is per-
formed immediately; otherwise, the decision maker decides whether to conduct a preventive
maintenance (PM) or not based on existing degradations. We assume that all maintenance
actions are instantaneous, which can be justified in the cases when the maintenance time is
relatively short comparing to the inspection intervals. We also assume that both corrective
and preventive maintenance are perfect, restoring the unit to a as-good-as-new state. Let ci

7
denote the inspection cost, cd denote the downtime cost per time unit, cp denote the preven-
tive maintenance cost, and cf denote the corrective maintenance cost (cf > cp ). The cost
incurred at time t has a discounting factor exp(−rt) where r is a fixed and known constant.
We want to find the optimal inspection interval δ and corresponding maintenance policy to
minimize the total operation costs.
According to the model assumptions, the inspections are carried out at equally spaced
times {0, δ, 2δ, · · · , kδ, · · · }. We denote τk as the unit age, which is the time since last
maintenance, at kth inspection; Yτk as the degradation level at age τk . Based on the prop-
erties of IG process, (τk , Yτk ) forms a discrete time continuous state Markov chain with
age-dependent transition probability. As a consequence, the maintenance policy can be for-
mulated as a Markov decision process (MDP). In more details, at kth decision epoch (inspec-
tion) k = 0, 1, · · · , the unit age and its degradation is revealed, denoted by (τk , Yτk ). They
form the state of the MDP. At each state, one of the three actions from {PM,CM,NULL}
can be chosen, where NULL denotes no action. If Yτk is larger than the failure threshold D,
CM needs to be conducted immediately with cost cf . The unit then restores to the as-good-
as-new state (0, 0) because both the age and degradation level becomes 0 afterwards. On
the other hand, if Yτk ≤ D and the action PM is chosen, PM is conducted immediately with
cost cp and the unit returns to the state (0, 0) as well. However, if action NULL is chosen,
no maintenance action is taken and the decision is deferred to the next inspection. Before
that, the unit may fail with unknown failure time Tk between τk and τk + δ. According to the
model assumption, the discounted downtime cost (relative to kδ) before (k + 1)th inspection
can be calculated as
τk +δ
cd (e−r(Tk −τk ) − e−rδ) )
Z
ρ(Tk ) = cd exp[−r(t − τk )]dt = , τk ≤ Tk ≤ τk + δ, (2)
Tk r

and ρ(Tk ) = 0 otherwise. Since the exact failure time is never known, we can define the
expected downtime cost based on current state instead, i.e., Wδ (τk , Yτk ) = E[ρ(Tk )|τk , Yτk ].
At each decision epoch k, we define the value function Vk,δ (u, v) for all u = 0, δ, 2δ, · · · ,
and v ∈ R+ . The value function Vk,δ (u, v) represents the minimum total discounted cost
starting from epoch k until the end of the horizon with initial state (u, v). In this paper, we
focus on the infinite horizon decision problem. As a result, the Vk,δ (u, v) does not depend on
k starting from the same state (u, v). We can therefore drop the subscript k in the remainder.
By definition, the value function should satisfy the Bellman equation (Puterman 2009)
(
min e−rδ Uδ (u, v) + Wδ (u, v), cp + Vδ (0, 0) , v ≤ D

Vδ (u, v) = , (3)
cf + Vδ (0, 0), v>D

8
where Uδ (u, v) = E[Vδ (u + δ, Yu+δ )|τk = u, Yτk = v] is the expected value function with one
period transition from the state (u, v). Instead of including the inspection cost ci in the
value function, we choose to consider them as a separate cost component to make the trade
off on inspection interval δ explicit. In fact, with inspection interval δ, the total discounted
inspection cost can be easily computed
∞
X ci
S(δ) = ci exp(−rkδ) = . (4)
k=0
1 − exp(−rδ)

It becomes clear that S(δ), the total inspection cost, is decreasing as δ increases. On the
other hand, Vδ (0, 0) represents the expected total maintenance cost, including CM, PM, and
downtime cost, starting from a new unit installation. Putting both costs together, we want
to find the optimal inspection interval δ and maintenance policy to minimize S(δ) + Vδ (0, 0).
Different from Elwany et al. (2011), we consider the inspection interval as a decision
variable. In addition, we account for the downtime cost due to improper operation in the
failure state, which is more natural when the failure could not be self reported. We want to
emphasize that the downtime cost is essential. If it is not considered, we can expect that
the optimal δ is ∞, a trivial and non-interesting solution. In the next, we take a two-step
approach to solve the optimization problem. We first fix δ and find the optimal maintenance
policy and corresponding Vδ (0, 0), then we find the δ that can minimize the total cost.

3.2 Structural maintenance policy with fixed inspection

In this part, we fix the inspection interval to be δ, and solve the MDP problem (3) to obtain
the optimal maintenance policy. For notation simplicity, we omit the subscript δ in this
subsection when there is no confusion.
Based on the results in Section 2, we know that even when Λ(t) = t, the heterogeneity and
uncertain drift parameter µ cause different stochastic orders in hYτ +δ |Yτ i. These stochastic
orders are crucial to help us analyze the property of different cost functions. We summarize
the key results as follows.

Lemma 3 The expected downtime cost W (u, v) is a non-decreasing function of v, and non-
increasing function of u.

Lemma 3 is intuitive. Since the downtime cost is only incurred if a failure occurs before next
inspection, its cost is determined by the conditional failure time distribution given current
state. When the current degradation level is larger, the unit is likely to fail at an earlier
time before next inspection, and thus incurs higher downtime cost. On the other hand, if the

9
unit reaches the same degradation level at a later stage, it degrades slower and is expected
to have a later failure time (if any). As a result, W (u, v) is increases with v while decreases
with u. The detailed proof to Lemma 3 is provided in Appendix A-IV.
Similarly, we can establish the monotone property of the cost function U (u, v). Remem-
ber U (u, v) = E[V (u + δ, Yu+δ )|τk = u, Yτk = v] and hYu+δ |Yu i is stochastically increasing
in Yu . Therefore, if V (u, v) is non-decreasing in v, we can conclude U (u, v) is also non-
decreasing in v according to Proposition A.1. The monotone relation with respect to u can
be established as well. The results are formally stated in Lemma 4 with proof provided in
Appendix A-V.

Lemma 4 The optimal value function V (u, v) is non-decreasing in v, and non-increasing

in u.

Lemma 4 enables us to explore the structure of the optimal maintenance policy. Because
of the monotone property of the value function, the optimal maintenance policy can be shown
to have a simplistic structure. This structural information not only provides insights on the
maintenance policies, but also reduces the computation time in numerical solutions.

Theorem 1 Given δ, the optimal maintenance policy that minimizes V (0, 0) is a monotone
control limit policy. Equivalently, there exists a non-decreasing sequence {ξk , k = 0, 1, · · · }
such that the optimal action at state (kδ, y) is PM if y > ξk , and NULL otherwise.

Proof The optimal action can be determined through the value functions. In particular,
when y ≤ D, the optimal action in state (kδ, y) is
(
PM, e−rδ U (kδ, y) + W (kδ, y) > cp + V (0, 0)
a(kδ, y) = .
NULL, Otherwise

If there exists ξk ≤ D such that e−rδ U (kδ, ξk ) + W (kδ, ξk ) > cp + V (0, 0), then ∀y ≥ ξk we
have e−rδ U (kδ, y) + W (kδ, y) ≥ cp + V (0, 0). This is because both U (kδ, y) and W (kδ, y) are
non-decreasing in y and cp + V (0, 0) is a constant. This result establishes that the policy is
a control limit policy.
In addition, if there exists j such that the e−rδ U (jδ, y) + W (jδ, y) ≥ cp + V (0, 0), then
for all k ≤ j, the inequality holds as well because U (kδ, y) and W (kδ, y) are non-increasing
in k. In other words, we should have ξk ≤ ξj ∀k ≤ j, which means {ξk , k = 0, 1, · · · } is a
non-decreasing sequence.

Theorem 1 shows some unique features when the heterogeneity is considered. In par-
ticular, we note that the control limits for PM is monotonically increasing in the unit ages.

10
This can be interpreted as an early screening strategy. In more details, a unit with high
degradation rate should be identified and replaced early to reduce the failure risk. Because
of the stochastic evolution, the unit with given rate µ might cross the threshold ξ1 at age t1
or cross the threshold ξ2 at t2 when ξ1 < ξ2 , t1 < t2 . To screen units with higher degradation
rates, it becomes natural to use non-decreasing ξk for different ages. In addition, the uncer-
tainty in hµ|Yt i reduces as t and Yt increase, which allows the maintenance policy to be less
conservative (or equivalently higher limit). As a comparison, when there is no heterogeneity
and the degradation rates of all units are the same, the control limits become constant over
age τ .

Corollary 1 With periodic inspection at kδ for k = 0, 1, · · · , if κ = ∞, which means no

heterogeneity in the population, the optimal maintenance policy is a constant control limit
policy ξk = ξ0 , ∀k.

The proof of Corollary 1 is provided in Appendix A-VII. Intuitively, when κ = ∞, the

degradation rate µ is a fixed constant rather than random. This means that the observed
degradation levels have no effect on µ. Because hYt |µi is a stationary process, the transition
probability in the MDP remains the same at different ages. Consequently, we would expect
a constant control limit at all ages. Corollary 1 can demonstrate the impact of heterogeneity
from a different perspective. When κ = ∞, the degradation rate µ is a constant. In this case,
when µ is larger, we need a lower (constant) control limit to balance the failure cost and
preventive maintenance cost. In addition, the optimal inspection interval tends to be smaller
to be more conservative. This trend is well demonstrated in Figure 2. It clearly shows that
even when the degradation rate is observable (no heterogeneity), the control limit reveals
monotone trend in terms of 1/µ. Not surprisingly, when the degradation rate is unobservable
(with heterogeneity), we expect a monotone control limit in age as a result of estimating µ
from the observations and finding the optimal control limit based on hµ|Yt i simultaneously.

Remark At the beginning of the section, we assume Λ(t) = t for notation simplicity. In fact,
all the results can be naturally extended to arbitrary monotone increasing functions Λ(t).
The key is to employ age-dependent inspection intervals δτ such that Λ(τk + δτk ) − Λ(τk ) ≤
Λ(τj + δτj ) − Λ(τj ), ∀τk ≥ τj . If the inspection interval is constant, Λ(t) can be any concave
functions. On the other hand, for an arbitrary Λ(t), we can carefully choose the inspection
times τk so that {Λ(τk )} is a sequence of equally spaced points. In this case, Λ(τk +δτk )−Λ(τk )
is a constant for all k. This design is also intuitively appealing. If the units degrades at an
increasing speed, it is natural to inspect more frequently to reduce the risk of failures. Similar
dynamic inspection schedule has also been used in (e.g. Grall et al. 2002b). For readers of

11
 2 9

Optimal Replacement Threshold

Optimal Inspection Interval
1.5 8

1 7

0.5 6
0.5 1 1.5 2
1/µ

Figure 2: Optimal control limit and optimal inspection in-

terval for homogeneous degradation process as a function of
1/µ.

interests, We provide a simplified proof to the Theorem 1 with concave Λ(t) and periodic
inspections in Appendix A-VIII. We also include a numerical example for demonstration.

3.3 Optimal inspection interval

To minimize the total cost, including inspection cost S(δ) and maintenance cost Vδ (0, 0), we
can search the optimal inspection

δ ∗ = arg min S(δ) + Vδ (0, 0), (5)

δ

which is only a one dimensional optimization problem. For any given δ, Vδ (0, 0) can be found
using value iteration algorithm (Appendix A-VI) or policy iteration algorithm. Derivative
free search methods can be used to find the optimal δ ∗ .
It is easy to show that S(δ) is decreasing in δ: less frequent inspections have lower cost.
In addition, we can also show that Vδ (0, 0) is an increasing function in δ. As δ → ∞, only
down time cost will be counted towards the total maintenance cost while S(δ) → 0. As a
result, the objective function approaches a constant

cd e−rT 
  
lim [S(δ) + Vδ (0, 0)] = E Y0 = 0 .
δ→∞ r 

Figure 3 demonstrates that as the inspection interval changes, S(δ) and Vδ (0, 0) change

12
 70
Inspection
60 Maintenance
Total

50

40

Cost 30

20

10

0
0 1 δ* 2 3 4 5
δ

Figure 3: Cost functions at different inspection intervals δ.

accordingly. It shows that while S(δ) approaches to 0 and Vδ (0, 0) approaches to a constant,
their sum has a minimum at δ ∗ , the optimal inspection interval. In addition, the optimal
δ ∗ is also very robust as the total cost is flat in the region near δ ∗ . In other words, a rough
estimate of the optimal inspection would not inflate the total operational cost significantly.

4 Numerical Studies
4.1 Optimal CBM policy
We let the unit degradation follow the IG process with random-drift model with Λ(t) = t
and η = 1. In addition, the degradation rate µ−1 follows a truncated normal distribution (1)
with ω = 1.2, and κ = 4. The unit fails when the degradation reaches the failure threshold
D = 10. In the numerical study, we consider the optimal replacement policy. Suppose the
unit is inspected periodically with inspection cost ci = 0.05. If the degradation level is less
than the failure threshold D yet larger than the control limit, a preventive replacement is
conducted with cost cp = 3. On the other hand, if the unit fails unexpectedly (degradation
exceeding the failure threshold D), a corrective replacement is required with cost cf = 10.
The unit downtime cost is cd = 1, and the cost discount factor is r = 0.01.
Given the parameters of the random-drift model and the cost parameters, the optimal
maintenance policy can be obtained using the value iteration algorithm or the monotone

13
 10
Failure threshold
*
δ =1.3
Replace δ=0.5
9 δ=2.5

Degradation
7

5
Not Replace

4
0 2 4 6 8 10
Age

Figure 4: Optimal maintenance policies when different in-

spection intervals are used.

policy iteration algorithm (Puterman 2009). To make the computation feasible, we discretize
the continues degradation states into fine intervals and use a time horizon that is long enough
to approximate the infinite horizon. This practice is commonly used in similar studies (e.g.,
Elwany et al. 2011). Figure 4 shows the optimal maintenance policies when different δs are
used. The optimal inspection interval is found to be δ ∗ = 1.4. Regardless the δ, Figure
4 clearly shows that the optimal control limits are monotone increasing in age τ . When a
smaller than δ ∗ inspection interval is used, the maintenance policy becomes less conservative
(higher limits). In contrast, when δ > δ ∗ , the control limit is more conservative.
It is also interesting to find how much improvement we can make when we consider the
degradation heterogeneity in the maintenance policy. To compare, we investigate the opti-
mal maintenance policy when the degradation follows IG process without heterogeneity. The
degradation rate parameter is µ0 = 0.876, which is the same as the expected degradation rate
from the heterogeneous degradation. Other model parameters and cost parameters are kept
the same. Figure 5 compares the optimal maintenance policy and inspection when the het-
∗
erogeneity is not considered (denoted by Πhomo , δhomo ) with the policy when the heterogeneity
∗
is considered (denoted by Πhetero , δhetero ). It shows that Πhomo has a constant control limit
independent of the unit’s age, as stated in Corollary 1. However, when the degradation has
population heterogeneity, using this simple control limit may result in sub-optimal results.
For example, if we use Πhomo for maintenance of the heterogeneous population, the expected
total discounted cost is 38.817. In contrast, the Πhetero has expected total discounted cost
36.264.

14
 9

8.5

7.5

Degradation
7

6.5

5.5

5
Πhetero, δ*hetero=1.3
4.5 *
Πhomo, δhomo=1.4
4
0 5 10 15
Age

Figure 5: Optimal maintenance policies and inspection in-

tervals when heterogeneity is and is not considered.

Table 1: Summary of the different model parameters in the

sensitivity analysis

ω κ ci cd cf cp
Base 1.2 4 0.05 1 10 3
Low 0.8 2 0.02 0.5 5 1
High 1.6 6 0.1 2 15 5

4.2 Sensitivity analysis of the optimal policy

We also conduct the sensitivity analysis of the parameters on the optimal CBM policies.
Based on the model studied in Section 4.1, we vary the parameter values and obtain the
corresponding optimal policies. Table 1 summarizes the parameter values used in the sen-
sitivity analysis. The Base row contains the parameter values used in Section 4.1, and the
Low and High rows list the alternative values used in the sensitivity study.
We first investigate how the the optimal policy changes with the different cost parameters.
By changing the cost parameters, we can obtain the optimal maintenance policies as well as
the optimal inspection interval δ ∗ , as summarized in Figure 6. Figure 6a shows that as the
CM cost cf increases, the policy becomes more conservative in terms of shorter inspection
interval and lower control limits. In other words, the preventive replacement becomes more
cost effective. In contrast, Figure 6b demonstrates the opposite. As cp increases, it becomes
less attractive to replace the unit before its failure. Therefore, the control limits increase
monotonically in cp . Nevertheless, the inspection interval still becomes shorter as the overall

15
 9 9

8 8

7 7
Degradation

Degradation
6 6

5 5

c =5.00, δ*=1.7 c =1, δ*=1.7

f p
*
4 cf=10.00, δ =1.3 4 cp=3, δ*=1.3
c =15.00, δ*=1.1 c =5, δ*=1.1
f p
3 3
0 5 10 15 0 2 4 6 8 10
Age Age

(a) CM Cost cf (b) PM Cost cp

9 9

8 8

7 7
Degradation

Degradation

6 6

5 5
* *
cd=0.5, δ =1.3 ci=0.02, δ =0.5
* *
4 cd=1.0, δ =1.3 4 ci=0.05, δ =1.3
cd=2.0, δ*=1.2 ci=0.10, δ*=2.0
3 3
0 2 4 6 8 10 0 2 4 6 8 10
Age Age

(c) Downtime Cost cd (d) Inspection Cost ci

Figure 6: Sensitivity of optimal policies with respect to dif-

ferent cost parameters

maintenance cost increases. Figure 6c shows that the downtime cost does not have much
impact on the optimal policy. When cd increases, the optimal inspection interval decreases
to reduce the expected downtime cost. Since the optimal policies in this example are far
from the failure threshold D, the downtime costs have minimal impact on the control limits.
Last but not least, Figure 6d shows the effects of inspection cost. Not surprisingly, as ci
increases, the optimal inspection interval becomes longer. To safeguard against the risk of
failures, the control limits also decrease to make the policy more conservative.
In addition to cost parameters, we also investigate the effects of model parameters,
particularly the ones related to the heterogeneity ω, κ. Figure 7 shows the optimal policies
for different values of ω and κ. These two parameters determine the distribution of the drift

16
 9 9

8 8

7 7
Degradation

Degradation
6 6

5 5

* *
ω=0.8, δ =0.8 κ=2.0, δ =1.0
4 4
ω=1.2, δ*=1.3 κ=4.0, δ*=1.3
ω=1.6, δ*=1.8 κ=6.0, δ*=1.4
3 3
0 5 10 15 0 5 10 15
Age Age

(a) Change in ω (b) Change in κ

Figure 7: Sensitivity of optimal policies to different random

effects parameters

rate µ in the degradation model. A larger ω leads to a smaller µ on average, and hence
slower degradation. A larger κ reduce the heterogeneity. Figure 7a clearly indicates that a
larger ω leads to higher control limits at early ages. This is because only limited degradation
observations are available during this period, and the decision relies more heavily on the
prior distribution of ω, κ. A higher ω (a lower µ) means the policy can be less conservative.
As the age increases, more observations are available. The effect of ω reduces in maintenance
decision. As a result, we can observe the control limits “converge” for different ω. rates. On
the other hand, a larger κ indicates a higher precision in µ, and hence a smaller heterogeneity.
Figure 7b shows that as κ increases, the jumps in the control limits ξk become smaller. This
is because κ represents the uncertainty in µ. For a smaller κ, the degradation process is
more age dependent. Therefore, control limits vary more at different ages. In contrast, for a
larger κ or even κ = ∞, the degradation rate is almost certain, depending much less on age
and observations. As a result, the control limits converge to a constant eventually.

5 Conclusion
In this paper, we investigated the optimal CBM policy when the system degradation con-
forms to the IG process with random-drift model. The random drifts are used to represent
heterogeneities commonly observed across the product population. The structure of the op-
timal policy was analytically analyzed. We found that the optimal replacement policy is the
monotone control limit policy. In addition, the optimal inspection interval is chosen to min-

17
imize the total operational cost. Extensive numerical studies are conducted to investigate
the sensitivity of optimal policies to different model parameters and cost parameters.
In the future, we can generalize this method to other maintenance scenarios, such as
imperfect inspections, dynamic inspections (Kim and Makis 2013), etc. We can also consider
imperfect repairs (Wu and Zuo 2010) rather than replacement and maintenance actions that
can lower the degradation rates, which are more realistic in system maintenance. We are
also interested in the maintenance policy of a system consisting of multiple components with
dependent degradation. Moreoever, due to the fast development of sensing technologies, the
inspection interval is not necessarily equal. The inspection schedule can be condition-based
as well, which should be more effective but more complex. Other interesting extensions
include models with competing risks, etc.

Acknowledgement
We would like to thank the editor, associate editor, and two anonymous referees for their
constructive comments and suggestions that have considerably im- proved the paper. Nan
Chen and Zhisheng Ye are partially supported by the National Research Foundation Singa-
pore under its Campus for Research Excellence and Technological Enterprise (CREATE).
Yisha Xiang is partially supported by National Natural Science Foundation of China under
Grant 71301171.

Appendix
A-I Useful propositions
The following results are derived from Shaked and Shanthikumar (2007).

Proposition A.1 If φ(·) is a monotone non-decreasing function, and X ≺ Y , then E[φ(X)] ≤

E[φ(Y )] provided the expectations exist.

Proposition A.2 Consider a family of distributions {Gθ (x)} indexed by θ, if Gθ (x) ≺

Gθ0 (x) whenever θ ≤ θ0 , and if F1 (θ) ≺ F2 (θ) are two distributions of θ, then
Z Z
Gθ (x)dF1 (θ) ≺ Gθ (x)dF2 (θ).

Proposition A.3 If two random variables satisfy X ≺LR Y , then X ≺ Y .

18
Proposition A.4 If X ≺LR Y , and φ(·) is any non-decreasing (non-increasing) function,
then φ(X) ≺LR (LR )φ(Y ).

A-II Proof of Lemma 1

We first consider a family of IG distributed random variables Y (µ) ∼ IG(µλ, ηλ2 ) with PDF
f (y|µ). For any 0 < µ1 < µ2 , we can get the ratio of their densities as
(s ),(s 2 )
η(y − µ2 λ)2 η(y − µ1 λ)2
 
f (y|µ2 ) ηλ2 ηλ
= exp − exp −
f (y|µ1 ) 2πy 3 2µ22 y 2πy 3 2µ21 y
    
η y y 1 1
= exp − − ηλ − . (A.1)
2 µ21 µ22 µ1 µ2

Since µ1 < µ2 , µ−2 −2

1 > µ2 . The right hand side (RHS) of (A.1) is an increasing function of
y when y ≥ 0. This indicates that Y (µ1 ) ≺LR Y (µ2 ). Because of the independent increment
property conditional on µ, we can establish that

S1. hYt+∆ − Yt |µ1 i ≺LR hYt+∆ − Yt |µ2 i when µ1 < µ2 .

In the second step, we consider the conditional distribution of µ given the latest observation
Yt . According to Section 2, the conditional distribution of hµ−1 |Yt = ys i s = 1, 2 follows a
truncated normal distribution. Therefore, when y1 < y2 the likelihood ratio of hµ−1 |Yt = y1 i
and hµ−1 |Yt = y2 i becomes

1 − Φ(−κ̃t,2 ω̃t,2 ) κ̃t,1 exp −κ̃2t,1 (µ−1 − ω̃t,1 )2 /2

 
f (µ−1 |Yt = y1 )
=  
f (µ−1 |Yt = y2 ) 1 − Φ(−κ̃t,1 ω̃t,1 ) κ̃t,2 exp −κ̃2t,2 (µ−1 − ω̃t,2 )2 /2
 2
κ̃t,2 − κ̃2t,1 −2

2 2 −1
= M · exp µ − (κ̃t,2 ω̃t,2 − κ̃t,1 ω̃t,1 )µ (A.2)
2
p
where κ̃t,s = ηys + κ2 and ω̃t,s = (ηΛt + ωκ2 )/κ2s , s = 1, 2, M is a positive constant not
depending on µ. Clearly, κ̃2t,1 ω̃t,1 = ηΛt + ωκ2 = κ̃2t,2 ω̃t,2 because both quantities do not
depend on Yt . Since κ̃2t,1 < κ̃2t,2 , the RHS of (A.2) is an increasing function of µ−1 . Therefore,
hµ−1 |Yt = y2 i ≺LR hµ−1 |Yt = y1 i. In addition, g(x) = x−1 , x > 0 is a decreasing function, we
can conclude

S2. hµ|Yt = y1 i ≺LR hµ|Yt = y2 i when y1 < y2

by Proposition A.4.

19
 In the third step, we show that the degradation increment conditional on Yt has stochastic
order. In more details, by integrating out µ, we have
Z
f (Yt+∆ − Yt = y|Yt ) = f (Yt+∆ − Yt = y|µ) · dF (µ|Yt ).
µ

Combining S1 and S2 and applying Proposition A.3 and Proposition A.2, we can conclude
that hYt+∆ − Yt |Yt = y1 i ≺ hYt+∆ − Yt |Yt = y2 i if y1 < y2 . Finally we note that

hYt+∆ |Yt = y1 i ≡ hYt+∆ − Yt |Yt = y1 i + y1 ≺ hYt+∆ − Yt |Yt = y1 i + y2

≺ hYt+∆ − Yt |Yt = y2 i + y2 ≡ hYk+1 |Yk = y2 i,

which concludes the proof.

A-III Proof of Lemma 2

By definition of the IG process, the degradation increment conditioning on µ follows IG
distribution. We denote λ1 = Λ(t1 + ∆1 ) − Λ(t1 ) and λ2 = Λ(t2 + ∆2 ) − Λ(t2 ). Then
hYt1 +∆1 − Yt1 |µ1 i follows the IG distribution IG(µ1 λ1 , ηλ21 ). Similarly, we have hYt2 +∆2 −
Yt2 |µ2 i ∼ IG(µ2 λ2 , ηλ22 ). As a result, their likelihood ratio becomes
(r ),(r 2 )
ηλ22 η(x − µ2 λ2 )2 η(x − µ1 λ1 )2
 
f (Yt2 +∆2 − Yt2 = x|µ2 ) ηλ1
= exp − exp −
f (Yt1 +∆1 − Yt1 = x|µ1 ) 2πx3 2µ22 x 2πx3 2µ21 x
   
η 1 1 η 2 2 1


= M · exp − − x− λ − λ1 , (A.3)
2 µ22 µ21 2 2 x

where again M is a positive multiple constant not depending on x. Since λ1 ≤ λ2 , and

µ1 < µ2 , (A.3) is a non-decreasing function of x. Therefore we can establish the likelihood
ratio ordering

S3. hYt1 +∆1 − Yt1 |µ1 i ≺LR hYt2 +∆2 − Yt2 |µ2 i when µ1 < µ2 .

In addition, similar to the proof of Lemma 1, we can compare the conditional distribution
of hµ−1 |Yt i:

f (µ−1 |Yt1 = y)
 2
κ̃2 − κ̃21 −2

2 2 −1
= M · exp µ − (κ̃2 ω̃2 − κ̃1 ω̃1 )µ , (A.4)
f (µ−1 |Yt2 = y) 2

where κ̃21 = κ̃22 = ηy + κ2 , κ̃22 ω̃2 = ηΛ2 + ωκ2 < ηΛ1 + ωκ2 = κ̃21 ω̃1 . Therefore (A.4) is
an increasing function of µ−1 . In other words, hµ−1 |Yt2 = yi ≺LR hµ−1 |Yt1 = yi if t1 > t2 .
Equivalently, we have

20
 S4. hµ|Yt1 = yi ≺LR hµ|Yt2 = yi when t1 > t2 .

By the definition of stochastic orders, we find

Z
P(Yt1 +∆1 − Yt1 ≤ x|Yt1 = y) = P(Yt1 +∆1 − Yt1 ≤ x|µ)dF (µ|Yt1 = y)
Z
≥ P(Yt2 +∆2 − Yt2 ≤ x|µ)dF (µ|Yt1 = y)
Z
≥ P(Yt2 +∆2 − Yt2 ≤ x|µ)dF (µ|Yt2 = y)
= P(Yt2 +∆2 − Yt2 ≤ x|Yt2 = y).

In other words, hYt1 +∆1 − Yt1 |Yt1 i ≺ hYt2 +∆2 − Yt2 |Yt2 i. The first inequality holds because S4.
The second inequality holds because S4 and Proposition A.2. Adding constant y on both
sides, we finally get hYt1 +∆1 |Yt1 = yi ≺ hYt2 +∆2 |Yt2 = yi, which concludes the proof.

A-IV Proof of Lemma 3

According to (2), given the failure time, the downtime cost is analytically available. Since
the exact failure is unobservable, we need to find its distribution first. Conditioning on the
state (τk , Yτk ) and the drift rate µ, the failure time Tk has CDF

P(Tk ≤ t|Yτk , µ) = P(Yt > D|Yτk , µ).

As a result, the stochastic ordering of hTk |Yτk , µi can be analyzed through stochastic ordering
of hYt |Yτk , µi. According to S1, we have hYt |Yτk , µ1 i ≺LR hYt |Yτk , µ2 i when µ1 < µ2 . Therefore

P(Tk ≤ t|Yτk , µ1 ) = P(Yt > D|Yτk , µ1 ) = 1 − P(Yt ≤ D|Yτk , µ1 )

≤ 1 − P(Yt ≤ D|Yτk , µ2 ) = P(Tk ≤ t|Yτk , µ2 ). (A.5)

Since the (A.5) is valid for all t ≥ τk , by definition we can conclude that

S5. hTk |Yτk , µ2 i ≺LR hTk |Yτk , µ1 i when µ1 < µ2 .

In addition, when the drift rate µ is the same, but degradation starts at two different levels
y1 < y2 , we have hYt |Yτk = y1 , µi ≺LR hYt |Yτk = y2 , µi when y1 < y2 . Following the similar
arguments in (A.5), we can obtain

S6. hTk |Yτk = y2 , µi ≺LR hTk |Yτk = y1 , µi when y1 < y2 .

21
From (2), we know ρ(Tk ) is a decreasing function of Tk . According to S5, S6, and Proposition
A.1,
E[ρ(Tk )|Yτk = y1 , µ1 ] ≤ E[ρ(Tk )|Yτk = y2 , µ1 ] ≤ E[ρ(Tk )|Yτk = y2 , µ2 ].

Moreover,
Z
W (τk , y1 ) ≡ E[ρ(Tk )|Yτk = y1 , µ]dF (µ|Yτk = y1 )
Z
≤ E[ρ(Tk )|Yτk = y2 , µ]dF (µ|Yτk = y1 )
Z
≤ E[ρ(Tk )|Yτk = y2 , µ]dF (µ|Yτk = y2 ) ≡ W (τk , y2 ). (A.6)

The second inequality holds because of S2, which was proved in Appendix A-II. This shows
that the downtime cost is non-decreasing in degradation level.
To prove the monotonicity in τk , we note that conditioning on µ, the degradation is
stationary. Therefore, hTk − τk |Yτk = y, µ2 i ≺LR hTj − τj |Yτj = y, µ1 i if µ1 < µ2 ∀k, j. This
leads to E[ρ(Tj )|Yτj = y, µ1 ] ≤ E[ρ(Tk )|Yτk = y, µ2 ]. We also notice that ρ(Tk ) only depends
on the Tk − τk , and S4 holds (shown in Appendix A-III). Therefore, for τk ≤ τl
Z
W (τk , y) ≡ E[ρ(Tk )|Yτk = y, µ]dF (µ|Yτk = y)
Z
= E[ρ(Tj )|Yτj = y, µ]dF (µ|Yτk = y)
Z
≥ E[ρ(Tj )|Yτj = y, µ]dF (µ|Yτj = y) ≡ W (τj , y). (A.7)

Collectively, (A.6) and (A.7) shows that the downtime cost W (u, v) is non-decreasing in v
and non-increasing in u.

A-V Proof of Lemma 4

We prove the lemma through the mathematical induction using the value iteration algorithm
depicted in Appendix A-VI. We denote the superscript s as the iteration number. Then,
V s (u, v), u = kδ k = 0, 1, 2, · · · , v ∈ R+ , is the value function at the sth iteration. When
s = 0, we set the initial value V 0 (u, v) = 0, ∀u, v, which is non-decreasing in v, and non-
increasing in u. Assume Lemma 4 holds at the sth iteration. Then at the (s + 1)th iteration,
when v ≤ D
V s+1 (τ, y) = min e−rδ U s (u, v) + W (u, v), cp + V s (0, 0) .

(A.8)

22
Since V s (u, v) is non-decreasing in v, and hYu+δ |Yu = yi is stochastically non-decreasing in
y (Lemma 1), U s (u, v) = E[V s (u + δ, Yu+δ )|u, v] is also non-decreasing in v. In the same
way, since V s (u, v) is non-increasing in u, and hYu+δ |Yu = yi is stochastically decreasing in
u (Lemma 2), we have

U s (u1 , v) ≡ E[V s (u1 + δ, Y )|u1 , v] ≥ E[V s (u2 + δ, Y )|u1 , v]

≥ E[V s (u2 + δ, Y )|u2 , v] ≡ U s (u2 , v) u1 < u2 . (A.9)

In summary, when v < D, both U s (u, v) and W (u, v) are non-decreasing in v and non-
increasing in u. Since cp + V s (0, 0) is a constant, each component in the RHS of (A.8) satisfy
the monotone property in u and v. Consequently, V s+1 (u, v) is non-decreasing in v and
non-increasing in u. On the other hand, when v > D, V s+1 (u, v) = cf + V s (0, 0), a constant.
Hence the induction hypothesis holds ∀u, v.
Using the mathematical induction, we have shown that V s (u, v) has the stated monotone
property for all iteration s. As s increases to infinity, V s (u, v) will converge to V (u, v) (Put-
erman 2009). The monotone property then naturally extends to V (u, v), which completes
the proof.

A-VI Value iteration algorithm

Data: Model parameters ω, κ, η, Λ(t), cost parameters cd , cf , cp , inspection interval δ
Result: Optimal value function V (τk , Yτk )
begin
Initialize: V 0 (τk , Yτk ) ←− 0;
Compute: P(Yτk+1 |Yτk , τk ) and W (τk , Yτk ) for all τk , Yτk ;
while |V s (τk , Yτk ) − V s−1 (τk , Yτk )| <  do
if Yτk > D then
V s+1 (τk , Yτk ) = cf + V s (0, 0);
else
V s+1 (τk , Yτk ) = e−rδ V s (τk+1 , Yτk+1 ) · P(Yτk+1 |Yτk , τk ) + W (τk , Yτk );
P

if V s+1 (τk , Yτk ) > cp + V s (0, 0) then

V s+1 (τk , Yτk ) = cp + V s (0, 0);
end
end
end
end

23
A-VII Proof of Corollary 1
If κ = ∞ for the distribution of µ, it indicates that no heterogeneity exists. The degradation
model reduces to the simple IG process with fixed parameters. From the proof of Lemma 1,
we can find that hYt+∆ −Yt |Yt = y1 i is identical in distribution to hYt+∆ −Yt |Yt = y2 i regardless
of the values y1 and y2 . As a result, we can easily show that hYt+∆ |Yt = y1 i ≺ hYt+∆ |Yt = y2 i
if y1 ≤ y2 because the increment has the same distribution and Yt+∆ therefore only depends
on the initial value Yt . Similarly, from the proof of Lemma 2, we can find that hµ|Yt1 = yi
is identical in distribution to hµ|Yt2 = yi. Further, because Λ(t) = t, hYt1 +∆ |Yt1 = yi is also
identical in distribution to hYt2 +∆ |Yt2 = yi.
Using the mathematical induction, we can show that the value function V (u, v) is con-
stant in u and non-decreasing in v. As a result, using the similar argument as in the proof
of Theorem 1, we can prove that the optimal policy is a control limit policy, and ξk is a
constant for all k.

A-VIII Proof to Theorem 1 (general case)

Theorem A.1 If the degradation Y (t) follows the IG process with concave shape function
Λ(t) and random drift parameter µ with distribution µ−1 ∼ T N (ω, κ−2 ), and the inspection
has constant interval δ, then the optimal maintenance policy that minimizes V (0, 0) is a
monotone control limit policy. Equivalently, there exists a non-decreasing sequence {ξk , k =
0, 1, · · · } such that the optimal action at state (kδ, y) is PM if y > ξk , and NULL otherwise.

Proof For concave shape function Λ(t) and constant inspection interval δ, it is easy to check
that Λ((k2 + 1)δ) − Λ(k2 δ) ≤ Λ((k1 + 1)δ) − Λ(k1 δ) for integers 0 ≤ k1 ≤ k2 . As a results,
conditions for both Lemma 1 and Lemma 2 are satisfied.
In examining the proof of Lemma 3, we can observe that the first part (i.e., W (u, v)
is non-decreasing in v) does not depend on Λ(t). Therefore, it holds without any change.
We only need to show the monotonicity in u. We note that conditioning on µ, we have
hYτk +∆ |Yτk , µ1 i ≺LR hYτj +∆ |Yτj , µ2 i for any τk ≤ τj , µ1 ≥ µ2 and ∆ > 0. Therefore

P(Tk − τk ≤ t|Yτk = y, µ1 ) = P(Yτk +t > D|Yτk = y, µ1 ) = 1 − P(Yτk +t ≤ D|Yτk = y, µ1 )

≤ 1 − P(Yτj +t ≤ D|Yτj = y, µ2 ) = P(Tj − τj ≤ t|Yτk , µ2 ).

By definition, hTk − τk |Yτk = y, µ1 i ≺ hTj − τj |Yτj = y, µ2 i for τk ≤ τj , µ1 ≥ µ2 . This leads to

E[ρ(Tj )|Yτj = y, µ2 ] ≤ E[ρ(Tk )|Yτk = y, µ1 ]. The rest follows the lines of proof in Appendix
A-IV. Similarly, Lemma 4 can be proved using induction without reference to the form of
Λ(t). Therefore it holds without changes. Following the arguments in proving Theorem 1, we

24
 10

Degradation
6

2
0 2 4 6 8 10
Age

Figure A.1: √The optimal control limits (black solid line)

when Λ(x) = 10x (red dashed line) with the same random
effects parameter as in Section 4.1. The optimal inspection
interval is 0.6 compared with 1.3 in the case of Λ(x) = x.

can see that this theorem under slightly more general condition holds as well. To conclude,
√
we use a numerical example to demonstrate the optimal policy when Λ(x) = 10x, as shown
in Figure A.1.

References
Albin, S. L. and Chao, S. (1992), “Preventive Replacement in Systems with Dependent
Components,” IEEE Transactions on Reliability, 41, 230–238.

Bloch-Mercier, S. (2002), “A Preventive Maintenance Policy with Sequential Checking Pro-

cedure for a Markov Deteriorating System,” European Journal of Operational Research,
142, 548–576.

Chen, C.-T., Chen, Y.-W., and Yuan, J. (2003), “On a Dynamic Preventive Maintenance
Policy for a System under Inspection,” Reliability Engineering & System Safety, 80, 41–47.

Chen, N. and Tsui, K. L. (2013), “Condition Monitoring and Residual Life Prediction Using
Degradation Signals: Revisited,” IIE Transactions, 45, 939–952.

Dieulle, L., Berenguer, C., Grall, A., and Roussignol, M. (2003), “Sequential Condition-

25
 Based Maintenance Scheduling for a Deteriorating System,” European Journal of Opera-
tional Research, 150, 451–461.

Elwany, A. H., Gebraeel, N. Z., and Maillart, L. M. (2011), “Structured Replacement Policies
for Components with Complex Degradation Processes and Dedicated Sensors,” Operations
Research, 59, 684–695.

Gebraeel, N., Lawley, M., Li, R., and Ryan, J. (2005), “Residual-Life Distributions from
Component Degradation Signals: A Bayesian Approach,” IIE Transactions, 37, 543–557.

Grall, A., Berenguer, C., and Dieulle, L. (2002a), “A Condition-Based Maintenance Policy
for Stochastically Deteriorating Systems,” Reliability Engineering & System Safty, 76,
167–180.

Grall, A., Dieulle, L., Berenguer, C., and Roussignol, M. (2002b), “Continuous-Time
Predictive-Maintenance Scheduling for a Deteriorating System,” IEEE Transactions on
Reliability, 51, 141–150.

Guo, C., Wang, W., Guo, B., and Si, X. (2013), “A Maintenance Optimization Model
for Mission-Oriented Systems based on Wiener Degradation,” Reliability Engineering &
System Safety, 111, 183–194.

Hontelez, J. A. M., Burger, H. H., and Wijnmalen, D. J. D. (1996), “Optimum Condition-

Based Maintenance Policies for Deteriorating Systems with Partial Information,” Relia-
bility Engineering & System Safety, 51, 267–274.

Jardine, A. K. S., Lin, D., and Banjevic, D. (2006), “A Review on Machinery Diagnostics
and Prognostics Implementing Condition-Based Maintenance,” Mechanical Systems and
Signal Processing, 20, 1483–1510.

Kim, M. J. and Makis, V. (2013), “Joint Optimization of Sampling and Control of Partially
Observable Failing Systems,” Operations Research, 61, 777–790.

Lawless, J. and Crowder, M. (2004), “Covariates and Random Effects in a Gamma Process
Model with Application to Degradation and Failure,” Lifetime Data Analysis, 10, 213–227.

Liao, H., Elsayed, E. A., and Chan, L.-Y. (2006), “Maintenance of Continuously Monitored
Degrading Systems,” European Journal of Operational Research, 175, 821–835.

Liao, H. and Tian, Z. (2013), “A Framework for Predicting the Remaining Useful Life of a
Single Unit under Time-Varying Operating Conditions,” IIE Transactions, 45, 964–980.

26
Liu, Z., Ma, X., Yang, J., and Zhao, Y. (2014), “Reliability Modeling for Systems with Mul-
tiple Degradation Processes Using Inverse Gaussian Process and Copulas,” Mathematical
Problems in Engineering, 2014.

Peng, W., Li, Y.-F., Yang, Y.-J., Huang, H.-Z., and Zuo, M. J. (2014), “Inverse Gaussian
Process Models for Degradation Analysis: A Bayesian Perspective,” Reliability Engineer-
ing & System Safety, 130, 175–189.

Puterman, M. L. (2009), Markov Decision Processes: Discrete Stochastic Dynamic Program-

ming, vol. 414, Hoboken, NJ: Wiley-Interscience.

Shaked, M. and Shanthikumar, J. G. (2007), Stochastic Orders, New York, NY: Springer.

van Noortwijk, J. M. (2009), “A Survey of the Application of Gamma Processes in Mainte-

nance,” Reliability Engineering & System Safety, 94, 2–21.

Wang, H. (2002), “A Survey of Maintenance Policies of Deteriorating Systems,” European

Journal of Operational Research, 139, 469–489.

Wang, X. (2010), “Wiener Processes with Random Effects for Degradation Data,” Journal
of Multivariate Analysis, 101, 340–351.

Wang, X. and Xu, D. (2010), “An Inverse Gaussian Process Model for Degradation Data,”
Technometrics, 52, 188–197.

Wu, S. and Zuo, M. J. (2010), “Linear and Nonlinear Preventive Maintenance Models,”
IEEE Transactions on Reliability, 59, 242–249.

Xiang, Y., Cassady, C. R., and Pohl, E. A. (2012), “Optimal maintenance policies for systems
subject to a Markovian operating environment,” Computers & Industrial Engineering, 62,
190 – 197.

Ye, Z. S., Chen, L. P., Tang, L. C., and Xie, M. (2014), “Accelerated Degradation Test
Planning Using the Inverse Gaussian Process,” IEEE Transactions on Reliability, 63, 750–
763.

Ye, Z. S. and Chen, N. (2014), “The Inverse Gaussian Process as a Degradation Model,”
Technometrics, 56, 302–311.

Ye, Z. S., Hong, Y., and Xie, Y. (2013), “How Do Heterogeneities in Operating Environments
Affect Field Failures Predictions and Test Planning?” The Annals of Applied Statistics,
7, 1837–2457.

27
Ye, Z. S. and Xie, M. (2014), “Stochastic modelling and analysis of degradation for highly
reliable products,” Applied Stochastic Models in Business and Industry, in press.

Yeh, R. H. (1997), “Optimal Inspection and Replacement Policies for Multi-State Deterio-
rating Systems,” European Journal of Operational Research, 96, 248–259.

28