Computer-Aided Civil and Infrastructure Engineering 00 (2011) 1–12

Determining an Optimal Maintenance Period

for Infrastructure Systems
Xueqing Zhang∗ & Hui Gao
Hong Kong University of Science & Technology, Clear Water Bay, Kowloon, Hong Kong, China

Abstract: This article proposes a maintenance policy tion. Maintenance actions can be classified into three
that sets a limit to the cumulative failure rate over the categories according to their restoration effects: per-
life cycle of an infrastructure system. Under this policy, fect, imperfect, and minimal (Wang, 2002). A perfect
there are three maintenance scenarios in an arbitrary pe- maintenance action (i.e., replacement) makes the sys-
riod of the life cycle: a maintenance action will be imple- tem as good as a new one. An imperfect maintenance
mented in scenario 1 in which the failure rate limit has action (e.g., minor rehabilitation) improves the system
not been reached and the system does not fail; a preven- by some degree but does not bring it to a perfect condi-
tive replacement will be implemented to renew the system tion. A minimal maintenance action (e.g., routine main-
in scenario 2 where the failure rate limit has been reached tenance) does not improve the condition of the system.
and the system does not fail; and a corrective replace- One important issue in the maintenance of a system is
ment will be implemented at the time of failure to renew how to model the effect of the maintenance action over
the system in scenario 3 where the system fails no matter the life cycle of the system. Generally, the maintenance
whether the failure rate limit has been reached or not. The effect can be measured in terms of a jump of the con-
maintenance effect is measured by the type 1 effective age dition state (e.g., Labi and Sinha, 2003 for road pave-
model. An optimization model is developed to determine ment maintenance), or a reduction of the failure rate
the optimal length of the maintenance period on the ba- (e.g., Chu and Durango-Cohen, 2008 for road pavement
sis of the proposed maintenance policy, with an objective overlay), or of the effective age (e.g., Martorell et al.,
to minimize the system’s life cycle cost per unit time that 1999 for infrastructure maintenance) of the system after
includes maintenance cost, failure loss, and the cost of a maintenance action is taken.
system unavailability. This optimization model and the Regarding the failure rate reduction model, in Lie
search algorithm that is consequently formulated are ap- and Chun (1986) and Nakagawa (1986), the failure rate
plied to the maintenance of bridge decks. Weibull distri- of the system is reset to zero after the completion of a
bution is used to model the deterioration process of the maintenance action in each maintenance period. In the
bridge deck. planning horizon, the failure rate increases faster in a
following maintenance period than that in a previous
maintenance period. That is, the slope of the failure rate
1 INTRODUCTION becomes steeper and steeper over the life cycle of the
system. In Chan and Shaw (1993), the failure rate is as-
The condition of an infrastructure system (hereinafter sumed to be reduced either by a fixed reduction or by a
referred to as the system) continuously deteriorates and proportional reduction after each maintenance action.
consequently its failure rate increases over time. The The concept of the effective age is introduced by Malik
failure of the system (i.e., the system is not able to per- (1979). Kijima (1989) developed two effective age mod-
form the desired functions) causes economic and so- els. Type 1 model assumes that a maintenance action
cietal loss. To reduce the failure rate of the system, can only reduce the additional age gained from the pre-
maintenance actions are needed periodically to restore vious action. Type 2 model assumes that a maintenance
the condition of the system and mitigate its deteriora- action can reduce the cumulative age gained from the
start point of the life cycle.
∗ To whom correspondence should be addressed. E-mail: zhangxq@ Maintenance optimization models are well-
ust.hk. developed for infrastructure systems in the past


C 2011 Computer-Aided Civil and Infrastructure Engineering.

DOI: 10.1111/j.1467-8667.2011.00739.x
2 Zhang & Gao

decades (Adeli and Ge, 1989; Vaurio, 1999; Sarma systems, and different maintenance effect models are
and Adeli, 1998; Sarma and Adeli, 2000; Sarma and first discussed and compared. Then, a periodic mainte-
Adeli, 2002; Ahmadkhanlou and Adeli, 2005; Sirca and nance policy involving a failure rate limit (FRL) is intro-
Adeli, 2005; Zhang, 2006; Kang et al., 2009; Schoefs duced for infrastructure systems. According to this pol-
et al., 2009; Deshpande et al., 2010; Zhang et al., icy, maintenance actions are periodically implemented
2010; Zhang and Gao, 2010; Cusson et al., 2011; Peng if no failure occurs over the life cycle of an infra-
et al., 2011). In general, such a model derives an optimal structure system and a preventive replacement (PR) is
maintenance policy in terms of the cost effectiveness carried out when the accumulative failure rate reaches
of maintenance actions. The effect of a maintenance a FRL to renew the system; a corrective replacement
action can be modeled by a jump of the performance (CR) will be undertaken to renew the system if an in-
state, a reduction of the failure rate, or a reduction of tervening failure happens. On the basis of the predeter-
the effective age. One common policy is the periodic mined maintenance policy, this article develops a cost
maintenance policy in which maintenance actions are optimization model to determine the optimal length of
carried out at fixed time intervals. General periodic the maintenance period such that the expected life cycle
maintenance policy models are discussed in Martorell cost (including maintenance cost, failure loss, and the
et al. (1999), Wang and Pham (1996), Wang (2002), and cost of system unavailability) per unit time is minimized.
Wu and Clements-Croome (2005). Another common Next, a search algorithm is developed using MATLAB
policy is the failure limit policy, which sets control to find the optimal maintenance period. After that, the
levels on the condition state, failure rate, or other per- optimization model and the search algorithm are ap-
formance indicators. Maintenance actions are carried plied to the maintenance of bridge decks. The results
out when the predetermined control levels are reached. are consequently compared and analyzed. Finally, this
Under this policy, the continuous monitoring may be article draws the conclusions.
needed to timely detect the achievement of the control
levels. For example, Breysse et al. (2007) recommend
adopting maintenance strategies for sewer systems on 2 FAILURE RATE AND LIFETIME
the basis of the service and ultimate limit state. Another DISTRIBUTION
example of failure limit policy can be found in Grall
et al. (2002). The failure rate is commonly used in the reliability the-
Maintenance optimization models are established on ory to describe the failure probability of a system in a
the basis of the structural condition assessment of in- specified time interval. The failure rate of a deteriorat-
frastructure systems. The structural condition of a sys- ing system is usually nonnegative and age-dependent,
tem usually can be classified into several discrete states. that is, it is constant or monotonically increasing over
For example, the Federal Highway Administration de- time (Barlow and Proschan, 1965). In a continuous time
veloped a rating system for the bridge deck in which model, the failure rate is defined as follows:
the conditions of bridge decks are rated from zero to
P(t |t ) F(t +t)−F(t) f (t)
nine, with nine representing the best condition (FHWA, λ(t)= lim = lim = (1)
1995). Breysse et al. (2007) classified the condition of t→0 t t→0 S(t) · t S(t)
sewer systems with two states (i.e., the service and ul- where λ(t) is the failure rate at time t; P(t|t) is the
timate limit state). The defects usually are detected by conditional probability that a failure occurs in the time
visual inspection. Besides, some minor or nondestruc- interval t given that the system is not failed at time
tive testing measures, such as carbonation testing, dust t; F(t) and f (t) are the cumulative density function and
sampling for chloride analysis, and cover checks, may probability density function (PDF) of the failure time
be carried out to measure the structural condition of of the system, respectively; and S(t) is the survival func-
the system (HyD, 2004). If the system has suffered from tion (the probability the system survives at time t),
other specific problems such as severe traffic impact, fire S(t) = 1 − F(t).
attack, or possibility of severe corrosion, more testing The lifetime distribution of an infrastructure system
should be carried out to assess the structural integrity can be obtained from the statistical analysis of life-
of the system. These test results can provide useful in- times of demolished systems and ages of existing sys-
formation for the physical measurement on structural tems. The Weibull distribution, defined by two param-
condition assessment. Hearn and Shim (1998) proposed eters, a shape parameter α and a scale parameter β, is
an approach to transform nondestructive test results to recommended to properly model the lifetime of infra-
the discrete condition states. structure systems (van Noortwijk and Klatter, 2004). In
In this article, the Weibull distribution is used a Weibull distribution, the shape parameter α controls
to model the lifetime distribution of infrastructure the shape of the distribution. By changing the values
 Determining an optimal maintenance period 3

of α, it can be fitted to a wide range of deteriora- the failure rate by the reduction of the effective age af-
tion processes including negative age-dependence, age- ter a maintenance action (Doyen and Gaudoin, 2004;
independence, and positive age-dependence. If α < 1, Martorell et al., 1999). There are two types of effective
the failure rate decreases over time (i.e., negative age- age models (Kijima, 1989). Type 1 model assumes that
dependence); if α = 1, the failure rate is constant over each maintenance action only reduces the age of the
time (i.e., age-independence) and the distribution of the system gained after the immediately previous mainte-
failure time is an exponential distribution; and if α > 1, nance, whereas type 2 model assumes that each mainte-
the failure rate increases over time (i.e., positive age- nance reduces the cumulative age gained from the first
dependence). If 1 < α < 2, the failure rate function is period. For type 1 model, the effective age instantly af-
a concave curve; if α = 2, it is a slope line; and if α > ter the ith maintenance is given as follows:
2, the curve of failure rate function is convex. The scale 
i
parameter β controls the spread of the distribution. A x̄i+ =(1 − ε)T1 + · · · +(1−ε)Ti =(1−ε) Tj , 0 ≤ ε ≤ 1
small β means that the failure rate increases quickly and j=1 (5)
failure will occur in a short time (Barlow and Proschan,
where x̄i+ is the effective age immediately after the i th
1965). The failure rate function of a Weibull distribution
maintenance; Ti is the length of the ith period in calen-
is expressed as follows:
  dar year; and ε is the improvement factor (ε = 1 means
α t α−1 a perfect maintenance, 0 < ε < 1 means an imperfect
λ(t) = (2)
β β maintenance, and ε = 0 means a minimal maintenance).
For type 2 model, the effective age instantly after the
ith maintenance is calculated as follows:
3 MAINTENANCE EFFECT MODELS x̄i+ = (1 − ε)i T1 + (1 − ε)i−1 T2 + · · · + (1 − ε)Ti

i
3.1 Failure rate reduction model = (1 − ε)i+1− j Tj , 0 ≤ ε ≤ 1 (6)
j=1
A maintenance action can in general reduce the failure
rate of the system (Chan and Shaw, 1993). This effect Let the failure rate in the first period be λ1 (t) and the
may be simply expressed as a proportional reduction of N periodic maintenance actions have a constant period
the failure rate instantly after the implementation of a of T, then the failure rate in the ith period for type 1
maintenance action (Doyen and Gaudoin, 2004): effective age model can be calculated as follows:
λi (t) = λ1 [t − (i − 1)T + (1 − ε)(i − 1)T]
λ+ (xi ) = (1 − ρ)λ− (xi ), 0 ≤ ρ ≤ 1 (3)
= λ1 [t − ε(i − 1)T], for i = 1, . . . , N (7)
where xi is the time of the ith maintenance; λ+ (xi ) is the
failure rate instantly after the ith maintenance; λ− (xi ) For type 2 effective age model, the failure rate in the
is the failure rate just before the ith maintenance; and ith period can be calculated as follows:
ρ is the improvement factor that denotes the effect of λi (t) = λ1 [t − (i − 1)T
a maintenance (ρ = 1 means a perfect maintenance,
0 < ρ < 1 means an imperfect maintenance, and 
i−1

ρ = 0 means a minimal maintenance). + (1 − ε)i− j T], for i = 1, . . . , N (8)

j=1
The interval between the (i – 1)th and the ith main-
tenance is defined as the ith maintenance period (here-
inafter referred to as the period). Suppose there are N
periodic maintenance actions that have a constant pe- 3.3 Comparison of maintenance effect models
riod of T, and the failure rate in the first period is λ1 (t), In this section, three maintenance effect models are
then the failure rate in the ith period can be calculated compared by varying the shape parameter α, while fix-
as follows: ing the scale parameter β and the maintenance period
λi (t) = λi−1 (t) − ρλi−1 [(i − 1)T], for i = 2, . . . , N T. Suppose, ρ = 0.5 for the failure rate reduction model
(4) and ε = 0.5 for both the type 1 and type 2 effective age
models, the failure rate functions for α = 1, 1.5, 2, 4
with the same β = 18 and T = 5 years are plotted in
3.2 Effective age model
Figure 1. When α = 1, the failure rate functions of the
Instead of using the direct reduction of the failure rate two types of effective age models are a same horizontal
after each maintenance as in the failure rate reduction line and the failure rate function of the failure rate re-
model, the effective age model expresses the change of duction model is a decreasing staircase (see Figure 1a).
4 Zhang & Gao

Fig. 1. Failure rate curves for α = 1, 1.5, 2, 4 with the same β = 18, T = 5. (a) α = 1, (b) α = 1.5, (c) α = 2, and (d) α = 4.

With the increase of α, the failure rate curve of type is less than 2 (see Figures 1a and b). This means that the
2 effective age model gradually leaves that of type 1 failure rate gradually decreases under imperfect mainte-
effective age model, and comes close to the curve of nance. Similarly, the failure rate of type 2 effective age
the failure rate reduction model (see Figure 1b). When model increases only slightly when the shape parameter
α = 2, the failure rate curve of type 2 effective age model is larger than 2 (see Figure 1d). This is also not realistic
is the same as that of the failure rate reduction model because a large shape parameter means a faster increase
(see Figure 1c). When α continues increasing, the fail- in the failure rate.
ure rate curve of the failure rate reduction model will
leave from that of the type 2 effective age model (see
Figure 1d). The failure rate of the failure rate reduction
4 FAILURE TIME MODEL
model increases quickly, whereas that of the type 2 ef-
fective age model increases slightly.
4.1 Effective age at the failure time of period i
In comparison of the failure rate functions for dif-
ferent shape parameters, it is found that type 1 effec- As shown in Figure 2, the system has a constant main-
tive age model is more appropriate to be used to mea- tenance period T. Suppose it fails in an arbitrary period
sure the maintenance effect for infrastructure systems i, τ i is defined as the time from the beginning of the ith
because the failure rate of an infrastructure system un- period and ti is the failure time in period i since the start
der imperfect maintenance usually increases over time. point of the life cycle. Considering that maintenance
In contrast, the failure rate reduction model provides a actions are taken at the end of each period, in the type 1
lower and lower failure rate when the shape parameter effective age model, the effective age at the failure time
 Determining an optimal maintenance period 5

ti
tm i tf

0 T 2T (i-1)T iT
failure occurs in ith period
M M M CR

1st period ith period

Case 1: the system fails in an arbitrary period i

tm tr

0 T 2T (i-1)T iT (i+1)T (N -1)T NT

M M M M M M PR

1st period ith period (i+1)th period N th period

Case 2: the system survives in the N periods

M: Maintenance Action CR: Corrective Replacement PR: Preventive Replacement

Fig. 2. Two cases of system failure time.

ti of period i is calculated as: effective age of time ti . It can be calculated as follows:

 t̄i
t̄i = ti − ε(i − 1)T i (ti ) = λi (t)dt
(1−ε)(i−1)T
= τi + (1 − ε)(i − 1)T, for i = 1, . . . , N, 0 ≤ ε ≤ 1  τi
(9) = λ[s+(1−ε)(i −1)T]ds, for i = 1, 2, . . . , N
0 (11)
where τ i is the time from the beginning of the ith period,
The survival function of the system at time ti in the
0 ≤ τi ≤ T; and ti is the failure time in period i since the
ith period can be calculated as (Barlow and Proschan,
start point of the life cycle, ti = τ i + (i − 1)T, (i − 1)T ≤
1965):
ti ≤ iT.
Si (ti ) = exp[− i (ti )]
  τi 
4.2 Conditional PDF of failure at time ti in period i = exp − λ[s + (1 − ε)(i − 1)T]ds ,
0
Assuming that the failure rate function of the system is
for i = 1, 2, . . . , N (12)
λ(t) and the total number of periods over the life cycle
of the system is N, in the type 1 effective age model, the The PDF of the failure time ti in the ith period is
failure rate at time ti of period i is reduced to the failure fi (ti ). According to Equation (1), it can be calculated as
rate at the effective age, t̄i . According to Equation (7), follows:
it can be calculated as: fi (ti ) = λi (ti )Si (ti ) = fi (τi )
λi (ti ) = λ[ti − ε(i − 1)T] = λ(t̄i ) = λ[τi + (1 − ε)(i − 1)T]
  τi 
= λ[τi + (1 − ε)(i − 1)T], for i = 1, 2, . . . , N
× exp − λ[s + (1 − ε)(i − 1)T]ds
(10) 0

for i = 1, 2, . . . , N (13)
The cumulative hazard at time ti of the ith period,
i (ti ), is equal to the hazard cumulated from the ef- The probability that the system survives in the first
fective age of the start point of the ith period to the period is F̄1 (T). Given the system survives in the first
6 Zhang & Gao

period, the probability that the system survives in the 5.2 Maintenance policy
second period, F̄(2T), is equal to F̄1 (T) F̄2 (T). There-
Compared to related policies in previous literature, the
fore, the probability that the system survives in the first
maintenance policy proposed in this article has the fol-
i periods, F̄(i T), can be iteratively obtained as follows:
lowing two features:

i
F̄(i T) = F̄ j (T), ∀i = 1, 2, . . . , N (14) 1. A limit is set to the cumulative failure rate over
j=1 the life cycle of the system. This limit is referred to
as the FRL. It can be determined by the estimated
where F̄ j (T) is the probability that the system survives replacement age of the system without receiving
in the jth period. It can be calculated as: maintenance actions.
 jT  T
2. There are three maintenance scenarios in an arbi-
F̄ j (T) = f j (t j )dt j = f j (τ j )dτ j trary period i: Scenario 1 (the FRL has not been
( j−1)T 0
 reached and the system does not fail)—a mainte-
T
= λ[τ j + (1 − ε)( j − 1)T] (15) nance action will be implemented at the end of
0 this period; Scenario 2 (the FRL has been reached
  τj  and the system does not fail)—a PR will be imple-
× exp − λ[s + (1 − ε)( j − 1)T]ds dτ j mented at the end of this period and consequently
0
the system is renewed; and Scenario 3 (the system
As shown in Figure 2, if the system fails in an ar- fails no matter whether the FRL has been reached
bitrary period i, the system has survived in the first or not)—a CR will be implemented immediately at
(i − 1) periods. The probability that the system survives the time of failure to renew the system.
in the first (i − 1) periods can be calculated from Equa-
tion (14). The conditional PDF of the system fails at Scenarios 1 and 2 are illustrated in Figure 3. The sys-
time ti in the period i can be calculated as follows: tem does not fail in the first N periods and the FRL will
be reached at the end of the Nth period. A maintenance
gi (ti ) = F̄[(i − 1)T] · fi (ti ), action is implemented in each of the first (N – 1) pe-
∀i = 1, 2, . . . , N, (i − 1)T ≤ ti ≤ i T (16) riods whereas a PR is executed at the end of the Nth
period. This PR renews the system and a new life cycle
From Equation (13), the Equation (16) can be rewrit- starts. Scenario 3 is illustrated in Figure 4. The system
ten as: fails in an arbitrary period i. In this case, a maintenance
gi (ti ) = gi (τi ) = F̄[(i − 1)T] fi (τi ), action is implemented in each of the first (i − 1) periods
∀i = 1, 2, . . . , N, 0 ≤ τi ≤ T (17) whereas a CR is executed in the ith period immediately
after the failure. This CR renews the system and a new
life cycle starts.

5 PROPOSED MAINTENANCE POLICY 5.3 Maximum number of maintenance periods

5.1 Assumptions The FRL determines the maximum number of mainte-

nance periods over the life cycle of the system. Using
The following assumptions are made in the proposed the type 1 effective age model, the maximum number of
maintenance policy: maintenance periods, N, can be uniquely determined by
1. Maintenance actions are periodically performed. solving the following equations:
They are identical and imperfect, i.e., they have the λ[(N − 1)T] = λ[(N − 1)T − ε(N − 2)T] ≤ FRL (18)
same effect and base cost; they cannot restore the
system as good as new.
λ(NT) = λ[NT − ε(N − 1)T] > FRL (19)
2. The base cost of each maintenance action is the
same. The maintenance cost of each period is dis-
counted to the start time point of the life cycle.
3. Failure of the system is immediately detected and 6 LENGTH OF LIFE CYCLE
the system is replaced consequently. This replace-
ment is hereinafter referred to as a CR. The life cycle of the system will be renewed in two cases
4. The life cycle of the system is defined as the du- as shown in Figure 2. In Case 1, the system fails in
ration beginning with a renewal and ending with a an arbitrary period i (i ≤ N). Maintenance actions are
replacement. implemented in each of the first (i – 1) periods, and a CR
 Determining an optimal maintenance period 7

Failure Rate
Original Failure Rate Function without Maintenance
Reduced Failure Rate in Type 1 Effective Age Model

Preventive
Replacement
Periodical Maintenance

0 T 2T 3T (i-1)T iT (N -1)T NT Time

Fig. 3. No failure in the life cycle.

Failure Rate

Original Failure Rate Function without Maintenance

Reduced Failure Rate in Type 1 Effective Age Model

Periodical Maintenance Corrective

Replacement

0 T 2T 3T (i-1)T iT Time

Fig. 4. Failure in the ith period.

is implemented in period i immediately after the failure. fails at time τ i in the period i given that it survives in the
The length of the life cycle is equal to the failure time first (i – 1) periods.
plus the downtime of the CR. The expected length of In Case 2, the system survives in the first N pe-
the life cycle in this case, E1 (L), is calculated as follows: riods. Maintenance actions are performed in each of
 T  2T the first (N – 1) periods and a PR is implemented
E1 (L) = (t1 + t f )g1 (t1 )dt1 + (t2 + t f )g2 (t2 )dt2 at the end of the Nth period. The length of the
0 T life cycle is equal to the length of the N periods
 NT plus the downtime of the PR. The expected length
+··· + (t N + t f )g N (t N )dt N of the life cycle in this case, E2 (L), is calculated as
(N−1)T
 follows:
T
= (τ1 + t f )g1 (τ1 )dτ1 (20) E2 (L) = (NT + tr )S(NT) (21)
0
 T where tr is the downtime of the PR; and S(NT) is
+··· + [τ N + (N − 1)T + t f ]g N (τ N )dτ N the probability that the system survives in the first N
0
periods.
N 
 T The probability that the system survives in the first N
= [τi + (i − 1)T + t f ]gi (τi )dτi periods is equal to 1 minus the cumulative probability
i=1 0
that the system fails in the first N periods:
where ti is the failure time in period i; tf is the downtime  N  T
of the CR; τ i is the time from the beginning of the ith S(NT) = 1 − gi (τi )dτi (22)
period; and gi (τ i ) is the conditional PDF of the system i=1 0
8 Zhang & Gao

According to the law of total probability, the the cost caused by a failure. The cost of system unavail-
expected length of the life cycle of the system is derived ability is the cost of lost production because of the main-
as follows taking into consideration Case 1 and Case 2: tenance downtime. This cost may be calculated by mul-
N  T tiplying the mean cost of lost production per unit time
E(L) = [τi + (i − 1)T + t f ]gi (τi )dτi by the expected time of unavailability in a life cycle
i=1 0 (Vaurio, 1999).
In Case 1, if the system fails in an arbitrary period
Case 1
+ (NT + tr )S(NT) i (i ≤ N), the life cycle cost includes a CR cost and a
(23) failure loss in period i, and the maintenance costs in the
Case 2
first (i – 1) periods. In Case 2, the life cycle cost includes
the maintenance costs in the first (N – 1) periods and
7 AVAILABLE AND UNAVAILABLE TIME a PR cost in the Nth period. In both the cases, the life
cycle cost also includes a cost of system unavailability.
The available time of a system is the total operation The expected total life cycle cost of the system, E(C),
time in a life cycle and the unavailable time is the total can be obtained from the following equation:
⎛ ⎞
time of a life cycle minus the total operation time in a
N  T  i−1
life cycle. Let tm denote the downtime of a maintenance E(C) = ⎝ CM + CFi ⎠ gi (τi ) dτi
j
action. Then, the available time in each period is T – i=1 0 j=0
tm . In Case 1 mentioned in the previous section, if the
Case 1
system fails in an arbitrary period i, the available time (26)
 N−1 
is equal to the failure time ti minus the total downtime 
of the maintenance actions in the first (i – 1) periods. In + i
CM + CRN S(NT) + CU E(U)
Case 2, the available time is equal to the length of the i=1
Cost of System
N periods minus the total downtime of the maintenance Case 2 Unavailability
actions in the first (N – 1) periods. The expected avail- j j
able time of the system in a life cycle can be obtained where CM is the maintenance cost in period j, CM =
from the following equation: Cm · (1 + r )− j T , CM
0
= 0; CFi is the failure cost in period
i, CF = (C f 1 + C f 2 )(1 + r )−ti ; CRN is the PR cost at the
i
N  T
end of the Nth period, CRN = Cr · (1 + r )−NT ; CU is the
E(A) = [τi + (i − 1)T − (i − 1)tm]gi (τi )dτi
0
mean cost of lost production per unit time; Cm , Cr , and
i=1
Cf 1 are the base costs of a maintenance action, PR, and
Case 1 CR, respectively; Cf 2 is the failure loss; and r is the dis-
+ [NT − (N − 1)tm]S(NT ) (24) count rate.
The expected life cycle cost per unit time of the sys-
Case 2
tem is defined as the expected total life cycle cost di-
The expected unavailable time over the life cycle can vided by the expected length of the life cycle:
be obtained as follows:
E(C)
E(U) = E(L) − E(A) (25) E(K) = (27)
E(L)

8 LIFE CYCLE COST 9 OPTIMIZATION MODEL

The total life cycle cost of the system includes mainte- The objective of the optimization model is to mini-
nance costs, failure loss, and the cost of system unavail- mize the expected life cycle cost per unit time subject
ability. Maintenance costs include the costs of all main- to the constraint of the FRL in the life cycle of the
tenance actions, and a PR or CR cost. Failure loss is system:
⎛ ⎞  N−1 
N  T i−1 
⎝ CM + CF ⎠ gi (τi )dτi +
j i
CM + CR S(NT) + CU E(U)
i N

i=1 0 j=0 i=1

min E(K) = N  T

[τi + (i − 1)T + t f ]gi (τi )dτi + (NT + tr )S(NT) (28)
i=1 0

s.t. λ[(N − 1)T] ≤ FRL

λ(NT) > FRL
 Determining an optimal maintenance period 9

terminative value, and the step length of T are prede-

Set the initial value of maintenance period, T
termined. In each iteration, the calculated E(K) will be
compared with that of the previous iteration to find the
minimum value of E(K), E(K)min . The minimum E(K)
Set the terminative value of maintenance period, T ter
at the end of the search is the optimal value, E(K)opt .
The value of T corresponding to E(K)opt is the optimal
maintenance period, T opt .
Set the step length of maintenance period, s

Calculate the initial value of expected 10 AN ILLUSTRATIVE EXAMPLE–BRIDGE

life cycle cost per unit of time, E(K) DECK MAINTENANCE

10.1 Lifetime distribution of bridge deck

Set E(K)min = E(K) The deck is an important bridge element that deterio-
rates over time because of many factors such as gen-
eral wear, fatigue loading, carbonation, chloride attack,
Set T = T + s sulfate attack, acids, alkali-aggregate reactions, over-
loading, and fire damage. The deterioration process of
bridge deck usually is impacted by both the chemi-
Calculate the new value of E(K) cal and mechanical deterioration. The chemical dete-
rioration process is positive age-dependent (i.e., it has
an increasing deterioration rate over time) whereas
No
E(K) < E(K)min the mechanical deterioration process is nearly age-
independent (i.e., it has a steady failure rate over time)
Yes (Mishalani and Madanat, 2002). In Hong Kong, the per-
formance of bridge deck is classified into five states (one
Set E(K)min = E(K) for best condition and five for failed condition) accord-
ing to the severity of defects, such as spalling, delami-
nation, and cracking (HyD, 2004). The periodic mainte-
No
T > T ter nance action is the asphalt concrete overlay.
The lifetime of the bridge deck is the duration that
Yes the bridge deck deteriorates from the best state to the
Set E(K)opt = E(K)min
failed state. It is impacted by various parameters, such
as material property, structure dimensions, reinforce-
ment, traffic loading, environmental condition, and ini-
tial bridge deck condition (Zhang, 2010). The lifetime
Set T opt = T with E(K)min
distribution of the bridge deck often is assumed to fol-
low a Weibull distribution. van Noortwijk and Klatter
Fig. 5. Optimal searching algorithm.
(2004) proved that the Weibull distribution fits very
well for the bridge lifetime distribution, especially for
In the above optimization model, the type 1 effective bridges of which the deterioration processes are primar-
age model is used to model the maintenance effect. As ily governed by the mechanical deterioration (Mishalani
discussed in Section 3, it is appropriate to use the type and Madanat, 2002).
1 effective age model for an infrastructure system as
this model reflects the fact of increasing failure rate of
10.2 Parameter estimation of Weibull distribution
an infrastructure system under imperfect maintenance
over its life cycle. The decision variables in the opti- The parameters of the Weibull distribution can be
mization model are the length of the maintenance pe- determined by the maximum-likelihood estimation
riod (T) and the number of maintenance periods (N). N method if there are sufficient observation data of the
is a function of T. Therefore, the objective is to find an lifetimes of bridge decks (van Noortwijk and Klatter,
optimal T that minimizes E(K). A search algorithm is 2004). If historical observation data are insufficient, ex-
developed using MATLAB to find the optimal T. The pert opinions can be solicited to derive these param-
algorithm is shown in Figure 5. The initial search value, eters. A group of experts can be invited to answer
10 Zhang & Gao

Table 1 Table 2
Parameters of Weibull distribution Optimal results of the optimization model

u (year) S(u) v (year) S(v) α β E(C) E(L) E(K)

T (years) N (HK $) (years) (HK $/year)
10 90% 25 25% 2.812 22.258
9 6 126,080 26.9 4,687.1

questions pertaining to their beliefs about the likeli-

hood of a bridge deck surviving for a certain time period
(Kleiner, 2001).
The survival function of the Weibull distribution can
be calculated as follows:
  α 
t
S(t) = exp − (29)
β

Let S(u) and S(v) be the probabilities that a bridge

deck, without receiving maintenance actions, will sur-
vive for at least u and v periods, respectively. The pa-
rameters α and β can be obtained as follows (Kleiner,
2001):

ln{ln[S(u)] ln[S(v)]}
α= (30)
ln(u/v)

u
β= (31) Fig. 6. Expected life cycle cost per unit time.
{− ln[S(u)]}1/α

The expert opinions regarding the survival probabil- and the maximum maintenance periods are 6 years. The
ities of the deterioration process are shown in Table 1. minimized life cycle cost per unit time is HK $4,687 per
The parameters of Weibull distribution calculated using year. The changes of E(K) for a maintenance period
Equations (30) and (31) are shown in Table 1. from 0 to 50 years is shown in Figure 6. It is seen from
Figure 6 that E(K) decreases when the maintenance
10.3 Assumption of the input data period increases from 0 to the optimal maintenance pe-
The necessary input data of the optimization model are riod, and then increases when the maintenance period
assumed as follows. The base costs of a maintenance continues to increase.
action, PR, and CR are: Cm = HK $4,000/m2 , Cr = If there is no maintenance action, the expected length
HK $10,000/m2 , and Cf 1 = HK $20,000/m2 , respectively; of life cycle is 19.8 years. The expected life cycle cost per
the downtimes of a maintenance action, PR, and CR unit time is HK $5,579 per year. The cumulative failure
are: tm = 20 days, tr = 70 days, and tf = 150 days, probabilities of the bridge deck at the end of each pe-
respectively; the failure loss Cf 2 = HK $150,000/m2 ; riod with and without receiving maintenance actions are
the mean cost of the system unavailability CU = HK shown in Table 3 and Figure 7. It is found that the cumu-
$50,000/m2 /year; the discount rate r = 3%; and the lative failure probabilities have been decreased when
improvement factor of the type 1 effective age model maintenance actions are taken. The observations prove
ε = 0.5. that the maintenance actions can decrease the failure
Assuming that the estimated replacement age of the probability over the life cycle and extend the lifetime of
bridge deck without receiving maintenance actions is the bridge deck.
30 years, the survival probability at the end of the esti- When there is no FRL, the optimal maintenance pe-
mated replacement age is 10% and the calculated FRL riod is 10 years. The expected life cycle cost per unit
is 0.217. time is HK $4,699 per year and the expected length of
life cycle is 26.5 years. This observation shows that the
proposed policy limits the maximum maintenance pe-
10.4 Results of the optimization model
riods such that the risks of failure occurring in the lat-
The optimal values of T and other results are shown ter stages in which the bridge deck is very old can be
in Table 2. The optimal maintenance period is 9 years reduced.
 Determining an optimal maintenance period 11

Table 3
Comparison of cumulative failure probabilities without and with receiving maintenance actions

T (years) 9 18 27 36 45 54
Without maintenance actions 0.075 0.423 0.821 0.979 0.999 1
With maintenance actions 0.075 0.268 0.544 0.792 0.935 0.987

1 basis of the proposed maintenance policy, with an ob-

Cumulative Failure Probability

0.9
jective to minimize the system’s life cycle cost (including
0.8
0.7 maintenance cost, failure loss, and the cost of system un-
0.6 availability) per unit time. This optimization model and
0.5
the search algorithm that is consequently formulated
0.4
0.3 are applied to the maintenance of bridge decks. The
0.2 Weibull distribution is used to model the deterioration
0.1
process of the bridge deck. The results show that the
0
0 9 18 27 36 45 54
proposed maintenance policy is appropriate for infra-
Years
structure management involving imperfect mainte-
nance. Under this policy, the expected lifetime of a
Cumulative failure probability without maintenance bridge deck can be extended and the expected life cy-
Cumulative failure probability with maintenance
cle cost per unit time can be decreased.
Fig. 7. Cumulative failure probability over the life cycle.

ACKNOWLEDGMENT
11 CONCLUSIONS
This study is sponsored by the Public Policy Research
The effect of a maintenance action for an infrastructure Grant HKUST6004-PPR-10 of the Hong Kong Re-
system can be measured by a reduction of the failure search Grant Council.
rate or of the effective age of the system immediately
after the maintenance action is taken. Modeling the life- REFERENCES
time of an infrastructure system as the Weibull distribu-
tion, this article has compared three maintenance effect Adeli, H. & Ge, Y. (1989), A dynamic programming method
models. The results show that it is appropriate to use the for analysis of bridges under multiple moving loads, Inter-
type 1 effective age model to measure the maintenance national Journal for Numerical Methods in Engineering, 28,
effect as the failure rate of an infrastructure system un- 1265–82.
der imperfect maintenance usually increases over the Ahmadkhanlou, F. & Adeli, H. (2005), Optimum cost design
of reinforced concrete slabs using neural dynamics model,
life cycle of the system. The type 1 effective age model Engineering Applications of Artificial Intelligence, 18(1),
reflects this characteristic regardless of the value of the 65–72.
shape parameter. Barlow, R. E. & Proschan, F. (1965), Mathematical Theory of
On the basis of the type 1 effective age model, a main- Reliability, SIAM, Philadelphia.
tenance policy can be derived. This article proposes Breysse, D., Vasconcelos, E. & Schoefs, F. (2007), Man-
agement strategies and improvement of performances of
a maintenance policy that sets a limit to the cumula- sewer networks, Computer-Aided Civil and Infrastructure
tive failure rate over the life cycle of the system. There Engineering, 22(7), 462–77.
are three maintenance scenarios in an arbitrary period Chan, J. & Shaw, L. (1993), Modeling repairable systems with
under this policy: a maintenance action will be imple- failure rates that depend on age and maintenance, IEEE
mented in Scenario 1 in which the FRL has not been Transactions on Reliability, 42(4), 566–71.
Chu, C. Y. & Durango-Cohen, P. L. (2008), Incorporating
reached and the system does not fail; a PR will be im- maintenance effectiveness in the estimation of dynamic
plemented to renew the system in Scenario 2 where the infrastructure performance models, Computer-Aided Civil
FRL has been reached and the system does not fail; and and Infrastructure Engineering, 23(3), 174–88.
a CR will be implemented at the time of failure to re- Cusson, D., Lounis, Z. & Daigle, L. (2011), Durability mon-
new the system in Scenario 3 where the system fails no itoring for improved service life predictions of concrete
bridge decks in corrosive environments, Computer-Aided
matter whether the FRL has been reached or not. Civil and Infrastructure Engineering, 26(7), 524–41.
An optimization model is developed to determine Deshpande, V. P., Damnjanovic, I. D. & Gardoni, P. (2010),
the optimal length of the maintenance period on the Reliability–based optimization models for scheduling pave-
12 Zhang & Gao

ment rehabilitation, Computer-Aided Civil and Infrastruc- Peng, F., Kang, S., Li, X., Ouyang, Y., Somani, K. & Acharya,
ture Engineering, 25(4), 227–37. D. (2011), A heuristic approach to railroad track mainte-
Doyen, L. & Gaudoin, O. (2004), Classes of imperfect repair nance scheduling problem, Computer-Aided Civil and In-
models based on reduction of failure intensity or virtual frastructure Engineering, 26(2), 129–45.
age, Reliability Engineering and System Safety, 84(1), 45–56. Sarma, K. & Adeli, H. (1998), Cost optimization of con-
Federal Highway Administration (FHWA). (1995), Recording crete structures, Journal of Structural Engineering, ASCE,
and Coding Guide for the Structure Inventory and Appraisal 124(5), 570–78.
of the Nation’s Bridges, FHWA-PD-96–001, U.S. Depart- Sarma, K. & Adeli, H. (2000), Cost optimization of steel struc-
ment of Transportation, Washington DC, USA. tures, Engineering Optimization, 32(6), 777–802.
Grall, A., Bérenguer, C. & Dieulle, L. (2002), A condition- Sarma, K. & Adeli, H. (2002), Life-cycle cost optimization of
based maintenance policy for stochastically deteriorating steel structures, International Journal for Numerical Meth-
systems, Reliability Engineering and System Safety, 76(2), ods in Engineering, 55(12), 1451–62.
167–80. Schoefs F., Aduriz X., Bernard O. & Capra B. (2009), Com-
Hearn, G. & Shim, H. (1998), Integration of bridge manage- parison of additional costs for several replacement strate-
ment systems and nondestructive evaluations, Journal of gies of randomly ageing reinforced concrete pipes, Com-
Infrastructure System, ASCE, 4(2), 49–55. puter Aided Civil and Infrastructure Engineering, 24(7),
Highways Department (HyD). (2004), Inspection Manual for 492–508.
Highway Structures, Highways Department, Hong Kong, Sirca, G. & Adeli, H. (2005), Cost optimization of prestressed
China. concrete bridges, Journal of Structural Engineering, ASCE,
Kang, M.W., Schonfeld, P. & Yang, N. (2009), Prescreening 131(3), 380–88.
and repairing in a genetic algorithm for highway alignment van Noortwijk, J. M. & Klatter, H. E. (2004), The use of life-
optimization, Computer-Aided Civil and Infrastructure time distributions in bridge maintenance and replacement
Engineering, 24(2), 109–19. modeling, Computers and Structures, 82(13–14), 1091–99.
Kijima, M. (1989), Some results for repairable systems Vaurio, J. K. (1999), Availability and cost functions for peri-
with general repair, Journal of Applied Probability, 26(1), odically inspected preventively maintained units, Reliability
89–102. Engineering and System Safety, 63(2), 133–40.
Kleiner, Y. (2001), Scheduling inspection and renewal of Wang, H. (2002), A survey of maintenance policies of deterio-
large infrastructure assets, Journal of Infrastructure System, rating systems, European Journal of Operational Research,
ASCE, 7(4), 136–43. 139(3), 469–89.
Labi, S. & Sinha, K. C. (2003), Measures of short-term ef- Wang, H. & Pham, H. (1996), Optimal maintenance policies
fectiveness of highway pavement maintenance, Journal of for several imperfect maintenance models, International
Transportation Engineering, ASCE, 129(6), 673–83. Journal of Systems Science, 27(6), 543–49.
Lie, C. H. & Chun, Y. H. (1986), An algorithm for preven- Wu, S. & Clements-Croome, D. (2005), Preventive mainte-
tive maintenance policy, IEEE Transactions on Reliability, nance models with random maintenance quality, Reliability
35(1), 71–75. Engineering and System Safety, 90(1), 99–105.
Malik, M. A. K. (1979), Reliable preventive maintenance Zhang, X. Q. (2006), Markov-based optimization model for
scheduling, AIIE Transactions, 11(3), 221–28. building facilities management, Journal of Construction En-
Martorell, S., Sanchez, A. & Serradell, V. (1999), Age- gineering and Management, 132(11), 1203–11.
dependent reliability model considering effects of mainte- Zhang, X. Q. (2010), Report of Study on Development
nance and working conditions, Reliability Engineering and of Highway Asset Management, Highways Department,
System Safety, 64(1), 19–31. Hong Kong, China.
Mishalani, R. & Madanat, S. (2002), Computation of in- Zhang, X. Q., Cheung, M. S. & Gao, H. (2010), Performance-
frastructure transition probabilities using stochastic dura- based infrastructure asset management strategy, Journal of
tion models, Journal of Infrastructure System, ASCE, 8(4), Disaster Prevention and Mitigation Engineering, 30, 407–
139–48. 12.
Nakagawa, T. (1986), Periodic and sequential preventive Zhang, X. Q. & Gao, H. (2010), Optimal performance-based
maintenance policies, Journal of Applied Probability, 23(2), building facility management, Computer-Aided Civil and
536–42. Infrastructure Engineering, 25(4), 269–84.