ARTICLE IN PRESS

Reliability Engineering and System Safety 92 (2007) 530–534

www.elsevier.com/locate/ress

Reliability-centered predictive maintenance scheduling for a

continuously monitored system subject to degradation
Xiaojun Zhoua,, Lifeng Xia, Jay Leeb
a
School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200030, China
b
University of Wisconsin, Milwaukee, WI 53211, USA
Received 26 August 2004; received in revised form 29 April 2005; accepted 12 January 2006
Available online 4 April 2006

Abstract

This paper tries to integrate sequential imperfect maintenance policy into condition-based predictive maintenance (CBPM). A
reliability-centered predictive maintenance policy is proposed for a continuously monitored system subject to degradation due to the
imperfect maintenance. It is assumed that the system hazard rate is a known function of the system condition and then can be derived
directly through CBPM. A hybrid hazard rate recursion rule based on the concept of age reduction factor and hazard rate increase factor
is built up to predict the evolution of the system reliability in different maintenance cycles. Whenever the system reliability reaches the
threshold R, an imperfect preventive maintenance (PM) is performed on the system. The optimal reliability threshold R is determined by
minimizing the cumulative maintenance cost per unit time in the residual life of the system which is based on simulation. Finally, a
discussion is presented to show how the optimal results depend on the different cost parameters.
r 2006 Published by Elsevier Ltd.

Keywords: Sequential imperfect maintenance; Condition-based predictive maintenance; Reliability; Optimization; Cost

1. Introduction usage and age. There are about eight main treatments to
modeling an imperfect PM action [5]. One of them is
Maintenance optimization is a popular issue to research- improvement factor method. This kind of method is very
ers. Since Barlow and Hunter gave the minimal repair useful in engineering because the maintenance decision is in
model in 1960 [1], a lot of optimal maintenance strategies terms of the system hazard rate or other reliability
have been developed and implemented for improving measures. Two types of improvement factors are devel-
system reliability, preventing system failures and reducing oped. Malik [6] introduces the concept of age reduction
maintenance costs. factor. If T i and hi ðtÞ for t 2 ð0; T i Þ, respectively, represent
Conventional preventive maintenance (PM) is scheduled the PM interval and the hazard rate function of the system
periodically on the basis of technician’s experience and it prior to the ith PM, the hazard rate function after the ith
often holds a same time interval T for PM actions [2–4]. PM becomes hi ðt þ ai T i Þ for t 2 ð0; T iþ1 Þ, where 0oai o1 is
However, PM is generally imperfect and it cannot restore the age reduction factor due to the imperfect PM action.
the system to as good as new. The age-T policy gives This implies that each imperfect PM changes the initial
unavoidably decreasing reliabilities at the PM actions for a hazard rate value right after the PM to hi ðai T i Þ, but not all
degradation system with imperfect maintenance effect. the way to zero (not new). Nakagawa [7] proposes another
This leads to the sequential imperfect PM policy. Under model based on the hazard rate increase factor. The hazard
this policy, the system is maintained at a decreasing rate function becomes bi hi ðtÞ for t 2 ð0; T iþ1 Þ after the ith
sequence of intervals, which is more practical since most PM, where bi 41 is the hazard rate increase factor. This
systems need more frequent maintenance with increased indicates that each PM resets the hazard rate function
value to zero (new), and makes the increase rate of the
Corresponding author. hazard rate function more and more higher. Because the

0951-8320/$ - see front matter r 2006 Published by Elsevier Ltd.

doi:10.1016/j.ress.2006.01.006
 ARTICLE IN PRESS
X. Zhou et al. / Reliability Engineering and System Safety 92 (2007) 530–534 531

improvement factor method is practical in engineering [5], hi ðtÞ system hazard rate function prior to the ith PM
it has been a research issue and many research results have Ti time interval for PM prior to the ith PM
been proposed [8–11]. R system reliability threshold for scheduled PM
Although the improvement factor method can be used as cup expected cost rate for unscheduled PM
a general treatment for imperfect maintenance in engineer- csp expected cost rate for scheduled PM
ing, how to get the system’s hazard rate curve hi ðtÞ is still a tp duration of PM, same for scheduled PM and
great topic to researchers. Some well-known modes (i.e. unscheduled PM
Weibull) [9,11] are always applied in the existing imperfect Costr additional replacement cost
PM models which may be inaccurate in practice for many cEr expected cost per unit time for the system in the
systems. residual life.
The advancement of condition-based maintenance
(CBM) makes it feasible to obtain the exact hazard rate Two kinds of PM action are considered in this paper. The
of the system [12,13]. By integrating prediction tools, CBM first is the scheduled PM action which is performed whenever
provides the assessment and prediction of the system the system reliability reaches the threshold. The second is the
hazard rate in current maintenance cycle based on the unscheduled PM action performed whenever the system fails
system condition collected through continuous monitoring, before the scheduled PM. It could be a corrective main-
with the aim of determining the required maintenance tenance, or a corrective replacement for the damaged system
action prior to any predicted failure. From this aspect, this component. Only one PM action, scheduled or unscheduled,
technique can be called condition-based predictive main- can be implemented in one maintenance cycle. It is assumed
tenance (CBPM ). It has been proved that CBPM can that the unscheduled PM action has the same imperfect effect
greatly improve system safety [14]. However, two issues as the scheduled one. However, the maintenance cost for the
still exist. On one hand, CBPM is traditionally costly and, unscheduled PM is much higher than the cost for the
therefore, it is only used for some important equipment (i.e. scheduled PM. For example, the unexpected system failure
generator, engine). Many researchers usually assume that may cause low system production and delay of delivery. After
the monitoring is not continuous and then try to find the the PM action, the system begins a new decreasing process
optimal inspection intervals to save the monitoring cost because of the imperfect maintenance effect.
[15–20]. Fortunately, with the development of intelligent Other assumptions are as follows:
maintenance system (IMS), the applications of the
embedded agent techniques (i.e. Watchdog AgentTM) and (1) The system undergoes relatively constant conditions of
the tether-free techniques (i.e. wireless) make it possible stress, environment and maintenance in the residual
to monitor the system continuously with much lower life.
cost [21,22]. On the other hand, a lot of CBPM models (2) The condition of the system is monitored and predicted
assume that the maintenance (either corrective or pre- continuously. The monitoring is perfect and has no
ventive) is always perfect and it can restore the system to as effect on system reliability.
good as new after PM [15,16,18–20,23]. This is not (3) The system hazard rate is a known function of the
practical in reality. condition. This implies that the hazard rate function
This paper tries to integrate sequential imperfect can be monitored and predicted continuously through
maintenance policy into CBPM. A system is considered CBPM.
which is subject to degradation and is monitored con-
tinuously and perfectly. Whenever the system reliability 2.1. Evolution of the hazard rate function
reaches the threshold R, an imperfect PM is performed
which leads a gradual decrease in maintenance intervals. In From the descriptions of imperfect maintenance models,
order to predict the evolution of the system reliability in it can be seen that the age reduction method has the
different maintenance cycles, a hybrid hazard rate recur- advantage of determining the initial failure rate value right
sion rule is adopted with the aim of combining the age after a PM, and the hazard rate increase method has the
reduction factor method and the hazard rate increase advantage of allowing the increase rate of the hazard rate
factor method. The optimal reliability threshold R is function to be higher after each PM. With the aim of
determined with an aim of minimizing the cumulative combining these two advantages, a hybrid hazard rate
maintenance cost per unit time in the residual life of the evolution rule is adopted based on the age reduction
system. method and the hazard rate increase method.
As shown in Fig. 1, the relationship between the hazard
2. Maintenance model rate functions before and after the ith PM can be defined as

Notations: hiþ1 ðtÞ ¼ bi hi ðt þ ai T i Þ for t 2 ð0; T iþ1 Þ, (1)

where 0oai o1 and bi 41 are the age reduction factor and
i ordinal of PM cycles, i ¼ 1; 2; . . . ; N the hazard rate increase factor, respectively, which need to
N PM cycle number be deduced from the history maintenance data of the
 ARTICLE IN PRESS
532 X. Zhou et al. / Reliability Engineering and System Safety 92 (2007) 530–534

Hybrid Model Hazard rate increase Model Assuming that the system is replaced just at the Nth PM
hi+1(t) = bihi(t+aiTi) hi+1(t) = bihi(t), bi>1, action and the duration for replacement is equal to the
t∈(0, Ti+1)
duration for PM, Eq. (5) can be rewritten as
Hazard Rate h(t)

hi(t) Nðcup ð ln RÞ þ csp ð1 þ ln RÞÞtp þ Costr

cEr ¼ , (6)
PN
Age Reduction Model ðT i Þ þ Ntp
hi+1(t) = hi(t+aiTi) i¼1
0<ai<1, t∈(0, Ti+1)
where Costr is the additional cost for replacement except
the PM cost. T i can be obtained from Eq. (3) and it is a
ith PM function of R. By minimizing cEr , the optimal system
Time t
reliability threshold for scheduled PM can be determined.
Fig. 1. Hybrid evolution model for system hazard rate. At the same time, the nominal value of (N, T i ) can be
decided which should be helpful for preparing the
scheduled PM activities. (N, T i ) may change if failure
system [6–8,24,25]. This hybrid recursion rule makes it occurs during the scheduled PM cycle. Under this
possible to predict the evolution of the system reliability in circumstance, (N, T i ) is recalculated from Eq. (6) based
different maintenance cycles. on the obtained system reliability threshold R.
If the working environment changes before the system is
2.2. Maintenance model formulation and optimization replaced, all the steps are reprocessed to determine the new
system reliability threshold.
In this model, a scheduled PM is performed whenever
the system reliability reaches the reliability threshold R.
Based on this policy, a reliability equation can be 2.3. Derivation of h1 ðtÞ through CBPM
constructed as
 Z T1   Z T2  According to the assumption, the hazard rate function in
exp  h1 ðtÞ dt ¼ exp  h2 ðtÞ dt the current maintenance cycle h1 ðtÞ can be derived from the
0 0 monitored and predicted system states. Commonly, the
 Z TN 
hazard rate function h1 ðtÞ derived through CBPM is
¼    ¼ exp  hN ðtÞ dt ¼ R, ð2Þ discrete. In order to get the cumulative risk of system
0
failure, Eqs. (1) and (3) can be rewritten as
where hi ðtÞ can be deduced from Eq. (1). Eq. (2) can be  
rewritten as ai T i T iþ1
hiþ1 ðkÞ ¼ bi hi k þ for k ¼ 0; 1; 2;    ; , (7)
Z T1 Z T2 Z TN Dt Dt
h1 ðtÞ dt ¼ h2 ðtÞ dt ¼    ¼ hN ðtÞ dt ¼  ln R,
0 0 0 T1 T2 TN
X
Dt X
Dt X
Dt
(3) h1 ðkÞDt ¼ h2 ðkÞDt ¼    ¼ hN ðkÞDt
RT
where 0 i hi ðtÞ dt represents the cumulative failure risk in k¼1 k¼1 k¼1
maintenance cycle i. This implies that the cumulative risk ¼  ln R, ð8Þ
of system failure in each maintenance cycle is equal to
 ln R. Since only one PM action is performed in one where Dt  T i is the inspection interval; Dt  T i
maintenance cycle, Eq. (3) also indicates that the prob- satisfies the assumption that the system is continuously
ability to implement a scheduled PM action is equal to monitored. The residual computation steps are the same.
1 þ ln R. Taking into account maintenance duration tp for
PM, the expected cost per unit time for maintenance cycle i 3. Numerical examples
can be evaluated as
cup tp ð ln RÞ þ csp tp ð1 þ ln RÞ In order to get the discrete hazard rate function which
cEri ¼ , (4) will be used in the following simulations, a sampling
T i þ tp
process is carried out in the special case when the hazard
where cup is the expected cost per unit time for unscheduled rate function h1 ðtÞ is Weibull. It is assumed that
PM and csp is the expected cost per unit time for scheduled h1 ðtÞ ¼ ðm=ZÞðt=ZÞm1 , where the shape parameter m ¼ 3,
PM. Then, the expected cumulative cost per unit time for the characteristic life parameter Z ¼ 100 and the sampling
the system in the residual life cycle is interval Dt ¼ 0:1. To compute the optimal system relia-
P
N bility threshold, the adjustment factors ðai ; bi Þ, the cost
ðcup tp ð ln RÞ þ csp tp ð1 þ ln RÞÞ parameters ðcup ; csp ; Costr Þand the duration for PM ðtp Þ
cEr ¼ i¼1 . (5) need to be known. Usually, maintenance engineers should
P
N
ðT i þ tp Þ be responsible for the determination of these parameters.
i¼1 For the selection of improvement factors ðai ; bi Þ, please
 ARTICLE IN PRESS
X. Zhou et al. / Reliability Engineering and System Safety 92 (2007) 530–534 533

refer to Refs. [6–8,24,25], where the improvement factors (3) The nominal time interval T i and the nominal PM
ðai ; bi Þ are introduced. number N increase when Costr increases.P This indicates
N
Here, let ai ¼ i=3i þ 7, bi ¼ 12i þ 1=11i þ 1, cup ¼ 150, that the effective life of the system i¼1 T i has a
csp ¼ 10, Costr ¼ 100 and tp ¼ 2. The optimization is positive relationship with the replacement cost.
processed based on simulation. The assumed search ranges
are R 2 ½0:40; 1:00 (step ¼ 0.01) and N 2 ½1; 20. The
simulation results are listed in Table 1. 4. Conclusions and future works
From Table 1, it can be seen that the lowest cost per unit
time in the system residual life is 1.8743 and the optimal A lot of existing CBPM models usually assume that only
system reliability threshold R for scheduled PM is 0.93. A perfect maintenance (either corrective or preventive) is
nominal scheduled replacement will be performed at the performed on the system and the system will restore to as
4th PM action. Fig. 2 gives the corresponding nominal time good as new after PM. This is not practical in reality.
intervals for scheduled PM. It shows that the reliability- The purpose of this paper is to integrate sequential
centered maintenance policy decreases the time intervals imperfect maintenance policy into CBPM. A reliability-
for scheduled PM as PM cycling. This also indicates that centered predictive maintenance policy is proposed for a
the PM is imperfect and the system is subject to a system subject to degradation due to the imperfect
degradation process. maintenance effect. It is assumed that the system is
Other simulation results are listed in Tables 2 and 3 with continuously monitored and the hazard rate function in
the variation of cup and Costr , respectively, while other current maintenance cycle can be derived directly through
parameters are fixed. It can be seen that CBPM. In order to predict the evolution of the system
reliability in different maintenance cycles, a hybrid hazard
(1) With the increase of the reliability threshold R, the rate recursion rule is built up which is more general than
nominal scheduled PM intervals exhibit a gradual the age reduction factor method and hazard rate increase
decrease. This implies that to keep higher reliability for factor method. Whenever the system reliability reaches the
the system, more frequent PM actions are needed.
(2) The optimal reliability threshold R gives a gradual
increase with the increase of unscheduled PM cost rate Table 2
Simulation results when csp ¼ 10, tp ¼ 2 and Costr ¼ 100
cup . This means that the more expensive the unsched-
uled PM, the more frequent the scheduled PM actions. cup R N T1 T2 T3 T4 cEr

90 0.88 4 50.3 44.0 36.0 28.7 1.5677

100 0.90 4 47.1 41.2 33.9 26.9 1.6286
150 0.93 4 41.6 36.3 29.8 23.7 1.8743
Table 1
200 0.95 4 37.1 32.4 26.6 21.1 2.0604
Simulation results when cup ¼ 150, csp ¼ 10, tp ¼ 2 and Costr ¼ 100

R/cEr/N 1 2 3 4 5 6 7 8

0.97 3.8830 2.5210 2.1430 2.0124 1.9808 1.9983 2.0416 2.1000 Table 3
0.96 3.6207 2.3845 2.0452 1.9342 1.9147 1.9394 1.9889 2.0528 Simulation results when cup ¼ 150, csp ¼ 10 and tp ¼ 2
0.95 3.4364 2.2956 1.9891 1.8966 1.8876 1.9208 1.9764 2.0464
0.94 3.3090 2.2420 1.9609 1.8801 1.8807 1.9203 1.9818 2.0568 Costr R N T1 T2 T3 T4 T5 T6 cEr
0.93 3.2183 2.2056 1.9434 1.8743 1.8826 1.9296 1.9981 2.0785
0.92 3.1436 2.1785 1.9347 1.8764 1.8932 1.9464 2.0202 2.1064 40 0.95 3 37.1 32.4 26.6 1.4014
0.91 3.0888 2.1640 1.9339 1.8828 1.9061 1.9649 2.0435 2.1333 100 0.93 4 41.6 36.3 29.8 23.7 1.8743
0.90 3.0448 2.1560 1.9384 1.8969 1.9263 1.9910 2.0748 2.1698 150 0.92 5 43.6 38.1 31.2 24.8 19.6 2.1921
0.89 3.0104 2.1538 1.9507 1.9168 1.9524 2.0223 2.1118 2.2115 250 0.90 6 47.1 41.2 33.9 26.9 21.3 17.0 2.7433

1
Reliability

0.90
PM cycle 1 PM cycle 2 PM cycle 3 PM cycle 4
T1=41.6 T2=36.3 T3=29.8 T4=23.7

Fig. 2. Nominal time intervals for scheduled PM when cup ¼ 150, csp ¼ 10, tp ¼ 2 and Costr ¼ 100.
 ARTICLE IN PRESS
534 X. Zhou et al. / Reliability Engineering and System Safety 92 (2007) 530–534

threshold R, a scheduled imperfect PM action is performed [8] Lie CH, Chun YH. An algorithm for preventive maintenance policy.
on the system. The optimal reliability threshold R is IEEE Trans Reliab 1986;R-35(l):71–5.
[9] Martorell S, Sanchez A, Serdarell V. Age-dependent reliability model
deduced by minimizing the cumulative maintenance cost
considering effects of maintenance and working conditions. Reliab
per unit time in the residual life of the system which is Eng Syst Saf 1999;64(1):19–31.
based on simulation. A discussion is presented to show [10] Levitin G, Lisnianski A. Optimization of imperfect preventive main-
how the optimal schedule depends on different cost tenance for multi-state systems. Reliab Eng Syst Saf 2000;67(2):193–203.
parameters. [11] Zhao YX. On preventive maintenance policy of a critical reliability
Simulation results show that the reliability-centered level for system subject to degradation. Reliab Eng Syst Saf 2003;
79(3):301–8.
predictive maintenance policy decreases the PM time [12] Jardine AKS, Makis V, Banjevic D, Braticevic D, Ennis M. A
intervals and it is more practical than the age-T policy. decision optimization model for condition-based maintenance. J Qual
On the other hand, the predicted nominal PM schedule Maint Eng 1998;4(2):115–21.
also makes it possible for the enterprises to perform PM [13] Jardine AKS, Joseph T, Banjevic D. Optimizing condition-based
actions with near-zero inventory for the spare parts. maintenance decisions for equipment subject to vibration monitoring.
J Qual Maint Eng 1999;5(3):192–202.
Further research still needs to be done based on this [14] Mobley RK. An introduction to predictive maintenance. New York:
maintenance model; for example, how to obtain the Butterworth-Heinemann; 1989.
improvements factors and when the presented model can [15] Hontelez JAM, Burger HH, Wijnmalen DJD. Optimum condition-
be applied effectively. based maintenance policies for deteriorating systems with partial
information. Reliab Eng Syst Saf 1996;51(3):267–74.
[16] Ruey HY. State-age-dependent maintenance policies for deteriorat-
Acknowledgment ing systems with Erlang sojourn time distributions. Reliab Eng Syst
Saf 1997;58(1):55–60.
[17] Chen DY, Trivedi KS. Closed-form analytical results for condition-
This research is supported by the National Natural
based maintenance. Reliab Eng Syst Saf 2002;76(1):43–51.
Science Foundation of China (Grant no. 50128504) and the [18] Wang W. A model to determine the optimal critical level and the
National High-Tech R&D Program for CIMS, China monitoring intervals in condition-based maintenance. Int J Prod Res
(Grant no. 2003AA414120). 2000;38(6):1425–36.
[19] Badı́a FG, Berrade MD, Campos CA. Optimal inspection and
preventive maintenance of units with revealed and unrevealed
References failures. Reliab Eng Syst Saf 2002;78(2):157–63.
[20] Grall A, Bérenguer C, Dieulle L. A condition-based maintenance
[1] Barlow RE, Hunter LC. Optimum preventive maintenance policies. policy for stochastically deteriorating systems. Reliab Eng Syst Saf
Oper Res 1960;8(1):90–100. 2002;76(2):167–80.
[2] Nakagawa T. Optimal policy of continuous and discrete replacement [21] Lee J. Teleservice engineering in manufacturing: challenges and
with minimal repair at failure. Nav Res Logist Q 1984;31(4):543–50. opportunities. Int J Mach Tools Manuf 1998;38(8):901–10.
[3] Canfield RV. Cost optimization of periodic preventive maintenance. [22] Center for Intelligent Maintenance System. http://www.uwm.edu/
IEEE Trans Reliab 1986;35(1):78–81. CEAS/IMS
[4] Sheu S, Griffith WS, Nakagawa T. Extended optimal replacement [23] Dieulle L, Bérenguer C, Grall A, Roussignol M. Sequential
model with random minimal repair costs. Eur J Oper Res 1995;85(3): condition-based maintenance scheduling for a deteriorating system.
636–49. Eur J Oper Res 2003;150(3):451–61.
[5] Pham H, Wang HZ. Imperfect maintenance. Eur J Oper Res 1996; [24] Jayabalan V, Chaudhuri D. Cost optimization of maintenance
94(3):425–38. scheduling for a system with assured reliability. IEEE Trans Reliab
[6] Malik MAK. Reliable preventive maintenance policy. AIIE Trans 1992;41(1):21–5.
1979;11(3):221–8. [25] Chan JK, Shaw L. Modeling repairable systems with failure rates that
[7] Nakagawa T. Sequential imperfect preventive maintenance policies. depend on age and maintenance. IEEE Trans Reliab 1993;42(4):
IEEE Trans Reliab 1988;37(3):295–8. 566–71.