2015 Book MaintenanceOvertimePoliciesInR PDF

Lecture Notes in Production Engineering
Toshio Nakagawa
Xufeng Zhao
Maintenance
Overtime Policies
in Reliability
Theory
Models with Random Working Cycles
More information about this series at http://www.springer.com/series/10642
Toshio Nakagawa Xufeng Zhao
•
Maintenance Overtime
Policies in Reliability Theory
Models with Random Working Cycles
123
Toshio Nakagawa Xufeng Zhao
Aichi Institute of Technology Aichi Institute of Technology
Toyota Toyota
Japan Japan
and
Qatar University
Doha
Qatar
ISSN 2194-0525 ISSN 2194-0533 (electronic)

ISBN 978-3-319-20812-1 ISBN 978-3-319-20813-8 (eBook)
DOI 10.1007/978-3-319-20813-8
Library of Congress Control Number: 2015943037
Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
Preface
Maintenance theory is basically how to make appropriate maintenances at suitable

times to prevent failures for an objective system. Classic maintenance theory was
mathematically established by Barlow and Proschan [1] 50 years ago. After that, a
great part of maintenance models and a variety of maintenance policies, such as
preventive replacement, preventive maintenance, inspection, etc., were summarized
systematically in [2, 3]. To understand maintenance theory and master reliability
engineering, we have to learn probability, statistics, and stochastic processes because
reliability phenomena occurs randomly. The elementary textbook for graduate stu-
dents, researchers, and engineers to learn readily reliability theory was published [4],
which was written in an easy style on stochastic processes, and all examples are
quoted fittingly from reliability and maintenance models.
It is well known generally in maintenance theory that an optimum replacement
policy should be non-random [1], i.e., preventive replacement is always done at a
deterministic time T which is taken to be a constant. However, most production and
computer systems have variable working cycles and processing times. For such
systems, it would be impossible or impractical in a strict periodic fashion to interrupt
their running work for scheduled maintenances [1]. From this viewpoint, random
maintenances of age replacement, periodic replacement, and inspection have been
proposed, and their optimum policies were discussed and compared with each other
theoretically and numerically [5]. As practical examples, random maintenance was
applied to parallel systems, scheduling problems, damage models, and so on.
The concepts of maintenance first, maintenance last, and maintenance overtime
were newly introduced in [5]. Especially, combining maintenances with time and
cycles, maintenance overtime is proposed. For example, age replacement overtime is
to replace a unit at the first completion of working cycles over time T. In this case, a
unit continues to work without interruption by maintenance, and its maintenance
considering working cycles is made simply without counting the number of cycles.
The main object of this Lecture Note is to investigate the relations of mainte-
nance overtime with other maintenance policies and to add some new interesting
results: Chap. 1 summarizes age replacement in which the unit is replaced before
failure at a planned time T and at a working cycle N. We take up age and random
v
vi Preface
replacements as the basic policies in Sect. 1.1. The expected cost rates are obtained,
optimum policies which minimize them are derived analytically, and two policies
are compared. Replacement overtime is proposed in Sect. 1.2, in which the unit is
replaced at time T or at the first completion of working cycles. Replacement
overtime with age and random replacements are compared analytically and
numerically in Sect. 1.3. When three replacement costs are the same, replacement
overtime is better than random replacement, however, it is not rather than age
replacement. Replacement overtime last is proposed in Sect. 1.4, in which the unit
is replaced at the first completion of working cycles over time T or at cycle N,
whichever occurs last. Replacement overtime last is better than random replace-
ment, however, it is not rather than replacement overtime. Compared to replace-
ment overtime first, if replacement time T is given and is small, overtime last is
better than overtime first, and conversely, if T is large, it is not rather than
replacement first. We consider age replacement for a finite interval S in Sect. 1.5
and for a job with a specified number N in Sect. 1.6. All results in the previous
sections are rewritten effectively when a finite interval and a specified number are
constant and random variables. Finally, we apply replacement overtime to a parallel
system with constant and random number of units in Sect. 1.7.
Chapter 2 summarizes periodic replacement in which the unit undergoes mini-
mal repair at each failure, and is replaced at a planned time T, at a working cycle N
or at a failure number K. First, we take up periodic and random replacements in
which the unit is replaced at time T or at cycle N, whichever occurs first, and two
policies are compared in Sect. 2.1. Replacement overtime is proposed in Sect. 2.2,
in which the unit is replaced at time T or at the first completion of working cycles,
and is compared with periodic and random replacements in Sect. 2.3. When three
replacement costs are the same, replacement overtime is better than random
replacement, however, it is not rather than periodic replacement. Replacement
overtime last is proposed in Sect. 2.4, in which the unit is replaced at the first
completion of working cycles over time T or at cycle N, whichever occurs last.
Replacement overtime last is better than random replacement, however, it is not
rather than replacement overtime. Furthermore, replacing cycle N with failure K,
we consider replacement overtime first and last in Sect. 2.5. Two overtime policies
are proposed in Sect. 2.6, in which the unit is replaced at the first failure over cycle
N and at the first working cycle over failure K. Finally, we take up two imperfect
preventive maintenance overtime policies in which the preventive maintenance is
done at the first completion of working cycles over time T in Sect. 2.7.
Chapter 3 investigates inspection first, inspection last, and inspection overtime
when the failure time is exponentially distributed. We take up periodic and random
inspections in which the unit is checked at periodic times kT and successive
working cycles Yj , and two polices are compared in Sect. 3.1. Inspection first,
inspection last, and inspection overtime are proposed, and are compared to other
policies in Sects. 3.2 and 3.3. When the failure time is not exponential, we show
how to obtain expected costs of inspection first, last and overtime in Sect. 3.4.
Preface vii
Finally, we consider the backup policies and make similar arguments to inspection
policies in Sects. 3.5 and 3.6.
Chapter 4 proposes generalized replacement policies with three variables in
which the unit is replaced at time T, at cycle N or at failure K. We take up
replacement first, replacement last, and replacement middle in Sect. 4.1, where
replacement middle is an entirely new model. Age replacement is the best one
among three policies when all of replacement costs are the same. Two kinds of
replacement overtime with working cycle and failure number are considered in
Sects. 4.2–4.4, and their expected costs are obtained. Finally, replacement first and
last with n variables are proposed in Sect. 4.5, in which the unit is replaced at time
T or at random times Y1 ; Y2 ; ; Yn .
New policies such as maintenance first, last, middle, and overtime are formally
proposed in this Lecture Note, some of which have been analytically optimized,
however, some of them have not been done yet. For example, it is difficult to derive
theoretically optimum T , N and K of replacement policies with three variables
when replacement costs are different, even if to compute simply them numerically.
Such problems would offer new theoretically topics in maintenance theory as fur-
ther studies.
Most systems fail due to many kinds of causes and factors such as calendar or
operating times, running or flight distance, amount of damage or crack, number of
uses, works, shocks or faults, and so on. Based on these, we have to form a fitting
schedule with several kinds of maintenance measures to avoid the above failures.
Furthermore, it should be an important problem to make suitable maintenances for
systems due to several kinds of failures, i.e., systems fail according to several failure
distributions. By modifying and extending the proposed maintenance policies to fit
existing conditions, their applications could be found in actual systems. Finally,
extended failure rates appearing in this Lecture Note are summarized in Appendix
[5, p. 227].
Toyota Toshio Nakagawa

May 2015 Xufeng Zhao
References
1. Barlow RE, Proschan F (1965) Mathematical theory of reliability. Wiley, New York
2. Nakagawa T (2005) Maintenance theory of reliability. Springer, London
3. Nakagawa T (2008) Advanced reliability models and maintenance policies. Springer, London
4. Nakagawa T (2011) Stochastic processes with applications to reliability theory. Springer,
London
5. Nakagawa T (2014) Random maintenance policies. Springer, London
Contents
1 Age Replacement Overtime. . . . . . . . ........... . . . . . . . . . . . 1

1.1 Age and Random Replacements. . ........... . . . . . . . . . . . 2
1.2 Replacement Overtime . . . . . . . . ........... . . . . . . . . . . . 6
1.3 Comparisons of Age and Random Replacements . . . . . . . . . . . . 8
1.4 Replacement Overtime Last. . . . . ........... . . . . . . . . . . . 12
1.5 Finite Interval . . . . . . . . . . . . . . ........... . . . . . . . . . . . 16
1.6 Working Number. . . . . . . . . . . . ........... . . . . . . . . . . . 24
1.7 Parallel System . . . . . . . . . . . . . ........... . . . . . . . . . . . 27
References . . . . . . . . . . . . . . . . . . . . ........... . . . . . . . . . . . 31
2 Periodic Replacement Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.1 Periodic and Random Replacements . . . . . . . . . . . . . . . . . . . . . 34
2.2 Replacement Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Comparisons of Periodic and Random Replacements . . . . . . . . . 41
2.4 Replacement Overtime Last. . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5 Replacement Overtime with Number of Failures . . . . . . . . . . . . 46
2.5.1 Replacement Overtime First with Number of Failures . . . 46
2.5.2 Replacement Overtime Last with Number of Failures. . . . 50
2.6 Replacement Overnumber . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.6.1 Replacement over Number N . . . . . . . . . . . . . . . . . . . . 52
2.6.2 Replacement over Number K . . . . . . . . . . . . . . . . . . . . 54
2.7 Preventive Maintenance Overtime . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3 Inspection Overtime . . . . . . . . . . . . . . . . ... .... . . . . . . . . . . . . 59

3.1 Periodic and Random Inspections. . . . ... .... . . . . . . . . . . . . 60
3.2 Inspection First and Last. . . . . . . . . . ... .... . . . . . . . . . . . . 64
3.2.1 Inspection First . . . . . . . . . . . ... .... . . . . . . . . . . . . 64
3.2.2 Inspection Last . . . . . . . . . . . ... .... . . . . . . . . . . . . 67
3.2.3 Comparison of Inspection First and Last. . . . . . . . . . . . . 69
ix
x Contents
3.3 Inspection Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.3.1 Comparisons with Other Policies . . . . . . . . . . . . . . . . . . 73
3.4 General Failure Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5 Periodic and Random Backup . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.6 Backup Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4 Replacement Overtime with Three Variables . . . . . . . . . . . . . . . . . 85

4.1 Replacement with Three Variables . . . . . . . . . . . . . . . . . . . . . . 86
4.2 Replacement Overtime with Working Cycle . . . . . . . . . . . . . . . 91
4.3 Replacement Overtime with Failure Number . . . . . . . . . . . . . . . 95
4.4 Modified Replacement with Three Variables . . . . . . . . . . . . . . . 97
4.5 General Replacement Policies . . . . . . . . . . . . . . . . . . . . . . . . . 100
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Chapter 1
Age Replacement Overtime
We consider a unit operating for a job with random working cycles, in which case,
it would be impossible or impractical to replace the unit in a strict or deterministic
fashion [1, p. 72], [2, p. 245]. From such a viewpoint, we have already proposed
replacement overtime in which the unit is replaced at the first completion of working
cycles over a planned time T [3, p. 34, 8] .
This chapter summarizes optimum policies for age replacement overtime based
on our research works and adds some new results: In Sect. 1.1, combining age
and random replacements, we introduce age replacement first in which the unit is
replaced before failure at time T or at working cycle N , whichever occurs first
[3, p. 28, 4–7] . We obtain the expected cost rate and derive optimum policy to min-
imize it. Using this result, we compare age replacement with time T and random
replacement with cycle N . It is theoretically shown that when both replacement
costs are the same, age replacement is better than random replacement, as previously
expected. The derived results and analytical methods in this section are fundamental
to all of those in this Lecture Note.
In Sect. 1.2, we take up replacement overtime in which the unit is replaced before
failure at the first completion of working cycles over a planned time T [3, p. 34, 8] .
The expected cost rate is obtained and an optimum time TO∗ to minimize it is derived.
In Sect. 1.3, we propose replacement overtime first and compare replacement over-
time with age and random replacements discussed in Sect. 1.1. It is shown that when
three replacement costs are the same, replacement overtime is better than random
replacement, however, age replacement is still best among these three policies. The
fact that replacement overtime is superior to random replacement would be interest-
ing theoretically and practically, because both preventive replacements are done at
random times.
In Sect. 1.4, as one of extended models of replacement overtime, we propose
replacement overtime last in which the unit is replaced before failure at the first
completion of working cycles over time T or at cycle N , whichever occurs last [3,
p. 46, 9] . The expected cost rate is obtained and an optimum policy to minimize it
is discussed. Furthermore, we compare replacement overtime last with replacement
overtime first studied in Sect. 1.3. It is of great interest that if a replacement time T is
© Springer International Publishing Switzerland 2015 1
T. Nakagawa and X. Zhao, Maintenance Overtime Policies in Reliability Theory,
Lecture Notes in Production Engineering, DOI 10.1007/978-3-319-20813-8_1
2 1 Age Replacement Overtime
less than an optimum TO∗ then replacement overtime last is better than replacement
overtime first, and conversely, if T is greater than TO∗ then replacement overtime first
is better than replacement overtime last. This means that if a specified replacement
time T is large, we should adopt replacement first, and if it is small, we should adopt
replacement last.
Next, we consider replacement overtime for a finite interval S in Sect. 1.5 [10]:
The expected cost rate is obtained, and the optimum policy to minimize it is derived.
It is shown that if a finite interval S is greater than an optimum time T A∗ of age
replacement then there exists a finite replacement time which is less than S. When
S is not constant but a random variable, we make similar discussions. Furthermore,
optimum replacement policies for constant and random working cycle N are dis-
cussed in Sect. 1.6. Finally, replacement overtime for a parallel system with n units
is considered and an optimum policy is derived analytically when the failure time is
exponential in Sect. 1.7 [3, p. 141, 11] .
Throughout this Chapter,
∞ it is assumed that the unit has a failure distribution F(t)
with finite mean μ ≡ 0 F(t)dt < ∞, where Φ(t) ≡ 1 − Φ(t) for any distribution
t
Φ(t). When F(t) has a density function f (t) ≡ dF(t)/dt, i.e., F(t) = 0 f (u)du,
the failure rate h(t) ≡ f (t)/F(t) for F(t) < 1 is supposed to increase strictly from
h(0) = 0 to h(∞) ≡ limt→∞ h(t), except that F(t) is exponential.
1.1 Age and Random Replacements
Suppose that a unit has to operate for a job with random working cycles Y j ( j =
1, 2, . . .) in Fig. 1.1, where Y0 ≡ 0, S j ≡ Y1 + Y2 + · · · + Y j ( j = 1, 2, . . .) and
S0 ≡ 0. It is assumed that Y j are independent and have an identical distribution
G(t) ≡ Pr{Y j ≤ t} with finite mean 1/θ (0 < θ < ∞), where G ( j) (t) ≡ Pr{Y1 +
Y2 + · · · + Y j ≤ t} ( j = 1, 2, . . .) is the j−fold Stieltjes convolution of G(t)
with itself and G (0) (t) ≡ 1 for t ≥ 0, and g ( j) (t) ≡ dG ( j) (t)/dt ( j = 1, 2, . . .),
g(t) ≡ dG(t)/dt and g (0) (t) ≡ 0.
As preventive replacement, the unit is planned to be replaced at time T (0 <
T ≤ ∞), at cycle N (N = 1, 2, . . .), or at failure, whichever occurs first. This is
called age replacement first [3, p. 42]. Then, the probability that the unit is replaced
at time T is
[1 − G (N ) (T )]F(T ), (1.1)
Fig. 1.1 A job with working Y1 Y2 Y3 Yj

cycles
0 S1 S2 S3 S j-1 Sj
1.1 Age and Random Replacements 3
the probability that it is replaced at cycle N is

T
F(t)dG (N ) (t), (1.2)
0
and the probability that it is replaced at failure is

T
[1 − G (N ) (t)]dF(t), (1.3)
0
where note that (1.1) + (1.2) + (1.3) = 1. The mean time to replacement is
T T
(N ) (N )
T [1 − G (T )]F(T ) + t F(t)dG (t) + t[1 − G (N ) (t)]dF(t)
0 0
T
= [1 − G (N ) (t)]F(t)dt. (1.4)
0
Therefore, the expected cost rate is

T
cT + (c F − cT ) 0 [1 − G (N ) (t)]dF(t)
T
+(c N − cT ) 0 F(t)dG (N ) (t)
C F (T, N ) = T , (1.5)
(N ) (t)]F(t)dt
0 [1 − G
where c F = replacement cost at failure, cT = replacement cost at time T , and c N =

replacement cost at cycle N for any N ≥ 1 with c F > cT and c F > c N .
When the unit is replaced only at time T for the standard age replacement,
c F − (c F − cT )F(T )
C A (T ) ≡ lim C F (T, N ) = T . (1.6)
N →∞ F(t)dt
0
An optimum policy which minimizes C A (T ) is [1, p. 87], [2, p. 74]:

(i) If h(∞) > c F /[(c F − cT )μ], then there exists a finite and unique T A∗ (0 < T A∗ <
∞) which satisfies
T cT
h(T ) F(t)dt − F(T ) = , (1.7)
0 c F − cT
and the resulting cost rate is
C A (T A∗ ) = (c F − cT )h(T A∗ ). (1.8)
(ii) If h(∞) ≤ c F /[(c F − cT )μ], then T A∗ = ∞, i.e., the unit is replaced only at
failure, and the expected cost rate is
cF
C A (∞) ≡ lim C A (T ) = . (1.9)
T →∞ μ
When the unit is replaced before failure only at cycle N for an age random
replacement,
∞
c F − (c F − c N ) 0 G (N ) (t)dF(t)
C R (N ) ≡ lim C F (T, N ) = ∞ (N = 1, 2, . . .).
(N ) (t)]F(t)dt
0 [1 − G
T →∞
(1.10)
An optimum policy which minimizes C R (N ) is [3, p. 44]:

(i) If Q 1 (N ) increases strictly to Q 1 (∞) and Q 1 (∞) > c F /[(c F − c N )μ], then
there exists a finite and unique minimum N R∗ (1 ≤ N R∗ < ∞) which satisfies
∞ ∞ cN
Q 1 (N ) [1 − G (N ) (t)]F(t)dt − [1 − G (N ) (t)]dF(t) ≥ ,
0 0 cF − cN
(1.11)
where
T
[G (N ) (t) − G (N +1) (t)]dF(t)
Q 1 (T, N ) ≡ 0T ≤ h(T ),
0 [G (N ) (t) − G (N +1) (t)]F(t)dt
and
∞
[G (N ) (t) − G (N +1) (t)]dF(t)
Q 1 (N ) ≡ lim Q 1 (T, N ) = 0∞ .
T →∞
0 [G (N ) (t) − G (N +1) (t)]F(t)dt
(ii) If Q 1 (∞) ≤ c F /[(c F − c N )μ], then N R∗ = ∞, and the expected cost rate is
given in (1.9).

In particular, when G(t) = 1 − e−θt , i.e., G (N ) (t) ≡ ∞j=N [(θt) /j!]e
j −θt (N =
0, 1, 2, . . .), from Appendix 2.1,

T
(θt) N e−θt dF(t)
Q 1 (T, N ) = 0T (1.12)
0 (θt) N e−θt F(t)dt
increases strictly with N to h(T ), and Q 1 (N ) increases strictly to h(∞). Therefore,

if h(∞) > c F /[(c F − c N )μ], then a finite N R∗ (1 ≤ N R∗ < ∞) satisfying (1.11)
exists.
1.1 Age and Random Replacements 5
Next, we discuss theoretically which policy is better to replace the unit at time T
or at cycle N . For this purpose, we find both optimum TF∗ and N F∗ which minimize
the expected cost rate C F (T, N ) in (1.5). Differentiating C F (T, N ) with respect to
T and setting it equal to zero,
T T
(N ) (N )
(c F − cT ) h(T ) [1 − G (t)]F(t)dt − [1 − G (t)]dF(t)
0 0
T T
− (cT − c N ) r N (T ) [1 − G (N ) (t)]F(t)dt − F(t)dG (N ) (t) = cT ,
0 0
(1.13)
where
g (N ) (t)
r N (t) ≡ (N = 1, 2, . . .).
1 − G (N ) (t)
From the inequality C F (T, N + 1) − C F (T, N ) ≥ 0,

T
(N )
[1 − G (t)]F(t) (c F − cT )[Q 1 (T, N ) − h(t)]
0
T (N ) (u) − G (N +1) (u)]
0 F(u)d[G
+ (cT − c N ) T + r N (t) dt ≥ cT . (1.14)
(N ) (u) − G (N +1) (u)]F(u)du
0 [G
In addition, substituting (1.13) for (1.14), (1.14) becomes
(c F − cT )[Q 1 (T, N ) − h(T )]

T
(N ) (t) − G (N +1) (t)]

0 F(t)d[G
+ (cT − c N ) T + r N (T ) ≥ 0. (1.15)
(N ) (t) − G (N +1) (t)]F(t)dt
0 [G
Thus, when cT ≤ c N , there does not exist any finite N F∗ for T > 0 because
Q 1 (T, N ) ≤ h(T ), i.e., N F∗ = ∞. In this case, the unit should be replaced only
at time T .
When G(t) = 1 − e−θt , i.e., G (N ) (t) = ∞j=N [(θt) /j!]e
j −θt (N = 0, 1, . . .) and
cT > c N , (1.15) becomes
(c F − cT )[Q 1 (T, N ) − h(T )]

(θT ) N e−θT F(T )
+ (cT − c N ) Q 1 (T, N ) + r N (T ) + T ≥ 0, (1.16)
where Q 1 (T, N ) is given in (1.12), and
θ(θt) N −1 /(N − 1)!

r N (t) ≡ N −1 (N = 1, 2, . . .),
j=0 [(θt) /j!]
j
which decreases with N to 0 from Appendix 1.1. Because Q 1 (T, N ) increases with
N to h(T ), there exits a finite N F∗ (1 ≤ N F∗ < ∞) which satisfies (1.16) for T > 0.
Furthermore, the left-hand side of (1.13) goes to
∞
[1 − G (N ) (t)]F(t) {(c F − cT )[h(∞) − h(t)] + (cT − c N )[r N (t) − θ]} dt
0
(1.17)
as T → ∞, because r N (T ) increases with T to θ. Therefore, if (1.17) is greater

than cT , then there exists a finite TF∗ (0 < TF∗ < ∞) which satisfies (1.13). It can be
clearly seen that if h(∞) = ∞, both finite TF∗ and N F∗ exist in case of cT > c N .
1.2 Replacement Overtime
It might be wise to replace an operating unit at the completion of some working

cycle even if T comes because it continues to operate for some job without stopping.
Suppose that the unit is replaced before failure at the first completion of working
cycles Y j ( j = 1, 2, . . .) over a planned time T (0 ≤ T < ∞). This is called age
replacement overtime.
The probability that the unit is replaced at the first completion of working cycles
over time T is
∞
T ∞
F(t + u)dG(u) dG ( j) (t), (1.18)
j=0 0 T −t
the probability that it is replaced at failure before time T is

∞
T
[G ( j) (t) − G ( j+1) (t)]dF(t) = F(T ), (1.19)
j=0 0
and the probability that it is replaced at failure over time T is

∞
T ∞
[F(t + u) − F(T )]dG(u) dG ( j) (t), (1.20)
j=0 0 T −t
1.2 Replacement Overtime 7
where note that (1.18) + (1.19) + (1.20) = 1. Thus, the mean time to replacement is
∞
T ∞ T
(t + u)F(t + u)dG(u) dG ( j) (t) + tdF(t)
j=0 0 T −t 0
T ∞
∞ t+u
+ ydF(y) dG(u) dG ( j) (t)
j=0 0 T −t T
T ∞ T
∞
= F(t)dt + G(u − t)F(u)du dG ( j) (t). (1.21)
0 j=0 0 T

T ∞
c F − (c F − c O ) ∞ ( j)
j=0 0 [ T −t F(t + u)dG(u)]dG (t)
C O (T ) = T ∞ T ∞ , (1.22)
F(t)dt + [ G(u − t)F(u)du]dG ( j) (t)
0 j=0 0 T
where c O = replacement cost over time T with c O < c F and c F is given in (1.5). In
particular,
C O (0) ≡ lim C O (T ) = C R (1)

T →0
in (1.10) when c O = c N , and
C O (∞) ≡ lim C O (T ) = C A (∞)

T →∞
in (1.9).
When G(t) = 1 − e−θt ,
∞
c F − (c F − c O ) T θe−θ(t−T ) F(t)dt
C O (T ) = T ∞ . (1.23)
−θ(t−T ) F(t)dt
0 F(t)dt + T e
We find an optimum TO∗ which minimizes C O (T ). Differentiating C O (T ) with

respect to T and setting it equal to zero,
T
1 (T ) cO
Q F(t)dt − F(T ) = , (1.24)
0 cF − cO
where
∞ N −θt dF(t)
T (θt) e
Q 1 (T, N ) ≡ ∞ ≥ h(T ) (N = 0, 1, 2, . . .),
N −θt F(t)dt
T (θt) e
and
∞ −θt
T e dF(t)
Q 1 (T ) ≡ Q 1 (T, 0) = ∞ .
−θt
T e F(t)dt
1 (T ) also
From Appendix 1.4, when the failure rate h(t) increases strictly to h(∞), Q
increases strictly with T to h(∞), and decreases with θ to h(T ) for 0 ≤ T < ∞.
An optimum policy which minimizes C O (T ) is [3, p. 36]
(i) If h(∞) > c F /[(c F −c O )μ], then there exists a finite and unique TO∗ (0 < TO∗ <
∞) which satisfies (1.24), and the resulting cost rate is
c F − (c F − c O )F(TO∗ )
1 (TO∗ ) =
C O (TO∗ ) = (c F − c O ) Q . (1.25)
TO∗
0 F(t)dt
(ii) If h(∞) ≤ c F /[(c F − c O )μ], then TO∗ = ∞, and the expected cost rate is given
in (1.9).
It is of great interest that when G(t) = 1 − e−θt , if h(∞) > c F /[(c F − ci )μ]
(i = T, N , O), the finite replacement times T A∗ , N R∗ and TO∗ exist. Comparing (1.24)
with (1.7) when c O = cT , we obtain TO∗ ≤ T A∗ , and TO∗ increases with θ to T A∗ given
in (1.7).
1.3 Comparisons of Age and Random Replacements
We compare theoretically replacement overtime to age replacement policies with time

T (0 < T < ∞) and cycle N (N = 1, 2, . . .), respectively, when cT = c N = c O
and h(∞) = ∞. In this case, all of finite T A∗ , N R∗ and TO∗ always exist. Furthermore,
because T A∗ is an optimum solution of minimizing C A (T ) in (1.6), C O (TO∗ ) is greater
than C A (T A∗ ) in (1.8), i.e., age replacement is better than replacement overtime. If
c O < cT , then replacement overtime might be better rather than age replacement,
and we would compute numerically C A (T A∗ ) in (1.8) and C O (TO∗ ) in (1.25), and
compare them for different costs c F , cT and c O .
Next, we compare replacement overtime with random replacement. Generally, it
would be much difficult to compare directly continuous replacement and discrete
replacement. For the above purpose, we propose the following extended model with
time T and cycle N to combine random replacement and overtime replacement,
which is called replacement overtime first: The unit is replaced before failure at
cycle N (N = 1, 2, . . .) before time T (0 < T ≤ ∞) or at the first completion of
working cycles over time T , whichever occurs first. Then, the probability that the
unit is replaced at cycle N is
1.3 Comparisons of Age and Random Replacements 9
T
F(t)dG (N ) (t), (1.26)
0
the probability that it is replaced at the first completion of working cycles over time
T is

N −1 T ∞
F(t + u)dG(u) dG ( j) (t), (1.27)
j=0 0 T −t

T
[1 − G (N ) (t)]dF(t), (1.28)
0

N −1 T ∞
[F(t + u) − F(T )]dG(u) dG ( j) (t), (1.29)
j=0 0 T −t
where note that (1.26) + (1.27) + (1.28) + (1.29) = 1. The mean time to replacement is
T
N −1 T ∞
(N )
t F(t)dG (t) + (t + u)F(t + u)dG(u) dG ( j) (t)
0 j=0 0 T −t
T
N −1 T ∞ t+u
(N )
+ t[1 − G (t)]dF(t) + ydF(y) dG(u) dG ( j) (t)
0 j=0 0 T −t T
T
N −1 T ∞
= [1 − G (N ) (t)]F(t)dt + G(u − t)F(u)du dG ( j) (t). (1.30)
0 j=0 0 T

T
c F − (c F − c N ) 0 F(t)dG (N ) (t)
−1 T ∞ ( j)
−(c F − c O ) Nj=0 0 [ T −t F(t + u)dG(u)]dG (t)
C O F (T, N ) = T , (1.31)
F(t)[1 − G (N ) (t)]dt
0
−1 T ∞ ( j)
+ Nj=0 0 [ T F(u)G(u − t)du]dG (t)
where c F , c N and c O are given in (1.5) and (1.22). In particular, lim T →∞ C O F

(T, N ) = C R (N ) in (1.10) and lim N →∞ C O F (T, N ) = C O (T ) in (1.22), which
means that this policy combines replacement overtime and random replacement.
When G(t) = 1 − e−θt (0 < θ < ∞) and c O = c N , (1.31) is rewritten as
C O F (T, N ) = − θ(c F − c O )
−2 T −θt F(t)dt
c F + (c F − c O )θ Nj=0 0 [(θt) /j!]e
j
+ N −1 ∞ T ,
−θt F(t)dt + −θt F(t)dt}
j=0 {[(θT ) /j!] T e 0 [(θt) /j!]e
j j
(1.32)
−1
where j=0 ≡ 0 and note that
T
(θT ) N ∞ −θt (θt) N −θt
e F(t)dt + e F(t)dt
N! T 0 N!

T θ(θt) N −1 ∞ −θu
t e F(u)du dt (N = 1, 2, . . .),
= 0∞ (N −1)!
−θt F(t)dt
0 e (N = 0).
We discuss optimum TO∗ F and N O∗ F which minimize C O F (T, N ) in (1.32).

Differentiating C O F (T, N ) with respect to T and setting it equal to zero,

N −1 T
N −1 T
1 (T ) (θt) j −θt (θt) j −θt cO
Q e F(t)dt − e dF(t) = . (1.33)
j=0 0 j! 0j=0
j! c F − cO
1 (T ) increases strictly to h(∞), the left-hand side of (1.33) increases

Recalling that Q
strictly with T to

N −1 ∞
(θt) j −θt
L 1 (N ) ≡ e F(t)[h(∞) − h(t)]dt.
0 j!
j=0
Therefore, if L 1 (N ) > c O /(c F − c O ), then there exists a finite and unique TO∗ F (0 <
TO∗ F < ∞) which satisfies (1.33), and the resulting cost rate is
1 (TO∗ F ).
C O F (TO∗ F , N ) = (c F − c O ) Q (1.34)
In addition, because the left-hand side of (1.33) increases with N to that of (1.24),
TO∗ F decreases with N to TO∗ given in (1.24). So that, from (1.34), optimum TO∗ F and
N O∗ F which minimize C O F (T, N ) is TO∗ F = TO∗ and N O∗ F = ∞, i.e., replacement
overtime is better than random replacement.
We can show by another method that replacement overtime is better than ran-
dom replacement when both replacement costs are the same. From the inequality
C O F (T, N + 1) − C O F (T, N ) ≥ 0,
1.3 Comparisons of Age and Random Replacements 11

N −1 ∞
(θT ) j T (θt) j −θt
Q 2 (T, N − 1) e−θt F(t)dt + e F(t)dt
j! T 0 j!
j=0

N −1 ∞
(θT ) j −θt
T (θt) j −θt cO
− e dF(t) + e dF(t) ≥ , (1.35)
j! T 0 j! cF − cO
j=0
where
T
N ∞ −θu dF(u)]dt
0 (θt) [ t e
Q 2 (T, N ) ≡ T .
N ∞ −θu F(u)du]dt
0 (θt) [ t e
1 (T ). So
It is shown in Appendix 2.4 that Q 2 (T, N ) increases strictly with N to Q
that, letting L 2 (T, N ) be the left-hand side of (1.35),
L 2 (T, N + 1) − L 2 (T, N ) = [Q 2 (T, N ) − Q 2 (T, N − 1)]

N
T
(θT ) j ∞ −θt (θt) j −θt
× e F(t)dt + e F(t)dt > 0,
j! T 0 j!
j=0
and
T
1 (T )
lim L 2 (T, N ) = Q F(t)dt − F(T ),
N →∞ 0
which agrees with the left-hand side of (1.24). Therefore, any finite N O∗ F does not
exist for optimum TO∗ , i.e., optimum TO∗ F and N O∗ F which minimize C O F (T, N ) in
(1.32) are TO∗ F = TO∗ and N O∗ F = ∞.
From the above discussions, if c N ≥ c O , then replacement overtime is better
than random replacement. However, if c N < c O , then random replacement might be
better rather than replacement overtime. In this case, we would compute numerically
C R (N R∗ ) in (1.10) and C O (TO∗ ) in (1.25), and compare them for different costs c N
and c O .
When F(t) = 1 − e−(t/10) and G(t) = 1 − e−t , Table 1.1 presents optimum T A∗ ,
2
∗ ∗
TO , N R and their expected cost rates. This indicates that when three replacement
costs are the same, C R (N R∗ ) > C O (TO∗ ) > C A (T A∗ ) and TO∗ < T A∗ , as shown in the
above results. This shows that TO∗ and C O (TO∗ ) approach to T A∗ and C A (T A∗ ) as ci /c F
is larger, N R∗ /θ is roughly the same as T A∗ , and also, TO∗ < T A∗ < TO∗ + 1/θ. So that, if
1/θ becomes smaller, both TO∗ and N R∗ /θ approach exactly to T A∗ . On the other hand,
when replacement costs are not the same, for example, when cT /c F = 0.02 and
c O /c F = 0.01, C O (TO∗ ) < C A (T A∗ ), and when c O /c F = 0.02 and c N /c F = 0.01,
C R (N R∗ ) < C O (TO∗ ). However, even when replacement costs are different, if they
are large, the three policies are not so different.
Table 1.1 Optimum T A∗ , TO∗ , N R∗ and their expected cost rates when F(t) = 1 − e−(t/10) and
2
G(t) = 1 − e−t for ci /c F (i = T, O, N )

ci /c F T A∗ C A (T A∗ )/c F TO∗ C O (TO∗ )/c F N R∗ C R (N R∗ )/c F
0.01 1.006 0.020 0.431 0.027 1 0.029
0.02 1.431 0.028 0.767 0.034 2 0.038
0.05 2.304 0.044 1.548 0.047 2 0.053
0.10 3.365 0.061 2.563 0.063 4 0.068
0.20 5.107 0.082 4.283 0.083 6 0.087
0.50 10.908 0.109 10.112 0.109 13 0.111
1.4 Replacement Overtime Last
We have already obtained the expected cost rate of replacement overtime first in
which the unit is replaced at cycle N before time T in Sect. 1.3. Next, we propose the
following extended replacement with time T (0 ≤ T < ∞) and cycle N (0, 1, 2, . . .):
The unit is replaced at the first completion of working cycles over time T or at cycle
N after time T , whichever occurs last, i.e., it is replaced at cycle N after time T or
over time T after N cycles. This is called replacement overtime last.
The probability that the unit is replaced at cycle N is
∞
F(t)dG (N ) (t), (1.36)
T
the probability that it is replaced at the first completion of working cycles over
time T is
∞
T ∞
F(t + u)dG(u) dG ( j) (t), (1.37)
j=N 0 T −t
F(T ), (1.38)

∞
T ∞
[F(t + u) − F(T )]dG(u) dG ( j) (t)
j=N 0 T −t
∞
+ [F(t) − F(T )]dG (N ) (t), (1.39)
T
1.4 Replacement Overtime Last 13
where note that (1.36)+(1.37)+(1.38)+(1.39) ≡ 1. The mean time to replacement is

∞ ∞
T ∞
t F(t)dG (N ) (t) + (t + u)F(t + u)dG(u) dG ( j) (t)
T j=N 0 T −t
T ∞ t
+ tdF(t) + udF(u) dG (N ) (t)
0 T T
∞
T ∞ t+u
j=N 0 T −t T
T ∞
= F(t)dt + [1 − G (N ) (t)]F(t)dt
0 T
∞
T ∞
+ G(u − t)F(u)du dG ( j) (t). (1.40)
j=N 0 T

∞
c F − (c F − c N ) T F(t)dG (N ) (t)
T ∞
−(c F − c O ) ∞ ( j)
j=N 0 [ T −t F(t + u)dG(u)]dG (t)
C O L (T, N ) = T ∞ . (1.41)
F(t)dt + T [1 − G (N ) (t)]F(t)dt
0 T ∞ ( j)
+ ∞ j=N 0 T G(u − t)F(u)du dG (t)
In particular,
lim C O L (T, N ) = lim C O F (T, N ) = C R (N ),

T →0 T →∞
lim C O L (T, N ) = lim C O F (T, N ) = C O (T ).
N →0 N →∞
When G(t) = 1 − e−θt and c N = c O ,
C O L (T, N ) = − θ(c F − c O )
T −2 ∞ −θt F(t)dt}
c F + (c F − c O )θ{ 0 F(t)dt + Nj=0 T [(θt) /j!]e
j
+ T N −1 ∞ ,
F(t)dt + j=0 T [(θt) j /j!]e−θt F(t)dt
0 ∞ −θt
+ ∞ j=N [(θT ) /j!] T e
j F(t)dt
(1.42)
where note that C O L (T, 0) = C O L (T, 1).

We find optimum TO∗ L and N O∗ L which minimize C O L (T, N ) in (1.42). Differen-
tiating C O L (T, N ) with respect to T and setting it equal to zero,

N −1 ∞
N −1 ∞
1 (T ) (θt) j −θt (θt) j −θt
Q e F(t)dt − e dF(t)
T j! T j!
j=0 j=0
T
1 (T ) cO
+Q F(t)dt − F(T ) = , (1.43)
0 cF − cO
1 (T ) is given in (1.24). Letting L 2 (T, N ) be the left-hand side of (1.43),

where Q
⎡ ⎤
1 (T ) T
dL 2 (T, N ) d Q
N −1 ∞
(θt) j
= ⎣ F(t)dt + e−θt F(t)dt ⎦
dT dT 0 T j!
j=0
∞
(θT ) j −θT 1 (T ) − h(T )] > 0,
+ e F(T )[ Q
j!
j=N

N −1 ∞
(θt) j −θt 1 (0) − h(t)]dt < 0,
L 2 (0, N ) ≡ lim L 2 (T, N ) = e F(t)[ Q
T →0 0 j!
j=0
L 2 (∞, N ) ≡ lim L 2 (T, N ) = μh(∞) − 1,
T →∞
because Q 1 (T ) > h(T ) and Q 1 (T, N ) increases strictly with N from Q 1 (T ) for
0 ≤ T < ∞ from Appendix 2.2.
Thus, L 2 (T, N ) increases strictly with T to μh(∞) − 1. Therefore, if h(∞) >
c F /[(c F − c O )μ], then there exists a finite and unique TO∗ L (0 < TO∗ L < ∞) which
satisfies (1.43), and the resulting cost rate is
1 (TO∗ L ).
C O L (TO∗ L , N ) = (c F − c O ) Q (1.44)
Next, prove that L 2 (T, N ) decreases with N . From Appendix 2.2, because
1 (T, N ) increases with N from Q
Q 1 (T ),
∞
1 (T ) (θt) N −θt
L 2 (T, N + 1) − L 2 (T, N ) = Q e F(t)dt
T N!
∞ (θt) N −θt
− e dF(t) < 0.
T N!

So that, L 2 (T, N ) decreases strictly with N (N ≥ 1) from Q 1 (T ) T F(t)dt − F(T ),
0
i.e., TO∗ L increases with N from TO∗ given in (1.24). Therefore, from (1.44), TO∗ L =
TO∗ and N O∗ L = 0 or 1, i.e., replacement overtime is better than random replacement.
From the inequality C O L (T, N + 1) − C O L (T, N ) ≥ 0,

⎡

N −1 ∞
2 (T, N − 1) ⎣
T (θt) j −θt
Q F(t)dt + e F(t)dt
0 T j!
j=0
⎤
∞ ∞
(θT ) j ∞ −θt (θT ) j ∞ −θt
+ e F(t)dt ⎦ − e dF(t)
j! T j! T
j=N j=N
∞
(θt) j −θt cO
− e dF(t) − 1 ≥ , (1.45)
T j! cF − cO
where
∞ ∞
(θt) N [ t e−θu dF(u)]dt
2 (T, N ) ≡ T
Q ,
∞ N ∞ −θu F(u)du]dt
T (θt) [ t e
and
∞ ∞ −θu
T [ t e dF(u)]dt
1 (T ).
Q 2 (T ) ≡ Q 2 (T, 0) = ∞ ∞ >Q
−θu
T [ t e F(u)du]dt
From Appendix 2.5, Q 2 (T, N ) increases strictly with T to h(∞) and also increases
2 (T ) to h(∞). So that, the left-hand side of (1.45) increases
strictly with N from Q
strictly with N from
T ∞ ∞
2 (T )
Q F(t)dt + e −θ(t−T )
F(t)dt + [1 − e−θ(t−T ) ]dF(t) − 1
0 T T
T
1 (T )
≥Q F(t)dt − F(T ),
0
which is given in (1.24). Therefore, the inequality (1.45) holds for any N and optimum
TO∗ , i.e., optimum TO∗ L and N O∗ L which minimize C O L (T, N ) in (1.42) are TO∗ L = TO∗
and N O∗ L = 0 or 1.
From the above discussions, we get the following conclusions: When T is con-
stantly given, if T > TO∗ then we should adopt replacement overtime first, and
conversely, if T < TO∗ then we should adopt replacement overtime last.
In general, it would be much difficult to compare theoretically replacement over-
time first and last. It would be possible to compute numerically optimum TF∗ and N F∗
which minimize C O F (T, N ) in (1.31), and TL∗ and N L∗ which minimize C O L (T, N )
in (1.41) for c N < c O , and compare them.
Next, we compare the expected cost rates C O F (T, N ) in (1.32) and C O L (T, N )
in (1.42) for given N ≥ 1. From the inequality (1.33) − (1.43) ≥ 0,

N −1 T
N −1 T
1 (T ) (θt) j −θt (θt) j −θt
Q e F(t)dt − e dF(t)
0 j! 0 j!
j=0 j=0

N −1 ∞
N −1 ∞
1 (T ) (θt) j −θt (θt) j −θt
− Q e F(t)dt − e dF(t)
j! j!
j=0 T j=0 T
T
1 (T )
+Q F(t)dt − F(T ) ≥ 0,
0
i.e.,

N −1 ∞
(θt) j −θt 1 (T )]dt
e F(t)[h(t) − Q
T j!
j=0
∞
T (θt) j −θt 1 (T ) − h(t)]dt,
≥ e F(t)[ Q (1.46)
0 j!
j=N
In addition, note that

∞ T −θt dF(t)
j=N 0 [(θt) /j!]e
j
1 (T ) > h(T ) >
Q ∞ T −θt F(t)dt
j=N 0 [(θt) /j!]e
j
for 0 < T < ∞. Thus, from Appendix 2.3, both sides of (1.46) are positive, its
left-hand side increases with N from 0, and conversely, its right-hand side decreases
with N to 0 for T (0 < T < ∞). So that, there exists a finite and unique minimum
N0 (1 ≤ N0 < ∞) which satisfies
∞T −θt dF(t) +
N −1 ∞ −θt dF(t)
j=N 0 [(θt) /j!]e j=0 T [(θt) /j!]e
j j
∞ T 1 (T ).
≥Q
−θt F(t)dt + N −1 ∞ −θt F(t)dt
j=N 0 [(θt) /j!]e j=0 T [(θt) /j!]e
j j
(1.47)
Therefore, if N ≥ N0 then the inequality (1.47) holds, and hence, TO∗ L ≥ TO∗ F , i.e.,
overtime first is better than overtime last. Conversely, if N < N O then TO∗ F > TO∗ L ,
i.e., overtime last is better than overtime first.
1.5 Finite Interval
Suppose that the unit has to operate for a job with random working cycles Y j ( j =
1, 2, . . .) in Fig. 1.1 for a finite interval [0, S] (0 < S < ∞) [10]. As preventive
replacement, the unit is replaced before failure at the first completion of working
1.5 Finite Interval 17
cycles over time T (0 < T ≤ S). In case of T = S, replacement should be done only
at time S. The probability that the unit is replaced at the first completion of working
cycles over time T in [T, S] is
∞
T S−t
F(t + u)dG(u) dG ( j) (t), (1.48)
j=0 0 T −t
the probability that it is replaced at time S is

∞
T
F(S) G(S − t)dG ( j) (t), (1.49)
j=0 0
F(T ), (1.50)

∞
T S−t
[F(t + u) − F(T )]dG(u) dG ( j) (t)
j=0 0 T −t
∞
T
+ [F(S) − F(T )]G(S − t)dG ( j) (t), (1.51)
j=0 0
where note that (1.48) + (1.49) + (1.50) + (1.51) = 1. The mean time to replacement is
∞
T S−t
(t + u)F(t + u)dG(u) dG ( j) (t)
j=0 0 T −t
∞
T T
+ S F(S) G(S − t)dG ( j) (t) + tdF(t)
j=0 0 0
∞
T S−t t+u
j=0 0 T −t T
∞ T S
+ ydF(y) G(S − t)dG ( j) (t)
j=0 0 T
T ∞
T S
= F(t)dt + F(u)G(u − t)du dG ( j) (t). (1.52)
0 j=0 0 T

T S−t
c F − (c F − c O ) ∞ ( j)
0 [ T −t F(t + u)dG(u)]dG (t)
∞ T
j=0 (
−(c F − c S )F(S) j=0 0 G(S − t)dG (t) j)
C O (T ; S) = T ∞ T S , (1.53)
( j)
0 F(t)dt + j=0 0 [ T F(u)G(u − t)du]dG (t)
where c S = replacement cost for time S with c S < c R and the other costs are given
in (1.22). In particular, when S = ∞, C O (T ; ∞) agrees with (1.22) and
c F − (c F − c S )F(S)
C O (S; S) = S = C A (S), (1.54)
0 F(t)dt
which is given in (1.6) when T = S and c S = cT .

When G(t) = 1 − e−θt and c O = c S ,
S
c F − (c F − c O ){F(S) + T [1 − e−θ(t−T ) ]dF(t)}
C O (T ; S) = T S , (1.55)
−θ(t−T ) F(t)dt
0 F(t)dt + T e
which agrees with (1.23) when S = ∞.

We find optimum TS∗ which minimizes C O (T ; S) for given S > 0. Differentiating
C O (T ; S) with respect to T and setting it equal to zero,
T
1 (T ; S) cO
Q F(t)dt − F(T ) = , (1.56)
0 cF − cO
where
S
e−θt dF(t)
1 (T ; S) ≡ T
Q ≤ h(S),
S
T e−θt F(t)dt
which increases strictly with T to h(S) from Appendix 1.5. Therefore, if

S cO
h(S) F(t)dt − F(S) > ,
0 cF − cO
then there exists a finite and unique TS∗ (0 < TS∗ < S) which satisfies (1.56), and the
resulting cost rate is
1 (TS∗ ; S).
C O (TS∗ ; S) = (c F − c O ) Q (1.57)
In addition, because Q1 (T ; S) increases with S to Q

1 (T ) in (1.24) from Appendix
1.5, TS∗ ≤ TO∗ and TS∗ decreases with S to TO∗ .
Conversely, if
S cO
h(S) F(t)dt − F(S) ≤ ,
0 cF − cO
then TS∗ = S, i.e., there is no replacement overtime to be done. Noting that the
left-hand side of (1.56) corresponds to (1.7) as T → S, if S ≤ T A∗ given in (1.7),
then TS∗ = S and the expected cost rate is given in (1.54). In other words, if a finite
interval S is less than optimum T A∗ , then it is not necessary to consider any preventive
replacement for a finite interval.
Next, suppose that S is a random variable with a general distribution L(t) ≡
Pr{S ≤ t}. Then, by using a similar method of obtaining C O (T ; S) in (1.53), the
probability that the unit is replaced at the first completion of working cycles over
time T is
∞
T ∞
F(t + u)L(t + u)dG(u) dG ( j) (t), (1.58)
j=0 0 T −t

T ∞
T ∞
F(t)dL(t) + F(u)G(u − t)dL(u) dG ( j) (t), (1.59)
0 j=0 0 T

T ∞
T ∞
L(t)dF(t) + G(u − t)L(u)dF(u) dG ( j) (t), (1.60)
0 j=0 0 T
∞
T ∞ T
( j)
(t + u)F(t + u)L(t + u)dG(u) dG (t) + t F(t)dL(t)
j=0 0 T −t 0
∞
T ∞ T
( j)
+ u F(u)G(u − t)dL(u) dG (t) + t L(t)dF(t)
j=0 0 T 0
∞ T ∞
+ uG(u − t)L(u)dF(u) dG ( j) (t)
j=0 0 T
T ∞
T ∞
= F(t)L(t)dt + F(u)G(u − t)L(u)du dG ( j) (t). (1.61)
0 j=0 0 T
C O (T ; L) =
T ∞
c F − (c F − c O ) ∞ ( j)
j=0 0 [ T −t F(t + u)L(t + u)dG(u)]dG (t)
T
−(c F − c S ){ 0 F(t)dL(t)
T ∞
+ ∞ j=0 0 [ T F(u)G(u − t)dL(u)]dG (t)}
( j)
T ∞ T ∞ , (1.62)
( j)
0 F(t)L(t)dt + j=0 0 [ T F(u)G(u − t)L(u)du]dG (t)
which agrees with (1.53) when L(t) ≡ 0 for t < S and ≡ 1 for t ≥ S.
When G(t) = 1 − e−θt , L(t) = 1 − e−lt (0 < l < ∞) and c O = c S ,
T −lt F(t)dt
c F + (c F − c O )θ 0 e
C O (T ; l) = T ∞ − (c F − c O )(θ + l).
0 e−lt F(t)dt + eθT T e−(θ+l)t F(t)dt
(1.63)
Differentiating C O (T ; l) with respect to T and setting it equal to zero,

T T
1 (T ; l) cO
Q e−lt F(t)dt − e−lt dF(t) = , (1.64)
0 0 cF − cO
where
∞ −(θ+l)t
T e dF(t)
Q 1 (T ; l) ≡ ∞ ≥ h(T ),
−(θ+l)t
T e F(t)dt
which agrees with Q 1 (T ; l) increases

1 (T ) in (1.24) when l = 0. Thus, because Q
strictly with T to h(∞) from Appendix 1.4, the left-hand side of (1.64) increases
strictly with T to
∞
[h(∞) − h(t)]e−lt F(t)dt. (1.65)
0
Therefore, if (1.65) is greater than c O /(c F −c O ), then there exists a finite and unique
Tl∗ (0 < Tl∗ < ∞) which satisfies (1.64), and the resulting cost rate is
1 (Tl∗ ; l).
C O (Tl∗ ; l) = (c F − c O ) Q (1.66)
When F(t) = 1 − e−(t/10) and G(t) = 1 − e−t , Table 1.2 presents optimum
2
TS , Tl∗ , and their expected cost rates when c O /c F = 0.2. When c O /c F = 0.2,
∗
T A∗ = 5.107. Thus, if S ≤ 5.107, TS∗ = S, and when S = 1/l = ∞, TS∗ = Tl∗ = TO∗
in Table 1.1. This indicates that Tl∗ ≥ TS∗ and C O (Tl∗ ; l) ≥ C O (TS∗ ; S), however,
their differences become smaller as S and 1/l become larger. This means that it an
operating interval is random, we should replace the unit at a little larger time than
Table 1.2 Optimum TS∗ , Tl∗ , and their expected cost rates when F(t) = 1−e−(t/10) , G(t) = 1−e−t
2
and c O /c F = 0.2
S = 1/l TS∗ C O (TS∗ ; S)/c F Tl∗ C O (Tl∗ ; l)/c F
5.0 5.000 0.080 5.421 0.099
10.0 4.289 0.083 4.809 0.090
15.0 4.283 0.083 4.624 0.087
20.0 4.283 0.083 4.536 0.086
30.0 4.283 0.083 4.449 0.085
40.0 4.283 0.083 4.407 0.084
50.0 4.283 0.083 4.382 0.084
∞ 4.238 0.083 4.283 0.083
that for a constant interval time. However, if this interval is large, we may replace
the unit as if it is infinity.
As one of extended policies, suppose that the unit is replaced before failure at the
first completion of working cycles over time T , at cycle N or at time S, whichever
occurs first. Then, the probability that the unit is replaced at the first completion of
working cycles over time T is

N −1 T S−t
F(t + u)dG(u) dG ( j) (t), (1.67)
j=0 0 T −t

T
F(t)dG (N ) (t), (1.68)
0

N −1 T
F(S) G(S − t)dG ( j) (t), (1.69)
j=0 0

T
[1 − G (N ) (t)]dF(t), (1.70)
0
and the probability that it is replaced at failure after time T is

N −1 T S−t
[F(t + u) − F(T )]dG(u) dG ( j) (t)
j=0 0 T −t
T
( j)
+ [F(S) − F(T )]G(S − t)dG (t) , (1.71)
0
where note that (1.67) + (1.68) + (1.69) + (1.70) + (1.71) = 1. The mean time to
replacement is

N −1 T S−t T
( j)
(t + u)F(t + u)dG(u) dG (t) + t F(t)dG (N ) (t)
j=0 0 T −t 0

N −1 T T
+ S F(S) G(S − t)dG ( j) (t) + t[1 − G (N ) (t)]dF(t)
j=0 0 0

N −1 T S−t t+u
j=0 0 T −t T
T S
+ ydF(y) G(S − t)dG ( j) (t)
0 T
T
N −1 T S
= F(t)[1 − G (N ) (t)]dt + F(u)G(u − t)du dG ( j) (t). (1.72)
0 j=0 0 T
C O F (T, N ; S) =
−1 T S−t ( j)
c F − (c F − c O ) Nj=0 0 [ T −t F(t + u)dG(u)]dG (t)
T
−(c F − c N ) 0 F(t)dG (N ) (t)
−1 T ( j)
−(c F − c S )F(S) Nj=0 0 G(S − t)dG (t)
T −1 T S ,
0 F(t)[1 − G (N ) (t)]dt + Nj=0 ( j)
0 [ T F(u)G(u − t)du]dG (t)
(1.73)
where c N = replacement cost at cycle N .

Similarly, when S is a random variable with L(t) = Pr{S ≤ t}, the expected cost
rate in (1.73) is rewritten as
C O F (T, N ; L) =
−1 T ∞ ( j)
c F − (c F − c O ) Nj=0 0 [ T −t F(t + u)L(t + u)dG(u)]dG (t)
T (N )
−(c F − c N ) 0 F(t)L(t)dG (t)
T
−(c F − c S ){ 0 F(t)[1 − G (N ) (t)]dL(t)
N −1 T ∞
+ j=0 0 [ T F(u)G(u − t)dL(u)]dG ( j) (t)}
T . (1.74)
F(t)L(t)[1 − G (N ) (t)]dt
−1 T ∞
0
+ Nj=0 ( j)
0 [ T F(u)G(u − t)L(u)du]dG (t)
Suppose that the unit is replaced over time T or at cycle N , whichever occurs
last. That is, letting t O and t N be the occurrence times of overtime T and cycle N ,
respectively, the unit is replaced before failure at max{t O , t N } or at time S, whichever

occurs first. Then, the probability that the unit is replaced over time T is
∞
T S−t
F(t + u)dG(u) dG ( j) (t), (1.75)
j=N 0 T −t

S
F(t)dG (N ) (t), (1.76)
T

⎡ ⎤
∞
T
F(S) ⎣1 − G (N ) (S) + G(S − t)dG ( j) (t)⎦ , (1.77)
j=N 0
F(T ), (1.78)
and the probability that it is replaced at failure after time T is

∞
T S−t
[F(t + u) − F(T )]dG(u) dG ( j) (t)
j=N 0 T −t
T S
+ [F(S) − F(T )]G(S − t)dG ( j) (t) + [1 − G (N ) (t)]dF(t), (1.79)
0 T
where note that (1.75) + (1.76) + (1.77) + (1.78) + (1.79) = 1. The mean time to
replacement is
∞
T S−t S
( j)
(t + u)F(t + u)dG(u) dG (t) + t F(t)dG (N ) (t)
j=N 0 T −t T
⎡ ⎤
∞
T T
+ S F(S) ⎣1 − G (N ) (S) + G(S − t)dG ( j)
(t)⎦ + tdF(t)
j=N 0 0
∞
T S−t t+u
j=N 0 T −t T
∞
T S S
( j)
+ ydF(y) G(S − t)dG (t) + t[1 − G (N ) (t)]dF(t)
j=N 0 T T
T S
= F(t)dt + F(t)[1 − G (N ) (t)]dt
0 T
∞
T S
+ F(u)G(u − t)du dG ( j) (t). (1.80)
j=N 0 T
C O L (T, N ; S) =
T S−t
c F − (c F − c O ) ∞ 0 [ T −t F(t + u)dG(u)]dG (t)
( j)
S j=N
−(c F − c N ) T F(t)dG (N ) (t)
T
−(c F − c S )F(S)[1 − G (N ) (S) + ∞ j=N 0 G(S − t)dG (t)]
( j)
T S . (1.81)
F(t)dt + F(t)[1 − G (N ) (t)]dt
0 T TS
+ ∞ j=N 0 [ T F(u)G(u − t)du]dG (t)
( j)
Clearly, C O L (0, N ; S) = C O F (∞, N ; S) and C O L (T, 0; S) = C O F (T, ∞; S).

Similarly, when S is a random variable with L(t) ≡ Pr{S ≤ t}, the expected cost
rate in (1.81) is rewritten as
C O L (T, N ; L) =
T ∞
c F − (c F − c O ) ∞ [ F(t + u)L(t + u)dG(u)]dG ( j) (t)
∞ j=N 0 T(N−t)
−(c F − c N ) T F(t)L(t)dG (t)
T
−(c F − c S ){ 0 F(t)[1 − G (N ) (t)]dL(t)
T ∞
+ ∞ j=N 0 [ T F(u)G(u − t)dL(u)]dG (t)}
( j)
T ∞ . (1.82)
F(t)dt + T F(t)[1 − G (N ) (t)]L(t)dt
0 T ∞
+ ∞ j=N 0 [ T F(u)G(u − t)L(u)du]dG (t)
( j)
We can make similar discussions of deriving optimum policies which mini-

mize C O F (T, N ; S) in (1.73), C O F (T, N ; L) in (1.74) and C O L (T, N ; S) in (1.81),
C O L (T, N ; L) in (1.82).
1.6 Working Number
Suppose the the unit operates for a job with a specified number N (1 ≤ N < ∞) of
working cycles. As preventive replacement, the unit is replaced before failure at the
first completion of working cycles over time T before the N th cycle [10].
1.6 Working Number 25
Then, the expected cost rate is, from (1.31),

−1 T ∞ ( j)
c F − (c F − c O ) Nj=0 0 [ T −t F(t + u)dG(u)]dG (t)
T (N )
−(c F − c N ) 0 F(t)dG (t)
C O (T ; N ) = T . (1.83)
F(t)[1 − G (N ) (t)]dt
−1 T ∞
0
+ Nj=0 ( j)
0 [ T F(u)G(u − t)du]dG (t)
When G(t) = 1 − e−θt and c O = c N , the expected cost rate is
C O (T ; N ) =
−1 ∞ T
c O + (c F − c O ) Nj=0 {[(θT ) j /j!] T e−θt dF(t) + 0 [(θt) j /j!]e−θt dF(t)}
N −1 ∞ .
−θt F(t)dt + T [(θt) j /j!]e−θt F(t)dt}
j=0 {[(θT ) /j!] T e
j
0
(1.84)
Differentiating C O (T ; N ) in (1.84) with respect to T and setting it equal to zero,

N −1 T
N −1 T
1 (T ) (θt) j −θt (θt) j −θt cO
Q e F(t)dt − e dF(t) = . (1.85)
0 j! 0 j! cF − cO
j=0 j=0
Therefore, if

N −1 ∞
(θt) j −θt cO
e [h(∞) − h(t)]F(t)dt > ,
0 j! cF − cO
j=0
then there exists a finite and unique TN∗ (0 < TN∗ < ∞) which satisfies (1.85) and
decreases with N to TO∗ in (1.24), and the resulting cost rate is
1 (TN∗ ).
C O (TN∗ ; N ) = (c F − c O ) Q (1.86)
Next, when the cycle number N is a random variable with a probability function
pn ≡ Pr{N = n} (n = 1, 2, . . .), the expected cost rate in (1.83) is rewritten as
C O (T ; p) =
n−1 T ∞
c F − (c F − c O ) ∞n=1 pn −t F(t + u)dG(u) dG ( j) (t)
T j=0 0 T
−(c F − c N ) ∞ (n)
n=1 pn 0 F(t)dG (t)
∞ T . (1.87)
pn 0 F(t)[1 − G (n) (t)]dt

n=1 n−1 T ∞
+ ∞ n=1 pn j=0 0 [ T F(u)G(u − t)du]dG (t)
( j)
When G(t) = 1 − e−θt , pn = [αn−1 /(n − 1)!]e−α (n = 1, 2, . . .) with mean

α + 1 and c O = c N , the expected cost rate is
∞
c O + (c F − c O ) ∞ {[(θT ) j /j!] T e−θt dF(t)
T j=0

+ 0 [(θt) j /j!]e−θt dF(t)} ∞ n= j (α /n!)e
n −α
C O (T ; α) = ∞ ∞ . (1.88)
{[(θT ) j /j!] T F(t)e−θt dt
T
j=0

+ 0 F(t)[(θt) j /j!]e−θt dt} ∞ n= j (α /n!)e
n −α
Clearly,
∞
c O + (c F − c O ) ∞j=0 0 [(θt) /j!]e
j −θt dF(t)[ ∞ (αn /n!)e−α ]
n= j
C O (∞; α) = ∞ ∞ ∞ .
−θt dt} n= j (α /n!)e−α
j=0 { 0 F(t)[(θt) /j!]e
j n
(1.89)
Differentiating C O (T ; α) with respect to T and setting it equal to zero,

∞
∞
∞
1 (T ) (θT ) j T (θt) j −θt αn −α
Q F(t)e−θt dt + F(t) e dt e
j! T 0 j! n!
j=0 n= j

∞
(θT ) j
∞ T (θt) j −θt

∞
αn −α cO
− e−θt dF(t) + e dF(t) e = ,
j! T 0 j! n! cF − cO
j=0 n= j
(1.90)
which increases strictly with T from 0. Therefore, if

∞
∞
∞
(θt) j −θt αn −α cO
[h(∞) − h(t)]F(t) e dt e > ,
0 j! n! cF − cO
j=0 n= j
then there exists a finite and unique T p∗ (0 < T p∗ < ∞) which satisfies (1.90) and
decreases with α, and the resulting cost rate is
1 (T p∗ ).
C O (T p∗ ; α) = (c F − c O ) Q (1.91)
2
TN , T p∗ , and their expected cost rates for N = α + 1. This indicates that T p∗ ≥ TN∗
∗
and C O (T p∗ ; α) ≥ C O (TN∗ ; N ), and both TN∗ and T p∗ approach to TO∗ in Table 1.1.
Comparing to Table 1.2, we have TN∗ ≤ TS∗ and T p∗ ≤ Tl∗ for N = N /θ = S.
1.7 Parallel System 27
Table 1.3 Optimum TN∗ , T p∗ , and their expected cost rates when F(t) = 1 − e−(t/10) , G(t) =
2
1 − e−t and c O /c F = 0.2

N =α+1 TN∗ C O (TN∗ ; N )/c F T p∗ C O (T p∗ ; α)/c F
5 4.494 0.086 4.705 0.089
10 4.284 0.083 4.310 0.083
15 4.283 0.083 4.284 0.083
20 4.283 0.083 4.283 0.083
30 4.283 0.083 4.283 0.083
40 4.283 0.083 4.283 0.083
50 4.283 0.083 4.283 0.083
∞ 4.283 0.083 4.283 0.083
1.7 Parallel System
We have considered until now only a one-unit system with a failure distribution F(t).
We could extend a one-unit system to redundant systems if their failure distributions
could be estimated statistically. For example, we consider a parallel system with n
units (n = 2, 3, . . .), whose failure distribution is F(t)n [3, p. 141, 11] . Then, by
replacing F(t) with F(t)n formally, the expected cost rates are rewritten as follows:
The expected cost rate of replacement overtime is, from (1.22),
C O (T ; n) =
∞ T ∞ ( j)
c F − (c F − c O )
j=0 0 { T −t [1 − F(t + u) ]dG(u)}dG (t)
n
T ∞ T ∞
. (1.92)
( j)
0 [1 − F(t) ]dt + j=0 0 { T [1 − F(u) ]G(u − t)du}dG (t)
n n
When F(t) = 1 − e−λt and G(t) = 1 − e−θt ,

T
c F − (c F − c O )θ 0 [1 − F(t)n ]dt
C O (T ; n) = T ∞ − (c F − c O )θ.
−θ(t−T ) [1 − F(t)n ]dt
0 [1 − F(t) ]dt + T e
n
(1.93)
∗ which minimizes C (T ; n). Clearly,

We find optimum TO;n O
cF
C O (0; n) = ∞ −θt [1 − F(t)n ]dt
− (c F − c O )θ,
0 e
cF
C O (∞; n) = ∞ .
0 [1 − F(t) ]dt
n
Differentiating C O (T ; n) with respect to T and setting it equal to zero,

T
n (T ) cO
Q [1 − F(t)n ]dt − F(T )n = , (1.94)
0 cF − cO
where
∞ −θt dF(t)n
n (T ) ≡ ∞ T e
Q (n = 2, 3, . . .),
T e−θt [1 − F(t)n ]dt
which increases strictly with T to λ when F(t) = 1 − e−λt from Appendix 2.6.
Clearly, when n = 1, Q 1 (T ) = λ for any T ≥ 0. Thus, the left-hand side of (1.94)
increase strictly with T from 0 to
∞
n
1
λ [1 − F(t)n ]dt − 1 = .
0 j
j=2

Therefore, if nj=2 (1/j) > c O /(c F − c O ), then there exists a finite and unique
∗ (0 < T ∗ < ∞) which satisfies (1.94), and the resulting cost rate is
TO;n O;n
∗
C O (TO;n n (T ∗ ).
; n) = (c F − c O ) Q (1.95)
O;n
Next, suppose that a parallel system with n (n ≥ 2) units has to operate for a job
with the completion of N working cycles. Then, the expected cost rate in (1.84) is
rewritten as
C O (T ; N , n) =
N −1 ∞ −θt T
c O + (c F − c O ) j=0 {[(θT ) /j!] T e
j dF(t)n + 0 [(θt) j /j!]e−θt dF(t)n }
N −1 ∞ .
{[(θT ) j /j!] T e−θt [1 − F(t)n ]dt
T
j=0
+ 0 [(θt) j /j!]e−θt [1 − F(t)n ]dt}
(1.96)
When F(t) = 1 − e−λt , we find optimum TN∗ ,n for specified N and n. Differen-
tiating C O (T ; N , n) with respect to T and setting it equal to zero,

N −1 T
N −1 T
n (T ) (θt) j −θt (θt) j −θt cO
Q e [1 − F(t) ]dt −
n
e dF(t)n = ,
j=0 0 j! 0 j!
j=0
c F − cO
(1.97)
1.7 Parallel System 29
whose left-hand side increases strictly with T from 0. Therefore, if

N −1 ∞ ∞
(θt) j −θt (θt) j −θt cO
λ e [1 − F(t)n ]dt − e dF(t) >
n
,
0 j! 0 j! cF − cO
j=0
then, there exists a finite and unique TN∗ ,n (0 < TN∗ ,n < ∞) which satisfies (1.97)
and TN∗ ,n decreases with N to TO;n
∗ given in (1.94), and the resulting cost rate is
n (TN∗ ,n ).
C O (TN∗ ,n ; N , n) = (c F − c O ) Q (1.98)
Furthermore, when N is a random variable with a probability function pi ≡

[αi−1 /(i − 1)!]e−α (i = 1, 2, . . .), the expected cost rate in (1.96) is
∞
c O + (c F − c O ) ∞ {[(θT ) j /j!] T e−θt dF(t)n
T j=0
∞
+ 0 [(θt) j /j!]e−θt dF(t)n } i= j (α /i!)e
i −α
C O (T ; p, n) = ∞ ∞ . (1.99)
{[(θT ) j /j!] T e−θt [1 − F(t)n ]dt
j=0 ∞
+ 0 [(θt) j /j!]e−θt [1 − F(t)n ]dt} i= −α
T
j (α /i!)e
i
Differentiating C O (T ; p, n) with respect to T and setting it equal to zero,

∞

∞
n (T )
T (θt) j −θt αi −α
Q e [1 − F(t)n ]dt e
0 j! i!
j=0 i= j
∞
T (θt) j ∞
αi cO
− e−θt dF(t)n e−α = , (1.100)
j! i! cF − cO
j=0 0 i= j
whose left-hand side increases strictly with T from 0. Therefore, if

∞
∞ ∞
∞
(θt) j −θt (θt) j −θt αi −α
λ e [1 − F(t)n ]dt − e dF(t)n e
0 j! 0 j! i!
j=0 i= j
cO
> ,
cF − cO
then there exists a finite and unique Tn,∗ p (0 < Tn,∗ p < ∞) which satisfies (1.100),
n (Tn,∗ p ).
C O (Tn,∗ p ; p, n) = (c F − c O ) Q (1.101)
On the other hand, when the number n of units for a parallel system has a proba-
bility function pk ≡ Pr{n = k} = (β k /k!)e−β (k = 0, 1, 2, . . . , 0 < β < ∞), the
failure distribution of a parallel system [3, p. 147, 12] is
∞
βk
Fβ (t) = e−β F(t)k = e−β F(t) .
k!
k=0
Thus, replacing F(t)n with e−β F(t) formally, the expected cost rate in (1.92) is
C O (T ; β) =
∞ T ∞ −β F(t+u) ]dG(u)}dG ( j) (t)
c F − (c F − c O ) j=0 0 { T −t [1 − e
T ∞ T ∞
.
−β F(t) ]dt + −β F(u) ]G(u − t)du}dG ( j) (t)
0 [1 − e j=0 0 { T [1 − e
(1.102)
When G(t) = 1 − e−θt , and F(t) = 1 − e−λt , i.e., Fβ (t) = exp(−βe−λt ),
C O (T ; β) =
T −λt )]dt
c F − (c F − c O )θ 0 [1 − exp(−βe
T ∞ − (c F − c O )θ.
0 [1 − exp(−βe−λt )]dt + T e−θ(t−T ) [1 − exp(−βe−λt )]dt
(1.103)
We find optimum Tβ∗ which minimizes C O (T ; β). Differentiating C O (T ; β) with

T
β (T ) cO
Q [1 − exp(−βe−λt )]dt − exp(−βe−λT ) = , (1.104)
0 cF − cO
where
∞ −θt ∞
T e dFβ (t) βλe−(θ+λ)t exp(−βe−λt )dt

Q β (T ) ≡ ∞ = T ∞ −θt ,
T e
−θt F (t)dt
β T e exp(−βe−λt )dt
which increases strictly with T from Q β (0) to λ (Appendix 2.7). Therefore, the
left-hand side of (1.104) increases strictly with T , and if
∞
1 1 − e−βx βj cF
dx = (−1) j−1 > ,
0 x j j! cF − cO
j=1
then there exists a finite and unique Tβ∗ (0 < Tβ∗ < ∞) which satisfies (1.104), and
the resulting cost rate is
β (T ∗ ).
C O (Tβ∗ ) = (c F − c O ) Q (1.105)
β
References 31
References
4. Chen M, Mizutani S, Nakagawa T (2010) Random and age replacement policies. Inter J Reliab
Qual Saf Eng 17:27–39
5. Nakagawa T, Zhao X, Yun WY (2011) Optimal age replacement and inspection policies with
random failure and replacement times. Inter J Reliab Qual Saf Eng 18:1–12
6. Zhao X, Nakagawa T (2012) Optimization problems of replacement first or last in reliability
theory. Euro J Oper Re 223:141–149
7. Zhao X, Mizutani S, Nakagawa T (2015) Which is better for replacement policies with con-
tinuous or discrete schedule time ? Euro J Oper Re 242:477–486
8. Zhao X, Qian C, Nakamura S (2014) Optimal age and periodic replacement with overtime
policies. J Reliab Qual Saf Eng 21:1450016 (14 pp)
9. Zhao X, Nakagawa T, Zuo M (2014) Optimal replacement last with continuous and discrete
policies. IEEE Trans Reliab 63:868–880
10. Mizutani S, Zhao X, Nakagawa T (2015) Overtime replacement policies with finite operating
interval and number. IEICE Trans Fundamentals (To appear)
11. Nakagawa T, Zhao X (2015) A survey of replacement policies for parallel systems with newly
proposed approaches. Inter J Performab Eng 11:321–328
12. Nakagawa T, Zhao X (2012) Optimization problems of a parallel system with a random number
of units. IEEE Trans Reliab 61:543–548
Chapter 2
Periodic Replacement Overtime
In this chapter, we suppose that the system is large and complex which consists of
many kinds of units, and it operates for a job with random working cycles introduced
in Chap. 1. The system undergoes minimal repairs at failures [1, p.96], [2, p.95],
[3] and can be quickly resumed after minimal repairs. As preventive replacement
policies, the system is planned to be replaced at time T , at working cycle N or at
number K of failures.
In Sect. 2.1, we suppose that the unit is replaced at time T or at working cycle N ,
whichever occurs first. Respective policies are called periodic replacement and ran-
dom replacement [4, p.53, 6–8] . The expected cost rates are obtained and optimum
TP∗ and N R∗ which minimize them are derived analytically. Furthermore, we compare
theoretically periodic replacement with time T and random replacement with cycle
N [9, 10]. It is shown that when both replacement costs for time T and cycle N are
the same, periodic replacement is better than random replacement.
In Sect. 2.2, we propose periodic replacement overtime [4, p.66, 11] in which
the unit is replaced at the first completion of working cycles over time T discussed in
Sect. 1.2. Optimum replacement time TO∗ which minimizes the expected cost rate is
derived analytically. In Sect. 2.3, to compare random replacement with replacement
overtime, we take up replacement overtime first in which the unit is replaced at cycle
N or over time T , whichever occurs first. When both replacement costs for cycle N
and overtime T are the same, it is also shown that replacement overtime is better
than random replacement.
In Sect. 2.4, we propose replacement overtime last in which the unit is replaced at
cycle N or over time T , whichever occurs last, and compare replacement overtime
first and last. It is of interest that if replacement number N is less than some number
N O , then overtime last is better than overtime first, and vice versa.
As one of modified replacement policies in Sect. 2.5, we consider another overtime
replacement in which the unit is replaced at failure K or at the first failure over time
T in order to operate continuously. Two overtime replacement first and last policies
are considered, and optimum policies which minimize the expected cost rates are
discussed [8]. Finally, we take up preventive maintenance overtime in which the unit
undergoes imperfect preventive maintenance [4, p.171, 6] when it finishes each
34 2 Periodic Replacement Overtime
work and is replaced at the first completion of working cycles over time T . The
expected cost rates for two kinds of imperfect preventive maintenances are obtained.
Throughout this chapter, it is assumed that working cycles Y j are independent
and
∞ have an identical distribution G(t) ≡ Pr{Y j ≤ t} with finite mean 1/θ ≡
G(t)dt, G ( j) (t) ( j = 1, 2, . . .) denotes the j−fold convolution of G(t) with
0
itself, G (0) (t) ≡ 1 for t ≥ 0, and M(t) ≡ ∞ ( j)
j=1 G (t). In addition, the unit has a
failure distribution F(t) with finite mean μ, a density function f (t) ≡ dF(t)/dt,
t the
failure rate h(t) ≡ f (t)/F(t), and the cumulative hazard rate H (t) ≡ 0 h(u)du,
which represents the expected number of failures in [0, t]. It is assumed that the
failure rate h(t) increases from h(0) = 0 to h(∞) ≡ limt→∞ h(t).
2.1 Periodic and Random Replacements
A new unit begins to operate at time 0 and undergoes minimal repairs at failures,
where the time for minimal repair is negligible. Suppose that the unit has to operate
for a job with random working cycles Y j ( j = 1, 2, . . .) defined in Sect. 1.1. As
preventive replacement, the unit is planned to be replaced at time T (0 < T ≤ ∞)
or at working cycle N (N = 1, 2, . . .), whichever occurs first. This is called periodic
replacement first. Then, the probability that the unit is replaced at cycle N is G (N ) (T ),
and the probability that it is replaced at time T is 1 − G (N ) (T ). Thus, the mean time
to replacement is
T T
T [1 − G (N ) (T )] + tdG (N ) (t) = [1 − G (N ) (t)]dt,
0 0
and the total expected number of failures until replacement is

T T
H (T )[1 − G (N ) (T )] + H (t)dG (N ) (t) = [1 − G (N ) (t)]h(t)dt.
0 0

T
cT + (c N − cT )G (N ) (T ) + c M 0 [1 − G (N ) (t)]h(t)dt
C F (T, N ) = T , (2.1)
[1 − G (N ) (t)]dt
0
where cT = replacement cost at time T , c N = replacement cost at cycle N , and

c M = minimal repair cost at each failure.
In particular, when the unit is replaced only at time T , which is called standard
periodic replacement,
cT + c M H (T )
C P (T ) ≡ C F (T, ∞) = lim C F (T, N ) = . (2.2)
N →∞ T
2.1 Periodic and Random Replacements 35
An optimum policy which minimizes C P (T ) is [1, p.102], [2, p.101]:

∞
(i) If 0 tdh(t) > cT /c M , then there exists a finite and unique TP∗ (0 < TP∗ < ∞)
which satisfies
T T
cT cT
[h(T ) − h(t)]dt = or tdh(t) = , (2.3)
0 cM 0 cM
C P (TP∗ ) = c M h(TP∗ ). (2.4)

∞
(ii) If 0 tdh(t) ≤ cT /c M , then TP∗ = ∞, i.e., the unit always undergoes minimal
repair at each failure, and the expected cost rate is
C P (∞) ≡ lim C P (T ) = c M h(∞). (2.5)

T →∞
When the unit is replaced only at cycle N , which is called periodic random
replacement [4, p.75],
∞
cN + cM [1 − G (N ) (t)]h(t)dt
C R (N ) ≡ lim C F (T, N ) = 0
(N = 1, 2, . . .).
T →∞ N /θ
(2.6)
We find optimum N R∗ which minimizes C R (N ). From the inequality C R (N + 1) −

C R (N ) ≥ 0,
∞
N cN
H1 (N ) − [1 − G (N ) (t)]h(t)dt ≥ ,
θ 0 cM
or
∞ cN
[1 − G (N ) (t)][H1 (N ) − h(t)]dt ≥ , (2.7)
0 cM
where
T
[G (N ) (t) − G (N +1) (t)]h(t)dt
H1 (T, N ) ≡ T
0
≤ h(T ),
(N ) (t) − G (N +1) (t)]dt
0 [G
∞
H1 (N ) ≡ lim H1 (T ; N ) = θ [G (N ) (t) − G (N +1) (t)]h(t)dt.
T →∞ 0
Thus, if H1 (N ) increases strictly to H1 (∞), then the left-hand side of (2.7) increases
strictly with N . Therefore, an optimum policy which minimizes C R (N ) is:
∞
(i) If H1 (N ) increases strictly to H1 (∞) and 0 [H1 (∞) − h(t)]dt > c N /c M , then
there exists a finite and unique minimum N R∗ (1 ≤ N R∗ < ∞) which satisfies
(2.7), and the resulting cost rate is
c M H1 (N R∗ − 1) < C R (N R∗ ) ≤ c M H1 (N R∗ ). (2.8)
∞
(ii) If 0 [H1 (∞) − h(t)]dt ≤ c N /c M , then N R∗ = ∞, and the expected cost rate
is given in (2.5).

When G(t) = 1 − e−θt , i.e., G (N ) (t) = ∞j=N [(θt) /j!]e
j −θt (N = 0, 1, 2, . . .),
from Appendix 3.1,
∞ N ∞
θ(θt) N −θt (θt) j −θt
H1 (N ) = e h(t)dt = e dh(t)
0 N! j!
j=0 0
increases strictly with N to h(∞). In this case, if

∞ cN
tdh(t) > ,
0 cM
then a finite and unique minimum N R∗ (1 ≤ N R∗ < ∞) exists.

2
∗ ∗
TP and N R , and their expected cost rates for ci /c M (i = T, N ). This indicates that
optimum TP∗ and N R∗ increase with ci /c M (i = T, N ), and TP∗ ≥ N R∗ /θ, however,
they are almost the same, and when cT = c N , their cost rates are C P (TP∗ ) < C R (N R∗ ).
When c N < cT , e.g., when cT = 0.2 and c N = 0.1, C P (TP∗ ) > C R (N R∗ ).
It was shown numerically that when both replacement costs are the same, periodic
replacement is better than random replacement. Next, we discuss theoretically which
policy is better to replace the unit at time T or at cycle N . For this purpose, we find
optimum TF∗ and N F∗ which minimize the expected cost rate C F (T, N ) in (2.1).
Differentiating C F (T, N ) with respect to T and setting it equal to zero,
Table 2.1 Optimum TP∗ and N R∗ , and their expected cost rates when F(t) = 1 − e−(t/10) and
2
G(t) = 1 − e−t
ci /c M TP∗ C P (TP∗ )/c M N R∗ C R (N R∗ )/c M
0.1 3.162 0.063 3 0.073
0.2 4.472 0.089 4 0.100
0.5 7.071 0.141 7 0.151
1.0 10.000 0.200 10 0.210
2.0 14.142 0.283 14 0.293
5.0 22.361 0.447 22 0.457
2.1 Periodic and Random Replacements 37
T
cM [1 − G (N ) (t)][h(T ) − h(t)]dt
0
T
− (cT − c N ) [1 − G (N ) (t)][r N (T ) − r N (t)]dt = cT , (2.9)
0
where r N (t) is given in (1.13). From the inequality C F (T, N + 1) − C F (T, N ) ≥ 0,

T
cM [1 − G (N ) (t)][H1 (T, N ) − h(t)]dt
0
T
(N ) G (N ) (T ) − G (N +1) (T )
+ (cT − c N ) [1 − G (t)] T + r N (t) dt ≥ cT .
(N ) (u) − G (N +1) (u)]du
0 0 [G
(2.10)
Substituting (2.9) for (2.10),

G (N ) (T ) − G (N +1) (T )
c M [H1 (T, N ) − h(T )] + (cT − c N ) T + r N (T ) ≥ 0.
(N ) (t) − G (N +1) (t)]dt
0 [G
(2.11)
Thus, when cT ≤ c N , there does not exist any finite optimum N F∗ for T > 0 because
H1 (T, N ) ≤ h(T ), i.e., N F∗ = ∞. In this case, the unit should be replaced only at
time T .
In particular, when G(t) = 1 − e−θt and cT > c N ,
θ(θT ) N −1 /(N − 1)!

r N (T ) = N −1
j=0 [(θT ) /j!]
j
decreases strictly with N from θ to 0 and increases strictly with T from 0 to θ for
N ≥ 2, r1 (T ) = θ for T ≥ 0 from Appendix 1.1, and
G (N ) (T ) − G (N +1) (T ) θ(θT ) N /(N )!

T = ∞
j=N +1 [(θT ) /j!]
(N ) (t) − G (N +1) (t)]dt j
0 [G
increases strictly with N to ∞ and decreases strictly with T from ∞ to 0 from

Appendix 1.1. Thus, because lim N →∞ H1 (T, N ) = h(T ), there exists a finite
N F∗ (1 ≤ N F∗ < ∞) which satisfies (2.11) for T > 0. Furthermore, the left-hand
side of (2.9) goes to
∞
cM [1 − G (N ) (t)][h(∞) − h(t)]dt − (cT − c N )(N − 1)
0
∞
as T → ∞. Therefore, if 0 [1−G (N ) (t)][h(∞)−h(t)]dt > [cT N −c N (N −1)]/c M ,
then there exists a finite TF∗ (0 < TF∗ < ∞) which satisfies (2.9). It can be clearly
seen that if h(∞) = ∞, both finite TF∗ and N F∗ exist in case of cT > c N .
On the other hand, suppose that the unit is replaced at time T or at cycle N ,
whichever occurs last. This is called periodic replacement last. Because the proba-
bility that the unit is replaced at cycle N is 1 − G (N ) (T ), and the probability that it
is replaced at time T is G (N ) (T ), the mean time to replacement is
∞ ∞
T G (N ) (T ) + tdG (N ) (t) = T + [1 − G (N ) (t)]dt, (2.12)
T T
and the expected number of failures until replacement is

∞ ∞
H (T )G (N ) (T ) + H (t)dG (N ) (t) = H (T ) + [1 − G (N ) (t)]h(t)dt.
T T
(2.13)
Then, by the similar method of obtaining (2.1), the expected cost rate is [4, p.79]
∞
c N + (cT − c N )G (N ) (T ) + c M {H (T ) + T [1 − G (N ) (t)]h(t)dt}
C L (T, N ) = ∞ .
T + T [1 − G (N ) (t)]dt
(2.14)
Clearly,
C L (0, N ) ≡ lim C L (T, N ) = C F (∞, N ) = C R (N )

T →0
in (2.6), and
C L (T, 0) ≡ lim C L (T, N ) = C F (T, ∞) = C P (T )

N →0
in (2.2). We could make similar discussions of deriving optimum policies to minimize

the expected cost rate C L (T, N ) in (2.14).
2.2 Replacement Overtime
Suppose that the unit is replaced at the first completion of working cycles Y j ( j =
1, 2, . . .) over time T (0 ≤ T ≤ ∞), which has been introduced in Sect. 1.2. Then,
the mean time to replacement is
2.2 Replacement Overtime 39
∞
T ∞
(t + u)dG(u) dG ( j) (t)
j=0 0 T −t
∞ T ∞
=T+ G(t)dt + G(u − t)du dM(t), (2.15)
T 0 T

∞
T ∞
H (t + u)dG(u) dG ( j) (t)
j=0 0 T −t
∞ T ∞
= H (T ) + G(t)h(t)dt + G(u − t)h(u)du dM(t), (2.16)
T 0 T
which agrees with (2.15) when H (t) = t, i.e., h(t) = 1.

∞ T ∞
c O + c M {H (T ) + T G(t)h(t)dt + 0 [ T G(u − t)h(u)du]dM(t)}
C O (T ) = ∞ T ∞ ,
T + T G(t)dt + 0 [ T G(u − t)du]dM(t)
(2.17)
where c O = replacement cost over time T and c M is given in (2.1). In particular,
C O (0) ≡ lim C O (T ) = C R (1)

T →0
in (2.6) when c O = c N , and
C O (∞) ≡ lim C O (T ) = C P (∞)

T →∞
in (2.5). Differentiating C O (T ) with respect to T and setting it equal to zero,

∞
∞
θG(t) T h(T + t) − H (T ) + G(u)[h(T + t) − h(u)]du
0 T
T ∞
cO
+ G(u − x)[h(T + t) − h(u)]du dM(x) dt = , (2.18)
0 T cM
∞
whose left-hand side increases strictly from 0 to 0 tdh(t). Therefore,
∞
if 0 tdh(t) > c O /c M , then there exists a finite and unique TO∗ (0 < TO∗ < ∞)
which satisfies (2.18), and the resulting cost rate is
∞
C O (TO∗ ) = c M θG(t)h(t + TO∗ )dt. (2.19)
0

∞
c O + c M [H (T ) + 0 e−θt h(t + T )dt]
C O (T ) = . (2.20)
T + 1/θ
From (2.18), optimum TO∗ satisfies

∞ cO
T θe−θt h(t + T )dt − H (T ) = , (2.21)
0 cM
∞
whose left-hand side increases strictly with T from 0 to 0 tdh(t), and decreases
strictly with θ to T h(T ) − H (T ).
Therefore, we have the optimum policy:
∞
(i) If 0 tdh(t) > c O /c M , then there exists a finite and unique TO∗ (0 < TO∗ < ∞)
∞ c O + c M H (TO∗ )
C O (TO∗ ) = c M θe−θt h(t + TO∗ )dt = . (2.22)
0 TO∗
∞
(ii) If 0 tdh(t) ≤ c O /c M , then TO∗ = ∞, and the expected cost rate is given in
(2.5).
Note that TO∗ decreases with 1/θ from TP∗ given in (2.3).
When F(t) = 1 − e−(t/10) and G(t) = 1 − e−θt , Table 2.2 presents optimum
2
TO∗ and its expected cost rate. This indicates that optimum TO∗ increases with c O /c M
and decreases with 1/θ from TP∗ . Compared to Table 2.1 when 1/θ = 1, TO∗ < TP∗ <
TO∗ + 1/θ and C R (N R∗ ) > C O (TO∗ ) > C P (TP∗ ), however, their differences are very
small as c O /c M becomes large. So that, if c N < c O < cT then random replacement
might be better than replacement overtime, and replacement overtime might be better
than periodic replacement, respectively.
Table 2.2 Optimum TO∗ and its expected cost rate when F(t) = 1 − e−(t/10) and G(t) = 1 − e−θt
2
c O /c M 1/θ = 1 1/θ = 2 1/θ = 5

TO∗ C O (TO∗ )/c M TO∗ C O (TO∗ )/c M TO∗ C O (TO∗ )/c M
0.1 2.317 0.066 1.742 0.075 0.916 0.118
0.2 3.583 0.091 2.899 0.098 1.709 0.134
0.5 6.141 0.143 5.348 0.147 3.661 0.173
1.0 9.050 0.201 8.198 0.204 6.182 0.224
2.0 13.177 0.284 12.283 0.286 10.002 0.300
5.0 21.383 0.448 20.450 0.449 17.915 0.458
2.3 Comparisons of Periodic and Random Replacements 41
2.3 Comparisons of Periodic and Random Replacements
Compare theoretically replacement overtime to periodic and random replacements

with time T (0 < T ≤ ∞) and cycle N (N = 1, 2, . . .) when cT = c N = c R = c O
and h(∞) = ∞. In this case, finite TP∗ , N R∗ and TO∗ always exist. In addition, because
TP∗ is an optimum solution of minimizing C P (T ) in (2.2), C O (TO∗ ) is greater than
C P (TP∗ ) from (2.22), i.e., periodic replacement is better than replacement overtime.
If c O < cT then replacement overtime might be rather than periodic replacement.
In this case, we could compute numerically C P (TP∗ ) in (2.4) and C O (TO∗ ) in (2.22),
and compare them.
We have already compared numerically random replacement and replacement
overtime in Tables 2.1 and 2.2. Next, we compare theoretically random replacement
and replacement overtime. For this purpose, we propose the following extended
replacement with time T and cycle N , which is called replacement overtime first:
The unit is replaced at cycle N (N = 1, 2, . . .) or over time T , whichever occurs
first. Then, the probability that the unit is replaced at cycle N is G (N ) (T ), and the
probability that it is replaced over time T is 1 − G (N ) (T ), where it is counted as
replacement done over time T when the N th working cycle occurs over time T .
Then, the mean time to replacement is
T
N −1 T ∞
tdG (N ) (t) + (t + u)dG(u) dG ( j) (t)
0 j=0 0 T −t
T
N −1 T ∞
(N )
= [1 − G (t)]dt + G(u − t)du dG ( j) (t), (2.23)
0 j=0 0 T
T
N −1 T ∞
(N )
H (t)dG (t) + H (t + u)dG(u) dG ( j) (t)
0 j=0 0 T −t
T
N −1 T ∞
= [1 − G (N ) (t)]h(t)dt + G(u − t)h(u)du dG ( j) (t), (2.24)
0 j=0 0 T
which agrees with (2.23) when h(t) = 1.

T
c O + (c N − c O )G (N ) (T ) + c M { 0 [1 − G (N ) (t)]h(t)dt
−1 T ∞ ( j)
+ Nj=0 0 [ T G(u − t)h(u)du]dG (t)}
C O F (T, N ) = T
N −1 T ∞
.
(N ) (t)]dt + ( j)
0 [1 − G j=0 0 [ T G(u − t)du]dG (t)
(2.25)
It can be clearly seen that C O F (∞, N ) = C R (N ) in (2.6) and C O F (T, ∞) = C O (T )

in (2.17).
When G(t) = 1 − e−θt (0 < θ < ∞), i.e., G (N ) (t) = ∞ j=N [(θt) /j!]e
j −θt , and
c O = c N , (2.25) is rewritten as
T
c O + c M { 0 [1 −G (N ) (t)]h(t)dt
∞
+ [1 − G (N ) (T )] T e−θ(t−T ) h(t)dt}
C O F (T, N ) = T . (2.26)
0 [1 − G (N ) (t)]dt + (1/θ)[1 − G (N ) (T )]
We discuss optimum TO∗ F and N O∗ F which minimize C O F (T, N ). In particular, when

N = 1, C O F (T, 1) = C R (1), and hence, TO∗ F = ∞. For N ≥ 2, differentiating
C O F (T, N ) with respect to T and setting it equal to zero,
∞ T T cO
θe−θ(t−T ) h(t)dt [1 − G (N ) (t)]dt − [1 − G (N ) (t)]h(t)dt = ,
T 0 0 cM
i.e.,
T ∞
(N ) −θu cO
[1 − G (t)] θe [h(u + T ) − h(t)]du dt = , (2.27)
0 0 cM
whose left-hand side increases strictly with T . Therefore, if

∞ cO
[1 − G (N ) (t)][h(∞) − h(t)]dt > ,
0 cM
then there exists a finite and unique TO∗ F (0 < TO∗ F < ∞) which satisfies (2.27),
∞
C O F (TO∗ F , N ) = c M θe−θt h(t + TO∗ F )dt. (2.28)
0
In addition, because the left-hand side of (2.27) increases with N , TO∗ F decreases
with N to TO∗ given in (2.21). So that, from (2.28), optimum TO∗ F and N O∗ F which
minimize C O F (T, N ) for N ≥ 2 is TO∗ F = TO∗ and N O∗ F = ∞.
From the above discussions, and from (2.6) and (2.22), if
C O (TO∗ ) < C R (1),
i.e.,
∞ ∞
cM e−θt h(t + TO∗ )dt < c O + c M e−θt h(t)dt,
0 0
2.3 Comparisons of Periodic and Random Replacements 43
or
∞ cO
e−θt [h(t + TO∗ ) − h(t)]dt < ,
0 cM
then replacement overtime is better than random replacement.

For example, when H (t) = (λt)2 , from (2.21), TO∗ satisfies
2λ2 ∗ cO
(λTO∗ )2 + TO = .
θ cM
Then, from (2.22),

∞ ∞
cO + cM e−θt h(t)dt − c M e−θt h(t + TO∗ )dt
0 0

2λ2 2λ2 ∗ 1
= cO + cM 2 − cM TO +
θ θ θ
∗ 2

∗ 2 2λTO 2λ 2λ2 ∗ 1
= c M (λTO ) + + 2 − TO + > 0,
θ θ θ θ
which shows that replacement overtime is better than random replacement.
2.4 Replacement Overtime Last
We have already obtained the expected cost rate of replacement overtime first in
which the unit is replaced at cycle N before time T in Sect. 2.3. Next, we propose
replacement overtime last in which the unit is replaced at cycle N or over time
T , whichever occurs last. The probability that the unit is replaced at cycle N is
1 − G (N ) (T ), and the probability that it is replaced over time T is G (N ) (T ). Then,
the mean time to replacement is
∞ ∞
T ∞
(N )
tdG (t) + (t + u)dG(u) dG ( j) (t)
T j=N 0 T −t
∞ ∞
T ∞
(N )
=T+ [1 − G (t)]dt + G(u − t)du dG ( j) (t), (2.29)
T j=N 0 T

∞ ∞
T ∞
H (t)dG (N ) (t) + H (t + u)dG(u) dG ( j) (t)
T j=N 0 T −t
∞ ∞
T ∞
= H (T ) + [1 − G (N ) (t)]h(t)dt + G(u − t)h(u)du dG ( j) (t).
T j=N 0 T
(2.30)
c O + (c N − c O)[1 − G (N ) (T )]
∞
+ c M {H (T ) + T [1 − G (N ) (t)]h(t)dt
∞ T ∞
+ j=N 0 [ T G(u − t)h(u)du]dG ( j) (t)}
C O L (T, N ) = ∞ T ∞ .
T + T [1 − G (N ) (t)]dt + ∞ ( j)
j=N 0 [ T G(u − t)du]dG (t)
(2.31)
It can be easily seen that C O L (0, N ) = C O F (∞, N ) = C R (N ) in (2.6) and

C O L (T, 0) = C O F (T, ∞) = C O (T ) in (2.17). Note that when c N = c O ,
C O L (T, 0) = C O L (T, 1), in which the unit is
always replaced over time T .
When G(t) = 1 − e−θt , i.e., G (N ) (t) = ∞ j=N [(θt) /j!]e
j −θt , and c = c ,
O N
(2.31) is rewritten as
∞
c O + c M {H(T ) + T [1 − G (N ) (t)]h(t)dt
∞
+G (N ) (T ) T e−θ(t−T ) h(t)dt}
C O L (T, N ) = ∞ . (2.32)
T + T [1 − G (N ) (t)]dt + (1/θ)G (N ) (T )
We discuss optimum TO∗ L and N O∗ L which minimize C O L (T, N ). Differentiating

C O L (T, N ) with respect to T and setting it equal to zero,
∞ ∞
θe−θt h(t + T )dt T + [1 − G (N ) (t)]dt
0 T
∞
cO
− H (T ) − [1 − G (N ) (t)]h(t)dt = ,
T cM
or
T ∞
−θu
θe
[h(u + T ) − h(t)]du dt
0 0
∞ ∞
(N ) −θu cO
+ [1 − G (t)] θe [h(u + T ) − h(t)]du dt = , (2.33)
T 0 cM
∞
whose left-hand increases strictly with T to 0 tdh(t). Therefore, if
∞ cO
tdh(t) > ,
0 cM
then there exists a finite and unique TO∗ L (0 ≤ TO∗ L < ∞) which satisfies (2.33),
∞
C O L (TO∗ L ) = cM θe−θt h(t + TO∗ L )dt. (2.34)
0
Note that C O L (TO∗ L ) agrees with C O F (TO∗ F ) in (2.28) when TO∗ L = TO∗ F .
Furthermore, we prove that the left-hand side of (2.33) decreases with N as
follows:
∞ ∞
(θt) N −θt
e θe−θu [h(u + T ) − h(t)]du dt
T N! 0
∞ ∞
(θt) −θt
N
−θ(u−T )
= e θe [h(u) − h(t)]du dt
T N! T
∞
(θt) N −θt t
= e − θe−θ(u−T ) [h(t) − h(u)]du
T N! T
∞
−θ(u−T )
+ θe [h(u) − h(t)]du dt. (2.35)
t
Furthermore,
∞ ∞
(θt) N −θt −θ(u−T )
e θe [h(u) − h(t)]du dt
T N! t
∞
−θ(t−T )
t (θu) N −θu
= θe e [h(t) − h(u)]du dt.
T T N!
Thus, (2.35) becomes

∞ t
(θu) N (θt) N

−θ(t+u−T )
[h(t) − h(u)]θe − du dt ≤ 0,
T T N! N!
which follows that the left-hand side of (2.33) decreases with N . This shows that
TO∗ L increases with N from TO∗ given in (2.21). So that, from (2.34), optimum TO∗ L
and N O∗ L which minimize C O L (T ; N ) is TO∗ L = TO∗ and N O∗ L = 0 or 1.
Next, we compare the expected cost rates C O F (T, N ) in (2.26) and C O L (T, N )
in (2.32) for a fixed N ≥ 1. From the inequality (2.27)–(2.33) ≥ 0,
∞ ∞
[1 − G (N ) (t)] θe−θu [h(t) − h(u + T )]du dt
T 0
T ∞
(N ) −θu
≥ G (t) θe [h(u + T ) − h(t)]du dt. (2.36)
0 0
Noting that from (2.35), its left-hand side increases with N from 0, and conversely,
its right-hand side decreases with N to 0. So that, there exists a finite and unique
minimum N O (1 ≤ N O < ∞) which satisfies
∞ T ∞
[1 − G (N ) (t)]h(t)dt + 0 G (N ) (t)h(t)dt
T
∞ T ≥ θe−θt h(t + T )dt. (2.37)
(N ) (N )
T [1 − G (t)]dt + 0 G (t)dt 0
Therefore, if N ≥ N O then the inequality (2.37) holds, and hence, TO∗ L ≥ TO∗ F , i.e.,
replacement overtime first is better than replacement overtime last. Conversely, if
N < N O then TO∗ F > TO∗ L , i.e., replacement overtime last is better than replacement
overtime first.
2.5 Replacement Overtime with Number of Failures
The unit is replaced at periodic times in standard periodic replacement as shown

in Sect. 2.1. However, some units should be replaced when they have failed rather
than a planned time [2, p.104] in order to operate continuously without stopping.
This section proposes two policies with the number of failures in which the unit is
replaced at a planned number K of failures or over time T , whichever occurs first or
last.
It is assumed that the unit undergoes minimal repair between replacements. Then,
failures occur at a nonhomogeneous Poisson process with mean value function H (t),
i.e., the probability that j failures occur exactly in [0, t] is p j (t) ≡ [H (t) j /j!]e−H (t)
( j = 0, 1, 2, . . .) [5, p.27], and the probability that more than j failures occur in
∞ j−1
[0, t] is i= j pi (t) = P j (t) and P j (t) ≡ 1 − P j (t) = i=0 pi (t). Note that
p0 (t) = e−H (t) = F(t) = P 1 (t), P j (0) = 0, P j (0) = 1, P0 (t) = 1, and P 0 (t) = 0.
2.5.1 Replacement Overtime First with Number of Failures
Suppose that the unit is replaced at failure K (K = 1, 2, . . .) or at the first failure

over time T (0 ≤ T ≤ ∞), whichever occurs first, i.e., it is replaced either at failure
K before time T or over time T before failure K .
The probability that the unit is replaced at failure K is ∞ p j (T ) = PK (T ),
K −1 j=K
and the probability that it is replaced over time T is j=0 p j (T ) = P K (T ). Thus,
the expected number of failure until replacement is

K −1
K −1
K PK (T ) + ( j + 1) p j (T ) = K − (K − 1 − j) p j (T )
j=0 j=0
T
= P K (t)h(t)dt + P K (T ), (2.38)
0
2.5 Replacement Overtime with Number of Failures 47
where note that any failure at replacement is always counted. Because the probability
that some failure occurs in (u, u + du] for u > t, given that a failure have occurred
at time t is f (u)du/F(t) [2, p.96], the mean time to replacement is
T T ∞
1
td PK (t) + udF(u) d P K (t)
0 0 F(t) T
T ∞
= P K (t)dt + P K (T ) e−[H (t)−H (T )] dt, (2.39)
0 T
which agrees with (2.38) when H (t) = t, i.e., h(t) = 1.

T
c O + (c K − c O )PK (T ) + c M [ 0 P K (t)h(t)dt + P K (T )]
C O F (T, K ) = T ∞ ,
−[H (t)−H (T )] dt
0 P K (t)dt + P K (T ) T e
(2.40)
where c O = replacement cost over time T , c K = replacement cost at failure K , and

c M = minimal repair cost at each failure. In particular, when T = ∞, i.e., the unit
is replaced only at failure K , the expected cost rate is, from [2, p.106],
cK + cM K
C(K ) ≡ lim C O F (T, K ) = ∞ (K = 1, 2, . . .). (2.41)
0 P K (t)dt
T →∞
If h(∞) > c K /c M then there exists a finite and unique minimum K ∗ (1 ≤ K ∗ < ∞)
which satisfies
∞
1 cK
∞ P K (t)dt − K ≥ , (2.42)
0 p K (t)dt 0
cM

cM cM
∞ < C(K ∗ ) ≤ ∞ . (2.43)
0 p K ∗ −1 (t)dt 0 p K ∗ (t)dt
On the other hand, when K = ∞, i.e., the unit is replaced only at the first failure
over time T , the expected cost rate is
c O + c M [H (T ) + 1]
C O F (T ) ≡ lim C O F (T, K ) = , (2.44)
K →∞ T + 1/Q(T )
where
1 F(T )
Q(T ) ≡ ∞ = ∞ ≥ h(T ),
T e−[H (t)−H (T )] dt T F(t)dt
which increases strictly with T from 1/μ to h(∞) [2, p.9] from Appendix 1.2. Clearly,
when T = 0, i.e., when the unit is always replaced at the first failure,
cO + cM
C O F (0) ≡ lim C O F (T ) = = C(1) (2.45)
T →0 μ
for c O = c K , and when T = ∞, i.e., there is no replacement to be made,
C O F (∞) ≡ lim C O F (T ) = c M h(∞) = C(∞). (2.46)

T →∞
We find optimum TO∗ which minimizes C O F (T ) in (2.44). Differentiating C O F (T )

with respect to T and setting it equal to zero,
cO
T Q(T ) − H (T ) = , (2.47)
cM
whose left-hand side increases strictly with T from 0 to

∞ ∞
[h(∞) − h(t)]dt = tdh(t).
0 0
Therefore, we have the following optimum policy:

∞
(i) If 0 tdh(t) > c O /c M , then there exists a finite and unique TO∗ (0 < TO∗ < ∞)
C O F (TO∗ ) = c M Q(TO∗ ). (2.48)

∞
(ii) If 0 tdh(t) ≤ c O /c M , then TO∗ = ∞, and the expected cost rate is given in
(2.46).
It can be easily seen that when h(∞) = ∞, a finite TO∗ (0 < TO∗ < ∞) always exits.
When F(t) = 1−e−(t/10) , Table 2.3 presents optimum K ∗ , TO∗ and their expected
2
cost rates. This indicates that optimum K ∗ and TO∗ increase with ci /c M (i = K , O)
and C O F (TO∗ ) < C(K ∗
∞). In this case, the mean time to replacement to the K ∗ th
failure is μ K ∗ = 1/ 0 p K ∗ (t)dt = 10(K ∗ + 0.5)/ (K ∗ ), and μ K ∗ > TO∗ >
μ K ∗ −1 , however, their differences are small as ci /c M is large.
Next, we derive optimum TF∗ and K F∗ which minimize C O F (T, K ) in (2.40) when
c O = c K and h(∞) = ∞. Differentiating C O F (T, K ) with respect to T and setting
it equal to zero,
T cO
P K (t)[Q(T ) − h(t)]dt = , (2.49)
0 cM
Table 2.3 Optimum TO∗ , K ∗ and their expected cost rates for ci /c M (i = K , O) when F(t) =
1 − e−(t/10)
2
ci /c M TO∗ C O F (TO∗ )/c M K ∗ C(K ∗ )/c M μ∗K

1.0 6.936 0.214 1 0.226 8.862
2.0 11.476 0.289 2 0.301 13.293
3.0 14.959 0.350 3 0.361 16.616
4.0 17.862 0.403 4 0.414 19.386
5.0 20.394 0.449 5 0.461 21.809
6.0 22.665 0.491 6 0.507 23.990
7.0 24.738 0.530 7 0.552 25.990
8.0 26.657 0.567 8 0.598 27.846
9.0 28.447 0.601 9 0.648 29.586
10.0 30.123 0.633 10 0.700 31.230
whose left-hand side increases strictly with T from 0 to ∞. Thus, there exists a finite
and unique TF∗ (0 < TF∗ < ∞) which satisfies (2.49), and the resulting cost rate is
C O F (TF∗ , K ) = c M Q(TF∗ ). (2.50)
Furthermore, noting that TF∗ decreases with K to TO∗ , optimum policy which mini-
mizes C O F (T, K ) is TF∗ = TO∗ given in (2.47) and K F∗ = ∞, i.e., the unit should be
replaced only over time TO∗ .
On the other hand, suppose that T (0 ≤ T < ∞) is fixed. From the inequality
C O F (T, K + 1) − C O F (T, K ) ≥ 0,
T
1
Q 2 (T, K − 1) P K (t)dt + P K (T )
0 Q(T )
T cO
− P K (t)h(t)dt − P K (T ) ≥ , (2.51)
0 cM
where
T T
p K −1 (t)h(t)dt p K (t)h(t)dt + p K (T )
Q 2 (T, K − 1) ≡ T 0
= T0 ,
0 p K −1 (t)[h(t)/Q(t)]dt 0 p K (t)dt + p K (T )/Q(T )
which increases strictly with K to Q(T ) from Appendix 3.2. Thus, the left-hand of
(2.51) increases strictly with K to T Q(T ) − H (T ), which agrees with that of (2.47).
Therefore, if T > TO∗ , then there exists a finite and unique minimum K F∗ (1 ≤ K F∗ <
∞) which satisfies (2.51), and conversely, if T ≤ TO∗ , then K F∗ = ∞.
2.5.2 Replacement Overtime Last with Number of Failures
Suppose that the unit is replaced at failure K (K = 0, 1, 2, . . .) or at the first failure

over time T (0 ≤ T ≤ ∞), whichever occurs last, i.e., it is replaced either at failure
K after time T or over time T after failure K . −1
The probability that the unit is replaced at failure K after time T is Kj=0 p j (T ) =
P K (T ), and the probability that it is replaced over time T after failure K is
∞
j=K p j (T ) = PK (T ). Thus, the expected number of failures until replacement is
∞
∞

K P K (T ) + ( j + 1) p j (T ) = K + ( j − K + 1) p j (T )
j=K j=K
∞
= H (T ) + P K (t)h(t)dt + PK (T )
T∞
= H (T ) + P K −1 (t)h(t)dt, (2.52)
T
and the mean time to replacement is

∞ ∞
T ∞
1
td PK (t) + udF(u) d P j (t)
T j=K 0 F(t) T
∞ ∞ ∞
T H (t) j
= td PK (t) + udF(u) h(t)dt
T T j!
j=K −1 0
∞ ∞
−[H (t)−H (T )]
=T+ P K (t)dt + PK (T ) e dt, (2.53)
T T
which agrees with (2.52) when H (t) = t.


c O + (c K − c O )P K (T ) + c M [H (T ) + T∞ P K (t)h(t)dt + PK (T )]
C O L (T, K ) = ∞ ∞ .
T + T P K (t)dt + PK (T ) T e−[H (t)−H (T )] dt
(2.54)
Clearly, C O L (T, 0) = C O L (T, 1), C O L (0, K ) = C O F (∞, K ) = C(K ) in (2.41)

for c O = c K , and C O L (T, 0) = C O F (T, ∞) = C O F (T ) in (2.44).
We find optimum TL∗ and K L∗ which minimize C O L (T, K ) in (2.54), when c O =
c K and h(∞) = ∞. Differentiating C O L (T, K ) with respect to T and setting it equal
to zero,
∞ ∞
cO
Q(T ) T + P K (t)dt − H (T ) + P K (t)h(t)dt = , (2.55)
T T cM
whose left-hand side increases strictly with T from
K −1 ∞
1
p j (t)dt − μ < 0
μ 0
j=0
to ∞. Thus, there exists a finite and unique minimum TL∗ (0 < TL∗ < ∞) which
satisfies (2.55), and the resulting cost rate is
C O L (TL∗ , K ) = c M Q(TL∗ ). (2.56)
Furthermore, letting L 1 (K , T ) be the left-hand side of (2.55),

∞
∞ p K (t)h(t)dt
L 1 (K , T ) − L 1 (K + 1, T ) = p K (t)dt ∞
T
− Q(T ) > 0,
T T p K (t)dt
∞ ∞
because T H (t) K dF(t)/ T H (t) K F(t)dt increases strictly with K from Q(T )
by similar proof in Appendix 3.3, i.e., L 1 (K , T ) decrease with K from that of
(2.47). Thus, TL∗ increases with K from TO∗ , and optimum policy which minimizes
C O L (T, K ) is TL∗ = TO∗ given in (2.47) and K L∗ = 0.
On the other hand, suppose that T (0 ≤ T < ∞) is fixed. From the inequality
C O L (T, K + 1) − C O L (T, K ) ≥ 0,
∞
2 (T, K − 1) T + PK (T )
Q P K (t)dt +
T Q(T )
∞
cO
− H (T ) + P K (t)h(t)dt + PK (T ) ≥ , (2.57)
T c M
where
∞
p K −1 (t)h(t)dt
2 (T, K − 1) ≡ ∞
Q T
.
T p K −1 (t)[h(t)/Q(t)]dt
The left-hand side of (2.57) increases strictly with K from

1
Q 2 (T, 0) T + − H (T ) − 1 > T Q(T ) − H (T )
Q(T )
2 (T, 0) > Q(T ) from Appendix 3.3. Letting TO be a solution of

in (2.47), because Q

2 (T, 0) T + 1 cO
Q − H (T ) − 1 = ,
Q(T ) cM
we have TO < TO∗ . Thus, if T ≥ TO , then K L∗ = 0 or 1, and conversely, if T < TO ,

then there exists a finite and unique K L∗ (1 ≤ K L∗ < ∞) which satisfies (2.57).
From the above interesting results for a fixed T that if T < TO then we should
adopt replacement overtime last, if TO ≤ T ≤ TO∗ then we should adopt replacement
overtime, and if T > TO∗ then we should adopt replacement overtime first.
2.6 Replacement Overnumber
In this section, we propose two overnumber policies in which the unit is replaced at
the first failure over number of cycle N and at the first working cycle over number
of failure K .
2.6.1 Replacement Over Number N
The unit is replaced at the first failure over number N (N = 0, 1, 2 . . .) of working

cycles. Then, the mean time to replacement is, when F(t) = 1 − e−H (t) ,
∞ 1
∞
udF(u) dG (N ) (t)
0 F(t) t
∞ ∞ ∞
1
= [1 − G (N ) (t)]dt + F(u)du dG (N ) (t)
0 0 F(t) t
∞
h(t)
=μ+ [1 − G (N ) (t)] dt. (2.58)
0 Q(t)
The expected number of failures until replacement is

∞ ∞
(N )
1+ H (t)dG (t) = 1 + [1 − G (N ) (t)]h(t)dt, (2.59)
0 0
which agrees with (2.58) when H (t) = t, i.e., h(t) = Q(t) = μ = 1.

∞
c N + c M {1 + 0 [1 − G (N ) (t)]h(t)dt}
C O (N ) = ∞ (N = 0, 1, 2, . . .). (2.60)
μ + 0 [1 − G (N ) (t)][h(t)/Q(t)]dt
When N = 0, i.e., the unit is replaced at the first failure,
cN + cM
C O (0) = = C(1), (2.61)
μ
2.6 Replacement Overnumber 53
in (2.45) when c N = c O , and when N = ∞, i.e., it always undergoes only minimal

repair,
C O (∞) ≡ lim C O (N ) = c M h(∞), (2.62)

N →∞
in (2.5).
Next, the unit is replaced at failure K (K = 1, 2, . . .) or at the first failure over
number N (N = 0, 1, 2, . . .), whichever occurs first. The probability that the unit is
replaced at failure K is
∞ ∞
[1 − G (N ) (t)]d PK (t) = PK (t)dG (N ) (t), (2.63)
0 0
and the probability that it is replaced over number N is

∞
P K (t)dG (N ) (t). (2.64)
0
The mean time to replacement is

∞ ∞ ∞
1
t[1 − G (N ) (t)]d PK (t) + P K (t) udF(u) dG (N ) (t)
0 F(t) t 0
∞ ∞
1
= P K (t)[1 − G (N ) (t)]dt + P K (t) dG (N ) (t)
0 0 Q(t)
∞
(N ) h(t)
=μ+ [1 − G (t)]P K −1 (t) dt, (2.65)
0 Q(t)
∞
K −1 ∞
(N )
K PK (t)dG (t) + ( j + 1) p j (t)dG (N ) (t)
0 j=0 0

K −1 ∞
=K− (K − 1 − j) p j (t)dG (N ) (t)
j=0 0

K −1 ∞
= P j (t)dG (N ) (t), (2.66)
j=0 0
which agrees with (2.65) when H (t) = t, i.e., h(t) = Q(t) = μ = 1.

C O F (N , K ) =
∞ (N ) (t) + c
K −1 ∞ (N ) (t)
c K − (c K − c N ) 0 P K (t)dG M j=0 0 P j (t)dG
∞ .
μ + 0 [1 − G (N ) (t)]P K −1 (t)[h(t)/Q(t)]dt
(2.67)
Clearly, C O F (N , ∞) = C O (N ) in (2.60) and C O F (∞, K ) = C(K ) in (2.41).
2.6.2 Replacement over Number K
The unit is replaced at the first completion of working cycles over number K (K =
0, 1, 2, . . .) of failures. Then, the mean time to replacement is
∞
∞ t ∞
( j)
(u + y)dG(y) dG (u) d PK (t)
j=0 0 0 t−u
∞ ∞
∞ t ∞
( j)
= P K (t)dt + G(y − u)dy dG (u) d PK (t). (2.68)
0 j=0 0 0 t

∞
∞
t ∞
[H (u + y) − H (t) + K ]dG(y) dG ( j) (u) d PK (t)
j=0 0 0 t−u
∞ ∞ t
∞
=K+ G(y − u)h(y)dy dG ( j) (u) d PK (t), (2.69)
j=0 0 0 t

∞ t ∞
c K + c M (K + ∞ ( j)
j=0 0 { 0 [ t G(y − u)h(y)dy]dG (u)}d PK (t))
C O (K ) = ∞
∞ ∞ t ∞
( j)
0 P K (t)dt + j=0 0 { 0 [ t G(y − u)dy]dG (u)}d PK (t)
(K = 0, 1, 2, . . .).
(2.70)
When K = 0, i.e., the unit is replaced at the first working cycle is

2.6 Replacement Overnumber 55
∞
cK + cM G(t)h(t)dt
C O (0) = 0
= C R (1)
1/θ
in (2.6) when c K = c N , and when K = ∞, the expected cost rate is given in (2.62).
When G(t) = 1 − e−θt , the expected cost rate in (2.70) is
C O (K ) =

c K + c M (K + 0∞ P K (t){ 0∞ θe−θu [h(u + t) − h(t)]du}dt + 0∞ e−θt h(t)dt)
∞ .
0 P K (t)dt + 1/θ
(2.71)
Next, the unit is replaced at cycle N (N = 1, 2, . . .) before failure K (K =

0, 1, 2, . . .) or at the first working cycle over number K of failures, whichever occurs
first. The probability that the unit is replaced at cycle N is
∞ ∞
P K (t)dG (N ) (t) = G (N ) (t)d PK (t), (2.72)
0 0
and the probability that it is replaced over number K is

∞ ∞
[1 − G (N ) (t)]d PK (t) = PK (t)dG (N ) (t). (2.73)
0 0

∞
t P K (t)dG (N ) (t)
0

N −1 ∞ t ∞
+ (u + y)dG(y) dG ( j) (u) d PK (t)
j=0 0 0 t−u
∞
= [1 − G (N ) (t)]P K (t)dt
0
N−1 ∞ t ∞
( j)
+ G(y − u)dy dG (u) d PK (t). (2.74)
j=0 0 0 t

N −1 ∞ t ∞
( j)
[H (u + y) − H (t) + K ]dG(y) dG (u) d PK (t)
j=0 0 0 t−u

K −1 ∞
+ j p j (t)dG (N ) (t)
j=0 0

K ∞
=K− (K − j) p j (t)dG (N ) (t)
j=0 0

N −1 ∞ t ∞
( j)
+ G(y − u)h(y)dy dG (u) d PK (t)
j=0 0 0 t
∞
= [1 − G (N ) (t)]P K (t)h(t)dt
0
N−1 ∞ t ∞
+ G(y − u)h(y)dy dG ( j) (u) d PK (t), (2.75)
j=0 0 0 t
which agrees with (2.74) when h(t) = t.

∞
c K − (c K − c N ) 0 P K (t)dG (N ) (t)
∞
+ c M ( 0 [1 − G (N ) (t)]P K (t)h(t)dt
−1 ∞ t ∞ ( j)
+ Nj=0 0 { 0 [ t G(y − u)h(y)dy]dG (u)}d PK (t))
C O F (K , N ) = ∞ .
[1 − G (N ) (t)]P K (t)dt
−1 ∞ t ∞
0
+ Nj=0 ( j)
0 { 0 [ t G(y − u)dy]dG (u)}d PK (t)
(2.76)
Clearly, C O F (K , ∞) = C O (K ) in (2.70), C O F (∞, N ) = C R (N ) in (2.6) when

c K = c N . When G(t) = 1 − e−θt ,
∞
c K − (c K − c N ) 0 [θ(θt) N −1 /(N − 1)!]e−θt P K (t)dt
−1 ∞ −θt P (t)h(t)dt
+ c M { Nj=0 0 [(θt) /j!]e
j
K
N −1 ∞ ∞ −θu
+ j=0 0 [(θt) /j!][ t e h(u)du]d PK (t)}
j
C O (K , N ) = N −1 ∞ . (2.77)
−θt P (t)dt
0 [(θt) /j!]e
j
j=0 K
−1 ∞ −θt d P (t)
+(1/θ) Nj=0 0 [(θt) /j!]e
j
K
In general, it would be be very difficult to discuss analytically optimum policies to

minimize C O (K ) in (2.71) and C O F (K , N ) in (2.76), which would be interesting
problems for further studies.
2.7 Preventive Maintenance Overtime
When the unit finishes each work of a job, we do some preventive maintenance (PM)
which is imperfect [2, p.171], [6]. It is assumed that the PM is done at the completion
of successive working cycles Y j and let b j denotes the imperfect PM factor after the
2.7 Preventive Maintenance Overtime 57
jth PM. The failure rate after the first PM becomes b1 h(t) when it was h(t) before
PM, i.e., the unit has the failure rate B j h(t) during ( j + 1)th working cycle, where
j
1 ≡ b0 < b1 ≤ b2 ≤ · · · , B j = i=0 bi ( j = 0, 1, . . .) and 1 = B0 < B1 < B2 · · ·
[2, p.194].
Suppose that the unit is replaced at the first completion of working cycles Y j ( j =
1, 2, . . .) over time T (0 ≤ T < ∞) introduced in Sect. 2.2. The mean time to
replacement and the expected number of failures before replacement for perfect PM
have been derived in (2.15) and (2.16), respectively. Because the unit has the failure
rate B j h(t) during the ( j + 1)th working cycle, the expected number of failures
before replacement in (2.16) is rewritten as
∞
T ∞
T
Bj [G ( j) (t) − G ( j+1) (t)]h(t)dt + G(u − t)h(u)du dG ( j) (t) .
j=0 0 0 T
(2.78)
Thus, from (2.15), the expected cost rate is

T
cM ∞ B j { 0 [G ( j) (t) − G ( j+1) (t)]h(t)dt
T ∞
j=0
+ 0 [ T G(u − t)h(u)du]dG ( j) (t)} + c P M(T ) + c O
C O M (T ) = T ∞ , (2.79)
T+ ∞ j=0 0 [ T G(u − t)du]dG (t)
( j)

where M(T ) ≡ ∞ ( j)
j=1 G (T ), c P = PM cost for the completion of each cycle with
c P ≤ c O , and c M and c O are given in (2.17).
In particular, when G(t) = 1 − e−θt ,
T
cM ∞ j=0 B j { 0 [(θt) /j!]e
j −θt h(t)dt
∞ −θt
+[(θT ) /j!] T e h(t)dt} + c P θT + c O
j
C O M (T ) = , (2.80)
T + 1/θ
which agrees with (2.20) when B j ≡ 1 and c P ≡ 0.

We find optimum TB∗ which minimizes C O M (T ). Differentiating C O M (T ) with

∞ ∞

1 (θT ) j
T+ θe−θt h(t + T )dt B j+1
θ 0 j!
j=0
∞
∞
T (θt) j −θt (θT ) j cO − c P
− Bj e h(t)dt + e−θt h(t + T )dt = .
0 j! j! 0 cM
j=0
(2.81)
Letting L(T ) be the left-hand side of (2.81), it increases strictly with T from
∞
L(0) = (B1 − B0 ) e−θt h(t)dt
0
to L(∞). Thus, if L(∞) > (c O − c P )/c M , then there exists a finite and unique
TB∗ (0 ≤ TB∗ < ∞) which satisfies (2.81), and the resulting cost rate is
∞
∞
C O M (TB∗ ) = cM B j+1 θe−θt h(TB∗ + t)dt. (2.82)
j=0 0
Next, it is assumed that when the PM is done at the jth working cycle, the age
t is reduced to a j t (0 < a j ≤ 1) where a0 ≡ 1 [2, p.192], i.e., the age becomes
j
(1 − a j )t units younger after each PM, where A j ≡ i=0 ai ( j = 0, 1, 2, . . .) and
1 = A0 > A1 > · · · . Then, replacing B j H (t) in (2.79) with H (A j t), the expected
cost rate is
T ( j)
cM ∞ j=0 { 0 [G (t) − G
( j+1) (t)]dH (A t)
j
T ∞ (
+ 0 [ T G(u − t)dH (A j u)]dG (t)} + c P M(T ) + c O
j)
O M (T ) =
C T ∞ . (2.83)
T+ ∞ j=0 0 [ T G(u − t)du]dG (t)
( j)
In particular, when G(t) = 1 − e−θt ,

T
cM ∞ j=0 { 0 [(θt) /j!]e
j −θt dH (A t)
j
∞
+[(θT ) j /j!] T e−θt dH (A j t)} + c P θT + c O
O M (T ) =
C . (2.84)
T + 1/θ
References
3. Tadj L, Ouali MS, Yacout S, Ait-Kadi S (eds) (2011) Replacement models with minimal repairs.
Springer, London
5. Nakagawa T (2011) Stochastic processes with applications to reliability theory. Springer, Lon-
don
6. Chen M, Nakamura S, Nakagawa T (2010) Replacement and preventive maintenance models
with random working times. IEICE Trans Fundam E 93-A:500–507
7. Zhao X, Nakagawa T (2014) Comparisons of periodic and random replacement policies:
Frenkel I et al (eds) Applied reliability engineering and risk analysis, probabilistic models
and statistical inference, Wiley, New York, pp 193–204
8. Zhao X, Al-Khalifa KN, Hamouda AMS, Nakagawa T (2015) First and last triggering event
approaches for replacement with minimal repairs. IEEE Trans Reliab (To Appear in)
9. Zhao X, Mizutani S, Nakagawa T (2015) Which is better for replacement policies with con-
tinuous or discrete schedule time ? Eur J Oper Res 242:477–486
10. Zhao X, Nakagawa T, Zuo M (2014) Optimal replacement last with continuous and discrete
policies. IEEE Trans Reliab 63:868–880
11. Zhao X, Qian C, Nakamura S (2014) Optimal age and periodic replacement with overtime
policies. J Reliab Qual Saf Eng 21, 1450016 (14 pages)
Chapter 3
Inspection Overtime
Most systems such as standby electric generators and defense systems, and some
units in complex and large-scale systems such as aircrafts and industry plants, have
to be checked at suitable times to search their faults and to detect their failures. This is
called inspection policy, which plays an important role in reliability theory. Optimum
policies which minimize the total expected cost until failure detection were derived
[1, p. 107] and were summarized [2, p. 201]. Some industry and computer systems
execute successive jobs and process. For such systems, it would be impossible or
impractical to make maintenances in a strict periodic fashion. From such a viewpoint,
random inspection policies in which a unit is checked at some completion of working
times were introduced and their optimum policies were discussed analytically and
numerically [3, p. 87, 6–9] .
Referring to the above reliability models, we first consider the standard periodic
and random inspection policies in Sect. 3.1. Next, when the failure time is exponen-
tial, we take up inspection first and inspection last [3, p. 101, 10] , and derive their
optimum policies in Sect. 3.2. Furthermore, combining periodic and random policies,
we propose inspection overtime in which a unit is checked at the first completion
of working times over time T in Sect. 3.3, and compare it with the other inspection
policies [3, p. 108, 10] . Finally, applying such inspection policies to a backup policy
in which failures are detected immediately and the backup recovery is executed until
the latest checking time, all results of inspection policies are rewritten for periodic
and random backup in Sect. 3.5 and for backup overtime in Sect. 3.6 [3, p. 115, 11,
12] . Furthermore, we try to form general inspection first and last with n variables,
obtain their expected costs, and derive optimum polices which minimize them when
the failure time is exponential in Sect. 3.5.
Throughout this chapter, we suppose that a unit has a failure distribution F(t)
∞
with finite μ ≡ 0 F(t)dt < ∞, where Φ(t) ≡ 1 − Φ(t) for any distribution. After
Sect. 2.2, the unit is assumed to have an exponential distribution F(t) = 1 − e−λt
(0 < λ < ∞).

60 3 Inspection Overtime
3.1 Periodic and Random Inspections
Suppose that the unit is checked at successive working times S j ≡ Y1 +Y2 +· · ·+Y j
( j = 1, 2, . . .) denoted in Sect. 1.1 and also at periodic times kT (k = 1, 2, . . .)
for a specified
∞ T (0 < T ≤ ∞), where G(t) = Pr{Y j ≤ t} with mean time
1/θ ≡ 0 G(t)dt (0 < θ < ∞). The inspection process will end when the failure
is certainly detected at either random or periodic checking times, whichever occurs
first, as shown in Fig. 3.1.
The probability that the failure is detected by periodic check is
⎧ ⎫
∞
(k+1)T ⎨∞ t ⎬
G[(k + 1)T − x]dG ( j) (x) dF(t), (3.1)
⎩ ⎭
k=0 kT j=0 0
and the probability that it is detected by random check is

⎛ ⎞
∞
(k+1)T ∞
t
⎝ {G[(k + 1)T − x] − G(t − x)}dG ( j) (x)⎠ dF(t), (3.2)
k=0 kT j=0 0
where note that (3.1) + (3.2) = 1.

Let cT be the cost for periodic check, cR be the cost for random check, and c D
be the downtime cost per unit of time for the time elapsed between a failure and its
detection at the next check. Then, the total expected cost until failure detection is
[2, p. 254]
∞
(k+1)T
∞
C(T ) = {(k + 1)cT + jcR + c D [(k + 1)T − t]}
k=0 kT j=0
t
× G[(k + 1)T − x]dG ( j) (x) dF(t)
0
(k-1)T Sj kT (k+1)T S j+1
T T
(k-2)T S j (k-1)T kT S j+1 (k+1)T
T T T
Check at periodic and random times Detection of failure Failure
Fig. 3.1 Process of periodic and random inspections

3.1 Periodic and Random Inspections 61
∞
(k+1)T
∞ t (k+1)T −x
+ [kcT + ( j + 1)cR
k=0 kT j=0 0 t−x

( j)
+ c D (x + y − t)]dG(y) dG (x) dF(t). (3.3)
In particular, when T = ∞, i.e., the unit is checked only by random check, the total
expected cost is
CR (∞) ≡ lim C(T )

T →∞
∞ ∞
= cR [1 + M(t)]dF(t) + c D F(t)G(t)dt
0 0
∞ ∞
+ [F(t + x) − F(x)]G(t)dt dM(x)
0 0
∞
cD
= cR + 1+ F(t)dM(t) − c D μ, (3.4)
θ 0
∞ ( j)
where M(t) ≡ j=1 G (t) represents the expected number of random checks
during (0, t].
Next, when G(t) = 1 − e−θt and F(t) = 1 − e−λt for λ < θ, (3.3) is simplified
as

cT cR θ cD λ e−λT − e−θT
C(T ) = + + cR − cT + 1− .
1 − e−λT λ θ θ − λ 1 − e−λT
(3.5)
Clearly,
C(0) ≡ lim C(T ) = ∞,

T →0

θ cD
C(∞) ≡ lim C(T ) = cR +1 + . (3.6)
T →∞ λ θ
We find optimum T ∗ which minimizes C(T ). Differentiating C(T ) with respect

to T and setting it equal to zero,
θ cT
[1 − e−(θ−λ)T ] − (1 − e−θT ) = , (3.7)
θ−λ cR − cT + c D /θ
whose left-hand side increases strictly with T from 0 to λ/(θ − λ). Therefore, if
cR + c D /θ > (θ/λ)cT , then there exists a finite and unique T ∗ (0 < T ∗ < ∞)
which satisfies (3.7). Conversely, if cR + c D /θ ≤ (θ/λ)cT , then T ∗ = ∞, i.e.,
periodic inspection is not needed, and the expected cost is given in (3.6).
In particular, when θ → 0, i.e., 1/θ → ∞,
cT + c D T cD
C P (T ) = − , (3.8)
1 − e−λT λ
which represents the standard inspection policy only with time T . Optimum TS∗
which minimizes C P (T ) satisfies
1 λT cT
(e − 1) − T = , (3.9)
λ cD
and the resulting cost is

c D λT ∗
C P (TS∗ ) = (e S − 1). (3.10)
λ
Because
1 λT 1 1 − e−θT
(e − 1) − T > [1 − e−(θ−λ)T ] −
λ θ−λ θ
for T, θ > 0, we have TS∗ < T ∗ when cR = cT .

When λ = θ, the total expected cost is

cT cD λT e−λT
C(T ) = + cR + cR − cT + 1− . (3.11)
1 − e−λT λ 1 − e−λT
If cR + c D /λ > cT , then there exists a finite an unique T ∗ which satisfies

cT
λT − (1 − e−λT ) = . (3.12)
cR − cT + c D /λ
Similarly, when λ > θ, optimum T ∗ satisfies uniquely
θ cT
[e(λ−θ)T − 1] − (1 − e−θT ) = . (3.13)
λ−θ cR − cT + c D /θ
This indicates that optimum T ∗ satisfies (3.13) at first and increases with 1/λ from 0,
becomes equal to a solution of (3.12) when 1/λ = 1/θ, and after that, satisfies (3.7)
and increases with 1/λ to ∞.
Next, suppose that the unit is checked at every N th (N = 1, 2, . . .) working times
S j N ( j = 1, 2, . . .), i.e., at the j N th number of works, and also at periodic times
kT (k = 1, 2, . . .), whichever occurs first. Then, the total expected cost until failure
detection
∞ is, by replacing formally G(t) and M(t) with G (N ) (t) and M (N ) (t) ≡
( j N ) (t) (N = 1, 2, . . .) in (3.3) and (3.4), respectively,
j=1 G
3.1 Periodic and Random Inspections 63
∞
(k+1)T
∞
C(T, N ) = {(k + 1)cT + jcR + c D [(k + 1)T − t]}
k=0 kT j=0
t
(N ) ( j N)
× 1−G [(k + 1)T − x] dG (x) dF(t)
0
∞ (k+1)T
∞ t (k+1)T −x
+ [kcT + ( j + 1)cR
k=0 kT j=0 0 t−x

+ c D (x + y − t)]dG (N ) (y) dG ( j N ) (x) dF(t), (3.14)
CR (N ) ≡ lim C(T, N )
T →∞
∞
N cD
= cR + 1+ F(t)dM (N ) (t) − c D μ (N = 1, 2, . . .).
θ 0
(3.15)
In addition, when F(t) = 1 − e−λt ,

∞ ∞
∞ [G ∗ (λ)] N
e−λt dM (N ) (t) = e−λt dG ( j N ) (t) = ,
0 1 − [G ∗ (λ)] N
j=1 0
∞
where G ∗ (s) is the LS transform of G(t), i.e., G ∗ (s) ≡ 0 e−st dG(t) for Re(s) > 0.
In this case, the expected cost in (3.15) is
cR + N c D /θ cD
CR (N ) = ∗
− (N = 1, 2, . . .). (3.16)
1 − [G (λ)] N λ
We find optimum number N ∗ which minimizes CR (N ). From the inequality

CR (N + 1) − CR (N ) ≥ 0,
N
j
1 cR
−N ≥ , (3.17)
G ∗ (λ) c D /θ
j=1
whose left-hand side increases strictly from 1/G ∗ (λ) − 1 to ∞. Therefore, there
exists a finite and unique minimum N ∗ (1 ≤ N ∗ < ∞) which satisfies (3.17). When
G(t) = 1 − e−θt , the expected cost in (3.16) is
cR + N c D /θ cD
CR (N ) = − , (3.18)
1 − [θ/(λ + θ)] N λ
and from (3.17), optimum N ∗ satisfies
N

λ j cR
1+ −N ≥ , (3.19)
θ c D /θ
j=1
whose left-hand increases strictly with N from λ/θ to ∞. Clearly, N ∗ increases with
1/λ from 1 to ∞ and decreases with 1/θ from ∞ to 1.
3.2 Inspection First and Last
As modified inspection policies, we propose the following two policies of inspec-

tion first and inspection last, and derive optimum policies which minimize the total
expected costs and compare them [3, p. 101, 10] .
3.2.1 Inspection First
Suppose that the unit is checked at time T (0 < T ≤ ∞) or at a random working

time Y j ( j = 1, 2, . . .), whichever occurs first, where Y j has an identical distribution
G(t) ≡ Pr{Y j ≤ t}. In this case, Z j ≡ min{T, Y j } forms a renewal process with an
interarrival distribution Pr{Z j ≤ t} = G(t) for t < T , 1 for t ≥ T .
It is assumed that the failure time has an exponential distribution F(t) = 1−e−λt .
Then, the probability that the unit does not fail and is checked at time T is
G(T )F(T ), (3.20)
the probability that it does not fail and is checked at time Y j is

T
F(t)dG(t), (3.21)
0
the probability that it fails and its failure is detected at time T is
G(T )F(T ), (3.22)
and the probability that it fails and its failure is detected at time Y j is
T
F(t)dG(t), (3.23)
0
where note that (3.20) + (3.21) + (3.22) + (3.23) = 1.

3.2 Inspection First and Last 65
From (3.20)–(3.23), the mean downtime l D from failure to its detection is given
by a renewal equation
T
l D ≡ G(T )F(T ) + F(t)dG(t) l D
0
T T t
+ (T − t)G(T )dF(t) + (t − u)dF(u) dG(t).
0 0 0
By solving the above renewal equation and arranging it,

T
G(t)F(t)dt
l D = 0T . (3.24)
0 G(t)dF(t)
In a similar way, the expected number MT of checks at time T until failure detection
is given by a renewal equation
T
MT = (1 + MT )G(T )F(T ) + MT F(t)dG(t) + G(T )F(T ),
0
i.e.,
G(T )
MT = T . (3.25)
0 G(t)dF(t)
The expected number M R of checks at time Y j until failure detection is given by a

renewal equation
T T
M R = (1 + M R ) F(t)dG(t) + M R G(T )F(T ) + F(t)dG(t),
0 0
i.e.,
G(T )
MR = T . (3.26)
0 G(t)dF(t)
Therefore, the total expected cost until failure detection is
C F (T ) = cT MT + c R M R + c D l D
T
cT G(T ) + c R G(T ) + c D 0 G(t)F(t)dt
= T , (3.27)
0 G(t)dF(t)
where cT , cR and c D are given in (3.3). When G(t) = 1 − e−θt ,

cT + (cR − cT + c D /θ)(1 − e−θT ) c D

C F (T ) = − . (3.28)
[λ/(θ + λ)][1 − e−(θ+λ)T ] λ
In particular,
cT + c D T cD
lim C F (T ) = −λT
− , (3.29)
θ→0 1−e λ
which agrees with C P (T ) in (3.8) and represents the expected cost of periodic inspec-
tion, and

θ cD
lim C F (T ) = cR +1 + , (3.30)
T →∞ λ θ
which agrees with C(∞) in (3.6). This policy includes periodic and random inspec-
tions discussed in Sect. 3.1.
We find optimum TF∗ which minimizes C F (T ) in (3.28) for cR + c D /θ > cT .
Differentiating C F (T ) with respect to T and setting it equal to zero,
θ λ cT
(eλT − 1) − (1 − e−θT ) = , (3.31)
θ+λ θ+λ cR − cT + c D /θ
whose left-hand side increases strictly from 0 to ∞. Thus, there exists a finite and
unique TF∗ (0 < TF∗ < ∞) which satisfies (3.31), and the resulting cost is
θ(cR − cT ) + c D λT ∗ c D
C F (TF∗ ) = e F− . (3.32)
λ λ
Comparing (3.31) with (3.7), (3.12) and (3.13), respectively,
θ λ θ
(eλT − 1) − (1 − e−θT ) > [1 − e−(θ−λ)T ] − (1 − e−θT ),
θ+λ θ+λ θ−λ
1 λT 1
(e − 1) − (1 − e−λT ) > λT − (1 − e−λT ),
2 2
θ λ θ
(eλT − 1) − (1 − e−θT ) > [e(λ−θ)T − 1] − (1 − e−θT )
θ+λ θ+λ λ−θ
for T > 0, we have TF∗ < T ∗ . Furthermore, when cR = cT , (3.31) is rewritten as
1 λ cT
(eλT − 1) − (1 − e−θT ) = , (3.33)
θ+λ θ(θ + λ) cD
whose left-hand side increases with 1/θ from 0 to the left-hand side of (3.9). Thus,
TF∗ decreases with 1/θ to TS∗ .
3.2.2 Inspection Last
Suppose that the unit is checked at time T (0 ≤ T < ∞) or at random working

time Y j ( j = 1, 2, . . .), whichever occurs last. In this case,
Z j ≡ max{T, Y j } forms
a renewal process with an interarrival distribution Pr{ Z j ≤ t} = 0 for t < T , and
G(t) for t ≥ T .
It is assumed that the failure time has an exponential distribution F(t) = 1−e−λt .
Then, the probability that the unit does not fail and is checked at time T is
G(T )F(T ), (3.34)
the probability that it does not fail and is checked at time Y j is

∞
F(t)dG(t), (3.35)
T
the probability that it fails and its failure is detected at time T is
G(T )F(T ), (3.36)
and the probability that it fails and its failure is detected at time Y j is
∞
F(t)dG(t), (3.37)
T
where note that (3.34) + (3.35) + (3.36) + (3.37) = 1.

From (3.34)–(3.37), the mean downtime l D from failure to its detection is given
by a renewal equation
∞
l D = G(T )F(T ) + F(t)dG(t) l D
T
T ∞ t
+ (T − t)G(T )dF(t) + (t − u)dF(u) dG(t).
0 T 0
By solving the above renewal equation,

T ∞
F(t)dt + T G(t)F(t)dt
lD = 0
∞ . (3.38)
1 − T G(t)dF(t)
In a similar way, the expected number MT of checks at time T until failure

detection is given by a renewal equation
∞
MT = (1 + MT )G(T )F(T ) + MT F(t)dG(t) + G(T )F(T ),
T
i.e.,
G(T )
MT = ∞ . (3.39)
1 − T G(t)dF(t)
The expected number M R of checks at time Y j until failure detection is given by a

renewal equation
∞ ∞
M R = (1 + M R ) F(t)dG(t) + M R G(T )F(T ) + F(t)dG(t),
T T
i.e.,
G(T )
MR = ∞ . (3.40)
1 − T G(t)dF(t)
Therefore, the total expected cost until failure detection is

T ∞
cT G(T ) + cR G(T ) + c D [ 0 F(t)dt + T G(t)F(t)dt]
C L (T ) = ∞ . (3.41)
1 − T G(t)dF(t)
cT (1 − e−θT ) + cR e−θT + (c D /θ)(θT + e−θT ) c D

C L (T ) = − . (3.42)
1 − e−λT + [λ/(θ + λ)]e−(θ+λ)T λ
In particular,
lim C L (T ) = lim C F (T ) = lim C(T ),

θ→∞ θ→0 θ→0
lim C L (T ) = lim C F (T ) = lim C(T ),
T →0 T →∞ T →∞
where C(T ) is given in (3.5).

We find optimum TL∗ which minimizes C L (T ) in (3.42). Differentiating C L (T )

cT − cR θ λT λ −θT θ
(e − 1) + (1 − e ) +
eθT − 1 λ θ+λ θ+λ

c D θ λT λ −θT
+ [e − (1 + λT )] − e = cT . (3.43)
θ λ θ+λ
When cT = cR , (3.43) becomes
1 λT λ cT
[e − (1 + λT )] − e−θT = , (3.44)
λ θ(θ + λ) cD
whose left-hand side increases strictly with T from −λ/[θ(θ + λ)] to ∞. Thus,
L (0 < T
there exists a finite and unique T L < ∞) which satisfies (3.44). Therefore,
optimum TL∗ which minimizes C L (T ) is TL∗ ≥ T L for cR ≥ cT and T ∗ < T L for
L
cR < cT . Comparing (3.44) and (3.9), we have T > T ∗ . Furthermore, noting that
S
L increases with
the left-hand side of (3.44) decreases with 1/θ from that of (3.9), T
1/θ from TS .∗
3.2.3 Comparison of Inspection First and Last
We compare optimum policies for inspection first and last when F(t) = 1 − e−λt ,
G(t) = 1 − e−θt and cT = cR . In this case, the expected cost of inspection first is,
from (3.28),
cT + (c D /θ)(1 − e−θT ) cD
C F (T ) = − , (3.45)
[λ/(θ + λ)][1 − e−(θ+λ)T ] λ
optimum TF∗ satisfies (3.33), and the resulting cost is, from (3.32),
c D λT ∗
C F (TF∗ ) = (e F − 1). (3.46)
λ
The expected cost of inspection last is, from (3.42)
cT + (c D /θ)(θT + e−θT ) cD
C L (T ) = −λT −(θ+λ)T
− , (3.47)
1−e + [λ/(θ + λ)]e λ
optimum TL∗ which minimizes it is given in (3.44), and the resulting cost is
c D λT ∗
C L (TL∗ ) = (e L − 1). (3.48)
λ
Noting that C F (TF∗ ) = C L (TL∗ ) = C S (TS∗ ) when TF∗ = TL∗ = TS∗ , which means
that the sizes of optimum checking times determine whether the inspection policies
could save cost or not. By comparing (3.9) and (3.33),
1 λT 1 λ
[e − (1 + λT )] > (eλT − 1) − (1 − e−θT ),
λ θ+λ θ(θ + λ)
which follows that TF∗ > TS∗ and TF∗ decreases with 1/θ to TS∗ . Similarly, by com-
paring (3.9) and (3.44), TL∗ > TS∗ and TL∗ increases with 1/θ from TS∗ . Therefore,
periodic inspection is better than both inspection first and last when cT = cR .
Next, we compare inspection first and last. Let
θ λ −θT
Q(T ) ≡ [eλT − (1 + λT )] − e
λ θ+λ
θ λ
− (eλT − 1) + (1 − e−θT )
θ+λ θ+λ
θ λ θ
= [eλT − (1 + λT )] + (1 − 2e−θT ) − (eλT − 1), (3.49)
λ θ+λ θ+λ
which increases strictly with T from −λ/(θ + λ) to ∞. Thus, there exists a finite
and unique TI (0 < TI < ∞) which satisfies Q(T ) = 0. Therefore, from (3.44) and
(3.33), if
θ λ cT
L(TI ) ≡ (eλTI − 1) − (1 − e−θTI ) < ,
θ+λ θ+λ c D /θ
then TF∗ < TL∗ , and hence, inspection first is better than inspection last, and con-
versely, if L(TI ) > cT /(c D /θ), then TL∗ < TF∗ , and hence, inspection last is better
than inspection first. Because Q(T ) increases with θ, TI decreases with θ to 0, i.e.,
TI increases with 1/θ from 0.
In addition, because TF∗ decreases to TS∗ and TL∗ increases from TS∗ with 1/θ, TF∗
becomes equal to TL∗ at some 1/θ I (0 < θ I < ∞). This 1/θ I is computed by solving
two simultaneous equations of T and 1/θ for given cT /c D : From (3.33) and (3.44),
1 λ cT
(eλT − 1) − (1 − e−θT ) = ,
θ+λ θ(θ + λ) cD
1 λT λ cT
[e − (1 + λT )] − e−θT = .
λ θ(θ + λ) cD
When 1/θ I is obtained, it can be shown that if 1/θ > 1/θ I , then replacement first is
better than replacement last, and vice versa.
Table 3.1 presents optimum TF∗ , TL∗ and TI , 1/θ I , L(TI )/θ for cT /c D and 1/θ
when 1/λ = 1. This indicates that TF∗ decreases with 1/θ to TS∗ and TL∗ increases
with 1/θ from TS∗ and both TF∗ and TL∗ increase with cT /c D . When cT /c D is small,
i.e., L(TI )/θ < cT /c D , TF∗ < TL∗ and inspection first is better than inspection last.
Conversely, when cT /c D is large, i.e., L(TI )/θ > cT /c D , TL∗ < TF∗ and inspection
last is better than inspection first. Furthermore, 1/θ I increases with cT /c D , and
inspection first is better than inspection last if 1/θ > 1/θ I . For example, when
cT /c D = 0.050, 1/θ I = 0.298. So that, if 1/θ > 0.298 then inspection first is better
than inspection last, and conversely, if 1/θ < 0.298 then inspection last is better than
inspection first. In particular, when 1/θ = 1/θ I , TF∗ = TL∗ and two policies are the
same. Because 1/θ I increases with cT /c D , when cT /c D is large, inspection last is
Table 3.1 Optimum TF∗ , TL∗ , TI and 1/θ I when F(t) = 1 − e−t
cT /c D 1/θ = 0.1 1/θ = 0.2 1/θ = 0.5 1/θ = ∞ 1/θ I
TF∗ TL∗ TF∗ TL∗ TF∗ TL∗ TS∗
0.001 0.0479 0.0939 0.0461 0.1698 0.0450 0.3746 0.0444 0.050
0.002 0.0697 0.1012 0.0660 0.1737 0.0639 0.3762 0.0626 0.059
0.005 0.1168 0.1216 0.1069 0.1850 0.1016 0.3811 0.0984 0.094
0.010 0.1764 0.1511 0.1553 0.2030 0.1446 0.3891 0.1382 0.133
0.020 0.2727 0.1993 0.2279 0.2362 0.2061 0.4048 0.1936 0.188
0.050 0.5004 0.3017 0.3859 0.3190 0.3307 0.4492 0.3004 0.298
0.100 0.7884 0.4165 0.5817 0.4239 0.4739 0.5161 0.4162 0.422
0.200 1.1939 0.5723 0.8744 0.5747 0.6787 0.6297 0.5722 0.600
0.500 1.8871 0.8577 1.4350 0.8580 1.0792 0.8785 0.8577 0.947
1.000 2.4932 1.1462 1.9742 1.1462 1.4985 1.1539 1.1462 1.338
TI 0.1259 0.2444 0.5643
L(TI )/θ 0.0057 0.0226 0.1400
better than inspection first. It is of interest that when 1/θ = 0.5 and cT /c D = 0.100,
1/θ I = 0.422 < TF∗ = 0.4739 < 1/θ = 0.5 < TL∗ = 0.5161, and inspection times
are almost the same.
3.3 Inspection Overtime
Suppose that the unit is checked at the first completion of working times over time
T (0 ≤ T < ∞) when F(t) = 1 − e−λt and G(t) = 1 − e−θt . Such inspection
procedures continue to be done until failure detection, which is called inspection
overtime. The probability that the unit does not fail at some checking interval is
∞
∞
T θ
F(u)dG(u − t) dG ( j) (t) = e−λT ,
θ+λ
j=0 0 T
and the probability that it fails at some interval is

∞
∞
T θ
F(u)dG(u − t) dG ( j) (t) = 1 − e−λT .
θ+λ
j=0 0 T
Thus, the mean downtime l D from failure to its detection is

∞
∞ u
θ T
lD = e−λT l D + (u − x)dF(x) dG(u − t) dG ( j) (t)
θ+λ
j=0 0 T 0
θ 1 1 θ
= e−λT l D + T + − + e−λT ,
θ+λ θ λ λ(θ + λ)
i.e.,
T + 1/θ − 1/λ + {θ/[λ(θ + λ)]}e−λT

lD = . (3.50)
1 − [θ/(θ + λ)]e−λT
The expected number MC of checking times until failure detection is
θ θ
MC = (1 + MC ) e−λT + 1 − e−λT ,
θ+λ θ+λ
i.e.,
1
MC = . (3.51)
1 − [θ/(θ + λ)]e−λT
Therefore, from (3.50) and (3.51), the total expected cost until failure detection
is
c O + c D (T + 1/θ) cD
C O (T ) = −λT
− , (3.52)
1 − [θ/(θ + λ)]e λ
where c O = checking cost over time T and c D is given in (3.3). Clearly,
C O (∞) ≡ lim C O (T ) = ∞,
T →∞

θ cD
C O (0) ≡ lim C O (T ) = c O +1 + ,
T →0 λ θ
which agrees with (3.6) when c O = cR .

We find optimum TO∗ which minimizes C O (T ) in (3.52). Differentiating C O (T )

1 1 cO
+ (eλT − 1) − T = , (3.53)
λ θ cD
3.3 Inspection Overtime 73
whose left-hand side increases strictly from 0 to ∞. Thus, there exists a finite and
unique TO∗ (0 < TO∗ < ∞) which satisfies (3.53), and the resulting cost is

cD λ λT ∗
C O (TO∗ ) = 1+ e O −1 . (3.54)
λ θ
3.3.1 Comparisons with Other Policies
We compare inspection overtime with periodic inspection in which the expected cost
is given in (3.8) when cT = c O . It can be easily shown from (3.53) that TO∗ decreases
with 1/θ from TS∗ to 0, and TO∗ < TS∗ . On the other hand,

1 λ(T ∗ +1/θ) ∗ 1 1 1 ∗
[e O − 1] − TO + > + (eλTO − 1) − TO∗ ,
λ θ λ θ
which implies that TO∗ +1/θ > TS∗ . So that, comparing (3.10) with (3.54), C P (TS∗ ) <
C O (TO∗ ), i.e., periodic inspection is better than inspection overtime when c O = cT .
Furthermore, we compare TO∗ with TF∗ when cT = c O : From (3.33) and (3.53),

θ+λ θ λ
(eλT − 1) − θT − (eλT − 1) + (1 − e−θT )
λ θ+λ θ+λ
λ2 T λ
> + (1 − e−θT ) > 0.
θ+λ θ+λ
Thus, TO∗ < TF∗ . Similarly, from (3.44) and (3.53), TO∗ < TL∗ .
Next, assuming that c O < cT , then from (3.10) and (3.54), if

1
cT + c D TS∗ > c O + c D TO∗ + ,
θ
then inspection overtime is better than periodic inspection. Therefore, we obtain cO

in case when C P (TS∗ ) = C O (TO∗ ) for given cT and c D . First, we compute TS∗ from
O which satisfies
(3.9) and C P (TS∗ ) from (3.10). Using TS∗ and C P (TS∗ ), we obtain T

1 1 1 cT
+ (eλTO − 1) + = TS∗ + , (3.55)
λ θ θ cD
and from (3.54),

cO cT O + 1 .
= TS∗ + − T (3.56)
cD cD θ
Table 3.2 Optimum TO∗ for cT = c O and

c O for c O < cT when F(t) = 1 − e−t
cT /c D 1/θ = 0.01 1/θ = 0.05 1/θ = 0.1
TO∗
c O /c D TO∗
c O /c D TO∗
c O /c D
0.001 0.0355 – 0.0170 – 0.0095 –
0.002 0.0534 0.0012 0.0303 – 0.0182 –
0.005 0.0889 0.0045 0.0606 – 0.0407 –
0.010 0.1285 0.0097 0.0972 0.0010 0.0713 –
0.020 0.1838 0.0198 0.1503 0.0133 0.1190 –
0.050 0.2906 0.0498 0.2550 0.0454 0.2181 0.0323
0.100 0.4064 0.0998 0.3698 0.0964 0.3299 0.0863
0.200 0.5624 0.1998 0.5250 0.1972 0.4830 0.1892
0.500 0.8478 0.4999 0.8098 0.4979 0.7658 0.4919
1.000 1.1363 0.9999 1.0980 0.9982 1.0531 0.9931
Table 3.2 presents optimum TO∗ and c O /c D for 1/θ, cT /c D and c O /c D when
F(t) = 1 − e−t . Optimum TO∗ and c O /c D increase with cT /c D and decrease with
1/θ. Compared to Table 3.1, TO∗ < TS∗ < TO∗ + 1/θ. This indicates that c O /c D
approaches to cT /c D as cT /c D becomes larger. In other words, if cT (= c O ) becomes
higher, then TO∗ and TS∗ become larger, and both inspection overtime and periodic
inspection are almost the same. That is, the checking cost for inspection overtime
approaches to that for periodic inspection because two inspections are coincident
with each other. If TO∗ + 1/θ ≥ TS∗ + cT /c D , then there does not exist any positive

c O , i.e., inspection overtime cannot be better than periodic inspection.
3.4 General Failure Times
In general, we cannot use the same technique to analyze inspection first, inspection
last and inspection overtime, when the failure time is not exponential, because only
an exponential distribution has a memoryless property [4, p. 13]. For a general failure
time, we consider a random inspection in which the unit is checked
∞ at random times
Y j ( j = 1, 2, . . .) with G(t) ≡ Pr{Y j ≤ t} and mean 1/θ ≡ 0 G(t)dt. Then, the
expected cost has been already given in (3.4).
First, consider inspection first in which the unit is checked at times T or Y j ,
whichever occurs first. Then, by setting Z j ≡ min{T, Y j }, it has a distribution

G(t) t < T,
G F (t) ≡ Pr{Z j ≤ t} =
1 t ≥ T,
3.4 General Failure Times 75
and has a mean time

∞ T
1
= G F (t)dt = G(t)dt.
θF 0 0

∞
cD
CR (G F ) = cR + 1+ F(t)dM F (t) − c D μ,
θF 0
∞ ( j)
where cR = checking cost at time Z j and Mi (t) ≡ j=1 G i (t) (i = F, L , O).
In particular, when F(t) = 1 − e−λt ,
∞ G ∗F (λ)
e−λt dM F (t) = M F∗ (λ) = ,
0 1 − G ∗F (λ)
∞
where Φ ∗ (λ) ≡ 0 e−λt dΦ(t) for λ > 0 and any function Φ(t). Because
∞ T
G ∗F (λ) = e−λt dG F (t) = e−λT + G(t)λe−λt dt,
0 0
we have
1
M F∗ (λ) + 1 = T .
0 G(t)λe−λt dt
Thus, the expected cost is

T
cR + c D 0 G(t)(1 − e−λt )dt
CR (G F ) = T ,
−λt dt
0 G(t)λe
which agrees with (3.27) when cR = cT and F(t) = 1 − e−λt .

Next, consider inspection last in which the unit is checked at times T or Y j ,
whichever occurs last. Then, setting Z j ≡ max{T, Y j }, it has a distribution

0 t < T,
G L (t) ≡
G(t) t ≥ T,
and has a mean time

∞ ∞
1
= G L (t)dt = T + G(t)dt.
θL 0 T

∞
cD
CR (G L ) = cR + 1+ F(t)dM L (t) − c D μ.
θL 0
In particular, when F(t) = 1 − e−λt ,

∞ ∞
G ∗L (λ) ≡ e −λt
dG L (t) = G(t)λe−λt dt,
0 T
and
1
M L∗ (λ) + 1 = ∞ .
1− T G(t)λe−λt dt

∞
cR + c D [T − (1 − e−λT )/λ + T G(t)(1 − e−λt )dt]
CR (G L ) = ∞ ,
1 − T G(t)λe−λt dt
which agrees with (3.41) when cR = cT and F(t) = 1 − e−λt .

Finally, consider inspection overtime in which the unit is checked at Z j with the
first completion of working times over time T . Then, Z j has a distribution

0 t < T,
G O (t) ≡ ∞ T
j=0 0 [G(t − u) − G(T − u)]dG ( j) (u) t ≥ T,
and has a mean time
1
∞ ∞
∞ T
= G O (t)dt = T + G(t − u)dG ( j) (u) dt.
θO 0 j=0 T 0
Thus, from (3.4), the expected cost is

∞
cD
C O (G O ) = cR + 1+ F(t)dM O (t) − c D μ.
θO 0
In particular, when F(t) = 1 − e−λt and G(t) = 1 − e−θt ,

∞
∞ T
G ∗O (λ) = [e−θ(T −u) − e−θ(t−u) ]dG ( j) (u) λe−λt dt
j=0 T 0
θ
= e−λT ,
θ+λ
and
1 ∞ ∞ T
−θ(t−u) ( j)
=T+ e dG (u) dt
θO
j=0 T 0
1
=T+ .
θ
cR + c D (T + 1/θ) cD
CR (G O ) = −λT
− ,
1 − [θ/(θ + λ)]e λ
which agrees with (3.52) when c O = cR .

From the above results, this can be extended to inspection first and last with n
variables in which the unit is checked at times T and Y1k , Y2k , . . . , Ynk (k = 1, 2, . . .),
each of which has distribution G i (t) ≡ Pr{Yik ≤ t} (i = 1, 2, . . . , n) with mean
time 1/θi (0 < θi < ∞) for k = 1, 2, . . .. When the unit is checked at time
T (0 < T ≤ ∞) or at random times Y1k , Y2k , . . . , Ynk , whichever occurs first,
denote that
n
1 − i=1 G i (t) t < T,
G Fn (t) ≡
1 t ≥ T,
which has a mean time

n
!
1 T
≡ G i (t) dt.
θ Fn 0 i=1
The expected cost is

∞
cD cD
C Fn (T ) = cR + 1+ F(t)dM Fn (t) − ,
θ Fn 0 μ
( j)
where Min (t) = ∞ j=1 G in (t) (i = F, L).
In particular, when F(t) = 1 − e−λt , because
n
!
∞ T
G ∗Fn (λ) ≡ e −λt
dG Fn (t) = 1 − G i (t)λe −λt
dt,
0 0 i=1
we have
∗ 1
1 + M Fn (λ) = T "n # .
0 i=1 G i (t)λe−λt dt

T "n #
cR + c D i=1 G i (t) dt cD
C Fn (T ) = T "n 0 # − . (3.57)
G i (t)λe −λt dt λ
0 i=1
∗ which minimizes C (T ) is given by a solution of

Optimum TFn Fn
!
T n $ % cR
G i (t) eλ(T −t) − 1 dt = , (3.58)
0 cD
i=1
∗ c D λT ∗
C Fn (TFn )= (e Fn − 1). (3.59)
λ
Similarly, when the unit is checked at times T and Y1k , Y2k , . . . , Ynk , whichever
occurs last, the expected cost is
∞" n #
cR + c D {T + T 1 − i=1 G i (t) dt} c D
C Ln (T ) = "
∞ n # − . (3.60)
1− T i=1 G i (t)λe
−λt dt λ
∗ which minimizes C (T ) is given by a solution of

Optimum TLn Ln
!
eλT − (1 + λT ) ∞ n $ % cR
− 1− G i (t) 1 − e−λ(t−T ) dt = , (3.61)
λ T cD
i=1
∗ c D λT ∗
C Ln (TLn )= (e Ln − 1). (3.62)
λ
Furthermore, when cT is the inspection cost for time T and c Ri (i = 1, 2, . . . , n)
is the inspection cost for random check Yik , respectively, the expected cost in (3.57)
is
"n # T "n #
cT G i (T )$ + c D 0 i=1%G i (t) dt
n T n
i=1
+ i=1 c Ri 0 j=1, j=i G j (t) dG i (t) cD
C Fn (T ) = T "n # − , (3.63)
−λt λ
0 i=1 G i (t)λe dt
and (3.60) is
"n # ∞ n
cT G i (T )$ + c D {T + T [1% − i=1 G i (t)]dt}
n ∞ n
i=1
+ i=1 c Ri T j=1, j=i G j (t) dG i (t) cD
C Ln (T ) = ∞ "n # − . (3.64)
1− T i=1 i G (t)λe −λt dt λ
It has been shown that the expected cost is easily obtained from (3.4) when the unit
is
∞ checked at a renewal
∞process. However, it would be very difficult to derive explicitly
0 F(t)dM(t) = 0 M(t)dF(t) for a general failure distribution F(t). Using the
numerical techniques [4, p. 64], it would be possible to compute numerically expected
costs and to discuss optimum policies.
3.5 Periodic and Random Backup
Suppose that the unit is checked at successive random times S j ( j = 1, 2, . . .), and
also at periodic times kT (k = 1, 2, . . .) for a specified T > 0 in Fig. 3.1. Any
failure can be detected immediately, and then the backup recovery is executed until
the latest checking time to restore the system consistency. The process ends at the
backup recovery and starts newly from this checking point [5, p. 123].
We introduce the following costs for the above process: When a failure occurs
at time t between kT and (k + 1)T or S j+1 , the backup recovery is executed from
the failure time t to the latest checking time kT . This incurs a loss cost c D (t − kT )
which includes all costs resulting from the working time from kT to t and the backup
recovery from t to kT . On the other hand, when a failure occurs at time t between
S j and (k + 1)T or S j+1 , this incurs a loss cost c D (t − S j ). Costs cT and cR are
the same, and also, the other assumptions are the same in inspection policy of the
previous sections.
The probability that the backup recovery is executed to periodic check due to
some failure is
⎡ ⎤
∞ (k+1)T ∞ kT
⎣ G(t − x)dG ( j) (x)⎦ dF(t), (3.65)
k=0 kT j=0 0
and the probability that it is executed to random check is

⎡ ⎤
∞
(k+1)T ∞
t
⎣ G(t − x)dG ( j) (x)⎦ dF(t), (3.66)
k=0 kT j=0 kT
where note that (3.65) + (3.66) = 1. Therefore, the total expected cost until the backup
recovery is
C B (T ) =
⎧ ⎫
∞ (k+1)T ⎨∞ kT ⎬
[cT k + cRj + c D (t − kT )]G(t − x)dG ( j) (x) dF(t)
⎩ ⎭
k=0 kT j=0 0
⎧ ⎫
∞
(k+1)T ⎨∞ t ⎬
+ [cT k + cRj + c D (t − x)]G(t − x)dG ( j) (x) dF(t)
⎩ ⎭
k=0 kT j=0kT
∞
∞
= cT F(kT ) + cR M(t)dF(t) +
cD μ
k=1 0
⎧ ⎡ ⎤
⎨ ∞ (k+1)T ∞ kT
−cD (kT ) ⎣ G(t − x)dG ( j) (x)⎦ dF(t)
⎩ kT 0
k=0 j=0
⎡ ⎤ ⎫
∞ (k+1)T ∞ t ⎬
+ ⎣ x G(t − x)dG ( j) (x)⎦ dF(t) . (3.67)
kT kT ⎭
k=0 j=0
In particular, when Y j ≡ ∞, i.e., G(t) ≡ 0 for any t ≥ 0, the unit is checked only
at periodic times kT (k = 1, 2, . . .), and the total expected cost in (3.67) is
∞

B (T ) = (cT −
C cD T ) F(kT ) +
c D μ, (3.68)
k=1
which agrees with (5.55) of [5, p. 95]. When T = ∞, the unit is checked only at
random times S j ( j = 1, 2, . . .), and the total expected cost is
∞
R (∞) = cR
C M(t)dF(t) + cD μ
0
⎡ ⎤
∞ ∞ t
−
cD ⎣ x G(t − x)dG ( j) (x)⎦ dF(t), (3.69)
0 j=0 0
which agrees with [3, p. 118].

When F(t) = 1 − e−λt , the total expected costs in (3.68) and (3.69) are, respec-
tively,
CB (T ) = cT −
cD T
+

cD
, (3.70)
λT
e −1 λ
R (∞) = cR θ +
C
cD
. (3.71)
λ θ+λ
3.5 Periodic and Random Backup 81
We find optimum TB∗ which minimizes C B (T ) in (3.70). Differentiating C

B (T )
with respect to T and setting it equals to zero,
1 − e−λT cT
T− = , (3.72)
λ
cD
whose left-hand increases strictly from 0 to ∞. Thus, there exists a finite and unique
TB∗ (0 < TB∗ < ∞) which satisfies (3.72), and the resulting cost is
B (TB∗ ) =
C
cD ∗
(1 − e−λTB ). (3.73)
λ
3.6 Backup Overtime
Suppose that the unit is checked at the first completion of working times over time
T (0 ≤ T < ∞) when F(t) = 1 − e−λt and G(t) = 1 − e−θt . The backup procedure
continues until the unit fails and its recovery completion is done at the latest checking
point, which is called backup overtime. The probability that the unit does not fail at
some checking interval is
∞
∞
T θ
F(u)dG(u − t) dG ( j) (t) = e−λT , (3.74)
θ+λ
j=0 0 T
and the probability that it fails at some checking interval is

∞
∞
T θ
F(u)dG(u − t) dG ( j) (t) = 1 − e−λT . (3.75)
θ+λ
j=0 0 T
Thus, the mean recovery time

l D from failure to the latest checking time is
∞ ∞ u
θ T
lD = e−λT
lD + xdF(x) dG(u − t) dG ( j) (t)
θ+λ
j=0 0 T 0
θ λ 1
= e−λT
lD + [1 + (θ + λ)T ]e−λT + [1 − (1 + λT )e−λT ],
θ+λ (θ + λ) 2 λ
i.e.,
[λ/(θ + λ)2 ][1 + (θ + λ)T ]e−λT + (1/λ)[1 − (1 + λT )e−λT ]

lD = . (3.76)
1 − [θ/(θ + λ)]e−λT
The expected number MC of checking times until backup recovery is
C )
C = (1 + M θ
M e−λT ,
θ+λ
i.e.,
C = [θ/(θ + λ)]e−λT
M . (3.77)
1 − [θ/(θ + λ)]e−λT
Therefore, from (3.76) and (3.77), the total expected cost until backup recovery
is
c D {[λ/(θ + λ)2 ][1 + (θ + λ)T ]e−λT

cO +
+ (1/λ)[1 − (1 + λT )e−λT ]}
O (T ) =
C , (3.78)
1 − [θ/(θ + λ)]e−λT
where c O = checking cost over time T and

c D is given in (3.67). Clearly,
O (∞) ≡ lim C
C O (T ) = c O + cD
,
T →∞ λ

θ
cD
C O (0) ≡ lim C O (T ) = c O +1 + .
T →0 λ θ+λ
We find optimum TO∗ (0 ≤ TO∗ < ∞) which minimizes C O (T ) in (3.78). Differ-

entiating C O (T ) with respect to T and setting it equal to zero,
θ cO
T− (1 − e−λT ) = , (3.79)
λ(θ + λ)
cD
whose left-hand increases strictly with T from 0 to ∞. Thus, there exists a finite and
unique TO∗ (0 < TO∗ < ∞) and the resulting cost rate is

∗ 1
c D TO∗ +
C O (TO ) = . (3.80)
θ+λ
In addition, the left-hand side of (3.79) increases strictly with 1/θ from T − (1 −
e−λT )/λ which agrees with that of (3.72). Thus, TO∗ decreases with 1/θ from TB∗ to
c O /
c D when c O = cT .
Next, we compare backup overtime with periodic backup when c O = cT : From
(3.72) and (3.79),
θ 1 − e−λT 1 − e−λT
T− (1 − e−λT ) − T + = > 0,
λ(θ + λ) λ θ+λ
3.6 Backup Overtime 83
Table 3.3 Optimum TO∗ and TB∗ when F(t) = 1 − e−t for ci /
c D (i = T, O)
ci /
cD 1/θ = 0.01 1/θ = 0.05 1/θ = 0.1 1/θ = 0
TO∗ TO∗ TO∗ TB∗
0.001 0.036 0.018 0.010 0.045
0.002 0.055 0.032 0.020 0.064
0.005 0.092 0.064 0.045 0.102
0.010 0.135 0.104 0.079 0.145
0.020 0.197 0.164 0.134 0.207
0.050 0.324 0.289 0.254 0.334
0.100 0.473 0.437 0.399 0.483
0.200 0.697 0.660 0.620 0.707
0.500 1.188 1.151 1.109 1.198
1.000 1.832 1.794 1.751 1.841
which follows that TO∗ < TB∗ . Similarly,

1 1 λ θ
T+ − 1 − exp λT + −T + (1 − e−λT ) > 0,
θ+λ λ θ+λ λ(θ + λ)
because of

1 θ
exp −λ T + > e−λT .
θ+λ θ+λ
Thus,
1
TO∗ < TB∗ < TO∗ + ,
θ+λ
which concludes from (3.73) and (3.80) that, periodic backup is better than backup
overtime when both checking costs are the same.
Table 3.3 presents optimum TO∗ and TB∗ , and TO∗ decreases with 1/θ from TB∗ to
c D , and TO∗ < TB∗ < TO∗ + 1/(θ + 1) and c O /
c O / c D < TO∗ < TB∗ . The differences
∗ ∗
of TO and TB become smaller as ci / c D is large and 1/θ is small.
References
4. Nakagawa T (2011) Stochastic processes with applications to reliability theory. Springer, Lon-
don
5. Nakagawa T (2008) Advanced reliability models and maintenance policies. Springer, London
6. Nakagawa T, Mizutani S, Chen M (2010) A summary of periodic and random inspection
policies. Reliab Eng Syst Saf 95:906–911
7. Nakagawa T, Zhao X, Yun WY (2011) Optimal age replacement and inspection policies with
random failure and replacement times. Inter J Reliab Qual Saf Eng 18:1–12
8. Zhao X, Chen M, Nakagawa T (2014) Optimal time and random inspection policies for com-
puter systems. Appl Math Inf Sci 8:413–417
9. Chen M, Zhao X, Nakamura S (2014) Periodic and random inspections for a computer system.
In: Nakamura S, et al (eds) Reliability modeling with applications. World Scientific, Singapore,
pp 249–267
10. Zhao X, Nakagawa T (2015) Optimal periodic and random inspections with first, last, and
overtime policies. Int J Syst Sci 46(9):1648–1660
11. Nakagawa T, Naruse K, Maeji S (2009) Random checkpoint models with N tandem tasks.
IEICE Trans Fundam E92-A: 1572–1577
12. Naruse K, Maeji S (2014) Optimal checkpoint times for database system. In: Nakamura S, et al
(eds) Reliability Modeling with Applications, World Scientific, Singapore
Chapter 4
Replacement Overtime with Three Variables
In Chaps. 1 and 2, we have considered several replacement policies in which the unit
is replaced at time T or at cycle N , and at time T or at failure K . In this chapter,
we propose modified replacement policies with three variables in which the unit is
replaced at time T , and at cycle N , or at failure K : In Sect. 4.1, the classical approach
of whichever occurs first and the newly proposed approach of whichever occurs last
[1] are employed into modeling. Furthermore, we propose a new replacement policy
in which the unit is replaced at time T , at cycle N , or at failure K , whichever occurs
middle [2]. For example, the unit is replaced at time T in case of {t N < T < t K } or
{t K < T < t N }, there t N and t K are denoted by the respective occurrence times of
cycle N and failure K .
In Sect. 4.2, the unit is supposed to be replaced at the first completion of working
cycle over time T , at cycle N , or at failure K , whichever occurs first and last.
Furthermore, in Sect. 4.3, when the unit is replaced at the first failure over time
T , the same replacement policies are considered. For the above four replacement
policies, the expected cost rates are obtained.
In Sect. 4.4, as modified replacement policies, we propose three policies in which
the unit is replaced at times of max{t N , t K } before time T or at times of min{t N , t K }
after time T . For such replacement policy, we consider two overtime policies in which
the unit is replaced at the first completion of working cycles and at the first failure
over time T . Finally, in Sect. 4.5, we propose general replacement first and last in
which the unit is replaced before failure at time T or at n random times, whichever
occurs first and last. We obtain the expected cost rates and derive optimum times to
minimize them. By combining variously and giving concretely them, we could make
a variety of multi-replacement polices.
Unfortunately, we cannot analyze theoretically optimum policies which minimize
the expected cost rates for different replacement costs. However, it would be possible
to compute numerically optimum policies by estimating statistically replacement
costs and fitting distributions. In addition, the proposed replacement models with
three variables would show new theoretical research subjects to students, researchers

86 4 Replacement Overtime with Three Variables
and engineers. Furthermore, the proposed replacement policies could be applied into
more complex systems by suitable modifications and extensions. Throughout this
chapter, we use the same notations as those in Chaps. 1 and 2.
4.1 Replacement with Three Variables
Suppose that a unit is replaced at a planned time T , at a number N of working

cycles, and at a number K of failures, and we count the number of failures when the
unit is replaced at failure K for the simplicity of discussions. Then, we consider the
following replacement policies which combine three replacements with T , N and
K , and discuss optimum policies with the same replacement costs.
(1) Replacement First
A unit is replaced at time T (0 < T ≤ ∞), at cycle N (N = 1, 2, . . .), or at failure
K (K = 1, 2, . . .), whichever occurs first, where t N and t K are denoted by the
respective occurrence times at a number of N cycles and at a number of K failures
introduced in Chap. 2, i.e., it is replaced at min{T, t N , t K }. Then, the probability that
the unit is replaced at time T is
[1 − G (N ) (T )]P K (T ), (4.1)

T
P K (t)dG (N ) (t), (4.2)
0
the probability that it is replaced at failure K is

T
[1 − G (N ) (t)]d PK (t), (4.3)
0
where (4.1) + (4.2) + (4.3) = 1. The mean time to replacement is

T
(N )
T [1 − G (T )]P K (T ) + t P K (t)dG (N ) (t)
0
T T
+ t[1 − G (N ) (t)]d PK (t) = [1 − G (N ) (t)]P K (t)dt. (4.4)
0 0

K −1
K −1 T
(N )
[1 − G (T )] j p j (T ) + j p j (t)dG (N ) (t)
j=0 j=0 0
4.1 Replacement with Three Variables 87
T
+K [1 − G (N ) (t)]d PK (t)
0

K T
=K− (K − j) [1 − G (N ) (t)]d p j (t)
j=0 0
T
= [1 − G (N ) (t)]P K (t)h(t)dt, (4.5)
0
where agrees with (4.4) when H (t) = t, i.e., h(t) = 1.

c K − (c K − cT )[1 − G (N ) (T )]P K (T )
T
−(c K − c N ) 0 P K (t)dG (N ) (t)
T
+c M 0 [1 − G (N ) (t)]P K (t)h(t)dt
C F (T, N , K ) = T , (4.6)
(N ) (t)]P (t)dt
0 [1 − G K
where cT = replacement cost for time T , c N = replacement cost for cycle N ,

c K = replacement cost for failure K , and c M = minimal repair cost for each failure.
Clearly, C F (T, N , ∞) = C F (T, N ) in (2.1).
(2) Replacement Last
A unit is replaced at time T (0 ≤ T < ∞), at cycle N (N = 0, 1, 2, . . .), or at failure
K (K = 0, 1, 2, . . .), whichever occurs last, i.e., it is replaced at max{T, t N , t K }.
Then, the probability that the unit is replaced at time T is
G (N ) (T )PK (T ), (4.7)

∞
PK (t)dG (N ) (t), (4.8)
T
and the probability that it is replaced at failure K is

∞
G (N ) (t)d PK (t), (4.9)
T
∞ ∞
T G (N ) (T )PK (T ) + t PK (t)dG (N ) (t) + t G (N ) (t)d PK (t)
T T
∞
=T+ [1 − G (N ) (t)PK (t)]dt. (4.10)
T

∞
∞
∞ ∞
G (N ) (T ) j p j (T ) + j p j (t)dG (N ) (t) + K G (N ) (t)d PK (t)
j=K j=K T T
∞
∞
=K+ (j − K) G (N ) (t)d p j (t)
j=K T
∞
= H (T ) + [1 − G (N ) (t)PK (t)]h(t)dt, (4.11)
T

c K − (c K − cT )G (N ) (T )PK (T )
∞
−(c K − c N ) T PK (t)dG (N ) (t)
∞
+c M {H (T ) + T [1 − G (N ) (t)PK (t)]h(t)dt}
C L (T, N , K ) = ∞ . (4.12)
T + T [1 − G (N ) (t)PK (t)]dt
Clearly, C L (T, N , 0) = C L (T, N ) in (2.14), and
C L (T, 0, 0) = C F (T, ∞, ∞) = C P (T ) in (2.2),

C L (0, N , 0) = C F (∞, N , ∞) = C R (N ) in (2.6),
C L (0, 0, K ) = C F (∞, ∞, K ) = C(K ) in (2.41).
(3) Replacement Middle

A unit is replaced at time T (0 ≤ T ≤ ∞), at cycle N (N = 0, 1, 2, . . .), or at
failure K (K = 0, 1, 2, . . .), whichever occurs middle, i.e., it is replaced at time T
in case of {t N < T < t K } or {t K < T < t N }, at cycle N in case of {T < t N < t K }
or {t K < t N < T }, and at failure K in case of {T < t K < t N } or {t N < t K < T }.
Then, the probability that the unit is replaced at time T is: In case of t N < T < t K ,
G (N ) (T )P K (T ), (4.13)
and in case of t K < T < t N ,
[1 − G (N ) (T )]PK (T ). (4.14)
The probability that it is replaced at cycle N is: In case of T < t N < t K ,

∞
P K (t)dG (N ) (t), (4.15)
T
and in case of t K < t N < T ,

T
PK (t)dG (N ) (t). (4.16)
0
The probability that it is replaced at failure K is: In case of {T < t K < t N },

∞
[1 − G (N ) (t)]d PK (t), (4.17)
T
and in case of t N < t K < T ,

T
G (N ) (t)d PK (t), (4.18)
0
where note that (4.13) + (4.14) + (4.15) + (4.16) + (4.17) + (4.18) =1. The mean time
to replacement is
∞
T [1 − G (N ) (T )]PK (T ) + G (N ) (T )P K (T ) + t P K (t)dG (N ) (t)
T
T ∞ T
(N ) (N )
+ t PK (t)dG (t) + t[1 − G (t)]d PK (t) + t G (N ) (t)d PK (t)
0 T 0
T ∞
(N )
= 1 − G (t)PK (t) dt + [1 − G (N ) (t)]P K (t)dt. (4.19)
0 T

∞

K −1
(N ) (N )
[1 − G (T )] j p j (T ) + G (T ) j p j (T )
j=K j=0

K −1 ∞ ∞
T
+ j p j (t)dG (N ) (t) + j p j (t)dG (N ) (t)
j=0 T j=K 0
∞ T
(N ) (N )
+K [1 − G (t)]d PK (t) + G (t)d PK (t)
T 0

K ∞
= H (T ) − (K − j) [1 − G (N ) (t)]d p j (t)
j=0 T
∞
T
− (j − K) G (N ) (t)d p j (t)
j=K 0
T ∞
= [1 − G (N ) (t)PK (t)]h(t)dt + [1 − G (N ) (t)]P K (t)h(t)dt, (4.20)
0 T

C M (T, N , K ) =
c K − (c K − cT ){[1 − G (N ) (T )]PK (T ) + G (N ) (T )P K (T )}
∞ T
−(c K − c N ){ T P K (t)dG (N ) (t) + 0 PK (t)dG (N ) (t)}
T ∞
c M { 0 [1 − G (N ) (t)PK (t)]h(t)dt + T [1 − G (N ) (t)]P K (t)h(t)dt}
T . (4.21)
(N ) (t)P (t)]dt + ∞ [1 − G (N ) (t)]P (t)dt
0 [1 − G K T K
Clearly,
C M (0, N , K ) = C F (∞, N , K ), C M (T, 0, K ) = C F (T, ∞, K ),

C M (T, N , 0) = C F (T, N , ∞), C M (∞, N , K ) = C L (0, N , K ),
C M (T, ∞, K ) = C L (T, 0, K ), C M (T, N , ∞) = C L (T, N , 0).
(4) Optimum Policies

We discuss optimum policies which minimize the expected cost rates when cT =
c N = c K and h(t) increases strictly with t from h(0) = 0 to h(∞) = ∞. First, we find
optimum TF∗ which minimizes C F (T, N , K ) in (4.6). Differentiating C F (T, N , K )
T cT
[1 − G (N ) (t)]P K (t)[h(T ) − h(t)]dt = , (4.22)
0 cM
and unique TF∗ (0 < TF∗ < ∞) which satisfies (4.22), and the resulting cost rate is
C F (TF∗ , N , K ) = c M h(TF∗ ). (4.23)
Furthermore, note that TF∗ decreases with N and K , because the left-hand side
of (4.22) increases with N and K . Therefore, from (4.23), optimum policy which
minimizes C F (T, N , K ) is TF∗ = T ∗ given in (2.3) and N = K = ∞, i.e., the unit
should be replaced only at time T ∗ .
Next, we find optimum TL∗ which minimizes C L (T, N , K ) in (4.12). Differenti-
ating C L (T, N , K ) with respect to T and setting it equal to zero,
T ∞ cT
[h(T ) − h(t)]dt − [1 − G (N ) (t)PK (t)][h(t) − h(T )]dt = , (4.24)
0 T cM
∞
whose left-hand side increases strictly with T from − 0 [1 − G (N ) (t)PK (t)]h(t)dt
to ∞. Thus, there exists a finite and unique TL∗ (0 < TL∗ < ∞) which satisfies (4.24),
C L (TL∗ , N , K ) = c M h(TL∗ ). (4.25)

Furthermore, note that TL∗ increases with N and K , because the left-hand side of
(4.24) decreases with N and K . Therefore, from (4.25), optimum policy which
minimizes C L (T, N , K ) is TL∗ = T ∗ given in (2.3) and N = K = 0, i.e., the
unit should be replaced only at T ∗ .
Finally, we find optimum TM∗ which minimizes C M (T, N , K ) in (4.21). Differ-
entiating C M (T, N , K ) with respect to T and setting it equal to zero,
T
[1 − G (N ) (t)PK (t)][h(T ) − h(t)]dt
0
∞ cT
− [1 − G (N ) (t)]P K (t)[h(t) − h(T )]dt = , (4.26)
T cM
∞
whose left-hand side increases strictly with T from − 0 [1− G (N ) (t)]P K (t) h(t)dt
to ∞. Thus, there exists a finite and unique TM∗ (0 < TM∗ < ∞) which satisfies (4.26),
C M (TM∗ , N , K ) = c M h(TM∗ ). (4.27)
Furthermore, noting that from (2.3) and (4.26),

T T
[h(T ) − h(t)]dt − [1 − G (N ) (t)PK (t)][h(T − h(t))]dt
0 0
∞
+ [1 − G (N ) (t)]P K (t)[h(t) − h(T )]dt
T
T
= G (N ) (t)PK (t)[h(T ) − h(t)]dt
0
∞
+ [1 − G (N ) (t)]P K (t)[h(t) − h(T )]dt ≥ 0,
T
whose inequality becomes equal to 0 as {N = 0, K = ∞} or {N =∞, K =0}.

Therefore, optimum TM∗ takes a small value T ∗ when {N = 0, K = ∞} or
{N = ∞, K = 0}, i.e., the unit should be replaced only at T ∗ given in (2.3).
The above results show that when all replacement costs are the same, standard
periodic replacement is better than replacement first, last and middle. However, if
replacement costs for cycle N and failure K would be lower than that for time T ,
replacement first, last and middle might be better than the standard replacement.
4.2 Replacement Overtime with Working Cycle
Suppose that a unit is replaced at the first completion of working cycles over time
T to be denoted as t O (t O > T ), at a number N of working cycles, and at a number
K of failures, and the unit is replaced over time T when the N th working cycle has
just finished at over time T . We obtain the expected cost rates and derive optimum
policies which minimize them when all replacement costs are the same.
(1) Replacement Overtime First
A unit is replaced over time T (0 < T ≤ ∞), at cycle N (N = 1, 2, . . .), or at failure
K (K = 1, 2, . . .), whichever occurs first, i.e., it is replaced at min{t O , t N , t K }. Then,
the probability that the unit is replaced over time T , at cycle N , and at failure K are
given in (4.1)–(4.3), respectively. The mean time to replacement is
T T
t P K (t)dG (N ) (t) + t[1 − G (N ) (t)]d PK (t)
0 0

N −1 T ∞
+ P K (T ) (t + u)dG(u) dG ( j) (t)
j=0 0 T −t
T
N −1 T ∞
(N )
= 1−G (t)]P K (t)[dt + P K (T ) G(u − t)du dG ( j) (t).
0 j=0 0 T
(4.28)

K −1 T
K −1
j p j (t)dG (N ) (t) + j p j (T )[1 − G (N ) (T )]
j=0 0 j=0
T
+K [1 − G (N ) (t)]d PK (t)
0

N −1 T ∞
+ P K (T ) [H (t + u) − H (T )]dG(u) dG ( j) (t)
j=0 0 T −t

K T
=K− (K − j) [1 − G (N ) (t)]d p j (t)
j=0 0
T ∞
+ P K (T ) [H (t + u) − H (T )]dG(u) dG (i) (t)

0 T −t
T
= [1 − G (N ) (t)P K (t)]h(t)dt
0

N −1 T ∞
+ P K (T ) G(u − t)h(u)du dG ( j) (t), (4.29)
j=0 0 T

4.2 Replacement Overtime with Working Cycle 93
COF (T, N , K ) =
T T
c O + (c N − c O ) 0 P K (t)dG (N ) (t) + (c K − c O ) 0 [1 − G (N ) (t)]d PK (t)
T
+c M { 0 [1 − G (N ) (t)P K (t)]h(t)dt
−1 T ∞ ( j)
+P K (T ) Nj=0 0 [ T G(u − t)h(u)du]dG (t)}
T
N −1 T ∞
,
(N ) (t)]P (t)dt + P (T ) ( j)
0 [1 − G K K j=0 0 [ T G(u − t)du]dG (t)
(4.30)
where c O = replacement cost for overtime T , c N , c K and c M are given in (4.6).

Clearly, COF (T, N , ∞) = COF (T, N ) in (2.25).
(2) Replacement Overtime Last
A unit is replaced over time T (0 ≤ T < ∞), at cycle N (N = 0, 1, 2, . . .), or at fail-
ure K (K = 0, 1, 2, . . .), whichever occurs last, i.e., it is replaced at max{t O , t N , t K }.
Then, the probability that the unit is replaced over time T , at cycle N , and at failure
K are given in (4.7)–(4.9), respectively. The mean time to replacement is
∞ ∞
(N )
t PK (t)dG (t) + t G (N ) (t)d PK (t)
T T
∞
T ∞
+ PK (T ) (t + u)dG(u) dG ( j) (t)
j=N 0 T −t
∞ ∞
T ∞
=T+ 1 − G (N ) (t)PK (t) dt + PK (T ) G(u − t)du dG ( j) (t).
T j=N 0 T
(4.31)
The expected number of failures is

∞
∞
∞ ∞
(N ) (N )
G (T ) j p j (T ) + j p j (t)dG (t) + K G (N ) (t)d PK (t)
j=K j=K T T
∞ T ∞
+ PK (T ) [H (t + u) − H (T )]dG(u) dG ( j) (t)
j=N 0 T −t
∞
= H (T ) + [1 − G (N ) (t)PK (t)]h(t)dt
T
∞
T ∞
+ PK (T ) G(u − t)h(u)du dG ( j) (t), (4.32)
j=N 0 T

COL (T, N , K ) =
∞
c K − (c K − c O)G (N ) (T )PK (T ) − (c K − c N ) T PK (t)dG (N ) (t)
∞
+ c M {H (T ) + T [1 − G (N ) (t)PK (t)]h(t)dt
∞ T ∞
+ PK (T ) j=N 0 [ T G(u − t)h(u)du]dG ( j) (t)}
∞ . (4.33)
T + T [1 − G (N ) (t)PK (t)]dt
∞ T ∞
+PK (T ) j=N 0 [ T G(u − t)du]dG ( j) (t)
Clearly, COL (T, N , 0) = COL (T, N ) in (2.31).

(3) Optimum Policies
We derive optimum policies which minimize the expected cost rates of replace-
ment overtime when c O = c N = c K , G(t) = 1 − e−θt (0 < θ < ∞) and h(t)
increases strictly with t from h(0) = 0 to h(∞) = ∞. Then, the expected cost rate
COF (T, N , K ) in (4.30) is rewritten as
T
c O + c M { 0 [1 − G (N ) (t)]P
∞ K (t)h(t)dt
+ [1 − G (N ) (T )]P K (T ) 0 e−θt h(t + T )dt}
COF (T, N , K ) = T ,
0 [1 − G (N ) (t)]P K (t)dt + [1 − G (N ) (T )](1/θ )P K (T )
(4.34)
∞
where G (N ) (t) = j=N [(θ t) /j!]e
j −θt (N = 0, 1, 2, . . .). Differentiating C
OF
(T, N , K ) with respect to T and setting it equal to zero,
∞ T
−θt
θe h(t + T )dt [1 − G (N ) (t)]P K (t)dt
0 0
T cO
− [1 − G (N ) (t)]P K (t)h(t)dt = , (4.35)
0 cM
whose left-hand side increases strictly with T from 0 to ∞. Thus, there exits a finite
and unique TOF OF
∞
∗
COF (TOF , N , K ) = cM θ e−θt h(t + TOF
∗
)dt. (4.36)
0
Furthermore, note that TOF∗ decreases with N and K , because the left-hand side
of (4.35) increases with N and K . Therefore, from (4.36), optimum policy which
∗ = T ∗ given in (2.21) and N = K = ∞, i.e., the
minimizes COF (T, N , K ) is TOF O
unit should be replaced only over time TO∗ .
4.2 Replacement Overtime with Working Cycle 95
Next, the expected cost rate COL (T, N , K ) in (4.33) is rewritten as

∞
c O + c M {H (T ) + T [1 − G (N ) (t)PK (t)]h(t)dt
∞
+PK (T )G (N ) (T ) 0 e−θt h(t + T )dt}
COL (T, N , K ) = ∞ . (4.37)
T + 0 [1 − G (N ) (t)PK (t)]dt + [G (N ) (T )/θ ]PK (T )
Differentiating COL (T, N , K ) with respect to T and setting it equal to zero,

∞ ∞
θ e−θt h(t + T )dt T + [1 − G (N ) (t)PK (t)]dt

0 T
∞
cO
− H (T ) + [1 − G (N ) (t)PK (t)]h(t)dt = , (4.38)
T cM
whose left-hand side increases strictly with T from

∞ ∞ ∞
−θt (N )
θe h(t)dt [1 − G (t)PK (t)]dt − [1 − G (N ) (t)PK (t)]h(t)dt
0 0 0
∗ (0 ≤ T ∗ < ∞) which satisfies

to ∞. Thus, there exists a finite and unique TOL OL
∞
∗
COL (TOL , N , K ) = cM θ e−θt h(t + TOL
∗
)dt. (4.39)
0
Furthermore, note that TOL ∗ increases with N and K , because the left-hand side
of (4.38) decreases with N and K . Therefore, from (4.39), optimum policy which
minimizes COL (T, N , K ) is TOL ∗ = T ∗ given in (2.21) and N = K = 0, i.e., the unit
O
should be replaced only over time TO∗ .
The above results show that when all replacement costs are the same, replacement
overtime in Sect. 2.2 is better than replacement overtime first and last.
4.3 Replacement Overtime with Failure Number
Suppose that a unit is replaced at the first failure over time T , at a number N of
working cycles and at a number K of failures, and the unit is replaced at failure K
when the K th failure has just occurred over time T .
(1) Replacement Overtime First
A unit is replaced over time T (0 ≤ T < ∞), at cycle N (N = 1, 2, . . .) or at failure
K (K = 1, 2, . . .), whichever occurs first. The probabilities that the unit is replaced
over time T , at cycle N and at failure K are given in (4.1)–(4.3). The mean time to
replacement is
T T
(N )
t P K (t)dG (t) + t[1 − G (N ) (t)]d PK (t)
0 0
T ∞
(N ) 1
+ T [1 − G (T )] udF(u) d P K (t)
0 F(t) T
T
= [1 − G (N ) (t)]P K (t)dt
0
∞
+ [1 − G (N ) (T )]P K (T ) e−[H (t)−H (T )] dt. (4.40)
T
The expected number of failures before replacement is, from (4.5),

T
[1 − G (N ) (t)]P K (t)h(t)dt + [1 − G (N ) (T )]P K (T ), (4.41)
0

c K − (c K − c O − c M )[1 − G (N ) (T )]P K (T )
T
− (c K − c N ) 0 P K (t)dG (N ) (t)
T
+ c M 0 [1 − G (N ) (t)]P K (t)h(t)dt

COF (T, N , K ) = T . (4.42)
(N ) (t)]P (t)dt
0 [1 − G K ∞
+ [1 − G (N ) (T )]P K (T ) T e−[H (t)−H (T )] dt
OF (T, ∞, K ) = C F (T, K ) in (2.40).

Clearly, C
(2) Replacement Overtime Last
A unit is replaced over time T (0 ≤ T < ∞), at cycle N (N = 0, 1, 2, . . .) or at
failure K (K = 0, 1, 2, . . .), whichever occurs last. The probabilities that the unit is
replaced over time T , at cycle N and at failure K are given in (4.7)–(4.9). The mean
time to replacement is
∞ ∞
t PK (t)dG (N ) (t) + t G (N ) (t)d PK (t)
T T
T ∞
(N ) 1
+ G (T ) udF(u) d PK (t)
0 F(t) T
∞
=T+ 1 − G (N ) (t)PK (t) dt
T
∞
(N )
+ G (T )PK (T ) e−[H (t)−H (T )] dt. (4.43)
T
4.3 Replacement Overtime with Failure Number 97
The expected number of failures is, from (4.11),

∞
H (T ) + [1 − G (N ) (t)PK (t)]h(t)dt + G (N ) (T )PK (T ), (4.44)
T

c K − (c K − cO − c M )G (N ) (T )PK (T )
∞
−(c K − c N ) T PK (t)dG (N ) (t)
∞
c {H (T ) + T [1 − G (N ) (t)PK (t)]h(t)dt}
OL (T, N , K ) = M
C ∞ . (4.45)
T + T [1 − G (N )(t)PK (t)]dt
∞
+ G (N ) (T )PK (T ) T e−[H (t)−H (T )] dt
OL (T, 0, K ) = C L (T, K ) in (2.54).

Clearly, C
4.4 Modified Replacement with Three Variables
As modified replacement policies with three variables, we propose the following two
policies in which the unit is replaced at max{t N , t K } before time T or at min{t N , t K }
after time T .
(1) Replacement with Time T
The unit is replaced at max{t N , t K } before time T (0 ≤ T ≤ ∞) or is replaced at
min{t N , t K } after time T . Furthermore, if either the N th working cycle is completed
or the K th failure occurs before time T , the unit is replaced at time T . This policy
corresponds to replacement middle in (3) of Sect. 4.1.
(2) Replacement Over Time T
In the previous modified (1), if either the N th working cycle is completed or the K th
failure occurs before time T , the unit is replaced at the first completion of working
cycles over time T . The probabilities that the unit is replaced at time T , at cycle N ,
and at failure K are given in (4.13)–(4.18). The mean time to replacement is
T ∞
(N )
t PK (t)dG (t) + t P K (t)dG (N ) (t)
0 T
T ∞
(N )
+ tG (t)d PK (t) + t[1 − G (N ) (t)]d PK (t)
0 T
∞
T ∞
+ P K (T ) (t + u)dG(u) dG ( j) (t)
j=N 0 T −t

N −1 T ∞
+ PK (T ) (t + u)dG(u) dG ( j) (t)
j=0 0 T −t
T ∞
= 1 − G (N ) (t)PK (t) dt + [1 − G (N ) (t)]P K (t)dt
0 T
∞
T ∞
+ P K (T ) G(u − t)du dG ( j) (t)
j=N 0 T

N −1 T ∞
+ PK (T ) G(u − t)du dG ( j) (t). (4.46)
j=0 0 T
The expected number of failures before replacement is
∞
T
K −1 ∞
j p j (t)dG (N ) (t) + j p j (t)dG (N ) (t)
j=K 0 j=0 T
T ∞
(N ) (N )
+K G (t)d PK (t) + [1 − G (t)]d PK (t)
0 T

K −1 ∞

+ G (N ) (T ) j p j (T ) + [1 − G (N ) (T )] j p j (T )
j=0 j=K
∞ T
∞
+ P K (T ) [H (t + u) − H (T )]dG(u) dG ( j) (t)
j=N 0 T −t

N −1 T ∞
+ PK (T ) [H (t + u) − H (T )]dG(u) dG ( j) (t)
j=0 0 T −t
∞ T
= [1 − G (N ) (t)]P K (t)h(t)dt + [1 − G (N ) (t)PK (t)]h(t)dt
T 0
∞
T ∞
+ P K (T ) G(u − t)h(u)du dG ( j) (t)
j=N 0 T

N −1 T ∞
+ PK (T ) G(u − t)h(u)du dG ( j) (t), (4.47)
j=0 0 T

4.4 Modified Replacement with Three Variables 99
COM (T, N , K ) =
c K − (c K − c O ){G (N ) (T )P K (T ) + [1 − G (N ) (T )]PK (T )}
T ∞
−(c K − c N )[ 0 PK (t)dG (N ) (t) + T P K (t)dG (N ) (t)]
∞ T
+ c M { T [1 − G (N ) (t)]P K (t)h(t)dt + 0 [1 − G (N ) (t)PK (t)]h(t)dt
T ∞
+ P K (T ) ∞ 0 [ T G(u − t)h(u)du]dG (t)
( j)
Nj=N
−1 T ∞ (
+ PK (T ) j=0 0 [ T G(u − t)h(u)du]dG j) (t)}
T .
[1 − G (N ) (t)P (t)]dt + ∞ [1 − G (N ) (t)]P (t)dt
K K
0 T ∞ T
+ P K (T ) ∞ j=N 0 [ T G(u − t)du]dG (t)
( j)
N −1 T ∞
+ PK (T ) j=0 0 [ T G(u − t)du]dG ( j) (t)
(4.48)
Clearly, COM (0, N , K ) = C M (0, N , K ) in (4.21) and COM (T, N , ∞) = COL (T, N )
in (2.31).
(3) Replacement Over Time T with Failure Number
In the previous model (2), if either the N th working cycle is completed or the K th
failure occurs before time T , the unit is replaced at the first failure over time T . The
probabilities that the unit is replaced over time T , at cycle N , and at failure K are
given in (4.13)–(4.18). The mean time to replacement is
T ∞
t PK (t)dG (N ) (t) + t P K (t)dG (N ) (t)
0 T
T ∞
(N )
+ tG (t)d PK (t) + t[1 − G (N ) (t)]d PK (t)
0 T
T ∞
(N ) 1
+ G (T ) udF(u) d P K (t)
0 F(t) T
T ∞
(N ) 1
+ [1 − G (T )] udF(u) d PK (t)
0 F(t) T
T ∞
= 1 − G (N ) (t)PK (t) dt + [1 − G (N ) (t)]P K (t)dt
0 T
∞
(N ) −[H (t)−H (T )]
+ G (T )P K (T ) e dt
T
∞
+ [1 − G (N ) (T )]PK (T ) e−[H (t)−H (T )] dt. (4.49)
T
∞
T
K −1 ∞
(N )
j p j (t)dG (t) + j p j (t)dG (N ) (t)
j=K 0 j=0 T
T ∞
+K G (N ) (t)d PK (t) + [1 − G (N ) (t)]d PK (t)

0 T

K −1 ∞

+ G (N ) (T ) j p j (T ) + [1 − G (N ) (T )] j p j (T )
j=0 j=K
T
= [1 − G (N ) (t)PK (t)]h(t)dt + G (N ) (T )P K (T )
0
∞
+ [1 − G (N ) (t)]P K (t)h(t)dt + [1 − G (N ) (T )]PK (T ), (4.50)
T

COM (T, N , K ) =
c K − (c K − c O − c M ){G (N ) (T )P K (T ) + [1 − G (N ) (T )]PK (T )}
T ∞
− (c K − c N )[ 0 PK (t)dG (N ) (t) + T P K (t)dG (N ) (t)]
T ∞
c M { 0 [1 − G (N ) (t)PK (t)]h(t)dt + T [1 − G (N ) (t)]P K (t)h(t)dt}
T .
(N ) (t)P (t)]dt + ∞ [1 − G (N ) (t)]P (t)dt
0 [1 − G K T K
∞
+ G (N ) (T )P K (T ) T e−[H

(t)−H (T )] dt
∞
+ [1 − G (N ) (T )]PK (T ) T e−[H (t)−H (T )] dt
(4.51)
Clearly, COM (∞, N , K ) = C M (∞, N , K ) in (4.21) and COM (T, ∞, K ) = COL

(T, K ) in (2.54).
4.5 General Replacement Policies
We extend replacement policies with three variables to n (n = 1, 2, . . .) variables:

The unit is replaced before failure at time T (0 < T ≤ ∞) or at random times
Y1 , Y2 , . . . , Yn , each of which has a general distribution G i (t) ≡ Pr{Yi < t} with
finite mean 1/θi (i = 1, 2, . . . , n). Let ci (i = 1, 2, . . . , n) be the replacement cost
at random times Yi with ci < c F .
(1) Replacement First
Suppose that the unit is replaced at time T or at Y1 , Y2 , . . . , Yn , whichever occurs
first. Then, the probability that the unit is replaced at time T is
4.5 General Replacement Policies 101
n

G i (T ) F(T ), (4.52)
i=1
the probability that it is replaced at time Yi (i = 1, 2, . . . , n) is

⎡ ⎤
T
n
F(t) ⎣ G j (t)⎦ dG i (t) (i = 1, 2, . . . , n), (4.53)
0 j=1, j=i

n
T
G i (t) dF(t), (4.54)
0 i=1
where note that

⎡ ⎤

n n
T
n
G i (T ) F(T ) + F(t) ⎣ G j (t)⎦ dG i (t)
i=1 i=1 0 j=1, j=i
n
T
+ G i (t) dF(t) = 1.
0 i=1

n ⎡ ⎤
n
T
n
T G i (T ) F(T ) + t F(t) ⎣ G j (t)⎦ dG i (t)
i=1 i=1 0 j=1, j=i
n
T T
n
+ t G i (t) dF(t) = F(t) G i (t) dt. (4.55)
0 i=1 0 i=1

n T n
cT [ i=1 G i (T )]F(T ) + c F 0 [ i=1 G i (t)]dF(t)
n T
+ i=1 ci 0 F(t)[ nj=1, j=i G j (t)]dG i (t)
C F (T ; G 1 , . . . , G n ) = T n .
0 F(t)[ i=1 G i (t)]dt
(4.56)
When the failure rate h(t) increases strictly with t from h(0) = 0 to h(∞) = ∞,
and cT = ci < c F , we find optimum TF,n ∗ which minimizes C (T ; G , . . . , G ).
F 1 n
Differentiating C F (T ; G 1 , . . . , G n ) with respect to T and setting it equal to zero,
n n
T T cT
h(T ) F(t) G i (t) dt − G i (t) dF(t) = , (4.57)
0 0 c F − cT
i=1 i=1
and unique TF,n F,n
∗ ∗
C F (TF,n ; G 1 , . . . , G n ) = (c F − cT )h(TF,n ). (4.58)
∗ increases with n to ∞.
Noting that the left-hand side of (4.57) decreases with n, TF,n
(2) Replacement Last
Suppose that the unit is replaced at time T or at Y1 , Y2 , . . . , Yn , whichever occurs
last. Then, the probability that the unit is replaced at time T is
n

G i (T ) F(T ), (4.59)
i=1
the probability that it is replaced at time Yi (i = 1, 2, . . . , n) is

⎡ ⎤
∞
n
F(t) ⎣ G j (t)⎦ dG i (t) (i = 1, 2, . . . , n), (4.60)
T j=1, j=i

∞
n
F(T ) + 1− G i (t) dF(t), (4.61)
T i=1
where note that

n ⎡ ⎤
n
∞
n
G i (T ) F(T ) + F(t) ⎣ G j (t)⎦ dG i (t)
i=1 i=1 T j=1, j=i

∞
n
+ F(T ) + 1− G i (t) dF(t) = 1.
T i=1

⎡ ⎤

n n
∞
n
T G i (T ) F(T ) + t F(t) ⎣ G j (t)⎦ dG i (t)
i=1 i=1 T j=1, j=i
4.5 General Replacement Policies 103

T ∞
n
+ tdF(t) + t 1− G i (t) dF(t)
0 T i=1

T ∞
n
= F(t)dt + F(t) 1 − G i (t) dt. (4.62)
0 T i=1
C L (T ; G 1 , . . . , G n ) =
n ∞ n
cT [ i=1 Gi (T )]F(T ) + c F {F(T ) + T [1 − i=1 G i (t)]dF(t)}
n ∞ n
+ i=1 ci T F(t)[ j=1, j=i G j (t)]dG i (t)
T ∞ n . (4.63)
0 F(t)dt + T F(t)[1 − i=1 G i (t)]dt
When the failure rate h(t) increases strictly with t from 0 to ∞, and cT = ci <
c F , we find optimum TL∗,n which minimizes C L (T ; G 1 , . . . , G n ). Differentiating
C L (T ; G 1 , . . . , G n ) with respect to T and setting it equal to zero,

T ∞
n
h(T ) F(t)dt + F(t) 1 − G i (t) dt
0 T i=1

∞
n
cT
− F(T ) − 1− G i (t) dF(t) = , (4.64)
T c F − cT
i=1
∞ n
whose left-hand side increases strictly with T from − 0 [1 − i=1 G i (t)]dF(t)
to ∞. Thus, there exists a finite and unique TL∗,n (0 < TL∗,n < ∞) which satisfies
C L (TL∗,n ; G 1 , . . . , G n ) = (c F − cT )h(TL∗,n ). (4.65)
Noting that the left-hand side of (4.64) increases with n, TL∗,n decreases with n.
References
1. Zhao X, Nakagawa T (2012) Optimization problems of replacement first or last in reliability

theory. Euro J Oper Res 223:141–149
2. Zhao X, Qian C, Nakamura S (2014) What is middle maintenance policy? In: Bae SJ et al.
(eds). The 6th Asia-Pacific international symposium on advanced reliability and maintenance
modeling, pp 610–617
Appendices
It is assumed that the failure rate h(t) ≡ f (t)/F(t) increases strictly with t from
h(0) = 0 to h(∞). Then, we have the following results [1, p.227]:
Appendix 1
1.1 For 0 < T < ∞ and N = 1, 2, . . .,
θ(θT ) N −1 /(N − 1)!

r N (T ) ≡ N −1
j=0 [(θT ) /j!]
j
increases strictly with T from 0 to θ for N ≥ 2, is θ for N = 1, and decreases

strictly with N from θ to 0. For 0 < T < ∞ and N = 0, 1, 2, . . .,
θ(θT ) N /N !
r N (T ) ≡ ∞

j=N +1 [(θT ) /j!]
j
decreases strictly with T from ∞ to 0 and increases strictly with N from

θ/(eθT −1) to ∞.
Proof When N = 1, r1 (T ) = θ, which is constant for any T > 0. When N ≥ 2,
lim r N (T ) = 0, lim r N (T ) = θ.
T →0 T →∞
Differentiating r N (T ) with respect to T ,

N −1
dr N (T ) θ2 [(θT ) j /j!](N − 1 − j) (θT ) N −2
j=0
= −1 > 0,
dT { Nj=0 [(θT ) j /j!]}2 (N − 1)!
which follows that r N (T ) increases strictly with T from 0 to θ.

Lecture Notes in Production Engineering, DOI 10.1007/978-3-319-20813-8
106 Appendices
For 0 < T < ∞,
r1 (T ) = θ, lim r N (T ) = 0,
N →∞
and
N
θ j=0 [(θT ) /j!]( j − N )
j
(θT ) N −1
r N +1 (T ) − r N (T ) = N −1
< 0,
{ j=0 [(θT ) j /j!] Nj=0 [(θT ) j /j!]} N!
which follows that r N (T ) decreases strictly with N from θ to 0.

r N (T ).
Similarly, it can be proved for
1.2 For 0 ≤ T < ∞,
1 F(T )
Q(T ) ≡ ∞ = ∞
T e−[H (t)−H (T )] dt T F(t)dt
increases strictly with T from 1/μ to h(∞).

Proof Note that
1
Q(T ) > h(T ), lim Q(T ) = , and lim Q(T ) = h(∞).
T →0 μ T →∞
Differentiating Q(T ) with respect to T ,

dQ(T ) F(T ) F(T )
= ∞ ∞ − h(T ) > 0,
dT T F(t)dt T F(t)dt
which follows that Q(T ) increases strictly with T from 1/μ to h(∞).
1.3 For 0 < T ≤ ∞ and 0 ≤ θ < ∞,
T
e−θt dF(t)
Q 1 (T ) ≡ 0T
0 e−θt F(t)dt
increases strictly with T from h(0) = 0 to

∞
e−θt dF(t)
0∞ ,
0 e−θt F(t)dt
Appendices 107
and decreases with θ from

F(T )
T
0 F(t)dt
to 0.
Proof Note that
Q 1 (T ) < h(T ), lim Q 1 (T ) = lim h(T ) = h(0) = 0.

T →0 T →0
Differentiating Q 1 (T ) with respect to T ,

dQ 1 (T ) e−θT F(T ) T
= T e−θt F(t)[h(T ) − h(t)]dt > 0,
dT [ 0 e−θt F(t)dt]2 0
which follows that Q 1 (T ) increases strictly with T from h(0) = 0 to

∞ −θt
e dF(t)
0∞ .
−θt F(t)dt
0 e
Similarly, differentiating Q 1 (T ) with respect to θ,

T T
dQ 1 (T ) 1 −θt
= T te F(t)dt e−θt dF(t)
dθ [ 0 e−θt F(t)dt]2 0 0
T T
−θt −θt
− te dF(t) e F(t)dt < 0,
0 0
which follows that Q 1 (T ) decreases with θ. Furthermore,

T −θt T
0 e dF(t) f (t)d(1 − e−θt ) f (0)
lim T = lim 0T = = h(0) = 0.
−θt −θt
0 F(t)d(1 − e )
θ→∞ θ→∞ F(0)
0 e F(t)dt
1.4 For 0 ≤ T < ∞ and 0 ≤ θ < ∞,

∞ −θt
T e dF(t)
Q 1 (T ) ≡ ∞
−θt
T e F(t)dt
increases strictly with T from Q 1 (∞) to h(∞) and decreases with θ from Q(T )
to h(T ).
Proof Note that
1 (T ) > h(T ),
Q 1 (T ) = lim h(T ) = h(∞).
lim Q
T →∞ T →∞
108 Appendices
1 (T ) with respect to T ,
Differentiating Q
1 (T ) ∞
dQ e−θT F(T )
= ∞ e−θt F(t)[h(t) − h(T )]dt > 0,
dT [ T e−θt F(t)dt]2 T
1 (T ) increases strictly with T from

which follows that Q
∞
e−θt dF(t)
0∞ = Q 1 (∞)
0 e−θt F(t)dt
1 (T ) with respect to θ,
to h(∞). Similarly, differentiating Q
1 (T ) ∞ ∞
dQ 1
= ∞ te−θt F(t)dt e−θt dF(t)
dθ [ T e−θt F(t)dt]2 T T
∞ ∞
−θt −θt
− te dF(t) e F(t)dt < 0,
T T
1 (T ) decreases with θ from Q(T ). Furthermore,

∞ −θt dF(t)
∞
T e f (t + T )d(1 − e−θt )
lim ∞ = lim 0∞ = h(T ).
θ→∞ −θt F(t)dt θ→∞ F(t + T )d(1 − e−θt )
T e 0
1 (T ), we have Q 1 (T ) < Q
Comparing Q(T ), Q 1 (T ) and Q 1 (T ) < Q(T ) for
0 < T < ∞ and 0 < θ < ∞.
1.5 For 0 ≤ T < S ≤ ∞ and 0 ≤ θ < ∞,
S −θt
e dF(t)

Q 1 (T, S) ≡ TS
−θt F(t)dt
T e
increases strictly with T from Q 1 (S) to h(S) and increases strictly with S from
1 (T ).
h(T ) to Q
Proof Note that
1 (T, S) < h(S),

h(T ) < Q lim = h(S),
lim Q(T, S) = h(T ).
T →S S→T
1 (T, S) with respect to T ,

Differentiating Q

e−θT F(T ) S
S e−θt F(t)[h(t) − h(T )]dt > 0,
[ T e−θt F(t)dt]2 T
Appendices 109
which follows that Q 1 (T, S) increases strictly with T from Q 1 (S) to h(S).
1 (T, S) increases strictly with S from h(T ) to Q
Similarly, Q 1 (T ), and Q 1 (S) <
1 (T, S) < Q
Q 1 (T ) for 0 < T < S < ∞.
Appendix 2
2.1 For 0 < T ≤ ∞, 0 < θ < ∞, and N = 0, 1, 2, . . .,

T
Q 1 (T, N ) ≡ 0T
increases strictly with T from h(0) = 0 to Q 1 (N ) and increases strictly with N

from Q 1 (T ) to h(T ), where
∞ T
(θt) N e−θt dF(t) e−θt dF(t)
Q 1 (N ) ≡ 0∞ , Q 1 (T ) ≡ 0T .
N −θt F(t)dt
0 (θt) e 0 e−θt F(t)dt
Proof First, note that
lim Q 1 (T, N ) = 0, 0 < Q 1 (T, N ) < h(T ).

T →0
Next, differentiating Q 1 (T, N ) with respect to T ,

dQ 1 (T, N ) (θT ) N e−θT F(T ) T
= T (θt) N e−θt F(t)[h(T ) − h(t)]dt > 0,
dT [ 0 (θt) N e−θt F(t)dt]2 0
which follows that Q 1 (T, N ) increases strictly with T from 0 to Q 1 (N ).

Similarly, denoting
T T
N +1 −θt
L 1 (T ) ≡ (θt) e dF(t) (θt) N e−θt F(t)dt
0 0
T T
− (θt) N e−θt dF(t) (θt) N +1 e−θt F(t)dt,
0 0
we have L 1 (0) = 0, and

T
L 1 (T ) N −θT
= (θT ) e F(T ) (θt) N e−θt F(t)(θT − θt)[h(T ) − h(t)]dt > 0,
0
which follows that Q 1 (T, N ) increases strictly with N . Clearly, because
lim Q 1 (T, N ) = h(T ),

N →∞
110 Appendices
we have
lim Q 1 (N ) = h(∞),
N →∞
and Q 1 (T ) < Q 1 (T, N ) < Q 1 (N ) for 0 < T < ∞ and 1 ≤ N < ∞.

2.2 For 0 ≤ T < ∞, 0 < θ < ∞ and N = 0, 1, 2, . . ..
∞
1 (T, N ) ≡ T
Q ∞
T (θt) N e−θt F(t)dt
increases strictly with T from Q 1 (N ) to h(∞) and increases strictly with N

from Q1 (T ) to h(∞), where
∞ −θt
T e dF(t)
Q 1 (T ) ≡ ∞ .
−θt
T e F(t)dt
1 (T, N ) = h(∞), h(T ) < Q

lim Q 1 (T, N ) < h(∞).
T →∞
1 (T, N ) with respect to T ,

Next, differentiating Q
1 (T, N ) ∞
dQ (θT ) N e−θT F(T )
= ∞ (θt) N F(t)[h(t) − h(T )]dt > 0,
dT [ T (θt) N e−θt F(t)dt]2 T
which follows that Q1 (T, N ) increases strictly with T from Q 1 (N ) to h(∞).
Similarly, denoting
∞ ∞
L 2 (T ) ≡ (θt) N +1 e−θt dF(t) (θt) N e−θt F(t)dt
T
∞ T ∞
N −θt
− (θt) e dF(t) (θt) N +1 e−θt F(t)dt,
T T
we have L 2 (∞) = 0, and

∞
L 2 (T ) N −θT
= (θT ) e F(T ) (θt) N e−θt F(t)(θt − θT )[h(T ) − h(t)]dt < 0,
T
1 (T, N ) increases strictly with N . Clearly, because for

T < T1 < ∞,
T1
lim TT = h(T1 ),
N →∞ 1
T (θt) N e−θt F(t)dt
Appendices 111
we have lim N →∞ Q 1 (T, N ),

1 (T, N ) = h(∞). Comparing to Q 1 (T, N ) and Q
we have Q 1 (T, N ) < h(T ) < Q 1 (T ) < Q
1 (T, N ) for 0 < T < ∞ and
1 ≤ N < ∞.
2.3 For 0 ≤ T < ∞, 0 ≤ θ < ∞ and N = 0, 1, 2, . . .,
N ∞ −θt dF(t)
j=0 T [(θt) /j!]e
j
N ∞ −θt F(t)dt
j=0 T [(θt) /j!]e
j
increases strictly with T from

N ∞ −θt dF(t)
j=0 0 [(θt) /j!]e
j
N ∞ −θt F(t)dt
j=0 0 [(θt) /j]e
j
1 (T ) to Q(T ).
to h(∞) and increases strictly with N from Q
Proof Differentiating
N ∞ −θt dF(t)
j=0 T [(θt) /j!]e
j
N ∞ −θt F(t)dt
j=0 T [(θt) /j!]e
j
with respect to T ,
N −θT F(T ) N ∞
j=0 [(θT ) /j!]e
j
(θt) j −θt
N ∞ e [h(t) − h(T )] > 0.
{ −θt F(t)dt}2
j=0 T [(θt) /j!]e
j j!
j=0 T
Furthermore, when
∞
[(θt) j /j!]e−θt dF(t)
1 (T, j) = T
Q ∞
T [(θt) j /j!]e−θt F(t)dt
increases strictly with j, for N ≥ 1,

∞ N ∞ −θt dF(t)
−θt dF(t)
j=0 T [(θt) /j!]e
j
T [(θt) /N !]e
N
∞ > N ∞ .
−θt F(t)dt
T [(θt) /N !]e
−θt F(t)dt
j=0 T [(θt) /j!]e
N j
Thus, using the results of Appendix 2.2, we can prove Appendix 2.3.
2.4 For 0 < T ≤ ∞, 0 < θ < ∞ and N = 0, 1, 2, . . .,
T
0 (θt) [ t e
Q 2 (T, N ) ≡ T
0 (θt) [ t e
112 Appendices
1 (0) to Q 2 (∞, N ) and increases strictly with

increases strictly with T from Q
N from Q 2 (T, 0) to Q1 (T ).
1 (0), Q
lim Q 2 (T, N ) = Q 1 (∞) = Q 1 (0) < Q 2 (T, N ) < Q
1 (T ).
T →0

∞
(θT ) N T e−θt F(t)dt
T ∞
{ 0 (θt) N [ t e−θu F(u)du]dt}2
T ∞
× (θt) N
e −θu
F(u)[ Q 1 (u) − Q 1 (t)]du dt > 0,
0 t
1 (0) to Q 2 (∞, N ).
which follows that Q 2 (T, N ) increases strictly with T from Q
Similarly, denoting
T ∞ T ∞
L 3 (T ) ≡ (θt) N +1 e−θu dF(u) dt (θt) N e−θu F(u)du dt
0 t 0 t
T ∞ T ∞
− (θt) N e−θu dF(u) dt (θt) N +1 e−θu F(u)du dt,
0 t 0 t
we have L 3 (0) = 0, and

∞
L 3 (T ) =(θT ) N
e−θt F(t)dt
T
T ∞
× 1 (T ) − Q
e−θu F(u)du (θt) N (θT − θt)[ Q 1 (t)]dt > 0,
0 t
1 (T ).
which follows that Q 2 (T, N ) increases strictly with N from Q 2 (T, 0) to Q
2.5 For 0 ≤ T < ∞, 0 < θ < ∞ and N = 0, 1, 2, . . .,
∞
T (θt) [ t e
Q 2 (T, N ) ≡ ∞
T (θt) [ t e
2 (0, N ) to h(∞) and increases strictly with N

from Q2 (T, 0) to h(∞).
2 (T, N ) = lim Q
lim Q 1 (T ) < Q
1 (T ) = h(∞), Q 2 (T, N ) < h(∞).
T →∞ T →∞
Appendices 113
2 (T, N ) with respect to T ,

∞
(θT ) N T e−θt F(t)dt
∞ ∞
{ T (θt) N [ t e−θu F(u)du]dt}2
∞ ∞
× (θt) N
e −θu
F(u)du [ Q 1 (t) − Q 1 (T )]dt > 0,
T t
which follows that Q 2 (0, N ) to h(∞).

2 (T, N ) increases strictly with T from Q
Similarly, denoting
∞ ∞ ∞ ∞
N +1 −θu −θu
L 4 (T ) ≡ (θt) e dF(u) dt (θt) N
e F(u)du dt
T
∞ t ∞
T
∞
t
∞
−θu N +1 −θu
− (θt) N e dF(u) dt (θt) e F(u)du dt,
T t T t
we have L 4 (∞) = 0 and

∞
L 4 (T ) =(θT ) N e−θt F(t)dt
T
∞ ∞
× 1 (T ) − Q
e−θu F(u)du (θt) N (θt − θT )[ Q 1 (t)]dt < 0,
T t
2 (T, N ) increases strictly with N from Q

which follows that Q 2 (T, 0). Further-
more,
2 (T, N ) = lim Q
lim Q 1 (T ) = h(∞).
N →∞ T →∞
2 (T, N ), we have Q 2 (T, N ) < Q

Comparing to Q 2 (T, N ) and Q 1 (T ) <
2 (T, N ) for 0 < T < ∞ and 1 ≤ N < ∞.
Q
2.6 For 0 ≤ T < ∞ and 0 ≤ θ < ∞, F(t) = 1 − e−λt and n = 2, 3, . . .,
∞ −θt
e dF(t)n
n (T ) ≡ ∞ T
Q −θt [1 − F(t)n ]dt
T e
n (0) to λ.
Proof First, note that the failure rate of a parallel system with n (n ≥ 2) units
when F(t) = 1 − e−λt ,
nλe−λt (1 − e−λt )n−1 nλ

h n (t) ≡ = n−1 ,
1 − (1 − e−λt )n j=0 (1 − e−λt )− j
114 Appendices
n (T ) with respect
which increases strictly with t to λ. Next, differentiating Q
to T ,
∞
e−θT [1 − F(T )n ]
∞ e−θt [1 − F(t)n ][h n (t) − h n (T )]dt > 0,
{ T e−θt [1 − F(t)n ]dt}2 T
which follows that Q n (T ) increases strictly with T from Q

n (0) to h n (∞) = λ,

and Q n (T ) > h n (T ).
2.7 For 0 ≤ T < ∞, 0 ≤ θ < ∞, 0 < β < ∞ and Fβ (t) = exp(−βe−λt ),
∞ −θt
T e dFβ (t)

Q β (T ) ≡ ∞
−θt F β (t)dt
T e
β (0) to λ.
βλe−λt
h β (t) =
exp(βe−λt ) − 1
increases strictly with t to λ, because for x = e−λt ,

d ex − 1 ex − 1
> 0, lim = 1.
dx x x→0 x
β (T ) with respect to T ,

e−θT F β (T ) ∞
∞ e−θt F β (t)[h β (t) − h β (T )]dt > 0,
{ T e−θt F β (t)dt}2 T
which follows that Q β (T ) increases strictly with T from Q

β (0) to h β (∞) = λ,

and Q β (T ) > h β (T ).
Appendix 3
3.1 For 0 < T ≤ ∞, 0 < θ < ∞ and N = 0, 1, 2, . . .,

T
(θt) N e−θt h(t)dt
H1 (T, N ) = 0 T
N −θt dt
0 (θt) e
increases strictly with T from h(0) = 0 to H1 (N ) and increases strictly with N

from H1 (T ) to h(T ), where
Appendices 115
∞ θ(θt) N −θt
H1 (N ) ≡ H1 (∞, N ) = e h(t)dt,
0 N!
T
θe−θt h(t)dt
H1 (T ) ≡ H1 (T, 0) = 0
.
1 − e−θT
lim H1 (T, N ) = 0, H1 (T, N ) < h(T ).

T →0
Next, differentiating H1 (T, N ) with respect to T ,

dH1 (T, N ) (θT ) N e−θT T
= T (θt) N e−θt [h(T ) − h(t)]dt > 0,
dT [ 0 (θt) N e−θT dt]2 0
which follows that H1 (T, N ) increases strictly with T from 0 to H1 (N ).

Similarly, denoting
T T
N +1 −θt
L 5 (T ) ≡ (θt) e h(t)dt (θt) N e−θt dt
0 0
T T
N −θt
− (θt) e h(t)dt (θt) N +1 e−θt dt,
0 0
we have L 5 (0) = 0,
T
L 5 (T ) = (θT ) N e−θT (θt) N e−θt (θT − θt)[h(T ) − h(t)]dt > 0,
0
which follows that H1 (T, N ) increases strictly with N from H1 (T ) to h(T ), and
H1 (T ) < H1 (T, N ) < H1 (N ).
3.2 For 0 < T ≤ ∞ and K = 0, 1, 2, . . .,
T
0 H (t) dF(t)
K
Q 2 (T, K ) = T ∞
0 H (t) K { t e−[H (u)−H (t)] du}dF(t)
∞
increases strictly with T from 1/μ to 1/ 0 p K +1 (t)dt and increases strictly
with K from Q 2 (T, 0) to Q(T ).
1 1 1
lim Q 2 (T, K ) = ∞ = , < Q 2 (T, K ) < Q(T ).
T →0
0 F(t)dt μ μ
Next, differentiating Q 2 (T, K ) with respect to T ,

116 Appendices
T
H (T ) K f (T ) 1 1
H (t) K − dF(t) > 0,
( 0T H (t) K { t∞ e−[H (u)−H (t)] du}dF(t))2 0 Q(t) Q(T )
which follows that Q 2 (T, K ) increases strictly with T from 1/μ. Furthermore,
∞
H (t) K dF(t) 1
lim Q 2 (T, K ) = ∞ 0 ∞
−[H (u)−H (t)]
= ∞ .
T →∞
0 H (t) { t e
K du}dF(t) 0 p K +1 (t)dt
Similarly, denoting
T T 1
L 6 (T ) ≡ H (t) K +1 dF(t) H (t) K dF(t)
0 0 Q(t)
T T 1
− H (t) K dF(t) H (t) K +1 dF(t),
0 0 Q(t)
we have L 6 (0) = 0,
T
1 1
L 6 (T ) = H (T ) K f (T ) H (t) K [H (T ) − H (t)] − > 0,
0 Q(t) Q(T )
which follows that Q 2 (T, K ) increases strictly with K from Q 2 (T, 0) to Q(T ).
3.3 For 0 ≤ T < ∞ and K = 0, 1, 2, . . .,
∞
H (t) K dF(t)
2 (T, K ) ≡ ∞
Q T ∞
T H (t) K { t e−[H (u)−H (t)] du}dF(t)
∞
increases strictly with T from 1/ 0 p K +1 (t)dt to h(∞) and increases strictly
with K from Q 2 (T, 0) to h(∞).
2 (T, K ) = h(∞), Q(T ) < Q

lim Q 2 (T, K ) < h(∞).
T →∞
2 (T, K ) with respect to T ,

2 (T, K )
dQ H (T ) K f (T )
= ∞ ∞
dT ( T H (t) K { t e−[H (u)−H (t)] du}dF(t))2
∞
1 1
× H (t) K − dF(t) > 0,
T Q(T ) Q(t)
∞
2 (T, K ) increases strictly with T from 1/
which follows that Q p K +1 (t)dt
0
to lim T →∞ Q(T ) = h(∞). Similarly, denoting
Appendices 117
∞ ∞ 1
L 7 (T ) = H (t) K +1 dF(t) H (t) K dF(t)
T Q(t)
∞ T ∞
− H (t) K dF(t) H (t) K +1 dF(t),
T T
we have L 7 (∞) = 0,
∞
1 1
L 7 (T ) = H (T ) K f (T ) H (t) K [H (t) − H (T )] − < 0,
T Q(t) Q(T )
2 (T, K ) increases strictly with K from Q

which follows that Q 2 (T, 0) to h(∞),
and Q 2 (T, K ) < Q(T ) < Q2 (T, K ) for 0 < T < ∞ and 0 ≤ K < ∞.
3.4 For 0 < T ≤ ∞, 0 < θ < ∞ and N = 0, 1, 2, . . .,

T
Q 3 (T, N ) ≡ T 0
0 (θt) N e−θt [h(t)/Q(t)]dt
increases strictly with T from 1/μ to Q 3 (N ) and increases strictly with N from
Q 3 (T, 0) to Q(T ), where
∞
Q 3 (N ) ≡ Q 3 (∞, N ) = ∞ 0 N −θt .
0 θt e [h(t)/Q(t)]dt
1 1 1
lim Q 3 (T, N ) = lim Q(T ) = ∞ = , < Q 3 (T, N ) < Q(T ).
T →0 T →0
0 F(t)dt μ μ
dQ 3 (T, N ) (θT ) N e−θT h(T )

= T
dT { 0 (θt) N e−θt [h(t)/Q(t)]dt}2
T
1 1
× (θt) N e−θt − dt > 0,
0 Q(t) Q(T )
which follows that Q 3 (T, N ) increases strictly with T from 1/μ to Q 3 (N ).

Similarly, denoting
T T
L 8 (T ) = (θt) N +1 e−θt h(t)dt (θt) N e−θt [h(t)/Q(t)]dt
0 0
T T
− (θt) N e−θt h(t)dt (θt) N +1 e−θt [h(t)/Q(t)]dt,
0 0
118 Appendices
we have L 8 (0) = 0,
T
1 1
L 8 (T ) = (θT ) N e−θT h(T ) (θt) N e−θt h(t)(θT − θt) − dt > 0,
0 Q(t) Q(T )
which follows that Q 3 (T, N ) increases strictly with N from Q 3 (T, 0) to Q(T ).
Reference

2015 Book MaintenanceOvertimePoliciesInR PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2015 Book MaintenanceOvertimePoliciesInR PDF

Uploaded by

Copyright:

Available Formats

Lecture Notes in Production Engineering

ISSN 2194-0525 ISSN 2194-0533 (electronic)

Library of Congress Control Number: 2015943037

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

Maintenance theory is basically how to make appropriate maintenances at suitable

Toyota Toshio Nakagawa

1 Age Replacement Overtime. . . . . . . . ........... . . . . . . . . . . . 1

2 Periodic Replacement Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3 Inspection Overtime . . . . . . . . . . . . . . . . ... .... . . . . . . . . . . . . 59

3.3 Inspection Overtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4 Replacement Overtime with Three Variables . . . . . . . . . . . . . . . . . 85

1.1 Age and Random Replacements

Fig. 1.1 A job with working Y1 Y2 Y3 Yj

the probability that it is replaced at cycle N is

and the probability that it is replaced at failure is

Therefore, the expected cost rate is

where c F = replacement cost at failure, cT = replacement cost at time T , and c N =

An optimum policy which minimizes C A (T ) is [1, p. 87], [2, p. 74]:

and the resulting cost rate is

An optimum policy which minimizes C R (N ) is [3, p. 44]:

0, 1, 2, . . .), from Appendix 2.1,

increases strictly with N to h(T ), and Q 1 (N ) increases strictly to h(∞). Therefore,

From the inequality C F (T, N + 1) − C F (T, N ) ≥ 0,

In addition, substituting (1.13) for (1.14), (1.14) becomes

(c F − cT )[Q 1 (T, N ) − h(T )]

(N ) (t) − G (N +1) (t)]

cT > c N , (1.15) becomes

(c F − cT )[Q 1 (T, N ) − h(T )]

where Q 1 (T, N ) is given in (1.12), and

θ(θt) N −1 /(N − 1)!

as T → ∞, because r N (T ) increases with T to θ. Therefore, if (1.17) is greater

1.2 Replacement Overtime

It might be wise to replace an operating unit at the completion of some working

the probability that it is replaced at failure before time T is

and the probability that it is replaced at failure over time T is

Therefore, the expected cost rate is

C O (0) ≡ lim C O (T ) = C R (1)

in (1.10) when c O = c N , and

C O (∞) ≡ lim C O (T ) = C A (∞)

We find an optimum TO∗ which minimizes C O (T ). Differentiating C O (T ) with

1.3 Comparisons of Age and Random Replacements

We compare theoretically replacement overtime to age replacement policies with time

the probability that it is replaced at failure before time T is

and the probability that it is replaced at failure over time T is

Therefore, the expected cost rate is

where c F , c N and c O are given in (1.5) and (1.22). In particular, lim T →∞ C O F

When G(t) = 1 − e−θt (0 < θ < ∞) and c O = c N , (1.31) is rewritten as

We discuss optimum TO∗ F and N O∗ F which minimize C O F (T, N ) in (1.32).

1 (T ) increases strictly to h(∞), the left-hand side of (1.33) increases

L 2 (T, N + 1) − L 2 (T, N ) = [Q 2 (T, N ) − Q 2 (T, N − 1)]

G(t) = 1 − e−t for ci /c F (i = T, O, N )

1.4 Replacement Overtime Last

the probability that it is replaced at failure before time T is

and the probability that it is replaced at failure over time T is

where note that (1.36)+(1.37)+(1.38)+(1.39) ≡ 1. The mean time to replacement is

Therefore, the expected cost rate is

lim C O L (T, N ) = lim C O F (T, N ) = C R (N ),

When G(t) = 1 − e−θt and c N = c O ,

where note that C O L (T, 0) = C O L (T, 1).

1 (T ) is given in (1.24). Letting L 2 (T, N ) be the left-hand side of (1.43),

From the inequality C O L (T, N + 1) − C O L (T, N ) ≥ 0,

1 (T ) increases strictly to h(∞), the left-hand side of (1.33) increases

1 (T ) is given in (1.24). Letting L 2 (T, N ) be the left-hand side of (1.43),

In addition, because Q1 (T ; S) increases with S to Q

which agrees with Q 1 (T ; l) increases