Professional Documents
Culture Documents
2015 Book MaintenanceOvertimePoliciesInR PDF
2015 Book MaintenanceOvertimePoliciesInR PDF
Toshio Nakagawa
Xufeng Zhao
Maintenance
Overtime Policies
in Reliability
Theory
Models with Random Working Cycles
Lecture Notes in Production Engineering
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
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].
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
ix
x Contents
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.
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)
0 S1 S2 S3 S j-1 Sj
1.1 Age and Random Replacements 3
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
c F − (c F − cT )F(T )
C A (T ) ≡ lim C F (T, N ) = T . (1.6)
N →∞ F(t)dt
0
C A (T A∗ ) = (c F − cT )h(T A∗ ). (1.8)
4 1 Age Replacement Overtime
(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)
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 =
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)
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
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)
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
where c O = replacement cost over time T with c O < c F and c F is given in (1.5). In
particular,
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
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
8 1 Age Replacement Overtime
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).
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
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
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).
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
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),
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.
12 1 Age Replacement Overtime
Table 1.1 Optimum T A∗ , TO∗ , N R∗ and their expected cost rates when F(t) = 1 − e−(t/10) and
2
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)
In particular,
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)
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
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.
1.4 Replacement Overtime Last 15
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,
16 1 Age Replacement Overtime
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
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.
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
F(T ), (1.50)
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
+ ydF(y) dG(u) dG ( j) (t)
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
18 1 Age Replacement Overtime
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
where
S
e−θt dF(t)
1 (T ; S) ≡ T
Q ≤ h(S),
S
T e−θt F(t)dt
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)
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
where note that (1.58) + (1.59) + (1.60) = 1. The mean time to replacement is
∞
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
20 1 Age Replacement Overtime
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)
where
∞ −(θ+l)t
T e dF(t)
Q 1 (T ; l) ≡ ∞ ≥ h(T ),
−(θ+l)t
T e F(t)dt
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
1.5 Finite Interval 21
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
N −1 T
F(S) G(S − t)dG ( j) (t), (1.69)
j=0 0
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
22 1 Age Replacement Overtime
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
+ ydF(y) dG(u) dG ( j) (t)
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)
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 ,
1.5 Finite Interval 23
F(T ), (1.78)
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
+ ydF(y) dG(u) dG ( j) (t)
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
24 1 Age Replacement Overtime
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)
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)
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
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)
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)
26 1 Age Replacement Overtime
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)
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)
When F(t) = 1 − e−(t/10) and G(t) = 1 − e−t , Table 1.3 presents optimum
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
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
(1.93)
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
28 1 Age Replacement Overtime
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
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)
then there exists a finite and unique Tn,∗ p (0 < Tn,∗ p < ∞) which satisfies (1.100),
and the resulting cost rate is
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
30 1 Age Replacement Overtime
∞
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)
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)
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
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 (2014) Random maintenance policies. Springer, London
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
© Springer International Publishing Switzerland 2015 33
T. Nakagawa and X. Zhao, Maintenance Overtime Policies in Reliability Theory,
Lecture Notes in Production Engineering, DOI 10.1007/978-3-319-20813-8_2
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).
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
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
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)
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:
36 2 Periodic Replacement Overtime
∞
(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, . . .),
∞ N ∞
θ(θt) N −θt (θt) j −θt
H1 (N ) = e h(t)dt = e dh(t)
0 N! j!
j=0 0
∗ ∗
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
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 ,
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
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,
in (2.6), and
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
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
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
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 )]
i.e.,
T ∞
(N ) −θu cO
[1 − G (t)] θe [h(u + T ) − h(t)]du dt = , (2.27)
0 0 cM
then there exists a finite and unique TO∗ F (0 < TO∗ F < ∞) which satisfies (2.27),
and the resulting cost rate is
∞
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
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
2λ2 ∗ cO
(λTO∗ )2 + TO = .
θ cM
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)
then there exists a finite and unique TO∗ L (0 ≤ TO∗ L < ∞) which satisfies (2.33),
and the resulting cost rate is
∞
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!
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
46 2 Periodic Replacement Overtime
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.
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
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
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
48 2 Periodic Replacement Overtime
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
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
2.5 Replacement Overtime with Number of Failures 49
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
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
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∗ = ∞.
50 2 Periodic Replacement Overtime
∞
∞
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
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
where
∞
p K −1 (t)h(t)dt
2 (T, K − 1) ≡ ∞
Q T
.
T p K −1 (t)[h(t)/Q(t)]dt
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 .
cN + cM
C O (0) = = C(1), (2.61)
μ
2.6 Replacement Overnumber 53
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
∞
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
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)
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
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)
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
56 2 Periodic Replacement Overtime
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
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)
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
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
58 2 Periodic Replacement Overtime
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)
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. Tadj L, Ouali MS, Yacout S, Ait-Kadi S (eds) (2011) Replacement models with minimal repairs.
Springer, London
4. Nakagawa T (2014) Random maintenance policies. 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 < ∞).
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
T T
(k-2)T S j (k-1)T kT S j+1 (k+1)T
T T T
∞
(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
Clearly,
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).
62 3 Inspection Overtime
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
1 λT 1 1 − e−θT
(e − 1) − T > [1 − e−(θ−λ)T ] −
λ θ−λ θ
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)
∞
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
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
64 3 Inspection Overtime
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.
and the probability that it fails and its failure is detected at time Y j is
T
F(t)dG(t), (3.23)
0
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
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)
i.e.,
G(T )
MR = T . (3.26)
0 G(t)dF(t)
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)
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
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 Inspection First and Last 67
and the probability that it fails and its failure is detected at time Y j is
∞
F(t)dG(t), (3.37)
T
i.e.,
G(T )
MT = ∞ . (3.39)
1 − T G(t)dF(t)
i.e.,
G(T )
MR = ∞ . (3.40)
1 − T G(t)dF(t)
In particular,
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 .∗
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)
70 3 Inspection Overtime
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
3.2 Inspection First and Last 71
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.
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
θ 1 1 θ
= e−λT l D + T + − + e−λT ,
θ+λ θ λ λ(θ + λ)
i.e.,
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
C O (∞) ≡ lim C O (T ) = ∞,
T →∞
θ cD
C O (0) ≡ lim C O (T ) = c O +1 + ,
T →0 λ θ
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
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
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.
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
∞ ( 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
and
1
M L∗ (λ) + 1 = ∞ .
1− T G(t)λe−λt dt
1
∞ ∞
∞ T
= G O (t)dt = T + G(t − u)dG ( j) (u) dt.
θO 0 j=0 T 0
θ
= e−λT ,
θ+λ
3.4 General Failure Times 77
and
1 ∞ ∞ T
−θ(t−u) ( j)
=T+ e dG (u) dt
θO
j=0 T 0
1
=T+ .
θ
Thus, the expected cost is
cR + c D (T + 1/θ) cD
CR (G O ) = −λT
− ,
1 − [θ/(θ + λ)]e λ
we have
∗ 1
1 + M Fn (λ) = T "n # .
0 i=1 G i (t)λe−λt dt
78 3 Inspection Overtime
!
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
!
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
3.4 General Failure Times 79
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.
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
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
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
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
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
∞ ∞ 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.,
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
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
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
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
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 (2014) Random maintenance policies. Springer, London
4. Nakagawa T (2011) Stochastic processes with applications to reliability theory. Springer, Lon-
don
84 3 Inspection Overtime
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
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.
[1 − G (N ) (T )]P K (T ), (4.1)
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
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
G (N ) (T )PK (T ), (4.7)
where note that (4.7) + (4.8) + (4.9) = 1. The mean time to replacement is
∞ ∞
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
88 4 Replacement Overtime with Three Variables
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
G (N ) (T )P K (T ), (4.13)
[1 − G (N ) (T )]PK (T ). (4.14)
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 ∞ ∞
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,
whose left-hand side increases strictly with T from 0 to ∞. Thus, there exists a finite
and unique TF∗ (0 < TF∗ < ∞) which satisfies (4.22), and the resulting cost rate is
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),
and the resulting cost rate is
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),
and the resulting cost rate is
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
92 4 Replacement Overtime with Three Variables
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 ∞
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)
(4.31)
∞ 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)
whose left-hand side increases strictly with T from 0 to ∞. Thus, there exits a finite
∗ (0 < T ∗ < ∞) which satisfies (4.35), and the resulting cost rate is
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
cO
− H (T ) + [1 − G (N ) (t)PK (t)]h(t)dt = , (4.38)
T cM
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.
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
96 4 Replacement Overtime with Three Variables
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
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
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
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
98 4 Replacement Overtime with Three Variables
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
∞
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
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
100 4 Replacement Overtime with Three Variables
∞
T
K −1 ∞
(N )
j p j (t)dG (t) + j p j (t)dG (N ) (t)
j=K 0 j=0 T
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)
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,
102 4 Replacement Overtime with Three Variables
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
whose left-hand side increases strictly with T from 0 to ∞. Thus, there exists a finite
∗ (0 < T ∗ < ∞) which satisfies (4.57), and the resulting cost rate is
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
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
(4.64), and the resulting cost rate is
Noting that the left-hand side of (4.64) increases with n, TL∗,n decreases with n.
References
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
θ(θT ) N /N !
r N (T ) ≡ ∞
j=N +1 [(θT ) /j!]
j
lim r N (T ) = 0, lim r N (T ) = θ.
T →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!
1 F(T )
Q(T ) ≡ ∞ = ∞
T e−[H (t)−H (T )] dt T F(t)dt
1
Q(T ) > h(T ), lim Q(T ) = , and lim Q(T ) = h(∞).
T →0 μ T →∞
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
to 0.
Proof Note that
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 ) 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 ), 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
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
we have
lim Q 1 (N ) = h(∞),
N →∞
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
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
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
N ∞ −θu dF(u)]dt
0 (θt) [ t e
Q 2 (T, N ) ≡ T
N ∞ −θu F(u)du]dt
0 (θt) [ t e
112 Appendices
1 (0), Q
lim Q 2 (T, N ) = Q 1 (∞) = Q 1 (0) < Q 2 (T, N ) < Q
1 (T ).
T →0
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
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, . . .,
∞
N ∞ −θu dF(u)]dt
T (θt) [ t e
Q 2 (T, N ) ≡ ∞
N ∞ −θu F(u)du]dt
T (θt) [ t e
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 ) = lim Q
lim Q 1 (T ) = h(∞).
N →∞ T →∞
n (0) to λ.
increases strictly with T from Q
Proof First, note that the failure rate of a parallel system with n (n ≥ 2) units
when F(t) = 1 − e−λt ,
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
β (0) to λ.
increases strictly with T from Q
Proof First, note that
βλe−λt
h β (t) =
exp(βe−λt ) − 1
β (T ) with respect to T ,
Next, differentiating Q
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
Appendix 3
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
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 )
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 )
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
∞
(θt) N e−θt h(t)dt
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
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