IEEE TRANSACTIONS ON RELIABILITY, VOL. 37, NO.

3,1988 AUGUST 295

Sequential Imperfect Preventive Maintenance Policies

Toshio Nakagawa The mean cost-rates of two models are obtained and
Aichi Institute of Technology, Toyota the optimal sequences { X k } are discussed. When the failure
time has a Weibull distribution, the optimal policies are
computed explicitly. Numerical examples are given.
Key Words - Sequential policy, Imperfect maintenance,
Minimal repair, Mean cost, Optimization Notation

Reader Aids - h(t),H(t) hazard rate, cumulative hazard of the system.

Purpose: Report derivations xk scheduled intervals between pm’s (k = 1 , 2, ...,N).
Special math needed for derivations: Calculations, probability N scheduled number where the system is replaced; it
Special math needed to use results: Same is replaced at time X I + xz + ... + xN.
Results useful to: Reliability theoreticians ak improvement factor in hazard rate in period k of
pm; 1 = a. < al 6 a2 6 ... 6 &+I.
Abstract - This paper introduces improvement factors in k- I

hazard rate and age for a sequential preventive maintenance @m) Ak I I U j ( k = 1 , 2,..., N ) ; I = A i < A z < ... < A N .
j=0
policy and analyzes two imperfect pm models: 1) pm reduces the
hazard rate while it increases with the number of pm’s, and 2) pm c1 cost of minimal repair at failure.
reduces the age. The pm is done at intervals x k (k = 1, 2,..., N) cz, c3 costs of scheduled pm and replacement.
and is imperfect. The optimal policies to minimize the mean cost- bk improvement factor in age of the system reduced
rates are discussed. The optimal pm sequences { x k } are computed by pm k; 0 = bo < bl < b2 < ... < bN-1.
for a Weibull distribution. dk ( 1 - bk)[(l - bk)/(l - bQ]”‘-’).
Yk age of the system immediately before pm k; Y k =
xk +
bk-1Xk-l -k ... 4- bk-1 bk-2 ... bzblX1, X k =
1 . INTRODUCTION Y k - bk-flk-1.
Ci(xl, xz, ..., xN) mean cost-rates of the system for
Nguyen & Murthy [ l l ] and Nakagawa [lo] have pro- models A & B.
posed a sequential policy where the preventive
maintenance (pm) is done at fixed intervals xk. This policy Other, standard notation is given in “Information for
could be applied to real systems because most systems need Readers & Authors” at rear of each issue.
more frequent maintenance with age. In one approxima-
tion model of imperfect pm, the system has different 2. MODEL A - HAZARD RATE
hazard rates in period k of pm, while they increase with the
number of pm’s. Assumptions
The first imperfect pm model where pm is imperfect 1 . The pm is done at fixed intervals xk ( k = 1 , 2 , ...,
with probability p was considered in [ 3 ] .Similar imperfect N - 1 ) and is replaced at pm N, ie, the system undergoes
pm and repair models were discussed in [1 , 4, 6 , 71. Lie & pm at successive times x1 < X I + xz < ... < X I + x2 + ...
Chun [5] introduced an improvement factor in hazard rate + xN-land is replaced at time X I + x z + ... + xN.
after maintenance. Canfield [2] considered system 2. The hazard rate h(t) after pm k becomes akh(t)when
degradation with time where pm restores the hazard func- it was h(t) in period k of pm, ie, the system has the hazard
tion to the same shape, while the level remains unchanged rate A&(t) in period k of pm where 1 = a. < al 6 ... 6
and the hazard function is monotone. aN-l.The hazard rate increases with the number of pm’s.
It is reasonable to postulate the system where pm 3 . Time returns to zero after replacement, ie, the
reduces the hazard rate or the age but it may seem severe to system is as good as new at replacement.
4. The system undergoes only minimal repair at
assume that the system has different hazard functions be-
failures between replacements; the hazard rate remains un-
tween pm’s. The improvement by pm depends on cost of changed by minimal repair (by definition of minimal
pm and/or age of the system. repair).
By introducing improvement factors [5, 91 in hazard 5 . The times for pm, repair, and replacement are
rate and age for a sequential pm policy [ l o , 1 1 1 , I consider negligible.
the following two pm policies: The pm is done at fixed in- 6 . The hazard rate h(t) is continuous and strictly in-
tervals xk (k = 1 , 2 , ..., N - 1 ) and is replaced at pm N . If creasing.
the system fails between pm’s, it undergoes only minimal
repair. The pm is imperfect: 1 ) the hazard rate after pm k Derivations
becomes akh(t)when it was h(t) in period k of pm, and 2)
the age after pm k reduces to bkt when it was t before pm. The mean cost-rate is:

0018-9529/88/0800-0295$01.00O1988 IEEE

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on July 20,2024 at 19:26:15 UTC from IEEE Xplore. Restrictions apply.
I

296 IEEE TRANSACTIONS ON RELIABILITY, VOL. 37, NO. 3,1988 AUGUST

Cl(X1, XZ, ..., X N ) =

[ c AkH(Xk)
c1
lk:

1
+ (N - 1) cz + c3 / ( X I + x2 + e.. -k X N ) .

By the method similar to [lo], the optimal policy is

(1)
Further, if h(t)is differentiable and h’(t)is also strictly in-
creasing then a solution to ( 5 ) is unique.
Therefore, substituting each Y k into (6), its equation
derived as follows: A necessary condition that a sequence becomes a function only of y N :
{ x k }be a minimum cost schedule is that dCl/dxk = 0 for rN-1 T N
each k. Thus, differentiating Cl(xl, xz, ..., xN)with respect
to xk an setting it equal to zero imply:

Alh(X1) = Azh(X2) = ... = A,h(x,), (2) = [(N - 1)CZ + c3]/c1, (7)

where each y k (k = 1, 2, ...,N - 1) is given by some func-

tion of y N . If there exists a solution y N to (7), then a se-
When the hazard rate h(t) is strictly increasing to infinity, quence {yk} minimizes the mean cost-rate Cz(y1, yz, ...,
we can specify the computing procedure for obtaining an YN) -
optimal schedule [101: Finally, suppose that y,, y,, ..., y N are determined
1 . Solve Akh(xk)= D and express xk (k = 1,2, .,N ) .. from ( 5 ) and (7). Then, from (6), the resulting cost-rate is
by a function of D. clh(yN),which is a function of N . To complete an optimal
2. Substituting xk into (3) and solve it with respect to pm schedule, we may seek an optimal number N which
D. minimizes h(yN).
3. Determine N which minimizes D. From the above discussions, we can specify the com-
puting procedure for obtaining the optimal pm schedule:

3. MODEL B - AGE 1. Solve (5) and expressyk (k = 1 , 2, ..., N - 1) by a

function of y N .
Assumptions 2. Substitute y k into (7) and solve it with respect to
YN.
1 , 3-6. Same as Model A. 3. Determine N which minimizes h(yN).
2. The age after pm k reduces to bkt when it was t 4. Compute xk (k = 1 , 2, ..., N ) from xk = yk-
before pm, ie, the system of age t becomes t(l - bk)units
of time younger at pm k,where 0 = bo < bl < bz < ... <
bk-LYk-1.

bN < 1.
4. NUMERICAL EXAMPLES IN WEIBULL CASE
The system is aged from bk-l(xk-l+ b k - g k - 2 + ... +
bk-Zbk-3 ... bzblxl)after (k - 1)-th pm to xk bk-l(~k-l + Suppose that the failure time of the system has weif(t,
+ b k - g k - 2 + ... + b k - Z b k - 3 ... bzblxl)before pm k, ie, p), ie, h(t) = 0ts-l for 0 > 1 .
from bk-t))k-lto y k . Thus, the mean cost-rate is - From the computing procedure of model A, solving
the equation Akh(xk)= D,we have

Substituting xk into (3) and arranging it imply:

Differentiating CzCyl, y2, ..., y N ) with respect to yk

and setting it to zero imply: (B - 1)/8

k=l
(9)

( k = 1 , 2,...,N - l), (5)

From [lo], an N * to minimize D is given by a unique
minimum such that -

Suppose that y N(0 < y N < m) is fixed. From assump-

tion 6, there exists some y k (0 < y k < y N )which satisfies
( 5 ) since -

since A k are increasing in k.

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on July 20,2024 at 19:26:15 UTC from IEEE Xplore. Restrictions apply.
 NAKAGAWA: SEQUENTIAL IMPERFECT PREVENTIVE MAINTENANCE POLICIES 297

Further, if Ak -
03 as k- 03 then a finite N*exists TABLE 1
Optimal N * and pm intervals for Weibull examples
uniquely in (lo), and the optimal intervals are given in (8)
and (9). Model A
For Model B, solving (3,we have CI/C2 = 3

c3/c2 2 5 10 20 40

N* 2 3 4 5 6

Substituting Y k into (7) and arranging it, x1 0.77 1.06 1.37 1.82 2.45
x2 0.52 0.71 0.92 1.21 1.64
x3 0.43 0.55 0.73 0.98
x4 0.31 0.42 0.56
X5 0.23 0.31
x6 0.17
Next, consider the problem which minimizes
Model B
N- 1
B(N) = [(N- l)C2 + C3]/
k=O
dk, (N = 1, 2, ...), N* 1 2 4 7 11
~~ ~

0.54 0.82 1.07 1.40 1.84

which is the same problem as minimizing h ( y ~ )ie,
, C2(y1, 0.82 0.43 0.56 0.74
y2,..., yN). From the inequalities B(N + 1) 2 B(N) and 0.28 0.36 0.48
B(N) < B(N - 1) (N= 1, 2, ...) where B(0) = 03, we 0.92 0.27 0.35
0.21 0.28
have
0.18 0.23
1.13 0.20
L2(N) c3/cz and Lz(N - 1) < c3/c2, (N= 1,2, ...), 0.17
0.15
(13) 0.14
1.45

5 . SUMMARY & CONCLUSIONS

Most practical systems would not have different
If dkis decreasing in k then Lz(N)is increasing in N. hazard functions defined in [lo, 111, but rather the same
Thus, there exists a unique minimum N*such that Lz(N) 2 hazard rate h(t) which may be reduced by pm. This paper
c3/c2 if L Z ( a )> c3/cZ. presents two imperfect pm policies by using a concept of
Show that dkis decreasing in k from the assumption improvement factors introduced in [5, 91. It is easy in a
that bk < bk+l. Let - Weibull distribution case to compute optimal pm intervals.
The example shows that x1 > xz > ... > xN for both
g(x) = (1 - X)B/(l - A.)! Model A and the previous model [lo], but x1 > xN > x2
for c3/c2 = 10, 20, 40 of Model B. This shows that it
Then, g(x) is decreasing from 1 to 0 on 0 < x < 1 for 8 > 1 , would be reasonable to do frequent pm with age, but it
and hence, would be better to do the last pm as late as possible because
the system should be replaced at next pm. It may be given
as a conclusion that Model B would be more practical than
which follows dk > dk+l.Further, if bk - 1 as k - 03 then
Model A and the previous model, as one approximation
model of imperfect pm where the system is younger at each
pm [8, 91.
lim dk = lim [~(x)I’’@-~)
= 0,
k- m x-1

REFERENCES
-
ie, Lz(N) 03 and a finite N*exists uniquely.
Therefore, if bk 1 as k- -
03 then an N * is a unique
(11 M. Brown, F. Proschan, “Imperfect repair”, J. Appl. Prob., vol
20, 1983 Dec, pp 851-859.
minimum such that L2(N)2 c~/cZ and the optimal intervals [2] R. V . Canfield, “Cost optimization of periodic preventive
WeXk = Y k - bk-flk-1 (k = 1, 2, ..., N 3 WhereYkandYw maintenance”, IEEE Trans. Reliability, vol R-35, 1986 Apr, pp
are given in (11) and (12). 78-81.
Table 1 gives the optimal numbers N * and the pm in- 13) P. K. W. Chan. T. Downs, “TWO criteria for preventive
maintenance”, IEEE Trans. Reliability, vol R-27, 1978 Oct, pp
tervals xl, xz, ..., x,. for c3/cZ = 2, 5, 10, 20, 40 where 272-273.
Cl/C2 = 3, 8 = 2, and Uk = [1 + k / ( k I)], bk = k / ( k 141 R. A. Fontenot, F. Proschan, “Some imperfect maintenance
+ 1) (k = 0, 1, 2, ...). models”. in Reliability Theorv and Models, Academic Press, 1984.

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on July 20,2024 at 19:26:15 UTC from IEEE Xplore. Restrictions apply.
298 IEEE TRANSACTIONS ON RELIABILITY, VOL. 37, NO. 3,1988 AUGUST

[ 5 ] C. H. Lie, Y. H. Chun, “An algorithm for preventive maintenance AUTHOR

policy”, IEEE Trans. Reliability, vol R-35, 1986 Apr, pp 71-75.
Dr. Toshio Nakagawa; Department of Industrial Engineering; Aichi In-
[6] D. N. P. Murthy, D. G. Nguyen, “Optimal age-policy with im-
perfect preventive maintenance”, IEEE Trans. Reliability, vol R-30, stitute of Technology; Toyota 470-03 JAPAN.
1981 Apr, pp 80-81. Toshio Nnkagnwn was born in Yokkaichi City, Mie Prefecture,
[7] T. Nakagawa, “Optimal policies when preventive maintenance is Japan on 1942 November 30. He received BSE and MS degrees from
imperfect”, IEEE Trans. Reliability, vol R-28, 1979 Oct, pp Nagoya Institute of Technology in 1965 and 1967, and a PhD degree from
331-332. Kyoto University in 1977. He was working in the Department of
[8] T. Nakagawa, “Mean time to failure with preventive maintenance”, Mathematics, Meijo University in 1967-1988. He is now Professor in the
IEEE Trans. Reliability, vol R-29, 1980 Oct, p 341. Department of Industrial Engineering, Aichi Institute of Technology in
[9] T. Nakagawa, “A summary of imperfect preventive maintenance Toyota. His research interests are optimization, failure-time distribu-
policies with minimal repair”, R . A . I. R . 0. Oper. Res., vol 14, tions, and computer systems in reliability theory.
1980 Aug, pp 249-255.
[lo] T. Nakagawa, “Periodic and sequential preventive maintenance
policies”, J. Appl. Prob., vol 23, 1986 Jun, pp 536-542.
[I 11 D. G. Nguyen, D. N. P. Murthy, “Optimal preventive maintenance Manuscript TR87-036 received 1987 March 19; revised 1988 March 16.
policies for repairable systems”, Oper. Res., vol29, 1981 Nov-Dec,
pp 1181-1194. IEEE Log Number 21371 4 TR b

2nd Annual Workshop: R&M in Computer-Aided Engineering

Reflecting Reliability & Maintainability Considerations in Designs
Generated by Computer-Aided Engineering (CAE) Techniques

Based on the outstanding success of the 1987 R&M in CAE Workshop, the IEEE Reliability Society, in cooperation
with the IEEE Computer Society, will sponsor a second Workshop aimed at assuring that R&M concerns are addressed
when the capabilities of engineering workstations are being defined. Last year’s Workshop posed four questions related to
what is needed, what exists, and what can be expected. This year’s Workshop will explore and update the answers by bring-
ing together R&M engineers and workstation designers to exchange information on the current capabilities of CAE worksta-
tions and the additional capabilities needed to respond to R&M requirements.

This will be an active Workshop, not a passive conference

Day 1 Focus:
What is the status of R&M in CAE?
What recent break-throughs have occurred for R&M in CAE?
What barriers remain to integrating R&M in CAE?
Day 2 Focus:
Find solutions for overcoming the barriers identified during Day 1. The participants will be from diverse backgrounds
and industries so that proposed solutions can be realistic.
Day 3 Focus:
Summarize the solutions reached during Day 2, and prepare plans for action. Present results of an industry survey on
R&M in CAE implementation and an overview of joint US-DoD/Industry studies and recommendations on this topic.

Date 1988 September 27-29 Fee: $425.00

Location Xerox Training Center Includes lodging for 3 nights, all meals &breaks, conference record,
Leesburg, Virginia USA and all workshop sessions. Make check payable to IEEE R&M-CAE
Period 2 % days Workshop, in US dollars.

Further information: Registration:

Henry Hartt R&M-CAE Workshop

V I T R O Corporation c/o J . W . Thomas, S T E M C O
14000 Georgia Avenue 8730 Georgia Ave; Suite 600
Silver Spring, Maryland 20906 USA Silver Spring, Maryland 20910 USA
phone: 301-231-1431 phone: 301 -585-0421

Authorized licensed use limited to: UNIVERSITY OF MANITOBA. Downloaded on July 20,2024 at 19:26:15 UTC from IEEE Xplore. Restrictions apply.