Mieroeleetron. Reliab., Vol. 29, No. 1, pp. 57-71, 1989. Printed in Great Britain.
0026-2714/8953.00+ .00 © 1989 Pergamon Pre-.spie
OPTIMUM SCHEDULING OF A NEW MAINTENANCE PROGRAM UNDER STOCHASTIC DEGRADATION B.D. SIVAZLIAN Department
of Industrial and Systems Engineering, University of Florida, Gainesville, Florida 32611, U.S.A.
(Received for publication
31 December
1987)
ABSTRACT In this paper we consider an intermittently
operating system which
is under periodic review and which is subject to random aging or degradation.
The optimal schedule when initiating a new maintenance
program is analyzed so as to minimize The scheduled maintenance rather than time. incorporate costs,
cost.
is based on the level of system degradation
The functional equation approach enables one to
such factors as the fixed cost of maintenance,
the discount
a given period.
I.
the long term discounted
factor and the probability
The analysis
the recurring
distribution
of aging in
is illustrated with an example.
INTRODUCTION We consider an intermittently
periodic
operating system which is under
review and which is subject to random degradation.
of degradation
The amount
during each review period is a random variable indepen-
dently and indentically
distributed
from period to period.
A typical
example would be a truck where the number of miles registered per month would measure
the level of degradation,
The system degradation known as recurring costs. system,
a major maintenance
necessary, program, assumed
induces increases
in operational
In order to compensate (or overhaul)
whose time is negligible
for the usage of the
relative to the review period,
the amount of degradation
initiated.
associated with each maintenance.
is
At the end of
since last maintenance
and if this amount exceeds a specified level, is immediately
if
The maintenance
to bring the system to a level as good as new.
maintenance
costs,
program is initiated,
at the beginning of a review period.
each review period, measured,
the review being month.
say R, a
There is a fixed cost K
We assume that a new maintenance 57
is
58
B, D. SIVAZL1AN
program is being contemplated and our objective is to determine at which level of R, say R 1 should the new maintenance program be initiated,. Also what should be the optimum level of R, say R 2, at which the new maintenance program should proceed in the future. Problems of maintenance can sometimes be categorized as replacement type problems.
The literature work has dealt with either continuous
review system or periodic review system operating under a variety of policies using undiscounted or discounted costs.
Among the more recent
publications one should note the work of Bergman and Klefsgo Okumoto and Elsayed (1986) and Nakagawa
(1983), Nakagawa (1986).
1 - following maintenance, new;
(1984) Nakagawa
(1983),
(1984), Tapiero
Most of these papers assume that: the system is brought to a level as good as
2 - the optimum maintenance policy consists in the determination of a single critical number; 5 - cost is undiscounted; 4 - the same maintenance policy is used over and over again. A dyadic policy was considered by Sivaz!ian and Iyer (1981) where two critical numbers had to be determined optimally. where following maintenance,
Another variant
the system is hysteretic,
that is, is
brought to a level which is not necessarily as good as new was studied by Brown, Mahoney and Sivazlian
(1982).
In general,
these variants are
considerably more complex than the standard problems and will not be considered here.
The problem under investigation assumes a discounted
cost and studies the optimal scheduling of a proposed new maintenance program over an existing one: the scheduled maintenance is based on the level of system degradation rather than time.
II.
PROBLEM STATEMENT We assume that time is divided into equal intervals known as
periods.
During each period the system deteriorates or degrades or
ages.
Following the last maintenance or overhaul,
the system age is
zero.
The aging could be measured on any selected scale, e.g., it could
be the number of hours of operation of the system, or it could be the number of miles registered. Assume that the amount of aging over successive time periods is a non negative continuous random variable,
identically and independently
O p t i m u m scheduling
59
distributed
from period to period with distribution
probability
density function @(~),0<~<-,
f
#(~)=
function 0(~),
and finite moments.
Thus
(i)
#(u)du
~0 The structure of the maintenance economic
factors affecting
policy usually will depend on the
the operation of the system.
Two such
factors must be identified.
a.
The cost associated
in
overhauling
the system which we shall denote
by the fixed cost K (although K could be a function of the age at which the system is overhauled).
b,
The cost associated in operating period.
If x denotes
the age of the system at the beginning of the
period following a decision, the end of the period. operating
the system for an additional
then the system will be age x + ~ at
Let C(x)dx be the recurring cost of
the system between ages x and x + dx.
cost of operating
the system for one period,
age x the beginning of the period,
h(x) =
In general,
iZfl
[C(u)du]
Then the expected
given that it is of
is
0 < x < ®
~(~)dE
L(x) will be a non-decreasing
(2)
function of x, assumed
differentiable. The maintenance period,
policy used herein is:
"When at the end of a
the system age exceeds a level R, overhaul
We shall denote by u, 0 < u < i. the discount rate measured
in
$/
is no new maintenance expected
($)(period). policy.
the system". factor or interest
We consider first the case when there
A functional equation
for the total
cost over an infinite horizon period is formulated and solved.
The optimal value R* of R is obtained. to solve the case when a new maintenance Finally an example is presented.
Next. we use the results derived program is to be instituted.
60
B.D. SIVAZLIAN
Ill.
i,
THE DISCOUNTED STOCHASTIC NEW MAINTENANCE PROGRAM
MODEL OVER INFINITE HORIZON
- CASE OF NO
The Functional Equation
age
T
/i
.l ! l
i
overhaul FIGURE I.
time
A Periodic Review Maintenance System
Let f(x) = total expected discounted cost over an infinite horizon period given that the initial system age is x, 0 & x < R; this is a conditional expected cost given x, which is also a function of R. Then f(x) satisfies the following functional equation (see figure I) R f(x) = L(x) + ~ J f(y) ~(y-x) dy J vx
+ a[Z + f(O)] { ~(y-x)dy,
(3)
JR To explain this functional equation, we note that the first term on the right-hand side represents the expected cost incurred during the first period.
The second term is the expected cost incurred when starting the
second period with the system age y, x < y < R, when there is no overhaul, discounted to time origin.
The third term is the expected
cost incurred when starting the second period with an overhauled system, that is age zero, discounted to time origin. Let
g(x) = L(x) + a [K + f(O)] f® @(y - x)dy
(4)
JR Then R
f(x) = g(x) ÷ ~ f
f(y) ~(y - x)dy x
(5)
O p t i m u m scheduling
61
The solution of this integral equation, which is the Volterra type, is
f(X)
=' g(x) + n=1~~n J~x g(u) o(n)(u -
0(n)(.)
where
refers
to
the
x)du
(6)
convolution of 0(.) with itself.
n th
Let O ( u ) d u = @(x)
(7)
Relation (4) can be writte9 g(x)
= L ( x ) + (~K 0(R - x ) + a f ( O )
Substituting (8) in (6) f(x)
:
L(x)
we
obtain,
+ ~K # ( R - x )
÷ af(O)
@(R - x )
(8)
#(R - x)
R
+ ~ n I [L(u) + aK O(R - u) + af(O) O(R - u)]o(n)(u - x)du n=l Jx
(9)
At x = 0 f(O)
L(O)
=
+
aK O(R)
af(O) )(R)
+
+n=l ~ ~n fl [L(u) +
aK 0(R
- u)]
o(n)(u)du
+ ~ n |~r(o) ~(R - u) ~(n)(u)du n=l
(io)
#~0
Hence
f(O)
R
m
r
11 - a ~ ( R ) - u Z a n L n=l
;(R- o, .'n'(u,du] R
= L(O)
+ aM O(R)
an
+ QK
Ii ;(R - u)o(n)(u)du
n=l
+
n=l~ ~n
(n)
~0 L(U) o(n)(u)du
Let T(U) MR 29: I-E
=
Z .no(n)(u) n=l
O~;u<~
(12)
B. D. SIVAZL~N
62
Then, the Laplace transform of x(u), say T(s), expressed in terms of the Laplace transform of 0(u), say $(s) is
~(S) =
~ an[$(S)] n n=l
uS(s)
(13)
. since l~$(s)l < i
i - a%(s)
Consider the expression X
M(x) = ~(x) ÷
~ n
I
n=l
~(x - U) ,(n)(u)du
0 ~
x <-
(14)
0
The Laplace transform of this expression is
$(S)][$(s)]n s
$(s) + ~ anFi
i M(S) = ~ -
S
n=l
--
-s
m
n
n=0
i - =%(s) =!÷_i.
=$(s)
s
s i - =$(s)
i s
i - ~ =
i
$(s)
s i - =$(s)
1 ~(s) s I - =%(s)
(i5)
Formal inversion yields
M(x)
= i
--1
-
(x
(X
i
x
.Z n
0(n)(u)du
0n=l
x
1
- -
(16)
[ x(u)du 00
Now compute the quantity R
1 - aS(R) = 1 - a [I - -1 -l- aa
T(u)du] 0
--(I- a)[i + ~ T(u)duI
Optimumscheduling
63
Using ( l h ) and (17), expression (11) can be written as R
(18)
f(O) [1 - aM(R)] = L(O) ÷ I L(u) T(u)du + aKM(R) J0
Solving for f(0),
L(O) +
|n(u) T(u)du
+ ~KN(R)
#o
f(o) =
1
-
aN(R)
R
L(O) ÷ =
-
K
+
oL(U) T(u)du + K
(19)
1 - aM(R)
Using (17) in (19), we get FR K + L(0) + L L ( u ) vu K + f(0) = R
T(u)du
(I - a) (i + I T(u)du) J0
PR K + L(O) + I L(u) T(u)du J0 (i - a) [z ÷ f(0)] = R
or
(20)
i + / T(u)du J0 Note that f(0) is a function of R and we are seeking for the value of R that minimizes f(0).
2.
The Optimal Value of R The optimal value of R, R*, can be obtained by differentiating
with respect to R and setting the result equal to zero. R
I *
(20)
Thus
R
T(uld
L(R) ~(R)
+ L(0) +
L(u) ~(u) du
T(R) = 0
(21)
#0
or R
R
L(R) - L(O) + L(R) JOIT(u) du - J/oL(U) T(U) du
=K
(22)
64
B . D . SIVAZLIAN
Proof: From equation (9)
f(x)
= L(x)
+ a[K + f ( O ) ]
Z an n=l
#(R - x )
L(u) ¢(n)(u - x)du
[
R
+ =[K + f ( O ) ]
r. a n=l
n
+(a
-
x)
÷
u) ®(n)(u - x)du
vX
Let v = u - x, then
f(x)
= L(x)
÷ a[K ÷ f(O)]
{
~(B
-
an
(R - x - v) ¢(n)(v)dv
n=l
R-X
+
I v0
L(x ÷ v) ~ a n 0(n)(v)dv n=l
U s i n g (12) and ( 1 4 ) ,
we g e t R-X
f(x)
= L(x)
* a [K ÷ f ( 0 ) ]
M(R - x ) ÷ /
L ( x * v) T ( v ) d v
!
vO
From (16) R-X
f(x) = L ( x ) +
a [i - 1 [- ~a
T(u)du] [K + f(0)] v0
~R-x J + I L(x + v) T(v)dv
(27)
J0 Differentiate
f(x) with respect to R, set the result equal to zero, and
obtain R-X ~f(x)
= =[K
+ L(R)
+ f(o)]
{ -
T(R - x )
= 0
{(1
[K ÷ f ( O ) ]
+,u,du
Thus
T(R -
x)
- a)
- L(R)}
R-x
(28) 0
Optimum scheduling
65
Factoring out and interchanging the order of integration yields B
u
L(R) - L(O) + fL'(u) f ,(v)dvdu = K qO
(23)
aO
Thus
L'(u)
1 +
dv
(24)
du = K
Let
L'(x) = ~(x)
and X
T(x) = loT(V) dv
then B
I ~(U) [1 + T(u)]du = K
(25)
0 Note that if L(x) is a strictly increasing function of x, then (25) has a unique solution in R. The formal process of differentiating (20) and setting the result equal to zero yields at optimality (denoted by *) em
K + L(0) + (i = a ) [ K
+ f*(O)]
L(U) r(u)du #0
=
I +
Ii
*~(u)du
= L(R')
3-
(26)
Equivalence Between the Minimization of the functions f(O) and f(x)
Proposition: If L(x) is a strictly increasing function of x, the unique root R=R" that minimizes f(O) with respect to R, minimizes also f(x) with respect to R.
Thus, the initial age x of the system does not affect the
determination of the optimal policy in this situation.
66
B.D. SIVAZL~N
From (26), it is clear that R=R* satisfies (28) since anal(0) ,~R=R* = 0 .
IV.
A VARIATION OF THE DISCOUNTED STOCHASTIC MODEL OVER INFINITE HORIZON - INSTITUTION OF A NEW MAINTENANCE PROGRAM
Assume that the old maintenance program is to be replaced by a new maintenance program having a different fixed cost and which may induce different degradation process on the system than the old one.
The
policy to be instituted is as follows:
If the initial age x of the system is less than RI, let the system operate until at the end of a review period its age exceeds for the first time R I.
Immediately, initiate the new maintenance program
on the system.
Let then the system operate until at the end of a
review period its age exceeds for the first time R 2 at which epoch the system is overhauled, and so forth.
If the initial age x of
the system is greater than RI, then initiate immediately the new maintenance program. Use index 1 to refer to the old maintenance program and index 2 to refer to the new maintenance program.
The following system of
functional equations are obtained:
fl(x) = Ll(X) + ~
I
Rlfl(y
@l(y
x)dy
x
0 < x < R1
(29)
0 w
(30)
1
f2(w) = L2(w) + u
I
R2f2(t
e2(t
w)dt
w
o K2 f2oll °2t w0t R2 It
is
evident
that
the general equation
(30) d e f i n e d
for
all
0 ~ w < R2
may be used in particular to obtain an expression for K 2 + f2(O) which is the expected cost associated to the second alternative (corresponding to initiating immediately a new maintenance policy). proposition, M'n22 f2(w) is equiavalent• to M'n22 f2(0).
Also, from our The objective is to
Optimumscheduling determine the optimal values of R 1 and R 2.
67
Despite the apparent
difficulty of this problem, it can be solved easily on the basis of the analysis performed in the previous section. From equation (30), using equation (25), the optimal value of R 2 is given as the solution to the equation
R2* ,
I
L2 (u) [1 + ~(u)]
du =
K2
(31)
00 where x
T2(X) :
f
T2(u)du
(32)
0
and
T2(u) =
(33)
E ~n ~2 (n)(u) n=l
Then, from (26):
(34)
(i - ~) [K 2 + f2*(O)] = L2(R2*) From (27) the formal solution to (29) is .RI-X [K2 + f2(O)]
r.RI-X
+ jI 0
LI(X + ~) Tl(~)d~
(35)
where
Tl(U) =
~ ~n$1(n)(u) n=l
(36)
Differentiating (35) formally with respect to R 2, we obtain (note fl(x) is a function of R 1 and R2)
~fl(x)
~-R2 - 0 -
~f2(0)
~-R2
and this yields for the optimal value of R 2, R2", defined by (31). Differentiating (35) formally with respect to R 1 and setting the result to zero yields
68
B . D . SIVAZLIAN
afl -- -
(1
-
~) [K2
+
f2(O)]
TI(R
1
-
x)
nl(a 1) TI(R 1
+
-
x)
= 0
At optimality
LI(RI*) = (i - a) [X 2 + f2*(0)] = L2(R2* ) Thus. having determined the optimal value of
(37)
R2*, the
optimal value of
RI* can be determined from (37) (see figure 2).
q,
hi(x) LI(RI*)
; L2(R2*) • . . . . . . . . . . . . . . . . . . .
I .........
I
) RI* FIGURE 2.
R2*
X
D e t e r m i n i n g t h e O p t i m a l V a l u e o f R1 f r o m t h e O p t i m a l V a l u e o f R2
It should be noted that the determination of R2* depends on the value of K 2 and L2(.). whereby the determination of RI* depends on LI(.) but is independent of K 1.
This is to be expected,
since the cost K 1
associated with the last maintenance program is a "sunk" cost, and does not affect any future decisions. We next provide an exsmple to illustrate the calculations.
V.
EXAMPLE:
AGING HAS AN EXPONENTIAL DISTRIBUTION
Assume first no new maintenance p r o g r a m .
0(~) = ke -x~.
k > 0, 0 < ~ < ®
C(u) -- au.
a > O. u > 0
Let
AND C(u) IS LINEAR
Optimum scheduling
69
Then
L(x)
and
=
(au d u ) ke -X~ dE
L' (x) =
= ~ x + a
V
a
~k ~k
k + s
z(s)
Also
ak
=
~k
k + s - ak
X(I-~)
+
s
k+s
Hence
r(u)
ake -k(l'a)u
=
Thus X
T(x) = [ a k e -X(l-c~)u du ~0
a
=
[i
- e -k(l-a)x]
Equation (25) becomes R 4~
Xa
J
J0
[i
+ ~ ~- a -
1 -
~
e -k(l-~)u] du = K
or
R~
Ii [I - ~e-k(l-a)U]du = Kk(l-a) a
Thus R* is the unique solution to the equation R* - ~
[i - e -k(l-a)R*] - Kk(1-a)a
or
e -k(1-~)R*
= 1 +
Let Z = k(l-a)R*.
Y=I NR 2 9 : 1 - F
+
2
k(1-~)R*
Then the intersection of the two curves
Kk2(1-e) 2 a~
Kk2(1-a)
Z G
70
B.D. S|VAZLIAN
and Y = e
-Z
will yield the value of Z.
This is graphically illustrated in figure 3-
Consider now the introduction of a new maintenance program, and assume that
@i(£) = kle-kl £
,
Cl(U) = alu
k I > O, 0 < £ <
a I > O, u > 0
@2(£) = k2e-k2£
,
k 2 > O, 0 < £ < ®
C2(u) = a2u
,
a 2 > O, u > 0
aI a2 Clearly. ml(x ) = ~ii × and Lm(x) = ~22 x
1 ÷
KX2(l-~)2 me
0 )
Z k(l-a)R*
FIGURE 3.
Solving the Equation
First we determine R2*.
e
-Z
= 1 +
Kk2(l-a) 2
Z
a~
(x
Let Z 2 = k2(l - ~)R2*, then Z 2 is the
unique solution to the equation
K2k~(l - 6) 2
Z2
e-Z2 = 1 + a2~
Once Z 2, hence R2*, is determined, El* is obtained from the equation
LI(RI*)
= L2(R2* )
or
aI a2 ~ii RI" = ~22 R2*
Optimumscheduling VI.
71
CONCLUSION
In this paper we considered an intermittently operating system which is under periodic review and which is subject to random degradation.
The optimal policy to schedule a new maintenance program
was analyzed so as to minimize the long term discounted cost.
The
scheduled maintenance was based on the level of system degradation rather than time.
The functional equation approach enabled us to
incorporate such factors as the fixed cost of maintenance,
the recurring
costs, the discount factor and the probability distribution of aging in a given period.
Although we have not addressed in this paper the
problem of a comparative study between alternative maintenance program, nevertheless the developed methodology should form the basis to perform future research in this area.
REFERENCES
[i]
Bergman, B. And B. Klefsgo, "TIT Transforms and Age Replacements with Discounted Costs", Naval Research Logistics Quarterly, 30, 631-639 (1983).
[2]
Brown, J. F., J. F. Mahoney and B. D. Sivazlian, "Hysteresis Repair in Discounted Replacement Problem", liE Transactions 15, 156-165 (1983).
[3]
Nakagawa, T., "Periodic Inspection Policy with Preventive Maintenance", Naval Research Logistics Quarterly, 31, 33-34 (1984).
[4]
Nakagawa, T., "Optimal Policy of Continuous and Discrete Replacement with Minimal Repair at Failure", Naval Research Logistics Quarterly, 31, 543-550 (1984).
[5]
Nakagawa, T., "Modified Discrete Preventive Maintenance Policies", Naval Research Logistics Quarterly, 35, 707-715 (1986).
[6]
Okumoto, K. and E. A. Elsayed, "An Optimum Group Maintenance Policy", Naval Research Logistics Quarterly, 30, 667-674 (1983).
[7]
Tapiero, C. S., "Continuous Quality Production and Machine Maintenance", Naval Research Logistics Quarterly, 33, 489-499 (1986).
[8]
Sivazlian, B. D. and S. N. Iyer, "A Dyadic Age-Replacement Policy for a Periodically Inspected Equipment Item Subject to Random Deterioration", European Journal of Operational Research, 6,
315-320 (1981).