Component Repairs: When To Perform And What To Do?

Diederik Lugtigheid, University of Toronto

Dragan Banjevic, University of Toronto
Andrew K. S. Jardine, University of Toronto

Keywords: repairable system, virtual age, component repairs, intensity model, proportional hazard model, proportional intensity
model.
represent the response of the component repair and operating
SUMMARY & CONCLUSIONS process, and they are those measures that reflect the
successfulness of the process, and whose desired benchmarks
In this paper, a new indicator is proposed, that reflects the form the goals of the process itself. These output variables can
state of a major component in function of the ages of the be minimum costs or downtime and/or maximum availability
(groups of) parts of which it consists. The contribution of the or reliability, among others.
ages of the parts to the state of the component is defined by
their weights. The indicator can be interpreted as the virtual
age of the component, and can therefore be used to define age- (Decisions)
reduction factors of different types of repair in a virtual age or
Repair Interval?
age-reduction process. The state indicator is used as the time What parts to replace?
Costs

scale in a proportional intensity model. In this way, the joint Controlled

When to replace comp.? System
Remaining Useful Life

impact of different repair strategies and covariates on the Repair &

Operating
Downtime
(Goals)
Health Information Availability
component failure intensity can be evaluated. This relationship Uncontrolled Operating Environment Process
Reliability
is then used to address the question of which parts to replace State of System

whenever a component comes in for repair and when to set the

Y = f(X)
preventive repair interval, in order to minimize the expected Key Process
Input Variables
Key Process
Output Variables
costs per unit time until the next repair. An example is given.

1. INTRODUCTION Figure 1 The Component Repair Process

In capital-intensive industries, maintenance costs form a The input to the process is represented by the Key Process
big part of the total operating & production costs. According Input Variables (KPIV), and they can be classified into two
to Campbell (1995), maintenance costs as a percentage of total groups: controlled and uncontrolled KPIV. Uncontrolled
value-added costs can be as high as 20-50% for mining, 15- KPIV influence the output of the process, yet cant be
25% for primary metal and 3-15% for processing and controlled. KPIV can be health information such as metal
manufacturing industries. From these figures, one would particles in the lubricating oil or vibration measurements,
expect that critical efforts be directed towards the optimization operating environment and the state of the system
of major component performance, as they account for a (component), among others. The controlled KPIV are the only
considerable part of these costs. However, in practice this is means by which the process can be steered by the user or
rarely the case. controller. In the case of non-repairable components, this is a
In contrast to most minor components and parts, major replacement strategy only (preventive or predictive), yet in the
components are repaired and not replaced upon failure and/or case of a repairable component, these are the repair interval or
at preventive/predictive intervals. Therefore, the optimization strategy, the repair itself (to what extent to repair the
of component repair/replacement strategies is more complex component) and the replacement interval or strategy (when to
in this case. Moreover, it does not only involve deciding when replace the component rather than to repair it).
to replace, but also when to repair, which in itself creates The key for optimizing the component repair and
another issue to be addressed, that is, to what extent to repair operating process is to quantify the relationship between KPIV
the component. The latter question is what makes the and KPOV. This relationship then will translate the goals or
optimization of major components complex, as there are many benchmarks of the KPOV into decisions regarding the
different ways to carry out a repair. controlled KPIV.
The component repair problem can be seen as an input- There have been several attempts to quantify this
output process, where decisions can be made regarding the relationship in the reliability literature. The majority of them
repair process itself that affect the posterior performance of use an intensity process (t ) of the counting process of
the component, as shown in Figure 1. component failures N (t ) , defined as follows
In this figure, the Key Process Output Variables (KPOV)

RAMS 2005 398 0-7803-8824-0/05/$20.00 2005 IEEE

 (t ) = lim Pr(at least one failure in [t , t + ) | H ) / consequently in an age-reduction factor. In section 3 it will be
0 t discussed how this indicator can be incorporated into the
(1)
= lim Pr( N (t + ) N (t ) = 1 | H ) / intensity function and how to estimate its parameters. Then, in
0 t section 4, the decision process will be discussed, followed by
where N (t ) represents the amount of component failures up an example in section 5. Conclusions will be addressed in
section 6.
to, but not including t . H t represents the entire history up to
time t that incorporates all available information, such as 2. DEFINTION OF THE COMPONENT STATE
repair and maintenance actions the component has undergone
so far, health information, etc. Informally speaking, (t ) can For major repairable components, a repair will be carried
be interpreted as the hazard function of the current local out when the component fails, or before failure, according to a
history starting at the last component repair, given the entire preventive or predictive strategy. In both cases, the component
history of the component (the information on all prior repairs may be repaired on-site or off-site. On-site repairs usually are
and health information) up to t . minor corrections. When the component failure is more
Intensity models for non-repairable components that severe, it must be taken out of operation (uninstalled) and
account for uncontrolled KPIV are frequently encountered in taken to the repair shop. In both cases, however, the steps to
the reliability literature. A common method is the Proportional undertake are the same: the component is disassembled,
Hazards Model (PHM) (Cox, 1972a). According to a recent evaluated and finally the repair is carried out. In the evaluation
review of case studies by the authors, the PHM has been process, it is determined which parts of the component have
applied successfully in many different industrial settings. See failed and must be replaced (assuming no redundancy), and
for example Vlok et al (2002). which non-failed parts will be replaced as well. The more
The Proportional Intensity Model (PIM) (Cox, 1972b), is parts of the component are replaced, the closer the component
more suitable for modeling repairable components and gets to its new state (good-as-new), and when only few parts
includes the PHM as a special case. The PIM allows for the are replaced, the state of the component will likely be close to
incorporation of concomitant information into the intensity its state just before the repair (bad-as-old). The state of the
process as does the PHM, but does not necessarily assume that component, therefore, reflects how good (new) or bad (old)
all component repairs bring the component back to its new or the component really is.
original condition. Besides the amount of replaced parts, their ages
Even though many papers use minimal or perfect repairs, (accumulated operating hours) should also affect the
the majority of component repairs is neither perfect nor components state. That is, replacing a part with relatively few
minimal, but brings the component back to a state somewhere accumulated operating hours will have less impact on the
in between (imperfect repair). This can also be accounted for components state than replacing the same part with more
by the PIM, by using a virtual age or age-reduction process, accumulated operating hours. In other words, the state of the
and letting the time scale in the baseline intensity be the component should be sensitive both to the amount of replaced
virtual age of the component. The virtual age or age-reduction parts as to their respective accumulated operating hours.
process (Malik, 1979 and Kijima, 1989), assumes that besides Also, components usually do not consist of identical parts.
the real age of the component, the component has a virtual age Therefore, some parts might be more important in terms of
that takes into account the effect of repair and maintenance their contribution to the components state than others, and
actions on the component. Every repair reduces the component therefore should have a higher weight within the
age by a certain age-reduction factor, where this factor is 0 or component. These weights can also be interpreted as the
1 in case of minimal and perfect repair, and between 0 and 1 fraction of the components workload absorbed by each part.
in case of imperfect repair. Some parts will absorb more of this workload and therefore
Although the PIM in combination with a virtual age have a higher weight than others. Also, the weights may
process seems an appealing method to describe the impact of represent the contribution of the parts to the overall intensity
repair activities and covariates on the component, the meaning of failure of the component.
of the age-reduction factor and how it can be related to Taking the above mentioned desired properties of the
component repairs in a more direct form has not been component state into consideration, a state indicator (here
addressed so far. At the same time, the majority of the papers called repair and maintenance indicator, or RMI) for a
that use this method assume only up to four possible repair component consisting of k parts, can be defined as,
k
activities, but in reality this number can be much bigger. Also,
RMI (t ) = j s j (t ) , (2)
the connection to what effectively happens during a repair has j =1
not been considered, i.e. a black box approach is typically
where i is the weight of part j, and s j (t ) is the accumulated
used.
In this paper some of these issues will be discussed, and operating time of part j currently in operation at moment t.
argued how they can be combined into a more flexible This indicator is monotone, increasing, and makes jumps at all
intensity model, and how this model could be useful for times when component repairs are carried out. These jumps
making optimal repair decisions. In section 2 repair and are downward, in case parts are replaced by newer ones, or
maintenance actions will be discussed, and how these actions may be upward, in case some parts are replaced by older
can be translated into a component-state indicator, and ones, although downward jumps will be more common in

RAMS 2005 399 0-7803-8824-0/05/$20.00 2005 IEEE

practice. The downward jumps will be greater when parts with age is greater than A. In this way, the weights of the
higher weights and/or more accumulated operating hours are components parts should be related to their frequencies of
replaced. If RMI (t + ) is the value of this indicator at time t failure; higher weights representing parts that more frequently
cause the component to fail.
just after repair at t , and RMI (t ) is the value just before t (in
The intensity function can be non-parametric, that is
case a repair takes place at t), the state-reduction of the repair without a specified baseline 0 ( x) , however, for ease of
can be expressed as one minus their ratio, and can be
interpreted as an age-reduction factor in a Kijima type II computation in the decision process, we prefer to use a
virtual age process (Kijima, 1989). The reduction factor parametric model. A flexible parametric model and frequently
defined here not necessarily remains constant, even though the (and successfully) applied is the model with Weibull baseline
same parts are being replaced. This is because the factor intensity or power-law,
1
depends not only on the amount of replaced parts versus the RMI (t )

(t ) = exp( ' x(t )) . (5)
amount of non-replaced parts, but also on their accumulated
operating hours.
For the estimation of the models parameters, the
In order to construct the state-indicator, information
maximum likelihood method can be used. The necessary data
is needed on the parts of which the component consists, and
to estimate these parameters are the moments of failure or
those that were replaced at each repair. For critical (read
suspension (those components that are still in operation at the
expensive) components that are being tracked carefully, this
moment of data collection, or taken out of operation not
information should be readily available through invoices of
because of failure), and information on the parts (which form
component repairs and/or maintenance records of the
part of the RMI indicator) that have been replaced in each
CMMS/ERP. However, for the application of this indicator
repair. In the next sections, the parametric intensity introduced
such detailed information is not always necessary and depends
in (5) will be used, and we will take into account only
on the objectives of the study. For example, parts that are
covariates that are constant through one history, with the
always replaced together could be combined into one new
history being the period between two subsequent repairs.
part. At the same time, in case the number of repair types of
These time-independent covariates can only change their
a component can be clearly identified and the objective is to
value at the beginning of the history.
determine when to carry out which type and not their
An example of a time-independent covariate is the
configuration itself, a part might be defined as the set of
operating condition. Whenever the component is in the repair
parts to be replaced in each repair type.
shop for repair, its destination after repair is known, and
The indicator presented here is the most
therefore also its operating condition, which is assumed to
straightforward one with the desired properties. Some more
remain fixed until the next repair.
complicated functional forms can be found in Lugtigheid et al
(2004). However, the idea is to represent the state of the
4. THE DECISION PROCESS
component in a representative, yet simple way and therefore
this functional form seems to be the first one to consider and
Using the RMI-model, the main questions to address
so it has been used in this paper.
are which parts of the component to replace in the actual
repair, and when to set the preventive repair interval for the
3. THE INTENSITY PROCESS
actual history, that is, the period from the component repair
until the next time the component is taken out of operation for
The state-indicator is a measure of real age or
repair. The problem is very complex in general, and different
workload done by the component in function of time and
approaches to the solution may be used. In this paper we will
repair activities. A natural way to incorporate this indicator
use a myopic approach, that is, we will consider the optimal
into an intensity function, is to use the state-indicator as the
repair strategy and preventive repair interval that minimizes
time scale in the intensity function, as follows,
the expected costs per unit time until the next repair, given the
(t ) = 0 ( RMI (t )) , (3) conditions at the beginning of the history. Later we will
where 0 ( x ) is a baseline intensity. address some other possible methods.
In case there are covariates available that influence the
intensity function, they could be incorporated using a PIM 4.1 Replacement Policy and Cost Structure
approach,
(t ) = 0 ( RMI (t )) exp( 'x(t )) , (4) A repair action is needed in two cases: on system
failure or whenever the system reaches its next preventive
where is a vector of regression coefficients and x(t ) the
repair interval t p . In each of these two cases certain parts of
vector of covariates. The covariates could be time-dependent
or time-independent. the component will be replaced. For simplicity, assume that
The effect of repairs on component failure intensity the replacement of part j costs C j , no matter whether the part
in this model is to reduce the components state, and therefore was in a failed state or not, which is not essential for the
indirectly to reduce the component failure intensity. If after proposed method. We also may assume an initial extra fixed
repair the component state is reduced to, say, A, then the cost of performing the repair that also may depend on whether
system will behave as if it were of age A, even though its real the replacement is after failure, or is planned, but for

RAMS 2005 400 0-7803-8824-0/05/$20.00 2005 IEEE

simplicity we will set it to zero. These two types of costs (due d = [d 1 , d 2 ,..., d k ] , where d j = 1 if the jth part should be
to the replacement of certain parts, and the extra cost) are replaced and 0 otherwise. It is assumed that the system does
incurred at the beginning of the history. However, the not have any redundancy, and that the system cannot function
additional (fixed) costs C Ff of experiencing a failure (cost with one of its parts in a failed state. This means that all failed
consequence of failure, such as downtime, damage to parts must be replaced in any case. Therefore,
environment, personnel, etc.) before the planned preventive j : f j = 1 d j = 1 , and then the decision to be made are the
shutdown should be taken into account at the end of the
values for d j such that f j = 0 .
interval (history). A (fixed) cost C Fp of performing the planned
The impact of every repair alternative is represented by
shutdown of the component at t p , if the component didnt fail +
the decrease in the RMI (t ) and consequently by the
before t p should also be taken into account and is also
decrease in the components failure intensity. Therefore, the
incurred at the end of the interval. It is assumed that cumulative failure and survival probabilities are completely
C Fp C Ff . Note that because of the myopic approach, the defined for every repair alternative. The optimal repair
costs of a possible repair of the component, regardless whether decision d * and optimal repair interval t *p can be calculated in
they are due to failure or preventive action, belong to the next
interval, and they are not accounted for in the cost two steps. First, the optimal repair interval t d and the
consequence of planned or unplanned stopping of the associated expected cost per unit time are computed for each
components operation. It is even possible that due to high repair alternative d . Then, the repair alternative and the
expected costs in the next and subsequent intervals, the associated repair interval with the lowest expected cost per
component is completely discarded, or replaced with a new unit time is selected. We will refer to this alternative and the
one of the same or different design. Also, it is possible that the
costs of repair and consequence of stopping change from preventive repair interval by d * and t *p .
history to history, which makes this myopic approach The immediate cost of repair for a given repair alternative
flexible and attractive. d that should be taken into account is the cost of the parts that
If is the set of parts that are replaced at a certain repair, are replaced, C j . With A = RMI (t + ) representing the
then the total cost C T , incurred at that interval, will be, j d

C j + C , in case of failure ending

f state of the component just after repair, the expected cost over
F
the actual history for a preventive repair interval t p is
CT = j (7)
C j + C F , in case of preventive ending
p
E[C T ] = C + S A (t p | x (t + )) C Fp + F A (t p | x(t + )) C Ff , (8)
j j d
j

The objective of the decision process is to find the set and where,
repair interval t p that minimize the expected costs per unit S A (u | x(t + )) = Pr(Tt > u | A, x(t + )) = 1 F A (u | x(t + ))
time until the next repair (over the current history). A+ u

The decision policy is called myopic because it does not = exp( ( s )ds)
take into account the consequences of the repair decision and A

interval beyond the actual history, and therefore by definition Tt = T is the time from repair at t to the first failure after t, and
is sub-optimal. The information that is available to the x(t + ) is the covariate vector just after repair at t. It is assumed
decision-maker (a tradesperson in the workshop) is the
that the covariate vector is fixed over the current history, as
information on the age of every part at the beginning of the
interval, and whether the part failed or not at the end of the discussed at the end of section 3. Then, for x(t + ) = x ,
exp( ' x )
previous history. The cost consequence of failure might S ( A + u)
depend on the parts that failed, but this would require a more S A (u | x) = 0
, (9)
complicated competing-risk model (in which the S0 ( A)
components failure intensity is the sum of the individual parts where
failure intensities) whose parameters may be difficult to u u
estimate due to lack of information. S0 (u ) = exp 0 ( s )ds = exp . (10)
0

4.2 Calculating the Optimal Policy The expected time until the next repair is
tp

The ages of the parts just before the possible repair at time E[min(T , t p )] = S A (u | x))du (11)
t are represented by the vector s(t ) = [ s1 (t ), s 2 (t ),..., s k (t )] , 0

and the ages of the parts just after the possible repair at time t As a criterion for the calculation of the optimal
by the vector s(t + ) = [ s1 (t + ), s 2 (t + ),..., s k (t + )] . The state replacement decision and the repair interval t p we will use the
vector of the component just before the repair is expected costs per unit time for the current history, defined as
f = [ f 1 , f 2 ,..., f k ] , where f j = 1 if the jth part is in failed E[CT ]
(d , t p ) = . (12)
state and 0 otherwise. The decision vector is E[min(T , t p )]

RAMS 2005 401 0-7803-8824-0/05/$20.00 2005 IEEE

 To find d * and t *p that minimize (d , t p ) , a search relatively cheap replacement cost of $1000, followed by part
3, part 2 and part 4 respectively.
algorithm can be applied in two steps. First, for every feasible
policy d , we find min (d , t p ) , by calculating (d , t p ) for Operating Condition 1 Operating Condition 2
tp
Decision tp* Costs/Hour Decision tp* Costs/Hour
[0,0,0,0] 869 3.4868 [0,0,0,0] 704 4.8882
0 < t , which gives t dp . Then, the optimal policy d * and [0,0,0,1] 1955 7.6738 [0,0,0,1] 1628 10.1533
[0,0,1,0] 1378 4.1635 [0,0,1,0] 1134 5.4924
repair interval t *p are obtained by finding min
d
(d , t pd ) . The [0,0,1,1] 2219 7.4400 [0,0,1,1] 1853 9.5339
[0,1,0,0] 1642 5.9474 [0,1,0,0] 1358 7.8947
algorithm can be made more efficient than just using a simple [0,1,0,1] 2371 9.1616 [0,1,0,1] 1987 11.9116
search, but this will not be discussed here. [0,1,1,0] 1958 5.9637 [0,1,1,0] 1629 7.6341
[0,1,1,1] 2593 8.6836 [0,1,1,1] 2174 10.9568
[1,0,0,0] 1377 2.7373 [1,0,0,0] 1141 3.4492
5. EXAMPLE [1,0,0,1] 2268 5.6922 [1,0,0,1] 1900 7.0017
[1,0,1,0] 1827 3.1569 [1,0,1,0] 1535 3.8045
[1,0,1,1] 2555 5.5851 [1,0,1,1] 2155 6.6674
As an example, we will take a component consisting [1,1,0,0] 2011 4.4247 [1,1,0,0] 1683 5.4254
of 4 parts, or groups of parts. The weights of the 4 components [1,1,0,1] 2650 6.8248 [1,1,0,1] 2229 8.2779
are 0.60, 0.10, 0.25 and 0.05 respectively. Their replacement [1,1,1,0] 2341 4.4941 [1,1,1,0] 1978 5.3382
costs are $1000, $3000, $1500 and $5000 respectively. Also, [1,1,1,1] 2907 6.5705 [1,1,1,1] 2460 7.7621

let us assume that C Ff = $8000 , C Fp = $1000 and that the Table 1 Optimal Decisions for First Repair
parameters of the power-law intensity are = 3.0 and Next, let us assume that the optimal decision is
= 3000 . We will assume only one time-independent implemented; that is, only part 1 is replaced, and let us assume
covariate, representing the operating condition, which takes that part 1 fails again after 600 hours of operation. The ages of
the value 0 for operating condition 1 and the value 1 for the parts at that time are [600, 1848, 1848, 1848]. Table 2
operating condition 2. The coefficient for this covariate is displays the results of the calculation.
assumed to be 0.50, and therefore the failure intensity In this case, the optimal repair decision is different for
associated with operating condition 2 is 1.65 times the failure operating conditions 1 and 2. For the more severe operating
intensity for operating condition 1. condition, in addition to part 1, part 3 has to be replaced, while
The component starts off with all parts having 0 for operating condition 1 only part 1 has to be replaced. This
accumulated operating hours, and the only feasible repair shows that the destination of the component (operating
policy is replacing no parts at all. In case of operating condition) has an influence on the optimal repair decision.
condition 1 the optimal repair interval is 1248 hours at a cost From this example it can be seen that the selection of
of $1.2123 per operating hour, and in case of operating which parts to replace in a repair can have great impact on the
condition 2 the optimal repair interval is 1057 hours at a cost expected costs per hour. The proposed model can identify
of $1.4321 per operating hour. what the best parts to replace are, taking into account time-
Now, let us assume that the component started off in independent covariates, in this case one, the operating
operating condition 1 and reaches the optimal repair interval condition. In the example, part 1 had the highest weight,
before failure. Then, at the time, all parts will have 1248 which indicates that the contribution to the components
operating hours, and none will be in a failed state. After repair, failure intensity of this part is significant compared to the
the component can be put back into operation in either other parts. As the cost of replacing this part is relatively
operating condition 1 or 2. Denoting the repair decision by a cheap, this part will be replaced most frequently, followed by
vector d, as introduced in section 4.2, the feasible decisions parts 3, 2 and 4 respectively.
with their results for both operating conditions can be seen in
Operating Condition 1 Operating Condition 2
Table 1. The table displays the optimal repair interval for each Decision tp* Costs/Hour Decision tp* Costs/Hour
feasible repair. It can be seen that the optimal repair decision [1,0,0,0] 1304 3.2465 [1,0,0,0] 1074 4.2145
for both operating conditions is to replace part 1 only, [1,0,0,1] 2208 6.3392 [1,0,0,1] 1844 7.9578
[1,0,1,0] 1792 3.3290 [1,0,1,0] 1500 4.0524
although the optimal repair interval for operating condition 1 [1,0,1,1] 2532 5.7427 [1,0,1,1] 2133 6.8917
is greater than that for operating condition 2. Also, the [1,1,0,0] 1952 4.8861 [1,1,0,0] 1627 6.1041
expected costs per hour are higher for operating condition 2. [1,1,0,1] 2606 7.3224 [1,1,0,1] 2188 9.0029
[1,1,1,0] 2329 4.5598 [1,1,1,0] 1966 5.4309
The table also shows that there is a considerable difference in [1,1,1,1] 2907 6.5705 [1,1,1,1] 2460 7.7621
costs/hour for different repair decisions for either operating
condition 1 or 2. Finally, it can be seen that although the Table 2 Optimal Decisions for Second Repair
optimal repair decision is the same for both operating
conditions, that this is not necessarily the case, as the ordered 6. CONCLUSIONS
list of feasible repair decisions (according to their expected
costs per hour) are different. This means that the components The proposed indicator establishes a relationship between
destination (operating condition 1 or 2) does have an influence the ages of the parts of the component, and its state. It reflects
on the decision what parts to replace in the repair. the real age of the component, and can be useful to relate
The optimal decision of replacing part 1 in both cases is repair strategies and the activities of which they consist to age-
logical, as the weight of part 1, 0.60, is high compared to its reduction factors. The state indicator can be useful by itself,

RAMS 2005 402 0-7803-8824-0/05/$20.00 2005 IEEE

for the comparison of components that appear to be similar, and repair and maintenance actions, 2004, IMA Journal
yet differ in the amount of repairs that they have undergone. of Management Mathematics, vol. 15, pp 89-110.
The use of the state indicator as the time scale in a
proportional intensity model seems very convenient, as it can BIOGRAPHIES
be used to evaluate the impact of different repair strategies on
component failure intensity, in conjunction with a covariate Diederik Lugtigheid, MSc
process. It has been shown by the example that different repair Department of Mechanical and Industrial Engineering
strategies can result in quite different expected costs per unit University of Toronto
time, and the proper repair decision therefore is a very 5 Kings College Road
important one. Toronto, Ontario, Canada, M5S 3G8
In this paper, a myopic approach is used. This approach
can be seen as the best approach given that no information on e-mail: lugtigh@mie.utoronto.ca
individual probabilities of part failure is available. In case we
would have this information available, it would be interesting Diederik Lugtigheid is a Ph.D. candidate of the Department of
to verify how close the myopic approach is to the real Mechanical and Industrial Engineering of the University of
optimal policy, using a finite or infinite horizon. Toronto. Before entering this program, he has been leading the
Reliability department of the Repair & Maintenance contract
7. ACKNOWLEDGMENTS area of Komatsu Mining Systems Chile for four years.

The authors wish to thank the Natural Sciences and Dragan Banjevic, PhD
Engineering Research Council (NSERC) of Canada, Materials Department of Mechanical and Industrial Engineering
and Manufacturing Ontario (MMO) of Canada, and the CBM University of Toronto
Consortium members for their financial support. 5 Kings College Road
Toronto, Ontario, Canada, M5S 3G8
REFERENCES
e-mail: banjev@mie.utoronto.ca
1. J.D. Campbell, Uptime: strategies for excellence in
maintenance management, 1995, p 2, Productivity Press, Dragan Banjevic is Project Director of the Condition-Based
Portland, Oregon, USA. Maintenance laboratory at the Department of Mechanical and
2. D.R. Cox, Regression Models and Life Tables, 1972a, Industrial Engineering of the University of Toronto. He
Journal of the Royal Statistical Society, vol. B34, pp 187- teaches in both the University's Department of Statistics and
220. the Department of Mechanical and Industrial Engineering.
3. D.R. Cox, The Statistical Analysis of Dependencies in
Point Processes, 1972b, in: Stochastic Point Processes, Andrew K. S. Jardine, PhD, CEng, PEng
P. A. Levis (ed), pp 55-66, Wiley, New York, USA. Department of Mechanical and Industrial Engineering
4. P.J. Vlok, J.L. Coetzee, D. Banjevic, A.K.S. Jardine, V. University of Toronto
Makis, Optimal component replacement decisions using 5 Kings College Road
vibration monitoring and the proportional-hazards Toronto, Ontario, Canada, M5S 3G8
model, 2002, Journal of the Operational Research
Society, vol. 53, pp 193-202. e-mail: jardine@mie.utoronto.ca
5. M. Kijima, Some results for repairable systems with
general repair, 1989, Journal of Applied Probability, vol. Andrew Jardine is Professor in the Department of Mechanical
26, pp 89-102. and Industrial Engineering and principal investigator in the
6. M.A.K. Malik, Reliable preventive maintenance Condition-Based Maintenance Laboratory at the University of
scheduling, 1979, AIIE Transactions, vol. R28, pp 331- Toronto, Canada. He also serves as a Senior Associate
332. Consultant to the Enterprise Asset Management Group of
7. D. Lugtigheid, D. Banjevic, A.K.S. Jardine, Modelling IBM's Business Consulting Services.
repairable system reliability with explanatory variables

RAMS 2005 403 0-7803-8824-0/05/$20.00 2005 IEEE