APPLICATIONS OF DELAy TIME THEORY TO

NIAINTENANCE PRACTICE OF COMPLEX

PLANT

W
IE

CHANSUN LEE
EV
PR

T. I. M. E. ResearchInstitute
Centre for Operational Researchand Applied Statistics
University of Salford, Salford, UK

Submitted in Partial Fulfihnent for the Degree of

Doctor of Philosophy, February 1999
 TABLE OF CONTENTS

List of Figures vi
List of Tables x
Acknowledgements xii
Abstract XiH

Chapter 1. INTRODUCTION
W

1.1 Maintenance Optin-dsation 1

IE
EV

1.2 Overview of the Thesis 4

PR

Chapter 2. LITERATURE REVIEW

2.1 Introduction 7
2.1 Single Component Maintenance Models 9
2.2.1 Basic Maintenance Model 9
2.2.2 Modified Inspection Model 15
2.3 Multi-Component Maintenance Models 18
2.3.1 Simple Inspection Model 18
2.3.2 Imperfect Maintenance Model 23
2.4 Summary of the Literature Review 29

ii
Chapter 3. THE DELAY TimE THEORY 30

3.1 Introduction 30

3.2 The Origin of Delay Time Concept 30

3-3 The Development of Delay Time Model 32

3.4 Basic Delay Time Model 38

3.5 Variants to the Basic Delay Time Model 42

3.5.1 Non-Homogeneous Defect Arrival Case 42

3.5.2 Imperfect Inspection Policy Case 43
3.5.3 A Modification of Downtime Model 46
3.6 The Parameter Estimation and Revising of the Delay Time Models 46
3.6.1 A SubjectiveParameter Estimation Method 46
3.6.2 Revising the Subjectively Estimated Parameter 48
3.6.3 The Objective Parameter Estimation Method 52
W

3.7 Discussion 60
IE
EV
PR

Chapter 4. ModeHing of Preventive Maintenance 61

4.1 Introduction 61
4.2 Nature of Preventive Maintenance Practice 63
4.3 Modelling of Preventive Maintenance 66
4.3.1 General SystemDescription for Modelling 66
4.3.2 Component I'M Model 68
4.3.3 SystemPM Model 69
4.3.4 Numerical Examples 72
4.4 Extending the Downtime Modelling 91
4.4.1 General Assumptions 91
4.4.2 Models of Inspection 92
4.4.3 Non-Negligible Downtime Model 95

iii
 4.4.4 Numerical Examples 99
4.5 Conclusions 106

Chapter 5. Parameter Estimation Options With and Without PM

Information 108

5.1 Introduction 108

5.2 Maintenance Data Generation using Simulation 109
5.3 Parameter Estimation Methods With and Without PM Information 118
5.3.1 Method A; When the Number of Failures in Each Working Day and

the Number of Defects Identified at PM Times are Available and not

Available 118
5.3.2 Method B; When Time of Each Failure and the Number of Defects
Identified at PM Times are Available and not Available 124
W

5.4 Evaluation of Estimated Parameter 129

IE
EV

5.5 Parameter Estimation Results based upon Simulation Tests 133

PR

5.5.1 Under Perfect Inspection 133

5.5.2 Under Imperfect Inspection 143
5.5 Discussion of Estimation Options and Conclusions 155

Chapter 6. A CASE STUDY OF MODELLING PLANT MAINTENANCE

FOR A TEA PRODUCTION MACHINE 158

6.1 Introduction 158

6.2 The Production Plant and Maintenance Practice 160
6.3 The Maintenance Data 161
6.4 Data Analysis 162
6.5 The Objective Parameter Estimation 168
6.5.1 General Model Assumptions 168

iv
 6.5.2 Likelihood Formulation 169

6.5.3 Resultsof the Model Fit 173

6.5.4 The PM Model and Results 176

6.6 The Subjective Parameter Estimation 179

6.6.1 The ParameterEstimation of delay time from subjective data 179

6.6.2 The PM Model and Results 184

6.7 The Snap-shot Survey and Analysis 187

6.8 Discussion and Conclusions 194

Chapter 7. CONCLUSION 196

REFERENCES 202
W
IE
EV
PR

V
 LIST OF FIGURES

Fir,r> iire
-- - Page

3.1. The delaytime concept 31

3.2. Delaytime processof basicmodel 38
3.3. Inspectionprocessof the delaytime model 40
3.4. Exampleof inspectionprocessof a defectarisingat u 44
3.5. Estimatingdelaytime parameterfor defectat a PM 47
3.6. Number of failureof prior modeland observed 49
3.7. Exampleof notationused 53
W
IE
EV

4.1. Cost curve of preventive maintenance 64

PR

4.2. The structure of a system 65

4.3. Expected total downtime per hour according to PM cycle length 75
4.4. Expected total cost per unit time 77
4.5. Sensitivity curve to systemmodel when di, = 1, d, =1 and di is considered
2
as a1 with the reduction factor q+ =0.5 78
4.6. Sensitivity curve to systemmodel when di, =2, di2 =3 and di is considered

as a3 with the reduction factor a,, =0.5 79

4.7. The expected total downtime for caseModel 1 85
4.8. The expected total cost for caseModel 1 85
4.9. The expected total downtime for caseModel 11 86
4.10. The expected total cost for caseModel 11 86
4.11. The expected total downtime for caseModel 111 87
4.12. The expected total cost for caseModel 111 87

vi
Figure Page

4.13. The expected total downtime for caseModel IV 88

4.14. ne expectedtotal cost for caseModel IV 88
4.15. Ihe expectedtotal downtime for all Models 89
4.16. The expected total cost for all Models 90

4.17. Actual failure process based on failure downtime in system 96

4.18. Relations of the number of failures and actual operating time 99

4.19. Expected downtime for basic model and revised model 103
4.20. Expected downtime for imperfect PM casemodel and revised model 104
4.21. Expected downtime for non-Hornogenous casemodel and revised model 105

5.1. Inverse-transformmethod for continuousrandomvariables 110

W

5.2. The sequenceof inter-arrival times 113

IE
EV

5.3. Flow chart for generating the data recording the number of failures per
PR

working day, or failure time data and number of defect identified at PNI 115
5.4. The failure process of a defect arising in (Ti., Tj) 119
5.5. The observed number of failures and defects over time (T, T. ) 122
,
5.6. The failure process of a fault arising in (T,,, T, ) 124
5.7. The failure process of a defect arising in (T. T. 125
-,,
5.8. Expected downtime againstinspection interval 130
5.9. Expected downtime againstinspection interval

(when true parameter values are X=0.9 and a=0.05 and estimated
0.5 and 0.06. ) 134
parameter values
5.10. Expected downtime againstinspection interval when parameter values

are% = 0.5, a=0.15 P

and rangesover 0.6,0.9,1.1 and 2.0 136

5.11. Comparison of NfethodA and NIethodB against the true number of

failures when number of PMs are 100 and delay time is Weibull distributed 140

vu
Figure Page

5.12. Comparison of Method A and Method B againstthe true number of

failures when number of PMs are 100 and delay time is a Weibull distribution 143
5.13. Comparison of Method A and Method B againstthe true number of
failures when number of PMs are 100 and delay time is exponential
distribution 146
5.14. Comparison of Method A and Method B againstthe true number of
failures when number of PMs are 100 and delay time distribution is a
Weibull 151
5.15. Comparison of Method A and Method B againstthe true number of
failures when number of PMs are 100 and delay time distribution is a
Weibull 154
W
IE

6.1. A flowchartof the productionprocessof a teaproductionline

EV

160
PR

6.2. Failure frequency of P1 line 162

6.3. Downtimes of P1 line 162
6.4. Trend analysisfor IMA failure data since the start of PM 165
6.5. Trend analysisfor IMA downtime data since the start of PNI 165
6.6. Ile effectivenessof PNI for the INIA (Number of failures) 167
6.7. 'Me effectivenessof PM for the INIA (Downtimes) 167
6.8. The failure process of a defect arising in (ti-1,ti) 170
6.9. Histogram of failures for IMA 176
6.10. Expected downtime againstPM cycle length 177
6.11. Expected downtime when r is 0.85,0.9 and 1.0 178
6.12. Failure statistic of INIA at P1 line 180
6.13. Histogram of the delay time distributions of failure types 180
6.14. Histogram of the delay time distributions of all failure types 183
6.15. Expected downtime againstPNI cycle length 186

viii
Figure Page

6.16. Expected downtime againstPM cycle length 189

6.17. Failure category for IMA machine 190

6.18. Causesof failures and downtimes for INIA machine 190

6.19. Nature of the rectification of failure for IMA machine 192

6.20. The means of prevention of failures and downtimes 192

6.21. Comparison of expected downtime when the fault rate is reduced 193

W
IE
EV
PR

ix
 UsT OF TABLES

Table Page

4.1. Resultsfor downtimeand cost for all models 91

5.1. Ile generateddatasetby the simulationprogram 113

5.2. Comparisonof historicaldataand simulationoutput 116
5.3. Parameterestimationfrom simulateddatawith PM information 135
5.4. Parameterestimationfrom simulateddatawithout PM information 135
5.5. Parameterestimationfrom simulateddatawith Weibull distribution
W
IE

when a=0.15 P
and rangesover 0.6 - 2.0 (with PM information)
EV

in perfect inspection case 137

PR

5.6. Parameterestimation from simulated data with Weibull distribution

when cc= 0.15 0

and rangesover 0.6 - 2.0 (without PM information)

in perfectinspectioncase 138

5.7. Parameterestimationfrom simulateddatawith Weibull distribution

A,
when and a=0.5, A2and, 8 rangesover 0.6 - 2.0 of casewith
PM information 142

5.8. Parameterestimationfrom simulateddatawith exponentialdelaytime

distributionandwith PM information,and r=0.1 - 0.9 144
5.9. Parameterestimationfrom simulateddatawith exponentialdelaytime
distributionin caseof no PM information when r=0.1 - 0.9 145
5.10. Parameter estimation from simulated data with Weibull delay time
distribution in caseof with PNI information when r= 0.1 - 0.9 148
5.11. Parameter estimation from simulated data with Wcibull delay time
distribution in caseof without PNI information when r=0.1 - 0.9 149

x
Table Page

5.12. Parameterestimation from simulated data with Weibull delay time

distribution, PM information, and probability r fixed 152
5.13. Parameterestimation from simulated data with Weibull delay time

distribution,no PM information,andprobability r fixed 153

6.1. The PM plans of the P1 production line 161

6.2. The number and downtime of failures of P1 line 163
6.3. The averagenumberand downtimesof failuressincethe lastPNI 166
6.4. Modelsand fitted valuesof parametersbasedupon IMA machinedata 175
6.5. The subjectively estimated number of delay times 182
6.5. Values of estimated parameter 184
W
IE
EV
PR

X1
 ACKNOWLEDGEMENTS

I would like to expressmy great thanks to my supervisor,ProfessorA. H. Christer, for

his excellent supervision, untiring interest, support, encouragement,and remarkable
help for this research.

To the staff in the Centre of OperationalResearchand Applied Statistics,visitors, and

friends, thank you for your very kind and untiring help and assistantover the period of
this study. I particularly thank Dr. W. Wang for his continued interest and assistance.
I also want thank Miss Susan Sharples and Mrs. E. Heaton for their very kind and
untiring helps.
W
IE

I am greatly indebted to my wife, Ok-Hee, and the children, Ho-Jin and Ho-Bum, as
EV

for during the four

PR

well as my parents their patience,understanding,and support past

years.

My last thanks are also due to Republic of Korea Navy for granting a scholarshipthat
enableme to undertakethis researchwork.

xii
 ABsTRAcT

This thesis is concerned with investigating and understanding the role and

consequence of different modelling options and parameter estimation options for

modelling a complex plant. As systems become more complicated and required new

technologies and methodologies, more sophisticated maintenance models and control

to
policies are need solve the maintenance problems. The initial chapter introduces the

review of previous work on a single component system and multi-component system.

Although in recent years there has been a shift in the maintenance literature from

consideration of single items to systems composed of several components, so far only a

few papers have tackled the modelling of actual multi-component plant. In the third

chapter, delay time concept and analysis technique have been presented. Of

particularly importance are parameter estimation methods, namely the objective method
W
IE

and the subjective methods. In the fourth chapter the component PM model and the
EV

PM for downtimes based PM policies are

PR

system model and costs upon various

discussed. The key options within maintenance modelling are to determine regular
Minspection periods for the system modelled as a whole, and to determine the periods
for the plant as a set of separate component models. An extension to the downtime

is
model presented for the case when the downtime due to failures within system is not

small, and impacts upon the estimate of the number of failures arising over a specified
time zone. In the following chapter, we address parameter estimation methods using

simulated data, and assess the ability of estimation techniques to capture the true
parameter values. Particular attention is paid to the problem arising during the

parameter estimating process because of the inadequate recording of PM data and

implied correlation between model parameters. Finally, a case study is presented of

maintenance modelling of production plant in a local company with view to improving

current practice. The model developed is based upon the delay time concept where
because of an absence of PM data, using the results of earlier chapters, the process

parameters and the delay time distribution were estimated from failure data only using

xm
the method of maximum likelihood. The modelling was repeated based upon

subjectiveassessments
of parameter,and considerableconsistencywith the objectively
based case obtained. For the plant study, modelling indicated the current PM
inspection program was ineffective. A snap-shotapproachis then applied to assess
other ways of reducing the downtime, and the possibility of improving the PM
inspectionpractice. This leadsto readily adaptedimprovements.

W
IE
EV
PR

xiv
 Chapter 1

Introduction

1.1 The Optimisation of Maintenance

During the last decade the importance of the maintenancefunction to technical

systemshas been increasingly realised. Until three or four decades ago, maintenance

was simply regardedas an unavoidableand difficult to control part of production. As

W

become and required new technologies and

IE

systems more complicated, automated

EV

methodologies, more sophisticated maintenance models and control policies are

PR

needed to solve the maintenance problems. This thesis is concerned with

investigating and understanding the role and consequenceof different modelling

options for modelling a complex plant.

One of the aims of research in maintenance management is to provide decision-

making tools for the maintenance manager. Operations Research / Management

Science techniques are among the tools which can help maintenance decision making.

They allow subjective decisions to be replaced by objective decisions, taking into

account accurately formulated objective functions and a complex set of constraints.

OR / MS techniques have long been used and appreciated in areas like production and
inventory management (Pintelon and Gelders, 1992).

Preventive maintenanceoptimisation models are mathematicalmodels which aim

to balancequantitatively the costs and benefits of preventive maintenancein order to
determine optimum policies. The models can deal with many aspects of the

maintenanceprocess,such as determining the right type of maintenance,the optimum

I
frequencyof execution,the best way of planning and scheduling,the best combination

of maintenance activities across plants, and the optimum design of maintenance

facilities. Thus in order to fulfil the maintenance objective, the industrial

organization needs management skills to integrate people, policies, equipment and

practice. It also needs adequateengineering and technological skills in order to

provide the best repair

possible preventive maintenance, and overhaul for increasingly

automaticproduction equipment.

in
There are many ways which maintenanceoptimisation models can applied be in

One of the main problems associated with maintenance is the

maintenanceareas.
determinationof inspection frequencies. Inspection is one of the key functions within
PM. The basic purpose of an inspection is to reveal the true state or condition of a

system. As a result of an inspection, a repair or replacementaction may be performed

to prevent further deterioration or failures of the system. Thus the development of

models for inspection decisions is mainly centred around determining the optimal
W

inspection interval which optimises the criteria of interest such as total expectedcost
IE
EV

or downtime per unit time.

PR

Of the first maintenancemodels which appearedin the sixties, many dealt with the

problem of finding optimal inspection policies for systems which are subject to
failures. In the beginning, the maintenancemodels were relatively simple in that they

considereda single component only, McCall (1965), Pierskalla and Voelker (1976),

Sherif and Smith (1981), and Valdez-flores and Feldman (1989). However, in recent

years there has been a shift in the reliability and maintenance literature from the

consideration of single units to modelling systems composed of several units, see

Thomas (1986) and Cho and Parlar (1991). Some maintenancemodels are concerned

with optimal maintenancepolicies for a system consisting of several units within a

machine, or many pieces of equipment, which may or may not dependon each other
in terms of economic or stochasticdependency. Stochastic dependenceimplies that

eachcomponent'scondition probability dependson the other components'. Economic

dependencemeansthat set-up costs can possibly be saved when several components

are jointly maintained instead of separately,and vice versa. Since combining such

2
forms of dependencemakes the models very complicated, most maintenancemodels

consideronly one of thesedependencies.

In practice, problems arise becausethere may be many relationships which exist

between componentsto be maintained. Modelling these relations directly and non-
large models, which are difficult to analyseas they suffer from the
selectively yields
dimensionality. Dekker (1995) recommendsa decomposition approach for
curse of
this problem. In such an approach one applies simple models for individual

input in a comprehensive model. Other

components and uses the outcomes as
in the implementation of maintenance policies for
problems are encountered
individual components. It can be profitable to combine maintenance activities,
thereby saving common preparationwork. Therefore, for modelling a complex plant,
the systemmodelled as a whole and as a set of componentmodels will be contrasted.

In maintenancemodelling, most of maintenancemodels assumethe availability of

data. If the maintenancerecords of failures and recorded findings at maintenance
W
IE

interventions, such as inspections,are available and sufficient in quantity and quality,

EV
PR

the delay time distribution can be estimatedby the statistical method, called objective
method, see Baker and Wang (1992,1993), Christer and Wang (1995), and Christer et
data is not always present. Since in
al (1995). However enoughsuitable and correct
delay time modelling it is essentialto obtain the estimatesof the delay time and initial

distribution, to overcomethe problem when there is a lack of data, a method for

point
estimating delay time parameterscalled subjective estimation, suggested by Christer

(1982), has been developedusing opinions of experts for estimation of the delay time

parameters. For obtaining an estimateof delay time f(h), we use a revised parameter
estimation method (seeWang, 1997) which is based upon the analysis of historic data

of failures and the delay time concept where the distribution of the delay time was

estimated from the subjective data obtained from the expert.

Most production systemshave numerousfailure types or modes. Therefore, when

one is optimising all maintenancefor a given system, there is a natural tendency to

consider every failure type or mode. Here we use the data which recorded by

3
machine operator as a failure mode. The distribution of the delay time has been

estimatedfrom the subjective data for eachfailure mode.

1.2 Overview of the Thesis

In Chapter 2 we review the existing literature on maintenancemodels relating to

the developmentsand applications of modelling. We distinguish between a single

component maintenancemodel and a multi-component maintenancemodel. These

models deal with the problem of finding optimal inspection policies for systemswhich
are subject to failures. Maintenance models for single components can be useful for

modelling the maintenanceof individual componentsthat are part of more complex

W

systems. Due partly to improvements in analytic techniques,the work on policies for

IE
EV

maintenanceand replacementof deteriorating componentshas recently been extended

PR

to consider systems comprised of several components which are dependent on or

independentof one another.

In Chapter 3 the concept and developments of the delay time modelling are

presented. A technique called delay time analysis has been initially developed for
modelling inspection policies for industrial inspection maintenance when the
equipmentis regularly inspected. In maintenancemodelling the successfuluse of the
delay time concept dependsupon how well the underlying delay time distribution can
be estimatedfrom available information sources. One of the key issuesin the delay
time modelling is the estimation of the delay time parameterswhich are usually the
rate of occurrenceof defects,the distribution of underlying delay time h of a defect,
and the probability of identifying and removing a defect at PM. Two basic
approaches to solve the associated estimation problems, namely subjective and
objective methods, have been presented using the information obtainable from
maintenanceengineerswho repair the machine.

4
 Chapter 4 describesan investigation of PM modelling which is concern with the
downtime and cost aspectsof various maintenancepolicies. The component PM

model and systemPM model are presentedin the casewhere downtime due to failure
is relatively small. The value of the approach is that it looks at the maintenance

scheduleproblem in its entirety, with account taken of the specification of equipment,

inspection plan and the nature of the plant. By developing two models, that is the

system model and a collection of sub-systems models, the effectiveness of

maintenancescheduling for a system can be analysed. We revise the downtime

to
model embrace the case when the downtime due to failures of system is not very
small. This can changethe failure processover the PM period (0,7). The actual

operating time over the calendar time (0,7) of the system is obtained, and the
downtime models are then extendedto be basedupon the actual operatingtime.

The parameterestimation options with and without PM information are presented

in Chapter 5. To investigate and verify parameterestimation methods a simulation
W

study is undertaken. We test parameterestimation methods in terms of the ability to

IE
EV

recapture known parameters,and use simulated data to check the consequencesof

PR

different volumes and types of data upon the accuracy of parameter estimates for

maintenancemodels. The given maintenancerecord data includes the failures times,

or number of failures per day, and the number of defects identified at PM. The
importance and value of having data on PM activities and inspection results is
highlighted.

In Chapter 6 we present two modelling studies of preventive maintenance(PM)

policy of production plant in a local company with a view to improving current

practice. An objective data based model is developed based upon the delay time

concept where, becauseof an absenceof PM data, the processparametersand the

delay time distribution were estimated from failure data only using the maximum
likelihood. Particular attention is paid to the problem arising during the parameter

estimating processbecauseof the inadequaterecording of PM data and the implied

correlation between model parameters. The case of data deficiency explored in the
study is important becauseit is a relatively generalsituation in practice. A subjective
data basedmethod carried out at the same company and the same plant parallels the

5
objective data based study. The two studies of the same problem provide a rare

opportunity to compare the model formats and parameter values resulting from the
two approachesand to consider the degree of consistency between the subsequent
decision consequencesof the two methods. The consistency is reassuring. In

addition, to reducethe further downtime in this casestudy, a snap-shottype of'surveys

techniqueis undertakenand presented.

We conclude the thesis in Chapter7 with discussionsand final remarks for further

research.

W
IE
EV
PR

6
 Chapter 2

LITERATURE REVIEW

2.1 Introduction

Opportunities for the application of Operational Researchmodelling to problems

of plant maintenancemanagementare numerous. Whenever a management decision

arises,the potential for OR modelling exists. As systems become more complicated

and required new technologies and methodologies, more sophisticatedmaintenance

W

Thus, it
IE

to
models and control policies are needed solve the maintenanceproblems.
EV

is worth noting that one expectsattention given to the areaof maintenancemodelling,

PR

in general, and industrial maintenancemodelling in particular, to grow over the next

decade.

Over the last few decades,numerouspapershave appearedin the literature which

deal with the problem of finding optimal inspection policies for systemswhich are

subject to failures. This phenomenonis indicated in various surveysof maintenance

models by McCall (1965), Pierskalla and Voelker (1976), Sherif and Smith (1981),

Sherif (1982), Thomas (1986), KraIj and Petrovic (1988), Valdez-flores and Feldman
(1989), Cho and Parlar (1991), Thomas atal. (1991) and Scarf (1997). Valdez-Flores

and Feldman (1989) used a classification scheme that catagorizes single-unit

maintenancemodels into four topical categories: inspection models, minimal repair
models,shock models and other replacementmodels.

Although a systemmay consistsof severalcomponents,it is sometimespractical to

consider the system as a single unit that behaves in such a way that individual

7
 do
components not directly affect the reliability of the system. Another important

to
reason consider single unit systems is because in practice there are many instances
in which it is difficult to obtain reliability data for smaller components;whereasdata
for the stochasticbehaviour of the entire systemis available or easierto obtain. Cho

and Parlar (1991) also used classification scheme used by Barlow and Proschan

(1965) to some extent as well as by MaCall (1965). They surveyedand categorized

multi-unit maintenancemodels into the five topical categories:machine interferenceI

repair models, group / block / canibalization / opportunistic maintenance models,

inventory and maintenancemodels, other maintenanceand replacementmodels, and
inspection / maintenance (preparedness maintenance) models. Multi-unit

maintenancemodels are concernedwith optimal maintenancepolicies for a system

consisting of several units of machines or many pieces of equipment, which may or

may not depend (i.
on each other e., economic / stochastic dependency). Stochastic
dependencemeansthat each component'scondition probability dependson the other

components'. whereas economic dependence implies an opportunity for a group

replacement of several components provided that a joint replacement of several

W
IE

components costs less than separate replacements of the individual components

EV

(Sethi, 1977). If all units in the system are economically and stochastically
PR

independentof one another, a maintenancepolicy for the single unit models may be

applied to the multi-unit maintenanceproblems. On the other hand, if any units in the

systemare economically or stochasticallydependentupon each other, then an optimal

decision on the repair or replacementof one unit is not necessarilyoptimal for the

whole system. A decision must be made to improve the whole system, rather than

any subsystem. Thus, in recent years there has been a shift in the reliability and
maintenanceliterature from consideration of single units to systems composed of
severalunits (Thomas, 1986).

In this Chapter, single component and multi-component systems maintenance

models for inspection will be presented.

8
2.2 Single Component Maintenance Models

2.2.1 Basic Maintenance Model

Maintenance can be defined as the combination of all technical and associated

administrative actionsintendedto retain an item or systemin, or restoreit to, a statein

which it can perform its required function (British StandardsBS 3811,1984). Most
plant including machinery, electronic system, components, vehicles, and buildings
wear out and fail. The time at which equipment fails however is not known in

advance,and it is likely to fail when in operation. These operating failures can be

quite expensive not only in repairing or replacing the item, but also becauseof the
disruption and delay to the operation of the system. Thus such items are often subject

to a replacement,maintenance,or inspection policy. Preventive maintenancewas

W

advocated as a means to reduce failures, unplanned downtime and even operating

IE
EV

cost.
PR

Maintenanceinvolves planned and unplannedactions carried out to retain a system

in or restore it to an acceptablecondition. Maintenancemodels usually assumethat
the condition of the systemis completely unknown unless an inspection is performed.
Every inspection is normally assumedto be perfect in the sensethat it revealsthe true

condition of the system without error. In the absenceof repairs or replacement

actions, the system evolves as a stochastic process. Optimal maintenancepolicies

aim to minimize the total expected cost or downtime per unit time for the most
effective use of systems.

Since the 1965 survey on maintenanceby McCall (1965), different authors have

produced many interesting and significant results for variations of inspection models.
The different models developed dependon the assumptionsmade regarding the time
horizon, the amount of information available, the nature of the
cost functions, the
objective of the models, the system's constraints, etc. However, many inspection

9
 models follow the premise of the basic model presentedby Barlow and Hunter (1960)
(seealso Barlow et al (1963)). The model assumptionsare generally as follows

(a) the stateof the systemis known only by inspection.

(b) inspectionsdo not degradethe systemand take negligible time.
(c) inspectionsare perfect in that any failure within the systemwill be identified.
(d) inspection ceasesupon discovery of failure and repair or replacement takes

place which returns the systemto as good as new.

(e) the systemcannot fail while being inspected.
(f) eachinspection cost is c, per unit time.
(g) the cost of leaving an undetectedfailure is c2 per unit time.
(h) the failure time distribution, F(t) of systemis known.

Hencethe total cost per inspection cycle is given by

W
IE

C(t, x) = c,n+ C2 (Xn - t)9

,(2.1)
EV
PR

where t is the time to failure, x=( xi, x2,... ) is the sequenceof inspection times
with x, < x2 < x3 .... and n is the inspection which detectsthe failure occurring at
time t, that is, x,, <t<x.. The optimal inspection policy x* is the one that
-,
minimizes the expectedtotal costs due to inspection and the cost due to leaving the
systemin the failed stateuntil it is detected, E[C(t, x )], where

E[C(t, x)] =1S. "'+'[c,(n + 1) + C2 (Xn+I

(t)Idt, (2.2)
x» - t)f
n--0

and satisfiesthe recurrencerelation

F(x)-F(x. c,
-X,. = f (x)
-1) n (2.3)
c2,

10

Reproduced with permission of copyright owner. Further reproduction prohibited without permission.