Professional Documents
Culture Documents
Author(s): K. A. K. Kobbacy
Source: The Journal of the Operational Research Society, Vol. 55, No. 2, Part Special Issue:
Intelligent Management Systems in Operations (Feb., 2004), pp. 139-146
Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society
Stable URL: http://www.jstor.org/stable/4101866
Accessed: 19-02-2018 13:09 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/4101866?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
Operational Research Society, Palgrave Macmillan Journals are collaborating with JSTOR
to digitize, preserve and extend access to The Journal of the Operational Research Society
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
Journal of the Operational Research Society (2004) 55, 139-146 ? 2004 Operational Research Society Ltd. All rights reserved. 0160-5682/04 $25.00
www.palgrave-journals.com/jors
In this paper the author reviews the development of an intelligent maintenance optimization system over the past 16
years. The paper starts with discussion of the initial motivation behind developing the system and the designs of the
early versions of a computer program to access maintenance history data and provide an analysis. The concept behind
this system was gradually developed to incorporate a rule base for the selection of a suitable model for preventive
maintenance (PM) scheduling and then to a fully developed knowledge-based system for decision support. The need to
incorporate case-based reasoning thus creating a hybrid system that can learn with use in addition to using elicited
knowledge from experts is discussed. The experience with system validation with two versions of the system is analysed.
The paper also reviews the extensive fundamental work on developing appropriate PM models that can deal with real
data patterns. Finally, the scope for future development is presented.
Journal of the Operational Research Society (2004) 55, 139-146. doi:10.1057/palgrave.jors.2601696
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
140 Journal of the Operational Research Society Vol. 55, No. 2
university and hence more academic support was at hand. Knowledge-based maintenance DSS
However, in both studies students were closely communicat-
Another MSc projected was carried out by Harper6 in the
ing with each other and the maintenance management staff
following year in order to continue the development of
as well as the academic supervisor. What follows is a short
the maintenance DSSs of previous studies.',2 This study
summary of the main features and findings of both studies.
was characterized by formalizing the concepts implemented
Wighti used the statistical package SAS to develop a
in the previous studies as well as consolidating the modelling
system to assist the management of the Maintenance
process and the statistical estimation and testing of
Division in deciding on how often, if at all, preventive
parameters. Thus, the definition of DSS was adequately
maintenance (PM) should be performed on the equipment of
discussed and the special feature of the DSS that it has
North Sea oil platforms. He identified three patterns for PM
a 'certain degree of intelligence' started to emerge. Handling
data and used a different model for each data pattern. For
of the censored data received special attention in this
components that have relatively extensive data to establish
study. Overall, the study was very successful in consoli-
lifetime distributions following PM and corrective work
dating the concept of an intelligent maintenance man-
(CO), he adopted a simplified version of a model developed
agement system that was to be further developed in the
by Watson3 (Stochastic model). If sufficient data were not
future.
available for the component, a version of Handlarski4 model
The next step in developing the proposed system was to
(Deterministic model) was used. A Geometric model was
attract funding from the research council and/or industry,
also adapted to model cases with history containing only
which was a time-consuming process. This time was used to
PMs with no failures. He applied his system to the Water
develop ideas and clarify design details of the proposed
Injection System. The study identified that a large propor-
system. A key paper7 containing these developed ideas was
tion of survival data was censored, although that was not
published. Kobbacy7 dealt with the following fundamental
accounted for in Watson's model.3 The system was able to
issues:
decide an optimal PM interval for 83% of the components.
Why there is a need for developing an intelligent
The Geometric model was applied more frequently than any
maintenance DSS?
other model (67% of items) revealing that the system may
have been over-maintained. Watson's model seems to have
(1) The main functional features expected of such a system.
more value as it was used in cases that represents 77% of the
(2) The main components and outline design of the system.
man-hours expended in maintenance, although 17% only by
(3) How to identify maintenance data patterns.
number of components.
(4) Factors affecting model selection.
The other study by Alzubaidi2 had similar objectives for
developing a simple DSS to evaluate the maintenance The configuration and main modules of the proposed
policies of a Main Oil Line Pump system with a view to
system are shown in Figure 1. The data base and model base
extending the application to other systems. The criterion was
contain the data and models used in analysis, respectively.
to maximize availability. The FORTRAN programmingThe knowledge system contains the rules that are used in
language was used to develop the DSS. Five specific data
selecting the model for a data set. The processor commu-
patterns were identified and a suitable model was identified
nicates with all system modules and with the user through
for each. These were: no failure, few failures, frequent
the user interface.
failures with or without PM, and variable PM cycle length.
The models used were similar to Wight'si study. The focus
of the study was on analysing the results of the different
models for components. Using data for all the components,
EXPERT I DATA BASE
it was possible to model the failure patterns for only 50% of
MODEL BASE
them, a figure closer to the finding of the subsequent studies
than that of Wight'. However the Geometric model was
KNOWLEDGE-
again the most frequent model used for cases with few or no SYSTEM INTERFACE
failures. This applied only for 43% of components. The
study was concluded with useful recommendations for
further research.
The most important finding from these two studies was PROCESSOR
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
KAH Kobbacy-Evolution of an intelligent maintenance optimization system 141
The salient functional features that are expected of (3) Analysis data module: analyses of formatted data,
the proposed system were articulated by Kobbacy7 and identify pattern, select model and produce results. This
marginally adjusted in subsequent studies.8'9 These features module contains the rule base and model base.
are: (4) Display results module: displays results in numerical
form.
(1) To access the history data from a maintenance data base.
(5) Plot results module: displays results graphically.
(2) To check the quality of data.
(3) To recognize characteristic data patterns.
(4) To query the user for additional information,When handling massive amounts of data for thousands of
judge-
ment and criterion. components, it is essential to ascertain its consistency and
(5) To select the most appropriate PM scheduling model validity. This feature is performed by the format data
for the decision analysis. module. Maintenance events are classed into four categories:
(6) To optimize the selected decision model, evaluate the invalid which are excluded, incomplete which are partially
current policy and propose optimal maintenance policy. used, absorbed which are merged into other events and valid
(7) To present the results of the analysis in a flexible format. events which are readily used in subsequent analysis. All
(8) To respond to user enquiries, perform 'What if?' deci- events that are not classified as valid are flagged.10
sion modelling and provide explanations of the recom- Knowledge representation in the form of procedural rules
mended decisions. is commonly used in OR applications due to its rapid
(9) To have learning capabilities. development, efficiency of execution and ease of integration.
(10) To have a user friendly Windows interface. The development of an "explicit" model selection rulebase is
Kobbacy7 detailed the factors affecting the model selec-a feature of IMOS, which distinguishes it from its DSS
tion process "rule base" including: predecessors. In the Harper6 study, for example, the code
representing the rules was embedded at various points in the
(1) History data; type, failure pattern and distribution ofFORTRAN program, which makes any changes or addition
failure data. to the rules rather cumbersome. In IMOS a list of rules for
(2) Current maintenance policies; inspection, repair, pm selecting a suitable model under certain conditions were
and replacement.
presented for the first time in maintenance literature. The
(3) Optimisation criterion; maximize availability, minimize
typical form of a rule is:8
cost, safety.
(4) Additional information including user's judgement and Rule 1: If Not matched
information.
and Number of PM events is large
and No failure events
and PM cycle length is not very variable
Then
Intelligent maintenance optimisation system
Apply Geometric Model I
By 1990, research funding was secured from major oil Matched
companies to develop a prototype Intelligent maintenance
optimisation system (IMOS). During this 2-year research The development of the rule base perhaps is the most
project, three main significant development steps were complex of all the development activities of IMOS. This is
achieved. These are:
because it requires thorough understanding of the field of
(1) Statistical checks of data. mathematical modelling in maintenance, which is vast and
extensive, and a grasp of how to relate identified data
(2) Development of an "explicit" model selection rule-base.
(3) System validation. patterns to these models. Figure 2 demonstrates this process.
The rule base protects models from being fed unsuitable
IMOS was based on the previous systems, that is,data.
the Hence, once a model is identified it is most likely that it
work of Wight,' Alzubaidi2 and Harper,6 but was restruc-
will have the appropriate data to produce results. The order
tured and written in C programming language in a PC in which rules appear in the rule base can have effect on the
environment. This is unlike all the previous systems that
modelling of a data set only if models were interchangeable.
were developed for use on a VAX mainframe computer. The
Testing of IMOS revealed that the only potential case of
structure of the system involves five modules:10
model overlap is between the deterministic and stochastic
(1) Main module: serves as a user interface, coordinates
models and the probability of such overlap in less than
0.5%. Thus, it was established that the order of the rules is
other modules and handles hardware system config-
uration. only significant for those two models and priority was given
(2) Format data modules: converts format of input to
data
the deterministic model (see Kobbacy et al8 for detailed
and checks and validate data. discussion).
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
142 Journal of the Operational Research Society Vol. 55, No. 2
1 /1 / 1\
S Identification of \\RP
Stochastic 6.6 12.8
SData Patterns
NHPP 1.6 1.6
NRP 2.3 2.3
Total Stochastic 10.5 16.4 5.6
Taxonomy of Knowledge
Models in --- - ----------------- Elicitation Geometric I 15.7 23.5 13.8
Geometric II 1.7 1.8 3.1
Literature- --- .------------------modeling
practitioners
experts & Weibull 1.7 3.7 5.3
Deterministic 1.8 1.9 3.9
No model suitable 68.6 52.7 68.3
Definition
a rule base and a case base. Table 1 shows in the last column
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
KAH Kobbacy-Evolution of an intelligent maintenance optimization system 143
New Case are more than one matching model, the system recommends
the first-ranked model from those retrieved by the Nearest
Mathematical models
Found ? >lApply Model
HIMOS and all its predecessors address one of the most
No intractable problems in maintenance optimization, namely
Retrieval scheduling PM interventions. Specifically, it is assumed a
Case Base Identify similar cases
complex engineering system with a large number of
components. Each component may be subject to PM routine
Adapta I Adaptation
tion Apply adaptation rules and repair upon failure. Details of system assumptions are
rules
discussed elsewhere.8'11 The models aim at deciding the
optimal PM interval. Up to IMOS, only five maintenance
Model
models were used (see Table 1). All these models assumed
found? No
stationary conditions, that is, no trends.
Yes Expert User
Selects a model During the development of HIMOS, special attention was
(manual mode)
given to developing models that can cover more practical
Expert
situations. Table 2 describes the optimization models that
are included in the model base of HIMOS. These models
User eess
Yes
No
were divided into stationary and non-stationary.
The stationary models in HIMOS were similar to those
Store new case Yes Adapt model Yes used in IMOS but with two main differences. The stochastic
in retrieved )
cases?
model was replaced by a generalized stationary model that
assumes general lifetime distributions following PM and CO.
No
Kobbacy and Jeon12 have discussed this model in detail
No Model suitable
including the selection of appropriate lifetime distributions
and the estimation of the renewal function. The second
Figure 3 Model selection in HIMOS in expert and automatic
difference is the introduction of a delay time model in order
modes.
to deal with situations where inspection is involved. There
are various types of delay time models that essentially
retrieved, the CBR system adapts the solution of the assume that there is delay time between the origination of
retrieved case to the needs of the current case. Thus at the faults and their propagation to cause failure. HIMOS
adaptation phase, CBR looks for prominent differences implements a basic delay time model,13 and employs the
between the current and retrieved cases and then applies objective parameter estimation method.14
rules that take those differences into account when suggest- The non-stationary models in HIMOS are applied in cases
ing a solution. HIMOS uses the Derivational Replay when trend has been found in severity/frequency of failures
Adaptation technique, which reuses the rules that generated or severity of PM. Kobbacy and Jeon15 proposes a
the original solution to produce a new solution for the generalized framework for selecting models in non-station-
current case. The adaptation rules are very similar to the ary situations. In general, non-stationary models in HIMOS
can be classed into non-homogeneous Poisson process
RBR rule base except that the critical values used are more
relaxed. (NHPP) or renewal process (RP) models. NHPP models
The next step in CBR is the evaluation of alternativeassume that inter-arrival times between failures are neither
models for solving the current case. This step can be independent nor identically distributed, although RP models
performed in two ways: for the expert and for the user. Forassume they are independent and identically distributed.
the expert, the system displays the candidate models as A practical situation which has been included in HIMOS
identified by the adaptation rules together with their featuresbut seldom mentioned in the literature is the case where
for the current case. The system also displays the identified equipment is subject to multiple PM routines. A model
cases from retrieval. The expert can then make judgements assumes a multiple PM interval, which is an integer multiple
about the model to select. For the non-expert user,of the single PM intervals and different renewal functions
the system provides recommended model(s) based on the following each type of PM.
evaluation process. This is carried out by comparing the In total, HIMOS now has 21 different models compared
results of adaptation with the results of retrieval. If there iswith five in IMOS. When excluding CBR, both IMOS and
no matching models then no model is recommended. If there HIMOS has the same success rate for modelling large-scale
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
144 Journal of the Operational Research Society Vol. 55, No. 2
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
KAH Kobbacy-Evolution of an intelligent maintenance optimization system 145
systems as shown in Table 1, although more components are Another area of development is to enhance the model base
seen to be modelled in the current case with the more exact to cover wider practical situations. Of particular interest is
stochastic models by HIMOS. When CBR is also intro-the adaptation of the powerful approaches using Cox's
duced, both the percentage of stochastic models usage andproportional hazards model (PHM) and proportional
cases where models are chosen, increases. intensities model (PIM). Kobbacy et a117 presents a most
promising approach to utilizing the PHM in developing
optimal PM schedules based on the full equipment history.
System validation
Percy et al'8 discuss the implementation of the PIM which
Both IMOS and HIMOS were validated using test cases.8,9 allows the direct incorporation of the explanatory variables
Selected test cases representing different data patterns were(covariates). An additional advantage for those two models
run through the system. The results were then comparedis that they do not require the common restrictive
with the recommendations of a panel of experts. Theassumptions used in modeling PM regarding the quality of
validation of IMOS involved the use of six data sets and a PM and CO such as renewal upon breakdown or minimal
repair. Other areas for enhancing the model base would be
panel of six experts, who also met to discuss the results. With
to extend the applications of the delay time models and to
HIMOS, eight data sets were used and a panel of five experts
was involved. In general, there was agreement between
consider PM routines applied to multi-component systems.
IMOS/HIMOS and the experts, The experts had a measureThe further development of an efficient user interface is
of disagreement in their advice as a result of making
considered to be of considerable importance as it will lead to
different assumptions in their analysis. Experts also made
better and efficient use of the system. Finally, it is important
useful suggestion for the operation of the system. Below to
is atest the operation of HIMOS in real-life situations
addressing the ability of the DSS to produce a decision
typical example of HIMOS and experts' recommendations.9
recommendation, the quality of the resulting decision, and
the actual use of HIMOS by management in practice. It is
Data set 3 also important to study the effect of extended use of the
system on its case base and modelling capabilities.
HIMOS Increase PM interval from 177 to 403 days
(CBR-RPOW model)
Expert A There is no evidence of trend. Increase PM Acknowledgments-I express my sincere appreciation to all who
interval but should not be allowed to approach participated in this research at the different stages of development.
600 days In particular, my thanks are to E Wight, H J Azubaidi, M A Harper,
N C Proudlove and J Jeon.
Expert B Unless failure has substantive safety or risk
association, PM could be extended.
Expert C Optimal PM interval is found to be 404 days
Expert D Increase PM interval References
Expert E Increase PM interval to 250 days
1 Wight E (1987). A decision support system to help calculate
planned maintenance intervals. MSc dissertation, Strathclyde
University.
2 Alzubaidi HJ (1987). Preventive maintenance of the top side
Future developments equipment of North Sea oil platforms. MSc dissertation,
Strathclyde University.
There is a wide scope for the development of HIMOS to 3 Watson C (1973). Is preventive maintenance worthwhile? In:
make it more effective to cover a wider range of maintenance Jardine AKS (ed). Maintenance, Replacement and Reliability.
situations and to extend its scope of applications. Pitman, London.
4 Handlarski J (1980). Mathematical analysis of preventive
HIMOS is essentially an intelligent optimization system
maintenance schemes. J Opl Res Soc 31: 227-237.
that aims at enhancing the efficiency of equipment main- 5 Kobbacy KAH (1988). An information oriented maintenance
tenance activities. A natural extension is to address the
management system. Presented at EURO/TIMS Joint Interna-
higher level maintenance decisions including the prioritiza-
tional Conference, Paris.
tion of equipment and the selection of the type 6of Harper MA (1988). Development of a knowledge-based main-
tenance decision support system. MSc dissertation, Strathclyde
maintenance intervention. Recent published research ad-
University.
dresses the latter problems including the use of decision grid
7 Kobbacy KAH (1992). The use of knowledge-based systems in
and other artificial intelligent techniques including fuzzy
evaluation and enhancement of maintenance routines. Int J
logic and analytical hierarchy process (AHP).16 There
Prod Econ 24: 243-248.
8 Kobbacy KAH, Proudlove NC and Murry MA (1995).
are current research plans to develop multiple hybrid
Towards an intelligent maintenance optimisation system. J
IMOS (MHIMOS) that merges the two approaches.
Opl Res Soc 46: 831-853.
That would help achieve not only efficient but also effec-
9 Kobbacy KAH and Jeon J (2000). The development of a
tive solutions to the maintenance problems of complex hybrids intelligent maintenance optimisation system (HIMOS).
engineering systems. J Opl Res Soc 52: 762-778.
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms
146 Journal of the Operational Research Society Vol. 55, No. 2
This content downloaded from 200.3.149.5 on Mon, 19 Feb 2018 13:09:27 UTC
All use subject to http://about.jstor.org/terms