th

11 International Conference on Nuclear Engineering Tokyo, JAPAN, April 20-23, 2003 ICONE11- 36359

KNOWLEDGE ENGINEERING AS A SUPPORT FOR DECISION MAKING IN PLANT OPERATION AND MAINTENANCE

Philippe HAÏK EDF – R&D Dept. O. P. P. 6, Quai Watier 78401 Chatou Cedex Phone: (33/0) 1 30 87 75 86 Fax: (33/0) 1 30 87 75 89 e-mail: [email protected]

Sylvain MAHÉ EDF – R&D Dept. O. P. P. 6, Quai Watier 78401 Chatou Cedex Phone: (33/0) 1 30 87 75 54 Fax: (33/0) 1 30 87 75 89 e-mail: [email protected]

Benoït RICARD EDF – R&D Dept. O. P. P. 6, Quai Watier 78401 Chatou Cedex Phone: (33/0) 1 30 87 75 59 Fax: (33/0) 1 30 87 75 89 e-mail: [email protected]

Keywords: knowledge engineering, knowledge management, knowledge modeling, rapid prototyping, automated decision assistance, telecooperation, energy, diagnosis, maintenance, information systems.

ABSTRACT Electricité de France (EDF) considers the preservation of expert know-how as a major issue, especially in the field of nuclear power plant operation and maintenance. This can be related to the growing interest corporate knowledge management has encountered during the past years, but is basically due to the key role played by expert knowledge and know-how in plant operation and maintenance tasks. Furthermore, in the past fifteen years a significant amount of work has been devoted to the development of knowledge-based systems for the diagnosis and maintenance support of plant components. During this period, several various needs have emerged from the field of plant operation and maintenance. As most of

these needs appeared to be based on the exploitation of technical knowledge and know-how achieved by EDF operational experts during the past years, we focused our work on knowledge modeling and reuse. This paper presents a general knowledge based system design approach derived from our past experience in the development of knowledge-based solutions for troubleshooting and maintenance support that meet plant operators’ needs. Our goal is to provide them with automated operation or maintenance assistance. This approach is targeted at sharing and using expertise on technical fields. It relies on the use of models as an efficient supporting framework for knowledge acquisition and knowledge reuse (via the implementation of operational tools that

1

Copyright © 2003 by JSME

•

make knowledge available to others and facilitate knowledge maintenance and consistency). This provided assistance relies on services that have been applied for diagnosis and predictive maintenance for such components as reactor coolant pumps or turbine-generators. Services range from reference-knowledge browsing to health assessment, diagnosis or maintenance schedule analysis ; these are available as web intranet services and are implemented using the eKCM software environment (eKnowledge for Competitive Maintenance) that has been jointly developed by EDF and Sharing Knowledge a software company. This article highlights the rationales that led us to use knowledge engineering - and distributed knowledge based systems - to support plant operation and maintenance ; it will present the modeling and application design methodology, the unified framework and the associated software framework we developed to formalize, express and process expert knowledge and know-how ; it focuses on some significant operational examples such as condition-based maintenance, life cycle management or plant decommissioning. This paper will also describe the way our solutions can be integrated in operational activities and the associated benefits. The article will expose the way we conduct such knowledge management projects to efficiently manage expert knowledge and to develop adapted operational tools.

• •

maintaining expertise along large periods of time : employee versatility, retirement of experienced people will occur long before the need for their expertise has disappeared… gather and share feed-back about rare events from different plants and derive generic lessons therefrom ; make part of expert knowledge usable from local plants by less skilled employees.

In this article, we first present the reasons that led us to use knowledge engineering techniques and tools to support plant operation and maintenance, then we will introduce what we consider to be the focal point of our approach, our knowledge modeling methodology and the associated application design methodology. The next part describes eKCM, the software environment we developed to support our methodology and the last one gives a few operational examples. 2. MANAGING KNOWLEDGE : A KEY ISSUE FOR PLANT OPERATION AND MAINTENANCE 2.1 What is the need for managing knowledge ? 2.1.1 Managing knowledge in a highly technical domain Nuclear plant operation and maintenance is a highly technical domain, and thus relies on the use of large amount of various technical knowledge ; more over knowledge preservation is required in order to avoid the serious danger that a loss of mastery of technical knowledge would represent. Variety is one of the main characteristic of the knowledge required for nuclear plant operation and maintenance in several ways. First of all, nuclear plant operation and maintenance implies various technical fields such as nuclear, mechanical, electrical or hydraulics. In the following parts of this paper, we will use the expression « nuclear knowledge » to refer to all the knowledge taken from these various domains and required for nuclear power plant design, operation, maintenance and deconstruction. Experts, engineers and technicians working in this context are often specialists of one of these fields, and there is a need to point out the links between concepts used in each field, in order to help specialists interact with a more deep understanding of one another’ concern of what is implied between them. Specialists are to often organized more specifically on one field applied on a range of systems or on one system.

1. INTRODUCTION Our R&D department has quite a wide experience in the development of knowledge-based solutions for troubleshooting and maintenance support (Porcheron et al., 1994 ; Porcheron et al., April 1997). This led us to focus on the main component of such solutions, namely « knowledge ». Specifically, we used our past experience to implement a general framework which we now use to design new application-specific knowledge-based development. We also developed a software architecture to help us turn those concepts into an operable form. Knowledge management activities are quite varied. While some focus on theoretical knowledge or the sharing of informal knowledge, we decided to focus our work on trying to derive operational applications and services from expert know-how, which is usually obtained from experience. Typical targets for our knowledge management activity include : • cross-domain transfer, for instance making possible to access relevant design knowledge when operating or maintaining a plant, using operation feed-back experience to assist deconstruction decision…

2

Copyright © 2003 by JSME

components of various natures. This is the case of the nuclear sector in which safety is a major concern to address. Safety requires in-depth mastering of sometimes complex processes. Contrary to most industrial activities that generally focus on a few key aspects, analyses in the nuclear sector are more exhaustively performed and this implies even more complexity. According to Ashby’s law (1956) there is also a greater need for knowledge (as increasing with the level of variety and complexity to master increasing)

Secondly, nuclear plant operation and maintenance imply various internal and external actors from several departments and from suppliers, partners and authorities, and thus there is also a need for sharing and exchanging representations and various pieces of information with their context. Thirdly, there is a need to help knowledge transfer from the place where it is possessed to the place where it is needed, that may occur on several steps of the nuclear plant life cycle. For example, some events occurring during operation can be useful for the deconstruction phase if they induce some area specific contamination.

2.2 Knowledge engineering for managing knowledge We have just pointed out some needs for managing knowledge for nuclear activities. Knowledge Engineering participates in responding to these needs in two directions. The first added-value is the more visible one, it corresponds to the explicit and tangible output of knowledge engineering activities, that is, some guides or decision supporting tools (cf. §2.2.1). The other added-value of knowledge engineering activities is less tangible, more indirect (i.e. not directly associated to deliverables) and concerns the progress of employees’ technical mastery (cf. §2.2.2.).

2.1.2 Managing knowledge throughout employees movement In any domain implying several actors (including the domain of nuclear plant operation and maintenance), mere job changes (such as retirement, functional mobility inside or outside the organization, reorganization etc.) can induce knowledge losses. This is the case when people leave a given activity context and when their mind is the only container of some pieces of knowledge that are required for a given activity, and this is also the case when people who knew where to look for required knowledge are no more implied in the considered activity. In a preventive and reactive way, there is a need to enhance knowledge exchanges between people, which leads to let pieces of knowledge be less sensible to people departures. Indeed, pieces of knowledge are shared by several employees instead of being possessed by only one person. A complementary way to preserve knowledge against departures is to organize knowledge transfers between people with apprenticeship, training, technical knowledge networks or any other way that leads to explicit or implicit knowledge exchange.

2.2.1 Direct deliverables of Knowledge Engineering tasks Knowledge engineering activities are usually performed in order to design and construct some reference guides, computer aided tools to support some steps of diagnosis and decisions processes, and training materials, etc. These products are based on identifying concepts, links between concepts, typologies of concepts and links between them in order to build some domain models. Models are useful to organize existing knowledge, its acquisition and its validation; they also allow to make it evolve and to structure its restitution in documents and tools.

2.1.3 Managing knowledge for competitive improvement In any domain, the purpose of managing knowledge is recognized as a key issue for improvement. Nuclear plant operation and maintenance are concerned with this general issue, as taking place in a competitive and worldwide environment, that shares some characteristics with other activity sectors. Managing knowledge consists in enhancing the use of knowledge pieces in performing activities, as far as knowledge can bring some advantage in reaching activity objectives.

2.2.2 Indirect advantage of Knowledge Engineering tasks regarding management of knowledge Building model is performed with experts, engineers and technicians who take an active part in a prototyping and refining approach. In parallel of building models more and more close to the reality of practice, experts, engineers and technicians build some new knowledge in their mind while analyzing the reasons that led to conceiving new versions of models. In fact, this leads to questions, deeper analysis, drawing links to context, etc. Also in parallel of building some direct deliverables, one important value of knowledge engineering is to give a way for professionals to

2.1.4 Managing knowledge under safety requirements Activity objectives are often more or less economical but may also have some other

3

Copyright © 2003 by JSME

allowing a step-by-step acquisition process (listing faults and their characteristics, then performing device decomposition, etc.) • sharing is facilitated by model structures that can be used as a « reading grids » to help newcomers master the way the various concepts are used within the knowledge domain, then interpret them • reuse is made possible by model genericity : as the model does not include any information related to a given use of the knowledge, and as it tries to be based on « deep » knowledge about the domain, its contents can be used in many ways for different applications • knowledge maintenance is allowed by adaptating or completing model structures : when knowledge evolves or appears to be incorrectly captured, it is relatively easy to define the structures to add, remove or modify and thus have the knowledge base evolve without having to redesign the whole base. Several models might be necessary to deal with different aspects of a domain (e.g. one model for the causal representation of faults, one model to describe the structure of the equipment, one model to describe the sequences of maintenance tasks, etc.). A trade-off often has to be found between model richness and complexity.

progress on their technical mastery, and especially on practice-based knowledge wider view on their field that is more difficult to meet with courses and continuing education. 3. KNOWLEDGE ENGINEERING AS A WAY TO SUPPORT PLANT OPERATION AND MAINTENANCE 3.1 Using knowledge modeling to preserve, share, reuse and maintain expert knowledge Most of the time, experts use their technical knowledge and know-how to perform the important operations they are responsible for. They also often have to deal with more trivial but time-consuming problems, for which their skill is not fully necessary but for the solution of which their knowledge is a basic element. Thus it is essential for a company such as EDF to be able to (1) preserve and (2) make this mainly tacit knowledge available beyond the limited set of current experts. As a consequence we try to achieve several things : • acquire knowledge, that is help experts to express their knowledge and put it down in a readable and manageable form (in other words, we try to explicit the tacit knowledge experts use while reasoning) • share it among various users • reuse it for instance in knowledge based applications • maintain it, have it evolve over time and ensure its consistency. To do this, our experience and the analysis of existing research (Hamscher et al., 1992 ; Dieng, 1990) led us to focus on our « knowledge modeling » activity which became the core of our work. By « modeling », we mean defining a formal representation of the field of knowledge considered. This formal representation relies on typing precisely the various categories of knowledge elements we deal with and describing the relations between them. At this stage we do not include any information related to how this knowledge is used to solve any problem. This description of knowledge tries to rely mainly on principles (e.g. explicitly representing causality, material composition, physics…) and to be at the right level of « depth ». This « strongly typed », purely descriptive representation of knowledge helps meeting our initial goals : • knowledge acquisition is supported by model structures that serve to organize acquisition and to improve completeness and consistency by

3.2 Methodological issues in knowledge modelling and application design 3.2.1 A unified knowledge

framework

for

expert

Conceptual choices Over the past few years, we designed a methodology and an accompanying unified framework for formalization and operationalization of expert knowledge [Porcheron et al., 1999]. Elaborating this methodology was based on our past experience in diagnosis expert systems and on a study of the requirements for efficient preservation and sharing of expert knowledge, together with research on how to enhance the reusability of proven techniques in order to improve consistency and reduce development costs. The resulting methodology relies on the definition (or reuse) of models which enforces knowledge structuring and allows rapid prototyping in order to have an efficient support for knowledge acquisition and validation. Our methodology and the associated framework are based on the following choices :

4

Copyright © 2003 by JSME

•

separation between knowledge modeling (domain description) and knowledge operationalization (knowledge processing) : models allow domain knowledge expression and services use the capitalized knowledge to produce specific results (diagnosis, decisions, etc.) ; formal modeling based on an entity/relation description of the domain to allow knowledge expression ; problem decomposition based on a multiple model description as a way to reduce model complexity, to facilitate model re-use and to allow model understanding and validation. All models related to a specific topic (a component such as the turbine generator or a field of application such as monitoring) are grouped together in a knowledge base ; every model included in a given knowledge base gives a specific point of view on the topic (for example, the functional decomposition or the known dysfunctions of the studied component) ; conception of generic models as domain ontologies to allow model re-use ; re-use of validated models and services.

from our past experience that includes the following elementary tasks : • design (or more or less direct reuse) of models, based on a first outline of the knowledge given by the experts • « filling in » of the model, i.e. using the model as a support for knowledge acquisition, with all relevant knowledge as expressed by experts • specification of the relevant algorithms when dedicated tools are required by end-users • implementation of the model (and algorithms whenever relevant) in a tool in order to validate the model and the represented knowledge and making it usable by end-users.

3.2.2 Using our knowledge modeling methodology for designing new knowledge management applications If we try to summarize what we do in the various projects we are taking part in, we can outline the following points : • gather some know-how or expertise in a well-structured, readable, accessible, reusable form ; • provide end-users (e.g. power plant operators) with operational tools using the previously mentioned know-how to assist them in their work.

A key point of our methodology is the fundamental role played by the modeling activity, namely the design (or reuse) of generic structures and relations that will be used to acquire and represent knowledge.

• •

• •

A more detailed description of the way these tasks are combined in our modeling and application design process can be found in Haïk (2002). Our methodology implies the cooperation of two skills (knowledge engineering and tool specification and development), which are neither independent nor separate. On the contrary, efficiency is obtained through their combined approach, especially when starting a project.

Another important element of our methodology is the use of « feed-back » processes between these different tasks. This means that developing a model, acquiring knowledge and validating these is essentially a cyclic process. The easiness with which a sample of acquired knowledge will fit into the model reveals the level of adequacy of the latter and eventually leads to its partial redesign.

In those projects, besides knowledge engineers, participants include experts, end-users and project managers : • experts need to endorse the proposed models in order to express relevant knowledge within the proposed framework. • end-users need to validate that the proposed means for accessing and using the knowledge are well suited to their needs. • project management needs to assess that knowledge acquisition is performed efficiently and that the final goals of the project are to be met on schedule.

Another particularity of our approach is that even when the project does not require the development of operational tools, our method includes the computer implementation of data structures representing the model and the modeled knowledge. This « prototyping » of a very basic knowledge storage and retrieval system is a way to facilitate model and knowledge-base construction and validation. Besides specification validation and feasibility demonstration when a software product is expected, the prototyping of computer structures enables the encoding and the representation of modeled knowledge. Therefore, it yields the following advantages : • it makes the model a graphical one, which helps understanding and validating it ;

To meet these expectations, and to proceed efficiently with the expected developments, we have come up with a development methodology derived

5

Copyright © 2003 by JSME

• it enables tracing and managing the acquired knowledge through time ; • it helps experts master and endorse the model, as they can have a good understanding of it of it and sometimes even « play » with it (conversely, static models on paper are often abstract and hard to get into when experts are not used to structural modeling) ; • it enforces the strictness of the model, as computer structures cannot accommodate a vague or inconsistent description ; • it constitutes a repository in which the knowledge base can be stored ; • it enables the easy restitution of knowledge in a manageable and readable form ; • it gives some visibility on the volume and interest of gathered knowledge ; • it can concretely show end-users and project managers what knowledge management can bring to their projects.

Thanks to its flexibility and to its modularity, eKCM constitutes a relevant tool for developing new knowledge management applications (building solutions meeting emerging needs or new types of knowledge). In fact, as it is the case for models and services, every eKCM software component can be re-used (directly or after modifications) to build a new software tool. eKCM includes the following functions/capabilities : • a generic representation structures which allows to formalize operational knowledge (we call them « models ») ; • « services » that use the capitalized knowledge in order to provide operational assistance ; • access to non-formalized knowledge (photos, documents…) or external databases. Actual services can range from basic knowledge browsing and visualization to automated document generation or more specific algorithms (e.g. automated diagnosis, cycle detection and analysis tool for Bayesian networks).

A precise definition of the underlying models shared by knowledge engineers and experts as well as the interaction capabilities of the software implementation enable a true cooperative knowledge acquisition process of implicit expert know-how.

eKCM-based applications are made available from a standard Web server (i.e. accessible via an Intranet or an Internet).

Whenever possible, we endeavor to define « generic » models that offer good reuse perspectives.

4.2 Architecture choices

4. eKCM, A SOFTWARE TOOL TO SUPPORT OUR METHODOLOGY AND TO DEVELOP OPERATIONAL SOFTWARE TOOLS

and

implementation

Architecture The eKCM tool runs under an intranet server (a standard HTTP commercial server compatible with Java servlets) to which users can connect via a web browser (Netscape, Internet Explorer…). The end-users can navigate through the eKCM available models and services like they do with standard web services ; so they can use the well-known exchange functions (copy, cut and paste functions, etc.) to export knowledge or results producted by the services to their standard word-processing, database or spreadsheet softwares. For applications that need highly interactive user interface, the « light » client (the web browser) can be replaced by an autonomous client application written in JAVA which allow complex and dynamic interface functions ; the global architecture of the eKCM-derived application remains the same. A synthetic description of the eKCM architecure is given in the following figure.

4.1

General overview To support our methodology and facilitate our new development we have designed and developed a software environment named eKCM (eKnowledge For Competitive Maintenance). This environment was jointly developed with Sharing Knowledge software company (Ricard et al., 2001). This framework was built to support and implement our methodology ; thus eKCM enables the representation of expert knowledge and supports its acquisition by means of generic models, and provides experts and end-users with software modules. These modules implement various operational services : automatic diagnosis for effective control and maintenance decision support, knowledge validation, production of handbooks for end-users, and expert training.

6

Copyright © 2003 by JSME

•

Knowledge base for Knowledge base for another equipment another equipment

Knowledge acquisition, case description…

Diagnostic Diagnostic module module

The two latter points are largely motivated by the fact that : • knowledge might be used on plants located on a large territory, yet maintenance and storage are preferably centralized ; • in order to reduce IT costs, application installation on end-user machines must be as lightweight as possible.

Turbine fault model

Automated

Automated diagnostic diagnostic unction

function

I N T R A N E T

Health

Health assessment assessment function

function

Generic

Generic fault fault model

model

Generic

Generic behavior behavior desc. description model model

Turbine behavior description model model

eKCM kernel eKCM kernel Universal services

HTTP HTTP server server ++ servlet servlet engine engine

Other knowledge Other knowledge elements on elements on turbines turbines

Basic KB KB editor Basic editor

Standard HTTP data access

•

Knowledge presentation, automated services

Fig. 1. the eKCM intranet knowledge server architecture NB : the mentioned models and services are only used to illustrate the eKCM architecture ; there are many other available models and services in the present eKCM version. A closer description of the modular design of eKCM can be presented in figure 2. It shows the main blocks of this framework. Plain grey modules are the one that are included in the generic environment. Textured blocks represent application specific modules (D/M stands for "diagnosis and maintenance").

D/M

User

Remote access (network)

to limit the constraints for end-user deployment

Knowledge base for Knowledge base for steam turbines steam turbines

• • • •

That led us to the following technical choices : Persistence (i.e. knowledge storage) : XML files or relational databases (ORACLE, ACCESS, etc.) Repository (knowledge access management) : software modules in JAVA Knowledge representation (models) and processing (services) : sets of JAVA classes matching specific writing requirements Network access management : commercial HTTP server and communication filtering through JAVA servlets User interface : HTML + Javascript and JAVA applets.

4.3 eKCM, a software tool to support our methodology As previously mentioned, the methodological approach we use to address new applications relies on the cooperation of two skills (knowledge engineering and application specification and development) that need to be considered jointly. In fact, our efficiency depends on their close cooperation, especially in the beginning of a knowledge management project. The use of a framework such as eKCM facilitates this cooperation and efficiency : • rapid computer prototyping of representation structures helps to avoid ambiguities or imprecision in models and facilitatesknowledge validation, • a computer-based repository assists in managing tractability of represented knowledge.

D/M

services models Knowl. Persisaccess tence

interface Common services

Fig. 2. eKCM modular architecture Implementation choices When dealing with the design and development of the eKCM software tool, we wanted : • not to rely on too specific a tool to avoid being «constrained» by choices made by a tool manufaturer and to guarantee stability over time. • to use standard and portable formalisms and languages • to make porting of existing models (i.e. structured knowledge elements) and definition of new ones easy. • to allow flexible and modular development • to allow a good management of Intranet capabilities

As previously said, a key point in our methodology is the fundamental role played by the modeling activity, via the design (or reuse) of generic structures and relations that will be used to acquire and represent knowledge ; this activity is greatly supported by the eKCM framework as operational eKCM models are easily and directly derived from the structural description of the domain knowledge in terms of concepts and relations. Another key point is the central role of the prototyping activity for which eKCM is also an efficient dedicated time-saving framework.

7

Copyright © 2003 by JSME

4.4 eKCM, a software development kit for prototyping or developing industrial software solutions eKCM has been used for several internal projects in EDF, most of which were dedicated to the monitoring, maintenance and diagnosis of important pieces of equipment. Thus, in its current state, eKCM - via its model and service libraries - provides various models and services which have been specified and build from the experience gathered in past developments. Model

Description

Associative fault model

This model enables the representation of possible faults for an equipment and their manifestations

Health assessment model

This model allows the representation of conditions on monitoring data which expose a possible internal degradation of a part of an equipment

Prototypical fault model

This model allows the representation of abnormal situations which can be the result of a fault in the equipment

FMEA model

This model allows to bind a material decomposition to failure modes and to associated information : blueprints, experimental feedback, gravity…

relevant selection of models and services together with an adjustment of the user interface. Of course, when dealing with a new material, the specific knowledge base has to be filled. 5. EXAMPLES 5.1

Diagnosis assistance : DIVA A first knowledge-using domain is technical diagnosis. The basic goal is to identify primary failure origins by a good interpretation of the observed behavior of a piece of equipment. In the specific field of turbine-generator monitoring and maintenance, DIVA is an example of a knowledge-based diagnosis assistant, for which we applied our methodology and software framework. DIVA is a knowledge base expert system dedicated to turbine failure diagnosis (Porcheron et al., April 1997). The previous version of DIVA, issued in 1995, allowed to validate the diagnostic algorithm and the associated knowledge base. However, it had been developed in a commercial specific (now unsupported) environment and needs a UNIX autonomous machine to be run. This version had not been optimized for client-server use. This proved to be unsatisfactory first as it required on every power plant specific computer skills and secondly as the management of the knowledge base become a tricky issue when there are as many bases as software installments.

Table 1. Available eKCM models Service

Description

FMEA walkthrough

This is a navigation service dedicated to the FMEA model. It is based on a specific filtering based on the type of knowledge. It makes visualisation of information attached to each component easier.

Fault form generation

From an associative model, the service established text forms presenting faults and symptoms in a structured way. Outputs can be displayed in the browser or sent to a MS-Word RTF document.

Diagnostic table generation

From an associative model, the service automatically generates an array symptoms vs. faults. This array can be filled interactively to screen possible and excluded faults given observed symptoms.

Interactive diagnosis

Through queries to the user, the system establishes a diagnosis, via a set of faults which can be a explanation of observed symptoms (Console, 1989 ; Porcheron et al. (1997).

Equipment health assessment

Through queries to the user who fills in monitoring data, the system gives estimates on the current state of main parts of a complex equipment (satisfactory, dubious, abnormal).

For all these reasons, it was decided to port DIVA to the eKCM environment to improve its maintenance and its deployment. DIVA is now accessible as a "traditional" web service via our corporate intranet.

Table 2. Available eKCM specific services Fig. 3. DIVA main window A quick look at the modules listed in tables 1 & 2 – which are not exhaustive – shows that eKCM holds a fairly complete set of elements. These elements can be used to build diagnosis and maintenance applications that can be applied to large monitored pieces of equipment. This can be achieved through a

8

Copyright © 2003 by JSME

The RCP condition-based maintenance tool (RCP-CBM tool) is aimed at capitalizing knowledge related to the degradation and, more specifically, at providing the following services to the maintenance operators : • access, in an exploitable form, to several information relevant to the activities associated with condition-based maintenance (FMECA, experience feedback, etc.); • the drawing up of a RCP «health assessment» on an operating cycle on the basis of operating data (leakage flow rate from the seals, temperature of the bearings, vibrations), • a diagnosis support facility (DIAPO) (Porcheron et al., 1994 ; Porcheron et al., April 1997) making it possible, in the event of the health assessment detecting an anomaly, to search for the fault behind the anomaly; • different support facilities based on a reliability analysis of the pump's different components (in particular, joints and bearings), making it possible, before a shutdown or inspection, to assess the impact of the different maintenance scenarios with respect to maintenance costs and availability of the unit.

Fig. 4. DIVA graphical representation of the in-progress diagnosis 5.2 Condition-based Maintenance : application to reactor coolant pumps The aim of condition-based maintenance is to reduce maintenance costs by optimizing maintenance inspections according to the real state of a component. To do so, regular analysis of an operating equipment must be performed. This requires expert knowledge in order to know what to look after and how to interpret unexpected behavior. As this expertise is not evenly distributed on every site, the idea is to : • identify, model and store relevant expert knowledge ; • develop practical tools to assist non-experts in using this expert knowledge. This has been put into operation on Reactor Coolant Pumps in PWR nuclear power plants. A goal was to provide the user with a tool that assists him in assessing the feasibility of deferring a BPMP (the Basic Preventive Maintenance Program) scheduled inspection by drawing up a health assessment of the considered component.

Fig. 6. Health assessment service

eKCM has been used to develop a distributed software tool for condition-based maintenance of the reactor coolant pumps.

5.3

Life cycle management Life cycle management includes the optimization of operating life of production assets by selecting best suited operation conditions and maintenance planning. This is a key issue for widescale plant managing industries such as EDF. Knowing what might occur when a given equipment gets old and what operations or maintenance decisions might influence these occurrences is definitely an expert task. We therefore developed the « degradation model », described here, to allow the description and the capitalization of this tacit knowledge related to the degradation and aging

Fig. 5. The login window of the RCP-CBM tool

9

Copyright © 2003 by JSME

processes of power-plant equipment as known by our experts and as observed.

and aging mechanisms relevant for a given component, we had to introduce many other concepts as shown in figure 8.

InServiceComponent (Component/Function/Context)

Figure 8 gives a more detailed description of the « degradation model » in which 2 parts can be found : • the upper part is related to the functional and structural decomposition of the studied component ; • the lower part gives a causal description of the degradation processes relevant to a given component.

endures RelevantDegradationMechanism (ex. Wearing) mayCause: Degradation law, kinetic, distribution, etc... DegradationMode (ex. thicknessLoss) initiates

This model was cooperatively constructed with domain experts ; thus used representation corresponds to the implicit representation they have of their domain.

Causal Model for Failure Propagation

FailureMode

PrecursoryEvent

PrecursoryEvent

PrecursoryEvent

5.4

Other applications In many internal projects, eKCM has been used as a rapid prototyping tool to quickly validate suggested solutions. For example, we have recently used eKCM to develop the « Monitoring Techniques Catalogue », a knowledge management tool which allows the capitalization and the share of useful pieces of information relative to the monitoring techniques and tools relevant for a set of components.

DreadedEvent

Fig. 7. A general overview of the « degradation model » Figure 7 gives a general overview of this model ; it highlights the existing links between a given component, its relevant degradation mechanisms and the associated potential failure modes ; this model also allows a detailed description of the degradation mechanisms via its causal part (the lower part of the figure). It also introduces the general concepts identified by the experts (during expert interviews) : in-service components, relevant degradation mechanisms, degradation modes, failure modes, precursory events (events that foretell unwanted events) and dreaded (unwanted but possible) events.

We also use eKCM - and the associated knowledge modeling methodology - to address other domain problems ; we actually use our methodology in the fields of risk management and nuclear power plant decommissioning. 6. CONCLUSIONS AND PERSPECTIVES Safe and efficient nuclear power generation relies on a large and various amount of knowledge which sets various challenges for EDF : • ensure sharing of this knowledge, often spread over space and time, within populations of various skills, types of work, technical domains ; • maintain expertise over long periods of time, dealing with career evolution and retirement of experienced people ; • learn reusable lessons from rare events and past actions ; • make knowledge-based assistance available to people of lower expertise level.

Causality Plant

Characterization

Environment (temperature, pH, etc.)

isPartOf

Specialization

System

ensures

worksIn time period t1, t2

isPartOf

ensures ensures

Equipment

ensures

Function

influencingParameters mayEndure

isPartOf

InServiceComponent

isIn *

GenericMechanism (ex. Wearing, Corrosion...)

Material

MaterialCharacteristics isCharacterisedBy (composition, properties, RelevantDegradationMechanism initial defaults, (ex. Wearing) makesRelevan etc.) t isTheSiegeOf mayCause : Degradation law, kinetic, distribution, etc. canLeadTo : kinetic, DegradationMode propagation law (ex. thicknessLoss) endures

Consequence Safety/Exploitation Unacceptable/alarming/unfavourable... AcceptationCriterion isImpactedBy

leadsTo qualifies DreadedEvent affects (component, equipment, system, etc.)

causes: degradation law, Kinetic, distributionc, etc.

causes: degradation law, Kinetic, distributionc, etc.

FailureMode (ex. Crack)

qualifies ObservationCriterion

identifies

Observation canLeadTo : kinetic, propagation law identifies PrecursoryEvent affects (component) (ex. Leak)

Option/Parade

induces

annonces

Our implication in the related necessary knowledge management activity led us to develop a methodology and a unified framework for expert knowledge formalization and processing (Ricard, 2002).

implementedBy ObservationMeans :

DreadedEvent affects (component, equipment, system, etc.)

(ex. US inspections every 6 months)

Fig. 8. A more detailed view of the « Degradation Model » Further discussions with the experts led to a much more detailed description of their domain : to allow a precise and complete description of the degradation

This approach is based on a « knowledge modeling » activity and can be summarized as follows :

10

Copyright © 2003 by JSME

•

•

•

plants », Colloque Européen de Sûreté de Fonctionnement (λµ13), European Safety and Reliability Conference (ESREL 2002), Lyon, France, March 18-21, 2002. L. Console, D. Theseider-Dupré, P Torasso,. « A theory of diagnosis for incomplete causal models » Proceedings. of 11th IJCAI, Detroit. 1989. R. Dieng – INRIA. Méthodes et outils d'acquisition des connaissances, Rapport de recherche n°1319, INRIA, novembre 1990. P. Haïk (EDF) and F. Guarnieri (Ecole des Mînes de Paris). « AFD, a new methodology for failure and risk assessment : presentation, illustrations and prospects », Colloque Européen de Sûreté de Fonctionnement (λµ13), European Safety and Reliability Conference (ESREL 2002), Lyon, France, March 18-21, 2002. P. Haïk, B. Ricard and S. Mahé - EDF. « Knowledge modeling as a support for knowledge acquisition, sharing, processing and maintenance ». PKAW 2002 : The 2002 Pacific Rim Knowledge Acquisition Workshop, Tokyo, Japan, August 2002. W. Hamscher, L. Console, J. de Kleer, (Eds). Readings in Model-based Diagnosis. Morgan Kaufmann Publishers, San Matao, CA. 1992. M. Porcheron, B. Ricard - EDF. « DIAPO, a case study in applying advanced AI techniques to the diagnosis of a complex system » 11th European Conference on Artificial Intelligence (ECAI’94) ; Amsterdam, Netherlands, 1994. M. Porcheron, B.Ricard - EDF. « An Application of Abductive Diagnostic Methods to a Real-World Problem » Proceedings of the eighth International Workshop on Principles of Diagnosis DX'97, Le Mont Saint-Michel, France, September 14-18, 1997. M.Porcheron, B.Ricard, A.Joussellin - EDF. « Diva and Diapo: two diagnostic knowledge-based systems used for French nuclear power plants » 5th International Topical Meeting on Nuclear Thermal Hydraulics, Operation and Safety (NUTHOS-5), Beijing (China), April 1997. M. Porcheron, B. Ricard – EDF. « Steps towards a unified framework for knowledge-based diagnosis support of nuclear power plant devices ». ISAP'99 (Intelligent System Application to Power Systems), Rio-de-Janeiro, April 1999 B. Ricard and P. Haïk – EDF. « Distributed knowledge based maintenance assistance for nuclear power plant components ». EPRI International Maintenance Conference, Houston (USA), Août 2001. B. Ricard – EDF. "Knowledge-Based Systems for Diagnosis and Maintenance Assistance". Revue de l'Electricité et de l'Electronique, n°6, June 2002.

models (strongly typed formal representations that represent the existing points of view available for a given domain of interest) are used to support the acquisition process and to facilitate the sharing, reuse and maintenance of the capitalized knowledge ; the four steps of the methodology (model design, acquisition and filling-in of models, solving methods specification - when relevant-, model and solving method implementation) are used to address new knowledge management applications, using feed-back processes to put experts and end-users at the center of the design and development processes ; the systematic use of « operational prototypes » facilitates exchanges with experts, knowledge acquisition, and model validation ; it is also a way to demonstrate the feasibility of the project and to convince the project manager.

Various internal projects dealing with expert knowledge have been held with our methodology. Among the benefits of our approach (methodology, unified framework and dedicated software environment) are : • improving efficiency in expert knowledge modeling and structuring ; • improving productivity
in expert knowledge acquisition and management ;
in the design and development of new services and applications that operates the capitalized knowledge ; • improving model and service reliability as a result of model and service reuse. Together with the preservation of key expertise, the methodology, combined with our software tool, facilitates the development of economically-competitive decision-support software tools. In a wider perspective, we are currently performing research work on processes and means to stimulate exchange and use of (explicit or tacit) knowledge between employees (amongst which experts) in order to increase the global use of knowledge and thus to increase the liability and the efficiency of our processes and production tools (Haïk, 2002 ; Bouzaïene, 2002). REFERENCES W. R. Ashby. « An Introduction to Cybernetics », Chapman & Hall, London, 1956. L. Bouzaïene, F. Péres (Ecole Centrale de Paris) and P. Haïk (EDF). « Maintenance experts crossed stimulation: application to the French nuclear power

11

Copyright © 2003 by JSME