KNOWLEDGE ENGINEERING FOR INDUSTRIAL EXPERT SYSTEMS

The inherent difficulties involved in the process of extracting knowledge from experts are discussed and identified. Such difficulties have resulted in few expert systems progressing beyond the prototyping stage. The conflicting terminology used to describe the whole process is examined and, as a result, knowledge engineering is defined as the appropriate term for the whole process. This is then further split into knowledge acquisition and system implementation. Finally, knowledge acquisition is further subdivided into knowledge elicitation and machine induction. The particular problems associated with the construction of expert systems in industrial control applications are discussed. Such systems are characterised by the nature of their user population, the type of support provided and whether they operate on-line or off-line. The importance of defining functionality and goals at the outset is stressed. The need for user models is also highlighted. the various techniques used in knowledge elicitation-interviews, questionnaires, observations, protocol analyses, teachback interviewing, walkthroughs and formal techniques-are briefly reviewed. The alternative approach using machine induction techniques is also discussed. An examination is made of the competing approaches involving bottom-up and top-down techniques. The benefits resulting from the application of cognitive task analyses rather than technology-driven approaches are also stressed. Current knowledge acquisition tools such as KRITON, KADS, ACQUIST, KEATS and ROGET are reviewed. Examples are given of the use of time-line techniques in power plant knowledge acquisition, knowledge and task analyses in the construction of a failure management expert system and of the use of inductive techniques in gas-oil separator design and satellite power systems control. In the latter case, the use of qualitative modelling is highlighted. The possibility of domain experts in industrial control carrying out their own knowledge engineering is examined but rejected as unlikely, unless better tools exist. The provision of better tools is identified as one of the key factors required to simplify the knowledge engineering process.