REPAIR: A Model-Based Diagnosis System

This paper describes the work of the REPAIR project, with particular emphasis on research into model-based diagnosis. The REPAIR project was a three year research project entitled "Qualitative Reasoning Systems for Physical Mechanism Diagnosis ", and was funded by the UK Alvey IT Directorate (ref. IKBS/061). The aim of the project was to investigate diagnostic reasoning processes in the context of physical mechanisms drawn from industry and engineering. This goal was pursued by building experimental software simulators that facilitate the modular construction of device models and allow the simulation of realistic device behaviour. Our initial experience suggested that diagnosis techniques based on heuristic rules, as found in commercial expert systems, have serious inherent limitations. Although there has been a degree of success in building heuristic-based diagnostic systems, these systems are usually specially configured and tailored for given applications. In these cases the success is achieved because the heuristic rules are highly tuned to the domain and any reasoning involves very little first-principles analysis of the fault symptoms. In contrast, we believe it is important to model directly the structural and functional aspects of mechanisms, in order to investigate how automatic reasoning tools may ascertain the causes of faults and failures when malfunctions occur. Frequently, the knowledge used to perform diagnosis bears little relation to the physical structures in the domain. While this may not matter in applications that have their domain based in theoretic or semi-fomal schemas (e.g. financial or business expert systems), knowledge of the physical structures is paramount to good understanding in engineering situations. As we are interested in diagnosing novel failures and generating causal prognoses and failure scenarios, we studied methods that offer more potential than conventional heuristic systems. Another impetus for the work is the need for systems that are much more generic to be developed. This is necessary if knowledge is to become more readily available as a commodity for solving problems. Our studies of model-based techniques have confirmed this approach as an important advance. Model-based diagnosis takes knowledge about a domain and about the way in which the structures in that domain operate, and uses that knowledge as a basis for diagnosis. Basing a diagnostic system on a model has significant potential advantages in term of system robustness, reuse and maintainability. The main disadvantage is that model-based diagnostic systems are mostly still at the research stage. This paper describes the ideas of model-based diagnosis and outlines the requirements of suitable models; in particular, the features of qualitative reasoning are found significant. The REPAIR project model-based system is briefly described. The advantages and limitations of the techniques employed are considered and the use of model-based reasoning systems as the basis of a more powerful general diagnostic architecture is examined.