We have prototyped a diagnostic system, Doctor, which employs case-based reasoning (CBR). The objective of the Doctor is to support service persons to repair home appliances, e.g., room airconditioners. The Doctor asks several questions to collect symptoms when an end-user asks for repair. Then defect candidates for the repair are listed up by retrieving previous cases which are similar to the collected Symptomns. Good casebases are essential to obtain accurate results in CBR. The Doctor has casebases, called ByType casebases, which are assorted according to product types. Therefore the diagnostic output of our system can reflect the defect trend of each product type. The ByType casebases are initailly built from the generic casebase which collects diagnostic rules experienced service persons have. The ByType casebases are afterword updated by adding service reports which consist of observed symptoms and repair actions, i.e., what was done to fix a fault. Our method is one way of integrating two knowledge sources, rules and cases. An advantage of our method is that a diagnostic system needs only one reasoning mechanism, CBR, to exploit the two knowledge sources because they are combined into a ByType casebase in advance. One of development items of a CBR system is the design of a heuristic function which evaluates the similarity of two cases. We introduced a metric Possibility which values how much effective a previous case is to repair a malfunction. The Possibility function considers the similarity of two sets of symptoms and the frequency of a previous case, i.e. how often a repair action was taken to fix the same set of symptoms. Therefore the doctor can list up defect candidates with their probability. In this paper, we describe how to build a generic casebase and the retrieval algorithm of our system. We also demonstrate the effectiveness of the Doctor with some experimental results of room air-conditioner diagnosis.
抄録全体を表示