This paper describes a knowledge acquisition method used in conjunction with a trouble-shooting system for marine engines. According to diagnostic knowledge, our consultation system can give important and adequate advices to the user in the case of trouble with marine engines. Diagnostic knowledge is usually given by human experts, but it is actually difficult to obtain such a large amount of specialized knowledge from human experts or knowledge engineers. In our method diagnostic knowledge for trouble-shooting is acquired from a data base or input data according to human expert's advices. To extract such knowledge, our system employs heuristic knowledge-extraction rules, which find causal relations between every part of a marine engine. The extracted knowledge is not always appropriate and correct because of errors induced by its generalization process. Therefore, our system employs a learning mechanism which corrects a process for extracting knowledge by modifying the extracted knowledge according to the advices of human experts. Our trouble-shooting system can automatically obtain diagnostic knowledge, which has the following characteristics. - Frames are used in a data base to represent facts about the domain. - Knowledge-extraction rules (KERs) represent causal relations among frames to acquire the diagnostic knowledge. - Knowledge about the operation of machinery and its states are given by human experts. The use of many specialized KERs are not practical because of the increasing number of KERs. In contrast the use of generalized KERs may result in erroneous knowledge. In our system, the knowledge is extracted by using the generalized KERs but at the same time the generalization is controlled by an outside teacher. If erroneous knowledge is found, the system corrects the process for extracting knowledge. Correction is made by altering the weight of slots and predicates in the KERs for reordering the priority of KERs, and by altering the weight of the values of slots in the domain knowledge frames for recognizing the operating conditions of the machinery. In our system, the rule base is constructed hierarchically by dividing domain-independent knowledge, domain-dependent knowledge, and control knowledge. Since the system uses the metarules of representing control knowledge separated from the domain knowledge, control strategy can easily be understood and control structure can be quickly adapted for other fields. Our knowledge acquisition method is particularly useful for constructing an initial large knowledge base, and also applicable for addition, correction, and update of knowledge. Furthermore, We believe that this system can be used for acquiring diagnostic knowledge about different domains by changing omly the domain-dependent knowledge and about different domains of machinery by changing only the data base, since the system uses the generalized KERs.
抄録全体を表示