For the maintenance of a switching system, many levels of expert's decisions may be required to find a correct cause and its appropriate repair procedure. Output messages generated by our switching system, called TDX-1A, for the maintenances have, in general, the following characteristics: 1) A bunch of maintenance messages generated only by even not many trouble/faults make it hard for an operator to maintain. Approximately 10 messages are generated by a fault. 2) To figure out a correct cause, its fault name and its repair procedure, several kinds of knowledges for the switching system architecture, the message interpretation, the trouble history, the system options etc. may be utilized. By analyzing maintenance messages, we found the correlation between output messages and the system architecture for an interpretation, and set up a model for a navigation of a fault diagnosis. Based on the correlation and the model, an inference model for the fault diagnosis was found and the knowledge base had been classified into sub-knowledge bases according to the navigation model. Furthermore, to improve the performance of the inference model, we used meta control knowledges and the dynamic rule weighting mechanism. The intelligent maintenance system was implemented on the lisp machine LAMBDA 2×2/Plus using the expert system tool, KEE(Knowledge Engineering Environment) and tested for a couple of months.