A self-learning fault-diagnosis system

Abstract

This paper describes a self-learning fault-diagnosis system which can find any inappropriate settings of its parameters and, hence, improve its own performance. It diagnoses faults based on a deep qualitative model of the process being monitored. From this qualitative model, the expected behaviour of the process can be generated and, if it differs from the actual one, then it is perceived that a fault (or faults) occurs in the process. The diagnosis of sensor failures is based on a set of heuristic rules, while the other component failures are diagnosed by comparing the expected behaviour under a hypothesis with the actual one. The inappropriate settings of any threshold values, for converting quantitative values to qualitative values and for determining symptoms in the perception of sensor failures, are considered as a major reason for failures in diagnosis. Once such a failure occurs, the diagnosis system will inspect itself and find any inappropriate parameters. This is achieved by examining the recorded problem-solving history and performing backwards tracing through the model of the fault-diagnosis system. The expected output of the fault-diagnosis system is propagated backwards through this model. Any threshold values which are responsible for not giving the expected output are examined, and the inappropriate parameters are found. This self-learning fault-diagnosis system has been implemented in an expert systems shell: EXTRAN, and is applied to a pilot-scale mixing process.