Qualitative Event Analysis for Fault Diagnosis

Abstract

A robust fault detection system for industrial processes is described in this paper. The expected behavior of the process is generated from a qualitative model and a fault is detected if discrepancies are noted between the expected behavior and the significant actual evolution. A survey of the state of the art points out the classical drawbacks of the threshold methods: model based fault detection is generally too sensitive to the inappropriate settings of any threshold values that can lead to failures in diagnosis. The simulator used to provide with the expected behavior of the process is a qualitative one. So we introduce qualitative notions and fuzzy analysis in the model-process comparison to avoid the threshold problems and to take into account the imprecision inherent to the modelling or to the measurement system. A causality analysis of the discrepancies observed in time enables to explain alarms series. The fault detection system has been applied on a chemical engineering process and some encouraging results are obtained.