On-line process fault diagnosis using neural network techniques

Abstract

A neural-network based on-line fault-diagnosis system for industrial processes is presented in this paper. A multi-layer feed-forward neural network is developed and trained with symptom-fault pairs from experience of the operation of a process or from simulation analysis of that process. The trained network can be used to diagnose faults in that it can associate the abnormalities in on-line measurements with corresponding faults. Compared with diagnosis systems based on expert-systems techniques, which have several limitations such as the time consuming nature of developing the knowledge base and the inability to cope with situations not presented in the knowledge base, the neural-network based fault-diagnosis system is easy to develop and performs robustly. The feasibility of applying such a diagnosis system to industrial processes is demonstrated by applying it to a pilot-scale mixing process and in a simulation study of a continuously-stirred tank reactor (CSTR) system. A series of experiments is carried out to investigate the performance of the neutral-network based on-line diagnosis system and it is shown that it can perform satisfactorily with partially incorrect and partially unavailable information. Therefore, to some extent, the system can tolerate measurement noise and model-plant mismatch.