AN INTEGRATED SCHEME FOR PROCESS FAULT DIAGNOSIS

Abstract

Fault detection and diagnosis is currently a very important problem in process automation because of the increasing demands on reliability and safety. Current techniques include model-based methods, expert systems and the use of artificial neural networks. A particular form of neural networks, the Back Propagation neural network (BPN), has been applied extensively to process fault diagnosis (PFD) problems. However, applying BPN to PFD problems shows some drawbacks. Performance is inadequate while working under certain non-ideal conditions, such as in the presence of new fault class or when multiple faults occur simultaneously. Multiple fault diagnosis by traditional model-based methods or expert systems are rather complicated and a new mechanism is required to simplify the multiple fault diagnostic procedure. In this thesis, the weakness of applying only BPN for PFD problems is investigated. The Fuzzy Adaptive Resonance Theory neural network (FARTN) is introduced to perform process variable change pattern identification while BPNs are used for fault level classification. Each BPN corresponds to a fault class. Whenever the change pattern related to a certain fault class is identified, the corresponding BPN is employed. In the case when an unforeseen change pattern appears, it is stored in FARTN for use in later diagnosis. A novel Al-based pattern analysis mechanism is developed for multiple fault diagnosis. This mechanism is able to decide whether or not a pattern is due to a new fault class and, if not, determine likely fault sources based on an existing knowledge of process faults. In a sense, the mechanism is a general-purpose means, i.e., its performance will not be effected by process models. By combining neural network and Al techniques, an integrated PFD scheme is proposed in the thesis. Simulations using this integrated scheme showed that the proposed method can work well even under situations when a new fault class emerges and when multiple faults occur simultaneously. The proposed scheme is implemented in a PC-based integrated simulation package. The software includes several independent simulation units and a user-friendly interface. It provides an useful means for understanding and simulating PFD problems.