Process fault detection and diagnosis using neural networks

The field of fault detection and diagnosis deals with the design of computer-based automated systems that can assist human operators. Presently, the task of fault detection and diagnosis depends primarily on human operators, who have important limitations such as stress, fatigue, and inattentiveness. A number of expert system based approaches have proposed in the literature for automated fault diagnosis. However, these systems cannot be rapidly deployed due to inherent drawbacks, such as the difficulty of knowledge acquisition, the inability of the system to learn, and the brittleness of the system outside its domain of expertise. A potential solution to these problems is the use of neural networks as demonstrated in this thesis. A detailed analysis of the application of neural networks for diagnosing process failures, in steady-state and dynamic chemical process systems, is presented in this thesis. The diagnosis of single and two-fault scenarios, the effect of fairly high levels of measurement noise, the ability to diagnose with partial data or missing hidden units, and the effect of the choice of fault-class representation by the output nodes, have been explored. The factors affecting fault classification ability have been elucidated using a geometric analysis. For diagnosing process failures during transients, a desired output discretization scheme reflecting the magnitude and type of the malfunction, and its time of occurrence has been devised. The studies presented in this thesis indicate that neural networks have the potential to diagnose various types of failures in chemical process systems. Further, networks trained only on single-fault patterns, correctly identified both faults from measurement patterns caused by two concurrent interacting faults, in many cases. In the other cases, at least one of the two faults was diagnosed. Two-fault generalization by networks with the same number of input units, was found to be dependent on the number of hidden units; while, recall and single-fault generalization were relatively insensitive to the hidden unit number. Neural networks were found to be quite tolerant to sensor faults, hidden node failures, and measurement noise. An implicit representation of the normal process state, in the output nodes of the network, was found to be superior to an explicit representation. The geometric analysis showed that the number of hidden units had a strong influence on the length of training needed for correct recall, and the number of input units was important for success in two-fault generalization