Chemical plant fault diagnosis by neural networks

Abstract

An analysis of learning and generalization characteristics of neural networks for diagnosing process failures was presented. Various feedforward neural network topologies were tested and compared. The single fault assumption was relaxed to include multiple causal origins of the symptoms. A chemical plant composed of a reactor and a distillation column was used as a case study. The performance during recall improves at first with an increase in the number of hidden units and with the amount of training, and then attains convergence. The algorithm of the Generalized Delta Rule (GDR) was used to train the networks by minimizing the sum of squares of residual according to the given convergence criterion. The obtained results show the applicability of the neural networks structure with hidden layers for process fault diagnosing. These results illustrate the feasibility of using neural networks for fault recognition and location.