On the nature of fault space classification structure developed by neural networks

Abstract

Recent research efforts have shown that backpropogation neural networks can be successfully used for fault diagnosis in steady-state and dynamic chemical processes. This approach relies on the non-linear classification properties of neural networks. The diagnostic ability depends on the discrimination of decision regions corresponding to various fault classes, by the neural network. The delineation of decision regions for a given number of fault classes, in turn, hinges on several factors, such as the choice of training data, the number of input and hidden nodes, and the extent of training. In this study, the impact of hidden units and input units on fault space structure and fault classification performance for a given set of training data, is examined. The studies were conducted using a simulation of a chemical reactor. It was observed that the recall and single-fault generalization performances of the neural networks are very good. However, to achieve such good results one needs to have a rich training data set. The results indicate that as the number of hidden units is increased, the amount of training required for correct recall decreases considerably, showing that the delineation of the fault space into different decision regions becomes easier. Also, with a greater number of hidden units, there is a small improvement in single-fault generalization. However, significant improvements in two-fault generalization performance are possible only with the addition of extra input units that convey more information about the state of the system being diagnosed.