Application of neural network techniques to fault diagnosis of a heat exchanger system

Abstract

Artificial neural networks, by their superior pattern-recognition ability, are well-suited for developing intelligent diagnostic tools for complex processes such as process plant operation. Fault diagnosis in a cross-flow tubular heat exchanger system is carried out by using three different paradigms - the Backpropagation (BP) network, the Recurrent Cascade-Correlation (RCC) network and the Self-Organising Map (SOM). The study focusses on two different fault scenarios which are simulated for the heat exchanger plant. The first deals with distinct fault states caused by equipment failure within the system whilst the other deals with fouling in the heat exchanger tubes. Training and performance results obtained from a comparative study using the three different networks are presented and practical issues concerning their role and implementation are discussed.