Adaptive neural network approach to on-line learning for process fault diagnosis

Abstract

The practical difficulties of collecting representative fault data for network training can result in performance limitations when a trained static neural network is imple mented on-line for fault diagnosis of dynamic processes. To address this problem, an adaptive neural network is described that continually monitors and improves its performance on-line as new fault information becomes available. New nodes are automatically added to the network to accommodate new process faults after detec tion, and on-line adaptation is achieved using recursive linear algorithms to train selected network parameters. Applications of the adaptive neural network to the detection and diagnosis of incipient faults in a simulated chemical process and a real laboratory process rig are described. Results are presented to illustrate the network operation and demonstrate the ability of the network to learn and diagnose a range of process faults successfully.