Application of Artificial Neural Networks in Process Fault Diagnosis

Abstract

The application of neural network methods in process fault detection and diagnosis is discussed. The diagnosis is based on the classification of process measurements. A realistic heat exchanger - continuous stirred tank reactor system is studied as a test case. The system has 14 noisy measurements and 10 fault situations are considered. The arrangement of different fault categories is visualized by the principal component analysis. Three different neural network architectures are applied to detect and diagnose the faults of the test process. The performance of different architectures is compared and the properties of the traditional nearest-neighbor classifier is discussed too. The multi-layer perceptron network yields reliable fault diagnosis, while the networks which are trained unsupervised are not able to diagnose all situations correctly. On the other hand, self-organizing networks can produce valuable information about relationships between different fault situations.