Industrial plants often have many process variable measurements available, which can be monitored for fault detection and diagnosis. Using all these variables as inputs to an artificial neural network for fault diagnosis can result in an impractically large network, with consequent long training times and high computational requirement during use. Principal component analysis (PCA) is investigated in this paper for generating a reduced number of variables to be used as neural network inputs for fault diagnosis. The main application described is to a real industrial nuclear fuel processing plant. A simulated chemical process was also used to assist the development of the techniques. Results in both applications demonstrate satisfactory fault diagnosis performance with a reduction in the number of neural network parameters of approximately 50 per cent using PCA. The paper also includes some introductory material on PCA and neural networks, and their application to process fault diagnosis.
