A neural network methodology for dynamic fault diagnosis has been proposed. Moving windows cut the dynamic data into overlapping pieces. Then the segmented data are presented to the networks for training and generalization purposes. Some unique features associated nith this methodology, namely the length of the moving window, the sampling rate, and the construction of the training data set, have been studied. The proposed method was successfully applied to a binary distillation process and showed superiority over the networks trained by steady state data.

No abstract