The existing fault diagnosis methods using neural networks learn fault modes from pairs of single-symptom-single-cause only. But in real plants, the effect of a fault propagates continuously from its origin; different sensor values reflect this. In the present study, a new method is suggested which can find the fault origin in spite of dynamic nature of its symptoms since time varying symptoms are used in the training patterns.
In order to analyze learning, recall and generalization characteristics of artificial neural networks for process fault diagnosis, a distillation column was studied as an object for fault diagnosis. In the case of time varying symptoms, the measurement patterns for training and testing the artificial neural network were obtained from dynamic simulation programs that model the behavior of the plants.
To train the artificial neural network, 25 malfunctions (each malfunction has 50 time varying symptoms) were simulated. Various fault symptoms different from trained symptoms(for examples slow appearance of fault origin, sensor noise, multiple faults etc.) were simulated to test the fault diagnosing capability of trained artificial neural network. The trained artificial neural network could accurately diagnose both single and multiple faults. Therefore the proposed fault diagnosis method can find the exact origin of the fault for which the symptom is propagated continuously with time.