On-Line Learning for Fault Classification Using an Adaptive Neuro-Fuzzy Network

Abstract An approach to on-line learning and classification of fault conditions for process fault diagnosis using an adaptive neuro-fuzzy network is described. A hierarchical network structure is used incorporating subnetworks for each fault class and local activation functions in the hidden layer. Hidden nodes and subnetworks are automatically added to the network to accommodate new process faults after detection. Network adaptation is achieved using a decision based learning algorithm to train localised network parameters and relationships with fuzzy logic are used to provide an interpretation of the network operation in the form of qualitative rules. Applications to adaptive learning of incipient faults on a multi-variable, chemical process simulation and a laboratory process are described. Results illustrate the network operation and demonstrate the capability of the network to successfully learn and classify a range of fault conditions.