Oxford University, Computing Laboratory

Although neural networks have been successfully applied to many real world applications, including robotics, image analysis, control, diagnosis, medical area, bioinformatics, they have always been criticized for their lack of transparency in human understandable terms. Neural networks have excellent capabilities to learn directly from data, but the knowledge captured in the structure and weights matrix of a neural network is not in a human comprehensible form such as if-then rules or logic formulas.

On the other hand Fuzzy Systems, based on Fuzzy Set Theory, can handle uncertain and imprecise information using a small number of fuzzy if-then rules but they don't possess capabilities of automated learning and therefore have difficulties in knowledge acquisition. Neural Networks process quantitative and numerical information while Fuzzy Systems process symbolic and qualitative information. Fuzzy Systems are appropriate if sufficient expert knowledge is available, while Neural Networks are useful if sufficient data are available or measurable.

Neuro-Fuzzy Systems combine the advantages of both approaches, exploiting the learning ability of neural networks and the reasoning ability and transparency of Fuzzy Systems. Neuro-Fuzzy systems can be used for automatic extraction of linguistic rules from training data and also for refining prior expert knowledge using additional data.

The talk presents different methods for combining neural networks and fuzzy systems and focuses on the use of Neuro-Fuzzy techniques in Fault Detection and Isolation. The diagnosis of faults in industrial devices consists of two sequential steps: fault detection and fault evaluation (classification or isolation). The big concern of the talk is how to build interpretable Neuro-Fuzzy models while still accurate in order to detect faults, and how to build transparent Neuro-Fuzzy classifiers for fault isolation. In this idea the talk discusses the key issues in Neuro-Fuzzy modeling such as structure identification and optimisation, more exactly how to find an appropriate number of rules and membership functions for a given transparency level. Some application results are presented throughout the talk.

DATE: May 30, 2002, Thursday @ 15:45

PLACE: EA-409