Neuro-fuzzy structure for rule generation and application in the fault diagnosis of technical processes

An attempt has been made within this approach to establish a neural structure which models the fuzzy inference mechanism and thus combines both the adaptive feature of neural networks and the transparency of fuzzy systems. This structure is utilized to extract an initial knowledge base applying a coincidence learning law. The base is then optimized to obtain better matching to the underlying problem through adjusting the attributes of input variables, and operator's parameters like relevance weights and conjunction degrees where it is shown how a perceptron with a sigmoid activity function can implement softened conjunction or disjunction. Basic features of the neuro-fuzzy scheme are illustrated using a benchmark example. The scheme is then successfully applied to extract rules for fault diagnosis with a turbocharger using estimation residuals as analytical symptoms.