Measurement of fuzzy values using artificial neural networks and fuzzy arithmetic

A methodology for monitoring complex systems utilizing artificial neural networks and fuzzy arithmetic is presented. It employs the notion of a virtual measuring device, i.e., a software-based instrument for the measurement of user-specified dynamic variables with operational significance. Neural networks are utilized for mapping a set of complex, temporal input patterns to a simplified set of membership functions. The resultant membership functions are supplied to a rule-based system, which draws a conclusion about the validity of the network responses based on the statistical characteristics of the original training signals. The notion of time is explicitly incorporated into the decision-making procedure, rendering the measuring process a dynamic one. The rule-based system is supplemented by fuzzy algebraic techniques composing an optimization algorithm capable of identifying the most appropriate shape and position of the resultant membership functions in the universe of discourse. Data obtained from an experimental nuclear reactor during a start-up period are utilized to demonstrate the excellent tolerance of the methodology to noisy and faulty signals