Virtual measurements using neural networks and fuzzy logic for complex system monitoring

A new approach is presented that demonstrates the potential of pretrained artificial neural networks (ANNs) as generators of membership functions for the purpose of monitoring nuclear reactor systems. ANNs provide a complex-to-simple mapping of reactor parameters in a process analogous to that of measurement. Through such virtual measurements the value of parameters with operational significance, e.g., valve-position, transient-type or performance can be determined. In the proposed methodology the output of a virtual measuring device is a membership function which uniquely and unambiguously represents the state of the system. Utilizing a fuzzy logic representation offers the advantage of describing the state of the system in a condensed form, developed through linguistic descriptions and convenient for application in monitoring, diagnostics and generally control algorithms. The membership functions generated from independently firing networks, belong to a set of prototypes that describe the universe of discourse. A computational technique based on the dissemblance index between two fuzzy numbers, has been devised for discriminating amongst different network responses. Information criteria have been utilized for best model selection in the development of a rule-based diagnostician that incorporates the time dimension in the stationary behavior of the ANNs. The rule-based system is supplemented with fuzzy algebraic techniques forming an optimization algorithm capable to continuously adjust the membership functions describing the system. The optimization algorithm modifies appropriately the shape and position of the resultant membership functions, following changes in the operating environment. The proposed methodology is applied to the problems of measuring the disc position of the secondary flow control valve of an experimental reactor and transient identification in a Nuclear Power Plant (NPP). The results obtained clearly demonstrate the enhanced noise tolerance of the methodology as well as its unique robustness in fault signals. The virtual measuring methodology has been proven capable of identifying the exact system state even in the extreme case, where input signals have been substituted for random time series. The model behavior shows that it is possible to develop virtual measuring devices through artificial neural networks mapping dynamic time series to a set of membership functions, and thus enhance the capability of monitoring systems. Furthermore, the application of optimization algorithms on the neural network responses resolves the issue of designing time-varying membership functions suitable for coping with changes in the operation environment.