A Hybrid Approach for Detecting and Isolating Faults in Nuclear Power Plant Interacting Systems

This paper presents a Fault Detection and Isolation (FDI) system that can detect and isolate nuclear power plant (NPP) faults occurring in interacting systems. The proposed methodology combines two tools: observer based residual generation and neural network pattern matching into a powerful, hybrid diagnostic system. A computer-based model of a commercial Boiling Water Reactor (BWR) is used as the reference plant. Two FDI methods are implemented on each of two BWR systems and their performance characteristics are compared. One method uses conventional neural network techniques that use parameter values for input, and a second, hybrid methodology, uses system models to create residuals for input to a neural network. Both FDI systems show good generalization abilities, but only the hybrid system decouples system interactions. Although implementation is impractical for all NPP systems, this hybrid technique is most useful in specific applications where operators have difficulty diagnosing faults in strongly interacting systems.