Simulation-Based Expert System for Nuclear Power Plant Diagnostics (Artificial Intelligence, Filtering).

Abstract

The application of expert systems to the diagnostics of nuclear power plant accidents is considered. In this work, dynamic simulations, Kalman filtering, pattern recognition, fuzzy diagnostics and artificial intelligence have been combined in a unique algorithm for diagnosing and analyzing nuclear plant transients targeted for use on-line and in real time. Knowledge-based reasoning is used to monitor plant data and hypothesize about the status of the plant. Fuzzy logic is employed as the inferencing mechanism and an implication scheme based on observations is developed and employed to h and le scenarios involving competing failures. Hypothesis testing is performed by simulating the behavior of faulted components using numerical models. A simulation filter has been developed based on the structure of the Kalman filter for systematically adjusting key model parameters to force agreement between the simulation and actual plant data. The unique feature of the simulation filter is that it operates only on the discrete time-series of inputs and associated outputs of a dynamic simulation program, thus admitting arbitrary system dynamics and being readily applicable to any system for which a simulation program for computing system states is available. The simulation filter is shown to be successful for estimating system states, identifying system parameters and computing fault magnitudes for studies involving nuclear reactor kinetics, pressurizer and steam generator dynamics. Pattern recognition is employed as a decision analysis technique for choosing among several hypotheses based on simulation results. An Artificial Intelligence framework based on a critical functions approach is used to deal with the complexity of a nuclear plant system. Detailed simulation results of various nuclear power plant accident scenarios are presented to demonstrate the performance and robustness properties of the diagnostic algorithm developed. The system is shown to be successful in diagnosing and identifying fault parameters for a normal reactor scram, loss-of-feedwater and small loss-of-coolant transients occurring both separately and combined in a scenario similar to the accident at Three Mile Isl and .