THE USE OF NEURAL NETWORKS AND A PROTOTYPE EXPERT-SYSTEM IN BWR ATWS ACCIDENTS DIAGNOSIS

Neural networks and a symbolic expert system are employed to form a pr ototype system in BWR anticipated transients without scram (ATWS) acci dents diagnosis. Unsupervised learning based on discovery of cluster s tructures (Pao's approach) and back propagation (BP) neural networks a re used to group and memorize different ATWS patterns. Multiple traini ng data sets derived from output files of the SABRE computer code are used in training the BP networks to cope with oscillations of reactor power and other parameters. Tests of neural networks recall correctnes s are performed, given the presence of time shift of sample data, rand om noise and incomplete information. The expert system can simulate th e ATWS strategy developed by Pennsylvannia Power & Light company. It u ndertakes diagnosis, using an event tree structure, and can execute th e trained BP networks. The expert system is also able to deal with tem porary loss of reactor water level information.