Traditional signal pattern recognition versus artificial neural networks for nuclear plant diagnostics

This paper introduces a new tool based on a traditional noise analysis tech niques for monitoring reactor components' signal condition. It also present s the performance of artificial neural networks for pattern recognition to the same set of reactor signals and provides a comparison of these two tech niques. Reactor pump signals from the Experimental Breeder Reactor (EBR-II) are utilized here. Collected signals such as pump power, pump speed, and p ump pressure are obtained from already installed sensors in the reactor. Th e signals utilized are collected signals as well sis generated signals simu lating the pump shaft degradation progress. From the study of time series a nalysis and regression modeling of these signals, a parameter related to de gradation and material buildup in the shaft is identified and used in the d evelopment of a monitoring tool. The results are then used as a benchmark a gainst which to test the performance of artificial neural networks as a too l for reactor diagnostics. Several neural networks are examined in this study, including Restricted Co ulomb Energy (RCE), Cascade Correlation, and Backpropagation paradigms of a rtificial neural networks. RCE is selected due to its unique design and spe ed, Backpropagation is selected because it is widely used and well accepted in the neural network research community, and Cascade correlation is selec ted because it overcomes some of the problems associated with the Backpropa gation paradigm. Similar study is performed using the Adaptive Resonance Th eory (ART) family of neural network paradigms. The results of this study indicate that artificial neural networks are simp ler techniques for pattern recognition than noise analysis techniques such as the one introduced here. Neural networks do not require prior fault rela ted parameter identification; they generate their own rules by learning fro m being shown examples. On the other hand, noise analysis and regression mo deling can provide very sensitive techniques for monitoring of a detected p roblem in a component. (C) 2001 Elsevier Science Ltd. All rights reserved.