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

This paper introduces a new tool based on a traditional noise analysis techniques for monitoring reactor components' signal condition. It also presents the performance of artificial neural networks for pattern recognition to the same set of reactor signals and provides a comparison of these two techniques. Reactor pump signals from the Experimental Breeder Reactor (EBR-II) are utilized here. Collected signals such as pump power, pump speed, and pump pressure are obtained from already installed sensors in the reactor. The signals utilized are collected signals as well sis generated signals simulating the pump shaft degradation progress. From the study of time series analysis and regression modeling of these signals, a parameter related to degradation and material buildup in the shaft is identified and used in the development of a monitoring tool. The results are then used as a benchmark against which to test the performance of artificial neural networks as a tool for reactor diagnostics. Several neural networks are examined in this study, including Restricted Coulomb Energy (RCE), Cascade Correlation, and Backpropagation paradigms of artificial neural networks. RCE is selected due to its unique design and speed, Backpropagation is selected because it is widely used and well accepted in the neural network research community, and Cascade correlation is selected because it overcomes some of the problems associated with the Backpropagation paradigm. Similar study is performed using the Adaptive Resonance Theory (ART) family of neural network paradigms. The results of this study indicate that artificial neural networks are simpler techniques for pattern recognition than noise analysis techniques such as the one introduced here. Neural networks do not require prior fault related parameter identification; they generate their own rules by learning from being shown examples. On the other hand, noise analysis and regression modeling can provide very sensitive techniques for monitoring of a detected problem in a component. (C) 2001 Elsevier Science Ltd. All rights reserved.