Neural Approaches to Blind Separation and Cumulant Analysis and Its Application to Diagnostics of Nuclear Power Plants

Abstract

The problem concerned is to explore the possibility of using artificial intelligence techniques, namely neural networks, and design the appropriate neural networkbased algorithm to detect signals of interest from multichannel data recordings. The problem finds application in diagnostic systems of nuclear power plant with liquidmetal fast breeder. The idea of a whole approach is to make an adaptive diagnostic system of acoustic monitoring of a steam generator unit. The system is based on neural network feature extraction and pattern recognition of multi-channel acoustic signals generated by a steam generator unit. In the background noise environment the diagnostic system must detect water leaks in sodium which may occur in the steam generator unit under monitoring. Unfortunately traditional linear techniques fail to solve the problem with extreme low signal to noise ratio that is the case in the application described above. The power spectra of acoustic background noise signal and leak signal are completely overlapped and there is no possibility to separate or extract signals used only secondorder techniques which are based on autoand crosscorrelation. Neural networks turn to be powerful tools in modeling underlying phenomena that are responsible for generation of real signals. Due to significantly nonlinear structure and adaptive learning behavior neural networks become very attractive and promising in solving such problems as described. Recently there appeared a new technique for multi-channel signal separation referred as Blind Source Separation (BSS) or Independent Component Analysis (ICA) treating original source signals as independent components which are then observed mixed in the multi-channel recordings. The aim is to combine the powerful of neural networks and the methods of blind separation to solve the problem of signal extraction with low signal to noise ratio and overlapped power spectra.