A neural-network approach for acoustic leak monitoring in pressurized plants with complicated topologies

Abstract

A neural-network approach has been developed for localizing leakages and estimating the leak rate in pressurized plants with complicated three-dimensional structures. Results are presented from experiments with simulated leaks at a VVER-440 reactor vessel head. As features for characterizing the occurrence and the location of a leak, RMS values of acoustic emission sensors and coherence values and power spectra of microphone signals were used. Three-layer perceptron networks were found to be best suited for leak localization and for estimation of leak rates. However, the estimation of leak rates required an additional neural network because a different normalization procedure was necessary for extracting features from the RMS values of the acoustic emission sensors. Perceptron networks with continuously valued outputs corresponding to the coordinates of the leak positions were useful for classifying even positions which had not been offered during training.