Location of continuous AE sources by sensory neural networks

Abstract

A brief classification of location problems which appear in acoustic emission (AE) analysis is given. Empirical treatment of corresponding inverse problems is explained and applied to location of sources which generate continuous AE signals. A continuous AE phenomenon is treated as a stochastic process which is represented by the source coordinates and the correlation function of the emitted sound. The empirical model of AE phenomenon is formed based on a set of samples. The model includes a network of AE sensors and a neural network (NN). During formation of the model, the AE signals are generated by sources at typical positions on a specimen. Recorded ultrasonic signals are transmitted to the NN together with the source coordinates. The first layer of NN determines the cross-correlation functions of signals and forms from them and source coordinates the data vectors. In the second layer, a set of prototype vectors is formed from the data vectors by a self-organized learning. After learning, the network is capable to locate the source based on detected sound. For this purpose, the sensors provide AE signals, while the NN determines the corresponding correlation function and associates to it the source coordinates. The association is performed by a non-parametric regression which is implemented in the third layer of NN.