NEURAL NETWORKS APPLIED TO FAULT DETECTION USING ACOUSTIC EMISSION

The mathematical formalism of Neural Networks is useful when considering approximate or noisy data, generalization of situations and pattern classification. In this report an application of the Hopfield Neural Network model [ 2 , 3 ] to the classification of faults in tubes using acoustic emission is presented. The way in which the synaptic matrix is obtained has been modified so that the pseudo-orthogonality restriction required in Hopfield's model is not neccesary [ 5 ]. Furthermore, as a consequence of the fact that the neural actualization is made at random, the results can differ when recovering information. Therefore an actualization of several copies of the same network [ 6 ] is proposed obtaining a global answer based on the theory of fuzzy sets applied to decision making [ 8 , 9 ]. Experimental results performed at an Acoustic Emission Laboratory are presented evaluate the location of an acoustic emission source in a tube which simulates a fault.