A novel neural network approach in nondestructive testing

This paper presents a neural network (NN) approach to nondestructive testing (NDT) applications. The hidden layer of the proposed NN model includes both a block of neurons of the standard sigmoidal type and a set of Gaussian neurons which implement a set of fuzzy rules that have been proved to be able to describe the problem under study in an approximate but reliable way. The final model is then a hybrid processor which exploits the advantages of both techniques. The effectiveness of the proposed approach is tested on two example problems of interest in electromagnetics. In the first problem, the prior knowledge on the problem needed to design the fuzzy part is supposed to be obtained by an expert. The second example analyzes the case in which there are no ideas about the problem to be solved. In this case an adaptive neuro-fuzzy inference system is introduced in which the rules are extracted from a set of input-output pairs by a learning procedure. In both cases the performance of the hybrid approach is superior to that of both a standard NN and a simple fuzzy inference system