HEAT-TREATMENT CLASSIFICATION OF STEELS WITH NONDESTRUCTIVE EDDY CURRENT TESTING USING NEURAL NETWORKS

Abstract

Eddy current-nondestructive techniques are increasingly present in industry because of the growing quality control demand. Three steel piece classifiers implemented with Artificial Neural Networks (ANNs) are compared in this article. An inductive sensor connected to a currents equipment was used to acquire the impedance measurements over two steel piece sets with different heat treatments. Two monofrequency ANNs and one multifrequency ANN were used to perform the classification task. As a result (i) the monofrequency ANNs obtained a 90% and the multifrequency ANN obtained a 99.9% of accuracy rate, (ii) theoretical computing workload of the best multifrequency ANN is between 33% and 50% lower than of the best monofrequency ANN, and (iii) the multifrequency ANN execution time was 22% smaller than the monofrequency ANN's time. Performed experiments suggest that (i) the classification systems based on the multifrequency ANN reach higher accuracy rates than the monofrequency ANN-based ones, (ii) the higher number of input variables of the multifrequency ANNs with respect to the monofrequency ANNs does not imply higher computing workload, and (iii) the higher number of input variables of the multifrequency ANNs with respect to the monofrequency ANNs does not imply higher execution time to perform one steel piece classification.