Innovative Machine Learning Approaches for Nondestructive Evaluation of Materials

This paper deals with a machine learning framework dedicated to nondestructive testing applications, in view of flaws detection and characterization. A supervised learning strategy is used on a training set made of characteristic features, extracted from eddy current testing (ECT) and ultrasounds testing (UT) signals. The approach is first presented and the key role of the feature extraction by means of Partial Least Squares is highlighted. Then, the performance of the proposed data-fusion approach, in terms of both localization and characterization, is compared to that of similar approaches exploiting one inspection technique only.