Optimized database for training neural networks used in non-destructive testing

Abstract

Neural network is applied for the non-destructive evaluation of in-material defects based on the measured or simulated data of an eddy-current testing experiment. The network is trained with the data measured for suitably selected defect prototypes, so that this selection constitute a consistent representation of the forward problem. This network performs in defect reconstruction better, than those networks trained with randomly or regularly selected defect prototypes of about the same number.