DEEP LEARNING BASED 2-D INVERSE SCATTERING ANALYSIS FOR ESTIMATION OF POSITION AND SIZE OF A DEFECT IN SOLID

Abstract

In the field of non-destructive evaluation using ultrasonic and electromagnetic waves, several researches on inverse scattering analysis to reconstruct a defect in materials have been studied since early times. In general, inverse scattering analysis methods using the Born and Kirchhoff approximations have been formulated using difficult mathematical theories. However, there are still many problems to be improved from the viewpoint of practical application in actual nondestructive evaluation. In this study, we attempt to develop a method to reconstruct the shape and location of a defect from received scattered waves using deep learning, which has been attracting attention as a basis for AI(Artificial Inteligence) in recent years. However, in this paper, as a first step in this type of study, the received scattered waves are simulated waveforms created by the time domain boundary element method. The effectiveness of the proposed method is discussed by showing the results of estimating the location and size of a defect in solids by deep learning using the scattered waves from a defect as supervising learning data. The deep learning results indicate that the accuracy of defect shape reconstruction decreases as the defects move away from the received elements.