Deep Learning for Scattered Waves Obtained by Time-domain Boundary Element Method and an Attempt to Classify Defect Types

Abstract

Laser ultrasonic visualization testing (LUVT) can visualize elastic wave propagation fields on the surface of a test specimen, and save them as image data. In general, LUVT inspectors make a judgement on the status of a defect by viewing the images obtained by LUVT. If AI can judge this process, instead of inspectors, LUVT might be a more effcient technique. Therefore, in this research, the deep learning, which is basis of AI creation, is carried out to determine the presence or absence, and type of a defect in images. The images required for the deep learning with the convolutional neural network (CNN) are numerically prepared using the time-domain boundary element method for simplicity. As numerical examples, some defect type classification problems are solved by using the created learning model. In addition, Grad-CAM is used to confirm the regions that the learning model uses to judge the defect types. The results for how the created deep learning model classifies the defect types in images may be useful for future application of AI to LUVT.