Crack Defect Characterization Using Raw Channel Data and DNN-Based Classifier

Plane-wave ultrasound imaging has been successfully utilized in various nondestructive testing applications, such as detection, localization, and sizing of surface-breaking backwall cracks in steel plates and pipes. This work addresses the problem of characterizing single-crack defects in terms of their three main attributes: lateral position, length (i.e., size), and angular tilt. We propose to use a simple multioutput classifier incorporating two deep neural networks (DNNs). Unlike other DNN-based solutions that rely on multiview beamformed image patches, our classifier takes 2D raw channel data frames as input.