Ultrasonic inspection of austenitic steel welds remains a complex endeavor despite significant progress in ultrasonic data acquisition and imaging methods including the multiview Total Focusing Method (TFM). However, the optimal TFM view can only be chosen with a priori knowledge of the defect position, size, and type. This selection becomes even more complex when the inspected material is anisotropic. This study aims to develop a method to provide information about the defect by processing the Full Matrix Capture (FMC) data before any imaging reconstruction is performed. The FMC provides a rich input for deep learning models but also results in high-dimensional, computationally demanding datasets. To overcome this issue, strategies for data reduction are compared: the Principal Component Analysis (PCA) and the reduction of the FMC by selecting a limited number of A-scans. The dataset comprises simulated FMCs of two-dimensional cracks inside rectangular isotropic steel blocks. The results suggest that both data reduction methods are relevant and that it is possible to provide useful input to the selection of the optimal TFM view. In the next step, the inspection scenario will be further complicated by adding other types of defects and introducing an anisotropic steel weld.
Antile et al. (Wed,) studied this question.