Flaw sizing with plane wave imaging (PWI) – total focusing method (TFM) and deep learning for reactor pressure vessel

Abstract

Developments in machine learning and deep convolutional networks (CNNs) have enabled automated assessment of nondestructive evaluation (NDE) data. Ultrasonic data is especially challenging for automated evaluation due to its complexity, multi-channel nature, and volume. Typical flaw signals have low signal to noise ratio, particularly diffraction signals critical for sizing. This study presents a proof-of-concept on the application of deep CNNs, specifically U-net and Swin-U-net, for flaw sizing in ultrasonic data from a nuclear test block with realistic flaw simulations. The segmentation CNNs extract flaw signals, enabling the identification of the deepest crack tip echo, mimicking human inspection. This mimics the process used by human inspectors. Two distinct CNNs are trained: U-net and a transformer-based Swin-U-net. A novel data reconstruction method is proposed that combines plane wave imaging (PWI), synthetic aperture focusing (SAFT) and total focusing method (TFM) to provide a unified volume reconstructed view. Both networks provide good segmentation performance allowing accurate sizing, despite noisy data and complex flaw signals.