Benchmarking U-net and Swin-U-net models for automatic flaw detection in ultrasonic testing for non-destructive testing applications

Abstract

This research focuses on the development and evaluation of deep convolutional neural networks (CNNs) for flaw detection and sizing in ultrasonic testing (UT) data from a nuclear power plant (NPP) reactor pressure vessel (RPV) wall mock-up test block. Two models were employed, U-net and Swin-U-net, trained to extract flaw signals. A novel data reconstruction method combining synthetic aperture focusing technique (SAFT) and total focusing method (TFM) was used to provide a three-dimensional view of the data in Napari image viewer.

The original goal was to improve the models’ code, but their results exceeded expectations. Instead, it was decided to evaluate the models’ outputs. The provided flaw data from the test block (true state) was compared to the results of the two models. True state data was matched with the data from models using similarity transformation. Both models demonstrated high accuracy in flaw detection, particularly in cladding and cladding boundary, where all flaws were found with high confidence, regardless of the probe and flaw orientation.

While U-Net exhibited higher sensitivity, it produced more false positives. Swin-U-Net showed cleaner outputs, simplifying evaluation. Flaw sizing accuracy of defects in the region of interest was generally good, with Swin-U-Net being more accurate. Near the boundaries, sizing was less reliable, and some flaws were barely registered. Data from one of the two scan directions allowed the models to produce significantly clearer outputs and make better sizing predictions