Automated weld defect segmentation from phased array ultrasonic data based on U-net architecture

Abstract

Ultrasonic inspection is an environmentally friendly and easily deployable nondestructive testing (NDT) method widely used for defect detection of critical components in the industry. Phased array ultrasonic testing (PAUT) is one of the most advanced ultrasonic inspection methods, which gives volume inspection with increased resolution and coverage, improving inspection efficiency. Because of the weld structure echoes and the abstract nature of ultrasound images, especially facing meters and feature-changing weld joints with different thicknesses and welding methods in shipbuilding, the analysis of the PAUT weld data still relies on experienced rater random inspections. Rating complex welded products process is lengthy, costly, and prone to introduce human error during the manual rating but challenges automatic detection. To automatically segment defects in PAUT data, this work shows a combination of PAUT data of ship weld and three-dimensional (3D) U-net architecture. Combining PAUT imaging principles with welding and scan processes, a PAUT volumetric image dataset, including different thicknesses and scan angles, is established. We pioneered the application of 3D U-net architecture to segment defects in PAUT volume data. We found that U-net architecture with two encoding stages will perform better in segmenting defects in PAUT data, and region-based loss mainly improves the accuracy. Furthermore, a lightweight U-net architecture containing skip-connection and residual blocks is proposed with precision and efficiency improvement. The validation results show that the proposed U-net architecture offers a feasible solution to the problem of segmenting defects from PAUT data with a Dice accuracy of 90.9 %. Segmentation results help to locate and measure defects. This method makes locating and sizing defects in PAUT weld data possible within a fraction of a second.