DeWDRoP: An Automated Framework for Detecting and Characterizing Welding Defects Using High-Density Point Clouds

Abstract

[Abstract] A stringent, stable, and accurate system for the detection of welding defects is essential in Industry 4.0 manufacturing environments, aiming for near-zero defects. This work introduces DeWDRoP, an offline system for detecting and characterizing welding defects using high-density point clouds (between 20 and 30 million points, with a resolution over 130 μ m). We propose a novel theoretical framework based on two key innovations: region-based seam extraction (RBSE), a segmentation method tailored to isolate welding seams (WSs) from high-density point clouds, and tolerance-based deviation analyzer (TBDA), a method that quantifies local deviations for geometry tolerance evaluation. These methods improve the identification, classification, and characterization of defects. The system is designed to augment the capabilities of expert nondestructive testing (NDT) inspectors, facilitating more precise and consistent evaluations. Experiments on real weld datasets confirm the framework’s suitability for high-quality manufacturing environments. It accurately identifies spatters, pores, and fillet weld asymmetries, achieving a recall rate of 98.48% and a Type II error of 1.52%.