Plate additive, seam-tracking technology based on feature segmentation

This study aims to investigate the automatic welding process in fusion welding and additive manufacturing, and proposes a plate additive, seam-tracking technology based on active vision. However, noise interference, such as arc light and spatter during welding, poses significant challenges to vision-based feature extraction. This paper uses a self-designed line-structured light vision sensor to capture weld images and applies the concept of deep learning feature segmentation for weld information extraction. The Pixel-wise Spatial Pyramid Network (PSPNet) is used as the fundamental framework for feature extraction, allowing for the simultaneous extraction of the centre line of the laser stripe and weld feature points. To address extreme sample imbalance challenges, online hard example mining (OHEM) and Dice Loss are employed as the loss functions. The results demonstrate that the algorithm's extracted feature points exhibit an error of no more than 0.70 mm compared to manually marked feature points in each dimension of the weld pass. Additionally, the average height error is not more than 0.50 mm, and the frame rates surpass 30FPS, meeting the requirements of high precision, stability and real-time performance in additive manufacturing.