This thesis presents a data engineering pipeline for robust semantic segmentation in welding scenes. The work is motivated by an industrial computer-vision problem: raw welding-video collections are large, repetitive, heterogeneous, and expensive to review manually, while model quality depends strongly on how videos are selected, organized, labeled, and staged before training. Instead of treating data preparation as secondary preprocessing, the thesis studies it as the main technical contribution. The goal is to convert welding videos into traceable and balanced training evidence for segmenting structures such as the weld pool, welding arc, and wire stickout. The implemented system covers the path from video intake to downstream evaluation.