This study aims to develop an online method for reconstructing vision-based weld-seam coordinates corrupted by arc light, spatter and fumes, with the aim of providing stable coordinate inputs for robotic trajectory correction.
A RANSAC-based quadratic model is first used for robust small-sample initialization. Subsequent coordinates are processed using a coordinated-turn extended Kalman filter and Mahalanobis-distance outlier gating. The corrected points are then reconstructed by segment-wise cubic B-splines with endpoint and tangential-direction constraints. The method is evaluated using controlled outliers, naturally occurring welding disturbances, closed-loop TCP tracking, end-to-end latency and quantitative weld-quality measurements.
Under naturally occurring welding disturbances, the method achieved a precision of 95.14%, a recall of 93.84% and an F1 score of 94.49% for outlier identification. Coordinate RMSEs for straight, arc and S-curve paths were reduced to 0.31, 0.38 and 0.46 mm, respectively. Closed-loop TCP tracking RMSE decreased by 46.8%–48.7%, and the complete vision-to-control chain had an average latency of 60.5 ms for regular frames and 63.6 ms when B-spline smoothing was triggered, both below the 100 ms sampling period. Weld-width variation and centerline deviation were also reduced, whereas the change in spatter count was not statistically significant.
The study is validated only on bead-on-plate welding of 5-mm Q235 steel at 10Hz visual sampling and 8mm/s speed, excluding assembled joints and multilayer welds. The constant-velocity Z-axis model may be inadequate for highly curved 3D seams. Future work should extend to adaptive noise estimation, higher speeds, complex spatial paths and real butt/fillet joints to generalize the framework.
The method reduces tool center point (TCP) tracking RMSE by ∼47% and improves weld-width consistency and centerline deviation, lowering rework and scrap costs. As a real-time add-on for existing robotic welding systems, it enables reliable automation under severe arc-light, spatter and fume disturbances. Its low latency (63.6ms) and modular architecture offer strong commercial potential for retrofitting industrial robots, attracting investment for cost-effective quality enhancement.
Consistent weld quality reduces structural failures, enhancing public safety in infrastructure and manufacturing. Automating seam tracking decreases welders’ direct exposure to intense arc light, toxic fumes and high heat, improving occupational health. It also lowers skill entry barriers while creating higher-level supervisory and maintenance roles, fostering safer, more attractive working environments and supporting sustainable industrial development.
The contribution lies in a staged online architecture that integrates robust initialization, turn-rate state estimation, uncertainty-based outlier gating and finite-buffer smoothing for streaming weld coordinates. Its effectiveness is demonstrated not only at the coordinate level, but also through closed-loop robot motion, system-latency and weld-quality evaluation.
