Laser welding is a key technology for e-mobility manufacturing, demanding extreme precision and zero-defect tolerance. However, the need for omnidirectional tracking of complex trajectories and the harsh, dynamic process environment often exceed the capabilities of conventional monitoring. This study presents a unified, AI-driven coaxial vision pipeline that integrates robust real-time seam tracking and surface quality evaluation into a single optical setup.

The first stage employs a hybrid deep learning approach for seam tracking. By decoupling semantic topology isolation from analytical geometry, the system achieves sub-pixel accuracy and high-speed inference on industrial edge hardware, effectively filtering optical artifacts such as scratches. Simultaneously, the second stage performs near-zero-latency surface inspection. To ensure robustness, a spatiotemporal tracking strategy first filters out transient process noise, such as flying spatters. The cleansed image data is then passed to a lightweight multi-class segmentation model, followed by an algorithmic evaluation to extract precise quality metrics from the solidified weld seam.

This methodology eliminates the need for redundant sensor hardware by synergizing adaptive control with immediate quality assessment in an edge-deployable framework. Comprehensive experimental trials have been completed, validating the system’s sub-pixel precision alongside strict real-time capabilities (inference < 5 ms). The presented results comprehensively benchmark the pipeline against established industry standards, demonstrating a highly scalable solution for next-generation manufacturing.