High-Speed Monitoring and Control of Forming Defects in Coarse-Wire MAG Welding Based on Multi-Frame Spatio-Temporal Neural Networks

To mitigate undercut and humping defects induced by arc instability in high-speed coarse-wire MAG welding, this paper proposes a lightweight multi-frame spatio-temporal convolutional network for adaptive defect control. The network takes sequential multi-frame images as input, adopts a cross-loss function, and incorporates a reward–penalty mechanism to enhance defect recognition accuracy. Model light-weighting is realized through spatio-temporal decomposed convolutions, a depth-wise separable bottleneck, and global adaptive pooling, resulting in a 61% reduction in parameter count. Experiments conducted under different groove configurations and welding angles demonstrate that the overall performance is optimal when the input length is set to five frames. Across four welding states, the model achieves a macro-averaged F1 score of 0.8315 and an AUC of 0.965, with a recall of 0.82 for critical defects (undercut/humping). The recognition accuracy for all four welding states exceeds 90%, with an overall loss of 0.05 and an average response time of 1.5ms. Furthermore, the closed-loop adaptive control exhibits stable behavior, ensuring real-time performance, strong robustness, and industrial deploy-ability. This work provides practical guidance for the design of future adaptive welding systems targeting high-speed coarse-wire MAG processes.