In laser welding, harsh environmental conditions such as arc light interference, spatter, and strong background noise pose significant challenges to accurate process monitoring and measurement. Traditional single-modality sensing methods are often inadequate for a comprehensive characterization of welding states and remain highly vulnerable to noise contamination. To overcome the limitations, a dynamic multimodal attention-based weighting (DMAW) network is proposed, which integrates visual and acoustic information to achieve more comprehensive welding state characterization. First, modality-specific feature extractors are trained using unsupervised knowledge distillation (KD) to learn welding-relevant semantic representations. Next, the extracted features are passed through a cross-attention module to enable intermodal interaction and suppress noise. Finally, a dynamic reliability-weighted fusion is proposed that adaptively adjusts modality contributions, thereby reducing measurement uncertainty and enhancing robustness under varying conditions. Experimental validation on a dedicated laser welding platform demonstrates that the proposed framework achieves superior accuracy and resilience in welding state monitoring, highlighting its potential as a reliable solution for intelligent industrial welding systems.
Fan et al. (Thu,) studied this question.