Multi-label welding defects refer to the simultaneous occurrence of multiple defect types, such as porosity and slag inclusion, during complex robot arc welding processes. Online identification of multi-label welding defects using molten pool images presents a significant challenge due to the coupling of multiple defect types, intra-category variations and inter-category resemblances, as well as limited computational capacity in real welding scenarios. Existing methods for online welding defects identification primarily focus on independently occurring single-label defects. However, research on identifying multi-label welding defects remains highly limited. To tackle these challenges, this paper proposes a novel lightweight network named WeldViT to accurately recognize coupled multi-label defects. Firstly, the WeldViT network based on the EfficientViT block is designed to model and learn the feature dependencies between multi-label welding defects. Secondly, an overlapping patch embedding module is improved by integrating an efficient channel attention mechanism, enhancing the WeldViT network’s capability to learn low-level visual representations. Subsequently, a welding molten pool image dataset containing ten categories of multi-label compound defects is established to assess the WeldViT network’s performance. Finally, the WeldViT model is thoroughly analyzed and compared against seven advanced lightweight models. The results indicate that the WeldViT model achieves superior performance over other models in terms of identification accuracy, model complexity, and processing speed, providing high reliability for the online identification of multi-label defects in real welding scenarios.