Skeleton-based action recognition for manufacturing assembly task through graph convolution network

In modern manufacturing, human participation in assembly processes is essential, despite advancements in automation. However, accurately recognizing human actions in these environments presents challenges due to complex spatial–temporal dependencies and dynamic joint relationships. Graph Convolution Networks (GCNs) are utilized widely for action recognition, but they have poor accuracy for modeling long-range node correlations. Also, current GCNs have limitations in extracting various features due to utilizing the same pattern extraction for all frames. To overcome these issues, this study presents a novel approach to skeleton-based action recognition for manufacturing tasks using a Dual-Attention Graph Convolution Network (DAGCN). The proposed model integrates a Parallel Attention-Graph Mixer (PAGM) and Temporal–Spatial Attention Integrator (TSAI), enhancing the capture of both global and local joint relations and addressing the dynamic nature of skeletal joint relationships. Extensive evaluations on benchmark datasets, including HA4M that specifically designed for assembly tasks, NTU RGB+D, Northwestern-UCLA, and NTU RGB+D120, reveal the superior performance of DAGCN over state-of-the-art methods in terms of accuracy and computational efficiency. Experimental results demonstrate that DAGCN outperforms state-of-the-art methods, achieving a Top-1 accuracy of 89.0% on the HA4M dataset. The results validate DAGCN’s effectiveness in recognizing fine-grained human actions in industrial settings, contributing to improved efficiency and safety in human–robot collaboration. The proposed model offers a scalable and computationally efficient solution for intelligent assembly monitoring and automation in smart manufacturing systems.