In industrial scenarios, incidents caused by unautho-rized activities such as smoking and phone usage by personnel have been occurring frequently, resulting in irreparable losses. Real-time detection and timely warning of unsafe actions among on-site personnel have significant implications for ensuring fac-tory safety and protecting the lives of individuals. In this paper, we propose an industrial scenarios personnel unsafe action recognition algorithm based on improved ST-GCN, adding the pose estimation algorithm (Alphapose) to improve the spatial-temporal graph convolutional network ST-GCN to enhance the action recognition accuracy; introducing the adaptive attention mechanism module SENet to enhance the feature extraction capability of the model for unsafe action; The YOLOX network is used to pre-process the input photos in order to reduce the influence of the factory's complex background on action recognition and increase the algorithm's recognition precision for employee activity. To address the limitation of limited data availability for industrial personnel unsafe action, a self-constructed dataset specific to industrial scenarios is established to alleviate the constraints on algorithm generalization. Based on this dataset, ablation comparative experiments are designed to analyze the improvement effects of different enhancement operations on the original algorithm. The experimental results demonstrate an overall increase in detection accuracy by 6.6 % after the proposed improvements. Specifically, the recognition accuracy for unsafe actions such as smoking and phone usage achieves 94.0 %. The experimental results confirm the effectiveness of the proposed detection algorithm enhancements in real industrial scenarios, validating their practicality and performance.