Abstract: At present, the monitoring video in coal mine production area is vague, the type of personnel actions is complex, and the accuracy of conventional action recognition methods is low. In order to solve the above problems, a coal mine personnel action recognition method based on dynamic attention and multi-layer perception graph convolutional network (DA-GCN) is proposed. The Openpose algorithm is used to extract the key points of the human body in the input video to obtain the key point information of the human body in 3 dimensions and 18 coordinates, reducing the interference of fuzzy background information. The spatial characteristics of the key points of the human body is extracted by dynamic multilayer perception graph convolution network (D-GCN), and the temporal characteristics of the key points of the human body is extracted by temporal convolutional network (TCN) so as to improve the generalization ability of the network for different actions. The dynamic attention mechanism is used to enhance the network's attention to action key frames and key skeletons to further mitigate the impact of poor video quality. The softmax classifier is used for action classification. Through scene analysis, underground actions are classified into five types, including standing, walking, sitting, crossing and operating equipment. The method constructs a Cumt-Action data set that applicable to coal mine scenes. The experimental results show that the highest accuracy rate of D-GCN in the Cumt-Action data set is 99.3%, and the highest recall rate is 98.6%. Compared with other algorithms, DA-GCN has higher recognition accuracy in both the Cumt-Action data set and the public data set NTU-RGBD.