Abstract: A multi-scale edge graph convolutional network for fine-grained skeleton behavior recognition is proposed to address the problems of over smooth structure, single feature scale, and incomplete capture of fine-grained features in traditional deep learning based skeleton behavior recognition methods. Firstly, the dynamic time warping algorithm is used to obtain edge node features, and the spatital edge graph convolution is used to capture the features between adjacent joints, achieving dynamic representation of human skeleton features; Then connect the features of different levels in a residual manner to capture the distance dependence between joints; Finally, a multi-scale positional attention model and channel refinement operation are introduced to enhance the model's ability to capture fine-grained local features. The experiment showed that the recognition accuracy on the NTU RGB+D dataset reached 93.1% (X-Sub) and 97.2% (X-View), respectively, and the recognition accuracy on the NTU RGB+D 120 dataset reached 89.9% (X-Sub) and 90.7% (X-View), respectively. The experimental results show that the proposed model has improved the recognition accuracy compared to most methods, verifying the effectiveness of the model.