Socialmedia is highly active and full of rumors while the information is widely propagation. Mastering the process information of rumor propagation plays an important role in promoting automated rumor detection. However, the existing methods ignore the rumor evolution and development process, resulting in the loss of details, which is not con-ducive to extracting key information in rumor detection. In this paper, we propose a novel hierarchical dynamic graph convolution network (H-DynGCN-Enh) to build a dynamic propagation graph for each news differentiation, while taking into account the bi-directional features of rumor propagation and dissemination, fully extracting the key information. In addition, we introduce attention-based graph pooling to design a feature enhancement strategy to avoid redundant information and increase the interaction between comments and original news. Finally, we use the multi-head self-attention mechanism to adjust the weight of dynamic information to obtain global information for detection. The results of two open real-world datasets demonstrate that our model has exceeded the most advanced baseline, the effectiveness of each module has also been fully proved in the ablation experiment.