Through two-layer aggregation on the edge server side and the cloud server side, hierarchical federated learning effectively improves the learning efficiency. However, the heterogeneity of data volume, computing resources and communication resources of edge nodes leads to the phenomenon that fast nodes wait for slow nodes in the training process, which results in the decline of training efficiency. To solve this problem, this paper proposes a semi-asynchronous hierarchical federated learning method based on computing resource coordination. Firstly, a grouping strategy is proposed to balance computing resources, and the edge nodes are grouped to realize the matching of computing load and computing resources. Then, a collaborative training method is designed to implement the collaborative training of devices based on task offloading within the group to achieve the matching of computing load and computing resources of device granularity. Finally, considering the instability of communication between edge server and cloud server, an adaptive asynchronous cloud aggregation strategy is designed to participate in the group set of cloud aggregation through deep reinforcement learning decision-making, so as to reduce the waiting time and optimize the benefits of cloud aggregation.