Dynamic interactions between human joints and bones convey significant information for skeleton-based abnormal gait recognition. Existing graph convolutional networks (GCNs)-based methods either only consider the locomotion information of the joints or treat the motion information of joints and bones independently, failing to explore the implicit dynamic interactions between joints and bones effectively. These interactions also contain rich and useful abnormal gait representation information. In this work, we propose a novel adaptive graph convolutional fusion network (AGCFN) for skeleton-based abnormal gait recognition. The joint motion information and bone motion information are modelled as a joint spatiotemporal gait graph and a bone spatiotemporal gait graph, respectively. Our AGCFN is designed to explore the interaction information between joints and bones through learning the inter-graph relationships between the above two gait graphs, so as to obtain more discriminative gait feature representations. Extensive experiments on our abnormal gait dataset demonstrate that the generalization performance of our model exceeds the state-of-the-art by a significant margin.