Figure skating as a representative of fine-grained action has the characteristics of large intra-class variance and small inter-class variance. Accurate identification of figure skating actions plays a pivotal role in advancing fine-grained action recognition research. In recent years, skeleton-based fine-grained action recognition has witnessed notable advancements. Contrasting with traditional RGB-based methods, skeleton-based techniques offer heightened resilience to variations in lighting, camera angles, and backgrounds. Leveraging Graph Convolutional Networks (GCN), we effectively extract spatial features from the inherent topology of the human skeleton. Consequently, we have adopted a skeleton-based approach in conjunction with GCN for recognizing figure skating actions, showcasing the efficacy of this methodology. Simultaneously, we introduce an effective data preprocessing and augmentation strategy tailored to the skeleton-based GCN model. This strategy encompasses two key techniques: Centralization and Region-cutout. Centralization addresses the challenge of losing centroids, which can lead to significant shifts in the coordinates of skeleton joints. Region-cutout randomly masks selected joint points, thereby introducing diversity into data features and bolstering the model’s robustness. To further improve the recognition performance, we used three models with different feature learning methods for feature fusion to improve fine-grained feature recognition. Experimental results conducted on the FSD dataset validate the efficacy of these approaches.