Abstract: Aiming at the problems of insufficient kinematic interpretability of graph topology modeling, lack of biomechanically meaningful feature representation, and weak anomaly localization ability in existing abnormal movement pattern detection methods, a graph convolution-based method for movement anomaly recognition was proposed. A kinematic feature representation was constructed by integrating pose features, inter-frame dynamic features, and steadiness features. The pose representation was decoupled to reduce redundancy and improve viewpoint robustness. Dynamic features derived from joint angular velocity captured instantaneous motion changes, while displacement variance over a time window quantified motion stability. Furthermore, a multi−topology graph convolution module was designed. It incorporated principles from kinematic chain theory to define multiple skeleton topologies. A topology−aware self−attention mechanism dynamically adjusted adjacency matrix weights based on input features, and graph convolution aggregated node information. Experiments on three public datasets demonstrate that the proposed method achieves superior abnormal recognition accuracy for most actions, reaching 93.93% on dataset UI−PRMD.Ablation studies confirm the effectiveness of each feature component. Visualizations show that the model accurately localizes abnormal joints, providing interpretable recognition results.