this paper introduces a deep learning-based sports training action recognition and feedback system to tackle issues such as motion deviation, unstable recognition and delayed correction in real-time training. The system uses a dual-channel perception layer to fuse RGB video and IMU data, extracts 33 skeletal keypoints and synchronized acceleration/gyroscope features, and forms a unified temporal tensor. A task decomposition module breaks down continuous movements into essential posture sequences by energy-based stage boundary reasoning and offers structured input for model learning. Based on the above, a hybrid Dilated Temporal Convolutional Network and Graph Attention Network (TCN+GAT) are used to address the problems of long-term temporal dependency and cross-joint spatial relationship. Node features contain pose coordinates and extremity IMU signals, and an attention-weighted loss gives more weight to joints with large angular excursions for faster convergence. A deviation-aware feedback mechanism maps the classification results and joint angle offsets to a dual feedback channel of voice prompts and vibration intensity, thereby forming a closed loop of “perception-reasoning-correction”. Experiments on a combined dataset of a Kinetics-400 subset and 4,760 self-collected training clips (push-ups, squats, lunges and jumping jacks) evaluate Accuracy, Angle-Score and end-to-end Latency. The accuracy of the proposed model is 92.4±0.5%, the Angle-Score is 87.9±0.6%, and the latency is ≤1.8s; it outperforms TCN-Only and TCN+LSTM baselines in both recognition and response speed. Based on the ablation studies, the graph attention module, angle refinement strategy and dual feedback design have improved both the stability and feedback hit rate. Based on the above results, it can be seen that the TCN+GAT architecture with RGB-IMU fusion can offer practical real-time guidance for sports training and provide a deployable solution for intelligent coaching systems.
