A wearable device data-driven artificial intelligence framework for personalized training guidance and its application in physical education teaching

Wearable sensing creates an opportunity to tailor physical education (PE), but a deployable system must reconcile heterogeneous sampling rates, individual physiological variation, safety constraints, and classroom-scale latency. This study presents a scenario-oriented, full-chain framework that links multi-source sensing, edge preprocessing, spatiotemporal feature fusion, multi-task assessment, and multi-objective Actor-Critic decision support. Its distinguishing design is the coupling of individual baseline calibration with explicit objectives for training effectiveness, injury-risk control, and plan adaptability; the contribution is therefore an integrated PE-specific design rather than a new learning paradigm. Wearable inputs are acquired at different sampling rates, including 1, 50, and 100 Hz. On the self-built campus dataset, the framework achieved an ability-assessment MAE of 2.18 (R 2 = 0.962), risk-warning recall of 95.8%, and plan adaptability of 94.7%. The system supported 200 concurrent users with a mean response time of 185 ms. In an 8-week controlled teaching study (n = 120), the intervention group showed larger improvements than the control group across the reported fitness outcomes (P < 0.05), including a 32.4% higher overall improvement efficiency; injury incidence was 1.7% versus 13.3%. These findings support the feasibility of the integrated framework in the evaluated university PE setting. Teacher awareness of allocation and the single-institution design limit causal interpretation and generalizability.