In response to the frequent occurrence of student sports injuries and the lack of real-time warning methods in campus sports scenarios, this paper proposes a student sports risk warning model based on multi-source data fusion. A multi-source acquisition system incorporating wearable physiological signals, video-based human postures, and environmental sensors is constructed, and an adaptive weighted early fusion strategy is proposed to dynamically adjust the contribution of each data source. At the feature extraction level, a multi-scale temporal convolutional network with a multi-head attention mechanism (MHA-MSTCN) and a residual contraction module is designed to capture multi-time-scale risk patterns and suppress motion noise. Furthermore, a dual-path heterogeneous fusion architecture is proposed, using graph attention networks to model joint coupling abnormalities in human motion chains, gated recurrent units to capture temporal drift of physiological parameters, and cross-branch cross-attention to achieve collaborative discrimination of spatial and temporal features. Unsupervised clustering of risk levels is achieved using K-means++ and the silhouette coefficient, and the warning trigger is completed by combining individualized dynamic thresholds and a multi-level delay-tolerant rule engine. Experimental results on a self-built dataset (120 subjects, 53.6 hours of sports data) show that the model has an accuracy of 92.3%, a recall rate of 89.7%, an AUC of 97.1%, a warning lead time of 2.31 seconds, a single-frame inference delay of 7.6 milliseconds, and is significantly superior to baseline models such as LSTM, CNN, and Transformer. Ablation experiments verify the effectiveness of each core module. This model provides a feasible technical solution for intelligent campus sports risk prevention and control.