The student mental health early warning system based on multimodal learning analysis integrates text, behavior and physiological data to design and implement an efficient mental health risk assessment scheme. The system adopts a fusion method of adaptive attention mechanism and dynamic weight allocation, combined with Transformer architecture and time series modeling, to build an accurate early warning model. Experiments verify the excellent performance of the system in terms of accuracy, F1 score and AUC, which is significantly better than single modality models and traditional machine learning methods, especially in the identification of high-risk students. Data analysis reveals the difference in contribution of each modality and verifies the robustness of the dynamic fusion strategy. In terms of system implementation, modular design and real-time reasoning capabilities ensure its practicality in college mental health management, and provide innovative technical support for intelligent psychological intervention.