An Edge Intelligence Method for Learning-State Recognition and Personalized Feedback in Higher Education

To address fluctuations in multimodal data quality, client heterogeneity, and insufficient feedback timeliness in university classrooms, this paper proposes a cloud-edge-device collaborative method, HE-EdgePF. The method integrates visual, audio, textual, and platform-behavior features through quality-aware dynamic gating and a lightweight temporal unit, conducts federated training with a shared backbone, personalized adapters, and sparse quantized aggregation, and establishes a edge-cloud-teacher hierarchical feedback mechanism. Experiments were performed on DAiSEE and the HE-MMLA dataset containing 486 students, 12 clients, and 118320 valid windows. The results show that the model achieves a Macro-F1 of 88.3%, outperforming MM-Transformer by 2.6 percentage points, while reducing the expected calibration error to 0.032. Even with 75% modality missingness, the Macro-F1 remains 78.4%. The per-round communication volume is 10.7 MB, and the inference latency on Jetson Orin Nano is 28.6 ms. In an 8-week teaching experiment, the HE-EdgePF group achieved a post-test score of 84.7, a normalized learning gain of 0.46, a learning-engagement index of 0.81, and a cognitive-load score reduced to 3.62. These results demonstrate the effectiveness of the proposed method in recognition accuracy, robustness to missing modalities, edge-deployment efficiency, and personalized instructional feedback.