Under the background of digital teaching transformation in colleges and universities, students ‘learning behavior presents multi-source, dynamic and implicit characteristics, and a single platform log is difficult to support accurate analysis and personalized support. This paper constructs a college students ‘behavior analysis and personalized learning support system based on multimodal data fusion. It collects learning platform logs, classroom visual behaviors, discussion texts, evaluation records, resource access and terminal signals, and constructs a unified learning behavior sequence through timestamp alignment, missing repair, anomaly suppression, robust normalization, text semantic coding and sliding time Windows. In the model layer, CNN local behavior feature extraction, BiLSTM bidirectional temporal dependency modeling and Attention key modality weighting mechanism are fused to realize the recognition of concentration, participation status and learning risk. Personalized feedback is generated by combining state prediction, risk scoring, resource matching and profile update. Based on the 16-week learning data of 1280 students, 24860 sequence samples were formed. The results show that the Accuracy of the model reaches 95.8%, the F1 is 95.1%, the AUC is 0.972, the recommendation accuracy is 92.6%, and the average response delay is 149 ms. The research provides technical reference and practical significance for college students ‘learning status perception, risk early warning and personalized support.
