A Lightweight Edge Intelligence Method for Student Engagement Pattern Recognition in Blended College English Education

In blended College English teaching, student engagement is characterized by cross-stage continuity, multidimensionality, and contextual fluctuation. Traditional questionnaires, teacher observations, and platform statistics are insufficient for continuous and fine-grained identification. To address problems such as the separation of online and offline data, unstable modality quality, and high edge-deployment cost, this paper proposes a lightweight edge intelligence method for student engagement pattern recognition. A five-dimensional engagement framework involving behavioral, cognitive, emotional, social, and agentic engagement is constructed. Classroom videos, WiFi CSI, platform logs, questionnaires, teacher evaluations, and language-task performance are integrated to generate multi-source soft labels. BEE-LiteNet is designed to implement five-dimensional score regression, overall-level classification, and engagement-pattern recognition through graph-temporal visual encoding, CSI and log sequence modeling, quality-aware fusion, and cross-stage attention. Knowledge distillation, structured pruning, and INT8 quantization are further adopted for compression and deployment. Experimental results show that the proposed method achieves a level Macro-F1 of 93.1%, a pattern Macro-F1 of 90.6%, and an MAE of 0.062, with 6.9M parameters and an edge inference latency of 14.8 ms, thereby supporting formative assessment and personalized teaching intervention.