Multimodal Data-Driven Intelligent Assessment System for Classroom Teaching Quality—Based on Deep Learning Methods and Practices

With the rapid advancement of intelligent sensing technologies and Electromagnetic Waves, Antennas and Propagation systems, efficient multimodal information fusion and temporal perception have become fundamental requirements for communication-oriented monitoring and intelligent decision-making. To address the challenges of heterogeneous data integration and subjective classroom assessment, this study proposes a Hierarchical Cross-modal Attention with Temporal Dynamics Perception Network (HCA-TDPNet) for multimodal teaching quality evaluation. The framework integrates Video Swin Transformer, Wav2Vec 2.0, and BERT to extract visual, acoustic, and semantic representations, while a multi-head cross-attention mechanism and a hierarchical temporal encoder composed of temporal convolutional networks and gated recurrent units enable deep cross-modal interaction and long-term temporal dependency modeling. The resulting multimodal representations are mapped to quantitative indicators of teaching objective achievement, teacher-student interaction quality, and student participation depth. Experimental validation based on public pre-training datasets and the SCB-Dataset demonstrates Pearson correlation coefficients of 0.893, 0.918, and 0.904 for the three evaluation dimensions, respectively, while reducing the assessment time of a single lesson to approximately 2 minutes, saving more than 95% compared with manual evaluation. Beyond educational applications, the proposed architecture provides an effective framework for heterogeneous information fusion, real-time multimodal perception, and adaptive semantic interaction, offering valuable engineering references for intelligent sensing, communication-oriented monitoring, and distributed information processing in Electromagnetic Waves, Antennas and Propagation systems.