SENSE: A Sensor-Centric Multimodal Evidence Chain for Teaching Evaluation in Smart Classrooms

Smart-classroom teaching evaluation increasingly relies on heterogeneous sensing sources, including classroom cameras, microphones, speech-recognition systems, and learning-platform interaction logs. Existing systems often treat these streams as separate analytical inputs, with insufficient attention to cross-modal alignment, missing-sensor robustness, and evidence-grounded large-language-model reasoning. This paper proposes SENSE (Sensor-Evidence Network for Smart-Classroom Evaluation), a sensor-centric framework that organizes classroom evaluation as an end-to-end evidence chain. Visual behavior, acoustic-semantic content, and learning-platform events are aligned in lesson-level time windows and converted into structured evidence packets. Available learning-outcome anchors provide contextual evidence to the semantic evaluator, whereas expert ratings are retained exclusively as external validation references. A DeepSeek-based semantic evaluator then generates engagement, instructional-content, interaction-quality, and teaching-feedback judgments under a fixed output schema. The framework was evaluated using 48 college-English lessons, 48 videos, 48 audio recordings, 4628 timestamped learning-platform interaction logs, and 960 valid student-window observation instances. Compared with a transparent rule-based operational baseline, SENSE reduced the engagement-evaluation error from 12.3% to 5.8%, reduced the content-coverage error from 14.5% to 6.2%, and increased the correlation with the mean expert score from 0.71 to 0.89. Under 30% random evidence removal, the correlation remained 0.80. The results indicate that teaching evaluation benefits from a multimodal sensor-evidence and semantic-reasoning process rather than a single behavior-recognition task.