Student Sentiment Analysis Based on Improved Residual Network Model in Smart Classrooms

Smart classroom is an important environmental and technical support for achieving personalised learning, and the emotional state of learners is an important component of personalised learning. To lift the accuracy and the applicability of facial expression recognition (FER) models, this study proposes a feature fusion network framework based on residual networks. The framework assesses the emotional activity of online teaching by identifying the expressions of teachers and students, and evaluates the effectiveness of online teaching from a dimension. The experimental results showed that the loss values when training and testing the research model decreased to below 0.1, converged after 100 iterations and tended to be stable. The accuracy of the training set reached over 99%, while the test set remained around 98.67%. The research model achieved FER accuracy of 98.96%, 88.06% and 71.14% on three datasets, respectively, which was superior to other models and had higher FER accuracy than the other two initial models. The proposed improved residual network model lays the technical foundation for real-time knowledge management in intelligent classrooms. This model analyses students’ emotional states to support adaptive learning environments, delivering personalised feedback and dynamic resource recommendations. Additionally, it enhances teachers’ decision-making by providing real-time insights into classroom engagement. These capabilities optimise knowledge-sharing processes and pedagogical strategies. The approach contributes to building a data-driven educational knowledge system, bridging learning evaluation and instructional intervention.