The campus public opinion monitoring system based on deep learning achieves efficient public opinion classification and sentiment analysis through distributed data collection, preprocessing, deep learning model construction and visual analysis. The system uses BERT pre-training model and multi-head self-attention mechanism for feature extraction, combined with the hybrid architecture of transformer and graph convolutional network, and optimizes performance through weighted cross entropy loss and graph convolution feature update. Experimental results show that the system achieves 90.2%, 89.5% and 89.8% in precision, recall and F1 score, which is significantly better than the baseline model. The visualization analysis of confusion matrix and sentiment intensity distribution further verifies the robustness of the model in fine-grained sentiment classification, providing reliable technical support for campus public opinion management.