Abstract: In order to address the strong heterogeneity of multi-source monitoring data, the concealment of risk precursors, and the high cost and risk of on-site emergency drills for coal mine ventilation and prevention disasters, this study constructs a multi-source fusion-driven intelligent early warning and virtual drilling model. Multi-source monitoring data, including gas concentration, air velocity, CO/CO₂ concentration, temperature and humidity, dust concentration, negative pressure, and equipment operating status, were collected and processed through missing-value imputation, outlier removal, normalization, and risk-level labeling. A collaborative modeling framework integrating CNN-LSTM, GNN and Transformer was developed to extract the local temporal features, spatial correlation features, and multivariate dependency features of ventilation and prevention disaster data. SHAP analysis and attention weights were further introduced to identify key risk-contributing factors. Meanwhile, virtual reality, digital twin, and human posture assessment technologies were integrated to establish an immersive drilling system for disaster scenario reconstruction and emergency behavior evaluation. The proposed model achieved stable identification of different risk levels. The CNN-LSTM model obtained a validation accuracy of approximately 0.889, and the recognition accuracy for low-, medium-, and high-risk levels were 89.04%, 84.50%, and 76.70%, respectively. Airflow fluctuation, mean gas concentration, dust peak value, and CO concentration gradient were identified as the main influential factors. The proposed model improves the accuracy of coal mine ventilation and prevention disaster risk identification and enhances early warning capability. The virtual drilling module strengthens the scenario-based application of warning results and can provide technical support for closed-loop optimization of mine disaster prevention, control, and safety management.