An indicator-gas-driven XGBoost model for tiered early warning of coal spontaneous combustion

Background As coal mining operations advanced into the more extensive underground mining phase, coal spontaneous combustion (sponcom) poses a growing threat to production safety. This study comprehensively investigates the combustion behaviour of different coal seams and develops an accurate hazard classification and prediction model. Methods Synchronous thermal analysis (STA) and temperature-programmed experiment were conducted on coal samples from varying depths at the Hecaogou Coal Mine. The combustion process was delineated into four distinct stages: moisture evaporation, oxygen absorption, pyrolysis and combustion, and burnout. Kinetic analysis using the Coats–Redfern method revealed that different coal sponcom propensity under different burial depths in the seam. To obtain coal mine hazard assessment, the Extreme Gradient Boosting (XGBoost) algorithm was implemented. This model, utilizing gas concentration and seam-related data, achieved a high prediction accuracy of 91.67%. Significant Findings The XGBoost framework demonstrated strong hold-out classification performance and uses tree-complexity controls, row subsampling, and feature subsampling to reduce overfitting risk.