Abstract
Although coal-mine accident reports contain intertwined risk factors, co-occurrence alone cannot establish causal direction or quantify the risk of accident occurrence. The symmetry of statistical co-occurrence and the potential asymmetry of conditional relationships highlight the difficulty of determining meaningful directions between risk factors. This study aims to identify factors associated with severity in reported coal-mine accidents and examine their conditional dependencies. Based on a sample of 800 accident investigation reports collected from 2010 to 2025, this study integrated text mining and expert review to extract and standardize 55 risk factors. Subsequently, the Apriori algorithm was employed to mine candidate associations, and expert prior knowledge was introduced to establish directional constraints, forming the basis for constructing a Bayesian network. The analysis extracted 274 association rules and constructed a network topology comprising 57 nodes and 187 directed edges. The comprehensive assessment revealed that factors such as ventilation, roof management, safety education and training, employee safety awareness, and gas accumulation have particularly significant impacts on accident severity. This study not only provides a quantitative, structured analytical framework for deconstructing the causes of complex accidents but also offers a scientific basis for coal mining enterprises to transition their safety management from passive post-event response to proactive, targeted pre-event control. It holds significant practical value for optimizing the allocation of safety resources and enhancing the intrinsic safety level of coal mines.