Journal of Safety and Environment

To enhance the specificity and causal reasoning capabilities of coal mine fire risk assessment, this study constructed an integrated evaluation model that combines rough set differential matrices with fuzzy Bayesian network. Based on the causation mechanisms of coal mine fires and relevant standards, we established an indicator system encompassing four dimensions—fire prevention, emergency response, risk monitoring, and management capacity—comprising 14 secondary indicators. We employed the rough set differential matrix method to reduce indicator attributes, identifying five key indicators from the initial set: ignition source management(A₃), ventilation system design and maintenance(A₄), emergency plan completeness(B₁), intelligent monitoring system application(C₂), and personnel training and safety awareness(D₂). This process, grounded in data from 45 coal mine accident cases and expert assessments, achieved objective indicator selection through the construction of a differential matrix, calculation of attribute discrimination, and determination of normalized weights. Building upon the identified key indicators, we applied fuzzy set theory to address uncertainties in expert evaluations. We converted expert linguistic assessments into triangular fuzzy numbers and derived fuzzy probabilities for each root node using weighted averaging and defuzzification. Using these as prior probabilities, we constructed a fuzzy Bayesian network structure. Through network parameter learning and conditional probability table configuration, we achieved forward probability prediction and backward causation diagnosis under conditions of uncertainty. Model application and validation results indicate that ventilation system design, emergency plan completeness, and intelligent monitoring system application are high-risk factors, each exhibiting a probability exceeding 50% of being in a “high” risk state. Reverse inference further confirms that when the system is in a high-risk state, the posterior probabilities of these nodes increase significantly, indicating they are key drivers of the risk state. Validation using the Chongqing Songzao Coal Mine “9.27” major fire accident case demonstrates that the model successfully raised the high-risk probability from a baseline state of 27.1% to 89.3% based on anomalies in key indicators. Through posterior analysis, it precisely quantified the risk contribution of each factor. This research integrates the objective attribute reduction capability of rough set differential matrices with the uncertainty reasoning capacity of fuzzy Bayesian network, effectively addressing shortcomings of traditional risk assessments, such as redundant indicators, subjective weighting, and inadequate handling of ambiguous information. The model not only significantly enhances the accuracy and interpretability of coal mine fire risk identification but also facilitates quantitative diagnosis and early warning of key causal factors. It provides a reliable decision support tool for coal mining enterprises to implement differentiated and precise fire risk management. The model and methodology developed herein offer a novel technical approach for tackling the dynamic assessment of multi-factor coupled risks within complex industrial systems.