The identification of the bursting liability of coal is a basic task for preventing and controlling rock bursts, and also the main basis for evaluating the level of rock burst risk. The current Chinese national standard "Classification and Determination Method for Bursting Liability of Coal" (GB/T 25217.2-2010) provides 73 combinations of evaluation results based on the fuzzy comprehensive judgment method, but there are still 8 combinations that are difficult to determine. In recent years, many scholars have conducted in-depth research on the discrimination of bursting liability levels of coal, and have proposed multiple classification indicators and methods. Wang Chao et al. ^[1] introduced the Mahalanobis distance discriminant analysis (DDA) method to evaluate the coal’s bursting liability, avoiding the influence of the correlation between evaluation indicators on the discrimination results. Zhou Jian et al. ^[2] used the AHP-entropy weight method to determine the weights of evaluation indicators and proposed an improved bursting liability classification model based on the uncertain measurement theory.

  Therefore, this paper introduces fuzzy set theory to describe the fuzziness between data and levels, and uses the Delphi-Random Forest combination weighting method to determine the weights of evaluation indicators, which reduces the interference of data discreteness on weights and makes the indicator weights more reasonable. Two types of membership functions, trapezoidal fuzzy numbers (TMF) and Gaussian fuzzy numbers (GMF), are used to quantitatively describe the fuzziness between indicator levels. Four fuzzy operators, Zadeh operator (ZO), maximum product operator (MMO), weighted average operator (WAO), and comprehensive restriction operator (CRO), are employed to synthesize the weights and membership degree of indicators. Two evaluation criteria, maximum membership principle (MMP) and confidence probability criterion (CIP), are used to evaluate the level of bursting liability. 16 fuzzy comprehensive evaluation models of bursting liability are established by comprehensively considering the three influencing factors, membership function, fuzzy operator, and evaluation criterion. The performance of these models is compared and analyzed through 127 sets of samples. Finally, the optimal model is selected and applied to engineering practice. The results show that the fuzzy comprehensive evaluation model, TMF-WAO-MMP model, based on trapezoidal fuzzy number, weighted average operator, and maximum membership principle is the best model, with a discrimination accuracy of 97.64%. The optimal model is applied to 10 engineering cases and the evaluation results are consistent with the actual situation, which verifies the reliability and validity of the model.