SUMMARY As an important method of 3-D seismic exploration, intelligent structural interpretation has been improved in recent years. Traditional human–computer interaction explanations typically assess fault risk on a hierarchical basis. However, the majority of intelligent fault interpretations utilize binary classification methods, which limits the utility and development of this technology to a certain extent. In order to break this limitation, a fault risk assessment scheme based on particle swarm optimization (PSO) and support vector machine (SVM) posterior probability model has been proposed. In this study, the fault exposure situations for 11 wells and 4 roadways of a coal mine in Yunnan Province, China, were identified. A total of 15 559 data points with known fault and non-fault information were obtained in the target coal seam. Based on these data, SVM fault recognition models were constructed. The key parameters of the SVM models were optimized by PSO, and the test accuracy of each model was above 87 per cent. Meanwhile, the SVM posterior probability model was introduced to evaluate the fault identification results in the study area. The SVM posterior probability model was used to convert the decision value of a standard SVM into a posterior probability value. Probability values in the range of [0, 0.5) are evaluated as non-fault points, [0.5, 0.7) as ‘unreliable’ fault points, [0.7, 0.9) as ‘less reliable’ fault points, and [0.9, 1] as ‘reliable’ fault points. By constructing data sets with different sample numbers and fault types, the influence of each factor on the evaluation of the SVM posterior probability model was analysed. The results show that the SVM posterior probability model can be used to hierarchically evaluate fault risk in the study area and provide decision makers with more accurate information for decision making.