Prediction of Pillar Stability for Underground Mines Using the Stochastic Gradient Boosting Technique

The prediction of pillar stability is of great importance because pillar failure can lead to large disasters. In this paper, a stochastic gradient boosting (SGB) model was applied to classify pillar stability. Five potentially relevant factors, including the pillar width, the pillar height, the ratio of the pillar width to the pillar height, the uniaxial compressive strength of the rock, and the pillar stress, were chosen to establish the evaluation index system. The 205 pillar samples were collected, and an SGB model was developed by training 80% of original data (165 samples), and the optimal parameter values of the model were achieved by the method of 10-fold cross-validation. The external testing set (with 40 samples) was used to validate the feasibility of the SGB model. The accuracy and kappa analysis, together with the three within-class classification metrics (recall, precision, and F-measure), and receiver operating characteristic curve were utilized to evaluate the performance of the optimum SGB, random forest (RF), support vector machine (SVM), and MLPNN models. The results revealed that the SGB model has higher credibility than the RF, SVM, and MLPNN models. The sensitivity of the parameters was investigated based on the relative variable importance, in which the pillar stress and the ratio of the pillar width to the pillar height were found to be the major influencing variables for pillar stability.