Statistical assessment of rock burst potential and contributions of considered predictor variables in the task

Abstract

Rock burst phenomenon is one of the major concerns in underground mining and excavating. Despite the significant progress in the field of phenomenon prediction, including data-driven researches, the need for developing reliable models is still felt. In this research, a database consisting of 188 distinct case histories was considered. Each case history contains some of the predictor variables “overburden thickness, maximum tangential stress in the boundary of opening, uniaxial compressive strength of rock, tensile strength of rock, stress ratio, brittleness ratio and elastic strain energy index” and one of the four defined classes (none or not-occurred, weak, moderate and strong) for the qualitative dependent variable “rock burst intensity”. After data preprocessing procedure (including outlier detection and substitution, factor analysis and nonlinear principal component analysis in which the two later did not lead to extract reasonable predictor components), 48 regression models were fitted for different original predictor variable arrangements, fitting datasets and link functions to obtain proper statistical equations for rock burst occurrence and intensity prediction. The best models, recognized based on fulfilling certain criteria and maximizing specific objectives, have proper classification accuracies, especially in comparison to those of the empirical criteria developed for rock burst prediction. Due to the presence pattern of the predictor variables in regression equations, a stepwise statistical strategy, consisting of normality condition evaluation, evaluation of mean equality and distribution function equality hypotheses and correlation analysis for predictor variables, was considered. The strategy led to conclude that the predictor variables “overburden thickness, tensile strength of rock and brittleness ratio” with the available case histories have insignificant contributions to predict rock burst phenomenon. By contrast, the predictor variables “maximum tangential stress, stress ratio, elastic strain energy index and uniaxial compressive strength of rock” have considerable contributions, decreasing in an ascending order. The application of non-dominated sorting genetic algorithm and multi-layer perceptron neural networks for solving the feature (i.e. variable) selection problem had a perfect adaption with the inferences.