基于机器学习可解释性算法的岩爆指标分析

为探究弹性能指数、应力系数、脆性系数、埋深四种岩爆指标与岩爆等级之间的相关关系，解决复杂机器学习算法的黑盒问题。本文引入LIME(local interpretable model agnostic explanations)算法，完善岩爆机器学习预测过程中的可解释性。文章搜集国内外190组岩爆实例工程构建数据集经过预处理后，通过9种机器学习算法比较获得最优算法并采用贝叶斯优化获得算法最优参数，建立岩爆预测模型。基于LIME可解释性算法，对四种岩爆指标进行相关、回归及阈值分析，最后采用弹性能指数及应力系数两种指标阈值对终南山隧道竖井工程进行岩爆预测。研究结果表明：(1) 岩爆等级与弹性能指数、应力系数呈线性相关，且弹性能指数线性关系更明显；(2) 岩爆等级与脆性系数、埋深呈非线性相关，且脆性系数非线性关系更明显；(3) 4个岩爆指标对岩爆等级影响程度依次为：弹性能指数、应力系数、埋深、脆性系数；(4) LIME算法可以准确地表达岩爆等级与岩爆指标之间的相关关系且得到的两种指标阈值与终南山隧道竖井工程实例具有一致性。

Analysis of Rockburst Indicators Based on Machine Learning Interpretability Algorithm

In order to explore the correlation between the elastic energy index, stress coefficient, brittleness coefficient and burial depth of four rockburst indexes and the rockburst grade, and to solve the black box problem of complex machine learning algorithms.. The LIME algorithm was introduced in this paper to improve the interpretability of machine learning in rock burst prediction. In this paper, 190 sets of rockburst engineering construction data sets at home and abroad were collected. After preprocessing, the optimal algorithm was obtained by comparing 9 kinds of machine learning algorithms. The optimal parameters of the algorithm were obtained by Bayesian optimization, and the rockburst prediction model was established. Four kinds of rockburst were analyzed based on LIME algorithm. The correlation analysis, regression analysis and threshold analysis were carried out on the indexes. Finally, the rockburst of Zhongnanshan tunnel shaft project was predicted by using two thresholds of elastic energy index and stress coefficient. The research results show that: (1) the rockburst grade is linearly related to the elastic energy index and stress coefficient, and the linear relationship of the elastic energy index is more obvious; (2) the rockburst grade is nonlinearly related to the brittleness coefficient and burial depth, and the brittleness coefficient The nonlinear relationship is more obvious; (3) the four rockburst indexes influence the rockburst grade in order: elastic energy index, stress coefficient, burial depth, and brittleness coefficient; (4) the LIME algorithm can accurately express the relationship between rockburst grade and rockburst grade. The correlation between the explosion indicators and the obtained thresholds of the two indicators are consistent with the Zhongnanshan tunnel shaft project example.