基于9种机器学习算法的岩爆预测研究

Rockburst prediction based on nine machine learning algorithms

岩爆预测是减轻乃至消除岩爆危害的基础。本文构建含325组岩爆案例的岩爆预测数据集,通过引入9种经典机器学习算法,建立9个考虑多因素的岩爆预测模型,发展并完善岩爆预测方法。模型建立过程中,采用空值填充、重采样等多项数据预处理技术对数据进行清洗、归一化及降维,解决了数据不均衡问题。通过特征提取与选择,获得岩爆预测的最优的特征组合,采用网格搜索交叉验证技术获得模型的最优参数。采用准确率、精确率、召回率、F1值、宏平均、微平均等指标对模型预测性能进行验证与评估,并与常用理论判据的分类性能进行了对比。检验结果表明,所建模型的预测效果较好,且远好于单纯依靠理论判据所得的结果。最后,采用本文模型对西藏多雄拉隧道进行了岩爆倾向性分析,预测结果与现场情况具有较好的一致性。

Rockburst prediction is the basis for mitigating and eliminating rockburst hazards. In this paper,a rockburst prediction dataset containing 325 sets of rockburst cases was constructed. Based on nine classical machine learning algorithms,nine rockburst comprehensive prediction models that considering multiple factors were established. In the process of model establishment,multiple data preprocessing techniques were used to clean,normalize and dimensionly reduce the dataset,which addressed the data-imbalance problem. The optimal feature combination of rockburst prediction was obtained by extracting and selecting features,and the optimal parameters of the models were obtained by using grid search cross-validation technique. The prediction performances of the models were verified and evaluated by using accuracy,precision,recall rate,F1,macro-average,micro-average and other indicators,and compared with the classification performance of the commonly used theoretical criteria. The results of the model performance evaluation show that the accuracy of the model built in this paper is much higher than that of the widely used theoretical criteria. Based on the established models,the rockburst prediction of Tibet Duoxiongla tunnel is carried out,and the results are in good agreement with the field situation.