基于数据预处理的岩爆等级预测模型及精度优化

Prediction Model and Accuracy Optimization of Rockburst Grade Based on Data Preprocessing

岩爆预测精度对岩体工程灾害预测具有重要的现实意义,精确有效的数据预处理是后续预测工作的基础。通过收集国内外471组岩爆案例建立岩爆数据库,选取围岩最大切向应力、抗压强度、抗拉强度和弹性能量指数作为特征指标,并结合10种机器学习算法构建预测模型。为消除样本中离群值对预测模型的干扰,将离群值清洗范围缩小至单一等级内,根据岩爆烈度等级逐级检测并处理离群值。提出自适应过采样(ADASYN)改善数据分布,在保留少数类样本数据特征的情况下对原始少数类数据进行样本合成,解决各岩爆等级样本不平衡问题。引入遗传算法(GA)对高稳定性模型参数寻优,并结合混淆矩阵和多个评价指标对模型深度评估。研究表明:ADASYN方法将模型综合准确率提升11.58%,并选出最优性能GA-XGBoost模型,预测准确率和加权平均F1值均达到93%;将模型应用于锦屏二级水电站、三山岛金矿和马路坪矿,预测结果与现场情况有较好的一致性,可为今后岩爆预测提供新方法。

The accuracy of rockburst prediction has an important practical significance for the prediction of rock mass engineering disasters.Accurate and effective data preprocessing is the basis of subsequent prediction work.The rockburst database was established by collecting 471groups of rockburst cases at home and abroad.The maximum tangential stress,compressive strength,tensile strength and elastic energy indexes of surrounding rocks were selected as the characteristic indexes,and the prediction model was constructed by combining 10machine learning algorithms.In order to eliminate the interference of outliers in the samples to the prediction model,the outlier cleaning range was reduced to a single level,and the outliers were detected and processed step by step according to the rockburst intensity level.An adaptive oversampling(ADASYN)was proposed to improve the data distribution,and the sample synthesis of the original minority class data was carried out under the condition of retaining the characteristics of the minority class sample data,so as to solve the problem of sample imbalance of each rock burst grade.The genetic algorithm(GA)was introduced to optimize the parameters of the high stability model,and the model was deeply evaluated by combining the confusion matrix and multiple evaluation indexes.The research shows that the ADASYN method improves the comprehensive accuracy of the model by 11.58%,and GA-XGBoost model has been selected as the optimal performance.The prediction accuracy and weighted average F1value reach 93%.The model was applied to the JinpingⅡHydropower Station,Sanshandao Gold Mine,and Maluping Mine,and the predicted results showed good consistency with the on-site conditions,providing a new method for predicting rock bursts in the future.