LOF与改进SMOTE算法组合的强烈岩爆预测

Severe rock burst prediction based on the combination of LOF and improved SMOTE algorithm

为解决岩爆数据集中存在离群值、强烈岩爆数目少,导致强烈岩爆预测准确率较低等问题,提出LOF(local outlier factor)与改进SMOTE(synthetic minority oversampling technique)算法组合的方法。首先搜集国内外305组岩爆案例构建原始岩爆数据集,使用平均化处理法对数据集进行无量纲化处理;其次用LOF算法剔除各岩爆等级中的离群值,改进SMOTE算法在强烈岩爆样本与中等岩爆样本边界处增加强烈岩爆样本数目,增加少数类样本数目,在数据预处理阶段改善数据集结构;最后用6种常用机器学习模型分别对原始岩爆数据集和预处理后的岩爆数据集进行预测,验证预处理过程的有效性。结果表明:经过预处理后的岩爆数据集,整体的岩爆预测准确率平均提高18.35%,强烈岩爆预测准确率平均提高44.55%。因此基于LOF与改进SMOTE算法组合对岩爆数据集进行预处理,改善岩爆数据分布结构,能有效提高强烈岩爆预测准确率。

In order to solve low accuracy of strong rock burst prediction resulted from outliers in the rock burst data set and the small number of strong rock bursts. Combination of local outlier factor(LOF) and improved synthetic minority oversampling technique(SMOTE) algorithm is proposed. Firstly,305 groups of rock burst cases collected at home and abroad are used to construct the original rock burst data set,and the averaging method is adopted for non-dimension of the data set. Secondly,the data set structure during data preprocessing stage is improved by using LOF algorithm to eliminate outliers in each rock burst level and through improved SMOTE algorithm to increasing the number of strong rock burst samples at the boundary between strong rock burst samples and medium rock burst samples. Finally,the original rock burst data set and the preprocessed rock burst data set are respectively predicted by six commonly used machine learning models to verify the effectiveness of the preprocessing stage. The results show that the pre-processed rock burst data set has an average increase of18.35% in the prediction accuracy of the overall rock burst and an average increase of 44.55% in the prediction accuracy of strong rock burst, indicating that combination of LOF and improved SMOTE can effectively improve the accuracy of strong rock burst prediction.