基于集成算法和普通机器学习算法的岩爆分级预测及选择

岩爆是国内外深部地下工程面临的巨大灾害，岩爆预测具有显著的现实意义。现阶段单一的机器学习算法准确率较低、泛化性不足，难以发挥各个算法优点。为满足岩爆预测与工程实际需要，提出基于集成算法和普通机器学习算法相互结合预测岩爆各个等级，充分发挥不同算法在某一岩爆等级预测的优势并形成互补。提出改进的Boosting、Bagging集成预测算法，与stacking、random forest、random subspace集成算法和普通学习算法诸如BP、贝叶斯算法（bayes）、k最近邻（KNN）、支持向量机（SVM）等在内的14种预测算法进行相互结合验证。基于国内外地下工程165组岩爆实例，选取围岩最大切向应力(MTS)，岩石单轴抗压强度(UCS)，岩石单轴抗拉强度(UTS)，岩石弹性能量指数(Wet) 构建岩爆预测体系，引入T–分布邻域嵌入(T-SNE)，对数据进行降维可视化。为避免算法预测岩爆过程中预测结果的偶然性，即出现预测准确率过高或过低的现象，采用在各个岩爆等级按照比例随机筛选训练集和测试集，确保数据集分类的严谨性；每次机器学习过程的数据都具有随机性，再采用10次运行结果后取各项预测平均值，评价算法在各个等级的准确率和算法整体的预测稳定性。结果表明：LDA 对 Ⅱ 级岩爆有更高的准确率，Bayes分类 Ⅳ 级岩爆效果最好，Adaboost.M1对 Ⅰ 级和 Ⅲ 级有最高的准确率。整体预测效果基于决策树的Bagging预测稳定性更好，预测精确率高。最后引入终南山隧道竖井工程案例，预测结果与现场实际工况较为一致，表明本文所建立算法的可靠性。

Rock burst classification prediction and selection based on integrated algorithm and general machine learning algorithm

Rockburst is a huge disaster faced by deep underground engineering at home and abroad, and rock burst prediction has significant practical significance. At present, a single machine learning algorithm has low accuracy and insufficient generalization, which makes it difficult to give full play to the advantages of each algorithm. In order to meet the needs of rockburst prediction and engineering practice, it is proposed to combine the integrated algorithm and ordinary machine learning algorithm to predict each grade of rockburst, giving full play to the advantages of different algorithms in predicting a certain rockburst grade and forming complementarities. The improved Boosting and Bagging integrated prediction algorithm is proposed, and 14 prediction algorithms, including stacking, random forest, random subspace integrated algorithm, and general learning algorithm, such as BP, Bayesian algorithm (Bayes), k-nearest neighbor (KNN), support vector machine (SVM), are combined to verify each other. Based on 165 groups of rock burst examples of underground projects at home and abroad, the maximum tangential stress (MTS), uniaxial compressive strength (UCS), uniaxial tensile strength (UTS), and elastic energy index (Wet) of rock are selected to build a rock burst prediction system. T-SNE is introduced to reduce the dimension of data. In order to avoid the contingency of prediction results in the process of rock burst prediction, that is, the prediction accuracy is too high or too low, this paper uses random screening of training sets and test sets in proportion at each rock burst level to ensure the preciseness of data set classification. And the data of each machine-learning process is random. After 10 times of running, take the average value of each prediction to evaluate the accuracy of the algorithm at each level and the overall prediction stability of the algorithm. The results show that LDA has higher accuracy for Grade II rockburst, Bayes classification Grade IV rockburst is the best, and AdaBoost M1 has the highest accuracy for Grade I and Grade III. The overall prediction effect of Bagging based on a decision tree has better stability and high prediction accuracy. Finally, the case of the Zhongnanshan tunnel shaft project is introduced, and the predicted results are consistent with the actual working conditions on site, indicating the reliability of the algorithm established in this paper.