基于3种机器学习模型的岩爆类型预测

岩爆类型预测是防治和控制硬岩矿山岩爆灾害的有效方式。基于国内外397组岩爆案例数据，规范训练集与测试集的数据预处理方式，采用模型参数优化及交叉验证技术获得最近邻、支持向量机与决策树模型最佳参数；对比分析主成分分析法（PCA）与过采样SMOTE对3种机器学习算法预测准确率的影响，并采用准确率、精确率、召回率、F1等指标对模型预测性能进行评估。结果表明：主成分分析对3种机器学习模型的预测准确率并无提升，不同岩爆类型的样本之间不具有较为明显的决策边界；过采样SMOTE算法仅对决策树模型有明显的提升，基于过采样建立的SMOTE-DT模型预测准确率为77.50%，高于仅对原始数据集进行标准化处理的KNN、SVM模型的68.75%与57.50%；SMOTE-DT在高估与低估岩爆类型表现优于KNN与SVM模型，对于四种岩爆类型的F1值均大于0.7，岩爆预测性能稳定可靠。此外，采用本文构建的3种机器学习模型对山西紫金金矿进行了岩爆类型预测，模型预测结果与现场观测结果相一致。本文构建的三种用于岩爆类型预测的机器学习模型避免了训练集信息泄露对测试集造成影响，研究结果为岩爆类型预测及规范机器学习模型训练过程提供了理论支撑。

Rockburst type prediction based on three machine learning models

Prediction of rockburst type is an effective way to prevent and control rockburst disasters in hard-rock mines. Based on 397 groups of rockburst case data at home and abroad, standardize the data preprocessing methods of train set and test set, and use model parameter optimization and cross validation technology to obtain the best parameters of nearest neighbor, support vector machine and decision tree models. The influence of principal component analysis (PCA) and oversampling SMOTE on the prediction accuracy of the three machine learning algorithms was compared and analyzed, and the prediction performance of the model was evaluated by using accuracy, precision, recall, F1 and other indicators. The results show that the prediction accuracy of the three machine learning models has not been improved by principal component analysis. No obvious decision boundary between samples of different rockburst types are observed. The oversampling SMOTE algorithm only significantly improves the decision tree model. The prediction accuracy of the SMOTE-DT model based on oversampling is 77.50%, which is higher than 68.75% and 57.50% of the KNN and SVM models that only standardize the original dataset; SMOTE-DT performs better than KNN and SVM models in overestimating and underestimating rockburst types. The F1 value of four rockburst types is greater than 0.7, and the rockburst prediction performance is stable and reliable. In addition, three machine learning models constructed in this article were used to predict the rockburst types of the gold mine in Shanxi Zijin, and the predicted results of the models are consistent with the on-site observation results.The three machine learning models constructed in this paper for rockburst type prediction avoid the influence of information leakage of train set on test set. The research results provide theoretical support for rockburst type prediction and standardizing the training process of machine learning model.