关键词: 局部线性嵌入(LLE), 基于密度的带噪声应用空间聚类(DBSCAN), 合成少数类过采样(SMOTE), 数据处理, 岩爆预测

Abstract:

To address issues of correlation prediction indicators, outliers, and data imbalance in original data in rockburst prediction, a rockburst prediction method based on LLE-DBSCAN-SMOTE for data processing was proposed. Firstly, the maximum tangential stress of surrounding rock {\sigma }_{\mathrm{\theta }}, uniaxial compressive strength of rock {\sigma }_{\mathrm{c}}, uniaxial tensile strength of rock {\sigma }_{\mathrm{t}}, elastic strain energy index W_{\mathrm{e}\mathrm{t}}, brittle coefficient {\sigma }_{\mathrm{c}} / {\sigma }_{\mathrm{t}}, stress coefficient {\sigma }_{\mathrm{\theta }} / {\sigma }_{\mathrm{c}}, and stress concentration value β characterizing the stress gradient of surrounding rock were selected to construct a rockburst prediction indicator system. Secondly, the LLE algorithm was used for data dimensionality reduction to eliminate the cross-correlation effect between indicators, and the DBSCAN algorithm was introduced to remove outliers. Then, the SMOTE technology was introduced for data balancing. Finally, three types of rockburst prediction models were proposed using Decision Tree (DT), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT) algorithms. The prediction accuracy of the data training models before and after processing was compared and analyzed. Moreover, engineering verification was performed through the measurement in the diversion tunnel of Jiangbian Hydropower Station. The results show that the prediction accuracy of the three types of algorithm models which reduce the prediction index from the 7 dimensions of the original data to the 4 dimensions and adopt the graded outlier processing is the highest among the similar models. The rockburst prediction of the Jiangbian Hydropower Station demonstrates that the proposed model significantly improves prediction accuracy compared to similar models using original rockburst data.

Key words: local linear embedding (LLE), density-based spatial clustering of applications with noise (DBSCAN), synthetic minority over-sampling technique (SMOTE), data processing, rockburst prediction