Discussion and selection of machine learning algorithm model for rockburst intensity grade prediction
LI Mingliang1，2，LI Kegang1，2，QIN Qingci1，2，WU Shunchuan1，2，LIU Yuedong1，3，LIU Bo1，2
(1. Kunming University of Science and Technology，School of Land and Resources Engineering，Kunming，Yunnan 650093，China；2.Yunnan Key Laboratory of Sino-German Blue Mining and Utilization of Special Underground Space，Kunming，Yunnan 650093，China；3.Yunnan Diqing Mining Industry Group，Shangrila，Yunnan 674400，China)
Abstract：In order to improve the accuracy of predicting rockburst intensity grade，the four coefficients(rock mass stress coefficient( )，rock brittleness coefficient( ) and elastic deformation energy coefficient( ) were selected to construct the rock explosive intensity level prediction index system. Based on 145 sets of rockburst case engineering data at home and abroad，adopting six machine learning algorithms to establish the prediction model of rock explosive intensity level respectively，combined with random cross-validation methods. the correlation coefficients were calculated using the principle of correlation coefficients. According to the correlation coefficients between variables，there is no strong correlation between them. Meanwhile process the engineering data of the original rock explosion case first and then standardize standardization，eliminating the influence of extreme values in the data on the model. The T-distributed neighborhood embedding(T-SNE) dimensionality reduction method is introduced to reduce the dimensionality of the data and visualize the data. Finally，the accuracy of the six established rockburst intensity grade prediction models were analyzed，discussed and evaluated. The research results show that：Based on the T-distributed neighborhood embedding(T-SNE) dimensionality reduction method，the results show that each rockburst intensity level has obvious aggregation phenomenon. Support SVM has high prediction accuracy for rock burst grade 1. For samples with rockburst intensity levels of 2 to 4，the linear discriminant model has a high prediction accuracy and also has a relatively stable model performance；the linear discriminant model(LDA) is applied to Jinping rockburst case projects，such as the secondary hydropower station，the riverside hydropower station，and the Cangling tunnel，have found that the LDA model prediction results are the same as the actual rockburst levels. The research results provide a good guide for rockburst prediction problems in geotechnical engineering.
2021, Vol. 40 
