Microseismic Intelligent Location And Rockburst Risk Prediction Based On Machine Learning

The outline of China’s 14 th five-year plan for scientific and technological innovation proposed that future domestic technological development would focus on deep-buried underground space,abyssal sea,aerospace industry,and other leading-edge technology domains.Compared with rock in shallow underground space,the surrounding rock in deepburied underground space suffered from the influence of high geostress,high temperature,and violent dynamic disturbance,which would be likely to cause complex geological disasters,such as rockburst,large deformation,and so on.Currently,rockburst had become an obstacle to the safe and efficient development of deep-buried underground space.As a kind of rock mass micro-fracture monitoring technology,microseismic monitoring technology has become an essential means of monitoring and early warning of underground engineering disasters.At present,the location method of microseismic monitoring mainly adopts the location method with time residual as the research object.The dependence on manual verification limited it.As a result,this method could not guarantee the timeliness of the location of the microseismic event,which may delay the best time for disaster warning.Besides,in the early warning of rockburst,the prediction of rockburst risk based on the evolution of microseismic parameters is greatly affected by subjective factors.Based on the above description,this paper aimed to build an intelligent early warning method for underground disasters based on microseismic monitoring,relying on two engineering examples of "Shuangjiangkou Underground Powerhouse" and "Yinhanjiwei Water Diversion Tunnel"."Intelligent positioning method based on microseismic waveforms" and "rockburst prediction method based on microseismic parameters" were proposed to resolve the microseismic location and rockburst disaster prediction problems,which would provide references for the intelligent early warning and evaluation of underground engineering disasters.(1)The dataset was established by microseismic events that happened around Shuangjiangkou underground powerhouse.We built the MS-location Net fully convolutional neural network.The waveform was used as the input of the neural network,and the threedimensional Gauss function of the study area was used as the neural network’s output.After oversampling the dataset,the epicenter error,depth error,and absolute error of 148 microseismic events in the test set were less than 5m,indicating that the MS-location Net proposed in this paper can effectively locate the microseismic event near the main powerhouse of Shuangjiangkou.(2)We chose the rockbursts that occurred within 260 days as the research scope according to the construction progress of the TBM of Yinhanjiwei conveyance tunnel.Two kinds of classical machine learning algorithms were introduced to build a deep-buried tunnel rockburst prediction model based on microseismic parameters.In this model,we selected the number of cumulative microseismic events,cumulative energy,cumulative apparent stress,cumulative apparent volume,b-value,maximum moment magnitude,and maximum displacement detected by the sensor in the day in a day as the input part of the model,and the rockburst intensity of the next day was used as the output part of the model.(3)We have tested datasets with different "oversampling" methods through SVM,Adaboost,GBDT,Random tree,and XGBoost algorithms.The best performance algorithm was Borderline-SMOTE 1-XGBoost algorithm compared with other algorithms’ accuracy and F1 metrics.The accuracy of Borderline-SMOTE 1-XGBoost algorithm was 0.865 after 5-fold cross validation and the accuracy of the algorithm was 0.9 in the divided test set.The macroF1 and micro-F1 values were 0.9 and 0.9,respectively.The results of this algorithm would provide a reference for the prediction of rockburst risk in Yinhanjiwei Project.