Research On Positioning Accuracy Optimization Of Microseismic Events Based On Deep Learning

In tunnel engineering,the clustering of microseismic events can cause rockburst and other disasters.Microseismic monitoring technology,as one of the main means of rockburst warning,predicts potential rockbursts by recording,processing,and analyzing the temporal and spatial distribution of microseismic events.Thus,the accuracy of microseismic event location directly affects the effectiveness of rockburst warning.However,microseismic monitoring equipment needs to continuously collect data in the monitoring area.In addition to the data generated by microseismic events,a large amount of useless data,including blasting,electrical noise,and mechanical noise,will also be generated during the monitoring process.Therefore,the detection of microseismic events can indirectly affect the location accuracy.Secondly,due to the geological conditions in tunnel engineering changing with the construction process,it is difficult to establish a velocity model in the arrival-time-difference location algorithm,which directly affects the location accuracy of microseismic events.With the continuous deepening of research,the application of deep learning technology in microseismic monitoring has become more and more prominent.In response to the above problems,this paper focuses on the optimization of microseismic event location accuracy from two aspects: microseismic event classification and velocity model selection.First,the classification problem of low signal-to-noise ratio microseismic events.To address the challenging issue of distinguishing certain features in the processing of low signal-to-noise ratio microseismic events using conventional machine learning and Convolutional Neural Networks(CNN),this study proposes a CNN-Swin Transformer Classification(CSTC)network that combines CNN and Transformer for microseismic event classification.To reduce feature input,three raw microseismic record waveform,frequency spectrum,and time-frequency spectrum are extracted and used as the inputs of the network.A CNN with residual and spatial attention mechanisms is used in the network frontend for deep feature extraction,and the extracted features are then input to Swin Transformers for classification.Through training,validation,and testing,the CSTC network achieved a classification accuracy of 99.14% by using only the three aforementioned microseismic features.The CSTC network was compared with Inception,Alex Net,VGG,Resnet,and Swin Transformer networks using synthetic microseismic records with different signal-to-noise ratios,and the experimental results showed that the CSTC network has better classification performance and noise resistance,which better meets the requirements of accurate microseismic event classification in tunnel engineering.This paper indirectly improves event localization accuracy by using the CSTC network to improve the classification of low signal-to-noise ratio microseismic events.Second,the problem of velocity model selection.We compared the velocity models currently used in tunnel engineering and selected the "source-station" velocity model with the highest location accuracy for study.Regarding the problem of velocity model selection when using multiple active sources in the "source-station" velocity model,we proposed using the Multi-Label K-Nearest Neighbor(MLKNN)algorithm for velocity model selection.Through simulation experiments,the "source-station" velocity model selected using MLKNN was used for microseismic event location.Compared with the uniform velocity model and the layered velocity model,the average location error was reduced from 13.28 m and 6.24 m to2.68 m,and the average location accuracy was increased by 79.8% and 57.1%,respectively.The experimental results demonstrate that the velocity model selection method based on the MLKNN algorithm can directly improve the location accuracy of microseismic events.Finally,the proposed method is validated in engineering applications.The microseismic event classification and velocity model selection methods proposed in the microseismic event location accuracy optimization technology research are verified using tunnel laboratory data and actual measurement data from the Baihetan hydropower station.A comparative experiment is conducted using the tunnel laboratory to verify the effectiveness of the velocity model selection method proposed in this paper,while the actual measurement data from the Baihetan hydropower station is used to validate the classification performance of the microseismic event classification method proposed in this paper and the optimization effect of the "source-station" velocity model on location accuracy.The results of both experiments show that in practical engineering applications,the proposed deep learning-based method can effectively improve the accuracy of microseismic event location.