Microseismic Localization Method of Transformer Coal Mine Based on Conditional Confrontation Enhancement
With the development of artificial intelligence technology and the widespread use of microseismic monitoring sys-tems in coal mines,more and more deep learning models are applied to solve the source localization problem of microseismic events in coal mines.However,the small amount of microseismic data and single data are not enough to train large and deep neural network models,and the small and shallow neural network models are not enough to characterize the source of microseismic events influ-enced by multiple factors,thus leading to the low localization accuracy and weak robustness of the localization models,and poor per-formance in practical production life,which seriously hinders the development of deep learning models in the field of microseismic localization.To address the above problems,a Transformer coal mine microseismic localization method based on conditional adver-sarial augmentation,CGAN-Transformer,is proposed,which firstly augments microseismic data with small and single data volume into microseismic data with large data volume and certain diversity by a network model of CGAN architecture.Secondly,transforms microseismic waveform data into microseismic data using Transformer encoder layer to convert microseismic waveform data into fea-ture data and then use its attention mechanism to further learn the deep-level features and complex inter-station dependencies of mi-croseismic waveform data,and also use Gaussian distribution random variables to offset the influence of different geological condi-tions on localization accuracy.Finally,by introducing a hybrid density output layer to obtain Gaussian distribution parameters,the optimal source location is calculated.The experimental results on a mining dataset in Chile and Liaoning verify the effectiveness of the method.The results show that both the epicenter error and the source error obtained by this method are better than other meth-ods,and the localization error is reduced by 38%and 12%on the two datasets,respectively,achieving the purpose of improving the source localization accuracy and the robustness of the localization model.
国家科技期刊平台
NSTL
万方数据
