Research On First Arrival Picking Method Of Microseismic Signal Based On Deep Learning

Accurate first arrival picking of microseismic signal is the basis of source location and source mechanism inversion.Because the microseismic events caused by hydraulic fracturing have weak energy and are easily masked by noise,it is more difficult to accurately identify and pick them up.In this paper,we propose a transfer learning method based on Attention U-net(Trans Attention U-net)for intelligent pickup of microseismic signals.The proposed method mainly includes two stages: pre-training stage and transfer learning stage.In the pre-training stage,the Attention U-net network with attention gates is used as the pre-training model,and 20000 labeled natural earthquake data are used as the source domain data to train the network,so that the network can learn the general features related to the earthquake arrival information.In the transfer learning stage,3500 microseismic data after data augmentation are used as the target domain data,and the structure and weight of the pre-trained model are transferred to the target domain.The strategy of combining freezing and fine-tuning is adopted to adjust the parameters of the pre-trained model,so that the network retains the general characteristics of seismic signals while retaining the general characteristics of seismic signals.We gradually learn specific features that target microseismic oncoming information.The model after transfer learning finally achieved 94% accuracy on the microseismic training set and validation set.A test set consisting of 300 three-component microseismic data is used to test the migration model.The results show that the precision and recall of the arrival of P wave and S wave predicted by the migration model are both about 92%.In this paper,a small amount of microseismic data(only 18% of the data of the pretraining model)is used as the transfer learning data set,and the arrival time information of P wave and S wave can still be accurately picked up by the model,which effectively solves the problem that small sample microseismic data sets are difficult to support model training.In addition,the network designed in this paper is an end-to-end network model,which basically does not need to set parameters manually,and effectively solves the problem of low automation of traditional methods.Compared with the direct application of the pre-trained model and the newly trained model,the transfer learning method in this paper can still achieve good performance even when limited microseismic samples are used.Compared with the traditional STA/LTA method and AIC method,the proposed method can still obtain better picking effect at low signal-to-noise ratio.