Abstract:  Microseismic event recognition is a key part of the data processing of hydraulic fracturing microseismic monitoring, but there are some limitations in effective event feature extraction and noise resistance for current microseismic event recognition methods based on deep learning, which cannot meet the requirements of subsequent microseismic source localization and inversion. In this paper, a microseismic event recognition method based on an improved U-Net model is proposed. In this method, residual shrinkage modules are introduced during the downsampling process of U-Net. The residual blocks are used to implement shortcut connections of the network structure to reduce the loss of feature information during model training, and the soft thresholding shrinkage technology is combined to weaken the interference of noise features in microseismic data and enhance model noise resistance. The attention gate mechanism is introduced during the upsampling process to weight the extracted microseismic data feature vector by gating signal, so that the model focuses on the area of the data containing microseismic events, thereby improving the extraction ability and identification accuracy of the model’s effective feature. The experimental results of synthetic and actual microseismic data show that the proposed method can adequately extract the characteristics of effective microseismic events. Compared with the traditional convolutional neural networks and residual networks, it  can recognize microseismic events more accurately,  the accuracy of test datasets is improved by 6.28% and 3.70% respectively.   Especially for  the  recognition accuracy of  microseismic signals with weak energy,it is higher than that of similar network models. And it has better noise resistance and generalization ability.

Key words: microseismic event recognition, U-Net, residual shrinking module, attention gate, signal processing