Phase arrival picking for bridging multi-source downhole microseismic data using deep transfer learning

Abstract

Abstract The phase arrival picking of the downhole microseismic dataset is a critical step in fracturing monitoring data processing. Recently, data-driven methods have been widely used in seismology studies, especially in seismic phase picking. The picking results heavily depend on whether large quantities of accurately labeled phase samples could be obtained to extract the characteristics of seismic waveforms. Also, there is a shortcoming of poor generalization ability in dealing with the cross-source transfer scenarios. In this paper, we propose a novel deep transfer learning method for microseismic phase arrival picking by fine-tuning one existing pretrained model based on a few phase samples. The pretrained model, which has been domain-adapted for phase picking, adopts 2D U-Net to both extract time and space features, thereby improving the overall picking accuracy. Moreover, the fully convolutional U-Net architecture has the ability to handle samples with variable sizes so could be used for bridging downhole microseismic data from different sources. The results of two transfer cases show that compared with the direct application of the pretrained model and a newly trained model, the proposed method could provide more satisfactory performance with only limited seismic phase samples. Also, our method significantly reduces the cost of labeling and saves time because of avoiding repeated training.