基于U-Net的井中多道联合微地震震相识别和初至拾取方法

微地震震相识别和初至拾取是水力压裂微地震监测资料处理中的两个关键步骤，其结果会对后续事件定位和压裂裂缝缝网解释产生重要影响.常规方法如STA/LTA法、模板匹配法、多道互相关法等需要提取有效信号与噪声间振幅、偏振、频率、波形相似性等方面的特征差异完成震相识别和拾取工作.本文基于深度学习技术的自动特征提取能力，根据井中微地震观测系统的多道数据源特点，提出基于U-Net的多道联合震相识别和初至拾取方法（MT-Net）.方法采用具有"逐采样点"识别能力的U-Net模型，模型训练阶段以具有不同信号特征的多道微地震监测记录作为输入，以P波、S波及噪声的概率分布标签作为输出，通过设置二维卷积操作使得道内与道间的波形信息同时被自适应地学习，以满足对相邻道间波形记录处理结果高度一致性的要求；测试阶段将连续记录中的分段波形馈入模型，通过设定P波、S波概率分布曲线阈值完成单震相、双震相和噪声的波形分类，同时对含有效震相的微地震事件完成初至拾取.实际微地震资料处理结果显示，本文方法与同样基于U-Net的单道方法（ST-Net）相比，显著降低了震相识别中低信噪比事件漏拾与误拾发生的概率；同时有效避免了部分单道发生严重的初至拾取结果偏差及P、S震相误拾等情况.本文方法的识别与拾取结果整体上达到了与多道互相关法接近的水平，可满足微地震监测资料处理中实时性和准确性的要求.

Multi-trace joint downhole microseismic phase detection and arrival picking method based on U-Net

Microseismic phase identification and arrival picking are two key steps in the data processing of hydraulic fracturing microseismic monitoring, the results of which have great influences on the subsequent event location and the interpretation of hydraulic fractures. To complete the phase identification and arrival picking, conventional methods (such as STA/LTA method, template matching method, multi-trace cross-correlation method, etc.) need to extract characteristic differences between valid signals and noise in amplitude, polarization, frequency, and waveform similarity. Based on the automatic characteristic extraction capability of deep learning technology, considering the characteristics of multi-trace data sources in the downhole microseismic observation system, this paper has proposed a U-Net-based multi-trace joint microseismic phase detection and arrival picking method (MT-Net). Our method adopts the U-Net model, which has the "point-by-point" detection ability. In the model training stage, multi-trace microseismic monitoring records with different signal characteristics are input to the model, the probability distribution labels of P waves, S waves and noises are used as model output. The two-dimensional convolution operation enables the "in-trace" and "inter-trace" waveform information to be adaptively learned meanwhile, achieving the goal of high consistency in the processing results between waveform records of adjacent traces. In the model testing stage, the segmented waveforms in the continuous record are fed into the model, and the waveform classification among single-phases, double-phases and noises is completed by setting the probability distribution thresholds of P waves and S waves. Meanwhile, the arrival picking of valid microseismic events is completed. The processing results of actual microseismic data show that compared with the single-trace method similarly based on U-Net (ST-Net), our method significantly reduces the probability of missed and mistaken events with low signal-to-noise ratio in the phase detection. Also, serious deviations of the picking results on some single traces and the false pickings between the P and S phases are effectively avoided by our method in the arrival picking. The overall picking result of our method has reached a level close to that of the multi-trace cross-correlation method, which can meet the requirements of real-time and accuracy in the microseismic monitoring data processing.