基于深度卷积神经网络的地震震相拾取方法研究

地震震相拾取是地震数据自动处理的首要环节,包括了信号检测、到时估计和震相识别等过程,震相拾取的准确性直接影响到后续事件关联处理的性能,影响观测报告的质量.为了提高震相拾取的准确性,进而提高观测报告质量,本文采用深度卷积神经网络方法来解决震相拾取问题,构建了多任务卷积神经网络模型,设计了分类和回归的联合损失函数,定义了基于加权的分类损失函数,以三分量地震台站的波形数据作为输入,同时实现对震相的检测识别和到时的精确估计.利用美国南加州地震台网的200万条震相和噪声数据对模型进行训练、验证和测试,对于测试集中直达波P、S震相识别的查全率达到98%以上,到时估计的标准偏差分别为0.067s,0.082s.利用迁移学习和数据增强,将模型用于对我国东北地区台网的6个台站13000条数据的训练、验证和测试中,对该数据集P、S震相查全率分别达到91.21%、85.65%.基于迁移训练后的模型,设计了用于连续数据的震相拾取方法,利用连续的地震数据对该算法进行了实际应用测试,并与国家数据中心和中国地震局的观测报告进行比对,该方法的震相检测识别率平均可达84.5%,验证了该方法在实际应用中的有效性.本文所提出的方法展示了深度神经网络在地震震相拾取中的优异性能,为地震震相和事件的检测识别提供了新的思路.

Research on the seismic phase picking method based on the deep convolution neural network

Picking seismic phase is the primary step of automatic seismic data processing, including signal detection, arrival time estimation and phase identification. The accuracy of seismic phase picking affects the performance of subsequent phase association processing and the quality of event bulletin. Scholars have carried out a lot of research on this issue, while further improvement is needed. In this work, the deep convolution neural network method was used to solve the seismic phase picking problem. A multi-task convolution neural network model was constructed. The joint loss function of classification and regression was designed. The weighted classification loss function was defined. The method took the waveform data of 3-component seismic stations as input, and realized the detection and identification of the seismic phase and the estimation of phase onset. The model was trained, verified and tested using 2 million phase and noise data from the Southern California Seismic Network (SCSN) of American. For the test set, the recall rate of direct P and S phase identification of direct waves reached over 98%, and the standard deviations of arrival time estimation for P and S waves were 0.067 and 0.082, respectively. Using transfer learning and data augmentation methods, the model was applied to a small dataset of some stations in the seismic network of northeast China, and the recall rates of direct P and S phases identification reached 91.21% and 85.65%, respectively. Based on the model after transfer learning, we designed a phase picking method for continuous data. The method was tested in practical application using continuous real waveform data. The results were compared with the National Data Center bulletin and the China Earthquake Administration bulletin. The phase detection rate of this method reached 84.5%, and the effectiveness of the method was verified. The method proposed in this paper demonstrates the excellent performance of deep neural network in seismic phase picking and provides a new idea for the detection and identification of seismic phases and events.