剪切波分裂是分析地震各向异性的一种重要手段，常规方法是利用网格搜索获取分裂参数，再通过不同方法的测量结果对比测量结果进行质量检测，这一过程会耗费大量计算时间。本文针对这一问题提出了一种利用深度卷积神经网络对剪切波分裂进行质量检测的新方法，对使用了Resnet残差结构的深度神经网络进行训练，直接对二分量剪切波波形数据的质量进行分类。整个过程为：神经网络通过卷积层提取波形特征，计算损失函数后反向传播训练模型参数，完成迭代训练后的模型对输入波形数据正向计算自动输出类型。本文利用川西台站接收到的实际数据以及随机生成的合成数据分别对该网络进行训练，均可以获得准确的分类结果。相比于通过多种剪切波分裂方法对比测量结果的质量检测方法，基于神经网络的方法可以省略网格搜索的计算过程直接判断质量类型，在运算速度上的优势明显，并可继续通过训练提高模型的精度，为提升剪切波分裂方法在数据处理过程中的操作效率提供帮助。

Shear-wave splitting is an important method to analyze seismic anisotropy. Some conventional methods use grid search to obtain splitting parameters,and then compare the measurement results of different methods for quality detection. This process consumes a lot of calculation time on poor quality data. In this paper,a new method for quality detection of shear wave splitting using deep neural networks is proposed for solving this problem. A deep neural network with a total of 9 convolutional layers based on the Resnet residual structure is constructed,which can directly classify the quality of two-component shear-wave waveform data. In the whole process,the neural network extracts the waveform features through the convolutional layer,calculates the loss function and then backpropagates to train model parameters. After the iterative training,the model could automatically classify the quality type by forward calculation for input waveform data. This paper uses synthetic data and actual data to train the network,and both can obtain accurate classification results. Compared with the results by some shear-wave splitting methods,the neural-network-based method can omit the calculation process of grid search and directly judge the quality type,which has obvious advantages in computing speed and can continue training to improve the accuracy of the model,and helps to improve the operational efficiency of the shear wave splitting method during data processing.

P315

川滇地区三维公共速度模型构建与评价（2018CSES0101）、中国科学技术大学创新团队项目《地幔底部的组成及动力学》（WK2080000078）共同资助