地面多分量地震数据P/S波分离的深度学习方法

多分量地震记录P/S波分离是多波地震数据处理的关键技术环节.常规方法大多依据两种波模式视速度或偏振特征的差异,基于信号分析或偏振投影实现模式解耦.在许多实际的地震-地质条件下,这些基于信号特征假设或表层参数模型的P/S波分离方法往往不太有效.为此,本文将各向同性介质条件下的地面多分量地震数据P/S波分离视为非线性的逐点预测问题,借助深度神经网络强大的特征提取能力进行求解.以国际标准模型为基础,提出了创建弹性参数样本库和P/S波分离标签数据集的有效方法.实验表明,丰富的训练样本保证了深度神经网络的泛化性能,在测试数据体上取得了明显优于经典的偏振投影分离方法的处理效果,而且摆脱了对表层介质参数的依赖性,为多分量地震数据反射PP波和PS波成像提供了有效的技术支撑.

P/S separation of multi-component seismograms using a deep learning method

P/S separation is a key step in multi-component seismic data processing. Based on the differences of apparent velocity or polarization direction of P- and S-waves, the multicomponent data can be separated using the signal processing or polarization projection methods. In practical seismic or geological situations, these conventional P/S separation methods that highly rely on the signal characteristics or surface physical properties usually fail to deliver satisfactory results. Accordingly, in this paper, we regard the P/S separation of multi-component seismic data in isotropic media as a nonlinear point-by-point prediction problem, which can be addressed by the deep neural network with its powerful ability of feature extraction. Based on the standard public geological models, we propose an effective approach to build the elastic parameter models and the P/S separation labels. The numerical examples show that the abundant training samples can improve the generalization of the deep neural network, with which the trained network model achieves a better performance than the classical polarization projection method on the target test data. The proposed P/S separation method gets rid of the dependence on surface elastic parameters, which provides an effective support for the reflection PP wave and PS wave imaging of multi-component seismic data.