面向地震全波场成像的深度学习方法研究

完整的反射地震成像应当利用来自全波场（一次波、表面多次波和层间多次波）的信息，地震全波场成像需要解决不同传播模式信号的识别、分离与成像问题。人工干预的处理难以实现信号快速精确的识别与成像。建立在Bayes框架下的稀疏和冗余信号的表达是现代信号分析理论的基础。深度学习方法基于Bayes理论，能够通过对海量标签数据的学习，识别特定传播模式的地震信号。基于这种判断，本项研究试图从深度学习对地震信号的表示与采样理论出发，研究卷积自编码器深度神经网络对地震信号的映射机制；通过模拟与人工处理的地震数据建立多次波与一次波的标签数据集；建立可对多次波与一次波分离的深度神经网络，对深度神经网络进行训练，实现多次波与一次波的智能分离与匹配成像；研究全波场反演成像的点扩散函数性质，实现点扩散函数的智能估计。本项研究聚焦在制约人工智能地震成像应用中的瓶颈问题，预期能够促进地震成像方法的智能化。

The complete reflection seismic imaging should use the information of the full wave fields (primary wave, surface multiple wave and interlayer multiple wave). The seismic full wave field imaging needs to solve the problem of identification, separation and imaging of signals with different propagation modes. It is difficult to realize the rapid and accurate signal identification and imaging by manual intervention. The expression of sparse and redundant signals built on the Bayes framework is the basis of modern signal analysis theory. Deep learning methods are based on Bayes theory, which can identify seismic signals with a specific propagation mode through the study of massive label data. Based on this judgment, this study attempts to study the mapping mechanism of convolutional auto-encoder (CAE) deep neural network to seismic signals on the basis of deep learning representation and sampling theory of seismic signals; to establish multiple wave and primary wave label data set; to establish a deep neural network that intelligently separates multiple waves and matches imaging from the primary wave; to train the deep neural network to form automatic separation of different modes of seismic waves; and to study the nature of point spread functions for full wave field inversion imaging, form an intelligent estimate of the point spread function. This research focuses on the bottlenecks that restrict the application of artificial intelligence seismic imaging and is expected to promote the intelligence of seismic imaging methods.