基于位移窗口自注意力网络和迁移学习的地震面波分离  

Separation of Surface Wave from Seismic Data by Swin Transformer and Transfer Learning

作　　者：杨晨睿 沈鸿雁 车晗 孙云鹏 刘帅 YANG Chenrui;SHEN Hongyan;CHE Han;SUN Yunpeng;LIU Shuai(College of Earth Sciences and Engineering,Xi’an Shiyou University,Xi’an,Shaanxi 710065,China;Key Laboratory of Shaanxi Province for Petroleum Accumulation Geology,Xi’an,Shaanxi 710065,China)

机构地区：[1]西安石油大学地球科学与工程学院,陕西西安710065 [2]陕西省油气成藏地质学重点实验室,陕西西安710065

出　　处：《西安石油大学学报（自然科学版）》2024年第6期39-50,共12页Journal of Xi’an Shiyou University（Natural Science Edition）

基　　金：陕西省自然科学基础研究计划项目(2017JZ007);陕西省重点研发计划项目(2022GY-148)。

摘　　要：反射地震资料中的面波也携带了丰富的地质信息,充分利用这类面波解决地质问题的前提条件是将其完整地分离出来。针对现有深度学习方法没有充分利用面波与体波表现在时空域图像中的分布位置和纹理细节等问题,提出一种基于深度学习的地震面波分离方法。通过融合位移窗口自注意力机制和U-net主干网络,构建出面波智能分离网络,并使用数据增广后的模拟数据与实际数据构建面波数据集来提升神经网络的泛化性;在充分利用面波全局数据特征的同时,为避免波场分离过程中的面波损伤问题,提出一种对边界、结构和纹理信息敏感的混合损失函数以进一步提高面波分离的质量。通过数值模拟地震记录测试了本文方法的正确性,进而将该方法应用于实际地震资料处理。研究结果表明,在模拟地震记录训练的模型基础上进行迁移学习,可进一步提升神经网络的泛化性;相较于低通滤波法和去噪卷积神经网络方法,本文方法分离的面波更为完整,能大幅度提高能量混叠区域的面波分离质量。The surface waves in reflected seismic data also carry abundant geological information.The prerequisite for making full use of these surface waves to solve geological problems is to separate them completely.To solve the problem that the existing deep learning methods do not make full use of the distribution position and texture details of surface wave and body waves in time-space domain images,a seismic surface wave separation method based on deep learning was proposed.By integrating displacement window self-attention mechanism and U-net backbone network,a surface wave intelligent separation network is constructed,and a surface wave dataset is constructed using augmented simulated data and actual data to improve the generalization of the neural network.In order to make full use of the global data features of surface waves and avoid surface wave damage during wave field separation,a mixed loss function sensitive to boundary,structure and texture information was constructed to further improve the quality of surface wave separation.The correctness of the method proposed in this paper was tested through numerical simulation of seismic records,and then applied to actual seismic data processing.The research results indicate that the transfer learning based on the model trained by simulated seismic records can further enhance the generalization of neural network.Compared with low-pass filtering and denoising convolutional neural network methods,the surface waves separated using the method proposed in this paper are more complete,and the method can significantly improve the quality of surface wave separation in energy aliasing regions.

关 键 词：地震信号处理 面波 波场分离与去噪 深度学习 窗口自注意力网络 U-net网络 迁移学习 

分 类 号：TE19[石油与天然气工程—油气勘探] P315[天文地球—地震学]