基于Attention-UNet网络的速度模型构建方法研究  

Research on the construction method of speed model based on Attention-UNet network

作　　者：孙德辉 王云专[1] 王莉利[2] SUN Dehui;WANG Yunzhuan;WANG Lili(School of Earth Sciences,Northeast Petroleum University,Daqing,Heilongjiang 163318,China;School of Computer and Information Technology,Northeast Petroleum University,Daqing,Heilongjiang 163318,China)

机构地区：[1]东北石油大学地球科学学院,大庆163318 [2]东北石油大学计算机与信息技术学院,大庆163318

出　　处：《物探化探计算技术》2024年第1期1-10,共10页Computing Techniques For Geophysical and Geochemical Exploration

基　　金：国家重点自然科学基金(41930431);东北石油大学引导性创新基金(2020YDL-03)。

摘　　要：随着油气资源的不断勘探开发,相对易开采的油气矿逐渐建成,地震勘探的研究重点也向地下更深、构造更复杂的区域转移。目前,传统的地震速度建模方法在稳定性、准确性和计算效率方面都面临挑战。因此,笔者利用将地震数据映射到速度模型的思路,提出了一种基于Attention-UNet网络的深度学习速度建模方法。采用的这种方法利用有限差分正演得到反射波形数据,将反射波形数据和对应的速度模型(标签)对作为Attention-UNet网络的输入,建立地震数据和速度模型之间的映射关系。网络训练后对新输入的地震数据进行速度模型的估计。数值实验结果表明,与传统的FWI相比,笔者提出的方法表现出良好的性能;基于Attention-UNet网络模型训练完成后,不需要经过大量的计算,就可以快速执行与训练集中速度结构相似的地下结构的速度建模,这比传统方法计算效率更高。该方法在建立大量速度模型时具有很好的推广价值。As oil and gas resources continue to be explored and developed,relatively easy-to-reach oil and gas plays are gradually being built up,and the focus of seismic exploration research has shifted to more profound and more structurally complex areas in the subsurface.Traditional seismic velocity modeling methods face stability,accuracy,and computational efficiency challenges.Therefore,this paper proposes a deep learning velocity modeling method based on attention to UNet networks to map seismic data to velocity models.The method obtains reflection wave data through finite-difference forward modeling.It uses the reflection wave data and the corresponding velocity model(label)pair as the input to the attention UNet network to establish the mapping relationship between the seismic data and the velocity model.After network training,the velocity model is estimated for the newly input seismic data.Numerical results show that the method also exhibits good performance compared to conventional FWI;once the training of the attention-based UNet network model is completed,velocity modeling of subsurface structures similar to the velocity structure in the training set can be performed quickly without extensive computation,providing higher computational efficiency than conventional methods.The method has good extension value in building many velocity models.

关 键 词：速度建模 注意力机制 UNet 全波形反演 

分 类 号：P631.4[天文地球—地质矿产勘探]