基于深度全卷积神经网络的地震波阻抗预测方法  

Seismic Impedance Inversion Method Based on Deep Fully Convolutional Network

作　　者：王泽峰 孙颖 许辉群[1] 赵桠松 Wang Zefeng;Sun Ying;Xu Huiqun;Zhao Yasong(School of Geophysics and Petroleum Resources,Yangtze University,Wuhan Hubei 430100,China;Petrochina Bohai Drilling Co.,LTD,Tianjin 300457,China)

机构地区：[1]长江大学地球物理与石油资源学院,湖北武汉430100 [2]中国石油渤海钻探油气合作开发分公司,天津300457

出　　处：《工程地球物理学报》2022年第3期386-392,共7页Chinese Journal of Engineering Geophysics

基　　金：中国石油创新基金(编号:2018D-5007-0301)。

摘　　要：波阻抗反演是高分辨率地震资料处理的最终表达形式,线性的地震波阻抗反演方法求解精度依赖于初始地质模型,为高效地求解得到完全非线性的反演结果,地震波阻抗反演迫切需要发展智能化的反演技术。鉴于此,本文在全卷积神经网络(Fully Convolutional Network,FCN)的基础上,提出一种深度全卷积神经网络的地震波阻抗预测方法,以此来实现波阻抗反演。其实现过程是通过深度全卷积神经网络对正演模型和对应的波阻抗标签训练建立非线性的映射关系,得到反演映射模型,进一步通过该反演映射模型预测地震波阻抗。正演数据测试和实际资料应用结果表明,该方法可以实现地震与波阻抗之间的非线性映射,为地震波阻抗反演提供一种智能化的新手段。Impedance inversion is the ultimate expression of high resolution seismic data processing.The accuracy of linear seismic impedance inversion method depends on the initial geological model.In order to efficiently obtain complete nonlinear inversion results,it is urgent to develop intelligent inversion technology.Based on fully convolutional network(FCN),the seismic impedance prediction method of deep convolutional neural network is proposed to realize impedance inversion.The realization process is to establish nonlinear mapping relationship between the forward model and the corresponding impedance label training by deep convolutional neural network,obtain the inversion mapping model,and then predict the seismic impedance by the inversion mapping model.According to the test results of the forward modeling data,this method can realize the nonlinear mapping between seismic and impedance,provide a new intelligent method for seismic impedance inversion and be further applied to the actual data of SMI work area.

关 键 词：地震波阻抗反演 非线性 全卷积神经网络 反演映射模型 

分 类 号：P631.4[天文地球—地质矿产勘探]