褶积神经网络高分辨率地震反演  

High resolution seismic inversion by convolutional neural network

作　　者：张繁昌[1] 刘汉卿 钮学民[2] 代荣获 

机构地区：[1]中国石油大学(华东)地球科学与技术学院,山东青岛266555 [2]中国石化胜利油田物探研究院,山东东营257022

出　　处：《石油地球物理勘探》2014年第6期1165-1169,5-6,共5页Oil Geophysical Prospecting

基　　金：国家"863"项目(2011AA060302);国家"973"项目(2013CB228604);中国石油大学(华东)研究生创新工程基金项目(YCX2014003)联合资助

摘　　要：随着地震勘探精细化要求的提高,薄层及横向变化大的复杂储层反演越来越重要。而当前反演方法大多基于褶积模型,分辨率较低。本文提出了基于褶积神经网络的反演方法,该方法完全由数据驱动,不受褶积模型的限制。褶积神经网络具有层状结构,其输入输出之间的映射关系用褶积算子来描述,而非内积算子。基于褶积神经网络结构,本文给出了映射算子的优化算法,并将其应用到地震反演中。应用结果表明,通过褶积神经网络地震反演,可以获得比常规稀疏脉冲反演分辨率更高的地层波阻抗剖面。With the requirements of high-accuracy seismic exploration,the inversion technique for thin beds and complex reservoirs with large lateral variation is becoming more and more important.However,the current inversion methods are mainly based on the convolutional model,bearing with poor resolution.In order to improve the resolution of inversion results,this paper presents an inversion method based on the convolutional neural network,which is totally driven by data and not re-stricted by convolutional model.The convolutional neural network has a layered structure,whose mapping relationship between its input and output is described by convolutional operators instead of inner product operators.Based on the convolutional neural network structure,the paper further provides the optimization algorithm for mapping operators and applies it to seismic inversion process.Application results show that the convolutional neural network inversion can get higher resolution impedance profile than the conventional sparse pulse inversion method.

关 键 词：褶积神经网络 高分辨率 映射算子 数据驱动 地震反演 

分 类 号：P631[天文地球—地质矿产勘探]