时域卷积神经网络地震波阻抗反演因素影响的研究  

Study on the influence of preprocessing and hyperparameters on temporal convolutional network seismic impedance inversion

作　　者：王泽峰 许辉群[1] 杨梦琼 赵桠松 王鹏[1] WANG ZeFeng;XU HuiQun;YANG MengQiong;ZHAO YaSong;WANG Peng(School of Geophysics and petroleum resources,Yangtze University,Wuhan 430100,China)

机构地区：[1]长江大学地球物理与石油资源学院,武汉430100

出　　处：《地球物理学进展》2022年第5期2062-2071,共10页Progress in Geophysics

基　　金：中国石油集团科学研究与技术开发项目资助(2021DJ3704)资助。

摘　　要：时域卷积神经网络(Temporal Convolutional Network,简称TCN)被用于地震波阻抗反演,在取得较好反演效果的同时,因其最初的成功实例主要是自然语言处理领域,而在地震反演的应用相对较少,且影响反演结果的因素众多,为了加快该方法在反演中的应用,基于前人研究的基础,注重对地震资料的预处理及网络超参数的选择进行系统分析.因此,笔者在Marmousi-2数据集实现TCN地震波阻抗反演的基础上,进一步研究噪声、归一化、随机采样三种数据预处理方法及TCN反演网络构建时学习率、dropout、批数量、通道数量四种超参数的选取对TCN地震波阻抗反演的影响,为TCN地震波阻抗反演提供预处理与超参数选择依据,并将所选取的预处理操作和超参数组合在TCN地震波阻抗反演中进行应用.该研究可为TCN地震波阻抗反演提供可行的质控手段,同时可为其他深度学习反演的数据预处理和超参数的选取提供参考,对加快基于深度学习的地震波阻抗反演应用进程具有一定的现实意义.Temporal Convolutional Network(TCN)is used for seismic impedance inversion,and while it achieves good inversion results,its initial successful examples are mainly in the field of natural language processing,while its application in seismic inversion is relatively small,and there are many factors affecting the inversion results,in order to accelerate the application of this method in inversion In order to accelerate the application of this method in inversion,we focus on the systematic analysis of the preprocessing of seismic data and the selection of network hyperparameters based on the foundation of previous studies.Therefore,based on the implementation of TCN seismic impedance inversion in the Marmousi-2 dataset,the author further investigates the effects of three data preprocessing methods of noise,normalization,and random sampling and four hyperparameters of learning rate,dropout,number of batches,and number of channels on TCN seismic impedance inversion during the construction of TCN inversion network to provide This study provides the basis for the selection of preprocessing and hyperparameters for TCN seismic impedance inversion,and applies the selected combination of preprocessing operations and hyperparameters in TCN seismic impedance inversion.This study can provide a feasible quality control tool for TCN seismic impedance inversion,and can also provide a reference for the selection of data preprocessing and hyperparameters for other deep learning inversions,which is of practical significance to accelerate the application process of deep learning-based seismic impedance inversion.

关 键 词：时域卷积神经网络 地震波阻抗反演 预处理 超参数 

分 类 号：P631[天文地球—地质矿产勘探]