王坤喜^1,,
胡天跃^1,,,
刘小舟¹,
王尚旭²,
魏建新²
1. 北京大学地球与空间科学学院, 北京 100871
2. 中国石油大学(北京)油气资源与探测国家重点实验室, 北京 102249

基金项目: 国家重点研发计划（2018YFA0702500），国家自然科学基金（41674122）和国家重点基础研究发展计划（2013CB228600）联合资助


详细信息

作者简介: 王坤喜, 男, 1993年生, 博士研究生, 主要从事地震资料处理方法研究.E-mail: kxwang@stu.pku.edu.cn

通讯作者: 胡天跃, 男, 1963年生, 教授, 主要从事应用地球物理的教学与科研工作.E-mail: tianyue@pku.edu.cn

中图分类号: P631

收稿日期:2021-01-04
修回日期:2021-10-19
上线日期:2021-11-10

Suppressing seismic multiples based on the deep neural network method with data augmentation training

WANG KunXi^1,,
HU TianYue^1,,,
LIU XiaoZhou¹,
WANG ShangXu²,
WEI JianXin²
1. School of Earth and Space Sciences, Peking University, Beijing 100871, China
2. State Key Laboratory of Petroleum Resources and Prospecting, China University of Petroleum, Beijing 102249, China


More Information

Corresponding author: HU TianYue,E-mail:tianyue@pku.edu.cn

MSC: P631

--> Received Date: 04 January 2021
Revised Date: 19 October 2021
Available Online: 10 November 2021


摘要
摘要:地震数据中存在的多次波影响偏移成像，误导地震资料的解释，因此通常视为相干噪声而被去除.为了对多次波进行智能化衰减，本文提出了一种基于数据增广训练的使用深度神经网络的多次波压制方法.设计的深度神经网络包括卷积编码和卷积解码过程，其中卷积编码过程学习全波场数据中的一次波特征，卷积解码过程利用这些特征来重构一次波并压制多次波和随机噪声.在训练阶段，旋转训练集并在输入数据中加入随机噪声构成增广训练数据集来提升神经网络的抗噪稳定性和泛化性，通过迁移学习让深度神经网络具备跨工区压制多次波的能力.简单模型与Sigsbee2B模型三套模拟数据的实例验证了本文方法在一次波重构和多次波压制中的有效性、稳定性和良好泛化性；一套崎岖海底模型地震物理模拟数据的应用实例表明本文方法具有应用于复杂条件下压制地震多次波的能力.
关键词: 深度神经网络/
数据增广训练/
多次波压制/
迁移学习

Abstract:Multiples of seismic data can affect migration imaging and mislead the interpretation of seismic data, so they are often treated as coherent noise to be removed. In order to intelligently attenuate multiples, this paper proposes a multiple suppression based on the deep neural network method with data augmentation training. The designed deep neural network includes convolutional encoding and convolutional decoding processes, where the convolutional encoding process learns the features of the primaries in the full wavefield data, and the convolutional decoding process uses these features to reconstruct the primaries and suppress the multiples and random noise. In the training phase, the training set is rotated and the input data is added with random noise to form the augmented training data set to improve the anti-noise stability and generalization of the neural network. Through transfer learning, the deep neural network has the ability to suppress multiples in other working areas. Three sets of synthetic data examples of the simple model and the Sigsbee2B model verify the effectiveness, stability and good generalization of the proposed method in primaries reconstruction and multiples suppression; An example of the seismic physical simulation of a marine rugged bed model shows that the proposed method can be applied to suppress seismic multiples under complex conditions.
Key words:Deep neural network/
Data augmented training/
Multiple suppression/
Transfer learning



PDF全文下载地址:

http://www.geophy.cn/data/article/export-pdf?id=8be1008c-e62d-48c8-95a7-da6d0d601c80