吉林大学学报(信息科学版) ›› 2025, Vol. 43 ›› Issue (3): 489-496.

• • 上一篇    下一篇

基于多尺度和注意力特征融合的地震去噪方法

 王瑞敏,杨文博,邓  聪,鲁统祥,张文祥   

  1. 中海石油(中国)有限公司湛江分公司,广东湛江524057
  • 收稿日期:2024-11-22 出版日期:2025-06-19 发布日期:2025-06-19
  • 通讯作者: 杨文博(1983— ), 男, 甘肃张掖人, 中海石油(中国)有限公司 湛江分公司物探工程师,主要从事地震采集处理、储层预测研究,(Tel)86-18665765155(E-mail)yangwb2@cnooc.com.cn。
  • 作者简介:王瑞敏(1984— ),女,山东聊城人,中海石油(中国)有限公司湛江分公司物探工程师,主要从事地震采集处理、信息技术研究, (Tel)86-13763063636(E-mail)wangrm2@ cnoon. com. cn
  • 基金资助:
    十四五冶重大科技基金资助项目(KJGG2022-0304)

Seismic Denoising Method of Multiscale and Attentional Feature Fusion

WANG Ruimin, YANG Wenbo, DENG Cong, LU Tongxiang, ZHANG Wenxiang    

  1. Zhanjiang Branch, China National Offshore Oil Corporation (China) Limited, Zhanjiang 524057, China
  • Received:2024-11-22 Online:2025-06-19 Published:2025-06-19

摘要: 针对现有卷积神经网络在强背景噪声下弱信号恢复不足的问题,提出了一个基于多尺度U-Net和注意力 融合机制的去噪网络(MAUnet: Multi-Scale U-Net and An Attention Fusion Mechanism)。 创新性地构建了双重机 制, 通过多尺度模块使网络学习不同尺度上的特征;通过注意力特征融合机制使网络可将浅层高频细节与深层 语义信息融合,增强网络的学习能力,实现特征互补。 实验结果表明,相较于对比方法,所提方法具有更好的 噪声衰减和弱信号恢复能力。

关键词: 卷积神经网络, 地震数据去噪, 多尺度特征提取, 注意力特征融合 

Abstract: Due to the limitation of environmental and economic factors, the collected seismic records usually have a lot of noise interference, which may cause some obstacles to the subsequent seismic data processing. Effectively attenuating noise is a key issue in seismology. In recent years, CNNs ( Convolutional Neural Networks) have achieved some success in the field of seismic data denoising. However, weak signal recovery in the presence of strong background noise is insufficient for existing convolutional neural networks. To address the above issues, a denoising network called MAUnet(Multi-Scale U-Net and An Attention Fusion Mechanism) is proposed. based on a multi-scale U-Net and an attention fusion mechanism. MAUnet innovatively introduces a dual-mechanism architecture, where a multi-scale module enables the network to learn features at different scales. And an attention-based feature fusion mechanism allows the network to combine shallow high-frequency details with deep semantic information, enhancing its learning capability and achieving feature complementarity. Experimental results demonstrate that our method has better noise attenuation and recovery capability for weak signals than competitive methods. 

Key words: convolutional neural network (CNN), seismic data denoising, multi-scale feature extraction, attention feature fusion

中图分类号: 

  • TP391.4