Double-phase-shift filtering method for harmonic elimination based on AR2U-Net  

基于AR2U-Net的双相移谐波压制滤波法

作　　者：Li Bo-Lin Wang Yan-Chun Yuan Hang Liu Xue-Qing 李波霖;王彦春;袁航;刘学清(中国地质大学(北京)地球物理与信息技术学院,北京100083;北京京能油气资源开发有限公司,北京100022)

机构地区：[1]School of Geophysics and Information Technology,China University of Geoscience,Beijing 100083,China [2]Beijing Oil and Gas Resources Development Company,Beijing 100022,China

出　　处：《Applied Geophysics》2022年第2期271-283,309,共14页应用地球物理（英文版）

基　　金：supported by the National Science and Technology Major Project of China(No.2016ZX05003-003).

摘　　要：The double-phase-shift filtering method,which is based on the traditional purephase-shift filtering method,is a novel approach to harmonic elimination that can be applied to more complicated signals such as white noise and slip-sweep.Nonetheless,any type of phase-shift filtering method necessitates a relationship between the frequency of fundamental sweep and time,which may cost necessitate an enormous amount of human and physical resources to achieve inaccurate results with low efficiency.This paper combines deep learning with harmonic elimination to produce a double-phase-shift filtering method based on AR2UNet,a type of U-Net with attention gates structure and recurrent residual blocks for improving accuracy and function while simplifying computational complexity.The input of the AR2UNet structure in this paper is seismic data of slip-sweep signals in vibroseis,and the output is signal frequency variation with the time of the fundamental waves,which are required to eliminate the harmonic waves and adjacent signals using a double-phase-shift method to obtain the fundamental sweep.The training sets and test sets are formed by forward models,and a Log-Cosh loss function is used to monitor the process,during which the results of AR2U-Net and traditional U-Net are compared to demonstrate the eminent function of AR2UNet.Following that,the outcomes’Log-Cosh loss functions and accuracy are also compared to validate the conclusion.AR2U-Net,when applied to raw data and combined with the doublephase-shift method,tends to polish the filtering effects and is worth promoting.双相移谐波压制滤波方法是在传统相移法滤波的基础上加以改进,能够应用于噪声、滑动扫描信号等复杂信号的一种新型谐波去除方法,但该方法在应用于实际资料时由于需要基波的时段信息和频段信息,而人工处理不仅耗费大量精力、效率低下,且精度不高。因此,本文将深度学习和谐波压制相结合,提出基于AR2U-Net的双相移谐波压制滤波方法,其中AR2U-Net神经网络是在传统U-Net神经网络中加入了注意门结构与递归残差块,使得网络在节省参数数量、简化计算复杂度的同时,提高了预测精度并强化了网络性能。本文AR2U-Net神经网络结构的输入为可控震源滑动扫描地震数据,输出为双相移法进行谐波压制所需要的基波信号的频率随时间的变化图,通过该图得到其相移曲线并使用双相移法还原基波信号。使用正演模拟构建了网络的训练集和测试集,训练过程中使用更平滑的Log-Cosh损失函数监控网络性能,并在训练集和测试集中分别与传统U-Net神经网络进行对比,验证了AR2U-Net的优良性能。将AR2U-Net与双相移法相结合应用于实际数据中,表明该方法滤除效果可以得到保证,具有一定的推广价值。

关 键 词：AR2U-Net harmonic elimination double-phase-shifts deep learning VIBROSEIS 

分 类 号：P631.4[天文地球—地质矿产勘探]