基于循环神经网络的大地电磁信号噪声压制研究  

Research on noise suppression of magnetotelluric signal based on recurrent neural network

作　　者：韩盈 安志国[1,3,4] 底青云 王中兴[1,3,4] 康利利 HAN Ying;AN ZhiGuo;DI QingYun;WANG ZhongXing;KANG LiLi(Engineering Laboratory for Deep Resources Equipment and Technology,Institute of Geology and Geophysics,Chinese Academy of Sciences,Beijing 100029,China;China Earthquake Networks Center,Beijing 100000,China;Innovation Academy for Earth Science,Chinese Academy of Sciences,Beijing 100029,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院地质与地球物理研究所,深地装备技术工程实验室,北京100029 [2]中国地震台网中心,北京100000 [3]中国科学院地球科学研究院,北京100029 [4]中国科学院大学,北京100049

出　　处：《地球物理学报》2023年第10期4317-4331,共15页Chinese Journal of Geophysics

基　　金：重点研发计划(2021YFB3202104,2021YFB2301305);国家自然科学基金(41974112)资助。

摘　　要：由于天然电磁场源信号微弱,观测数据极易受到噪声干扰,严重影响反演和解释结果.传统去噪方法依赖于人工对时间序列和功率谱的筛选,去噪效率低,主观性强.本文提出利用循环神经网络对大地电磁时域信号进行特征噪声的识别和提取,进而重构出去噪后的大地电磁信号.在对大地电磁时域信号进行大量分析的基础上,对噪声进行分类并搭建含噪信号数据库,利用该数据库训练了两个循环神经网络,并选取长短时记忆单元优化循环神经网络结构,分别实现含噪数据段筛选和噪声形态提取.对仿真和实测数据分别进行了测试,循环神经网络均能准确筛选出大地电磁信号中的噪声段,本方法在避免人为操作主观性的同时提高了工作效率,视电阻率和相位曲线质量得到明显改善.As the natural electromagnetic signal is weak,the observation data is highly susceptible to noise interference,which seriously affects the inversion and interpretation results.Traditional denoising methods rely on manual screening of time series and power spectra,resulting in low denoising efficiency and strong subjectivity.In this paper,Recurrent Neural Network(RNN)is used to identify and extract the characteristic noise of magnetotelluric time domain signal,and then reconstruct the magnetotelluric signal.On the basis of a lot of analysis of magnetotelluric time-domain signals,noise is classified and a noisy signal database is built.We trained two neural networks using this database,and choose long-short term memory units to optimize the networks,respectively to screen noisy data segments and extract noise patterns.The simulated and measured data are tested respectively,and the RNN can accurately screen out the noise segment in the magnetotelluric signal.This method avoids the subjectivity of manual operation and improves the work efficiency,and the quality of apparent resistivity and phase curve is significantly improved.

关 键 词：大地电磁 循环神经网络 强干扰 去噪 机器学习 

分 类 号：P631[天文地球—地质矿产勘探]