基于深度残差网络与MVMD的多通道地磁信号处理

Multi-channel geomagnetic signal processing based on deep residual network and MVMD

地磁数据在地震预报、空间天气监测、矿产资源勘查、地球深部构造探索等领域具有重要价值.但现有的地磁台站观测数据受到人文噪声污染的问题日益严重,给地球内部高精度成像带来了极大困难.为此,我们将深度残差网络(residual network,ResNet)与多元变分模态分解(multivariate variational mode decomposition,MVMD)引入到地磁信号的处理,提出一种新颖的多通道地磁信号处理方法.首先,利用深度残差网络对大量人工标记的数据集进行训练,得到基于残差网络的地磁信噪识别模型;然后,利用训练好的模型识别出观测信号中的含噪片段;之后,利用MVMD对含噪片段进行多通道的信噪分离,得到去噪后的片段;最后,用去噪后的片段代替原始观测信号中的含噪片段,得到完整的高质量信号.为验证方法的有效性,我们设计了仿真实验,结果表明所提方法可以将观测信号的信噪比提高约15 dB,相对于变分模态分解(variational mode decomposition,VMD)、互补集合经验模态分解(complementary ensemble empirical mode decomposition,CEEMD)、数学形态滤波(mathematical morphological filtering,MMF)、小波去噪(Wavelet)等方法具有较明显的优势,且适合多通道信号的同时处理.我们将所提方法应用于菲律宾海及西太平洋的海底观测的地磁数据,结果表明所提方法的识别精度约为98%,并能够极大改善信号的质量.去噪后的信号与相邻台站的同时段高质量信号的相似度由去噪前的94.75%提升到了97.34%,表明处理结果是可靠的,使用我们的方法有望提高地磁数据成像的精度及可靠性.

The geomagnetic data are of great value in earthquake prediction,space weather monitoring,mineral resources exploration,and deep structure exploration of the earth.However,the geomagnetic data are increasingly polluted by cultural noise,which greatly complicates the high-precision imaging of the earth's interior.Therefore,we extend the deep residual network(ResNet)and multivariate variational mode decomposition(MVMD)to the processing of geomagnetic signals and propose a novel multi-channel geomagnetic signal processing method.Firstly,a large number of manually labeled data sets are trained by ResNet to obtain a signal-to-noise recognition model.Then the trained model is used to identify the noisy fragments from the raw observation signal.Hereafter,MVMD is adopted to perform multi-channel signal-to-noise separation on noisy segments,and the denoised segments are obtained.Finally,the noisy segments in the original observation signal are replaced by the denoised segments to obtain a complete high-quality signal.To verify the effectiveness of the method,we designed simulation experiments.The results show that the proposed method can improve the signal-to-noise ratio of the observed signal by about 15 dB,which has obvious advantages over VMD,complementary ensemble empirical mode decomposition(CEEMD),mathematical morphological filtering(MMF),and Wavelet,and is suitable for the batching processing of multi-channel signals.We apply the proposed method to the geomagnetic data observed in the Philippine Sea and the Western Pacific Ocean.The results show that the recognition accuracy of the proposed method is about 98%,and can greatly improve the signal quality.The normalized cross-correlation between the processed signal and the high-quality signal of the adjacent station at the same time has increased from 94.75%before denoising to 97.34%,indicating that the result is reliable.Our method is expected to improve the accuracy and reliability of geomagnetic data imaging.