基于字典学习的音频大地电磁数据处理

Denoising AMT data based on dictionary learning

音频大地电磁(Audio Magnetotelluric,AMT)信号常常受到持续性人文噪声影响,这类噪声使用远参考法和Robust阻抗估计等常规方法往往效果不佳.为此,本文从噪声的规律与特征出发,提出一种新的AMT数据处理方法.首先通过字典学习方法从观测数据中自主学习到人文噪声的特征结构,构建冗余字典,然后利用学习到的冗余字典,分离出AMT数据中的人文噪声.为验证方法的有效性,首先进行了合成数据的仿真试验,然后在四川凉山进行了针对性的野外试验研究,最后将本文方法应用于庐枞矿集区实测数据的处理.结果表明,本文方法能够快速、准确地分离出AMT信号中的人文干扰,保留有用信号,显著改善AMT数据质量.

The problem that audio magnetotelluric(AMT)data is susceptible to cultural noise is far from being solved.Especially when contaminated by persistent cultural noise,the use of remote reference,robust estimate or other conventional methods can hardly acquire a good result.To remove the persistent cultural noise,this paper proposes a novel method based on dictionary learning.First of all,the features of the cultural noises are learned from observational data through a dictionary learning algorithm,and then the cultural noises are separated by using the learned dictionary.We apply our procedure to synthetic and real data with strong cultural noise and compare different processing methods to that proposed in this paper.As a conclusion,the presented method can quickly and accurately extract the strong cultural noise from raw AMT data and preserve the useful signal completely,and the results acquired by using the filtered data are improved significantly with respect to previous work.