基于ε-SVR算法的大地电磁测深资料去噪  

Noise elimination for magnetotelluric sounding data based on ε-SVR algorithm

作　　者：程怀蒙 张胜业[1] 

机构地区：[1]中国地质大学地球物理与空间信息学院,武汉430074

出　　处：《地球物理学进展》2014年第2期668-673,共6页Progress in Geophysics

摘　　要：消除噪声干扰对大地电磁测深资料的影响是大地电磁(MT)工作中的首要问题.基于结构风险最小化原则的支持向量机能够解决小样本情况下非线性函数拟合的通用性和推广性问题,是求复杂的非线性拟合函数的一种有效技术.本文首先介绍了ε-SVR(ε不敏感损失函数—支持向量机回归)的原理及SVR相关参数的选择,在此基础上,利用该算法对大地电磁测深实测资料进行回归处理,并与当前常用的去噪方法(Robust变换结合人工筛选)进行对比,结果表明ε-SVR算法可以较好地消除MT测深曲线所受噪声影响,提高工作效率.同时给原始数据加入10%噪声,并利用该算法对加噪后的数据进行回归处理,加噪前后拟合结果的绝对误差的均方差为9.454,相对误差的均方差为1.61%,证明该模型具有良好的泛化能力和稳健性.Eliminate the noise impact of magnetotelluric sounding data is the most important issue in the MT work, currently the more popular approach on magnetotelluric sounding data processing is to use Robust transform with a combination of artificial selection methods to improve data quality, and sometimes need to use the remote reference, this method effect were satisfactory in the practical application, but when the amount of data is a large number and the interference is serious, the artificial selection will be time-consuming and labor-intensive, this need to high personal qualities. Statistical learning theory is a basic theoretical and mathematical framework which specialized study for machine learning the law in the case of small samples, and is also the best theory for small sample statistical estimation and forecasting study,the support vector machine based on the principle of structural risk minimization can solve the nonlinear function fitting versatility and promotion issues in the case of small sample, which is an effective technology for seeking the complex nonlinear fitting functions. This paper first introduced the principle of ε-support vector machine regression （ε-SVR） and the selection of SVR relevant parameters. On this basis, the regression processing on the measured magnetotelluric sounding data was did by this algorithm, and the current commonly denoising method （Robust transform combined with artificial selection） was compared. The result shows that ε-SVR algorithm can significantly eliminate the MT sounding curves suffered noise impact, improve work efficiency. Finally, 10% noise points was added into the original data, which was carried on regression processing by this algorithm. The absolute mean square error and relative mean square error of the fitting results before adding noise and after were 9.454 and 1.61% respectively. It proves that the model has good generalization ability and robustness.

关 键 词：大地电磁测深 ε不敏感损失函数 支持向量机 回归 去噪 

分 类 号：P631[天文地球—地质矿产勘探]