基于深度学习的大地电磁二维反演方法  

Two-dimensional magnetotelluric inversion method based on deep learning

作　　者：王方 熊杰[1] 田慧潇 李思平 康佳帅 WANG Fang;XIONG Jie;TIAN Huixiao;LI Siping;KANG Jiashuai(School of Electronic Information,Yangtze University,Jingzhou Hubei 434023,China)

机构地区：[1]长江大学电子信息学院,湖北荆州434023

出　　处：《地质科技通报》2024年第2期344-354,共11页Bulletin of Geological Science and Technology

基　　金：国家自然科学基金项目(62273060,61673006);长江大学大学生创新创业项目(Yz2022055)。

摘　　要：如何通过大地电磁测深反演方法来提高数据解释的精度一直都是大地电磁测深研究领域的重要课题。针对大地电磁传统反演方法存在的初始模型依赖、易陷入局部最优的问题,提出了一种基于深度学习的大地电磁二维反演方法。该方法首先设计GoogLeNetINV神经网络;接着构造多种地电模型,在TM模式下通过正演得到视电阻率数据,组成训练数据集;然后用训练数据集训练该神经网络并调整网络参数;最后,将视电阻率数据输入已训练好的GoogLeNetINV神经网络直接得到反演结果。实验结果表明,该方法能快速、准确地反演出“未学习”过地电模型的位置和电阻率数据,具有较好的泛化能力;使用噪声数据测试仍能取得良好的反演结果,有一定的抗噪声能力。将该神经网络应用于Bendigo Zone实际数据资料处理中,反演得到的电阻率模型与地震解释一致,因此基于深度学习的大地电磁反演方法能有效解决大地电磁反演问题。[Objective]The inversion of magnetotelluric sounding data to improve the accuracy of data interpretation has always been an essential topic in magnetotelluric sounding.[Methods] To address the problems of traditional magnetotelluric inversion methods,such as the dependence of the initial model and the ease of falling into a local optimum,this paper proposes a magnetotelluric inversion method based on deep learning.The method begins with the design of the GoogLeNetINV neural network.Then,various geoelectric models are constructed,and apparent resistivity data are extracted via forward modelling in the TM mode,constituting the training dataset.Additionally,the neural network is trained with the training dataset,and the network parameters are adjusted.Finally,the apparent resistivity data are input into the trained GoogLeNetINV neural network to directly obtain the inversion result.[Results] The experimental results reveal that the location and resistivity data of the "unlearned" geoelectric model can be inverted quickly and accurately,and the model has good generalization ability.The use of noise data can still yield good inversion results and a certain anti-noise ability.[Conclusion] The neural network is applied to the field data processing of the Bendigo Zone,and the resistivity model derived through inversion is consistent with the seismic interpretation.Consequently,the magnetotelluric inversion method based on deep learning can effectively solve the magnetotelluric inversion problem.

关 键 词：深度学习 大地电磁反演 神经网络 GoogLeNetINV 

分 类 号：P631.325[天文地球—地质矿产勘探]