基于改进DenseNet的大地电磁智能反演  

Intelligent inversion of magnetotelluric data based on improved DenseNet

作　　者：姚禹 张志厚 YAO Yu;ZHANG Zhi-Hou(China Railway Design Cooperation,TianJin 300143,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]中国铁路设计集团有限公司,天津300143 [2]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《物探与化探》2024年第3期759-767,共9页Geophysical and Geochemical Exploration

基　　金：中国铁路设计集团有限公司地质勘察设计研究院内部课题“基于三维工程地质建模方法的综合勘察技术应用研究”(2022A02264005)。

摘　　要：大地电磁测深法是隧道勘查中的一种重要手段。反演技术能够将大地电磁数据转换为地电参数从而帮助地质人员解释地质资料。传统的反演方法存在时效性差、依赖初始模型设置等弊端。本研究将深度学习技术应用于一维大地电磁反演之中。首先,本研究搭建了一种改进的DenseNet网络模型并进行训练,在其完成训练之后对各种电阻率变化地层的地质模型进行反演,其计算速度快,准确率高;之后,对提出的改进DenseNet网络进行鲁棒性测试,结果表明该网络结构对于噪声数据也能取得良好的反演效果;最后,将这项人工智能技术应用于黄山地区洪家前隧道大地电磁数据的反演中,得到的物探成果与地质调研成果相匹配,并且根据反演结果给出了相关的施工建议。Magnetotelluric(MT)sounding is a vital exploration method in tunnel engineering.Inversion methods can assist geologists in interpreting geological data by converting MT data into geoelectric parameters.However,conventional inversion methods exhibit inferior timeliness and reliance on initial model settings.In this study,deep learning was applied to the one-dimensional inversion of magnetotelluric data.First,an improved DenseNet model was constructed and trained to invert geological models of various resistivity-variable strata,yielding a fast computational speed and high accuracy.Then,the robustness of the improved DenseNet model was tested,suggesting that its network structure can achieve satisfactory inversion results for noisy data.Finally,this artificial intelligence technique was applied to the MT data inversion of the Hongjiaqian tunnel in the Huangshan area,obtaining geophysical exploration results that match the geological research results.Additionally,relevant construction recommendations were given based on the inversion results.

关 键 词：大地电磁 智能反演 深度学习 隧道工程 

分 类 号：P631[天文地球—地质矿产勘探]