深度学习在视电阻率快速反演中的研究与应用  

Research and application of deep learning in rapid inversion of apparent resistivity

作　　者：余传涛[1,2] 李子伦 薛俊杰 YU ChuanTao;LI ZiLun;XUE JunJie(College of Mining Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Institute of Geology and Geophysics,Chinese Academy of Sciences,Beijing 100029,China;Institutions of Earth Science,Chinese Academy of Sciences,Beijing 100029,China;University of Chinese Academy of Sciences,Beijing 100029,China)

机构地区：[1]太原理工大学矿业工程学院,太原030024 [2]中国科学院地质与地球物理研究所,北京100029 [3]中国科学院地球科学研究院,北京100029 [4]中国科学院大学地球与行星科学学院,北京100029

出　　处：《地球物理学报》2024年第11期4385-4399,共15页Chinese Journal of Geophysics

基　　金：山西省自然科学基金项目(202303021221050);山西省回国留学人员科研资助项目(2022-076)资助。

摘　　要：在电法勘探中, 通过测量视电阻率反演地下介质真实电阻率值, 会受到异常体的大小、角度, 测量装置的电流、排布, 周围围岩分布以及地形等诸多因素影响.传统反演方法将非线性近似看作线性处理, 造成反演结果的不确定性和存在多解性.本文通过深度学习的非线性拟合能力, 实现对视电阻率值的高效反演.参照RepVGG下采样架构, 搭建了未使用残差连接结构的全卷积神经网络ARESInvNet, 训练完成后的神经网络在训练集和验证集上的准确率为99%, 进行结构重参数化优化后, 相较于优化操作之前, 在CPU上的反演时间缩短44%, 内存占用减少50%;对测试集上层状介质视电阻率数据进行反演, 反演的准确率达到了98%;在大小10%以内的高斯白噪声视电阻率数据上进行反演, 测试集的准确率为96%, 证明ARESInvNet有很好的抗干扰能力;通过对实测视电阻率数据的反演, 表明神经网络能够准确地反映地层界面位置和地形起伏形态, 可为实际电法勘探工作提供快速准确的反演结果.In the field of electrical exploration, the inversion of actual resistivity values of the subsurface medium through the measured apparent resistivity is subject to various factors, including anomaly size and angle, current and arrangement of measurement devices, surrounding rock and topography. Traditional inversion methods treat non-linear approximations as linear processing, leading to uncertainty and multiple solutions to the inversion results. This paper achieves an efficient inversion of apparent resistivity values, utilizing the non-linear fitting capabilities of deep learning. Buildering a fully convolutional neural network, ARESInvNet, without residual connections, based on the downsampling architecture of RepVGG, the trained network achieves 99% accuracy on both the training and validation sets, applying network structural re-parameterization led to a 44% reduction in inversion time on the CPU and a 50% decrease in memory usage compared to that before the optimization operation. The inversion accuracy of the layered medium apparent resistivity data on the test set reached 98%. Inversion on Gaussian white noise apparent resistivity data within 10% of the size, with an accuracy of 96% on the test set, indicating its effective resistance against interference. The network reliably reflects the position of stratigraphic interface location and topographic relief pattern through the inversion of measured apparent resistivity data, thereby providing fast and accurate inversion results for real-world electrical exploration applications.

关 键 词：深度学习 视电阻率 反演 卷积神经网络 数值模拟 

分 类 号：P631[天文地球—地质矿产勘探]