地球物理学进展 ›› 2021, Vol. 36 ›› Issue (1): 325-337.doi: 10.6038/pg2021EE0222

• 应用地球物理学Ⅰ(油气及金属矿产地球物理勘探) • 上一篇    下一篇

基于全卷积神经网络的磁异常及磁梯度异常反演

张志厚1,2(), 路润琪1, 廖晓龙1, 徐正宣1,3, 乔中坤4, 范祥泰1, 姚禹1, 石泽玉1, 刘鹏飞1, 陆三明5   

  1. 1.西南交通大学地球科学与环境工程学院,成都 611756
    2.西南交通大学高速铁路线路工程教育部重点实验室,成都 610031
    3.中铁二院地勘岩土工程有限责任公司,成都 610000
    4.吉林大学地球探测科学与技术学院,长春 130021
    5.安徽省公益性地质调查管理中心,合肥 230001
  • 收稿日期:2020-06-28 修回日期:2020-11-17 出版日期:2021-02-20 发布日期:2021-03-11
  • 作者简介:张志厚,男,1983年生,博士,副教授,主要从事地球物理研究. E-mail: logicprimer@163.com
  • 基金资助:
    四川省科技厅科技计划项目(2019YFG0460);四川省科技厅科技计划项目(2019YFG0001);四川省科技厅科技计划项目(2020YFG0303);四川省科技厅科技计划项目(2021YJ0031);中国中铁股份有限公司科技研究开发计划项目(CZ01-重点-05);国家重点研发计划项目(2018YFC1505401)

Inversion of magnetic anomaly and magnetic gradient anomaly based on fully convolution network

ZHANG ZhiHou1,2(), LU RunQi1, LIAO XiaoLong1, XU ZhengXuan1,3, QIAO ZhongKun4, FAN XiangTai1, YAO Yu1, SHI ZeYu1, LIU PengFei1, LU SanMing5   

  1. 1. Faculty of Geosciences and Environmental Engineering, Southwest Jiaotong University, Chengdu 611756, China
    2. MOE Key Laboratory of High-Speed Railway Engineering, Southwest Jiaotong University, Chengdu 610031, China
    3. China Railway Eryuan Geotechnical Engineering Co.,Ltd., Chengdu 610000, China
    4. College of Geo-Exploration Sciences and Technology,Jilin University, Changchun 130021, China
    5. Public Geological Survey Management Center of Anhui Province, Hefei 230001 China
  • Received:2020-06-28 Revised:2020-11-17 Online:2021-02-20 Published:2021-03-11

摘要:

地球物理反演问题具有病态性、不适定性,传统的线性反演方法面临着次优逼近和初始模型选择等挑战,为了提高磁场数据反演的精度,受深度学习卓越的非线性映射能力的启发,本文提出了一种基于全卷积神经网络的磁异常及磁梯度异常反演方法.文中首先提出了一种基于网格点几何格架的磁异常及磁梯度异常的空间域快速正演算法,这为本文全卷积神经网络反演算法的实现奠定了基础;随后对大量不同剩余磁化强度模型进行正演计算获得样本数据集,将正演数据作为输入层,磁化强度模型作为输出层,并基于U-net网络结构设计了一种端到端的网络结构(MagInvNet),再对该网络结构进行监督学习与参数优化;最后进行反演预测.三组模型试验表明,MagInvNet网络能够快速准确识别出磁异常体的位置与形状,并且能够准确的反演出异常体磁化强度的大小,对于含噪声数据,其反演结果的质量不会降低.本文最后利用安徽霍邱铁矿的航磁数据验证了文中方法的有效性.

关键词: 全卷积神经网络, 端到端, 快速正演与反演, 磁异常及磁梯度异常

Abstract:

There exist ill-condition and ill-posed properties in the inversion problem of geophysics. The traditional linear inversion methods are faced with the challenges of suboptimal approximation and initial model selection. In order to improve the accuracy of magnetic field data inversion, inspired by the excellent nonlinear mapping ability of deep learning, an inversion method of magnetic anomaly and magnetic gradient anomaly based on fully convolutional network is proposed in this paper. This paper proceeds as follow: Firstly, a fast forward algorithm of magnetic anomaly and magnetic gradient anomaly in spatial domain based on grid point geometric trellis is presented, which lays a foundation for the implementation of fully convolution neural network inversion algorithm. Then, sample data set is obtained by forward modeling to an ocean of models with different remanent magnetization; Next, an end-to-end network structure (MagInvNet) based on the basic framework of U-net is designed, where the input layer is forward data and the output layer is remanent magnetization model; and then the network structure is trained and optimized. Finally, the inversion prediction is carried out. As three groups of model tests show, MagInvNet network can not only quickly and accurately identify the position and shape of the magnetic anomaly body, but also accurately invert the magnetization of the anomalous body. As for data with noise, the quality of inversion result will not be influenced. In the end, the aeromagnetic data of Huoqiu Iron Mine in Anhui province is used to verify the effectiveness of the method.

Key words: Fully Convolutional Network (FCN), End-to-end, Fast forward and inversion, Magnetic anomaly and magnetic gradient anomaly

中图分类号: