张志厚^1,2,,
廖晓龙¹,
曹云勇³,
侯振隆^4,5,
范祥泰¹,
徐正宣^1,3,
路润琪¹,
冯涛³,
姚禹¹,
石泽玉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. 中铁二院成都地勘岩土工程有限责任公司, 成都 610000
4. 东北大学深部金属矿山安全开采教育部重点实验室, 沈阳 110819
5. 东北大学资源与土木工程学院, 沈阳 110819

基金项目: 四川省科技厅科技计划项目(2020YFG0303，21YYJC3115，2019YFG0460，2019YFG0001)、中国中铁股份有限公司科技研究开发计划项目(CZ01-重点-05)和国家重点研发计划项目(2018YFC1505401)联合资助


详细信息

作者简介: 张志厚, 男, 1983年生, 博士, 主要从事重磁数据处理、电磁数据处理及深度学习反演研究.E-mail: logicprimer@163.com

中图分类号: P631

收稿日期:2020-08-07
修回日期:2020-12-30
上线日期:2021-04-10

Joint gravity and gravity gradient inversion based on deep learning

ZHANG ZhiHou^1,2,,
LIAO XiaoLong¹,
CAO YunYong³,
HOU ZhenLong^4,5,
FAN XiangTai¹,
XU ZhengXuan^1,3,
LU RunQi¹,
FENG Tao³,
YAO Yu¹,
SHI ZeYu¹
1. Faculty of Geosciences and Environmental Engineering, Southwest Jiaotong University, Chengdu 611756, China
2. Ministry of Education Key Laboratory of High-speed Railway Engineering, Southwest Jiaotong University, Chengdu 610031, China
3. Chengdu Geological Survey Geotechnical Engineering Co, Ltd, Chengdu 610000, China
4. Key Laboratory of Ministry of Education on Safe Mining of Deep Metal Mines, Northeastern University, Shenyang 110819, China
5. School of Resources and Civil Engineering, Northeastern University, Shenyang 110819, China


MSC: P631

--> Received Date: 07 August 2020
Revised Date: 30 December 2020
Available Online: 10 April 2021


摘要
摘要:高效高精度的反演算法在重力大数据时代背景下显得尤为重要，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深度学习的重力异常及重力梯度异常的联合反演方法.文中首先提出了一种基于网格点几何格架的重力异常及重力梯度异常的空间域快速正演算法，这为本文深度学习反演算法的实现奠定了基础；其次对大量的不同密度模型进行正演计算获得样本数据集；然后设计了一种端到端的深度学习网络结构(GraInvNet)，再利用样本数据对该网络结构进行训练；最后进行反演预测.组合模型试验表明，多维度数据联合反演相比单一分量反演其结果更"聚焦"，且与模型边界高度吻合，并且对于复杂模型的姿态与物性预测具有极为显著的优势，以及对于含噪声数据的反演，其质量也不会降低；Vinton岩丘实测重力数据也验证了文中方法的有效性；从而证明了深度学习在重力数据的高效高精度反演方面具有的巨大潜力.
关键词: 重力异常与重力梯度异常/
全卷积神经网络/
快速正演/
联合反演

Abstract:In the era of big data, high-efficient and high-precise inversion algorithms of gravity data become particularly important. Inspired by the excellent nonlinear mapping capability of deep learning, we propose a joint inversion method of gravity anomaly and gravity gradient anomaly based on deep learning. The main contents of the paper are as follow: Firstly, a fast forward algorithm in spatial domain of gravity anomaly and gravity gradient anomaly based on grid point geometric grid is put forward, which establishes a foundation to realize a new deep-learning inversion algorithm; Secondly, sample data sets are constructed by forward calculation of a great quantity models of different densities; Thirdly, an end-to-end deep learning network GraInvNet is creatively designed, which is then trained using the sample data sets; Finally, the inversion forecast is carried out. As shown by the model tests above, the results of joint inversion of multi-dimensional data are more focused than that of its counterpart using single component, and also more consistent with the model boundary. Moreover, the former possesses a significant advantage in predicting the buried depth, shape and physical properties of complex models, particularly in the inversion of noisy data. The validity of the method is further verified by the results of field data inversion using the gravity data of Vinton Salt-Rock, which, thus, effectively proves to us the great potential of deep learning in high-efficient and high-precise inversion of gravity data.
Key words:Gravity and gravity gradient anomaly/
Full convolutional neural network/
Fast forward/
Integrated inversion



PDF全文下载地址:

http://www.geophy.cn/data/article/export-pdf?id=6076842cc23e6710c26d5b3a