地球物理学进展 ›› 2019, Vol. 34 ›› Issue (6): 2256-2261.doi: 10.6038/pg2019CC0367

• 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3D半密度卷积神经网络的断裂检测

段艳廷1, 2, 郑晓东2, *, 胡莲莲2, 吴朝东1   

  1. 1. 北京大学地球与空间科学学院,北京 100871;
    2. 中国石油勘探开发研究院,北京 100083
  • 收稿日期:2019-01-30 修回日期:2019-08-23 出版日期:2019-12-20 发布日期:2019-12-29
  • 通讯作者: 郑晓东,男,教授级高级工程师,主要从事地震解释和储集层预测方面的研究工作. (E-mail: zxd@petrochina.com.cn)
  • 作者简介:段艳廷,男,1990年生,博士研究生,主要从事地震解释和储层预测方法研究工作. (E-mail: ytduan@pku.edu.cn)
  • 基金资助:
    国家重点研发计划重点专项(2016YFC0601107)和中国石油集团科学技术研究院有限公司超前技术研究项目(2017ycq11)联合资助.

Fault detection based on 3D semi-dense convolutional neural network

DUAN Yan-ting1, 2, ZHENG Xiao-dong2, *, HU Lian-lian2, WU Chao-dong1   

  1. 1. School of Earth and Space Sciences, Peking University, Beijing 100871, China;
    2. Research Institute of Petroleum Exploration and Development, PetroChina, Beijing 100083, China
  • Received:2019-01-30 Revised:2019-08-23 Online:2019-12-20 Published:2019-12-29

摘要: 基于传统相干属性的断层检测方法易受地层倾角的影响,为了提高断层检测的精度,本文提出了一种基于改进的半密度卷积网络的断裂检测方法.在密度卷积网络模型基础上,去除了池化层,并将之前所有的卷积层与第一个全连接层连接,然后传给下一个全连接层,改进后的网络结构模型既兼顾了不同尺度的信息,又保持了空间分辨率.实际地震数据测试表明,改进的方法可以在弱监督标签条件下实现高精度的断层检测,且断裂检测模型具有一定的迁移能力.

关键词: 深度学习, 3D半密度卷积网络, 断裂检测, 弱标签

Abstract: The fault detection methods based on traditional coherence attributes are sensitive to the dipping structures. To improve the accuracy of fault detection, a fault detection method based on semi-dense convolutional neural network is proposed in this paper. In the basic structure of dense convolutional neural network, the improved semi-dense convolutional neural network removes the pooling layer, and connect all convolutional layers directly to the end of network, which can take into consideration different scales of information and keep the spatial resolution. The improved semi-dense convolutional neural network can be divided into two parts: the training stage and the testing stage. In the training stage, the subvolume is randomly selected from the seismic data set as the training sample. Then, the training sample and the corresponding label are used to train the network, so that it has strong robustness and generalization ability. The testing stage is mainly for fault detection of unlabeled seismic data set. So we apply our method to two set of field data including the3D seismic data set in the Songliao basin, Northeastern China and the 3D seismic data set in the North Tarim Basin, Northwestern China. The test of actual seismic data shows that the improved method can accurately predict faults under the weak supervised label conditions, and the fault detection model has a part of transfer learning ability.

Key words: Deep learning, 3D semi-dense convolutional neural network, Fault detection, Weak labels

中图分类号: