西南石油大学学报(自然科学版) ›› 2020, Vol. 42 ›› Issue (6): 16-25.DOI: 10.11885/j.issn.1674-5086.2020.06.02.02

• 油气田人工智能技术与应用专刊 • 上一篇    下一篇

基于深度学习算法不同数据集的地震反演实验

黄旭日, 代月, 徐云贵, 唐静   

  1. 西南石油大学地球科学与技术学院, 四川 成都 610500
  • 收稿日期:2020-06-02 发布日期:2020-12-21
  • 通讯作者: 黄旭日,E-mail:xrhuang@sunrisepst.com
  • 作者简介:黄旭日,1965年生,男,汉族,广西灵川人,教授,博士生导师,主要从事油藏地质、地球物理等多专业综合技术的研究工作。E-mail:xrhuang@sunrisepst.com;代月,1997年生,女,汉族,四川内江人,硕士研究生,主要从事地震智能油藏预测等方面的研究工作。E-mail:dy_swpu@163.com;徐云贵,1977年生,男,汉族,湖北武汉人,教授,博士生导师,主要从事各向异性岩石物理和地震响应、三维四维地震属性反演和储层描述等方面的研究工作。E-mail:yungui.xu@fox-mail.com;唐静,1988年生,女,汉族,重庆开县人,讲师,主要从事地震反演方面的研究工作。E-mail:tjgucas@163.com

Seismic Inversion Experiments Based on Deep Learning Algorithm Using Different Datasets

HUANG Xuri, DAI Yue, XU Yungui, TANG Jing   

  1. School of Geosciences and Technology, Southwest Petroleum University, Chengdu, Sichuan 610500, China
  • Received:2020-06-02 Published:2020-12-21

摘要: 近年来,人工智能中的深度学习技术在地震数据处理、反演和解释领域显示出许多优势。以往的研究表明,深度学习与地震反演相结合的方法比传统方法更有效。利用深度学习技术有可能得到更高分辨率的结果,这对油藏开发至关重要。通过设计地质模型进行采样以获取不同大小数据集,基于卷积神经网络(CNN)研究了不同训练数据集的地震反演应用效果,实验表明,该神经网络的预测精度在一定范围内随训练集的增加而增加,得到了对神经网络模型构建的关键数据集大小占全数据集的比例。此外,通过对地震数据加入不同比例的噪声并对CNN进行训练,结果表明本文所设计的CNN具有良好的抗噪和泛化能力。

关键词: 人工智能, 深度学习, 卷积神经网络, 地震反演, 地质建模

Abstract: Recently deep learning of artificial intelligence demonstrates certain advantages for seismic processing, interpretation and inversion. Previous studies show that the combination of deep learning and seismic inversion could generate more robust results than traditional methods. The deep learning technique could achieve results of high resolution which is critical for reservoir development. This paper investigates the effects of different training datasets used in seismic inversion based on Convolutional Neural Network (CNN) by designing a reservoir model and its corresponding seismic response. The result shows that the prediction accuracy of this Neural Network increases with the increase of training datasets size in a certain range. The relationship of inversion quality and ratio between the entire datasets and the training data is demonstrated. In addition, different levels of noises in seismic are tested for CNN training. The results demonstrate the generalization and anti-noise ability of the designed CNN.

Key words: artificial intelligence, deep learning, convolutional neural network, seismic inversion, geological modeling

中图分类号: