基于深度神经网络的断层高分辨率识别方法  

Fault High-Resolution Recognition Method Based on Deep Neural Network

作　　者：丰超 潘建国[1] 李闯[1] 姚清洲[1] 刘军[1] Feng Chao;Pan Jianguo;Li Chuang;Yao Qingzhou;Liu jun(Research Institute of Petroleum Exploration&Development-Northwest,PetroChina,Lanzhou 730020,China)

机构地区：[1]中国石油勘探开发研究院西北分院,甘肃兰州730020

出　　处：《地球科学》2023年第8期3044-3052,共9页Earth Science

基　　金：中国石油天然气股份有限公司十四五上游领域前瞻性基础性项目《海相碳酸盐岩成藏理论与勘探技术研究》(No.2021DJ05)。

摘　　要：传统地震属性断层识别技术多基于数据不连续性识别断裂,干扰因素多,越来越难以满足深层精细勘探的需求.为了提高断层识别精度,提出一种断层高分辨率智能识别方法,在深度学习方法从地震数据预测断层属性的基础之上,建立高分辨率与低分辨率断层标签库,训练深度神经网络,获得高分辨率检测模型.通过模型与实际数据证实,方法解决了深度学习中卷积神经网络存在上采样造成高频损失,使断层分辨率有所下降的问题,提高了分辨能力,模拟数据均方根误差下降40.02%.方法不仅相对传统算法更加准确地检测了断层特征,而且比一般的深度学习断层识别分辨率高.Fine fault identification is of great significance to improve the efficiency of exploration and development.Traditional seismic attribute fault recognition technologies identify fractures based on data discontinuity,and there are many interfering factors,making it more and more difficult to meet the needs of fine exploration in deep area.In order to improve the fault identification accuracy,this paper proposes a high-resolution intelligent identification method of fault.Based on the deep learning method to predict fault attributes from seismic data,a high-resolution and low-resolution fault label library is established,and a deep neural network is trained.It is confirmed by the model and actual data that the method solves the problem of high-frequency loss caused by up sampling in the convolutional neural network in deep learning,which reduces the resolution of faults,and improves the resolution ability.The root mean square error of the simulated data decreased by 40.02%.Compared with traditional algorithms,the method not only detects fault features more accurately,but also has a higher resolution than common deep learning fault recognition.

关 键 词：卷积神经网络 Unet网络 断层识别 深度学习 

分 类 号：P631[天文地球—地质矿产勘探]