基于深度信念网络的多采样点岩性识别  

Lithology recognition of multi-sampling points based on deep belief network

作　　者：李国和[1,2,3] 郑阳 李莹 吴卫江 洪云峰[3] 周晓明 LI Guo-he;ZHENG Yang;LI Ying;WU Wei-jiang;HONG Yun-feng;ZHOU Xiao-ming(Beijing Key Lab of Data Mining for Petroleum Data,China University of Petroleum,Beijing 102249,China;College of Geophysics and Information Engineering,China University of Petroleum,Beijing 102249,China;PanPass Institute of Digital Identification Management and lnternet of Things,Beijing 100029,China)

机构地区：[1]中国石油大学(北京)石油数据挖掘北京市重点实验室,北京102249 [2]中国石油大学(北京)地球物理与信息工程学院,北京102249 [3]石大兆信数字身份管理与物联网技术研究院,北京100029

出　　处：《地球物理学进展》2018年第4期1660-1665,共6页Progress in Geophysics

基　　金：国家自然科学基金(60473125;61701213);油气国家重点专项子课题(G-5800-08-ZS-WX);中国石油(CNPC)石油科技中青年创新基金(05E7013);中国石油大学(北京)克拉玛依校区科研启动基金(RCYJ2016B-03-001);福建省教育厅中青年一般项目(JA15300;JAT170349);产学合作协同育人项目(201702098015)联合资助

摘　　要：岩性识别是储层预测中的一个重要环节.一方面,传统的机器学习算法缺少特征自动提取的过程,且不能有效利用地震数据局部特征预测储层;另一方面,采用单一采样点作为输入,缺失相邻数据关联关系反映层位信息.针对此不足,本文以多个相邻采样点的地震数据作为输入和测井岩性数据作为输出,利用受限玻尔兹曼机(RBM)对多采样点地震数据进行特征提取,逐层堆叠受限玻尔兹曼机(RBM)构建深度信念网络(DBN),并采用随机梯度下降算法对误差进行反向传递学习,最终构建岩性识别模型.以多点地震数据为输入,利用该模型实现地层岩性识别.通过多种智能建模方法实验对比,证实了多个采样点作为输入,隐含利用了部分地层信息,有效地提高了岩性识别的精度.Lithology recognition is an important part of reservoir prediction. On one hand the traditional machine learning algorithm lacks the process of automatic feature extraction, which cannot effectively utilize the local features of seismic data for the rock formation recognition,on the other the adoption of single point sampling as input loses the stratum information presented by the relation of multi-point seismic data. In order to resolve this drawback,the input of the model is extended from single-point sampling to multi-point sampling and the output of model is welllogging lithologic data. Restricted Boltzmann Machines（ RBMs） are used to extract the features of seismic data from multi-sampling points,and then Deep Belief Networks（ DBNs） are built up by layer-by-layer stacking RBMs to construct lithologic recognition model, learning by means of the stochastic gradient descent algorithm for the error backward propagation. The models are used to recognize the stratum lithology with the multi-point seismic data as input. By the comparison the models with other intelligent models for lithologic recognition,it is proved more efficient by experiments that the multiple sampling points of seismic data as inputs implicitly use the stratum information partly to improve the accuracy of lithologic recognition.

关 键 词：受限玻尔兹曼 深度信念网络 岩性识别 

分 类 号：P631[天文地球—地质矿产勘探]