融合SMGC-ECAs-Resnet的致密砂岩岩相识别方法研究  

Study on lithofacies identification method of tight sandstonewith fusion SMGC-ECAs-Resnet

作　　者：田枫[1] 王鑫[1] 刘芳[1] 刘宗堡[2] 刘涛 唐莎莎 刘悦 张世祺 TIAN Feng;WANG Xin;LIU Fang;LIU Zongbao;LIU Tao;TANG Shasha;LIU Yue;ZHANG Shiqi(College of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China;School of Earth Sciences,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318 [2]东北石油大学地球科学学院,黑龙江大庆163318

出　　处：《重庆理工大学学报（自然科学）》2024年第8期164-172,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金面上项目(42172161);黑龙江省省属本科高校基本科研业务费项目(2022TSTD-03);黑龙江省自然科学基金项目(LH2021F004);黑龙江省哲学社会科学研究规划年度项目(22EDE389)。

摘　　要：针对目前深度学习模型通常只能提取单一尺度岩相特征,无法获得多尺度信息且没有充分适应测井曲线自身形态特点影响岩相识别的问题,基于深度学习以Resnet50为基础网络,设计开发多尺度特征提取模块SMGC(strip-pooling and multi-scale group convolution),并加入改进的ECAs(efficient channel attention strengthen)注意力模块增强对测井曲线条形纹理信息关注度,提出一种SMGC-ECAs-Resnet致密砂岩测井曲线岩相识别方法。以松辽盆地三肇凹陷扶余油层为例,对测井曲线数据预处理构建图像数据集,利用SMGC-ECAs-Resnet网络模型对其进行识别得到分类结果,设置对比试验和鲁棒性实验证明模型有效性。结果表明:所提出的SMGC-ECAs-Resnet网络岩相识别准确率达到91.9%,为复杂致密砂岩岩相的测井识别提供了新的方法。The task of lithofacies identification and division is important in the exploration and evaluation of tight reservoirs.Currently,the deep learning model usually only extracts single-scale lithofacies features,which fails to obtain multi-scale information and does not fully adapt to the influence of the morphological characteristics of logging curves on lithofacies identification.Based on deep learning and Resnet50-based network,a multi-scale feature extraction module SMGC(Strip-pooling and Multi-scale Group Convolution)is designed and developed.An improved ECAs(Efficient Channel Attention Strengthen)attention module is added to propose a lithofacies identification method of SMGC-ECAs-Resnet tight sandstone logging curve.Fuyu oil layer in Sanzhao sag of Songliao Basin is taken as an example.First,the image data set is built by preprocessing the logging curve data.Then,the SMGC-ECAs-Resnet network model is employed to identify and obtain the classification results.Finally,the validity of the model is verified by comparative and robustness experiments.Our experiments show the proposed SMGC-ECAs-Resnet network reaches the optimal 91.9%in lithofacies recognition accuracy demonstrating its effectiveness in logging identification of complex tight sandstone lithofacies.

关 键 词：深度学习 多尺度 注意力机制 致密砂岩 岩相识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]