基于深度学习的页岩微观结构智能分割与定量表征方法——以北部湾盆地流沙港组二段为例  

Intelligent segmentation and quantitative characterization of shale microstructure based on deep learning:A case study of second member of Liushagang Formation in Beibu Gulf Basin

作　　者：刘成 万欢[1,2,3] 马立涛 唐鑫[4] 龚胜利 史长林[1,2,3] 李洋冰 陈相如 黄英 白瑞婷[1,2,3] 崔书姮 LIU Cheng;WAN Huan;MA Litao;TANG Xin;GONG Shengli;SHI Changlin;LI Yangbing;CHEN Xiangru;HUANG Ying;BAI Ruiting;CUI Shuheng(National Engineering Research Center of Offshore Oil and Gas Exploration,Beijing 100029,China;CNOOC Energy Technology&Services Limited,Engineering Technology Branch,Tianjin 300452,China;CNOOC Energy Technology&Services Limited,Key Laboratory for Exploration&Development of Unconventional Resources,Tianjin 300457,China;Chongqing Three Gorges University,Chongqing 404100,China)

机构地区：[1]海洋油气勘探国家工程研究中心,北京100029 [2]中海油能源发展股份有限公司工程技术分公司,天津300452 [3]中海油能源发展股份有限公司非常规勘探开发重点实验室,天津300457 [4]重庆三峡学院,重庆404100

出　　处：《中国海上油气》2024年第6期232-243,共12页China Offshore Oil and Gas

基　　金：海洋油气勘探国家工程研究中心2022年度开放基金课题“油页岩微观孔隙结构表征及原油可动性研究(编号:CCL2022RCPS0796 RQN)”部分研究成果。

摘　　要：页岩微观结构智能表征方法研究是当前的研究热点之一,也是页岩储集空间定量表征的重要手段,对页岩油勘探开发有重要的意义。以中国北部湾盆地流沙港组二段低成熟度页岩为研究对象,基于深度学习理论开发的Shale-net深度学习模型以及CT、FIB-SEM等物理实验数据,印证了模型的可靠性及图像处理的有效性。结果表明,该深度学习模型预测区域与真实区域平均交并比值(mIoU)最大可达到0.8998,即准确率可达89.98%;Shale-net深度学习模型对孔裂隙、有机质、黏土矿物、石英以及黄铁矿的分割效果较好,mIoU最大值分别为0.9438、0.9529、0.8592、0.8446、0.9800,其对孔裂隙、有机质以及黄铁矿的分割准确率均超过了90%,对特征复杂的黏土矿物及石英的分割准确率也达到85%左右;此外,从分割效果可以看出Shale-net深度学习模型能更加准确地分割不同的物质,证明Shale-net深度学习模型在页岩微观结构智能表征方面比传统的阈值分割与分水岭法效果更好;不同方法定量表征的结果表明Shale-net深度学习模型的定量结果与数字岩心更接近,证明Shale-net深度学习模型具有更优的油页岩储层结构及矿物识别能力,该模型可作为页岩微观结构智能分割与定量表征的有效手段。The research on the intelligent characterization method of shale microstructure is one of the current research hotspots,and it is an important means for the spatial quantitative characterization of shale reservoirs and is of great significance for shale oil exploration and development.In this study,the low-maturity shale of the second member of the Liushagang Formation in the Beibu Gulf Basin of China was taken as the research object.The Shale-net deep learning model developed based on the deep learning theory and the physical experiment data such as CT and FIB-SEM confirmed the reliability of the model and the effectiveness of image processing.The results show that the mean intersection over union(mIoU)between the predicted region and the real region of the deep learning model can reach a maximum of 0.8998,or in other words,the accuracy can reach 89.98%.The Shale-net deep learning model has a good segmentation effect on pore fractures,organic matter,clay minerals,quartz,and pyrite.The maximum values of mIoU are 0.9438,0.9529,0.8592,0.8446,and 0.9800,respectively,and the segmentation accuracy of pore fractures,organic matter,and pyrite exceeds 90%.The segmentation accuracy of clay minerals and quartz with complex characteristics also reaches about 85%.In addition,it can be seen from the segmentation effect that the Shale-net deep learning model can segment different substances more accurately,which proves that the Shale-net deep learning model has better performance than the traditional threshold segmentation and watershed methods in the intelligent characterization of shale microstructure.The results of quantitative characterization by different methods show that the quantitative results of the Shale-net deep learning model are closer to those of digital cores.It is proven that the Shale-net deep learning model has a better identification ability for oil shale reservoir structures and minerals.This model can be used as an effective means for intelligent segmentation and quantitative characterization of shale mic

关 键 词：北部湾盆地 深度学习 语义分割 智能表征 数字岩心 

分 类 号：TE132.14[石油与天然气工程—油气勘探]