电阻率测井成像图井壁裂缝智能识别与分割方法  

Intelligent identification and segmentation method of wellbore fractures in resistivity logging imaging map

作　　者：夏文鹤[1] 朱喆昊 韩玉娇 杨燚恺 林永学[2] 吴雄军[2] XIA Wenhe;ZHU Zhehao;HAN Yujiao;YANG Yikai;LIN Yongxue;WU Xiongjun(School of Electrical Engineering and Information,Southwest Petroleum University,Chengdu,Sichuan 610500,China;SINOPEC Petroleum Engineering Technology Research Institute,Beijing 102200,China)

机构地区：[1]西南石油大学电气信息学院,四川成都610500 [2]中国石化石油工程技术研究院,北京102200

出　　处：《石油地球物理勘探》2023年第5期1042-1052,共11页Oil Geophysical Prospecting

基　　金：国家重点研发项目“井筒稳定性闭环响应机制与智能调控方法”(2019YFA0708303);中国石油—西南石油大学创新联合体项目“深井复杂地层钻井方式优选及提速工艺技术”(2020CX040103)联合资助。

摘　　要：目前测井图像裂缝识别、处理过程工作量巨大,人工识别主观性强、稳定性差,为此,提出将计算机视觉技术和深度学习框架引入测井成像图分析解读领域,构建新型裂缝形态智能识别网络模型,实现了电阻率测井成像图中井壁裂缝区域的智能识别与分割标注。首先,通过多尺度空洞卷积结合注意力机制提取电阻率测井井壁成像图中浅层和深层特征,并将深、浅层特征进行多尺度融合,形成更具表征能力的新特征。然后,根据该特征进行像素点二分类,完成每个像素点的前景、背景类型识别,若干个前景分类的像素点对应裂缝区域的轮廓。多尺度特征融合模型从微观角度充分保留了裂缝区域图像轮廓细节,裂缝区域关联像素点识别分类准确率接近80%。最后,进一步借鉴人眼视觉相似度评价体系,从宏观角度设计裂缝轮廓智能识别性能评价算法。评价结果表明,当视觉相似度感受评级为Ⅱ级时,训练集和测试集图像中与人工识别结果基本一致的裂缝区域分别达到81.3%和80.0%,说明所提方法可替代人工解释完成裂缝的识别和标注工作,能大幅减少图像分析工作量,细致勾勒出裂缝区域轮廓线。同时,有利于及时、迅速地判断井筒、井壁稳定性,为后续裂缝区域的智能定量评价、计算提供技术支撑。In view of the huge workload,strong subjectivity in artificial identification,and poor stability in the fracture identification and processing of logging images,this paper introduces the computer vision technology and deep learning framework into the analysis and interpretation of logging images,builds an intelligent identification and segmentation network model of fracture morphology,and intelligently identifies wellbore fractures in resistivity logging images.First,the model extracts the shallow and deep features of wellbore images through multi⁃scale dilated convolution and attention mechanism,and multi⁃scale fusion of shallow and deep features is conducted to form new features with more representation ability.According to the new features,the two pixel classification is carried out to complete foreground and background type identification of each pixel in the logging images.Several pixels classified as foreground present the contour of the fractured area.The multi⁃scale feature fusion model can fully retain more contour details of the fracture image from the micro perspective,and the identification and classification accuracy of each fracture pixel reaches almost 80.0%.Finally,by drawing lessons from the evaluation system of human eye visual similarity,a performance evaluation algorithm is designed for intelligently identifying fracture contour from the macro perspective.The evaluation results show that when the visual similarity perception rating is grade II,81.3%and 80.0%of identification results in the fracture region in the training set and test set images are basically consistent with the artificial identification results.The results indicate that the proposed method can replace artificial interpretation to complete fracture identification and marking,greatly reduce the image analysis workload and carefully outline the fracture contour.Meanwhile,it is conducive to the rapid and timely judgment of wellbore and shaft stability,thus providing technical support for subsequent intelligent quantitative e

关 键 词：电阻率成像测井 井壁裂缝智能识别与分割 计算机视觉 多尺度特征融合 注意力机制 

分 类 号：P631[天文地球—地质矿产勘探]