基于岩屑录井图像的井壁稳定性智能预测方法  

An intelligent prediction method for wellbore stability based on drilling cuttings logging images

作　　者：夏文鹤[1] 唐印东 李皋 韩玉娇 林永学[3] 吴雄军[3] XIA Wenhe;TANG Yindong;LI Gao;HAN Yujiao;LIN Yongxue;WU Xiongjun(School of Electrical Engineering and Information,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Petroleum Engineering School,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Sinopec Research Institute of Petroleum Engineering,Beijing 100101,China)

机构地区：[1]西南石油大学电气信息学院 [2]西南石油大学石油与天然气工程学院 [3]中国石化石油工程技术研究院

出　　处：《天然气工业》2023年第12期71-83,共13页Natural Gas Industry

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

摘　　要：钻井现场通常利用岩石力学的分析结果对井壁稳定性进行预测,但其时效性普遍较差。为此,利用实时的岩屑录井图像资料建立了包括16种掉块形状和岩性的图像样本库,并以深度学习网络的高效特征提取技术为基础,建立了一种基于掉块图像特征的井壁失稳类型分析模型,针对钻井返出砂样图像中的掉块图像进行形状和岩性识别,以判定钻进地层和井壁失稳的类型。研究结果表明:①使用ShuffleNetV2网络作为智能系统基础架构,在单元模块中引入了XConv卷积核并行分支和SimAM注意力机制模块,强化了网络对掉块图像标志性特征信息的关注度;②对ShuffleNetV2网络中的Stage 2、Stage 3和Stage 4进行了多通道特征融合算法的设计,保留了掉块轮廓关键特征,最终改进的ShuffleNetV2网络模型对掉块形状及岩性的识别准确率为90.56%。结论认为,现场应用的效果验证了该方法的可靠性,从砂样图像输入到结果输出用时低于1 s,识别结果与地质资料以及施工过程的工况基本吻合,该方法能满足现场对井壁稳定状况快速感知的现实需求。On drilling sites,the analysis results of rock mechanics ae usually applied to predict wellbore stability,but this method generally has low time efficiency.In this paper,an image sample library including 16 types of falling block shape and lithology is established using the real-time drilling cuttings logging image data.A wellbore instability type analysis model based on features of falling block images is established on the basis of the efficient feature extraction technology of deep leaning network.In addition,the falling block images in the images of the cuttings returned while drilling are analyzed for shape and lithology identification,so as to determine the types of encountered strata and wellbore instability.And the following research results are obtained.First,ShuffleNetV2 network is used as the intelligent system infrastructure,and XConv convolutional kernel parallel branch and SimAM attention mechanism module are introduced into the unit module,which enhances the network's attention to the landmark feature information of falling block images.Second,a multi-channel feature fusion algorithm is adopted in the design of Stage 2,Stage 3 and Stage 4 of the ShuffleNetV2 network,which retains the key profile features of the falling block.As a result,the improved ShuffleNetV2 network model achieves an accuracy of 90.56%in identifying the shape and lithology of falling block.In conclusion,the on-site application results have verified the reliability of this method.The time from the input of returned cuttings images to the output of results is less than 1 second,and the recognition results are basically consistent with geological data and construction process conditions.This fully demonstrates that this method can meet the actual needs of rapid perception of wellbore stability on site.

关 键 词：岩屑录井图像 轻量化网络 单元结构 SimAm注意力机制 多通道特征融合 井壁稳定性 

分 类 号：TE142[石油与天然气工程—油气勘探]