基于生成对抗神经网络的岩性识别方法与应用  

Lithology Identification Method and Application Based on Generative Adversarial Neural Network

作　　者：尹琼 YIN Qiong(Faculty of Metallurgy and Mining,Kunming Metallurgical College,Kunming,Yunnan 650033,China)

机构地区：[1]昆明冶金高等专科学校冶金与矿业学院,云南昆明650033

出　　处：《测井技术》2025年第1期57-67,共11页Well Logging Technology

基　　金：国家自然科学基金项目“悬臂式抗滑桩嫁接微型群桩组合加固滑坡的承载机理研究”(42267020)。

摘　　要：岩性识别是储层评价的基础,是储层参数计算和油藏评价开发的关键。测井资料中包含大量的地层信息,是岩性识别的基础资料。然而,测井资料在解释岩性时会受到多种因素的影响,导致结果存在多解性,采用传统方法进行测井岩性识别时,其精度常常难以满足要求。利用机器学习法在数据分析和建模方面的优势,采用两步法的岩性识别策略和生成对抗神经网络进行类别不均衡的测井岩性识别。以阜康凹陷二叠系和三叠系的砂岩储层测井资料为基础,通过测井曲线的相关性和测井岩石物理分析,筛选出声波时差、井径、中子、密度、自然伽马、地层电阻率和自然电位7种测井曲线为特征输入,识别泥岩、砂质泥岩、细砂岩、中砂岩、砂砾岩和砾岩6种岩性,取得了较好的识别效果。根据对比试验发现,两步法岩性识别比单步法岩性识别准确率提高4.21%。此外,两步法生成对抗神经网络模型的岩性识别准确率高于随机森林、支持向量机、分布式梯度增强库和长短期记忆网络模型4.72%~7.19%,模型的整体识别准确率达到83.44%,该方法在岩性识别领域中有较好的发展前景。Lithology identification is the basis of reservoir evaluation and the key to reservoir parameter calculation and reservoir evaluation and development.Well logging data contains a lot of formation information,which is the basic data for lithology identification.However,the lithology interpretation of logging data is affected by many factors,which results in multiple solutions.The accuracy of lithology identification using traditional methods is often difficult to meet the requirements.Based on the advantages of machine learning in data analysis and modeling,two-step lithology identification strategy and generative adversarial neural network are used to identify the lithology of logging with unbalanced categories.Based on the logging data of Permian and Triassic sandstone reservoirs in Fukang depression,seven logging curves,including acoustic time difference,borehole diameter,neutron,density,natural gamma ray,formation resistivity and spontaneous potential,are selected as characteristic inputs through correlation of logging curves and petrophysical analysis.Six lithologies,namely mudstone,sandy mudstone,fine sandstone,medium sandstone,sand conglomerate and conglomerate,are identified.A good recognition effect has been obtained.According to the comparative test,the accuracy of two-step lithology identification is 4.21%higher than that of single step lithology identification.In addition,the accuracy of the two-step adjunct network model is 4.72%~7.19%higher than that of random forest,support vector machine,distributed gradient enhanced library and long short-term memory network model,and the overall recognition accuracy of the model reaches 83.44%.This method has a better development prospect in the field of lithology identification.

关 键 词：岩性识别 机器学习 测井数据 生成对抗神经网络 样本不均衡 阜康凹陷 

分 类 号：P631.84[天文地球—地质矿产勘探]