摘要
岩石矿物组分含量是地球物理勘探开发中的重点关注对象.在岩心与地层元素测井资料较少的情况下,如何提高矿物组分含量参数的预测精度显得尤为关键.本文采用深度学习方法,利用常规测井曲线对来自于地层元素测井获得的矿物组分含量进行预测.首先基于残差网络(ResNet)框架,利用一维卷积核和池化核构建了卷积神经网络模型.模型采用自然伽马、自然电位、井径、阵列感应电阻率、三孔隙度以及光电吸收截面指数测井参数作为输入,地层元素测井获得的矿物组分含量作为输出.随后对所搭建卷积神经网络进行了训练,建立了输入与输出之间的实际映射关系.最后,利用测试数据集和真实地层资料,对所建立的卷积神经网络进行了精度检验,并与人工神经网络和多元线性回归的评价结果进行了比较.结果显示,卷积神经网络在测试数据集上的总体预测数值相关性为0.90,明显优于人工神经网络的0.68与多元线性回归的0.51.通过处理实际测井资料,进一步验证了该方法的预测优越性和鲁棒性,以及其在地层参数评价方向的良好应用前景.
The composition contents of various minerals in the rock is a key concern in geophysical exploration and development. In general, accurate quantitative evaluation of rock mineral composition content can be performed by core analysis and formation element logging data. However, in practical applications, it is often challenging to obtain sufficient scale of the above data for field-wide mineral composition content evaluation. Therefore, it is especially critical to improve the prediction accuracy of mineral composition content with the support of other basic data. In this paper, the deep learning technique is used to predict the mineral composition content obtained from the formation element logging by using the conventional logging data. Firstly, based on the ResNet network architecture, a convolutional neural network model is constructed using one-dimensional convolution kernels and pooling kernels. The logging parameters of gamma, spontaneous potential, caliper curves, array induction resistivity, three porosity logging data and photoelectric absorption cross-section index are considered as input data of the model, and the mineral composition content obtained from formation element logging is used as output. We train the constructed convolutional neural network based on logging data from the Ordos Basin YanChang Formation and establish the mapping relationship between input and output. Finally, using the test data set and real stratigraphic data, the accuracy of the established convolutional neural network is verified and compared with the evaluation results of artificial neural network and multiple linear regression. The results demonstrate that the correlation between the predicted and actual results of the convolutional neural network on the test dataset is 0. 90,which is significantly better than 0. 68for the artificial neural network and 0. 51 for the multiple linear regression. By processing the actual logging data,the prediction superiority and robustness of the method are further verified,as well as its good application prospect in formation parameter evaluation.
作者
王玥天
毛志强
胡琮
李高仁
何伟
马凤情
WANG YueTian;MAO ZhiQiang;HU Cong;LI GaoRen;HE Wei;MA FengQing(State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum(Beijing),Beijing 102249,China;Research Institute of Exploration and Development,PetroChina Changqing Oilfield Company,Xi'an 710018,China)
出处
《地球物理学进展》
CSCD
北大核心
2023年第2期748-758,共11页
Progress in Geophysics
基金
国家科技重大专项(2016ZX05050-008)
中国石油长庆油田公司科学研究与技术开发项目(ZY20-XA407-TPFW713)联合资助。
关键词
深度学习
测井评价
岩石矿物组分含量
卷积神经网络
Deep learning
Logging evaluation
Rock mineral composition
Convolutional neural network
