小样本卷积神经网络井震映射反演  

Mapping Inversion of Logging and Seismic Data Based on Convolutional Neural Networks with Small Samples

作　　者：安振芳 张进[1,2] 张建中 邢磊[1,2] 黄忠来 AN Zhenfang;ZHANG Jin;ZHANG Jianzhong;XING Lei;HUANG Zhonglai(Key Laboratory of Ministry of Education for Submarine Geosciences and Prospecting Techniques,Ocean University of China,Qingdao,Shandong 266100,China;Marine Mineral Resources Evaluation and Prospecting Technology Laboratory,Qingdao National Laboratory for Marine Science and Technology,Qingdao,Shandong 266071,China)

机构地区：[1]中国海洋大学海底科学与探测技术教育部重点实验室,山东青岛266100 [2]青岛海洋科学与技术国家实验室,海洋矿产资源评价与探测技术实验室,山东青岛266071

出　　处：《西安石油大学学报（自然科学版）》2020年第4期30-38,共9页Journal of Xi’an Shiyou University（Natural Science Edition）

基　　金：国家重点研发计划项目(2017YFC0307401)。

摘　　要：针对常规卷积神经网络(Convolutional Neural Networks,CNN)在井少地区因无法获得大量测井数据而易于出现过拟合现象的问题,提出了一种小样本CNN井震映射反演方法。通过网络结构优化设计,选出了最佳的网络层数、卷积核大小、特征图规模和激活函数,并将优选出的最好网络模型应用于实际资料反演。实际应用表明,小样本CNN井震映射反演方法可以防止过拟合、提高泛化能力和反演精度,为精细刻画薄互层油气藏的空间展布提供了一项智能化的新技术。Conventional convolutional neural networks(CNN)are prone to over fitting because a large number of logging data cannot be obtained in the area with few wells.Therefore,a small sample CNN mapping inversion method based on well logging-seismic data is proposed.The optimal network layer number,convolution kernel size,feature map scale and activation function are obtained through the optimization of network structure,and the optimal network model is applied to the inversion of actual data.The application results show that the small sample CNN mapping inversion method based on logging-seismic data can prevent over fitting,improve the generalization ability and inversion accuracy,which provides a new intelligent technology for fine describing the spatial distribution of thin interbedded reservoirs.

关 键 词：井震联合反演 卷积神经网络 小样本学习 

分 类 号：P618.130.21[天文地球—矿床学]