摘要
针对微电阻率电成像测井仪器的特点导致测井成像呈现规律性空白带的问题,本文提出一种融合多尺度多层级特征的无监督填充模型及全井段填充框架用于填充空白带。填充模型采用UNet架构,利用非空白带区域电阻率数据自身的统计先验基于MAE损失进行无监督训练填充,主要通过以下2个措施对传统UNet进行改进:在编码器中引入MSR-Conv,提升单层网络的多尺度表征能力;在编解码特征连接环节引入多层级编码特征融合模块与信息引导模块,丰富上采样的特征尺度,减少解码过程中的信息丢失。实验结果表明:相较UNet,本文所提模型在自然场景数据集上的视觉效果与客观指标均有明显提升,其中PE降低了19.03%,SSIM提升了2.9%,PSNR提升了4.66%。全井段填充框架应用填充模型分段训练填充空白带电阻率数据后再合并,实现端到端填充单口井的微电阻率成像测井空白带,填充结果具有一定的鲁棒性,贴合实际生产场景。
Aiming at the problem that the characteristics of micro-resistivity imaging logging tools lead to the regular blank zone of the measured wellbore image,this paper proposes a filling model based on unsupervised learning framework,which integrates multi-scale and multi-level features,and a full-well section filling framework to fill the blank zone.The filling model adopts the UNet architecture,and uses the statistical prior of the non-blank zone resistivity data itself to perform unsupervised training filling based on MAE loss.The model is improved mainly through the following two measures:The multi-scale residual convolution is introduced into the encoder to improve the multi-scale representation ability of the single-layer network;The multi-layer feature fusion module and information guidance module are introduced in the encoding and decoding feature connection link to enrich the feature scale of upsampling and reduce the information loss in the decoding process.The experimental results show that compared with UNet,the visual effect and objective indicators of the model proposed in this paper are significantly improved on the natural scene dataset.PE is reduced by 19.03%,SSIM is increased by 2.9%,and PSNR is increased by 4.66%.The whole well section filling framework applies the filling model to train the filling blank zone resistivity data in sections and then merge them to realize the end-to-end filling of the micro-resistivity imaging logging blank zone of a single well.The filling results have certain robustness and fit the actual production scene.
作者
曾祥安
朱丹丹
周昊
徐朝晖
Zeng Xiang'an;Zhu Dandan;Zhou Hao;Xu Chaohui(College of Information Science and Engineering/College of Artificial Intelligence,China University of Petroleum(Beijing),Beijing 102249,China;Beiiing Key Laboratory of Petroleum Data Mining,China University of Petroleum(Beling),Beijing 102249,China)
出处
《电子测量技术》
北大核心
2024年第8期171-180,共10页
Electronic Measurement Technology
基金
国家重点研发计划(2019YFA0708304)
中国石油科技创新基金(2022DQ02-0609)
中国石油天然气集团有限公司-中国石油大学(北京)战略合作科技专项(ZLZX2020-03)
中国石油大学(北京)科研基金(ZX20200100)项目资助。
关键词
电成像测井
空白带填充
无监督学习
特征融合
全井段填充框架
electrical imaging logging
blank zone filling
unsupervised learning
multi-scale residual convolution
feature fusion
