基于地质知识蒸馏学习的油气储集层识别方法  

Method of oil and gas reservoir detection based on geological knowledge distillation learning

作　　者：李徵 刘淇[1] 王喆锋 郑毅[2] 林霞[3] 怀宝兴 米兰 陈恩红[1] Zhi LI;Qi LIU;Zhefeng WANG;Yi ZHENG;Xia LIN;Baoxing HUAI;Lan MI;Enhong CHEN(Anhui Province Key Laboratory of Big Data Analysis and Applicationn,University of Science and Technology of China,Hefei 230027,China;Huawei Technologies Co.,Ltd.,Hangzhou China;Research Institute of Petroleum Exploration and Development,Beijing 100083,China)

机构地区：[1]中国科学技术大学大数据分析与应用安徽省重点实验室,合肥230027 [2]华为技术有限公司,杭州310051 [3]中国石油勘探开发研究院,北京100083

出　　处：《中国科学：信息科学》2021年第1期40-55,共16页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61922073,61672483,U1605251)资助项目。

摘　　要：油气储集层识别是石油能源企业在勘测和开发业务中核心的任务之一.长期以来,油气行业一直依靠专家人工分析海量测井数据以对地下油气储集层进行定性分析,虽然专家解释结论有着很高的精准度,但是时间与经济成本都十分高昂.近些年来,随着以深度学习为代表的人工智能技术的迅速发展,智能油气储集层识别技术成为学术界和工业界共同关注的问题.然而,真实工业环境存在严重的传感数据不一致问题,给传统的监督学习模型带来巨大的挑战.本文针对传感器不一致情境中油气储集层识别任务展开研究,提出多尺度地质知识蒸馏网络的方法.首先,该方法提出一种多尺度特征自注意力融合机制来学习地质信息的多尺度动态表征.其次,该方法设计一种地质知识蒸馏学习模型,从非一致传感数据中学习额外的地质知识,进一步提升模型准确度.最后,在真实数据集上进行大量实验,结果充分证明本文提出的模型在油气储集层识别任务上的有效性和鲁棒性.Oil and gas reservoir detection is one of the major tasks of petroleum energy companies in the exploration and production process.The oil and gas industry has long relied on the expert manual analysis of massive logging data to perform qualitative analyses of oil and gas reservoirs.Although experts’interpretations are highly accurate,the time and economic costs are considerably high.With the rapid development of artificial intelligence technologies such as deep learning in recent years,intelligent oil and gas reservoir detection methods have become a focus in the academia and industry.However,sensor data in real industrial scenarios present serious inconsistencies,which bring great challenges to traditional supervised learning models.This paper presents a focused study on the oil and gas reservoir detection task in the context of sensor inconsistencies and proposes a geological knowledge distillation multiscale network approach.This method proposes a multiscale feature fusion mechanism based on self-attention to learn the multiscale dynamic representation of geological information.Then,the model designs a geological knowledge distillation learning framework to learn additional geological knowledge from inconsistent sensor data.This step further improves the model’s accuracy.A large number of experiments on real industrial datasets are subsequently performed.The results fully prove the effectiveness and robustness of the proposed model in oil and gas reservoir detection.

关 键 词：油气储集层识别 地质知识 蒸馏学习 传感数据 深度神经网络 

分 类 号：P618.13[天文地球—矿床学]