摘要
录、测井数据在储层流体识别,尤其是随钻阶段的流体识别中起着重要作用。录、测井数据体依赖于区域内井的数量,海上油气勘探在大数据维度上来说样本相对较少,导致在利用机器学习对储层流体进行识别时受限于有标签数据量太小,存在过拟合及泛化能力差等问题。针对上述问题,提出一种融合半监督学习(Self-Train)与马尔科夫链蒙特卡洛(MCMC)算法的流体识别方法。利用少量有标签数据初步训练神经网络模型;结合半监督学习算法为无标签数据生成机器标签(伪标签),然后使用马尔科夫链蒙特卡洛法随机采样量化模型预测的不确定性,筛选置信度高的机器标签,以扩充高质量训练数据集,最终结合筛选后的机器标签与原有标签数据,采用自适应训练方法调整利用有标签数据建立的神经网络模型,构建适用于小样本条件的录、测井数据储层流体识别模型。对新钻井进行模型验证,符合率达到85%以上,应用效果较好。MCMC方法筛选机器标签后建立的储层识别模型,提升了随钻流体解释模型的准确率与泛化能力,为井场快速识别随钻流体提供了有效的技术支撑。
Mud logging and well logging data play an important role in reservoir fluid identification,especially during the drilling stage.The data volumes of mud logging and well logging data depend on the number of wells in the area,and the number of samples is relatively small in terms of big data dimensions of offshore oil and gas exploration,which limits the machine learning of reservoir fluid identification due to the small amount of labeled data and leads to overfitting and poor generalization ability issues.To address the above problems,this paper proposes a fluid identification method that combines semi-supervised learning(SelfTrain)with Markov Chain Monte Carlo(MCMC).First,train the neural network model using a small amount of labeled data.Second,combining semi-supervised self learning algorithms to generate machine labels(pseudo labels)for unlabeled data.Then,using MCMC method to randomly sample and quantify the uncertainty predicted by the model,machine labels with high confidence coefficient are selected to expand the high-quality training dataset.Finally,by combining the screened machine tag with the original label data,and adopting adaptive training method to adjust and use the neural network model that is established with labeled data,a reservoir fluid identification model is created for mud logging and well logging data suitable for few-shot conditions.The model validation for new drilling wells achieved a coincidence rate of over 85%,demonstrating well application results.The reservoir identification model established after screening machine tags using the MCMC method improved the accuracy and generalization ability of the fluid interpretation model while drilling,providing effective technical support for rapid identification of fluids while drilling at the well site.
作者
张文颖
毛敏
袁胜斌
ZHANG Wenying;MAO Min;YUAN Shengbin(China France Bohai Geoservices Co.,Ltd.,Tianjin 300457,China)
出处
《录井工程》
2025年第4期29-35,共7页
Mud Logging Engineering
关键词
马尔科夫链蒙特卡洛
半监督学习
小样本
神经网络
流体识别
Markov Chain Monte Carlo
semi-supervised learning
few-shot
neural network
fluid identification
