基于贝叶斯神经网络的多臂测井套损检测方法-《计算机技术与发展》

文章信息/Info

Title:: Multi-arm Logging Casing Damage Detection Method Based on Bayesian Neural Network

Author(s):: CAO Mao-jun; WU Sheng-kun; School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China

Keywords:: multi-arm well diameter; casing damage detection; Bayesian neural network; variational inference; uncertainty

摘要:: 针对传统多臂井径测井套损检测过程中,测井资料人工解释准确性不高,管柱重要信息容易遗漏等问题,结合大庆油田某工区多臂井径测井数据,提出了一种基于贝叶斯神经网络的多臂测井套损检测方法。该方法可在对原始测井曲线方位校正、缺失值填充以及对常见套损类别进行曲线数据截取汇总的基础上,形成多臂井径数据集,同时对数据集进行归一化处理并以此作为训练数据进行套损检测实验。对比发现,在多臂井径测井套损检测问题上,采用的 MC Dropout 变分推理方法训练的贝叶斯神经网络,相较 BP 神经网络、随机森林、BayesByBackprop 和 SGLD 变分推理方法训练的贝叶斯神经网络,在性能和鲁棒性方面表现更优异。实验表明:该方法在多臂测井套损检测中有效性更高,平均准确率达到95. 67% ,较传统人工解释方法提升明显,并能给出可解释性更佳的分类结果不确定性,极大地提升了衡量检测结果的可信程度。

Abstract:: In response to the problems of low accuracy in manual interpretation of logging data and easy omission of important information in the traditional multi arm caliper logging casing damage detection process,we propose a multi arm caliper logging casing damage detection method based on Bayesian neural network,combined with multi arm caliper logging data from a certain work area in Daqing Oilfield. This method can form a multi arm wellbore diameter dataset based on correcting the orientation of the original logging curve,filling in missing values,and summarizing curve data for common casing damage categories. At the same time,the dataset is normalized and used as training data for casing damage detection experiments. Comparison shows that in the detection of casing damage in multi arm caliper logging,the Bayesian neural network trained using the MC Dropout variational inference method performs better in performance and robustness compared to BP neural network,random forest,the Bayesian neural network trained using BayesByBackprop and SGLD variational inference methods. The experiment shows that the proposed method is more effective in detecting casing damage in multi arm logging,with an average accuracy of 95. 67% ,which is significantly improved compared to traditional manual interpretation methods. It can also provide better interpretability for the uncertainty of classification results,greatly improving the credibility of measuring detection results.