基于半监督学习的抽油机井故障诊断方法研究  

Research on Semi-Supervised Learning-Based Working Condition Diagnosis Method for Sucker Rod Pumping Wells

作　　者：何岩峰[1] 王相[1] 褚宪翔 邵志伟 李明[1] HE Yanfeng;WANG Xiang;CHU Xianxiang;SHAO Zhiwei;LI Ming(School of Petroleum and Natural Gas Engineering,Changzhou University,Changzhou,Jiangsu 213164,China;No.2 Oil Production Plant of Sinopec Jiangsu Oilfield Company,Huaian,Jiangsu 211600,China)

机构地区：[1]常州大学石油与天然气工程学院 [2]中国石化江苏油田分公司采油二厂

出　　处：《钻采工艺》2025年第1期228-237,共10页Drilling & Production Technology

基　　金：国家自然科学基金项目“基于深度学习的抽油机井工况智能诊断方法研究”(编号:52204027)。

摘　　要：近年来,基于深度学习的油井故障诊断方法取得了显著进展,但这类方法的性能高度依赖标注样本集的数量和质量,且深度学习模型经过训练后可诊断的故障类型被固定,增加新类型需要重新收集样本并进行再训练,灵活性不足。为解决上述问题,文章提出了一种基于半监督学习的抽油机井故障诊断方法。该方法利用VGG19与小批量K均值相结合对大量示功图进行自动聚类分析,通过对聚类结果实施批量标注,能够有效提升样本分类的科学性及标注效率。在此基础上,构建基于欧氏距离的K近邻算法实现故障诊断,避免了深度学习方法中繁琐的模型训练及参数调优过程,同时支持样本集动态更新。基于矿场实际数据的实验结果显示,所提出的半监督学习诊断方法可达到与深度学习方法相当的准确率(均超过90%),但在前期准备阶段所需的时间成本减少了90%以上。更重要的是,当面对新出现的故障类型时,本方法能够快速响应并适应,极大地增强了抽油机井故障诊断系统的灵活性。In recent years,deep learning-based working condition diagnosis methods for oil wells have made significant progress.However,the performance of these methods heavily depends on the quantity and quality of labeled samples.In addition,the predefined working condition types that can be diagnosed by the deep learning model after training are fixed,and adding new types requires re-collecting samples and retraining,which is inflexible.To address these issues,this paper proposes a Semi-Supervised learning-based working condition diagnosis method for pumping wells.This method combines VGG19 with Mini-Batch K-means to perform automatic clustering analysis on a large number of dynamometer cards,enabling batch labeling of the clustering results to improve the scientific rigor and efficiency of sample classification.Based on this,a classifier using Euclidean distance of K-Nearest Neighbor is constructed to diagnose working condition,thus avoiding the cumbersome model training and parameter tuning processes required in deep learning methods,while also supporting dynamic updates of the sample set.Experimental results on actual field data show that the proposed semi-supervised learning diagnosis method achieves an accuracy comparable to that of deep learning methods(both exceeding 90%),while reducing time costs in the initial preparation stage by more than 90%.More importantly,when faced with newly emerging fault types,this method can quickly respond and adapt,greatly enhancing the flexibility of the pumping well working condition diagnosis system.

关 键 词：抽油机井 故障诊断 示功图 机器学习 半监督学习 

分 类 号：TE933.1[石油与天然气工程—石油机械设备] TP18[自动化与计算机技术—控制理论与控制工程]