基于机器学习的电潜泵工况诊断

为了减少电潜泵井电流卡片工况识别分析时的人为误差，建立了使用实时电流数据的基于机器学习的工况诊断模型。首先使用特征工程的方法，对电潜泵运行过程中的电流数据提取特征值；其次使用主成分分析法对特征值进行无监督降维聚类，并将聚类后的结果与实际工况进行对比证明聚类的有效性；然后使用降维后的带标签数据，建立逻辑回归模型；最后将未经训练的数据代入模型并进行误差分析。对A油田56口电潜泵井高密度实时电流数据进行了基于机器学习完整流程的工况诊断，结果表明，该模型在降低计算复杂度的同时，成功实现了正常工况、泵抽空、过载停泵、频繁短周期运行等4种常见工况的分类识别，诊断准确度、精确度、召回率均在80%以上，F₁分数85%，达到了期望的分类效果，证明了应用机器学习方法，使用实时电流数据对电潜泵工况诊断的可行性和可靠性。

Working condition diagnosis of electric submersible pump based on machine learning

The working condition diagnosis model of the electric submersible pump (ESP) based on machine learning (ML) using real-time current data was established to reduce human error in the working condition recognition and analysis of ESPs with current cards. Firstly, the feature engineering (FE) method was used to acquire the eigenvalues of ESP current. Secondly, the principal component analysis (PCA) method was used for the unsupervised dimensionality reduction clustering of eigenvalues, and the results of clustering were compared with the actual working conditions to prove the effectiveness of clustering. Thirdly, the logistic regression (LR) model was established using the labeled data after dimensionality reduction. Finally, the untrained data was substituted into the established model for error analysis. The complete working condition diagnosis process based on ML using high-density real-time current data was conducted for 56 ESP wells in Oilfield A. The results show that the model successfully realizes the classification and recognition of four common working conditions, including normal, pump depletion, overload shutdown, and frequent short-cycle operation. The accuracy, precision, and recall of the diagnosis reach more than 80%, while the F₁ score reaches 85%, which meets the requirement of classification. The feasibility and reliability of working condition diagnosis of ESPs based on ML using real-time current data are proved.