基于拉普拉斯特征映射与加权极限学习机的电动潜油离心泵故障诊断方法  

Electric Submersible Pump Fault Diagnosis Based on Laplacian Eigenmaps and Weighted Extreme Learning Machine

作　　者：许泽坤 付军 高小永 张誉 李强[1] 檀朝东[1] XU Zekun;FU Jun;GAO Xiaoyong;ZHANG Yu;LI Qiang;TAN Chaodong(Department of Automation,China University of Petroleum(Beijing),Beijing 102249,China;Engineering Technology Branch of CNOOC Energy Development Co.,Ltd.,Tianjin 300452,China)

机构地区：[1]中国石油大学(北京)自动化系,北京102249 [2]中海油能源发展股份有限公司工程技术分公司,天津300452

出　　处：《控制与信息技术》2024年第2期117-125,共9页CONTROL AND INFORMATION TECHNOLOGY

基　　金：国家自然科学基金项目(22178383,21706282);中国石油大学(北京)科研基金项目(2462020BJRC004,2462020YXZZ023)。

摘　　要：电动潜油离心泵(简称“潜油电泵”)采油技术在非自喷高产井和高含水井中应用广泛,但其在运作过程中易发生设备故障,后续维护会触发长时间停机,可能造成无法估量的经济损失。目前对潜油电泵故障的诊断主要依赖现场技术人员的经验,无法快速及时地自动诊断分析。为此,文章提出了一种结合拉普拉斯特征映射与加权极限学习机的潜油电泵故障诊断模型。针对潜油电泵采集的数据存在严重不平衡性问题,其首先通过加权极限学习机建立故障诊断模型;然后,为解决算法学习不充分、加权策略会带来计算成本高和应用于高纬度特征空间的效果差等问题,其引入拉普拉斯特征映射方法对模型进一步优化;最后,在TE化工过程数据集上验证了所提方法的有效性,并在潜油电泵实时故障数据集上对该算法的实用性进行实验验证。结果显示,本文算法的分类平均准确率、最大准确率及G-mean相比支持向量机、决策树、BP算法、极限学习机以及加权极限学习机的平均提升了10%以上,验证了本文方法的有效性。Electric submersible pump(ESP)oil production technology is widely used in non-flowing high-yield wells and high water-cut wells,but equipment faults are prone to occur during operation,and subsequent maintenance may trigger long-term downtime,which may cause incalculable economic losses.At present,ESP fault diagnosis mainly depends on the experience of field technicians,and automatic diagnosis and analysis cannot be done quickly and timely.Therefore,this paper proposes an ESP fault diagnosis model based on Laplacian eigenmaps and weighted extreme learning machine.In response to the serious imbalance in the data collected by ESP,firstly,a fault diagnosis model is established using a weighted extreme learning machine;Then,to solve the problems of insufficient algorithm learning,high computational costs caused by weighted strategies,and poor performance in applying to high-dimensional feature spaces,the Laplacian eigenmaps method is introduced to further optimize the model;finally,the effectiveness of the proposed method was validated on the TE chemical process dataset,and the practicality of the algorithm was experimentally validated on the real-time fault dataset of electric submersible pump.The results show that the classification average accuracy,maximum accuracy,and G-mean of the algorithm proposed in this paper are improved by more than 10%on average compared with those of the support vector machine,decision tree,backpropagation(BP)algorithm,extreme learning machine,and weighted extreme learning machine,thus confirming the effectiveness of the proposed method.

关 键 词：不平衡数据集 故障诊断 加权极限学习机 流形学习 拉普拉斯特征映射 

分 类 号：TH165.3[机械工程—机械制造及自动化]