基于LSTM-PNN神经网络的电潜泵故障诊断方法  

Fault Prediction Method for Electric Submersible Pumps Based on LSTM-PNN Neural Network

作　　者：周逸飞 刘新福[1] 曹砚锋 于继飞 欧阳铁兵 刘春花[3] 周伟 ZHOU Yifei;LIU Xinfu;CAO Yanfeng;YU Jifei;OUYANG Tiebing;LIU Chunhua;ZHOU Wei(Key Lab of Industrial Fluid Energy Conservation and Pollution Control(Ministry of Education),Qingdao University of Technology,Qingdao Shandong 266520,China;CNOOC Research Institute Ltd.,Beijing 100028,China;College of Mechanical and Electronic Engineering,China University of Petroleum(East China),Qingdao Shandong 266580,China)

机构地区：[1]青岛理工大学工业流体节能与污染控制教育部重点实验室,山东青岛266520 [2]中海油研究总院有限责任公司,北京100028 [3]中国石油大学(华东)机电工程学院,山东青岛266580

出　　处：《机床与液压》2024年第19期209-215,共7页Machine Tool & Hydraulics

基　　金：国家自然科学基金面上项目(52074161,52005281);泰山学者工程专项项目(tsqn202211177);山东省高等学校青创人才引育计划项目(2021-青创-30613019);山东省自然科学基金面上项目(ZR2022ME173);中海石油(中国)有限公司北京研究中心项目(CCL2023RCPS0237RSN,CCL2023RCPS0319RSN)。

摘　　要：针对电潜螺杆泵故障预测中发生故障难以及时发现、发现难以准确判别故障类型等问题,提出一种基于深度学习长短期记忆网络(LSTM)结合概率神经网络(PNN)的电潜螺杆泵故障预测方法。以LSTM网络为回归模型,使用时间序列法预测故障信号的未来趋势,利用小波包分解螺杆泵的故障信号,提取其中的故障特征,再结合油压、产量等多个工作参数,构建电潜螺杆泵的故障特征向量,并凭借PNN网络判别预测信号故障类型。收集新疆油田120组故障数据作为数据集对预测模型进行训练,从中取出90组数据作为故障数据库对模型进行训练,取出30组数据作为测试组测试模型准确率,使用LSTM-PNN神经网络预测模型分别对两组数据进行电潜螺杆泵故障预测。结果表明:预测前提取故障信号特征可有效提高电潜螺杆泵的故障预测精度,较常规电潜螺杆泵故障预测方法,LSTM-PNN网络预测具有更高的准确率且准确率提升了3%~16%。A fault prediction method for electric submersible screw pumps was proposed,addressing the challenges of timely fault detection and accurate fault type identification.The method combined long short-term memory networks(LSTM) and probabilistic neural networks(PNN).The LSTM network was employed as a regression model to predict the future trends of fault signals using time series analysis.The faulty signals of the screw pump were processed using wavelet packet decomposition to extract the fault features.Multiple operating parameters such as oil pressure and production yield were combined to construct the fault feature vector for the electric submersible screw pump.The PNN network was then utilized to classify and identify the predicted fault signals.A dataset of 120 sets of failure data from the Xinjiang Oilfield was collected for training the prediction model,and 90 sets of data were taken out as a fault database to train the model,30 sets of data were selected as the test set to evaluate the accuracy of the model.The LSTM-PNN neural network prediction model was applied to predict the faults in the electric submersible screw pumps using the two groups of data separately.The results show that performing fault feature extraction on the fault signals can effectively improve the accuracy of fault prediction for electric submersible screw pumps.Compared to traditional methods of fault prediction,the LSTM-PNN network demonstrates better predictive performance and its accuracy increases from 3% to 16%.

关 键 词：电潜螺杆泵 小波包分解 故障诊断 长短期记忆神经网络 概率神经网络 

分 类 号：TE933[石油与天然气工程—石油机械设备]