摘要
离心泵是工业中能量转换和流体输送的核心设备,其部件滚动轴承的可靠性对整个机组的安全运行尤为关键。为了解决目前滚动轴承寿命预测问题,对滚动轴承剩余寿命的最佳预测方案进行了研究。首先,从数据驱动和试验出发,利用试验台采集所得的离心泵轴承正常及故障状态下的数据,分析了时域、频域、时频域各特征在不同工况中的表现差异,发现了时域特征、频域特征、小波包分解能量特征、完全自适应噪声完备集合经验模态分解(CEEMDAN)能量特征可以捕捉到不同工况下的故障信息;然后,以单调性、趋势性指标加权分数为依据,结合特征的敏感性分析结果,优选出了轴承在全寿命周期中表现突出的12个特征,经核主成分分析(KPCA)-长短期记忆网络(LSTM)降维处理后,构建出了能够表征离心泵轴承退化过程的一维特征量;最后,对比分析了LSTM网络、反向传播(BP)网络和卷积神经(CNN)网络的预测效果。研究结果表明:LSTM网络的均方根误差(RMSE)为0.402,平均绝对百分比误差(MAPE)为0.332,预测精度在三者中最好,模型平均训练时间为12.6 s,可见LSTM网络在预测精度及模型训练时间上更具优势。
Centrifugal pumps are essential equipment for energy conversion and fluid transfer in the industry,and the reliability of their component rolling bearings is crucial for the safe operation of the entire unit.To address the current challenge of predicting the life of rolling bearings,a study was conducted to determine the best prediction method for the remaining life of rolling bearings.Firstly,the performance differences of characteristics in the time domain,frequency domain,and time-frequency domain under various working conditions were analysed.The analysis data collected from a test bench under normal and fault conditions of centrifugal pump bearings were utilized.It was found that the fault information under different working conditions was captured by time domain characteristics,frequency domain characteristics,wavelet packet decomposition energy characteristics,and fully adaptive noise complete ensemble empirical mode decomposition(CEEMDAN) energy characteristics.Subsequently,based on the weighted scores of monotonicity and trend indicators,along with the sensitivity analysis results of the features,12 features with outstanding performance throughout the bearing's life cycle were optimized.After dimensionality reduction treatment using kernel principal component analysis(KPCA)-long short-term memory network(LSTM),one-dimensional characteristic quantities were constructed to characterize the degradation process of centrifugal pump bearings.Finally,the prediction effects of the LSTM network,backpropagation(BP) network,and convolutional neural network(CNN) network were compared and analysed.The research results indicate that the root mean square error(RMSE) of the LSTM network is 0.402,and the mean absolute percentage error(MAPE) is 0.332,showcasing the best prediction accuracy among the three models.Additionally,the model average training time is 12.6 s,further demonstrating the LSTM network's advantages in prediction accuracy and on model training time.
作者
苏皓南
黄倩
胡波
付强
朱荣生
SU Haonan;HUANG Qian;HU Bo;FU Qiang;ZHU Rongsheng(National Research Center of Pumps and System Engineering Technology,Jiangsu University,Zhenjiang 212013,China;China Nuclear Power Engineering Co.,Ltd.,Beijing 100840,China)
出处
《机电工程》
CAS
北大核心
2024年第6期941-955,共15页
Journal of Mechanical & Electrical Engineering
基金
国家自然科学基金联合基金重点资助项目(U20A20292)。
关键词
叶片式泵
滚动轴承
完全自适应噪声完备集合经验模态分解
核主成分分析
长短期记忆网络
轴承退化过程
vane pump
rolling bearing
fully adaptive noise complete ensemble empirical mode decomposition(CEEMDAN)
kernel principal component analysis(KPCA)
long short-term memory network(LSTM)
bearing degradation process
