基于二维图像和卷积神经网络的阀冷系统主循环泵故障诊断

Fault Diagnosis of Main Circulation Pump in Converter Valve Cooling System Based on Two-dimensional Images and Convolutional Neural Network

特高压直流输电工程中,及时发现并排除换流阀冷却系统的主循环泵的故障,对保障换流阀的稳定运行具有重要意义,为此针对主循环泵在故障时产生的振动信号,提出一种基于二维图像和卷积神经网络的阀冷系统主循环泵故障诊断方法。首先,通过变分模态分解(variational mode decomposition,VMD)联合奇异值分解(singular value decomposition,SVD)对振动信号进行去噪处理:使用VMD分解轴向、竖直径向和水平径向的振动信号,基于相关系数法获取最优本征模态分量;使用SVD对分量信号滤波后,通过分量空间重构获取去噪后的振动信号。然后,通过格拉姆矩阵将时序振动信号转换为振动图像,提取振动信号的时空特征。最后,将轴向、竖直径向和水平径向振动图像多通道并行输入AlexNet深度卷积神经网络,通过卷积层和池化层实现多层次特征融合,提高故障诊断准确率。分析结果表明,该模型故障诊断精度为91%,优于多层感知机算法、一维卷积神经网络和浅层卷积神经网络,可以为阀冷系统主循环泵的故障诊断提供方法基础,为现场人员安排计划检修提供理论依据。

In the construction of ultra-high voltage transmission projects,it is of great significance for ensuring the stable operation of the converter valve detecting and eliminating the faults in the main circulation pump of the converter valve cooling system in time.Therefore,in response to the vibration signals generated by the main circulation pump during faults,this paper proposes a fault diagnosis method in the converter valve cooling systems based on two-dimensional images and convolutional neural networks(CNNs).Firstly,the vibration signal is denoised through variational mode decomposition(VMD) combined with singular value decomposition(SVD).The VMD method is used to decompose the axial,vertical,and longitudinal vibration signals,and the optimal intrinsic mode functions based on the correlation coefficient method is obtained.After filtering the component signal by using SVD,the vibration signals with noise removed are obtained through component reconstruction.Then,the vibration signal is converted into a vibration image through the Gram angle field method,and the spatiotemporal features of the vibration signal are extracted.Finally,the axial,vertical,and longitudinal vibration images are input into AlexNet deep CNN,and multi-level feature fusion is achieved through convolutional and pooling layers,which can improve the accuracy of fault diagnosis.The results show that the fault diagnosis accuracy of this model is 91%,which is superior to traditional machine learning algorithms such as one-dimensional CNNs and LeNet.It can provide a methodological basis for the fault diagnosis of the main circulation pump in the valve cooling system and theoretical basis for the arrangement of planned maintenance.