卷积神经网络在轴向柱塞泵故障诊断的应用

Application of Convolutional Neural Networks in Fault Diagnosis of Axial Piston Pump

柱塞泵作为液压系统的重要元件,其性能好坏直接影响液压系统的运转,因此柱塞泵的故障诊断一直是工程机械故障诊断的热点。针对传统的故障诊断方法需要人工设计并提取信号特征,信号特征提取不完善等问题,提出运用卷积神经网络对轴向柱塞泵进行故障诊断。在柱塞泵正常状态、松靴、配流盘磨损、滑靴磨损、中心弹簧失效五种工作状态下,采集柱塞泵的振动信号,将振动信号转化为频谱图与时频图,并加以标签标记,生成样本数据输入到卷积神经网络、深度置信网络、堆叠自动编码器进行不同网络性能比较。研究结果表明,当样本数据选择小波变换时频图,卷积神经网络相对于深度置信网络、堆叠自动编码器在轴向柱塞泵故障诊断方面具有更高的准确率,为92.17%。

As an important component of the hydraulic system,the performance of the piston pump directly affects the operation of the hydraulic system,so the fault diagnosis of the piston pump has always been a hot spot in the fault diagnosis of construction machin-ery.Aiming at the problems of traditional fault diagnosis methods that need to design and extract the signal features manually,and the signal feature extraction is not perfect,it is proposed to use Convolutional Neural Network to diagnose axial piston pump faults.The vibration signals of the piston pump are collected under five working conditions:normal state,loose shoe,valve plate wear,shoe wear and central spring failure.The vibration signal is converted into frequency spectrum and time-frequency diagram,and labeled.The sample data is generated and input into Convolutional Neural Network.Deep Belief Networks and Stacked Auto Encoder to com-pare the performance of different networks.The results show that the Convolutional Neural Network has a higher accuracy rate of 92.17%in fault diagnosis of axial piston pump than the Deep Belief Networks and Stacked Auto Encoder when the sample data select the wavelet transform time-frequency diagram.