基于改进VMD和1DCNN的泵注系统轴承故障诊断

Fault Diagnosis of Bearings for Pumping System Based on Improved VMD and 1DCNN

针对泵注系统轴承振动信号传递路径长,故障特征微弱而导致传统机器学习方法难以进行故障诊断的问题,提出一种基于改进变分模态分解(IVMD)和一维卷积神经网络(1DCNN)的泵注系统轴承故障诊断模型。首先,通过多个传感器采集泵注系统轴承振动信号;然后,利用Elastic回归替换Ridge回归构造IVMD并将轴承振动信号自动分解,采用改进峭度指标进行筛选重构实现振动信号的有效降噪;最后,将降噪后的重构信号输入适用于工业多传感器系统的1DCNN进行自动特征提取和故障诊断。试验结果表明,IVMD-1DCNN模型的故障诊断准确率达99.27%,相比于其他方法具有较大优势。另外,对1DCNN学习到的卷积核进行可视化,也在一定程度探讨了深度学习故障诊断的可解释性问题。

Aimed at the problem that traditional machine learning methods are difficult to diagnose the faults due to long vibration signal transmission path and weak fault characteristics of bearings for pumping system, a fault diagnosis model of the bearings is proposed based on improved variational mode decomposition(IVMD) and one-dimensional convolutional neural network(1DCNN). Firstly, the vibration signals are collected through multiple sensors. Secondly, the Ridge regression is replaced by Elastic regression to construct IVMD, and the vibration signals are automatically decomposed, and the improved kurtosis index is used for screening and reconstruction to achieve effective noise reduction of the vibration signals. Finally, the reconstructed signals after noise reduction are fed into 1DCNN suitable for industrial multi-sensor systems for automatic feature extraction and fault diagnosis. The experimental results show that the fault diagnosis accuracy of IVMD-1DCNN model reaches 99.27%, which has great advantages over other methods. In addition, the visualization of convolution kernel learned by 1DCNN also discusses the interpretability of deep learning fault diagnosis to a certain extent.