基于VMD与多尺度一维卷积神经网络的故障诊断方法

A fault diagnosis method based on VMD and multi⁃scale one⁃dimensional convolutional neural network

针对旋转机械设备在多工况、小样本状态下故障诊断精度不高的问题,提出一种基于VMD与多尺度一维卷积神经网络(M1DCNN)的故障诊断方法。该方法首先利用VMD对原始振动信号进行分解,并以峭度为指标,筛选出峭度值最大的分量进行包络分析;然后构建包含多个不同尺度卷积核通道的卷积神经网络,并采用多尺度卷积核提取不同尺度下包络信号的特征信息,进而对故障进行识别。将该方法应用于齿轮箱中的齿轮和滚动轴承的振动数据分析,结果表明:该方法在多工况、小样本情况下均有较高的故障识别精度,且模型具有较强的泛化性能;同时,与单通道卷积神经网络(1DCNN)的对比分析表明,所搭建的多尺度卷积神经网络能更全面地提取信号特征,因而具有更高的诊断精度。

The fault diagnosis accuracy of rotating machinery equipment is not high under the conditions of multiple working conditions and small samples,so a fault diagnosis method based on VMD(variational mode decomposition)and M1DCNN(multi⁃scale one⁃dimensional convolutional neural network)is proposed.In this method,VMD is used to decompose the original vibration signal,and the component with the largest kurtosis value is selected by taking the kurtosis as the index to perform envelope analysis.Then,a convolution neural network with multiple convolution kernel channels at different scales is constructed,and the multi⁃scale convolution kernel is utilized to extract the characteristic information of envelope signals at different scales to identify faults.The proposed method was applied to the vibration data analysis of gears and rolling bearings in the gearbox.The analysis results show that the proposed method has high fault recognition accuracy in the case of multi⁃working conditions and small samples,and also the model possesses strong generalization performance.Additionally,the results of contrastive analysis with one⁃dimensional convolutional neural network(1DCNN)show that the constructed multi⁃scale convolutional neural network can extract signal features more comprehensively,and thus has higher diagnosis accuracy.