基于数据映射和胶囊网络的轴承故障诊断方法

Bearing fault diagnosis method based on data mapping and CapsNet

传统深度学习模型自适应提取振动信号故障特征,实现端到端轴承故障诊断。然而,振动监测信号为非常复杂的非稳态时序信号,若深度网络直接以原始振动信号为输入,数据之间的非线性耦合作用会极大影响模型对故障特征的提取效率。目的为降低故障信号之间强非线性耦合作用,解决卷积神经网络对空间约束信息丢失的问题,提升轴承故障诊断性能,方法本文提出一种基于数据映射和胶囊网络(capsule network,CapsNet)的轴承故障诊断方法。首先,将图像处理领域中具有细化颜色特征能力的颜色空间模型(color names,CN)引入故障数据预处理中,将原始低维空间数据映射至高维空间,提升故障数据空间区分度;其次,针对映射后数据维度较高且具有一定冗余,影响故障诊断效率的问题,引入主成分分析(principal compo⁃nent analysis,PCA)法提取故障数据主元信息,降低数据维度;最后,考虑到胶囊网络有效提取空间约束信息的能力,将CapsNet作为故障诊断的骨干网络对故障特征进行识别和分类。结果使用CWRU、XJTU-SY数据集对该方法进行验证,实验结果表明,该方法在两种数据集上故障诊断准确率均达98%以上,与其他基于深度学习的故障诊断方法进行对比,该方法的诊断性能具有一定优势。结论本文方法可对故障数据进行有效解耦,提升数据之间的空间区分度,获得较高的轴承故障诊断精度。

Conventional deep learning models adaptively extract fault features from vibration signals to realize end-to-end bearing fault diagnosis.However,the vibration monitoring signal is a very complex nonstationary time series signal,and if the deep network directly takes the original vibration signal as input,the nonlinear coupling effect between the data will greatly affect the efficiency of the model for fault feature extraction.Objectives To reduce the strong nonlinear coupling effect between fault signals,and to solving the problem of the convolutional neural network on the loss of spatial constraint information so as to im⁃prove the performance of bearing fault diagnosis,Methods a bearing fault diagnosis method based on data mapping and capsule network(CapsNet)was proposed.Firstly,the color space model(color names,CN),which could refine color features in the image processing field,was introduced into the fault data preprocess⁃ing to map the original low-dimensional space data to the high-dimensional space and improve the spatial differentiation of the fault data.Secondly,to address the problem of high dimensionality and redundancy of the mapped data that affected the efficiency of fault diagnosis,principal component analysis(PCA)was in⁃troduced to extract the main meta-information of the fault data,which reduced the dimensionality of the data.Finally,considering the ability of the capsule network to effectively extract spatial constraint information,CapsNet was used as the backbone network for fault diagnosis to identify and classify fault features.Results The method was validated using the Case Western Reserve University(CWRU)and Xi’an Jiaotong Univer⁃sity(XJTU-SY)bearing datasets,the experimental results showed that the method achieved a fault diagno⁃sis accuracy of more than 98%on both datasets,and the diagnostic performance of the method had certain advantages when compared with other deep learning-based fault diagnosis methods.Conclusions The pro⁃posed bearing fault diagnosis method could effectively decouple the fault data,improve the spatial differen⁃tiation between the data,and then obtain higher bearing fault diagnosis accuracy.