Fault Diagnosis Method of Bearing Based on Manifold Neighborhood Preserving Embedded Distribution Alignment
The following challenges remain in the application of data-driven bearing fault diagnosis methods:the decrease in fault diagnosis accuracy is caused by the distribution differences between the samples with same fault type under variable operating condi-tions;the lack of sufficient and comprehensive fault data under actual operating conditions results in weak generalization ability of the trained fault diagnosis model.In view of this problem,a cross-domain fault diagnosis method for bearings based on manifold transfer learning was proposed,including four steps:the empirical mode decomposition was used to process the original vibration signal and ex-tract statistical features;then,the domain adaptation feature evaluation method was proposed based on naive Bayesian classification ac-curacy and domain differences,to select features with strong domain adaptation ability from the original set;the proposed manifold neighborhood preserving embedding distribution alignment was used to process the feature sets of the source domains and target do-mains,to reduce the distribution differences between domains,and a domain invariant classifier was trained under the principle of struc-tural risk minimization to achieve cross domain fault diagnosis of bearings;finally,two different types of bearing fault data were used to conduct cross domain fault diagnosis experimental analysis.The experimental results show that the proposed domain adaptive feature evaluation method can effectively select statistical features that are more conducive to domain adaptation processing,and improve the accuracy of cross domain fault diagnosis;the performance of the proposed manifold neighborhood preserving embedding distribution alignment method is significantly better than other classical feature transfer learning methods,achieving ideal cross domain fault diagno-sis performance.The fault diagnosis model constructed on this basis has the highest cross domain fault diagnosis accuracy of 100%and 95.17%under two types of bearing fault data,which is significantly better than other comparative models.
fault diagnosistransfer learningfeatures extractiondomain adaptationmanifold transfer learning
NETL
NSTL
万方数据
