Bearing Fault Diagnosis Of Small Sample Based On Compressed Sensing Reconstruction And Transferred Dictionary

With the development of industrial modernization,the safety and reliability of complex mechanical equipment have been paid more and more attention.The health status of the rolling bearing,which is one of the most common and easily damaged components,is related to the safe operation of the whole machinery.Most of the current intelligent fault diagnosis methods require the same distribution of training and test data.However,in practice,considering the safety and cost,the real fault samples are extremely scarce and it is difficult to meet the requirements of deep learning diagnosis methods,resulting in the performance of theoretical methods being greatly affected in practical applications.This paper combines compressed sensing,dictionary learning,and transfer learning theory to investigate the rolling bearing fault diagnosis methods based on small sample data enhancement.Firstly,while the actual application scenario is in a small sample situation where labeled data is scarce.This paper introduces CS(compressed sensing)to propose a data enhancement method for increasing samples of bearing faults.It first extractes the valid fault information by CS transformation,then applies multiple CS construction methods to restore the original and can obtain many fault signals which possesses both effective information and more diversity,finally the mixed dataset is used to train the deep neural network of fault diagnosis to improve the accuracy of fault classification.The comparison with traditional data augmentation methods shows that it can better improve the fault recognition performance under the circumstance of small samples.Secondly,considering that the sample volume of the source domain in the laboratory is much larger than that of the target domain in practice,and the source domain information is related to the target domain information in some respects,the paper proposes a data enhancement algorithm that combines dictionary learning and transferred reconstruction.At first,a large number of source domain data are applied to generate a source domain dictionary to obtain the shared representation coefficients,and then a small number of target domain samples are used to fine-tune the shared representation coefficients and update a specific source domain dictionary to generate a transferred dictionary to bridge the information transfer from the source domain to the target domain.Finally,a large number of new samples with target domain characteristics are created from the integration of source domain representation coefficients and transferred dictionary,which increases the volume of fault signals in the target domain.On this basis,the diagnostic model is trained to improve the fault recognition performance.Finally,the small sample data enhancement method studied in this paper is verified on several public datasets,and good diagnostic results are obtained.The paper also analyses the effects of key parameters in the method on the diagnostic results.The method in this paper puts forward new ideas for solving the problems of small sample fault diagnosis,and has certain significance to practical application.