Ensemble weighted DJPMMD based deep transfer metric learning method for fault diagnosis of bearing under variable working conditions

Abstract

Intelligent fault diagnosis of bearings under variable working conditions remains a challenging task. Although the deep transfer learning model can effectively diagnose the faults of bearing under variable working conditions, the diagnosis accuracy and stability of the individual deep transfer learning model are not high because of the instability of the individual diagnosis model and the difficulty in mining the hidden fault information contained in the original signal. An ensemble Yu norm-based deep transfer metric learning (Yu_DTML) combined with the improved weighted soft voting method based on the information entropy (IWSVMIE) and variational mode decomposition (VMD) is proposed to improve the diagnosis accuracy and stability. The preferred intrinsic mode function obtained through the VMD are transformed into frequency domain amplitude spectra through fast Fourier transform, and then input into multiple individual Yu_DTML models for training, respectively. Thereinto, to improve the transferability of each same fault class and the discriminability of different fault classes between different domains, the weighted discriminative joint probability maximum mean difference is introduced into the Yu_DTML model to measure the data distribution discrepancy between the source domain and the target domain. Finally, to make use of compensation of different individual Yu_DTML models, the IWSVMIE is proposed to fuse the preliminary results produced by all individual Yu_DTML models to obtain the final diagnosis result. The validation on three transfer diagnosis tasks of bearing under variable working conditions demonstrates that the proposed ensemble Yu_DTML based on IWSVMIE can effectively diagnose the faults with higher accuracy and stronger stability compared with other ensemble transfer learning methods.