Deep dynamic adaptation network: a deep transfer learning framework for rolling bearing fault diagnosis under variable working conditions

Many cross-domain bearings fault diagnosis approaches have been developed by researchers. However, how to reduce the shift of training and test data remains a big challenge. To this end, a new deep dynamic adaptation network (DDAN) is developed for fault diagnosis. DDAN simultaneously takes advantage of stacked sparse autoencoder (SSAE), correlation alignment (CORAL), dynamic distribution adaptation (DDA) and domain-invariant classifier. Firstly, multiple domain feature extraction approach is developed to extract diverse features from raw signal, and then an unsupervised SSAE network as feature extractor to extract deep features from diverse original features. Secondly, CORAL reduces shift via matching the second-order statistics of training and test data. Finally, DDAN exploits the principles of structural risk minimization and DDA to learn an adaptive domain-invariant classifier for fault transfer diagnosis. Paderborn University (PU) and Case Western Reserve University (CWRU) bearing datasets were used to verify performance of the DDAN network. Comparing the performances with the best deep adaptation network (DAN), the average accuracy of DDAN is improved by 2.11%, and the SD is decreased by 1.76% on CWRU bearings dataset. Comparing the performances with best deep CORAL network, the average accuracy of DDAN is increased by 1.74%, and the SD is decreased by 2.31% on PU bearings dataset. The experimental results reveal that DDAN network can accurately diagnose fault type and effectively eliminate distribution divergence.