Small sample bearing fault diagnosis based on compressed sensing reconstruction and dictionary transfer

Here, aiming at the problem of serious shortage of label samples for intelligent bearing fault diagnosis in practical applications, a data augmentation algorithm combining compressed sensing, dictionary learning and transfer was proposed for small sample bearing fault diagnosis study. Firstly, source domain label data were used to generate a specific source domain dictionary through wavelet packet dictionary learning and optimization method, obtain shared representation coefficients and acquire intrinsic fault information. Afterwards, a small amount of target domain signals was used to fine-tune shared representation coefficients, update the source domain dictionary and generate a transfer dictionary. Finally, a large number of new samples with target domain features were generated through shared representation coefficients and the transfer dictionary to realize data augmentation. The commonly used deep fault diagnosis network was used to verify the diagnostic performance of the proposed data augmentation algorithm here. The results showed that signals generated with the proposed method have effective information of faults, and can achieve better diagnostic performance when used for model training and recognition; the proposed method can provide a new approach for small sample fault diagnosis problems. © 2024 Chinese Vibration Engineering Society. All rights reserved.