An unsupervised domain adaptation approach with enhanced transferability and discriminability for bearing fault diagnosis under few-shot samples

As a key component widely used in electric multiple units (EMU), fault diagnosis of EMU bearing is an important link. Typically, labeled data from different conditions provides the most usable domain knowledge. However, many devices face the bottleneck of lacking sufficient labeled data under special conditions, known as few-shot samples distribution. Although unsupervised domain adaptation (UDA) can solve the above problems, existing models achieve sample transfer mainly by learning domain-invariant features in the source and target domains. Moreover, learning domain-invariant features does not necessarily guarantee sufficient discriminability and transferability of the sample. In turn, the samples transfer and the fault discrimination will be greatly affected. In this study, we propose an unsupervised domain adaptation approach with enhanced transferability and discriminability (ETDS-UDA) for bearing fault diagnosis of EMU under few-shot samples. First, we construct an efficient feature extractor (MiniNet) for fault feature extraction. Then, we construct ETDS-UDA based on UDA model by designing strategies that enhance simultaneously transferability and discriminability. Finally, we also propose a balanced strategy and a discriminative feature learning strategy to further optimize the final fault diagnosis. Ultimately, multiple results verify the performance of ETDS-UDA in EMU bearing fault diagnosis under few-shot samples.