Multi-adversarial domain transfer network with adaptive threshold generation pseudolabels for fault diagnosis under strong interference

In practical industrial scenarios, the fault diagnosis for rolling bearings faces the challenges of scarce fault data and strong interference. The adversarial transfer learning can address those issues, but it often relies on global feature alignment, overlooking local and class-level features, leading to poor performance under strong interference. To overcome these difficulties, this paper proposes a fault diagnosis method based on multi-adversarial domain transfer network with adaptive threshold generation pseudolabels (MADTN-ATGP). A two-stage adversarial transfer learning strategy with dual feature extractors is proposed to address the difficulty of extracting transferable features under strong interference. The high-quality target domain pseudolabels are generated by an improved adaptive threshold strategy. Finally, global and local alignments are achieved by adversarial training of global/local domain discriminators, and the local maximum mean discrepancy (LMMD) is introduced to construct a joint loss function to achieve class-level alignment. The transfer tasks from interference-free data to interference data collected from an external interference simulation experimental platform are verified to demonstrate the superiority of the proposed method, offering a reliable solution for fault diagnosis under strong interference.