Scholarlink

<p>Domain adaptation methods based on domain adversarial training have shown prominent application prospects in cross-condition fault diagnosis. However, the existing domain-adversarial approaches do not consider both domain-level and class-level discrepancies in target domain samples simultaneously, resulting in poor discrimination ability of the fault diagnosis model for targets in specific tasks. For this reason, a fault diagnosis method based on a model with the dynamic discriminator-free adversarial mechanism (DDAM) and layered decoding (LD) is proposed. First, a residual network (ResNet) with an implicit domain discriminator based on the nuclear-norm Wasserstein discrepancy (NWD) with feature dynamic-alignment mechanism on marginal and conditional distribution is constructed to achieve domain discrimination and adjusts the correlation between domain-level and class-level features to ensure sufficient alignment. Compared to current methods, the proposed DDAM has a simpler network structure and is more robust during training, while still achieving joint distribution alignment. Additionally, the layered decoding module is introduced into the network to enhance the extraction capabilities of common features in source and target domains to further improve the accuracy of cross-domain classification. Experiments on public bearing datasets and self-made bearing datasets demonstrate the effectiveness of the proposed method for cross-domain rolling bearing fault diagnosis, and its superiority over other representative methods has also been demonstrated by comparative experiments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-domain rolling bearing fault diagnosis based on dynamic discriminator-free domain adversarial residual network

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Xuejin Gao,
  • Yumeng Zhao,
  • Yue Liu,
  • Huayun Han,
  • Huihui Gao,
  • Yongsheng Qi

摘要

Domain adaptation methods based on domain adversarial training have shown prominent application prospects in cross-condition fault diagnosis. However, the existing domain-adversarial approaches do not consider both domain-level and class-level discrepancies in target domain samples simultaneously, resulting in poor discrimination ability of the fault diagnosis model for targets in specific tasks. For this reason, a fault diagnosis method based on a model with the dynamic discriminator-free adversarial mechanism (DDAM) and layered decoding (LD) is proposed. First, a residual network (ResNet) with an implicit domain discriminator based on the nuclear-norm Wasserstein discrepancy (NWD) with feature dynamic-alignment mechanism on marginal and conditional distribution is constructed to achieve domain discrimination and adjusts the correlation between domain-level and class-level features to ensure sufficient alignment. Compared to current methods, the proposed DDAM has a simpler network structure and is more robust during training, while still achieving joint distribution alignment. Additionally, the layered decoding module is introduced into the network to enhance the extraction capabilities of common features in source and target domains to further improve the accuracy of cross-domain classification. Experiments on public bearing datasets and self-made bearing datasets demonstrate the effectiveness of the proposed method for cross-domain rolling bearing fault diagnosis, and its superiority over other representative methods has also been demonstrated by comparative experiments.