A target-domain augmentation and class-wise alignment framework for few-shot cross-condition bearing fault diagnosis
- 1Industrial Technology Center, North China Institute of Aerospace Engineering, Langfang, China.
- 2Hebei Province Technology Innovation Center for Aerospace Composite Structural Design and Manufacturing, North China Institute of Aerospace Engineering, Langfang, China.
- 3School of Mechanical and Electrical Engineering, North China Institute of Aerospace Engineering, Langfang, China.
Abstract
Cross-condition bearing fault diagnosis often has to adapt to a new operating condition with only a few labelled target samples. This setting creates two coupled problems. First, the sparse target support set cannot describe the target fault distribution sufficiently. Second, the same fault class may occupy different feature regions across source and target conditions, which weakens source-domain pretrained classifiers. Existing methods usually improve few-shot learning or reduce domain shift, but they rarely enrich the labelled target-class representation before class-wise source-target alignment. To address this limitation, this paper proposes a target-domain augmentation and class-wise alignment framework for few-shot cross-condition bearing fault diagnosis. The rime optimization algorithm (RIME) is first used to optimize the variational mode decomposition (VMD) parameters K and α, and fault-sensitive components are selected using envelope entropy and adjacent modal-frequency overlap. The reconstructed signal is converted into symmetric dot pattern (SDP) images to provide a noise-suppressed polar representation. The fast generative adversarial network (FastGAN) then generates auxiliary target-domain SDP samples from the limited labelled target support set, making each target fault class less sparse. On this basis, class-wise maximum mean discrepancy (CMMD) is embedded into a Swin Transformer classifier to align source and target features within each known fault category. Experiments on the Case Western Reserve University (CWRU) and Xi’an Jiaotong University–Shenyang Blower Works (XJTU-SY) datasets show that the proposed framework improves target-domain few-shot adaptation, achieving 93.33% and 97.58% accuracy under the G20 and G50 XJTU-SY target-augmentation settings, respectively.
Related
- Rolling bearing fault diagnosis method based on dual attention multi-scale convolutional network
- An innovative framework based on GMV-ALFF multi-modal MRI and Gaussian data augmentation for psychiatric disorder classification
- iAMP-deep: an alignment-free AI framework for insect antimicrobial peptide discovery
- Disentangling non-specific and condition-specific transcriptomic responses: a framework from heat stress in yeast
- Semantic Alignment of EEG and Text: A Contrastive Learning Framework for Decoding Imagined Speech
BibTeX
@article{hu2026targetdomain,
title = {{A target-domain augmentation and class-wise alignment framework for few-shot cross-condition bearing fault diagnosis}},
author = {Yijia Hu and Ying Gu and Yanjie Su and Chuan Zhao},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361486},
url = {https://doi.org/10.1371/journal.pone.0361486},
note = {CC BY},
}
RIS
TY - JOUR
TI - A target-domain augmentation and class-wise alignment framework for few-shot cross-condition bearing fault diagnosis
AU - Yijia Hu
AU - Ying Gu
AU - Yanjie Su
AU - Chuan Zhao
PY - 2026
DA - 2026/10/29
JO - PLOS One
DO - 10.1371/journal.pone.0361486
UR - https://doi.org/10.1371/journal.pone.0361486
AB - Cross-condition bearing fault diagnosis often has to adapt to a new operating condition with only a few labelled target samples. This setting creates two coupled problems. First, the sparse target support set cannot describe the target fault distribution sufficiently. Second, the same fault class may occupy different feature regions across source and target conditions, which weakens source-domain pretrained classifiers. Existing methods usually improve few-shot learning or reduce domain shift, but they rarely enrich the labelled target-class representation before class-wise source-target alignment. To address this limitation, this paper proposes a target-domain augmentation and class-wise alignment framework for few-shot cross-condition bearing fault diagnosis. The rime optimization algorithm (RIME) is first used to optimize the variational mode decomposition (VMD) parameters K and α, and fault-sensitive components are selected using envelope entropy and adjacent modal-frequency overlap. The reconstructed signal is converted into symmetric dot pattern (SDP) images to provide a noise-suppressed polar representation. The fast generative adversarial network (FastGAN) then generates auxiliary target-domain SDP samples from the limited labelled target support set, making each target fault class less sparse. On this basis, class-wise maximum mean discrepancy (CMMD) is embedded into a Swin Transformer classifier to align source and target features within each known fault category. Experiments on the Case Western Reserve University (CWRU) and Xi’an Jiaotong University–Shenyang Blower Works (XJTU-SY) datasets show that the proposed framework improves target-domain few-shot adaptation, achieving 93.33% and 97.58% accuracy under the G20 and G50 XJTU-SY target-augmentation settings, respectively.
SN - 1932-6203
ER -