Contrastive single-domain generalized diagnosis framework for bearing fault diagnosis under anonymous target operating conditions
Measurement Science and Technology · 27.03.2026 · 37 (12) · 126109
DOI: 10.1088/1361-6501/ae5409 ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 27.03.2026. Verfasst von Weiwei Qian und 1 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben.
Abstract (Originalsprache)
Existing contrastive learning-based intelligent fault diagnosis methods face three fundamental algorithmic challenges when generalizing from a single source domain to anonymous target operating conditions: (1) insufficient simulation of real-world signal variability, as conventional augmentation employs fixed-scale slicing and single perturbation types, failing to capture joint variations in rotational speed, environmental noise, and sensor characteristics; (2) instability in contrastive representation learning, where standard Siamese encoders lack constrained update mechanisms, leading to feature drift under domain shift; and (3) inadequate suppression of ambiguous hard samples, which blur class boundaries and degrade cross-domain generalization. To address these limitations, this paper proposes the contrastive single-domain generalized diagnosis (CSDGD) framework. First, a dual-level cascaded dynamic data augmentation library is designed to jointly model multi-scale temporal dynamics and diverse physical disturbances: variable-scale slicing simulates rotational speed variations, while difficulty-coordinated random augmentations (noise, quantization, drift, flipping, Mixup) mimic environmental and sensor-induced interference. Second, a three-stage progressive learning architecture with exponential moving average-constrained twin encoders is introduced to stabilize contrastive pre-training and ensure smooth transition to supervised fine-tuning. Third, an adaptive hard-sample refinement mechanism, driven by pseudo-label guided clustering and margin-based optimization, explicitly clarifies decision boundaries by mining and leveraging ambiguous samples. On the CWRU dataset, CSDGD achieves an average accuracy of 97.84% across 12 transfer tasks, significantly outperforming existing methods. On the more challenging Zhenjiang dataset with wide rotational speed variations, it attains an average accuracy of 89.93% over 42 tasks, demonstrating superior generalization capability and robustness under variable operating conditions.
Zitieren
APA
Qian, W., & Su, W. (2026). Contrastive single-domain generalized diagnosis framework for bearing fault diagnosis under anonymous target operating conditions. Measurement Science and Technology, 37(12), 126109. https://doi.org/10.1088/1361-6501/ae5409
Vancouver
Qian W, Su W. Contrastive single-domain generalized diagnosis framework for bearing fault diagnosis under anonymous target operating conditions. Measurement Science and Technology. 2026;37(12):126109. doi:10.1088/1361-6501/ae5409
BibTeX
@article{Qian2026,
author = {Qian, Weiwei and Su, Wentao},
title = {Contrastive single-domain generalized diagnosis framework for bearing fault diagnosis under anonymous target operating conditions},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {12},
pages = {126109},
doi = {10.1088/1361-6501/ae5409}
}
EUVIMED