Order spectrum correction-based imbalanced single-domain generalization fault diagnosis for bearings under varying operating conditions

Abstract

Bearing fault diagnosis under varying operating conditions is crucial for ensuring the reliability of rotating machinery. However, the task is hindered by the unavailability of target-domain data during training, the limited diversity of single-source-domain data, and severe class imbalance. These issues substantially degrade the generalization capability of diagnostic models, and existing single-domain generalization approaches largely overlook the influence of class imbalance. To address these challenges, we propose a novel framework termed order spectrum correction-based imbalanced single-domain generalization (OISDG). OISDG comprises three key components. First, a condition-aware spectral correction module generates domain-invariant order spectral representations by suppressing condition-induced distortions. Second, an uncertainty-aware intra-class mixup strategy enriches minority-class representations by synthesizing informative same-class samples. Third, an uncertainty-aware contrastive module adaptively adjusts anchor weights and temperatures based on normalized uncertainty to enhance intra-class compactness and inter-class separability. Experiments on three benchmark bearing datasets demonstrate that OISDG achieves over 95% accuracy and 91% F-score, outperforming state-of-the-art methods. These results verify that OISDG provides a robust and generalizable solution for fault diagnosis under varying operating conditions.