Bearings are critical components in rotating machinery, whose reliability directly influences equipment performance. However, maintaining reliable equipment operation poses significant challenges for fault data collection. The scarcity of fault data produces strongly imbalanced samples, which degrade the diagnostic performance. To address this issue, a digital twin (DT)-driven dual-stage domain adaptation feature augmentation method is proposed to mitigate data imbalance, which achieve the fault diagnosis under imbalanced samples. First, considering the structures and operating conditions of bearings, a general DT model with multiple degrees of freedom is established to obtain simulated vibration data for minority classes. Second, the improved deep subdomain adaptation network optimized by margin-aware regularization is designed to extract shallow domain-invariant features from simulated and measured vibration samples, achieving the first stage of domain adaptation. Third, the adversarial learning-based noise generation network (ALNGN) is innovatively proposed to learn the distribution of fault-irrelevant noise inherent in measured features, achieving domain adaptation in the second stage. The ALNGN generates fault-irrelevant noise added to simulated features and constructs synthetic features that closely resemble measured features. Finally, synthetic features are utilized to augment the limited measured features of minority classes for bearing fault diagnosis. Experiments indicate that the proposed method effectively alleviates data imbalance and enhances diagnostic performance.
Fan et al. (Thu,) studied this question.