Fault-sensitive component transplantation for data augmentation in class-imbalanced bearing fault diagnosis

Class imbalance remains a major challenge in data-driven bearing fault diagnosis because fault observations are often far less abundant than normal operating data. Existing augmentation methods typically generate or modify complete fault waveforms, which places a considerable modeling burden on the limited minority samples and may entangle fault-related information with background and sample-specific variations. To address this limitation, we propose Fault-Sensitive Component Transplantation Augmentation (FSCTA), a component-level augmentation framework for class-imbalanced bearing fault diagnosis. Rather than synthesizing complete fault waveforms, FSCTA performs component-level augmentation by extracting transferable fault-sensitive components from observed fault samples and transplanting them onto measured normal observations. In this way, scarce fault data are primarily used to characterize fault-related content, while the abundant normal data provide realistic background variability. Experiments on three bearing datasets under multiple imbalance ratios show that FSCTA consistently improves diagnostic performance across different downstream classifiers, with particularly clear advantages under severe class imbalance. Ablation and signal-level evaluations further verify the contribution of the component-transplantation strategy and show that the augmented samples retain diagnostically relevant distributional and spectral characteristics. These results suggest that exploiting the complementary roles of scarce fault observations and abundant normal data provides an effective alternative to complete-waveform generation for imbalanced bearing fault diagnosis.