A dual-domain shapelet-guided conditional generative framework for imbalanced bearing fault diagnosis

Abstract In bearing fault diagnosis, the scarcity of fault samples creates severe class imbalance and makes minority fault distributions difficult to learn reliably. Existing data-level methods mainly rely on interpolation, neighborhood relationships, or distribution modeling strategies to synthesize minority samples, which may fail to preserve complex temporal structures and fault-related local patterns. Although GAN-based methods provide flexible distribution modeling capabilities, their conditioning mechanisms often lack explicit representation of class-discriminative local patterns across temporal and spectral domains. To address these limitations, we propose ShapeGAN, a dual-domain shapelet-guided conditional generative framework for minority fault sample augmentation. Specifically, ShapeGAN first discovers class-specific temporal shapelets from minority and majority samples and derives the spectral representation of each retained segment using the discrete Fourier transform (DFT). A Conditional Mixing Module samples these paired representations, which are integrated by a bidirectional cross-attention generator. Training then proceeds from single-shapelet to multi-shapelet conditioning through the proposed Two-Stage Mixed Training strategy. Experiments on 24 imbalanced bearing fault scenarios show that ShapeGAN achieves the highest average Macro-F1 and G-Mean with the MLP classifier, while its average AUC remains comparable to that of the strongest baseline. With LSTM and ResNet, ShapeGAN achieves the best average performance across all three metrics, demonstrating the effectiveness of the proposed augmentation method across different downstream classifier architectures.