Physics-guided conditional diffusion model for small-sample and class-imbalanced bearing fault diagnosis

Deep learning has become a mainstream approach for fault diagnosis. However, its application in industrial settings remains constrained by data scarcity. Although generative models have been introduced for data augmentation, challenges such as mode collapse and fault-frequency-band drift are still observed in existing methods. To address these issues, a physics-guided conditional diffusion model (PG-CDM) is proposed to generate high-quality fault data under small-sample conditions. Specifically, a dual-path physics guidance (DPPG) strategy is introduced at the input stage of the denoising U-Net. A theory-derived fault-band mask is concatenated with the noisy time-frequency map. In this way, the theoretical fault-frequency regions are anchored on the time-frequency plane, thereby improving consistency with bearing kinematics. In addition, a physics-constrained adaptive group normalization (PC-AdaGN) module is introduced during feature propagation. Physical parameters, such as the theoretical fundamental frequency, are mapped into a prior vector. The prior vector is fused with latent features to modulate affine transformations in residual blocks. This design aligns the energy distribution of weak periodic impulsive components with the expected bearing kinematics. Furthermore, a physics-aware coordinate attention (PA-CA) mechanism is integrated into the multi-scale encoding-decoding pathway. The model is guided to focus on critical fault-frequency regions, and weak periodic impulsive signatures are reconstructed. Experiments are conducted on the laboratory dataset, and the effectiveness of PG-CDM is validated. Fault-frequency-band drift in generated samples is effectively suppressed by PG-CDM. Higher fault-related feature fidelity is achieved compared with mainstream generative models. Under extreme class-imbalance ratios, classification accuracy above 90% is achieved. Overall, this study provides a physics-guided data augmentation framework for industrial fault diagnosis under small-sample and class-imbalanced conditions.