Reliable fault diagnosis of rotating machinery is essential for stable and safe operation of industrial systems. However, challenges such as significant data distribution discrepancies and intra-domain shifts caused by latent noise in the unlabeled target domain pose major barriers to the practical application of deep learning based fault diagnosis. These issues are particularly evident for samples near the decision boundary. Such samples often exhibit higher uncertainty and reduce the separability of the target feature space. To address these challenges, this paper proposes an uncertainty-aware domain adaptation framework for bearing fault diagnosis. First, we propose a perturbation-based module to measure sample uncertainty. It jointly uses perturbation sensitivity and historical information, enabling stable and reliable uncertainty estimation. Second, all domain alignment designs are adaptively guided by quantifying and exploiting sample uncertainty throughout training. For inter-domain alignment, we introduce an uncertainty-weighted discrepancy measure to improve the accuracy of cross-domain comparison. For the often-neglected intra-domain modeling, we further propose an uncertainty ranking based dual-matrix mechanism to improve pseudo-label reliability and preserve the geometric structure of the target domain. Experiments on a public dataset and an experimental bearing dataset demonstrate that the proposed method consistently outperforms state-of-the-art fault diagnosis methods in terms of diagnostic accuracy and reduce sample uncertainty.

Accurate fault diagnosis of rotating machinery is essential for maintaining the safe and stable operation of industrial systems, yet the availability of fault data is often severely limited in practical scenarios. Under such conditions, existing generative augmentation approaches face fundamental challenges, particularly due to unreliable initialization from data-agnostic noise and poorly constrained generation processes. To address these issues, this paper proposes a prior-driven flow matching framework (PDFM) for fault data augmentation under limited data conditions. PDFM reformulates fault sample generation as a constrained transport problem, where samples are initialized from data-supported source regions and continuously mapped toward the target fault distribution. Specifically, a prototype-based prior is introduced to construct a data-dependent source distribution, reducing the discrepancy between initial samples and the target data. Based on this structured source, flow matching is employed to learn a continuous and deterministic transport process, enabling stable and well-constrained generation without relying on adversarial training or stochastic denoising. By jointly improving generation initialization and transport dynamics, PDFM provides higher quality generated samples for downstream fault diagnosis. Extensive experiments on two bearing fault datasets demonstrate that PDFM consistently outperforms representative and state-of-the-art generative methods and significantly improves diagnosis performance under limited data conditions.