A physics-aware model with contrastive language-image pre-training for few-shot fault diagnosis in smart manufacturing

Few-shot fault diagnosis of industrial rotating equipment is challenging because fault samples are scarce and weak signatures are often masked by noise. This paper proposes a Physics-Aware model with Contrastive Language-Image Pre-training (PA-CLIP) for few-shot fault diagnosis. The framework integrates a texture-enhanced visual branch, a mechanism-embedded physical branch, and a physics-regularised multimodal alignment strategy. Cyclic Spectral Coherence maps are used to represent vibration signals, while a consistent self-attention module enhances weak fault textures. An Adaptive Kinematic Filter embeds bearing kinematic priors into the network structure, and a Physics-Regularised Multimodal Loss preserves physical consistency during feature alignment. Experiments on the WHUT and CWRU datasets show that PA-CLIP consistently outperforms conventional baselines under few-shot conditions. On the noise-intensive WHUT dataset, it achieves 91.48% accuracy in the 1-shot setting and 98.83% in the full-shot setting. Ablation studies confirm the contributions of the visual branch, physical branch and PRM Loss. The framework also provides physically interpretable current-state health assessment and shows potential for lightweight edge deployment in industrial maintenance.