Multi-granularity attribute refinement for compound fault diagnosis with temporal multi-resolution diffusion

Abstract Diagnosing compound faults in electric traction systems is challenging due to scarce labeled data, which renders standard supervised learning ineffective for unseen faults. Additionally, signal degradation caused by noise and complex pattern coupling further impede identification. To address these challenges, this paper proposes a zero-shot attribute transfer framework (ZSATF) for compound fault diagnosis without training examples of the target fault categories, in which attributes serve as a knowledge bridge between seen and unseen faults. Specifically, a temporal multi-resolution diffusion network is first introduced to reconstruct clean signals from noisy environments. Subsequently, an attribute refiner enhances the representation and inter-attribute relationships of fault patterns. Finally, an attribute-guided feature alignment module project signal features into the refined attribute space, enabling the identification of unseen compound faults by matching features with attribute-based definitions. Experimental results on CRH2 traction data demonstrate that ZSATF outperforms existing fault diagnosis methods.