A zero-shot method for cross-condition compound fault diagnosis based on an Attribute-Conditioned Diffusion Model and Distribution Shift Calibration

Achieving zero-shot diagnosis of unknown compound faults under conditions of extreme sample scarcity while ensuring cross-condition generalization capability is of great significance for promoting the engineering application of intelligent diagnosis methods. Although sample generation-based generalized zero-shot learning provides a feasible approach for recognizing unknown compound faults, existing methods still face two major limitations: first, reliance on linear interpolation in high-dimensional feature spaces to construct compound fault prototypes causes synthetic features to deviate significantly from the real data manifold as dimensionality increases; second, generative models introduce prior and conditional distribution shifts, yet current research lacks effective calibration mechanisms, thereby constraining model generalization performance. To address these challenges, this paper proposes a diagnostic framework integrating an Attribute-Conditioned Diffusion Model (ACDM) and a Distribution Shift Calibration (DSC) method. Theoretically, it quantitatively demonstrates that linear intervention operations in existing methods lead to a significant increase in the expected distance between synthetic features and real compound fault features with growing dimensionality, and qualitatively explains the inevitable introduction of prior and conditional distribution shifts by generative models. Building on this, the ACDM is constructed, which overcomes the limitations of high-dimensional linear intervention by decoupling semantic representation from the high-dimensional feature space to a low-dimensional orthogonal attribute space, thereby achieving the generation of high-quality compound fault samples under zero-shot conditions. Furthermore, the DSC mechanism is designed, which systematically calibrates the prior and conditional distribution shifts in the generated data by jointly optimizing a conditional distribution alignment loss and a class separation loss. Experimental results show that the proposed method outperforms existing approaches in both same-condition and cross-condition diagnosis on multiple public and laboratory-built datasets, and provides a novel theoretical solution for zero-shot fault diagnosis. The code used in this paper is available at: https://github.com/EVANGELION-NO1/ACDM-DSC .