Toward cross-domain zero-shot diagnosis: Integrating fault type and severity assessment in complex mechanical systems

Fault diagnosis under cross-domain scenarios and severity assessment remain understudied for complex mechanical systems. This paper proposes a cross-domain zero-shot diagnosis framework that utilizes historical data to identify new fault classes and assess their severity. First, a fine-grained competitive encoder is designed to enhance feature discriminability and mitigate domain discrepancies through layer-wise competitive screening and feature reconstruction. Second, we introduce dynamic-static attribute representations, where static attributes describe fault categories and dynamic attributes characterize severity levels. This dual representation overcomes the rigidity of fixed attribute vectors in conventional zero-shot learning. Finally, a simple residual network learns the nonlinear mappings between signals and semantic attributes in a shared latent space, enabling generalized diagnosis of unseen faults. Extensive experiments on two industrial circuit breakers demonstrate that the proposed framework achieves an average accuracy of 86.87% and performs fault severity assessment with a mean absolute error of 13.48%, proving its practicality for predictive maintenance.