Bio-inspired zero-shot learning with hierarchical fault-trait knowledge transfer for cross-domain diagnosis

Fault diagnosis methods for complex mechanical systems often face challenges in adapting to unseen fault types and varying operating conditions. Existing techniques typically suffer from limited cross-domain adaptability and weak zero-shot recognition capabilities. To overcome these limitations, this paper proposes a bio-inspired zero-shot learning framework that incorporates hierarchical fault-trait knowledge transfer and domain adaptation. A biomimetic network is designed to quantify domain discrepancies and generate transferable domain factors, using a hierarchical fault-trait description framework inspired by the differentiation process from general to specialized features. This approach enables fault diagnosis of previously unseen fault types by leveraging labeled data from the source domain only. The results demonstrate the effectiveness of the proposed method for cross-domain zero-shot diagnosis and suggest its potential applicability to complex mechanical systems.