A shapelet-driven distillation generation method for generalized zero-shot learning in compound fault diagnosis

Acquiring comprehensive compound fault data for rotating machinery is costly and limited, posing significant challenges for deep learning-based intelligent diagnosis. Zero-shot learning (ZSL) addresses this issue by leveraging seen-class data to construct class-level semantics, enabling the recognition of unseen compound faults without target samples. However, existing ZSL methods suffer from semantic deficiencies, limited robustness, and weak orthogonality, reducing their effectiveness when single and compound faults coexist. Therefore, this paper proposes a shapelet-driven distillation generation (SDDG) method within the generalized zero-shot learning (GZSL) framework to enhance fault recognition and improve generalization to unseen fault classes. SDDG enhances semantic discrimination and robustness, while mitigating inter-class interference, by incorporating sampling order optimization and location sensitivity metrics into an information gain-driven shapelet screening process. Furthermore, a co-optimization strategy for feature generation and distribution refinement is introduced, ensuring representation consistency and improved generalization through feature alignment loss, logical representation space mapping, and classification loss. Experimental results on public and lab-built datasets demonstrate that SDDG outperforms state-of-the-art methods, achieving superior performance in compound fault diagnosis under the GZSL framework.