Generative Model Combining Cross-Modal Reconstruction and Semantic Alignment for Generalized Zero-Shot Fault Diagnosis

Zero-shot fault diagnosis (ZSFD) serves as a successful paradigm for identifying unseen fault categories. However, the performance of ZSFD significantly deteriorates under generalized zero-shot conditions, i.e., to diagnosis both seen and unseen classes of faults. To address the issue, this paper proposes a generative model combining cross-modal reconstruction and semantic alignment (CRSA-GM). CRSA-GM initially encodes features of samples and attribute vectors from different modalities into latent vectors for reconstruction. A Barlow matrix is devised to measure the alignment between sample features and attribute vectors. Subsequently, a center-guided dual-modal generative model is proposed based on reconstructed data, considering both fault samples in feature space and semantic vectors in attribute space for synthetic sample generation. An incremental training strategy is also adopted, aiming at training the attribute learner in stages to enhance training efficiency. The process on the the three-phase transmission system demonstrates that the proposed CRSA-GM exhibits remarkable superiority and effectiveness in both zero-shot and generalized zeroshot fault diagnosis.