A Noncontact Open-Set Fault Diagnosis Method Based on Latent Space Disentanglement and Prototype Representation

With the widespread application of permanent magnet synchronous motors (PMSMs) in modern industry, transportation, and aerospace, ensuring operational stability and reliable fault detection has become increasingly crucial. Under complex operating conditions, potential faults may arise, making timely detection essential for safety and reliability. Traditional fault detection methods often rely heavily on manually labeled data or predefined fault patterns, limiting their adaptability. To address these challenges, this paper proposes a non-contact open-set fault diagnosis method based on latent space disentanglement and prototype representation. Unlike conventional VAE based methods that employ a single entangled latent variable, the proposed framework introduces a dual-variable latent space, explicitly separating class related features from stochastic rotational speed transition noise, thereby enhancing discriminative feature learning for open-set recognition. Furthermore, integrating prototype representation into the latent space tightens intraclass distributions and establishes adaptive distance based thresholds for unknown class detection. Experimental results on field collected PMSM datasets demonstrate that the proposed method achieves an open-set accuracy (OACC) of up to 99.07%, a closed-set accuracy (CACC) of 97.42%, and an AUROC of 99.95%. These results empirically validate the accuracy and robustness of the proposed method.