An uncertainty-aware prototype learning framework with structural constraints for open-world semi-supervised fault...

Real-world industrial fault diagnosis faces challenges from unknown fault types and limited labeled data, where existing methods often suffer from prototype collapse and unreliable clustering. This paper proposes an uncertainty-aware prototype learning framework with structural constraints for open-world semi-supervised fault diagnosis (OpenUPS). It introduces prototypes based on simplex equiangular tight frame to enforce uniformly distributed and maximally separated class centers, effectively preventing collapse under limited supervision. To address the varying reliability of unlabeled data, an uncertainty-aware contrastive strategy adaptively selects informative pairs, enabling robust alignment of seen classes and progressive clustering of novel faults. Experiments on the Tennessee Eastman process and a real-world polyester esterification process demonstrate that OpenUPS outperforms existing methods, achieving strong generalization and adaptability for open-world industrial fault diagnosis.