Exemplar-free class incremental learning for rotating machinery fault diagnosis via adaptive prototype correction and separation network

Class incremental learning technologies have been extensively studied in rotating machinery fault diagnosis to continuously learn and integrate new diagnosis knowledge. However, existing approaches usually require retaining some historical samples to replay previous diagnosis knowledge when learning new fault categories, which not only poses data privacy leakage but also wastes significant storage and training resources. To address these challenges, a novel Adaptive Prototype Correction and Separation Network (APCSN) is developed for exemplar-free incremental fault diagnosis. In the APCSN, an optimal transport theory-based prototype correction module is designed to adaptively transport historical category prototypes to the new representation space, effectively mitigating the prototype drift caused by model evolution. In addition, a hybrid contrastive learning module that incorporates the old category prototypes and new category features is designed to enhance intra-category feature compactness and inter-category feature separability, thus alleviating the feature overlap between new and old categories. Experiments conducted on the bearing dataset and wind turbine gearbox dataset demonstrate that the APCSN attains the diagnosis accuracy of 99.01% and 97.36%, respectively, outperforming state-of-the-art methods. The results indicate that the APCSN exhibits superior performance in incremental fault diagnosis without reserved old samples.