Few-shot class-incremental fault diagnosis of bearings via pseudo-incremental learning and dual-model complementarity

In real-world industrial scenarios, the rapid evolution of new fault patterns under complex operating conditions imposes great challenges for class-incremental learning (CIL), such as catastrophic forgetting of previously learned knowledge and poor generalization caused by overfitting under limited data. To address these issues, this study proposes an intelligent diagnostic framework based on pseudo-incremental learning and dual-model coordination. The framework first constructs a generalizable feature representation space through pretraining a base model on fundamental fault classes. A perturbation-based sample augmentation strategy is then employed to generate a structured dataset. A dual-objective optimization framework, which incorporates coordination tasks (between old and new classes) and specialized training tasks for new classes, is then used to enable effective integration of evolving fault modes. Finally, a lightweight parameter fine-tuning mechanism is applied. Experimental results on multiple bearing datasets demonstrate that the proposed method exhibits excellent dynamic adaptability and robustness in few-shot class-incremental learning (FSCIL) scenarios.