Gradient distribution correction replay network for dual-imbalance continual fault diagnosis of rotating machinery

In intelligent fault diagnosis of rotating machinery, continual learning supports the continuous identification of new fault types from non-stationary data by retaining previously acquired knowledge. However, in real-world monitoring with non-uniformly distributed data, continual learning methods face a dual-imbalance challenge: disparities in sample sizes among fault types within a task (intra-task imbalance) and distributional shifts between stored exemplars and new samples across tasks (inter-task imbalance). To tackle these issues, a novel gradient distribution correction replay network (GDCRN) is proposed for continual fault diagnosis under fault-type increments and dual-imbalance scenarios. For intra-task imbalance, we introduce a gradient adaptive weighting strategy that dynamically adjusts the learning contribution of each fault type through accumulated gradients, thereby boosting the model’s sensitivity to neglected fault types. To mitigate inter-task imbalance, we design a novel distribution-aligned knowledge distillation loss that preserves the information of the original training distribution while alleviating forgetting caused by distributional shifts. Furthermore, a decoupled gradient distribution correction module is constructed to further balance the learning process of new and old tasks. Validation on two distinct train wheelset bearing datasets demonstrates that the GDCRN achieves superior performance with enhanced robustness and effectiveness in diagnostic tasks under fault-type increments and dual-imbalance scenarios.