Task similarity-based continual learning for multi-phase environments and its application in few-shot fault diagnosis

During the operation of industrial machinery, few-shot faults that are difficult to diagnose due to data scarcity and task heterogeneity are frequently encountered. Traditional deep learning methods struggle with these challenges, as they require large labeled datasets and lack adaptability to evolving fault patterns. Continual learning provides a promising solution by enabling models to learn sequentially while mitigating catastrophic forgetting. A task similarity-based continual learning (TSCL) fault diagnosis framework that incorporates feature replay and loss allocation strategies is proposed, enhancing knowledge retention and transfer across tasks in multi-phase environments. The feature replay mechanism identifies key samples from the previous phase based on feature similarity and replays them, projecting all samples into the feature space of the same probability distribution. Additionally, a loss allocation mechanism based on parameter importance is proposed that evaluates the significance of each parameter in previous phases and assigns appropriate update magnitudes, thereby enhancing the ability to retain previous task knowledge of the model. Experimental validations on four public datasets demonstrate that, in few-shot learning and multi-stage scenarios, the proposed method outperforms mainstream comparative approaches. In particular, on the excavator dataset from an industrial application, TSCL exhibits excellent stability and high accuracy in multi-phase learning, with a marked improvement in memory retention for previous tasks.