Dynamic branch layer fusion: A new continual learning method for rotating machinery fault diagnosis

Abstract

In real-world environments, the critical components of rotating machinery often encounter various new fault types because of complex operating conditions. The replay-based continual learning method in fault diagnosis mitigates catastrophic forgetting associated with the introduction of previous fault samples. However, the retention of previous samples during the training of new tasks creates an imbalance in the distribution of dataset and limits the mitigation of catastrophic forgetting. A new continual learning method based on dynamic branch layer fusion is proposed and applied to the diagnosis scenarios with imbalanced dataset. In particular, the proposed method builds a branch layer for each old task to retain the old knowledge upon the arrival of a new task, then the branch layers fusion structure is designed to solve the problem of model growth. Additionally, a two-stage training process encompassing model adaptation and fusion is proposed. On this basis, integration loss is used to optimize the learning of models for all types across tasks. Finally, the assembly of the old and new models is achieved through distillation loss, enhancing the reliability of models on all tasks. Experimental results indicate that the catastrophic forgetting problem prevalent in imbalanced dataset can be effectively alleviated by the proposed method.