Knowledge Library Network for Non-Exemplar Incremental Remaining Useful Life Prediction

Aiming at the difference of various incremental remaining useful life (RUL) prediction tasks, a novel nonexemplar incremental learning (IL) method is proposed, named knowledge library network (KLN). First, a multilevel focal self-attention (MLFSA) is developed to model the multiscale temporal dependencies and extract the key features. On this basis, a dual domain fusion temporal convolutional network (DDFTCN) is designed to mine the degradation information from both time-domain and frequency-domain data. Next, a task similarity judgment strategy specifically tailored for IL scenarios is proposed to assess the discrepancies between different tasks. Based on the task differences, a network augmentation rule is proposed for implementing the feature extraction network augmentation or task branch network augmentation. The feature extraction network augmentation is employed for the tasks with significant differences, while the task branch network augmentation is utilized for the tasks with small differences. Especially for the task branch network augmentation, a retention and absorption distillation strategy is proposed to enhance the model's capability to retain previous knowledge while adapting to new task. The network augmentation network rule is integrated into DDFTCN to form the KLN. Finally, the experimental results on the life-cycle bearing and gear datasets demonstrate that DDFTCN achieve an average accuracy of 93.6% across both datasets, surpassing the existing prediction methods. The experimental results on the incremental RUL prediction tasks validate the superiority of KLN over the typical and state-of-the-art IL methods.