A new lifelong learning method based on dual distillation for bearing diagnosis with incremental fault types

Abstract

In the rapidly evolving industrial environment, bearings may develop new fault types, posing significant challenges to deep learning-based intelligent fault diagnosis models. These models often suffer from catastrophic forgetting when encountering unknown fault types, resulting in performance degradation. Lifelong learning strategies offer a solution by enabling models to retain old knowledge while acquiring new information. However, traditional replay-based lifelong learning methods typically involve risks of privacy leakage and escalating storage costs. To address these issues, this study proposes a novel lifelong learning method called lifelong learning based on dual distillation (LLDD), which integrates a dual-distillation mechanism comprising dataset distillation and feature distillation, and introduces an equiangular basis vector (EBV) classifier. The dataset distillation technique compresses the dataset of each task into a small number of synthetic data that capture the essential information of the task, serving as replay exemplars. This approach reduces reliance on original data and storage costs. Feature distillation ensures that the model’s representations do not deviate significantly from previous ones. The proposed method effectively prevents an increase in the number of model parameters during the lifelong learning process by incorporating the EBV classifier, thereby maintaining model complexity stability. The performance of LLDD is validated on two bearing diagnosis cases with incremental fault types. Results demonstrate that the proposed method surpasses other lifelong learning methods in performance and memory efficiency.