A New Incremental Training Framework Based on Dynamic Weight Allocation for Intelligent Fault Diagnosis

Abstract

Intelligent fault diagnosis (IFD) has become a research primary concern in prognostic and health management. In engineering, it is hard to collect all fault mode data in advance. However, the existing IFD-trained models usually do not have the ability to continue to learn and expand. It is also costly to retrain IFD models when new fault modes data arrives. To solve this drawback, this paper proposes a new incremental training framework for IFD model updating. The framework increases the model's ability to diagnose new fault modes by adding new linear classification nodes to the original model. Cross-entropy and knowledge distillation loss are used to avoid catastrophic forgetting, and a new dynamic weight allocation strategy is introduced to solve the stability plasticity dilemma. Finally, stable and reliable incremental training and dynamic updating of the IFD model are realized. The proposed method is applied to incremental fault diagnosis of bearings. The results show that the IFD model applied with the proposed framework has high accuracy in incremental diagnosis tasks, which provides a new solution for the expansion of the IFD model.