Adaptive incremental diagnosis model for intelligent fault diagnosis with dynamic weight correction

Intelligent fault diagnosis (IFD) has become a research hotspot in the fields of prognostics and health management. Existing mechanical IFD methods cannot continuously learn and integrate new diagnostic knowledge. In engineering, new fault data is continuously collected over time, and it is costly to retrain IFD models when new fault mode data arrives. To solve this problem, this paper proposes a new adaptive incremental diagnosis model (AIDM) with incremental capabilities. The AIDM is composed of a feature extraction module, an exemplar library, and a series of linear classifiers. By adding new output nodes and adopting knowledge distillation loss, the quick reconstruction and updating of AIDM can be realized on the premise of avoiding catastrophic forgetting. In addition, to solve the stability-plasticity dilemma, a new dynamic weight correction algorithm is proposed to dynamically adjust the biased weight of different linear classifiers. In this way, the stable and reliable incremental training and dynamic updating of IFD models are realized. Finally, the proposed method is verified on bearings and gearboxes. The results show that the proposed AIDM has outstanding performance in incremental diagnosis tasks, which provides a new solution for the adaptive updating of the IFD model.