Fault diagnosis of planetary gearboxes under class-incremental scenarios based on hybrid knowledge distillation

Abstract The knowledge distillation algorithm in incremental learning enables the intelligent diagnosis model for planetary gearboxes to gradually expand the diagnosable fault classes. However, the standard knowledge distillation method still faces two primary challenges. One is that it is difficult to effectively suppress semantic drift, and the other is that it overlooks the mutual interference among highly similar faults in adjacent learning phases. To overcome the above challenges, we propose a novel method called hybrid knowledge distillation. To improve the diagnosis performance of planetary gearboxes under the scenario of continuously expanding fault classes. First, we propose memory proofreading to constrain semantic drift, thereby avoiding the exacerbation of catastrophic forgetting. Then, we introduce feature distillation to learn finer-grained features, reducing confusion among highly similar fault classes, and alleviating catastrophic forgetting caused by them. Finally, we propose advantage distillation to enhance the plasticity of the model. The effectiveness of the proposed method is verified by 10 fault classes of planetary gearboxes. The average incremental accuracy is 96.99%, and the average incremental forgetting is only 2.73%. Compared with the representative continual learning methods learning without forgetting, synaptic intelligence and incremental classifier and representation learning, the proposed method has the best performance and a large increase. It also adapts well to different class-incremental scenarios. The results show that proposed method can effectively alleviate catastrophic forgetting and well balance the stability and plasticity of the model.