Model Evolution Mechanism for Incremental Fault Diagnosis

Abstract

As the key issue of equipment operation and maintenance, fault diagnosis has attracted extensive attention, in which deep learning is often applied to the construction of fault diagnosis models. However, the changing operation state of industrial systems will lead to the decline of model performance. Therefore, it is necessary and urgent to explore the dynamic update of models. This work establishes an incremental fault diagnosis (IFD) framework, and further proposes a model evolution mechanism based on adaptive knowledge distillation and representative exemplar selection (MEMAR). Firstly, when new samples pour in, knowledge distillation (KD) is used to mine the potential correlation between old and new samples, with adaptive constraints on parameter optimization for model updating. Secondly, genetic algorithm (GA) is applied to selectively retain representative exemplar subsets, so as to enhance the model's ability to deal with dynamic data. Moreover, in order to further verify the effectiveness of MEMAR, the discrimination performance of the fault diagnosis model after multi-stage incremental learning is tested. The experimental results show that the proposed MEMAR can reduce the storage space of old samples and training time. Besides, it enables the fault diagnosis model based on deep learning to have excellent diagnostic performance for new and old samples.