Finding trustworthy neighbors: Graph aided federated learning for few-shot industrial fault diagnosis with data heterogeneity

Abstract

Federated fault diagnosis has drawn increased attention recently, which makes use of datasets from different clients with data privacy. However, data distribution varies across clients since facilities often work under different conditions, and thus clients may have the few-shot problem on different fault type. Personalized federated learning (PFL) methods have been proposed to analyze client relationship related with data distribution and establishes multiple models suitable for different clients. Nevertheless, such relationship analysis adopts a single description, which may be disturbed by clients with the few-shot problem and cannot be adapted to new clients with new dataset and unknown client relationships. In this paper, a graph aided federated learning method with a few-shot node inhibition mechanism (GAFL) is proposed to solve the problem by finding trustworthy neighbors for all clients. Collaboration graphs are designed in pair-wise and category-wise levels to describe client relationships so that the varied data distribution under different fault type could be distinguished. After proper initialization, graphs and personalized models are trained alternately with the designed few-shot node inhibition mechanism, where information flow from neighbors with the few-shot problem are controlled and less trusted during the training process. Moreover, an incremental learning strategy is designed to address new clients, in which graphs and model updates are restricted to a small range, so that old models are less interfered. The proposed GAFL is evaluated on the real-world datasets in load bearings, demonstrating the superiority of the proposed method for few-shot fault diagnosis problem with data heterogeneity.