Source-free domain adaptation for cross-domain remaining useful life prediction: A distributed federated learning perspective

As the complexity of industrial equipment continues to increase, determining the remaining useful life (RUL) with high precision holds substantial significance for maintaining intricate industrial systems. The development of cross-domain prognostic approaches without source domain data necessitates thorough investigation, given the inherent distribution shifts among edge devices’ degradation patterns and the imperative of preserving data security protocols. Furthermore, convolutional neural network, and long short-term memory network perform insufficiently when processing complex structurally dependent data. Consequently, this paper proposes a distributed RUL prediction approach based on graph convolutional neural network. Specifically, this paper designs a differential attention graph convolutional neural network that can focus on key areas in degradation data. Furthermore, considering the privacy and security of degradation data, this paper designs a two-stage decision boundary adjustment approach to achieve source-free RUL prediction under cross-domain conditions. On this basis, the study introduces a federated consensus mechanism that implements progressive weight calibration aligned with distributed training dynamics in edge computing environments, which can effectively reduce overfitting, and improve the generalization ability. Experimental validation on NASA’s publicly available aircraft engine degradation dataset confirms the operational efficacy of the proposed approach.