Multi-hop graph pooling adversarial network for cross-domain remaining useful life prediction: A distributed federated learning perspective

Accurate remaining useful life (RUL) prediction has gained increasing attention in modern maintenance management. Considering the data privacy requirements of distributed multi-client collaborative training and the phenomenon of domain drift, how to accomplish the RUL prediction for distributed federation under cross-domain conditions needs in-depth research. In this context, the paper constructs a multi-hop graph pooling adversarial network based on distributed federated learning (MHGPAN-DFL) for the RUL prediction. In particular, this paper designs a multi-hop graph pooling adversarial network, which can decrease domain differences through adversarial transfer while achieving global modeling for input data. Furthermore, this paper designs a predictive model consistency strategy based on distributed federated learning. It dynamically assigns model weights to promote the generalization ability based on ensuring the privacy and security of local data in each client. This study confirms the efficacy of the proposed approach adopting the NASA aircraft turbofan engine dataset, and the bearing degradation dataset provided by Xi'an Jiaotong University.