Remaining useful life prediction of turbofan engine using global health degradation representation in federated learning

In recent years, deep neural networks have been widely applied in remaining useful life (RUL) prediction, and good prognostic performance has been achieved. However, existing centralized learning methods often ignore data privacy, modeling efficiency, and common feature of learning tasks. This paper presents a new RUL pre-diction method using global health degradation representation (GHDR) in federated learning (FL) framework named GHDR-FL, which aims to extract GHDR from distributed datasets and build personalized models for multiple clients. Specifically, GHDR is an aggregation of shallow features learned by the FL server and clients jointly. The head of each model is the unique superstructure, which is adopted to extract high-level features from the GHDR and local data on the client side. With the GHDR and unique superstructures, RUL prediction models customized for different operating conditions and fault modes can be built simultaneously in the FL. A degra-dation dataset of turbofan engines is used to evaluate the proposed method. The experimental results show that the GHDR-FL has high accuracy than the centralized learning methods, and the ready-made GHDR has strong versatility.