Federated Denoising Transformer-Recurrent Regressor for Remaining Useful Life Prediction of Lithium-Ion Batteries

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries (LIBs), which serve as the core components of energy systems, is of paramount importance for ensuring safety and optimizing maintenance strategies. However, existing RUL prediction methods encounter three significant hurdles: insufficient data privacy protection, susceptibility to signal noise, and the complexity of modeling nonlinear degradation processes. To overcome these challenges, this paper introduces a federated denoising transformer-recurrent regressor (Fed-DTRR) for LIB RUL prediction. Leveraging the federated learning (FL) framework, Fed-DTRR enables collaborative training on multisource heterogeneous data while safeguarding participants’ raw data privacy through a dynamic weighted averaging aggregation strategy according to client model performance. The core denoising transformer recurrent regressor (DTRR) architecture within Fed-DTRR addresses the issues of noise and nonlinearity by integrating a Gated Recurrent Unit (GRU) to robustly extract temporal features from noisy signals, a Transformer module to capture long-range nonlinear degradation patterns, and a regression output layer to achieve precise RUL estimation. Experimental evaluations on the NASA and CALCE data sets reveal that Fed-DTRR outperforms traditional methods such as RNN and LSTM, as evidenced by significantly lower mean absolute error (MAE), root mean squared rrror (RMSE), and relative error (RE) values, thereby validating its enhanced capabilities in privacy preservation, noise immunity, and nonlinear modeling.