A Federated Transfer Learning Framework for Lithium-Ion Battery State of Health Estimation Based on Fast-Charging Segments

Abstract

Accurately estimating the state of health (SOH) of lithium-ion batteries (LIBs) using limited segments of fast-charging data is essential for effective battery management in electric vehicles (EVs). However, the task is complicated by two main challenges: insufficient training data from each target battery, requiring personalized models; and privacy risks associated with centralized data aggregation. To address these issues, this work proposes a two-stage federated transfer learning (TL) framework. In the first stage, federated learning (FL) enables multiple distributed batteries to collaboratively train a global model by sharing only model parameters, preserving privacy while learning generalized knowledge. In the second stage, this global model is fine-tuned using a small amount of local data from the target battery, resulting in a personalized model that captures individual battery characteristics. The framework is built on a lightweight convolutional neural network (CNN) enhanced with an efficient channel attention (ECA) mechanism, enabling accurate mapping from fast-charging segments to SOH values. Experimental results on a public fast-charging battery dataset show that the proposed method significantly outperforms both local-only models and conventional FL approaches without personalization. It achieves a root-mean-square error (RMSE) of just 1.13%, demonstrating its effectiveness in accurately predicting SOH, preserving privacy, and potential for real-world battery management systems.