Privacy-Preserving Federated Learning for State-of-Health Prediction of Lithium-Ion Batteries With Randomized Quantization-Enabled Data Augmentation

Abstract

Lithium-ion batteries are extensively used in electric vehicles as efficient energy storage systems, making accurate prediction of their state of health (SOH) crucial. In practice, accurate SOH prediction requires large amounts of battery data for model training. However, individual battery users often lack sufficient data, and centralized learning with multiple users raise data privacy concerns, limiting leverage of distributed data resources. As an emerging technology, federated learning enables multiple clients to collaboratively train models while preserving data privacy. However, insufficient local data at each client still pose significant challenges for collaborative SOH prediction modeling. To address these challenges, we propose a novel collaborative battery SOH prediction modeling framework with randomized quantization-enabled data augmentation. The framework first proposes randomized quantization as a data augmentation strategy to expand each user’s local data and improve the accuracy of local models. Subsequently, federated learning is proposed with the Swin Transformer, serving as both the local and global model to enable privacy-preserving collaborative training. The proposed method is evaluated on three battery aging datasets, demonstrating its effectiveness in achieving high prediction accuracy while protecting data privacy. This framework provides a new perspective for advancing SOH prediction of lithium-ion batteries with multiple users.