Privacy-Preserving Federated Learning Framework for LithiumIon Battery SOH Prediction Using Hybrid CNN-BiLSTM Models

Abstract

Lithium-ion batteries are widely used in electric vehicles and portable electronic devices, where accurate State of Health (SOH) prediction is essential for ensuring safety, reliability, and efficient energy management. This paper proposes a deep learning-based framework for SOH estimation using hybrid architectures including CNN1D, CNN-LSTM, CNN-GRU, and an enhanced Bidirectional LSTMCNN model to effectively capture nonlinear degradation patterns from battery data. The system utilizes NASA battery datasets (B5, B6, and B7) and applies preprocessing techniques such as normalization and Pearson correlation for feature optimization.To address data privacy concerns and improve model generalization, a federated learning approach is integrated, enabling decentralized training across multiple clients without sharing raw data. Experimental results demonstrate that CNN1D performs effectively for certain datasets, while the proposed Bidirectional LSTM-CNN model achieves superior performance with lower RMSE and MAPE values. The proposed framework enhances prediction accuracy, preserves data privacy, and provides a scalable solution for real-world battery management systems, supporting improved battery lifespan and operational safety.