PID-NN: A Dual-Branch Neural Network With Physics-Informed Constraints for State-of-Health Estimation of Lithium-Ion Batteries

Le résumé fourni par la source

Accurate prediction of the state of health (SOH) of lithium-ion batteries is crucial for ensuring the safety of on-board battery management systems (BMS) and for optimizing their lifetime. However, existing data-driven approaches face challenges in multi-scale feature extraction and lack physical interpretability, particularly in jointly modeling short-term local fluctuations and long-term global degradation trends. This paper proposes a dual-branch architecture framework, PID-NN, that integrates convolutional neural networks (CNNs) with the state-space model Mamba. The two branches are adaptively fused at each time step via a Softmax-Gate, enabling effective integration of local and global degradation information. Additionally, a set of physics-informed loss terms grounded in electrochemical principles is introduced, and employ a dynamic block sampling strategy to enhance training efficiency and SOH prediction accuracy. Experiments conducted on four public datasets, covering 387 battery cells, demonstrate that the proposed method achieves a mean absolute percentage error (MAPE) of 0.79%, outperforming mainstream baseline models. Furthermore, PID-NN offers enhanced interpretability, stability, and few-shot transferability, indicating strong potential for practical deployment in next-generation battery health diagnostics.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.