An Attention-Based Multi-Feature Fusion Physics-Informed Neural Network for State-of-Health Estimation of Lithium-Ion Batteries

This study proposes an Attention Mechanism–Multi-Feature Fusion Physics-Informed Neural Network (AM-MFF-PINN) for accurate and physically consistent estimation of the State of Health (SOH) of lithium-ion batteries in practical battery management systems (BMSs). The model integrates multi-domain features, including time-domain, frequency–domain, and wavelet–domain indicators, to capture both macroscopic degradation trends and microscopic dynamical behaviors under varying operating conditions. A dual-correlation feature selection strategy that combines the Pearson correlation coefficient and the maximal information coefficient (MIC) is adopted to automatically retain the most degradation-sensitive variables, while a dynamic loss balancing mechanism adaptively coordinates data-fitting and physics-based constraints to ensure robust convergence. Experimental results on the Xi’an Jiaotong University (XJTU) and Tongji University (TJU) datasets demonstrate that AM-MFF-PINN achieves superior performance, with a mean absolute error (MAE) of approximately 0.002, a root mean square error (RMSE) of about 0.004, and a coefficient of determination (R2) of 0.99 for the XJTU dataset, and an MAE of 0.005, an RMSE of 0.006, and an R2 of 0.97 for the TJU dataset. These results indicate that the proposed method can provide reliable SOH estimates across different chemistries, temperatures, and charging protocols, using only standard charging data that are readily available in on-board and stationary BMSs. Therefore, AM-MFF-PINN offers a generalizable and practically deployable evaluation methodology to support early fault warning, predictive maintenance, and life-cycle optimization of lithium-ion batteries in electric vehicles and energy storage systems.