Physics-informed neural network for co-estimation of state of health, remaining useful life, and short-term degradation path in Lithium-ion batteries

Lithium-ion batteries gradually degrade over time due to various internal and external factors. This degradation introduces significant safety and reliability risks, highlighting the importance of battery health management as a critical area of research. However, a major challenge remains to develop a universal health management approach that accommodates various battery materials, operating environments, and diverse tasks. To address this, we present a novel multi-task health management approach that combines a multi-task processing framework with a physics-informed neural network. By leveraging the co-design of shared and task-specific parameters alongside physics-informed feature extraction, the approach efficiently integrates the tasks of state of health estimation, remaining useful life prediction, and short-term degradation path forecasting. Specifically, the method captures voltage and current data before and after the constant voltage charging phase, and employs an improved transformer model to extract temporal information. An adaptive weighting method is then applied to integrate the task losses effectively. Experimental results demonstrate that the mean absolute percentage error (MAPE) of state of health (SOH) estimation is 0.75 %, the normalized deviation of short-term degradation path (S-DP) prediction is approximately 0.01, and the mean absolute error (MAE) of remaining useful life (RUL) prediction is 104 cycles. Model comparative experiments, sensitivity analyses of the training sample size, and transfer learning demonstrate that the proposed framework not only substantially improves prediction accuracy but also showcases strong generalization capabilities and practical applicability. This provides a novel research perspective for advancing battery management technologies.