Physics-informed neural network and momentum contrastive learning for battery state of health estimation

Abstract

Estimating the State of health (SoH) of lithium-ion batteries is essential for ensuring their safe and efficient operation across various applications. Traditional approaches often struggle to balance accuracy, physical consistency and data efficiency. This paper proposes a novel combination model of Physics-Informed Neural Network and Momentum Contrastive Learning for Battery State of Health Estimation that associates the interpretability of physics-based model with the representational power of contrastive learning. Our innovation lies in developing a unified optimization strategy that carefully balances an estimation physics-informed architecture and the power of contrastive learning. To specifically improve the physics-informed network, we leverage a shared feature encoder to improve representation learning for accurate SoH estimation. For contrastive learning, we design a physics-guided data augmentation strategy with a shared encoder, which generates realistic variations of battery degradation patterns and a momentum encoder architecture, which stabilizes the learning process. Extensive experiments on the NASA lithium-ion battery datasets demonstrate that our model achieves superior performance over state-of-the-art baselines such CNN, BPINN, Informer and XGBoost-ARIMA, achieving a mean absolute error (MAE) average of 0.095% and a root mean squared error (RMSE) average of 0.117% across all batteries. The associations of physics constraints with contrastive learning improve prediction accuracy and enhance model generalization across different battery types and operating conditions, addressing key limitations in existing battery health estimation approaches.