Physics-Based Neural Network for Unsupervised Health Assessment of Lithium-Ion Batteries

In the health assessment of lithium-ion batteries, physically interpretable health indicators (HIs) are extensively utilized to characterize the intrinsic degradation mechanisms. However, the deployment of sensors in actual operating environment is often constrained, which means it is challenging to obtain a diverse array of observation signals, hindering the construction of interpretable HIs. Furthermore, the experiment over full degradation cycles entails substantial time and material costs, resulting in insufficient datasets for algorithm training. To address these challenges, a physics-based neural network for unsupervised health assessment of lithium-ion batteries is proposed, which employs a feature extraction network based on the equivalent circuit model to acquire degraded features with physical significance, while leveraging the degradation dynamic model to guide the construction of HI through the physics-informed neural network (PINN). This integration seeks to reconstruct the neural network utilizing physical prior information, ensuring the accuracy and interpretability of health assessment under limited data. The approach is validated on a public dataset of lithium-ion batteries and demonstrates superior performance compared to existing physics-informed neural networks.