Real-Time Physics-Aware Battery Health Monitoring from Partial Charging Profiles via Physics-Informed Neural Networks

TL;DR

This work tackles real-time battery health diagnostics by embedding aging-related parameters into a physics-informed neural network built on the single-particle model (P-PINNSPM). It achieves rapid internal-state inference (~30 s) and a 47× speedup over finite-volume methods, while significantly improving SOH estimation when internal parameters are included (up to ~87% relative improvement over voltage-only baselines) and enabling extrapolation to unseen aging states. The approach demonstrates strong generalization across unseen SOH levels, charging profiles, and operating conditions, and remains practical for Level-2 charging with tail-end data. Collectively, the results show that physics-informed machine learning can deliver real-time, data-efficient, and physics-consistent battery management insights, advancing next-generation BMS capabilities.

Abstract

Monitoring battery health is essential for ensuring safe and efficient operation. However, there is an inherent trade-off between assessment speed and diagnostic depth-specifically, between rapid overall health estimation and precise identification of internal degradation states. Capturing detailed internal battery information efficiently remains a major challenge, yet such insights are key to understanding the various degradation mechanisms. To address this, we develop a parameterized physics-informed neural network (P-PINNSPM) over the key aging-related parameter space for a single particle model. The model can accurately predict internal battery variables across the parameter space and identifies internal parameters in about 30 seconds-achieving a 47x speedup over the finite volume method-while maintaining high accuracy. These parameters improve the battery state-of-health (SOH) estimation accuracy by at least 60.61%, compared to models without parameter incorporation. Moreover, they enable extrapolation to unseen SOH levels and support robust estimation across diverse charging profiles and operating conditions. Our results demonstrate the strong potential of physics-informed machine learning to advance real-time, data-efficient, and physics-aware battery management systems.

Figures (8)

• Figure 1: The technical framework for real-time battery health diagnostic with internal physical status.
• Figure 2: Battery capacity fade and parameter sensitivity. (a): Battery capacity fade with respect to RPT cycle number; (b): Sensitivity analysis for independent parameters.
• Figure 3: Prediction performance of P-PINNSPM for concentration and terminal voltage across selected parameter combinations.
• Figure 4: Comparison of parameter identification results obtained by P-PINNSPM and PyBaMM.
• Figure 5: Extrapolation performance for unseen SOH values. (a): Estimation of SOH below 85% using identified physical parameters; (b): Estimation of SOH below 85% using voltage curves; (c): Estimation of SOH below 90% using physical parameters; (d): Estimatino of SOH below 90% using voltage curves; (e): Illustration of a potential application scenario enabled by the extrapolation capability, such as second-life battery diagnostics.