Physics-Informed Multi-scale spatiotemporal transformer for robust state estimation and fault localization in Lithium-ion batteries

Reliable state estimation and fault diagnosis are critical for battery operation. However, existing data-driven methods often suffer from overfitting and poor interpretability, especially when data is scarce. We propose PIMST, a Physics-Informed Multi-scale Spatiotemporal Transformer that integrates electrochemical mechanisms with deep learning. This approach enforces physical consistency and improves robustness. We incorporate a physics-constrained loss derived from the Thevenin Equivalent Circuit Model, which embeds Kirchhoff’s laws and Ampere-hour integration directly into the network. Additionally, our architecture employs multi-scale temporal attention to capture both long-term degradation and high-frequency dynamics, while a Graph Attention Network models inter-cell coupling. Evaluations on NASA, Oxford, and CALCE datasets demonstrate that PIMST achieves a 0.42% RMSE for State of Health estimation and 98.7% accuracy for fault diagnosis. The model significantly outperforms baseline methods in few-shot settings. These results establish PIMST as a reliable and physically grounded framework for intelligent Battery Management Systems.