Electrochemically constrained transformer with steady-state probability for lithium-ion battery risk evaluation

Accurate health assessment and early risk warning are essential for ensuring the safety and reliability of lithium-ion batteries. However, the complexity and uncertainty of real-world operating conditions hinder anomaly identification, physically consistent state prediction, and dynamic risk evaluation. To address these challenges, this study proposes a mechanism–data fusion framework integrating steady-state probabilistic analysis, electrochemically constrained Transformer model, and Bayesian dynamic inference. The framework employs a steady-state probability model to identify abnormal intervals and critical cells under unsupervised conditions, while a Transformer model constrained by a single-particle electrochemical model is developed to improve prediction accuracy and electrochemical consistency. Building upon anomaly indicators and prediction residuals, a multi-scale dynamic risk evaluation framework is established to characterize risk evolution and quantify system-level risk. Validation using real-world vehicle operation data demonstrates that the proposed framework achieves 92.3% accuracy in abnormal-cell identification and significantly outperforms benchmark models in battery state prediction, with a maximum error reduction of 67.3%. Moreover, the framework provides an early warning approximately 100 s before the onset of abnormal voltage fluctuations. This work highlights the potential of integrating electrochemical mechanisms with data-driven intelligence for battery safety management and offers a practical solution for next-generation intelligent battery monitoring systems.