Physics-Informed Nonlinear Extension Techniques for Robust Joint State Estimation of Li-Ion Batteries

To address the challenges of poor noise immunity and limited generalization performance in Li-ion battery modeling and state estimation (SE), a novel robust framework for parameter identification (PI) and joint estimation of state of charge (SOC) and surface temperature is proposed in this study by leveraging physical information and nonlinear extension techniques. Initially, a robust forgetting factor recursive maximum total correntropy algorithm is developed for PI, providing a solid foundation for SE under noisy conditions. Subsequently, a robust SOC estimation method is formulated by embedding the maximum correntropy criterion (MCC) with an adaptive kernel width into the square-root cubature Kalman filter, effectively replacing the conventional mean square error with MCC to enhance noise resilience. Next, a multidimensional feature input set is constructed using the PI results, including total internal resistance as auxiliary physical information, along with SOC estimates and raw measurements. A subinput structure is further designed using partial correlation analysis, and then the extreme learning machines are utilized to project the subinputs into a high-dimensional (HD) feature space to extract latent correlation features. Finally, by integrating nonlinear extended features with raw data in parallel, the input to the bidirectional gated recurrent unit model is generated, enabling simultaneous extraction of global representations from both HD and low-dimensional spaces. Experimental results demonstrate that the proposed method outperforms existing advanced approaches in SE under strong noise interference and complex operational conditions.