Ultrasonic enhanced hierarchical deep learning framework for advanced LiFePO4 battery multi-state joint estimation

With the rapid development and increasing complexity of battery storage systems, achieving comprehensive and precise battery management requires transitioning from single-state estimation to accurate multi-state joint estimation. However, accurate multi-state joint estimation for LiFePO₄ batteries remains challenging due to two key factors: the flat external voltage curve leads to weak observability of internal states and independent estimators fail to capture the strong coupling between multi-states. To address these issues, we introduce ultrasound to obtain in-situ and in-operando information about the battery's internal physical and electrochemical states, significantly enhancing multi-states observability. Nonlinear correlation analysis reveals that ultrasonic time-of-flight (ToF) and spectral features show much stronger correlations with battery states than traditional external features. Furthermore, we propose a hierarchical deep learning framework with attention mechanisms to fully leverage the correlations between multi-states to improve the joint estimation. The estimation results demonstrate that the ultrasonic features and the hierarchical deep learning framework comprehensively enhance the core temperature (T_core), state-of-charge (SoC), and remaining discharge time (RDT) joint estimation of LiFePO₄ batteries. Compared to the traditional external features and independent estimators, the proposed framework achieves the RMSE of 0.198 °C (T_core), 1.045 % (SoC), and 208.5 s (RDT), resulting in RMSE reductions of 59 %, 52 %, and 61 %, respectively. This study pioneeringly introduces ultrasonic tests in multi-state joint estimation with high accuracy and low computational complexity, showing great potential in advanced battery management systems.