Experimental Evaluation of a Cascaded SoC–SoH Deep Learning Framework for Lithium-Ion Battery Management Systems

Accurate estimation of battery state indicators, particularly State of Charge (SoC) and State of Health (SoH), remains a fundamental requirement for advanced BMS, directly affecting safety, reliability, operational efficiency, and battery lifetime. Recent advances in deep learning have demonstrated promising results for SoC estimation and SoH classification individually; however, the impact of SoC estimation errors on subsequent SoH classification has received limited attention. This work investigates whether accurately estimated SoC information can effectively support data-driven SoH classification and quantifies the propagation of uncertainty across a cascaded battery diagnostic framework. A two-stage deep learning architecture combining LSTM-based SoC estimation and ResNet-based SoH classification is proposed and evaluated using three lithium-ion battery datasets. The SoC estimation stage is assessed using two experimental datasets (LG Chem E66 and Panasonic NCR18650PF), while the SoH classification and uncertainty-propagation analysis are conducted using the GS Yuasa LEV50 dataset generated from a validated battery simulation model. To improve temporal feature extraction and noise robustness, signal preprocessing based on Butterworth and Exponential Moving Average filters was incorporated. Experimental results achieved SoC estimation errors as low as 1.46% and an SoH classification accuracy of 94.23% when reference SoC values were used. Under realistic operating conditions, where SoH classification relied on estimated SoC values, the accuracy decreased to 79.65%. These findings provide a quantitative assessment of uncertainty propagation in integrated SoC-SoH pipelines and demonstrate that highly accurate SoC estimation alone does not necessarily guarantee equivalent SoH classification performance. The study highlights the challenges and limitations of cascaded data-driven battery diagnostics and provides insight for the development of more reliable intelligent BMS solutions.