Machine Learning-Based Electrode-Level State-of-Health Estimation for NMC/Graphite Battery Cells

Accurate health diagnostics of lithium-ion batteries are critical for ensuring safe, reliable, and prolonged battery operation. This study presents a data-driven approach to estimating electrode-level state of health (eSOH) using a deep neural network (DNN), enabling the assessment of the loss of active material (LAM) in both electrodes and the loss of lithium inventory (LLI). To construct the DNN models, essential features are extracted from the differential voltage and incremental capacity analyses of the open-circuit voltage (OCV), derived from a mechanistic model for a nickel-manganese-cobalt (NMC)/graphite battery. The DNNs are trained and tested using various hyperparameters, leveraging the OCV data. Specifically, accuracy and robustness are considered in the model selection and evaluation process to balance precision and generalization ability. In addition, feature reduction is performed considering three aspects: the number of features, robustness to noise, and the SOC window for data acquisition. Based on the analysis, three DNN models with reduced futures are meticulously chosen and extensively evaluated against the baseline all (12)-feature model, demonstrating their performance. Finally, a correction method utilizing the entropy change of the electrodes is proposed to improve the estimation accuracy in the presence of temperature variations from the training phase.