Battery State of Health Estimation from Discharge Voltage Segments Using an Artificial Neural Network

Abstract

Battery state of health (SOH) estimation is imperative for preventive maintenance, replacement, and end-of-life prediction of lithium ion batteries. Herein, we introduce a data-driven approach to state of health (SOH) prediction for battery cells using a Deep Neural Network (DNN). Our DNN model, trained on short discharge curve segments, outperforms Multilayer Perceptron (MLP) and Support Vector Regression (SVR) models. The Mutual Information (MI) score guides the selection of voltage range and width for model training, reflecting nonlinear degradation characteristics. A transfer learning strategy is applied for outlier cells, initially training on normal cells and fine-tuning with outlier cells, resulting in improved SOH predictions, particularly at higher cycles. The study finds that increasing the segment width reduces SOH prediction error, with the smallest segment of 0.05 V demonstrating good performance (RMSE of 0.0246), decreasing to 0.0142 at a width of 0.2 V. For outlier cells, transfer learning leads to a 48% reduction in RMSE. The partial segment-based approach offers potential for rapid SOH prediction in laboratory and field applications, enhancing efficiency in the development process.