Transfer Learning for Battery Health Prediction under Varying Operating Conditions

Abstract

Battery health estimation and lifetime prediction are essential for the safe and reliable operation of Li-ion batteries. In practical applications, predictive models are often required to generalize across different cells or cell groups, where operating conditions, materials, and degradation behaviours may vary. This thesis investigates battery health modelling from a cross-condition perspective under constant-current operating protocols, where charge C-rate, discharge C-rate, and depth of discharge define the domain structure. The study first examines state of health (SOH) estimation using features extracted from incremental capacity (IC) curves. Although SOH estimation provides useful diagnostic information, it shows limited domain divergence in the present dataset, mainly because most available samples are concentrated in the early ageing stage where degradation trajectories across-conditions remain similar. Therefore, while still considering SOH estimation as an important diagnostic task, this thesis also considers remaining useful life (RUL) prediction because it can construct sufficiently large domain divergence to support the transfer learning objective. A transfer learning framework is then developed for cross-condition RUL prediction. The model is first trained on a source cell group and subsequently adapted to a target group through fine-tuning. Optuna with the Tree-structured Parzen Estimator (TPE) is employed to optimize model structure and learning hyperparameters, and a sensitivity analysis is conducted to examine the influence of early-stage data availability on prediction performance. The results show that direct cross-condition prediction is challenging and that models trained without transfer learning have limited generalization capability. Finetuning improves RUL prediction on the target group, demonstrating the practical value of transfer learning for cross-condition battery prognostics. The results also indicate that prediction performance is affected by feature selection, domain definition, and the quantity and selection of cells used for fine-tuning.