When Labels Shift: Few-Shot Fine-Tuning Versus Unsupervised Domain Adaptation for Cross-Dataset Battery Cycle-Life Prediction

Abstract

Lithium-ion batteries are widely deployed in electric vehicles and energy-storage systems, and accurate early prediction of cycle life is critical for safety, warranty cost, and second-life planning. Models trained on one dataset, however, rarely transfer to new chemistries, test protocols, or laboratories, and re-collecting full-life labels for every new scenario is prohibitively slow. This study investigates which transfer paradigm is effective for cross-dataset battery cycle-life prediction under such distribution shift. A CNN-LSTM regressor and two public datasets, Stanford (41 NMC532 cells) and Tongji (130 NCA/NCM cells), are used to run a six-experiment ladder covering in-domain baselines, direct transfer, few-shot fine-tuning, multi-kernel MMD alignment, MMD pre-training plus fine-tuning, and adversarial DANN, in both transfer directions, with strict isolation between adaptation and test cells and from-scratch controls for every supervised variant. Direct transfer collapses (MAE 517 cycles), driven mainly by a threefold gap in mean cycle life (220 versus 685 cycles). Fine-tuning with only five target cells reduces MAE by 81% to about 98 cycles, while unsupervised alignment methods (MMD 394, DANN 474) clearly underperform, and MMD pre-training adds no gain. The underperformance of feature alignment is attributed to label shift between the two fleets, and a pre-train-then-fine-tune recipe is recommended for transferable battery prognostics. Mixed-domain random splits are also shown to overstate model quality through domain leakage, so per-domain evaluation is advised.