Test-Time Learning-Enhanced Transformer for Lithium-Ion Battery SOH Prediction

Accurate state-of-health (SOH) estimation is essential for battery management systems, but conventional data-driven models remain fixed after training and may lose accuracy when deployment conditions change. This paper proposes a Test-Time Learning-Enhanced Transformer (TTL–Transformer) for hetero- geneous LCO, NMC, and LFP battery data. The key novelty is that the model remains adaptable after deployment: as new unlabelled voltage, current, and temperature measurements become available, the CNN denoising module can be periodically updated to suit new batteries, ageing stages, temperatures, operating profiles, and measurement conditions. Meanwhile, the Transformer encoder and SOH regression head remain frozen, preserving the supervised SOH mapping and avoiding full-model retraining. Only 2.17% of the complete model is updated during this process. Under pooled multi-chemistry evaluation, the plain Transformer achieves an average RMSE of 1.20%. Adding the denoiser reduces RMSE to 1.05%, while post-deployment adaptation further reduces it to 0.79%, giving an overall improvement of 34.4%. The model also achieves RMSE values of 0.56–0.73% in matched-chemistry evaluation. These results show that the proposed architecture enables lightweight, label-free adaptation to changing conditions after deployment without modifying the trained SOH predictor.