Transfer learning-based SOH estimation from partial charging data with polarization-aware modeling

Accurate state-of-health (SOH) estimation from incomplete charging data remains difficult when batteries are operated under variable charging protocols and when the prior polarization history is unknown. Many existing data-driven methods either rely on handcrafted health indicators or assume relatively fixed sampling conditions, which limits their robustness when charging segments differ in location, duration, and current profile. In this study, we develop a transfer-learning framework in which a long short-term memory (LSTM) network is first pretrained on a voltage-prediction task and is then fine-tuned for SOH estimation. The motivation is that accurate voltage prediction requires the hidden state to encode the latent polarization dynamics that connect recent current-voltage history to future battery response. Using the public Severson fast-charging dataset of 124 commercial lithium iron phosphate/graphite cells, we evaluate the proposed framework on 20%-capacity charging segments sampled at varying starting positions. Under the matched in-house protocol used in this study, transfer learning reduces the SOH estimation mean absolute error from 1.75% to 0.91% and the root mean square error from 2.35% to 1.30% relative to direct LSTM training. We further clarify the comparison protocol, discuss the scope and limitations of cross-paper comparisons, and position the method against recent partial-charging, CNN-based, Transformer-based, and transfer-learning studies. The results support the value of mechanism-informed pretraining for improving data efficiency and robustness, while also showing that broader benchmark reruns under a unified protocol remain an important next step for future work.