Selective model transfer for lithium-ion battery state of health estimation based on a weighted similarity metric

Accurate estimation of lithium-ion battery state of health (SOH) is essential for safe and reliable electric vehicle operation. Most data-driven methods rely on full-life cycling or discharge measurements, which are unavailable in realistic deployments where only early-life charging data can be collected. This work proposes a similarity-aware transfer learning framework that determines when model adaptation should be applied to improve SOH estimation accuracy. A weighted similarity metric, derived from feature contributions of a source LSTM model, quantifies degradation proximity using early charging statistics and serves as a gate to decide whether transfer or scratch learning is more appropriate. Experiments across three datasets (MIT, NASA, Oxford) show that charging-only SOH prediction can achieve near-lossless accuracy under aligned conditions (MIT RMSE ≈ 0.185%), and that transfer provides improvement only when similarity is high, yielding a 72.9% normalized gain. When similarity is low, transfer degrades performance and scratch learning remains superior. The proposed framework provides a lightweight mechanism for safe, conditional SOH deployment and establishes a foundation for future domain-alignment pipelines in heterogeneous electric vehicle fleets.