<p>Accurate state-of-health (SOH) estimation of lithium-ion batteries under partial charging conditions remains challenging due to the inflexibility of existing methods requiring fixed voltage intervals. To address this challenge, this work proposes a novel ensemble learning framework capable of handling arbitrary voltage segments (as short as 0.01 V) by integrating fractional-order differential features, mutual information theory, and adaptive feature reuse. First, multiple base learners are independently trained on distinct voltage sub-intervals extracted from the constant-current charging phase. Their outputs are then concatenated and processed through a meta-model to obtain the final result. To enhance feature diversity, both integer-order and fractional-order differential curves are introduced, while mutual information theory is employed to identify the most discriminative feature set. The optimal base and meta-models are selected via cross-validation. Experiments on batteries with divergent aging patterns demonstrate excellent performance, achieving MAE and RMSE below 0.02 across all voltage intervals. This outperforms conventional approaches and maintains computational efficiency for real-world deployment.</p>