State-of-health estimation for battery packs of real-world electric vehicles with cell-to-pack transfer learning

Transfer learning has emerged as a powerful tool for state-of-health (SOH) estimation of lithium-ion batteries. However, existing studies focus mainly on knowledge transfer from cells to cells under laboratory controlled conditions. A critical gap remains in adapting these methods to battery packs, where complex real-world conditions pose significant challenges. To bridge this gap, we propose a novel transfer learning framework that enables SOH estimation for battery packs by leveraging abundant single-cell data and limited labeled pack data. First, a transfer learning framework from cell to pack is designed to utilize the cell aging knowledge and enhance the generality of this method. Second, a fused aging feature vector is constructed, integrating health indicators from both charging curves (extracted via an autoencoder) and pack-level inconsistency information (selected through feature engineering). Third, an LSTM network pre-trained on single-cell data and fine-tuned with merely 10% early-stage labeled pack data estimates SOH using these fused features. Finally, validation on real-world electric bus aging data yields a 2.78% RMSE and 0.9083 R2, confirming the method’s efficacy and robustness under practical operating conditions.