Cell-to-pack transfer learning for battery health estimation using domain-adversarial parallel networks

Accurate estimation of the state of health (SOH) for lithium-ion battery packs is essential for ensuring the safety, performance of electric vehicles. However, cell inconsistency and limited labeled pack-level data significantly hinder reliable estimation. This study presents a cell-to-pack transfer learning framework for SOH estimation using a domain-adversarial parallel spatiotemporal network. First, robust health features, including voltage tangent angles and their corresponding time, are extracted from partial charging curves and validated via Spearman correlation. A parallel long short-term memory–Informer network is then constructed to jointly capture local temporal dependencies and global aging trends, with its hyperparameters optimized using Bayesian optimization. To realize effective knowledge transfer across domains, a domain adversarial training strategy with a gradient reversal layer is introduced, aligning the feature distributions between individual cells (source domain) and the battery pack (target domain) without explicit modeling of cell inconsistencies. Finally, a fine-tuning strategy with frozen shallow layers and hierarchical learning rates adapts the model to target data with minimal supervision. Experimental results on a four-cell series-connected battery pack demonstrate superior accuracy, achieving a maximum error of 2.29%. The proposed method offers a data-efficient and generalizable solution for battery SOH estimation under real-world constraints.