Bridging the feature gap: Heterogeneous transfer learning for lithium-ion battery health estimation using one-shot data

Data-driven approaches for estimating the state of health (SOH) of lithium-ion batteries (LiBs) have garnered substantial attention. Due to the variability in actual operating conditions and the diversity of battery types, distribution discrepancies between source and target domains are inevitable. Transfer learning (TL) has emerged as a promising technique to address these issues. However, most existing TL methods based on feature extraction require identical features in both source and target domains. To tackle this challenge, this paper proposes a heterogeneous-feature transfer learning (HFTL) method for scenarios where feature space of source and target domains are inconsistent. A two-step strategy for robust migration is proposed. In the first step, feature distribution correction is performed using the Box-Cox transformation (BCT). The second step involves feature alignment through a rotation matrix (RM), which incorporates maximum mean discrepancy (MMD) with the aid of auxiliary supervisory information. Additionally, a sliding window incremental ridge (SWIR) model is employed for capturing the knee point and continuous optimization during battery degradation. The proposed method achieved an average RMSE of 0.0133 and MAPE of 0.0121 across 16,983 cross-validation tests. Notably, it reduces MAPE by up to 60.44% compared to existing methods and enables real-time SOH estimation within 0.25 seconds. The above methods leverage the advantage of small sample sizes, offering a novel approach for the secondary utilization of LIBs.