State-of-health estimation of lithium-ion batteries using end-of-charge features with cross-domain validation

Abstract

Conventional lithium-ion battery state-of-health (SOH) estimation methods often rely on complete charge–discharge cycles or are restricted to specific battery chemistries and protocols, hindering their generalization across diverse real-world scenarios. This study proposes an SOH estimation framework based on an end-of-charge (EOC) fingerprint window: a voltage-current region anchored to the charge cutoff and independent of dataset-specific distributions, yielding statistical features of its high-voltage response and low-current decay. To capture both intra-cycle local variation and cross-cycle degradation trends, a hybrid network, CCT-Net, integrates a convolutional neural network (CNN) for local patterns, a convolutional block attention module (CBAM) for adaptive feature weighting, and a Transformer for long-range dependencies. For interpretability, Shapley additive explanations (SHAP) quantify each feature's contribution and its evolution across degradation stages. On the XJTU dataset, CCT-Net attains an average mean absolute error (MAE) of 0.45% across six operating conditions. Across 15 cross-domain scenarios spanning four datasets, the nickel‑cobalt‑manganese oxide (NCM) and lithium iron phosphate (LFP) chemistries, capacities from 1.1 to 20 Ah, and diverse protocols, domain-adversarial transfer learning attains an average MAE of 1.11% in the full-trajectory setting using one target-domain adaptation cell. Under a stricter early-life setting that uses only the first 10% of each target cell's cycles, the average MAE is 4.5% across 7 representative tasks. This work provides an interpretable and transfer-oriented approach to cross-domain SOH estimation from partial charging data.