State of health estimation of lithium-ion batteries based on real-world charging behavior analysis and phase space topology of fragmented data

The fragmented and randomized nature of real-world electric vehicle fast-charging data poses significant challenges for accurate battery state of health (SOH) estimation. To address data truncation, this study proposes a novel diagnostic strategy for extracting nonlinear topological health indicators from partial multistage constant current (MSCC) segments via 2D phase space reconstruction (PSR). The degradation mapping is then established using a Denoising Aggregation Transformer (DA-Transformer). This architecture integrates a feature purification mapping block (FPM-Block) for latent noise suppression with a global self-attention mechanism for temporal modeling. The model’s generalization is systematically evaluated under identical-condition and cross-condition validations. Experimental results demonstrate that the DA-Transformer achieves high-precision SOH estimation using only randomized local segments. Notably, under stringent cross-condition validation against entirely unseen MSCC protocols, the framework maintains strong topological invariance within the 40%–80% state of charge (SOC) intervals. The model successfully restricts the root-mean-square error (RMSE) to 0.0059–0.0063, demonstrating highly competitive accuracy and robustness for practical deployment.