Abstract
Accurate lithium-ion battery state of health estimation is critical for safety and range anxiety mitigation. Existing methods often lack interpretability in the extraction of feature fragments and fail to model spatial correlations between features. To address these gaps, this paper introduces a novel framework centered on interpretable feature engineering and synergistic spatial-temporal learning. The core novelty lies in using incremental capacity (IC) analysis on charging data, captured by onboard sensors, to dynamically select a 0.1 V voltage window based on IC peaks, ensuring the extracted voltage and capacity fragments are physically meaningful. These fragments are then transformed into graph-structured data, enabling a graph attention network and a bi-directional gated recurrent unit to effectively capture both spatial dependencies and temporal degradation trends, with a residual connection optimizing the network. Validation on two public benchmark datasets demonstrates the model's superiority, achieving an average mean absolute error of 0.561% and a root mean square error of 0.783%. Furthermore, the model exhibits a low computational footprint, requiring only 1.68 MFLOPs per inference, and its fast inference time of 17.55 ms on an embedded platform confirms its feasibility for practical deployment.