Electrochemical guided time frequency transformer for battery 2027 Applied E

Accurate estimation of battery state of health (SOH) is essential for ensuring the safety, longevity, and reliable operation of energy storage systems. However, existing data-driven SOH estimators still face two coupled limitations. First, many methods rely on fixed or manually selected aging features, which may become unreliable when degradation-sensitive regions shift across battery chemistries and operating protocols. Second, localized degradation-sensitive responses are often modeled as isolated feature vectors or purely sequential inputs, making it difficult to capture non-stationary variations in local-response patterns and inter-fingerprint dependencies within each cycle. To address these limitations, we propose an Electrochemical-Guided Time-Frequency Transformer (EG-Former). EG-Former first introduces a battery-specific fingerprint detection module that adaptively extracts incremental capacity (IC) peaks or current-switching points as compact degradation-sensitive markers. A graph convolutional embedding then encodes statistical dependencies among localized fingerprint regions, enabling structured representation of coupled intra-cycle responses. An adaptive spatial Transformer further models global.