Noninvasive electrolyte-linked aging diagnosis of LiFePO4 cells from post-charge voltage relaxation using a feature-enhanced deep learning approach

Electrolyte-linked aging (i.e., aging behavior associated with electrolyte degradation) in lithium iron phosphate (LiFePO 4 ; LFP) cells alters ion-transport properties and interfacial kinetics, leading to progressive performance degradation. We propose a noninvasive aging diagnosis framework based on post-charge open-circuit voltage relaxation analysis using feature mode decomposition (FMD) and a dual-input deep temporal architecture integrating one-dimensional convolutional neural networks, bidirectional long short-term memory, and multihead attention. FMD generates frequency-localized modes, and modes FMD3–FMD5 (mid-to-low-frequency bands below 0.03 Hz) are selected to capture electrolyte-sensitive relaxation dynamics. Compact spectral and contextual descriptors are incorporated through squeeze-and-excitation gating and pre-attention integration to enhance cross-scale representation. On 60-Ah cycled LFP cells (120 cells evaluated using five cell-disjoint folds), the model achieves a mean absolute error of 0.00725 (0.72%), root mean square error of 0.00992, and R 2 of 0.9856, with strong calibration performance (calibrated expected calibration error of 0.00646 and 100% coverage). Paired tests with Benjamini–Hochberg correction confirm statistically significant improvements over baseline models. Zero-shot evaluation on 2.5-Ah cells reveals predictable degradation under capacity shift, and lightweight calibration mitigates reliability loss. The proposed framework operates in real time, requires only a voltage input, and excludes state-of-charge features to prevent shortcut leakage, making it suitable for practical deployment in battery management systems. • Post-charge voltage relaxation enables noninvasive LiFePO 4 cell aging diagnosis • The state of charge is excluded to prevent protocol shortcuts and leakage • A filter bank ensures stable feature-mode decomposition • Cross-fitted residual-isotonic-to-linear calibration tightens reliability • OOD gaps for capacity/protocol can be alleviated via lightweight calibrators