Fault diagnosis of lithium-ion batteries using dynamic multi-time-scale decomposition of relaxation behavior

초록

Lithium-ion batteries are widely used in electric vehicles (EVs) and energy storage systems (ESSs) because of their high energy density and reliable performance. However, aging and manufacturing defects can compromise safety, efficiency, and lifetime, highlighting the need for early fault diagnosis in battery management systems (BMSs). Conventional state-evaluation methods based on pulse-derived direct-current internal resistance (DCIR) show substantial overlap between normal cells and those with electrode misalignment defects, limiting single-threshold early detection. To address this limitation, this study proposes an early anomaly detection framework that exploits voltage relaxation behavior following pulse currents during fast charging, using an SOC-windowed charge–rest protocol. Relaxation voltage time series are transformed into image snapshots via time-delay embedding and Hankel–Gramian Angular Field (GAF) mapping to better capture system dynamics. Dynamic Mode Decomposition (DMD) decomposes the response into dominant temporal and spatial modes across multiple time scales. Texture features are extracted from spatial-mode images using Local Binary Pattern (LBP) histograms and statistical descriptors, then projected into a PCA-based latent space constructed exclusively from normal-cell data. Hotelling's T2, Q statistics, and a combined fault score are used for anomaly assessment, with performance evaluated using five-fold cross-validation. Results show limited detection at low SOC (30–35%) due to phase-transition effects, whereas robust detection is achieved at 40–55% SOC, with accuracy, precision, recall, and F1-score consistently above 0.9. Requiring only voltage–current data and short rest periods, the proposed method demonstrates the feasibility of a practical, sensor-free approach for early detection of electrode misalignment defects under controlled charging–relaxation conditions. © 2026 The Authors

키워드