Battery usage-agnostic multi-task diagnostics using contrastive learning and knowledge-guided voltage relaxation

초록

Modern battery systems increasingly require various reliable tasks, especially in real-world settings where cell history and usage conditions are unknown. To address these challenges, we propose a unified framework that performs capacity estimation, state of health diagnosis, and cathode classification using only voltage relaxation data. A knowledge-guided equivalent circuit model (KG-ECM) is introduced to extract physically interpretable features, providing richer degradation information than conventional statistical descriptors. From the multi-task learning architecture, the model learns a shared and discriminative embedding that enables joint SOH diagnosis and cathode identification. Our results show that the proposed KG-ECM significantly improves capacity prediction accuracy, achieving an RMSE of 0.0026 in the single-task setting and 0.0108 under multi-task learning. The framework additionally attains 94.6% SOH classification accuracy and 99.6% cathode identification accuracy. Using only 10 relaxation cycles, the proposed method provides an interpretable and usage-agnostic solution suitable for real-world ESS, EV, and second-life battery applications. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드