Remaining Useful Capacity Estimation of Lithium-Ion Batteries Based on Electrochemical Impedance Spectroscopy Features and Convolutional Neural Networks

Accurate capacity estimation of lithium-ion batteries is crucial for reliable battery management, while electrochemical impedance spectroscopy (EIS) provides a non-destructive way to characterize battery degradation. However, existing EIS-based methods often directly use full-spectrum data, which introduces redundant information, increases model complexity, and weakens degradation-sensitive frequency characteristics. To address these issues, this study proposes a degradation mechanism oriented and frequency-aware framework for battery remaining capacity estimation. The main innovation of this work is threefold. First, a capacity-sensitive frequency band optimization strategy is developed to identify the degradation-relevant impedance magnitude range of 0.01–1000 Hz, enabling compact input construction while preserving dominant aging information. Second, a hybrid CNN–BiLSTM–Attention architecture is designed to jointly capture local spectral variation, long-range cross-frequency dependency, and the relative importance of different frequency regions. Third, by combining frequency compression with hybrid sequence learning, the proposed method improves prediction accuracy while reducing computational burden, making it more suitable for lightweight deployment. Experimental results show that the proposed framework outperforms CNN, CNN–LSTM, and full-spectrum input schemes, and maintains stable predictive performance under varying temperature and SOC conditions. These results demonstrate that the proposed method provides an accurate, lightweight, and physically meaningful solution for EIS-based lithium-ion battery capacity estimation.