A hybrid deep learning model for lithium-ion battery state-of-health estimation using electrochemical impedance spectroscopy

State of health (SOH) of lithium-ion batteries (LIBs) is a significant indicator of performance, and it is being adopted by new energy companies with increasing frequency. The precise estimation of SOH serves as a critical prerequisite for guaranteeing operational safety and reliability, while facilitating effective battery management strategies. Nevertheless, the real-world operating conditions of LIBs are both complex and ever-changing. This poses a significant challenge for SOH estimation methods developed based on laboratory conditions, hindering their practical application. Hereby, we propose an efficient end-to-end convolutional neural network (CNN) and an attention (ATT) mechanism model, CNN-ATT, for SOH estimation in LIBs using electrochemical impedance spectroscopy (EIS) data. The model integrates a CNN for local feature extraction and an ATT to capture long-term dependencies, forming a multiview information perceptron (MIP). Validated with experimental data from the university of cambridge, the CNN-ATT model achieves high accuracy, with RMSE, MAE, and MAPE below 0.02 and an average R 2 exceeding 0.9. Results demonstrate robust performance across varying temperatures, outperforming traditional CNN and CNN-LSTM methods. The study highlights the significance of different EIS data for SOH estimation, offering a reliable, non-invasive solution for battery health monitoring in real-world applications. This study proposes CNN-ATT, a novel deep neural network integrating CNNs and a Multi-Head Attention mechanism for lithium-ion battery SOH estimation. It establishes a direct mapping from EIS data to battery capacity, theoretically overcoming limitations associated with discharge-dependent methods by relying solely on EIS. The study validates that using comprehensive EIS features (real part, imaginary part, modulus, phase angle) instead of just the real and imaginary parts improves accuracy, reducing all error metrics. Furthermore, the proposed model demonstrates over 30 % error reduction and enhanced robustness across varying temperatures compared to traditional models, indicating a promising solution adaptable to diverse conditions and operating scenarios. • Unlike traditional SOH estimation, this method is independent of vehicle operating conditions. • The hybrid CNN-attention model enables accurate SOH estimation under varying temperatures. • The proposed model reduces SOH estimation errors by over 30% compared with existing models. • A parallel architecture simplifies the model while improving accuracy over traditional RNNs.