State estimation of lithium-ion batteries via electrochemical features and self-attention networks

The State of Health (SOH) is a crucial metric for evaluating the performance and reliability of batteries. Accurate SOH estimation is vital for implementing fault early warning, optimizing charge-discharge strategies, extending battery life, and reducing system risks. This paper proposes a machine learning approach based on Self-Attention Networks (SAN) to estimate the SOH of lithium-ion batteries using laboratory-measured aging datasets. Aging experiments were designed under various operating conditions to obtain datasets from eight batteries, including voltage, current, charge-discharge capacity, and impedance. First, raw data such as voltage, current, and impedance are preprocessed and used for feature construction to generate time-series sequences. Regarding the network architecture, a self-attention mechanism is introduced to capture long-term dependencies, combined with a convolutional module to enhance local feature extraction capabilities. Furthermore, a cross-cycle attention submodule is designed to highlight data segments that contribute most significantly to the degradation state across different cycles, thereby enhancing the model’s sensitivity to key aging stages. Comparative analysis with LSTM and other methods demonstrates that the proposed method achieves higher accuracy than traditional approaches, yielding a Root Mean Square Error (RMSE) of less than 0.0033 across different test sets. Proposed method effectively mitigates the issue of information dilution in long-sequence prediction often found in traditional methods and strengthens representation capability the multi-stage aging behavior of batteries.