Abstract
<jats:p>During lengthy service, lithium-ion batteries inevitably suffer performance decline; for this reason, a precise estimation of their state of health(SOH) becomes essential to the safe running of battery management systems. To address challenges including complex equivalent circuit model structures, difficult parameter identification, the lack of physical meaning in data-driven features, and insufficient generalization across multiple temperature conditions, this paper presents a SOH estimation method integrating electrochemical impedance modeling and deep learning. Based on the aging mechanisms of lithium-ion batteries, this work develops an improved equivalent circuit model with a streamlined structure. The model fits electrochemical impedance spectroscopy (EIS) data with high accuracy, which raises the success rate of parameter identification and improves computational efficiency. Two impedance features most strongly correlated with state of health (SOH) are selected from the model parameters using Pearson correlation analysis, removing redundant information and strengthening feature interpretability. A hybrid deep network architecture that combines a convolutional neural network (CNN) with a bidirectional long short-term memory (BiLSTM) network is then designed. This architecture extracts local spatial features from the impedance parameters and captures long-range temporal dependencies, enabling high-precision and robust SOH estimation under complex operating conditions. Experimental results show that under multiple temperature conditions, the proposed method achieves a mean absolute percentage error below 0.94%. Compared with a direct data-driven approach that does not include impedance modeling, the estimation error decreases by 20.34%, 16.47%, and 13.16% at 25°C, 35°C, and 45°C, respectively. These findings demonstrate that the proposed method features easy implementation of parameter identification, short data measurement time, high SOH estimation accuracy, and good temperature adaptability.在长时间的使用中，锂离子电池不可避免地会出现性能下降；因此，对电池健康状态（SOH）的精确估计对于电池管理系统的安全运行至关重要。为了解决等效电路模型结构复杂、参数识别困难、数据驱动特征缺乏物理意义以及在多种温度条件下泛化不足等挑战，本文提出了一种集成电化学阻抗建模和深度学习的SOH估计方法。基于锂离子电池的老化机理，本文提出了一种改进的流线型结构等效电路模型。该模型对电化学阻抗谱（EIS）数据拟合精度较高，提高了参数辨识的成功率，提高了计算效率。通过Pearson相关分析，从模型参数中选择两个与健康状态（SOH）相关性最强的阻抗特征，去除冗余信息，增强特征可解释性。然后设计了一种将卷积神经网络（CNN）与双向长短期记忆（BiLSTM）网络相结合的混合深度网络架构。该架构从阻抗参数中提取局部空间特征，并捕获长期时间依赖性，从而在复杂操作条件下实现高精度和鲁棒的SOH估计。实验结果表明，在多种温度条件下，该方法的平均绝对百分比误差在0.94%以下。与不包含阻抗建模的直接数据驱动方法相比，在25°C、35°C和45°C时，估计误差分别降低了20.34%、16.47%和13.16%。结果表明，该方法参数辨识容易实现，数据测量时间短，SOH估计精度高，具有较好的温度适应性。</jats:p>