【摘要】 准确估计锂离子电池的健康状态（state of health, SOH）是新能源汽车安全高效运行的重要保障，对电池荷电状态及剩余使用寿命估计具有重要意义。针对锂离子电池SOH估计精度低，本文提出一种基于充电电压特征的锂电池SOH混合深度学习估计模型。首先基于充电电压片段提取若干与传感器测量参数及其间接计算相关的特征，筛选出4个与健康状态密切相关的关键特征参数作为模型的输入。其次，构建卷积神经网络-长短期记忆神经网络（CNN-LSTM）混合模型，进一步基于长鼻浣熊优化算法（COA）对模型中的超参数进行优化，以提升估计模型的准确性和鲁棒性。从单体、整车多层面进行了SOH估计模型验证，并与灰狼优化算法（GWO）-CNN-LSTM、粒子群优化算法（PSO）-CNN-LSTM、CNN-LSTM 3种估计模型进行对比，COA-CNN-LSTM估计模型的精度与准确性最优，其均方根误差、平均绝对误差、平均绝对百分比误差3个误差指标均在0.80%、0.73%、0.84%以内，验证了所构建的估计模型在不同容量、不同老化路径、不同化学体系下电池的适用性与准确性。本文构建的COA-CNN-LSTM混合深度学习模型，基于间接健康特征参数与直接特征参数相结合，可以更好地表征电池性能衰退，利用部分充电电压片段获得的健康特征参数进行SOH估计时，该模型仍能具备较高的估计精度。更多还原

【Abstract】 Accurately estimating the State of Health(SOH) of lithium-ion batteries is an important guarantee for the safe and efficient operation of new energy vehicles, and it is of great significance for the estimation of battery charging state and remaining service life. For the low accuracy of SOH estimation of lithium-ion batteries, in this paper a hybrid deep learning estimation model of SOH of lithium batteries based on charging voltage features is proposed. Firstly, based on the charging voltage segments, a few features related to the sensor measurement parameters and their indirect calculation are extracted, and four key feature parameters closely related to the state of health are screened out as input to the model. Secondly, a convolutional neural network-long and short-term memory neural network(CNN-LSTM) hybrid model is constructed, and the model hyperparameters are further optimized based on the coati optimization algorithm(COA) to enhance the accuracy and robustness of the estimation model. The SOH estimation model is validated at multiple levels from single unit and whole vehicle, and compared with the three estimation models of Gray Wolf Optimization Algorithm(GWO)-CNN-LSTM, Particle Swarm Optimization Algorithm(PSO)-CNN-LSTM, CNN-LSTM, and CNN-LSTM, and the COA-CNN-LSTM estimation model is the most optimal in terms of precision and accuracy, and its root mean square error, average absolute error, and average absolute percentage error are all within 0.80%, 0.73% and 0.84%, which verifies the applicability and accuracy of the constructed estimation model in batteries with different capacities, different aging path and different chemical systems. The COA-CNN-LSTM hybrid deep learning model constructed in this paper can better characterize the battery performance decline based on the combination of indirect health feature parameters and direct feature parameters, and the model can still have high estimation accuracy when SOH estimation is performed by using the health feature parameters obtained from partial charging voltage segments.更多还原