The accurate estimation of state of health (SOH) for lithium-ion batteries is significant to improve the reliability and safety of batteries in operation. However, many existing studies on battery SOH estimation are conducted on the premise of large sizable labeled training data acquisition without considering the time cost and experimental cost. To solve such issues, this article proposes a novel capacity prediction method for SOH estimation based on the battery equivalent circuit model (ECM), deep learning, and transfer learning. First, an actual charge–discharge experiment is carried out, and a simulation of the corresponding cycling process is conducted for virtual data acquisition using the battery equivalent model. Second, a convolutional neural network (CNN)-based feature extraction network is selected by conducting a performance comparison. Then, a capacity estimation model consisting of a feature extraction network, regressor, and feature alignment metric calculation modules is generated. Several transfer learning methods are chosen for feature alignment metric calculation. Finally, a capacity estimation performance comparison is done for the final selection of the feature alignment metric calculation methods. The results illustrate that the capacity prediction model established using virtual data and the generative adversarial network (GAN)-based transfer learning method has ideal prediction performance (with the 0.0941 of the maximum test error in all capacity estimation situation).

中文翻译：

---

用于电池状态监测的混合电池等效电路模型、深度学习和迁移学习

准确估计锂离子电池的健康状态（SOH）对于提高电池运行的可靠性和安全性具有重要意义。然而，现有的许多关于电池SOH估计的研究都是在大量标记训练数据获取的前提下进行的，而没有考虑时间成本和实验成本。为了解决这些问题，本文提出了一种基于电池等效电路模型 (ECM)、深度学习和迁移学习的 SOH 估计容量预测新方法。首先进行实际充放电实验，利用电池等效模型模拟相应的循环过程进行虚拟数据采集。其次，通过性能比较选择基于卷积神经网络 (CNN) 的特征提取网络。然后，生成由特征提取网络、回归器和特征对齐度量计算模块组成的容量估计模型。选择了几种迁移学习方法来计算特征对齐度量。最后对最终选择的特征对齐度量计算方法进行了容量估计性能比较。结果表明，使用虚拟数据和基于生成对抗网络（GAN）的迁移学习方法建立的容量预测模型具有理想的预测性能（所有容量估计情况下的最大测试误差为0.0941）。选择了几种迁移学习方法来计算特征对齐度量。最后对最终选择的特征对齐度量计算方法进行了容量估计性能比较。结果表明，使用虚拟数据和基于生成对抗网络（GAN）的迁移学习方法建立的容量预测模型具有理想的预测性能（所有容量估计情况下的最大测试误差为0.0941）。选择了几种迁移学习方法来计算特征对齐度量。最后对最终选择的特征对齐度量计算方法进行了容量估计性能比较。结果表明，使用虚拟数据和基于生成对抗网络（GAN）的迁移学习方法建立的容量预测模型具有理想的预测性能（所有容量估计情况下的最大测试误差为0.0941）。