This study develops a methodology by capturing both the battery aging state
and degradation rate for improved life prediction performance. The aging state
is indicated by six physical features of an equivalent circuit model that are
extracted from the voltage relaxation data. And the degradation rate is
captured by two features extracted from the differences between the voltage
relaxation curves within a moving window (for life prediction), or the
differences between the capacity vs. voltage curves at different cycles (for
life classification). Two machine learning models, which are constructed based
on Gaussian Processes, are used to describe the relationships between these
physical features and battery lifetimes for the life prediction and
classification, respectively. The methodology is validated with the aging data
of 74 battery cells of three different types. Experimental results show that
based on only 3-12 minutes' sampling data, the method with novel features
predicts accurate battery lifetimes, with the prediction accuracy improved by
up to 67.09% compared with the benchmark method. And the batteries are
classified into three groups (long, medium, and short) with an overall accuracy
larger than 90% based on only two adjacent cycles' information, enabling the
highly efficient regrouping of retired batteries.

通过捕捉电池的老化状态和降解速率，本研究提出了一种方法来提高电池寿命的预测性能。该方法通过从电压松弛数据中提取等效电路模型的六个物理特征来指示老化状态，并通过从移动窗口内电压松弛曲线的差异或不同周期的容量 - 电压曲线的差异提取两个特征来捕捉降解速率。基于高斯过程构建了两个机器学习模型，分别用于描述这些物理特征与电池寿命之间的关系，用于寿命预测和分类。该方法在三种不同类型的 74 个电池单元的老化数据上进行了验证。实验结果显示，仅使用 3-12 分钟的采样数据，具有新特征的方法可以准确预测电池寿命，与基准方法相比，预测准确率提高了 67.09%。仅使用两个相邻周期的信息，电池被分为三组（长寿、中寿、短寿），整体准确率大于 90%，实现了退役电池的高效重新分组。