The safety of battery packs is greatly affected by individual abnormal cells. However, it is challenging to diagnose abnormal aging batteries in the early stages due to the low abnormality rate and imperceptible initial performance deviations. This paper proposes a feature engineering and deep learning (DL)-based method for abnormal aging prognosis and end-of-life (EOL) prediction. The mathematical model of dimensionless indicators (DIs) is applied to partial voltage-capacity (V-Q) of one cycle to construct the optimal DIs with high sensitivity to abnormal aging. Taking the optimal DIs and partial V-Q as inputs, an abnormal aging prognosis model is established, and an ablation study is designed to verify the necessity of the selected DIs. Finally, the gated recurrent unit (GRU) is utilized to establish EOL prediction models for normal and abnormal degradation batteries, respectively. The proposed method is verified by a Lithium Cobalt Oxide (LiCoO) battery dataset. The results indicate that the 100% prognosis of abnormal degradation batteries is achieved and the EOL prediction accuracy is less than 3.7%. This work highlights the rapid abnormal battery detection using data of one cycle without excessive battery testing, which contributes to the rational deployment of batteries and reduces the probability of failures during operation.

中文翻译：

---

特征工程与深度学习相结合的锂离子电池老化异常检测

电池组的安全性受个别电芯异常影响较大。然而，由于异常率低且初始性能偏差难以察觉，因此在早期阶段诊断异常老化电池具有挑战性。本文提出了一种基于特征工程和深度学习（DL）的方法，用于异常衰老预测和寿命终止（EOL）预测。将无量纲指标（DI）的数学模型应用于一个周期的部分电压容量（VQ），以构建对异常老化具有高敏感性的最佳DI。以最优DI和部分VQ作为输入，建立异常衰老预后模型，并设计消融研究来验证所选DI的必要性。最后，利用门控循环单元（GRU）分别建立正常和异常退化电池的EOL预测模型。所提出的方法通过锂钴氧化物（LiCoO）电池数据集进行了验证。结果表明，异常退化电池实现了100%的预测，EOL预测精度小于3.7%。这项工作突出了利用一个周期的数据快速检测电池异常情况，无需进行过多的电池测试，有助于电池的合理部署，降低运行过程中出现故障的概率。