Lithium-ion batteries (LiBs) degrade slightly until the knee onset, after which the deterioration accelerates to end of life (EOL). The knee onset, which marks the initiation of the accelerated degradation rate, is crucial in providing an early warning of the battery's performance changes. However, there is only limited literature on online knee onset identification. Furthermore, it is good to perform such identification using easily collected measurements. To solve these challenges, an online knee onset identification method is developed by exploiting the temporal information within the discharge data. First, the temporal dynamics embedded in the discharge voltage cycles from the slight degradation stage are extracted by the dynamic time warping. Second, the anomaly is exposed by Matrix Profile during subsequence similarity search. The knee onset is detected when the temporal dynamics of the new cycle exceed the control limit and the profile index indicates a change in regime. Finally, the identified knee onset is utilized to categorize the battery into long-range or short-range categories by its strong correlation with the battery's EOL cycles. With the support of the battery categorization and the training data acquired under the same statistic distribution, the proposed SOH estimation model achieves enhanced estimation results with a root mean squared error as low as 0.22%.

翻译：锂离子电池在拐点出现前会轻微退化，此后劣化加速直至寿命终结。拐点标志着加速退化速率的开始，对电池性能变化的早期预警至关重要。然而，目前关于在线拐点识别的文献十分有限。此外，利用易于采集的测量数据进行此类识别是理想选择。为解决这些挑战，本文开发了一种利用放电数据中时间信息的在线拐点识别方法。首先，通过动态时间规整提取轻微退化阶段放电电压周期中嵌入的时间动态特性；其次，利用矩阵特征在子序列相似性搜索过程中暴露异常。当新周期的时间动态特性超过控制限且剖面索引指示状态变化时，即可检测到拐点。最后，利用已识别的拐点与电池寿命周期之间的强相关性，将电池分类为长寿命或短寿命类别。在电池分类及相同统计分布下获取的训练数据支持下，所提出的健康状态估计模型实现了低至0.22%均方根误差的增强估计结果。