Battery Capacity Knee Identification Using Unsupervised Time Series Segmentation

Capacity knees have been observed in experimental tests of commercial lithium-ion cells of various chemistry types under different operating conditions. Their occurrence can have a significant impact on safety and profitability in battery applications. To address concerns arising from possible knee occurrence in battery applications, this work proposes an algorithm to identify capacity knees as well as their onset from capacity fade curves. The proposed capacity knee identification algorithm is validated on both synthetic degradation data and experimental degradation data of two different battery chemistries, and is also benchmarked to the state-of-the-art knee identification algorithm in the literature. The results demonstrate that our proposed capacity knee identification algorithm could successfully identify capacity knees when the state-of-the-art knee identification algorithm failed. The results can contribute to a better understanding of capacity knees and the proposed capacity knee identification algorithm can be used to, for example, systematically evaluate the knee prediction performance of both model-based methods, and data-driven methods and facilitate better classification of retired automotive batteries from safety and profitability perspectives.