Early Quality Classification and Prediction of Battery Cycle Life in Production Using Machine Learning

Abstract

An accurate determination of the product quality is one of the key challenges in lithium-ion battery (LIB) production. Since LIBs are complex, electrochemical systems, conventional quality control measures such as aging are time-intensive and costly. This paper presents the applicability of machine learning approaches for an early quality prediction and a classification of cells in production. Using inline measurement data of 29 NMC111/graphite pouch cells, linear regression models and artificial neural networks (ANNs) were compared regarding their prediction accuracy. From comprehensive electrochemical impedance spectroscopy (EIS) and cycling datasets, a total of 24 features were extracted, combined, and analyzed. The best ANN achieved a test error of 10.1% at an observation time of less than two days. For a classification into two cycle life groups, a maximum accuracy of 97% was reached. Moreover, a reliable classification of high-lifetime cells was achieved using only EIS measurements during wetting. The results highlight the great potential of data-driven models for the prediction of LIB quality in production as well as their implementation to increase the throughput and the overall cell quality.