Increasing the efficiency of li-ion battery cycle life testing with a partial-machine learning based end of life prediction

Extensive cycle life studies are conducted in the laboratory to understand the aging behavior of a li-ion battery under various environmental and operational conditions. However, designing an optimal test plan is not straightforward and usually leads to test plans with test points providing little information gain. This paper proposes a machine learning-enhanced testing procedure, which stops the expensive cycling of a subset of tested batteries and replaces their physical test result with a machine learning-based prediction. The core component of the proposed testing procedure is a test subset selection algorithm, which selects a subset of batteries during the cycling whose aging can be predicted with low errors. This is done by extracting aging features during the initial cycling period and feeding them into the Kennard-Stone algorithm to select batteries that have a similar distribution of their aging features as batteries that are physically cycled. Based on three different aging studies and four different machine learning-based predictors, it is shown that the proposed selection approach helps to reliably find subsets that can be predicted with significantly lower prediction errors than random selections. A feature set with capacity-, internal resistance-, and incremental capacity-based features is proposed, which helps to select battery subsets with low prediction errors across all three tested aging studies. Furthermore, it is shown that a higher dissimilarity between a battery and the batteries within the training set also results in a higher prediction error for the respective battery. Therefore, a dissimilarity -based threshold is defined to effectively select the most batteries for the prediction. For the three investigated aging studies, the proposed ML-enhanced testing approach allows to stop the cycling of 30% to 75% of all batteries while ensuring EOL predictions with < 10% MAPE.

Automated summary

This paper proposes a machine learning-enhanced testing procedure that selects and predicts battery subsets to reduce testing costs and improve prediction accuracy.