Comparison among Machine Learning Models Applied in Lithiumion Battery Internal Short Circuit Detection

Abstract

The world is experimenting a decarbonization process, mainly through lithium-ion-based solutions. Nonetheless, catastrophic events have negatively affected the social acceptance of lithium-ion-based solutions. One of the most interesting projects regarding catastrophic event prevention is the internal short-circuit detection. This paper proposes to detect it using different machine-learning algorithms such as random forest and combination of random forest with neural network-based algorithms through time-instant classification and historical feature classification. The hyper-parameters have been optimized through grid-search. The selected algorithms have been trained thanks to synthetically generated data using a first-order electrical equivalent circuit model. The performance of the generated models has been verified and compared thanks to testing and validation data sets taken from the synthetically generated data. Afterward, the most accurate internal short circuit detection algorithm was selected and validated through laboratory-level data. The selected cell in this study is SLPB526495HE, a pouch cell of 3.7Ah. The generated data are time series of voltage and current, which are the variables that will be available in a real application. The results demonstrate an accuracy above 90% in detecting an internal short circuit in the most interesting cases. The validation with laboratory data has shown that an accuracy of 90% can be achieved. This paper provides learned lessons on the process of developing the internal short circuit detection machine-learning model, highlighting the potential they possess to detect accurately internal short circuits.