Unlocking Unlabeled Battery Data

Abstract

The transport sector is expanding rapidly and with it the use of lithium-ion batteries. Accurately estimating a battery’s State of Health (SoH) is crucial to ensure optimal usage and in turn optimize both the safety and the environmental impact of the battery. Traditional machine learning relies on labeled data, which is limited in the field of SoH estimation. Furthermore, the field holds a significant gap between laboratory and field data. This project aims to investigate how self-supervised and semisupervised machine learning methods can be used for SoH estimation. This is done through upstream self-supervised learning of unlabeled data using autoencoders, and downstream supervised estimation of SoH. Furthermore, a multi-chemistry battery dataset containing both field and laboratory data was curated. To further bridge the gap between field and laboratory data a fine-tuning task was performed. The proposed semi-supervised hybrid architecture consists of a convolutional autoencoder combined with an LSTM head. This resulted in SoH predictions with an MAE of 0.0196 and an RMSE of 0.0313. The amount of labeled training data could be significantly reduced whilst maintaining accurate results. The model was successfully fine-tuned to produce solid early predictions for SoH on the field data. This project provides a machine learning pipeline which both lowers the need for labeled data and helps bridge the laboratory-field gap in SoH estimation. Future work with a focus on data engineering and acquiring higher quality field data could further build on these results.