A Comprehensive Investigation on Data Redundancy of Large Battery Degradation Datasets

Machine learning (ML) methods have demonstrated superiority in state of health (SOH) estimation of lithium ion batteries. Previous work has focused on collecting extensive data to train ML models; however, the richness of the information within these datasets remains unknown. This study systematically investigates redundancy in three datasets to quantitatively assess their information richness. We propose an iterative filtering (IF) method to remove redundant data and evaluate model performance after training on various sizes of the filtered data. The IF method proves more effective than random filtering, which randomly selects a proportion of data from the dataset. Results suggest that the root mean squared error (RMSE) increases by only 10% after removing up to 95% of the training data. The removed data results in lower errors than the test data split from the original dataset, further demonstrating that the removed data lacks new ageing information. Additionally, we demonstrate the generalizability of the proposed data filtering method by training one ML model on data filtered by another ML model. This work highlights that information richness is more important than dataset volume.