Retired Lithium-Ion Batteries Screening via Feature Tokeniser-Transformer Considering Data Imbalance

There is an imbalance in the retired battery data, primarily because the majority of the batteries are still in usable condition, leading to a severely skewed data distribution. This imbalance can significantly impact the performance of deep learning models, causing the classification results to be biased toward the majority class. To address the above problems, we propose a novel method for screening retired lithium-ion batteries based on the Feature Tokeniser-transformer (FT-transformer) and the synthetic minority oversampling technique (SMOTE). First, time series and internal resistance features are extracted based on partial charging voltage-SOC curves and direct current pulses. Considering the imbalance of the data distribution, some samples are added using SMOTE to balance the sample distribution. Then, the FT-transformer is used for retired battery multiclassification. The proposed method has been validated on our laboratory's self-collected and MIT public datasets, demonstrating higher accuracy and stronger stability.