Detecting abnormality of battery decline for unbalanced samples via ensemble learning optimization

Aiming at the challenges of one single algorithm's limited performance on unbalanced samples and restricted analysis dimensions in battery risk detection, this paper proposes an early abnormal decline battery diagnosis method based on feature engineering and ensemble learning optimized convolutional neural network (CNN) applicable to unbalanced datasets. Initially, comprehensive dimensionless indicators (DI) are derived from the discharge voltage-capacity (V-Q) data, and the Pearson correlation coefficient (PCC) is then conducted to precisely screen out the optimal DI subset that is highly sensitive to abnormal battery decline. Subsequently, an ensemble CNN-based model for diagnosing abnormal decline batteries is constructed. By integrating the prediction results of multiple CNN models, ensemble learning can leverage the strengths of each model across different categories. It effectively balances the model's ability to recognize both minority and majority classes, thereby enhancing the model's adaptability and generalization when dealing with class-imbalanced data. Ultimately, one single CNN model is adopted as a benchmark to highlight the advantages of the ensemble CNN model in addressing the classification problem posed by class-imbalanced datasets. The proposed method is validated using a class-imbalanced Lithium Cobalt Oxide (LCO) battery dataset. The results demonstrate that the ensemble CNN-based method achieves a 100 % accuracy rate in diagnosing abnormal decline batteries.