Electrochemical impedance spectroscopy-based deep ensemble for lithium-ion batteries remaining useful life prediction considering data imbalance and uncertainty

Lithium-ion batteries (LIBs) are widely deployed in modern life. To ensure their safe and reliable operation, it is crucial to investigate capacity degradation mechanisms and accurately predict remaining useful life (RUL). However, existing model-based and deep learning (DL) methods on RUL prediction of LIBs often employ charging and discharging curves that contain limited information about internal states, and are subject to model imprecisions, data imbalance, or ignorance of associated uncertainties. To achieve accurate and reliable RUL prediction over the entire lifetime of LIBs, a novel deep ensemble framework considering data imbalance and uncertainty is proposed in this study, which adopts electrochemical impedance spectroscopy (EIS)-a non-destructive, powerful, and informative measurement-as input. Besides, to alleviate aleatoric uncertainty and facilitate accurate prediction, a multi-source information fusion strategy is developed, which takes physical parameters indicating battery internal states and the features extracted from capacity curves as additional model inputs. Moreover, the regularization strategy that effectively utilizes current and voltage curves is explored to further boost the framework prediction performance. Experiments on the public Cambridge EIS dataset verify the effectiveness of the proposed framework and the two strategies, yielding root mean square error of 9.53, 7.84, and 3.62, respectively, which are superior to state-of-the-art methods, and three experiment dataset configurations are designed to demonstrate their generalization capability in accurately predicting RUL.