A multi-stage augmentative generalization learning prediction model for lithium-ion battery remaining useful life under uncertain working conditions

While data-driven methods have improved remaining useful life (RUL) prediction of Lithium-ion batteries (LIB), their reliance on large datasets limits performance under uncertain conditions without degradation data. This study introduces the concept of domain generalization (DG) to tackle the problem, aiming to trains models based on historical conditions that generalize well to uncertain scenarios. However, traditional DG methods often fail to the LIB RUL prediction task due to limited degradation diversity and their neglect of fine-grained degradation information. To overcome these drawbacks, this study proposes a novel multi-stage augmentative generalization learning (MSAGL) algorithm that constructs a generalized prediction model using degradation data from only one historical working condition. Specifically, addressing the lack of fine-grained degradation representation, a temporal distribution partition scheme is designed to adaptively divide the degradation process into multiple stages. Meanwhile, a diverse augmented generalization network is developed to generate varied augmented working conditions, thereby alleviating the limitations caused by insufficient historical degradation diversity. The evaluation indicates that the suggested approach yields estimation errors (RMSE) under 0.07 and 0.06 for capacity prediction tasks on the NASA and CALCE datasets, respectively, while requiring degradation data from only one historical condition.