Adaptive stage segmentation and model selection for data-driven online remaining useful life prediction of lithium-ion batteries

Le résumé fourni par la source

Abstract Capacity degradation of lithium-ion batteries usually exhibits multi-stage characteristics. Existing studies frequently adopt change-point detection for stage segmentation, yet most ignore the degradation mechanism connections across stages, making it hard to dynamically match models with stage-specific degradation patterns. To address this issue, this paper proposes an adaptive degradation stage segmentation and model selection framework for online remaining useful life (RUL) prediction of lithium-ion batteries. First, a likelihood function is constructed based on the evolution of capacity degradation data, and two‐stage segmentation is realized by locating the extreme point of the likelihood evolution curve. Second, by fusing historical change-point detection results and using a likelihood-based quantitative metric, the optimal model is adaptively assigned to each stage, achieving dynamic alignment between models and degradation patterns. Then, an extended Kalman filter—based fusion method is proposed to estimate capacity degradation trajectories by combining real-time observations with prior model information. Integrated with Bayesian parameter updating, the framework enables accurate online RUL prediction. Finally, the proposed method is verified on public datasets including the MIT-Stanford battery dataset and the CALCE dataset. Experimental results demonstrate that the proposed approach outperforms comparative models in terms of RUL prediction accuracy and reliability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.