A progressive learning residuals based on multivariate Mamba and adaptive singular value decomposition method for remaining useful life prediction of lithium-ion batteries

The performance and remaining useful life (RUL) of lithium-ion (Li-ion) batteries, which are critical components in contemporary electronic devices, have been extensively studied in both scientific research and industry. However, existing RUL prediction models typically do not adequately address the exploration of potential correlates affecting battery capacity degradation. Therefore, this paper proposes a hybrid architectural model that integrates the strengths of Multivariate Mamba and Progressively learned residual module with multivariate adaptive singular value decomposition (MaSVD). The model establishes the relationships between variables using a Token-based approach, emphasizes the correlation among variables in Li-ion batteries, and incorporates a state-space model (SSM) via Mamba’s selection mechanism. This integration effectively captures both long-term and short-term dependencies of the batteries, effectively compensates for the inability of existing RUL prediction models to capture subtle characterization information during battery degradation. Furthermore, this paper introduces a novel modification to the attention mechanism by employing a subtraction method in place of the aggregation method typically used in Transformer architectures. This modification, combined with a step-by-step learning approach using multiple MaSVDs, mitigates the overfitting issue and reveals underlying feature relationships in battery data. By progressively learning from the residual outputs of supervisory signals, the proposed model enables multiple neural blocks to extract nuanced relationships among battery attributes. The model demonstrated in this study outperforms existing state-of-the-art technologies regarding prediction accuracy and stability. The minimum mean absolute error prediction accuracy has reached 97%, and the minimum root mean square error prediction accuracy has reached 95%.