Bidirectional Mamba network with multi-scale feature fusion and sparse-channel mixture of experts for battery state of charge estimation

The accurate end-to-end state of charge (SOC) estimation for lithium-ion batteries via deep learning remains a critical focus of current research. However, mainstream models face three major challenges: the architectural inability to simultaneously capture local and global features of input data, the frequent neglect of the significant impact of reverse temporal information on estimation, and the difficulty in balancing computational efficiency and performance during inference. To address these issues, this study proposes a new end to end estimation model, termed CMFE-SBM-SCMoE, which fully utilizes bidirectional temporal information, integrates local and global features, and avoids excessive computational overhead. Specifically, the Channel Attention-based Multi-scale Feature Extraction (CMFE) module extracts local spatial multi-scale features based on multi-layer convolution and channel attention mechanisms; the Spatial Attention-based Bidirectional Mamba (SBM) module captures long-term dependencies through parameter-shared bidirectional Mamba blocks and enhances global feature correlation via spatial attention; the Sparse-Channel Mixture of Experts (SCMoE) module activates a small number of channel features and expert networks through a sparse activation mechanism, reducing computational cost while maintaining accuracy. Experiments using public lithium-ion battery datasets under complex conditions show that under different operating conditions, the average maximum absolute error does not exceed 2.65 %, and the model significantly outperforms mainstream and state-of-the-art models in SOC estimation accuracy across almost all temperatures and working conditions, effectively achieving a balance between performance and inference efficiency.