<p>Accurately assessing the state of health of lithium-ion batteries is crucial for ensuring the safe and stable operation of electric transportation systems and clean-energy storage platforms. Current research primarily concentrates on time-domain representations, whereas frequency-domain information, including inherent periodicity and disturbance structures, is rarely incorporated in a systematic manner. Consequently, model stability suffers when processing complex signals that contain oscillations or structural disturbances. The Mamba3B framework integrates a Bayesian gating mechanism with a three-branch heterogeneous processing strategy, forming a unified model for robust representation learning. The method first applies sliding-window partitioning and the fast Fourier transform to reduce noise and extract the frequency-domain characteristics of the time series. The extracted features are then divided into three processing branches, each modeling the features from both time-domain and frequency-domain perspectives. Finally, Bayesian-gate fusion integrates the outputs from the multiple branches and provides uncertainty estimates for the predictions. Experiments conducted on the CALCE and A123 public battery datasets demonstrate that this model outperforms existing methods in both RMSE and MAE metrics, significantly enhancing the accuracy and stability of State of Health estimation. This research offers novel insights for the further development of lithium-ion battery state-of-health estimation methods.</p>