<p>Accurately evaluating the health condition of battery packs deployed in electronic equipment is essential for ensuring both the safe operation and energy efficiency of such systems. A reliable SOH estimation helps prevent unexpected failures and supports optimal battery management. However, many existing methods suffer from limited generalization capabilities and high computational demands, which pose challenges for efficient implementation. Additionally, these methods are susceptible to noise in real-world signals, and this susceptibility can significantly reduce estimation accuracy and system reliability. In response to the above challenges, this work introduces an SOH estimation approach that integrates Bayesian networks with an enhanced Mamba model. The method improves robustness under operation variations and external disturbances through Bayesian uncertainty modeling, while efficient feature extraction and denoising are achieved using discrete wavelet transform (DWT) and time window segmentation. Experimental evaluations on the CALCE and A123 battery datasets demonstrate that the proposed approach effectively captures complex nonlinear degradation patterns. Moreover, its RMSE and MAE results are consistently lower than those obtained by leading existing methods. Our findings indicate that this approach delivers both broad applicability and computational efficiency for diagnosing battery health, opening up a novel pathway for robust SOH estimation in real-world applications.</p>