<p>Maglev train levitation systems operating under prolonged and complex disturbances often experience parameter variations, which can significantly impact the accuracy of system response prediction, parameter identification, and control. To address these challenges, this study proposes a Physics-Informed Machine Learning (PIML) framework, integrating Physics-Informed Neural Networks (PINNs) and Bayesian Physics-Informed Neural Networks (B-PINNs), for robust inverse problem solving under noisy measurement conditions. While PINNs are effective for scenarios with minor disturbances, such as sensor noise, B-PINNs extend this capability to handle larger disturbances, such as external disturbance forces, while providing uncertainty quantification for inverse problem solutions. Unlike conventional methods like least squares, which rely on precise mathematical models, our approach leverages the integration of physical laws with data-driven techniques. Starting with PINNs, which combine physical principles with neural networks for inverse problem solving, we advance to B-PINNs, which incorporate Bayesian inference to quantify uncertainties and enhance robustness. The proposed B-PINNs framework successfully resolves inverse problems using sparse and noisy data, simultaneously estimating system parameters and reconstructing system states with quantified uncertainties. Numerical simulations demonstrate the superior ability of the proposed framework to handle sensor noise and disturbance force while maintaining physical consistency. This work highlights the potential of PIML for addressing complex inverse problems in maglev train levitation systems under real-world noisy and uncertain conditions.</p>