Accurate identification of vehicle dynamic states is fundamental for ensuring safety, running stability, and ride comfort in railway operations, yet remains challenging due to the coupling of multiple excitations and inherent system uncertainties. This study introduces a physics-guided state identification framework that integrates neural networks with physical consistency constraints. A scaled vibration test bench was constructed using similarity principles to provide controlled excitations and high-fidelity response measurements, and a corresponding multi-degree-of-freedom dynamic model was established in SIMPACK. A CNN–LSTM network was trained using combined experimental and simulation data, while a Kalman-inspired module was incorporated to enforce adherence to dynamic laws. Results from both bench tests and numerical simulations demonstrate that the proposed hybrid approach achieves significantly higher accuracy and robustness than purely analytical or purely data-driven methods when applied to complex operating conditions.