Abstract: To address the limitations of inaccurate empirical formulas, the difficulty of accurately identifying excitation sources through numerical simulation, and the inability to reliably predict high-frequency vibrations during aircraft development, this study proposes a machine learning approach based on existing flight data for vibration prediction. To account for variations in flight speed and propulsion types, multiple flight parameters and state variables related to aerodynamic and propulsion-induced excitation sources were integrated into the training of the vibration prediction model. Vibration measurements from different locations on the aircraft were used as target variables to capture the influence of structural configuration. The input parameters were tuned according to prediction errors to optimize the multi-physical-parameter prediction model. Validation with a subset of the measurement data showed that the prediction accuracy for vibrations in different directions exceeded that of conventional empirical formula methods by more than a factor of six, demonstrating the potential of this approach for further research and practical application.