This paper proposes a novel on-line rotor system condition monitoring approach using nonlinear data-driven modelling and model frequency analysis. First, the dynamic process model of the vibration transmission path between the vibration measurement points of two fulcrum structures is established by utilizing nonlinear data-driven modelling. Then, the unique frequency properties are extracted from the established model to reveal, in real time, the health condition of the rotor system. Finally, using the frequency properties as features, the unsupervised learning technology is applied to the on-line monitoring of the rotor system. Compared to conventional condition monitoring methods, the proposed approach can output an early warning 26 min before a shaft fracture occurs, without generating false alarms. Consequently, this approach can greatly enhance diagnostic accuracy, demonstrating its potential to contribute to the advancement of rotor system condition monitoring techniques.