Abstract: As the core equipment of industrial systems, the accuracy of the dynamic model of rotating machinery directly affects the reliability of fault diagnosis and health management. To address the issues of parameter mismatch and large modeling errors in traditional physics-based modeling under practical operating conditions, a bidirectional model correction method combining “physical constraints and data-driven approaches” is proposed, which is based on a Multi-Layer Perceptron (MLP). First, the Sobol sensitivity analysis method is employed to select key correction parameters, and Poisson disk sampling is used to achieve uniform sampling in the high-dimensional parameter space, thereby generating datasets of structural and bearing parameters. Subsequently, an MLP neural network is constructed to perform joint optimization of these parameters, enabling high-precision mapping between the dynamic model and the measured responses. Compared with the traditional Response Surface Method (RSM), the proposed method demonstrates clear advantages in terms of accuracy and generalization capability. Taking a typical dual-disc rotor system as the study object, both simulation and experimental validations are carried out. The results indicate that the proposed method can significantly reduce the discrepancies between the finite element model and the actual system in key dynamic indicators such as critical speed, while also accurately reproducing the resonance responses and imbalance dynamics observed in the test bench. The corrected model achieves a critical speed identification error of less than 2%. This method features automation, high accuracy, and strong engineering adaptability, providing effective technical support for digital twin modeling and intelligent diagnosis of rotating machinery, and showing promising prospects for practical engineering applications.