A component dynamic modeling based bi-level optimization strategy, called the global stage and the local stage, is proposed to improve the iteration speed of the dynamic optimization of large-scale mechanical structures. Node information and element information, corresponding to the overall dimension and cross-section size, are isolated into two different design spaces. At the global stage，the nodal positions and the dynamic performance of the overall structure are optimized and the optimized nodal positions are transferred to the local stage. At the local stage the sectional sizes and the dynamic characteristics of the component or substructure are optimized and the optimized sectional sizes are transferred to the global stage. The optimization alternates between the two stages until the final convergence achieved. The component modal synthesis method is employed to determine the objective of the local stage optimization and to measure its effect on the dynamic performance of the overall structure. The dynamic optimization of the mental structure of a practical large-span escalator proves the feasibility and efficiency of the proposed method.