As a virtual representation of a specific physical asset, the digital twin has great potential for realizing the life cycle maintenance management of a dynamic system. Nevertheless, the dynamic stress concentration is generated since the state of the dynamic system changes over time. This generation of dynamic stress concentration has hindered the exploitation of the digital twin to reflect the dynamic behaviors of systems in practical engineering applications. In this context, this paper is interested in achieving real-time performance prediction of dynamic systems by developing a new digital twin framework that includes simulation data, measuring data, multi-level fusion modeling (M-LFM), visualization techniques, and fatigue analysis. To leverage its capacity, the M-LFM method combines the advantages of different surrogate models and integrates simulation and measured data, which can improve the prediction accuracy of dynamic stress concentration. A telescopic boom crane is used as an example to verify the proposed framework for stress prediction and fatigue analysis of the complex dynamic system. The results show that the M-LFM method has better performance in the computational efficiency and calculation accuracy of the stress prediction compared with the polynomial response surface method and the kriging method. In other words, the proposed framework can leverage the advantages of digital twins in a dynamic system: damage monitoring, safety assessment, and other aspects and then promote the development of digital twins in industrial fields.