Abstract: In northern China, typical residential buildings are predominantly single-story masonry structures characterized by a multi-bay single-depth layout with numerous door and window openings in the longitudinal walls. Masonry dwellings are widely distributed across suburban areas in northern China owing to their simple construction techniques and low cost. During urban renewal processes, these masonry structures often undergo renovation and adaptive reuse. Owing to differences in the loading history, material degradation, or retrofitting interventions, the inner and outer surfaces of their walls may exhibit inconsistent crack patterns and propagation paths that significantly increases the difficulty of damage identification and assessment. The complexity of the material properties, constitutive relationships, and geometric configurations of masonry structures renders the seismic performance analysis process cumbersome and complicated. Moreover, the existing methods fail to effectively identify and track the dynamic evolution of cracks in these structures. Because cracks serve as critical indicators for the damage assessment of masonry structures, this limitation introduces considerable errors in post-earthquake evaluations of masonry buildings. In terms of damage simulation for masonry structures, although traditional finite element programs can predict structural damage evolution to a certain extent, they often rely on static finite element models. Moreover, they lack a data exchange mechanism with the physical structure that limits their ability to predict the performance state of the structure throughout its life cycle; therefore, tracking the dynamic evolution of cracks is impossible. A digital twin, characterized by real-time data integration, dynamic updating, and high-fidelity simulation capabilities, offers transformative potential for structural health monitoring. By correlating sensor monitoring data with structural models and processing the monitoring data based on multisource data fusion algorithms, the accuracy and efficiency of structural health state assessments can be improved. To improve the efficiency and accuracy of seismic damage assessments, this study proposes a dynamic crack detection method for masonry structures based on digital twins and deep learning. First, to accurately capture the nonlinear mechanical behavior of the structure, a dynamically updatable digital twin model of the masonry structure was established by integrating structural monitoring data with refined finite element modeling. Based on shaking table tests, the authentic structural response and damage evolution data of the masonry structure model were collected, providing precise data references for the calibration and validation of the digital twin model. Subsequently, by comparing the crack evolution observed in the masonry structure during seismic wave excitation with the damage progression simulated by the digital twin model, the deep learning model, Retcrack, was employed to detect and segment cracks corresponding to specific damage states, thereby circumventing the inaccuracy caused by crack closure under dynamic loading that hindered precise measurements. The results demonstrated that the model could accurately simulate the structural response and damage states under seismic actions. Through a comparative analysis of crack evolution in masonry structures subjected to seismic waves using both the digital twin models and shaking table tests, the proposed method effectively tracked crack development patterns, achieving an accuracy of 0.962 and a mean intersection over union (mIoU) of 0.577 for dynamic crack tracking and detection in masonry structures.