The traditional automobile assembly process has problems such as low efficiency, poor assembly quality, and difficulty in real-time monitoring and adjustment. To overcome these challenges, this study constructs a digital twin-driven virtual assembly system for intelligent automobile manufacturing. In the system perception layer, physical entity data such as equipment operating status, component location, and assembly force are collected. The data transmission layer transfers the data collected by the perception layer to the virtual model layer securely and quickly. The virtual model layer uses Blender to digitize the automobile design drawings and build a preliminary three-dimensional model of the automobile product based on the process specifications. Then, Gazebo is used to assign physical properties to the 3D model and simulate the assembly process to form a digital twin model that contains geometric shapes, physical properties, and assembly process information. At the application layer, the system provides functions such as assembly process monitoring, process optimization, and quality prediction. Through interaction with the virtual model layer, real-time visual monitoring of the assembly process is achieved. When the actual assembly data is fed back to the virtual model, the assembly process is dynamically adjusted and optimized. Among all 15 assembly tasks, the average assembly time of the virtual assembly system is lower than that of the traditional assembly method, and the average qualified rate of assembled products has increased by about 7.75%. These data show that the virtual assembly system can improve the level of intelligent automobile manufacturing, achieve high efficiency, precision and coordination in the assembly process, and provide a new solution for intelligent automobile manufacturing.