The construction industry is one of the most perilous sectors globally, with significantly higher fatality rates compared to other industries. Collision with heavy construction machines is a major cause of worker injuries and fatalities. This research proposes a digital twin-enabled anti-collision system to provide collision warnings and improve construction safety. A five-dimensional digital twin model encompassing the physical layer, digital data layer, virtual entity layer, service layer, and connection layer is established. The physical layer designs multi-sensor fusion approach to respectively collect image data, depth data, and motion data, achieving comprehensive perception of construction sites. The digital data layer designs multi-modal analysis methods to obtain on-site information such as category, position and motion status. The virtual entity layer integrates geometry model, physical model, behavior model, and rule model to achieve the synchronization between physical entities and digital entities. It also divides construction sites into dangerous zones, warning zones, and safe zones to perform collision warnings based on the status of construction robots and workers. The services layer provides supervision, motion states monitoring, zone prediction, and data management for the DT system. The connections layer establishes the information interconnection and synchronization between these layers. Case study is conducted to verify its performance in visualization, motion states monitoring and collision warning. It is proven that the proposed method can achieve a prediction accuracy of over 96% for dangerous zones and over 95% for warning zones. The proposed digital twin-enabled anti-collision system can greatly improve the safety for construction sites.