Based on long-term monitoring data in Nanjing No.3 Yangtze River Bridge, a long-term trend assessment and dynamic early warning method on the girder deflection of the steel cable-stayed bridge with statistical theory was proposed. First of all, the variation characteristics of the deflection with temperature were analyzed. The data of 0:00-1:00 at night was selected. The influence of vehicle load was removed, but the integral temperature effect was considered. The reference value of the deflection due to dead load was obtained. Secondly, the assessment was implemented by the linear moderate model and grey correlation analysis according to actual fact. Taking the data from 2007 to 2010 as an example, the results show that the established methods are consistent with long-term trend and helpful for the further inspection of bridge defects. Finally, two-stage (yellow and red) early warning line was set up based on a certain guarantee rate. Considering the influence of traffic volume variance and cumulative damage, the dynamic warning line was established. Two samples of period before and after the snowstorm in 2008 and traffic volume growth were taken for verification. The results show that the proposed dynamic warning line not only can be effectively applied to the daily operation or emergency situation, but also achieve synchronous update according to the actual condition of the bridge.