Based on the Cangkou cable-stayed bridge of Qingdao Bay bridge, the multi-scale finite element model was used to analyze and identify the local crack damage of the steel box girder. Firstly, a multi-scale model of steel box girder was established by using small-scale shell elements and large-scale beam elements, and the whole bridge model was modified according to actual monitoring data. Then, the small damage simulation of the vulnerable section of small-scale steel box girder was carried out, and the variation law and damage sensitivity of the section strain response were studied. Finally, the BP neural network method was used to identify the location and extent of the damage. The analysis results show that the multi-scale model makes the small damage simulation and calculation efficiency get better coordination; the strain along the direction of the bridge has better sensitivity to small damage of the steel box girder, and the strain change has significant localized characteristics; the BP network method of strain along the bridge direction can better identify the damage location and damage degree, and has strong anti-noise ability based on the reasonable strain point arrangement.