The accurate identification of rock mixed ratio is the main technical bottleneck to realize intelligent coal caving. In this paper, the working face of a Mine is taken as the engineering background, and the image data set of coal gangue accumulation body is constructed by means of field investigation, numerical calculation and laboratory experiment. The typical semantic segmentation network model of DeepLab v3+ is trained. Numerical simulation is carried out based on the three-dimensional reconstruction block. The quantitative relationship between the projected area rock mixed ratio and the volume rock mixed ratio of coal gangue accumulation body on scraper conveyor under different top coal thickness conditions is studied, and the prediction model of volume rock mixed ratio of coal flow is established. The results show that the accuracy rate of the DeepLab v3+ model is 97.68 %, which can meet the accurate and rapid identification of the rock mixed ratio in the projection area of the coal gangue accumulation body. The determination coefficient R2 of the volume rock mixed ratio prediction model is 0.9828, and the prediction effect is good.