Aimed at the problems of high labor intensity, low efficiency and high safety risk of manual coal caving in the process of top coal caving mining, this paper proposes a method of coal caving state control based on U-Net. This method is based on image segmentation technology, and provides accurate coal caving state control signal for the realization of intelligent control function of top coal caving mining by studying the state of coal caving falling to scraper conveyor. Firstly, the foreground segmentation of the coal flow area in the moving state in the video of coal flow on the rear scraper conveyor is completed through the mixed Gaussian model, and the gangue area in the foreground segmented coal flow image is semantically segmented; Then, compare the pixel area of the foreground segmentation part and the semantic segmentation part to obtain the percentage of gangue on the two-dimensional image in the coal flow video stream; Finally, based on the workers’ experience, the images containing gangue are classified into two categories, and the threshold value of closing the door is calculated from the proportion of gangue corresponding to the images containing gangue. This method is simple, reliable and robust. Taking the 81309 working face of Shendong Baode Coal Mine as an example, the effectiveness of this method is verified. The model training is carried out with 300 fully-mechanized coal gangue mixed images as the data set, and the gangue target area segmentation test is carried out with 50 images as the test set. The accuracy is 96.04%; According to the workers’ experience, 300 images containing waste are classified into two categories and the threshold value of “closing the door after seeing the waste” is calculated, which provides a reliable door closing control signal for the realization of the intelligent control function of the top coal caving mining. The experimental results show that the intelligent control function of top coal caving mining based on this method can replace the manual coal caving of workers. Therefore, the application of this method improves the efficiency of top coal caving mining, ensures the safety of mine production, and contributes to high-quality unmanned top coal caving mining.