The gradation and fineness modulus of manufactured sand are important quality indicators in industrial sand production. To solve the problem that the traditional method of grading detection of manufactured sand cannot be implemented online under the actual working condition. This article combines experimental research to propose an online detection method for manufactured sand grading based on deep learning. Firstly, the images of the stacked manufactured sand on the conveyor belt are collected. Then, the manufactured sand images are segmented by the convolution neural network ( CNN) . Finally, the gradation and fineness modulus of manufactured sand is computerized online by the image processing technology. Comparative experimental results show that the mask R-CNN instance segmentation model can effectively segment intact particles in the manufactured sand stacking scenario. The equivalent elliptical Feret short diameter is used as the equivalent particle size parameter and the area gradation is used as the gradation characterization parameter. The maximum repeatability error values of online detection of two groups of fineness modulus manufactured sand are 0. 03 and 0. 05. The maximum repeatability error values of particle size interval are 2. 97% and 3. 43% . Compared with traditional detection methods, this method is feasible and can meet the requirements of online detection in industrial sand production.