Abstract: The sensor-based intelligent dry ore separation technology has attracted much attention in the field of coal recent years. Its advantages of energy saving, waterlessness, intelligence, high-precision and easeof use have a broad market prospectin the links of downhole gangue discharging, pre-enrichment in the separation plant and gangue re-selection with the advantages of cost reduction, energy saving and consumption reduction. In order to explore the classification accuracy and application potential of deep learning in the intelligent dry separation of coal based on machine vision, a multi-product coal image classification method was proposed based on deep learning. On the basis of obtaining anthracite images of low-ash coal, medium-low ash coal, high-ash coal and gangue products, the position transformation operation was used to enhance the data of the coal particle image set, providing a large amount of training data for the deep learning model. Inception_v3 neural network model based on Tensorflow and Keras framework was adopted for migration learning. The initial network parameters and weightswere trained on the Inception_v3 neural network model by using ImgNet large data sets, and were transferred to the coal multi-product classification model, and the model was processed by Dropout to prevent over-fitting, and finally established a deep learning classification model for anthracite multi-products, and used the loss value of the training process and the trend graph of accuracy acc to determine whether the training process has converged. Through the comparison and analysis of multiple classifiers, the results show that the test accuracy and verification accuracy of the deep learning classification model reach more than 90%, and slightly decreases as the number of products increases. The verification accuracies of multi-product are 93.5%, 91.3% and 90.1% respectively, which are superior to the accuracy of conventional KNN, RF and SVM classification models. Among them, the accuracy of SVM classification model was higher, followed by RF, and KNN classification accuracy was the lowest. The analysis results indicate that deep learning can effectively distinguish the feature differences between products and has great potential in the application of image classification of coal products.