Research On Coal Gangue Sorting Method At Mobile End Based On Deep Learning

Coal is the main energy source in China,occupying a large proportion of the energy structure.Gangue is the waste product associated with the coal mining process.The existence of gangue will reduce the heat of coal combustion,and the existence of gangue will cause environmental pollution and increase the transportation cost of raw coal,so it is necessary to separate the coal and gangue.Traditional manual gangue sorting is not only strong but also subjective and unstable;mechanical gangue sorting also has different degrees of pollution.As a new type of gangue automatic sorting equipment,machine vision gangue sorting robots have gained attention.To overcome the shortcomings of the traditional machine vision gangue sorting robot,the design of a mobile gangue sorting robot is proposed.This paper thoroughly investigates the visual gangue identification and positioning algorithm and pneumatic separation method,as well as their design and implementation work on the mobile end.The following are the contributions and work in this paper:(1)Research of lightweight convolutional neural network,model compression,and coal and gangue identification at mobile device.To address the problems of high complexity of traditional lightweight convolutional neural network models,slow identification speed at the mobile end,small sample data set training,and poor recognition effect,we propose an efficient coal and gangue identification method at the mobile end,E-Mobile Netv3.Using an efficient channel domain attention mechanism to replace the attention mechanism SENet in the Mobile Netv3 network to enhance the information interaction between networks across channels,avoid the side effects of channel dimensionality reduction,and improve the model identification accuracy.The SMU activation function and fully connected layer are used to optimize the network structure for small sample data sets,enhance the training convergence speed,and lower the complexity of the model.An experimental setup is built to train,deploy,and test the identification effect of the model on the mobile end.(2)based on the object detection network to develop research on coal and gangue identification and localization methods.For the traditional object detection algorithm with high complexity,slow inference speed,and poor recognition and localization of coal and gangue under complex conditions,a variety of methods are used to improve the network.For the recognition and localization of different sizes of coal and gangue,the network is improved by embedding a hybrid domain attention mechanism to enhance the spatial and channel feature extraction of the network for key object regions.For the detection of coal and gangue with different light intensities,the model training set data is expanded by using data augmentation,and the robustness of the network for coal and gangue identification and localization under different light conditions is improved by setting several different light enhancement coefficients.For the stacking of coal and gangue,the accuracy value of identifying multiple objects with different mixup fusion coefficients is explored.In addition,the model improves the network using a phantom network,Meat-ACON activation function,and a decoupling head.For the localization method of the network,anchor-based and anchor-free are compared and analyzed,and the anchor-free coal and gangue localization method is finally selected to output the actual location information of coal and gangue on the conveyor belt.(3)Research on dynamic control of coal and gangue separation method based on visual measurement of solenoid valve.For the study of the coal and gangue separation method at the mobile end,the pneumatic separation method is used to separate the coal and gangue after identification and positioning through the deployment of a lightweight compression target detection network at the mobile end,visual identification and positioning of coal and gangue,output the position and size information of coal and gangue on the conveyor belt,analyze and obtain the starting action position and action range of the solenoid valve,dynamically control multiple solenoid valves on and off,and inject high pressure gas to realize the separation of coal and gangue blocks.(4)Experimental verification of coal and gangue identification and separation method at mobile device.Build the mobile end of the coal gangue identification and separation device,collect coal and gangue pictures,build a coal and gangue data set,through the training of the model,verify the improved model for different sizes,different lighting conditions,with or without stacking situation of coal and gangue identification and location,and pneumatic separation of coal and gangue under over.Figure 50 Table 10 Reference 83...