Research And Implementation Of Coal And Gangue Identification Based On Deep Learning

The identification of coal rock interface and coal gangue mixing degree in fully mechanized coal caving is the key to the breakthrough of intelligent coal mining technology,which is of great importance to improve the intelligence degree of coal face,production efficiency and recovery rate.Therefore,the study of automatic accurate identification of coal and rock has always been the focus and hotspot of research in the field of coal mining technology.It is related to the further improvement of the level of coal mining technology and the guarantee of safe and green mining in coal mines.At present,there is no accurate and efficient non-contact automatic identification equipment for coal and rock in China.For coal gangue recognition in the visible light leak and problem of fault detection and low recognition rate,and the high radiation,visible light exists problem of high cost and low recognition rate,this topic research design implementation of detection algorithm to accurately identify the coal gangue,facilitate the automatic and intelligent level of coal waste sorting identification.In the visible light,this thesis proposes the coal and garbage recognition network structure of semantic segmentation deep learning network.Upsampling operation is repeatedly used in U-NET decoding part to effectively integrate the low-level feature information into the high-level information and use fewer parameters to achieve efficient information fusion.At the same time,we used Photoshop for image annotation,which took only 1/2 of the time of traditional annotation tools.Finally,the robustness of the model is further improved by data enhancement of coal and gangue samples.The results show that our model can cope well with the cases of missed detection and wrong detection under the premise of improving the accuracy.In the invisible light,this thesis firstly uses the terahertz technology to image the coal and gangue.The image processing technology is used to effectively partition the terahertz coal and gangue video.The image classification network Efficient classification of the terahertz coal and gangue image,the experiment shows that the recognition rate of the terahertz coal and gangue image is up to 100%.And the combination of image segmentation and image processing technology realizes the real-time segmentation and recognition of terahertz coal-gangue video.The research work of this thesis is mainly divided into the following contents:(1)Under visible light,the improved semantic segmentation network is used to identify coal and gangue.Firstly,a data set of coal and gangue samples was established,and the randomly placed coal and gangue samples were collected by a high-definition camera.Photoshop image processing software was used to label the data set and enhance the data,so as to build a semantic segmentation database of coal and gangue.Then inspired by U-net and U-Net++,U-Net++ decoding part combines the mid-tier feature is obtained by peer and low-level features fusion,we will be the middle tier features directly using the lower sampling operations on many times,to get more information and retain more of the original features,at the same time reduces the quantity and information redundancy;Finally,the semantic segmentation model is trained.In addition to setting the same and reasonable training parameters for all the models,in order to prove the robustness of the model in this thesis,we use three different learning rates to conduct training tests.(2)Terahertz technology is used to identify coal and gangue under invisible light.Firstly,the data set of coal and gangue samples was created,and the linear scanning imaging of the coal and gangue samples was performed by the high-speed linear array terahertz imaging system.The data set of coal and gangue image classification was constructed by taking random and multiple screenshots of the imaging video and enhancing the data results.Then,the terahertz coal-and-gangue recognition algorithm was constructed.On the one hand,the traditional image processing method was used to obtain the coal-and-gangue image segmentation video by threshold segmentation and closing operation,and on the other hand,the lightweight image classification network was constructed to efficient Net-b0.Then,the image classification network training,the existing popular image classification network using their pre-training weight and the same parameter Settings for training.Finally,the image processing algorithm is combined with the image classification network to realize the real-time segmentation and recognition of terahertz coal-gangue video.