Deep Learning Based Quantity Estimation Algorithm For Vibrating Screen

Swirl machine is the key equipment in coal washing process,and its separation efficiency and stability directly affect the efficiency and benefit of coal preparation plant.However,the swirl machine is blocked in the actual production process,which leads to the interruption of the production process and brings great economic losses to the enterprise.The physical structure of the swirl machine is special,and its internal operating state cannot be directly observed,but the input and output can be observed.The existing swirl machine blockage detection is based on manual visual inspection of the amount of material discharged and combined with the amount of material input for identification.The level of automation is low.The application of image processing technology to the estimation of the amount of material discharged from the swirl machine is helpful to improve the intelligent identification of the swirl machine blockage.In this thesis,the swirl machine discharge estimation algorithm based on deep learning is studied.The main research contents are as follows :(1)Affected by light,air dust and vibrating screen vibration factors,the sieve surface image of the video monitoring system of the coal preparation plant is blurred.Aiming at the problem of loss of restored image details in existing image deblurring methods,this thesis designs a SS-Deblur GANv2 deblurring network that combines shuffle attention and switchable normalization.The network first adds a feature aggregation module after the 1×1 convolution of the generator,that is,shuffle attention,to enhance the feature information to improve the recovery quality of the detail texture part.Then,a switchable normalization method is introduced into the generator and discriminator instead of the instance normalization method to determine the appropriate normalization operation for each normalization layer of the network to improve the generalization ability and adaptability of the model.The deblurring experiments are carried out on synthetic blurred images and real blurred images.The results show that the SS-Deblur GANv2 network designed in this thesis can reduce the ringing effect of the restored image while removing the blur of the sieve surface image,and better retain the edge details of the material contour.(2)Affected by factors such as the small size of the material on the sieve surface of the vibrating screen and equipment stains,the material characteristics of the sieve surface are relatively vague.Aiming at the problem of missed detection of small objects in existing object detection algorithms,this thesis designs an object detection network MFI-YOLOv7 based on multi-dimensional feature interaction.The network first introduces omni-dimensional dynamic convolution in the backbone network to capture more global feature information to improve the feature extraction ability of convolution.Then,the CARAFE-FPN feature fusion structure is designed at the Neck layer to expand the receptive field of the model to enhance the fusion of feature information.In addition,Focal L1 Loss is integrated into the prediction layer CIo U Loss,and Focal-CIo U Loss is designed to solve the problem of class imbalance between high and low quality samples,and to measure the distance between the prediction results and the real objects more accurately.The comparative experiments of different algorithms for different scenes to be detected show that the MFI-YOLOv7 algorithm designed in this thesis can more accurately estimate the amount of material on the sieve surface and effectively reduce the missed detection of small objects.This thesis has 33 figures,12 tables and 87 references.