Research On Coal Quantity Identification Method Of Belt Conveyor Based On Machine Vision And Deep Learning

In the process of coal mining,the belt conveyor is the main equipment for transporting raw coal,and it is also the main equipment that consumes electricity.After starting,it always keeps running at a constant high speed,which is prone to the phenomenon of "big horse-drawn trolley" with low load and high speed,or even no-load high speed,wasting a lot of power.Adjusting the belt speed according to real-time changes in coal volume can effectively alleviate the above problems.Therefore,the real-time and accurate detection of coal volume becomes the key to energy-saving speed regulation of belt conveyors.The traditional detection methods are affected by many factors,the detection accuracy is unstable,and considering that the speed regulation process of the belt conveyor is not step-less speed regulation,this thesis proposes a coal quantity recognition method based on machine vision and deep learning.The coal flow image is classified into categories by the coal quantity recognition model,and the belt speed is adjusted according to the coal quantity,which provides the basis for the energy-saving speed regulation of the belt conveyor.The main research work of this thesis is as follows:Aiming at the characteristics of materials transported by belt conveyors,the method of coal flow image acquisition is carried out based on machine vision.The structure of the belt conveyor and the morphological characteristics of the transported materials are analyzed.In view of the problem that the directly collected coal image is close to the background color of the belt and is easy to confuse,this thesis uses the coal flow cross-section images to make the data set;analyzes the monocular vision and binocular vision.The acquisition scheme,combined with the imaging theory of machine vision,selects the image acquisition scheme of monocular vision combined with laser line stripes to enhance the image characteristics of the surface texture of coal flow;finally,research on the selection of acquisition equipment is carried out,through the ROI setting,the influence of the image edge information on the later image processing is reduced.Aiming at the current lack of coal flow cross-section data sets,a coal flow cross-section data set with a total of 5,000 images in five categories was established.In order to make the network have a better training effect,the data augmentation strategy was used to further expand the data set,and finally the data set was expanded to 16,850,including 13,480 training data sets and 3,370 test data sets.During the acquisition of the coal flow cross-section image,median filtering was used to remove the noise of the original image.The growing laser fringe thinning algorithm refines the laser fringes;for the laser fringe fracture problem,the Cardinal spline curve is used to repair the fringes.Finally,the coal flow image processed by the above algorithm is compared with the baseline image at no load to form a coal flow cross-section image.Aiming at the problems of large training task and low recognition accuracy for directly constructing a coal quantity recognition model,a method based on transfer learning and improved Inception-Res Net-v2 network was used to build a coal quantity recognition model.Through the research on classical image recognition models and transfer learning methods,the Inception-Res Net-v2 network pre-trained on large-scale datasets is selected as the feature extractor for coal flow cross-section images.On the basis of this network,the method of feature fusion is used to improve the model.On the basis of changing the output layer of the Inception-Res Net-v2 network,feature fusion processing is performed on each layer from the Input to the Pre Aux Logits of the network,and the shallow image is realized.The fusion of features and deep image features improves the recognition accuracy of the model.The network optimization experiment and the dynamic coal quantity identification experiment are researched.Firstly,through multiple sets of network optimization experiments,the optimal parameters for model training are selected.Secondly,through multiple sets of comparative experiments,the recognition performance of the improved model is verified.In the comparison experiment with the classic model,the Top-1accuracy rate of the improved Inception-Res Net-v2 model is as high as 97.26%,and the recognition performance is the best.Finally,the trained model is embedded into the experimental platform of the belt conveying system to verify the coal quantity recognition effect under different transportation volumes and different belt speeds.The experimental results show that the average accuracy of coal quantity recognition is93.42%,the processing time of single frame image is 0.143 s,and the dynamic recognition performance is good when the belt conveyor is running at medium and low speed.