Research On Recognition Of Coal And Gangue Based On Multispectral Imaging And Deep Learning

Coal is known as "the food of industry" and plays an important role in energy,chemical,and other fields.With the continuous development of science and technology,intelligence in the mining industry is the inevitable trend of development The intelligent separation of coal and gangue can promote the intelligent process in the mining field and the precise identification of coal and gangue is an important prerequisite for intelligent separation.There are some deficiencies in the traditional method of coal-gangue identification(consumption of water resources,radiation,interference by the environment,etc.),which can not meet the requirements of intelligent mine for the precise,efficient and green selection.Multispectral imaging(MSI),as an efficient non-destructive testing method,is widely used in agriculture,food science,biomedicine,and other fields.In this paper,MSI is used to accurately identify coal and gangue,focusing on the accurate recognition of coal and gangue,image information recognition,spectral information recognition,heterogeneous fusion recognition,and anti-interference,and so on,it researches the identification model for coal and gangue.The main research contents of this thesis are summarized as follows:(1)It studies the problem of wavelength selection for multispectral recognition of coal and gangue.Using the multispectral image information of coal and gangue in the range of 675-975 nm(25 wavelengths),the image features of coal and gangue are extracted based on the HOG,LBP and Haar.Finally,the identification model of coal and gangue is constructed by feeding into grid search support vector machine(GS-SVM),genetic algorithm SVM(GA-SVM),and particle swarm optimization SVM(PSO-SVM)classifier,which is used to select the best wavelength for identification,and the combination strategy of feature extraction and classifier is determined.The experiment shows that the image information of the ninth wavelength(773.776 nm)of the multispectral data of coal and gangue can achieve better classification results,and the highest recognition accuracy of the test set(96.25%)and training set(99.375%)can be obtained using the combination strategy of LBP and GS-SVM,at this time,C=8,g=0.17678.(2)It studies the problem of the recognition of coal and gangue based on the image information from multispectral data.Using the multispectral image information of coal and gangue at 773.776nm,based on the Plain network,Inception network and Residual network,and with the help of SGD,Adam,Adamax and Nadam optimizer,the identification model of coal and gangue is constructed to determine the optimal structure and parameters of the two-dimensional convolutional neural network(2D-CNN).At the same time,considering that the principal component analysis network(PCANet)is a simplified deep learning model based on CNN,the two-stage PCANet model is used to construct the identification model and compared with the 2D-CNN recognition model.The experiment shows that using the 2D-CNN model with Adamax as the optimizer and two two-dimensional Residual units can maximize the average recognition rate(98.75%)and minimize the average loss(0.0328).The recognition model of coal and gangue based on the two-stage PCANet can also achieve an accuracy of 98.75%.At this time,the model parameters are:k1=k2=3,L1=L2=2,the histogram block size is set to[8 12],and the block overlap ratio is set to 0.(3)It studies the problem of the recognition for coal and gangue based on the spectral information from multispectral data.Using the multispectral spectral information of coal and gangue at all 25 wavelengths,based on the two common one-dimensional convolutional neural network(1D CNN)structures,and with the help of SGD,Adam,Adamax and Nadam optimizer,the identification model of coal and gangue is constructed by using ReLU and its improved function as activation function,which is used to determine the optimal structure and parameters of the 1D-CNN.The experiment shows that the model can maximize the average recognition rate(98.75%)and minimize the average loss(0.0382)by using Nadam as the optimizer for 1D-CNN-B,which contains three one-dimensional convolution units B and using PReLU as the activation function.(4)It studies the problem of heterogeneous fusion recognition for coal and gangue based on the image information and spectral information from multispectral data Using the image information of coal and gangue at the ninth wavelength and the spectral information at all 25 wavelengths,based on the basic structure of the heterogeneous fusion network,ReLU,PReLU,and its combination are used as the activation functions,and the identification model is constructed by Adamax and Nadam optimizer,the identification model is respectively constructed to determine the optimal structure and parameters of heterogeneous fusion network.The experiment shows that the heterogeneous fusion network using Nadam as the optimizer and PReLU as the activation function can achieve better classification effect(the average recognition rate is 99.17%,and the average loss is 0.0211).(5)It studies the anti-interference problem of the identification model for coal and gangue.To verify the robustness of t the identification model for coal and gangue proposed in this paper,two sets of anti-interference experiments are designed,and noise signals such as Gaussian noise,Poisson noise,and Salt and pepper noise were added to verify the anti-interference ability of the identification model.The experiments show that the recognition method for coal and gangue based on CNN has a strong anti-interference ability when the same or different environmental disturbances occur,and the recognition method using the combination strategy of image feature extraction and SVM classifier has poor anti-interference ability for the different environmental disturbance.This paper provides a new idea for the identification of coal and gangue,which combines MSI with CNN and its improved structure in deep learning to build accurate identification models.Aiming at the recognition problem of coal and gangue using MSI,compared with the traditional recognition method,the CNN identification model optimized by structures and parameters has the characteristics of high accuracy and anti-interference.With the help of the trinity of "theoretical analysis experimental simulation and numerical simulation",the model is constructed from the industry demand of the coal gangue identification,and then the intelligent separation of coal and gangue is assisted,which has a certain role in promoting the intelligent construction process of the coal industry.Figure[54]table[27]reference[178]...