Research Of On-line Classification Of Multicoal Density Level Based On Machine Vision

Coal resource is abundant in China,but as the mining time is getting longer,coal resources are becoming poor and hard to separate.With the progress of science and technology,the coal dry sorting technology based on high-tech sensor is emerging.However,due to the high cost,radiation,slow response speed,difficult identification of ore monomer and so on,the application and popularization of most advanced detection technologies in the field of dry ore separation are limited while the intelligent dry separation technology based on machine vision can overcome the above problems.Thus,combining the advantages of machine vision technology in coal dry dressing technology and the application status of machine vision technology in coal dressing industry,this paper proposes an online classification technology of multi-coal density level based on machine vision.In this paper,gas coal,coking coal and anthracite with different coal-forming environments and coal-burning degrees were taken as research objects.Coal samples were prepared into four density level of<1.4g/cm~3,1.4?1.6g/cm~3,1.6?1.8g/cm~3,>1.8g/cm~3 in each coal by floating and sinking test,and a set of dynamic coal sample image acquisition system was established.Through the maceral analysis of the experimental coal and the surface image analysis of coal particles,it was proved that the surface color,luster,texture and other apparent characteristics of coal particles are closely related to the density level of coal particles.The causes of the differences in the apparent characteristics of various coal were analyzed.At the same time,it was pointed out that there is the same variation rule between the apparent characteristics and density levels of different coal types,which laid a theoretical foundation for the subsequent extraction of the surface characteristic parameters of coal particles and the establishment of the generic classification model of multiple coal types.A method for feature extraction and screening of coal surface image was presented.Image preprocessing techniques such as threshold segmentation,binarization,morphology,area marking and pixel index were used to ensure the effectiveness of coal particle surface feature quantification.Three color and luster characteristics based on RGB color model were extracted by pseudo color map and histogram analysis.Two texture analysis methods,wavelet transform and gray co-occurrence matrix,were used to extract 24 texture features based on HSV color model.Boxplots-correlation calculation and PSO-SVM were used to screen the features,and their performance was evaluated according to the classification effect of each classification model corresponding to the two methods.The subsequent results showed that the PSO-SVM feature screening method was more effective.The prediction model of coal particle density level was established.Five pattern recognition algorithms of SVM,Random Forest,BP neural network,SIMCA and KNN were used to establish a coal particle density prediction model in three coal species,and the classification results of each classification model under each classification standard were obtained after 10 experiments.And then the classification results of each classification model under each classification standard were obtained.Considering the stability,timeliness and prediction accuracy of each model,a classification decision of"competitive voting"with"multi-algorithm fusion"was made by selecting three models with good comprehensive performance,namely SVM,Random Forest and SIMCA.Finally,based on the analysis of coal and rock components and surface characteristics and the variation law of characteristic parameters of each coal type with density level,a generic classification model of multiple coal types was established,and the classification results of each classification model under each classification standard were obtained.The results showed that the total classification accuracy of competitive voting was the highest,reaching 84.78%in the four categories and 90.89%in the three categories,and the classification accuracy of each model is higher in the two categories.The test results showed that the multi-coal generic classification model based on raw coal data in different coal-forming environments and coal-burning degrees had high accuracy and strong robustness,which provides a certain theoretical basis and technical support for the application of machine vision technology in coal dry dressing technology.