| Coal and iron ore play an important role in meeting the world's energy needs.The information of coal type and place of origin and iron ore brand and place of origin provide technical support for the quality evaluation of coal and iron ore and the collection and administration of import and export tax.It is essential to inspect the quality,type and origin of coal and iron ore before entering the customs.Due to the high dependence of coal and iron ore in China and the huge amount of coal and iron ore imported every year,the traditional chemical detection method with high energy consumption and long detection cycle is not benefited to rapid customs clearance at ports.Near infrared spectroscopy is widely used in coal field due to its advantages of simple sample preparation and rapid nondestructive analysis.In this paper,the differences of near infrared data of coal samples of different types and origin are mainly used to establish the classification model of coal samples and origin areas combined with stoichiometry algorithm.Since X-ray fluorescence and Raman spectroscopy can provide the information of element content and composition structure of samples respectively,the two technologies are applied to brand iron ore identification in this paper.By expanding the near infrared database of coal,X-ray fluorescence and Raman database of iron ore,it can provide reference for further expanding the research scope and realizing the intelligent discrimination of coal type/origin and iron ore brand.The main contents of this paper are as follows:(1)The diffuse reflectance NIR spectral of lignite,bituminous and anthracite was collected and compared,the correlation between the NIR spectral characteristics of coal and coal species was excavated.The coal species discriminant model based on stepd-Fisher discriminant analysis was established.The samples were representative samples of 410 batches of imported coal from nine countries including Australia,Russia and Indonesia.Spectrogram analysis showed that the absorbance,spectral slope and characteristic peak of NIR spectra of different coal samples is different.Data mining was conducted based on sample composition information,X-ray diffraction and near infrared spectroscopy.It was found that the absorbance of near infrared spectroscopy was positively correlated with fixed carbon content in coal,and the spectral slope was negatively correlated with coal aromatization.The increase of coal aromatization resulted in the increase of absorption coefficient in long wavelength direction and the decrease of spectral slope.The absorption peak of spectral characteristics could be mainly assigned to the characteristic information of water and hydrogen groups in organic substances,and the intensity of characteristic peak depends on the content of water and volatile matter in coal.Principal component analysis(PCA)was used to reduce the dimension of the data,and the spectral variables were reduced from 1557 to 394.The principal components were discriminated step by step.The screened 10 principal components(PC1-PC10)were used as the model input variables instead of the original data,and Fisher discriminant analysis model was established step by step.70% samples of each coal species were selected as the modeling set,and 30% samples were selected as the prediction set to evaluate the model stability.The results showed that the validation accuracy of modeling sample is 98%,cross-validation accuracy is 97.8%,and test sample validation accuracy is 99.1%.PCA load diagram showed that PC1 and PC2 were related to volatile and moisture content in coal.Fisher discriminant function 1(57.7%)had the strongest correlation with PC1,and Fisher discriminant function 2(42.3%)had the strongest correlation with PC2.The result suggested that the difference of volatile and water content in different coal species could be used as an internal basis for coal species identification by NIR spectroscopy.(2)Diffuse reflectance near infrared spectra(NIR)of 222 batches of imported bituminous coal samples from Russia,Australia,Indonesia,Mongolia and Canada were compared and analyzed in terms of absorbance,spectral slope and main characteristic absorption peaks.Principal component analysis(PCA)and t-distributed stochastic neighborhood embedding algorithm(t-SNE)were used to visually analyze the difference of near infrared data of bituminous coal from different countries.The results showed that whether based on PCA or t-SNE algorithm,Indonesian bituminous coal is clearly separated from the other four countries,while the scattered points of Russian bituminous coal,Australian bituminous coal,Mongolian bituminous coal and Canadian bituminous coal cross each other and are difficult to distinguish due to similar characteristics.Compared with PCA,t-SNE algorithm has more concentrated distribution among similar samples,more obvious dispersion among different like samples,and greater feature difference.In order to verify the effectiveness of the unsupervised algorithm,PCA and t-SNE dimensionality reduction data were combined with stepwise discriminant-Fisher discriminant analysis and back-propagation artificial neural network(BP-ANN),respectively,to establish the traceability model of bituminous coal from different countries.The classification results showed that the method based on principal component analysis-back propagation artificial neural network(PCA-BP-ANN)has the best classification effect,and the accuracy of modeling sample set,validation sample set and test sample set is 95.6%,94.3% and 81.5%,respectively.(3)X-ray fluorescence and Raman spectrum data of 202 iron ore samples imported from 14 brands from Australia,South Africa and Brazil were collected and analyzed.X-ray fluorescence data,Raman spectral data,and fusion modeling of Raman spectral data and X-ray fluorescence data were used.80% of the samples from each brand were selected as the modeling set,and the remaining 20% were used as the prediction set to evaluate the model stability.Elemental content of all samples was determined by X-ray fluorescence spectrometry without standard sample analysis,and the contents of 11 elements,including Fe,O,Si,Ca,Al,Mn,Ti,Mg,P,K and S were screened out by stepwise discriminant Fisher discriminant analysis as effective variables to establish a two-dimensional discriminant model.The accuracy of modeling set,cross validation set and prediction set was 96.5%,94.1% and 98.2%,respectively.After dimensionality reduction of Raman spectral data by principal component analysis(PCA),the spectral variables were changed from 1768 to 201.The first 10 principal components were further used as input variables to establish a stepwise discriminant-Fisher discriminant analysis model.As the main phase of iron ore brands used in the experiment is hematite with similar composition,the modeling set,cross validation set and prediction set based on single Raman spectrum only obtain 82.1%,68.0% and 72.3% accuracy,respectively.The obtained Raman spectral data and X-ray fluorescence data were fused,and the iron ore brand recognition model was established by combining step-to-fisher discriminant analysis.Due to the complementary effect of element content and phase information,the discriminant model based on Raman spectroscopy-X-ray fluorescence fusion data obtained the optimal results,and the accuracy of modeling set,cross validation set and prediction set was 98.4%,95.1% and 100.0%,respectively. |
PDF Full Text Request