Quantitative Inversion Model Research Based On The Visible And Near-Infrared Spectrum For Skarn-Type Iron Ore

Mineral resources are an important component of the basic industries of China’s national economy,especially iron ore,which has a self-evident impact on the economic development of China.Silica-type deposits are one of the important types of iron ore deposits in China,and their reserves account for about 11%of the total reserves of iron ore deposits in the country.The rich ore accounts for about half of the identified iron-rich ores in the country and has certain research value.Nowadays,how to achieve reasonable,efficient and accurate mining of iron ore resources has become an urgent problem.The traditional ore grade determination method is based on chemical analysis,but due to its long period,high cost,can not complete the in-situ instant determination of ore grade,relative to the ore allocation process there is a lag effect,can not effectively reduce the loss depletion rate of ore mining,so the development of in-situ iron ore grade determination technology based on visible-near infrared spectral analysis is an effective way to solve this problem.In this paper,the Red Ridge silica-type iron ore was selected as the test object of this study,and 225 massive silica-type iron ores were obtained as the experimental samples,which were analyzed by chemical analysis and visible-near infrared spectroscopy test to obtain the iron grade chemical analysis data and visible-near infrared spectroscopy data.The raw data were firstly smoothed(unprocessed data)and used as the data source to synthesize the visible-NIR spectral characteristics of the siliciclastic iron ore,then the unprocessed data were preprocessed by six basic transformations and seven dimensionality reduction algorithms,and the basic transformations and dimensionality reduction algorithms were combined to generate 42 preprocessing combinations to preprocess the unprocessed data.Finally,based on Random Forest(RF),BP(Back Propagation)neural network,Extreme Learning Machine(ELM),Partial Least Squares(PLS),Support Vector Machine(SVM)Five algorithms were used to establish the quantitative inverse model of metallic iron grade of silica-type ores,and the stability,accuracy and credibility of the model were evaluated by three indicators:coefficient of determination(R2),root mean square error(RMSE)and mean relative error(MRE),respectively.The following main conclusions were drawn.1)The five iron grade inversion models established using the random forest method,BP neural network method,extreme learning machine method,partial least squares method,and support vector machine for unprocessed spectral data,respectively,and the five iron grade inversion models for unprocessed spectral data,respectively,have the highest accuracy with the BP neural network method,with a coefficient of determination R2 of 0.93,root mean square error RMSE of 0.046,and mean The second is the support vector machine method,and the worst is the partial least squares method.2)The unprocessed spectral data were preprocessed using the derivative method,inverse and inverse logarithm,multiple scattering correction,standard normal transformation,and continuous statistical removal method,and the inverse model of iron grade was established by the random forest method,BP neural network method,extreme learning machine method,partial least squares method,and support vector machine algorithm,respectively,among which,the inverse logarithm method and the combined extreme learning machine algorithm built the The model accuracy was the highest,with a coefficient of determination R2 of 0.93,root mean square error RMSE of 0.048,and mean relative error MRE of 0.16,but there was little improvement in modeling accuracy over the unprocessed data.In particular,the models built by the support vector machine algorithm inversion model for the spectral fundamental transformed processed data were overall better than the other models.3)Genetic algorithm,multidimensional scaling method,linear discriminant analysis,local linear embedding algorithm,isometric mapping algorithm,Laplace feature mapping algorithm,and principal component analysis were used to preprocess the original spectral data,and the iron grade inversion model was built based on random forest method,BP neural network method,extreme learning machine method,partial least squares method,and support vector machine algorithm for the preprocessed data,among which the principal The accuracy of the model built by the component analysis method and BP neural network algorithm is the highest,with the coefficient of determination R2 of 0.94,root mean square error RMSE of 0.045 and mean relative error MRE of 0.13,followed by the highest accuracy of the model built by the limit learning machine modeling after the multidimensional scaling method.The dimensionality reduction algorithm not only improves the inversion prediction accuracy but also increases the model building speed.4)For the hyperspectral data of smectite type iron ore,a pre-processing combination algorithm combining spectral fundamental transformation method and dimensionality reduction algorithm was used to compare 210 inversion results,and it was found that the quantitative grade inversion model built by ELM limit learning machine and BP neural network was the best after MSC and SNV processing and PCA dimensionality reduction of the spectral data.Among them,the combined MSC-PCA-ELM pre-processing algorithm has the highest inversion accuracy,with the coefficient of determination R2 improved from 0.93 to 0.99,root mean square error RMSE reduced from 0.046 to 0.0057,and relative error MRE reduced from 0.15 to 0.02,which has significantly improved the prediction accuracy.This study provides an effective method for real-time and rapid analysis of smectite iron ore grade,which is of great practical significance for realizing efficient mining of smectite-type iron ore.