Convolutional neural network assisted laser-induced breakdown spectroscopy for determination of calcium,magnesium,silicon and aluminum in iron ore
The quantitative analysis of calcium(Ca),magnesium(Mg),silicon(Si)and aluminum(Al)in iron ores by laser-induced breakdown spectroscopy(LIBS)is helpful for rapid evaluation of iron ore quality.However,due to the influence of laser energy fluctuation,matrix effect and spectral interference and other factors,LIBS combined with single variable quantitative analysis of Ca,Mg,Si,and Al in iron ores has the application challenges of large error and low accuracy.Multivariate analysis of the original LIBS spectra can effectively improve the quantitative performance of LIBS.In this study,a one-dimensional convolutional neural network(CNN)model based on LIBS spectra was established for the quantitative analysis of Ca(in CaO),Mg(in MgO),Si(in SiO2)and Al(in Al2O3)in iron ores.Total 628 representative samples of iron ore from 35 brands in 8 countries were collected.The determination coefficient(R2)and root mean square error(RMSE)were used to evaluate the model performance.The influence of normalization method of LIBS spectra of iron ore on model performance was compared,including the feature normalization,spectral normalization and internal standard normalization.The results showed that the normalization preprocessing had a minor impact on the contents of Mg and Al,while the spectral normalization was more suitable for the analysis of Ca content analysis,and the feature normalization was more suitable for the analysis of Si content.The model parameters had a great influence on the model performance.The number of convolution cores,the size of convolution cores and the batch size were optimized,respectively.The results showed that when the number of convolution cores was 24,the size was 50,and the batch size was 256,the predictive model for Si content achieved R2 and RMSE of 0.962 6 and 0.469 8％,respectively.When the number of convolution cores was 12,the size was 60,and the batch size was 256,the predictive model for Al content a-chieved R2 and RMSE of 0.949 4 and 0.132 4％,respectively.When the number of convolution cores was 24,the size was 60,and the batch size was 128,the predictive model for Ca content achieved R2 and RMSE of 0.967 0 and 0.077 6％,respectively.When the number of convolution cores was 12,the size was 60,and the batch size was 256,the predictive model for Mg content achieved R2 and RMSE of 0.999 2 and 0.075 3％,re-spectively.The constructed optimal models with partial least squares(PLS),support vector machine(SVM),random forest(RF)and variable importance-backpropagation-artificial neural network(VI-BP-ANN)were used for method comparison.The results showed that the CNN model exhibited better predic-tion performance with the lowest RMSE and the highest R2.It indicated that CNN-assisted LIBS was appli-cable for the determination of Ca,Mg,Si,and Al contents in iron ores.
NETL
NSTL
万方数据
