Accurate quantification of alkalinity of sintered ore by random forest model based on PCA and variable importance...

The alkalinity of sintered ore has an important impact on the quality, output, and energy consumption of blast furnace smelting, and there is an urgent need for a method for accurate quantifying of the alkalinity of sintered ore. The present work explores the combination of the laser-induced breakdown spectroscopy (LIBS) technique and random forest (RF) based on principal component analysis (PCA) and variable importance for the quantitative analysis of the alkalinity of sintered ore. Sixteen sintered ore samples were used in this study, and the characteristic lines of LIBS spectra for sintered ore samples can be identified based on the National Institute of Standards and Technology (NIST) database. At first, abnormal spectra are identified and rejected by PCA coupled with Mahalanobis distance (MD). Then, the input variable for the RF calibration model is optimized according to the variable importance threshold obtained by the RF model, and two RF model parameters of ${{n}_{\rm tree}}$n_tree and ${{m}_{\rm try}}$m_try are determined by out-of-bag estimate. Finally, the PCA-VI-RF model is built under the optimal model parameters. In order to verify the predictive ability of the quantitative model, the PCA-VI-RF model prediction results were compared with the RF model, partial least-squares model, and least-squares support vector machine model. The result demonstrated that PCA-VI-RF shows better analytical performance than other methods. Compared with the RF model with the original spectrum as input, the averaged relative errors of test results decreased from 5.82% to 3.94%, coefficients of determination (${R^2}$R²) of the test set increased from 0.8957 to 0.9814, and the root mean square error decreased from 0.1502% to 0.0860%. The speed of modeling and prediction has also been greatly improved, and the modeling time was reduced from 4675.56 to 16.86 s. The stability of the PCA-VI-RF model was verified by the relative standard deviation (RSD) of the test data prediction results, and the RSD reached below 4.74%. This study shows LIBS combining PCA-VI-RF is an effective method for accurate quantification of the alkalinity of sintered ore. It has great significance for the potential application of real-time online analysis of the alkalinity of sintered ore.