Research And Its Application In Metallurgical Analysis Of Chemometrics Methods In Laser-induced Breakdown Spectroscopy

Steel and slag are the significant product in steelmaking, and its components directly influence the quality and performance of steels. In order to ensure smooth operating of steelmaking process, an effective method is necessary to carry out quick and in-site analysis. Conventional analytical techniques bears the disadvantages of complicated sample preparation, time-consuming and are unable to online analysis, which seriously hindered the automation and intelligent of metallurgical industry.Laser induced breakdown spectroscopy (LIBS) is a kind of quantitative analysis tool based on atomic emission spectrum of emerging technology with the advantages of simple sample preparation, rapid, on-line and real-time analysis. So, there is a wide range of applications in the process control of steel-making and the on-site analysis of slag. Due to the LIBS signal is easily effected by matrix effect, self absorption effect and other factors, its accuracy of quantitative and qualitative analysis is seriously affected. Chemometrics methods can be employed to overcome the influence from these factors, and are used to extract useful information from the complex spectral data and construct a more robust analysis model. In this paper, several chemometrics methods combined with LIBS technology are applied for the quantitative and qualitative analysis of steel and slag samples, and are used to optimize and improve its analytical performance. This work is divided into six chapters, the main research contents as follows:In the first chapter, we introduces the general situation, basic principle, advantages and disadvantages and application fields of LIBS technique, and then reviews the research progress on chemometrics in LIBS analysis, finally states the background, significance and content on this work.In the second chapter, the random forest regression(RFR) model is constructed to quantitative analysis simultaneously five elements(Si, Mn, Cr, Ni and Cu) of steels. Normalized LIBS spectra(220-400 nm) of steels as input variables are employed to construct PLS, SVM and RFR calibration model. RFR model shows a better results of quantitative analysis, and it has a better predictive performance.In the third chapter, LSSVM algorithm coupled with LIBS technique is employed to quantitative analysis of seven major components(CaO, SiO2, Al2O3, MgO, Fe2O3, MnO2 and TiO2) of slag. The whole LIBS spectrum preprocessed by wavelet denoise as the input variables is used to construct PLS and LSSVM calibration model. The parameters(radial basis kernel function parameter γ and σ2) on LSSVM model are optimized by Grid search approach. Under the optimized model parameters, the external validation of two calibration model is completed by RMSE and correlation coefficient(R2) as evaluation indicator with slag samples of testset. It shows that LSSVM model has a well forecasting performance.In the fourth chapter, RF combined with LIBS technology is applied for classification and identification of steel brand. Normalized LIBS spectrum(220-400nm) as the input variables are employed to construct PLS-DA, SVM and RF training model on the classification of steel brand. It shows that PLS-DA, SVM and RF model have lower predictive error rates, however, there is a lowest average error rates for RF model, and RF has a better classification performance for steels.In the fifth chapter, RF based on variable importance(VIRF)combined with LIBS technique is successfully used for classification of 60 slag samples. Normalization by maximum intensity of LIBS spectra (200-500 nm) as the input variables is used to construct the VIRF classification model of slags, the averaged OOB error, sensitivity, specificity and accuracy as evaluation parameters are used to evaluate the VIRF model. We compared with the classification result of the PLS-DA, SVM, RF and VIRF model, VIRF model shows a better classification results and classification performance.In the sixth chapter, Fast-ICA-LSSVM combined with LIBS technology is successfully applied for slag classification. The whole LIBS spectra dressed by Fast-ICA combined with normalized by maximum spectral intensity as input variable is used to construct LSSVM training model. Grid global optimization and 5-folds cross validation are used to optimize the model parameters on LSSVM. In order to verify the prediction ability of Fast-ICA-LSSVM classification model, we compares the classification results based on LSSVM and Fast-ICA-LSSVM model. The results show that Fast-ICA-LSSVM model can overcome the influence of factors such as matrix effect and has a better prediction effect.