Research On Laser-induced Breakdown Spectroscopy Quantitative Analysis Algorithm And Application Based On Ensemble Learning

Laser-Induced Breakdown Spectroscopy,LIBS is a new chemical analysis technology.Due to its various advantages,it is widely used in material identification as well as the quantitative and qualitative analysis of mixtures currently.Among them,the research on the quantitative analysis method of LIBS is the current research hotspot,which can help scientific researchers to quickly determine the element composition and calculate the concentration information of the element content in mixtures.Therefore,the in-depth study of LIBS quantitative analysis and the improvement of its quantitative predictive accuracy are considered critical from both the scientific and practical views.The paper summarizes the related research progress in the field of LIBS quantitative analysis and introduces the methodology of LIBS quantitative analysis based on Artificial Neural Networks,Support Vector Regression and Random Forest Regression.Moreover,this paper analyzes the advantages and disadvantages of these methods as well as their application scenario.When using a single learner as the LIBS quantitative analysis model,it is easily affected by some factors including the matrix effect that exists in the samples,the possible overfitting problem,and the random errors.To address those issues,based on ensemble learning,this paper proposes a two-layer structure LIBS quantitative analysis method to help further improve the predictive accuracy of LIBS quantitative analysis.This method first selects some suitable base learners by the process of base learner selection.The first layer of the method consists of several selected base learners.The prediction of the validation set in the process of cross-validation of the first-level base learner is used as the input feature of the secondary metalearner,and the meta-learner is trained accordingly,and the final well-trained model is used as a quantitative analysis model.Meanwhile,in the model training process,a feature selection process,which combines the Genetic Algorithm and Sequential Forward Selection,is adopted to select features in the spectrum.Thus,the generalization ability and the predictive accuracy of the model is improved.The experiment using two related LIBS database proves the validity and feasibility of the proposed method.The experimental result shows that the proposed method outperformed other methods mentioned before,which indicates a better generalization ability and higher predictive accuracy.According to the actual application demands,we designed a spectrum analysis and measurement system based on Andor Spectrometer.The software is based on the C# WPF framework,including modules like spectrometer configuration,chart display,element selection,quantitative analysis,sample settings as well as plasma modules.Users can change configurations of the spectrometer and software for specific experimental needs easily.Meanwhile,the system integrates the proposed method and the calibration-curve method into the software for users to compare and analyze the final experimental result.