Study On The Improvement Of The Stability Of LIBS Water Quality Detection By The Selection Of Characteristic Spectral Lines

When laser-induced breakdown spectroscopy(LIBS)is used for substance detection,the spectrum contains multiple atomic lines or ion lines of all excitable elements in the measurement sample.The spectral line information is complex and easy to generate.Cross-overlap interference.Especially for water quality detection,due to the influence of sputtering and fluctuation of liquid samples,there are a lot of redundant features in the spectrum,which will affect the stability of quantitative analysis.In this paper,the Ca Cl2 solution was used as the experimental sample,and a LIBS online detection platform with liquid jet as the injection method was built to study the improvement of the stability of LIBS water quality detection by the selection of characteristic spectral lines.The main contents include:(1)When LIBS is used for water quality detection,there are problems such as high baseline,complex noise,and inconsistent spectral line intensity dimensions in the spectrum.In order to reduce the error and improve the detection stability,the original spectra were pre-processed such as baseline correction,noise removal,data de-dimensionalization,variance filtering and F-test.(2)Due to serious water sputtering and large fluctuations,the artificial selection of characteristic spectral lines combined with univariate model prediction results are not ideal.In order to improve the stability of LIBS water quality detection,the CART regression tree is used to select the characteristic spectrum.Through the variable selection of the CART regression tree,the number of characteristic variables is reduced from 100 to 6,and the compression rate of variables reaches 94%,which significantly reduces irrelevant spectral line interference.The R~2,RMSEC,RMSEP and ARE of the CART regression tree model were 0.9975,0.0035 wt.%,0.0061 wt.%and 2.500%,respectively,and the ARSD was reduced to 3.21%.Comparing the prediction results of the univariate model and partial least squares regression,the CART regression tree model not only has better feature explanatory ability,but also has a lower relative standard deviation,indicating that it has higher prediction stability.(3)The stability of a single CART regression tree model is not high,and when there are many input features,it is prone to overfitting.In order to solve the above problems and improve the stability of LIBS in water quality detection,this chapter introduces the Gradient-Boost Decision Tree(GBDT)to analyze the Ca content in the mixed solution.The characteristic spectrum selection results and quantitative analysis stability of a single CART regression tree,Random Forest Regressor(RFR)and GBDT were compared.The feature reduction rate of GBDT reaches 94.8%,and it has better feature discrimination ability.The fitting coefficients R~2,RMSEC,and RMSEP of the GBDT model were 0.999,0.001wt.%,and 0.003wt.%,respectively.The ARE reached 2.45%,and the ARSD was reduced to 3.37%.The quantitative analysis results were significantly better than CART and RFR.The above results prove the feasibility of combining the GBDT model with LIBS for water quality detection.The GBDT model based on the Boosting idea can not only accurately select the characteristic spectral lines,but also significantly improve the stability of water quality detection.This paper studies the instability of LIBS when it is used for water quality detection,and introduces CART regression tree and GBDT model to select LIBS characteristic spectral lines.By comparing the prediction results,it is shown that the CART regression tree and the GBDT model can effectively improve the stability of LIBS water quality detection,and provide effective support for the development of LIBS technology in the field of water quality detection.