Study On Laser-induced Breakdown Spectroscopy Using Feature Selection

Laser-Induced Breakdown Spectroscopy(LIBS)is an atomic spectroscopy technology with the advantages of strong adaptability,multiple-element detection,less consumption and fast analysis.LIBS has shown potential power in the field of chemical analysis.However,it is still faced with the challenges of complex effects on the spectrum,matrix effects and poor quantitative accuracy.In this dissertation,the following works are carried out from the study of influential factors,shielding mechanism,spectrum pretreatment and quantitative analysis.(1)Taking the plasma emission produced by 120 picoseconds laser as an example.The effects of pulse energy,laser wavelength,effective focusing distance and delay time of the spectrometer on emission intensity were analyzed macroscopically.And the relationship between the cumulative pulse number,ablation depth and spectral intensity was studied microscopically.The results show that the strongest emission can be obtained when the focal point is located below the target surface.The spectral intensity is proportional to the pulse energy,but the increased intensity is closely related to the pulse wavelength.Compared with1064 nm,532nm laser can effectively alleviate the fluctuation of effective focusing distance caused by the roughness of the target surface.The stability of spectral intensity is affected by the number of shooting pulses and pulse energy.(2)The physical mechanism of shielding effects in picosecond laser-induced plasma was explored.The intrinsic relationship between emission,plasma temperature and electron density and plasma shielding was studied.When the focal point is just on the surface of the target,the shielding effect is intensified.The more high-temperature molten matter will take away by the shock wave,resulting in the lowest spectral intensity,electron density and plasma temperature.When the focal point is above the target surface,the pulse directly breaks down the fresh air.The above three factors increased again due to the effects of the high-temperature and high-pressure shock wave.Finally,the possibility of LTE existing in plasma is verified by the Mc Whirter criterion,plasma relaxation time and diffusion length,and the approximate optical thin is determined by the actual ratio of spectral intensity.It is proved that the estimated plasma temperature and electron density are reasonable.(3)Considering the mathematical characteristics of LIBS,an asymmetric penalty term is used to maintain the positivity and sparsity of the spectral peak,and a symmetric penalty term is used to constrain the smoothness of the baseline.An adaptive method for fitting baseline is proposed based on the traditional penalized least squares,and then verified by simulated and experimental spectra respectively.The results of the simulated spectra show that the proposed method can fit the baseline correctly.With the increased power in added noise signal,the proposed method achieved the best robustness and reduced the fitting deviation by 2-4 times compared with Mpls,Backcor and Airpls.The results of the measured spectra indicated that the proposed method can solve all kinds of baseline drift formed by spectral line overlap in different degrees,and show strong adaptability to the change of wavelength resolution.(4)Given the scarcity of natural copper ore/concentrate,23 pieces of ore tablets with different contents are developed based on 11 kinds of raw materials.The basic characteristics of these ore spectra were analyzed,and 12 atomic and 17 ion lines were selected.To achieve accurate copper content in raw ore,the above 29 lines were used to realize the feature selection and quantitative analysis by Elastic Net.Compared with the traditional/improved methods of OLS and PLSR,PSO-SVR,the three generalized linear models,Lasso(Least Absolute Shrinkage and Selection Operator),Ridge and Elastic Net,can effectively improve the prediction performance.Among eight methods,the proposed Elastic Net obtained the best prediction accuracy and generalization performance in the test set.The values of R~2,root mean square error(RMSE)and mean absolute percentage errors(MAPE)are 0.9798,1.1971,and8.21%,respectively.Compared with traditional PLSR,the RMSE and MAPE are reduced by24.92%and 35.15%,respectively.This dissertation takes typical complex samples such as raw ore as the research object,the proposed quantitative method of feature selection is conducive to linking mathematical-statistical models with the physical mechanism of LIBS quantification,and improves the interpretability of quantitative analysis models.