Research On Improving The Effect Of Spectral Detection Based On Machine Learning Algorithm

Laser induced breakdown spectroscopy(LIBS)is a simple and safe spectral analysis technique.It can analyze the gas,liquid and solid samples without or simple pretreatment,laser ablation of the sample size is small,can quickly analyze all the elements in the sample at the same time.Due to the influence of matrix effect,self-absorption effect of spectral lines and mutual interference between spectral lines of different elements,there is a large error between the detection result obtained by analyzing soil elements with LIBS technology and the actual value.In order to improve the detection effect,this paper proposes to apply the machine learning algorithm to the quantitative analysis of LIBS technology,and constructs several algorithm models,with a view to extending the technology to other detection applications.In this paper,three types of quantitative analysis models including Support vector machine(SVM),Least squares support vector machine(LSSVM)and Relevance vector machine(RVM)are constructed,and their detection effects on soil element concentration are analyzed and discussed.In order to make the results more accurate and convincing,three target elements were selected,including Ba,Cr,Ni,to comprehensively analyze the detection effect of the established machine learning algorithm model.Firstly,a quantitative analysis model of SVM was established.Before modeling,spectral data were normalized to the full spectrum,and the peak intensity was used as the model input,to compare with the detection effect of traditional quantitative analysis methods.From the perspective of the overall root-mean-square error,the quantitative analysis model of SVM is not as good as the traditional quantitative analysis method in the detection effect of high concentration Ba.In the detection results of Cr and Ni elements,the results of the two methods were similar.In most of the predictions,the SVM model is slightly better than the traditional quantitative analysis method for the relative standard deviation of repeated measurements,indicating that the accuracy of SVM detection is higher.Secondly,the least squares support vector machine quantitative analysis model was established.The spectral data were normalized before modeling,and the peak intensity was taken as the model input.The predicted results were compared with the results obtained by the traditional quantitative analysis method and the support vector machine model.The results show that the relative error of the prediction results of the LSSVM model is greatly reduced,and the detection accuracy is far higher than that of the traditional quantitative analysis method and the support vector machine model.The overall prediction accuracy is good,but the precision of the prediction results is somewhat low.Thirdly,the peak intensity and spectral peak area of the normalized data were used as the model inputs respectively to establish the RVM model and analyze its influence on the detection effect of the model.It was found that compared with the detection results of the RVM model established with the peak strength as the model input,the prediction effect of the RVM model was higher in accuracy than that of the SVM model and LSSVM model,but lower than that of the LSSVM model.It is higher in precision than SVM and LSSVM.The training and testing results of the model with spectral peak area as input are better than those with peak intensity as input.In addition,the RVM model was established with the spectral peak area after full spectrum normalization and the spectral peak area after non-full spectrum normalization as input respectively.The precision of the model was better with the spectral peak area after full spectrum normalization as input,but the prediction accuracy was not significantly improved.