Research On EDXRF Spectrum Analysis Based On On-line Detection Of Steel Slag

Energy dispersive X-ray fluorescence(EDXRF)spectroscopy is a non-destructive multielement analysis method that has been widely used in industrial real-time online testing in recent years.However,due to the complicated scene of industrial material analysis,elemental characteristic X-rays have been disturbed in varying degrees during the occurrence and measurement phases,resulting in a large amount of signal noise and baseline interference in the obtained spectrum.Therefore,this paper studies the processing method of on-line EDXRF spectrum analysis for industrial materials from three aspects: energy spectrum noise reduction,baseline correction and quantitative analysis.In the spectrum noise reduction method,the comprehensive value of smoothness coefficient,signal-to-noise ratio and cosine distance is taken as the evaluation criterion.Firstly,the Kalman filtering technique is used to reduce the noise.The maximum expected iteration algorithm is used to estimate the parameters.The optimal number of iterations is about 50 and the noise reduction value is 1.891.Then,the original spectrum is denoised by wavelet transform(WT)method.The optimal values of wavelet base and wavelet decomposition level are obtained through experiments.The optimal parameters are as follows: bior2.2,4 layers,The noise reduction value is 1.825.The advantage of kalman filtering algorithm is self-adaptation,but the kalman filtering speed is not ideal due to the maximum expected iterative algorithm.Though wavelet analysis is theoretically non-adaptive,this problem has no obvious performance in actual noise reduction,And wavelet analysis because the calculation is simple and fast.So this article uses wavelet analysis(bior2.4,4)as the energy spectrum noise reduction algorithm.Based on the previous section,a comparative study on the weighted peak clipping least-squares baseline fitting method and wavelet analysis is carried out.Weighted peak clipping least-squares baseline fitting method has better adaptability,while the wavelet analysis method is subject to the selection of its wavelet base and other parameters.Therefore,it is considered that weighted peak clipping least-squares baseline fitting method is more suitable for the actual spectrum Line processing.Based on the wavelet analysis of noise reduction and weighted peak clipping least-squares method,a multiple linear regression quantitative analysis model based on full-spectrum data was established.Among them,two kinds of parameter estimation methods,ordinary least square and partial least square,were compared.The predictive-reference curves of silicon,calcium and iron obtained by partial least squares method were 0.99912?0.999924 and 0.99608 respectively.The average relative errors of the prediction sets are 1.885%,0.702% and 10.332% respectively,which is better than the ordinary least squares estimation method.In addition to the multiple linear regression model based on the full-spectrum data,this paper presents a quantitative analysis model based on random forest model.Obtained by experiment,the optimal parameters for the model of silicon and calcium have determined: 500 decision trees and 5 sampling;The optimal training parameters for determining the iron element model are as follows: 500 decision trees and 10 sampling.The average relative errors of the random forest quantitative analysis models of silicon,calcium and iron are 0.626%,0.162% and 0.392% respectively.Slow training is its shortcomings in online applications,but once the model is determined,the variable evaluation can be used as a basis for variable screening to enhance the effectiveness of other modeling methods.Therefore,a combinatorial modeling scheme based on stochastic forest plus least squares estimation is proposed.The variables of screening for the three elements of silicon,calcium and iron are determined as follows: 336,272,200 respectively;the average relative errors of the prediction set are 1.382%,0.609% and 2.792%,the effect is between the random forest model and the least square multiple linear regression model,with the advantages of reduced error,fast and flexible.