The Study On Energy Dispersive X-ray Fluorescence Spectra Modeling Methods Based On Wavelet Transform

Energy dispersive X-ray fluorescence (EDXRF) spectrometry can use for thesimultaneous determination of various heavy metals.The advantages of XRF are rapid,accurate, simple pre-treatment, low cost, field testing, etc., what makes it widely usedin mining, smelting, import and export inspection and quarantine areas. EDXRF hasalso been widely used in the analysis of heavy metals in the soil at present, and how toaccurately establish the quantitative detection of heavy metal model has become thefocus problem. Due to the complexcomposition of soil, X-ray will interact withelements to produce interference during the detection, what makes a lot of noise andbaseline interference exist in the original spectrum. Thus, before creating heavy metaldetection model, we need to find a reasonable method of spectral pre-processing toreduce interference and then to improve the accuracy and stability of the detectionmodel.In this paper, the wavelet transform (WT) method was used in noise reductionbaseline correction of the original spectrum. The wavelet basis and decompositionlevels were chosen through experiments before the pretreatment on the opticalspectrum. The fast Fourier transform (FFT) method, Savitzky-Golay (S-G) filtermethod and moving average (MA) method were also used to process the spectrum,and the processing results were compared with WT method. It shows that the WTmethod is superior to other methods in noise reduction. Baseline correction wascarried out based on the spectral de-noising, using the principle of "stripping peak" toachieve rapid separation of peaks, in order to obtain the baseline and correct it. Afterthe processing by WT, the spectrum before and after processing were compared, theresultsshow that the position and shape of the peak is not changed, which makes itsuitable to process the actual process line.The quantitative detection models of heavy metals in soil were established based on the processed spectra, three kinds of quantitative models were discussed in thispaper. The first is based on polynomial regression: the modeling results were preparedwith spectrum before and after pretreatment, the coefficient of determinationincreased and the detection limit decreased after pretreatment, indicating thatpre-treatment of the spectrum can effectively improve the accuracy and stability of theinstrument. The second is based on Levenberg-Marquardt back propagation artificialneural network (LM-BP-ANN) theory. First, theoptimal value of neurons number inthe hidden layer, learning rate and number of iterations were determined by training,then we finally got the quantitative models of Cr, Cu, Zn, As, Pb. The correlationcoefficients (R) were higher than0.98for all the training sets, calibration sets and theprediction sets, indicating good modeling effects. The last is based on support vectormachine (SVM): genetic algorithm(GA) was used to optimizeand train SVM model,the results showed that modeling after (-1,1) normalized data, R were higher than0.95for allvalidation models of all the five kinds of heavy metals, indicating that themodels were successfully established.