A new technique for baseline calibration of soil X-ray fluorescence spectra based on enhanced generative adversarial networks combined with transfer learning

Abstract

Obtaining accurate characteristic spectra and the net peak area is crucial in X-ray fluorescence (XRF) quantitative analysis. To improve the calculation accuracy of the net peak area of the characteristic X-ray spectrum, carrying out the baseline calibration is necessary before resolving the spectrum. First, this article proposes the use of an enhanced generative adversarial network (EGAN) depth network model for baseline calibration of XRF spectra. This method directly takes a clean spectrum after deducting the background as the optimal target for EGAN deep network training, thus eliminating the traditional cumbersome baseline fitting process. It can directly and quickly obtain the XRF spectrum after baseline calibration, and the entire process requires neither fitting the background nor knowing the background data. Further, to improve the generalization ability of the EGAN model, we introduce transfer learning in the model training process. We use the existing alloy sample spectrogram data to pretrain the EGAN model generator and then migrate the pretrained XRF-EGAN generator model to the baseline calibration task of 57 national standard soil sample spectra collected. During the model training process, a cross-validation method is used to train and test the model’s effectiveness. Finally, experiments are conducted on simulated spectra and actual XRF spectra to verify the accuracy and adaptability of the proposed method. Instrument calibration curves for peak counting and element concentration are established to verify the effectiveness of baseline calibration. The coefficient of determination R² of the calibration curves for elements Cu, Zn, Mn and Cr is increased to 0.998, 0.988, 0.922 and 0.991, respectively. The results indicate that our proposed method can effectively estimate the pure spectrum after deducting the background. In addition, this method can also be applied to the baseline correction of other similar spectral signals.