Genetic algorithm-optimized BP neural network model for prediction of soil heavy metal content in XRF

Abstract

In the quantitative analysis of XRF (X-ray fluorescence spectroscopy) for predicting trace heavy metal elements in soil, there are large errors in data analysis due to background interference and signal interference of neighboring elements. This paper proposed a hybrid algorithm, which optimizes the weights and thresholds of the BP neural network by using Genetic Algorithm, so as to study the intrinsic relationship between the element's component information and content. The presented hybrid algorithm is used for quantitative analysis of four heavy metal elements Pb, Mn, Cr, and Cu in soil. Firstly, the X fluorescence spectra of 57 standard soil samples were obtained using a portable X fluorescence analyzer, and the discrete wavelet was used for denoising and background deduction to improve the signal-to-noise ratio. Then, the component information of four heavy metals and some other elements that might interfere with them are obtained by using Compton's normalization method, which are respectively taken as the inputs of the model, and the actual contents of Pb, Mn, Cr, and Cu are respectively taken as the outputs, so as to establish a three-layer BP neural network with one implicit layer. Finally, the training results of the model show that there is a good overlap between the predicted data and the expected one which verifies that the method is effective for the quantitative analysis of trace heavy metal elements in the soil.