Quantitative analysis of Cu, Zn, and Pb elements in ores by X-ray fluorescence using a hierarchical convolutional network with attention excitation

Abstract

Precise determination of heavy metal element concentrations in ores is vital for sustainable resource utilization, environmental protection, and industrial applications. X-ray fluorescence spectroscopy (XRF) has emerged as a preferred technique owing to its non-destructive, rapid, and on-site analytical capabilities. However, challenges such as matrix effects, spectral line interference, and instrumental noise often limit its accuracy. In this paper, a novel deep learning model, the Hierarchical Convolutional Network with Attention Excitation (HCNAE) is developed to enhance the prediction of heavy metal element quantification, copper (Cu), zinc (Zn), and lead (Pb), in ores using XRF spectra. First, ore spectra were acquired using a handheld XRF analyzer. Second, to address challenges such as spectral continuity, inter-channel correlations, and matrix effects, a HCNAE was developed. The HCNAE model incorporates hierarchical convolutional layers for global and local feature extraction and a squeeze-and-excitation (SE) mechanism for dynamic channel recalibration. Finally, the model integrates feature extraction, attention mechanisms, and regression tasks in an end-to-end framework, enabling the accurate concentration estimation of heavy metal elements. The performance of the model was compared with six widely used machine learning and deep learning algorithms to ensure a comprehensive evaluation. The HCNAE achieved coefficients of determination of 0.9961, 0.9715, and 0.9894 for Cu, Zn, and Pb, respectively. The results demonstrate the effectiveness of the HCNAE in mitigating matrix effects and spectral interference in XRF, offering accurate predictions even under challenging conditions. This study presents HCNAE as a scalable and innovative solution for heavy metal element quantification in ores, providing a strong foundation for advancements in mining and geological research.