Related Experiment Video
Updated: Jan 18, 2026
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
Multi-scale convolutional attention GRU network combined with improved CARS strategy to determine key elements in
Zenan Xiao1, Fusheng Li1, Wanqi Yang1
1School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 611731, PR China; Yangtze Delta Region Institute (Huzhou), University of Electronic Science and Technology of China, Huzhou, Zhejiang, 313001, PR China; Laboratory for Microwave Spatial Intelligence and Cloud Platform, Yangtze Delta Region Institute (Huzhou), University of Electronic Science and Technology of China, Huzhou, Zhejiang, 313001, PR China.
Background:
X-ray fluorescence (XRF) technology is a promising method for estimating the metal element content in ores, which helps in understanding ore composition and optimizing mining and processing strategies. However, due to the presence of a large number of redundant features in XRF spectra, traditional quantitative analysis models struggle to effectively capture the nonlinear relationship between element concentration and spectral information of XRF, making it more difficult to accurately predict metal element concentrations. Thus, analyzing ore element concentrations by XRF remains a significant challenge.
Results:
This study proposes a novel spectral variable selection method and a new model for quantitative analysis of metal elements in ore XRF. The variable selection method is composed of a heuristic optimization-based competitive adaptive re-weighted sampling method (CARS). The proposed quantitative analysis method integrates multi-scale convolutional attention with gated recurrent unit network (MSCA-GRU). Firstly, multi-scale convolutional attention uesd to reshape spectral feature importance and subsequently utilizes GRU network to extract potential nonlinear relationships in spectral variables. The experimental results show that the heuristic optimization-based CARS method outperforms the traditional CARS algorithm by effectively removing redundant features while retaining useful information. Furthermore, the MSCA-GRU model, compared to other advanced machine learning models, achieves high-precision quantitative analysis for Cu, Zn, and Pb, reaching the coefficient of determination (R2) of 0.9982, 0.9985, and 0.9970, respectively.
Significance:
The proposed MSCA-GRU method combined with heuristic optimization-based CARS includes variable selection and quantitative analysis algorithms, effectively improving the accuracy of key elements quantification in ores. It provides a new method for accurately determining the concentration of metal elements, helping to simplify mineral exploration processes.
