Scholarlink

<p>Prompt Gamma Neutron Activation Analysis (PGNAA) enables real-time multi-element analysis of bulk materials but suffers from severe peak overlap, high noise, and complex backgrounds. To address these challenges, we propose SE-Autoencoder-MBR Net, an end-to-end deep learning model integrating Squeeze-and-Excitation (SE) attention, a symmetric autoencoder, and multi-branch heteroscedastic regression. The SE module recalibrates spectral features to enhance characteristic peaks of Ca, Si, Fe, and Al, while the autoencoder enforces spectral morphology preservation and implicit denoising. Crucially, the multi-branch regression outputs element concentrations alongside input-dependent uncertainty estimates, enabling adaptive task weighting via heteroscedastic negative log-likelihood. Experiments on a Monte Carlo–simulated cement dataset—verified for spectral fidelity against experimental measurements—demonstrate that the method outperforms CNN and PLS baselines, achieving an average R<sup>2</sup> &gt; 0.97 and RMSE &lt; 0.47. Ablation studies confirm the contribution of each module, highlighting the framework’s potential for accurate, robust, and uncertainty-aware industrial monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on PGNAA online multi-element quantitative method based on SE-Auto-MBR neural network

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • You-Jian Zhang,
  • Yan Zhang,
  • Hao-ran Zhang,
  • Xuan-di Hu,
  • Shi-Liang Liu,
  • Fang-zheng Luo,
  • Ren-Bo Wang

摘要

Prompt Gamma Neutron Activation Analysis (PGNAA) enables real-time multi-element analysis of bulk materials but suffers from severe peak overlap, high noise, and complex backgrounds. To address these challenges, we propose SE-Autoencoder-MBR Net, an end-to-end deep learning model integrating Squeeze-and-Excitation (SE) attention, a symmetric autoencoder, and multi-branch heteroscedastic regression. The SE module recalibrates spectral features to enhance characteristic peaks of Ca, Si, Fe, and Al, while the autoencoder enforces spectral morphology preservation and implicit denoising. Crucially, the multi-branch regression outputs element concentrations alongside input-dependent uncertainty estimates, enabling adaptive task weighting via heteroscedastic negative log-likelihood. Experiments on a Monte Carlo–simulated cement dataset—verified for spectral fidelity against experimental measurements—demonstrate that the method outperforms CNN and PLS baselines, achieving an average R2 > 0.97 and RMSE < 0.47. Ablation studies confirm the contribution of each module, highlighting the framework’s potential for accurate, robust, and uncertainty-aware industrial monitoring.