Speaker
摘要
基于LaBr₃(Ce)探测器的伽马能谱现场快速测量是核应急响应与环境辐射监测的关键技术手段，但其较低的能量分辨率导致复杂放射性场景中出现严重谱峰重叠与弱峰淹没，传统分步进行平滑-本底扣除流程的谱分析方法及商用解谱软件GammaVision在低峰康比及低信噪比条件下存在显著的弱峰面积丢失问题，制约了动态工况下的核素快速定性定量分析。本研究提出的4-Level-Wavelet ResAttention模型将离散小波变换（DWT）多尺度分解嵌入端到端深度学习框架，通过LevelNetRes残差模块实现弱峰层级特征高维提取，Level-attention Transformer实现跨尺度特征自适应融合，并针对弱峰能谱分析的峰型特征保留关键需求，设计了联合约束峰位精度、峰型对称性及峰面积误差的多目标加权损失函数，提升低峰康比条件下的弱峰提取精度与鲁棒性。实验表明，与传统方法相比，新模型在低峰康比条件下净峰谱均方误差与峰型变形程度分别降低95.82%和97.71%，峰面积误差降低89.33%；在重叠峰分析中面积比误差降低91.76%；在峰康比低至1.34—2.97的极端实测场景中，峰形保真度为传统方法的7.94%，峰面积误差降低17.47%。该模型为低能量分辨率探测器的复杂弱峰能谱解析提供了高精度、高鲁棒性的技术方案，可有效支撑复杂放射性核素的现场快速识别与定量分析。
Abstract
Rapid in-situ gamma spectroscopy based on LaBr₃(Ce) detectors is critical for nuclear emergency response and environmental radiation monitoring. However, their limited energy resolution leads to severe peak overlap and weak-peak masking in complex radioactive scenarios. Conventional stepwise smoothing and background subtraction methods, as well as the commercial spectrum analysis software GammaVision, suffer from significant weak-peak area loss under low Peak-to-Compton ratio and low signal-to-noise ratio conditions, hindering rapid nuclide identification and quantification in dynamic environments. To address this issue, this study proposes a 4-Level Wavelet ResAttention model that embeds Discrete Wavelet Transform (DWT) multi-scale decomposition into an end-to-end deep learning framework. The LevelNetRes module extracts high-dimensional hierarchical features of weak peaks, while the Level-attention Transformer enables adaptive cross-scale feature fusion. A multi-objective weighted loss function is further designed to jointly constrain peak position accuracy, peak symmetry, and peak area error, thereby improving extraction accuracy and robustness. Experimental results demonstrate that, compared with conventional methods, the proposed model reduces the relative mean squared error and peak shape distortion of net spectra by 95.82% and 97.71%, respectively, and decreases peak area error by 89.33% under low peak-to-Compton conditions. In overlapping peak analysis, the area ratio error is reduced by 91.76%. In extreme measured scenarios with peak-to-Compton ratios as low as 1.34–2.97, peak fidelity reaches 7.94 times that of conventional methods, while peak area error is reduced by 17.47%. This model provides a high-precision and robust solution for analyzing complex weak-peak spectra from low-resolution detectors, supporting rapid in-situ identification and quantification of radioactive nuclides.
| 关键词 | 伽马能谱分析；弱峰提取；离散小波变换；ResNet；Transformer；净谱预测；现场快速测量；环境辐射监测 |
|---|---|
| Keywords | Gamma-ray spectral analysis; Weak-peak extraction; Discrete Wavelet Transform; ResNet; Transformer; Net spectrum prediction; Rapid in-situ measurement; Environmental radiation monitoring |
