At border crossings around the world, plastic scintillator-based radiation portal monitors (RPMs) are employed to detect the presence of illicit radioactive materials in large trailer trucks. However, the RPM system shows a low energy resolution owing to the large size and physical characteristics of plastic scintillators; and thus, the identification of illicit artificial isotopes from naturally occurring radioactive material is difficult. This study aims to develop an advanced algorithm for radionuclide identification with commercial RPMs based on commercial plastic scintillators to reduce the occurrence of frequent nuisance alarms. Subsequently, machine learning models, namely, a convolutional neural network (CNN) was applied. The spectral distributions of energy weighted spectra were used as features of the CNN model. The energy spectra of Cs, Co, Ra, and K measured under static and moving conditions were used to implement the identification model. To evaluate the performance of the implemented model, the F-score was used. The trained CNN model correctly identified most of the radionuclides. That is, despite the theoretical Compton edge energies of Co and K being similar, the spectral distributions of K are distinctively different from those of Co. The result demonstrates that the CNN model-based identification algorithm performs robust radionuclide identification, thereby reducing the frequency of nuisance alarms at border crossings. Furthermore, considering that the actual cases of cargo passing by the RPMs are becoming more complicated, the algorithm would need to be continuously improved and trained with more complex scenarios in the future.

中文翻译：

---

辐射门户监测系统中基于卷积神经网络的放射性核素识别算法的开发


在世界各地的过境点，采用基于塑料闪烁体的辐射入口监视器 (RPM) 来检测大型拖车中是否存在非法放射性物质。然而，由于塑料闪烁体的大尺寸和物理特性，RPM系统表现出较低的能量分辨率；因此，从天然存在的放射性物质中识别非法人造同位素是很困难的。本研究旨在开发一种基于商业塑料闪烁体的商业 RPM 放射性核素识别的先进算法，以减少频繁滋扰警报的发生。随后，应用了机器学习模型，即卷积神经网络（CNN）。能量加权光谱的光谱分布被用作CNN模型的特征。使用静态和移动条件下测量的 Cs、Co、Ra 和 K 的能谱来实现识别模型。为了评估所实施模型的性能，使用了 F 分数。经过训练的 CNN 模型正确识别了大部分放射性核素。也就是说，尽管Co和K的理论康普顿边缘能量相似，但K的谱分布与Co的谱分布明显不同。结果表明，基于CNN模型的识别算法具有鲁棒的放射性核素识别能力，从而降低了放射性核素识别的频率。边境口岸的滋扰警报。此外，考虑到货物经过RPM的实际情况变得越来越复杂，未来算法需要不断改进和训练更复杂的场景。