Responders need tools to rapidly detect and identify airborne alpha radioactivity during consequence management scenarios. Traditional continuous air monitoring systems used for this purpose compute the net counts in various energy windows to determine the presence of specified isotopes, such as ²³⁵U, ²³⁹Pu, and ²⁴¹Am. These calculations rely on having a well-calibrated detector, which is challenging in low-background environments. Here, an alternative approach of using artificial neural networks to classify alpha spectra is presented. Two network architectures, fully connected and convolutional neural networks (CNNs), were trained to classify alpha spectra into four categories: background and background plus the three isotopes above. Sources were injected into measured background at various fractions of the derived response level (DRL) corresponding to early-phase Protective Action Guides. The convolutional network identifies all sources at 1% of the DRL with average probability of detection of 95% and false alarm probability of 1%. Further, the network identifies sources ranging between 0.25% and 1% of the DRL with higher than 80% probability of detection and lower than 7% false alarm probability. Most significantly, the network performance improves in low-count background conditions, increasing its minimum probability of detection to 93% and reducing the false alarm probabilities to lower than 0.25%. These results show that, once trained on datasets representing a range of detection scenarios, artificial neural networks can accurately identify alpha isotopes of interest without the need for detector calibration.

响应者需要工具来在后果管理场景中快速检测和识别空气中的 alpha 放射性。用于此目的的传统连续空气监测系统计算各种能量窗口中的净计数，以确定特定同位素的存在，例如²³⁵ U、²³⁹ Pu 和²⁴¹是。这些计算依赖于校准良好的检测器，这在低背景环境中具有挑战性。在这里，提出了一种使用人工神经网络对 alpha 谱进行分类的替代方法。训练了两种网络架构，即全连接神经网络和卷积神经网络 (CNN)，将 alpha 光谱分为四类：背景和背景加上上述三种同位素。源被注入到与早期保护行动指南相对应的衍生响应水平 (DRL) 的各个部分的测量背景中。卷积网络以 1% 的 DRL 识别所有源，平均检测概率为 95%，误报概率为 1%。此外，网络识别范围在 0 之间的源。25% 和 1% 的 DRL 具有高于 80% 的检测概率和低于 7% 的误​​报概率。最重要的是，网络性能在低计数背景条件下得到改善，将其最小检测概率提高到 93%，并将误报概率降低到 0.25% 以下。这些结果表明，一旦在代表一系列检测场景的数据集上进行训练，人工神经网络就可以准确识别感兴趣的 α 同位素，而无需检测器校准。