快速、准确的放射性核素识别可有效地对放射性危险源进行及时的监测预警,对保护人们远离放射源的威胁具有重要意义。该文基于卷积神经网络研究了放射性核素γ能谱的识别。通过溴化镧能谱仪采集16种放射性核素的γ能谱数据,并通过改变放射性核素γ能谱的计数和能谱漂移程度,创建生成大量单核素和双核素γ能谱训练数据,利用自搭建的卷积神经网络开展放射性核素识别模型训练。实验采集其中9种核素及其双核素的混合能谱对核素识别模型开展验证,结果表明：在剂量率约为0.5 μSv/h、测量采集时间为60 s时,模型的识别准确率可达92.63%,满足在低剂量率下对放射性核素进行快速识别筛查的需求。

[Objective] Rapid and reliable radionuclide identification can enable rapid monitoring and early warning of radioactive sources, which is essential for safeguarding people from the threat of radioactive materials. However, distinctive peak matching algorithms are not suitable for low gross count gamma-ray spectrum identification, especially when there are overlapped peaks in a spectrum. To improve the identification performance for low gross count gamma-ray spectra, this study creates a radionuclide identification model based on convolutional neural networks that can better identify the spectra obtained at low dose rates.[Methods] Firstly, a gamma-ray spectrum dataset was created. The gamma-ray spectra of 16 radionuclides were obtained at a dose rate of about 0.5μSv/h using a LaBr₃ spectrometer with measuring energy ranging from 30 to 3000keV, a resolution of about 5% at 662 keV, and a measured acquisition time about 100s. Secondly, a training dataset was developed. To train the model, a huge number of gamma-ray spectra of 16 radionuclides and their two mixed radionuclides were generated. We created 1100 data points for each type of gamma-ray spectra by varying the gross count and energy drift. Thus, a total of 149 600 gamma-ray spectrum data were generated. Among them, 80% of the data were randomly selected for model training and the remaining 20% for model crossvalidation. Finally, the convolutional neural networks was constructed. The random searching approach was used to search hyperparameters of the model using the Keras-Tuner tool for determining the ideal architecture of convolutional neural networks. The convolutional layer filter numbers were 96, 128, 32, 256, and 256 in order. The activation function for convolutional layers was the rectified linear unit. Furthermore, the neuron number of the hidden layer was 480, and the learning rate was 0.000 029 6. At last, the spectra labels were encoded using the one-hot format, and the softmax function was used as the activation function for the model's output layer. The model parameters were optimized using the Adam optimizer by employing crossentropy as the loss function. We obtained the radionuclide identification model after 100 epochs of training.[Results] To estimate the identification performance of the model under the condition that a dose rate was about 0.5 μSv/h and the measurement acquisition time was up to 120 s, we acquired 1 333 gamma-ray spectra from nine single radionuclides and their two mixed radionuclides using the LaBr₃ spectrometer. The nine radionuclides were ²⁴¹Am, ¹³³Ba, ¹³⁷Cs, ¹³¹I, ²²⁶Ra, ²³²Th, ⁵⁷Co, ²³⁵U, and ⁶⁰Co. The model was used to identify these spectra and the results showed that the model's accuracy was 90.11% with the acquisition time of 30s, and the accuracy was increased to 92.63% with the acquisition time of 60s.[Conclusions] In this study, we propose a radionuclide identification model based on convolutional neural networks. Analyses show that the model can effectively identify various radionuclides' gamma-ray spectra in a short period of time at a low dose rate.