基于深度学习模型的伽马能谱解析方法

Deep Learning Model Based Gamma Spectrum Analysis Method

为提高伽马能谱解析精度,建立专用深度学习模型,含12个残差卷积模块、51个神经网络层、超107个参数;独特设计模型输出,使其直接预测核素出射谱,突破对预设核素库的依赖。选择自建全身计数器测量人体放射性作为实验场景,基于蒙特卡罗模拟构造了数据集,测试实验表明,深度学习模型核素识别率93.3%、活度计算平均误差8.6%,相较峰分析法的62.3%、28.3%,能谱重建法的78.2%、18.7%,浅层ANN模型的81.3%、14.8%,优势明显。实测实验进一步验证了上述结论。所建立方法借助深度学习的多层次特征提取能力与高数值稳定性,实现了全谱信息与伽马射线能量、数量间的复杂映射,具备高准确性、通用性,未来可为多种应用提供技术基础。

In view of the limitations of existing gamma spectrum analysis methods on nuclides identification and activity evaluation, a special deep learning model was proposed which consists of 51 layers and more than 10~7 parameters. Based on multitude residual convolutional modules, this model can extract characteristics of whole gamma spectrum hierarchically and comprehensively and keep its numerical stability at the same time. The output of this model was also specially designed so that it can predict the energy and quantity of the gamma rays emitted by nuclides directly, no longer rely on the preset nuclide library. After model construction, the testing experiment was carried out. A NaI-type Whole-Body-Counter was chosen to measure the gamma spectrum of human body. The corresponding digital model was constructed and lots of simulated spectra were generated by Monte Carlo simulation method. For training, the data set was acquired by setting the energy and quantity of gamma ray source particles randomly in each spectrum, and for testing, 9 radionuclides were selected to determine the source particle setting during testing data set simulation. When testing, besides the constructed deep learning model, three existing methods were also tested for comparison. Results show, deep learning model performs best with 93.3% nuclide identification rate and 8.6% average activity predicting error, while traditional peak-analysis method identifies the least nuclides(62.3% identification rate), and gives most inaccurate activity values(28.3% average error), and spectrum-reconstruction method and shallow ANN model also show their limitation when analysis is carried out, with 78.2%, 81.3% identification rate and 18.7%, 14.8% average activity error respectively. In real measurement experiment, a human physical phantom containing134Cs,^(137)Cs,57Co, and60Co was measured for 10 times, and results of the four involved methods were compared. Result indicates that deep learning model identifies 6 gamma rays contained in the spectrum correctly and predicts the activities of four nuclides with less than 10% error. In contrast, peak-analysis method incorrectly treats the scattering structure in the low energy spectrum region as a gamma ray related peak, and the activity evaluation errors of the three compared methods were all relatively high. The reasons of the performance difference among the involved methods were further discussed. The peak-analysis method utilizes only the peak region characteristic in the spectrum ignoring the rest part, so the weak and overlapped peaks are easy to be missed and false peak structures can also be identified incorrectly. Meanwhile, its activity assessing process involves continuum subtraction and net counts fitting procedures, which could introduce high uncertainty. Spectrum-reconstruction method though reconstructs the nuclide emitting spectrum based on whole measured spectrum, and it cannot ensure its accuracy due to the numerical difficulty of the inverse-solving problem. Shallow ANN model, when applied to nuclides identification among limited nuclide categories, shows good results in previous research. However, it’s unable to retain the performance when predicting more complex information in this paper, because of its very limited ability on characterizing and learning. The results of testing based on simulated as well as measured spectrum confirm the accuracy and reliability of the deep learning model for gamma spectrum analysis. Based on enhanced characteristic extracting ability and highly numerical stability, the proposed method is possible to be implemented in various radiation detecting applications in future.