Data for: Rapid nuclide identification algorithm based on convolutional neural network

Figure 1. Illustrate of CNN architecture Figure 2. Examples of the simulated 60Co gamma-ray spectra with different count rates of peaks Figure. 3. Four transformation methods from vector mapping to a matrix Figure. 4. Convergence curves for each transformation method Figure. 5. Example of the transformed 60Co gamma-ray spectrum using Hilbert curve Figure. 6. ROC curve for the predicted results; (b) shows the details of (a) Figure. 7. P-R curve for the predicted results; (b) shows the details of (a) Figure. 8. 60Co spectra of different gross counts. (a) 9648 gross counts and (b) 1884 gross counts. Figure. 9. (a) Mixed spectrum of 137Cs and 60Co and (b) mixed spectrum of 238Pu, 137Cs, and 60Co Figure. 10. Illustration of drift gamma-ray spectra and original spectrum Figure. 11. Measured 137Cs gamma-ray spectra for evaluating CNN and BPNN Figure. 12. Accuracy of nuclide identification for (a) BPNN and (b) CNN Figure. 13. Examples spectra with a low count rate. main.m contains CNN and BPNN training. My2Dtransformation.m is a funtion for 2D transiformation, contains Hilbert and Z-order curves, vertical and horizon scanning. trainingset.mat contain all training simulated spectra for CNN and BPNN.

图1：卷积神经网络（Convolutional Neural Network, CNN）架构示意图
图2：不同峰计数率的模拟60Coγ射线能谱示例
图3：从向量映射至矩阵的四种变换方法
图4：各变换方法的收敛曲线
图5：采用希尔伯特曲线（Hilbert curve）变换后的60Coγ射线能谱示例
图6：预测结果的受试者工作特征（Receiver Operating Characteristic, ROC）曲线；(b)为(a)的细节展示
图7：预测结果的精确率-召回率（Precision-Recall, P-R）曲线；(b)为(a)的细节展示
图8：不同总计数的60Co能谱：(a) 总计数为9648，(b) 总计数为1884
图9：(a) 137Cs与60Co的混合能谱，(b) 238Pu、137Cs与60Co的混合能谱
图10：漂移γ射线能谱与原始能谱示意图
图11：用于评估卷积神经网络（CNN）与反向传播神经网络（Backpropagation Neural Network, BPNN）的实测137Csγ射线能谱
图12：核素识别准确率：(a) 反向传播神经网络（BPNN），(b) 卷积神经网络（CNN）
图13：低计数率能谱示例
main.m 包含卷积神经网络（CNN）与反向传播神经网络（BPNN）的训练代码
My2Dtransformation.m 为二维变换函数，支持希尔伯特曲线、Z阶曲线、垂直扫描与水平扫描四种变换方式
trainingset.mat 包含用于卷积神经网络（CNN）与反向传播神经网络（BPNN）训练的全部模拟能谱数据集