| At present,the research on the reconstruction of the radioactive distribution mainly includes two aspects:the radiation field dose evaluation and the nondestructive assay of radioactive distribution in radioactive waste.The semiempirical formulas used in the radiation field dose evaluation of the nuclear facilities mainly depend on the accuracy of various buildup factors.While,the non-destructive assay technology related to radioactive distribution of nuclear waste is costly,and timeconsuming by means of complicated devices and high activity radioactive sources,which means the application range of the non-destructive assay is immensely confined.Thus,a passive non-destructive assay technology based on convolutional neural network is given in this paper,to identify the nuclides and accurately reconstruct the radioactive distribution in solid materials.This method can reconstruct the radioactive distribution of the solid waste with given geometry and material information by merely detecting photon counting.In this study,the data set of "Radioactive Distribution——Detector Count" was made by Monte Carlo method for 58Co,60Co,51Cr,137Cs,152Eu,59Fe,54Mn,124Sb,65Zn and 95Zr nuclides in nuclear waste canister,and the convolutional neural network model was established according to the characteristics of the radioactive distribution.The model of neural network was trained based on the established data set.Then,the network model was verified by a self-designed example,and the radioactive distributions of all nuclides in nuclear waste canister were reconstructed.The reconstruction results show good performance in two evaluation parameters:average relative error and maximum relative error.In addition,by comparing the reconstruction results of different nuclides in nuclear waste canister with five different detector arrangements,an optimized design of 34 points arrangement was explored to coordinate the balance between the detector arrangements and the reconstruction accuracy.According to this optimized design,the data set of nuclides in nuclear waste container,such as 58Co,60Co,51Cr,137Cs,152Eu,59Fe,54Mn,124Sb,65Zn and 95Zr,was built.The convolution neural network model was designed based on the data set to reconstruct the radioactive distribution of self-designed examples,and the results were compared with the expected values of radioactive distribution.The results show that the average relative errors of reconstructed distributions of these radioactive nuclides are all below 5%,while the maximum relative errors are 7%-12%.The reconstruction of radioactive distribution is relatively accurate.This work has practical guiding significance for the reconstruction of radioactive distribution.Compared with traditional methods,convolution neural network holds manifest superiorities in the cost,time-consuming,device simplicity and reconstruction accuracy.Given the above,it can be widely used in decommissioning of nuclear facilities,emergency radiation monitoring and other fields. |
