Research On Nuclide Recognition Algorithm Based On Convolutional Neural Network

With the development and wide application of nuclear technology,?energy spectrum analysis technology plays an important role in environmental radioactivity research,nuclear physics research,geological exploration,etc.,which is of great significance to national security and social stability.The?-ray spectrum analysis technology mainly identifies the species of nuclides and analyzes the radioactivity of the nuclides based on the?-ray spectrum data.The application of nuclear technology from special fields to today's wide applications in the medical,industrial,and transportation fields has made it necessary to perform more accurate and accurate nuclide identification.NaI detectors have become the most widely used detectors due to their low price and high detection efficiency.However,this type of detector has low energy resolution,poor recognition ability for weak peaks,and difficulty in distinguishing the spectrum of energy spectra caused by nuclei with similar peaks.Aiming at the problem that the?-ray spectrum identification method generated by the traditional NaI detector having low accuracy rate for complex?-ray spectrum analysis,an artificial neural network?ANN?nuclides identification method is proposed.The artificial neural network is a mathematical model that mimics the structure and function of the biological neuron network,which consists of artificial neurons connected to each other.It can be learned and summarized by known data,and is widely used in economics,information,medicine,automation,etc.In all major fields,it has great development potential and application market.The Convolutional Neural Network?CNN?is a widely used form of deep artificial neural network.Its local connection,pooling operation and weight sharing can effectively reduce the complexity of the network and reduce the parameters of the model.And it makes the model have a certain degree of invariance to translation,scaling and distortion,has strong adaptability,and being easy to train and optimize.Convolutional neural networks perform better in signal and information processing tasks than standard fully-connected neural networks,and it improves the traditional shallow neural network learning ability and basically solves issues like gradient disappearance.The main content of this paper is:using Monte Carlo simulation software MCNP5to simulate the gamma energy spectrum response of NaI detector to nine kinds of point sources(¹³⁷Cs,²⁴¹Am,⁶⁰Co,etc.)under four different conditions:the number of different point sources?this article Set to 1³ quantities?,different point source combinations,different distances,and different particle numbers.The spectrum of the simulation results is analyzed,the gamma ray emission law is studied,and the simple spectrum and the complex spectrum are compared and analyzed,which provides a basis for the research of?-ray and nuclides identification algorithms.And based on MPI,the parallel calculation of MCNP5 program can effectively improve the calculation efficiency of the program and shorten the calculation time.The convolutional neural network is introduced to identify the nuclide in the gamma ray spectrum analysis.The four sets of gamma spectroscopy data of Monte Carlo simulation results are divided into training set and testing set respectively,and then one-dimensional based on the Keras deep learning library of python front end.Convolutional neural network consisting of a series of convolutional layers?including pooling?and fully connected layers,the convolutional layer is used for feature extraction,and the fully connected layer is used as a classifier.The activation function uses the softmax function to perform sigmoid-type when performing multi-classification tasks,and functions when performing multi-tag tasks,and uses dropout to reduce the risk of overfitting.Normalize the input data,convert the tag into a unique heat code,and then input the network for GPU acceleration training,which greatly reduces the training time and effectively shortens the model hyper-parameter debugging cycle.The?energy spectrum was analyzed,and the recognition rate was analyzed for different situations of 1³ points sources,whose recognition rate was 100%.The nucleus recognition rate of different smoothness?spectrum was compared and analyzed,whose spectrum recognition rate is 98.9%.The generalization performance of convolutional neural networks for various cases is explored,and the results show that the generalization performance of convolutional neural networks is very good.