Research On Nuclide Recognition Based On Recurrent Neural Network And Attention Mechanism

In the context of the severe international nuclear deterrence situation,emergency response to nuclear leakage,environmental radiation monitoring and many other radionuclide detection backgrounds,nuclide identification technology is one of the key means to identify the types of radionuclides.In a real environment,radionuclide detection is often affected by factors such as the level of its own radioactivity and background noise,making it more difficult to identify the nuclide.Therefore,studying how to quickly and accurately identify species of nuclides in real noise environments has very great theoretical significance and practical application value.This paper studies the radionuclide identification method,which does not rely on the traditional peak-seeking nuclide identification process,but focuses on the research of the rapid nuclide identification method based on the neural network and the research of the multi-nuclide identification method,aiming to improve the nuclide identification Speed and accuracy.The main work content of this paper includes the following aspects:(1)Aiming at the problem of determining the existence of nuclides with low levels of radioactivity in a short time,research on the rapid identification of nuclides based on neural networks.A fast nuclide recognition method based on Bi-directional Long Short-Term Memory Neural Network(Bi-LSTM NN)is proposed.Through the study of energy spectrum data smoothing,normalization and two-dimensional conversion,and build a nuclide recognition model based on bi-directional long and short-term memory neural network to achieve the purpose of quickly and accurately judging whether radionuclides exist in the environment.By comparing with the traditional peak-finding nuclide identification method,the average identification speed is2.1645 s,while the average identification speed of the proposed method is 11.2ms.Compared with the nuclide identification method based on BP neural network,the average missing alarm rate is 35.9% and the average false alarm rate is 50.45%,while the proposed method is 3.66%and 1.86% respectively.The experimental results show that the proposed method performs better in the efficiency of rapid nuclide recognition.(2)Aiming at the problem of low recognition accuracy caused by the mixed overlapping peaks of multiple radionuclides,a neural network-based multi-nuclide recognition method is studied.A multi-nuclides recognition method based on Bi-directional Gated Recurrent Unit Attention(Bi-GRU-Attention)is proposed.After energy spectrum data preprocessing,the bidirectional feature extraction of each time state is proposed,and The attention mechanism is used to further enhance the effect of multi-nuclides recognition.The experiment uses the energy spectrum data set simulated by MNCP4 to group the mixed nuclide energy spectrum for nuclide identification.The results show that the proposed method performs better in multi-nuclide identification,with Macro-P reaching 95.76%,compared with Macro-F1 The multi-nuclides recognition method based on convolutional neural network improves by 12.55%.(3)In order to verify the application effect of the proposed nuclide identification method under real noise environment,this paper designs and builds a radionuclide real-time warning system based on recurrent neural network.After experimental tests,the results show that the system proposed in this paper can The radionuclide responds quickly and can identify a variety of mixed radionuclides.