Explainable Neural Network Algorithm for Rapid Radionuclide Identification Under Low Count Gamma-Ray Spectrum Data

Real-time and automatic radioisotope identification using gamma spectrum is an important issue in the field of nuclear safety. It is widely used in vehicle mounted radioisotope monitoring, Marine radioisotope monitoring and nuclear decommissioning verification scenarios. At present, the focus of radionuclide identification is fast and stable recognition under low count conditions. In this paper, a radionuclide recognition method based explainable artificial neural network is proposed, and a synthetic gamma spectrum data set is created. The data set contains gamma-ray spectra of 12 different types of radionuclides, which were obtained by Geant4 Monte Carlo simulation software and gaussian broadening of the detector. Data Augmentation was achieved by simulating gamma spectra at different measuring times, different measuring distances and different ambient temperatures. The training results of the neural network optimized by hyperparameter show that it has a high accuracy on the test set with shorter measurement time, longer measurement distance and larger energy spectrum drift range, which provides a method for rapid identification of nuclides in the case of low count. Using t-SNE dimension reduction technology, the twelve dimensions data output by the neural network is reduced to two dimensions for feature visualization, which vividly explains and verifies the recognition results of neural network.