Study on n/γ Discrimination Method Based on Deep Metric Learning

Abstract: Precise identification of neutrons and gamma rays in complex radiation environments is a key technology in the fields of nuclear material detection and nuclear safety. The current mainstream end-to-end neural network classification models have the problem of insufficient generalization ability when facing the jitter of the temporal characteristics of electronic waveforms. Therefore, an n/γ discrimination method based on deep metric learning is proposed. An improved convolutional neural network structure is designed based on LENet5. Combined with the constraint of the triplet loss function, the discriminative feature space of neutron and gamma nucleus pulses is obtained, achieving efficient discrimination of neutrons and gamma rays. Training was conducted using a mixed dataset of data collected by the hardware core-imitating pulse generator and software simulation data. Quantitative tests were carried out respectively on the simulation data and the measured pulse data with jitter of the electronic waveform time characteristics, and comparisons were made with the traditional charge comparison method and the traditional CNN discrimination method. The results show that in the simulation data, the quality factor (FOM) of the method proposed in this paper has increased by 247.7% compared with the charge comparison method. The discrimination error rate in the measured pulse data is reduced by 70% compared with the traditional CNN discrimination method. This method can effectively solve the problem of low generalization ability in the traditional CNN discrimination method, providing a new idea for the high-precision particle discrimination technology in the radiation detection system.