Neutron-gamma discrimination based on STFT-DFF model and FPGA implementation

In the field of neutron-gamma discrimination, traditional discrimination algorithms exhibit poor performance in the low-energy region, while machine learning-based discrimination algorithm face challenges such as difficulty in hardware deployment and poor real-time discrimination. To address these issues, this study proposes a machine learning-based discrimination algorithm named STFT-DFF, which integrates the Short-Time Fourier Transform (STFT) and Dynamic Feature Fusion (DFF), and realizes real-time discrimination deployment on an Field-Programmable Gate Array (FPGA). First, the algorithm converts pulse signals into one-dimensional time-frequency feature vectors using STFT to extract the time-frequency feature differences between neutrons and gamma-rays. Then, a lightweight DFF model is utilized to generate dynamic weights through an attention mechanism, which adaptively enhances key time-frequency features and suppresses noise interference. Two types of datasets were used in this study for comparative experiments: one is the time-of-flight labeled ²³⁸Pu-⁹Be standard dataset, and the other is the ²⁴¹Am-⁹Be neutron source dataset independently measured using an EJ309 detector. Experimental results show that under both datasets, the Figure of Merit (FOM) value of the STFT-DFF model is significantly improved compared with the Charge Comparison Method (CCM). Moreover, after deployment on an FPGA, the real-time pulse discrimination processing efficiency of the algorithm can reach more than 20,000 pulses per second, demonstrating excellent real-time processing capability.