Pulse pile-up recognition using multi-module DenseNet in neutron-gamma discrimination

Neutron-gamma discrimination is crucial for various applications in nuclear science and technology. Currently, the majority of research is focused on pulse shape discrimination, and conventional methods achieve a certain level of accuracy in conventional neutron-gamma discrimination scenarios. However, under high-count-rate conditions, neutron-gamma signals tend to pile-up, resulting in pulse shape changes, that significantly affect the accuracy of conventional methods. In recent years, neural network technology has been shown to be effective for signal waveform recognition. In this study, two Multi-Module DenseNet network structures were designed: Multi-module DenseNet (MMDenseNet) and Multi-module DenseNet with base layer Reuse (MMDenseNet-R). The accuracy and F1-score of MMDenseNet/MMDenseNet-R for recognizing piled-up pulses at different pile-up degrees and noise levels was evaluated using DenseNet and ResNet as comparison networks. Among the various pile-up cases examined in this study, MMDenseNet/MMDenseNet-R consistently outperformed ResNet and DenseNet, showing clear superiority over conventional pulse shape discrimination methods. MMDenseNet/MMDenseNet-R achieved high-precision pulse piled-up recognition under various pile-up conditions through their modular design, thereby improving the usage of piled-up pulses during detection. These network architectures are expected to acquire more valid signals in complex neutron fields, further optimizing the accuracy of particle detection.