Discriminating between full- and partial-energy deposition events in a high-purity germanium detection system can achieve the subtraction of background events and reduce the system’s minimum detectable activity (MDA). Traditional anticoincidence techniques and pulse shape discrimination techniques can both achieve event type discrimination, but the former has a complex equipment structure and is difficult to maintain, and the latter still has room for improvement in recognition accuracy. This paper proposes a method using convolutional neural networks (CNNs) to learn and recognize pulse waveform features for event type discrimination. We studied the impact of space-filling curve encoding methods and convolution kernel size on the training effect of the model. The results show that the zigzag curve encoding method results in a higher discrimination accuracy of the model; when the convolution kernel size is 9 × 9, the model has the best convergence speed and training effect. Finally, we compared the performance of fully connected neural network (FC network), 1D-CNN, and 2D-CNN in event type discrimination using MDA as the indicator. The results show that compared to FC network, CNNs perform better in reducing the MDA across the entire energy range, especially 2D-CNN, which performs better in the medium and low energy range compared to 1D-CNN. After using 2D-CNN for event discrimination and background suppression, the MDA is reduced by 21.23% (122 keV), 16.32% (244 keV), 8.43% (344 keV), 13.76% (662 keV), 5.54 % (1173 keV), and 26.35% (1332 keV).
Read full abstract