Improving the performance of pulse shape discrimination using feature selection preprocessing to scintillation counters

To meet the nuclear monitoring needs of neutron detection, a feature selection preprocessing was proposed to improve the performance of pulse shape discrimination (PSD) for scintillation counters. Feature selection was implemented by constructing a new pulse shape parameter (PSP) through feature weighting. These weights were assigned by leveraging the classification information, including the model difference and the standard deviations of each model, extracted from the reference pulse model (RPM) of typical neutron and gamma events. Two feature weighting methods were used, the model difference thresholding (MDT) for maximizing inter-class separability, and the model difference combined standard deviation (MDSD) to balance inter-class separability with intra-class compactness. To examine the method, the digitized signal waveforms from fast digitizer were analyzed, and the PSPs of PSD methods with and without feature selection were compared, including the charge comparison method (CCM) and the neutron gamma model analysis (NGMA). The figure of merit (FoM) was used to assess the discrimination performance, and the results indicated that this feature selection preprocessing can significantly improve the classification performance. Moreover, the optimization of classification performance is primarily contributed by inter-class separability in the low-energy region, whereas in the high-energy region, it comes from intra-class compactness.