Automatic classification of multi-channel PSD results combining unsupervised and supervised learning

Fast neutron radiography is a nondestructive testing technique used in various fields, such as homeland security. Fast neutrons can be detected using either a thermal neutron capture detection system with moderators or a recoil proton-based neutron scattering detection system. In particular, recoil proton-based neutron detection systems are more suitable for radiography systems. This study proposes a machine-learning-based algorithm that can automatically classify neutron signals obtained from multichannel organic plastic scintillators with silicon photomultipliers for radiography acquisition. It covers the entire energy range above the pulse height threshold of 150 keVee in the pulse shape discrimination (PSD) results. We categorized the PSD results into two regions: the unsupervised learning region (more than 1.7 MeVee) and the supervised learning region (less than 1.7 MeVee). Neutron signals in the supervised learning region were effectively classified by exploiting the characteristics of the data distribution in the unsupervised learning region. The algorithm was applied to the neutron and gamma-ray signals obtained by time-of-flight measurements, and an excellent classification performance was demonstrated with the area under the receiver operating characteristic curve of 0.9904.