A methodology for alpha/beta particles identification in Liquid Scintillation using a three-channel Convolutional Neural Network

摘要: To mitigate the dark pulse and attain low background level measurements, the liquid scintillation counter (LSC) is generally equipped with two or three photomultiplier tubes (PMT) for coincidence measurements. However, traditional identification method in the LSC only utilize the anode pulse from a single PMT to identify alpha/beta particles, which limits their ability to identify particles. We developed a three-channel Convolutional Neural Network (TCNN) model, which integrates pulses from three PMT anodes to identify particle categories. Anode pulses are organized into a shape (3,512) and subsequently fed into the TCNN for alpha/beta pulse discrimination. To train and validate TCNN, we prepared two samples: 241Am sample as alpha emitter and 90Sr/90Y sample as beta emitter. In the validation set, TCNN performed significantly better than traditional convolutional neural networks (CNN) in identifying alpha/beta pulses, achieving accuracy, recall, and F1 score of 99.44%, 99.23%, and 99.34%, respectively. We also prepared a mixed-emitter sample exhibiting a β activity of approximately 172Bq and an α activity of 98Bq to evaluate the impact of the TCNN on the spectral performance in practical applications. Firstly, the category of the pulse from the sample is identified by the TCNN, and then it’s height is recorded in an α-MCA spectrum or β-MCA spectrum according to the identified category. The alpha particle peak in the α-MCA spectrum is used to evaluate spectral performance. The optimal detection limit for the alpha particle peak is 0.3337 cps, which shows a sensitivity increase of 31.16% compared to the CNN method. This indicates that the TCNN can effectively utilize the three-channel pulses to enhance the ability to distinguish between alpha and beta particles when analyzing both simultaneously, thereby significantly improving the sensitivity of the detector.