Research On N/γ Discrimination Method Of EJ309 Liquid Scintillator Detector

Neutron detection is an important part of nuclear physics,particle physics,nuclear medicine and radiation protection.It is widely used in military,aerospace and energy.It can provide technical support for explosive detection,cancer treatment,resource exploration,engine temperature measurement,petroleum logging,imaging security inspection and material analysis.Organic scintillation detectors have high neutron detection efficiency,short decay time,and are easy to be made into various sizes and shapes,so they are widely used in neutron detection.Gamma rays are released by nuclear reactions that produce neutrons,or spontaneous nuclear fission.Gamma rays are also produced by the capture reaction and inelastic scattering of neutrons with surrounding matter.Because organic scintillators are sensitive to both neutrons（n）andγrays（γ）,n/γdiscrimination is a prerequisite for neutron detection using organic scintillators.N/γdiscrimination is usually achieved by analyzing the pulse shape.Traditional n/γdiscrimination algorithms include charge comparison method,rise time method,pulse gradient analysis,frequency gradient analysis and wavelet transform.With the development and wide application of machine learning technology,some machine learning algorithms such as artificial neural network,clustering,support vector machine are also used to realize n/γdiscrimination.In order to meet the requirements of neutron detection and neutron energy spectrum measurement in laboratory,the problem of organic scintillation n/γdiscrimination needs to be solved.In this paper,EJ309 liquid scintillation detector is used to detect ²⁴¹Am-Be neutron source,²⁵²Cf neutron source and ¹³⁷Csγray source.DT5730SB digitizer was used to collect detector pulse signal.The study of n/γdiscrimination has three aspects:first,the mechanism of scintillation,the action process of neutrons in organic scintillation and the principle of n/γdiscrimination are studied.The pulse signal characteristics of EJ309 liquid scintillation detector in n/γmixed radiation field are studied.It lays a theoretical foundation for the subsequent design and implementation of n/γdiscrimination algorithm.Second,the main principles of traditional n/γdiscrimination methods and machine learning algorithms are analyzed.Analyze the main characteristics of different machine learning algorithms.Combined with the above principle,the n/γdiscrimination application of the algorithm is realized.Third,traditional algorithms such as charge comparison method and machine learning algorithms such as clustering and artificial neural network are used to realize n/γdiscrimination.The former was evaluated by quality factor（FOM）.The latter adopts the evaluation indexes of accuracy,recall rate and F1 score.The discrimination error rate（DER）was selected as an index to compare the same type of n/γdiscrimination models.Among the traditional algorithms,charge comparison method has the best discrimination effect among charge comparison method,pulse gradient method and frequency gradient method.FOM values of ²⁵²Cf,1Ci ²⁴¹Am-Be and 20Ci ²⁴¹Am-Be neutron sources are 1.81,1.64 and 1.74.In machine learning,artificial neural network discrimination models based on Dropout regularization techniques have the advantage of fast n/γdiscrimination and are insensitive to the number of neutrons andγrays.The disadvantage is that the algorithm effect is highly correlated with the training data set.Compared with decision tree and support vector machine,this model can quickly identify single particle,and its accuracy is higher.The F1 score in each data set is above0.995.The discrimination performance in different data sets is also more stable and training is not easy to overfit.It can realize rapid identification in the fixed identification scenes such as the verification/calibration of neutron dose monitoring equipment and reference radiation field of radionuclide neutrons.The PCA-GMM discriminant model based on principal component analysis and Gaussian mixture model has the advantage that it can be directly applied to mixed radiation field without prior training.The disadvantage is that only fixed data sets can be clustered,requiring offline discrimination.The PCA-GMM discrimination model has the highest discrimination accuracy among several unsupervised algorithms,with F1 scores above 0.96.It still has good discrimination effect on the detection data in different experimental conditions,and has the strongest generalization ability compared with the supervised algorithm and other unsupervised algorithms.It can meet the discrimination requirements in different environments and experimental conditions,and realize the n/γdiscrimination model that can be applied to different discrimination scenes.