Study on neutron-gamma discrimination methods based on GMM-KNN and LabVIEW implementation

Machine learning algorithms are considered as effective methods for improving the effectiveness of neutron-gamma (n-gamma) discrimination. This study proposed an intelligent discrimination method that combined a Gaussian mixture model (GMM) with the K-nearest neighbor (KNN) algorithm, referred to as GMM-KNN. First, the unlabeled training and test data were categorized into three energy ranges: 0-25 keV, 25-100 keV, and 100-2100 keV. Second, GMM-KNN achieved small-batch clustering in three energy intervals with only the tail integral Q(tail) and total integral Q(total) as the pulse features. Subsequently, we selected the pulses with a probability greater than 99% from the GMM clustering results to construct the training set. Finally, we improved the KNN algorithm such that GMM-KNN realized the classification and regression algorithms through the LabVIEW language. The outputs of GMM-KNN were the category or regression predictions. The proposed GMM-KNN constructed the training set using unlabeled real pulse data and realized n-gamma discrimination of Am-241-Be pulses using the LabVIEW program. The experimental results demonstrated the high robustness and flexibility of GMM-KNN. Even when using only 1/4 of the training set, the execution time of GMM-KNN was only 2021 ms, with a difference of only 0.13% compared with the results obtained on the full training set. Furthermore, GMM-KNN outperformed the charge comparison method in terms of accuracy, and correctly classified 5.52% of the ambiguous pulses. In addition, the GMM-KNN regressor achieved a higher figure of merit (FOM), with FOM values of 0.877, 1.262, and 1.020, corresponding to the three energy ranges, with a 32.08% improvement in 0-25 keV. In conclusion, the GMM-KNN algorithm demonstrates accurate and readily deployable real-time n-gamma discrimination performance, rendering it suitable for on-site analysis.