Neutron spectrum unfolding based on generalized regression neural networks for neutron fluence and neutron ambient dose equivalent estimations.

Abstract

Neutron fluence and neutron ambient dose equivalent, H*(10), are important physical quantities for neutron radiation protection and monitoring. They can be deduced from neutron spectrum, which is usually measured by multisphere system with proper unfolding methods. Novel unfolding methods on the basis of artificial intelligence, mainly artificial neural networks (ANNs), have been researched and developed. However, without normalization on network inputs, ANNs can not be applied to accommodate demands of various neutron field measurements for neutron spectrum unfolding in practice, because the neutron spectra for training the ANNs are mostly extracted from IAEA (2001), the integrals of which over neutron energy are unit fluences. Moreover, derived from an unfolded normalized spectrum, the true values of neutron fluence and H*(10) are never to know. In this work, three normalization methods-zero-mean normalization method, min-max normalization method, and maximum-divided normalization method were used to process with the inputs of generalized regression neural networks (GRNNs), and a new method was proposed for neutron fluence and H*(10) estimations derived from unfolded neutron spectrum based on GRNNs for the first time. Sixty-three neutron spectra were unfolded based on GRNNs with use of three normalization methods, and the corresponding neutron fluences and H*(10) were obtained and compared. From the testing results, the GRNNs with the maximum-divided method is most effective to unfold neutron spectrum and to evaluate neutron fluence and H*(10). The feasibility of the method was further studied through experiments by using Bonner sphere spectrometer in well characterized 241Am-Be neutron field.