Neural networks for γ-spectra analysis

Artificial neural networks (ANNs) have a powerful representational input multi-output mapping problem, e.g. in clustering, pattern recognition and identification areas, particularly when combined with some a priori knowledge and statistical point of view. They can be useful in spectrometry for the uranium enrichment measurement methods, where numerous approaches like model fitting or expert analysis are limited. These depend on the radiation measured: the methods most widely used developed over the past 20 years were based on the counting of the 185,7 keV peak with a sodium iodide scintillation detector or the 163,4 keV peak of ²³⁵U. But these methods depend critically of the source detector geometry. A means of improving the above conventional methods is to reduce the region of interest: it is possible by focusing at the region called KαX where the three elementary components are present. The measurement of these components in mixtures leads to the isotope ratio ²³⁵U/²³⁵U+²³⁶U+²³⁸ U. In this paper the authors explore statistical orientations and their consequences for “neural” parameters. The authors show that these decisions are induced by a log-linear model; a special case of a GLIM (Generalized LInear Model) and correspond to a maximum likelihood estimation problem