International Society for Industrial Process Tomography

Machine Learning and Algebraic Reconstruction Methods
for Gamma-ray Spectral Analysis

Icaro V. M. Moreira^1*, Silvio de Barros Melo¹, Ilker Meric², Carlos Costa Dantas³, Geir Anton Johansen²

¹Centro de Informatica - UFPE, Recife, Brazil

²Department of Electrical Engineer - HVL, Bergen, Norway

³Departamento de Energia Nuclear - Recife, Brazil

^*Email: ivmm@cin.ufpe.br

ABSTRACT

Prompt Gamma-ray Neutron Activation Analysis (PGNAA), a technique widely applied in several industry areas, is based on the use of nuclear reaction in a sample induced by neutron beam bombardment, followed by measurements of the emitted gamma ray to identify and quantify different elements in the sample. Due to the complexity and amount of noise present in this data, PGNAA is usually applied in controlled environments where assumptions can be made about the sample composition. This article proposes a solution composed of two approaches which have the potential to address the main problems found in this field. The first one applies a machine learning algorithm, multi-label classifier, to identify the elements in the sample, allowing us to remove the a priori knowledge about the sample composition, typically required in current solutions. The second one is the use of algebraic reconstruction methods to quantify the amount of each component in the sample, where these methods are already widely applied in industrial tomography because of their performance and their capability of handling ill-conditioned and noise data. The method proposed was able to correctly identify all elements in the sample and achieved an almost perfect estimation of the weight fraction of each one of these elements for a data set composed of simulated spectra with different concentrations of ²⁴¹Am, ²²Na, and ¹³⁷Cs.

Keywords Machine Learning; Multi-label Classification; Computed Tomography; Algebraic Reconstruction Methods

Industrial ApplicationOil and Gas