Automated isotope identification and quantification using artificial neural networks

Current radioisotope identification devices struggle to identify and quantify isotopes in low-resolution gamma-ray spectra in a wide range of realistic conditions. Trained gamma-ray spectroscopists typically rely on intuition when identifying isotopes in spectra. A trained gamma-ray spectroscopist can inject their intuition into pattern recognition algorithms by creating training datasets and intelligently choosing a machine learning model for a task. Algorithms based on feature extraction such as peak finding or ROI algorithms work well for well-calibrated high resolution detectors. For low-resolution detectors, it may be more beneficial to use algorithms that incorporate more abstract features of the spectrum. To investigate this, we simulated datasets and used them to train artificial neural networks (ANNs) for identification and quantification tasks using gamma-ray spectra. Because the datasets were simulated, this method can be extended to a variety of gamma-ray spectroscopy tasks. Models we investigated include dense, convolutional, and autoencoder ANNs. In this work we introduce annsa, an open source Python package capable of creating gamma-ray spectroscopy training datasets and applying machine learning models to solve spectroscopic tasks. Using annsa, we found that identification performance in simulated spectra was sensitive to the source-to-background ratio, detector gain setting, and shielding. Performance was less sensitive to the source-detector height and detector resolution. We demonstrate annsa's capabilities on a source interdiction classification problem, outperforming a peak-based Bayesian classifier for source identification. We also demonstrate annsa on a uranium enrichment quantification problem which shows an accuracy useful for homeland security applications.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Nuclear, Plasma, Radiolgc Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Kamuda, Mark M.

Contributors dc:contributor

• Huff, Kathryn
• Hasagawa-Johnson, Mark
• Kozlowski, Tomasz
• Sullivan, Clair
• Uddin, Rizwan

• gamma-ray spectroscopy
• neural networks
• machine learning

Statement dc:rights

• Copyright 2019 Mark Kamuda

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/106234

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/106234