Application of neural networks in the analyses of gamma spectra collected during UAV flights

SCK•CEN Mentor

Camps Johan, jcamps@sckcen.be, +32 (0)14 33 27 61

Expert group

Crisis Management and Decision Support

Introduction

The analyses of gamma-ray spectra,  i.e. the identification and quantification of the different radionuclides observed, continuous to be an active field of research. Traditionally, this is done by the analyses of the full energy peaks observed in the spectrum, for identification sometimes in combination with other spectral features, such as Compton edges, sum peaks, etc… . Depending on the energy resolution of the detector system, overlapping peaks can be an important challenge in the analyses of the spectra, especially when many radionuclides are observed and spectra become complex. Also spectra with low statistics, influenced by background, are a challenge for analyses, including the determination of a Minimal Detectable Activity in case of non-detection. With the introduction of Artificial Intelligence (AI) techniques, such as artificial neural networks, it has been demonstrated that these techniques can be very interesting in relation with spectral analyses [1,2]. Because such machine learning algorithms use abstract features of the spectrum, as for example the shape of overlapping peaks and the Compton continuum, they seem a natural choice for analyzing radionuclide mixtures or the presence of artificial radionuclides in a natural background. This last situation is something typically encountered in environmental measurements and especially when the measurement time is limited.

Radiological measurements using UAVs or drones, collect in general spectra for short acquisition times to allow, if the radiological information is combined with the location, the construction of radiological maps (e.g. contamination maps, the localization of source(s), ...). Acquisition times are often limited to a few seconds (1 - 5 s). Even when combining multiple spectra, the total time in which the detector is near the source during a UAV survey can be very limited. Consequently, identification and also quantification of artificial radionuclides in such surveys is a challenging task. In this master thesis we will explore the use of artificial neural networks to identify single radionuclides in typical UAV surveys collected with a CsI scintillator detector under different natural backgrounds. 

[1] V. Pilato, F. Tola, J.M. Martinez, M. Huver, Application of neural networks to quantitative spectrometry analysis, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, Volume 422, Issues 1–3, 1999, Pages 423-427, ISSN 0168-9002, https://doi.org/10.1016/S0168-9002(98)01110-3.

[2] M. Kamuda, J. Stinnett and C. J. Sullivan, "Automated Isotope Identification Algorithm Using Artificial Neural Networks," in IEEE Transactions on Nuclear Science, vol. 64, no. 7, pp. 1858-1864, July 2017, doi: 10.1109/TNS.2017.2693152.

 

Objective

The following objectives are foreseen during the master thesis (changes are possible if in the course of the research motivated insights are gained to walk a different route):

1/ a literature review on gamma-ray spectral analyses using AI techniques, especially artificial neural networks. Focus can be on methods appllied for scintilation detectors and for environmental measurements, in which a varying natural background can be encountered. If available literature, also the application of such techniques in low counting statistics can be explored;

2/ delineation of the research question: e.g. in relation to single/multiple radionuclides (which radionuclides), which background conditions, identfication or identification and quantification of radionuclides, including problem of MDA, ... with the goal to define the training and validation dataset;

3/ construction of an artificial neural network (preferable using Phyton);

4/ collecting of training dataset (for this some experimental work will be required setting-up an automated system for collecting spectra with sources at for example different distances/positions);

5/ training of artifical neural network and optimization of the network;

6/ validation of method using real UAV data (which will be collected especially for this goal or for which data from previous campaigns will be used);

7/ writing thesis and presentation of work.

The minimum diploma level of the candidate needs to be

Academic bachelor

The candidate needs to have a background in

Physics , Informatics , Mathematics , Radiation physics and radiation detection