Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
A hybrid deep learning based algorithm for gamma spectroscopy analysis
Thesis (PhD)--Stellenbosch University, 2025.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613954545549312 |
|---|---|
| access_status_str | Open Access |
| author | Buckton, Calib |
| author2 | Wyngaardt, Shaun |
| author_browse | Buckton, Calib Wyngaardt, Shaun |
| author_facet | Wyngaardt, Shaun Buckton, Calib |
| author_sort | Buckton, Calib |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (PhD)--Stellenbosch University, 2025.
|
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/132091 |
| institution | Stellenbosch University (South Africa) |
| last_indexed | 2026-06-10T12:44:21.236Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/132091 A hybrid deep learning based algorithm for gamma spectroscopy analysis Buckton, Calib Wyngaardt, Shaun Ngxande, Mkhuseli Stellenbosch University. Faculty of Science. Dept. of Physics. Gamma ray spectrometry -- Data processing Algorithms -- Evaluation Deep learning (Machine learning) Scintillation counters -- Data processing Sodium iodide -- Optical properties Neural networks (Computer science) UCTD Thesis (PhD)--Stellenbosch University, 2025. Buckton, C. 2025. A Hybrid Deep Learning based Algorithm for Gamma Spectroscopy Analysis. Unpublished doctoral dissertation. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/8db1d4f0-1872-4d33-87fd-8b3527a44704 ENGLISH ABSTRACT: There exists a select group of unstable nuclei which undergo radioactive decay by emission of highly-energetic photons or γ-rays. An effective tool for analysing γ-emitting radioactive sources, γ spectroscopy is a common experimental technique used in high-energy physics and environmental monitoring of radiation. Analysis of various naturally-occurring radionuclides and their decay products through γ spectroscopy has provided a practice for health and safety regulation. It is a process which has seen a large amount of improvement in efficiency and optimization over the past decade. This has often been partly due to developments in detection and analysis which build upon existing experimental methods. Recently, there have been developments in machine learning based approaches, for automated and efficient detection and radioisotope "fingerprinting". That is, these methods make use of the characteristic spectra measured from radioactive isotopes to identify them. These methods have commonly consisted mostly of a computational component, a deep-learning algorithm known generally as a deep neural network (DNN), which is trained on a representative dataset of energy spectra from different isotopes, often simulation-based. Becoming increasingly more efficient, a type of network called convolutional neural networks (CNNs) are most prominent in this application, due to their ability to effectively learn useful features from high-quality data and being able to classify or recognise radioactive species with very good accuracy. However, a challenge which can occur is when there are multiple sources present in a sample of data, as opposed to just a single source, where the accuracy of the network will often decrease as the number of sources increases. Additionally, developing a CNN model which is computationally efficient, but also accurate and robust to noise and other environmental influences can be difficult. This work sees the development of custom deep-learning algorithms for analysis of energy spectra resulting from γ-emitting radionuclides. Apart from the ability to accurately predict the presence of different isotopes from spectral data, a CNN possesses the ability to quantify the proportion or contribution of multiple sources from a mixture of isotopes. Another model provides a denoising method by predicting a spectrum which contains less statistical noise than the provided spectrum. Here, a real-time capable network can improve the signal-to-noise ratio of a given spectrum and apply some smoothing. Finally, an algorithm will be able to separate a multiple-isotope sample spectrum into the individual constituents, as a form of blind source separation. This new analysis method offers the ability to observe the different statistical contributions of multiple isotopes in a spectrum, and effectively isolate each isotope’s measured spectrum. The algorithms are computationally efficient, and trained and tested on simulated data from a Monte Carlo simulation, designed after a simple γ-spectroscopy experiment. Hence, a comprehensible dataset on a select few radionuclides is also developed. AFRIKAANSE OPSOMMING: Daar bestaan ’n uitgesoekte groep onstabiele kerne wat radioaktiewe verval ondergaan deur die vrystelling van hoë fotone of γ-strale. ’n Effektiewe hulpmid- del vir die ontleding van γ-uitstralende radioaktiewe bronne, γ spektroskopie is ’n algemene eksperimentele tegniek wat gebruik word in hoë-energie fisika en omgewingsmonitering van bestraling. Ontleding van verskeie natuurlik- voorkomende radionukliede en hul vervalprodukte deur γ-spektroskopie het ’n praktyk vir gesondheid- en veiligheidsregulering verskaf. Dit is ’n proses wat die afgelope dekade ’n groot mate van verbetering in doeltreffendheid en optimali- sering gesien het. Dit was dikwels deels te wyte aan ontwikkelings in opsporing en analise wat voortbou op bestaande eksperimentele metodes. Onlangs was daar ontwikkelings in masjienleer-gebaseerde benaderings, vir outomatiese en doeltreffende opsporing en radio-isotoop "vingerafdrukke". Hierdie metodes bestaan gewoonlik meestal uit ’n berekeningskomponent, ’n diep-leer-algoritme wat algemeen bekend staan as ’n diep neurale netwerk (DNN), wat opgelei word op ’n verteenwoordigende datastel van energiespektra van verskillende isotope, dikwels simulasie-gebaseer. ’n Tipe netwerk genaamd "convolutional neural networks"(CNNs) word al hoe meer doeltreffend, as gevolg van hul vermoë om effektief nuttige kenmerke van hoë kwaliteit data te leer en radioaktiewe spesies met baie goeie akkuraatheid te klassifiseer of te herken. ’n Uitdaging wat egter kan voorkom, is wanneer daar verskeie bronne in ’n steekproef van data teen- woordig is, in teenstelling met net ’n enkele bron, waar die akkuraatheid van die netwerk dikwels sal afneem namate die aantal bronne toeneem. Gevolglik kan die ontwikkeling van ’n CNN-model wat rekenaardoeltreffend is, maar ook akkuraat en robuust is vir geraas en ander omgewingsinvloede moeilik wees. Hierdie werk behels die ontwikkeling van ’n pasgemaakte diepleeralgoritme vir ontleding van energiespektra wat voortspruit uit γ-emitterende radionukliede. Afgesien van die vermoë om die teenwoordigheid van verskillende isotope akkuraat uit spektrale data te voorspel, beskik dit oor die vermoë om die proporsie of bydrae van veelvuldige bronne uit ’n mengsel van isotope te kwantifiseer, asook om ’n denoiseringsmetode te verskaf deur ’n spektrum te voorspel wat minder statistiese bevat geraas as die huidige spektrum. Laastens sal die algoritme in staat wees om ’n meervoudige-isotoopmonsterspektrum in die individuele bestanddele te skei, as ’n vorm van blindebronskeiding. Die algoritme is rekenkundig doeltreffend, en opgelei en getoets op gesimuleerde data van ’n Monte Carlo-simulasie, ontwerp na ’n eenvoudige γ-spektroskopie- eksperiment. Gevolglik word ’n verstaanbare datastel oor ’n uitgesoekte paar radionukliede ontwikkel. Doctoral 2025-05-23T07:30:48Z 2025-05-23T07:30:48Z 2025-03 Thesis https://scholar.sun.ac.za/handle/10019.1/132091 Stellenbosch University xvi, 124 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Gamma ray spectrometry -- Data processing Algorithms -- Evaluation Deep learning (Machine learning) Scintillation counters -- Data processing Sodium iodide -- Optical properties Neural networks (Computer science) UCTD Buckton, Calib A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title | A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title_full | A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title_fullStr | A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title_full_unstemmed | A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title_short | A hybrid deep learning based algorithm for gamma spectroscopy analysis |
| title_sort | hybrid deep learning based algorithm for gamma spectroscopy analysis |
| topic | Gamma ray spectrometry -- Data processing Algorithms -- Evaluation Deep learning (Machine learning) Scintillation counters -- Data processing Sodium iodide -- Optical properties Neural networks (Computer science) UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/132091 |
| work_keys_str_mv | AT bucktoncalib ahybriddeeplearningbasedalgorithmforgammaspectroscopyanalysis AT bucktoncalib hybriddeeplearningbasedalgorithmforgammaspectroscopyanalysis |