Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
A comparison framework for deep learning RFI detection algorithms in radio astronomy
Thesis (MEng)--Stellenbosch University, 2024.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2024
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613781941551104 |
|---|---|
| access_status_str | Open Access |
| author | Du Toit, Charl |
| author2 | Ludick, Danie |
| author_browse | Du Toit, Charl Ludick, Danie |
| author_facet | Ludick, Danie Du Toit, Charl |
| author_sort | Du Toit, Charl |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2024. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/130199 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:41:36.774Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2024 |
| publishDateRange | 2024 |
| publishDateSort | 2024 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/130199 A comparison framework for deep learning RFI detection algorithms in radio astronomy Du Toit, Charl Ludick, Danie Grobler, Trienko Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Radio -- Interference Electromagnetic waves Radio astronomy Neural networks (Computer science) Demodulation (Electronics) UCTD Thesis (MEng)--Stellenbosch University, 2024. ENGLISH ABSTRACT: Radio frequency interference (RFI) are man made electromagnetic signals that disrupt the signal of interest. The detection of RFI is an essential step in the radio astronomy reduction pipeline. Many statistical and deep learning detection algorithms approach the problem from an image processing background, by identifying anomalies in the spectrograms of cross-correlated radio antenna measurements. In this thesis we compared supervised fully convolutional neural networks (CNNs) in terms of classification accuracy and computational complexity. We compared these CNNs (U-Net, AC-Unet, R-Net, RFI-Net, DSC-Dual- Resunet) to the popular SumThreshold method, as implemented in the AOFlagger software. In addition, we developed two new CNNs: ASPP-Unet and DSC-Mono-Resunet. We used observations from the Low Frequency Array (LOFAR) and simulated observations from the Hydrogen Epoch of Reionization Array (HERA) to construct our datasets. We compared loss functions (Mean Squared Error (MSE), Binary Crossentropy (BCE), Dice and log-cosh-dice), regularization techniques, training methods and trends between the datasets. The best regularization technique employed both L2 kernel regularization and dropout. We selected the Dice loss function since it maximised the F1 score. However, the MSE loss function would be a better alternative if a higher recall is desired. We found that training these CNNs with unreliable AOFlagger reference outputs performed better than AOFlagger, when tested with reliable reference outputs. The low capacity CNNs struggled to learn the AOFlagger labels (especially when the labels are worng) and as such performed better during testing. The R-Net7 architecture is the most computationally efficient and achieves the highest accuracy if the test set is out of distribution (OOD). RFI-Net achieves the highest accuracy if the test set is within distribution. The two architectures we developed did not yield any significant benefit. As a training method, models were initially trained with AOFlagger reference outputs and then fine tuned (retrained) with little hand-labeled (LOFAR), or pixel-perfect simulation (HERA), reference outputs. Fine tuning proved to be an effective method in this case as there were little hand labeled data available. Fine tuning was more effective for the HERA data compared to the LOFAR data due to the weaker AOFlagger strategy employed, the low variance of the dataset and the easily identifiable RFI features present in the simulated data. AFRIKAANSE OPSOMMING: Radiofrekwensie-interferensie (RFI) is mensgemaakte elketromagnetiese seine wat die sein van belang ontwrig. Die opsporing van RFI is ’n noodsaaklike stap in die radiosterrekunde verwerkings pyplyn. Baie statistiese en diep-leer-opsporingsalgoritmes los die probleem op deur beeldverwerking te gebruik, m.a.w. deur uitskieters in die spektrogramme wat verkry is met behulp van kruisgekorreleerde antenna-afmetings te identifiseer. In hierdie tesis het ons gerigte volledige konvolusionele neurale netwerke (KVN’s) vergelyk in terme van klassifikasie akkuraatheid en bewerkingskompleksiteit. Ons het hierdie KVN’s (U-Net, ACUnet, R-Net, RFI-Net, DSC-Dual-Resunet) vergelyk met die gewilde SomDrempel-metode, soos ge¨ımplementeer in die AOFlagger-sagteware. Ons het twee nuwe KVN’s ontwikkel: ASPP-Unet en DSC-Mono-Resunet. Ons het waarnemings van die Lae Frekwensie Skikking (LOFAR) en gesimuleerde waarnemings van die Waterstof Era Herioniseering Skikking(HERA) gebruik om ons datastelle op te stel. Ons het verliesfunksies (Gemiddelde Kwadratiese Fout (MSE), Binˆere Kruisentropie (BCE), Dobbelsteen en log-kosh-dice), regularisasietegnieke, opleidingsmetodes en tendense tussen die datastelle vergelyk. Die beste regularisasietegniek het beide L2 kern-regularisasie en uitval (dropout) gebruik. Ons het die Dobbelsteen verliesfunksie gekies aangesien dit die F1-score gemaksimeer het. Die MSE verliesffunksie sou egter ’n beter alternatief wees indien ’n ho¨er herroepring (recall) verkies word. Ons het bevind dat opleiding van hierdie KVN’s met onbetroubare AOFlagger-verwysings-uitsette beter presteer het as AOFlagger, indien dit getoets is met betroubare verwysings-uitsette. Die KVN’s met ’n lae kapasiteit het gesukkel om AOFlagger-etikette aan te leer (veral wanneer AOFLAGGER verkeerd is) en dus het dit beter presteer tydens toetsing. Die R-Net7-argitektuur benodig die minste bewerkings van al die argitekture en behaal die hoogste akkuraatheid as die toetsstel buite distribusie is. RFI-Net behaal die hoogste akkuraatheid as die toetsstel binne distribusie is. Die twee argitekture wat ons ontwikkel het, het geen beduidende voordeel opgelewer nie. As ’n afrigtingsmetode, is modelle aanvanklik afgerig met AOFlagger-verwysings-uitsette en daarna fynafgestem (heropgelei) met min handgemerkte (LOFAR) of pixelperfekte simulasie (HERA) verwysings-uitsette. Die resultate het gewys dat Fynafstelling in hierdie geval effektief was, omdat daar min handgemerkte data beskikbaar was. Fynafstelling was meer effektief vir die HERA-data in vergelyking met die LOFAR-data as gevolg van die swakker AOFlagger-strategie wat gebruik is, die lae intrinsieke variansie van die datastel, en die maklik identifiseerbare RFI-kenmerke wat in die gesimuleerde data teenwoordig was. Masters 2024-02-29T07:05:27Z 2024-04-26T08:54:51Z 2024-02-29T07:05:27Z 2024-04-26T08:54:51Z 2024-03 Thesis https://scholar.sun.ac.za/handle/10019.1/130199 en_ZA en_ZA Stellenbosch University xxi, 165 pages : illustrations. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Radio -- Interference Electromagnetic waves Radio astronomy Neural networks (Computer science) Demodulation (Electronics) UCTD Du Toit, Charl A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title | A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title_full | A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title_fullStr | A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title_full_unstemmed | A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title_short | A comparison framework for deep learning RFI detection algorithms in radio astronomy |
| title_sort | comparison framework for deep learning rfi detection algorithms in radio astronomy |
| topic | Radio -- Interference Electromagnetic waves Radio astronomy Neural networks (Computer science) Demodulation (Electronics) UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/130199 |
| work_keys_str_mv | AT dutoitcharl acomparisonframeworkfordeeplearningrfidetectionalgorithmsinradioastronomy AT dutoitcharl comparisonframeworkfordeeplearningrfidetectionalgorithmsinradioastronomy |