Now showing 1 - 1 of 1
- PublicationIdentifying Muon-Produced Cherenkov Ring Images in VERITAS Data Using Convolutional Neural Networks(University College Dublin. School of Physics, 2021)In this thesis, the use of convolutional neural networks (CNNs) for identifying muon ring images from the VERITAS gamma-ray telescope is investigated. Muon images are important to identify due to their use in the calibration of the telescope and being a source of background which can sometimes be misidentified as being produced by a gamma ray. Currently muon images are identified using an algorithm in the VEGAS analysis software which was developed for VERITAS. It is hoped that CNNs may be able to improve upon this algorithm and increase the efficiency of the detection of these images. In order to train a convolutional neural network for this purpose, a labelled dataset is required. Multiple datasets were generated and tested for this purpose, using VEGAS-labelled data, simulated data, and data from the Muon Hunters 2 citizen science project. Extra information in the simulations was used to generate labels for the simulated data, while volunteer votes were used to label the Muon Hunters 2 data. For this purpose the volunteer votes were first analysed in order to determine how best to use them to generate hard labels for the images. It was found by applying the trained models to an expert-labelled test dataset that CNNs are more effective than the current algorithm at identifying muon images. Particularly the models trained on Muon Hunters 2 data produced the highest accuracy and AUROC (area under receiver operating characteristic) values, with AUROC being an indication of how well the model can separate muon and non-muon images across a range of output decision boundaries. A model trained on Muon Hunters 2 data identified approximately 30 times the number of muon images that the current algorithm does when applied to an independent expert-labelled test dataset. This was achieved with the output boundary upon which classifications were based adjusted so as to eliminate false positives from the model predictions. One of the benefits of the CNN model is its ability to identify various types of rings other than the full, central rings identified by the current algorithm. For calibration purposes however, more work is required in order to allow these less than full rings to be used. More work is also required in order to be able to separate out muon images which are usable for calibration from those which aren't, without using existing algorithms which may also eliminate some usable images.