Presentasjon av masteroppgave: Maren Rasmussen
Data selection for intensity mapping experiments using machine learning
The CO Mapping Array Pathfinder (COMAP) experiment is an intensity mapping experiment, currently in an early stage where learning how to process the data is one of the top points on the priority list. This includes the process of data selection, where bad data are discarded. In the COMAP experiment, two important challenges are weather contaminated observations and observations containing spikes. This thesis aims to use machine learning in the form of one-dimensional convolutional neural networks for sorting out data that have been contaminated by the weather. It also aims to develop an algorithm for detecting spikes in the data.
Through the thesis we test different preprocessing techniques of the raw telescope data before feeding it into a convolutional neural network. We also fine-tune the convolutional neural network, and find that we can change the cutoff value for predicting bad weather to ensure that few bad weather samples are classified as good weather. We have found that with a cutoff value of 0.23 for the probability of bad weather, we get a training accuracy of 93.2%, a validation accuracy of 95.5%, and a testing accuracy of 97.0%. Hence we have shown that machine learning can be used for distinguishing between the observations that are contaminated by the weather, and the observations that are not. We have also found that 37.9% of the data are somewhat contaminated by the weather. By using a robust spike detection algorithm, called Smoothed z-score, we further have found that 66.3% of the data contain at least one spike. Most of the data contain only a few spikes, but some contain more than 2000 spikes. The spikes come in a variety of widths and amplitudes. Finally, we have looked at cross-spectrum of our current COMAP data, and have seen that there are still some challenges that need to be sorted out.
Medveileder: Professor Ingunn Katrine Wehus, Institutt for teoretisk astrofysikk, UiO
Intern sensor: Professor David Fonesca Mota, Institutt for teoretisk astrofysikk, UiO
Ekstern sensor: Forsker Elina Keihänen, Universitetet i Helsinki