Machine Learning for Radio Frequency Interference Flagging

Harrison, Kyle

Machine Learning for Radio Frequency Interference Flagging

Master Thesis

2021

Abstract

The field of radio frequency interference (RFI) flagging involves the identification of corrupted data within radio astronomy measurements. This work explores the application of supervised machine learning algorithms for RFI flagging, trained on real measurement data and simulated data with simulated RFI. The goal of this work is to investigate the prediction of RFI using specific machine learning algorithms; Naive Bayes Classifier, K-Nearest Neighbours Classifier, Random Forest Classifier, the U-Net convolution neural network and the Multilayer Perceptron. These algorithms are trained on real data, in which the ground truth includes inherent false positives, and simulated data where the ground truth positions of RFI are absolute. This is done through the use of time/frequency spectrogram data, relating to radio astronomy measurements, using the magnitudes and phases of each available polarization. Predictions for unseen test data are compared between algorithms, different implementations of those algorithms and each dataset. A specific implementation for data pre-processing is designed and implemented, utilizing a two dimensional filtering technique for feature construction. The goal of this method is intended to implement a means of injecting a form of spatial information of nearby time/frequency samples for each sample in a spectrogram. The inclusion of this spacial information, which is relevant to broadband bursts and narrowband persistent RFI, is hypothesised to increase the level of information present in the processed dataset. The use of feature construction using filtering techniques, demonstrates a noticeable improvement in the machine learning methods where each sample is treated individually during training and inference.

Keywords

Electrical Engineering

Reference:

Collections

Masters

Full item page