Deep Learning for Semantic Embedding and Anomaly Detection in LOFAR Data

Voncina, Klemen (2023) Deep Learning for Semantic Embedding and Anomaly Detection in LOFAR Data. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text Master_Thesis_Klemen_Voncina.pdf Download (20MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (134kB)

Abstract

This thesis explores the application of several self-supervised representation learning techniques for anomaly detection in spectral data. These techniques include five distinct methods, four deep learning-based auto-encoder models, and a linear transform method known as Independent Com- ponent Analysis (ICA). The data used in this research is obtained from an aperture synthesis radio telescope, specifically LOFAR. The primary aim is to evaluate the potential of these techniques in creating semantically meaningful embeddings. These embeddings are expected to aid in tasks such as data inspection, classification, and unsupervised anomaly detection. To achieve this aim, the methods are trained in a self-supervised manner and evaluated on a limited, labeled dataset curated for this study. One of the deep learning models tested includes a novel cascading convolutional auto-encoder architecture, uniquely adapted to account for time-frequency data characteristics and to include spa- tial context within its encoding. The proposed method shows best average classification scores across Random Forest and Gaussian Naive Bayes classifiers in long time-scale downsampled observations (43.83%). We demonstrate second best average accuracy in short time-scale downsampled obser- vations (76.14%). It is beaten out in the combined embedding by ICA, however still shows itself to be the best all-round method with an average accuracy over all test conditions of 55.49%. The next best method appears t

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Schomaker, L.R.B.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	17 Aug 2023 07:02
Last Modified:	17 Aug 2023 07:02
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31201

Actions (login required)

View Item