UC Berkeley

Modern radio astronomy has brought forth an era of data explosion. With advances in instrumentation and computational power, new host of algorithms for data analysis become possible and necessary. Techniques of data reduction and interference rejection are crucial to signal detection for a wide range of scientific questions. In 21\,cm cosmology, a faint and diffuse signal from primordial hydrogen requires extracting every last bit of sensitivity from the instruments. In search of fast radio burst and signals of extra-terrestrial technology we need to sort through large volumes of data while effectively distinguishing candidate signals from noise and interference. Overcoming these challenges requires novel intuition of the experiments as well as applications of state-of-the art methods of statistics.

Redundant arrays are a kind of interferometers designed to detect the 21cm power spectrum. The spectrum itself is naturally linked to the square of the visibilities, giving rise to a set of theoretical formulations that skips the need for forming images. Such formulation, however, does not naturally include all sensitivity of the interferometer. Thus in this thesis we develop a technique to complete this formulation and provide the remaining sensitivities.

Fast radio bursts (FRB) are a type of millisecond-duration transient signal seen in spectrograms with a quadratically dispersed shape. At the time of the writing of this thesis, the physical nature of the sources is yet unknown. One of the repeating sources is FRB121102. Using deep learning, we detect the most number of signals from FRB121102 during a single observation. The 93 pulses provide new wealth of statistical properties, most notably periodicity. Whether the emission is periodic provides an insightful look on the nature of the source. With our novel method of periodicity search, we show that the received pulses are likely not periodic with any period longer than 10 milliseconds.

Radio frequency search for extraterrestrial intelligence (SETI) is one of the most open-ended questions in astronomy. With properties of the candidate signal unknown, SETI needs intelligent methods that automatically extracts information from large volumes of data. In parallel with new developments in radio astronomy, machine learning techniques, in particular deep learning has successfully revolutionized applications in computer vision, speech processing, text processing, and signal processing. In this thesis, we demonstrate a range of data processing techniques, with focus on deep learning, to signal detection in SETI. We showcase a blind detection for repeating signals, a semi-supervised representation learning framework, as well as a generative model for anomaly detection and interference rejection.