UCLA

In this work, I describe significant advancements to the signal detection and Radio Frequency Interference (RFI) identification capabilities of modern radio technosignature detection algorithms. These improvements are presented alongside the results of the analysis of four annual UCLA radio technosignatures searches spanning 2016-2019. First, I describe the UCLA SETI Group’s initial versions of the signal detection and RFI identification algorithms, which were able to detect approximately 850,000 candidate signals within a frequency range of 1.15-1.73 GHz over ~2 hours of observations with the 100 m diameter Green Bank Telescope in 2016. Next, I describe an improved candidate signal detection algorithm that detected approximately 6 million signals in a 2017 search for technosignatures with identical observational parameters. Importantly, I show that the common practice of ignoring frequency space around candidate detections can reduce the number of signals detected by a factor of four or more and presents significant problems when estimating figures of merit or upper limits on the prevalence of technosignatures. I then present further improvements to these detection algorithms, which introduce the use of the topographic prominence for detection purposes and nearly double the signal detection count of some previously analyzed data sets. I also describe improvements to direction-of-origin filter algorithms, which are designed to remove most of the signals attributable to RFI from the data. The updated algorithms ensure a unique link between signals observed in separate scans. Finally, I present a novel machine-learning-based RFI mitigation algorithm, which helps address a major remaining challenge in the search for radio technosignatures. Specifically, I describe the design and deployment of a Convolutional Neural Network (CNN) that can determine whether or not a signal detected in one scan is also present in another scan. This CNN-based filter outperforms both a baseline 2D correlation model as well as existing filters over a range of metrics and reduces the number of signals requiring visual inspection after the application of traditional filters by a factor of 6-16 in nominal situations.