Searching for short-timescale radio anomalies using nonlinear dimensionality reduction techniques

TL;DR

This study employs nonlinear dimensionality reduction and machine learning to identify anomalous radio signals in archival telescope data, successfully detecting 202 events and demonstrating the method's potential for future radio surveys.

Contribution

Introduces a novel unsupervised machine learning pipeline combining ResNet and UMAP for detecting radio anomalies in high-time-resolution data.

Findings

Successfully detected 202 anomalous radio events.

UMAP effectively identifies anomalies in high-time-resolution datasets.

Pipeline is not suitable for standard dispersed pulse searches.

Abstract

We have searched for anomalous events using 2,520 hours of archival observations from Murriyang, CSIRO's Parkes radio telescope. These observations were originally undertaken to search for pulsars. We used a machine-learning algorithm based on ResNet and Uniform Manifold Approximation and Projection (UMAP) in order to identify parts of the data stream that potentially contain anomalous signals. Many of these anomalous events are radio frequency interference, which were subsequently filtered using multibeam information. We detected 202 anomalous events and provide their positions and event times. Our results show that the UMAP unsupervised machine learning pipeline effectively identifies anomalous signals in high-time-resolution datasets, highlighting its potential for use in future surveys. However, the pipeline is not applicable for standard searches for dispersed single pulses. We classify the detected events and, in particular, we are currently unable to determine the possible origin of events that last multiple seconds. For these we encourage follow-up observations.