Machine learning classification of continuous gravitational-wave signal candidates

Filip Morawski, Michał Bejger, Paweł Ciecieląg

Continuous gravitational waves (GWs) emitted by non-axisymmetric, rotating neutron stars are expected to be detected in the near future by LIGO, Virgo and Kagra interferometers. Although the GW waveform is well known, its small amplitude makes it difficult to register. Searches for continuous GWs, among them the F-statistic method used here, are based on matched-filtering i.e., evaluation of signal templates on a grid of parameters, resulting in distributions of candidate GW signals.

In our work we present the application of machine learning in the analysis of the distributions of F-statistic signal candidates belonging to one of three distinct classes: pure Gaussian noise, astrophysical continuous GW, and stationary-frequency detector artifacts (lines). We study the robustness of candidates' classification with the use of one-dimensional convolutional neural networks, demonstrating their benefits in the context of the analysis of continuous GWs. We also show limits to the signal-to-noise ratio of the signal our method is able to correctly identify, and the ability of generalization from the training data.

Proceedings of the Polish Astronomical Society, vol. 10, 33-39 (2020)

Download full article as PDF file: