A machine learning algorithm for minute-long Burst searches

Minute-long Gravitational Wave (GW) transients are events lasting from few to
hundreds of seconds. In opposition to compact binary mergers, their GW signals
cover a wide range of poorly understood astrophysical phenomena such as
accretion disk instabilities and magnetar flares. The lack of accurate and
rapidly generated gravitational-wave emission models prevents the use of
matched filtering methods. Such events are thus probed through the
template-free excess-power method, consisting in searching for a local excess
of power in the time-frequency space correlated between detectors. The problem
can be viewed as a search for high-value clustered pixels within an image,
which has been generally tackled by deep learning algorithms such as
Convolutional Neural Networks (CNNs). In this work, we use a CNN as a anomaly
detection tool for the long-duration searches. We show that it can reach a
pixel-wise detection despite trained with minimal assumptions, while being able
to retrieve both astrophysical signals and noise transients originating from
instrumental coupling within the detectors. We also note that our neural
network can extrapolate and connect partially disjoint signal tracks in the
time-frequency plane.