Pulsars are spinning neutron stars which emit an electromagnetic beam. We expect pulsars to slowly decrease their rotational frequency due to the radiation emission. However, sudden increases of the rotational frequency have been observed from different pulsars. These events are called “glitches”, and are followed by a relaxation phase with timescales from days to months. Gravitational-wave (GW) emission may follow these peculiar events, including long-duration transient continuous waves (tCWs) lasting hours to months. These are modeled similarly to continuous waves but are limited in time. Previous studies have searched for tCWs from glitching pulsars with matched filtering techniques and by computing a detection statistic, the F-statistic, maximized over a set of transient parameters like the duration and start time of the potential signals. This method is very sensitive, but the computational costs can easily increase when widening the frequency and spindown search bands and the duration of the potential signals.
In order to reduce computational and human effort, we present a procedure for detecting potential tCWs using Convolutional Neural Networks (CNNs). CNNs have proven to be valid networks for detecting various CW signals, but have never been tested on tCWs from glitching pulsars. For our initial configuration, we train the CNN on F-statistic “atoms”, i.e. quantities computed during the matched filtering step from signal/noise data. This still constrains the frequency evolution of the signal to be CW-like, but already allows for flexible amplitude evolution and significant speed-up compared to the traditional method. In the future, we also plan to implement a second CNN with input the frequency-time maps, which in this case can search for unmodeled tCWs both in frequency and amplitude evolution, which we expect to be a further improvement to the speed and performance of the search.