• machine learning
• pulsars
• continuous gravitational waves

Pulsars are among the most promising sources that ground- based gravi-
tational wave detectors will be able to detect within the next years. Their
gravitational wave emission is expected to be continuous but weaker than that
of coalescence of compact objects. To reach reasonable values of signal-to-noise
ratio for the detection, however, one has to increase considerably the coherent
observation time, making extremely difficult a full sampling of the whole pa-
rameter space. One solution comes from incoherent methods, where the whole
observation time is divided in sub-segment, or stacks, which are then analyzed
coherently.
One key limitation for detecting continuous periodic gravitational waves from
pulsars is thus the available computing power. In this context, we propose a
machine learning algorithm with the purpose to offer an efficient and yet rapid
method for characterizing continuous gravitational waves. Our algorithm is
based on the analysis of the time-frequency space, where the pattern that a
continuous gravitational wave produces is peculiar. Frequency of the emitted
signal is affected by the intrinsic spindown rate, which gives the line a negative
slope, and the motions of the Earth, that produce a characteristic oscillating
pattern. In this work, we propose a Convolutional Neural Network trained
specifically to recognize a line in this time-frequency plane, with the tasks of
classifying between noise and signal and infer the unknown values of frequency
and spindown. We evaluate the performance of this architecture on simulations
at different signal-to-noise ratio and for different values of frequency and fre-
quency derivatives, and we test its ability to generalize to data for which it
was not trained on. We test our network on more detailed signals produced
by a new dedicated Python simulated software. Finally, we present the results
showing how this approach can be promising to detect these kind of continuous
gravitational wave emission.