Assessing the impact of non-Gaussian noise on convolutional neural networks that search for continuous gravitational waves

We present a convolutional neural network that is capable of searching for continuous gravitational waves, quasimonochromatic, persistent signals arising from asymmetrically rotating neutron stars, in approximately one year of simulated data that is plagued by nonstationary, narrow-band disturbances, i.e., lines. Our network has learned to classify the input strain data into four categories: (1) only Gaussian noise, (2) an astrophysical signal injected into Gaussian noise, (3) a line embedded in Gaussian noise, and (4) an astrophysical signal contaminated by both Gaussian noise and line noise. In our algorithm, different frequencies are treated independently; therefore, our network is robust against sets of evenly spaced lines, i.e., combs, and we only need to consider a perfectly sinusoidal line in this work. We find that our neural network can distinguish between astrophysical signals and lines with high accuracy. In a frequency band without line noise, the sensitivity depth of our network is about ${\mathcal{D}}^{95%}\ensuremath{\simeq}43.9$ with a false alarm probability of $\ensuremath{\sim}0.5%$, while in the presence of line noise, we can maintain a false alarm probability of $\ensuremath{\sim}10%$ and achieve ${\mathcal{D}}^{95%}\ensuremath{\simeq}3.62$ when the line noise amplitude is ${h}_{0}^{\mathrm{line}}/\sqrt{{S}_{\mathrm{n}}({f}_{k})}=1.0$. The network is robust against the time derivative of the frequency $\stackrel{\ifmmode \dot{}\else \textperiodcentered \fi{}}{f}$ of a gravitational-wave signal, i.e., the spin-down, and can handle $|\stackrel{\ifmmode \dot{}\else \textperiodcentered \fi{}}{f}|\ensuremath{\lesssim}{10}^{\ensuremath{-}12}\text{ }\text{ }\mathrm{Hz}/\mathrm{s}$, even though our training sets only include signals with $\stackrel{\ifmmode \dot{}\else \textperiodcentered \fi{}}{f}=0$. We evaluate the computational cost of our method to be $\mathcal{O}({10}^{19})$ floating point operations, and compare it to those from standard all-sky searches, putting aside differences between covered parameter spaces. Our results show that our method is more efficient by 1 or 2 orders of magnitude than standard searches. Although our neural network takes about $\mathcal{O}({10}^{8})\text{ }\text{ }\mathrm{sec}$ to employ using our current facilities [a single graphics processing unit (GPU) of GTX1080Ti], we expect that it can be reduced to an acceptable level by utilizing a larger number of improved GPUs.