Validation of denoising system using non-harmonic analysis and denoising convolutional neural network for removal of Gaussian noise from gravitational waves observed by LIGO

Abstract

The third observation (O3) run of the Laser Interferometer Gravitational-Wave Observatory (LIGO) and Virgo was conducted between April 1, 2019, and March 27, 2020. The observed signal contained gravitational waves (GWs) and two types of noise: stationary and non-stationary noise. Both noise types must be removed to properly observe GWs. In this study, we characterized and imaged Gaussian noise, which is stationary noise, in the time–frequency domain and removed it from the observed signals using a denoising convolutional neural network (DnCNN). Furthermore, we verified its removal using non-harmonic analysis (NHA), which has a higher resolution than conventional methods such as short-time Fourier transform (STFT) and wavelet transform. The window length of the NHA and the visualization rate were varied to facilitate the removal. The system used was adapted to the model signal, i.e., the calculated GW and Gaussian noise, and the data observed by LIGO. This validation did not cover the data observed by Virgo. In experiments conducted with the model signal, the peak signal-to-noise ratio (PSNR) increased by up to 12.16 dB (optimal SNR = 8). In the data observed by LIGO, the Gaussian noise was removed, whereas the GWs and glitches remained. Thus, we were able to eliminate the Gaussian noise in both the model signal and the observed signal. This system can potentially be used to discover astronomical events in the time–frequency domain that cannot be estimated using matched filtering.