Toward the detection of gravitational waves under non-Gaussian noises II. Independent component analysis

Abstract

We introduce a new analysis method to deal with stationary non-Gaussian noises in gravitational wave detectors in terms of the independent component analysis. First, we consider the simplest case where the detector outputs are linear combinations of the inputs, consisting of signals and various noises, and show that this method may be helpful to increase the signal-to-noise ratio. Next, we take into account the time delay between the inputs and the outputs. Finally, we extend our method to nonlinearly correlated noises and show that our method can identify the coupling coefficients and remove non-Gaussian noises. Although we focus on gravitational wave data analysis, our methods are applicable to the detection of any signals under non-Gaussian noises.

Noise subtraction from KAGRA O3GK data using Independent Component Analysis

“…Then we reformulate the frequency dependent model discussed in Ref. [25] to the multiple components and consider its implementation to the data analysis.…”

“…ICA is a method of signal processing which separates mixture of signals into statistically independent components by making use of non-Gaussianity of the sources. While the principle of ICA based on the statistical independence is so general that it is applicable to the nonlinearly coupled system [25], it makes the calculation more transparent compared to the machine learning. In the previous study [23], we demonstrated that the seismic noise can be partially subtracted by the simplest model of ICA.…”

“…In the previous study [23], we demonstrated that the seismic noise can be partially subtracted by the simplest model of ICA. Note that the model assuming instantaneous linear mixing, was implemented in two different ways and the optimal case turned out to be the time domain Wiener filtering [24,25].…”

“…As for the ICA model, convoluted linear mixing in time domain is considered to include the frequency dependence of the coupling. The optimal subtraction in this case was shown to be the Wiener filtering for each mode [25] and here the discussion is generalized to multiple components. One can understand that the method estimates the so-called transfer function of the noise for each frequency.…”

“…One can understand that the method estimates the so-called transfer function of the noise for each frequency. Among two possible ways of the estimation [25,14], we here utilize the one whose computational cost is lower and performance was investigated with the LIGO O2 data resulting in successful noise subtraction [15]. In order to ensure the validity of the method, we first apply it to the data in which acoustic noise was artificially injected and check whether the injected amount of noise is subtracted.…”