Data mining and machine learning improve gravitational-wave detector sensitivity

Gabriele Vajente

Phys. Rev. D 105, 102005 – Published 20 May 2022

Abstract

Application of data mining and machine learning techniques can significantly improve the sensitivity of current interferometric gravitational-wave detectors. Such instruments are complex multi-input single-output systems, with close-to-linear dynamics and hundreds of active feedback control loops. We show how the application of brute-force data-mining techniques allows us to discover correlations between auxiliary monitoring channels and the main gravitational-wave output channel. We also discuss the result of the application of a parametric and time-domain noise subtraction algorithm, that allows a significant improvement of the detector sensitivity at frequencies below 30 Hz.

Received 8 March 2022
Accepted 10 May 2022

DOI:https://doi.org/10.1103/PhysRevD.105.102005

Physics Subject Headings (PhySH)

Gravitational wave detection Gravitational waves

Gravitational wave detectors

Gravitation, Cosmology & Astrophysics

Authors & Affiliations

Gabriele Vajente ^*

LIGO Laboratory, California Institute of Technology, Pasadena, California 91101, USA

^*vajente@caltech.edu

Article Text (Subscription Required)

Click to Expand

References (Subscription Required)

Click to Expand

Issue

Vol. 105, Iss. 10 — 15 May 2022

Reuse & Permissions

Access Options

Article Available via CHORUS

Download Accepted Manuscript

Author publication services for translation and copyediting assistance advertisement

Images

Figure 1
Summary table with the channels that have the highest coherence for each frequency. In this case the main gravitational-wave strain channel is called DCS-CALIB_STRAIN. The title also shows the times used in the analysis, both in UTC time and in GPS time (number of seconds elapsed since 0 hours Sunday 6 January 1980). Each row corresponds to a frequency bin, as indicated in the first column. The other columns show the coherence with a specific channel at that frequency. The name of the channel and the value of the coherence is indicated in each cell. For each row (frequency bin) the channels are listed in order of decreasing coherence. The cell background color gives a visual representation of the coherence value: red corresponds to one and white corresponds to zero.
Reuse & Permissions
Figure 2
Example of one of the plots generated automatically by the brute-force coherence computation. The top panel shows the coherence between the main strain channel DCS-CALIB_STRAIN and one of the angular feedback control loop channels ASC-CHARD_P_OUT_DQ. The blue trace shows the coherence as a function of frequency, and the red dashed line shows the confidence level $C_{α}$ . The bottom panel shows the spectrum of the main channel, that is the square root of the power spectral density $S [h, h] (f)$ and the coherence-based projection of the auxiliary channel noise.
Reuse & Permissions
Figure 3
Two example signals from the Advanced LIGO detector. The solid curves show the power spectral densities (PSD), while the dashed curves represent the integrated power, showing at each frequency the amount of power contained in the signal above that frequency. Note that those traces are plotted with the same vertical axis as the PSD, but are in units of signal power instead of signal power spectral density. Both signals have 99% of their power below about 3 Hz. The signals shown here are error points of two feedback control loops in the angular sensing and control (ASC) and longitudinal sensing and control (LSC) systems.
Reuse & Permissions
Figure 4
An example of the plots produced by the brute-force coherence algorithm when used to look for modulated couplings. The top panel shows the coherence of the detector main output channel with a monitor of the power supply, and with the same channel modulated with angular control error channels, as in equation (6). The nonmodulated channel show high coherence only at the main power supply 60 Hz line, while the modulated channels show coherence only in correspondence of the sidebands of the main peak, as expected in the case of a nonstationary coupling. The bottom panel shows the coherence-based noise projection for each channel, as in Eq. (3).
Reuse & Permissions
Figure 5
Result of some stationary and nonstationary noise subtractions for the Advanced LIGO Hanford detector during the O3 observing run [2]. The two insets show zoomed-in versions of the main plot, highlighting some areas where nonstationary coupling was visible around some prominent lines (the 60 Hz power supply and some angular control calibration lines around 20 Hz). The broadband improvement at frequencies below 30 Hz is due to the subtraction of linearly coupled angular and longitudinal feedback control channels. Each trace shows the power spectral density of the detector output, calibrated in gravitational-wave strain. The blue trace is the original channel, before any subtraction. Each other trace shows the cumulative effect of additional noise subtractions: the orange trace shows the effect of removing the 60 Hz power line sidebands only; the green trace shows the effect of removing the angular calibration lines and sidebands from the data of the orange trace; finally the red trace shows the effect of all noise subtractions together.
Reuse & Permissions
Figure 6
Comparison of the detector range for binary black hole coalescence signals with and without the noise subtraction. The range is the maximum distance at which a signal could be detected with a signal-to-noise ratio of 8, averaged over sky position and source orientation [36]. The range is shown as a function of the component mass, for equal-mass binaries with no spin, integrating from 10 Hz. The bottom panel shows the improvement obtained with the noise subtraction. The effect is larger for the more massive binary systems.
Reuse & Permissions

Physical Review D

covering particles, fields, gravitation, and cosmology