LIGO Document T1800288-v4
- In the Advanced LIGO observation runs, detection of gravitational waves is directly dependent on the sensitivity of the detectors. Transient noise, called "glitches," not only affects the general sensitivity of the detectors along with continuous noise, but also mimics and obscures real gravitational waves in the calibrated strain data channel. One machine learning software package used to classify these glitches and identify their sources, GravitySpy, is successful when the spectrogram of the glitch has a very distinct and unique shape. However, GravitySpy's spectrogram of one of the most common types of glitches, called a "blip," has an indistinct shape due to so few cycles being in-band, and tends to ring off template signals of binary black hole mergers, making it especially necessary to eliminate blips for future observing runs. Additionally, auxiliary channels other than the calibrated strain can pick up glitches and supplement GravitySpy's search for sources, but blip glitches are very infrequently witnessed by other channels and therefore still have unknown sources. The focus of this paper is to examine blip glitches in a Q-transform spectrogram with different parameters than those used by GravitySpy to determine if there are sub-classifications of blips that might have identifiable sources, and then use Convolutional Neural Networks to sub-classify these blips. Manual searches of short-duration Q-transform spectrograms of random blip glitches, nearly indistinguishable in GravitySpy's spectrograms, reveal six distinct possible sub-classes. The implementation of Convolutional Neural Networks has provided compelling evidence of distinguishable differences between these hypothesized sub-classes, confirming the assumption that there are multiple types of blips and aiding in future searches for blip sources.
- held from 23 Aug 2018 to 24 Aug 2018 in Caltech SCR, West Bridge 351, TeamSpeak
DCC Version 3.3.0, contact
Document Database Administrators