Glitch noise classification in KAGRA O3GK observing data using unsupervised machine learning

TL;DR

The paper tackles glitch noise classification in KAGRA O3GK data using an unsupervised learning pipeline. A convolutional variational autoencoder learns latent representations from time-frequency spectrograms of Omicron-triggered glitches, which are then projected to 3D with UMAP and clustered with spectral clustering to reveal distinct glitch shapes. Eight glitch classes are identified, with a dominant teardrop-shaped class and several smaller categories, highlighting differences from LIGO Gravity Spy in scope and morphology. This label-free approach enables scalable glitch taxonomy that can aid detector upgrades and will be applicable to O4 data to study evolution of glitch topology in response to instrument changes.

Abstract

Gravitational wave interferometers are disrupted by various types of nonstationary noise, referred to as glitch noise, that affect data analysis and interferometer sensitivity. The accurate identification and classification of glitch noise are essential for improving the reliability of gravitational wave observations. In this study, we demonstrated the effectiveness of unsupervised machine learning for classifying images with nonstationary noise in the KAGRA O3GK data. Using a variational autoencoder (VAE) combined with spectral clustering, we identified eight distinct glitch noise categories. The latent variables obtained from VAE were dimensionally compressed, visualized in three-dimensional space, and classified using spectral clustering to better understand the glitch noise characteristics of KAGRA during the O3GK period. Our results highlight the potential of unsupervised learning for efficient glitch noise classification, which may in turn potentially facilitate interferometer upgrades and the development of future third-generation gravitational wave observatories.

Figures (7)

• Figure 1: Data analysis flow in this study. (a) Datastream acquired by the KAGRA interferometer were processed using the Omicron pipeline to identify glitching times. (b) Creation of a time-frequency spectrogram for each trigger. (c) Conversion of the spectrograms to grayscale. (d) Integration of four time ranges into a composite image. (e) Training the created dataset VAE to extract latent variables. (f) Dimensional compression into 3D space using UMAP. (g) Glitch noise shape classification using spectral clustering.
• Figure 2: Glitch noise visualization in 3D space using UMAP with full O3GK data. Additionally, the figures show glitch noise classified using spectral clustering and color-coded by class. The difference between the color-coded figures is the difference in the number of classes divided by spectral clustering. From left to right in the bottom row, the figures are classified into 6, 8, and 10 classes.
• Figure 3: Average Davies--Bouldin index values for $k$-means++ and spectral clustering as a function of the number of clusters. Each clustering experiment was repeated 20 times with different random seeds to account for stochastic variability, and the mean index was computed. Error bars represent the standard error of the mean. Blue and orange are the results of $k$-means++ and spectral clustering, respectively.
• Figure 4: Latent variables of full O3GK data were dimensionally compressed into a 3D space using UMAP and then classified. Each glitch noise image illustrates the characteristic glitch noise shape for each cluster. The glitch noise image has time windows of 0.5 s and 4 s on the left and right sides, respectively, for each class.
• Figure 5: Latent variables of full O3GK data plotted in a 3D space using UMAP and divided and color-coded using $k$-means++. The number of divisions is eight.