GSpyNetTreeS: a machine learning solution for glitch localization in time and frequency

TL;DR

GSpyNetTreeS extends GSpyNetTree with a YOLO-like architecture to automatically detect, classify, and localize glitches in time-frequency representations of GW detector data. Using a 1-second Q-transform-based image representation and a carefully designed loss that combines presence, localization, and classification, the method achieves robust performance on common glitch classes and injected GW signals, enabling reproducible, automatic event validation. The approach shows strong IOU-aligned localization for most glitches and GW, with room for improvement in underrepresented classes, and it offers a practical path toward reducing human intervention in LVK’s data-quality vetting pipeline. Overall, GSpyNetTreeS provides a fast, scalable tool that can feed directly into glitch mitigation workflows and improve the robustness and reproducibility of gravitational-wave analyses in future observing runs.

Abstract

Data from ground-based gravitational wave detectors are often contaminated by non-Gaussian instrumental artifacts or detector noise transients. Unbiased source property estimation relies on the ability to correctly identify and characterize these artifacts and remove them if necessary. To this end, the LIGO-Virgo-KAGRA Collaboration has implemented candidate vetting for all significant candidates to identify the presence of artifacts and assess the need for mitigation. The current candidate vetting process requires human experts to identify the frequency ranges and the time windows associated with any data quality issues present. Differences in judgment between human experts may cause inconsistency, making results difficult to reproduce across gravitational wave events. We present GSpyNetTreeS, an extension to GSpyNetTree based on the You Only Look Once algorithm, for the automatic detection, classification, and time-frequency localization of detector noise transients. As a proof of concept, we tested GSpyNetTreeS's performance on the data collected by the LIGO detectors during the third observing run for gravitational waves as well as common detector glitch classes included in GSpyNetTree: Blip, Low frequency blip, Low frequency line and Scratchy. We also demonstrated that GSpyNetTreeS is capable of accurately identifying common glitch classes and capturing the frequency and time information associated with detected detector noise transients, establishing its potential as an automatic event validation tool for LIGO-Virgo-KAGRA's observing runs.

Figures (8)

• Figure 1: The architecture of GSpyNetTreeS. The input to GSpyNetTreeS is a time-frequency representation image of GW data. GSpyNetTreeS consists of two major components. The first component is a simplified Darknet with 24 Darknet convolutional layers. The Darknet generates two sets of outputs from its 15th and 24th layer. The outputs will be further processed by the second component of GSpyNetTreeS, which consists of a few Darknet convolutional layers, an up-sampling layer and a concatenate layer. The outputs are then passed onto two output layers, each of which is specialized for regions of excess power of different sizes (see Section \ref{['sec:training']}). The lower panel shows the layers used to construct a Darknet convolutional layer (DCL).
• Figure 2: Images of time-frequency representation of a time series from LIGO Hanford. The images show three regions of excess power that meet the criteria defined in Section \ref{['sec:training']}. A simulated GW signal from a binary black hole merger with component masses 35.6 and 40.3 $\mathrm{M}_{\odot}$ is shown in the blue time-frequency box. The other two time-frequency boxes highlight real glitches observed by the detector: a Blip (green box) and Low frequency line (red box). The upper panel shows the 'ground truth': manually drawn bounding boxes, with human expert judgment used to assess the most appropriate time-frequency bounds. The lower panel shows the corresponding bounding boxes predicted by the algorithm, including predicted frequency ranges and time windows. $\hat{y}^{\mathrm{e}}_{ijk}$ indicates the confidence of GSpyNetTreeS that the predicted bounding boxes surround a region of excess power (see Section \ref{['sec:loss']}).
• Figure 3: A histogram showing the distribution of pixel values for all images of time-frequency representation of GW detector data in the training samples. The vertical line indicates the threshold for loud pixels (i.e., 10.5), which is greater than $98\%$ of all pixels.
• Figure 4: A scatter plot showing the distribution of the widths and heights in pixels for all the bounding boxes in the images for training. K-means clustering is applied to group bounding boxes into different clusters, with cluster members shown in different colours. The centroids of the clusters are then taken as the typical size of anchor boxes, which are used as reference boxes for GSpyNetTreeS.
• Figure 5: The definition of IOU. Given two boxes $\mathrm{B}^{1}$ and $\mathrm{B}^{2}$, the shaded areas in the top and bottom panels show the overlapping area between the two boxes $\mathrm{A_{overlap}}$ and the combined area of the boxes $\mathrm{A_{union}}$ respectively.