GW-YOLO: Multi-transient segmentation in LIGO using computer vision

TL;DR

The paper tackles the challenge of disentangling overlapping gravitational-wave signals and non-astrophysical glitches in real time. It proposes GW-YOLO, a YOLOv8-based instance segmentation framework applied to Q-transform spectrograms to simultaneously classify and localize signals and noise with pixel-level masks. Key results show $50\%$ detection efficiency for BBH signals at $SNR=15$ when overlapped with glitches, and $50\%$ efficiency for BNS signals at $SNR=30$, demonstrating robustness to overlaps and enabling automated event validation and glitch mitigation. This approach promises real-time, scalable improvements to gravitational-wave data analysis and downstream low-latency workflows.

Abstract

Time series data and their time-frequency representation from gravitational-wave interferometers present multiple opportunities for the use of artificial intelligence methods associated with signal and image processing. Closely connected with this is the real-time aspect associated with gravitational-wave interferometers and the astrophysical observations they perform; the discovery potential of these instruments can be significantly enhanced when data processing can be achieved in O(1s) timescales. In this work, we introduce a novel signal and noise identification tool based on the YOLO (You Only Look Once) object detection framework. For its application into gravitational waves, we will refer to it as GW-YOLO. This tool can provide scene identification capabilities and essential information regarding whether an observed transient is any combination of noise and signal. Additionally, it supplies detailed time-frequency coordinates of the detected objects in the form of pixel masks, an essential property that can be used to understand and characterize astrophysical sources, as well as instrumental noise. The simultaneous identification of noise and signal, combined with precise pixel-level localization, represents a significant advancement in gravitational-wave data analysis. Our approach yields a 50\% detection efficiency for binary black hole signals at a signal-to-noise ratio (SNR) of 15 when such signals overlap with transient noise artifacts. When noise artifacts overlap with binary neutron star signals, our algorithm attains 50\% detection efficiency at an SNR of 30. This presents the first quantitative assessment of the ability to detect astrophysical events overlapping with realistic, instrument noise present in gravitational-wave interferometers.

Figures (11)

• Figure 1: Time-frequency spectrogram of the binary neutron star event GW170817 abbott2017gw170817. The astrophysical event corresponds to the time-frequency track sloping up in frequency as time elapses (and labeled as "chirp"). A loud glitch at the time of this event that adversely impacted the quality of data at the LIGO-Livingston detector can also be seen dominating the low frequency (and beyond) response of the instrument (and labeled as "glitch"). Such overlaps or even in near proximity of noise and signal can challenge significantly the detection and parameter estimation of astrophysical events.
• Figure 2: Time frequency spectrograms of two sample glitch categories "extremely loud" on the left and "scattered light" on the right as identified by the GravitySpy algorithm Zevin:2016qwy. These are two of multiple transient noise classes that show up in the LIGO detector data.
• Figure 3: Left: Spectrograms (in the left column panels) depicting both "chirp" and "glitch" signals present in the data. These are generated from the combined timeseries of chirps and glitches. These spectrograms are then annotated as shown in right column panels with the labels corresponding to "chirp" and "noise". These annotated images form the dataset on which the segmentation model is then trained.
• Figure 4: Recall values for the detection of gravitational-wave chirp signals (BBH on left plot and BNS on right) with and without the addition of transient noise at model confidence value of 0.48. This is the value at which F1 score is maximized. The light shade band shows the $95\%$ confidence interval. As the SNR increases, the signals become more visible in the data, leading to higher recall values. Addition of transient noise adversely impacts the morphological appearance of the chirp signals in the data, thus causing a reduction in recall values. Fig. \ref{['fig:bns_chirp_12_15']} and Fig. \ref{['fig:chirp_scatter_noise']} show and explain this concept in greater detail. Despite the addition of transient noise, the model is able to correctly identify a substantial fraction of chirps in the data in the mid to high SNR bands.
• Figure 5: Algorithmic performance on low SNR events. Left: This spectrogram contains a BNS chirp in the SNR band $12 - 15$. Our segmentation model did not identify the presence of any "chirp" in this image due to the absence of any morphological features resembling those of a chirp signal. Right: This spectrogram is also based on a BNS chirp timeseries data in the same SNR band. Since a signal albeit a faint one showing a rise in frequency is visible here, the model was able to identify and localize the presence of "chirp" in this data.