DeepExtractor: Time-domain reconstruction of signals and glitches in gravitational wave data with deep learning

TL;DR

DeepExtractor tackles the challenge of reconstructing gravitational-wave signals and glitches in noisy detector data by learning to isolate and subtract the background noise in the time–frequency domain. It uses a PSD-informed U-Net operating on complex STFT spectrograms to predict the noise component $\hat{n}(t,f)$, yielding $\hat{g}(t,f)=h(t,f)-\hat{n}(t,f)$ without needing explicit waveform training. Across fully simulated glitches, real O3 backgrounds with transfer learning, and unseen glitch classes, DeepExtractor achieves a median mismatch as low as $0.9\%$, outperforms BayesWave in glitch recovery, and delivers over $10^4\times$ speedups on CPU. It also demonstrates model-agnostic signal reconstruction for three O3 events, highlighting potential for real-time glitch mitigation, though future work is needed to separate overlapping signal and glitch components and to leverage multi-detector coherence.

Abstract

Gravitational wave (GW) detectors, such as LIGO, Virgo, and KAGRA, detect faint signals from distant astrophysical events. However, their high sensitivity also makes them susceptible to background noise, which can obscure these signals. This noise often includes transient artifacts called 'glitches', that can mimic genuine astrophysical signals or mask their true characteristics. In this study, we present DeepExtractor, a deep learning framework that is designed to reconstruct signals and glitches with power exceeding interferometer noise, regardless of their source. We design DeepExtractor to model the inherent noise distribution of GW detectors, following conventional assumptions that the noise is Gaussian and stationary over short time scales. It operates by predicting and subtracting the noise component of the data, retaining only the clean reconstruction of signal or glitch. We focus on applications related to glitches and validate DeepExtractor's effectiveness through three experiments: (1) reconstructing simulated glitches injected into simulated detector noise, (2) comparing its performance with the state-of-the-art BayesWave algorithm, and (3) analyzing real data from the Gravity Spy dataset to demonstrate effective glitch subtraction from LIGO strain data. We further demonstrate its potential by reconstructing three real GW events from LIGO's third observing run, without being trained on GW waveforms. Our proposed model achieves a median mismatch of only 0.9% for simulated glitches, outperforming several deep learning baselines. Additionally, DeepExtractor surpasses BayesWave in glitch recovery, offering a dramatic computational speedup by reconstructing one glitch sample in approximately 0.1 seconds on a CPU, compared to BayesWave's processing time of approximately one hour per glitch.

Figures (17)

• Figure 1: An overview of DeepExtractor's reconstruction approach. Two seconds of detector strain $h(t)$ is processed into magnitude and phase spectrogramns using an STFT. This is fed through our network, which maps the instance to the magnitude and phase spectrograms of the underlying background noise. We then use an inverse STFT to yield the corresponding time series $n(t)$. An estimation of the underlying signal or glitch can be simply calculated as $h(t) - n(t)$.
• Figure 2: The U-Net architecture featured in DeepExtractor applied to batches (bs) of STFT data ($\mathbb{R}^{2\times h \times w}$), illustrating its characteristic ‘U’-shaped structure. The network processes both the magnitude and phase components of the STFT simultaneously through two input and output channels.
• Figure 3: Examples of each of the five training glitch classes; chirp, sine, sine-gaussian, gaussian pulse and ringdown. For each $2\,$s training sample, anywhere between 1 and 30 of these signals (selected randomly) are injected into the Gaussian background sample.
• Figure 4: Magnitude and phase STFT spectrograms for noise+glitch input (top) and noise-only target (bottom) used to train DeepExtractor.
• Figure 5: Power spectral density (PSD) of (a) real LIGO Hanford detector noise and (b) the same PSD whitened and our simulated white noise. In comparison, it is visible that our simulated noise misses instrumental or environmental lines commonly found in the GW detectors.