Extracting non-Gaussian features in gravitational wave observation data using self-supervised learning
使用自监督学习提取引力波观测数据中的非高斯特征
We propose a self-supervised learning model to denoise gravitational wave (GW) signals in the time series strain data without relying on waveform information. Denoising GW data is a crucial intermediate process for machine-learning-based data analysis techniques, as it can simplify the model for downstream tasks such as detections and parameter estimations. We use the blind-spot neural network and train it with whitened strain data with GW signals injected as both input data and target. Under the assumption of a Gaussian noise model, our model successfully denoises over 90% of GW signals from binary black hole (BBH) and neutron star--black hole mergers when the signal-to-noise ratio (SNR) is larger than 13 and 16 and denoises over 50% of binary neutron star signals when the SNR is larger than 23, with an overlap greater than 0.2, over which the denoised signal can have visually distinguishable patterns. Our model also denoises over 90% of BBH signals in H1, L1, and V1 data detected in the O1, O2, and O3 observation runs when the network SNR is larger than 11, 12, and 20, respectively. We also test the model's potential to extract glitch features, loud inspiral compact binary coalescence signals a few seconds before the merger, and unseen core-collapse supernova signals during training.
展开 ▾首次将自监督学习与盲点网络范式引入引力波去噪,在无干净波形参考的条件下能同时提取CBC、glitch甚至未训练的CCSN信号特征。
以往引力波去噪方法(如字典学习Torres-Forné+ 2016、监督学习Wei & Huerta 2020及各种去噪自编码器)依赖干净波形作为参考,只能复原已知模板信号。自监督去噪中的盲点网络(如Krull+ 2018和Laine+ 2019)开创了仅用噪声数据训练的可能。本文借鉴Laine+ 2019的高质量盲点架构,结合高斯噪声似然训练WaveNet,在注入CBC信号的应变数据上学会了提取非高斯特征,并在真实BBH事件中成功去噪,同时展现出提取glitch、inspiral及未训练CCSN的能力。未来可优化模型以应对长持续时间、低幅度信号(如BNS),并融入连续波等多种源,作为通用特征提取器接入下游检测与参数估计管线。
预印本 2024-03-07 · 刊出 2025-03-07 · 收录 2026-10-12