Gravitational wave signal denoising and merger time prediction with a deep neural network
基于深度神经网络的引力波信号去噪与并合时间预测
The mergers of massive black hole binaries could generate rich electromagnetic emissions, which allow us to probe the environments surrounding these massive black holes and gain deeper insights into the high energy astrophysics. However, due to the short timescale of binary mergers, it is crucial to predict the time of the merger in advance to devise detailed observational plans. The overwhelming noise and slow accumulation of the signal-to-noise ratio in the inspiral phase make this task particularly challenging. To address this issue, we propose a novel deep neural denoising network in this study, capable of denoising a 30-day inspiral phase signal. Following the denoising process, we perform the detection and merger time prediction based on the denoised signals. Our results demonstrate that, for a 30-day inspiral phase data with a signal-to-noise ratio between 10 and 50 occurring no more than 10 days before the merger, our absolute prediction error for the merger time is generally within 24 h.
展开 ▾将音频分离中的 WaveUNet 降维与 Sepformer 噪声分离相结合,首次对长达 30 天的旋进引力波信号进行高效去噪,并在此基础上直接预测并合时间,预报误差多数小于 24 小时。
大质量黑洞双星并合前最后几小时的信号贡献约 99% 的总信噪比 Chen+ 2024,旋进信号极弱,传统匹配滤波方法因需庞大模板库而计算开销巨大 Cornish+ 2022;近期基于深度学习的预检测模型 Ruan & Guo 2024 虽能提前数小时至数天识别信号,但高检测阈值限制了灵敏度,且未给出并合时间。本研究借鉴 WaveUNet 的降维架构 Stoller+ 2018 与 Sepformer 的噪声分离思想 Subakan+ 2021,提出 WaveUformer‑CNN,在去噪波形上同时完成信号检测与并合时间预测,显著提升了低信噪比旋进段的早期预警能力。未来工作将提升对低啁啾质量信号的泛化能力,处理多信号重叠场景,并验证模型在数据空缺、仪器非平稳及更高阶波形模式等更真实条件下的鲁棒性。
预印本 2024-10-11 · 刊出 2025-03-14 · 收录 2026-10-12