arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2407.02391astro-ph.IM
收藏

神经网络方法搜索长瞬态引力波

Neural network method to search for long transient gravitational waves

Francesca Attadio, Leonardo Ricca, Marco Serra, Cristiano Palomba, Pia Astone, Simone Dall'Osso, Stefano Dal Pra, Sabrina D'Antonio, Matteo Di Giovanni, Luca D'… 展开作者

Francesca Attadio, Leonardo Ricca, Marco Serra, Cristiano Palomba, Pia Astone, Simone Dall'Osso, Stefano Dal Pra, Sabrina D'Antonio, Matteo Di Giovanni, Luca D'Onofrio, Paola Leaci, Federico Muciaccia, Lorenzo Pierini, Francesco Safai Tehrani

AI总结:

提出基于卷积神经网络和去噪器的组合方法,以低成本高效搜索长瞬态引力波信号,在LIGO模拟噪声中实现2%误报率和90%效率,并验证对频率演化幂律的鲁棒性。

摘要:

我们提出了一种新方法,用于在干涉仪探测器数据中搜索长瞬态引力波信号,例如由快速旋转的新生磁星产生的信号。标准搜索技术在计算上不可行(匹配滤波)或要求极高(次优半相干方法)。我们探索了利用机器学习范式的不同方法,以定义一种快速且廉价的过程。我们使用卷积神经网络开发了一个分类器,能够区分信号的存在与否。为了补充分类并增强其有效性,我们还开发了一个去噪器。我们根据LIGO干涉仪的设计噪声曲线,使用模拟有色噪声研究了两个网络的性能。我们表明,两个模型的结合对于提高检测机会至关重要。实际上,随着信号初始振幅从$10^{-22}$降低到$10^{-23}$,分类任务变得更加困难。特别是,如果不使用去噪器,我们无法正确标记初始振幅为$2 \times 10^{-23}$的信号。通过研究组合网络的性能,我们找到了单个干涉仪搜索误报率(2%)和效率(90%)之间的良好折衷。此外,我们证明了我们的方法对于描述信号频率随时间演化的幂律变化具有鲁棒性。我们的结果突出了该方法在生成长瞬态信号触发方面计算成本低的特点。本工作所进行的研究为进一步改进奠定了基础,目的是开发一个能够对长瞬态信号进行系统搜索的流程。

英文摘要:

We present a new method to search for long transient gravitational waves signals, like those expected from fast spinning newborn magnetars, in interferometric detector data. Standard search techniques are computationally unfeasible (matched filtering) or very demanding (sub-optimal semi-coherent methods). We explored a different approach by means of machine learning paradigms, to define a fast and inexpensive procedure. We used convolutional neural networks to develop a classifier that is able to discriminate between the presence or the absence of a signal. To complement the classification and enhance its effectiveness, we also developed a denoiser. We studied the performance of both networks with simulated colored noise, according to the design noise curve of LIGO interferometers. We show that the combination of the two models is crucial to increase the chance of detection. Indeed, as we decreased the signal initial amplitude (from $10^{-22}$ down to $10^{-23}$) the classification task became more difficult. In particular, we could not correctly tag signals with an initial amplitude of $2 \times 10^{-23}$ without using the denoiser. By studying the performance of the combined networks, we found a good compromise between the search false alarm rate (2$\%$) and efficiency (90$\%$) for a single interferometer. In addition, we demonstrated that our method is robust with respect to changes in the power law describing the time evolution of the signal frequency. Our results highlight the computationally low cost of this method to generate triggers for long transient signals. The study carried out in this work lays the foundations for further improvements, with the purpose of developing a pipeline able to perform systematic searches of long transient signals.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑