深度学习在引力波数据处理中的应用初探

截至2018-01-16,LIGO已成功探测引力波事件6次.可以预期,引力波探测事件会越来越多,引力波天文学会很快进入到大数据阶段.深度学习在大数据处理方面近年来得到迅速发展.它在数据处理速度,准确度等方面都表现出极大的优势.深度学习在引力波数据处理中的应用讨论还不多.本文引入此问题,并对其进行初步研究.引力波数据最大的特点是强噪声、弱信号.现行的数据处理方法是利用匹配滤波的方式把引力波信号从强噪声中挖掘出来.同时,匹配滤波方法还可以确定引力波源的性质,定量确定其参数.匹配滤波方法的弱点是计算量巨大.这导致数据处理速度很慢.对于将来的大数据引力波天文学,这更将是一个巨大的隐患.匹配滤波方法的另一个潜在问题是,完备准确的理论波形模板是其工作的前提条件.这个潜在问题的后果是很难找到理论预期之外的引力波信号.深度学习的数据处理方法有可能在这些问题上提供出路.同时,深度学习也会遇到其自身的若干困难和问题.本文将从网络结构、训练数据制备、训练优化、对信号识别的泛化能力、对数据的特征图表示以及对特征数据遮挡的响应等方面来展开讨论.

Initial study on the application of deep learning to the Gravitational Wave data analysis

Till 2018-01-16,LIGO has successfully detected 6 gravitational wave events.It is expected that there will be more and more gravitational wave events be detected,and the gravitational wave astronomy will quickly start its big data phase.Deep learning has developed rapidly in recent years especially on big data processing.It has shown great advantages in the speed of data processing,accuracy and so on.There is few discussion about the application of deep learning to gravitational wave data analysis.On the one hand,the current article will introduce this problem.On the other hand,we will give a preliminary exploration of this problem.The most outstanding character of gravitational wave data is weak signal hided in strong noise.The current method of gravitational wave data analysis is using the matched filtering to excavate the gravitational wave signal from the strong noise.At the same time,the properties of the gravitational wave source can be determined by the matched filtering method,and the parameters of the source can be determined quantitatively.The weakness of the matched filtering method is that the amount of computation cost is huge and the speed of data processing is slow.In the face of future big data gravitational wave astronomy,this will be a great hidden challenge.Another potential problem of the matched filtering method is that a complete and accurate theoretical waveform template is a prerequisite making sure the matched filtering technique works.The consequence of this potential problem is that it is difficult for us to find the gravitational wave signals which is out of theoretical expectation.The data processing methods of deep learning can provide a way for solving these problems.At the same time,deep learning will encounter some of its own difficulties and problems.This article will discuss the aspects of network structure,training data preparation,training optimization,generalization ability of signal recognition,representation of data character and response to feature data occlusion.