A convolutional neural network to distinguish glitches from minute-long gravitational wave transients Vincent Boudart1

2025-04-28 0 0 2.59MB 12 页 10玖币

侵权投诉

A convolutional neural network to distinguish glitches from minute-long

gravitational wave transients

Vincent Boudart

1∗

STAR Institute, Bˆatiment B5, Universit´e de Li`ege, Sart Tilman B4000 Li`ege, Belgium

Gravitational wave bursts are transient signals distinct from compact binary mergers that arise

from a wide variety of astrophysical phenomena. Because most of these phenomena are poorly

modeled, the use of traditional search methods such as matched ﬁltering is excluded. Bursts include

short (

10 seconds) and long (from 10 to a few hundreds of seconds) duration signals for which the

detection is constrained by environmental and instrumental transient noises called glitches. Glitches

contaminate burst searches, reducing the amount of useful data and limiting the sensitivity of current

algorithms. It is therefore of primordial importance to locate and distinguish them from potential

burst signals. In this paper, we propose to train a convolutional neural network to detect glitches in

the time-frequency space of the cross-correlated LIGO noise. We show that our network is retrieving

more than 95% of the glitches while being trained only on a subset of the existing glitch classes

highlighting the sensitivity of the network to completely new glitch classes.

INTRODUCTION

Gravitational waves (GW) have been detected on

September 14, 2015 [

] by the Advanced LIGO [

]

detectors, revealing the collision of two black holes

for the ﬁrst time. Since then, the Advanced LIGO

and the Advanced Virgo [

] detectors have observed

more than 90 compact binary coalescence (CBC)

events [

], among which black hole-neutron star [

]

and binary neutron star collisions [

]. In light of the

planned sensitivity improvement of the Advanced

LIGO and Advanced Virgo detectors, a new family

of gravitational wave sources, known as unmodeled

GW transients or bursts, is a prime target candidate

for the next observing run. Bursts include a wide

range of astrophysical phenomena for which accurate

waveforms are not accessible. The computational

resources required to build a template bank covering

a wide range of complex and highly turbulent events

prevents us to use matched ﬁltering methods as in

CBC searches [

]. Some of the expected progenitors

of gravitational wave transients are supernovae [

fallback accretion events [

], accretion-disk instabilities

[

], nonaxisymmetric deformations in magnetars [

accretion-disk instabilities [

] as well as gamma-ray

bursts [

]. Two classes of bursts are identiﬁed: short

(

10 seconds) and long (from 10 to a few hundreds

of seconds). In this paper, we present a new machine

learning tool that complements our previous work [

]

and discriminates transient noises happening in the

detectors from long-duration burst signals.

The main approach to detect burst events while

making minimal assumptions on the targeted signals

∗vboudart@uliege.be

relies on the excess-of-power method. It consists in

searching for excess of power in the time-frequency

space of single or multiple detector data, i.e. to ﬁnd

narrow time-evolving frequency curves. This problem

has already been tackled by diﬀerent groups who built

the current generation of pipelines, namely PySTAM-

PAS [

], cocoA [

], the two diﬀerent versions of

STAMP-AS, Zebragard and Lonetrack [

], the

long-duration conﬁguration of coherent WaveBurst

(cWB) [18] and X-SphRad [19].

One of the main hindrance in burst searches is

glitches. Glitches are transient noises caused by

instrumental or environmental sources [

] that

appear in the detector data in large quantities. Several

families of glitches have been reported [

], showing

diﬀerent time-frequency morphologies. Glitches limit

the sensitivity of the searches and can hinder GW de-

tections. Therefore, all of the aforementioned pipelines

deal with glitches either in pre- or post-processing

steps. In a previous work [

], we trained a neural

network with chirp signals having random parameters

and showed that our methodology can be used to

detect minute-long GW transients. However, it can

also recover glitches fairly and a visual inspection is

needed to discriminate them from chirp signals. This

work aims at removing the false-alarms caused by

glitches through a convolutional neural network.

Convolutional neural networks (CNNs) have been

recently used in Burst detection [

]. The authors in

[

] have built a 1 dimensional CNN to detect generic

short duration signals from the strain data of LIGO

and Virgo detectors. CNNs have shown promising

results in the identiﬁcation and classiﬁcation of GW

bursts from supernovae [

], in the detection of

binary black hole mergers [26] as well as long-duration

arXiv:2210.04588v2 [gr-qc] 11 Oct 2022

transients from isolated neutron stars [

] or as

early alert systems for binary neutron star collisions

[

]. CNNs are widely used for pattern recognition

[

] and classiﬁcation tasks [

–

]. Their pow-

erful capability to identify shapes and structures

has lead to the deﬁnition of Generative Adversarial

Networks [

], allowing to generate new samples by

learning the underlying distribution of the original data.

In Sec. I, we describe how glitches have been selected

to constitute the training set and how we highlight them

in the cross-correlated TF maps. Details about the

architecture of our classiﬁer and the training method

are given in Sec. II. We then show the results of the

training in Sec. III. Section IV is dedicated to large

scale tests comparable to the analyses conducted during

burst searches. Future prospects and conclusions are

given in Sec. V.

I. METHODOLOGY

Our search for minute-long bursts is based on the

excess-of-power method [

]. We make use of correlated

spectrograms, also referred to as time-frequency (TF)

maps, as described in [

]. In order to distinguish

glitches from possible burst signals, we will train a

neural network to identify them in the spectrograms.

As both can be present in a single TF map, we need to

consider the following cases : (1) a glitch is present in

the map, (2) a burst signal is present in the map, (3)

both or (4) none of them show up in the spectrogram.

Accordingly we will build 4 diﬀerent data sets to

include all the possible scenarios in the training phase.

The fourth scenario consists in building a dataset

with background TF maps. The data from Hanford

(H1) and Livingston (L1) from the ﬁrst half of the third

observing run (O3a) are ﬁrst whitened [

] prior to be

correlated. Using time-slides [

], we then generate

10000 spectrograms with a time resolution of 6 seconds

and a frequency resolution of 2 Hz. As the TF maps

span 1000 seconds and 2048 Hz, their size is 166

1025.

Since we aim to apply our classiﬁer on ALBUS’ output,

the size of the TF maps is chosen to be identical to [

A. Chirp generation

A methodology to recognize minute-long burst sig-

nals using machine learning techniques with very few

assumptions has been proposed in our previous work

[

]. This approach consists in using the Scipy library

[

] to generate chirp signals in the time domain with

random parameters, covering the whole time-frequency

parameter space. Figure 1shows some examples of gen-

erated chirps. As has been shown [

], this allows to

train a neural network with no prior assumption on the

targeted signals while conﬁdently identifying minute-

long burst models. Chirps are injected into noise with

9 levels of visibility, deﬁned as :

V=X

i,j

Sij −Nij (1)

where

Nij

is a noise-only spectrogram and

Sij

refers

to the same spectrogram in which a signal has been

injected. The sum is carried over all the pixels (

i, j

) in

the map. The deﬁnition of the visibility is particularly

useful to ensure chirps to be visible in the TF maps,

preventing the network to be fooled during the training

phase. The visibility can also be seen as a measure of

the anomalousness of the input TF maps. We choose

9 intensity levels in order to cover a quite large inten-

sity range, as seen in Figure 15 in the Appendix. We

use this intensity criterion to build our second dataset,

containing 10000 samples.

B. Glitch selection

During the second observing run (O2), glitches

happened roughly at a rate of 1 every min in the

detectors [

]. Although it amounts to a considerable

volume of contaminated data, they barely show up in

cross-correlated spectrograms. Indeed, both glitches

have to fall into overlapping time bins while showing

a suﬃciently high signal-to-noise ratio (SNR) and

sharing some frequency bandwidth. Even if these

conditions greatly reduce the amount of glitches that

contaminate our search, several thousands of glitches

can be found out of a couple of millions of TF maps

generated during the background searches.

To constitute our dataset with glitches, we need a

way to inject several glitch classes into time frequency

maps. However, the only tool that is currently available

to produce realistic glitches can only generate blip

glitches [

]. Blips are one the 23 classes that have

been characterized by Gravity Spy [

]. They

have a frequency between 0 and 256 Hz [

] which

would limit the detection bandwidth of the classiﬁer if

used exclusively in our dataset. Therefore, we have to

rely on the glitches detected so far to constitute the

training set. We thus select glitches that have been

recorded by Gravity Spy during O3a [

]. We load

the data around the GPS time of the chosen glitch in

each single detector (Hanford H1 and Livingston L1)

FIG. 1. Examples of chirp signals.

and shift them so that they fall into the same time bin.

In this way, we maximize the probability of ﬁnding

cross-correlated glitches that appear clearly in the TF

maps. Moreover, glitches showing higher SNR do not

always lead to stronger cross-correlated signals in the

TF maps. To circumvent these problems, we choose 7

glitch classes with SNR ranging from 20 to 10000 in

both Hanford and Livingston data. This will ensure

some variability in the results of the cross-correlation.

Table Isummarizes the useful information.

Glitch classes

Blip, Low Frequency Burst,

Scattered Light, Tomte,

Whistle, Extremely Loud,

Koi Fish

SNR ranges

20-30, 30-40, 40-50, 50-100,

100-150, 150-200, 200-300,

300-500, 500-10000

Number per range 30 (if possible)

Injection time between 50s and 950s

Total H1: 1110 L1: 1260

TABLE I. Information about the glitches selected from H1

and L1.

The total number of selected glitches is 1110 for

H1 and 1260 for L1. Then, we randomly choose one

glitch from each detector and build the resulting time-

frequency map. We reproduce this procedure 50000

times. To evaluate if the cross-correlation of the cho-

sen glitches has lead to a visible glitch in the output

spectrogram, we employ ALBUS, the neural network

dedicated to burst detection [

]. We showed that AL-

BUS can recover glitches as well as chirp signals. We

use its output map to introduce a score quantifying

the anomalousness present in the original spectrogram,

called anomaly score (AS). This score is deﬁned as :

AS =X

i,j

Oi,j if Oi,j >0.5 max(O) (2)

where

is the ALBUS output map and

and

indicate the time and frequency dimensions. The

anomaly score can be thought of the sum over the

pixels remaining after applying an intensity cut to

the output map. This threshold has been chosen to

exclude all the values close to zero, as they are quite

numerous given the size of the TF maps and can have

an impact on the ﬁnal anomaly score. The anomaly

score can also be used to rank detected signals as

seen in Figure 2where an extended glitch shows a

higher score compared to a glitch that spends a small

frequency range.

After visual inspection, background maps without

any glitch have a maximum score around 6.5, as seen in

Figure 3where 10000 images have been processed. All

the 15 background images with scores above 8 show a

correlated glitch. We thus set the threshold to conﬁrm

the presence of a correlated glitch to 8 in order to leave a

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Aconvolutionalneuralnetworktodistinguishglitchesfromminute-longgravitationalwavetransientsVincentBoudart11STARInstitute,B^atimentB5,UniversitedeLiege,SartTilmanB4000Liege,BelgiumGravitationalwaveburstsaretransientsignalsdistinctfromcompactbinarymergersthatarisefromawidevarietyofastrophysicalphen...

展开>> 收起<<

A convolutional neural network to distinguish glitches from minute-long gravitational wave transients Vincent Boudart1.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A convolutional neural network to distinguish glitches from minute-long gravitational wave transients Vincent Boudart1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: