MG15 - Talk detail

Back to previous page

 Participant

Institution

Session

Accepted

Yes

Order

Time

Talk

Title

Abstract

We propose a method to analyze gravitational-wave data in the time-frequency plane by using an Artificial Neural Network performing classification as an image recognition task. We have built a Deep Convolutional Neural Network that is able to simultaneously process the data from the three interferometric antennas of LIGO Hanford, LIGO Livingston and Virgo. Our model is optimized to search for long $\mathcal{O}(\text{days})$ gravitational-wave transients, but the classifier can also be straighforwardly trained to detect other types of signals, such as supernova core-collapse. The method proposed is characterized by a very fast computation time during the prediction phase, thus enablig the possibility to build a low-latency trigger to allow faster messaging between gravitational-waves detectors and their electromagnetic counterparts. The low-latency trigger can also contribute to reduce the computational burden of the current offline analysis pipeline and follow-up. The classifier is able to reach more than 90\% detection efficiency and less than 1\% false alarm rate with a signal \emph{time domain} strain much smaller than the equivalent gaussian white noise \emph{time domain} standard deviation $h_S(t) \sim 4 \cdot 10^{-4} \sigma(h_N(t))$. The \emph{frequency domain} signal+noise strain is thus just above the median of the sensitivity curve.

Pdf file

Back to previous page