Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SiGMa-Net II: Distinguishing Binary Black Holes from Glitches

Soorya Narayan, Anupreeta More, Sunil Choudhary, Sudhagar Suyamprakasam, Sukanta Bose

TL;DR

This work tackles the challenge of rapidly distinguishing binary black hole (BBH) merger signals from non-Gaussian glitches in gravitational-wave data. It advances SiGMa-Net by using Sine-Gaussian Projection (SGP) maps as inputs and applying transfer learning with InceptionNetV3 to differentiate BBHs from six glitch types observed in O3 LIGO data. The approach achieves an accuracy of $87\%$ and a TPR of $0.83$ at an FPR of $0.1$ on test data, and demonstrates robustness when applied to real BBH events from O1–O3; it also achieves a substantial speedup in SGP map generation via GPU acceleration. The results indicate that SGP maps combined with CNNs can provide fast, scalable identification to augment standard GW search pipelines and motivate future expansion to more glitches and additional detectors.

Abstract

With increasing sensitivity of the gravitational wave (GW) detectors, we expect a significant rise in the detectable GW events. To process, analyse and identify such large amounts of GW signals arising from mergers of Binary Black Holes (BBH), we need both speed and accuracy. In the search for (massive) BBH signals, the biggest hurdle is posed by the various non-gaussian noise transients called glitches. Compared to our previous work, which used a simple convolutional neural network to distinguish BBHs from Blip glitches, this work uses transfer learning with InceptionNetV3 to distinguish BBHs from six types of most popular glitches from the third observing run of LIGO. While the glitches are real and identified via GravitySpy, the BBH signals are simulated and then injected into the real detector noise for each of the two LIGO detectors. We generate Sine-Gaussian Projection (SGP) maps by cross-correlating data with Sine-Gaussian functions of varied quality factors ($Q$) and central frequencies ($f_0$) and projected on the $Q$ - $f_0$ plane. We find that SGP maps make it easier to distinguish BBHs from glitches that look very similar to BBHs in the Time-Frequency maps like the Blips, while also maintaining significant morphological differences between BBHs and the more frequent glitches - Scattered Light and Fast Scattering. Our network has an accuracy of $87%$, a TPR of 0.83 for an FPR of 0.1 on our test dataset. It is also robust, retaining its level of accuracy, when tested on real BBH events identified in the first three observing runs of LIGO. Our proposed method shows the viability of using the SGP maps and neural networks for fast identification of GW events improving the efficiency of standard search pipelines.

SiGMa-Net II: Distinguishing Binary Black Holes from Glitches

TL;DR

This work tackles the challenge of rapidly distinguishing binary black hole (BBH) merger signals from non-Gaussian glitches in gravitational-wave data. It advances SiGMa-Net by using Sine-Gaussian Projection (SGP) maps as inputs and applying transfer learning with InceptionNetV3 to differentiate BBHs from six glitch types observed in O3 LIGO data. The approach achieves an accuracy of and a TPR of at an FPR of on test data, and demonstrates robustness when applied to real BBH events from O1–O3; it also achieves a substantial speedup in SGP map generation via GPU acceleration. The results indicate that SGP maps combined with CNNs can provide fast, scalable identification to augment standard GW search pipelines and motivate future expansion to more glitches and additional detectors.

Abstract

With increasing sensitivity of the gravitational wave (GW) detectors, we expect a significant rise in the detectable GW events. To process, analyse and identify such large amounts of GW signals arising from mergers of Binary Black Holes (BBH), we need both speed and accuracy. In the search for (massive) BBH signals, the biggest hurdle is posed by the various non-gaussian noise transients called glitches. Compared to our previous work, which used a simple convolutional neural network to distinguish BBHs from Blip glitches, this work uses transfer learning with InceptionNetV3 to distinguish BBHs from six types of most popular glitches from the third observing run of LIGO. While the glitches are real and identified via GravitySpy, the BBH signals are simulated and then injected into the real detector noise for each of the two LIGO detectors. We generate Sine-Gaussian Projection (SGP) maps by cross-correlating data with Sine-Gaussian functions of varied quality factors () and central frequencies () and projected on the - plane. We find that SGP maps make it easier to distinguish BBHs from glitches that look very similar to BBHs in the Time-Frequency maps like the Blips, while also maintaining significant morphological differences between BBHs and the more frequent glitches - Scattered Light and Fast Scattering. Our network has an accuracy of , a TPR of 0.83 for an FPR of 0.1 on our test dataset. It is also robust, retaining its level of accuracy, when tested on real BBH events identified in the first three observing runs of LIGO. Our proposed method shows the viability of using the SGP maps and neural networks for fast identification of GW events improving the efficiency of standard search pipelines.
Paper Structure (11 sections, 5 figures, 2 tables)

This paper contains 11 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Sine-Gaussian Projection Maps
  4. Datasets
  5. Populating the Datasets
  6. The Network
  7. Speeding Up SGP Generation Process
  8. Results
  9. Performance on various datasets
  10. Performance on the real GW events
  11. Summary and Conclusions

Figures (5)

  • Figure 1: Sine Gaussian Projections of the siz different types of glitches used to train the network. The Blip and Koifish glitches span the entire frequency axis of the space, while the other glitches are localized to the low-frequency region of the space. Fast scattering glitches extend along the Q-factor axis while the other glitches typically stay in the small Q-factor region.
  • Figure 2: Example images used in SiGMa-Net. Panel (a) shows the event GW190706_222641. Both detectors show similar projections, indicating the astrophysical nature of the source of the signal. The left half of the image shows the H1 projection and the right half the L1 projection. Panel (b) shows a Blip sample used to train the network. The Blip glitch is detected in the H1 detector, and the corresponding data from L1 shows random noise. The two halves are labelled as "H1" and "L1", and a dashed vertical line down the middle demarcates the boundary between the two projections.
  • Figure 3: The network architecture of SiGMa-Net. The pre-trained weights are frozen, and the final dense layers are replaced with custom, trainable dense layers.
  • Figure 4: The ROC curves for various test datasets which are detailed in Section \ref{['sec:results']}. The black dot shows the selected threshold for this work, $0.4$.The network performance is similar for the Test and Control datasets and are better than others with the former dataset comprising simulated BBHs and the latter real BBH events.
  • Figure 5: Performance of the network, for a threshold of 0.4, when classifying real BBH events binned by their total masses and SNRs.The All Runs dataset contains all of the BBH events from the O1, O2 and O3 observing runs. The fractions indicate correctly classified events out of the total events in a given bin. The colors depict the accuracy of the network and seem to change slightly across the datasets, but only because of the low statistics.