Deep learning network to distinguish binary black hole signals from short-duration noise transients

Blip glitches, a type of short-duration noise transient in the LIGO-Virgo data, are a nuisance for the binary black hole (BBH) searches. They affect the BBH search sensitivity significantly because their time-domain morphologies are very similar, and that creates difficulty in vetoing them. In this work, we construct a deep-learning neural network to efficiently distinguish BBH signals from blip glitches. We introduce sine-Gaussian projection (SGP) maps, which are projections of gravitational wave (GW) frequency-domain data snippets on a basis of sine-Gaussians defined by the quality factor and central frequency. We feed the SGP maps to our deep-learning neural network, which classifies the BBH signals and blips. Whereas only simulated BBH signals are used for training, both simulated and real BBH signals are used for testing. For glitches only blips from real LIGO data are used for both testing and training. We show that our network significantly improves the identification of the BBH signals in comparison to the results obtained using traditional-${\chi }^2$ and sine-Gaussian ${\chi }^2$. For example, our network improves the sensitivity by 75% at a false-positive probability of ${10}^{-2}$ for BBHs with total mass in the range $[80,140]M_{\odot }$ and SNR in the range [3, 8]. When tested on real GW events, it correctly identifies 95% of the events in GWTC-3. The computation time for classification is a few minutes for thousands of SGP maps on a single core. With further optimization in the next version of our algorithm, we expect a further reduction in the computational cost. Our proposed method can potentially improve the veto process in the LIGO-Virgo GW data analysis and conceivably support identifying GW signals in low-latency pipelines.

Figure 1

Examples of time-frequency maps of some of the blips (right panels) and BBH events (left panels) showing the similarities and differences of the two in this two-dimensional representation.

Figure 2

The sine-Gaussian projection map of an H1 blip from O2 run with SNR 16 (left) and a simulated BBH event injected in real noise from H1 O2 run with SNR 16 (right). Here the SNR for blips is calculated in the same way as for BBH signals. The color bar represents value of projections as calculated using Eq. (2).

Figure 3

Here we show an example of a simulated BBH input sample fed to the network. The H1 (left) and L1 (right) SGP maps are placed adjacent to each other. Axis labels and color bars are omitted. Here, the $\mathrm{SNR}\approx 10$ for H1 and $\approx 9$ for L1.

Figure 4

Similar to Fig. 3, except that a blip input sample is fed to the network here. The H1 (left) and L1 (right) SGP maps are kept adjacent to each other. We can see a blip is present in H1 (left) with an $\mathrm{SNR}\approx 11$.

Figure 5

Schematic diagram of our neural network. Each layer uses a relu activation function except for the last one, which uses a sigmoid function. The final output is the probability that the input image corresponds to a BBH signal. Here, “$\mathrm{Conv}k$” stands for $k$th convolutional layer, “$\mathrm{MaxP}k$” stands for $k$th max-pooling layer, “Flatten” denotes the flattening of 2D matrix, “$\mathrm{Dropout}(x)$” layer is to tackle overfitting by dropping some of the neurons, and “$\mathrm{Dense}(k)$” is fully connected layer with $k$ neurons [47].

Figure 6

Comparison of ROC curves for CNN with those for two other methods, namely the traditional ${\chi }^2$ [50] and the sine-Gaussian ${\chi }^2$ [14]. The curves are shown for low- and high-SNR bins (solid and dotted, respectively). On the left, we show the performance at low-mass BBH signals, which is similar to that of the high-mass BBH signals on the right. The gray shaded region has more uncertainty for the low SNRs (dotted curve) due to limited sample size. The solid circle marks the TPP-FPP for a threshold of 0.8.

Figure 7

Comparison of ROC curves for CNN with those for two other methods. Same as Fig. 6 except the performance is shown for low and high mass ratios of BBH signals (left and right, respectively).

Figure 8

F1 score of our neural network for a threshold of 0.8. Scores are shown for low and high total mass of the BBH signals ($x$ axis) further split by their SNRs ($y$ axis).

Figure 9

Detection efficiency of our network on the real BBH events from GWTC-3. The format is the same as in Fig. 8 and the threshold is also set to 0.8.