Gravitational-wave signal recognition of LIGO data by deep learning

The deep learning method has developed very fast as a tool for data analysis in recent years. Moreover, as a technique, it is quite promising as a way to analyze gravitational-wave detection data. Multiple works in the literature have already used deep learning to process simulated gravitational-wave data. In this paper, we apply deep learning to LIGO data. In order to improve the weak signal recognition, we design a new structure of the convolutional neural network (CNN). The key feature of our new CNN structure is the sensing layer. This layer mimics matched filtering but is different from the usual matched-filtering technique. Usually, the matched-filtering technique uses a full template bank to match the data. However, our sensing layer only uses tens of waveforms. Our new convolutional neural network admits comparable accuracy and efficiency of signal recognition compared to other deep learning works published in the literature. Based on our new CNN, we can clearly recognize the 11 confirmed gravitational-wave events included in O1 and O2. In addition, we find about 2000 gravitational-wave triggers in O1 data.

Figure 1

Structure of the matched-filtering convolutional neural network (MFCNN).

Figure 2

SNR distribution of the training data set and testing data set for the simulated signals.

Figure 3

The total mass and mass ratio of training or test data and templates. The 11 GW events for both O1 and O2 are also shown.

Figure 4

(a) The ROC curves for test data sets containing signals with matched-filtering SNR 6, 8, 10, and 12, respectively. We plot the true alarm probability versus the false alarm probability estimated by our matched-filtering CNN network. (b) Efficiency curves for false alarm probabilities 0.1, 0.01, and 0.001, respectively. The true alarm probability is plotted as a function of the optimal SNR. In this figure, we use $p_c=0.5$, $N_t=35$ and $T=5$ seconds.

Figure 5

Comparison of different superparameters of the MFCNN network. (a) Comparison for different numbers $N_t$ of templates. Note that $\mathrm{SNR}=10$ and $T=5$ seconds are used. (b) Comparison for different data segment duration $T$. Here, $\mathrm{SNR}=10$ and $N_t=35$ are used.

Figure 6

The output confidence values of our MFCNN near the three GW events of O1.

Figure 7

The output confidence values of our MFCNN trained with O1 noise near the eight GW events of O2.

Figure 8

The output confidence values of our MFCNN near the three GW events of O2 newly reported by [64].

Figure 9

The spectrograms based on the Q-transform of the two detectors for the trigger on $2016-01-12$. The left diagram is for the Hanford detector, marked with H1. The right diagram is for the Livingston detector, marked with L1.

Figure 10

Comparison of distribution with respect to the MFCNN output $p$ value for triggers, all O1 data, and O1 data without triggers. Note we have chosen $p_c=0.5$, so the distribution for triggers starts from 0.5.