Neural network time-series classifiers for gravitational-wave searches in single-detector periods

The noise class corresponds to segments that are free of known GW signals, glitches (see next section) or hardware injections ¹⁰ . All segments in the dataset passed the first criterion, as no GW signals were confidently detected by standard pipeline over the selected period ¹¹ . Overall, a total of 1000 000 samples are obtained in this way from the one-month O1 dataset. Of these, 250 000 are only used to build the signal class for training as discussed in section 2.3, other 250 000 make up the noise samples for training and 500 000 are used both to generate the noise and signal samples for testing.

A database of glitches is created using two different sources: the unmodelled transient search coherent WaveBurst (cWB) [57, 58] and the citizen science project Gravity Spy [59].

The cWB pipeline is an open-source software package designed to search for a wide range of GW transients without prior knowledge of the signal waveform. To evaluate the analysis background, cWB uses a resampling technique [57] that involves applying non-physical time shifts to the data before analysis. Loud, i.e. high signal-to-noise ratio (SNR), background triggers resulting from this procedure are good candidates for glitches. The loudest triggers in LIGO Livingston with an SNR higher than 5.8 were selected (258 480 glitches). This list was complemented with 8244 glitches from the Gravity Spy database. These glitches are identified by the Omicron trigger pipeline [60] with an SNR above 7.5 and can be downloaded from zenodo [61] selecting the GPS time between 1132 444 817 and 1135 036 817, i.e. in the one month we are considering in this step of the analysis. The timestamps and duration of the identified glitches from these two sources are collected in a single list, which is then used to label the one-second data segments from the O1 observing run. If the glitch duration is shorter than 1 s, the associated segment is labelled as a glitch. Note that the glitch has a random position within the one-second window. If the glitch duration is longer than 1 s, all segments that overlap with that glitch duration are labelled as glitch. Only the glitches whose time belongs to the data segments available on GWOSC (i.e. when the detector is on observing mode) are considered. In many cases, the glitches are closer in time than one second, so multiple glitches can fall in the same one-second segment.

The number of glitches over a given period is determined by the occurrence frequency of those noise artefacts, typically tenths of seconds to tenths of minutes depending on the period, and on the typical glitch amplitude. As glitches occur sporadically, this represents a small fraction of the total observation time, thus resulting in a smaller number of samples in the glitch class. From the one-month O1 data, a total of $150\,000$ segments receive the glitch label.

The samples from the signal class are produced by adding simulated GW signals from BBH systems to the one-month O1 data in periods without known GW signals or hardware injections. For the training set, the data segments used to generate samples of the signal class are not utilised for the noise class nor the glitch class, while for the testing set, the same data segments are used for both the noise and signal classes. To generate the astrophysical signals, the waveform model SEOBNRv4 [62] is employed, with a lower frequency cutoff of 30 Hz. The simulated signals are sampled, whitened, and band-pass filtered in the same manner as the data segments.

The masses of the BBH used for generating the simulated signals in the class signal are chosen to ensure that they fall within the mass range observed by the LVK and that the signals are short enough to be contained within the one-second data segments. Specifically, the component masses m₁ and m₂ are chosen randomly, with the constraint that $m_1 \gt m_2 \unicode{x2A7E} 10 M_{\odot}$ and the total mass $M = m_1 + m_2$ is uniformly distributed in $33 M_{\odot} \unicode{x2A7D} M \unicode{x2A7D} 60 M_{\odot}$. We consider non-spinning BH, so the dimensionless spin magnitudes χ₁ and χ₂ are set to 0. The phase at coalescence and the polarisation angle are drawn uniformly in $(0, 2\pi)$, and the inclination angle in $(0, \pi)$. Since the focus is on a single detector, the right ascension and declination are not particularly important and are thus fixed to zero.

The amplitude of the added signals is computed such that the corresponding optimal SNR ${\rho_\mathrm{opt}}$ is uniformly distributed between 8 and 20. Following [63], it is defined as

$$\begin{equation} {\rho^{2}_\mathrm{opt}} = 4\int\nolimits_0^\infty {{{\vert \tilde h\left(f{\,}\right){\vert ^2}} \over {{S_n}\left(f{\,}\right)}}{\mathrm{d}}f}, \end{equation} \tag{ 1 }$$

where $\tilde h(f)$ denotes the Fourier transform of the template h(t) and $S_n(f)$ is the power spectral density of the detector noise. To generate the signals, a fiducial luminosity distance $d_\mathrm{L}$ of 100 Mpc is initially chosen, and then scaled to obtain the desired ${\rho_\mathrm{opt}}$. The final values of $d_\mathrm{L}$ range from 1 to 1300 Mpc approximately.

The simulated signals are added at a random position within the segment while ensuring the merger part of the signal is completely contained in the segment. More precisely, the whole signal is shifted in such a way that the merger time is in the interval [0.25 s, 0.8 s]. A total of 750 000 signal samples are generated.

Overall, the training set consists of $250\,000$ segments for the noise class, the same number for the signal class, and $70\,000$ for the glitch class. A 20% fraction of the training set is allocated for validation. The testing set, used to evaluate the classifier, comprises $500\,000$ samples for both the noise and signal classes, and $80\,000$ for the glitch class. This ensures sufficient statistical data for characterising the classifier's performance. In total, the training and testing datasets comprise 1650 000 one-second segments, with 45% for the noise class, 45% for the signal class, and 10% for the glitch class. This amounts to a storage space of 26 gigabytes. Out of the total number of segments, 28% is utilised for training, 7% for validation, and 65% for testing.

CNNs were first introduced for image classification [42]. They are now used for a wide variety of tasks, including the detection of GW signals [43–48]. In this study, we tested a range of CNNs similar to those considered in previous works.

We limited ourselves to shallow networks with five layers, four convolutional layers, and one final fully connected embedding layer. For simplicity, we only report here on the best-performing CNN, whose structure is detailed in table 1.

The convolutional layers are defined by the number of output filters, the length of the 1D convolution window (kernel size), the stride length of the convolution, and the activation function. The dense layer only requires the definition of the activation function. The input of inner convolutional layers is downsampled with a max pooling operation over a window size indicated in the table. The output of convolutional layers is processed by a dropout layer that randomly sets the input units to 0 with the frequency rate specified in the table. A global average pooling, followed by a dropout with a rate of 10%, is applied to the output of the last convolutional layer.

TCN [51, 68] is a neural network architecture specifically developed for sequence modelling problems. TCN has been shown to outperform generic state-of-the-art architectures over a diverse range of tasks and datasets. The TCN architecture is based on causal convolutions, where an output at time t is only convolved with past inputs from the previous layer. This allows the network to collect information from further in the past, using a combination of deeper networks (augmented with residual layers) and dilated convolutions where the filter is 'dilated' by inserting gaps between the filter coefficients. The size of the gaps is fixed to d − 1 where d is referred to as the dilation factor.

In this study, we have tuned the hyperparameters of the TCN model to find a compromise between the best performance and a reasonable training time. We ended up using a network with a TCN layer consisting of N = 6 dilated convolutional layers with 32 filters, a kernel size of k = 16, default values of dilation factors $d_{k = 1 \ldots 6} = (1, 2, 4, 8, 16, 32)$ for the six convolutional layers, and a dropout rate of 0.1. The output of the TCN layer goes into a final dropout layer with a rate of 0.5, and a dense embedding layer closes the model.

A key parameter that governs the training efficiency is the receptive field, which is the size of the region in the input data that produces a given feature in the output. The receptive field of the TCN can be expressed as $R = 1 \,+ \,2 \,(k-1)\, d_\mathrm{tot}$ where $d_\mathrm{tot} = \sum d_k$ [51]. With the above configuration, we have R ≈ 1900. The data used in this work have a sampling rate of 2048 Hz so each segment of data has 2048 points. The training with TCN is effective when R is much larger than the length of the input sequence [51]. To satisfy this constraint, only for this model, it is necessary to downsample the input data to 1024 Hz, therefore producing an input vector of size 1024.

IT [52] is a deep network ensemble designed specifically for time series classification. It leverages the concept of residual networks and incorporates Inception modules [69]. In a nutshell, the Inception module first produces a one-dimensional summary of the input multivariate time series (this is the 'bottleneck' layer), and then convolves this summary through multiple filters of different lengths, leading to a multivariate output that provides inherently multi-resolution features. The module output is finally reduced by max pooling (pool size of 4) before passing to the next module.

The IT architecture is composed of five ResNet networks with a sequence of depth d inception modules, with two residual blocks. The outputs of the five models are combined through a global average pooling and a final softmax layer, used to produce the classification probabilities for the different classes. In this study we have used the standard implementation of IT provided by the authors [52] with networks of depth d = 10, each with a bottleneck size of 32 processed through 32 filters with kernel sizes 20, 40 and 80.

The three classifiers are optimised using the training set described in section 2 to minimise the categorical cross-entropy loss function. The default implementations of the Adam optimiser are utilised, with a batch size of 24 [64]. The training procedure is repeated 10 times with different (random) initialisations of the model weights and dropouts, and the instance exhibiting the best receiver operating characteristic (ROC) curve on the testing dataset (as explained in section 4) is chosen. Note that this evaluation cannot be done with the validation dataset, as it does not provide enough statistics to compute the ROC in the relevant regime of low false alarm rates.

Throughout the training process, the model's area under the ROC curve [70] is evaluated on the validation data, and the model with the highest value is ultimately selected. The CNN, TCN, and IT models are trained for 50, 150, and 20 epochs, respectively. The best models are obtained at the 24th epoch for CNN, the 34th epoch for TCN, and the 5th epoch for IT ¹³ . On the Tesla K40d GPU we used, the training times per epoch were 220 s for CNN, 1000 s for TCN, and 3320 s for IT.

The final objective is to detect with high confidence the segments with a true astrophysical signal, i.e. to classify them as signal and to reject the other segments as noise or glitch ¹⁴ . We aim to constrain false alarms to a rate of two per day (similar to the current online search pipelines). This implies that we should reject all but one noise or glitch segment from the testing set in $1.7 \times 10^5$ trials.

The classifiers output the probability of class membership for each of the classes, that is three numbers between 0 and 1, summing to 1. The final detection is performed by applying a threshold to the membership probability $P_\mathrm{s}$ assigned to the signal class, which thus defines our decision statistic. The class membership probability is computed by the softmax activation function applied to the raw output (the 'logits tensor') of the fully connected embedding layer which concludes the classifiers. Because of the high-confidence level required, this threshold is very close to 1, thus requiring attention to the numerical precision for the evaluation of the membership probability. (This issue related to the precision of floating-point arithmetics was already noted in [48]). This has consequences on the way the classification loss is computed from the membership probability at the training stage. We found that the categorical cross-entropy loss should be directly computed from the logits tensor rather than from the class membership probability after the softmax transformation. The impact of this numerical issue is shown later in section 4.3.

We first assess the noise rejection capabilities of the classifiers. Figure 2 compares the distributions of the decision statistic $P_\mathrm{s}$ (the membership probability assigned to the signal class) when the input segment belongs to each of the three classes. The $P_\mathrm{s}$ distributions obtained with samples from the noise or glitch classes have very similar shapes, reflecting the intrinsic similarity of those two classes (see above). The best classifier is the one that provides the greatest contrast between the distributions of the $P_\mathrm{s}$ statistic obtained in the presence of a signal (blue) versus noise or glitch (red dot-dashed and black).

Zoom In Zoom Out Reset image size

The distributions obtained for the noise or glitch classes exhibit maxima at zero for the TCN and IT classifiers, while the maximum is shifted to around 0.1 for CNN. Moving from the peak to $P_\mathrm{s}$ higher values, the distributions show a monotonic decay until the last bin $P_\mathrm{s} = 1$ where distributions exhibit an count increase which is more prominent for TCN ¹⁵ . The TCN classifier appears to reach the lowest background $\lesssim$10⁻² in normalised count units.

Since our objective is to achieve high-confidence classification, we are primarily interested in the immediate vicinity of $P_\mathrm{s} = 1$. This motivates us to reparameterise the $P_\mathrm{s}$ statistic as $\lambda : = -\log_{10}(1-P_\mathrm{s})$. While $P_\mathrm{s}$ ranges from 0 to 1, λ can theoretically take values across the entire real line. Our main focus lies in the range of large λ values, that is $\lambda \gtrsim 7$ because the computations are performed in single precision ¹⁶ . The most stringent criterion is to require $P_\mathrm{s} = 1$ at machine precision, which corresponds to $\lambda = \infty$. The number of noise and glitch samples in the testing set that satisfy this selection criterion is 0, 1, and 2 for the CNN, TCN, and IT classifiers, respectively. Such rejection power (between 0 and 2 false alarms in $5.8 \times 10^5$ trials) is in agreement with the false-alarm rate targeted initially.

We proceed to assess the classifiers' ability to extract signals. Figure 2 illustrates the distributions of the decision statistic $P_\mathrm{s}$ when the input segment belongs to the signal class, represented in blue. As anticipated, all distributions exhibit a peak at $P_\mathrm{s} = 1$. However, the peak appears narrower for the IT classifier. To focus on the region of interest near $P_\mathrm{s} = 1$, we employ the λ reparametrisation, as depicted in figure 3. This figure also incorporates the dependencies on the SNR of the injected GW signal and the chirp mass ${\cal M}$ of the source binary. The distributions are computed separately for three ranges of chirp mass ${\cal M}$: low, mid, and high, corresponding to ${\cal M}$ values between 13 and 17 $M_{\odot}$, 17 and 21 $M_{\odot}$, and 21 and 26 $M_{\odot}$, respectively. The histograms on the right-hand side are computed with the samples of the ${\textsf{signal}}$ class that are classified with $P_\mathrm{s} = 1$, showing their distribution in terms of SNR for the three chirp mass ranges.

Zoom In Zoom Out Reset image size

It is worth noting that we have either $\lambda \lesssim 7.5$ or $\lambda = +\infty$ (i.e. $P_\mathrm{s} = 1$). ¹⁷ In a sense, the latter case seems to accumulate all samples with $\lambda \gtrsim 7.5$. From the left column of the figure, it is apparent that the IT classifier assigns larger λ (or $P_\mathrm{s}$) values more uniformly over the full range of chirp mass and to lower SNR. In contrast, the CNN fails to do so for the lower chirp mass interval shown in blue. This is confirmed by the histograms in the right column, which indicate that the TCN and IT classifiers have a higher overall count (approximately 76%, compared to 65% for CNN) and extend to lower SNR values. In general, the histograms with $P_\mathrm{s} = 1$ tend to be more populated at large SNR. For the samples classified with $P_\mathrm{s} \lt 1$, the correlation between the SNR and λ is visible for IT and less clear for the other classifiers, for which the chirp mass seems to play a more important rôle.

To fully characterise the performance of the classifier, the noise rejection and signal extraction capabilities have to be evaluated jointly. This can be done by computing the ROC curves [70]. The classification efficiency $S_\mathrm{th}/S_\mathrm{tot}$ and false alarm rate $N_\mathrm{th}/N_\mathrm{tot}$ are evaluated from the testing set, with $S_\mathrm{th}$ and $N_\mathrm{th}$, the number of signal samples and noise and glitch samples with a $P_\mathrm{s}$ value above some threshold, and $S_\mathrm{tot}$ and $N_\mathrm{tot}$ the total number of samples for each category. Note that since each sample has a duration of 1 s, $N_\mathrm{tot}$ is intended here as the total duration in seconds of noise and glitch samples, so $N_\mathrm{th}/N_\mathrm{tot}$ is measured in $\mathrm{s}^{-1}$. By varying the threshold, one obtains the ROC curves in figure 4 which displays the classification efficiency versus the false alarm rate. The TCN and IT classifiers appear to have similar ROC curves and show a clear improvement with respect to CNN. Figure 4 also shows the ROC computed for two instances of the IT architecture, when the categorical cross-entropy loss is calculated from the logits tensor (green) and when it is calculated from the class membership probability after the softmax transformation (red), see section 3.5 for an explanation and discussion. The shaded area represents the range between the best and worst models among the ten instances computed at training. When softmax is used, the uncertainty in the performance is larger (the shaded area is wider) and the classification accuracy reached at small false alarm rates is lower.

Zoom In Zoom Out Reset image size

Figure 5 shows the classification efficiency for a given false alarm rate set to 10⁻⁵ s⁻¹ (that is 2 per day), as a function of the injected SNR as defined in equation (1). The classifiers TCN and IT give similar efficiencies and surpass uniformly over CNN. Note that the efficiency shown in this figure is averaged over the full chirp mass range and thus does not show the differences evidenced in figure 3. Overall, signals with SNR = 10 can be detected at the considered significance level with a good probability, larger than 50%.

Zoom In Zoom Out Reset image size

Another standard metric to assess the performance of the classifier is the sensitive distance, see e.g. [71]. This metric can be computed by running the classifiers over a new test set for the signal class generated using the same configuration as above, but with sources distributed uniformly in volume instead of uniformly in SNR ¹⁸ . We obtain a sensitive distance of about 500 Mpc for the IT, TCN and CNN classifiers for a false alarm rate of 10⁻⁵ s⁻¹. The sensitive distances measured by [49] (dataset 4, using O3a data from LIGO Livingston and Hanford) range from 700 to 1700 Mpc approximately with the same requirement on the false alarm rate. In this range, we have ignored the method 'TPI FSU Jena' [48] whose sensitive distance drops close to 0 Mpc when applied to real data. This comparison is only indicative as the search sensitivities evaluated in [49] are for a two-detector network.

We first investigate how the three events detected in the O1 data [8] by matched filtering searches are classified by the considered models. The statistic $P_\mathrm{s}$ is evaluated for different positions of the one-second window, that is for different time delays Δt between the start of the analysis window and the merger time. This definition implies that, for $\Delta t = -1$ s, the analysis window only includes the initial part of the signal (inspiral), whereas, for $\Delta t = 0$ s, the analysis window starts at the merger time and thus only includes the final part (merger and ringdown). Figure 6 shows the evaluation of $P_\mathrm{s}$ between those two extreme cases for the IT model for GW150914, GW151012 and GW151226 (see also appendix).

Zoom In Zoom Out Reset image size

As expected, when the merger is not included in the analysis window, the classifier is not able to detect the presence of the signal. GW150914 appears to be loud enough to be always identified with $P_\mathrm{s} = 1$, regardless of its position in the time window, even if it is partially visible. GW151012 is only detected with $P_\mathrm{s} \sim 0.9$ when the merger is at the centre of the analysis window. GW151226 is not detected (i.e. $P_\mathrm{s}$ is always below 0.2). This is expected as the binary component masses are outside the range used to generate the astrophysical signals in the signal class of the training data. Both GW151012 and GW151226 have single detector optimal SNRs for Livingston from parameter-estimation analyses lower than the minimum value of 8 we used to train the network (namely, $5.8^{+1.2}_{-1.2}$ for GW151012 and $6.9^{+1.2}_{-1.1}$ for GW151226 according to table V of [8]).

We analysed all the remaining L1 data in O1 excluding the month we used to train and test the classifiers (see section 2). This corresponds to the period between GPS = 1126 051 217 (9 December 2015 00:00:00 UTC) and GPS = 1132 444 817 (25 November 2015 00:00:00 UTC) and between GPS = 1135 036 817 (25 December 2015 00:00:00 UTC) and GPS = 1137 254 417 (19 January 2016 16:00:00 UTC). In this period we excluded the intervals of ±1 s around the chirp time of the three known events (see previous section). This amounts to a total of 4216 489 s (about 49 days), of which 1054 564 s (about 12 days) are single-detector times, corresponding to 25% of the total. This data set is whitened following the same procedure used to produce the training set (the ASD was calculated from periods of non-interrupted data taking with 26 s minimum and 146 978 s maximum). The data are then divided into non-overlapping one-second segments that are processed through the three classifiers. For each, we used the best-performing model on the testing data. The processing time for the full data set is about 4 h per model on NVIDIA Tesla V100S GPUs, but most of this time is taken to load the data, the extraction of the model predictions takes about 8 min for CNN, 18 min for TCN and 52 min for IT. No data quality information was used, so this analysis is solely based on the GW strain data.

Figure 7 shows the distribution of the $\lambda = - \log_{10} (1-P_\mathrm{s})$ statistic obtained with the IT classifier (similar plots can be found in appendix for the other models). To find GW signals in these data we apply the most restrictive selection cut, by requiring $P_\mathrm{s} = 1$ (at machine precision). In other words, detected events correspond to segments classified with $P_\mathrm{s} = 1$. Given the result of the previous section, this implies that, with this selection cut, GW150914 would be the only detected event among the three known detections during O1. We recall that this selection cut corresponds to a false-alarm rate of $\lesssim 4 \times 10^{-6}$ s⁻¹ (that is one false alarm per 3 days) and a classification efficiency of 76% when estimated on the testing set, see sections 4.1 and 4.2. Based on these results, we estimate from basic counting statistics that the maximum number of false alarms expected for this analysis should be 29, 43 and 55 for CNN, TCN and IT respectively at 95% level for the full data set, and 9, 13 and 16 when restricting to the single-detector part.

Zoom In Zoom Out Reset image size

For the IT classifier, a total of nine segments pass the selection cut, with two occurring in single detector time at GPS = 1131 289 775 (11 November 2015 15:09:18 UTC) and GPS = 1135 945 474 (4 January 2016 12:24:17 UTC). For the CNN and TCN classifier, we obtain 4 and 105 segments passing the cut, with 2 and 14 falling in single detector periods. The results are thus consistent with the expectations for CNN and IT, while there is a clear excess with TCN. We have observed that a significant fraction of the triggers comes from two time intervals around 20 October 2015. Our interpretation is that the data from those periods could differ in nature from those of the training set, and TCN may be sensitive to this difference.

Interestingly there is only one segment passing the selection cut for all three classifiers: GPS = 1135 945 474 (4 January 2016 12:24:17 UTC) which we investigate further in the next section. As single-detector searches cannot employ statistical resampling techniques with time shifts [73], we can only provide an upper limit on the false alarm rate for this detection. The upper limit is estimated to be 1 event every 49 days, based on the available data from the three-month analysis period. This segment on 4 January 2016 corresponds to the event identified in the Livingston detector data during the O1 single-detector periods using a standard matched-filtering-based search, as reported in [36]. However, this candidate is subsequently eliminated by the authors of [36] after examining the residual obtained by subtracting the best-fit waveform from the data, since excess power is observed in the residual at frequencies below 80 Hz.

We have performed a number of detailed checks of the 4 January 2016 event. We have performed a 'visual' inspection with the time-frequency constant-Q transform [55, 74]. Figure 8 provides a time–frequency representation of the entire segment. A transient is visible ∼0.37 s after the start of the segment, at a frequency of about 150 Hz. In the magnified view, the shape of the transient is clearly indicative of a frequency modulated chirp-like transient.

Zoom In Zoom Out Reset image size

The Gravity Spy database [72] has marked this specific GPS time classified as being an instrumental artefact of the 'Blip' type. The term refers to a well identified family of instrument glitches whose origin is still largely unknown (see, e.g. [14, 75] for more details). Generally, 'Blip' glitches do not exhibit a chirping frequency (see figure 1 of [76] for a typical example). To complement this initial inspection, figure 7 gives in pink the statistic λ (or equivalently $P_\mathrm{s}$) of the 600 blip glitches listed in Gravity Spy overlapping with the part of the O1 dataset being analysed. The resulting distribution is compatible with the overall background distribution. The 4 January segment appears to be an outlier with respect to the blip glitches identified in the data.

Further, we checked if the transient signal can be fitted by a GW waveform model associated to a compact binary merger. To do so, we ran the Bayesian inference library Bilby [77] and used the IMRPhenomXPHM waveform model [78]. It is assumed that the component spins are co-aligned with the orbital momentum. For the rest of the source parameters, generic and agnostic priors are assumed, along with a standard Λ-CDM cosmology model with $H_0 = 67.9\,\mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1}$ [79]. The analysis did not include a marginalisation over calibration uncertainties. The analysis results in a signal-versus-noise log Bayes factor of 47. The estimated time of arrival of the merger at the detector is GPS = $1135\,945\,474.373^{+0.076}_{-0.07}$ and the measured optimal SNR is $11.34^{+1.8}_{-1.6}$.

Figure 9 shows the result of the fit in the time domain, by comparing the whitened data in orange to the reconstructed waveform corresponding to the maximum likelihood fit (green), and from the posterior mean (blue), shown with the 90% credible belt. We do not visually identify any notable residual after subtraction of the best fitting waveform as shown in figure 10. As an independent check of the nature of the signal, figure 9 also includes the waveform estimate produced by the denoising convolutional autoencoder described in [80] (dashed red). The reconstructed waveforms are in good agreement, following a similar phase evolution, except for the initial and final parts of the signal, where the denoiser's reconstruction is not optimal because of its low-frequency cut-off (see [80]), and lower SNR of the signal. In addition, we note that the denoising autoencoder was trained on the waveform family SEOBNRv4 which is different than the one used for Bilby (IMRPhenomXPHM), which may contribute to differences.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Following [80] we calculate the SNR of the denoised waveform using equation (1) and obtain a value of 9.7. Bacon et al [80] investigates the ability of the denoising algorithm to denoise data in the presence of glitches depending on the denoised SNR, see figure 5 of [80]. A value of 9.7 is shown to be sufficiently high to hint an astrophysical origin of the signal.

This candidate signal was also identified by [36] using standard matched filtering techniques, but the event was subsequently downgraded to a noise artefact due to an excess observed in the residual. Figures 8 and 10 are produced with similar dynamic scale, colour range and time-frequency resolution as in [36]. Therefore, the mismatch either comes from a difference in the parameters estimated for the best fitting waveform (unfortunately, not included in [36]), from the difference in the waveform model (IMRPhenomPv2 was used by [36]) or both.

Above checks are all compatible with the event being of astrophysical origin. The corner plot in figure 11 displays the posterior distribution of the source parameters including the binary component masses, spins and source distance. Since only one detector is available, the source direction is not localised in the sky. The 90% credible intervals for those parameters are: the measured (redshifted) chirp mass ${\cal M} = 30.18^{+12.3}_{-7.3} M_\odot$, the (redshifted) component masses $m_1 = 50.7^{+10.4}_{-8.9}\,M_{\odot}$ and $m_2 = 24.4^{+20.2}_{-9.3}\,M_{\odot}$, the binary effective spin $\chi_{\mathrm{eff}} = 0.06^{+0.4}_{-0.5}$ and the luminosity distance $d_\mathrm{L} = 564^{+812}_{-338}\, \mathrm{Mpc}$; see [81] for a definition of those physical parameters. Overall, these values are consistent with the observed population of BBH to date.

Zoom In Zoom Out Reset image size