The characterization of Virgo data and its impact on gravitational-wave searches

Seismic activity is probably the most pervasive source of noise in Virgo, affecting the detector in many different ways. Almost every Virgo sub-system is sensitive to sufficiently large vibrations. Seismic noise can produce a large variety of glitches which are very difficult to track. Loud seismic glitches due to violent shocks or earthquakes are likely to produce noise in the DF. If this happens, the data recorded during a seismic event is rejected and so the noise coupling is less relevant. The low frequency signals collected by the multiple seismometers and accelerometers on site are used to define DQ flags for large seismic activity. Several frequency bands are monitored at all times, from 0.25 up to 16 Hz. A priori, such low-frequency seismic glitches should not be an obstacle for the transient GW searches whose frequency band usually starts above 40 Hz. However, seismic noise is often up-converted in frequency, for example through scattered light mechanisms as described in [56]. For example, the seismic glitch presented in figure 3 was detected by the seismometers at about 8 Hz and is seen in the DF signal at much higher frequency (∼200 Hz).

Bad weather conditions can increase the seismic activity and cause significantly deteriorated data quality. In such conditions, the Omega pipeline shows an excess of triggers at low frequency (typically below 100 Hz). In the case of very bad weather, the Omega trigger rate below 100 Hz can increase by a factor 5 to 7. During the winter, VSR2 showed many periods of high seismic noise. Substantial efforts were devoted to studying the resulting glitches [56]. One family of scattered-light glitches was characterized by no visible glitch in the Omega time–frequency maps of the seismic sensors. This fact indicates that a nonlinear coupling was in action. The time–frequency shape of these triggers is very well-recognizable (see the second plot of figure 3). It consists of a series of arch-shaped glitches that can last several seconds. The glitches are caused by light scattered by e.g. the tower walls or the suspended baffles moving with the micro-seismic motion of the ground. When the micro-seismic activity is large, higher harmonics can be seen, probably due to multiple-bounce optical paths. In such conditions, several rows of arch-shaped glitches can be seen in time–frequency maps. This noise is well-modeled and the frequency of the arches is proportional to the velocity of the scattering object. Tests showed that the position sensors installed at the top stage of the suspensions are well suited to measure the velocity. The scattered-light glitches can be rejected when thresholding on the measured velocity. When applying the DQ flag created in this way, 8% of the science time is lost but 2/3 of the scattered-light glitches were vetoed. The coupling mechanism for the scattered-light glitches was understood during VSR2 when it was noticed that most of the scattered light was re-injected into the beam at the level of the west-end optical bench. For VSR3 the number of scattered light glitches decreased because of the lower transmission of the new end mirrors and the installation of absorbing baffles in the west-end tube.

Acoustic disturbances can mechanically affect the Virgo systems and produce glitches. Acoustically-isolating enclosures have been installed around each optical bench in order to limit the acoustic coupling with the environment. However, acoustic pollution can either be produced inside the enclosure or can get inside through mechanical vibrations. To monitor acoustic noise, each building is equipped with several microphones. Most of the time, the acoustic disturbances originate from mechanical devices located near the interferometer which can be mitigated. However, acoustic noise can also have an external origin which cannot be controlled or suppressed. Several times during the day, airplanes or helicopters fly over Virgo and they are seen in the DF signal. These glitches can be clearly identified by the typical Doppler shift at about 100 Hz seen in time–frequency maps of the detector output (see an example in figure 3) [57].

Electrical cables represent a major source of noise coupling since they can propagate an electrical disturbance throughout the Virgo site. The Virgo sub-systems are usually designed to be electrically isolated from the environment. However the 50 Hz mains frequency (European standard) can couple into the detector and transmit magnetic transients. For example, during VSR2, it was noticed that a family of glitches was periodically produced roughly every 15 min. This effect was identified as electrical coupling of an air-conditioning unit switching on and off. During the winter period of VSR2 a loud glitch was produced every day at 8 am due to the heating system that switched on at the beginning of the day and drew a significant amount of current. Such glitches can be vetoed by using magnetic sensors that are sensitive to electrical transients. Electrical glitches are usually corrected by breaking the electrical noise path. In the two specific cases here, the glitches disappeared after upgrades to the detector electronics.

One critical element of the Virgo detector is the main laser injection system. This system contains many control loops to stabilize the laser power and frequency. Failures in these control systems caused various families of glitches. During VSR3, the laser power stabilization control loop was experiencing saturations due to a mis-tuned gain. This created strong broadband glitches from tens to thousands of Hz (see figure 3). This was fixed a few days after the problem was discovered. In the meantime a specific DQ flag was built to monitor the control loop channel and to efficiently exclude the glitches from the data attributed to the control failure.

Most of the laser light propagation is done in a high-quality vacuum. However the beam propagates through air in some parts of the detector, for instance on the injection and detection benches. Some disturbances, due to dust crossing the beam, for example, create glitches which are difficult to veto. The laser light propagation can also be disturbed by unexpected events like spiders building webs or bugs flying through the beam. It is possible to limit such pollution by protecting the laser path with plastic covers. Some of the remaining glitches can be vetoed by using the photo-diode signals of the secondary beams which are not sensitive to GW signals. Many DQ flags were created in this way. A very efficient veto was introduced in VSR1 which relies on the fact that a real GW event seen in the in-phase demodulated DF channel should not be visible in the quadrature channel if the demodulated phase is well-tuned. This PQ veto [58] has been extensively used to eliminate these potential 'dust events' in the Virgo data.

The alignment of the main optical beam is critical in order to maintain the detector in operation. Sophisticated feedback systems are required to continuously control the optical component angular degrees of freedom and to optimize the laser beam alignment [59, 60]. In the Virgo sensitivity curves shown in figure 2(a) several spectral lines are known to correspond with resonances of some optical mounts of the detection bench (165, 210, 420, 495 and 840 Hz) and are due to light scattered by these optical components. In principle they should not be seen as glitches unless they suddenly vary in amplitude which can happen when the interoferometer alignment conditions change. This effect of non-stationary lines is a well-known source of glitches to which transient GW searches are very sensitive. In Virgo, alignment glitches represent a quite large fraction of Omega triggers (about 25%). In the case of bad weather, alignment fluctuations are even larger. As a result, the fraction of glitches due to alignment reaches 40% and the amplitude of the glitches increases. Alignment signals can be used to build DQ flags to suppress these alignment glitches. For VSR2 and VSR3, large deviations of the mirror angular positions were flagged. This allowed for the removal of as much as half of the alignment glitches.

The TCS [46] was installed in Virgo between VSR1 and VSR2. A TCS instability can directly influence the DF signal by producing a thermal or a radiation pressure disturbance at the mirror level. During VSR2, the TCS laser has been stabilized, reducing the number of glitches. However, it was necessary to build a specific DQ flag, using the channel monitoring the TCS power, to veto the remaining glitches. Figure 3 shows an example of a TCS glitch that is vetoed by a DQ flag.

Very loud glitches can be produced by the saturation of different Virgo active systems. For instance, every photo-diode must operate within its nominal range (±10 V). Specific DQ flags have been introduced to reject noise transients whenever a photo-diode voltage is out-of-range. Similarly, the mirror coil driver currents are monitored to check for saturations.

In data taken two years before the first science run, we identified a nonlinear coupling between the dark fringe and the laser frequency noise. Laser frequency noise usually lies well below the shot noise level at high frequencies (see figure 2(b)). Every 27 s, broadband glitches were visible in the DF signal. This period corresponds to a mechanical resonance in the lower part of the mirror suspension. The periodic noise increase was correlated with the extremal angular tilt of the Fabry–Perot cavity's mirrors. When the mirrors are badly aligned the coupling of the laser frequency noise increases. To cure this problem, the mirrors' alignment control loops have been greatly improved. A veto using the direct measurement of the laser frequency noise in the DF signal (a line at 1111 Hz was injected in the laser frequency control system) was created to efficiently eliminate all of the periods containing this noise [61].

The Virgo detector has many piezo-electric drivers used to control various elements of the beam path. At the beginning of VSR1, one of the four piezos of the beam monitoring system was malfunctioning, causing the input beam to jitter [62]. This jitter can couple to interferometer asymmetries and was a source of glitches in the DF signal. The typical frequency of theses glitches was around 150 Hz. This problem was discovered during the first month of the run and the piezo was replaced two months later. A similar problem occurred during VSR3 at the output mode cleaner. A piezo voltage was found noisy for several hours. The faulty piezo elements were fixed but DQ flags, based on the control channels, were defined in order to completely exclude the glitches from the data recorded while the piezo was faulty.

During VSR2, some glitches were observed with the distinctive feature of an abrupt step in the h(t) time series which resulted in a loud broadband disturbance. Theses glitches were demonstrated to be associated with an excitation of the internal modes of the west-input (WI) or west-end (WE) mirrors (depending on the glitches), identified by their accurately known frequencies. The glitches were interpreted as a sudden displacement of the surface of those mirrors, of unknown origin. In the case of the WI mirror, the glitches appeared after the magnets glued on the back of the mirror were replaced using a type of glue that had not been used before for that purpose, suggesting the possibility of a creeping mechanism in the hygroscopic glue. It was not possible to firmly confirm this suspicion, and the cause of the WE mirror noise still lacks a convincing explanation. It was impossible to safely veto those glitches, due to the lack of independent auxiliary information.

The external temperature can also be an indirect source of glitches. The steel vacuum tubes, in which the laser travels, have poor thermal isolation. Hence, external temperature variations are very likely to mechanically stress the tube through contraction and expansion. During VSR4, it was understood that when the expansion/contraction force exceeds the static friction which holds the tube on its support, a sudden shock occurs and a mechanical vibration propagates along the tube. This effect is the strongest around noon and midnight when the temperature gradient is the largest. Seismometers have been installed to track the noise propagation and it was found that the noise source was the tube between the IMC and the injection tower. The resulting glitches are produced in the DF signal at about 80 and 160 Hz. A seismometer placed on the injection tower allowed us to flag these glitches with high efficiency.

High-frequency electromagnetic noise overlapping with the laser modulation frequency (6.26 MHz) can be picked up by the DF photo-diode signal before demodulation. It can then enter the detector's sensitive band after demodulation. High-frequency electromagnetic transients are generated, for example, by fast switching electronic devices (i.e. power supplies with a typical switching rate of 100 kHz and above), and by data flow to/from digital devices (the clock rate of communication protocols is typically in the MHz range). During VSR2 the 6.26 MHz modulation signal was intermittently polluted by a large amount of glitches that were also seen in the DF signal. The origin of this noise has never been identified, mostly because of its intermittent nature. The noise was suspected to originate from serial transmission devices. It was possible to build a DQ flag based on the modulation signal to remove the glitches seen in the DF signal. The beginning of VSR3 showed a large excess of Omega glitches at high frequency (at 1 kHz and beyond). This was identified as a result of a coupling between the modulated DF signal and an electromagnetic field whose frequency was close to the modulation frequency (see section 5.2.2 for more details).

The Virgo interferometer is kept at its working point by various digital control loops. The control servos dedicated to longitudinal control are fast control loops running at 10 kHz and any digital problem occurring in theses systems can directly affect the DF signal. One example is a set of loud glitches in VSR1 that were due to a loss of synchronization in the control system. This led to dropped samples between the global control system (which provides the 10 kHz signals for the interferometer's longitudinal control) and the digital signal processing board in charge of filtering the correction signal before it is sent to a mirror's coil. Combined with a strong but harmless 5 kHz oscillation that is sometimes present in the control signals, the dropped samples produced loud glitches which were vetoed offline by searching for missing samples within the 5 kHz noisy time periods.

All vetoes, except the PQ veto [58], are derived from channels that are assumed to be independent of the DF (which may contain a GW signal). By accident, a veto can dismiss a genuine GW signal, but the probability of such an event must be small and follow the Poisson probability of coincidence between two random processes. To test that a veto is safe, fake GW signals are injected into the interferometer by applying a force on one mirror of one Fabry–Perot cavity to mimic the path of a GW event (hardware injections). Different types of signals are injected, but to test the veto safety, the very loud (SNR ∼ 100) GW burst signals were used (Sine Gaussian waveforms with a frequency between 50 and 1300 Hz). These hardware injections, grouped by 10, are regularly performed at a rate that varies between once a day and once each three days, depending on the science run. We count the number N_flagged of vetoed hardware injections. This number is compared to the expected number of hardware injections accidentally vetoed:

$$\begin{equation} N_{\rm flagged}^{\rm exp} = \frac{T_{f}}{T_{\rm tot}} \times N_{GW} , \end{equation} \tag{ 1 }$$

where T_f is the time rejected by the flag, T_tot is the total science time and N_GW is total number of hardware injections performed during T_tot. The Poisson probability to have N_flagged or more events when N^exp_flagged are expected is simply

$$\begin{equation} p (N \ge N_{\rm flagged}) = \sum _{n=N_{\rm flagged}}^{n= \infty } P\big(n,N_{\rm flagged}^{\rm exp}\big), \end{equation} \tag{ 2 }$$

where P(n, λ) is the Poisson distribution of mean λ. This defines the probability that the veto is safe. Setting a threshold on this quantity provides an automatic means to determine which veto is unsafe. Two thresholds were considered: when the probability is lower than 10⁻⁵, the flag is considered unsafe. When the probability is below 10⁻³, all flagged hardware injections are manually inspected to determine if this low probability is due to the fact that a long DQ flag segment has vetoed several hardware injections belonging to the same series since the hardware injections are grouped by 10, each separated by 5 seconds. It has been checked that a priori unsafe flags based, for instance, on channels that are known to contain a fraction of a GW signal have a probability well below 10⁻⁵. On the other hand, all flags, a priori safe but with a probability between 10⁻⁵ and 10⁻³ were found to be safe, the low probability being due to the effect explained above.

A data quality flag is said to have good performance if it is able to veto glitches affecting an analysis pipeline without vetoing long periods without noise transients. DQ flag performance is measured by considering a set of N_t triggers spanning a large frequency band and the science period T_tot of the GW transient searches (the mean trigger rate is R = N_t/T_tot). Each DQ flag is characterized by the number N_seg of disjoint time segments and the total time T_f rejected by the flag. Three figures of merit, discussed in details in [67], are used to measure the flag performance:

• (i)  
  The efficiency () measures the percentage of triggers vetoed by a DQ flag and is given by N_f/N_t, where N_f counts the number of flagged triggers.
• (ii)  
  The use-percentage (UP) gives the fraction of DQ segments which are actually used to veto triggers and is given by N_use/N_seg, where N_use is the number of segments used to reject at least one trigger. When this number is close to 1, the flagged time period certainly contains a glitch that the DQ flag was designed to veto.
• (iii)  
  The dead-time (D = T_f/T_tot) is the percentage of science time rejected by a flag.

It is often convenient to compare the efficiency to the dead-time in order to make sure the flagging is not random. In case of random flagging, N_f = R × T_f and /D = 1. On the contrary, if the DQ flag is highly selective for glitches, /D > 1. Finally, a DQ flag has /D < 1 if it tends to systematically flag periods of time where no triggers can be found. These figures of merit must be used with care and have limitations. In particular, they are average numbers and they may be biased by large variations of trigger rate or by the segment structure of the DQ flag.

A DQ flag's performance and the level of understanding of its corresponding noise source determine at which stage of a GW search the DQ flag should be applied. DQ flags are divided into three categories: CAT1–3, defined in table 2. When a severe malfunction prevents the detector from working in normal operating conditions, the corresponding period must be discarded from the GW searches. Such CAT1 DQ flags are used to re-define the science segments on which analysis pipelines are run. CAT2 DQ flags are characterized by high performance resulting from a good description of the noise source and its coupling with the DF. CAT2 flags can be applied with confidence to the output of transient GW searches. CAT3 are effective at removing transient noise from the data, but in the presence of a weak physical coupling, caution is exercised when using these flags. Furthermore, CAT3 vetoes typically have an overall larger dead-time than CAT2 flags (∼10%). For these reasons, transient GW searches are usually performed in two steps: in the case of CBC searches the search output, with both CAT2 and CAT3 flags applied, is first considered to make statements about the significance of GW candidates or, in the absence of a detection, to derive upper limits on the GW event rate. If no GW candidate is observed, GW candidates after CAT2 flags have been applied are considered. This search is less sensitive, since the noise background is higher, but it allows to ensure that a significant GW event has not been vetoed accidentally because of a large dead-time veto. For the burst searches, all triggers after CAT2 flags are taken into account. CAT3 flags are then used when computing their significance [15]. Some additional DQ flags are uncategorized, because of very low performances or highly uncertain coupling mechanisms. These flags are only considered during the follow-up procedure if a GW event is found to be significant [68]. Further studies of auxiliary channels at the time of the event may rule out an astrophysical origin for the event.

The definition of flags for transient glitches is an iterative process. Most DQ flags are produced online, but from one run to another, the noise coupling can change as new noise sources appear and existing noise is mitigated. The next step is to estimate the performance of the DQ flags used by each data analysis search in order to veto transient features and to determine what noise sources remain after the application of DQ flags. This will be discussed in section 6.1.

All the machinery operating frequencies constitute a seismic background due to the engine vibrations. The on-site seismic sensors reveal a 'forest' of spectral lines up to 600 Hz, whose amplitude roughly decreases as f⁻², meaning they have a roughly constant energy content. As explained in section 4.2.1, seismic disturbances are likely to couple to the DF through a variety of mechanisms. It is often not possible to disentangle all of the spectral lines and to link them to a specific noise source. However some couplings were identified and are explained below.

One well-known vibration noise path is located in the injection bench where the laser beam travels a few meters through optical components for shaping and alignment purposes before entering the interferometer. Vibrating optics add angular jitter noise to the beam. Moreover the Virgo in-air input bench has large quality factor (20–40) resonant modes around 15–20 Hz and 45 Hz which are associated with the small rigidity of the supporting legs. These frequencies happen to exactly match the vibration noises of cooling fans (around 45 Hz) and of some vacuum motor fans (18 Hz); the noise is therefore amplified. Mitigation was attempted before VSR2 by moving fan-cooled electronic racks to a separate acoustically-isolated room. Existing optical mounts were also replaced with more rigid ones. By doing this, the resonant modes were shifted to higher frequencies where the vibration noise is weaker. Between VSR1 and VSR2, the accuracy of the interferometer global alignment [60] was improved which also significantly helped reduce alignment noise due to vibrations.

Scattered light often results from vibrating objects such as lenses, vacuum link windows or vacuum pipe walls. As discussed in section 4.2.1, scattered light can be an important source of noise. If the vibration is periodic, a spectral line will be visible in the DF signal. As an illustration, the lower-left plot of figure 4 shows an example of a noise line caused by scattered light. The line at 45.5 Hz is associated with the vibrations of turbo pump cooling fans which propagate to the vacuum tank walls. This noise was mitigated 90 days after the start of VSR2 by seismically isolating the fans from the vacuum tank.

During VSR1, another source of scattered light was discovered in the detection system; a glass window used to isolate the detection vacuum compartment from the rest of the interferometer was acting as an efficient transducer of seismic and acoustic noise from the external environment to the detector. To cure this problem before VSR2 the window was removed and replaced with a larger aperture pipe with an associated cryogenic pump. Similarly, during VSR2, some light was scattered back into the interferometer by the main beam output window. Improving the quality of the window anti-reflection coating reduced the noise to a negligible level.

Sometimes the noise path of a spectral line cannot be identified. In this case NoEMi can provide useful hints e.g. by connecting a noise line to a noise source based on coincidences with auxiliary channels. For example, during VSR3, the correlation between the frequency variations of a line in the DF and a line detected in a seismometer allowed for the identification of the coupling with the vibration of an air-conditioning fan (see upper row of figure 4). This specific spectral line has been moved out of the detector's sensitive band by reducing the fan rotation speed.

Electromagnetic (EM) fields produced by electrical systems are likely to contribute to the noise spectral lines, especially through their magnetic component. The noise strength will depend on the intensity of the field, its frequency and the source distance. Virgo is mostly sensitive to low frequency EM fields (frequencies less than a few hundreds of Hz) which couple directly into the detector bandwidth. At higher frequencies, radio-frequency EM fields are the main source of noise, since a 6.26 MHz frequency is used to phase-modulate the Virgo laser beam. Hence, it is important to keep the frequency region within ± 10 kHz around the modulation frequency as free as possible of EM noise.

The main magnetic noise entry path is located at the level of the actuators used to control the mirror positions. These actuators are made of coils and magnets which can be disturbed by the presence of a magnetic field and its gradient. During VSR1 small magnetic sources (e.g. power supplies and cooling fans used for the mirrors local control electronics) were located a few meters from the mirrors. The magnetic field radiated by these small magnetic-dipole-like sources decays quickly with distance (as d⁻³), and so it was sufficient to move them away by a few meters to reduce their effect to a negligible level. After VSR1, to further reduce the magnetic coupling through the actuators, the magnets were replaced with new and less intense ones. It was also determined that the mirror recoil mass, made of aluminum, had an amplification effect on the magnetic field gradient. For VSR3, new recoil masses, made from a dielectric material, were installed.

During VSR3, another magnetic noise path was discovered in the fans used to cool down the electronics which compute the mirror position correction. The magnetic field radiated by the fan motors induced a noisy current in the correction signals sent to the actuators which was converted into mirror displacements. This was solved by increasing the distance between the fans and the electronics without compromising the cooling efficiency. The lower-left plot of figure 4 illustrates this problem by showing a frequency line around 44.3 Hz which results from the coupling between the fans and the DF. The line disappeared after 70 days, due to the noise mitigation actions.

As discussed in section 4.2.13, the high-frequency EM fields are likely to couple with the modulated DF signal. During VSR3, an example of a coupling mechanism was identified: environmental sensor ADCs were using a 300 kHz bit-rate serial communication protocol, and the 20th harmonic of this frequency (∼6 MHz) lies a few kHz from Virgo main modulation frequency. A radio-frequency antenna showed a fluctuating line which was seen in time coincidence with the DF signal (see figure 5). An unexpected solution consisted of increasing the temperature of the room hosting the serial link server by 2°. This slightly changed the clock oscillator rate and was sufficient to shift the 20th harmonic spectral line out of the detector's bandwidth.

Zoom In Zoom Out Reset image size

Another family of spectral lines is composed of very narrow (W < 1 mHz) and stationary lines at multiples of a few fundamental frequencies associated with digital systems in the detector. For instance, several ADC boards used during VSR1 contained a 10 Hz internal clock, which produced a comb of lines spaced by exactly 10 Hz in the frequency domain. These lines were very intense and covered the whole frequency range of interest for CW searches (between 10 Hz and 2 kHz). A test consisting of switching off the ADCs during data taking confirmed that they were the source of the disturbance, although the noise coupling with the DF was not clearly understood. After the end of VSR1 these ADCs were replaced and almost all the 10 Hz noise lines disappeared in the subsequent runs. Harmonics of 1 Hz were also observed during VSR1. These lines were concentrated at frequencies below 100 Hz and tests indicated that the noise source was probably the same as for the 10 Hz harmonics since this frequency comb also disappeared nearly completely after the ADC replacement. The bottom-right plot of figure 4 shows some remaining harmonics which were still present in VSR2 data despite of the ADC fix.

Another well-known comb of 10.278 Hz harmonics with a digital origin is present in all Virgo runs. The source of the lines has been recently identified in digital modules used to control the mirror coil drivers. There is a strong indication that the coupling mechanism is of EM nature.

The strongest lines in the Virgo spectrum are often surrounded by a dense forest of sidebands. This effect was identified as a result of a coupling with the SA and suspension mechanical modes. For example, figure 6 shows the lines observed on both sides of the 444 Hz injected calibration line and table 3 lists the sideband frequencies associated to each identified mode. The exact mechanism that produces the sidebands is not known, but is likely due to some nonlinearity of the interferometer.

Zoom In Zoom Out Reset image size

Using the category definition and figures of merit described in section 4.3.2, we estimate for each run the performance of the DQ flags used by GW burst and CBC searches, using respectively the triggers delivered by single detector Omega and MBTA online analyses. The list of DQ flags and their category assignment is then prepared for burst [15], CBC low-mass [13] and CBC high-mass searches [69]. The prescription can be different for each analysis. For instance, a DQ flag can be prescribed as CAT2 for burst and high-mass searches while it is used as CAT3 for the low-mass search.

Figure 7 illustrates the DQ flag performance for Omega triggers. These results show significant differences between the Virgo science runs. The first run, VSR1, is characterized by a low number of DQ flags (about 20) but they are able to reject a large fraction of loud events. In fact, most of the rejection is obtained by only one flag and the corresponding noise excess occurred in a single night of VSR1 when the laser power stabilization failed due to a blown fuse. The DQ flag was categorized as CAT2 even though the science time should have been re-defined by removing this noisy period from the start (CAT1). If one takes into account this correction, the sample of triggers to be considered is shown by the dashed white histogram in figure 7. With this consideration, the trigger rejection is limited and mostly effective for high-SNR events. Moreover, the initial trigger rate of VSR1 is very high (2.1 Hz for Omega) and, after applying the DQ flags, remains quite large (10 times larger than in VSR2 for SNR  >  10).

Zoom In Zoom Out Reset image size

VSR2 started with a greatly improved knowledge of the detector and of its response to noise. Furthermore the detector glitch rate decreased by a factor of 4 with respect to VSR1; this facilitated the noise investigations. This resulted in a significant increase of the number of DQ flags (more than a hundred), explaining the larger dead-time. This also translates into a larger glitch rejection efficiency: ≃ 70% for SNR  >  8 while it was only ≃ 10% for VSR1. VSR2 is also characterized by a rejection efficiency which covers a wider range of SNR. Low-SNR events are removed with a non-negligible efficiency which is important for multiple-detector analyses since it is likely that some of the numerous low-SNR events will combine to produce the most significant coincidences.

VSR3 data quality was not as good as VSR2 mostly because of the contrast loss issue explained in section 2. Consequently, the detector had to be set on a new working point which increased scattered-light effects. This created a significant number of glitches seen by Omega at frequencies above 500 Hz for which it was not possible to design a DQ flag, explaining the degraded overall performance of VSR3 DQ flags. As explained in section 4.3.2, if the ratio between the efficiency and the dead-time is larger than 1, then DQ flags target glitches with good accuracy. These numbers can be derived from the second row of plots in figure 7. For example, if we consider triggers with SNR  >  10, CAT2 and 3 flags give /D = 6.9, 6.8 and 3.4 for VSR1, 2 and 3 respectively.

Glitch families are often characterized by a given frequency which can be measured by Omega. It is therefore possible to sort glitch families and to study the ability of a DQ flag to eliminate them. The frequency plots shown in figure 7 (lower-row) present good flagging efficiencies in specific frequency bins. For example, the piezo glitches of VSR1 detailed in section 4.2.10 are visible at a frequency of ∼ 140 Hz and are efficiently rejected. In the VSR2 plot, the efficiency histogram for triggers with SNR  >  5 exhibits higher efficiency values for frequencies of 210, 420, 495 and 840 Hz which correspond to alignment glitches described in section 4.2.6. Finally, the efficiency peaks visible on the VSR3 plot between 500 and 1100 Hz are mainly explained by the good performance of the laser power stabilization flag defined in section 4.2.4.

For each trigger, the search pipeline computes a SNR statistic after applying a coincidence (CBC) or coherence (burst) test to determine if the trigger is present in more than one detector. To measure the background rate of events in the search due to noise, data from the detectors in the network is time-shifted (by an amount greater than the gravitational-wave travel time difference between observatories) and then re-analyzed. Many different shifts are performed to obtain an accurate measure of the background rate in the search. The significance of a candidate GW trigger is characterized by its false alarm rate (FAR), which is computed by comparing the SNR of the candidate trigger to the background. An excess of noise events in a detector can cause the distribution of the background to have a significant non-Gaussian tail at high SNR, thus reducing the significance of GW events. It is therefore very important to remove loud background events by mitigating them in the detector, or excluding them in the analysis with vetoes. Reducing this non-Gaussian tail in the background is the primary goal of detector characterization, as it increases the astrophysical sensitivity of the search.

Using several detectors in coincidence presents many advantages. The most important one is to reduce the number of background triggers and hence decrease the FAR of GW signals. Initially, it was believed that the coincidence between detectors would be sufficient to reduce the detector noise to its Gaussian component. In fact, it has been realized that searches are limited by accidental coincidences of transient glitches. Thus, noise investigations and DQ flags are very important to improve the sensitivity of the searches. Since VSR1, Virgo data has been used in coincidence with the three LIGO detectors offering multiple coincidence schemes, from 2 to 4 detectors. As an alternative to a basic coincidence between detectors, LIGO and Virgo data can also be combined coherently [70], taking into account the individual detector's antenna patterns. This approach provides an optimal detection efficiency since the network is not limited by the least sensitive detector (at least when combining more than two detectors).

For a network analysis, the performance of DQ flags can differ from what has been obtained with single detector triggers, as presented in section 6.1.1. To study the effect of Virgo DQ flags on multi-detector searches, we chose to consider the coincident CBC low-mass analysis [13] and the coherent all-sky burst search [15]. Only a subset of the data used in the published analyses has been considered to quantify the DQ flags impact. Moreover, only background triggers will be presented in the following. Finally, LIGO DQ flags are never applied in the following studies (except CAT1).

The CBC low-mass analysis makes use of a χ² discriminatory test [71] to efficiently reject glitches whose waveform does not match the expected CBC signal. After having selected single detector triggers with a SNR larger than 5.5, a preliminary cut is applied in order to reject events strongly disfavored by the χ² test. For triggers with SNR below 12, an additional cut is performed based on the behavior of the χ² time series near the trigger time [72]. For the remaining triggers, a reweighted SNR [13] is calculated by down-weighting the SNR progressively with the reduced-χ² when reduced-χ² > 1. Reweighted SNRs obtained for each detector are summed in quadrature to form the ranking statistic used in the CBC search. To evaluate the Virgo contribution to the CBC statistic and the impact of the DQ flags, both Virgo SNR and Virgo reweighted SNR variables can be considered.

The upper plots on figure 8 show, for VSR2 data, how the Virgo DQ flags perform on low-mass CBC triggers which are coincident in two detectors (Virgo and one of the LIGO detectors). The combination of CAT2 and 3 flags is able to remove background triggers with an efficiency of 22.6%. The efficiency increases rapidly with the SNR measured in Virgo: = 92.9% for SNR  >  10 which proves the ability of Virgo DQ flags to remove the loudest CBC triggers. One can note that the loudest triggers are removed by CAT3 flags. These events are found to result from a strong laser disturbance for which a DQ flag was designed. The performance of this flag is too limited to be categorized as CAT2. Many events flagged by a Virgo DQ flag are already disfavored by a high χ² value and thus ranked with a low value of reweighted SNR. Nevertheless, the upper-right plot of figure 8 shows that Virgo DQ flags have a non-negligible impact on noise events with large reweighted SNR. For example, when considering CBC triggers with a reweighted SNR above 8, approximately 60% of triggers are removed by Virgo DQ flags. In general, the SNR of the loudest background event allows us to measure the sensitivity of a detector, since a GW candidate must be louder than this to be considered significant. The use of Virgo DQ flags reduced the reweighted SNR of the loudest event from 9.5 to 8.7, leading to an astrophysical volume that is 1.3 times larger than the search without DQ flags. Although here we only consider the sensitivity of the Virgo detector, and the astrophysical sensitivity depends on all the detectors in the network, this increase in reach is a clear indication of the power of data quality and vetoes.

Zoom In Zoom Out Reset image size

The initial distribution of CBC triggers visible on the upper-left plot of figure 8 presents some structures which are understood. First, the SNR distribution shows a steep break at SNR = 12. This effect results from the analysis feature which consists of applying the χ² cuts with a discontinuity at SNR = 12. Three populations of glitches then dominate the SNR distribution. The loudest events (SNR  >  150) correspond to laser disturbances described in section 4.2.4. The large bump with SNR  >  40 results from an excess of TCS glitches (see section 4.2.7) which are removed by specific DQ flags. Events below SNR = 12 (which can also be seen as a bump with Omega triggers in the VSR2 plot of figure 7) are mostly produced by scattered-light mechanisms described in section 4.2.1 and 4.2.6. The DQ flags based on the ground motion velocity and alignment signals are able to remove this population. Moreover, this population of glitches is also characterized by χ² ∼ 1 (i.e. large reweighted SNR), so these DQ flags are probably the most important flags to improve the sensitivity of the CBC search.

The same study has been performed on triple-coincident events (Virgo and two of the LIGO detectors), as can be seen on the lower row of figure 8. The overall performance remains about the same on triple-coincident CBC events: 30.8% of efficiency. Requiring a triple coincidence offers an even more stringent way to suppress the background than double coincidence, but the reduction concerns all categories of glitches and thus it does not affect the DQ flags rejection efficiency. The lower-right plot of figure 8 shows that the loudest triple background event of the CBC search is not removed by a Virgo DQ flag.

The generic GW burst searches are designed to look for a large variety of transient signals, spanning the full frequency bandwidth of the detectors and without a precise model of waveforms. They are therefore sensitive to a larger number of glitch types than CBC searches and cannot make use of consistency tests such as the χ² test. The all-sky search [14, 15] has been performed by several analysis algorithms. Here we use the latest results obtained with the coherent wave-burst (cWB) pipeline [73] which combines coherently the detectors' strain amplitudes. In the cWB search, the network parameters can be derived from a likelihood method based on a network SNR estimator [74] and can be used to characterize and reject noise transients. Finally, events are ranked as a function of the correlated amplitude ρ, which measures the degree of correlation between the detectors for an event. Virgo DQ flags have a significant impact on cWB triggers and greatly improve the search sensitivity. To study this impact, the cWB pipeline was run over two months of VSR2 data (November and December 2009). The winter season of VSR2 was chosen because these data were the most affected by noise.

In the upper row of figure 9, cWB events obtained with a two detector network (Virgo and one of the LIGO detector) are shown. The left plot shows the impact of DQ flags on the distribution of Virgo SNR which measures the Virgo contribution in the coherent data stream. Firstly, a collection of selection cuts based on the likelihood parameters are implemented in the cWB algorithm which excludes detector glitches incompatible with signals expected from the detector network. This allows for the suppression of the loudest (and most obvious) Virgo glitches. Even with this analysis feature, Virgo DQ flags still efficiently reject part of the remaining triggers. The overall veto efficiency of CAT2 and 3 flags is 60.4%. For SNR  >  10, 89.5% of cWB triggers are rejected by Virgo data quality flags. The DQ flag rejection efficiency can be derived from figure 9 for any Virgo SNR or ρ threshold when neglecting the DQ flags dead-time (∼10%). Unlike the CBC analysis, where only a few DQ flags were performing the majority of the rejection, all Virgo DQ flags contribute to the background suppression in the cWB search.

Zoom In Zoom Out Reset image size

As can be seen in the upper-right plot of figure 9, the Virgo DQ flags are less efficient to remove events ranked with a high ρ mostly because, for these events, the LIGO data strains preponderantly contribute to the coherent stream. Nonetheless, the number of high-ρ events is reduced by Virgo DQ flags. For example, if one fixes the FAR to 1 event per 6 years (rate ≃ 5 × 10⁻⁹ Hz), the cWB network selection cuts allow to lower the ρ threshold by 10% and Virgo DQ flags offer an additional 20% of reduction. This represents a gain of sensitive volume of about a factor 2. Such an improvement should be compared with the ideal case corresponding to Gaussian detector noise (also shown on figure 9). Data quality work is increasingly challenging upon approaching this limit. Understanding the glitch production and coupling mechanisms is much more difficult at lower SNRs.

The same study has been performed on cWB triggers produced with a three detectors coherent data stream and results are presented in the lower row of figure 9. As expected, in this configuration, the search is more sensitive since, for a comparable FAR, the ρ threshold can be reduced with respect to the two detector case. For example, with a FAR of 1 event per 6 years, adding a third detector in the network allows for a 30% reduction of the ρ threshold (i.e. the sensitive volume gets twice larger). This threshold can be further lowered by about 10% by the use of Virgo DQ flags (i.e. the sensitive volume gets 30% larger).

During VSR3, the online data quality monitoring took on a new and important dimension. Transient GW searches using LIGO-Virgo data were performed online and alerts were sent to telescopes in order to observe a possible EM counterpart which would increase the detection confidence of a GW event [75]. Therefore, the data quality information had to be provided with a very low latency in order to exclude obviously false GW candidates (noise glitches) which would have otherwise been sent to telescopes.

For VSR2 an online architecture, based on tools used for the data acquisition system, was set up to provide DQ flags with a latency of about 30 s. These flags were stored in the LIGO and Virgo databases. In parallel, the DQ flags were monitored which allowed scientists in the control room during VSR3 to rapidly check data quality to make the decision whether or not to send an alert for prompt EM follow-up.

The main requirements for online DQ flags are: the reliability of the online production system, the possibility of using the processing algorithm both online and offline, and the ability to provide a complete data quality information while at low latencies. During VSR2 and VSR3, the online DQ production did not encounter major problems and had a duty cycle similar to the Virgo data acquisition system (above 99.8%). The algorithms producing the DQ flags used generic I/O libraries and thus have also reprocessed missing segments. Finally, the most difficult part of the DQ flag production concerns the confidence of the data quality information provided with low latency. A software architecture has been created to provide online DQ monitoring. This allowed for the selection of the most reliable flags in order to veto events before sending alerts to telescopes. To improve this architecture and to provide accurate online DQ flags will be one of the main challenges for advanced Virgo [47].

One strong constraint on the online DQ flags is the daily variation of the glitch rate and glitch types, depending on e.g. the detector working point or the weather conditions. Online DQ flags performance can vary significantly if they are not tuned on the fly. Automatization of such tuning will be an important step to provide the required reliability of DQ flags for the advanced Virgo online analyses.

The study of data quality is a challenging task and many families of glitches have origins which have not been identified. For instance, many Omega scans performed on the VSR2 data show a recurrent glitch around 60 Hz that always has the same morphology in the time–frequency plane. It is very likely that these glitches have a common source of noise. However, no explanation for these glitches has been found.

The lack of understanding of a noise source and of the coupling to the DF is, in most cases, due to the fact that no auxiliary channel is correlated with the DF glitches. There are three possible scenarios which can result in unknown glitch families:

• (i)  
  The detector or the environment is not fully monitored: the noise source and the coupling mechanism cannot be detected by any of the current sensors. This explains why no auxiliary channel has been found to be sensitive to this noise.
• (ii)  
  The sensitive channel is actually operational, but it is also sensitive to many other kinds of noise which do not affect the GW data. In that case, the effective signal component is swamped by uninteresting noise and it is highly unlikely that this channel will be identified as useful for glitch flagging.
• (iii)  
  The current flagging procedure mostly relies on a glitch-to-glitch method. Only a few examples of DQ flags are defined by more advanced approaches (for example the scattered-light glitches) resulting from a complete understanding of the noise path. In the future, it may be necessary to explore more nonlinear coupling hypotheses (see section 7 for further discussions).

As can be seen on figure 10, most of the remaining glitches do not seem to be associated with a given permanent noise source that could have been associated with a specific frequency band. After applying the DQ flags, the low frequency region remains the most contaminated: triggers with a frequency below 300 Hz represent 89% of the remaining triggers. The '60 Hz glitches' mentioned above represent about 12% of the remaining low frequency glitches. Figure 10 also displays a sudden drop of the trigger rate at mid-run. On 5 October 2009, a short commissioning break occurred during which several actions were performed (dust cover installation, laser and TCS maintenance) and the exact reason for the glitch rate reduction has never been well-established..

Zoom In Zoom Out Reset image size

Given the sensitivity of the first generation of interferometers, only a few known pulsars are astrophysically relevant for close examination [17, 19]. For these pulsars, even in the case of a null detection, it is possible to approach and possibly beat the so-called spin-down limit. To achieve this goal, it is important to make sure that no noise spectral line crosses the frequency band of these targeted pulsars. This task was performed by the NoEMi software described in section 5.1. In table 4, known pulsars monitored in the last Virgo science runs are listed.

During VSR2 a non-stationary noise line affected the sensitivity of the Virgo detector at the frequency of the Vela pulsar (22.38 Hz) as shown on the left plot of figure 11. The disturbance caused a loss of sensitivity of about 20% [19]. Running the NoEMi coincidence analysis on the auxiliary channels led to evidence that the disturbance was correlated with a line (actually a doublet of lines), clearly visible in the data of an accelerometer monitoring the vibrations of the TCS optical benches (see right plot on figure 11). Although a satisfactory description of the noise coupling mechanism was not achieved, the source of the disturbance was identified as being two chillers (pumps that circulate a cooling fluid for the TCS laser) located near the TCS room. The rotation frequency of the chiller engine was indeed 22.4 Hz. The vibration was probably transmitted to the TCS bench through the cooling pipes. During VSR3 the noise line was no longer visible in the DF, although it was still present in the accelerometer. It is assumed therefore that the line was hidden under the detector noise, which at the Vela frequency was 2 to 3 times worse with respect to VSR2. A small but indicative coherence was indeed found between the DF and the accelerometer data. To remove the disturbance away from the Vela band a variable frequency drive was installed during VSR3 to change the rotation frequency of the chiller engines, as can be seen in the right plot of figure 11.

Zoom In Zoom Out Reset image size

All-sky searches produce a list of CW 'candidates', characterized by a position in the sky, a signal frequency and one or more frequency derivatives (spin-down). A follow-up of those candidates is performed in the next step of the analysis [76–78]. More precisely, candidate events are selected by thresholding on a quantity characterizing the candidate significance (using Hough maps built in the source parameters space). If a noise line is present in the data, it shows up as a collection of fake candidates. Even a very narrow and constant frequency line produces multiple candidates in a frequency band around it and for various spin-down values. This effect is even larger in the case of broader lines or a forest of narrow lines, like the sidebands described in section 5.2.4. A sufficiently high threshold on the line significance helps to maintain a reasonable number of candidates but reduces the sensitivity of the search. For example, figure 12 shows the number of CW candidates selected during VSR1 in the 410–422 Hz and 438–450 Hz frequency bands. Two excesses of candidates are clearly visible. The first one, around 444 Hz, is associated with a calibration line and its sidebands, discussed in section 5.2.4. The second excess, around 416 Hz, corresponds to the 10th harmonic of a 41.618 Hz noise line and its sidebands. There are strong indications that this noise line is due to vibrations of the external injection optical bench producing some beam jitter [79].

Zoom In Zoom Out Reset image size

To reduce the number of signal candidates, it is crucial to produce lists of frequency intervals affected by noise disturbances described in section 5.2. One should add to this list all the lines associated with the intrinsic resonances of the interferometer as well as the injected lines used for calibration and control. To achieve this task, all the lines detected by NoEMi are reviewed and identified one by one. This work is still in progress. Table 5 presents the current status of the lines identification in VSR3 data. 962 lines have been identified and about 400 lines still remain to be classified. Once this work is finished, the frequency bin corresponding to each identified line will be discarded before running the all-sky CW analysis.