Recently, some of LIGO’s detchar experts investigated the connection between Tomte glitches from Gravity Spy and the bias voltage on a particular circuit. You can read the full alog of this investigation here. Below is a summary written by TJ Massinger, one of the detchar experts who does a lot of work with Gravity Spy data!
During recent commissioning work at the LIGO Livingston Observatory, it was noticed that glitches were occurring while the bias voltage on a circuit (an electrostatic actuator) was adjusted. Inspection of these glitches using the same time-frequency visualization that Gravity Spy uses showed that they looked qualitatively similar to glitches classified by users as the “tomte” class.
Using the GravitySpy classifier, it was found that these recent glitches are also classified as tomte glitches. To quantify the similarity of the bias voltage glitches to the tomte glitches seen in the second observing run (O2), GravitySpy was used to gather a collection of previously identified tomte glitches in O2 with parameters similar to those seen when adjusting the bias voltage. Upon comparing these populations, their time- and frequency-domain morphology was found to be nearly identical, suggesting that the population of tomte glitches in O2 might be understood by continuing to investigate the glitches that occur when adjusting bias voltages.
We are excited to bring you new data from our Engineering Run 13! Data taken during Engineering Runs are meant to test not only some of our new upgrades to the detectors but some of our software (like Gravity Spy).
Dates: 10 am Central Time Dec 14 to 8 am Central Time Dec 18 (N.B. Due to some issues at the sites science ready data was not available until Saturday December 15)
What Detectors Are Running
Originally, we anticipated only having Hanford and Virgo due to critical repairs at Livingston. These repairs, however, completed yesterday and after a short delay Livingston has joined ER13. At first, we will be streaming in the data live for Hanford and Livingston over the weekend, and at a later date will add Virgo ER13 data to the Virgo only workflow.
The sensitivity of the LHO detector has increased its range to detect binary neutron stars from 80Mpc to 90Mpc, LLO has increased to 100Mpc and Virgo has nearly doubled its range from 25 to 43Mpc. A number of different glitch classes have arisen and the engineering run is a golden opportunity to identify and eliminate these so we can be rid of them for the year long O3 run which is anticipated to start in March 2019.
Some of the changes at both LHO and LLO that have led to this improvement include squeezed light and a new 70W laser amplifier that will improve LIGO’s quantum noise limit. In addition, Acoustic Mode Dampers will damp internal modes of the test masses to reduce parametric instability (light interacting with mirrors as positive feedback). Also, there was a change of several test masses to improve their coatings (especially for green light) and to remove a point absorber at LHO.
We look forward to your collections of interesting new glitches and for determining the cause of the new excess noise sources!
The Gravity Spy Team.
Hey Gravity Spiers,
We are really excited to finally introduce a new detector, workflow structure, and tool this week. First, we present a new workflow containing glitches from the Virgo detector in Pisa, Italy. Second, we are changing the level structure to speed up our user training. Finally, we are bringing you an auxiliary web tool to help the search for unique and novel glitches.
Virgo differs from Hanford and Livingston in a few ways including the length of the arms, the apparatus holding up the test masses, and the suspensions. We anticipate there will be a number of interesting new glitches in Virgo. For some glitches, such as Scattered Light, they will appear different but have the same cause. For other glitches, such as the Violin Mode Harmonics, they will be the same source but at different frequencies due to the different suspension system. Below we demonstrate a few novel glitches you may find along the way while classifying Virgo glitches, including a new class we have called Fireball (bottom right).
New Workflow Structure
For those of you familiar with Gravity Spy’s training method, we intend to utilize pre-labelled images to help train new users in the classification task. In addition, in order to facilitate training, we introduce a different number of new families of glitches in different levels, culminating in Level 4 where all 22 classes are introduced. However, after some feedback, as well as looking at the data, we learned that getting from level 4 to level 5 was taking longer than anticipated. We decided that this was due to too many new classes being introduced between level 3 and 4. Specifically, the amount of pre-labelled images users were seeing was spread out across too many new classes causing the number of classifications a user must complete before seeing pre-labelled data to sky rocket. Therefore, we are adding another intermediate level that has 15 classes. We believe this will cause users to see pre-labelled images of the new classes faster and, in turn, move through the levels faster. In total, with the addition of Virgo, there are now 7 levels in Gravity Spy (see image above).
This restructuring of the workflows may cause some users to start on levels lower then they may expect. This can be due to a number of factors, and we encourage all users to simply charge ahead with classifying on whichever level they find themselves on. You should experience a fairly rapid promotion through the levels.
We want to thank all of our volunteers for their continued efforts on Gravity Spy and we appreciate all the feedback we have received. We look forward to seeing what novel Virgo glitches you are able to find! As always please reach out to me with all leveling issues. We hope this restructuring proves an effective method to boost training.
Gravity Spy Tools
With the introduction of the new Virgo workflow, we anticipate there being a number of novel glitches, some that will look like what you may have seen in Hanford and Livingston, and some very different. In an effort to help facilitate the generation of large collections of novel glitches, especially when we are not sure what to expect with Virgo, we are introducing a new supplementary tool for Gravity Spy, gravityspytools. For an idea of how to use this tool please watch the linked video. The goal of this tool is to maximize the impact of a new machine learning algorithm that the Gravity Spy team has developed called DIRECT. This algorithm utilizes transfer learning in order to learn what makes gravity spy images similar and dissimilar from each other. This allows every Gravity Spy image to be abstracted into a feature space containing 200 points. It is in this feature space that we calculate distances from one images to another. An interface to do this is provided on gravityspytools called the “Similarity Search.” It takes as input one sample from Gravity Spy and as output the closest samples in the feature space based on distance. An attempt to visualize in three dimensions what the set of known images (such as blip, whistle, etc) looks like in this 200 dimensional feature space is shown above.
We want to thank all of our volunteers for their continued efforts on Gravity Spy and we appreciate all the feedback we have received. Please let us know how you find using the gravityspytools! As always please reach out to me with features you would like to see!
Hello Gravity Spiers!
At long last and after much demand, we have added audio examples of what the Gravity Spy glitches sound like to our field guide! As many of you may know, the frequencies of gravitational waves (and frequencies of glitches) detectable by LIGO are similar to the frequencies of sound (i.e. the pitches) that humans can hear. Therefore, LIGO scientists oftentimes convert our signals to sound!* The MP3s are embedded directly into the field guide, so you should be able to play them straight from there.
Some glitch categories may be hard to distinguish above the background noise, whereas others you should hear quite distinctly. Either way, having a good set of headphones will help hear the subtle features of the glitches better.
Big thanks to our LIGO collaborator Derek Davis for putting together these glitch sounds! Head on over to the Gravity Spy field guide to take a listen to the sounds of glitches!
-Mike / the GSpy team
*There are a few changes done to the data that make the sounds easier to hear. Other than the standard whitening and band-passing, for the glitches in our field guide we also linearly shift the frequency up by 60Hz to make the sounds (especially the low frequency glitches) more in our audible range. Also, a filter that can be described as an “inverse A-weighting” filter is utilized. The basic idea of this filter is to account for the fact that our ears are less sensitive to particularly low and high frequencies. Since the drop off starts around 200 Hz, this affects a decent number of glitches. By increasing the loudness of these lower frequencies, we make it so that features of similar intensity in an omega scan are ideally heard equally loud, no matter their frequency.
Thanks to the hard work by GravitySpy Citizen Scientists, we now have more than 53,000 retired images, images that have had enough consistent classifications by citizens that we are quite sure they are correctly categorized. New images are retired every day, so this data set is always growing. Using this set of retired images, along with information from machine learning image analysis, has allowed us to get a clearer picture of which glitches appear often and rarely in LIGO Hanford (H1) and LIGO Livingston (L1).
Since image retirement relies on classifications by citizens, the images that get retired the fastest and most often are those that are the clearest, the ones that contain only one sort of glitch, and look like the example images. Thus using only retired images may not be a good measure of the total number of glitches. Because of this caveat, we also looked at glitches that the machine learning algorithm identified as belonging to a given category with a confidence of 90% or above.
Looking at the summary information for categorization done by these two methods for LIGO’s two Observing Runs, O1 and O2, we found that there are a handful of glitch categories that are never or almost never found in H1 and L1.
Above is a summary of all the retired glitch categories at Hanford in O1 that had at least one glitch in them. As one can see, this is not every glitch category that we have; 1080 Line, 1400 Ripple, Chirp, Helix, None of the Above, Tomte, Violin Mode Harmonic, and Whistle don’t show up at all! However, None of the Above likely does not show up because it is such a diverse category that it is challenging to have enough consistency to retire such glitches or reach high machine learning confidence. Also, chirps are found infrequently, but when they are, they are very interesting! Paired Doves and Wandering Line show up only twice each out of the 4,448 retired glitches. Here are the pair of Paired Doves:
For Hanford, we have determined that Paired Doves and Helix happen infrequently. Similarly for Livingston, 1080 Line, 1400 Ripple, Air Compressor, Helix, Light Modulation, Paired Doves, Tomte, and Wandering Line are very rarely used.
For the interested reader, below are links to screenshots of the full summary of retired and machine learning categorizations.
ML Confidence 0.9
Again, thank you all for your continued hard work in classifying these glitches. Without you all, this wouldn’t have been possible!
The GravitySpy team
There is some debate on whether Koi fish and blip glitches are part of the same morphology distinguished mainly by loudness or amplitude of the signal. Andy Lundgren has shown that removing calibration lines as part of “data cleaning” makes them look much more similar. A possible explanation of this is that for loud enough glitches, the calibration lines may be causing the Q-transform’s whitening filter to ring at those frequencies, creating the fins on the koi fish. Whitening is a process that removes stationary differences in the loudness of Individual frequencies in the signal.
Fig 1 – Q-transform of typical Koi fish glitch
Fig 2 – Q-transform plot of the same data as above, but with calibration lines removed prior to calculating the Q-transform.
Beverly Berger has suggested that the amplitude of the glitches may be an effective way of discriminating between Koi fish and blips, with Koi fish being louder. To dig into that a little further we plotted the frequency and amplitude of the two glitch types during the 02 run.
Fig 3 – Frequency distribution of blip glitches during 02
Fig 4 – Frequency distribution of Koi fish glitches during 02
The plots of frequency distributions show significant overlap between the two glitch types. Koi fish have a lower peak frequency due to the shape but not enough of a difference to help in classification.
Fig 5 – Amplitude distribution of blip glitches during 02
Fig 6 – Amplitude distribution of Koi fish blitzes during 02
When we look at the amplitude differences between Koi fish and blips we see a pretty sharp dividing line around 10-21, especially during the first 21 weeks. The gap starting at week 23 is a time when there was significant commissioning. There was also a significant increase in range at Livingston around this time.
It is still unclear whether Koi fish are the same as blip glitches, only louder. We also have not been able to identify what exactly causes either glitch. This is just an interesting observation that we thought we would share with our colleagues at Gravity Spy.