Acoustic
Detection and Triangulation of Thunders around the LIGO Detector
Lonique Coots
SURF student from The University of Texas at Austin
(Mentor: Szabolcs Marka)
Objective
Several environmental noise
sources can have significant effect on the extremely sensitive LIGO detector.
To veto signals due to environmental origin, we have to monitor, identify
and record noteworthy local events, such as earthquakes, lightning, thunder,
rain, planes, etc. In this project we will use sensitive outside
microphones to detect the acoustic signatures (thunder) of local lightning
events, estimate their relative location and determine their strength at
the detector.
Abstract, major goals and milestones
We recorded and identified thunders
in the vicinity (R~O(20Km)) of the LIGO detector. In order to do
this we set up low frequency, pre amplified microphones at the end stations
and the LVEA. We found that using a simple threshold on the band
limited RMS data, we can construct an on-line trigger for each individual
microphone signal. In case of double or triple coincidence within
the allowed time window (~20s), we will cross correlate the time series
pairs and determine the exact time delay between the sites. Based on this
information we will provide real time data of individual triggers, signal
strength at the buildings. If the strength and thunder rate allows, we
will locate the origin of the thunder and display it on a real time map
and enter the trigger into the LIGO LDAS event database.
Note 1: It is clear that these goals are ambitious
and the fully automatic implementation may (and probably will) require
more time than allowed by the SURF interval. To be realistic, we set the
minimal goal of demonstrating the viability of the method (in practice),
while developing a working prototypes to cover all major milestones and
hardware issues. The actual assembly of the subsystems and the development
of the C++ code necessary for the fully automatic system will be implemented
later.
Note 2: This project and the LHO SURF project lead by
Dr. Daniel Sigg targeting the detection of lightning strikes based on their
RF signature is complementary and coordinated. The coincident detection
of lightning in the RF and acoustic channels will allow us to draw a more
exhaustive picture, including direction and distance information.
Outside consultants:
We established first contact with the Trilon Technology,
the developers of ShotSpotter(TM) (www.shotspotter.com)
system, used by the L.A. Police to locate illegal gunshots. Their system
is based on an extensive array of microphones mounted on top of buildings
at urban areas. Since their needs are fairly similar to ours and Rob Calhoun
their chief programmer were very open, we can expect significant help from
them.
Hardware
After initial tests with recorded and real thunder
it was obvious that we will not able to get good results with the existing
inside microphones. Even if we count on significantly decreased noise inside
during real operations, the higher frequency (sensitive band of microphone)
acoustic isolation of the building makes the analysis much harder and is
likely to decrease our range/sensitivity significantly. It is also a good
idea to monitor lower frequency environmental disturbances in the acoustic
band since they can have significant effect on the detector. Therefore,
we decided to set up an outside microphone at the top of each building
which will become the part of the PEM system. To hit two birds with one
stone, we will try to install microphones with low frequency sensitivity
(microbarographs), which complement our present system and more advantageous
for us detecting thunders. We are currently designing the microphone housing
and the amplifier. This will allow us to suppress the effects from unwanted
sources (e.g. rain droplets hitting the enclosure) and to galvanically
isolate the microphone from the rest of the DAQ (extra security measure).
Weekly log of major events:
Week 1
We collected thunder recordings from the net and burnt
a CD. We tested the inside PEM microphones as well as our temporary outside
microphones at the end stations by playing back recorded thunders from
a CD. The acoustic isolation of the building is very efficient in removing
the higher frequencies (as they should). Due to this attenuation and the
high noise within the building during the measurement, it is not feasible
to use the internal microphones. We examined the recorded signals and worked
on finding an optimal triggering algorithm.
Week 2
We finally had our first thunderstorm (Tuesday (06/18/01))
and we also had an outside microphone in working order by then. Our conclusion
is supported by the recordings made during thunderstorm; the outside microphones
recorded the thunders very nicely, while the internal microphones recorded
practically nothing. This doesn't mean that the low frequency components
of the thunder do not penetrate the wall and can have significant effects.
We decided that we need to install outside microphones for the time being.
As a plus we were able to work with real data and had a chance to evaluate
whether our previous ideas about the trigger algorithms (based on recordings
with limited frequency content) are useful and/or optimal. We started to
work on the preamplifier design and microphone selection process.
Week 3
We tested the frequency response of our microphones and
concluded that in this price range flat response is non-existent and no
microphones are alike. It is very likely that higher quality low frequency
microphones would serve us better and they would enable us to produce more
precise measurements. One of the problems that lie ahead is matching
thunders between microphones, which can be greatly complicated by mediocre
sensors.
We recorded the sound of a thunderstorm with two outside
microphones. The preliminary results are presented below. They are very
encouraging.
Week 4 - 5
This week we continued the work on our triggering
algorithms. We took time out to understand the basic principles of
C programming and Fourier Transforms. Looking for a strategy to approach
our data, we found that using a simple threshold on the band limited RMS
data we can construct an on-line trigger for each individual microphone
signal.
Week 6
This week we explored different ways
to prep our data in our program. We also looked for appropriate low
frequency, pre amplified microphones. The microphones we have been
working with were tested and found to be inaccurate in comparison to each
other. We found that just as low frequency components of seismic
waves spread through the earth from their points of origin, so can acoustic
pressure waves travel long distances through the atmosphere. These
far traveling disturbances occupy the frequency range below that of human
hearing and are termed infrasound. We decided that an infrasound
microphone would yield the most accurate results.
Week 7
Working in Matlab, we used the specgram
feature to examine our thunders frequency range (figure 4).
We found our frequency to be most dominate from 0 to 200 Hz. This
is significant to our microphone set-up (figure 1). We chose
a pressure sensor (low pressure differential, gage, vacuum gage/amplified)
(figure 3) with features including low pressure measurement, PCB
terminals on opposite side from the ports, and fully signal conditioned.
Now that we have our microphone, we tested it's accuracy at low frequencies.
We set it up with a low pass filter, a high pass filter and a resistor.
When analyzing the signal, we saw that the microphone would respond better
to waving your hand in front of it instead of loud high frequency claps.
We also found that people opening and closing doors 30 ft away had a significant
effect. Our signal would peak low frequency noise far away but not
to high frequency noise close by. We found it to be sensitive to up to
1 KHz and down to .05 Hz.
The program we are working on to locate the thunder
source is able to use a circular buffer. We are in the process of
setting up a trigger to use with our correlation function. We have
been able to produce functions that perform Fourier Transforms and Correlations
on our signal.
Week 8
The week we have be installing and testing our microphones.
We set up working outside low frequency microphones at each of the end
stations. Our low frequency microphone is powered by a 12V power
supply, and then passes our signal through a low pass filter (1KHz) and
a high pass filter (.03Hz) before it travels to LIGO DAQ. To reduce
the noise made by rain, we covered our microphones with a thick layer of
soft foam . At the X end station, we have installed a high frequency
microphone as well as the low frequency microphone to do a comparison between
the two (figure 5). Thunder was recorded by both microphones.
The low frequency was a good choice because it has a lower background noise
and more distinct peaks for the thunders. This will help when triggering
the microphones for correlation. The high frequency microphone was
more sensitive to other noise including cars and rain. The microphone
inside of the end station was not able to record any of the storm.
Another good aspect of the microphone being outside
is the ability to detect movement at the end stations. We were checking
our data for a thunder storm overnight and found that at the X end station
there was some noise. After researching what produced the noise,
we found that the guard went to the end station. This was a good
observation because during the engineering run being perform at the moment
there has to be a log of everyone that has been out to the end stations.
Since our microphones are sensitive enough to detect noise produced by
cars, it's easy to keep a precise log.
Figure 1. Above is a schematic of our test location for the microphones
and a suggested final location for the microphones. We are able to
judge our microphones signal distance accuracy by knowing that the distance
between the two end stations is 5.65 Km. We predicted that anything
over a 16 second time delay would be out of their range and therefore negligible.
Figure 2 Above is the microphone set-up we have out at
the X end station. On the left is the low frequency microphone with
a flow control system which consist of a valve and a T. The low frequency
microphone (~.1Hz - ~200Hz) picks up a signal from the thunder and then
sends it through a filter system (.03Hz - 1KHz) and amplifier before reaching
the LIGO DAQ. On the right is a high frequency microphone (~20Hz
- ~20KHz) with a pre-amp that sends a signal to LIGO DAQ.
Figure 3. The above is a detailed drawing of the differential
pressure sensor. It detects a signal by balancing the pressure inside
the fore volume with the pressure inside the backing volume. A thin
diaphragm in the middle allows it to measure very little differences in
pressure in the surrounding atmosphere and makes it sensitive to low frequencies.
To measure these small differences in pressure a solid state device.
The value of the leak is very large compared with the connector and has
primary control over the low pass cutoff of the microphone. The connector
will substantially control the high pass cutoff for the microphone
Both the connection to the outside and the leak between the two volumes
can be characterized by their acoustic resistance, which determines the
rate at which volume flow can occur under a given pressure differential.
Thus we have Ra =
where Ra represents acoustic resistance, dV/dt a rate of volume transfer
through a channel, and DP the pressure differential between the ends of
the channel. For a cylindrical channel, the acoustic resistance can be
estimated by the Hagen-Poiseuille relation:
where L is the length of the channel, h the molecular shear viscosity of
air, and a the radius of the channel. The value h is a function of air
temperature, but a reasonable average value can be taken as 1.8 x 10 -4
poise (poise = g cm -1 s -1 ). The acoustic resistance is often stated
in ohms (ohm = g cm -4 s -1 ). The applicability of the formula (2) is
dependent upon several conditions including laminar flow within the channel
and the absence of significant end effects on the channel; these conditions
should be satisfied here.
Figure 4 This specgram of our thunder
computes the windowed discrete-time Fourier transform using a sliding window.
The dominate frequency is between 0 to 200 Hz. The time is
measured in seconds.
Figure 5. The plot on top is the high frequency microphone
detection of thunderstorm and the one below is the low frequency detection
of the same thunderstorm. The noise level is greatly reduced by the
low frequency microphone and the thunder peaks are very distinct.
Observations
On 8/5/01 at 23:14:00 (UTC) we had a cluster of
thunderstorms in the vicinity of LIGO. LIGO DAQ recorded data strips
from two low frequency microphones, a high frequency microphone and a magnetometer.
The magnetometer had a peak signal every time lightning struck within the
area. This allows us to place certain peaks in the signal from the
microphones with real lightning thus making sure that the signal is thunder
and not other background noise. When comparing the low frequency
microphones, we noticed for some of the storm the Y end station microphone
did not record thunder that was recorded at the other end station.
This is from the fact that our microphones are set up on the ground behind
the end stations. The building itself is acting like a barrier for
the signal. If the microphones were located on the roof then they
would not be as direction sensitive.
Matlab specgrams of two low frequency microphones during 8/5/01 thunderstorm
(figure 6a). The data on the right is from the low frequency
microphone at the Y end station; on the left is from the low frequency
microphone located at the X end station. The Y end station
microphone had dominate frequencies ranging from 0 to 350 Hz, while the
x end station microphone had frequencies ranging from 0 to 260 Hz.
Cross Correlation gives obvious time delays for the first 300 seconds,
but afterwards produces multiple peaks because another storm enters the
area and allows more thunder peaks into the microphones. Since there
are peaks from different storms the cross correlation isn't as clear, as
shown in figure 7f . Figures 7a-7f are a series of graphs representing
a sequence of cross correlations during a 360s data set. They progress
minute by minute for 6 minutes. They allow you to see how the storm
progresses around LIGO. Peaks at larger scale values
represent negative lags. We cut the graphs off at 18.75s because
correlations peaking at higher values would be negligible.
The peaks at these high values will be out of the microphones range.
The main peak in the graphs a-e is approximately at max. time delay around
6s. When analysing the last graph, we see that the correlation produces
more than one peak. This is because there was one than one storm
at that time. The correlation is able to make a number of time delays
because of all the random noise happening at once. We believe that
the first peak is an accurate representation of the time delay because
of judgments made by eye. This would mean that there was a time delay
of 2 seconds for the storm that was becoming dominate at that time.
 |
 |
Figure 6a. Matlab specgrams of two low frequency microphones
during 8/5/01 thunderstorm. The data on the right is from the low
frequency microphone at the Y end station; on the left is from the low
frequency microphone located at the X end station. The Y end station
microphone had dominate frequencies ranging from 0 to 350 Hz, while the
x end station microphone had frequencies ranging from 0 to 260 Hz.
Cross Correlation gives obvious time delays for the first 300 seconds,
but afterwards produces multiple peaks because another storm enters the
area and allows more thunder peaks into the microphones. Since there
are peaks from different storms the cross correlation isn't as clear, as
shown in figure 6f .
Figure 6b. Above is a superimposed specgram of of our two low
frequency microphone signals. This allows us to have an estimated
correlation of our thunders. We estimated an approximate time delay
of 6 seconds. When we superimpose the last 60 seconds we find that
there are more than one way for the peaks to line up. When running
a correlation on the last 60 seconds, it is obvious that there is more
than one time delay. Thus leading us to the conclusion that there
are multiple peaks of thunder but not from the same storm. It seems
that another storm is heading in and becoming the dominate feature.
Figure 7a. Cross correlation of 1st minute with a peak
around 7.5s.
Figure 7b. Cross correlation of 2nd minute with a peak around
5.31s.
Figure 7c. Cross correlation of 3rd minute with a peak around
6.56s.
Figure 7d. Cross correlation of 4th minute with a peak
around 6.25s.
Figure 7e. Cross correlation of 5th minute with a peak
around 6.25s.
Figures 7a-7f. Above is a series of graphs representing
a sequence of cross correlations during a 360s data set. They progress
minute by minute for 6 minutes. They allow you to see how the storm
progresses around LIGO. Peaks at larger scale values
represent negative lags. We cut the graphs off at 18.75s because
correlations peaking at higher values would be negligible.
The peaks at these high values will be out of the microphones range.
The main peak in the graphs a-e is approximately at max. time delay around
6s. When analysing the last graph, we see that the correlation produces
more than one peak. This is because there was one than one storm
at that time. The correlation is able to make a number of time delays
because of all the random noise happening at once. We believe that
the first peak is an accurate representation of the time delay because
of judgments made by eye. This would mean that there was a time delay
of 2 seconds for the storm that was becoming dominate at that time.
Figure SpecA. This Spectrogram is of the low frequency outside
microphone located at the X End Station. It has dominate frequencies
raning from 0 to 275 Hz. The microphone produces clear and distinct
peaks with little background noise.
Figure SpecB. This Spectrogram is of the low frequency outside
microphone located at the Y End Station. It has dominate frequencies
raning from 0 to 350 Hz. The microphone produces clear and distinct
peaks with little background noise.
Figure SpecC. This Spectrogram is of the high frequency outside
microphone located at the X End Station. It has dominate frequencies
raning from 0 to 600 Hz. The microphone produces clear and distinct
peaks with more background noise than the low frequency microphones.
Figure SpecD. This Spectrogram is of the high frequency inside
microphone located at the X End Station. It has dominate frequencies
raning from 0 to 300 Hz. The microphone does not produce clear and
distinct peaks, but you can still tell that there is thunder outside.
There is a lot of background noise and would be more if the inside fans
were to be turned on. Also note that there are 60 Hz lines caused
by being power through a wall socket. Compared to the high frequency outside
microphone powered by batteries (figure SpecC) this is alot harder
to read.
Figure SpecE. This Spectrogram is of the high frequency inside
microphone located at the Y End Station. It has dominate frequencies
raning from 0 to 250 Hz. The microphone does not produce clear and
distinct peaks, but you can still tell that there is thunder outside.
There is a lot of background noise and would be more if the inside fans
were to be turned on. Also note that there are 60 Hz lines caused
by being power through a wall socket. Compared to the high frequency outside
microphone powered by batteries (figure SpecC) this is alot harder
to read.
Figure SpecF. This Spectrogram is of the magnetometer located
at the X End Station. It peaks instantly when lightening appears
within the LIGO area.
Figure SpecG. This Spectrogram is of the magnetometer located
at the TriNet Vault. It peaks instantly when lightening appears within
the LIGO area.
Conclusion
In Conclusion, the microphones should be place at each of the end stations
and the LVEA. They should be located on top of the roof so they are
not direction sensitive with the end station acting like a wall barrier.
When installed they should be put in a protected housing unit with a thick
soft foam insulation to reduce background noise produced by rain.
The microphones should be treated the same in every respect to get more
accurate results (i.e. same tube length for resistance, housing, location
on top of building, etc..). The microphones should send their signal
through a filter system with a low pass filter of 350 Hz and a high pass
filter of .03 Hz. Even with a built in pre-amp, the signal should
travel through another amp for better results. When anyalizing the
data, the data string should be average by 128 and take the RMS to give
results with less background noise. Correlation should be run in
a continus string with an observation time of a minute. This seems
to an appropriate time because it needs to be longer than 16.63 secs and
small enough not to include to many thunders at once. The reasoning
behind the 16.63 secs comes from the fact that the thunder takes that long
to travel from end station to the other. It doesn't need to be longer
than a minute because if there are a lot of storms in the vicinity of LIGO
then the correlation will not have distinct time delays in each frame.
This is interesting because we know that when the is only one storm our
accuracy will be higher than if there is more than one. Once all
three microphones are installed it will be easy to get a general area of
the thunder storm by way of triangulation (see
figure 8).
Preliminary Results based on the coincident detection of thunders
at both end stations
On 06/26/01 we observed a thunderstorm ~ 18 km north of
the observatory. Microphones were put outside each of the end stations
near the ground. Raindrops caused a significant amount of background
noise therefore we covered our microphones with a thick layer of soft foam.
The layer of foam over the microphones significantly lowered the
noise which in turn made the thunders more visible in our data. We were
able to record a strip of data that was easy to comprehend. This allowed
us to derive the arrival time difference between the end stations.
The thunder signatures were very clearly recorded at both sensors (Figure
10). We measured a time difference of 10.7s with a scatter of
0.5s from thunder to thunder. We were able to do this simply by looking
at the data by eye and are now working on an algorithm for triggering and
measuring the arrival time difference between the microphones. The
arrival time difference was measured by determining the delay between
characteristic features of the corresponding peaks (Figure 10).
Not only do the respective peaks correspond with each other but their time
delay is very stable from thunder to thunder and the structure of paired
peaks is surprisingly similar. We believe that this will be useful
in finding an optimal triggering algorithm.
Promptly after the storm, we downloaded lightning locator
information from www.weather.com. We
were able to obtain only the lightning map of SE United States. The
map was fairly hard to scale; therefore we had to do an estimate of distance
away from the LIGO site. We found that the center of the storm was
about 18 km north of the corner station. Knowing this and the length of
the arms we estimated a time difference of 10.5s (Figure 8), assuming
that the thunders originated from the center of the lightning storm.
This is in excellent agreement with our measurement. The assumption that
the thunders came from the closest edge of the storm gave us ~6.9 second
delay, which is incompatible with our observation. Based on our estimated
~0.5 second accuracy, it is fair to conclude that the surprisingly good
coincidence between the measured delay time and time difference computed
for the storm center is real. This also means that we should have a very
good resolution (in the order of +/- 1Km for fairly distant thunders).
We expect that more accurate coherence measurements and the use of 3 microphones
instead of 2 will increase our resolution.
Figure 8. Location and approximate extent of the storm relative to
the LLO detector (left)
Figure 8b. "Back-of-the envelope" arrival time difference estimate,
based on the www.weather.com lightning map (right)
Figure 9 Lightning storms detected at the South-Eastern United States
at the time of our coincidence measurement.
Figure 10. (left) A six minute segment of the thunderstorm
recorded at both end stations. The delay between the arrival times is clearly
shown. The first coincident event is magnified on Figure 10 (right). A
simple (by eye) comparison between the curves gives 10.7 +/- 0.5 second
delay between the corresponding transients.
Conclusion
Based on this preliminary measurement, it is expected
that we will be able to identify coincident thunders at all three stations
(X,Y end and LVEA) with an accuracy better than 0.5sec when using more
sophisticated tools. Our preliminary measurement assumes that the
thunder frequency is low enough that individual thunders are visible. The
data also indicates that we should be able to see farther than 20Km around
LIGO, especially when using the higher quality microphones which will be
mounted at the top of the buildings. The present microphones are simply
placed behind the buildings, and they are not the best quality therefore
we are in the process of finding an alternate suitable low frequency microphone.
These will help us to see further with better resolution. Care also must
be taken with the mounting of the microphones since thunder and rain comes
hand in hand and rain noise can significantly deteriorate the signal quality.
