Cocos Island Infrasound May Be Key to Locating MH370

There is an array of eight infrasound recorders at Cocos Keeling West Island that was continuously collecting data during the flight of MH370. That data has been unavailable to the public. Further study was previously dismissed after separate analyses in early 2014 by multiple governmental research teams that found no infrasound evidence of flight MH370. A closer look at those reports shows flaws in the methodology that may have caused key information to be dismissed.


Two days after flight MH370 went missing, the head of the CTBTO noted during a press briefing that their infrasound monitoring stations would be most suitable for detecting an explosion or impact of the aircraft. One day later, CTBTO put out a Mar 11 press release and an initial report from their data center that no infrasound detections were found. The search area at that early date was focused on the Gulf of Thailand and Straits of Malacca.

On Mar 14, the CTBTO put out a call for researchers to investigate the data for signs of the missing flight. Unfortunately, the invitation was only to their member state research institutions, who are under strict contract which prohibits sharing the data. CTBTO has declined to contract with independent researchers or release data from March 2014. In 2017 they began releasing infrasound and hydrophone data to the public seismic networks, including prior years. For the Cocos Island array, that data set frustratingly starts on April 4, 2014.

From their report, it appears that the CTBTO only examined automated event reports above a certain noise threshold. Their search started at the takeoff time of the plane and was centered on the location where the plane transponder stopped, consistent with looking for a mid-air explosion. Cocos Island was marked on a map but outside the likely 2000 km potentially detectable range. While the single plot from an infrasound array in Japan appears to cover 24 hours UTC on March 7, late events from the Cocos array were not mentioned and probably ignored at the time. Hubert Foy of wrote a detailed analysis of the CTBTO infrasound report within weeks.

In later discussions with the Australian search team about taking a closer look at the Cocos Island infrasound data for weak Doppler signatures, they seemed convinced that the data had already been examined in detail by experts and dismissed.

In July 2018, comment discussion on an MH370 blog focused partly on the CTBTO infrasound, and a link reference by member David revealed a one page informal poster report by LLNL on the Cocos Island infrasound. A conference agenda page that lists the paper shows the presentation date to be May 21, 2014. The report undoubtedly was provided in support to the search authorities, and was a basis for dismissing further inquiry.

A Closer Look at the Poster Report

The LLNL study design is well founded, using the creative idea of checking whether twice daily A320 jet flights to and from the Cocos Island airport were detectable on the IM.I06 infrasound array located within a few kilometers of the runway. The local flight schedules provided an approximate time for each takeoff and landing for Christmas Island shuttles, and two days of data on eight sensors were analyzed. They found no infrasound events matching the known aircraft activity. This negative finding would seem to be in conflict with the statements from the CTBTO that jet traffic is routinely detected by the infrasound monitoring network.

Ultimately, a simple time zone error flawed the analysis. While the poster listed the correct offsets for local time, UTC+6:30 for Cocos CCK and UTC+7:00 for Christmas CXT, those offsets were erroneously added instead of subtracted from local time to get UTC. The result was that the infrasound was being examined for expected flight traffic in the middle of the night when the airport was closed.

The resolution is limited when zooming into the poster on the plotted waveforms for each station, but the detail is enough to confirm the error.

Recordings on Mar 7 and 8, 2014 from eight infrasound stations on Cocos Keeling West Island near the airport.

The time zone on the infrasound data would be UTC. Notice the increased daily noise activity starting after 01:00 UTC, which is 7:30 AM local time. This is more prominent on the March 8 plots, as the March 7 data has a higher noise level. It is most likely due to wind noise at the sensors, which could also account for the slow burst of noise around 00:00 UTC on March 8th. Waves on nearby beaches are another source.

The takeoffs from Cocos CCK would of course be louder than landings. Terminal departures are scheduled for 13:43 and 15:59 local time, or 07:13 and 09:29 UTC (not 22:37 and 00:50 as listed in the poster.) Allowing a few minutes to taxi, takeoffs might be around 07:18 and 09:40 UTC. Spikes of activity can be seen around those times on the various plots.

Checking with Seismometers.

The known flight timing can also be checked using another method of infrasound detection. Cocos Island has an active seismometer II.COCO with available data, and Christmas Island has both AU.XMIS and AU.XMI.

Here is a seismometer data plot showing what appears to be accurate timing for taxi, runup, and takeoff of flight VOZ1917 at 07:15:47 UTC.

Takeoff from COCO of flight VOZ1917 at 07:15:47 UTC.

There are frequency changes that would be caused by a combination of throttling up and Doppler shift as the plane picked up speed. The seismometer would be far more sensitive to vibrations through the wheels than the air, but the trailing off in intensity over a minute would indicate infrasound. Wind was about 15 kt from 130 deg, so the early noise burst may have been the plane turning around near the seismometer at the northern end of the runway.

A second plot of the later flight shows it taking off one minute before the scheduled gate departure time.

Takeoff from COCO of flight VOZ1917 at 09:28:00 UTC.

It looks like the crew and passengers got going ahead of schedule for both flights. The spectrogram of the second takeoff shows a different pattern of frequency variations, possibly wind gusts, but along a clearly increasing trend. XMI on Christmas Island shows similar takeoffs. The landings are also detectable.

Looking for a Blessing in Disguise

The real interest is in the ability of the infrasound array to detect the flyby of jet aircraft. Although the raw infrasound data is inaccessible and the negative reports may have delayed a proper analysis, the zoomed poster plot can still be utilized. A recent writeup details a possible flyby of Cocos Island by MH370 at 22:22:22 UTC on March 7, by using the active seismometer COCO at the airport as an infrasound detector.

A tentative confirmation can be seen in the prominent spike near that time on four of the plots. Interpolating the pixels, it appears closer to 22:40 UTC. The shift might be a cropping artifact of the reduced plot size. A seismometer spectrogram of the 22:20 to 22:50 time range shows only the flyby candidate event, plus what could be 4-8 Hz wave noise on the nearby shore.

Future analysis of the raw infrasound data will hopefully show far more detail on Doppler shift and closest approach timing for deriving altitude, velocity, and possibly even the flight heading.

April 2019 Update: Testing the Detection of Jet Aircraft with the Cocos Island Infrasound Array

With data for IM.I06 publicly available for dates after April 3, 2014, is is possible to execute the experiment attempted in the LANL poster, but for an available day. There is an advantage to using a recent day, because an ADS-B receiver has since been installed on Cocos Island, which can track the time and location of nearby commercial aircraft with GPS accuracy.

A search for the flight schedule of YPCC/CCK Cocos Island Airport found that tracking was available on the FlightRadar24 site for two flights on April 12, 2019. Flight VA1913 from Perth touched down at 05:35 UTC, and flight VA1917 left the runway at 07:02 UTC for Christmas Island. There are eight stations in the IM.I06 infrasound array. Downloading data for that day found one station H3 of the eight was missing, providing recordings from seven stations.

A first quick attempt at visualization was to use triads of stations in the array with the BearingOverTime plotting tool (not shown) used for hydrophones. The principle is the same, using known lat/lon for each station position and the known local speed of sound in air. Shown are normalized spectrograms from three receivers H2, H4, and H6 in the infrasound array.

Spectrogram for flyby-landing infrasound (not hydrophone) triad I06-H2,H4,H6.
Spectrogram for jet takeoff infrasound (not hydrophone) triad I06-H2,H4,H6.

The spectrograms show that the infrasound environment is cluttered by noise on this day from wind noise plus waves crashing on the nearby beach. Nullschool modeled wind speed is 15kn from ESE, similar to on the MH370 date. Wave height is 2.7m from SSE. The flyby spectrogram is centered on the closest approach, about a minute before touchdown. The takeoff spectrogram is centered on the wheels-up time. Faint rising spectral tones around 2-3 Hz can be glimpsed in both plots.

Using simple Time Difference Of Arrival (TDOA) methods to plot the back azimuth over time did not provide a useful demonstration. The plots are dominated by the beach waves. An effort was undertaken to utilize the entire I06 array by applying a beam-forming method. This was accomplished in Python using open source tools.

The basic software beamforming approach can be described in brief. Each of the 21 pairs from 7 available stations is cross correlated and the results saved for reference. Peaks in the X-corr data represent the sampling delay as a similar waveform pattern passes across two separated microphones. The widest spacing in the array is 1500 meters in the N-S direction; the shortest is 200 meters. A digital cartographic map is built where a set of calculations are pre-computed for each pixel. The lat/lon pixel coordinates give a distance to each station microphone, and thus an expected delay time at the speed of sound. The map delays index into the saved cross correlations as lookup tables, generating spatial signal intensities into the map in a single Numpy array operation. The map generation is quick enough to generate a 300 frame animation in a few minutes of single core processing. Following is an infrasound visualization of the recorded jet takeoff. (The red color in the map shows spatial information. The radial distance of cyan colored data is proportional to the speed of sound.)

This is firm confirmation that the Cocos Island infrasound array is capable of picking up jet traffic details. There are several things to note about the infrasound map animation. The time is centered on takeoff, so the first third of the movie is local background noise. A .kmz file provides the I06 array locations on the map for reference in Google Earth. While distant sources are not well ranged on the map, the azimuth appears to be very accurate. A repeating source at bearing 193 points to two shore buildings very close to the array, but may also be waves on the beach (2.7m from SSE).

The throttle-up of the jet engines seen starting at 07:01:00 is on a bearing of 171 degrees from the center of the array, which points within meters of the north end of the runway. About ten seconds later, the bearing is showing movement of the jet down the runway. At 7:01:45, the bearing stabilizes at about 162 degrees, which would require takeoff and points near the end of the runway. The plane holds that heading away from the array and quickly fades from the animation by 07:02:15 UTC.

That fadeout may not indicate the full range of the infrasound array. The fade could be explained as a side effect of the TDOA method being used for the beamforming summation, which works best with stationary sources. When a source is in motion and varying in intensity, there may be a significant shift in the perceived delay between two station microphones. Different doppler shifts at each microphone would disrupt the phase correlations.

The color coding on the animation is based on shifts in the expected delay. This shows an effect at narrow shifts, but the signals do not sum properly among the 21 pairs and the beamforming is quickly lost as the plane picks up speed. This could be compensated for by constructing a visualization designed for specific speed vectors by color coding them as map layers. An indication that the infrasound from the plane is still detectable is to look at the appearance in the animation at 07:02:28 and 07:02:45 of artifacts toward the NW with a progressively shifting azimuth. They may be beach waves, but none have that magnitude anywhere else in this time period. These artifacts are possibly the points where summations from multiple station pairs coincidentally line up with the speed vectors of the plane.

The fadeout of a fast moving source may explain why the jet approach and flyby within 1 km of the array is barely detectable in the following infrasound plot of the landing:

There are strong artifacts washing over the array animation until the plane has touched down and can be seen slowing at 05:35:30 UTC by the shifting signal azimuth 167 which points mid-runway.

Detailed ADS-B GPS tracking as .kml files of the VA1917 takeoff and VA1913 landing have been obtained from FlightRadar24 for comparison with the infrasound response.

Longer range detection should be quite feasible. A larger correlation window to reduce noise shows evidence of the inbound flight VA1913 final approach track passing about 25 km NE of the array. Fine tuning of beamforming for the speed vectors could detect a more distant flyby. Flight tracking sites show that Cocos Island frequently gets flyovers from high altitude jets on paths from Europe and the mideast to Australia. These may be useful for testing other beamforming methods. It may also explain why there were no reports of an MH370 flyover from residents of Cocos Island, since flyovers there are common at all hours.

Using a different beamforming method that searches the Frequency-Wavenumber space for best fit over time (FK analysis) provides a bit more detail on the VA1913 landing and local noise sources above 2 Hz.

Frequency-Wavenumber (FK) analysis plot of flight VA1913 landing at YPCC.

This plot uses a wide 30 second window to reduce noise, which masks weaker sources. It does clearly show the movement along the runway at 05:35. Also revealed are rolling waves on the beaches, as identified by the northward shifts of the strongest event at 05:41 and similar weaker sources near bearing 185 to 195 (-175 to -165). The “slow” subplot shows the inverse of the speed of sound. The median value of 2.857 equates to a sound speed of 350 m/sec. Aside from noise, lower slowness would indicate the angle of incidence of a source from above or below the array, where the apparent velocity is higher. An inverse number is used to avoid the case of a source directly above/below the array with zero slowness having an infinite apparent velocity. Slowness values over 2.9 on the plot would indicate a wavefront traveling slower than the speed of sound in air. They might be artifacts of the motion of the aircraft or jet exhaust.

Further development of the infrasound analysis tools could bring out more flight path details if an MH370 flyby is detected. That work is on hold until the proprietary infrasound data from March 7, 2014 becomes available. The prospects for detecting weaker signals around an MH370 flyby should be good, since there was less environmental noise at Cocos Island predawn at 04:52 AM than during these daytime flight comparisons.

Curiosity continues to bring incidental insights related to the acoustic data. For example, an accurate estimate was needed for the speed of sound. It can be obtained by accumulating cross-correlations between station pairs over several hours and measuring the upsampled delay peaks. This resulted in a consistency check for the array which shows only minor variations in speed among the 21 spatial pairings. The mean speed comes out to 349.46 meters/sec for midday on April 12, 2019. This corresponds to an average slowness of 2.86 on the FK plot above. Since the sound speed is dependent on temperature, it can be accurately calculated as a hot 29.97 degrees C = 85.95 F at Cocos West Island. The infrasound array turns out to be a very accurate recording thermometer, which might be useful for correcting historical local instrument log data. The relative station locations were corrected to within a meter using Google Earth, so the speed variations seen within the array could be due to sunny vs shady sound paths.

The Cocos Island infrasound array appears to be perfectly viable for detecting a flyby and confirming the nature of the event detected by the island seismometer. The seismometer shows that there were no other significant acoustic events there besides the 22:22:22 candidate during the time between the fourth and fifth pings. This closely matches when the plane would have crossed the radial distance of Cocos Island from the satellite.

[Jan 2020 Update]

A higher resolution same from one of hydrophones was published in the LLNL poster. It shows a signal peak at 22:46:40 UTC.

Cocos Island infrasound peak at 22:46:40 UTC from the LLNL poster.

The infrasound records are proprietary to CTBTO, but spectral analysis of the II.COCO seismometer around 22:46 shows a short 3.6 Hz event lasting only a few seconds. The infrasound peak appears to last much longer, and may represent a very low frequency pressure wave arriving from the passing aircraft, similar to a wake vortex. Ultra low frequency infrasound has minimal atteuation in the atmosphere, and can travel long distances – hundreds or thousands of km for loud events like bolides or explosions.