Acoustic Evidence for an MH370 Flyby at Cocos Island

Cocos Keeling West Island is an isolated atoll in the Southern Indian Ocean with an airport that routinely handles large commercial jet aircraft. It also has continuously recording infrasound detectors and seismometers. A straight line flight path to the previous MH370 search area would have passed far to the west of the island. As the search area moves north along the 7th arc, it approaches a point where MH370 may have passed close enough to be detected. Examining publicly available seismometer data reveals a unique signal at approximately the expected time for an MH370 flyby. Timing and indications of Doppler shift show that the plane may have passed within several km of the airport.


The infrasound detectors on Cocos Keeling island would be most likely to detect the passing of the plane, with an array of eight active stations recording data for the CTBTO to monitor and locate nuclear events. Unfortunately, past requests for the raw IM.I06 recordings have been denied. The CTBTO data center recently made a generous release of data to the public IRIS/FSDN seismic data networks, but the available IM.I06 infrasound record is only from April 3, 2014 to current, starting three weeks after the loss of MH370. The data is only accessible to institutions with an established CTBTO contract, which restricts them from sharing it. Researchers with access could use the infrasound to verify the findings in this report, and ideally the CTBTO will release all recordings around Mar 8, 2014 to the public data archives.

Within days of the MH370 flight, the CTBTO held a press conference and released a report on analysis of the infrasound. It appears that only automated impulse event detection records were examined, which have a relatively high threshold to avoid the clutter of cultural noise. They were looking for an explosion or impact and found none.

As an alternative to the infrasound array, data is publicly available from seismometer II.COCO located at the Cocos YPCC/CCK airport. The quality of the recording is good with no gaps, but there is minor noise from waves rolling by on the nearby beach, and local blowers or pumps. The airport was closed at the time of interest, just before dawn.

Using a seismometer to detect aircraft is not far-fetched. A detailed  2015 study titled “Helicopter vs. Volcanic Tremor” reports that helicopters can be reliably detected as far as 40 km from multiple seismometers, and also provides graphical examples and formulas for the Doppler shift. Jet aircraft have a completely different sound profile, but as anyone who lives near an airport knows, the rumble of a passing jet can be heard for several minutes.

Expected Timing

The well-investigated MH370 SATCOM handshake timing has produced arcs on the globe showing where the plane could have been, given the expected distance from the Inmarsat 3F-1 satellite in geostationary orbit. The 2014 Mar 8 00:19:37 UTC final handshake at the 7th arc has defined the current search area. Cocos island lies between the 4th and 5th arcs denoting handshakes at Mar 7 21:41:27 and 22:41:22. The plane is generally assumed to have been on a steady course. An expected flyby time can be approximated by picking a path between the arcs that intersects Cocos, and interpolating the time based on distances from the island to the arcs.  Various paths passing between the arcs past the island would deviate only slightly from the expected flyby time.  Using a great circle path derived from an approximately 480kt cruising speed in previous operational reports that passes by Cocos gives an expected flyby time of 22:20:02 UTC.

A Unique Event

Exploring the COCO seismic recording finds an event centered on around 22:22:22 UTC, just 140 seconds later than expected. Rising just above the noise, it is the strongest event after hours of relative quiet and has a broad higher frequency energy component that sets it apart from the repetitive beach waves.

One-hour spectrogram of the COCO seismometer around flyby time.

The first figure is a spectrogram of the II.COCO seismometer showing a spectral energy plot for one hour surrounding the event. The three axes of the instrument are assigned to the RGB channels as: Red = east, Green=north, Blue=vertical. (Click to see a higher resolution .pdf image).

The slowly descending red line at around 7 Hz is not a Doppler signal. It might be a fishing boat heading out to sea pre-dawn, as it recedes over time. The yellow pulses at 8 Hz are likely an on-demand pump. The event of interest is around the central green then red spikes.

Zoomed spectrogram detail at flyby time.

These plots are showing detection of vibration on a north-south axis followed about one minute later by slightly stronger motion on the east-west axis. This peaking of the signal in different directions could be caused by resonance of the jet noise at different angles with the walls of the vault housing the seismometer. [Please see the March 2021 update below for details on how the 80 second gap and broad frequency components indicate MH370 flying over the atoll’s coral reef north and then east of the airport seismometer.]

The peaking is fortunate, as it not only brings the signal above the noise, but provides a way to estimate the distance to the flight path. The maximal distance would be from a flight path that is oriented at 45 degrees from a rectangular vault. A course at either a shallower or steeper angle would pass closer to the detector. Higher airspeed would also imply a greater distance given the timing between the signal peaks.  The flight altitude would have little affect on the separation time between the peaks, causing the same delay for both. Only the maximum horizontal distance to the flight path is measured this way.

A constant heading flight path starting west of Nicobar at waypoint APASI requires some speed variations to match each arc timing with an average speed of around 360 kt = 667 kph = 11.1 km/min. Using this as the maximum speed for a flyby, the one minute flight distance over a right hypotenuse would be twice the horizontal distance from the island. This approximation implies that the flight path passed within 6 km of the airport. If the plane continued on that constant heading of around 168.14 degrees, it would intersect the 7th arc north of Gulden Draak seamount. It would be very close to where the original Curtin event H01 hydrophone bearing 301.6 meets the 7th arc. 

Doppler Hints

A Doppler shift in the signal is barely detectable as diagonal lines descending past the peaks. The Doppler shift could be useful for estimating the altitude of the plane. If the plane passed close enough, the Doppler signal would appear as an inverse sigmoid or reverse S-shaped curve. What appears in the spectrogram is a linear central portion of the curve surrounding the peaks. As the plane would be flying tangent to the detector at the point of closest approach, there would be no Doppler shift at that instant, with the inflection point on the curve matching the frequency of the source. The rate of shift is proportional to the frequency, and is determined by the velocity and the closest distance to the source.

Most Doppler applications are about determining the velocity of an object from the maximum frequency shift found when an object is directly approaching or receding. Here we can only detect the slope of the Doppler shift at closest approach. Using estimates of the velocity results in estimates of the distance at those speeds. In fact, this Doppler method would have been used by SARSAT search and rescue if any one of the ELT (Emergency Locator Transmitter) devices on the plane had deployed after impact. A satellite passing overhead in Low Earth Orbit would detect the beacon. Knowing the precise speed of the satellite gives an estimate of the distance at closest approach. The book, “Satellite Systems for Personal Applications: Concepts and Theory” (Ch, p. 197) provides the differential equation for determining minimum range using the Doppler slope method:

closest range = velocity^2 / (wavelength * slope)

A visual estimate of the most apparent Doppler slope diagonal in the spectrogram has a center frequency of about 5.3 Hz and slope of 0.50 Hz/minute. This allows some rough estimates for the flyby distance at different velocities.

Doppler shift example plot for f=5.3Hz, distance 10-50 km, velocity = 550 kt

Previous estimates of the plane flying at Long Range Cruise (LRC) speed of mach 0.82 (82% of the speed of sound, or about 550 knots), cause a very large Doppler shift. Note that the bandwidth of the Cocos Island seismic and infrasound sensors is only 10 Hz, while a 5.3 Hz source signal on approach at 550 kt would be fully shifted to over 30 Hz. It can also be seen that a near pass of the island would cause a very sharp Doppler slope at that speed.

Doppler shift example plot for f=5.3Hz, distance 10-50 km, velocity = 200 kt

A closer match to the slope seen in the spectrogram would be at a flight speed of under 200 kt, with more distance to the aircraft. This implies that either the plane was flying slowly at a higher altitude, or the assumptions above for estimating a flyby within 6 km were flawed.

Doppler shift example plot for f=5.3Hz, distance 10-50 km, velocity = 150 kt

Operation at a flight ceiling of FL360 or about 11 km altitude would require a high velocity, which would mean a longer distance from the airport to match the Doppler slope. The plane is far less likely to be detectable at more than 20 km.

The visible slope of 0.5 Hz/minute would match the plane flying slowly near its approach velocity of around 140 kt, about 15 km total distance from the Cocos airport. If the 6 km horizontal distance is accurate, it would mean the plane was flying very high for that speed.

Seismic Analysis

The three channels of seismometers like II.COCO typically record motion in east-west (E), north-south (N), and vertical (Z) directions. Besides detecting timing differences of different seismic phases, “polarization analysis” uses the eigenvector rotation of the sampled data to determine the 3D direction of motion over time. This provides strike (azimuth) and dip (inclination) info for a seismic source. A customized plot has been developed from first principles for visualizing the seismic data, showing channel cross correlation, Fourier and instantaneous frequency spectra, the waveform and envelope, along with a chart of rolling polarization and statistical measures. (Click for a full size .pdf file.)

Seismic analysis at flyby time, with azimuth estimate in blue.

There is an ambiguity of 180 degrees on each axis, which can generally be resolved by assuming the origin is below ground. An aerial source breaks that assumption, but even an ambiguous forward/back azimuth (shown in blue) is useful for checking a flight path past the airport. The azimuth readings should be considered only an approximation. The results reflect the dominant signal, which at times may be a noise source like waves rolling down the nearby beach.

The azimuth plot wanders, but appears to show that the plane approached from the NW on the approximate bearing 330 of the runway (15/33), interspersed with 90 degree resonances. The angle shifted west to near 280 before flipping at about 22:23:00 (just after the second resonance peak) to the south, then receded toward the same flight heading 150 as the approach.

Seismometer Details

The orientation of the seismometer in the vault is questionable. Most seismic records include axis orientation E/N/Z in the channel label. The COCO channels are labeled BH1, BH2, and BHZ, with no consistent standard for channel 1 being N or E. Different catalogs list the COCO seismometer as being rotated by various angles at different times. For the 2014 time period, catalogs and embedded data variously report it as both 16 and 106 degrees from true north. Magnetic declination at Cocos is near nil.

A 2017 check using the DLOPy python/obspy tool for orienting Ocean Bottom Seismometers (doi:10.1785/0120160165) reported an orientation of 104.73 (or 14.73) degrees from 40 months of reference seismic events.

The first unpublished draft of this report had the channels swapped. Mid-2018, the data was reacquired through ObsPy and corrected for phase and drift. It was relabeled as ENZ order (BH1=BHN) with nil (unknown/minimal) azimuth correction, based on comparison with reference seismic events on Mar 7-9, 2014.

The proposed vault resonance with strong N-S energy at 22:21:20 counterintuitively reads an azimuth near 90/280 degrees. The azimuth is determined by phase differences, not amplitudes. The lesser E-W wall motion resonance is still physically tied to the N-S wall in a way that probably causes a larger signal in either wall to register as a 90/280 degree origin.

A further consideration is that the seismometer may not be oriented with the vault. Satellite imagery shows that the Cocos airport buildings of the Meteorological Office near the approximate seismometer coordinates (also shown in the station photos) are not aligned on either N-S or with the runway, but at around 45 degrees from north. There is a small building 160 meters NNE of the main building that is oriented at about 15 degrees from north. This also approximates the closest approach seen just as planes touch down on the runway.


Local residents did not report hearing a fly-by in the pre-dawn hours, which might imply that the plane was flying high, or gliding past the airport with engines idling. The plane would then need to throttle up to reach the 7th Arc. This suggests pilot control, and therefore the airport as an alternate for landing.

The conditions in the cockpit are unknown, but there has been speculation that the initial diversion was caused by a windshield heater fire. Visibility may have been limited, and electrical problems may have disabled IFR instrument landing ability. [N.B. 2019 – There is no IFR glide slope facility at Cocos or Christmas Island airports.] Electrical problems could also have prevented dumping of fuel. Large jetliners are not designed to land with a heavy fuel load, especially on shorter runways.

After extinguishing a fire, the plane may have descended to an altitude where oxygen was not required, and flown at near HOLD speed (maximum endurance) to an airport that might be safest for a daylight landing. This would also give the crew time to attempt repairs and plan for landing a disabled plane.

Approaching the airport before dawn, a very slow flyby at a high altitude makes sense to evaluate the conditions for a landing. There were only scattered clouds but the airport was closed, and possibly without lights. The conditions in the cockpit and the ability to maneuver or land the plane may have degraded since the decision to select Cocos YPCC as a waypoint.

The heading of the plane after passing Cocos airport might be derived by estimating the flight speed for crossing the remaining arcs.


Sincere condolences go out to families of those lost on flight MH370. Please understand that details are still coming to light, and the focus here has been on pursuing the research leads rather than progress reports. It is hoped that the result of future infrasound analysis will be the required “credible specific location” for continuing the search.

— Ed Anderson


Mid-2018 Update:

A draft of this report was created in October 2017, but the full significance wasn’t realized during investigation of other seismic data leads to support the findings. The figures have since been revised using corrected azimuth calculations and improved spectral plot methods. New findings have been added.

July 2018 update: was previously unaware that Lawrence Livermore National Labs researchers examined the infrasound data in 2014. An unindexed report was recently revealed on the LLNL site as an informal .pdf poster titled, “Infrasound and the search for MH370” by R.Kane, S.Lehman, and J.Candy. The authors attempt to match infrasound recordings with four Airbus A320 takeoffs and landings at Cocos CCK airport for flights to/from Christmas Island CXT airport on Mar 7th and 8th 2014. They conclude, “We don’t see anything to start with in the data with aircraft activity!” That report was possibly the basis for infrasound being dismissed as useful in the previous ATSB summary report.

Unfortunately, the LLNL study has a major flaw. The time zone corrections were added rather than subtracted from local time. This caused the examination of time ranges when the airport was closed. Each aircraft takeoff and landing has a prominent signature at the proper times on seismometers at both Cocos and Christmas island airports.

The LLNL poster does have a small but zoom-able graphic representation of signal amplitude at each Cocos infrasound station over the surrounding two days. They show a daily cultural noise cycle, spikes around the corrected takeoff/landing times, and a strong isolated peak on five stations near the proposed flyby time. ]

[p.s. From 2018 to 2020, one of the LLNL researchers (now retired) made attempts to take a second look, but gave up after making no progress at getting the recordings through his LLNL contacts.]

March 2021 update:

It was previously guessed that the unique signature of first a N-S energy pulse followed 80 seconds later by an E-W pulse might have been caused by resonance with the wall of the seismometer vault. One problem with this idea is that a resonance should have very distinct banding in the spectrogram for the fundamental frequency and any harmonics. Instead a broad range of higher frequencies is seen. The second insight comes from tracking other low flying lanes over land with seismometers. The sound appears to travel from the ground nearest the plane at high speed through the crust to the seismometer. The doppler shift through the ground is also quite different from through the air, as if the recorder were tracking along with the plane.

A more plausible explanation for the signal at Cocos Island is that it is a coral atoll, and water disperses the sound of a jet flying overhead rather than transmitting it into the seabed. The two stronger bursts would come as the pressure wave from the plane passes first over the land reef to the north of the airport, then over the water in the shallow lagoon, and over the narrow reef to the east.

A first estimate of the turn radius was for a 25 degree bank autopilot turn at the 377 knot averaged flight path speed near Cocos. This made for a sharper turn that spent more time over the land to the north, and the arc length divided by the time gave a lower speed of 325 knots. Recomputing for a 20 degree bank at 325 kts gave the turn path shown on the map. Computing the 6.4 nmi path over 80sec (.022 hr) = 314 knots. A very similar turn could be flown at a 15 degree bank (as estimated for autopilot turns earlier in the flight) and 315 knots. It would have a slightly shorter pass over the reef to the N, and the path length should give a matching speed very close to 315 knots. The altitude should not matter, as it would only cause a delay and reduce the intensity.

Map showing an MH370 autopilot turn past the Cocos Island Airport waypoint, crossing atoll reefs N and E.

The faint doppler shift around 5.6 Hz at closest passage between the bursts still reflects the far weaker sound transmitted through the air when the plane is over water. Below is a recently computed Doppler plot that should better match the speed of MH370 if the above map is accurate.

If the 315-325 knot speed is accurate, it also indicates that the plane slowed down from its averaged speed from Arc 4 to Cocos and Cocos to Arc 5, then sped up again. If engines were idled for a descent, that might explain why there no reports in the wee hours of a jet at oxygen altitude.