Wavelet Homepage                               Table of Contents

7 Cockpit Voice Recorder Analysis  

            Keeping with our puzzle analogy, the electro-magnetic, triboelectric inputs to the signal are the pieces the puzzle, and the wavelet analysis technique provides a valid technique for analyzing the signal, or piecing together the puzzle.  The tests performed by the previous research group, and the nominal engine tape provide excellent references to direct the analysis of the actual CVR data.

            Although the CVR records four channels of data, the bulk of the analysis is focused on the silent track which functions as the public address system when in use, but is for the most part a latent wire when the P.A. system is not in use.  The reason for this is simple.  Research has shown that wires that have a high impedance are much more susceptible to the triboelectric effect.  A P.A. system that is not turned on will act like a non-sourced (open) circuit with almost infinite impedance.  For this reason, we expect the triboelectric effect to be much stronger in the silent track.

The CVR tape was originally digitized at 44,100 samples per second.  To better accommodate the wavelet transform, Matlab was utilized to re-sample the data at 32,768 samples per second.  A time-series analysis revealed that the silent track contained six events that deviated greatly from the average range of the silent track, as well as hundreds of lower amplitude “spike” transients identical to the one that preceeded the main transient in the nominal engine case.  The six events of significance can be placed into three categories: an initial event, four middle events, and the final catastrophic event.  Given in Figure 43 is the time-series of the silent track with the space between each event condensed to facilitate plotting.  Figure 44 is a representation of all events with a magnitude over 0.005 V on a proper timeline.  The Spike_finder.m program written by this group produced this plot.  The large spikes at the beginning and end of the spike plot are the initial and final event respectively.  The middle jumps represent the “middle” events and everything in between is a  “spike” transient.

Figure 43: Time-Series of Significant Events



Figure 44: Plot of Time versus Power for all Events over 0.005 V on Silent Track


  7.1 Initial Event

The initial event occurred just 32.5 seconds into the recording.  It is a high amplitude, short duration pulse followed by one second of high frequency “sound”.  What sets this event apart from the other events is its amplitude of 0.06V, and the fact that it seems to have two distinct parts, an initial pulse, and follow up noise.  Despite the fact that the event occurs at the beginning of the tape, its amplitude alone (compared to the norm of the tape) makes it worthy of consideration in any hypothesis of the cause of the crash.  One significant property of this first event is that it is very strong on the silent track and very weak on the CAM, pilot, and copilot tracks.  On the other hand, less powerful events on the silent track manifest more significantly on the CAM, pilot, and copilot tracks.

Figure 45 shows the entire initial event, and the boxed area in Figure 45 is shown in a close up view in Figure 46.  Figure 46 clearly shows a building trend in the initial event that is not present in the known recorder effect shown in Figure 47.  The significance of this building cannot be overemphasized.  We now have several examples of impulses, recorder effect and even a break, but nothing that resembles this resonating appearance.  The only question is what stopped the building?

Figure 45: Time-Series of Initial Event


Figure 46: Close up View of Boxed up Session in Initial Event



Figure 47: Known Tape Start Transient



            The Matlab wavelet toolbox was used to create a Coiflet wavelet representation of the initial event. That representation is given in Figure 48.  Notice that the pattern of the first section goes from high frequency to lower frequency while increasing its strength.  This reinforces the idea that the event was a building event.


Figure 48: Coiflet Wavelet Decomposition of Initial Event


What is significant about the remainder of the initial event is that it resembles the high frequency “ringing” seen after the break of the support strut of the Star-Lite.  This ringing coupled with the end of the building trend is strong evidence that the initial event is a break.  At this time, there is not enough evidence to conclude where the break was in a bolt, or a load bearing member.

  7.2 Middle Events

The next four events all lasted roughly 0.4 seconds and occurred at: 21:18.2, 21:30.0, 22:08.2, 22:30.0 minutes into the recording.  The four events are listed together because they all seem to be the same event occurring multiple times.  Each “event” appears to be a two “pop” occurrence, with a louder pop (0.035V) followed by a softer one (0.015V).  Although two of the events are masked by conversation on the pilot/copilot tracks, the other two events are obvious on all four recordings.  Figure 49 shows one of the middle events up close.

Figure 49: Time-Series of "Middle" Event


            What is immediately obvious is the event’s similarity to the transient on the nominal tape.  Indeed all four transients match the same pattern in both time and frequency as the nominal tape transient.  What is also significant is that the second “pop” on all transients is out of phase with the first “pop”.  The fact that all events match the same pattern so closely can only mean that they were all caused by the same event.  Figures 50 through 57 detail each transient and give their Coiflet wavelet decompositions.

Figure 50: Time-Series of First Middle Event

Figure 51: Coiflet Wavelet Decomposition of First Middle Event

Figure 52: Time-Series of Second Middle Event

Figure 53: Coiflet Wavelet Decomposition of Second Middle Event


Figure 54: Time-Series of Third Middle Event

Figure 55: Coiflet Wavelet Decomposition of Third Middle Event

Figure 56: Time-Series of Fourth Middle Event

Figure 57: Coiflet Wavelet Decomposition of Fourth Middle Event


The fact that all four events appear on all four tracks is strong evidence that the events are not caused by triboelectric effects.  The lower frequencies in the events also suggests an electro-magnetic disturbance caused the transients.  Unfortunately, due to difficulties with digitizing the nominal tape, it is impossible at this time to coincide the transient to a physical event.  The four transients on the Beech 1900C recording do coincide somewhat with engine out tests, but not directly enough to enable us to identify them as the cause.

            Looking closer at the CVR, we noticed that there were other events similar to the four middle events only of lesser magnitude.  An example of one of these events is given in Figure 58.

Figure 58: Example of Event Similar to Middle Event but Lower in Amplitude


            The most logical cause for these events is that they were the result of changes in power to the engine resulting in changes in the magnetic field surrounding the generator.  The smaller transients are the result of smaller changes in power.

  7.3 Final Catastrophic Event

At 32inutes and 22 seconds into the tape, the final catastrophic event occurs that causes the crash of the plane.  While this event has been noted in previous reports, its obvious importance makes it the key event in the tape.  This event is a broad-spectrum impulse reaching 0.1V in amplitude, and is the strongest event on all channels.  The recording stops immediately after the impulse as the recording device itself cuts off.  What is especially unique about the final event is that it differs on all four channels.  Given below in Figures 59 through 62 are the time-series plots of the final event over all four tracks.

Figure 59: Close up of Final Event on Silent Track


Figure 60: Final Event on CAM Track


Figure 61: Final Event on Pilot 1 Track


Figure 62: Final Event on Pilot 2 Track


            There are several points of interest in the four tracks.  Notice in Figure 62 that there is high frequency ringing similar to the strut break.  This is better seen through a Haar decomposition as given in Figure 63.

Figure 63: Haar Wavelet Decomposition of Pilot2 Track During Final Event


            The Ringing in the Pilot2 track is significant in that it most likely was an acoustic occurrence that came through the co-pilots microphone.  However with no acoustic example available to compare the event to, no conclusions can be made from the event.  The only example we have is on a silent track.  The silent track Coiflet decomposition of the final event is given in Figure 64.

Figure 64: Coiflet Wavelet Decomposition of Final Event

            Notice that the final event is an impulse of large magnitude coupled with a low frequency occurrence just after the impulse.  The impulse itself is not the most important thing in Figure 64.  What is important is the low frequency energy in the eleventh level leading up to the final event.  Given the power of the final impulse, the low frequency energy would have to be very powerful to be seen on the same plot.  Previously, Dr. Stearman had calculated the whirls flutter frequency of the engine to be 0.8 Hz.  This did not take into account the fact that the engine is in flight idle at the time of the final event.  Dr. Stearman has estimated that the whirl flutter frequency of the engine at idle would be around 3 Hz.  The eleventh decomposition level represents the 4 Hz-8 Hz band for the Coiflet wavelet plot.

            Based on this evidence we can conclude that a whirl flutter event did occur just before the final event. 

  7.4 Spike Transients

            The nominal tape showed one spike transient over the entire tape.  Figure 65 shows ten seconds of data from the silent track on the Beech 1900C that represents the normal amount of spike transients present throughout the recording.  What is most significant is that no spike transients exist in the short amount of time preceding the initial event, and that the spike transients appear only on the silent track.


Figure 65: Example of Spike Transient on Silent Track


            The spike transients present in the silent track of the cockpit voice recorder do not have one unique frequency band in which they all reside.  In fact some are very band isolated and some resemble impulses that exist over a wide range of frequencies.  This makes them very difficult to classify, but if they are a result of the movement of the plane as it flies, different movements could produce different spike transients.  Again what is unique about all the spike transients is that they are not represented on the other tracks.  This is strong evidence that they are produced by the triboelectric effect.  Figures 66 through 68 give examples of the wide range of spike transients present throughout the tape.

Figure 66: Band Limited Spike Transient


Figure 67: Mid-Range Frequency Spike


Figure 68: Impulse Like - Wide Band Spike


            While the form of the spike transients cannot tell us much, the fact that they exist only on the silent track may be something in itself.  The initial event is believed to be a break and exists only on the silent track, so we can say with reasonable certainty that the transient events are mechanical in nature.  Dr. Stearman has theorized that the spike transients are the result of the two ends of a broken member banging together.  If a broken member is indeed banging against itself, the difference in the frequency ranges of the spike could be the result of the member hitting at different angles.

  7.5 Conclusions from Signal Analysis

At this point we have determined all the pieces of the puzzle, or sources of input.  We have clarified the affects of those sources through research, and brought the pieces of the puzzle into focus with the wavelet decomposition technique.  All that remains is to assemble the pieces and look at the picture we have formed.

            The importance of the initial event cannot be overstated.  The fact that it has characteristics of a “build to break” event is paramount.  There is little doubt that the event is indeed a break.  The fact that the spike transients start only after the initial event is important, but given that there is less than 30 seconds before the initial event it cannot be stated as conclusive evidence.  However, the lack of similar events on the nominal tape lends credit to the idea that the spike transients signify a problem with the plane.

            Based on this data a plausible scenario forms.  An unknown event began in some member of the engine mount or possibly the mount itself began to flutter.  This member broke as shown in the initial event, or in the case of the entire truss shaking it broke one of its supporting bolts.  This break caused the numerous spike transients located throughout the tape.

            Although the engine mount was damaged, the pilot and co-pilot did not realize it as (based on the audio) the plane functioned normally.  That is until the final event.  The significant amount of energy in what is believed to be the whirl flutter frequency cannot be ignored, and is most like what finally caused the engine mount to break free.  This whirl flutter is what caused the crash of the Beech 1900C.

            On a larger level, we have shown that the triboelectric effect is visible, and can be identified as existing solely on the silent track.  We have also shown that the triboelectric effect results from mechanical events, thus we have found a method to distinguish between mechanical and electrical events.


  7.6 Defense of Wavelet Analysis Technique

As was noted in Table 1, a previous design group performed a wavelet analysis on the CVR.  There were several reasons why our group did not pick up where the previous group left off but instead chose to use different wavelets for our analysis, and we feel the need to validate our technique. 

1.                          Incomplete Research.  The previous group’s report lists the triboelectric effect as the sole source of input to the CVR wires, but initial research such as that discussed in sections 3.1 and 3.2 showed that this was not true.  The paper also states that the twisted pair accounts for the triboelectric effect, but research found that the purpose of the twisted pair is to cancel out the triboelectric effect.

2.                          Improper Wavelet selection.  The Morlet wavelet was designed to be used on continuous data, and will not produce correct results on discretely sampled data.

3.                          Unproven code.  A code downloaded from the web was used and the group stated it did not know how the inputs affected the output.

4.                          Unexplained Claims.  The paper states that for reasons unknown the code has a linear relationship with the entered wave number and frequency.  The formula for the wave number is 1/frequency.

            As an example of the differences in results the two wavelets yield, a sample second was run through the Morlet and the Coiflet wavelet in the Matlab toolbox.  Figures 69 through 71 give the results of the Morlet and the Coiflet wavelet transforms of the data in figure 70.


Figure 69: Morlet Wavelet Transfer of Middle Event


Figure 70: Time-Series of Middle Event


Figure 71: Coiflet Wavelet Decomposition of Middle Event


Wavelet Homepage                               Table of Contents