NCAR In Situ Vertical winds-based EDR Estimation Algorithm ... · This system estimates values of the cube rate of the eddy dissipation rate, referred to as the EDR, an atmospheric

Ver 0.9, 20130321 ©1991-2013 UCAR RAL All Rights Reserved

NCAR In Situ Vertical winds-based EDR

Estimation Algorithm

Description

March 21, 2013

Authors: Dr. Gregory Meymaris and Dr. Robert Sharman

Affiliation: University Corporation for Atmospheric Research (UCAR)

NCAR In Situ Turbulence Algorithm Description

Ver 0.9, 20130321 ©1991-2013 UCAR RAL All Rights Reserved 1

Table of Contents Introduction ......................................................................................................................... 2

EDR Algorithm Overview .................................................................................................. 3

Quality Control Algorithm .............................................................................................. 4

Vertical Wind Calculation .............................................................................................. 8

Maximum Likelihood Estimation ................................................................................... 9

Data Sampling Rates ..................................................................................................... 16

Event Based Triggering Logic .......................................................................................... 16

References ......................................................................................................................... 20



Introduction

Historically, there has been one main source for turbulence observations, namely pilot

reports (PIREPS). There have been problems using PIREPS including:

• time/position fixes are often far away from where/when the turbulence occurred

• the turbulence intensity reported is subjective, not measured

• the turbulence intensity reported is not atmospheric but rather aircraft-centric

• there are a dearth of NULL reports

These deficiencies make PIREPs ill-suited for providing maps of atmospheric (i.e., aircraft

independent) turbulence levels over aircraft flight routes. To address these difficulties, NCAR,

under sponsorship from the FAA Aviation Weather Research Program (AWRP), has developed

an in situ turbulence monitoring system (Cornman et al. 1995, 2004) which has now been

implemented on three major U.S. commercial air carriers composed of three different aircraft

types. These reports are routine, completely automated, are aircraft independent, and are

accurate to within at least about 10 km and 1 min, and are therefore ideal for operational

purposes (Farrar et al., 2013) and turbulence case studies (e.g., Lane et al., 2012).

This system estimates values of the cube rate of the eddy dissipation rate, referred to as

the EDR, an atmospheric turbulence intensity metric. The EDR estimation algorithm is coded as

a C library which is then loaded on board the Aircraft Conditioning and Monitoring System

(ACMS) or other suitable onboard computer. Once on board the aircraft, the software runs

continuously when the aircraft is off the ground, producing mean and peak EDR over the last

minute of flight once every minute. While for the original accelerometer based version, running

currently on United B757-200s, all one minute mean and peak EDR estimates are downlinked,

the newer vertical wind-based version, running currently on Delta B737NG and (recently on)



Delta B767-300/400, and soon on some Southwest B737NGs, has the capability to report only

turbulence events plus occasional (e.g. 15 minute), routine reports. The EDR algorithm

described in this document pertains only to the vertical wind based algorithm and not the older

accelerometer based one. (described in Cornman et al., 1995)..

While there are a few adjustable parameters that are aircraft dependent, the EDR itself is

aircraft independent. This is important since an aircraft dependent turbulence metric would need

to be accompanied by aircraft data, such as airspeed, altitude, mass, and type in order for others

to be able to interpret it. While this is certainly possible, it is easier to just report one aircraft

independent quantity. An important component of the EDR algorithm is sophisticated quality

control (QC), which is implemented onboard as part of the software package, but may have a

ground-based component as well.

This document describes the vertical wind based EDR algorithm and the event triggering

logic.

EDR Algorithm Overview

The in situ turbulence algorithm is a vertical winds-based maximum-likelihood algorithm

which can be conceptually thought of as having 3 components: the vertical wind calculation, the

maximum likelihood EDR estimation, and, integrated into the previous two, a quality control

algorithm. The algorithm takes as inputs (see Table 1) the left and right angle of attack ( 1 and

2 ; deg.), true airspeed (V; m s-1

), roll (; deg.), pitch (θ; deg), and inertial vertical velocity ( h ;

ms-1

; positive up) and outputs, once per minute, the mean and peak EDR (1/3̂ ; m

2/3 s

-1), along

with associated quality control metrics. Pressure altitude, h (m), and vertical acceleration, gw, are

not used in the algorithm, but are used in the tuning.



Parameter Symbol Required for

Algorithm

Units Minimum

Sampling

Frequency

(Hz)

Minimum

precision

UTC date and

time

NA No NA 4 0.125 s

Aircraft type NA Yes1 N/A

True Airspeed VT Yes m s-1

4 0.036 m s-1

Auto-pilot

(on/off)

No 4

Left vane angle

of attack (+

nose up)

1 Yes deg 4 0.05 deg

Right vane

angle of attack

(+ nose up)

2 Yes deg 4 0.05 deg

Pitch + nose up) Yes deg 4 0.02 deg

Body pitch rate No deg s-1

4 0.02 deg s-1

Roll (+ right

wing down Yes deg 4 0.02 deg

Inertial Vertical

Velocity (+

upward)

IVV Yes m s-1

4 0.00508 m s-1

Barometric

Altitude

Z No feet MSL 1 1 ft

Vertical

Acceleration

gw No g 4 0.003 g

Current aircraft

weight

M No lbs 1 100 lbs

Latitude No deg 1 0.0002 deg

Longitude No deg 1 0.0002 deg

Wind Speed No knots 1 1 knot

Wind Direction No deg 1 1 deg Table 1: List of required parameters and update frequencies for vertical winds-based algorithm including running real-

time, tuning, and verification.

Quality Control Algorithm

The quality control algorithm is crucial as minor spikes or steps in the data can result in

substantial EDR values, and, in our experience, these spikes are more frequent (often several per

1 Via adaptable parameter initialization constant values.



flight) than the elevated turbulence events. In other words, failure to mitigate these could result

in a high (turbulence) false alarm rate.

The overarching idea for the quality control algorithm is that (generally) each value of

each data field is assigned a confidence (0 representing low and 1 representing high confidence)

and a QC flag (0/1 indicating no problem/problem found). Initially, confidences are set to 1 and

QC flags are set to 0. If a problem is identified with a value, then its confidence is lowered and

its QC flag is set. There are several ways that problems are identified. First, the flag from the

databus may be set, indicating that the data value, as indicated on the bus, is bad. Second, the

algorithm inputs and the (derived) vertical winds (w) are bounds checked. Third, derived

quantities such as w have an inherited confidence from their input fields. Fourth, V and w are

processed through the median QC algorithm described below. If the data is found to be

suspicious for the first two ways, the confidences are set to 0 and the QC flag is set to 1. The

third and fourth require further explanation.

There are several derived high-rate quantities in the algorithm: the averaged angle of

attack ( a ), the body-axis angle of attack ( b ), and w. The general paradigm is that if an input

field’s flag is set then the derived field’s is also and thus the initial QC flag for the derived field

is simply the logical OR of its inputs. The derived field’s confidence is initially set to the

minimum input confidence discounted by a factor. Namely, the derived confidence is

j i

i j

c f c

where ci are the enumerated input confidences, arg min ii

j c and f is the piecewise

linear function with vertices (0,0), (0.5,1), (1,1). a is a special case because it is nominally just

the average of two values that should be about the same. So if one is bad but the other is good,

then the good can be used and the bad ignored. So, for a , if both input confidences are above



threshold (0.5) then the confidence is the average of the two input confidences. If only one input

has an above threshold confidence, then the resultant confidence is the confidence of the good

one scaled by a penalty (nominally 0.8). If both inputs are below threshold then the confidence

for a is set to 0.

The median QC algorithm is designed to detect spikes and steps. There are three separate

components. The first component, the so-called median QC z-confidence, is similar to the

median absolute deviation (MAD) method discussed in Iglewicz and Hoaglin, 1993. First, the

5th

, 10th

, 15th

, 20th

, 25th

, 30th

, 50th

, 70th

, 75th

, 80th

, 85th

, 90th

, and 95th

percentiles are computed on

a sliding window of data. Currently the window for 8 Hz data is 31. The value to be quality

controlled is always the center value. Values in the window that have below threshold

confidence, as input into the median QC algorithm, are not used to compute the percentiles. The

idea for z-confidence median QC confidence is to compute a robust version of the z-statistic.

Namely,

50

n

x xz

R

(1)

where 100n n nR P P , and Pn is the nth

percentile. z is then input into a piece-wise linear

confidence map (vertices may in general depend on the field, but the ones used for V and w are

(0,1), (1.21,1), 0,2.21), (∞,0)) to generate the so-called z-confidence. Robust statistics are used

to mitigate the effects of large outliers in the window that would significantly bias the mean and

standard deviation. The choice of n is guided by the balance of two competing issues. If n is

chosen to be small (say 5), then most points (nominally, all but the highest and lowest two) in the

window are used to determine the percentile ranges. Thus, if there is a short, high intensity

turbulence event within the window, the value of Rn will be larger, ensuring that z is small. This



is the desired effect in the case that the point being evaluated (center) is part of the turbulence.

On the other hand, if a larger number of points are spikes or constitute a step, then the small

value of z is deleterious. In other words, small values of n have fewer false alarms for spikes, but

also a lower detection rate. Contrariwise, larger values of n have more false alarms for

spikes/steps with a higher detection rate. To improve the skill, instead of picking one value, Rn is

computed for n at 5, 10, 15, 20, 25, and 30. The adjacent ratios are computed (namely R5/R10,

R10/R15, R15/R20, R20/R25, and R25/R30). The condition 5 /n n RR R T (TR is nominally 4.5) is

checked for 10,1 ,35 0n . If the condition is true for any n, then the Rn for the largest such n

is used to compute z. If all ratios are below threshold, then R5, is used. This procedure will tend

to eliminate spikes and steps from the calculation of Rn. This principal behind this is that in the

case of good winds data these percentile-ranges decrease more slowly than the case of the

existence of data quality problems. This will not work if there are a very large number of

spikes/steps or if the data gradually transitions from nominal into a step. However, these cases

appear to be quite rare.

Note that the thresholds for the piecewise-linear confidence map are the same even if R5

is not used. This is done for simplicity, but it has the benefit that if a different Rn is picked then

R5, the result is a generally larger z, and hence, a lower the resultant confidence.

Additionally, two more confidences are computed and combined with to arrive at the

final median QC confidence. The first, the so-called percentile-range confidence, is computed

by applying a piecewise-linear confidence map to Rn, with larger values of Rn associated with

low confidences. This helps guard against the case where the z-confidence is low because there

are a large number of spikes or a long step that results in a large Rn. The second, the so-called

percent-good confidence, is computed by applying a piecewise-linear confidence map to the



percentage of points in the window with higher than threshold (nominally 0.5) confidence as

determined before applying the median QC algorithm. If this was done afterwards then the

algorithm would frequently get stuck in a state determining all data as bad. The final median QC

confidence is simply the product of the three intermediate confidences (z-, percentile-range, and

percent-good confidences).

After the initial input bounds check and also after the median QC algorithm, data with

confidence below the confidence threshold is linearly interpolated over as long as the number of

“bad” data points in a row is 3 or less. The confidences at those places are also linearly

interpolated. The flags, however, are left set to 1.

Vertical Wind Calculation

To compute the vertical winds from the inputs, the body-axis angle of attack ( b ) need to

be computed. First, if available, the left and right angle of attack ( 1 and 2 ) are averaged ( a

). If one of them has been found to be bad (below threshold confidence), then the other is used.

If both are bad, then it is flagged as bad and the confidence is set to 0. The body-axis angle of

attack is computed by 1 0b aa a .

Determining, and 1a is done by considering the following. If the aircraft is not changing

altitude and is flying under nominally smooth conditions, then, on average, b . The vertical

wind will cause this to not be instantaneously true. So given several flight hours of data,

preferably including aircraft flying on different days and different wind conditions, a least-

squares linear fit is performed for the following model 1 0ca a . Care should be taken to

limit the data to straight and level flight under smooth conditions. This is accomplished by



requiring: 30,000h ft , 10.5h ms , 1 .025wg (where gw is the vertical accelerometer in

g’s), and 1 .

Finally, the vertical winds are computed from V, b , θ, , and h (Parks, et. al., 1985;

Cornman, et.al., 2004):

sin cos cocos s sinb bw V h (2)

Note that there could be a pitch rate correction factor, like that in Lenschow (1972), namely

c180

osM

, where M is the sensor moment arm in meters, but in practice this term is

negligible and leads to nearly identical EDR estimates, under most circumstances. The

computation of w is performed at the common sampling rate (fs) for all of the data fields.

Maximum Likelihood Estimation

The spectral-domain based maximum likelihood estimation method of estimating ε is

described in Smalikho (1997). The basic calculation is to divide, frequency bin by frequency

bin, the empirical wind power spectrum by the theoretical power spectrum, where

2/3 1 m4/3

s-2

, over a certain range of frequencies, and then average the ratios to compute 2/3̂ .

To compute the empirical wind power spectrum, the data are first linearly de-trended,

using a least-squares fit.

21 0floord mk kw w k (3)

for , 10, mk , where 10 sm f (corresponding to 10 s of wind data). 1 0, and are linear

and constant (resp.) coefficients from the least-squares best fit, computed using the normal

equations, with x-values 2floor mk for , 10, mk , and y-values kw . Note that much of



the computations in the normal equations rely only on the x-values, which are fixed, and thus can

be precomputed.

Next, a power normalized Tukey (tapered cosine; Harris, 1978) window is applied. The

m-point Tukey window can be formulated as

12 1

1

112

1

1

1

cos 0

2

2

cos 1

kM

m k

M

k

k M

k

m M

M m M

k m

(4)

where floor 0.1 0.2M m . In the more traditional formulation of the Tukey window, this

corresponds to a taper factor of about 0.2. The power normalized Tukey window is

1

21

0

kk m

m j

j

(5)

The linear de-trended and windowed vertical wind time-series is

k

dW

k

d

kw w (6)

Linear de-trending and the window function application mitigate spectral leakage, though

at the cost of corrupting the wind spectrum at lower frequencies. The Fourier transform is now

applied:

21

2 /

0

2ˆm

dW i

k j

js

jk mS w ef m

(7)

where ,50, sfk , and i is the complex imaginary number. Note that the kth

frequency

corresponds to /skf m Hz. This computation is typically done by via fast Fourier transform,

appropriately scaled and halved. Finally, 2/3̂ is computed by



2/3

ˆ1ˆ

1

h

l

k

k

k kh l k

S

k k S

(8)

where kl and kh are the lower and upper index bounds over which the average is taken. These

correspond to the frequencies fl and fh.

Several possible models could be used. The von Kármán spectrum (e.g. Frehlich, et. al.,

2001) would be a likely choice except that it does not take into account the finite-length, discrete

time sampling, or the window function. The model spectrum needs to take into account the

procedure for generating the empirical spectrum or else the result can be biased and will not be

truly a maximum likelihood method. The form chosen is as follows. First, the von Kármán

transverse autocorrelation function (Frehlich, et. al., 2001), given by

2 / /NN DG r LB r G r L (9)

is sampled at / skV f , where , 10, mk , L is a length scale (nominally 669 m; corresponds

to an integral length scale of 500 m), σ2 is the variance of the velocity component, and

2/3 1/3

1/3

2

1/ 3

xG x K x

(10)

4/3

2/31/32 1/ 3D

xG x K x

(11)

Here, Γ is the Gamma function and K is the modified Bessel function. Note that G and GD are

not technically defined for 0x , so the value used there is defined by their limit at 0, namely,

0 1G and 0 0DG . The relationship between σ2 and ε is given in Cornman, et. al. (1995)

as

2 2/3 1/2 2/3

1/ 39

55 5 / 6oA L

(12)



where 1.6A (A has been estimated to be between 1.5 to 1.7; Kristensen and Lenschow 1987;

Mann 1994; Cornman et al. 1995). Next, the biased autocorrelation function for the

(normalized) Tukey window, , must be computed:

1

0

1 m k

k j j k

j

Tm

(13)

Finally, the average periodogram from the model spectrum can be computed (Frehlich, et. al.,

2001):

1

2 /

0

2Re / 0

mva ijk m

k j NN s NN

js

S T B jV f e Bf

(14)

Sva

is the exact form of the spectrum model given that a frozen von Kármán wind field is

sampled m times with a / sV f m spacing, including that a window, , is applied to the data

before the power spectrum is computed. Note that the only aircraft dependent parameter for vaS

is V. So, in practice, a 2-D table (frequency index vs. V) of vaS for different true airspeeds can

be computed offline and simply stored in the software. We compute and store for

65,70 26, , 5V m s-1

, which is sufficient for the aircraft used to date. Some newer aircraft

may need to have the upper range extended. At run-time, the vaS used is taken from the median

V over the same time interval as the vertical winds, and interpolating the table.

Four issues are, thus far, unaccounted for: the unknown true length scale, the linear de-

trending of the data, the noise signal, and the fact that the vertical winds are typically computed

from filtered fields. The first three are all handled the same way, namely by use of the cutoffs kl

and kh in eqn (8). fl and fh, the frequencies corresponding to indices kl and kh (resp.) are set to 0.5

and 3.5 Hz for the 8 Hz implementations. The effects due to the unknown outer length scale and



the linear-detrending of the data are largely limited to the lowest frequencies. Noise mostly

affects the higher frequencies since the turbulence spectrum has the lowest power there.

The aircraft fields used to compute the winds are typically filtered in the analog-to-digital

(A/D) conversion. In the case of commercial aircraft the details of the filtering are often hard to

obtain. Additionally, w is a non-linear function of the input fields (see eqn. (2)), though most of

the time, for small angle fluctuations, it is approximately linear. It is important to know the

model as well as possible to calibrate (i.e. minimize bias) the algorithm.

Also, it is generally desirable to have one piece of software that can be deployed on

different aircraft types; failure to do so drives up the cost of deployment. Avionics boxes, like

the aircraft monitoring system (ACMS) have limited processing capabilities and thus directly

computing the wind spectrum model there has not been feasible. Even if it were feasible, there is

still the issue that one might encounter an aircraft that employs different filtering techniques that

are not captured by the implementation. Passing a table of model values via an initialization

routine, was also considered but ultimately rejected because of the risk that not all systems may

be able to pass approximately 1600 (spectra for about 40 values of V each with about 40

frequency bins) values to the insitu library. To address these issues, we developed a so-called

pseudo-maximum likelihood, i.e. the wrong model is used (e.g. Gong and Samaniego, 1981;

Gourieroux, et. al., 1984). Instead of using the tailored filtered, aliased, von Kármán model, the

unfiltered aliased, von Kármán model is used. This will lead to a biased estimate of ε2/3

.

However, the bias due to using a different model is fixed. Namely, the relative bias can be

shown to be given by:

1

1

h

l

k vaf

k

vak kh l k

S

k k S (15)



The correction factor, γ, is set to the inverse of (15). Assuming the independence of the errors of

ˆkS for different k, the variance can thus be shown to be increased by the multiplicative factor:

22

1

h

l

k vaf

k

vak kh l k

S

k k S

(16)

Thus, if the exact correct model is known, the bias induced from using Sva

can be

corrected, at the cost having an increased variance. This allows the software to always use the

same model, namely Sva

, regardless of the actual correct model. Note that if the two models

differ greatly, the increase in variance can be significant. One would expect the significant

deviations between Sva

and Svaf

to be at the higher frequencies since A/D filters are low-pass, and

thus fh can be adjusted lower to exclude these frequencies. See Table 2 for the correction factor

and associated variance increase for an example scenario

fs 8 Hz

V 250 m s-1

Li 500 m

Filter Butterworth

Filter order 2

Stop-band cutoff 3 Hz

fl 0.5 Hz

fh 3.5 Hz

γ 1.73

Variance Increase

after correction

1.19

Table 2: Example bias and variance from using Sva ranther than S

vaf for the approximated filter from a Boeing 767-

300/400 with V=250 m s-1.

One issue with this is that tweaking the filter parameters slightly can lead to fairly

different correction factors, and the “true” filter settings are not known. To address this we note



that the 2-pole Butterworth with 3 Hz stop-band cutoff would have almost no difference (1.02) in

correction factor to that of the 4-pole if the frequency bands is, say, 0.5 to 1.5 Hz. Thus, the

following procedure can be performed to arrive at the correction factor. Over a large dataset of

aircraft data with some turbulence events, compute 2/3ˆvaf ( 2/3̂ based on vafS ) with no correction

factor, using an approximate filter, but only over a limited band of frequencies in the pass-band,

say 0.5 to 1.5 Hz. Second, compute 2/3ˆva ( 2/3̂ based on vaS ) over the full band of desired

frequencies (e.g. 0.5 Hz to 3.5 Hz for 8sf Hz). Lastly, compute the slope of the errors-in-

variables linear regression (Parks, et. al., 1989). This should approximate γ. In our case, we

obtained correction factors between 1.61 and 1.76 for various settings of lower bounds for the

data used in the fit as well as estimated variances. This range agrees nicely with the correction

factor 1.73 determined in Table 2.

Finally, the (corrected, if necessary) 2/3̂ is computed every 5 s. The confidence is the

minimum of two confidences: the average of the input w confidences, and the so-called, number-

of-good-inputs confidence. The latter is computed by counting the number times the flag for w

is set. Since w is derived from all of the inputs, if any of the inputs’ flags are set, then the

derived w flag will be set. This count will be a number between 0 and 10fs. A piece-wise linear

function is applied to this count with vertices (0,1), ( 0.0625·10 sf ,1), ( 0.1250·10 sf ,0), and

(10fs,0) to obtain the confidence. Then, once per minute, the peak and confidence weighted are

computed. The confidence weighted mean is computed using only those 2/3̂ with above

confidence threshold (nominally 0.5). The confidence of the mean is simply the average of the

input confidences where confidences below threshold are set to 0. The peak is the largest 2/3̂

above confidence threshold, and its corresponding confidence is simply the confidence of that



particular 2/3̂ discounted by 1/12 (since there are 12

2/3̂ measurements per minute) for every

2/3̂ that was greater but had a below confidence threshold. The square root of the mean and

peak are reported along with their associated confidences.

Data Sampling Rates

Care must be taken with respect to the data rates. Ideally, all input fields should be

sampled at the rate fs with a proper de-aliasing filter applied. Thus if t was sampled at 2fs, proper

resampling techniques should be used (e.g. interpolate to 10fs, apply de-aliasing (low-pass) filter

with stop-band cutoff at 0.4 sf , and then downsample to fs). Alternatively, having one or more of

the fields available at a data rate less than fs, can have deleterious effects on the spectral shape of

the winds. Practically speaking, the critical fields appear to be V, and b as they contribute the

most to the spectral content of the wind spectrum, though a rigorous analysis of this needs to be

performed. Recommended minimum data sampling rates are provided in Table 1.

Event Based Triggering Logic

Ideally, all once per minute turbulence measurements would be downlinked, but, using

current standard air-to-ground communications (e.g. VHF, SATCOM), this would result in

substantial costs to downlink mostly null turbulence (smooth) reports. Some null turbulence

reports are helpful. For example, a dispatcher can see where aircraft are experiencing smooth

conditions and possibly route other aircraft into that area. Additionally, forecasting models need

both positive (turbulence) and negative (smooth) events to be skillful. To try to balance these

needs, a report triggering logic was developed and implemented to decide which and when EDR

measurements should be reported.

2/3̂



The reporting scheme augments routine reports (sometimes referred to as heartbeats),

spaced 15 to 20 minutes and consisting of just one mean/peak EDR measurement, with

turbulence event-based triggering logic.

Turbulence Event Reports

There are 3 types of minimum events.

Type 1: Higher intensity experienced.

Type 2: Fairly consistent medium intensity experienced over the last 6 minutes.

Type 3: Consistent lower intensity experienced over the last 6 minutes.

Figure 1: Cartoon illustrating the three different types of events that would trigger a downlink report.

More formally:

Type 1: peak EDR ≥ T1 from the last minute. 6 minutes later a followup report is

also generated.

Type 2: peak EDR ≥ T2 for at least 3 out of the last 6 minutes. All 6

measurements are reported. 6 minutes later a followup report is also generated.



Type 3: mean EDR ≥ T3 for at least 4 out of the last 6 minutes. All 6 minutes are

reported. No followup report is reported.

Followup reports are generated after type 1 and 2 events so that users can see “both

sides” of the event (before and after). Otherwise, users would only see data before the event.

Followup reports always contain 6 measurements.

In all implementations thus far,

1

2

3

0.18

0.12

0.06

T

T

T

(17)

In our experience thus far, the bulk of the event triggered reports are Type 1 and their associated

followups, though the proportions will vary according to the turbulence climatology over the

route structure of the airline.

Further, the logic is designed to limit the possibility that mean/peak EDR measurements

are reported more than once. To see the detailed triggering logic, see the document In Situ

Turbulence – Event Reporting Algorithm at http://www.ral.ucar.edu/projects/in_situ_software.

Routine Reports

The goal for downlinkng routine reports is to provide some null turbulence data. Also,

some turbulence information can be inferred between reports (whether routine or event-based).

Namely, the turbulence was not strong enough to trigger and even-based report, and thus the

peak EDR had to be less than 0.18. Using this information, it is possible to fill in the gaps. The

times of these missing EDRs is easy to approximate since EDR measurements are spaced by 1

minute. The 3-D positions can be linearly interpolated, which generally should lead to good

results. It is possible to use the Aircraft Situational Display to Industry data stream to do a better

job.

http://www.ral.ucar.edu/projects/in_situ_software



The first routine report is generated immediately after the first mean/peak EDR is

computed, which ideally should be 1 minute after the take-off. Routine reports are generated

thereafter at a configurable but constant time interval, ideally less than 20 minutes. The only

time routine reports are not generated is if an event-based report is being generated at the same

time. If there are any unreported mean/peak EDRs at wheels-down, the last routine report is

generated and downlinked.



References

Cornman, L. B., C. S. Morse, and G. Cunning, 1995: Real-time estimation of atmospheric

turbulence severity from in-situ aircraft measurements. J. Aircraft, 32, 171-177.

Cornman, L. B., G. Meymaris and M. Limber, 2004: An update on the FAA Aviation weather

Research Program’s in situ turbulence measurement and reporting system. 11th Conf. of

Aviation, Range and Aerospace Meteorology, Hyannis, MA.

Harris, Fredric J., 1978: On the use of Windows for Harmonic Analysis with the Discrete Fourier

Transform. Proceedings of the IEEE 66 (1): 51-83.

Farrar, T. J., 2013: Integration of Real-time In Situ Turbulence Reports Into Airline Flight

Operations. 16th Conf. of Aviation, Range and Aerospace Meteorology, Austin, TX.

Frehlich R. J., 2000: Simulation of coherent Doppler lidar performance for space-based

platforms. J. Appl. Meteor., 39, 245–262.

Frehlich R. J., L. Cornman, and R. Sharman, 2001: Simulation of three dimensional turbulent

velocity fields. J. Appl. Meteor., 40, 246–258.

Gong, G., F. J. Samaniego, 1981: Pseudo Maximum Likelihood Estimation: Theory and

Applications. The Annals of Statistics , 9 (4), 861-869

Gourieroux, C; A. Monfort; and A. Trognon, 1984: Pseudo Maximum Likelihood Methods:

Theory. Econometrica, 52, 681-700.

Iglewicz, B., D. Hoaglin, 1993. How to detect and handle outliers. ASQC Quality Press [got this

reference from a thesis paper – have not actually seen this book].

Lane, T. P., R. D. Sharman, S. B. Trier, R. G. Fovell, and J. K. Williams, 2012. Recent advances

in the understanding of near-cloud turbulence. Bull. Amer. Meteor. Soc, 93, 499-515.

Lenschow, D.H. 1972: The measurement of air velocity and temperature using the NCAR

Buffalo Aircraft Measuring System. NCAR Tech. Note NCAR/TN-74+EDD, pp. 39.

Parks, E. K., R. C. Wingrove, R. E. Bach, and R. S. Mehta, 1985: Identification of Vortex-

Induced Clear Air Turbulence Using Airlines Flight Records. J. Aircr., 22 (2), 124-129.

Press, William H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, 1989: Numerical

Recipes: The Art of Scientific Computing. Cambridge: Cambridge Univ. Press, 666-670.

Smalikho, I. N., 1997: Accuracy of turbulent energy dissipation rate estimation from wind

velocity temporal spectrum. Atmos. Oceanic Opt., 10, 898–904.

Documents

NCAR In Situ Vertical winds-based EDR Estimation Algorithm ... · This system estimates values of the cube rate of the eddy dissipation rate, referred to as the EDR, an atmospheric