RESEARCH ARTI CLE

A note on wind-tunnel turbulence measurements with DPIV

G. R. Spedding A. Hedenstrom L. C. Johansson
Received: 23 January 2008 / Revised: 3 October 2008 / Accepted: 7 October 2008 / Published online: 26 October 2008
Springer-Verlag 2008
Abstract At moderate Reynolds numbers (10
4
B Re B
10
5
), the performance of lifting surfaces is strongly affec-
ted by the potential for laminar boundary layer separation
and subsequent reattachment and the use of high-quality,
low-turbulence wind tunnels is essential in characterising
ight at comparatively small scales (where the wing chord
may be from 1 to 5 cm in length) and low speeds (on the
order of 10 m/s). Measurement of the existing turbulence
levels in such facilities is hard and has not been achieved
using DPIV methods due to the relatively small bandwidth
of measurable velocities. A series of experiments is
reported here where DPIV sampling parameters are driven
beyond their normal range in an attempt to measure tur-
bulence levels in a low turbulence wind tunnel. The results
show that DPIV can measure the background turbulence,
and therefore its instantaneous structure. The measure-
ments also reveal certain challenges in investigating the
aerodynamic performance of small-scale ying devices.
1 Introduction
The purpose of a wind tunnel is to generate a uniform,
steady mean ow, U, over a small part of it, known as the
test section. Ideally, the ow should not vary in any of the
spatial directions {x, y, z} (streamwise, spanwise, and
vertical) and should not vary with time. Wind tunnels are
constructed so that departures from this ideal are small.
The departures that do exist are commonly termed turbu-
lence, though they are not simply related to the usual uid
mechanical denition of turbulence (Tennekes and Lumley
1972). Small, or even moderate levels of turbulence (we
will retain the traditional wind tunnel usage of this term for
the moment) need not be detrimental; one can argue that
the higher the turbulence, the higher the effective Reynolds
number of the wind tunnel ow (see Barlow et al. 1999,
p. 227). However, this connection is not very straight-for-
ward, even if the so-called turbulence characteristics do
model the full-scale application correctly.
Because wind tunnels are engineered to produce very
small variations about the steady mean U, it is hard to
measure the remaining uctuations that do exist. Two
methods are in common use. The rst uses the known sharp
reduction in drag around a sphere at some transition Rey-
nolds number (see Barlow et al. 1999). This transition
Reynolds number is a function of ambient turbulence, and so
a wind tunnel can be characterised by a turbulence factor,
TF =
385000
R
obs
; (1)
where R
obs
is the observed transition Reynolds number, and
the number 385000 is the Reynolds number at which
transition occurs in turbulence-free conditions (as measured
fromatmospheric ight tests). More precise measurement of
turbulence properties can be accomplished using hot-wire
G. R. Spedding
Department of Aerospace and Mechanical Engineering,
University of Southern California,
Los Angeles, CA 90089, USA
Present Address:
G. R. Spedding (&)
Department of Mechanical and Aeronautical Engineering,
University of Pretoria, Pretoria 0002, South Africa
e-mail: geoff.spedding@up.ac.za
A. Hedenstrom L. C. Johansson
Department of Theoretical Ecology,
Lund University, Lund, Sweden
e-mail: anders.hedenstrom@teorekol.lu.se
L. C. Johansson
e-mail: christoffer.johansson@teorekol.lu.se
1 3
Exp Fluids (2009) 46:527537
DOI 10.1007/s00348-008-0578-1
anemometry. Operated in strong overheat, the hot-wire set-
up can be tuned for maximum sensitivity to uctuations
about the mean at any given speed, and this kind of method is
used for more serious descriptions. Dryden et al. (1937) and
Dryden and Kuethe (1929) remain standard references for
wind tunnel turbulence measurements, where hot-wire and
turbulence sphere measurements are compared. Hot-wire
rakes and traverse systems allow maps of average
uctuating velocity components to be measured.
Nevertheless, it is still common practice to characterise a
wind tunnel with one single number,
percent turbulence =
q
U
100; (2)
where q is either the root mean square of the uctuating
streamwise velocity component, u - U, or the rms of all
measurable uctuating components in
q = (u U)
2
v
2
w
2
1
2
; (3)
averaged over some time interval and over some spatial
region.
The signicance of the magnitude of the uctuations, q,
depends on the characteristics of the ow. If the length and
time scales of interest in the test system differ greatly from
those of the remaining disturbances in the tunnel then the
latter might be unimportant. One ow conguration that is
notoriously sensitive to residual uctuations is where ini-
tially laminar boundary layers approach conditions for
separation. This condition is found in low-speed airfoils,
whose performance is exquisitely sensitive to details of
separation and possible reattachment (Lissaman 1983).
In the Reynolds number range Re
c
: Uc/m = [10
4
10
5
],
where c is a streamwise chord length and m is the kinematic
viscosity, variations in measured drag can be a factor of 2
different in different facilities (see Lyon et al. 1998, p. 222,
for examples). These differences are usually attributed to
different behaviour of separation bubbles due to different
ambient turbulence characteristics. In any study of airfoils
or wings over this range of Re, a correct description of the
wind tunnel uctuations is clearly necessary. Nevertheless,
the descriptions that are given rarely stray beyond
the single number specication in Eq. 2 (cf. Selig and
McGranahan 2004).
Digital Particle Image Velocimetry (DPIV) methods are
ubiquitous in experimental uid mechanics, but cannot
easily be used to measure wind tunnel turbulence because
the measurement bandwidth is so low. If the largest mean
particle image displacement on the sensor is 5 pixels, and
the smallest resolvable displacement is 1/20 pixel, then the
bandwidth is 100. Fluctuating velocity magnitudes in high-
quality wind tunnels are very low, relative to the mean
speed, with q/U B 0.05%, which is 20 times smaller than the
smallest resolvable DPIV measurement under standard
operating procedures. Even with some care, reasonable
turbulence measurements are elusive, and Spedding et al.
(2003a) reported how a systematic decrease in observed
turbulence level with increasing DPIV delay time, dt,
showed that no reliable turbulence measurement could be
reached. Note that the measurement of the residual, low-
level turbulence in a clean wind tunnel facility poses a
slightly different set of problems than commonly addressed
in application of PIV to larger amplitude turbulence prob-
lems (e.g. Fincham and Spedding 1997; Adrian et al. 2000;
Scarano 2003; Agrawal 2005; Adeyinka and Naterer 2007;
Tanaka and Eaton 2007; Lavoie et al. 2007), although the
common denominator remains the low bandwidth of PIV
methods. The comparatively low velocity bandwidth
(compared with a hot-wire, for example, or with a numerical
simulation) is a shared characteristic of all PIV algorithms
and so the results from this paper are also quite general.
This paper reports on results of an investigation into the
feasibility of making wind tunnel turbulence measurements
with DPIV methods. It uses an analogue of the hot-wire
overheat principle by using very large values of dt that are
compensated for by measuring only uctuating displace-
ments and not mean quantities. This is possible because the
ow is otherwise extremely uniform. The instantaneous
spatial structure of wind tunnel turbulence can be shown,
and some important operating principles of wind tunnels in
nominally low-turbulence regimes are found.
2 Materials and methods
2.1 A low-turbulence wind tunnel for bird ight
The wind tunnel used at Lund (Fig. 1) is a standard, closed-
circuit, recirculating conguration, with two special mod-
ications that make it especially valuable for research on
bird ight. The entire tunnel can be tilted at any angle from
?8 (nose down in the test section) to -6, allowing
experiments on climbing or descending ight, and most
important, on steady gliding ight. The tunnel also has an
open section just aft of the test section. This makes it very
practical for inserting and removing live animals (or xed
models), and it also removes streamwise pressure gradients
there since the ow is open to atmospheric pressure. One
further property makes it well-suited for bird/bat ight
studies, which is the low turbulence level over most of the
test section. The tunnel turbulence properties were care-
fully measured by single and x-wire hot wire anemometry
and streamwise and cross-stream plane maps were made
of the magnitude of the average uctuating velocity.
These properties are documented in Pennycuick et al.
(1997). The average turbulence levels, q/U, were approx-
imately 0.035%, which is a satisfactorily low number, so
528 Exp Fluids (2009) 46:527537
1 3
that results from the tunnel can be reliable and compared
with other top quality research wind tunnels. Recall that
this is particularly important for 10
4
B Re
c
B 10
5
, which is
exactly the range in which small birds and bats operate.
Extensive studies have been made on live, trained (but
otherwise unrestrained) birds and bats in this tunnel (e.g.
Pennycuick et al. 2000; Park et al. 2001; Rosen and
Hedenstrom 2001; Spedding et al. 2003b; Hedenstrom
et al. 2006; Muijres et al. 2008; Hedenstrom and Spedding
2008; Henningsson et al. 2008), the latter publications
using custom DPIV methods on a two-camera (or equiva-
lent) system that removed the mean ow and allowed the
DPIV bandwidth to be concentrated on disturbance quan-
tities. In spite of this, turbulence measurements were not
successful when DPIV exposures were set close to appro-
priate settings for the bird wake studies. Although the wake
disturbances could be shown to be signicantly above
measurement noise, the background turbulence was not.
2.2 Experimental procedure
The ow was seeded by a commercial fog generator (JEM
ZR12-AL) placed inside the rst diffuser, about 1 m
downstream of the bellmouth and covered by a fairing to
deect the ow smoothly. The ability of an individual
smoke particle with diameter, d & 1 lm, and density, q, to
follow a ow with characteristic length and velocity scales
of L and U can be expressed through the Stokes number,
St =
qd
2
UC
c
18lL
;
where l is the uid viscosity, and C
c
is a slip correction
factor (Hinds 1999) that has a value close to one for such
particle diameters. If St 1, then particle inertia is
small relative to variations in the ow streamlines. For
U = 10 m/s and L = 2 cm (as suggested by results to
appear in the following sections), then St & 2 9 10
-3
, and
so the criterion is satised and the smoke particles can be
assumed to follow the air ow.
The wind tunnel was run for at least one hour before any
experiment (to remove both temperature and velocity
uctuations) and then lled with smoke over a period of
about 30 min. During experiments, 2030 s bursts of
smoke were introduced at regular intervals to maintain the
particle density (most particularly particles that generate
large, bright images; such particles easily meet the 1.5
pixel minimum diameter required for a reasonable sub-
pixel estimator to work.). The times between these smoke
rells and each experiment were kept constant. The parti-
cles were illuminated by pulses from a dual-head Nd:YAG
laser (Spectra-Physics PIVII) and imaged onto a Redlake
Megaplus II ES4020 12-bit CCD sensor (2,048
2
pixels)
with Nikkor 50 mm f1.416 lens. Streams of image pairs
were saved directly on a SCSI RAID disk array through an
IO Industries DVR Express CL160 interface, controlled by
Video Savant software from the same company. Image
sequences were saved uncompressed, and analysed using
the Correlation Imaging Velocimetry (CIV) algorithms
described in Fincham and Spedding (1997) and Fincham
and Delerce (2000). Correlation boxes were typically 30
pixels on a side, corresponding to 5 mm in the ow. This is
comparable to the laser sheet thickness of 34 mm. In the
CIV algorithms the correlation box size is unrelated to the
search domain size, which is modied by adding a large
pre-computed shift as described in the following section.
2.3 Measuring low turbulence levels with DPIV
The basic idea of the original shifted 2-camera system is
the same, in principle, as performing a compensating mean
Fig. 1 The wind tunnel
conguration includes a pivot
about which the wind tunnel can
rotate, and an open test section
where objects such as live birds
can easily be inserted and
removed
Exp Fluids (2009) 46:527537 529
1 3
shift in cross-correlation box for the second image in a pair
(the 2-camera system loses no pixels, and is signicantly
cheaper and more exible). The principle is shown in
Fig. 2. Adding a constant shift in one or more directions to
remove a mean ow component in DPIV calculations is a
standard option in most commercial codes, and it is a built-
in component of the CIV algorithms used here. We now
investigate the behaviour as dt is allowed to become very
long, so that the compensating mean ow pixel displace-
ments can be a signicant fraction of the total number of
pixels in the mean ow direction. Such a procedure can
work, provided the light sheet is extremely carefully
aligned with the mean ow, and provided the out-of-plane
velocity component is small by comparison. These condi-
tions can be met in a low-turbulence wind tunnel.
2.4 Analysis and notation
Estimates of the in-plane velocity components are made
either through a tting of a spline-tted autocorrelation
function to the computed cross-correlation function, or
from a direct functional tting of splined particle image
windows. The displacement vector is located half way
between the geometric centre of the initial image correla-
tion window and its nal estimated destination. The raw
velocity eld is interpolated back onto a regular rectangular
grid through a smoothing thin-shell spline and gradient
quantities are computed analytically from the spline coef-
cients themselves. Although errors in the velocity
estimates are not constant fractions of the displacements,
the most likely uncertainty in velocity estimates is not
likely to exceed 1% and for the gradient quantities the
gure is 5%. Fincham and Spedding 1997 and Fincham and
Delerce (2000) may be consulted for further details. Note
that although the CIV algorithms can be shown to be
among the best available (e.g. Piirto et al. 2005), none of
the results herein depend on details of the method. They all
ultimately stem from the limited resolution at small scales
that is common to all PIV methods.
The disturbance velocity, u, is described in terms of its
components {u, v, w} in directions {x, y, z}:
u = ~ u

U
v = ~ v
w = ~ w
9
>
=
>
;
; (4)
where the tildes denote the original, measured quantities
and

U is the steady, time-averaged mean speed over the
whole observation window, and over each image sequence.

U is thus calculated from the prescribed integer pixel shift

in the CIV calculation, plus the mean of any fractional part
computed in CIV. Measurements are made over a vertical
plane, aligned in the streamwise direction, with size
DX 9 DZ. The rms quantities for one single velocity eld
estimate are
u
/
=

1
mn
X
m
i=1
X
n
j=1
u
2
(i; j)
v
u
u
t
w
/
=

1
mn
X
m
i=1
X
n
j=1
w
2
(i; j)
v
u
u
t
9
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
;
; (5)
where the indices i, j cover the rectangular array of
measurements in x and z, respectively. u
/
and w
/
are
therefore spatial averages (also averaged over the effective
exposure time, dt). The sampling rate for successive
velocity eld estimates is 10 Hz, dictated by the pulse-
pair repetition rate of the laser. When time averages are
made over a number of frames, they are denoted u x; z ( ) and
w x; z ( ) for the spatial maps,
u(x; z) =
1
N
X
N
i=1
u
i
(x; z)
w(x; z) =
1
N
X
N
i=1
w
i
(x; z)
9
>
>
>
>
=
>
>
>
>
;
; (6)
where the u and w components are averaged locally at each
grid location over a time series of length N so u and w are
the temporal deviations from

U at each interrogation point.
If the uctuations in u and w are normally distributed about
a stationary

U then as N ? ?, u
/
; w
/
0: It will also
prove convenient to dene spatially averaged values for
U(t) and W(t) at each timestep:
U(t) =
1
mn
X
m
i=1
X
n
j=1
u(i; j)
W(t) =
1
mn
X
m
i=1
X
n
j=1
w(i; j)
9
>
>
>
>
=
>
>
>
>
;
: (7)
b1
b1 b1
Ut
ut
b2
Fig. 2 a (left) and b (right) are sampled with a time interval, dt,
between them. The correlation box b1 in a is rst moved by a pixel
distance Udt.s, where s is a constant in units of pixels/cm. The new
location of b1 is b1
/
, and it is from here that displacements are
calculated by cross-correlation with candidate locations in a local
neighbourhood. Here box b2 is shown as the most likely displacement
location in b. dt can now be expanded so that pixel distances udt.s are
optimised
530 Exp Fluids (2009) 46:527537
1 3
In all experiments reported here, the wind tunnel speed
was set at a nominal value of 9.37 m/s, which is close to
the cruising speed of the small birds studied in the Lund
wind tunnel.
3 Results
3.1 Correction of lens distortion effects
In choosing large dt values, measurement of u
/
and w
/
become possible, but so do measurement errors of certain
kinds. In particular, the particle images now move from
one part of the camera lens to a signicantly distant
part, and the difference in lens distortion characteristics
increases in signicance with increasing dt. This can be
seen in Fig. 3, which shows contours of measured u and w;
averaged over 10 velocity elds captured during 1 s. The
contour patterns are coherent, and have magnitudes that
exceed the likely value of the turbulence levels. They are
characteristic of displacement across a pincushion pattern
lens distortion, and have the same shape as those found by
Spedding et al. (2003a) and subsequently used as correc-
tion surfaces. The magnitude of these distortion patterns
depends on dt, and here dt = 2,000 ls, a comparatively
large value used to illustrate the phenomenon. However,
anticipating use of such large values in these experiments,
it is clearly necessary to compensate for this systematic
error. Thus, for each dt, we dene a mean residual eld,
calculated from the time-average of all frames in the data
set (from 10 to 100),
u
r
; w
r
= u; w; (8)
and then the best estimate of the uctuating velocities is
u = ~ u U u
r
w = ~ w w
r
)
: (9)
A typical point-by-point convergence rate for the
residual eld is 0.72 after 10 timesteps, 0.85 after 20 steps
and 0.96 after 50 steps. The mean residual eld can be
tted by various order polynomial and spline surfaces, and
these smoother functions can be used as a better estimate of
the likely true residual eld, but doing so has no effect on
the measured statistics and we prefer to retain the com-
puted mean eld as the most direct approach.
3.2 Long DPIV exposures for turbulence measurement
The effect of varying dt on measured turbulence intensities
is shown in Fig. 4. The apparent turbulence intensity
magnitude is a strong function of dt. If the measurement
were actually of turbulence, then its magnitude would be
independent of dt. Instead, the result is an artefact of the
nite resolution in measuring small displacements. Since
Fig. 3 Contours of time-averaged u (top) and w (bottom). Contour
values are given in %U. The DPIV exposure time, dt = 2,000 ls
1000 1200 1400 1600 1800 2000
0.035
0.040
0.045
0.050
0.055
0.060
0.065
0.070
0.075
0.080
(
u
'
/
U
)
o
b
s
,


(
w
'
/
U
)
o
b
s

(
%
)
t (s)
u'
w'
Fig. 4 Observed turbulence intensity in streamwise (u
/
, circles) and
vertical components (w
/
, triangles), as a function of DPIV exposure
time, dt. Each case is composed of a sequence of experiments with
increasing dt (arrows left to right) and then decreasing dt
Exp Fluids (2009) 46:527537 531
1 3
this algorithm error (which is usually dominated by some
form of peak-locking errorsee for example Fincham and
Spedding 1997; Westerweel 1997; Prasad et al. 1992) is
xed in pixels (we may call it e
a
), then as dt increases, so
does the particle image displacement D, and so the relative
error e
a
/D decreases. The spurious dependence of q/U on dt
can be used as a diagnostic of whether a DPIV method can
be said to be measuring the ow physics or primarily
discretisation errors.
The measurements show signs of convergence towards
the lower right corner of Fig. 4, and that value of 0.035% is
the same as previous hot-wire measurements in this wind
tunnel (Pennycuick et al. 1997). Note that dt values are
much larger than normally used during an experiment,
where typically 80 B dt B 200 ls. Figure 4 also shows
that the results are independent of the elapsed time during
the experiment and there is no hysteresis between the
increasing and decreasing dt parts. Although Fig. 4 shows
results that are related somehow to disturbance levels in the
uid, it is not clear that the values at the longest value of
dt = 2,000 ls (on the right side of Fig. 4) are not still
signicantly affected by measurement error.
Figure 5 reports the results of an experiment using
extremely long exposure times. Measured values of u
/
/U
and w
/
/U continue to decrease as dt rises to 4,000 ls, but
after dt = 2,500 ls, u
/
/U stabilises briey and then uc-
tuates with increasing amplitude with further increases in
dt. Beyond this point u
/
/U and w
/
/U behave differently.
While u
/
/U varies around a mean value of 0.035%its last
reliably measured valuew
/
/U stabilises between 0.02 and
0.025%. The reasons for this behaviour will be discussed
shortly, but for the moment, we note that, thus far, the best
estimate of the turbulence level for this tunnel, if indeed
it must be reduced to a single number, is approximately
q=U = u
/
=U ( )
2
2 w
/
=U ( )
2

1=2
= 0.046%, where we
assume that uctuations normal to the mean ow have the
same magnitude (v
/
& w
/
).
3.3 The structure of wind tunnel turbulence
If we are now measuring mostly true velocity uctuations,
and not measurement noise, then the disturbance velocity
elds, estimated over some window of {x, z} (which
shrinks with increasing dt) contain information about the
spatial structure. This structure (and the associated turbu-
lence measurement itself) will reect the true variations
provided it is correct to imagine that it has convected along
with the ow over a time dt and streamwise distance Udt.
Figure 6 shows the spanwise or cross-stream vorticity,
x
y
=
ow
ox

ou
oz
;
and the divergence in the plane of measurement,
D
y
=
ou
ox

ow
oz
:
The two quantities have approximately equal magnitude,
and show no preferred orientation.
Figure 7 shows the spectral distribution of kinetic
energy over the range that can be resolved by the DPIV
experiment. The energy is summed over circular shells of
constant magnitude wavenumber k [ [ = k
2
x
k
2
z

1=2
: The
result shown is an average over 100 velocity elds taken
over 10 s. Most energy is at moderate length scales, from
1 B |k| B 4.5, and at higher wavenumbers the kinetic
energy has dropped an order of magnitude from its peak
value. k = 4.5 rad/s corresponds to a length scale of about
1.5 cm. The magnitude and distribution of energy in the
uctuating velocity components is equal in the streamwise
and vertical direction. The uctuations at higher wave-
numbers for |k| [8 rad/s (or wavelengths \0.8 cm, which
is four times the grid spacing) are almost certainly affected
by the nite resolution of the CIV algorithm.
In this paper, the instantaneous spatial uctuations in the
otherwise uniform wind-tunnel ow are averaged for
measurements of u
/
and w
/
. Hot-wire data construct the
same measure from temporal uctuations at a single point
and the two methods are equivalent if the standard Taylor-
hypothesis is invoked. Temporal uctuations over time
scales on the order of seconds can be measured from long
sequences of DPIV images, and one such sequence is
shown in Fig. 8. This gure shows that excursions of the
mean streamwise velocity, U(t) (dened in Eq. 7), of the
order of 1 cm/s, or about 0.1% of the mean, can be reached
2000 3000 4000 5000 6000 7000 8000
0.02
0.03
0.04
0.05
0.06
0.07
(
u
'
/
U
)
o
b
s
,


(
w
'
/
U
)
o
b
s

(
%
)
t (s)
u'
w'
Fig. 5 Observed turbulence intensities (streamwise component, u
/
in
circles, vertical component, w
/
in triangles) as function of long dt. At
the arrow, the average u
/
measure stabilises, but with large amplitude
uctuations about that value
532 Exp Fluids (2009) 46:527537
1 3
in occasional bursts that have a timescale of between 3
and 4 s (the tunnel circulation time, T
C
& 30 s). There
are no corresponding uctuations in the mean vertical
velocity, W.
These variations in mean streamwise velocity could
originate from the tunnel fan speed controller, or they
could be due to variations in boundary layer thickness on
the tunnel walls. The 0.1% corrections are small by most
standards, but they can give a signicant non-zero mean
component to the instantaneous velocity elds if they are
calculated from a mean over the whole time sequence.
Indeed, this explains the increased and erratic uctuations
in u
/
/U for large dt in Fig. 5. The peak values are caused by
regions where the entire window has some small mean
value away from the global mean. We return to the
Fig. 6 Spanwise vorticity (left)
and divergence in the plane of
measurement (right). The eld
of view {DX, DY} =
14.5 9 17 cm. DX\DY
because the x-range is reduced
by the 131 pixel shift required to
remove the mean ow for
dt = 2,400 ls. The 20-step
colour bar is mapped to 1/s
for both elds
Fig. 7 Time-averaged energy spectra of the streamwise (left) and
vertical (right) velocity uctuations. The time-averaged spectra over
100 realisations are shown by the solid lines and banded by dashed
lines at 1r for each |k|. The thick straight lines have |k|
-5/3
slope
Fig. 8 Above: mean streamwise disturbance velocity, U(t) (solid line)
and vertical disturbance velocity, W(t) (dashed line) from spatial
averages at each timestep. Below: the corrected, uctuating wind
tunnel turbulence levels (u
//
/U is solid line, w
//
/U is dashed). The data
come from one sequence of 200 images, with dt = 2,500 ls
Exp Fluids (2009) 46:527537 533
1 3
question of whether such motions should be counted in
overall turbulence measures in the next section.
At each timestep, the local mean U(t) and W(t) (Eq. 7)
may be removed from the {u, w}(x, z) elds (Eq. 9),
u
//
(t) =

1
mn
X
m
i=1
X
n
j=1
u(i; j) U(t) ( )
2
v
u
u
t
w
//
(t) =

1
mn
X
m
i=1
X
n
j=1
w(i; j) W(t) ( )
2
v
u
u
t
9
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
;
; (10)
so that only spatial variation in the observation window
contributes to the total. The lower part of Fig. 8 shows
corrected time traces of u
//
/U and w
//
/U over the same time
interval. Both measures vary around a mean value of about
0.036%. u
//
/U is weakly correlated with |U|, but it is not
clear whether this variation is an artefact of the data
reduction. Accelerations in U, for example, could be
responsible for increased magnitude of u
//
, but they would
also contribute to an extra residual in the u(x, z, t) - U(t)
calculation.
If the same correction is applied to data in Fig. 5, the
result is Fig. 9, where the smoothness of u
//
and its
agreement with w
//
has improved signicantly. However,
once again, there is a point, at dt & 4,000 ls, where the
curves of u
//
/U and w
//
/U begin to separate and the variation
in the result increases. Now, the long disturbance dis-
placements (after the mean has been subtracted) of the
groups of image particles responsible for the measured {u
//
,
w
//
} convect those particles to slightly different parts of the
lens distortion correction surface (Fig. 3). This effect is
taken into account to leading order by applying the
correction of an averagely displaced group of particles, but
small departures from this mean, due to local velocity
gradients and accelerations can give spurious structure to
the pattern of {u
//
, w
//
}, and affect the mean values. These
spurious values will have some structure similar to the lens
distortion curves of Fig. 3, since they are caused by relative
motion on this surface. It is an effect that will always be
present, and it is amplied as dt increases. A careful
inspection of D
z
(x, z) in Fig. 6b shows that the divergence
eld has some small remnant of circular symmetry, and the
measured values are a projection onto this symmetry.
The best estimates of u
//
/U and w
//
/U now appear to lie
between 0.012 and 0.020%, which are signicantly below
the previous estimates derived from Fig. 5. If the uctua-
tions normal to U are again assumed to be isotropic, q/
U = 0.026%. Pennycuick et al. (1997) reported mean tur-
bulence values of 0.035% (ranging from 0.031 to 0.037)
over the same region of the same wind tunnel. Our lower
estimate which compensates for momentary accelerations
in the global ow, U(t), is q/U = 0.026 0.006%, while
the upper limit is q/U = 0.046 0.006%. Since the hot-
wire estimates can be affected by mechanical vibrations,
Pennycuick et al. noted that their measurements should be
regarded as upper bounds. This brings up a more general
question of how wind tunnel turbulence measurements can
be calculated and reported.
4 Discussion
4.1 The basic problem
Having worked through the examples in these experiments,
we may summarise the basic measurement problem using
convenient numbers that are similar to those used here.
Suppose the mean owspeed, U = 10 m/s and the tunnel
turbulence levels are given as 0.05%. Then u
/
= 0.5 cm/s. A
DPIV exposure time, dt = 2,000 ls, gives a displacement
due to turbulent motions of u
/
dt = 10
-3
cm. If a 2,048
2
pixel array covers a 20 9 20 cm observation window, the
scaling factor, s & 100 pixel/cm. A displacement of
10
-3
cm is thus equivalent to dpix = 0.1 pixel. If the min-
imum expected error in the estimated pixel displacement is
0.02 pixel, the required measurement is only ve times
above the noise. Alternatively, if we stipulate that the pixel
displacements due to turbulent uctuations must be ten
times the smallest measurable displacement, d(pix)
min
, then
a criterion for successful wind tunnel turbulence measure-
ment can be given as
10d pix ( )
min
_u
/
dt:s:
Given an expected turbulence level, u
/
, the experimenter
must select an appropriate combination of DPIV exposure
2000 3000 4000 5000 6000 7000 8000
0.01
0.02
0.03
0.04
(
u
'
'
/
U
)
o
b
s
,


(
w
'
'
/
U
)
o
b
s

(
%
)
t (s)
u'
w'
Fig. 9 Corrected turbulence intensities (streamwise component, u
//
in
circles, vertical component, w
//
in triangles) from Fig. 5. Plausible
values of the true turbulence level lie between the horizontal bars
534 Exp Fluids (2009) 46:527537
1 3
time, dt, and magnication, s, onto the pixel array. Neither
parameter can be increased without limit because errors
due to optical nonlinearities essentially multiply both
terms.
The required dt values of 2,0004,000 ls in the example
above are at least one order of magnitude above typical
DPIV settings for ow measurements in such circum-
stances (given the same assumptions on sensor geometry,
optical magnication and mean ow speeds), which cannot
measure wind tunnel turbulence. The better the ow quality
in the wind tunnel, the more acute the problem. Wind
tunnels with four times higher turbulent uctuation mag-
nitudes are not four times easier to measure because the
higher turbulence levels increase the likelihood of particles
leaving even a perfectly aligned light sheet during the large
required dt.
4.2 Reporting turbulence levels
A wing (or wing pair) of 5 cm chord ying at 10 m/s has a
ow transit time of 5 ms. Fluctuations in the mean ow
with periods of the order of 24 s therefore occur over at
least 400 transit times, and so might reasonably be regar-
ded as slow variations in U with little dynamical
signicance. Arguing thus, an appropriate estimate of the
wind tunnel turbulence levels is 0.026%. On the other
hand, if the control manoeuvres required by a gliding bird
over the course of 10 s are being studied, then the higher
value of 0.046%, which includes these longer-timescale
variations, is more correct. In reporting rms turbulence
values from hot-wire data, the mean value is always
removed from the measured signal. In various statistical
detrending operations, other long time period/low fre-
quency components might be removed also, and removing
the linear part of any global curve t is also common, for
example. The complete reporting of a wind tunnel turbu-
lence level should therefore specify the effective range of
time scales or frequencies over which it has been
calculated.
4.3 Turbulence estimates from DPIV
The instantaneous structure of the uctuating vorticity and
divergence elds is shown in Fig. 6 for this wind tunnel,
operating at one speed. The wavenumber range of energetic
scales is given in Fig. 7. Together, they show that the
uctuating velocity eld has no signicant coherent large
scale structure that is observable over this range of scales,
and that most turbulence energy is contained within a
wavenumber range of 1.56 rad/cm (the corresponding
wavelengths are 14.2 cm). The frame size DX = 14.5 cm
and the grid spacing is 0.25 cm so the range of measurable
wavenumbers in these gures is 0.4312.5 rad/cm.
The Taylor hypothesis is routinely invoked by those
making time-resolved turbulence measurements at a single
point, and if convected along by the mean ow past a xed
measurement point (such as a hot wire), the frozen ow of
Fig. 6 would appear as frequencies
f = kU=2p 64 Hz--1:86 kHz [ [: (11)
The most energetic wavenumbers of 1.5 B |k| B 6 rad/
cm lie between 220 and 890 Hz. In order to obtain these
data, DPIV exposure times, dt, of up to 4,000 ls were used.
Based on this effective exposure time alone, the Nyquist
frequency representing the highest accessible frequency,
f
Ny
= 1/(2dt) = 125 Hz. Figures 5 and 9 suggest that the
minimum usable dt = 2,500 ls, where f
Ny
is still 200 Hz.
The f
Ny
determined by dt is in a reference frame moving
with the mean ow, which is not the same as Eq. 11 above.
Whether it is sufcient to resolve dynamically important
time- and space-scales in the ow depends on the x(k)
relation for that ow. The basic point is that the larger the
value of dt, the lower the upper limit of resolvable fre-
quencies, and so dt cannot be increased without limit. Note
that while the upper limit in observable frequencies affects
the correct interpretation of the ow structure (in Fig. 6)
and spectral distribution (Fig. 7), it does not necessarily
affect the global (single number) statistics, as aliased fre-
quencies still appear in the measured signal without loss of
energy.
The estimated values of q/U = 0.026 and 0.046
0.006%, with and without low frequency correction, are
consistent with those previously measured in hot wire
experiments in the same wind tunnel, so the DPIV methods
for measuring them appear to be reasonable.
4.4 Experiments in nominally low turbulence wind
tunnels at transitional Reynolds number
Flows involving marginal stability of laminar boundary
layers are very susceptible to ambient turbulence, and to
small variations in model geometry and boundary condi-
tions. Such is the case for tests on small-scale ying
devices which may be human-engineered micro air-vehi-
cles or naturally evolved birds or bats. Even though neither
device normally ies through perfectly still air (if there
were such a thing), generalisable and repeatable test results
require that the background turbulence levels be low and
also that the turbulence characteristics be well-docu-
mented. In addition to global averages, the spatial and
temporal structure of the uctuations should be character-
ised. This paper shows that DPIV methods can be used to
measure the tunnel properties, given suitable operating
conditions and particularly careful tuning of the DPIV
parameters. If adequate wind tunnel turbulence properties
are measured this way, they may yet be signicantly
Exp Fluids (2009) 46:527537 535
1 3
different when the test and/or diffuser section contains
mounting struts, camera cables, mirrors or other common
clutter, and should be measured again (Selig and McGra-
nahan (2004), for example, show turbulence levels and
spectra, together with spatial variation of the mean ow
with the tunnel empty and with measurement apparatus
installed).
Finally, we briey note an application of the long-
exposure DPIV method to detect a signicantly non-uni-
form background ow that would not otherwise be
measured. Otherwise uniform, low-turbulence background
conditions can be signicantly disrupted in practical animal
ight experiments, when various markers and platforms or
perches for regulation of ight are placed inside the wind
tunnel. In some instances, safety nets are installed not only
in the downstream diffuser section (a necessity!), but also
in the front of the test section. These are often claimed to
have little noticeable effect on ight performance, but now
that a sufciently sensitive measurement technique has
been developed, the effect on the ow can be seen directly.
Figure 10 shows the disturbance velocity eld caused by
the upstream net superimposed on colour-coded cross-
stream vorticity contours. There is a regular array of easily
detectable, horizontal, laminar wakes, separated by oppo-
site signed vorticity layers. The instantaneous wake
structures (Fig. 10a) themselves show some signs of semi-
regular variation in the streamwise direction. The time-
averaged vorticity eld (Fig. 10b) is taken from 20 inde-
pendent timesteps and shows exactly the same regular
variation in the vertical direction, but the horizontal vari-
ance is signicantly reduced. The background turbulence
cannot be seen on this scale because it is many times (about
30) smaller in magnitude. Small lens distortion effects can
be seen at the grid edges, particularly at the bottom of the
images as they could not be removed using the usual
averaging methods employed for the tunnel turbulence
estimates.
The net itself is composed of a 27 mm square mesh of
threads with thread diameter 15.1 0.4 lm. The Reynolds
number based on thread diameter, d, Re
d
= 7 for U = 7 m/
s. The mean vertical spacing between wakes in Fig. 10b is
27.1 0.03 mm, which is equal to the mesh spacing.
Measurements were taken approximately 1 m downstream
of the start of the test section, where the mesh was strung,
and so x/d & 66,000. The low Reynolds number and
comparatively low background turbulence levels lead to a
perfectly preserved wake image of the mesh, far down-
stream of its origin. Pennycuick et al. (1997) previously
documented an increase in global turbulence levels in the
test section due to an upstream (but different) net. Here we
see that a net actually imposes a rather stable, laminar
pattern on the ow, which may or may not be aerody-
namically important, but its presence ought to be noted,
particularly in view of the sensitivity of wing performance
to small disturbances.
5 Summary
A single example case is described to demonstrate that,
provided measurements are made with care and precision,
the low turbulence levels in high quality low-speed wind
tunnels can be estimated with DPIV methods. The uctu-
ation magnitudes are consistent with independent hot-wire
Fig. 10 Instantaneous (a) and time-averaged (b) spanwise vorticity,
x
y
(x, z), for empty test section with upstream net. The eld of view
{DX, DY} = 14.5 9 17 cm, dt = 2,500 ls and vector displacements
are shown 16 times their original length. Re
d
= 7, based on net thread
diameter
536 Exp Fluids (2009) 46:527537
1 3
estimates, and the spatial patterns of the disturbance
quantities in the measurement plane can also be found. The
window of usable operating parameters is constrained by
the requirement for large dts on one hand, and higher-
order corrections as disturbance quantities occupy different
locations in the lens distortion correction surface on the
other. Given perfect lighting and particle seeding, further
improvements will benet both from better optics and
further advances in DPIV algorithms. A DPIV system that
is correctly tuned to measure very small disturbance
quantities is a valuable indicator of actual operating con-
ditions, and the correct specications of these (including
the range of scales and frequencies over which global
statistics are computed) will be particularly important in
future experiments on the aerodynamics of birds, planes,
bats and gliders at moderate Reynolds number.
References
Adeyinka OB, Naterer GF (2007) Measured turbulent entropy
production with large eddy particle image velocimetry. Exp
Fluids 42:881891
Adrian RJ, Christensen KT, Liu Z-C (2000) Analysis and interpre-
tation of instantaneous turbulent velocity elds. Exp Fluids
29:275290
Agrawal A (2005) Measurement of spectrum with particle image
velocimetry. Exp Fluids 39:836840
Barlow JB, Rae WH, Pope A (1999) Low-speed wind tunnel testing.
Wiley, New York
Dryden HL, Kuethe AM (1929) Effect of turbulence in wind tunnel
measurements. NACA Report# 342
Dryden HL, Schubauer GB, Mack WC, Skramstad HK (1937)
Measurements of intensity and scale of wind tunnel turbulence
and their relation to the critical Reynolds number of spheres.
NACA Report# 581
Fincham AM, Spedding GR (1997) Low-cost, high resolution DPIV
for measurement of turbulent uid ow. Exp Fluids 23:449462
Fincham AM, Delerce G (2000) Advanced optimization of correlation
imaging velocimetry algorithms. Exp Fluids 29:S13S22
Hedenstrom A, Rosen M, Spedding GR (2006) Vortex wakes
generated by robins Erithacus rubecula during free ight in a
wind tunnel. J R Soc Interface 3:263276
Hedenstrom A, Spedding GR (2008) Beyond robins: aerodynamic
analyses of animal ight. J R Soc Interface 5:595601
Henningsson P, Spedding GR, Hedenstrom A (2008) Vortex wake
and ight kinematics of a swift in cruising ight in a wind
tunnel. J Exp Biol 211:717730
Hinds WC (1999) Aerosol technology: properties, behavior and
measurement of airborne particles. Wiley, New York
Lavoie P, Avallone G, De Gregorio F, Romano GP, Antonia RA
(2007) Spatial resolution of PIV for the measurement of
turbulence. Exp Fluids 43:3951
Lissaman PBS (1983) Low Reynolds number airfoils. Ann Rev Fluid
Mech 15:223239
Lyon CA, Broeren AP, Giguere P, Gopalarathnam A, Selig MS
(1998) Summary of low-speed airfoil data, vol 3. SoarTech, H.A.
Stokely, Virginia Beach
Muijres FT, Johansson LC, Bareld R, Wolf M, Spedding GR,
Hedenstrom A (2008) Leading-edge vortices increase lift in bat
ight. Science 319:12501253
Park KJ, Rosen M, Hedenstrom A (2001) Flight kinematics of the
barn swallow (Hirundo rustica) over a wide range of speeds in a
wind tunnel. J Exp Biol 204:27412750
Pennycuick CJ, Alerstam T, Hedenstrom A (1997) A new low-
turbulence wind tunnel for bird ight experiments at Lund
University, Sweden. J Exp Biol 200:14411449
Pennycuick CJ, Hedenstrom A, Rosen M (2000) Horizontal ight of a
swallow (Hirundo rustica) observed in a wind tunnel, with a new
method for directly measuring mechanical power. J Exp Biol
203:17551765
Piirto M, Eloranta H, Saarenrinne P, Karvinen R (2005) A compar-
ative study of ve different PIV interrogation algorithms. Exp
Fluids 39:571588
Prasad AK, Adrian RJ, Landreth CC, Offutt PW (1992) Effect of
resolution on the speed and accuracy of particle image veloc-
imetry interrogation. Exp Fluids 13:105116
Rosen M, Hedenstrom A (2001) Gliding ight in a jackdaw: a wind
tunnel study. J Exp Biol 204:11531166
Scarano F (2003) Theory of non-isotropic spatial resolution in PIV.
Exp Fluids 35:268277
Selig MS, McGranahan BD (2004) Wind tunnel aerodynamic tests of
six airfoils for use on small wind turbines. NREL Report SR-
500-34515
Spedding GR, Hedenstrom A, Rosen M (2003a) Quantitative studies
of the wakes of freely-ying birds in a low turbulence wind
tunnel. Exp Fluids 34:291303
Spedding GR, Rosen M, Hedenstrom A (2003b) A family of vortex
wakes generated by a thrush nightingale in free ight over its
entire range of ight speeds. J Exp Biol 206:23132344
Tanaka T, Eaton JK (2007) A correction method for measuring
turbulence kinetic energy dissipation rate by PIV. Exp Fluids
42:893902
Tennekes H, Lumley G (1972) A rst course in turbulence. MIT
Press, Cambridge
Westerweel J (1997) Fundamentals of digital particle image veloc-
imetry. Meas Sci Tech 8:13791392
Exp Fluids (2009) 46:527537 537
1 3