Professional Documents
Culture Documents
Foucaut2004 PDF
Foucaut2004 PDF
PIV optimization for the study of turbulent flow using spectral analysis
This content has been downloaded from IOPscience. Please scroll down to see the full text.
(http://iopscience.iop.org/0957-0233/15/6/003)
View the table of contents for this issue, or go to the journal homepage for more
Download details:
IP Address: 35.8.11.2
This content was downloaded on 14/06/2015 at 16:20
Abstract
Particle image velocimetry (PIV) is a measurement technique which is well
adapted to the study of the structure of turbulent flows, as it allows us to
obtain quantitative information on the spatial structure of the velocity field.
This contribution presents an experimental approach to characterize the
measurement noise of a PIV system and the spatial response of such a
method. This approach is based on a specific spectral analysis of the
velocity vector field deduced from several PIV experiments. This study was
done in two steps. The first step was to measure the noise level of PIV and
to determine a model for the PIV transfer function from a series of
displacement fields measured in a quiet liquid. This model shows the effect
of the interrogation window size and introduces a spectral noise density
which is constant for a given recording set-up. The second step was to
compute spectra from velocity fields obtained in a turbulent boundary layer
in a plane parallel to the wall. These spectra show that PIV behaves as a
band pass filter. This series of experiments allows us to build a model for the
prediction of the PIV spectrum knowing the real one. This model confirms
that the PIV noise is white. It allows us to optimize the interrogation
window size in order to obtain the best compromise between the spectral
response and the spatial resolution. The rms value of the noise can be
estimated from the noise density, allowing us to quantify the measurement
accuracy. The improvement of sub-pixel window shift is also discussed,
leading to a small decrease in the noise level. An analysis is proposed to
identify the main sources of noise: particles cut by the border of the
interrogation window, isolated particles, etc.
Keywords: PIV, spectrum, measurement noise, transfer function
x2, y2 particle location at the second exposure Without looking directly at the spectrum, Willert and
X, Y interrogation window (IW) size Gharib (1991) have discussed the low-pass filtering effect of
z0 laser sheet thickness PIV, due to the averaging over the interrogation windows (IW).
x, y spatial resolution They showed that the spatial cut-off wavelength is twice the
t PIV delay window size according to the Nyquist criterion. They deduced
laser wavelength these results from a spatial wavelength response which looks
u rms value of the noise in the x direction like an unsigned sinc function. Recently, Lourenco and
v rms value of the noise in the y direction Krothapalli (2000) introduced the averaging effect of PIV by
noise parameter t2 x2
x2 x1 t1 u dt x u dx
1 noise parameter in the x direction umeas = = 1 (1)
2 noise parameter in the y direction t t x
where x1 and x2 are the successive locations of the particles
1. Introduction being moved during the PIV delay t. The spatial increment
x should be the dynamic range of the particle motion.
The use of particle image velocimetry (PIV) is now Equation (1) is in agreement with the Willert and Gharib result,
widespread for the measurement of spatial phenomena in which explains that the curvature of the particle trajectory
flows (Stanislas et al 1998). In PIV, the largest scales of between two successive exposures is lost by PIV. The main
the flow which are accessible are limited by the field size. consequence is that the measured velocity is filtered. Lourenco
The smallest scales are also not captured due to the filtering and Krothapalli (2000) were the first to introduce a sinc
effect of the interrogation window. However, the use of function which is the Fourier transform of a gate function
PIV to understand turbulence is still a domain of research of length x, in order to take into account the filtering
which offers great possibility of progress. During the last effect due to the truncation embedded in the PIV analysis.
ten years, the PIV method has undergone a strong evolution. This sinc function is obtained by using the Fourier function
This revolution concerns pre-processing, accuracy, spatial u = exp(ikx) which allows us to deduce the transfer function
resolution and noise (Lourenco and Krothapalli 2000), as well T = uumeas
real
= sin(kx/2)
kx/2
. This transfer function corresponds to
as post-processing (Raffel et al 1998), with tools such as the low-pass filtering effect of PIV which is one of the main
double spatial correlations, pattern recognition and proper limitations of this technique. The cut-off wave number is then
orthogonal decomposition (Carlier et al 2000). But what is of the order of 1/x.
the spatial frequency measured by means of the PIV system? Guichard et al (1998) generated synthetic images from
What is the transfer function of PIV? What is the effect of DNS results. They compared spectra from DNS and PIV
the measurement noise on the turbulence spectra? These are analysis. They also evidenced the low-pass filtering effect due
several points which have not been deeply investigated and to the PIV method. But they could not evidence the evolution
will be addressed in the present contribution. of the cut-off frequency with the inverse of the IW size. The
In the study of turbulence, the flow characterization by obtained evolution was affected by the fact that the noise
means of a spectral analysis is a useful tool to determine the level increases when the window size decreases and modifies
energy contained at each turbulent scale. The standard method the cut-off frequency. Moreover, they observed that the PIV
to do so is to put a hot wire in the flow and to compute the spectrum is in poorer agreement with the DNS spectrum when
power spectrum from the recorded signal. This method gives the window size decreases. Finally they had some difficulty
the temporal spectrum and needs a Taylor hypothesis, based coming to a conclusion on the ability of PIV to measure a
on local isotropy of turbulence, to determine the spatial one spectrum with good accuracy. In the present contribution, a
(Hinze 1975). One advantage of PIV as compared to single- comparison of PIV and hot wire anemometry (HWA) spectra
point measurement techniques is to give access directly to the will be performed.
spatial spectrum. Westerweel et al (1993) were the first to As observed by Guichard et al (1998), the measurement
compare power spectra from PIV measurements in pipe flows noise due to the complete PIV chain is an important element
and from DNS results. They showed that the PIV spectrum to characterize the PIV spectral response. Allano et al (1998)
is in agreement with the DNS up to a cut-off wave number. explain that the noise origins can be classified into four
The value of this cut-off wave number depends on the signal- categories: physical (light sheet profile, particle optical and
to-noise ratio (SNR) which, for turbulent flows, is the ratio dynamical characteristics), specificity of the flow (velocity
between the rms values of fluctuations and the measurement gradient, out-of-plane motion, wall reflection, seeding
noise. If, due to the noise, the SNR decreases, the cut-off homogeneity, etc), instrumental (camera and digitizer),
wave number decreases. Later, Westerweel et al (1997), analysis method (peak fitting, algorithm). Using the
using experiments of grid-generated turbulence, showed the Europiv SIG algorithm (Lecordier and Westerweel 2004),
improvement brought by discrete image shifting on the power Foucaut et al (2004) tested part of these noise sources
spectra computed from digital PIV. They did a theoretical study by computing bias and rms errors in the case of sub-
of the noise reduction due to image shifting as a function of pixel displacement. They obtained a series of conclusions
turbulence intensity. They compared power spectra with and which will be used in this contribution. Willert and Gharib
without image shifting to validate the theoretical study. It was (1991) also studied the noise with a series of images
shown that the noise level in the Fourier space can be decreased obtained by recording a cloud of particles. From the rms
by an order of magnitude when the image shifting technique value of the displacement, they studied the effect of the
is used. seeding density (from 6 103 up to 7.3 102 particles
1047
J M Foucaut et al
per pixel) and of the displacement between both exposures response of a device should be to measure the response to an
(from 0.05 to 10 pixels). They proposed that the noise could impulse signal or to a white noise, but this kind of experiment
be essentially due to the particle images which are truncated would be very difficult to conceive with the complete PIV
by the edge of the interrogation window, to the particle density chain. A first step of the spectral characterization is thus
in the analysis area and to the unpaired particle images. Raffel to determine the response to a zero input signal. This is
et al (1998) give the evolution of the noise as a function of obtained by means of a motionless record. This should be
the particle image diameter, using a Monte Carlo simulation sensitive at least to the effect of the background, speckle and
of synthetic images. Like Willert and Gharib (1991) and camera noise, together with the effect of the shape difference
Westerweel et al (1997), they used a three-point Gaussian- between the two pulses. The noise due to particle motion will
peak fit. They show that an optimum diameter of the order be studied afterwards.
of 2.2 pixels can be found for cross-correlation and of 1.6
pixels for auto-correlation. When the window size or the 2.1. PIV experiment in a fluid at rest
number of pairs of particle images in the window decreases,
the noise increases. When the diameter of the images increases A motionless record is supposed to be obtained by means of
the peak-fitting algorithm also loses some accuracy and the a PIV experiment carried out in a glass aquarium filled with
noise increases. Another effect appears when the particle water at rest. The x-axis is horizontal and the y-axis vertical.
image diameter becomes too small. This effect tends to The PIV delay was chosen as small as possible in order
bias the displacement towards integer values. Following to minimize any particle displacement (eventual convection)
Raffel et al (1998), this peak-locking appears due to the between the exposures (25 s). The camera used was a Pulnix
peak-fitting which is not accurate enough. Peak-locking progressive scan TM9701. Its CCD has a size of 768
effects also appear depending on the sensor geometry, which 484 pixels2 of 11.6 13.6 m2 each. The video signal was
influences the spatial resolution. This source of peak-locking digitized at a rate of 30 im s1 by means of an ICPCI board
is described by Westerweel (1998) who specifies that the plugged into a PC computer. A 50 mm focal lens was used
particle images have to be larger than 2 pixels to reduce it. To with f # = 8. The magnification was 0.11. For seeding, the
minimize peak-locking error without affecting the accuracy, particles naturally contained in water were used. Their
a solution proposed by Guichard et al (1998) and Scarano diameter is about 1 m. The light sheet was generated by
and Reithmuller (2000) is to use an iterative sub-pixel shift, a BMI pulsed YAG laser with 2 250 mJ of energy at 15 Hz.
based on an interpolation of the particle images. Foucaut et al The sheet thickness in the field of view was about 2 mm.
(2004) tested these methods, evidenced the efficiency and Following Adrian (1991), the depth of the field was about
showed that the Whittaker interpolation proposed by Scarano 14 mm which avoids any problem to focus the particle images.
and Reithmuller (2000) presents the best accuracy. Basically, In this configuration, the Airy disc diameter is of the order
peak-locking is a systematic periodic bias, with a period of of 12 m which gives a size of the order of 1.5 pixels (Adrian
1 pixel. Its effect, which is to attract the displacements towards 1997). By looking at the particle images, the image size
integer pixel values, leads to important distortions in the appeared to be of the order of 2.53 pixels along x by 1.5
probability density function (PDF) but, when the displacement 2 pixels along y. These differences are attributed to the transfer
dynamic range is large enough, it has a negligible effect on the functions of the CCD. Following Westerweel (1997) the
rms value (Foucaut et al 2004). Finally, it should be mentioned maximum optical spatial frequency is W = [(M +1)f #]1 =
that Westerweel (1997) demonstrated that the bandwidth of 211 mm1 which implies, according to the Nyquist criterion,
digital PIV is limited due to the resolution of CCD cameras a resolution of at least 422 pixels per mm. This is much higher
(their maximum optical spatial frequency is indeed generally than the capability of the CCD which is 77 mm1 in length and
of the order of 100 mm1). This limitation increases the noise 90 mm1 in width. An estimation of particle images number
and, according to Lourenco and Krothapalli (2000), could be was made, which leads to an order of 0.02 ppp (particles
one source of peak-locking. per pixel). This value is a bit too small, the optimal value
In the first part of the present contribution, an experimental should be around 0.04 ppp (Willert and Gharib 1991). Raffel
approach based on spectral analysis of the vector field will et al (1998) discussed the influence of the background level
be presented to characterize the PIV noise response. A on the noise. The present experiment was done in a dark
characterization of the measurement noise will be presented. room which contributes to giving a background level close to
In the second part, a comparison of PIV and hot wire zero. The analysis was made by the classical cross-correlation
anemometry spectra will be done (Bruun 1995). A model method without shift (the displacement being close to zero)
taking into account the noise will be proposed to predict the with a three-point Gaussian peak fitting in each direction. The
PIV spectrum from the HWA one. In this part, two experiments displacement being zero, the results were weakly affected by
will be detailed. The first one corresponds to two single peak-locking, which tends to reduce the noise intensity.
exposure images using a video camera. The second one is By this method, the systematic and random errors can be
done with a digital photographic camera which records double estimated taking into account the full PIV chain. The mean
exposure images (auto-correlation). This tends to increase the values and standard deviation of displacements along x and y
noise level (Keane and Adrian 1992). were obtained. These results are given in table 1. Whatever
the analysis window size is, a systematic error of the order of
2. Spectral response of PIV 0.02 pixel can be observed in the displacement along x. This
value increases slightly with the window size. This systematic
In this paragraph, the PIV spectral response is studied error is probably due to the transfer function of the video
experimentally. A well-established way to obtain the spectral camera, the data transfer of which is made along x. This
1048
PIV optimization for the study of turbulent flow using spectral analysis
Table 1. Systematic and random errors computed from no motion PIV maps.
Window Number of Mean displacement Mean displacement Standard deviation Standard deviation
sizes recorded fields along x (pixel) along y (pixel) u (pixel) v (pixel)
16 16 125 0.016 0.000 0.087 0.050
32 32 125 0.023 0.000 0.055 0.029
64 64 125 0.026 0.000 0.033 0.015
0.1 0.01
0.01
0.001
E11 (pix3)
E 22 (pix )
3
0.001
Overlapping 50%
Overlapping 75% Overlapping 50%
0.0001
0.0001
Overlapping 87.5% Overlapping 75%
Equation (2) Overlapping 87.5%
Equation (2)
kmin
0.00001 kmin
0.001 0.01 0.1 1 0.00001
0.001 0.01 0.1 1
k (rad/pix) k (rad/pix)
Figure 1. Power spectra of displacements along x from no motion Figure 2. Power spectra of displacements along y from no motion
PIV maps. PIV maps.
10
phenomenon is probably specific to the PIV chain used. The
Cardinal sine function
standard deviation decreases when the window size increases slope - 2
according to the results of all the authors mentioned in the 1
introduction. Its values along x are systematically about two - 3db
times larger than the vertical ones. This is in agreement with
2
(sinc k)
k
From the 125 vector fields recorded, power spectra were
computed along x, varying the window sizes X (along x) and Figure 3. Characteristics of square cardinal sine function.
Y (along y), this for different overlapping values. For that
equation is: 1/k2. At k0 = 0.39 the function starts to depart
purpose, a Fourier transform was computed by FFT from each
from 1 (by 5%). For a decrease of 3 dB, one gets a value
vector line. The power spectrum is classically computed by
of kc = 1.4 which corresponds to the cut-off frequency of the
the product of the Fourier transform with its conjugate value
filter. In figures 1 and 2, the use of an overlap higher than 50%
divided by the length of the field. The spectra are averaged
does not give a better frequency response but is interesting to
along y according to an homogeneity hypothesis. The mean
evidence the sinc function. In the first paragraph, a cut-off
spectrum of each field is also averaged on the 125 realizations
wave number based on the mean particle displacement 1/x
to obtain a better convergence.
was mentioned. Here, due to the windowing effect, a cut-off
Figures 1 and 2 present the power spectra respectively
wave number of 1.4 (kc X = 2.8) is obtained. In fact, if the
along x and y in the case of a 32 32 interrogation window.
displacement is smaller than the window size (generally four
The spectra are plotted as a function of k, where k is the wave
to three times smaller) the windowing cut-off wave number is
number taken along x. The analysis could be performed along y
the relevant one.
without change. The results are given for three different values
If a PIV record is considered as a continuous signal, the
of the overlap (50, 75 and 87.5%). As can be seen, the obtained
analysis using a sampling window can thus be considered as an
spectra correspond to a white noise multiplied by a squared
averaging procedure over this window. This leads, by Fourier
sinc function: (sinc(kX/2))2 . The sinc function is indeed the
transform, to a multiplication by a sinc function. According
Fourier transform of a gate function which corresponds to
to Lourenco and Krothapalli (2000), the power spectra of the
the interrogation window. As figures 1 and 2 show power
PIV noise can then be expressed as
spectra, the sinc function is squared. Figure 3 gives some
characteristics of the square sinc function: sinc2 (k). This sin(kX/2) 2
Eii = Enoise (2)
function presents an envelope (given by the thin line) whose kX/2
1049
J M Foucaut et al
0.01 1
0.1
Y E 22 (k) (pix )
0.001
4
E 22 (k) (pix )
3
0.01
64x32 32x64
32x32 32x32
0.0001 32x16
16x32
Equation (2) 0.001 Equation (5)
kmin X/2
0.00001 0.0001
0.01 0.1 1 k c X/2 10 0.1 1 10
k X/2 (rad) k X/2 (rad)
Figure 4. Power spectra of displacements along x, influence of X. Figure 5. Power spectra of displacements along x, influence of Y.
E22 versus kX/2. Y E22 versus kX/2.
where Enoise is the white noise level obtained by fitting the Enoise is of the same order as earlier (17.6 103 pixel3). The
spectrum with equation (2). This model is also plotted in abscissa representation seems universal when X varies, which
figures 1 and 2 with Enoise = 17.5 103 pixel3 and 4.4 leads to a cut-off wave number given by
103 pixel3 respectively for E11 and E22. The noise level is
kc X = 2.8. (3)
thus about four times higher for E11 than for E22 in agreement
with the values of the standard deviation in table 1. As can This result is in agreement with the fact that the standard
be seen, there is a small gap at high frequency between the deviation u of the displacement decreases when the window
spectra and the model around the point where the sinc function size increases. u is indeed the square root of the
goes to zero. However, if this function is disconvoluted from integral of the power spectra: u2 = E11 (k) dk =
the measured spectra, the signal obtained is very close to a 2 0 E11 (k) dk. In the present case, equation (2) leads to
white noise. 2 2
u2 = 2 0 Enoise sin(u)
u
du. If an overlap of 75% is used, as
At the lowest wave numbers, E11 presents some variations the spectrum shows a second lobe, the integration is performed
which lead to a gap with the white noise (this is not the case 2 2
on [0, 2]. From I = 0 sin(u) du = 1.492, u can be
for E22). This gap is probably also due to the transfer function u
directly obtained by
of the video camera. This brings a disturbance in the spectra
which can be linked to the systematic error observed in table 1. 4Enoise I
In the following, the results presented correspond to E22 and u = (4)
X
are generalized to Eii .
u is thus proportional to the inverse of the square root of the
Theoretically, the lower significant frequency resolved
window length. With 50% of overlap, the integral term has to
is the inverse of the field length which, in this case, gives
be limited to and I = 1.418.
kmin = 8.5 103 rad/pixel. This value is shown in figure 1
The effect of the window size along y is presented in
by a vertical dashed line. The number of vectors varies from
figure 5. The spectra are calculated with a window size X of
48 to 192 depending on the overlap used.
32 pixels and three different values of Y: 16, 32 and 64 pixels.
The PIV measurement transfer function is thus
In this case the convergence of the spectra can be influenced
comparable to a low-pass filter which is characterized by a
by the variable number of vector lines in the y-direction. The
square sinc function due to the window effect. It presents a
adequate number of fields used should minimize this problem.
cut-off wave number of 2.8/X, X being the physical window
When Y varies, the spectra are still in agreement with the model
size, whatever the overlap is.
of equation (2) but the noise level Enoise increases when the
From figures 1 and 2, it is clear that an overlap larger
window size Y increases. The Enoise values are 35, 17.5 and
than 50% is of no interest as far as the spectral resolution is
8.8 103 pixel3 respectively for Y = 16, 32 and 64 pixels.
concerned. But an overlap of 75% will be used hereafter in
It can be remarked that Enoise varies with the inverse of Y.
order to detect easily the sinc function. Figure 4 presents the
A new representation can then be proposed of Y E 22 (kX/2)
influence of the window size X which varies from 16 to 64
versus kX/2. Figure 5 shows the spectra in this representation.
pixels. Y is kept at 32 pixels in order to have no effect on
A model deduced from equation (2) is also plotted, it is
the convergence of the spectra. Only linked to the field size
expressed as
along x, the value of kmin is the same as in figure 1. The effect
of the interrogation window size is only to change the length sin(kX/2) 2
of the sinc function. In figure 4, the results are plotted in a Y E 22 = . (5)
kX/2
representation where k has been scaled by the half window
As can be seen, the spectra for different Y values superimpose
size X/2. Strictly speaking, the fact of multiplying k by X/2
and the model of equation (5) is in good agreement with them.
corresponds to a change of variable which leads to E22 (k) =
The constant which has been introduced is given by
E22 (kX/2) X/2. In this figure, the spectra follow the model
of equation (2) whatever the window size is. The noise level = Enoise Y. (6)
1050
PIV optimization for the study of turbulent flow using spectral analysis
1 1
0.1 0.1
Y E 22 (k) (pix )
Y E 22 (k) (pix )
4
4
0.01 0.01
64x64 Classical
32x32 Padding
16x16 0.001 Equation (5)
0.001
Equation (5)
0.0001 0.0001
0.01 0.1 1 1.4 10 0.1 1 10
k X/2 (rad) k X/2 (rad)
Figure 6. Power spectra of displacements along x, influence of the Figure 7. Power spectra of displacements along x, influence of zero
window size. padding.
1051
J M Foucaut et al
1 1.E+00
1.E-01
0.1
Y E 22 (k) (pix )
4
1.E-02
No periodicity
E 11
0.01 Sinus
FFT+ Interger shift 1.E-03 Squared sinus
Equation (8)
Sym. sinus
Equation (9)
0.001 1.E-04 Straigth line
FFT + Whittaker
Sym. straight line
Exact solution
1.E-05
0.0001 1 10 100
0.1 1 10 k
k X/2 (rad) (a)
E 11 (m /s )
Sym. straight line
based algorithm. This comparison is done with 32 32 2
3
1.E-05 Slope -5/3
window size. The shape of each spectrum is the same but
the FFT-based correlation gives the best result as far as the 1.E-06
noise level is concerned. The values of Enoise are 0.145, 0.30
and 0.57 respectively for FFT-based, normalized and non- 1.E-07
normalized direct correlation. This phenomenon is explained kmin
by Foucaut et al (2004). It is due to the particle images cut by 1.E-08
10 100 1000 10000
the border of the first interrogation window. These images are k (rad/m)
not truncated in the same way by the second one which moves (b)
when direct correlation is used. A solution is to compute the
direct correlation on limited windows in order to avoid this Figure 9. Comparison of different treatments performed on a sin
phenomenon (Foucaut et al 2004). This method then gives wave (a) and on PIV data (b) to compute spectrum.
the same spectrum as FFT-based cross-correlation.
It is now well known that sub-pixel shift reduces the samples was 1260 000. The Kolmogoroff frequency being
peak-locking effect and then improves the accuracy. Foucaut of the order of 1 kHz, a filter was used to avoid the folding effect
et al (2004) showed that the Whittaker interpolation used with a cut-off frequency of 2 kHz. The hot wire length was
by Scarano and Reithmuller (2000) is well adapted to the 0.5 mm, corresponding to four wall units which is comparable
PIV processing. They explained that this technique is better to the Kolmogoroff scale. The spectra were obtained by
because it respects the sampling theorem. Figure 8 also shows computing the Fourier transform of the HWA signal with the
a spectrum obtained from an analysis with sub-pixel shift same FFT algorithm as for PIV (average of about 300 spectra
using the Whittaker interpolator (see Scarano and Reithmuller of 4096 samples). A Taylor hypothesis was used to convert
(2000)). This method tends to avoid the peak-locking effect. from time to space (Hinze 1975).
As seen in this figure, the noise level increases slightly to about For PIV, the measurements were performed in a small
4%. This result confirms that the peak-locking effect has no region of a turbulent flow. The signal is thus not periodic.
strong effect in the spectral domain (Foucaut et al 2004). This non-periodicity leads to discontinuities when the FFT is
used to compute the spectrum. If the signal is comparable in
size with one of the largest structures of the turbulent flow, the
3. Comparison of PIV and HWA power spectra
discontinuity can increase the noise enough to the order of the
3.1. PIV experiment with a small dynamic range signal. In this case, a treatment to make the sample periodic
is necessary. Figure 9(a) shows the spectrum computed
In order to better characterize the measurement noise and the from a signal corresponding to a small part of a sin wave
PIV response to a turbulent signal, an experiment has been (0 < x < /2) f (x) = sin(x) + 13 1
sin(x/13). The exact
carried out in a boundary layer wind tunnel (Carlier 2001). solution gives two Dirac functions for k = 1 and k = 13.
This experiment allows us to compare hot wire anemometry If no treatment of periodicity is done, the noise due to the
spectra with PIV spectra. The boundary layer is turbulent. The discontinuity is of the same order as the signal. Different
Reynolds number based on momentum thickness is 7800. The treatments were tested: multiplying the signal by a sine
boundary layer thickness is about 0.3 m and the free stream function, subtracting the straight line which passes through
velocity is 3 m s1. The HWA measurements were obtained the first and the last samples of the signal. These two methods
with a classical method by recording the instantaneous velocity improve the computation but are not enough, as the derivative
during a long period (300 s) with a sampling frequency of of the signal has also to be periodic. Possible solutions are:
2200 Hz (Shannon theorem was fulfilled). The number of to multiply the signal by a squared sine function or to use a
1052
PIV optimization for the study of turbulent flow using spectral analysis
U 0.4
Integer shift
0.35 Whittaker
u(y)
0.3 HWA
PDF(u)
Seeding 0.25
0.2
0.1
45 mirror
0.05
Spherical Camera
0
lens
-4 -3 -2 -1 0 1 2 3 4
Cylindrical
lens u/u'
Pulsed laser
1.E-03
Figure 10. PIV experimental set-up.
1.E-04
E 11 (m /s )
symmetry around the last point of the sample when the sine
2
3
1053
J M Foucaut et al
1.E-02 1.E-02
1.E-03 1.E-03
1.E-04 1.E-04
E 11 (m /s )
E 11 (m /s )
2
2
3
3
1.E-05 1.E-05
HWA spectrum
1.E-06 1.E-06 HWA spectrum
PIV spectrum
PIV spectrum
Equation (10)
1.E-07 1.E-07 Equation (10)
Slope -2
kmin kc kmin kmax kc
1.E-08 1.E-08
1 10 100 1000 10000 1 10 100 1000 10000
k (rad/m) k (rad/m)
Figure 13. Power spectra of velocity along x, 64 64 interrogation Figure 14. Power spectra of velocity along x, 32 32 interrogation
window. window, high magnification.
1.E-02
up to about kX = 3.8 which corresponds to an attenuation
by the squared sinc function of 6 dB. Figure 13 presents the 1.E-03
comparison between HWA and PIV spectra with 64 64
interrogation windows. In this case, the PIV spectrum is 1.E-04
E 11 (m /s )
2
1.E-05
kc = 430 rad m1. Looking at figure 13, it is clear that the HWA spectrum
filtering effect due to this interrogation window size is too 1.E-06 PIV spectrum
strong. The model given by equation (10) is also presented in Equation (10)
this figure. Enoise is computed from the previous value using 1.E-07 slope -2
= Enoise X which is considered as a constant. kmin kmax kc
For the present experiment, the value of is 1.18 1.E-08
800
1 10 100 1000 10000
108 m4 s2. In pixel units, a value of = 2.432 pixel4 is
k (rad/m)
obtained which is higher than in the experiment of section 2.
The difference will be discussed further downstream. As Figure 15. Power spectra of velocity along x, 64 64 interrogation
for the 32 32 pixel2 window size analysis, the model window, high magnification.
is in agreement with the PIV spectrum approximately up
to kX = 3.8. Around and above this value, a significant number of particles was down to 0.025 ppp. The PIV delay
difference exists, as in the experiment of section 2. The was 100 s.
PIV spectrum does not go to zero. Here it tends towards the Figures 14 and 15 present the same results as figures 12
2 envelop of the sinc function. Thus, the proposed model and 13 but with a magnification of 0.32. The comparison
supports the idea that the difference between PIV and HWA of figures 12 and 14 clearly shows the objective: for the
is due to the combined effect of PIV noise and the use of an same interrogation window size in pixels, the sinc function
interrogation window for the PIV analysis. The same result is extended towards the high frequencies (it goes to zero now
as in section 2 can be deduced. The PIV method behaves as a at 5733 rad m1 instead of 1916 rad m1 and both kmin and kc
band pass filter. The cut-off wave numbers kc can be deduced are tripled taking the values 240 rad m1 and 2550 rad m1
respectively.
from equation (2), X being the interrogation window size in
As can be seen from figure 14, the PIV spectrum is in
physical units. This cut-off, indicated by the long dashed line
agreement with the hot wire over a small range. In the covered
in figures 12 and 13, decreases when the interrogation window
bandwidth (kmin < k < kc), the model of equation (10) follows
size increases. As observed in both figures, the wave number
this spectrum quite well with = 5.8 pixel4. This value of is
kmax increases first when the window size increases up to
about two times higher than the value for low magnification.
kmax = kc. This is due to the fact that, for a given PIV
However, given the magnification, the value of in physical
recording set-up, is constant and thus Enoise increases when
units is 7.86 1010 m4 s2 which is about 15 times
X decreases. Above this limit, the noise plays a less important
smaller. Consequently, Enoise is about five times smaller
role and kmax = kc.
(7.2 107 m3 s2 for a 32 32 window). At the wave number
kmin; the signal amplitude is about 5.7 times smaller which is
3.1.2. High magnification. A solution often suggested to comparable. But, at the cut-off wave number (analysis with
resolve the high frequency part of the spectrum with PIV is to 64 64 IW), the signal amplitude decreases down to a value
increase the optical magnification. Thus, the same experiment about 20 times smaller. In the Fourier space, the signal-to-
was repeated in the same installation but with a magnification noise ratio (signal amplitude over Enoise computable for each
three times higher than the previous one. The f # number frequency) is thus much poorer in the second experiment. In
was 16. The maximum optical spatial frequency is then figure 15 the PIV spectrum is in agreement with the hot wire
89 mm1 which is very close to the capability of the CCD up to kc = kmax. This window size seems to be close to the
sensor. The particle image size was nearly the same. The limit under which the noise increases and kmax decreases.
1054
PIV optimization for the study of turbulent flow using spectral analysis
1.E-02 1.E-02
1.E-03 1.E-03
1.E-04 1.E-04
E 11 (m /s )
E 11 (m /s )
2
2
PIV spectrum
3
3
kc kmin kc
1.E-08 1.E-08
1 10 100 1000 10000 1 10 100 1000 10000
k (rad/m) k (rad/m)
Figure 16. Interrogation window optimization method from HWA Figure 18. Interrogation window optimization method from PIV
spectrum. spectrum analysed with 24 24 pixel2 window size.
1.E-02
If no a priori spectral information is available, an
1.E-03 alternative is proposed in figure 18. An analysis with a
small window size (24 24 in this figure) has to be used,
1.E-04 which shows a high noise level. From the PIV spectrum, the
E 11 (m /s )
3 2
1055
J M Foucaut et al
E 11 (m /s )
pulses (difference beam shape and divergence). From a study
2
3
on synthetic images, Foucaut et al (2004) show that the order 1.E-05
of out-of-plane velocity in a turbulent flow can beestimated HWA spectrum
PIV spectrum
using the out-of-plane turbulent intensity as W = 2 w 2 Mt 1.E-06
Equation (10)
where M is the magnification in pixel m1. According to Keane
1.E-07 slope -2
and Adrian (1990) the value of W should verify W/z0 < 0.2
kmin kmax kc
in order to limit the noise where z0 is the laser sheet thickness. 1.E-08
800
This ratio is respectively of 0.08 and 0.05 for the low and 1 10 100 1000 10000
1056
PIV optimization for the study of turbulent flow using spectral analysis
1.E-02 4. Conclusion
1.E-03 In this contribution, the PIV frequency response has been
studied by an experimental approach. A PIV experiment in
1.E-04
quiet water was carried out in order to measure the part of the
E 11 (m /s )
3 2
1.E-05
noise which is not due to the particle motion. This experiment
has allowed us to obtain a universal representation and a model
1.E-06 HWA spectrum of the noise spectrum (equation (5)). This model shows that
PIV spectrum the noise is white and that the PIV transfer function is a
Equation (10)
1.E-07 squared cardinal sine function. This function presents a cut-
kmin kc off frequency kc = 2.8/X which depends only on the window
1.E-08
1 10 100 1000 10000
size in the direction where the spectrum is computed. The
k (rad/m) noise level in one dimension of the Fourier domain depends
only on the window size in the other direction (equation (6)).
Figure 20. Power spectra of velocity along x, optimized 44 44 A spectral noise density coefficient can be introduced which
interrogation window.
is only a function of the PIV recording configuration. This
density coefficient allows us to compute the standard deviation
a large range of wave number: from 80 to 1360 rad/pixel. of the spectral noise knowing the IW size. This standard
The Kodak camera allows us to cover almost the same range deviation is in good agreement with the experimental result
(401200 rad/pixel), but only one experiment was necessary and the noise behaviour confirms the idea of Willert and Gharib
due to the large size of the CCD sensor. The noise level (1991) that most of it is due to the particle images cut by the
u is of the order of 0.05 pixel for the noise experiment, of IW border.
0.1 pixel in the case of the low magnification experiment In a turbulent flow, the experiments performed show that
(Pulnix camera) and of 0.2 pixel in the case of the large the PIV transfer function is conserved and that the turbulent
dynamics experiment (Kodak camera). The difference particle motion does not change deeply the nature of the
between the first two experiments has already been discussed. noise. If the turbulence spectrum is known (in the present
The results with the Kodak camera are due to the double case deduced from hot wire measurement), a model can be
exposure mode. The fact of using a large dynamic range proposed (equation (10)), which is coherent with the model
(about 14 pixels) can increase the out-of-plane component for zero displacement and which allows us to deduce the noise
effect. The estimation of W/z0 is 0.2 which is smaller than the level and thus the density from a PIV spectrum computation.
acceptable limit (0.25) given by Keane and Adrian (1990). The This model shows again that the noise is white and that its level
wall distance being about three times higher, the error due to depends strongly on the recording set-up. The results show
the gradient is negligible. The accuracy of the last experiment a significant increase of when the particles are moving. A
(1.32%) seems to be the best due to the large dynamic range careful analysis of the possible supplementary noise sources
which was used here. shows that the peak locking is negligible if the dynamic range
The following step of the present study would be is larger than three pixels. A significant source of noise appears
to perform the PIV analysis using an image deformation to be the small scale motion of particles inside the interrogation
algorithm (Lecordier and Trinite 2003). Lecordier and Trinite window. Sub-pixel shifting brings a slight improvement if the
did compare spectra computed from synthetic images built dynamic range of turbulence is small.
from DNS results with and without the deformation algorithm. Finally, the model proposed has allowed us to determine
The improvement concerning the noise level seems significant. an optimal interrogation window size which is the best
They evidence also the strong effect of the out-of-plane compromise between spectral noise and spatial resolution.
component on the noise level in the spectral space. This optimization brings a noticeable improvement to the
1057
J M Foucaut et al
spectrum when compared to the HWA. One important Carlier J 2001 Etude des structures coherentes dune couche limite
conclusion of this study is that, for the assessment of turbulente a grand nombre de Reynolds PhD Thesis Univ. de
Lille, France
turbulence, PIV is limited by the size of the CCD on the
Carlier J, Foucaut J M and Stanislas M 2000 Experimental Study of
low wave number side and by the noise level on the high wave Near Wall Turbulence Using PIV (Rouen: EUROMECH 411)
number side. Thus if it seems useful to reduce the field of view Foucaut J M, Milliat B, Perenne N and Stanislas M 2004
to try to resolve the small scales, it is better first to reduce the Characterization of different PIV algorithms using the
noise. Part of it can be minimized by a careful adjustment of the EUROPIV Synthetic Image Generator and real images from a
turbulent boundary layer Proc. EUROPIV 2 Workshop on
recording parameters, but some gain can also be expected from
Particle Image Velocimetry (Berlin: Springer) pp 16386
the processing algorithms. From a practical point of view, one at press
should emphasize the interest of equation (7) which means that Gui L and Merzkirch W 1998 Generating arbitrarily sized
in a turbulent flow, if one can deduce from the comparison interrogation windows for correlation-based analysis of particle
between the PIV spectrum and the real one (measured by HWA image velocimetry recording Exp. Fluids 24 669
Guichard L, Lecordier B and Reveillon J 1998 Evaluation des
or modelled), one has access to the standard deviation of the
algorithmes utilises en PIV grace a la simulation numerique
measurement noise. This allows us to estimate the accuracy of directe 6 Congres Francophone de Velocimetrie Laser
the measurement. If no spectrum is available, a method based (Saint-Louis) pp F5.18
on an analysis using small windows is proposed. This method Hinze J O 1975 Turbulence (New York: McGraw-Hill) pp 190
allows us to estimate the value of at the cut-off wave number Keane R D and Adrian R J 1990 Optimisation of particle image
velocimetersPart I: double pulsed systems Meas. Sci.
when the signal is negligible as compared to the noise level.
Technol. 1 120215
Finally, it should be mentioned that the present Keane R D and Adrian R J 1992 Theory of cross-correlation
experimental contribution is of course partial and that it analysis of PIV images Appl. Sci. Res. 49 191215
would be worth developing it by a detailed analysis, based Lecordier B and Trinite M 2003 Advanced PIV algorithms with
on synthetic images of turbulent flow, of the contribution of image distortionvalidation and comparison from synthetic
images of turbulent flow Proc. 5th Int. Symp. on PIV (Busan,
the various noise sources. Such a study was already started
Korea) p 3250
by Foucaut et al (2004) in the case of series of uniform Lecordier B and Westerweel J 2004 The EUROPIV Synthetic Image
displacement. Generator Proc. EUROPIV 2 Workshop on Particle Image
Another development would be to look at the effect of the Velocimetry (Berlin: Springer) pp 14562 at press
advanced PIV algorithms which take into account the velocity Lourenco L and Krothapalli A 2000 True resolution PIV, a mesh
free second order accurate algorithm Proc. 10th Int. Symp.
gradients inside the interrogation window.
Appl. Laser Tech. Fluid Mech. (Lisbon, Portugal) p 13.5
Nogueira J, Lecuona A and Rodriguez P A 2001 New source of
Acknowledgments peak locking related to the window size: analysis and its
removal 4th Int. Workshop on PIV (Gottingen) paper no 1013
This work has been performed under the EUROPIV2 project. Raffel M, Willert C and Kompenhans J 1998 Particle Image
Velocimetry, A Practical Guide (Berlin: Springer)
EUROPIV2 (a joint program to improve PIV performance for Roth G I and Katz J 2001 Five techniques for increasing the speed
industry and research) is a collaboration between LML URA and accuracy of PIV interrogation Meas. Sci. Technol. 12
CNRS 8107, Dassault Aviation, DASA, ITAP, CIRA, DLR, 23845
ISL, NLR, ONERA and the universities of Delft, Madrid, Scarano F and Reithmuller M L 2000 Advances in iterative
Oldenburg, Rome, Rouen (CORIA URA CNRS 230), St multigride PIV image processing Exp. Fluids 29 s5160
Stanislas M, Carlier J, Foucaut J M and Dupont P 1998
Etienne (TSI URA CNRS 842), Zaragoza. The project Experimental study of a high Reynolds number turbulent
is managed by LML URA CNRS 8107 and is funded by boundary layer using DPIV Proc. 8th Int. Symp. Appl. Laser
the European Union within the 5th framework (contract no. Tech. Fluid Mech. (Lisbon, Portugal) pp 11.2.18
G4RD-CT-2000-00190). Westerweel J 1997 Fundamentals of digital particle image
velocimetry Meas. Sci. Technol. 8 137992
Westerweel J 1998 Effect of sensor geometry on the performance of
References PIV interrogation Proc. 8th Int. Symp. Appl. Laser Tech. Fluid
Mech. (Lisbon, Portugal) pp 1.2.18
Adrian R J 1991 Particle-imaging techniques for experimental fluid Westerweel J, Adrian R J, Eggels J G M and Nieuwstadt F T M
mechanics Ann. Rev. Fluid Mech. 23 261303 1993 Measurement with particle image velocimetry of fully
Adrian R J 1997 Dynamic ranges of velocity and spatial resolution developed turbulent pipe flow at low Reynolds number Proc.
of particle image velocimetry Meas. Sci. Technol. 8 13938 Int. Symp. Appl. Laser Tech. Fluid Mech. (Lisbon, Portugal)
Allano D, Lecordier B, Lecordier J C and Trinite M 1998 no 1.2 pp 285307
Qualification dune chane de mesure piv digitale source de Westerweel J, Daribi D and Gharib M 1997 The effect of discrete
bruit et estimation des erreurs 6 Congres Francophone de window offset on the accuracy of cross-correlation analysis of
Velocimetrie Laser (Saint-Louis) pp F6.18 digital PIV recording Exp. Fluids 23 208
Bruun H H 1995 Hot-wire Anemometry (Oxford: Oxford University Willert C E and Gharib M 1991 Digital particle image velocimetry
Press) Exp. Fluids 10 18193
1058